38.

PEIXOTO’S THEOREM 385

38 Peixoto’s theorem

We consider compact 2-differentiable mfd 𝑀 . We introduce a 𝐶 1 -topology in 𝒳 (𝑀 ).

As you read from the Wikipedia article of Peixoto413 Peixoto’s theorem was a spring-
board for the study of dynamical systems, Smale was interested in Peixoto’s work,
and defined the Morse-Smale system. Then to cover more generic dynamics, he in-
troduced the horseshoe dynamical system and then the Axiom A system.
In this section, we taste the original proof of Peixoto’s theorem.

38.1 𝜀-homeomorphism
A homeomorphism that does not move points more than 𝜀 is called an 𝜀-homeomorphism.

413
Usually, M M Peixoto, “Structural stability on two-dimensional manifolds” Topology, 1 101
(1962) is cited, but a large chunk of the proof is in M C Peixoto and M M Peixoto, “Structural
stability in the plane with enlarged boundary conditions” Ann Acad Bras Sci 31 135 (1959) [MCP
(1921-1960, the first Brazilian woman to receive a doctorate in mathematics) is his wife]. Read
Wikipedia M M Peixoto (1920-):
“Once, while talking with his mentor, Solomon Lefschetz, Mauricio Peixoto commented that no one
cared about structural stability of dynamical systems and that was the main problem in working
with it. But to Peixoto’s surprise Lefschetz’s answer was no less than “No Mauricio, this is no
trouble, this is your luck. Try to work as hard and as fast as you can on this subject because
the day will come when you will not understand a single word of what they will be saying about
structural stability; this happened to me in topology.” Lefschetz’s support was very important to
Peixoto at the time. In 1957, Peixoto went to research the subject with Lefschetz at the Princeton
University, where he spent uncountable hours talking to the Russian professor about Mathematics
and other subjects. Despite of the great age difference (Peixoto was 36 years old and Lefschetz 73),
they became good friends.
With Lefschetz incentive, Peixoto wrote his first paper on structural stability, that would be
later published on the Annals of Mathematics, of which Lefschetz was editor. In 1958, they went
to the International Mathematical Congress, in Edinburgh, Scotland, where Lefschetz introduced
Peixoto to the Russian mathematician Lev Pontryagin, whose work on dynamical systems was used
by Peixoto as a basis for his studies. Pontryagin, though, showed no interest whatsoever in Peixoto’s
work.
Back to Princeton, Peixoto met Steve Smale, the mathematician that would later become a
reference in dynamical systems. Smale was interested in Peixoto’s work and realized he could
extend his own based on it. Their contact intensified and, when Peixoto came back to Brazil,
the American mathematician spent six months at the Instituto Nacional de Matemática Pura e
Aplicada (Institute of Pure and Applied Mathematics or IMPA) at Rio de Janeiro. Through Smale,
Peixoto would meet the French mathematician René Thom, who would help Peixoto to formulate
his theorem, that was finalized during Thom’s visit to IMPA.”
386

38.2 Structural stable vector field

𝑋 ∈ 𝒳 (𝑀 ) is structurally stable, if there is a nbh 𝑈 of 𝑋 such that any 𝑇 ∈ 𝑈 is
homeomorphic to 𝑋.
The original definition by Pontryagin and Andronov required 𝜀-homeomorphy in-
stead of the simple homeomorphy. Peixoto demonstrated that homeomorphy implies
𝜀-homeomorphy at least for 𝒳 (𝑀 ) with 𝑀 being 2-mfd.
The ‘smallness of perturbation is usually in 𝐶 1 -topology, since 𝐶 0 -small pertur-
bations are “too strong and may destroy any singularity or closed orbit.”

38.3 Peixoto’s theorem

The Topology paper contains two theorems:
THEOREM 1. In order that the vector field 𝑋 (at least 𝐶 1 ) be structurally stable
on 𝑀 it is necessary and sufficient that the following conditions be satisfied:
(1) there is only a finite number of singularities, all hyperbolic;
(2) the 𝛼 and 𝜔-limit sets of every trajectory can only be singularities or closed or-
bits;
(3) no trajectory connects saddle points;
(4) there is only a finite number of closed orbits, all hyperbolic.
THEOREM 2. The set Σ of all structurally stable systems is open and dense in
𝒳 (𝑀 ).
Here, the openness can be shown in any dimension, so the 2D specialty is the
denseness.

38.4 How about higher dimensions?

Does something like 38.3 hold for 𝑑 ≥ 3? The Morse-Smale system 38.5 was intro-
duced to consider this problem.
Theorem414 Morse-Smale systems are dense in Diff(𝑀 ) for any 𝑑.
Theorem415 For Diff(𝑀 ) for compact 𝑀 the Morse-Smale systems is structurally
stable.
However, there are structurally stable non-MS systems, the converse is not true.
Incidentally,
Theorem416 In Diff𝑟 (𝑀 ) structurally stable systems are 𝐶 0 -dense.

414
Palis, Topology 8 385 (1969).
415
Palis-Smale
416
Shub
38. PEIXOTO’S THEOREM 387

38.5 Morse-Smale systems417

A dynamical system is called a Morse-Smale system, if
(MS1) Ω is finite (so Ω consists of periodic orbits).418
(MS2) Periodic points are all hyperbolic.
(MS3) If 𝑝, 𝑞 ∈ Ω, then 𝑊 𝑠 (𝑝) ⊤
∩ 𝑊 𝑢 (𝑞).419
Gradient dynamical systems are Morse-Smale.420
With this terminology
Peixoto’s theorem A compact 2-mfd 𝐶 1 -vector field is structural stable iff it is
Morse-Smale.

38.6 Outline of the proof of 38.3

⇐
In this proof actually an 𝜀-homeomorphism is constructed to relate perturbed sys-
tems. Thus, Theorem 1 implies that structural stability implies 𝜀-structural stability.
To show vector fields allowing the flow satisfying (1)-(4) are structurally stable,
we must construct homeomorphisms between the original and the perturbed fields.
This is rather tedious and is shown in the AABS paper quoted above.
⇒
Starting from any vector field 𝑋, with a series of approximation lemmas it is shown
that 𝑋 is approximated by 𝑌 (i.e., there is a vector field 𝑌 in any nbh of 𝑋) that
satisfies (1)-(4).
Thus, if 𝑋 is structurally stable, then it must be homeomorphic to its perturbed
version, but a perturbed version must satisfy (1)-(4), so 𝑋 itself must have satisfied
(1)-(4), because these properties are homeomorphism invariant.
Theorem 2 is shown almost simultaneously: Structural stable vector fields make
an open set (in any dimension, actually), but the proof of ⇒ implies such vector fields
are dense (exists in any neighborhood of any vector field on a compact 2-manifold).

417
M Shub: http://www.scholarpedia.org/article/Morse-Smale_systems.
418
Thus, we may say ∪𝑗 𝑊 𝑠 (𝑃𝑗 ) = 𝑀 , ∪𝑗 𝑊 𝑢 (𝑃𝑗 ) = 𝑀 .
419
Because of (MS1) transversality in this case means: no saddle connection nor homoclinic
tangency. That is, 𝑊 𝑠 (𝑝) ⊤
∩ 𝑊 𝑢 (𝑞) means they are not in contact except at separatrices.
420
See also K. R. Meyer Energy Functions for Morse Smale Systems, Am J Math 90 1031 (1968).
388

38.7 Outline of the proof of ⇐421

To explain this several definitions must be introduced:
Definition. A region on which the vector field is homeomorphic to one of the fol-
lowing (i)-(iii) in Fig. 38.1 is called ‘parallel region.’ Note that parallel regions are
arcwisely connected.

Figure 38.1: Parallel regions

Definition. A trajectory is called an ordinary trajectory, if it is in (i) the strip
region. Other trajectories are called separatrices; limit cycles, homo and heteroclinic
orbits, etc., are the examples.
Definition. A region 𝑀 ∖ {separatrices} is called the canonical region. Perhaps it
is better to define separatrix set as the complement of the totality of parallel regions
in 𝑀 .
(i) Canonical regions are classified into five types (I)-(V).
(ii) On each canonical region, flows are structurally stable. This is shown by con-
structing homeomorphisms.
(iii) From these homeos a homeo for 𝑀 is constructed (which is actually 𝜀-homeo).

38.8 There are five types of canonical regions

Trajectories in a single canonical region 𝑅 share 𝛼 and 𝜔 limits (Fig. 38.2).
Note that 𝜕𝑅 consists of separatrices, so when a band of parallel trajectories is
extended none of the trajectories inside the band cannot reach saddles. There-
fore, all the members of the bands must share the same destination.

Figure 38.2: Connecting two trajectories on 𝑅. The connecting curve 𝜎 must always be in 𝑅,
so it is singly connected.
Thus, 𝑅 has one 𝛼 and one 𝜔 limit set on its boundary. Each boundary
421
In MCP+MMP
38. PEIXOTO’S THEOREM 389

connecting these limit points cannot have more than one saddle because of
38.3(3), so we have only the following 5 types of 𝑅s (Fig. 38.3):

Figure 38.3: Five types of canonical regions

Their stability against perturbation is obvious.

38.9 Structural stability of separatrix set

Let 𝑋 (separatrix set Λ) and 𝑋 ˜ (separatrix set Λ̃) are vector fields within 𝛿 in
1
the 𝐶 -norm. Then, Λ and Λ̃ are one-to-one correspondent:
(i) Corresponding separatrices are of the same type (homeomorphic),
(ii) If a subset of Λ is the boundary of a canonical region 𝑅, then for the cor-
responding canonical region 𝑅 ˜ its boundary is a homeomorphic subset of Λ̃.
This follows from 38.3(1)-(4) and the usual discussion of the stability of
hyperbolic structures (+, if you wish, degree-theoretical arguments).

38.10 Strategy to prove the structural stability of 𝑋

To show the structural stability we must construct a homeomorphism between
𝑋 and 𝑋.˜ The strategy is as follows. Remove small discs 𝐷𝑖 containing fixed
points and tubular neighborhoods 𝑆𝑖 of limit cycles from 𝑀 to make 𝑁 .
(i) For each canonical region 𝑅 on 𝑅* = 𝑅 ∩ 𝑁 a homeomorphism is con-
structed.
(ii) Homeomorphisms are constructed on 𝐷𝑖 and 𝑆𝑖 .
390

(iii) Connect all these homeo to make a homeo for 𝑋.

Notice that both 𝑋 and 𝑋 ˜ live on the same manifold 𝑀 and All these local
sets, 𝑅s, 𝐷s and 𝑆s for 𝑋 are perturbed (deformed a bit) to become the coun-
˜ Since 𝐷 and 𝑆 are built around hyperbolic limit sets, they are
terparts of 𝑋.
˜ Thus,
stable under perturbation and, e.g., for a 𝐷, there is its counterpart 𝐷.
you can imagine that the structures due to 𝑋 and their perturbed counterparts
are overlapping on 𝑀 ; deformations and displacements are small.
Therefore, first, local homeomorphisms are constructed on these local sets.
This is 38.11 and 38.12.

38.11 Construction of homeo on canonical regions

We wish to show (i) in 38.10.
For type V region 𝑅* (Fig. 38.3), we wish to do the following (Fig. 38.4).
We make a 𝜀-homeomorphism for 𝑅* and 𝑅 ˜ * by constructing 𝜙 and 𝜙′ to a
square.422 The homeo we need is just 𝜙 ∘ 𝜙′−1 .

Figure 38.4: Construction of homeo on type V region 𝑅*

In this case, the smooth parallel portion has no problem, since 𝑋 and 𝑋 ˜ are
close. The only (slightly) unclear situations are around the saddle points 𝐾1
and 𝐾2 . This is clear if we note that if we take sufficiently small disk 𝐷𝑖 around
𝐾𝑖 , then the lengths 𝐿 of the trajectories of 𝑋 ˜ in 𝐷𝑖 can be as small as we

422
Precisely speaking, between 𝑅 and the square, homeo can be constructed, but not 𝜀-homeo.
Thus, we should say that we use the square to construct homeos, and then we must adjust them
so that 𝜙 ∘ 𝜙′−1 is an 𝜀-homeo.
38. PEIXOTO’S THEOREM 391

wish. Therefore, displacement of the trajectories due to perturbation must be

˜ displacement + 𝐿, so we can make this as small
smaller than the sum of 𝐷 − 𝐷
as we wish.
Types IV and III are considered as ‘degenerate’ version of type V as illus-
trated in Fig. 38.5.423

Figure 38.5: Construction of homeo on type IV region and III (bottom) is reduced to that of
type V.
Type II should be understood as a portion of V.
For Type I, instead of a square in the type V case, we make maps to a cir-
cular annulus.
All these maps are 𝜀-homeo.

38.12 Construction of homeo on nbh of separatrices

Since saddle points are not in the nonwandering set of the system in our case,
we have only to consider fixed points for 𝐷𝑖 . The situation is exactly the case
of Type I.
For limit cycles we can introduce ‘polar coordinates’ specified by the crossing
position on 𝜎 and the distance along the curve from 𝜎 in the tubular nbh as
illustrated in Fig. 38.6. Using such coordinates, we can construct a homeo.

38.13 Completion of homeomorphism

We construct a global homeo by patching the homeos on 𝑅* , 𝐷𝑖 and 𝑆𝑖 . To do
so we must glue these along the boundaries, so we need appropriate local defor-
mations. These deformations can be homeos, so we can adjust all the homeos
423
If you wish to be precise, for IV you introduce a coordinate system with a periodic boundary
condition along the dotted curve direction, and glue side edges (containing 𝐾s) continuously.
392

Figure 38.6: Coordinate system near a limit cycle

with these local deformation homeos and patch them continuously over 𝑀 to
construct a global homeo.
Historically, in this way, Peixoto’s theorem restricted on a disk was proved
first.424

38.14 Outline of the proof of ⇒

As is noted in 38.6, ⇒ is proved by a chain of approximation lemmas.
Start from any 𝑋.
(A) 𝑋 may be approximated by (= 𝑋 has in its sufficiently small nbh) 𝑌1 that satis-
fies 38.3(1), i.e., with finitely many hyperbolic fixed points. Note that we do not pay
any attention to other members of the non-wandering set Ω of 𝑋. Thus, te resultant
𝑌1 may have a ‘horrible’ invariant set 𝜇.425
Definition. A closed invariant set whose genuine subset is not a closed invariant set
is called a minimal set. A minimal set that is neither a fixed point nor a limit cycle
is called a nontrivial minimal set. [Note its closedness.]
(B) 𝑌1 may be approximated by 𝑌1′ without nontrivial minimal sets. Here, minimal
sets are converted to saddle connections and periodic orbits.
(C) 𝑌1′ may be approximated by 𝑌2 with fixed points (sink, source and saddle), closed
orbits and saddle connections as nonwandering sets.
(D) 𝑌2 may be approximated by 𝑌3 satisfying 38.3(1)-(3): We can sever all the saddle
connections, but perhaps new periodic trajectories are created. Note that periodic

424
M M Peixoto, On structural stability, Ann Math 69, 199 (1959).
425
You might say on 2-mfd not much complicated dynamics is possible. We know chaos requires
three dimensions, and putting trajectories in 2-mfd is like putting noodles on a tray. Indeed on
𝑇 2 there is no flow with positive KS entropy. However, Denjoy constructed a highly non trivial
invariant set whose ‘cross section’ is related to a Cantor set. If the genus 𝑔 of 2-mfd increases (𝑔
is the number of holes: 𝑔(𝑆 2 ) = 0, 𝑔(𝑇 2 ) = 1, etc. Adding a handle increases 𝑔 by one), then 𝑀
is riddled with holes, so perhaps trajectories could come back through holes and mingle with the
staying trajectories in a nontrivial fashion. Thus, it is safe to assume such 𝜇 exists in the flow due
to 𝑋 (and so in 𝑌1 ).
38. PEIXOTO’S THEOREM 393

trajectories need not be limit cycles.

(E) 𝑌3 may be approximated by 𝑌4 satisfying 38.3(1)-(4). This step uses Whitney’s
embedding theorem and Weierstrass’ polynomial approximation of continuous func-
tions.

38.15 Any vector field may be approximated by a vector field with

finitely many hyperbolic fixed points
We can make all the singularities as simple by arbitrarily small perturbations.
Simple singularities are isolated (and since 𝑀 is compact, there are only finitely
many of them), so we may handle them separately. A nonhyperbolic fixed point
can be converted by sufficiently small deformation to a hyperbolic fixed point.
Thus, 𝑋 may be approximated by a vector field 𝑌1 satisfying 38.3(1).

38.16 Perturbatively nontrivial minimal sets can be killed: outline

Suppose 𝑌1 has a nontrivial minimal set 𝜇. Since there are only finitely many
saddle points, we consider the following two cases (recall that usually dynami-
cal systems are defined on the time range (−∞, +∞))
(A) There is no trajectory connecting 𝜇 and a saddle.426
(B) There is a trajectory connecting 𝜇 and a saddle.
For case (A) we can show (38.17) that for any 𝑝 ∈ 𝜇 there is a coordinate
nbh such that all the trajectories leaving it return to it,427 and the trajectory
going through 𝑝 can be converted into a periodic orbit. For case (B) such a
trajectory can be converted into a saddle connection (38.18).
After these perturbations, if nontrivial minimal sets still remain, we repeat
the procedure. It will be shown (38.19) that only finitely many repetition is
required, so 𝑌1 can be perturbed with an arbitrarily small perturbation into 𝑌1′
without nontrivial minimal set.

38.17 Case A: no saddle connection

For (A) in 38.16 𝑌1 has a closed orbit passing through 𝜇. This closed orbit
does not bound any cell (see Fig. 38.7) [Closing lemma].

426
As we will see, this happens only if 𝑀 = 𝑇 2 pf a Klein bottle, so the system actually has no
singularity at all.
427
Since 𝜇 is non trivial, there is a trajectory returning to any nbh of 𝑝 infinite times.
394

Figure 38.7: Two kinds of closed orbits on 2-mfd.

To show this take a local coordinate square ‘abcd’ around a point P in 𝜇

(Fig. 38.8).

Figure 38.8: Local coordinates for 𝜇; Right illustrates the impossible case for (A)
Since 𝜇 is minimal, at least one trajectory leaving from edge ab must come
back to 𝜇 through edge cd.428 The boundary trajectory between the trajecto-
ries coming back to 𝜇 and those not coming back must go to a saddle point (See
Fig. 38.8 Right).429 However, our assumption is that such 𝜇-saddle connection
does not exist. Therefore, all the trajectories leaving 𝜇 must come back to 𝜇
(infinitely many times to different points on edge cd), since 𝜇 is nontrivial.
Also these orbits cannot encircle a cell (see Fig. 38.7 Right); 𝜇 is a closed
set, so its boundary cannot be a cycle; otherwise, since 𝜇 is a closed set, this
boundary cycle belongs to 𝜇, destroying its minimal nature.
Since trajectories do not cross each other, the ordering of the trajectories is
preserved (or reversed). Consider the perturbation illustrated in Fig. 38.9. We
can create a limit cycle going through P.

Needless to say, this may not totally erase 𝜇; some potion(s) may survive as
nontrivial minimal sets. Thus, we need to repeat the procedure.

428
The local square around P is so chosen that the trajectories foliate the square. This is possible,
because there must be a trajectory coming back to any nbh of P infinitely many times, and 𝑀 is a
2-mfd, so near P the returning portion of the trajectory must be almost parallel.
429
A formal proof is lengthy. Lemma 3 in the original.
38. PEIXOTO’S THEOREM 395

Figure 38.9: Closing lemma

38.18 Case B: when 𝜇-saddle connection exists

For (B) in 38.16 the 𝜇-saddle point connection can be converted to a saddle-
saddle connection.
Actually, we have only the case of Fig. 38.10 Left. Suppose there is a 𝜇-
saddle connection in the future direction (the right-side edge situation in Fig.
38.10 Left). Then, above and below the connection, the trajectories have dif-
ferent fate. However, these trajectories come back to this coordinate square 𝑅.
Then, reverse the time, we must say there must be a saddle on the left side
just as illustrated in the figure. This saddle 𝛾1 cannot be the same saddle as
𝛾2 ., because if so the trajectories must cross the unstable manifold of 𝛾1 .
We may assume that 𝜔(𝛾1 ) = 𝛼(𝛾2 ) = 𝑝. If not, we have a situation like Fig.
38.10 Right, and the curly bracketed portion may be subjected to a surgery as
performed in case (A).
If this surgery fails to connect 𝛾1 and 𝛾2 , then the boundary between the
curly bracketed portion and its outside must go to a saddle point as illustrated
in Fig. 38.8 Right. Thus, we get the situation as Fig. 38.10 Left. When 𝛾1 and
𝛾2 are sufficiently close, we can short-circuit them with a small perturbation to
create a saddle connection.

Figure 38.10: Saddle connection case

38.19 Finite repetition of surgeries totally kill nontrivial minimal

sets
396

Since the total number of saddle point in 𝑌1 is finite, so the number of surgeries
in (B) must be finite.
If (A) does not end with finitely many surgeries, we have very many closed
orbits. Since 𝑀 is a 2-mfd, with a finitely many cuts 𝑀 can be expanded into a
2-disk.430 Since trajectories cannot cross closed orbits, all the orbits must be on
this disk. Suppose we still have some 𝜇 on this disk, since all the saddle points
have been used up to make saddle connections, it produces an orbit encircling
a cell, an impossibility as already discussed in 38.17.
Thus, the procedure explained above ends with finite repetitions and an ar-
bitrarily small perturbation can convert 𝑋 into a vector field with finitely many
hyperbolic fixed points without any nontrivial minimal set (but with too many
periodic trajectories and possibly bands of periodic trajectories (as centers)).

38.20 Almost homoclinic orbits are converted to homoclinic orbits

(C) has almost been demonstrated, but there can be a situation where an al-
most homoclinic situation occurs for a saddle (Fig. 38.11), because the situation
does not produce a nontrivial minimal set.

Figure 38.11: With a little perturbation a homoclinic orbit may be created.

If we look at 𝑅, since there is trajectory coming back to any nbh of P, so we can
convert this into a homoclinic trajectory. Since the number of saddle points
is finite, just as discussed in 38.19, we have only to repeat such a procedure
finite times and the perturbation to 𝑌2 is complete.

38.21 Perturbation can remove all the saddle connections

Graphically, we can exhibit the situation as in Fig. 38.12 [Actually such a dia-
gram is called a graph].

430
Genus 𝑔 2-mfd may be cut open to a disk (or a polygon) with 2𝑔 cuts, just as 𝑇 2 is converted
to a square by 2 cuts.
38. PEIXOTO’S THEOREM 397

Figure 38.12: Graphic representation of saddle connections

There are two situations:

(1) 𝑆1 𝑆2 is not a part of a large graph.
(2) 𝑆1 𝑆2 is a part of a large graph
For (1) the following perturbation does not produce a saddle connection
anew.

Figure 38.13: Surgery of saddle connections

However, in case (2) a new saddle connection might be formed (Fig. 38.14),
but, in this case, if perturbed further n (pushed down further in the figure),
the saddle connection would be reconnected to a periodic orbit. Since there are
only finitely many saddle points, this can be done by arbitrarily small pertur-
bation, so we may virtually ignore such cases.

Figure 38.14: Possible emergence of a new saddle connection, but this can be killed by the small
perturbation indicated by an arrow

In any case, we can repeat the above procedure finite times. All the saddle
connections will be gone. Thus 𝑌3 has been produced from 𝑋. That is, except
for the condition for periodic trajectories we are done.
398

38.22 Periodic orbits are made hyperbolic by perturbation

To make periodic orbits hyperbolic, we apply the perturbation illustrated in
Fig. 38.15 to box 𝑅:

Figure 38.15: Make orbits hyperbolic

Also marginal periodic orbits can be understood as a degenerate version of
two hyperbolic orbits. Thus, We can make all the separatrices hyperbolic (the
stabilized 𝑌3 ). However, still there can be infinitely many closed orbits (like a
center).
Peixoto embedded the stabilized 𝑌3 into R5 using Whitney’s embedding the-
orem, and then made a polynomial approximation of the vector field (using
Weierstrass’ theorem). That is, he made an analytic vector field, which has
only isolated periodic trajectories and periodic bands. Then, he constructed
a vector field homeomorphic to this polynomial approximation on 𝑀 . The
resultant vector field is no more analytic, but the topological features are all
preserved. Thus, the vector field satisfies (1) and (3) in 38.3 and isolated peri-
odic trajectories/bands. However, nontrivial minimal sets 𝜇 may reappear by
the approximation procedure.
If this further satisfies (2), that is, no new 𝜇 shows up, then there are two
cases:
(A) All the orbits are closed.
(B) Otherwise.
For (A) if 𝑀 is not a result of gluing a square (i.e., 𝑇 2 or Klein’s bottle),
then we have something like Fig. 38.16.

Figure 38.16: Expanded 𝑀 should not have a hole; the holes mean these peripheries are glued
to make a handle.
38. PEIXOTO’S THEOREM 399

This implies there are saddle points, contradicting the assumption that there
are only closed orbits. If 𝑇 2 , for example, we can apply ‘bunching’ as exhibited
in Fig. 38.15 to make a limit cycle.
For (B) even if we have a band of closed orbits (no returning orbits), the
boundary of the band (black dots in Fig. 38.17) goes to a saddle (or comes
from a saddle) just as in Fig. 38.8 Right, so this contradicts 38.3(3), because
𝑌3 has already been constructed to satisfy this and polynomial approximation
does not alter this.431

Figure 38.17: Killing the bands

If (2) is not satisfied, that is, there is nontrivial minimal set 𝜇, then repeat
the above argument (38.16). If, after this repetition, the number of orbits is
finite, we are done. If not, we repeat the polynomial approximation, because
polynomial fields allow only isolated or band of periodic orbits. Since the poly-
nomial approximation maintains all the simple closed orbits, the newly added
simple trajectories during the repetitions are maintained by each polynomial
procedure step. However, as discussed in 38.19, we can repeat the procedure
only finitely many times, and 𝜇 will not show up eventually. We are done.

431
need polynomial field check