Ergod. Th. & Dynam. Sys.

(1990), 10, 531-554

Printed in Great Britain

On Cherry flows
MARCO MARTENSt AND SEBASTIAN VAN STRIENt
Mathematics Department, Technical University of Delft, Julianalaan 132,
2628 BL Delft, The Netherlands
WELINGTON DE MELO
IMP A, Estrada Dona Castorina 110, 22460 Rio de Janeiro RJ, Brazil
PEDRO MENDES*
Departamento de Matemdtica, ICEx-UFMG, Cidade Universitdria, Pampulha,
30000 Belo Horizonte MG, Brazil
{Received 1 June 1988 and revised 28 July 1989)

Abstract. The purpose of this research is to describe all smooth vector fields on the
torus T2 with a finite number of singularities, no periodic orbits and no saddle-
connections. In this paper we are able to complete the description within the class
of vector fields which are area contracting near all singularities. In particular we
give a large class of analytic vector fields on the torus T2 which have non-trivial
recurrence and also sinks.
This result for vectorfieldsfollows from a result dealing with continuous monotone
circle mappings which are possibly constant on a finite number of intervals.

1. Statement of results for vectorfieldson T2

In this paper we are interested in recurrence of vector fields on surfaces. From the
Poincare-Bendixson theorem the dynamics of vector fields on the sphere is clear:
the w-limit of every point is equal to a point or to a closed orbit. On the torus T2
the situation is more interesting. If a vector field on T2 has periodic orbits then
there are again no non-trivial recurrent orbits, because then the situation reduces
to a planar one.
In his celebrated 1932 paper, Denjoy studied vector fields on the torus T2 without
singularities and without periodic orbits. He showed that if a vector field X on the
torus is Cx, then such a vector field either has a periodic orbit or every orbit of X
is dense in T2. In 1963, A. J. Schwartz extended Denjoy's result and showed that
any minimal set of a Cx vector field on a compact connected, two dimensional
manifold M is either a singleton (consisting of a singularity), or a single closed
orbit, or all of M in which case M = T2. (Here a set L is called minimal if it is a
t Supported by Netherlands Organisation for the Advancement of Pure Research (N.W.O.).
+ Since most of the results in this paper were proved independently by M.M. & S.v.S and by W. de M.
& P.M. it was decided to publish this paper jointly.
* Partially supported by the Brazilian CNPq, grants 300376/79/0 and 150013/87/MA/Fv.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 532 M. Martens et al.

closed, non-empty set which is invariant under the vector field such that L contains
no smaller set with these properties.) From this it follows that the a>-limit of any
point is equal to a periodic orbit, to M = T2 or contains a singularity. But because
no minimal set can strictly contain a singularity, the result of Schwartz gives a rather
incomplete description of the dynamics of vector fields with singularities.
The next simplest vector fields on the torus have
(i) a finite number of singularities all of which are hyperbolic;
(ii) no periodic orbits;
(iii) no saddle-connections;
(iv) no sources.
We shall call a vector field which satisfies properties (i)-(iv) a Cherry vector field,
because Cherry gave in 1938 an example of an analytic vector field on T2 satisfying
these properties with a sink and also a recurrent orbit. His example is quite specific,
but shows that Cherry vector fields can have recurrent behaviour.
In this paper we will generalise Denjoy's result to all Cherry vector fields which
have negative divergence at each saddle-point. This last condition is unnecessary if
X is a Cherry vectorfieldwith at most one saddle-point. In this case, index arguments
give that X can have only one other singularity. We may assume that the divergence
in these singularities is of opposite signs, because otherwise either X or -X would
satisfy the conditions from before. So in this case we can assume that X has precisely
one sink, no sources and one saddle-point. At this saddle-point the divergence of
X is positive. This situation can be treated using the result of C. Gutierrez, A. Lins
and W. de Melo described in § 6.
Theorem A implies that there exists a large class of vector fields which display
the same phenomenon as Cherry's example.
Let 38 be the class of C^ Cherry vector fields X on T2 such that at each singularity
p the divergence of X is smaller or equal to 0. Furthermore, let 53' be the class of
Cx Cherry vector fields X on T2 with precisely one hyperbolic sink and one
saddle-point; moreover assume that X has positive divergence at this saddle-point.
(As we noted above, for each vector field X with only hyperbolic singularities and
one saddle-point either X e 38 u 53' or -X e 53 u 38'.)
Denote the flow of X through x by t-*X,(x). The a and a> limit set of x are
defined as a(x) = {y; 3f,, -»oo with X_,u{x)->y}, respectively w(x) = {y; 3tn-*oo
with X,ir(x)-»_y}. We say that x is recurrent if xea(x)<ua)(x).
THEOREM A. Let X e 53 u 38'. Then there exists an infinite set Lcz T2 such that every
point xeL is recurrent, co(x) = a(x) = L. Moreover, for every xeT2 which is not
contained in the unstable manifold of a saddle point one has a(x) = L and for every
point xeT2 which is not contained in the stable manifold of a sink or of a saddle, x e L.
Remark 1. The condition that X e 33 u 38' implies that X cannot have any sources.
Remark 2. It is not very difficult to construct many vector fields which satisfy the
assumptions of Theorem A.
Remark 3. Of course Denjoy's result also holds for C2 vector fields. Our result also
holds for C 2 vector fields, but in this case we have to require that there exist C2

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 533

linearising coordinates near every saddle-point of X, or alternatively that X is C 6

near saddle-points (see the Appendix).

The next result says that vector fields X e 38 can be described by circle maps.

T H E O R E M B. Let X e 38 u 38'. Then there exists a neighbourhood U of X such that

for every Y e Uone can associate a continuous order-preserving return-map fY :S' -> Sl
with the following properties:
(a) / y depends continuously on Ye U;
(b) if Y, Z e U have no periodic orbits and no saddle-connections then they are
equivalent iffY and fz are topologically conjugate circle maps.

In fact, these circle maps may be constant on some intervals. Also they will satisfy
some smoothness conditions and some non-flatness conditions at the boundary of
the intervals where they are constant. In the next section we will study a special
class si of circle mappings and show that if these maps have no periodic points
then they are determined by a finite number of parameters, see Theorem D. In § 6
we will show that the maps/ x are contained in this class si, provided some transition
map is sufficiently smooth and non-flat. In the Appendix we will show that these
conditions hold for all smooth vector fields. (These conditions would follow immedi-
ately if we assumed that X was C2 linearizable at singularities. However, it is rather
pretty to see how natural the conditions that one obtains without assuming lineariza-
bility correspond to the conditions needed in the class si.)

2. Statements of results for circle maps

Let S' = R(mod 1) a n d / : S' -> S1 be a continuous, order-preserving, degree one map
which is possibly constant on a finite number of intervals. Since/is order preserving
it has a well-defined rotation number p ( / ) e [ 0 , 1 ) . p ( / ) is irrational if and only if
/ has no periodic points. It is well known that if p(f) is irrational then / is
semi-conjugate to the rotation Rl>(f):x-+x + p(f) mod 1. The semi-conjugacy is
continuous, monotone, has degree one and maps orbits of/ to orbits of R,,in.
Assume that/: S' -* S1 is continuous, order-preserving, degree one and everywhere
1
C except possibly in a finite number of points. Let K, be the closure of the set of
points x with either
- / i s not C' in x, or
- / i s constant on some neighbourhood of x, or
- Df(x) = 0.
Let C, be the closure of the set of points where / is locally constant.
We say that such a map / is in si if / satisfies the conditions from above, K, has
finitely many components and also the following smoothness and non-flatness
conditions. For each point xoeS'\K, there exists a neighbourhood U of x0 such
t h a t / is a local C1 diffeomorphism on U and Df has bounded variation on U.
f is constant on each component of Kf. (In particular/ is everywhere C 1 except
possibly in points xedK,.) Moreover we require that for every point xoedKr, f
satisfies the following non-flatness condition. There exist neighbourhoods U, V of

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 534 M. Martens et al.

x0 respectively f(x0), a constant a > 1, and two C' difleomorphisms <£ r :((-l, 1), 0)-»
(U,x) and 0 , : ( / ( t / ) , / ( x ) ) ^ ( ( - l , l),0) such that D<t>, and D<f>r have bounded
variation and </>, ° / ° </>r(x) is equal to 0, or ±|x|".
Similarly we say t h a t / e M' if / has degree one, is everywhere strictly increasing
except on one interval K, where it is constant and / | ( S ' — Kf) is C'. Moreover at
the boundary points of K, = [a, b] the (one-sided) derivative of/ is infinite and
there exists e > 0 such that log Df\ [a - e, a) and log Df\ (b, b + e] are monotone.

Remark. If/ is C2 on Sl\Kf, and the limits limx^XoX<SK/ D2f(x) are non-zero for
each point xoedKt, then/e si. However, we will not restrict ourselves to such maps,
because, in applying our results to vector fields on the torus, we get that singularities
of the vector field of saddle-type give rise to critical points of a corresponding circle
map of the form |x|*, where 5 > 0 and is the absolute value of the ratio of the
eigenvalues at the saddle-point and need not be an integer.
We can also consider the slightly more general class of maps such that there exists
neighbourhoods VcClos (V)c U of K, such that the m a p s / | ( S ' \ V), <£,| U and
<l>r\U are absolutely continuous and the maps log Df\(S\V), log D<£, \U and
log D<j>r\ U (which therefore exist almost everywhere) are almost everywhere equal
to maps with bounded variation. The proofs of our results go through without much
change; we indicate the changes needed in the remark following Proposition 3.4.
We say that / is a wandering interval of/: S1 -» S1 if
- / " ( / ) n / m ( / ) = 0 , Vn, m > 0 with n*m;
- there exists no n > 0 such that/"(/)<= Cf.

THEOREM C. Let f be in si^jsi' with irrational rotation number. Then f has no

wandering intervals.
COROLLARY. For every x e Sl\[_)nJ,oj^"(Cf) both the forward and the backward orbit
ofx is dense in S\\^Jn,,of~"(Cr). In particular there exists a non-finite recurrent orbit
off.
Proof of Corollary. Since/has no wandering intervals, the semi-conjugacy h, between
/ a n d /?,,(/-, is constant only on preimages of Cf. Since S'\LJ n »o/~"(Q) cannot be
empty the Corollary follows. •
Remark 1. J. C. Yoccoz has proved Theorem C for maps / which are strictly
monotone (isolated critical points are still allowed). However, if a vector field on
T2 has sinks then the corresponding return-map is constant on some interval. So
the result of Yoccoz cannot be applied to vector fields. Unfortunately the proof of
his result cannot be 'directly generalized to maps which are allowed to be constant
on some intervals, see the remarks below: Definition 3.1, Lemma 3.7 and in § 5.
Remark 2. Consider order-preserving maps f:S'->S' of degree one which are
constant on at least one interval and which are discontinuous in at least one point.
Let K, be the closure of the set of points where / is locally constant or is discon-
tinuous. Assume that / satisfies the same smoothness conditions as the maps from
M (except that / is not continuous). Then Gutierrez has shown that either p(f) is

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 535

rational or some iterate of one of the components of Kf is mapped into some other
component of Kf.
So the main thing which is missing in the general description of these maps is
the case where a > 1 in some points in dK, and a < 1 in other points of dKf.
From this result one can easily show that the topological conjugacy class of/ is
determined by a finite number of invariants. Indeed, if p(f) is irrational, then there
exists an order-preserving map hf: S1 -» 5' such that h, °f= Rp ° hf, where Rf,(x) =
x + p (mod 1) and p = p(f), see [He]. For simplicity write R = RP. Clearly hf is
constant on the components of Cf.
T a k e / e ^ u ^ ' . If p(f) is irrational and hr is constant on some interval / then
R"(hf(I)) is disjoint for all n > 0 . Hence the intervals/"(/) are mutually disjoint
for all n and from Theorem C there exists n > 0 with /"(/)<= Cf. It follows that
hfl(hf(x)) is a non-trivial interval if and only if xe\^JnsOf~"(Cr). Let d be the
metric on 5 1 =IR (mod 1) induced by U.
THEOREM D. Let f gesi^j sd' with irrational rotation number such that the number
of components of Cf is the same as that of Cg. Call this number k. Assume that
(2.1)
and that we can order the components C) and C'g of C, respectively Cg cyclically on
S' such that

Here c}=/j r (C}) and cg = hg(Cg). Then f and g are conjugate, i.e. there exists a
homeomorphism h: S1 -» S 1 such that h ° / = g ° h.
Proof of Theorem D. We begin by taking an arbitrary order-preserving homeomorph-
ism k:C,-* Cg such that k(C)) = C'g. Then, let h, and hg be the order-preserving
maps such that hr °f = Rp ° hf and hg°g = Rl,°hg. Choosing 6 so that Re maps
h,(Cj) onto hg(Cg), one gets from (2.2) that Ro(hf{Cf)) = hg(Cg). Moreover taking
hr = RH° hh the following diagram commutes

S1 ——* S1

S1 - > S'
and that hj(C ; • ) = c' . So -
hr(x ), for**
*(*) = •
n
u •(

k(x) forxe

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 536 M. Martens et al.

defines a strictly order-preserving homeomorphism from S'\Un2i/~"(Q) to

S'NLUi g"(Cg) such that h °f=goh. Finally, for xe\Jn^r"(C,), let n > l be
the minimal number such that/"(x)e Cf and define li(x) = g~" ° h °fn. In this way
we have defined a homeomorphism h such that h ° f = g ° h. •
In the Proof of Theorem C, we will need a simple Proposition which produces
collections of disjoint intervals. Since it does notfitlogically in the next two sections
we state and prove it here.
PROPOSITION 2.1. Let {W:\i = 1 , . . . , m] be a collection of intervals in S 1 with the
following two properties.
(i) every point in S 1 is in at most three elements of the collection;
(ii) Wj c Wk implies i = k.
Then there exists a partition of the collection
{W,\ i = 0 , l , . . . , m } = / l | u A 2 u - • - u A j ,
such that Ak consists of mutually disjoint intervals for each k = 1, 2 , . . . , 5.
Proof. We prove this Proposition inductively for each m s 1. The Proposition is
trivially true for m = 1. Suppose that it is proved inductively for m - 1 . Then we can
write
{W;-;i = l , 2 , . . . , i n - l } = U A)
j=l,2,...,5

where A) consists of mutually disjoint intervals. From property (ii) it follows that
dWmn Wi7i0 when Wm n W , ^ 0 . So assumption (i) implies that there exists at
most four is with l < / < w such that Wm has a non-empty intersection with W).
Hence there e x i s t s / e { 1,2,... ,5} such that Wm does not intersect any interval
from A'y. So by letting As = A) for j e { 1 , 2 , . . . , 5}, j *f and A,- = A'ru{ WJ we
get the required partition of {Wt; i = 1 , 2 , . . . , m). This proves the Proposition by
induction. •
The organization of the paper is as follows. In the next three sections we will
prove Theorem C. In § 6 we will prove Theorems A and B, except a technical Lemma
which is proved in an Appendix.
It will be convenient to introduce the following notation. For an interval / let |/|
denote the length of /. Moreover let d be the metric on S ' = R (mod 1) induced
by U.

3. Some analytic tools

In this section we will develop some analytic tools which give estimates on the
shape of non-linearity of/" on intervals T such that T,f(T),...,/"(T) are disjoint.
Definition 3.1. Let g: S 1 -* S 1 be a continuous order-preserving map and /, T intervals
in S 1 with Clos ( / ) c int (T). Then

D(T, / ) = . . . , where Land R are the components of T \ / ;

\L\-\R\

B(g,T,I) =

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 537

Remark. J. C. Yoccoz uses the operator B0(g, T) = lim,_ r B(g, T, I).

Definition 3.2. Let g: T-* S1 be a C 3 -map on the interval I c S1. Then the operator

g'(x) 2\g\x)
is called the Schwarzian derivative.

It is easy to see that the operator B is multiplicative:

B(g o h, T, I) = B(g, h(T), /)(/)) x B(h, T, I). (3.1)
The proof of the following proposition can be found in [MS].

PROPOSITION 3.3. Let g:T-*T be a C3 diffeomorphism on the open interval T such

that for every x e T, Sg(x)<0. Let Clos(/)<=int(T). Then
B(g,T,I)>h (3.2)
Let Yc Clos( Y) c int (°U) be interval neighbourhoods of Kf such that the number
of components of Y and °U are both the same as #Kf.

PROPOSITION 3.4. Let fesi. Then there exists V<oo with the following property. Let
5 = {T,, T2,..., Tn} be a collection consisting of n intervals in S 1 with the following
properties. Every point of S' is contained in at most three intervals Tt, T,;<= 7} implies
i =j and none of the intervals T, contains a component of°U\Yor contains points of
K,. Then

t log B(/;r,,/,) a - 5 - V. (3.3)

Proof.Let V,=Var (log Df\(S\Y)). Since on S'\Y the m a p / i s a C 1 diffeomorph-

ism and Df has bounded variation, Vf < oo. Moreover on each component <%, of °U,
f has the form f(x) = <f>u ° <f>a ° <f>ri, where <t>a{x) = ±\x\" as above. Here </>,, and
4>rj are C' diffeomorphisms such that log D</>,, and log D<f>r, have bounded variations
on %. Hence V, =X, Var (log (D^,,,)) and Vf =!,-Var(log(D^ M -)) are both finite.
Let V—V-+Vi+Vr.
Let /, = {i; 7] n T = 0 } and I2 = {i; 7] c %}. Since the intervals 7] never contain
a component of °U\Y, /, u / 2 = {1,2,..., n).
First assume that i e I,. Let L, and /?, be the components of 7]\/,. For «;, u, e 7]
let (u(, V,-) be the open interval connecting M, and vt. Because/ is C 1 on S\Y the
mean-value theorem is valid on Tt. Using this in the definition of B(f Th /,-) we
find that there exist /f G L,, r, e /?;, m, e /, and T, e T, such that

....
(D/(/)D/(J (3.4)
and
m,e(/,,r,). (3.5)

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 538 M. Martens et al.

From (3.4) one has

|log B(f, Tt, 7,)|s|log D/(m,)-log D/(/,)| + |log D/(T,)-log D/(r,)| (3.6)
and also
|log B(f, Ti3 /,)|s|log D/(m,)-log D/(r,)| + |log D/(T,)-log D/(/,)|. (3.7)
Rename the points /,, mu rt, T, in increasing order a,, bt, ciy d{. From (3.5) one gets
that either (/,, w,)n(T,, r,) = 0 or (T,,/,)n(m y , r,) = 0 , and so we can use either
(3.6) or (3.7) and get
|log B(f, Tit 7,)|s|log D/(fe,)-log D/(a,)| + |log Df(d,)~\og Df(c,)\

and therefore
log B(f, Th 7f) > -Var (log Df\ Tt). (3.8)
Now consider i e I2. Then T{ is contained in some component ^ , of °U. (and does
not intersect Kf) and so / has the form f(x) = <j>u ° <j>a ° <f>ri. Hence
B(f, rit /,) = Bit,,,, r:,vi) x B(4>a, T;, /;) x B ( ^ , 1 ; T,, /,).
Here T? = </.„ ° ^,,(7,), T!; = ^.r,,(T;), /? = *„<>*,,,(/,) and /; = «/.,,,(/,). Since the
Schwarzian derivative of (/>„ is less or equal to 0 (because a > l ) one gets
B(<f)a, r ; , / ; ) > l . Hence, as above,
log B(f, T,, /,-)alog B(</»,,,, r r , / n + 0 + log B(^r,,, T,-, /,)
2=-{Var(logD0 l . i |Tr) + Var(logD0r,i|Tj)}. (3.9)
Since every point of S' is contained in at most three intervals, using Proposition
2.1, one can write {7^,; i = 1, 2 , . . . , n} = A, u A2 • • • u As where A, consists of a
collection of mutually disjoint intervals. Therefore also the collection of intervals
T" corresponding to Tt e A, consists of disjoint intervals. Hence from (3.8) and (3.9)
one gets

The Proposition follows. •

Remark. In eq. (3.4) in this Proposition we have used that Df exists and has bounded
variation on S\T. Of course it would have been sufficient to assume that / is
absolutely continuous, log Df\(S'\T) is almost everywhere equal to a map with
bounded variation. In this case we could have still found m, e lt, /, 6 i,,, r, e R, and
r, e T) such that Df exists in these points and such that

ana

Hence (3.6), (3.7) and (3.8) would still hold. Similarly it suffices to assume that <£,,
4>r are absolutely continuous and log D<£, and log D<f>r are almost everywhere equal
to maps with bounded variation.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 539

LEMMA 3.5. Letfe sd. Then there exists Ao> 0 such that if I < intervals in Sl
such that
(a) W\ I consists of one component H;
(b) | H | < | / | ;
(c) InKf = 0,
then

(3.10)

Proof. Let us prove this Proposition by contradiction. So suppose that there exists
a sequence of intervals /„, Wn, Hn = Wn\In satisfying (a), (b) and (c) above such that

lim ;-»0. (3.11)

-H/(H n )|/|H n |
Since |W n |<|/ n | this implies that |Wn|-»0 and * Kf. By taking a subsequence
we may assume that Wn -» p e K,.
Since / G ^ we may assume that p is a critical point and near p, f is of the form
f=<f>l°<t>°<f>r where <f>a is equal t o x ° o r O o n x > 0 and to 0 or — \x\a o n x < 0 a s
before. Hence
|/(/„)[/1/„ | minPfr m i n P 0 |^ n (J^)|/|/^|
|/(// n )|/|H n | supD^, supD^ f |^ a (Hi)|/|H;|" "
1
Here H^ = <j>r{Hn) and /^ = <£r(/n). Since <^>r is a C difleomorphism, writing Xn =
l#!i|/|/' n | we get HmsupK n <oo and therefore there exists Ke(0,oo) such that
Kn < K, Vn >fl. Let us distinguish between the following three possibilities.
Case 1. /„ is between p and Hn.
Writing / ; = [an, bn], H'n = [bn, c j c (0,1) one has

\4>aV'n)\/\rH\
\<t>AH'H)\/\H'H\
b"n
lbn
Uan -baA
-an. m. m y-(bnr~\
cn -bn + K)bn-bn J
cn-bn]
K
(3.13)
(l + K T - i
Here, the second inequality follows because a > 1 implies that the function
= a-x" is increasing. TTie third inequality follows from this and because
\H'n\^K'n-\In\ gives
cn^bn + Kn- (bn - a,,) < 6n + Kn • bn s (1 + K) • bn.
Using (3.13) in (3.12) gives a contradiction in (3.11).
Case 2. H,, is between p and /„, but Hn n K, = 0.
Since the derivative of <pa is increasing we have in this case
\4>AIn)/\I'n\
(3.14)
\<f>t,(H'n)\/\H'n\>h
Using (3.14) in (3.12) we again get a contradiction with (3.11).

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 540 M. Martens et al.

Case 3. Hn is between p and /„, but Hn n K, # 0 .

If //„ intersects an interval component of Kr, then we can choose the maximal
interval Wn with /„ <= Wn <= Wn such that Wn n K, = 0 . Let //„ = Wn\In and H'n =
<t>r{H'n). By construction | / / ^ | < | / / ; | and <i>a{H'n) = 4>a{H'n), and therefore
\<t>a(I'n)\/\I'A
\<t>a{H'n)\/\H'n\ \<j>a{H'n)\/\Hn\-
As in case 2 this last term is bounded away from zero and we get a contradiction
as before.
So we may assume that Hn n K, is an isolated point. So writing H'n = [-an, bn],
I'n = [K, cn], with an, bn, cn > 0 one has

\<t>a(I'n)\/\I'n\
\<t>a{H',,)/\Hn\
C"n-b»n
.cn -bn\
'b"n
b +a
Cn

K <
-ban
.cn -bn\ m (3.15)

bn an.
Here the first inequality follows from
x+u x u
<—h— for all x, y, u, v > 0.
y+v y v
The last three inequalities follow from the concavity of xa and because bn, a n s
\H'n\<K- \I'n\< K- cn. So using (3.15) in (3.12) we get again a contradiction with
(3.11).
Since cases 1, 2 and 3 cover all possibilities we have proved the Proposition by
contradiction.
•
1
PROPOSITION 3.6. Let fe sd and /, The intervals in S with Clos
int (T) and L
and R be the components of T\I. Let ye (0,1). / /

'W\ and
then
B(fT,I)>y2.
Proof.

\f(L)\y y+y2-d + yd
_
\R\ \R\\L\ \L\
Here

\R\\L\
d=

\R\ \L\

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 541

and the last inequality follows since y e (0,1) and therefore (0, OO)BX-»
(1 + yx)/(l + x) is always greater than y. The result follows. •
PROPOSITION 3.7. Letfe si. Then there exists A, > 0 such that if I, T are intervals in
S1 with Clos(/) c int (T) and L and R the components of T\I such that
(a) m*\I\ or \R\*\I\;
(b) InKf = 0 ;
then
B(f,T,I)*Ax. (3.16)
Proof. Let Ao be the number from Proposition 3.5. We may assume that Aoe (0,1).
We will prove the Proposition for A, = | • (Ao)2. By possibly renaming L and R, we
may consider the case that | / ? | < | / | . Then from Proposition 3.5 we get

\f(R)\/\R\
and hence

If \L\>\I\ then it follows from this a n d | K | < | / | that

and the Proposition is proved. So assume that | L | < | / | . Then applying Proposition
3.5 again we get

«"-
and it follows from (3.17), (3.18) and Proposition 3.6 that B(f, T, I)>A0- AQ = A,.
D
Remark. In general the interval T we need to consider may intersect or even overlap
a component of K,. It is easy to check that there exists no universal lower bound
for B0(f, T) (as in (3.16)) in this case. Here Bu is the operator of Yoccoz mentioned
in the remark below Definition 3.1. Therefore we have to replace the operator BQ
by B.
Let Yc. °U be the neighbourhoods of Kf from above.
THEOREM 3.8. Let / e i . Then there exists e>0 and Bo such that for any intervals
/<= Clos (/)<= int (T) in S', and any n > 0 , satisfying the following conditions
(a) | L | < e - | / | and \R\<e- |/|; (Here L and R are the component ofT\I.)
(b) every point in S1 is contained in at most three members of the sequence T,

(c) f(I)nKr = 0,ie{O,l,...,n-\} then

B(f",T,I)>Bt). (3.19)
Proof. Let Ax be the number from Proposition 3.7 and assume that At < 1 and
e"5V < 1. Let V be the number from Proposition 3.3. We may assume that e~5V < 1.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 542 M. Martens et al.

LetB 0 = {A, e-5Vy2#Kr+l and e = VB0/3. Let 0 < f, < t2< • • •< t, < n -1 be integers
such that either/'(T) contains a component of °U\Y, or f'(T)n Kf9i0. From the
disjointness property of the orbit of T we get s== \2#Kf. From the definition of f,
one has that/'(T) does not contain a component of °U\T for i < f, and because/
does not have periodic points we have that f'(T)<=f'(T) implies i=j. Using this
and the disjointness property of/'(T), Proposition 3.4 gives
B ( / \ T,/)>e- 5 V .
So

D mm
|/'-a)||/''(K)l \L\\R\-
This implies

!/()! !/()! !/()

I yf / r»\l I ft / r M I /*r / r \l
r n.
0
1
2

So \f'{R)\^\f'(I)\ or | / ' ' ( i ) | ^ | / ' ' ( / ) | . Hence we can apply Proposition 3.7 and
get B{fJ\T),f'<{I))>Ax and hence
B(/'. + 1 , r , / ) > B ( / , / ' ' ( 7 - ) , / ' 1 ( / ) ) x B ( / \ T , / ) > e - 5 V - A 1 . (3.20)
In the same way we get

and hence, using (3.20),

From this we get as before \fh(R)\^\f'2(I)\ or |/'*(L)|s|/'2(J)|. Hence we can

apply Proposition 3.7 again and get

This procedure has to be repeated at most 12 • #/£, times, i.e., s < 12 • #Kr. Since
[e" 5V - A , ] 1 2 # K ' > Bo we get by induction that for each fc< s, |/'^(/J)| s |/'^(/)| or
| / ' ^ ( L ) | < | / ' K / ) | and hence
B(f+\ T, / ) > [ e - 5 V - A , ] l 2 # K ' .
Using Proposition 3.4 we can handle the last piece from f., + 1 to n and get
1
, T, I)^[e-5V- AXV2*K< • e-5V. •

4. 77ie topological situation

Throughout this section we will assume that / has a wandering interval /. We may
assume that / is a maximal wandering interval, i.e. not contained in a bigger
wandering interval. Since the forward orbit of / is disjoint and Kr has at most a
finite number of components, by possibly replacing / by some iterate of / we can
assume that / ' ( / ) n K, = 0 . We want to show that there exist iterates qn and
neighbourhoods Tn of / such that f» is contracting on / but strongly expanding
on Tn\I. In order to obtain a contradiction from this we need to show we can
control the non-linearity of f"\ Tn. From the previous section it follows that this

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 543

can be done if the intervals Tn, f(Tn),... ,fn(Tn) are more or less disjoint. This
disjointness will follow from the fact that / preserves ordering on S\ The proofs
of the results in this section can be found in section V of [He].
In fact, because / is order-preserving and has no periodic points, there exists an
order-preserving and continuous map h:S'-*S' such that h°f=Rp°h, where
P~p(f) and Rp is the rigid rotation over p (see [He]). In particular, the order of
points in an orbit of/ is the same as the order of points in an orbit of Rp. Since
we keep p fixed in this section, we write R~ Rp. Take some x e S1 and inductively
define q(n) to be the time of the nth closest approach:
q(0) = 1;
q(n +1) = min {/ e N; i > q(n), d(x, Rj(x)) < d(x, Rqin)(x))}.
Remark. These numbers q(n) are independent of x (they only depend on p(/)).
PROPOSITION 4.1. d(I,fqln)(I))^0 as |«|^oo ;
In the next two Propositions we will describe the way the sequence fin\l) is
ordered in S1.
PROPOSITION 4.2. The intervals {fi2n)(I); n >0} approach I from the right and
qi2 +l
{f " \l)\ M>0} approach I from the left (or vice versa).
From Propositions 4.1 and 4.2 we know that the sequence {fi2n>(I)}, n^O,
accumulates arbitrarily close to one side of /. In the next Proposition the order of
the intervals {/'(/); q(2n)< /< q(2n + 2)} is described.
Let Q(2n) be the set of integers teN, such that q(2n)< t<q(2n+2) and such
t h a t / ' ( / ) is between/« (2 " +2l (/) and/" (2n) (7).
PROPOSITION 4.3. te Q(2n) if and only if there exists ief^J such that t is of the form
t = q(2n) + iq(2n + l) andq(2n)< t<q(2n + 2). In particular, there exists a(2n + 2) e
N such that q(2n+2) = q(2n) + a(2n + 2) • q(2n + l).
PROPOSITION 4.4. Let Tn be the smallest interval containing f~qin)(I), I andfq{n\l).
Then every point in S1 is an at most three intervals of the sequence

5. Conclusion of the proof of Theorem C for maps fe si

PROPOSITION. 5.1. Let f&st and assume that I is a wandering interval off. There
exists n0 < oo such that if n> n0 then
I /-i/(2n)+a(2n+2)-</(2n+l)/ r^l — l /"'' I 2 " + 2 'C/^I>-I f1^2n+2)~1^2"+lUf\\ f< 1\

Proof Let Tn be the smallest interval containing/-" (2 " +2 '(7), / and/"" <2 " +1) (7).
Since Tn a [/« (2n+ "(/), /-* < 2 " + l l (/)] One has from Proposition 4.4 that each
point of S' is contained in at most three of the intervals Tn, f(Tn),
f2(Tn),...,f"an+2)-[(Tn). Moreover from Proposition 4.1 the length of the two
components Ln and Rn of Tn\I tends to zero as n -> oo. Hence we can apply Theorem

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 544 M. Martens et al.

3.8 and get that there exist Bo>0 and «0<oo such that for all n>n0,
B(/* (2tl+2) , Tn, I) > Bo- Since f"a"+2)(Rn) 3/^"+2>-</<2"+o(/) o n e g e t s
(2 +2)
|run+2,(RJ| j llnl|r - ( T n ) | |L n ||i? n |
2 (2n+2)
Bo • \I\ ' "'' "' |/" (L n )l " Bo • |/| 3 '
For large n the last term tends to 0 and since q(2n + 2) = q{2n) + a(2n + 2) • q(2n +1),
the Proposition follows. •
PROPOSITION 5.2. Let fe si and I be a wandering interval off. Then there exists
sequences {nk} and {ink} with 0< ink <a(2nk+2) such that for i= ink,

(/n particular, iffe si has a wandering interval then there are infinitely many integers
n > 0 SMC/J f/iaf a(2n)>l.)
Proof. For M > n0 (see Proposition 5.1) we have the following. Suppose that for some
n>n0 we cannot find an integer 0<i<a(2n + 2) such that (5.2) is satisfied. Then
eq. (5.1) implies that for all 0< z<a(2n + 2) one has

Hence, because q(2n + 2) = q(2n) + a(2« + 2) • ^(2« + l),

|/* ( 2 - + 2 ) (/)|>|/" 2 n ) (/)|. (5.3)
But since f'(I) are mutually disjoint, Xi |/'(^)l — 1- So there must exist a sequence
nk-*<x> such that (5.3) does not hold and for each of these nk there exists 0 < i <
0(2^ + 2) such that (5.2) holds. •
Now we can finish the proof of Theorem C.
The proof of Theorem C for maps fe si. Let nk and ink be as in Proposition 5.2,
q(n) = q(2nk) + ink • q(2nk + l) and Kn the smallest interval containing/<"2"l + "(7),
I and / " ( 2 ^ + l l (/). As before let Ln, Rn be the components of Kn\I. Proposition
4.4 guarantees the disjointness properties needed in order to apply Theorem 3.8 for
f>{n)\Kn. Fromeq. (5.2) and since B(f"(n), Kn, / ) > Bo,
\Ln\\Rn\y\Ln\\Rn\ r ir in) (/)i \f"in)(D\ \rin\n\ \rin\n\
|7|2 |7|2 L i / ^ d u ir ( n > (i?)i i/" n ) (L)i ir'-uj

Since \Ln\, \Rn\^0 as /c->oo we get a contradiction. •

Remark. The difference with the case Yoccoz considers (where there are no intervals
where/ is constant) is that he can use the operator Bo. This allows him to choose
Ln and Rn to be intervals in the backward orbit of /. In this way he can compare
\f(n){Ln)\ and \Rn\. This makes his proof easier.

6. The proof of Theorem C for maps in si'

The result in this section is a version of a result from unpublished work of C.
Gutierrez, A. Lins and W. de Melo.
Assume that / e si'. This means that / has degree one, is everywhere increasing
except on one interval Kr where it is constant and f\(S'-Kr) is C\ Moreover

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 545

outside every neighbourhood of K, is log Df of bounded variation, at the boundary

points of Kj = [a, b] the (one-sided) derivative of/ is infinite and there exists e > 0
such that log Df\[a - e, a) and log Df\ (b, b + e] are monotone.
Let A = S'\LUo.r'(*,)•
THEOREM 6.1. Let f&si'. Assume that the rotation number of f is irrational. Then
there exists NeN such that for all x e A,
DfN(x)>l.
This Theorem implies Theorem C for maps in si':
COROLLARY./e si' (with irrational rotation number) cannot have wandering intervals.
Proof. This immediately follows from the expansion of fN | A and the disjointness
of wandering intervals. •

The most important step in proving Theorem 6.1 is the following Proposition.
PROPOSITION 6.2. Letfe si'. Assume that the rotation number off is irrational. Then
for each x e A there exists N e N such that
DfN(x)>l.
Proof. Choose x e A , i.e., f'(x)£ K, for all i > 0 . Let a = -p{f). We first consider
the rotation Ra : S' -» S1. From [He] (V.8.4) we get that for each y e S1 and each n e N,

*U Ri([y,RHy)~\)L)''u Ri(\.y,Rl"+'(y)]) = S\ (6.1)

J=0 7=0
where the interiors of the above intervals are pairwise disjoint. Using the semi-
conjugacy between/" 1 and Ra and applying (6.1) t o / ' we get that x is contained
in one of the two unions corresponding to those in (6.1). In particular for all keN,
there exist n > k and jeN such that
x e / - ' ( U), {int (U), int ( / ( [ / ) ) , . . . , int (/"'(I/))} are pairwise disjoint, (6.2)
where U is the smallest interval containing Kf and f~q"(Kr). Because
x&\Ji5:af~'{Kj:) one has (for fixed x) thatj-^oo as fc-»oo. Let e > 0 be the number
from the definition of the class of maps sd' and let U be a e neighbourhood of Kt.
Let
V = Var (log Df\ IT), A = max (log Df\ Ur), B = min (log Df\ Uc).
We may assume that Df\(U\Kf)>l (and therefore log Df\(U\K,)>0). Further-
more since K,, f'(Kr), f~2(Kt),... is disjoint we may assume that k € I^J is so big
that

Now choose yef~J(Kf) such that Dfi(y) = \Kf\/\f''(Kl)\. Let T be the smallest
interval containing x and y. Notice that xef~'(U) and since U => K, and ye f~'(K,)
one has T<=/~'(t/). Moreover, since xi\^}ixUf'(Kr) the interval T is a proper
subset off'(U). From (6.2) this gives that
{TJ(T),... ,/'•(T)} is pairwise disjoint. (6.3)

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 546 M. Martens et al.

This disjointness and the fact that / is orientation-preserving then implies that the
points {x,/(x),... ,fJ(x)} and {y,f(y),... ,f'(y)} are alternatively distributed over
S1 (more precisely, between each two points of the first set there is a point of the
second set and vice versa). Choose the usual covering R -> S1 = K mod \,t-*t mod 1,
and for a point zeSx let z be one of the points in R covering ze S\ Write the sets
{ / ' ( * ) ; i = 0 , . . . , y } = {x, ;/• = ( ) , . . . , . / } ,

{fi(y);i =o , - - - J } ={y,\i = o , . . . j } ,
in such a way that there are points x,, pt, a, b covering x;, yh a, b such that
a<yo=fl(y)<b<xo<yl<xl<y2<- • •<pJ<xJ<a + \.
Choose 0< 5< r<j such that
ys<b + e<ys+l<- • -yr<a + \-e<yr+u
Let N=j+1.

log Df"(x)=Z log Df(x,)

= t log Df(x,)- I lo
i=0 i=l

Note that there are j+l of the form logD/(Xi) and only j terms of the form
logD/(j>,). Now we will rewrite (6.6) by pairing the term logD/(x,) either with
log Df(yi) or with log D/(>>,+,). Indeed, notice that
a<yo=fJ(y)<b<xo<yl<xx<y2<- • • <x,_, <y, < b + e
<Ps+i<x,+i<y!i+2<- • •

y f + 1 <x r + 1 <- • •<pJ<Xj<d + \. (6.5)

Using that log Df is decreasing on (b, b + e] and increasing on [a — e, a) and using
(6.4), one gets

logD/ N (x)> I logD/(x,)- Z log Df(y,) +

i=0 i=l

= V [log D/(x,) - log D/( yi+,)] + log D/(x.v) - log D/( yx+,)

(6.6)
+ logD/(xf)+ i [log Df(xi)-\og Df(yi)]+

>0 + \ogDf(xs)-\ogDf(ys+l)+

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 547

Now, distinguishing between x, < b + e or xv s: b + e, one has

log Df(xs)-\ogDf(ys+l)>min(0-A, B-A)>-\B\-A (6.7a)
and similarly
logD/U_,)-logD/(>' r ) + l o g D / U ) > m i n ( B - A + B , B - A + 0)>-2|B|-A.
(6.7b)
Using (6.7) in (6.6) and the definition of V gives
\og DfN(x)>0+(-\B\- A)- V+(-2\B\-A)+ V + 2A + 3|B|>0.
This finishes the proof of the Proposition.
Proof of Theorem 6.1. Let A' = S ' \ L U o / ~ ' int (Kf). We claim that for all xeA'
there exists an interval U of x and a number NeP*J such that DfN(y)> 1 for all
ye Un A. Indeed if x e A this follows from Proposition 6.2, since fN is C 1 in a
neighbourhood of x If x e A'\A then there exists « > 0 such that f"(x)edKf. Let
N = n + 1. Since DfN(y)-*<x> as .ye A and y-»x the claim also holds for x e A'\A.
Since A' is compact one can cover this set with a finite number of these neighbour-
hoods Uj as above such that for each i there exists N, such that DfN' | (t/, n A) > 1.
Let

p, = min {min {log DfN< | (t/, n A)}}> 0,

p2= min {min{log D / ' | A}}.

On,ls.N'

Since p, >0 we can choose N so that

N-N'
Pi" N, +P2>0.
Now we will show that DfN(x)> 1 for all x e A. So choose x e A. Then there exists
a sequence i, such that x e t/,,, / N ' . ( x ) e uh, fN;+N^(x)e t/,3, / N '. + N ' 2 + N . 3 (x)e
t/j,, Letfc,(x,N) be the number of times [/, appears in the sequence X , / N M ( X ) ,
/ ' V ' 2 ( x ) , . . . , / ( V , + N 2 + +yvw(x) where / is the largest number such that N,,+ N,2 +
N

• • • + Nh < N -1. By definition of N',

N - (N,, + 7V,2 + • • • + Nh) < N'. (6.7)
Now write

N-N'
2>p- ^ , +p2>l.

Here the second inequality holds because of (6.7) and because of the definition of
p 2 . This finishes the proof of Theorem 6.1. •

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 548 M. Martens et al.

7. The proof of Theorems A and B

Consider the class 38 of Cx Cherry vector fields on T2 such that at each singularity
p, div (X)(p) < 0. Furthermore, let ffl' be the class of Cx Cherry vector fields X on
T2 with precisely one hyperbolic sink and one saddle-point and such that X has
positive divergence at this saddle-point. We will denote the flow through a point x
by t-* X,(x). Let Sing (X) be the set of singularities of X.
PROPOSITION 7.1. Let X eSftu 5&'. Then there exists a closed Cx curve 2 on
2
T \Sing (X) without self-intersections and with the following properties
(a) £ is everywhere transversal to X;
(b) 2, is not retractable to a point.
Proof. It is enough to show that there exists a recurrent orbit y which is non-trivial
(i.e. not equal to a point or a closed curve), see for example page 144 of [PaMe].
Because by assumption / has no periodic orbits, the theorem of Denjoy implies
that if X has no singularites then every orbit of X is recurrent and we are done.
So assume that X has at least one singularity. Since the Euler characteric of T2 is
zero, this implies that X has at least one saddle-point. Let us call a stable separatrix
a component of Ws(p)\{p} where p is some saddle-point. We claim that there exists
a stable separatrix y such that the a-limit set of y, a(y), contains y. Indeed, take
a stable separatrix y0. Because of the Denjoy-Schwartz theorem a(y0) must contain
at least one singularity p. Since all singularities of X are sinks or saddles, p is a
saddle. Let y, be the stable separatrix of p. Because X has no saddle-connections,
«(To) ^ yi • By induction we define in this way a sequence of stable separatrices
such that «(%,) => yn+1 for all n >0. Since there are only finitely many separatrices
there exist n > 0 , m > 0 such that yn = yn+m. One has that
a(yn) => « ( r » + i ) = > • • • = > a ( y n + m - i ) => y n + m = yn.

Hence yn is a non-trivial recurrent orbit. As we remarked above this implies the

existence of a transversal circle. •
Let X e 58 u 38' and let 2 be the closed transversal to X on T2 from the previous
Proposition. Notice that T 2 \2 is an annulus I x ( 0 , 1 ) and we can write T2 =
2x[0, l ] / ~ , where (s,0)~(s, 1). Consider X as a flow on T7' = 2 x [0,1 ] where we
identify Sx{0} and 1x{l}. Since X has no sources it follows that
(c) for every x e 1 x {0} which is not contained in the stable manifold of one of
the saddles or one of the sinks, there exists a / > 0 such that X,(x)e2x{l};
(d) there exists at least one xe1x{0} such that X,(x)e1x{l} for some f >0.
Now let 2 and 2' be two closed curves transversal to X. Denote the points x e X
such that there exists / > 0 with X,(x) e 2' for some t > 0 by 2 0 . For x e l0, let t{x)
be the minimal / > 0 such that X,(x)e1' and define the map/:2 0 -»2' by/(x) =
X,(v,(x). This map is called the transition map between £ and 2'.
Now let 2' = 2 be the section from Proposition 7.1 and take the transition / from
2 to 2'. This map is called the return-map to 2.
Since orbits of X cannot intersect, / is order preserving. Let / be a component
of (S\2()) x {0}. Take x e dl. Since the basin of a sink is open, x must be contained

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 549

in the stable manifold of some saddle-point p e S x ( 0 , l ) . Since pedl and X has

no saddle-connections, W(p) intersects 2 x {1} in some point v and for every point
in ue(2\/)x{0}, near dl there exists f > 0 such that X,(u) intersects I x { l } near
w. (This can be seen by considering the backward orbits of X | (2 x [0,1]) of points
in I x { l } near v intersecting 2x{0}. This set consists of a neighbourhood of dl in
S.) In particular \\mx^aIxel f{x) consists of one single point and we can define /
on components of 2 \ 2 0 to be constant. From the smooth dependence on initial
condition,/is then everywhere continuous, and as smooth as the vector field outside
boundary points of S\2 () . From Theorem 7.1 in the Appendix,/ also satisfies the
non-flatness conditions from the introduction. It follows t h a t / e .si
Remark. Let X be the section from Proposition 7.1, and/:2 0 -»S its transition map.
Now T2\L is an annulus A, see figure 7.1. It is easy to see that the orbits through
boundary points of S o go directly to a saddle without leaving the annulus. From
this one gets that S\S 0 consits of finitely many components. Let / <= 2 be one of
the components.
(i) / is not equal to a point. Indeed if / consists of one point then this point
goes to some saddle-point and both unstable separatrices of this saddle-point
intersect 2. Let D c A be the region bounded by 2 and these two unstable
separatrices, see figure 7.1. But since X has no periodic orbits and no sources, this
contradicts the Poincare-Bendixon theorem.

FIGURE 7.1. This situation cannot occur for X e 38 u 3)'.

(ii) Let / = [a, ft]. Ifp is the saddle-point such that a e W(p) then alsofte Ws(p).
Indeed, let y be the stable separatrix of p which does not contain a. From the
assumption on X and Poincare-Bendixon theorem it follows again that y cannot
be contained in the annulus. So let c be the intersection of the closure of y n A and
2. It is easy to see that c =ft.From all this it follows that the situation is as drawn
in figure 7.2.
Now we are in the position to prove Theorems A and B.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 550 M. Martens et al.

FIGURE 7.2. A typical situation for i e 9 u 3)'.

Proof of Theorem A. Since to a vector field X we associated a m a p / e si, and since

X has no periodic orbits and has no saddle-connections, p(f) is irrational.
But if p ( / ) is irrational, we can apply the Corollary to Theorem C t o / : S ' = I ^ 2 s
S1 and the orbit of every point in {x;/ n (x)e2 0 ,Vn>0} is dense in this set. But
now notice that every point in 2 \ 2 0 is contained in the basin of a periodic attractor
or in the stable manifold of a saddle-point. The statement follows. •
Proof of Theorem B. Take X as above and let 2 be the section from Proposition
7.1. There exists a neighbourhood U of X such that for each Ye U, 2 is also a
section of V. Associate fY '• 2 -» 2 to K as above. From the construction fY depends
continuously on Ye U. If there exists a conjugacy between fx, fY :2-»2 then one
can construct an equivalence between X and Y in the same way as in the proof of
the structural stability of Morse-Smale flows on surfaces, see for example, [PaMe].

8. Appendix: The transition map near singularities

Let X be a Cx vectorfieldon R2 which has a hyperbolic singularity at 0 of saddle-type
with eigenvalues A>0>/u.. Let />, and p2 be points in Wv(0) respectively W{0).
Furthermore let 2; be a C 2 curve through p, which is transversal to X. If we choose
2, sufficiently small, then for every x in one of the components of 2, -{/>,} there
exists / > 0 such that X,(x)e2 2 . Call this component 2^ and let l(x)eR be the
smallest number so that X,{x>(x)e12 and define T:I.'I->12 by

We call this map the transition map.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 551

Since 2, are C 2 curves, it is a well defined notion to say that maps <f> : S , ->R and
«/r :!R ->S 2 have a derivative which has bounded variation.
In this section we want to show that T : 2 , -»2 2 is equal to a map (f>a, up to maps
whose derivatives have bounded variations. Here a = \/x\/X > 0 , and <f>a is defined
by <t>a = ±\z\a.
THEOREM 8.1. Let X be a Cx vector field on U2 which has a hyperbolic singularity at
0 of saddle-type with eigenvalues A > 0 > //,. Lef a = \fj.\/X. There exist maps 4>, i/»:IR+-»
R+ which are C\ such that D<f> and Difj have bounded variation and such that the
map T from above is of the form
T(x) = <p<>(t>aoil>(x). (8.1)
Proof.
Step 1. Let us first show that (8.1) is true if X is linear near 0.
If 2, are the straight lines {x = 1}, {y = 1} and 7*{x = 1M,, = )( the transition map from
{x = 1} and {y = 1}, then (8.1) follows from explicit integration (and choosing the
natural parametrisations of {x = 1}, {y = 1}. If 2, are different C2 curves (8.1) still
holds. In fact consider the transition map T, fromS, to{x =1}, resp. T2:{y = l}-ȣ 2 .
Then T S| v 2 = T2° 7*(x«i},(, = i}° Tx. Since X is linear (and in particular theflowmap
is C2) it follows from the implicit function theorem that the maps Tt are C2. This
last argument also shows that if (8.1) is true for one choice of C2 curves 2,, then
it is also true for any other choice of C2 curves 2, as above.
Step 2. X is C2 linearisable if |A|g{|/*|, 2\fi\, ^ | } .
According to the Linearisation Theorem of Sternberg (see [Ster], [Bel], [BD] or
for example Theorem 4 in [Sto]) there exists a C2 coordinate system <f> near 0 such
that <}>^X is linear if
|A|*|M|,2|MI,5|MI- (8-2)
That is, in this case

<bj,X(x, v) = A • x hu •y —.
* dx dy
Using step 1 we are finished if (8.2) holds.
Step 3. A normal form for X when |/t| = 2|A| or |/x| = |A|.
Let us now see what happens if (8.2) fails. We deal with the cases that |/u.| = 2|A|
and that |/u.| = |A|. The case that |A| = 2|/t| is similar. By considering a multiple of
the vector field we may assume that X is of the form
) = x-—-2y—+o\x\2. (8.3a)
dx dy
^ y ^ \ x \ . (8.3b)
dx dy
Here x = (x,y).
LEMMA 8.2. Let X be a Cx vector field.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 552 M. Martens et al.

(a) IfX is of the form (8.3a), then there exists a C°° change of coordinates <f> such
that

<t>*(X) = x(l + ax2y) — -2y{\ + bx2y) — + o|x| 6 . (8.4a)

dx dy
(b) IfX is of the form (8.3b), then there exists a C* change of coordinates <f> such
that

* y(l bxy)o\x\4. (8.4b)

aX ay
Proof. Let X, be the vector field on R2 which has the same 1-jet in 0 as X and whose
coefficients are linear. Let [X,, -]h:Hh^Hh be the linear map which assigns to
each homogeneous vector field Ye Hh of degree h the Lie-product [X,, Y] (which
is again in Hh). For each h e N choose Gh so that Hh = Im ([X, ,-]h) + Gh. Accord-
ing to Theorem 2.1 of Takens [Ta] there exists for each / < oo, a Cx diffeomorphism
* such that $ # ( X ) is of the form

where g, e G\ i = 2, 3 , . . . , / .
Let us determine Im ([X,, -];,). Taking /3 G {1, 2},

dx dy' dx] dx'

dx dy dy] dy
h
For p = 2, we can choose G = 0 for h < 6, h ^ 4, and

Case (a) follows.

For P = 1, we can choose G'1 = 0 for h < 4, /i 5^ 3, and
d d
^G =\x'-y
I •> ,xy -
\ dx dy
Case (b) follows and the proof of the Lemma is finished. •
Step 4. A C' + 'bounded variation' linearisation of X.
From step 1, we may assume that X is of the form

X = x(\ + a- x2y)—-2y{\ + b • x2y) —+ o\x\k; (8.5a)

dx dy
or
y)— -y(l + b- x- y)— +o\x\\ (8.5b)
dx dy

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 On Cherry flows 553

Step 4A. Let us first consider the case of (8.5a). Consider the C' coordinate
transformation ^(x, y) = (x, y) defined by:
=
X X — CLX y l o f i Ixj

y = y-bx2y2log\y\.
Then {x = l} = {x = l} and {y = \} = {y = \}. Moreover, the transformation between
y and y coordinates on the line {x = 1} = { X = 1}, {(x, y); x= 1}-»{(X, y); x = 1}, is
equal to y*-^y = y-by2 log \y\, which has a derivative of bounded variation on
[-1,1]. Similarly y>-*y is also C 1 and has a derivative of bounded variation. The
same holds for the coordinate transformation {(x,y);y = l}-*{(x,y);y = l}. Since
T{x=,)-<v=|| is equal to 7*(.$ = i(i<,-=1) composed with y>-*y on the left and with x<-»x
on the right, it suffices therefore to prove the result for W*X. Let us calculate y*X.

'dx) \ 'dx) dx \ 'dy) dy

= x(l + ax y)( 1 - a3x2>> log |x| - ax2y) -2y{\ + bx2y)(-ax3 log |x|) + o|x|6
2

= x - axyy log |x| + o|x|6+ o|x|6 = x + o|x|6,

and

A ( A U )
dy) \ dx) dx \ dy) dy
= x( 1 + ax2y)(-2bxy2 log \y\) -2y(l + bx2y)(\ -2bx2y log \y\ - bx2y)
+ o\x\6
= -2y + 2bx2y2 log |^| + o(|x|6) + o|x|6 = - 2 y + o\x\6.
Here x = (x,y). Hence

^ XxA:2yz+o\x\.
y Ul
* dx dy
Using the notation of Theorem 2 of [Sto] we take <J = 2, r = 6, and it follows that
there exists a transformation of Holder class C\ where s = c r ( r - l ) / l + o- = 10/3
linearising V%X at 0. (We use here the vector field analogue of these linearisation
results for diffeomorphisms, see the end of the introduction in [Sto].) Using step
1, the result follows for (8.5a).
Step4B. Now we consider the case of (8.5b). Consider the C1 coordinate transforma-
tion ¥(x, y) = (x, y) defined by:
x = x — ax2y log |x|,
y = y-bxy2\og\y\.
Then as before {(x, y); x = 1} = {(x, v); x = 1} and {(x, y); y = 1} = {(x, y); y = l} and
the coordinate transformation y>-*y and x>-+x on these curves are C' and have
derivatives which have bounded variation. So as before it suffices to prove the result

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.
 554 M. Martens et al.

for ^ X . Let us calculate ^ X .

dx) \ dx) dx \ dy) dy

= x(l + axy){\ -alxy log |x| -axy) -y{\ + bxy)(-ax2 log |x|) + o\x\
= x~ax3y log |x| + o|x|4+ o|x|4 = x + o|x|4,
and

'dy) \ dx) dx \ 'dy/ dy

= x(l + axy)(-by2 log \y\)-y(l + bxy)(l -2bxy log \y\ - bxy) + o\x\4

Hence

Using the notation of Theorem 2 of [Sto] we take cr= 1, r = 4 and it follows that
there exists a transformation of Holder class C\ where s = (r + cr)/(l + cr)=§
linearising ^ X at 0. Using step 1, the result follows.
Since we had reduced everything to these two cases we arefinishedwith the proof
of Theorem 8.1. •
6
Remark. All the results in this section are valid if X is C .

REFERENCES
[Bel] G. R. Belitskii. Equivalence and normal forms of germs of smooth mappings. Russian Math.
Surveys 33 (1978), 107-177.
[BD] P. Bonckaert and F. Dumoriter. On a linearization theorem of Sternberg for germs of
diffeomorphisms. Math. Z. 185 (1984), 115-135.
[Ch] T. M. Cherry. Analytic quasi-periodic curves of discontinuous type on the torus. Proc. London
Math. Soc. 44 (1938), 175-215.
[Den] A. Denjoy. Sur les courbes definies par les equations differentielles a la surface du tore. J. de
Math. Pures et Appl. 11 (1932), 333-375.
[Ha] G. Hall. A C°° Denjoy counter example. Ergod. Th. & Dynam. Sys. 1 (1981), 261-272.
[He] M. R. Herman. Sur la conjugaison differentiable des diffeomorphismes du cercle a des rotations.
Publ. Math. I.H.E.S. 49 (1979), 5-233.
[MSI] W. de Melo and S. van Strien. A structure theorem in one dimensional dynamics. Ann. Math.
189 (1989), 519-546.
[MS2] W. de Melo and S. van Strien. One-dimensional dynamics: The Schwarzian derivative and
beyond. Bull. Amer. Math. Soc. 18 (1988), 159-162.
[PaMe] J. Palis and W. de Melo. Geometric Theory of Dynamical Systems, an Introduction. Springer
Verlag: New York, 1982.
[Sch] A. J. Schwartz. A generalization of a Poincare-Bendixson Theorem to closed two dimensional
manifolds. Amer. J. Math. 85 (1963), 453-458.
[Ster] S. Sternberg. On the structure of local homeomorphisms in Euclidean n-space, II. Amer. J.
Math. 80(1958), 623-631.
[Sto] D. Stowe. Linearisation in two dimensions. J. Diff. Eq. 63 (1986) 183-226.
[Ta] F. Takens. Singularities of vector fields. Publ. Math. I.H.E.S. 43 (1974), 47-100.
[Y] J. C. Yoccoz. II n'y a pas de contre-exemples de Denjoy analytique. C.R. Acad. Sci. Paris 298,
ser. 1, no. 7 (1984), 141-144.

Downloaded from https://www.cambridge.org/core. 25 Apr 2022 at 21:37:05, subject to the Cambridge Core terms of use.