BF 02392981

Acta Math.
, 168 (1992), 273-318
Julia-Fatou-Sullivan theory
for real one-dimensional dynamics
by
M. MARTENS(1) w. DE MELO and S. VAN STRIEN
Technical University of Delft IMPA Technical University of Delft

Delft, The Netherlands Rio de Janeiro, Brazil Delft, The Netherlands
Introduction
Our aim is to show that the Julia-Fatou--Sullivan structure theory for the dynamics of
rational maps is also valid for smooth endomorphisms of the circle (and of the interval)
under extremely mild smoothness and non-flatness conditions.
In order to stress the similarity between real and complex one-dimensional dynam-
ics let us recall the main results from the Fatou-Julia-Sullivan theory. If f is a rational
map then there is a dynamical decomposition of the Riemann sphere into the disjoint
union of two totally invariant (i.e., both forward and backward invariant) sets
J(f), F(f). Here, F(f) is the domain of normality of the family of iterates o f f , and is
called the Fatou set. Its complement, which is called the Julia set off, is a compact set,
which contains all the complications of the dynamics off. The connected components
of the open set F(f) are mapped onto each other by f. Hence the orbit of a component
of F(f) is the union of some components of F(f). Julia proved in the beginning of the
century that if a component of F(f) is periodic and contains an attracting periodic point
then the orbit of this component must contain a critical point. Sullivan, in the remark-
able paper [Su], showed via quasi-conformal deformations that the components are
eventually periodic and fall into finitely many orbits.
Let N be either the circle S J or a compact interval of the real line and f." N---~N be a
smooth endomorphism. A critical point of f is a point where the derivative vanishes. A
critical point is non-flat if some (higher) derivative is non-zero. A critical point is an
inflection point if it has a neighbourhood where f is monotone. Otherwise it is called a
turning point. Assume f is not a homeomorphism ( i f f is a homeomorphism then these
maps correspond to degree + 1 rational maps of the Riemann sphere and the situation is
(1) Supported by N.W.O.

274 M. MARTENS, W. DE MELO AND S. VAN STRIEN
simple). I f f h a s turning points then we define the singular set off, Sing(f), as the union
of the set of turning points of f and the boundary points of N. If f has no turning points
thenfis a circle map of degree >t2 (or <~-2) and we define Sing(f) to be the set of fixed
points o f f . Finally we define the Julia set of f to be the a-limit set of Sing(f), i.e.,
x E a(Sing(f)) iff there exists y i ~ x and a sequence ni--~oo such that f~i(y i) E Sing(f).
Denote this set by J(f). It is forward invariant. Thus its complement, called the Fatou
set F(f), is backward invariant: f - l ( F ( f ) ) c F ( f ) . In general the Fatou set is not
forward invariant but if U is a connected component of F ( f ) which does not contain a
turning point t h e n f ( U ) is also a component of F(f).
MAIN THEOREM. Let f'. N - ~ N be a smooth map such that all its critical points are
non-fiat. Then:
(1) All the connected components o f F ( f ) are eventually periodic (i.e., eventually
mapped into a periodic component o f F ( f ) ) ;
(2) The number o f periodic components o f F ( f ) is finite.
Part (1) of the above theorem can be reformulated by stating that there are no
wandering intervals for such maps. By a wandering interval we mean an interval J such
that all forward iterates of J are disjoint and such that the a~-limit set of J is not a
periodic orbit. The first result in this direction was obtained by Denjoy in 1932, [D]. He
proved that a C 2 diffeomorphism of the circle does not have wandering intervals (more
precisely, his proof applies to all C t diffeomorphisms f: S1--~S such that l o g D f has
bounded variation). His proof relies on a detailed understanding of the dynamics of
rotations and on the control of the distortion of iterates of the map on intervals whose
iterates are all disjoint. Later, in 1963, A. Schwartz [Sc] gave a different proof of
Denjoy's result. His proof does not rely on precise dynamical properties but requires
that logDf is Lipschitz. For maps with critical points the techniques of Denjoy and
Schwartz cannot be used. In 1979, J. Guckenheimer was able to deal with critical
points in some special cases. He proved in [Gul] the non-existence of wandering
intervals for unimodal maps of the interval with negative Schwarzian derivative and no
inflection points, see also [Mi]. J. C. Yoccoz, [Y], proved the non-existence of
wandering intervals for C = homeomorphisms of the circle having only non-flat critical
points. He combines techniques of Denjoy away from the critical points with some
analytical estimates near the critical points which are related to the Schwarzian
derivative. In [MS1] the same result was proven for smooth unimodal maps (not
necessarily having negative Schwarzian derivative) with a non-flat critical point and
also for maps satisfying the so-called Misiurewicz condition. In [MS1] the main tool is
JULIA--FATOU--SULLIVAN THEORY FOR REAL ONE-DIMENSIONAL DYNAMICS 275
the control of the distortion of the cross-ratio under iterates. This control implies that
under some disjointness assumptions the diffeomorphic inverse branches of iterates of
a smooth map behaves very much like univalent holomorphic maps. This similarity is
clear from the minimum principle and the Koebe distortion principle for real maps, see
w2 and w4 of this paper. In 1988, Blokh and Lyubich proved in [L], [BL] the non-
existence of wandering intervals for smooth maps whose only critical points are turning
points. They introduced some very nice and powerful new topological tools generaliz-
ing those of [Gul] and used the analytical tools developed in [MS1]. Our proof of the
first part of the Main Theorem combines the analytical tools developed in [MS1] and
[MMMS] (which allow for inflection points) with an extension of the topological
ingredients of [L] and [BL]. In fact many of the ideas of w 1 6 7 in this paper are
simplifications and modifications of results contained in [L] and [BL]. We should note
however that for the proof in ILl and [BL] it is necessary that the maps are C 2 and that
our proof of part (1) of the Main Theorem also works for piecewise linear maps (see
w7). To be more specific our proof applies to all diffeomorphisms of the circle for which
Denjoy's results hold. So Theorem A is a natural extension of Denjoy's original ideas
to maps with critical points.
The first contribution to an analogue of part (2) of the Main Theorem for rational
maps on the Riemann sphere is due to Julia. He proved in [J] (see also [Fa]) that if the
orbit of a periodic domain contains an attracting point then it must also contain a
critical point. Hence the number of orbits of these periodic domains must be bounded
by the number of critical points. In [Si], Singer introduced for the first time the
Schwarzian derivative in one dimensional dynamics and proved the same result of Julia
for maps with negative Schwarzian derivative. Mafir, using estimates related to Denjoy
and Schwartz proved that (2) holds for maps of the circle without critical points.
Instead of (2) we will prove a still stronger result. We show that for each smooth map
satisfying the hypothesis of the Main Theorem there exists O>0 such that if p is a
periodic points of sufficiently high period n then [Dfn(p)[>~l+o. This last estimate is
new even for maps from the family x---~ax(1-x).
Let us finish this introduction by stating a corollary of the Main Theorem.
COROLLARY. (For the proof see [Me].) In the space ~ r o f C unimodal maps o f the
interval [ - 1 , 1] endowed with the C r topology, r>~3 the set of structurally stable maps is
open and dense.
The above corollary is the analogue of Mafir-Sad-Sullivan Theorem on the density

of structural stable rational maps on the Riemann sphere. Note however that it does not
follow that structurally stable maps satisfy the Axiom A condition.
COROLLARY. Let f : [ - 1 , I]--->[-1, 1] be a C 2 unimodal map which is infinitely

renormalisable and whose turning point is non-flat. Then f has an attracting Cantor set
and its basin contains intervals.
Proof. By Theorem B the turning point o f f cannot be approximated by periodic

attractors. Therefore after renormalizing a finite number of times, the new renormal-
ized map has no periodic attractors. So for this new map all points, except those which
are eventually mapped into periodic orbits, tend to the closure of the forward orbit of
the turning point. Q.E.D.
Many of D. Sullivan's ideas were an important source of inspiration for our work.
We are grateful to him for inviting the last two authors to C.U.N.Y. and his interest in
this work. Also we would like to thank the mathematics department of the University
of Warwick at which this paper was finished.
w1. Definitions and statement of the theorems

Let N be a smooth compact l-dimensional manifold (i.e., a finite union of closed
intervals or circles). We say that p 6 N is a periodic point of f of period n iffn(p)=p and
fi(p)4=p for 0<i<n. J c N is a periodic interval of period n if f f ( J ) c J and fi(J) NJ = ~ for
0<i<n. An interval J c N is a wandering interval for f i r (i)fi(j)tl f J ( j ) = o for all O<~i<j
and (ii) the w-limit of J, to(J)={x;3ni--~oo and yEJ such that ff'(y)--~x}, is not a
periodic orbit. (In fact, if it is a periodic orbit then it necessarily has to be a (possibly
one-sided) attracting periodic orbit.) If f is not a homeomorphism and has a finite
number of turning points then a component of J of the Fatou set of f is either eventually
periodic (i.e., eventually mapped into a periodic component of F ( f ) ) or a wandering
interval. If U is a periodic component of period n>0 of the Fatou set of f (i.e.
f n ( U ) c U), then either the orbit of U contains a critical point o f f or ffl U is monotone
a n d f ~ has a non-repelling fixed point in the closure of U. Therefore our Main Theorem
follows from the following two theorems.
THEOREM A. Let f'. N--+N be a C | map whose critical points are non-flat. Then f
has no wandering interval.
THEOREM B. Let f: N-->N be a C ~ map whose critical points are non-flat. Then
there exists no and p>0 such that for every periodic points o f f of period n>~noone has
[Ofn(p)[ ~ 1+ 0.
The assumptions in these theorems are satisfied for analytic maps f: N-->N. In fact
Theorem B implies that analytic maps can have at most a finite number of non-repelling
periodic orbits (unless f o r f2 is the identity map).
The non-flatness conditions in Theorems A and B cannot be dropped. Indeed if

one considers C ~ maps with a flat critical point then Theorems A and B are false. Such
maps may have wandering intervals and an infinite number of attractors. The first
example of such a map was given by Hall, [Ha], but see also [SI], [Me]. Moreover,
Theorem A and B are actually proved under much weaker smoothness assumptions.
Theorem A is for example also proved for general continuous piecewise linear maps (in
fact it is proved under the same smoothness conditions as the result of [D] for circle
diffeomorphism).
More precisely we will prove
THEOREM A'. Let f'. N-->N be in the clase M (defined presently). Then f has no
wandering interval.
Here M is the class of absolutely continuous maps f: N--->N such that the following
two conditions are satisfied:
(AI) There exists a finite set Ky such that for each neighbourhood of U of the set
Kf, the map N \ U g x ~ l o g ] D f ( x ) ] (which exists almost everywhere because f is
absolutely continuous) extends to a map of bounded variation on N \ U;
(A2) For each xoEKy there exists a~>l, a neighbourhood U(xo) of x0 and a
homeomorphism ~: U(xo)---~(- 1, 1) such that q~(x0)=0, ~, ~-~ are absolutely continuous,
(-1, 1 ) 9 x ~ logD~(x) (which exists almost everywhere) can be extended to a map of
bounded variation, and such that
f ( x ) = +ldp(x)la+f(xo), VxE U(xo).
This number a is called the order of XoE Kf.

Notice that condition (A2) gives a non-flatness condition at the set of critical
points. (If x0 is an inflection point we can also allow that f(x)=+ldp(x)l~'+f(Xo) on one
side of x0 and f(x)= +_l~(x)l~e+f(Xo) on the other side, where al, a2~>l need not be the
same. On the other hand, if x0 is a turning point then we need the order or flatness to be
the same on both sides; indeed, otherwise the involution r from Section 4 below would
not be Lipschitz.) This class M contains the class of C | maps with non-flat critical
points (and therefore Theorem A' implies Theorem A) and also for example the class of
continuous piecewise linear maps. For these piecewise linear maps even the non-
existence of wandering intervals whose orbits stays away from turning points was
previously unknown. The classical proof based on Denjoy-Schwartz requires log IDfl
to be Lipschitz, see [Sc], and also [CE], [Str2]. Our proof relies on a disjointness result
for backward iterates of certain intervals, see Theorem 6.4. In this sense our proof is
closer to the original ideas of Denjoy than the later developments of Schwartz.
Instead of Theorem B we will prove the following stronger result:
THEOREM B'. Let f: N--*N be in the clase ~ (defined presently). Then there exists
no and (~>0 such that for eoery periodic point o f f o f period n>~no one has
[Dfn(p)l >! l + o .
The class ~3 is somewhat smaller than the class M. Indeed, f E ~ if and only if:
(B1) f i s C2;
(B2) Let Kf be the set of critical points off. Then for every Xo E Kf, there exist a > 1
and a C 2 coordinate system q~: U(xo)--->(-1, 1) on a neighbourhood U(xo) of x0 such that
dp(xo)=O and
f(x) = +f~b(x)ta+f(xo), V x e U(xo).
Clearly ~ includes the class of C ~ maps with non-tiat critical points (and therefore
Theorem B' implies Theorem B). Moreover, it is not difficult to prove that condition
(B2) is satisfied if for each critical point c there exists k>~2 such t h a t f i s C T M at c and
Dfk(c):#O. It is not clear whether we can weaken the condition t h a t f i s Cz in the proof
of Theorem B'.
The numbers no and 0 from Theorem B obviously depend in an essential way on

the maps f. Indeed, there exists a sequence of analytic maps f~ converging to an
analytic map f such that fn has an attracting periodic orbit of period >n. (Take for
example f to be a quadratic map with an eventually periodic critical point.) Neverthe-
less, our proof of Theorem B gives more uniform estimates. More precisely, if ~ is a
compact family of maps in ~3 then there exists Q>0 and no E N such that iffE ~ a n d p is
a periodic point of period n>~no o f f then one of the following possibilities hold.
(i) p is an attracting periodic point whose immediate basin of attraction contains a
critical point;
JULIA--FATOU--SULLIVAN THEORY FOR REAL O N E - D I M E N S I O N A L DYNAMICS 279
(ii) p is in the boundary of the immediate basin of a periodic attractor which

attracts a critical point;
(iii) [Ofn(p)[>-I+0.
In particular the number of periodic orbits of maps in X of type (i) and (ii) is
bounded by the number of critical points.
Notation. We will use the following notation. Jn will denote fn(j). If I, J are
intervals in the same component of N, let [I, J] be the (smallest) convex hull o f / a n d J.
(Even if this component of N is equal to S 1 it will always be clear which interval we
mean. I f / a n d J are in different components then we define [I, J] =N.) (I, J] denotes the
set [I, J ] \ I . Similarly define [I, J) and (I, J). The Lebesgue measure of a measurable
set I=N is denoted by III.
2. Analytical estimates on diffeomorphic branches o f f n when f E ~:

the Koebe and the minimum principle
In Denjoy's theory for circle diffeomorphisms the main technical tool is the control of
the distortion of iterates of the map restricted to some interval under some disjointness
assumptions on the iterates of this interval. Here the distortion of a differentiable m a p f
on an interval T is defined as the maximal ratio of the absolute values of the derivative
in two different points. This number measures the non-linearity of the map. Another
way to present the same concept is to consider pairs of intervals L, R c T, intersecting at
a common boundary point, and the distortion of the ratio D(L, R) = ILI/IRt by the map f,
i.e., the number D(f, L, R)=D(f(L),f(R))/D(L, R). It is easy to see that the distortion of
a differentiable map f i n the interval N is bounded if and only if there is an upperbound
for D(f,L,R) for any pair of intervals L , R c T .
If a map f has critical points we cannot hope to get a bound for its non-linearity.
Hence, instead of the distortion of a pair of consecutive intervals, or three consecutive
points, we have to analyze the distortion of a more complicate configuration called the
cross-ratio of four points.
2.1. Definition. Let J e T be open and bounded intervals in N such that T \ J

consists of intervals L and R. Define the cross ratio of these intervals as
I11 ITI
D(T, J) - ILl IRI'
(where II t denotes the length of an interval I). If g: T ~ N is continuous and monotone
define
B(g, T, J) = D(g(T), g(J))

D(T, J)
We notice that if f "IT is monotone and continuous then

n--I
B ( f n, T, J) = N B(f'f~(T)'fi(J))"
i=O
Of course this cross-ratio is related to the hyperbolic metric. Indeed, let T be an

open and bounded interval on N. For x, y E T let
+ [LUJIIJURI = +Log(I+D(T,J))
Or(X, y) = Log ILl In[ 2
where J is the interval bounded by the points x, y. Then Pr is a mettle in T and the group
of isometries of this metric is exactly the group d~r of all M6bius transformations that
map T onto T. Furthermore, the group d~r acts transitively on T, namely, given x, y E T,
there exists an isometry ~ E d~t such that ~(x)--y. If we take T = ( - 1,1) and x, y ~ T then
pr(x, y) is exactly the hyperbolic distance between the two points (x, 0), (y, 0) in the
unit disc. Therefore we shall call Pr the hyperbolic metric of the interval T.
As is well known, holomorphic maps of the unit disc contract the hyperbolic
metric. Moreover holomorphic maps of the disc have many very powerful properties:
for example there is a universal bound for the non-linearity of these maps on a smaller
disc (this follows from the Koebe Lemma). We shall show that diffeomorphisms of the
interval T which contract the hyperbolic metric Or satisfy similar properties.
One way to check that a map contracts the hyperbolic metric on an interval is
through the Schwarzian derivative.
2.2. Definition. Let g: T--~N be a C3-map on the interval T=N. Then
Sg(x)= 3( )2
g'(x) 2 \g-~--~/
is called the Schwarzian derivative of g.
2.3. PROPOSITION. Let g: T-*N be a C3 diffeomorphism on the open interval T e N

such that for every x E T, Sg(x)<O. Let cl(I)cint(T). Then
B(g, T, I) > 1.
Furthermore if Sg(x)<.O, Vx E T, then B(g, T, I)>~1.

Proof. The proof is well known and can be found in for example [MS].
In particular a diffeomorphism g: T--->T' having negative Schwarzian derivative

expands the hyperbolic metric: Qr(g(x), g(y))>Qr(x, y) for all x, y E T and x * y . There-
fore the inverse of g contracts these metrics.
So our aim is to show that inverses of diffeomorphisms g: T--~T' with negative
Schwarzian derivative behave in many ways as holomorphic maps on a disc in the
complex plane.
Unfortunately the condition that a map has negative Schwarzian derivative is not
very natural: it is not preserved under coordinate changes, has no dynamic interpreta-
tion and it excludes a large class of maps.
Therefore we will not only consider maps with negative Schwarzian derivative but
maps with some smoothness properties. By studying the distortion of the cross-ratio
under iteration, we will first show in this section that the iterates of a smooth map
restricted to an interval where it is a diffeomorphism, does not contract the metric too
much provided some disjointness assumption on the iterates of the interval under
consideration. Then we will show that in this case one obtains analogues of the Koebe
and Maximum Principle for conformal maps (the analogy is with respect to the inverse
of the maps). These results will hold for m a p s f E ~. In w4 the case is considered t h a t f n
is a homeomorphism which may have critical points of inflection type and when f E ~t.
In the following theorem we give a lower bound for B ( f ~, T, J) i f f E ~, see also in
[MS1] or in [Str3], see also [Str2] and [MS2] (where also a different cross-ratio is
considered).
2.4. THEOREM. Let f E ~. Then there exists a bounded continuous function

a: [0, ~)---~R+ such that tr(t)--~O as t--->Owith the following property. I f T is an interval
such that fruiT is diffeomorphism then for any interval J~-T with cl(J)cint(T):
m-1
(2.1) B ( f m, T, J ) ~ e x p { - a ( r ) - ~ Ifi(T)l}.
i=O
Here lr=maxi=o..... m-l lfi(r)l 9
Proof. Since B ( f n, T, J)=l-l~__201B(f,fi(T),fi(J)), it suffices to show that there exists

a constant CoE(0,~) and an increasing continuous function tr:[0,~)---~[0, C0) with
lim/__,0tr(t) =0 such that
B(f, T, J) ~>exp{-o(ITI). Irl)
19-928283 Acta Mathematica 168. Imprim6 le 24 avril 1992

for all intervals J = T c N such that Df(x)*O for all xE T. It is enough to prove that
(2.2) B(f, ~r,j ) - 1/> -o(l~q). ITI

for some (other) function tr as above. Let V~-cl(V)c U be a neighbourhood of Kf such
that each component U' of U contains a unique critical point and such that f is of the
formf(x)=+l~(x)l~+f(xo), for all xE U' where a~>l and ~: U'--~(- I, I) is a diffeomor-
phism.
Case 1. T c N \ V. In this case write T=[a, d] and M=[b, c]. Then

f(c)-f(b) . f(d)-f(a) _ f(b)-f(a) . f(d)-f(c)
c-b d-a b-a d-c
B(f, T, J ) - 1 =
f(b)-f(a) . f(d)-f(c)
b-a d-c
>-1 f(c)-f(b) f(d)-f(a) f(b)-f(a).f(d)-f(c),
~-~" ~-b " d-a - b-a d-c
where K=inf{ [Df(x)l; x E N \ V}. Writing
f ( a + x) = f(a) +It(a, x). x
one gets
f ( d ) - f ( a ) _ It(a, d - a ) ,
d-a
f ( b ) - f ( a ) _ It(a, b - a ) .
b-a
Also
f ( c ) - f ( b ) = It(a, c - a ) . ( c - a ) - I t ( a , b - a ) . ( - c + b - a + c )
= [It(a, c - a ) - # ( a , b-a)] (c-a)+It(a, b-a)" ( c - b )
and therefore
f ( c ) - f ( b ) = It(a, b-a)-~ It(a' c - a ) - I t ( a , b - a ) . ( c - a )

c-b c-b
and similarly
f ( d ) - f ( c ) _ It(a, d - a ) q it(a, d - a ) - g ( a , c - a ) . ( c - a ) .
d-c d-c
Hence
B(f, T, J ) - I >I--=.
- 1 ix(a, d - a ) ix(a, b-a)+ ix(a, c-a)-ix(a, b-a) . (c_a) ~
K~ \ c-b
-ix(a, b-a). (ix(a, d-a)-~ ix(a' d-a)-ix(a' c-a) "
= -_J_l2.K ic_al, ix(a, d - a ) . Ix(a, c-a)-ix(a,c_b b-a)
-ix(a, b-a) 9 ix(a, d-a)-ix(a,d_c c-a) .
Hence
J ) - 1 ~'--~
B(f, T,ITI - 1 ix(a, d - a ) . ix(a, c-a)-ix(a,c_b b-a)
-ix(a, b - a ) . ix(a, d-a)-ix(a, c-a) .

d-c
Since f is C 2
a(t) = sup ix(a, d-a) 9 ix(a, c-a) -ix(a, b-a) -ix(a, b-a) 9 ix(a, d-a)-ix(a, c-a) [
ITI<~t c-b d-c I
is a bounded increasing function with o(t)-~O as t--,0. From this the result follows.
Case 2. T~ U. Since a ~ I, the map ~ ( x ) = x ~ has Schwarzian derivative <~0. Then

Proposition 2.3 implies B(qS~,~(T), ~(U))~>I. Therefore
B(f, T, U)= B(dp,, dp(T), dp(U)).B(dp, T, U)>>-B(dp, T, U).
Since q~ is C 2 and inf{lDep(x)l;x E T} is bounded from below, B(r T, U)~>I-o(ITI)-ITI

follows as in Case 1. Combining this gives (2.2).
Case 3. If T contains a component of U \ V then (B(f, T, M ) - 1)/IT I is bounded

from below since ITI is bounded from below. Combining Cases 1, 2 and 3 this finishes
the proof of Theorem 2.4. Q.E.D.
Remark. We should emphasise that it is really essential in this theorem t h a t f i s C 2.

It is not sufficient that f ' is Lipschitz. This is in contrast with the usual bounded non-
linearity results in the theory of A. Denjoy and of A. Schwartz. If in addition f " is
Lipschitz then it is proved in [MS2] that there exists C0>0 such that
B ( f m, T,J)>~exp -Co. x If"(T)l 2 ~>exp -Co.r. fi(T) .

i=0
The previous theorem shows how to find lower bounds for B ( f m, T, J). From the
next results the usefulness of these lower bounds will become clear. The first of these
are similar to well known properties of (quasi)-conformal mappings.
2.5. "MINIMUM PRINCIPLE." Let Te-N and g: T---~g(T)~N be a C l diffeomor-

phism with T=[a, b]cN. Let xE(a, b). If for any J*~T*cT.
B(g, T*, J*) >I C> 0
then
(2.3) [Og(x)l ~ C a min{IOg(a)l, IOg(b)l}.
Proof. The proof of this principle is given already in [MS1]. In order to be self-
contained let us give it here too. Let us consider the following two operators:
Bo(g, T*) = Ig(T*)12 1

IT*I 2 IOg(a*)lIOg(b*)l
IOg(x)l Ig(Z)l
Bl(g, T, x) = IZl
Ig(t)l Ig(R)l '
ILl Igl
where T*=[a*, b*]c T and L and R are the connected components of T-{x}. Observe
that
Bo(g, T*) = lim B(g, T*,J), Bl(g, T,x) = limB(g, T,J).

J---~T * J--~x
Hence Bo(g, L), Bo(g, R), Bl(g, T, x)~>C>0. Since
Bo(g, L), Bo(g, R) >IC
we have:
Cl ,,a,I IO.(x,I. Cl ,,x,I Io,( 'l

Since Bl(g, T, x)~>C>0 we have
lOg(x)[ ~ >I C Ig(Z)l Ig(R)I

ILl IRI "
Since glT is a diffeomorphism,
min/lg(t)l Ig(R)l~ ~< Ig(T)l ~<max~lg(t)l Ig(R)ll

l I L l ' [RI J ~ [ ~T ' ~ J'
Then
C2(Ig(L)I/ILI.Ig(R)I/IRI~2 C2min~(ig(L)[~2, (ig(R)l)~

IDg(x)12>~ \ ~ ] t \ - - ~ / \ IRI / J
>t C 3min { lOg( a)l IOg(x)[, IOg( b )l [Og(x)[} .
Hence IDg(x) l>~C 3min( IDg(a)l IOg(b) l}. Q.E.D.
For maps with negative Schwarzian derivative one version of the next principle was
first used and proved in [Strl] and later rediscovered by S. Johnson and J. Gucken-
heimer, see [Gu2]. See also [Str2] and [Str3].
In order to state this principle it is convenient to introduce the following termino-

logy. Let U c V be two intervals. We say that V is a 6-scaled neighbourhood o f U if each
component of V \ U has length 6[U I. Similarly we define when V contains a 6-scaled
neighbourhood o f U.
2.6. "KOEBE PRINCIPLE." For each C,v>0 there exists K(C,v)<oo with the
following property. Let g: T-->g(T)cN be a C l diffeomorphism where T is a subinterval
of N. Assume that for any intervals J* and T* with J * = T * c T one has
B(g, T*, J*) >I C> O.
Let M c T and assume that g(T) contains a v-scaled neighbourhood of g(M) then
1 ~ Ig'(x)l ~< K(C, r), Vx, y E M.

(2.4) K(C, v-------~ [g'(Y)l
Proof. See [Str3] for a more general statement of this principle. In order to be
complete we will include a proof of (2.4) here. By rescaling we may assume that
M=g(M)=[O, 1] and that g is increasing. Let a, bE T be such that a < 0 < l < b and that
286 M, MARTENS, W. DE MELO AND S. VAN STRIEN
g ( a ) = - r and g(b)=l+r, L=[a, 0], M=[0, 1] and R=[1, b]. As in the proof of (2.3) one
has
_< 1 { [g(M)[ ~2
(2.5) IDg(0)[ IDg(1)l ~ - ~ \ - - ~ / 9
Similarly
Ig(L )l / ILl "[g(M)l / IMI
IDg(O)l~ C.
[g(L UM)I/IL OMI
Using Ig(M)I=IMI=1 and

Ig(L)l ILUM[
ILl [g(LuM)[ ILl l+r l+r'
this gives
Cr
(2.6) [Dg(O)[ >I l +----~"
Similarly
(2.7) IOg(l)l >I C----L-r.
l+r
Combining (2.5)-(2.7) gives that there exists K'<oo such that
(2.8) 4_, ~< [Dg(0)[, [De(l) I ~<K'.

K
Using the Minimum Principle one obtains that for each x E [0, 1],
C3
(2.9) ]Dg(x)l >>-~.
K'
Let U=[O,x] and V=[x, 1]. Since g is a diffeomorphism one has either Ig(U)]/IUI <<
.
Jg(M)i/JMl=l or Ig(V)I/IVJ<.Ig(M)J/IMI=I. If the former holds then
[Ig(U)lllUI]2 >-C
IDg(O)[ IDg(x)l
gives
]Dg(x)[ [Dg(O) I <~1 . 1 .
Using (2.6) this gives

IDg(x)] <~_1 ~l + z
C Cr
From this and (2.9) it follows that there exists K " < ~ such that
(2.10) 17 <~Og(x)l <~ K".

K,---
Therefore
<<. IDg(x)] <~(K") 2, Vx, y E M.

(IC) 2 [Dg(y)l
Q.E.D.
The next two results will play an important role in proving that the periodic points
of high period of a map in ~3 are repelling and are concerned with the situation that
B(g, T, M ) - 1 is positive and bounded from below. The next result states that whenever
an interval is mapped monotonically over itself with expansion of the cross-ratios then
the map is really 'bending' and therefore at some point expanding.
2.7. "EXPANSION PRINCIPLE." Let T be an interval in N and g: T---~g(T)cN a C l

diffeomorphism. For every e>0 and 8>0 there exists p>0 such that the following holds.
Let T, M be intervals such that both components of T \ M have at least length 6" ITI. If
B(g, T, M) >- 1+ e and g(T)~ T then there exists 0 E T with Dg(O)~ 1+Q.
Proof. Let ~>0 be so small that (l+e)(1-~)2~>l+ 89
Case 1. First suppose that Ig(L)l/ILl>~l-~ and Ig(R)l/IRl>~l-~. Then, using

B(g, T,M)>-I+e, we get
Ig(T)l Ig(M)l >>,(l+e)(1-~) 2 I> 1+-~-1e.

ITI IMI 2
So at least one of the terms Ig(T)I/ITI, Ig(M)I/IM I is greater or equal than 1V~e/2.
Using the mean value theorem we obtain in either case 0 E T so that
(2.11) tDg(O)l >- 1VT-~e/2.
Case 2. Suppose that Ig(L)I/ILI<I-~. The case that Ig(R)I/IRI<I-~ is proved

similarly. Then, using ILl, IRI-->61TI, we get
Ig(MUR)I Ig(T)l-lg(t)l ITl-lg(t)l

IMURI IMI+IRI IMI+IRI
> IZl-(1-~)fZl > 1+8~.
IMI+IRI
288 M. MARTENS, W . DE MELO A N D S. VAN STRIEN
The mean value theorem then gives 0 E M UR such that
(2.12) [Dg(O)I >>-I +6e.
Together, Case 1 and Case 2 prove the result. Q.E.D.
We will need to use the previous lemma in the case where g is an iterate f f o f f .
Hence we will need that B ( f f , T, J) is strictly bigger than one. In order to get such a
lower bound for B ( f f , T, J) we will use the following result which is based on the non-
flatness of the critical point. (This lemma does not hold for piecewise linear maps.)
2.8. LEMMA. Let gE ~ and c a critical point o f g. Then for every r>0 there exists a
neighbourhood U o f c and a number ~>0 such that if T is an interval in U \ {c) with
ITl>~rl[c,T)l and M the middle third interval in T then
(2.13) B(g, T , M ) ~ 1+~.
Proof. We may assume that N = [ - I , 1]. Let B(g, T)=B(g, T, M) where M is the
middle third interval in T. Let b*c, k the order of the critical point and define
IX -- C I
q~o: [c, b]---> [0, 11 by dPb(X)= ib_c I ,
[Y--g(c)l
~Pb:[g(c), g(b)] ~ [0, 1] by ~Ob(y) = ig(b)_g(c)l,
gb: [0, 1]--->[0, 1] by go(x) = ~ o g o ~ l ( x )
and g~(x)=x k. Since ~b and ~Pb are affine transformations and the crossratio are
invariant under affine transformations, one has for any interval T=[a, b]~[c, b]
(2.14) /l(g, T) =/~(gb, [ l a - c l

Let ~: [c, b]x[0, 1/(1 +r)]---~R be defined by
dp(x, y) = B(gx, [y, 11).
Since g(x)=+[Jp(x)lk+g(xo) where ~ is a C 2 diffeomorphism and k > l one has that gb

depends continuously on b and gb tends in the C o topology to gc as b---~c where gc is the
function gc(x)=x k. Therefore ~ is continuous.
Notice that since k~>2, the map gc has negative Schwarzian derivative and there-
fore B(gc, V, U ) > I for any intervals U c V c [ 0 , 1] with cl(U)~int(V). Therefore the map
[0, 1/(l+z-)] By--->~(gc,[y, I])-1 has a positive minimum, say 2~>0. So ~(0,y)>~l+2~
for all y E [0, 1/(l+z)]. Let Y0 be such that
(2.15) qb(x,y)~>l+~, V(x,y)E [C,Yo]X[O, l-~r ].
Now let U be a neighbourhood of c with diameter ~<Y0- If I=(a, b) is an interval in

U\{c} and III/ I[c, l)l~>r, then lb-al/Ic-al>~r. Therefore la-cl/Ib-cl E [0, 1/(i+r)]
and from (2.15),
~(g,i)=~(gb,(la--cl ~_{, [a--c[\
This finishes the proof of the lemma. Q.E.D.
3. Analytical estimates on monotone branches o f f n when f E ,~:

the Macroscopic Koebe Principle
In this section we will analyze the distortion of the cross-ratio for iterates of a map
fEM. We will face two new difficulties. First we have less differentiability as in the
previous section and, therefore, we will have to rely on more disjointness assumptions
in order to get good bounds for the distortion of the cross ratio of iterates. The second
type of problems is that we will need later some estimates on the cross-ratios when the
iterates of the interval may contain critical points of inflection type. So we will prove
here a version of the Koebe Principle for branches o f f n which are only monotone, and
not necessarily diffeomorphic.
The basic strategy is the same as before: we need to estimate the contraction of the
cross ratio under high iterates o f f on an interval T. Because in this section f E M, we
only have that logDfhas bounded variation. Therefore we need to split up the iterates
of T into collections of disjoint intervals. So we start by discussing some disjointness
properties of families of intervals.
3.1. Definition. The intersection multiplicity of a finite collection of intervals in N

is the maximal number of intervals in this collection whose interior has a non-empty
intersection.
3.2. PROPOSITION. Let ~ be a finite collection of intervals in N with intersection

multiplicity at most p then there exists a partition of the collection
~V'=A l UA2U... UA2p,
such that A k consists o f mutually disjoint intervals for k= 1,2 ..... 2p.
Proof. Clearly we may assume that N is connected. So we only need to consider

the cases that N is equal to an interval or a circle.
P r o o f if N = [ - 1 , 1]. Of course we may assume that all the intervals in ~r are open.
We claim that if N = [ - 1 , 1] then there are p classes
At,A2, ...,Ap,
which form the desired partition of ~V. Indeed let ~ be some collection of intervals in
[ - 1, 1]. F o r I E ~ let next{l, ,~ } be some interval I' E ~ such that (i) I' does not intersect
I and is to the right o f / a n d (ii) there is no interval J E ~ satisfying (i) which is closer to I
(if there is no interval to the right of I take next{l, ~ } to be the empty set). Similarly
~ ( t ) is some interval in ~ such that there is no interval in ~ which has points to the
right of ~ ( ~ ) . F o r k = l , 2 ..... p we define inductively Ak as follows. L e t A0=O and
suppose that we have defined by induction Ak_ r If k<~p, and ~k=~//'\t.Ji=l ..... k_lA,.=~
then let A k = ~ . If ~k is non-empty then take an interval lkE ~k SO that there is no
interval in ~k containing points to the left of Ik. Then define
Ak,, = {Ik},
A,,,+ 1 = A,,, U next (~(Ak,,), ( 7/V\ (A o U ... UA,_ I tJA~,,)) ) ,
and
A k -- U Ak, n"
n>~l
Then Ak is a collection of intervals with disjoint interiors and the collections At ..... Ap
are mutually disjoint. Now we will show that ~ O ... OAp. Suppose this is not the
case and there exists an interval I in the collection ?,F\(AI tJ ... OAp). Then
IE ~ U ... UAk_ 1 UAk)
for each k = l ..... p and since I ~ A k it follows from the inductive definition above that
there exist for i= 1,2 ..... p, Ti E Ai such that the left boundary point o f Ti is to the left of
(or equal to) the left boundary point x of I. But then since all these intervals are open,
points just to the right of this boundary point x are contained in I as well as in Tk,
k= 1,2 ..... p. This contradicts the assumption that the intersection multiplicity of ~ is
at most p.
P r o o f i f N = S 1. I f N = S l then choose some x E S 1 and let Ii ..... L be intervals in ~V

which contain x. Then r<,p. Since s l \ { x } - ( - l , 1 ) c [ - 1 , 1] it follows from the pre-
vious case that ~/r .... , L} can be disjointly decomposed into collections A1 ..... Ap
of disjoint intervals. Then ~ 1U... UAp Ut,lk=1...... {Ik} has the desired properties.
Q.E.D.
The main result of this section is a version of the Koebe Principle for maps fn
restricted to an interval T such that f"lT is monotone (but may have critical points of
inflection type) and such that the collection T,f(T) .... ,f"-l(T) has low intersection
multiplicity.
3.3. MACROSCOPIC KOEBE PRINCIPLE. Let f E M . Then there exists a strictly

positive function B0: (0, 1)---~(0, 1) such that for any pair o f intervals M e T , any n>~Oand
any 0 < e < l , satisfying the following conditions:
(a) f~(T) contains an e-scaled neighbourhood o f f ' ( M ) ;
(b) the intersection multiplicity o f {T,f(T) .... ,f~-l(T)} is at most 17;
(c)fi(M)flKf=f~, iE {0, 1..... n - l }
one has:
T is a Bo(e)-scaled neighbourhood o f M.
For the proof of this theorem we will need a number of lemmas. Let f E M. Let
~ c c l ( ~ ) ~ i n t ( q / ) be interval neighbourhoods of Kf such that the number of compo-
nents of W and ~ are both the same as # K s.
3.4. LEMMA. Let f be a map in M and -=={T1, T2..... T~} a collection o f intervals in
N and Mi~Ti. Assume that the intersection multiplicity of E is at most 17 and that none
of the intervals Ti contains a component o f ~ \ W or contains points of Kf. Then there
exists V<oo such that
n
(3.1) X log B(f/, T/, Mi) >~- V .

i=l
Proof. Since f E M , the restriction of log lDf I to N \ W exists almost everywhere

and log IDfl has bounded variation on this set. Let
V~= Var(log[IOfl(N~ ~r)]) < oo.

For each xiE Ks let Ui be the component of ~ which contains x~. Since f E M,
f(x)=Cp~o r where r +Ix[ ~' and ai>~l. Here r U i - ~ ( - 1 , 1) is a homeo-
morphism and log [D~;[ exists almost everywhere and has bounded variation. Hence
V'=E~Var(loglDdpi[) is finite. Let V=34(Vs+V').
Let II={i;TiA~V=~) and I 2 = { i ; T i c ~ ) . Since the intervals T; never contain a
component of ~ \ ~ V , Ii UI2={1,2 ..... n).
First assume that i EIl. Let L; and R; be the components of Tg\Mi. For
u v v~E T~ let (u;, vi) be the open interval connecting u; and vi. Because D f exists almost
everywhere there exist miEMi, liELi, riERi and r;E T i such that D f exists in these
points and such that [f(Mi)l/[Mil>~lDf(mi)[, If(Ti)l/ITil>-lDf(r~)l, If(Z~)l/IZ~l<~lDf(OI and
If(R;)[/[Ri[<<.IDf(Q[. Using this in the definition of B(f, Ti, Mi) we find that there exist
li E L~, ri E Ri, mi E Mi and r i E T i such that
/IDf(m,)l IDf(r;)l
(3.2) log B(f, T i, Mi)>--l~ [~ ~ y/[t r / 'J q \ /
and
(3.3) mi E (li, ri).
From (3.2) and the choice of the points li, mi, ri, r; one has
(3.4) log B(f, T;, M~) I> - {[log [Df(m~)[-log IDf(O[[+[log IDf(ri)l-log IDf(r,)l [}
and also
(3.5) log B(f, Tg, Mi) t> - {Ilog IDf(mi)l-log IDf(ri)ll+ Ilog IDf(rg)l-log [Df(lg) I I).
Rename the points Ii, m v r i, r i in increasing order a~, b i, c i, d v From (3.3) one gets that
either (li, mi) fl (ri, ri)=~ or (ri, Ii) fl (m i, ri)=~, and so we can use either (3.4) or (3.5) and
get
log B(f, Ti, M i) >I - { l l o g [Df(bi)[-log IDf(ai)l 1)+llog IDf(d)l-log IDf(ci)[ l)

(3.6)
>I - V ar(log lDf lTil)
and therefore
(3.7) logB(f, T i, Mi) >>--Var(log IDflLI).

Now consider i E 12. Then T,. is contained in some component Ui of ~ (and does not
intersect Kf) and so f has the form f(x) =dpaio r Hence
B(f, T/, Mi) = B(qJai, T~, M~)xB(ePi, Ti, Mi).
Here T~=dpi(Ti) and M~=dpi(Mi). Since the Schwarzian derivative of $~, is less or equal
to 0 (because ai~>l) one gets B($~? T~,M~)~>I. Hence, as above,
loge(f, T/, M,) ~ O+loge($,, T/, M i)

(3.8)
~> Var(log ]DqbilT/I).
-
Since the intersection multiplicity of E is at most 17, using Proposition 3.2, one can
write "--=At UA2... UA34where Aj consists of a collection of mutually disjoint intervals.
Hence from (3.7) and (3.8) one gets
(3.9) ~ logB(f, Tv M~) >i -34. (Vf+ V').

i=1
The lemma follows. Q.E.D.
3.5. LEMMA. Let f E M. Then there exists A0>0 such that if le-W are intervals in N
such that
(a) W \ I consists of one component H;
(b) [at~llI;
(c) InKs=O,
then
(3.10) If(I)l If(H)I

Ii I >-'A0" IH I
Proof. The proof of this lemma is elementary and can be found in [MMMS].
Q.E.D.
3.6. LEMMA. Let f E M and I, T be intervals with cl(I)cint(T) and L and R be the
components of T \ I . Let 7E(0, 1). If
If(I)l >17" I I I and If(I)l >17" [I_]_[

If~R)l IRI If~L)l ILl'
then
B(f, T, I) >-),2.
Proof. Let
d = {\ l-~-~l
Ill Ill ~/(
: / \ -III~ + ~ ) 9
Then
If(I)l ~_ If(I)l If(I)] F If(I)]
B(f, T, I) = 9 If(R)[ If(R)l If(L)l If(L)l /> ~,+72"d _ 7 l+y.d ~> •.
Izl~ II1 II1~ III l+d l+d
IRI IRI ILl ILl
The last inequality follows since y E (0, 1) and therefore (l+Tx)/(l+x) is always greater
than 7 for x>0. The result follows. Q.E.D.
3.7. LEMMA. Let f E M. Then there exists A I > 0 such that i l l T are intervals with
cl(I)cint(T) and L and R the components o f T \ I such that
(a) It[,<[I[ or [R[_<[I[;
(b) In Ks=O;
then
B(f, T, I) ~>AI.
Proof. LetAo be the number from Lemma 3.5. We may assume that AoEO, 1). We
will prove the lemma for A] =]. (A0)2. By possibly renaming L and R, we may consider
the case that [RI~<[I[. Then from Lemma 3.5 we get
If(I)[/[I[
(3.1 l) [f(R)t/IR[ >IA o,
and hence
If(Z)l ILl ILl
B(f'T'I)>~ If(t)l IZl "A~176 IT--(.
If ILI~>[II then it follows from this and IR[~<II I that

ILl ILl
B ( f , T, I ) ~ A o 9 ILI+III+IRI >~A~ " ILl+ILl+ILl - A o 93 ~ > A ]
and the lemma is proved. So assume that [LI~<[I1. Then applying Lemma 3.5 again we
get
(3.12) If(I)l/lII ~ Ao,

If(L)l/ILl
and it follows from (3.11), (3.12) and Lemma 3.6 that B(f, T, I)>~Ao.Ao>A~. Q.E.D.
Proof of Theorem 3.3. Let eE(0, 1). Let m<~n be the smallest number such that
Ifm(L)l>~elfm(M)l and If'~(R)l>-elfm(M)l. Let Al be the number from Lemma 3.7 and
assume A~<I. Let V>0 be the number from Lemma 3.4. Let
B 1= [A1151~*/r +1 . e-V.
We claim that
B ( f m, T, M) >1B~.
Indeed let t(1)<t(2)<...<t(s)<m be the integers t<m such thatf'(T) contains a compo-
nent of ~ \ ~ V or such that f ' ( T ) n K i , ~ . Since the intersection multiplicity of
{T,f(T), ...,fn-l(T)} is at most 17 one gets s < 5 1 # K s. From the choice of m either
[ft(~176176 or [ft~176176 So from Lemma 3.7
B( f ,ft(~ T),ff(~ )) >1A, .
From Lemma 3.4
E log B(f, fJ(T),if(M)) >I - V.

j<~m-l,j~ {t(l) ..... t(s)}
Hence the claim follows. From the claim and the definition of the operator B the
theorem easily follows. Q.E.D.
4. Some simplifications and the induction assumption

The proof of Theorem A will go by induction on the number of turning points off. So
let M u be the collection of all endomorphisms of N in M with f(ON)cON and with
precisely d turning points. So we will prove inductively that
u
(Indd) maps in tJ M i have no wandering intervals.
i=0
Because the proof of Theorem A goes by induction on the number of turning points we
have to consider a more general situation: the manifold N is not necessarily connected
(but does consist of a finite number of components). This does not give, however, a
more general result. In fact if a map f: N---~N has a wandering interval then one of the
connected components N' of N is periodic of period s and the map fs: N'---~N' has a
wandering interval. Therefore, if we prove the theorem for connected manifolds we get
also the same theorem for non-connected manifolds. However, if f has d turning points,
the m a p f s may have more than d turning points. This is the main reason to start with a
disconnected manifold. So if Indd_ 1 holds then we may use the following fact: if there
exists a finite disjoint union o f intervals which is inoariant by f and contains a
wandering interval then the union o f these intervals contains at least d turning points.
By extending f r o a slightly bigger interval we may assume that
f(ON) ~ 8N.
It suffices to prove the theorem for maps f such that none of the turning points is
contained in (the closure) of a wandering interval. Indeed, otherwise f also has a
wandering interval W containing a turning point c in its interior. Now modify the m a p f
in a small neighbourhood V c W of c to a map gE ~t such that f = g on N - V for which
glV has a unique turning point in c and g(c) is not contained in a wandering interval.
Thenf(W) is still a wandering interval for g (the forward iterates off(W) u n d e r f a n d g
are the same because these iterates never enter W) and g(c) is not contained in a
wandering interval of g. Repeating this procedure for every wandering interval which
contains a turning point in its closure we get a map g which still has wandering intervals
but such that none of its turning points is contained in the closure of a wandering
interval. So it suffices to prove the theorem for g.
Furthermore, for each turning point c of f there exists a neighbourhood Sc of c and
a continuous involution r: Sc--~S~ (which is not the identity) such thatf(r(x))=f(x) for all
x E So. From the non-flatness conditions for maps in M the map r is Lipschitz.
5. The pullback of space: the Koebe/Contraction Principle

In this section we will start the proof of the non-existence of wandering intervals for
maps in ~/. This will be proved by contradiction. More precisely, suppose that J is a
wandering interval and that J is not contained in a larger wandering interval. The
strategy of the proof is to show that there exists necessarily an interval I which strictly
contains J such that infn~0 Ifn(I)l=0. Using the next principle, which also can be found
in [L], it follows that I is a wandering interval (since fn(J) does not converge to a
periodic orbit the same holds for f~(I)), contradicting the maximality of J.
5.1. CONTRACTION PRINCIPLE. Suppose I is an interval such that
inf If"(I)l = 0
n~>0
then I is a wandering interval or there exists a periodic orbit t~ such that fk(I)---~t7 as
k---~. In particular, if I contains a wandering interval J then it is also a wandering
interval.
Proof. Let or ['Jn~0 fn(int(l)) 9 Then or is forward invariant.

First suppose that there exists a component U of ~r and n>0 such that
fn(U) fl U~=~. Since ~ is forward invariant this i m p l i e s F ( U ) c U. There are three cases.
(1) U is an interval which contains a fixed point p o f f n: U--~U in its interior. In this
case some iterate of I contains this fixed point of f " in its closure and since
infk>>.olfk(I)[=O this fixed point o f f " must attract Lfk(I)--->O(p) as k-->oo. So we are
finished in this case.
(2) U is an interval and there exists no fixed point as in (1). Then cl(U) contains in
its boundary an attracting fixed point p of i f : cl(U)--~cl(U). Iff"(U)~=U then every
point in cl(U) is asymptotic to O(p), to(I)=O(p). I f f n ( U ) = U then the boundary point
{q) = a U \ { p ) is a repelling periodic point and infk~olfk(I)]=O implies that no iterate of
I contains q in its closure. Since every point in int(U) is asymptotic to O(p) this implies
that fk(I)--+O(p). Again the result follows.
(3) U is a circle. Then there exists a finite collection nl<...<nr of positive integers
such that Ll~=lf"'(int(I)) covers S 1. But since inf,~0lf"(I)l=0 this implies that there exists
an integer n>nr such that f"(I) is strictly contained in f~i(int(I)) for some i= 1..... r. But
t h e n f n-"' has an attracting fixed point in f'(int(l)) which attracts I. So again the result
follows.
Now assume that for every component U of ~ one has fn(U)fl U = ~ for all n~>l.
Since ~ is forward invariant and this holds for each component, this implies that
fn(U)flfm(U)=r for all n>m~O. It follows that U and therefore I is a wandering
interval (or asymptotic to a periodic orbit). Q.E.D.
5.2. Definition. The pullback of Pn~Jn is the sequence of intervals
{eil i=0, 1..... n)
where Pi-1 is the maximal interval containing Ji-1 such that
f ( P i - 1 ) c Pi
20-928283 Acta Mathematica 168. Imprim6 le 24 avril 1992

for each i= I ..... n. The integers i for which Pi contains a turning point in its closure are
called the cutting times. It will turn out to be useful to call n also a cutting time. This
pullback is monotone iffnIP0 is monotone and fn(Po)=P. Furthermore it is said to be
diffeomorphic iffnlPo is a diffeomorphism and f~(Po)=P and unimodal if Pi contains at
most one turning point for each i.
All the pullbacks we will consider are unimodal because of the following lemma:
5.3. LEMMA. There exists r/>0 such that if Pn~J, and IP~I~,J then the pullback of
P~ is unimodal.
Proof. Follows from the Contraction Principle and since by assumption no wan-
dering interval contains two turning points. Q.E.D.
5.4. CONTRACTION/KOEBE PRINCIPLE. For each e>0 and each p EN there exists
No(e, p) with the following properties. Let Pn~Jn contain an e-scaled neighbourhood of
Jn. If the pullback of P~ has intersection multiplicity <<,pthen n<<-No(e,p).
Proof. Let p=min(e, 1/2) and / ~ P ~ be a p-scaled neighbourhood of Jn. Let

m(O)<m(1)<...<m(l)=n be the cutting times of the pullback of Pn. Since the intersec-
tion multiplicity of the pullback is <<,p,l<~p9d. Let/~3 be the components of/~i\J,. For
t=0 ..... l - 1, let P~(t) be the component which contains the turning point. Now the map
f~-m<t-l)+l: Pm(t-1)+ l -"> P n
is monotone. Since/~ is a p-scaled neighbourhood of J~ it follows from the Macroscop-

ic Koebe Principle that both components of Pm<t_l)+l\Jm<t_l)+~ have length at least
B0(p). From this and the non-flatness of the turning points we get that there exists a
universal constant C E (0, I) such that
IPx( -,I >I c" Bo(p)" IL.(H)[.
Because Pm<t-~) is symmetric around a turning point, we also have that IVm(,_,)
~+ 1I>
]Jm(l-l)l" Therefore
p+
I m(H)l I> C'Bo(P)'IJ,~<I-I)["
Repeating this l~p. d times we get, letting g(x)= C. Bo(x),
(*) IPol I> g'(e) IJI.

Hence there exists an interval P which does not depend on n and which strictly
contains the wandering interval J such that If~(e)l<<.(l+2p)lJ~l<<.21Jnl for each n as
above. Now if there exists no upperbound No(e,p) for n then we would get
If~'(P)l<.2lJn,I--,ofor some sequence ni---~oo and since P contains J this would imply
from the Contraction Principle that P is also a wandering interval. But this contradicts
the maximality of J. Q.E.D.
6. Disjointness of orbits of intervals

In this section we will give some upperbounds on the intersection multiplicity of orbits
of intervals. These bounds are needed in order to apply the Macroscopic Koebe
Principle. Let Jn=fn(J). We will define a natural neighbourhood Tn of J~ such that its
monotone pullback has good disjointness properties. (For those familiar with the circle
homeomorphisms these neighbourhoods will coincide with the neighbourhood
[fqk-~(J),fqk-2+(ak-1)qk-l(J)] of fqk(J) when n=qk and with
If qk_2+(i-1)qk_l(j),fqk_~+{i+1)qk-l(j)]
of fqk-2+iqk-~(J) when n=qk_2+iqk_ l and l<.i<~ak--1. )
6.1. Definition. We say that J~l and J~2 have the same orientation if the forward
iterate of f which sends one of these intervals in the other is orientation preserving. If
n E N, we say that Jk is a predecessor of J~ if O<-k<n, if Jk and Jn have the same
orientation and if J,~(Jk,.In) and O<~s<n implies that Js and Jn have different orienta-
tions. IfJn has a predecessor to its left (right) then we denote the corresponding iterate
by L(n) (respectively R(n)). Jn has a successor J~+~ if
(1) J~-a is predecessor of J~ (with 0<a~<n);
(2) fal[Jn_a, Jn+a] is monotone, orientation preserving, and its image contains no
predecessor of J~ (if L(n) and R(n) both exist and ff for example n - a = L ( n ) then this
implies that ft[Jn_a, Jn+a]C[Jn,JR(n));
(3) if JkC(Jn, Jn+~) and k=0,1 ..... n § then the intervals Jk and J~+a have
different orientations.
Next we define the natural neighbourhood of Jn to be the biggest open interval
containing J~ which contains no neighbourhood of a predecessor or successor.
Remark. Of course J , can have at most one predecessor on each side and it has a
predecessor to, say, its right if there exists an interval Js with s<n to the right of Jn with
the same orientation as J~. Moreover, as we will see in the lemma below an interval J~
has at most one successor; so denote this successor by s(n). Therefore, if Jn has two
predecessors and no successor then the natural neighbourhood T. of J. is equal to
Tn= [JL~.), JR~.)] and if it has a successor then it is equal to
T~ = [J/.~.), J~r or 7". = [Jsr JRr
6.2. LEMMA. For every nEN, J~ can have at most one successor.
Proof. Suppose that J~ had two successors. Then it also has two predecessors, JL~n)
and JRr By definition f.-Lr [j/.r j~]._. [j~, JRr and f.-Rr [j., JRr ~ [JLr J~] are
both monotone. Hence J is attracted to a periodic point, a contradiction. Q.E.D.
6.3. LEMMA. Assume that the interval .In has two predecessors JL~n),JR~.), and a
successor Js~.). I f this successor is to the right of J. then the predecessors of Jsr are J~
and JR~) and if Jsr has a successor then this successor is between Js~.) and JRr
Proof. The left predecessor of s(n) is n by the definition of s(n). The right
predecessor of s(n) is certainly defined because JRCn)and Jsr have the same orientation
and R(n)<n<s(n). Let us show that R(n) is this predecessor. If this was not the case
then there exists k<s(n) so that JkC(J~.), JR(.)) and so that Jk and J~(~) have the same
orientation. Because of the definition of R(n) this implies certainly that k>n. Because
k<s(n) this implies that O ~ k - n < s ( n ) - n = n - L ( n ) and L(n)+k-n<n. From the first of
these two inequalities and the definition of s(n) it follows that f k - . is monotone and
orientation preserving on H=[JL~.), J.]. In particular JL(.)+k-. has the same orientation
as J.. From this, the definition of L(n) and R(n) and the second of these inequalities it
follows JL~.)+k-n cannot be between JL~.) and JR~n)" It follows that
f k-n(H) = [JLc-)' J/c] = [am,Js(.)] = f "-L(n)(H).
In particular, f~.-z~.))-~k-.) maps fk-"(H) monotonically into itself, and hence J would be
attracted to a periodic attractor, a contradiction.
Let us finally show that s(n) cannot have a successor to its left. Indeed if it did,
then by definition fs(n)-R(n) would map [J~.), Jm.)] monotonically into [J., J~.)]. From
the definition of s(n), f~")-" maps [J., J~.)] monotonically into [J~n)'JR~.)]" Combining
this gives that J is attracted to a periodic attractor. With this contradiction the proof of
this lemma is completed. Q.E.D,
Remark. The previous lemma implies that if J. has a successor Js~.) and Js~.) also
has a successor J ~ . ) ) then s(n)-n=a=s(s(n))-s(n) and J~.) is between J. and J~s~.)). So
continuing this there exists a maximal integer k such that J,+~(,) is a successor of J'(r)for
i=0, 1 ..... k - 1. In this case the intervals
L(r), J,(,(r))
..... Lk(r)(n)
lie ordered and f"[[J,, Jk(r) ] is monotone. So f " acts as a translation on these intervals.
6.4. THEOREM. Let n ~ N and assume that Jr has two predecessors JRcr) and JL(r).
Let Mr be an open interval contained either in [,In,JR~r)] or in [JL(r), Jr]. Assume that
{Mto, Mt0+l ..... Mr} is a monotone pullback o f Mr. I f the intersection multiplicity o f this
piece is at least 2p and p>-2 then there exists tE {to ..... n} such that
(1) Js(t),J2(t) ..... J~_2(t) are defined;
(2) n=sP(t) and J i(t) is contained in M r f o r j = p ..... 2 p - 2 .
COROLLARY. I f the pullback o f an interval T=J, with T contained in the natural

neighbourhood Tr is monotone then the intersection multiplicity o f this pullback is at
most 11. Similarly, if TDJ, and sk(n) does not exist then the monotone pullback o f t has
at most intersection multiplicity 2k+4.
Proof o f Corollary. Consider the pullback of TN [J,, JR(,)] and Tn [JL(n), Jr] sepa-
rately, ff the intersection multiplicity of T is at least 12 then the pullback of either
Tn [Jr, JR(,)] or Tn [JL(n), J,] has intersection multiplicity t>6. So take p = 3 and the
previous theorem implies that p + 1~<2p-2 and Js(,) is contained in Tn [J,, Jmr)], which
is impossible since TcTr. The second statement follows also immediately. (Note that
2 k + 4 ~ 6 for k~>l.) Q.E.D.
Proof of Theorem 6.4. In order to be definite assume that Mnc[Jr, Jk(r)]. By

assumption there exists a point y which is contained in 2p of the intervals Mto..... M r.
Of course this implies that for at least p of these intervals Mi the corresponding
intervals Ji all lie on the same side of y. So let N>~p be the maximal number of distinct
integers t<~i(1) ..... i(N)<,n such that each of the intervals Mm),..., Mi(u) contains this
point y, Ji(j) all lie on one side of y and that these are labeled so that
[Jio),Y] ~ [Ji(z), Y] D...D [Ji(N), Y]"
Since Mi has a common endpoint with J; this implies that the intervals Ji(l) .... , J . m all
have the same orientation.
Claim 1. i(1)<i(2)<...<i(N) and we may assume that i(N)=n. Furthermore Ji(l)

cannot be contained in ft(Mim) for t= 1..... i(N)-i(1) i f f ' is orientation preserving on
Mi(1). In particular if we take a=i(N)-i(N-1) then fa m a p s [Ji(1),Ji(N_l)] (which is

contained in Mi,)) monotonically into (Jm), Ji(m].
Proof of Claim I. For j E (2, 3, ..., N} the interval Ji(j) is contained in [Ji~-l), y]c
Mi(j_l) and Ji(j) and Ji(j-l) have the same orientation. Since fn-i(j-l): Mi(j_l)___>Mnc
[.In,JR(n)]is monotone this implies that Jg(j)+~-e(j-~)has the same orientation as Jn and is
between J~ and JR(.)" Since JR(~) is a predecessor of J~ this implies that
i(j)+n-i(j-l)>n and therefore i(j)>i(j-1). Let us show that we may assume that
i(N)=n. Indeed from what we have s h o w n f ~-i(N) is monotone on M~(j) f o r j = l ..... N.
In particular letting i'(j)=i(j)+(n-i(N)), and taking the images of these intervals under
this map we get that Mm) ..... M:(N) all contain one point y', Ji'(y) all lie on one side o f y '
and that these are labeled so that [Ji,fl),y]D[Ji,(2),y]~...=[Ji,(N),y]. Since i'(N)=n we
may as well assume that i(N)=n. The first statement of the claim follows and it follows
t h a t f a is monotone and orientation preserving on Mi,)=[Ji, ~,J;(N-1)]" Now Ji0) cannot
be contained in ft(Mm))=Mi(j)+t for t= 1..... i(N)-i(1) because otherwise J~(N)-twould
be contained in MI(N)=M. and would have the same orientation as J.. But this is
impossible because M.~[J., JR(.)] and JR(~) is a predecessor of J.. Q.E.D.
Claim 2. If k<n, Jk and .In have the same orientation and Jk~[Ji,),Jn] then
kE {i(1), i(2) ..... fiN)}. Furthermore if we let a=i(N)-i(N-l) then i(j+l)-i(j)=a for
j=l,2 ..... N-1.
Proof. Let j<n be maximal such that Jkc[Ji(.i), y]~Mi(.i ). Therefore f,~-i(j) maps Jk
into M.. Because Jk, Ji(j) and Jn all have the same orientation, it follows that Jk+.-i(j)
also has the same orientation. This implies that k>~i(j). Suppose k>i(j). If Mk:PJ~ then
MkcMi(j) and fk-i(~~ maps Me(:) monotonically into Mk~M~(:). This implies that J is
attracted by a periodic orbit, contradiction. Hence, Mk~J,, and by the maximality of
N, kE {frO)..... i(N)}. This proves the first statement of the claim. From Claim 1, fa
maps [Jm)' J~(N-~)] monotonically and orientation preserving into (Ji(l),Je(m]. It follows
from this and the first part of this claim that i(j)+a=i(j+l) for j = l ..... N - 2 . Thus
Claim 2 is proved.
Define i(N+j)=n+j.a f o r j = l ..... N. As we have shown in the previous claim this

formula holds for j negative. So we get
i(N+j)=n+j.a for j = - N + I , - N + 2 ..... N-1,N.

Now consider the interval
/4= [Ji(,. J,N)].
The m a p fn-i(1)=f(N-1)a maps HcMm) monotonically and orientation preserving into

M.c[J., Jr(n)]" In particular f a maps [Ji(l),Ji(2N_2)] monotonically and orientation onto
[Jm), Ji(2N-l)]" Therefore Ji~j) lies between Ji~j-l) and Ji~i+l)for Ij[<.N-2. Furthermore if
ft(H) n [Ji(l), JR(n)]*f3 for some t= 1..... ( N - 1) a thenft[H is either orientation reversing
or t is a multiple of a. Indeed, assume that f'lH is orientation preserving and that this
intersection is non-empty. By Claim 1, ft(H)cft(M.l)) does not contain Ji.). S~
must be contained in [Ji(1), JR(n)]" But ft(Ji(1))c-ft(H ) cannot be contained in [J., JR~.)]
since JR~.) is a predecessor of J.. So ft(Jm))c [Jill), J.] and it follows from the previous
claim that t must be a multiple of a. Furthermore consider Jk with k<n and with the
same orientation as J.. This interval cannot be contained in [J.1), J.] (see Claim 2) and
neither in [J., JR~.)] from the definition of JR~.). Combining all this shows that the only
intervals Jk with k<.i(2N-1) inside [Jm),JR~.)] with the same orientation as Jn are
intervals of the form k=i(1)+j.a. It follows that J,j+~) is a successor of Ji~j) for
j-- 1..... 2 N - 2 . Q.E.D.
7. Wandering intervals accumulate on turning points

In this section we are going to prove that the w-limit set of a wandering interval
contains at least one turning point and thus prove that Ind0 holds. From this is follows
in particular that maps without turning points, e.g. circle homeomorphisms in M,
cannot have wandering intervals. So the proof in this section includes the classical
proof of Denjoy.
Suppose we have a map f: N---,N which has a wandering interval J which stays
away from the turning points off. But then we can modify the m a p f n e a r these turning
points without affecting the orbit of J. So change f so that each maximal interval on
whichfis monotone is mapped b y f o n t o a component of N. Once we have done this we
may assume that every pullback is monotone.
7.1. PROPOSITION. There exists no such that if Jn, n>>-no,has two predecessors JL~n)
and JR(.) and and [JLr then J. has a successor and [J,(.)[<~/.[.
Remark. Suppose for example that J,(.) is to the right of J.. Then Lemma 6.3
implies that the interval Js(.) has two predecessors, namely J. and JR(.) and that it has no
left successor, From the conclusion of this proposition it follows that the assumptions
of this proposition are again satisfied for J., where n'=s(n). So one can apply the
proposition infinitely often!
Proof. Let n0=N0(1; 11) be as in the Koebe/Contraction Principle and let n>-no.
Suppose s(n) is not defined and let Tn= [JL(,), JR(,)] be the natural neighbourhood of J,.
By the corollary to Theorem 6.4 the monotone pullback of T, has intersection
multiplicity bounded by 11. Because the intervals IJL(J, Jm,)l>~lJ,[ this contradicts the
Koebe/Contraction Principle. Hence s(n) is defined. From Lemma 6.3 L(s(n))=n and
R(s(n))=R(n). Furthermore s(n) has no left successor. Now the natural neighbourhood
of Jn is Tn=[JL(,),Js(,)] and again by the corollary to Theorem 6.4 the monotone
pullback of T, has intersection multiplicity bounded by 11. Since IJL(,)[>lJnl it follows
from the Koebe/Contraction Principle that [Jst,)[<[(Jn,Js(,)][< [J,l" Q.E.D.
7.2. THEOREM. l f f E M and f has a wandering interval the to-limit set of J contains
a turning point.
Proof. Let us first show that to(J) cannot be finite. Indeed otherwise to(J) contains
a periodic point p of, say, period k and there exists a neighbourhood U of p such that
Un to(J)= {p}. Furthermore there exist a neighbourhood V~-U o f p such thatfk(v)c U
and an integer n' for which J , , c V and such that J,~-V whenever n>-n ' and J, n U * ~ .
Sincefk(v)c U this implies by induction that J, cVUf(V)tJ.., tJfk-l(V) for all n>-n' and
therefore that to(J)=O(p). Hence to(J) is attracted to a periodic orbit, a contradiction.
Therefore, and since all intervals Ji are disjoint, there exists arbitrarily large
integers l, r<n such that Jl, Jr, J, are in he same component of N, have the same
orientation, J,c(Jt, Jr), such that
IJ~[ <~min(}Jtl, IJrl)
and, for i=O, 1..... n - 1,
Ji n (Jr, Jr) implies that Ji has a different orientation.
Assume that l, r, n are bigger than the number no from above. It follows that Ji and Jr are
the predecessors of J,. So we can apply Proposition 7.1 and hence J, has infinitely
many successors Jk(,), k= 1,2 ..... From the description in Lemma 6.3, all these succes-
sors are contained in [art,Jr], they either all lie to the right of the previous one or all to
the left. Moreover sk(n)--sk-l(n) is independent of k. It follows that as k tends to infinity
these intervals Jk(,) converge to a fixed point o f f where a=s(n)-n. Hence this fixed
point is an attracting fixed point with J in its basin, a contradiction. Q.E.D.
Now we know that iterates of wandering intervals tend to some turning point. In
order to analyze the metric properties of iterates of a wandering interval we will pay
special attention to the moments where the iterates get closest to some turning point.
This is formalised in the following definition.
7.3. Definition. If c is a turning point in to(J),J, c S , and m~n, then we say that J ,
is the m-closest approach to c if (J,,v(J,))n(Lli<_,,Ji)=f3. Similarly J~ is the closest
approach to c if it is the n-closest approach to c. Now fix a turning point c in w(J) and
let N(c) be the collection of integers i 6. N with
Ji'~S~ and [Ji,'K(Ji)]['l( I.J J j ) = ~ .

o<~j<i
From now on let
N(c) = {n(1), n(2) .... }
where n(1)<n(2)<... We call J,(1), J,(2) .... the sequence of closest approach to c.
7.4. LEMMA. If Jn(k) has a successor J~r

then s(n(k))=n(k+ 1) and Jnr is be-
tween J~(k) and c. Furthermore, if there exists an integer j such that s(j) and s2(j)) are
both defined and such that j<n(k)<s(j) then
n(k+ 1) = s(n(k)).
Proof. In order to be definite assume that J~k) is to the left of c. Since Jn~k)is a
closest interval to c, any predecessor of J,~k) to its right must also be to the right of c.
Therefore c 6. [J~k), JR~,~k))]and there can be no successor of J~k) to its left. Hence if J~tk)
has a successor then it must be to its right and because fft"~k))-"tk)l[Jun),J,~,)] is
monotone it even must be between J~tk) and c. So if s(n(k))*n(k+l) then
n(k+ 1)<s(n(k)). Consider H---[JL~k)), J~k~] and let a=n(k)-L(n(k))=s(n(k))-n(k). Then
f2~ is monotone on H and a>n(k+l)-n(k). Since L(n(k))+n(k+l)-n(k)<n(k) the
interval JL(n(k))+n(k+l)_n(k)cfn(k+l)-n(k)(H) is not contained in [Jn(k), r(J~<k))] whereas by
assumption J~k+l)Cf"~k+l)-~k)(H) is contained in this interval [Jn~k),r(J~<k))]" It follows
that either J~k) or V(J~r is contained in f~k+l)-~<k)(H). Hence for t=a-(n(k+ 1)-n(k))
one has 0 < t < a and
(*) ft(J,,r cfa(H) = [J~r Jsr162 c [J~r c].
Since Js~,~k))is the successor of Jn~k)this implies t h a t f t is orientation reversing on H. So

JLfn~k))+t is to the right of Jn~k)" But since L(n(k))+t<n(k) this interval cannot be
contained in [J~(k),r(J~tk))]" Therefore, and because of (*), i f ( H ) contains c; this contra-
dicts the monotonicity offaIH. This proves the first statement of this lemma.
21-928283 Acta Mathematica 168. Imprim~ le 24 avril 1992

Let us now prove the second statement. According to the first part it is enough
to show that s(n(k)) is defined. Now let a = s ( j ) - j then L ( j ) = j - a , s(j)=j+a and
s(s(j))=j+2a. Since fa is monotone on [Lj_ a, Lj+2a] and n ( k ) - j < s ( j ) - j = a it follows
that J,~k) is contained in T= [Jn<k)-a,J~tk)+a] and that fa is monotone on T. Furthermore
there is no predecessor of J~<k) in fa(T) because otherwise there would be a predecessor
of Jj+a in fa([zj, Zj+2a]) , contradicting that s2(j) exists. So property (I) and (2) of the
definition of successor of J~tk) hold. Finally there is also no interval Jt in [J~k), Jn<k)+~]
with t<n(k)+a and with the same orientation as J,<k) because Jt+~s~j)-~k)) would have the
same orientation as "/so) and be contained in [J,~j), Js<y)+a], contradicting the definition of
s2(j). Q.E.D.
8. Topological properties of unimodal pullback's

From now on we will assume
a
(Inda_ 1) maps in 13 ~ i have no wandering intervals
i=O
and try to show that this implies Inda. Throughout this section we will consider
properties of pullbacks of two intervals. The first of these intervals is Q~(k)-~J~<k):this is
the interval in M \ { c } such that
f(Q.<k)) = [ f(Jn(k-l)),f(Jn(k+ i)) ] 9
Similarly let
O_..<k)= Q.<k)0 [J.<k+l), "g(Jn(k + l)) ]"
In this section we will describe the structure of the unimodal pullback P0 ..... P.~k) of
intervals P.~k) which are contained in Q.~k) or in Q.~k).
In the next result it is shown that the intervals from the pullback of Q.~k) meet the
turning points in a periodic way.
8.1. STRUCTURE THEOREM. Let Pn(k)~Jn(k) be an interval which is contained in

Qn~k). Let m(0)<m(1)... <m(l)=n(k) be the cutting times of its unimodal pullback
Po ..... Pn~k) and let cj denote the turning point in Pmtj)" Then we have the following
properties.
(I) I f iE {0 ..... l - d } then Jm~o is a m ( i + d ) - I closest approach to ci;
(2) I f iE { l - d + l ..... l} then Jmto is a n(k) closest approach to ci;
i=0 ..... l - d and {Cl_d+ 1. . . . . el} are distinct;

(3) c i = c i + d f o r
(4) I f Jmo~+j=(Jmo~,r(Jm(i))) , i E {0 ..... l - 1 } and m(i)<m(i)+j<.n(k) then
P-(O+j= [Jm(i),Z(J.(i))];
(5) Pm(i+d)C[Jm03,r(Jm(i))]~Pm(i) for i=O, 1..... l - d and therefore fm(i+d)-m(O m a p s
Pm(O into itself.
Proof of(4). If property (4) does not hold then the closure of J~(o is contained in
the interior of Pm(o+j" SO the closure of Jm(o+,(k)-(m(O+j~is contained in the interior of P,(k)"
From the definition of P,(k) this implies m(i)+n(k)-m(i)-j>~n(k) and therefore j~<0, a
contradiction.
Proof o f (1) and (2). Let us just prove (1). Statement (2) is proved in exactly the
same way. Suppose by contradiction that there exists lE {0 ..... m ( i + d ) - l } such that
Jtc(Jm(o ' V(Jm<,3))" Then Jl+n(k)_m(i)=fn(k)-m(i)(Ji)c fn(k)-m(i)(Pm(i))cen(k) and l+m(i). Hence
l>m(i). From statement (4) we know that Pt=Pm(o. SO fl-m(O maps Pmo~ into Pt=Pm(o.
Because l<m(i+d) the map f: ~ it-m(o-~dr ~.* m ( i~) l l~ v t =,t-re(O-1CqU
~'at=0 O d ~,--m(O!~ has at most d - 1 turn-
ing points. Since Jm(o is a wandering interval of this map, we get a contradiction with
the induction hypothesis.
Proof of(3). Suppose there are i-d<j<.i<.s with ci=cj. From (1) we get that Jm(o
and Jm(j) are both m(j+d) closest approaches to ci=cj. Hence because re(i), m ( j ) <
m(j+d) this implies i=j. S i n c e f h a s precisely d turning points one gets that Ct_d+I ..... Ct
are distinct and that ci=ci+ d for iE (0 ..... l - d } .
Proof of (5). The proof of statement (5) follows immediately from the other state-
ments. Q.E.D.
The following theorem shows that we can even take monotone pullbacks of
intervals which contain topologically rather large sets. A unimodal version of this
theorem was already used by J. Guckenheimer for his proof of the non-existence of
wandering intervals for unimodal maps with negative Schwarzian derivative.
8.2. MONOTONE EXTENSION THEOREM. Let P,(k)=Q,(k) and {P0,Pl ..... P,(k)} be
its unimodal pullback. Let m(O)<m(1)<...<m(l)=n(k) be the cutting times and ci the
turning point in Proof"Let Hm(z~J be the maximal interval such that f m(~is monotone on
Hm(i). Let R,~(O be the component o f Pm(o\Jm(o which contains ci and Lm(o the other
component. I f the number l o f cutting times o f the pullback is at lest d+ 1 then
fra(i)(nm(i)) = [Lm0~, ci],
fm(i-l)(Hm(o) = [Lm(i_l), Ci-l)
and
fm(O-m(i-l)(Lmr 1)) c Lm(0
for i=0, 1..... l - d .
Proof. Let us first show that
(*) fm(i+ l)-rn(i)(Pm(i) ) ::3 [Jm(i+ l), Ci+l]
for i=0, 1,2 ..... l-d. Suppose by contradiction that there exists iE {0, 1..... l - d } with
Pm(i+l)~fm(~+~)-'n(O(Pm(o)9 Ci+l. By statement (5) of the previous theorem fm(i+,l)-m(O
maps Pm<o into itself. Now fm(~+l)-m(O(Pm(o)~)C~+1 implies that
m(i+d)-m(O-1
T= k3 ft(Pm(O)
t=0
does not contain c~.+~. Hence f maps T into itself and has at most d - 1 turning points.
Since JCPm( 0 it follows from the induction hypothesis that J is not a wandering interval,
a contradiction. This proves (*). Furthermore
(**) fm(i+l)-m(O: Lra(O....>Lm(i+D is monotone and onto
for i=0, 1,2 ..... l - d because otherwise there exists such an integer i with
fro(i+ l)-m(O(Rm(o) = Lm(i + 1)
and then as before

m(i+d)-m(O- l
T= U f'(Rm( o UJm(o)
t=0
contains at most d - 1 turning points and f maps this interval into itself. Since J is
contained in T this contradicts the induction hypothesis. It follows from (*) and (**)
that fm~i+l)-m(O maps [Lm(o, ci] monotonically over [Lm(i+l), ci+l]. The theorem clearly
follows. Q.E.D.
Next we give two results about the disjointness of unimodal pullbacks of intervals
in Qn(k) and in 0n(k)" The first result deals with the unimodal pullback of Qn(k).
8.3. THEOREM. Let m(0)<m(1)<...<m(l)=n(k) be the cutting times of the unimo-

dal pullback of Q,(k). Then:
(1) l<.d- 1.
(2) For every O<.j<~l the second successor sE(m(j)) is not defined.
(3) The intersection multiplicity of the unimodal pullback of Q.(k) is universally
bounded (in fact by 12d).
Proof. Suppose by contradiction that l~d. By statement (5) of Theorem 8.1 it

follows that Qm<o=Qm(k)is contained in Qm<l-d)" Hence f maps
m(l)-m(I-d)- 1
I.J ft(am(l_d) )
t=O
into itself and since Q.(k)contains no turning point this map has at most d - 1 turning
points. This contradicts Indd_ r So let us prove statement 2 by assuming by contradic-
tion that there exists j 6 { 0 , 1..... l} for which s(m(j)) and s(s(m(j))) are defined.
Because re(j) is n(k)-closest we get s(m(j))>n(k). Hence from Lemma 7.4 we get
that n(k+l)=s(n(k)) and s(n(k))-n(k)=s(j)-j. Because the closure of Jstm(j))is con-
tained in Qm(j) we get that the closure of J,(k+l)=f~(k)-m(J)(Jso,~j))) is contained in
fn~k)-m(J)(Qm(j))cQn(k) which contradicts the definition of Q,~k). So let us prove statement
(3). From statement (2) and the corollary of Theorem 6.4 it follows that the intersection
multiplicity of {Pm~j)..... Pm(j+~)} for j = - l , 0 , 1 ..... l - 1 (where we let m ( - 1 ) = 0 ) is
bounded by 11. Since l<.d-1 the theorem follows. Q.E.D.
8.4. THEOREM. Assume n(k)>n(k-1)+(n(k-1)-n(k-2)). Let m(0)<m(1)<...<

m(l)=n(k) be the cutting times of the unimodal pullback of Q~<k)and l>-2d. Then
s(n(k-1))=s(m(l-d)) and st(m(j)) are both not defined if j 6 { l - l d ..... l-ld+ d - 1 } and
l~ 2. Furthermore, if j E { l - l d ..... l - l d + d - 1 } , the intersection multiplicity of
{ am(j) ..... Om(j+ 1)}
is bounded by 4+2l. Similarly the intersection multiplicity of (Qo, Q~..... am(0)} is

bounded by 5+2[1/d]. (Here [l/d] is the smallest integer smaller than lid.)
Proof. If s(n(k-1)) were defined then Jn(k_2)C[JL(n(k_,)), Jn(k_l) ) and n(k)=s(n(k-1)).

Therefore, for a=n(k)-n(k- 1)=s(n(k- 1 ) ) - n ( k - 1),
J.~k-2)+a = [J.(k-,), J~<k)).

But since n(k-2)+a<n(k-1)+a=n(k) this implies that n(k-2)+(n(k)-n(k-1))=

n(k-2)+a must be equal to n(k-I) and this contradicts the assumption of the theorem.
Let us show that if l~>2 and j E {l-ld ..... l - l d + d - 1 ) then sl(m(j)) is not defined.
So suppose by contradiction sl(m(j)) exists. We claim that then si+l(m(j))>n(k-1)>
si(m(j)) for some i=0, 1, ..., l - 1. Indeed, since m(O)<n(k- 1) we may otherwise assume
that st-l(m(j))<n(k - 1). But then Jmcj+~t-l)a) is the (l-1)th successor of Jm~j) and, again
by Lemma 7.4, the successor of Jm~j+~t-1)a) must be between J,~j+~t-1)a) and ci. By
statement (2) of Theorem 8.1 this implies that st(m(j))>n(k - 1). This proves the claim.
Hence Lemma 7.4 and si+l(m(j))>n(k - l)>si(rn(j)) imply that s(n(k-1)) exists and so
we get a contradiction.
The disjointness statements immediately follow from the Corollary of Theorem
6.4. Q.E.D.
9. The non-existence of wandering intervals

In Section 7 we have proved Ind0 and so Theorem A follows from
Inda_ 1 =,. Ind a.
So let us assume that Inda_ 1 holds and that there exists a map f E ~a which has a
wandering interval J. Asume that J is maximal in the sense that J is not contained in
any strictly larger wandering intervals. From the Contraction Principle this implies that
In.I
----->1 if n---->~
IJI
where H, is the maximal interval containing J on which f " is monotone.
From Section 7 we know that J accumulates at a turning point, say c. Consider the
sequence of closest approach to c, {Jn<k)}k~>0"
9.1. Tr~EOREM. There exists ko such that for all k~ko
n ( k ) - n ( k - 1) ~<n(k- 1)-n(k-2).
COROLLARY. J is not a wandering interval.
Proof of Corollary. Since n(k)-n(k-1)<.n(k-1)-n(k-2) it follows that n(k)-

n(k-1) is eventually equal to some integer a for all k sufficiently large. In particular
since Jn~k-~) and J,r tends to c it follows that c is an attractive fixed point o f f which
attracts J. Hence J is not a wandering interval. Q.E.D.
Proof of Theorem 9.1. Since the intervals J,(k) are disjoint there exist arbitrarily
large k such that IJn(k_l)l>lJn(k)l. Let On(k) be as before. If IJ,(k)l<lJ,(k+l)l then
Because f i s non-fiat in the critical point c this implies that Q,(k)is a 1/2-
scaled neighbourhood of J,(k) (if k is large). By Theorem 8.3 the intersection multiplicity
of the unimodal pullback of Qn(k)is at most 12d and therefore the Koebe/Contraction
Principle gives a contradiction when n(k) is large. So we have shown that k~lJ~(k)l is
monotone decreasing for k large.
Let us now show that n(k)<~n(k-1)+(n(k-1)-n(k-2)) for k large. So assume by
contradiction n(k)> n(k- 1) + (n(k- 1) - n(k- 2)). Consider the unimodal pullback of 0,(k)"
If l<.2d then from Theorem 8.4 the intersection multiplicity of {Q0..... 0,(k)} is uniform-
ly bounded. Because Q,(k) contains a 1/2-scaled neighbourhood of J,(k) this gives a
contradiction with the Koebe/Contraction Principle.
This implies that for k large the number of cutting times I is at least 2d+ 1. Now let
Li and R; be as in Theorem 8.2. From Theorem 8.2 and for j = 1,2 .... l-d, there exists
an interval H=J which is mapped by fro(j) monotonically o n t o [Lm(j), Cj_d]. We claim
thatf"(J)(H) does not contain r(J,,(j)). Indeed, since Theorem 8.2 gives
fm(J)(H) ~ [Lm(j), Cm(j)]
one would otherwise have
fm(J)(H) ~ [L,n(j), ~(Jm(j))].
If we take jE {l-2d ..... l-d} then s2(m(j)) does not exist by Theorem 8.4, and since
fro(j) is monotone on H we can apply the Corollary of Theorem 6.4 and the intersection
multiplicity of H,f(H) ..... fm(J)(H) is at most 8. Furthermore, from the definition
0n(k), [Lm(j), cj-a] contains [J,~(j-a), cj] or it contains [r(Jm(j_a)), cj]. Since the length of
the closest approach intervals decreases one has tJm(j_a)l>[Jm(j)l and therefore fm(J)(H)
contains a 1/2-scaled neighbourhood offm(J)(J). Therefore we get a contradiction with
the Koebe/Contraction Principle and this proves the claim.
Now let Fro(j)= [Jm(j), r(Jm(j))]. By definition fm(J)lH is monotone and by Theorem
8 . 2 fm(j-1)(H)~[Jm(j_l),Cj_l) and fm(j)-m(j-l) m a p s [Jm(j_l),Cj_l) monotonically over
[Jm(j), cj]. By the previous claim, the image of [Jm(j-l), Cj-O under this map is contained
in Fm(j) and therefore we get
fm(j)-m(j-1)(Fm(j_l) ) c--Fm(j )
for all.iE {l-2d ..... l - d } . In particular
fn(k-l)-n(k-Z)(Fn(k_2)) = F.(k- I)"
Therefore
Since n(k-1)+(n(k-1)-n(k-2))>n(k-1) and J.<k) is a closest approach this gives

n(k- l ) + ( n ( k - 1)-n(k-2))>~n(k), a contradiction. Q.E.D.
10. The proof of Theorem B: finiteness of attractors
In this section we will prove Theorem B. If f." N--~N is a diffeomorphism then the
period of periodic orbits of f is bounded. So Theorem B holds trivially. So from now on
assume that f is not a diffeomorphism and that ffi O. Of course some points in Xf may
be attracted by periodic orbits so let ~ be an upper bound for the period of this
attracting orbits.
In this section we have to show that one has some expansion near periodic orbits.
For this it will be convenient to consider the orientation preserving period of a periodic
orbit. More precisely let ~7 be a periodic orbit of period k>max(ti, 300). Then p E ~7
implies Dfk(p)4=O. Let n=2k if Dfk(p)<O and n=k otherwise.
The main idea of the proof of Theorem B is to choose p E ~7and get on both sides of
p points 01 and 02, very close to p, with Dff(oi)>>-l+2~. Using the Minimum Principle
we will get Dfn(p)>>-l+Q for large n and we are done.
For p E ~7define Tp to be the maximal open interval such that both components of
T p \ { p } contain at most one point of •. (So the closure of Tp contains at most 5 points
of ~7.) The interval Tq, qE~, is a direct neighbours of Tp if TqNTp=O and cl(Tp) and
cl(Tq) have one point in common.
10.1. LEMMA. There exists a number r > 0 such that for each periodic orbit ~7of
period ~>max(t~, 300) there exists p E ~7such that:
(I) Tp has direct neighbours on both sides;
(2) [Tq~l~>2rlTpl and ITq~l~>2rlTplwhere Tq~ and Tq~ are the two direct neighbours
of Tp.
Proof. Let s E ~ be such that (i) #(cl(Ts)N 6 ) = 5 and (ii) ITs]<~ITq[ for all q with
#(cl(Tq) N~7)=5. If T~ has neighbours on both sides then we take p=s and we are done.
Otherwise N = [ - 1 , 1], and then let T t and T r be the smallest open intervals containing
points of respectively ( - 1} and ( 1} such that #(cl(T t) N r = g(cl(T r) 0 (7) = 5. Because
Ts has no two neighbours TscT t or T~cT r. Since the interval cl(T~) contains five points
of (7, and n is at most twice the period of p, there are at most 5 x 2 x ( 5 + 5 ) = I00 integers
t, O<~t<~nsuch t h a t f / maps a point in cl(T,) n r a point of cl(T r) Ucl(T t) A 6. So there
exists a 0<~t<~101 such that ft(cl(T,)A r is between T t and T r. Let p be the middle point
offt(cl(Ts) A 0). Let S = m a x ( I, SUpx~lvtDf(x)l}. Since Tpcft(Ts), we get
IZpl-< If'(Z~)l-< (sup IOf(x)l )' ITsl-< S'IT, I.

\x~N /
As t~<101 and ILI-<ITJ, i=1,2 the lemma is proved. Q.E.D.
Let p E ~7be the point from Lemma 10.1 and as before let n be the period or twice
the period of p. Let Un be the interval around Tp such that
(10. I) U~ is a 2v-scaled neighbourhood of Tp.
So Un=Tqt UTp OTq2. Let J be the maximal interval around p satisfying:
(10.2) fn(J) ~ U~
and
(10.3) fn[j is an orientation preserving diffeomorphism.

Let jt and j r be the components of J \ ( p } and let uin be the component of Un\{p}
which contains j i for iE {l, r}. Let ~ be as in the beginning of this section.
10.2. LEMMA. l f n > ~ w e have
(10.4) J c Tp,
(10.5) f~(ji) = j i , i = l, r.
Proof. We claim that JNP consists of at most 2 points. Indeed, take I to be the
maximal interval containing p such that OIcP and such that f~[I is a diffeomorphism.
Since f is not a diffeomorphism, I is an interval (i.e. not equal to sl). Then fkll is a
diffeomorphism for all k~>0 and since I is maximal, fi(1)fi I4=0 implies that fi(I)=L In
particular the boundary points of I (which are in P) cannot be mapped into int(I) and so
I contains at most two points of P. (This argument also shows that if I contains two
points of P then f~ interchanges these two points.) This proves statement (10.4). If
(10.5) does not hold then f n ( j i ) ~ j i and a critical point o f f is attracted by a periodic
orbit of period n. This implies n<.h. Q.E.D.
The main step in the proof of Theorem B is the proof of the following:
10.3. PROPOSITION. There exist a number Q>0 and an integer no such that if the
number n corresponding to p is greater than no then there exist oi E j i, i=l, r, such that
Df"(O ~)i> 1+2•.
Before proving Proposition 10.3 we will make a few remarks. Let r be the number
from Lemma 10.1. If there exists OiEJ i with Dfn(Oi)>-l+r then we are done with J~. So
from now on we may assume that
(10.6) O<Df"(x)< l + r , V x E J i.
By the previous lemma, for n>~ and i=l, r, J i ~ f " ( J i ) c u i . T h e n f " ( J i ) = u i is impossi-
ble because otherwise
If"(ji) l I> 2 r l L l + l J ' l > (1 +2r),

Isil IJ;I
a contradiction with (10.6). Hence in this case
f n ( j i ) c U i.
Let {U0..... Un} be the diffeomorphic pullback of U,~f"(J): Ui is the maximal interval
containing fi(j) which is mapped by f diffeomorphically into Ui+r Since fnlj is a
diffeomorphism this is well defined and Uk~fk(J). Furthermore from the maximality of
J one has Uo=J.
10.4. LEMMA. (i) There is a universal upperbound for the intersection multiplicity
of { Uo, U1..... Un} (in fact it is at most 74);
(ii) For every e>0 there exists no such that if n>no then IUkl<~efor all k=0, 1..... n.
Proof. Let Uko) N... N Uk(r)9 X with k(1)<...<k(r)<~n. Because fn-k(Uk)cint[Tq,, Tq2]
andf(Uk)=Uk+ l we get that
fn-k(r)(x) ~ Uk(l)+n_k(r) rl ... [1 U n.

Therefore #{k; Uk n int[Tql, Tq2]~=O}>~r.Hence statement (i) follows from
(10.7) ~{k[Uk n int[Tq,, Tq~], ~} <~74.
Let us prove (10.7). Notice that int[Tql, Tqfl contains 11 points of 6. So there are at most
22 integers O<.k<n such that fk(p)E int[Tqi, Tq~]. Hence there are at most 22 integers
O<.k<n such that U, cint[Tql, Tq2]. Now let a and b be the boundary points of [Tq~,Tq2],
so a, b E 6. If U, is not contained in int[Tql, Tq:] but has a non-empty intersection with
this set then f"-k(a) or f"-k(b)Eint[Tql, Tqfl. Because cl([Tq,, Tq~]) contains only 13
points of 6, there exist at most 2 x 2 x 13=52 integers O<_k<n with this property. This
implies inequality (10.7) and finishes the proof of (i).
Statement (ii) follows from the Contraction Principle and the fact that Uk contains
at most 5 points of 6. Q.E.D.
Proof of Proposition 10.3. Let CE(0, 1) be so that if T=I are intervals, fiT is an
diffeomorphism and f(T) is an e-scaled neighbourhood of f(I) then T is a Ce-scaled
neighbourhood of I. Since f is non-fiat at its critical points such a constant exists.
Let m(1)<m(2)<...<m(l) be the 'cutting' times of the pullback, i.e., the integers
for which Uj contains a turning point in its closure. Let Lj and Rj be the components of
Uj\fi(J j) and ~ = f i ( f ) and let Rm(k) be the component which contains Ck in its bound-
ary. Now U. contains a r scaled neighbourhood off"(Ji). Since f.-m(t)+l: Um(l)+l..._~Un is
a diffeomorphism it follows from the Macroscopic Koebe Principle that there exists a
positive function B0 (which only depends on f a n d the intersection multiplicity 74 from
Lemma 10.6) such that Um~t)+lcontains a B0(r)-scaled neighbourhood of Jm(I)+l" i NOW
i
let g(x)=C'Bo(x). If Um~n contains a C.Bo(z)=g(r)-scaled neighbourhood of Jm(I)we
repeat this procedure and we get from the Macroscopic Koebe principle again that
Urn(l_1)+1contains a B0(g(r)))-scaled neighbourhood of Jm(H)+r i If U,.(l_l) contains a
gZ(r)= C. B0(g(r)))-scaled neighbourhood of J~t-1) then we repeat this procedure again.
Since U=J this procedure must stop however. Say it stops at m(r) where r<.l and then
(by the definition of C above)
IRm(r)[< gr (r)lJ imCr)l'
where, since intersection multiplicity is at most 74, one has r<.74.#Kf.

Now let M' be the middle third interval of Jm(r) i and M c J i be such that
fm~r)(M)=M'. From the Macroscopic Koebe principle ji is b-scaled neighbourhood of
M. From Lemma 10.4, IU, I is small if n is large. So we can apply Lemma 2.8 and get
some universal constant ~>0 such that
B(f, fm~r)(Ji),M') I> 1+~.

Let 2 be such that (1-2)2(1 +~)~>1 + ~ . From the disjointness property of the orbit o f J ~
we get
B(fm('),Ji, M) >I 1-2,

B(fn-m(~)-l,f~(~)+l(Ji),f(M')) >t 1-~.,
for n large enough. Hence
B(f",J i, M) >I I + + ~ ,
for n large enough. Because of the First Expansion Principle, see Theorem 1.3, it
suffices to show that the length of both components o f j i ~ M is at least 6-]J;I. But since
fn-m(r)-I has bounded dlstorUon
" " onf(Jm(r)),
i sincefis non-flat at critical points and since
i
the components of Jm(r)~g have the same length as M', there exists a universal
constant fl such that the length of both components o f f n (J i X.M)=f
__ n - r a ( r )
(Jm(r)\M)
i t
is
at least ft. [fn(ji)l. However, by Lemma 10.2, f n ( j ~ j i and by assumption ]Dfnl~<l+z -
on ji. It follows that both components of JiN~M have at least size (fl/(l+r)).lJi[.
Q.E.D.
Proof of Theorem B. Let ~7and as before let n bet he period or twice the period of
8. Assume that n>no where no is as in Proposition 10.3. So there exist two points 01, 02
such that p E T=[O1, 02],
Df"(Oi) I> 1+2Q, i = l, r.

For n large, Ifi(T)l is small for all iE{0, 1..... n} and the orbits has intersection
multiplicity ~<74. Therefore we get
[B(f n, T*, J*)] 3~>1~200
for all intervals J*cT*cT, provided n is large. Now we apply the Minimum Principle.
Then
IDf"(x)] >I[infB(f n, T*, j,)]3 (I +2Q) ~> 1+Q

for all xE T. So Df"(p)~l+Q. This proves Theorem B. Q.E.D.
JULIA--FATOU--SULLIVANTHEORY FOR REAL ONE-DIMENSIONAL DYNAMICS 317
R e f e r e n c e s
[BL] BLOKa, A. M. & LYUmCH, M. Ju., Non-existence of wandering intervals and struc-
ture of topological attractors of one dimensional dynamical systems 2. The smooth
case. Ergodic Theory Dynamical Systems, 9 (1989), 751-758.
[CE] COLLET, P. & ECKMANN, J., Iterated Maps of the Interval as Dynamical Systems.
Birkhfiuser, 1980.
[131 DEmov, A., Sur les courbes definie par les equations differentielles a la surface du
tore. J. Math. Pures Appl. 9, 11 (1932), 333-375.
[Fa] FATOU, P., Sur les equations fonctionnelles. Bull. Soc. Math. France, 47 (1919),
161-271; 48 (1920), 33-94; 48 (1920), 208-314.
[Gul] GUCKENHEIMER, J., Sensitive dependence on initial conditions for one dimensional
maps. Comm. Math. Phys., 70 (1979), 133-160.
[Gu2] Limit sets of S-unimodal maps with zero entropy. Comm. Math. Phys., 110
- -
(1987), 655-659.
[Ha] HALL, C. R., A C | Denjoy counterexample. Ergodic Theory Dynamical Systems, 1
(1981), 261-272.
[Jl JULIA, G., Memoire sur l'iteration des fonctions rationnelles. J. de Math., 8 (1918),
47-245.
[L] LVOBICH, M. Ju., Non-existence of wandering intervals and structure of topological
attractors of one dimensional dynamical systems 1. The case of negative Schwar-
zian derivative. Ergodic Theory Dynamical Systems, 9 (1989), 737-750.
[Ma] MA~, R., Hyperbolicity, sinks and measure in one dimensional dynamics. Comm.
Math. Phys., 100 (1985), 495-524; Erratum. Comm. Math. Phys., 112 (1987),
721-724.
[Me] DE MELO, W., A finiteness problem for one dimensional maps. Proc. Amer. Math.
Soc., 101 (1987), 721-727.
[Mi] MISIUREWICZ, M., Absolutely continuous measure for certain maps of an interval.
Publ. Math. LH.E.S., 53 (1981), 17-51.
[MS1] DE MELO, W. & VANSTVaEN,S. J., A structure theorem in one-dimensional dynamics.
Ann. of Math., 129 (1989), 519-546.
[MS2] Schwarzian derivative and beyond. Bull. Amer. Math. Soc., 18 (1988), 159-162.
- -
[MMMS] MARTENS,M., DE MELO,W., MENDES,P. & VANSTRIEN,S., Cherry flows on the torus:
towards a classification. Ergodic Theory Dynamical Systems, 10 (1990), 531-554.
[MT] MILNOR, J. & THURSTON, W., On iterated maps of the interval: I, II. Preprint
Princeton, 1977.
[Si] SINGER, D., Stable orbits and bifurcations of maps of the interval. SlAM J. Appl.
Math., 35 (1978), 260-267.
[Sc] SCHWARTZ, A., A generalization of a Poincarr--Bendixon theorem to closed two
dimensional manifolds. Amer. J. Math., 85 (1963), 453--458.
[si] SI-~RKOVSraL A. N. & IVANOV, A. F., C| of an interval with attracting
cycles of arbitrary large periods. Ukrain. Mat. Zh., 4 (1983), 537-539.
[Strl] VAN STmEN, S., On the bifurcations creating horseshoes, in Dynamical Systems and
Turbulence, Warwick, 1980. Lecture Notes in Mathematics, 898 (1981), 316-351.
Springer-Verlag, Berlin.
[Str2] - - Smooth dynamics on the interval, in New Direction in Chaos. Cambridge Univer-
sity Press, Cambridge, 1987, pp. 57-119.
[Str3] Hyperbolicity and invariant measures for general C2 interval maps satisfying the
- -
Misiurewicz condition. Comm. Math. Phys., 128 (1990), 437-496.

[Su] SULLIVAN, D., Quasiconformal homeomorphisms and dynamics I: a solution of
Fatou-Julia problem on wandering domains. Ann. of Math., 122 (1985), 401-418;

II: Structural stability implies hyperbolicity for Kleinian groups. Acta Math., 155
(1985), 243-260.
[Y] Yoccoz, J. C., Il n'y a pas de contre-examples de Denjoy analytique. C. R. Acad. Sci.
Paris S~r. I Math., 298 (1984), 141-144.
Received January 19, 1989

Received in revised form January 7, 1991

BF 02392981

Uploaded by

Copyright:

Available Formats

BF 02392981

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

BF 02392981

Uploaded by

Copyright:

Available Formats

Acta Math.

, 168 (1992), 273-318

M. MARTENS(1) w. DE MELO and S. VAN STRIEN

Technical University of Delft IMPA Technical University of Delft

(1) Supported by N.W.O.

The above corollary is the analogue of Mafir-Sad-Sullivan Theorem on the density

COROLLARY. Let f : [ - 1 , I]--->[-1, 1] be a C 2 unimodal map which is infinitely

Proof. By Theorem B the turning point o f f cannot be approximated by periodic

w1. Definitions and statement of the theorems

The non-flatness conditions in Theorems A and B cannot be dropped. Indeed if

More precisely we will prove

f ( x ) = +ldp(x)la+f(xo), VxE U(xo).

This number a is called the order of XoE Kf.

Instead of Theorem B we will prove the following stronger result:

f(x) = +f~b(x)ta+f(xo), V x e U(xo).

The numbers no and 0 from Theorem B obviously depend in an essential way on

(ii) p is in the boundary of the immediate basin of a periodic attractor which

2. Analytical estimates on diffeomorphic branches o f f n when f E ~:

2.1. Definition. Let J e T be open and bounded intervals in N such that T \ J

B(g, T, J) = D(g(T), g(J))

We notice that if f "IT is monotone and continuous then

Of course this cross-ratio is related to the hyperbolic metric. Indeed, let T be an

2.2. Definition. Let g: T--~N be a C3-map on the interval T=N. Then

is called the Schwarzian derivative of g.

2.3. PROPOSITION. Let g: T-*N be a C3 diffeomorphism on the open interval T e N

Furthermore if Sg(x)<.O, Vx E T, then B(g, T, I)>~1.

In particular a diffeomorphism g: T--->T' having negative Schwarzian derivative

2.4. THEOREM. Let f E ~. Then there exists a bounded continuous function

Here lr=maxi=o..... m-l lfi(r)l 9

Proof. Since B ( f n, T, J)=l-l~__201B(f,fi(T),fi(J)), it suffices to show that there exists

B(f, T, J) ~>exp{-o(ITI). Irl)

19-928283 Acta Mathematica 168. Imprim6 le 24 avril 1992

(2.2) B(f, ~r,j ) - 1/> -o(l~q). ITI

Case 1. T c N \ V. In this case write T=[a, d] and M=[b, c]. Then

where K=inf{ [Df(x)l; x E N \ V}. Writing

f ( a + x) = f(a) +It(a, x). x

f ( c ) - f ( b ) = It(a, b-a)-~ It(a' c - a ) - I t ( a , b - a ) . ( c - a )

= -_J_l2.K ic_al, ix(a, d - a ) . Ix(a, c-a)-ix(a,c_b b-a)

-ix(a, b-a) 9 ix(a, d-a)-ix(a,d_c c-a) .

-ix(a, b - a ) . ix(a, d-a)-ix(a, c-a) .

Case 2. T~ U. Since a ~ I, the map ~ ( x ) = x ~ has Schwarzian derivative <~0. Then

B(f, T, U)= B(dp,, dp(T), dp(U)).B(dp, T, U)>>-B(dp, T, U).

Since q~ is C 2 and inf{lDep(x)l;x E T} is bounded from below, B(r T, U)~>I-o(ITI)-ITI

Case 3. If T contains a component of U \ V then (B(f, T, M ) - 1)/IT I is bounded

Remark. We should emphasise that it is really essential in this theorem t h a t f i s C 2.

B ( f m, T,J)>~exp -Co. x If"(T)l 2 ~>exp -Co.r. fi(T) .

2.5. "MINIMUM PRINCIPLE." Let Te-N and g: T---~g(T)~N be a C l diffeomor-

B(g, T*, J*) >I C> 0

(2.3) [Og(x)l ~ C a min{IOg(a)l, IOg(b)l}.

Bo(g, T*) = Ig(T*)12 1

Bo(g, T*) = lim B(g, T*,J), Bl(g, T,x) = limB(g, T,J).

Hence Bo(g, L), Bo(g, R), Bl(g, T, x)~>C>0. Since

Bo(g, L), Bo(g, R) >IC

Cl ,,a,I IO.(x,I. Cl ,,x,I Io,( 'l

Since Bl(g, T, x)~>C>0 we have

lOg(x)[ ~ >I C Ig(Z)l Ig(R)I

min/lg(t)l Ig(R)l~ ~< Ig(T)l ~<max~lg(t)l Ig(R)ll

C2(Ig(L)I/ILI.Ig(R)I/IRI~2 C2min~(ig(L)[~2, (ig(R)l)~

Hence IDg(x) l>~C 3min( IDg(a)l IOg(b) l}. Q.E.D.

B(g, T, J) >I C> 0

Bo(g, T) = Ig(T)12 1

Bo(g, T) = lim B(g, T,J), Bl(g, T,x) = limB(g, T,J).

B(g, T, J) >I C> O.