(London Mathematical Society Lecture Note Series 274) Tan Lei - The Mandelbrot Set, Theme and Variations-Cambridge University Press (2000)

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES
Managing Editor: Professor N.J. Hitchin, Mathematical Institute,

University of Oxford, 24-29 St Giles, Oxford 0X1 3LB, United Kingdom
The titles below are available from booksellers, or, in case of difficulty, from Cambridge University Press.
46 p-adic Analysis: a short course on recent work, N. KOBLITZ

59 Applicable differential geometry, M. CRAMPIN & F.A.E. PIRANI
66 Several complex variables and complex manifolds II, M.J. FIELD
86 Topological topics, I.M. JAMES (ed)
88 FPF ring theory, C. FAITH & S. PAGE
90 Polytopes and symmetry, S.A. ROBERTSON
93 Aspects of topology, I.M. JAMES & E.H. KRONHEIMER (eds)
96 Diophantine equations over function fields, R.C. MASON
97 Varieties of constructive mathematics, D.S. BRIDGES & F. RICHMAN
99 Methods of differential geometry in algebraic topology, M. KAROUBI & C. LERUSTE
100 Stopping time techniques for analysts and probabilists, L. EGGHE
105 A local spectral theory for closed operators, I. ERDELYI & WANG SHENGWANG
107 Compactification of Siegel moduli schemes, C.-L. CHAI
109 Diophantine analysis, J. LOXTON & A. VAN DER POORTEN (eds)
113 Lectures on the asymptotic theory of ideals, D. REES
114 Lectures on Bochner-Riesz means, K.M. DAVIS & Y.-C. CHANG
116 Representations of algebras, P.J. WEBB(ed)
119 Triangulated categories in the representation theory of finite-dimensional algebras, D. HAPPEL
121 Proceedings of Groups - St Andrews 1985, E. ROBERTSON & C. CAMPBELL (eds)
128 Descriptive set theory and the structure of sets of uniqueness, A.S. KECHRIS & A. LOUVEAU
130 Model theory and modules, M. PREST
131 Algebraic, extremal & metric combinatorics, M.-M. DEZA, P. FRANKL & I.G. ROSENBERG (eds)
138 AnalysisatUrbana.il, E. BERKSON, T. PECK, & J. UHL (eds)
139 Advances in homotopy theory, S. SALAMON, B. STEER & W. SUTHERLAND (eds)
140 Geometric aspects of Banach spaces, E.M. PEINADOR & A. RODES (eds)
141 Surveys in combinatorics 1989, J. SIEMONS (ed)
144 Introduction to uniform spaces, I.M. JAMES
146 Cohen-Macaulay modules over Cohen-Macaulay rings, Y. YOSHINO
148 Helices and vector bundles, A.N. RUDAKOV et al
149 Solitons, nonlinear evolution equations and inverse scattering, M. ABLOWITZ & P. CLARKSON
150 Geometry of low-dimensional manifolds 1, S. DONALDSON & C.B. THOMAS (eds)
151 Geometry of low-dimensional manifolds 2, S. DONALDSON & C.B. THOMAS (eds)
152 Oligomorphic permutation groups, P. CAMERON
153 L-functions and arithmetic, J. COATES & M.J. TAYLOR (eds)
155 Classification theories of polarized varieties, TAKAO FUJITA
158 Geometry of Banach spaces, P.F.X. MULLER & W. SCHACHERMAYER (eds)
159 Groups St Andrews 1989 volume 1, CM. CAMPBELL & E.F. ROBERTSON (eds)
160 Groups St Andrews 1989 volume 2, CM. CAMPBELL & E.F. ROBERTSON (eds)
161 Lectures on block theory, BURKHARD KULSHAMMER
163 Topics in varieties of group representations, S.M. VOVSI
164 Quasi-symmetric designs, M.S. SHRIKANDE & S.S. SANE
166 Surveys in combinatorics, 1991, A.D. KEEDWELL (ed)
168 Representations of algebras, H. TACHIKAWA & S. BRENNER (eds)
169 Boolean function complexity, M.S. PATERSON (ed)
170 Manifolds with singularities and the Adams-Novikov spectral sequence, B. BOTVINNIK
171 Squares, A.R. RAJWADE
172 Algebraic varieties, GEORGE R. KEMPF
173 Discrete groups and geometry, W.J. HARVEY & C MACLACHLAN (eds)
174 Lectures on mechanics, J.E. MARSDEN
175 Adams memorial symposium on algebraic topology 1, N. RAY & G. WALKER (eds)
176 Adams memorial symposium on algebraic topology 2, N. RAY & G. WALKER (eds)
177 Applications of categories in computer science, M. FOURMAN, P. JOHNSTONE & A. PITTS (eds)
178 Lower K-and L-theory, A.RANICKI
179 Complex projective geometry, G. ELLINGSRUD et al
180 Lectures on ergodic theory and Pesin theory on compact manifolds, M. POLLICOTT
181 Geometric group theory I, G.A. NIBLO & M.A. ROLLER (eds)
182 Geometric group theory II, G.A. NIBLO & M.A. ROLLER (eds)
183 Shintani zeta functions, A. YUKIE
184 Arithmetical functions, W. SCHWARZ & J. SPILKER
185 Representations of solvable groups, O. MANZ & T.R. WOLF
186 Complexity: knots, colourings and counting, D.J.A. WELSH
187 Surveys in combinatorics, 1993, K. WALKER (ed)
188 Local analysis for the odd order theorem, H. BENDER & G. GLAUBERMAN
189 Locally presentable and accessible categories, J. ADAMEK & J. ROSICKY
190 Polynomial invariants of finite groups, D.J.BENSON
191 Finite geometry and combinatorics, F. DE CLERCK et al
192 Symplectic geometry, D. SALAMON (ed)
194 Independent random variables and rearrangement invariant spaces, M. BRAVERMAN
195 Arithmetic of blowup algebras, WOLMER VASCONCELOS
196 Microlocal analysis for differential operators, A. GRIG1S & J. SJOSTRAND
197 Two-dimensional homotopy and combinatorial group theory, C. HOG-ANGELONl et al
198 The algebraic characterization of geometric 4-manifolds, J.A. HILLMAN
199 Invariant potential theory in the unit ball of C", MANFRED STOLL
200 The Grothendieck theory of dessins d'enfant, L. SCHNEPS (ed)
201 Singularities, JEAN-PAUL BRASSELET (ed)
202 The technique of pseudodifferential operators, H.O. CORDES
203 Hochschild cohomology of von Neumann algebras, A. SINCLAIR & R. SMITH
204 Combinatorial and geometric group theory, A.J. DUNCAN, N.D. GILBERT & J. HOWIE (eds)
205 Ergodic theory and its connections with harmonic analysis, K. PETERSEN & I. SALAMA (eds)
207 Groups of Lie type and their geometries, W.M. KANTOR & L. DI MARTINO (eds)
208 Vector bundles in algebraic geometry, N.J. HITCHIN, P. NEWSTEAD & W.M. OXBURY (eds)
209 Arithmetic of diagonal hypersurfaces over finite fields, F.Q. GOUVEA & N. YUI
210 Hilbert C*-modules, E.C. LANCE
211 Groups 93 Galway / St Andrews I, C M . CAMPBELL et al (eds)
212 Groups 93 Galway / St Andrews II, C M . CAMPBELL et al (eds)
214 Generalised Euler-Jacobi inversion formula and asymptotics beyond all orders, V. KOWALENKO et al
215 Number theory 1992-93, S. DAVID (ed)
216 Stochastic partial differential equations, A. ETHERIDGE (ed)
217 Quadratic forms with applications to algebraic geometry and topology, A. PFISTER
218 Surveys in combinatorics, 1995, PETER ROWLINSON (ed)
220 Algebraic set theory, A. JOYAL & I. MOERDIJK
221 Harmonic approximation, S.J. GARDINER
222 Advances in linear logic, J.-Y. GIRARD, Y. LAFONT & L. REGNIER (eds)
223 Analytic semigroups and semilinear initial boundary value problems, KAZUAKITAIRA
224 Computability, enumerability, unsolvability, S.B. COOPER, T.A. SLAMAN & S.S. WAINER (eds)
225 A mathematical introduction to string theory, S. ALBEVERIO, J. JOST, S. PAYCHA, S. SCARLATTI
226 Novikov conjectures, index theorems and rigidity I, S. FERRY, A. RANICKI & J. ROSENBERG (eds)
227 Novikov conjectures, index theorems and rigidity II, S. FERRY, A. RANICKI & J. ROSENBERG (eds)
228 Ergodic theory of Zd actions, M. POLLICOTT & K. SCHMIDT (eds)
229 Ergodicity for infinite dimensional systems, G. DA PRATO & J. ZABCZYK
230 Prolegomena to a middlebrow arithmetic of curves of genus 2, J. W.S. CASSELS & E. V. FLYNN
231 Semigroup theory and its applications, K.H. HOFMANN & M.W. MISLOVE (eds)
232 The descriptive set theory of Polish group actions, H. BECKER & A.S. KECHRIS
233 Finite fields and applications, S. COHEN & H. NIEDERREITER (eds)
234 Introduction to subfactors, V. JONES & V.S. SUNDER
235 Number theory 1993-94, S. DAVID (ed)
236 The James forest, H. FETTER & B. GAMBOA DE BUEN
237 Sieve methods, exponential sums, and their applications in number theory, G.R.H. GREAVES et al
238 Representation theory and algebraic geometry, A. MARTSINKOVSKY & G. TODOROV (eds)
239 Clifford algebras and spinors, P. LOUNESTO
240 Stable groups, FRANK O. WAGNER
241 Surveys in combinatorics, 1997, R.A. BAILEY (ed)
242 Geometric Galois actions I, L. SCHNEPS & P. LOCHAK (eds)
243 Geometric Galois actions II, L. SCHNEPS & P. LOCHAK (eds)
244 Model theory of groups and automorphism groups, D. EVANS (ed)
245 Geometry, combinatorial designs and related structures, J.W.P. HIRSCHFELD et al
246 p-Automorphisms of finite p-groups, E.I. KHUKHRO
247 Analytic number theory, Y. MOTOHASHI (ed)
248 Tame topology and o-minimal structures, LOU VAN DEN DRIES
249 The atlas of finite groups: ten years on, ROBERT CURTIS & ROBERT WILSON (eds)
250 Characters and blocks of finite groups, G. NAVARRO
251 Grobner bases and applications, B. BUCHBERGER & F. WINKLER (eds)
252 Geometry and cohomology in group theory, P. KROPHOLLER, G. NIBLO, R. STOHR (eds)
253 The ^-Schur algebra, S. DONKIN
254 Galois representations in arithmetic algebraic geometry, A.J. SCHOLL & R.L. TAYLOR (eds)
255 Symmetries and integrability of difference equations, P. A. CLARKSON & F.W. NIJHOFF (eds)
256 Aspects of Galois theory, HELMUT V6LKLEIN et al
257 An introduction to noncommutative differential geometry and its physical applications 2ed, J. MADORE
258 Sets and proofs, S.B. COOPER & J. TRUSS (eds)
259 Models and computability, S.B. COOPER & J. TRUSS (eds)
260 Groups St Andrews 1997 in Bath, I, C M . CAMPBELL et al
261 Groups St Andrews 1997 in Bath, II, C M . CAMPBELL et al
263 Singularity theory, BILL BRUCE & DAVID MOND (eds)
264 New trends in algebraic geometry, K. HULEK, F. CATANESE, C. PETERS & M. REID (eds)
265 Elliptic curves in cryptography, I. BLAKE, G. SEROUSSI & N. SMART
267 Surveys in combinatorics, 1999, J.D. LAMB & D.A. PREECE (eds)
268 Spectral asymptotics in the semi-classical limit, M. DIMASSI & J. SJOSTRAND
269 Ergodic theory and topological dynamics of group actions on homogeneous spaces, B. BEKKA & M. MAYER
270 Analysis on Lie Groups, N. T. VAROPOULOS & S. MUSTAPHA
London Mathematical Society Lecture Note Series. 274
The Mandelbrot Set,

Theme and Variations
Edited by
Tan Lei
Universite de Cergy-Pontoise
| CAMBRIDGE
UNIVERSITY PRESS
cambridge university press
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo, Mexico City
Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by

Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521774765
© Cambridge University Press 2000
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2000
A catalogue record for this publication is available from the British Library

isbn 978-0-521-77476-5 Paperback
Cambridge University Press has no responsibility for the persistence or

accuracy of URLs for external or third-party internet websites referred to in
this publication, and does not guarantee that any content on such websites is,
or will remain, accurate or appropriate. Information regarding prices, travel
timetables, and other factual information given in this work is correct at
the time of first printing but Cambridge University Press does not guarantee
the accuracy of such information thereafter.
Contents
Introduction vii
Tan L.
Preface xiii
J. Hubbard
The Mandelbrot set is universal 1
C. McMullen
Baby Mandelbrot sets are born in cauliflowers 19
A. Douady, with the participation of X. Buff, R. Devaney &; P. Sentenac
Modulation dans l'ensemble de Mandelbrot 37
P. Haissinsky
Local connectivity of Julia sets: expository lectures 67
J. Milnor
Holomorphic motions and puzzles (following M. Shishikura) 117
P. Roesch
Local properties of the Mandelbrot set at parabolic points 133
Tan L.
Convergence of rational rays in parameter spaces 161
C. Petersen and G. Ryd
Bounded recurrence of critical points and Jakobson's Theorem 173
S. Luzzatto
The Herman-Swiatek Theorems with applications 211
C. Petersen
Perturbation d'une fonction linearisable 227
H. Jellouli
Indice holomorphe et multiplicateur 253
H. Jellouli
vi Contents
An alternative proof of Marie's theorem on non-expanding Julia sets 265

M. Shishikura and Tan L.
Geometry and dimension of Julia sets 281
Yin Y.-C.
On a theorem of M. Rees for the matings of polynomials 289
M. Shishikura
Le theoreme d'integrabilite des structures presque complexes 307
A. Douady, d'apres des notes de X. Buff
Bifurcation of parabolic fixed points 325
M. Shishikura
References 365
Introduction
Tan Lei
Complex dynamical systems is a fascinating field. It has both the visual

appeal of endlessly varying and beautiful fractals, and the intellectual appeal
of sophisticated mathematical developments. Its theoretical challenges have
attracted mathematicians from around the world.
A central object of study in complex dynamics is the Mandelbrot set M, a
fractal shape that classifies the dynamics of the quadratic polynomials fc(z) —
z2 + c. The Mandelbrot has a remarkably simple definition:
M = {c £ C : f™(z) remains bounded as n —> oo.}.
Nevertheless M exhibits a rich geometric and combinatorial structure, with

many intriguing details and many remaining mysteries. Although it is defined
in terms of quadratic polynomials, the Mandelbrot set reappears in virtually
every other family of rational maps, as can be observed in computer experi-
ments.
The mathematical theory of the Mandelbrot set and related objects has
undergone rapid development during the last two decades, sparked by the
pioneering work of Douady and Hubbard. This volume provides a coherent
perspective on the present state of the field.
Its articles range from the systematic exposition of current knowledge
about the Mandelbrot set, to the latest research in complex dynamics. In
addition to presenting new work, this collection documents for the first time
important results hitherto unpublished or difficult to find in the literature.
1 A detailed description of the contents of

this volume
The Preface, by J. Hubbard, gives an intriguing first-hand account of devel-
opments in the field during the period 1976-1982. It recounts the discovery
of the Mandelbrot set and the introduction of basic tools, such as the Rie-
mann mapping to the exterior of M, Hubbard trees, quasiconformal map-
pings, polynomial-like mappings, holomorphic motions, Thurston's theory,
etc.
Part one: Universality of the Mandelbrot set
With the advent of computers, it has been easy to experimentally observe
copies of Mandelbrot set in many families of complex analytic dynamical
systems, but it was a mathematical challenge to explain this universality.
vii
viii Tan Lei
The first results in this direction were obtained by Douady-Hubbard using

polynomial-like mappings. This volume contains three articles about the
subject:
[McMullen] shows that small Mandelbrot sets are dense in the bifurcation
locus for any holomorphic family of rational maps. As a consequence, one
obtains general estimates for the dimension of the bifurcation locus, using
Shishikura's result that the Hausdorff dimension of the boundary of M is 2.
[Douady-Buff-Devaney-Sentenac] studies the birth of infinitely many baby
Mandelbrot sets in parameter space near a rational map with a parabolic
point. This paper shows each baby Mandelbrot set sits in the heart of a nest
of imploded cauliflowers.
[Hai'ssinsky] explains why small copies of the Mandelbrot set appear within
M itself.
Part two: Quadratic Julia sets and the Mandelbrot set
One of the central open questions concerning the Mandelbrot set is the
following: is M a locally connected set? A positive answer would imply
that there is a complete topological description of M, and in particular that
hyperbolic dynamical systems are dense in the quadratic family. Similarly,
one can completely answer questions about the topology of J(/ c ) when the
Julia set is known to be locally connected.
A major breakthrough, due to Yoccoz, states that J(/ c ) is locally con-
nected and M is locally connected at c, for certain special values of c E M.
These are the values where fc has no indifferent points and is not renorrnal-
izable. Yoccoz's original proof is not published, but this volume includes two
articles, [Milnor] and [Roesch], treating its central points:
[Milnor] shows that the Julia set of fc is locally connected, using Yoccoz
puzzles;
[Roesch] continues from [Milnor] and shows that M is locally connected
at such parameters c. The proof given is a new argument, due to Shishikura,
using holomorphic motions.
Turn to the case where fc does have an indifferent point, [Tan] proves that
M is locally connected when fc has a parabolic point of multiplier 1. This
paper also surveys more general results on the local connectivity of M when
c has an indifferent point.
The landing behavior of external rays is often a starting point in the
study of local connectivity. In this volume, [Petersen-Ryd] gives a new and
elementary proof that external rays to M with rational angle do land, and
describes the dynamics of fc for the endpoint c.
Dynamical systems on the interval or the circle are closely related to dy-
namics in one complex variable. This volume includes two papers on real
dynamics:
[Luzzatto] presents an annoted account of Jakobson's theorem, stating that
Introduction ix
the set of c 6 E such that fc{z) has chaotic dynamics has positive Lebesgue
measure. More precisely, c = - 2 is a point of Lebesgue density of the set of
c 6 R such that fc has an absolutely continuous invariant measure;
[Petersen] presents the Herman-Swiatek's theorem, stating that an ana-
lytic circle homeomorphism with a critical point and irrational rotation num-
ber 6 is quasi-symmetrically conjugate to a rigid rotation if and only if 9
is of bounded type. Using this theorem, [Petersen] also establishes the new
result that there exist quadratic Siegel polynomials with a rotation numbers
of unbounded type whose Julia sets are locally connected.
The delicate theory of bifurcations of indifferent period points is studied
in two related papers:
[Jelloulil] shows that small perturbations of quadratic Siegel polynomials
with Diophantine rotation number are conjugate to their linear parts up to
an arbitrarily small error term. This work bears on the study of the Lebesgue
measure of nearby filled Julia sets;
[Jellouli2] discusses various important quantities and their relations ap-
pearing in the perturbation of a quadratic polynomial with a parabolic fixed
point. These include the multiplier of the periodic orbit coming from the
perturbation, its first and second derivatives with respect to the parameter,
and the holomorphic indices.
Part three: Julia set of rational maps
A rational map / tends to be expanding on its Julia set J{f), away from
the orbits of recurrent critical points. This phenomenon, observed already by
Fatou and Julia, was made precise in a strong form by Ricardo Mane. In this
volume:
[Shishikura-Tan] presents a new proof of Mane's result, showing that any
point x € */(/), which is neither a parabolic periodic point nor a limit point
of a recurrent critical orbit, has a neighborhood which is contracted by the
family {/-"} n e N .
[Yin] shows, as an application, that if all critical points in J(f) are non-
recurrent, then •/(/) is 'shallow' or 'porous'; consequently its Hausdorff di-
mension is less than two (as first shown by Urbanski).
The combinatorics of general rational maps, even of degree two, is still not
well-understood. However many interesting rational maps can be constructed
by mating pairs of polynomials of the same degree. If the polynomials are crit-
ically finite, then the existence of their mating can be reduced to a topological
problem using Thurston's theory. In this volume:
[Shishikural] provides a proof of a result, first developed by Mary Rees,
showing that if a mating / exists combinatorially, then in fact / is topologi-
cally conjugate to a natural quotient of the dynamics of the two polynomials.
Part four: Foundational results
x Tan Lei
[Douady] gives a new proof of Ahlfors-Bers theorem on integrability of

measurable complex structures. This theorem is of central importance for
the surgery and deformation of holomorphic dynamical systems.
[Shishikura2] gives a thorough treatment of the theory of parabolic im-
plosions (developed by Douady and Lavaurs from Ecalle-Voronin's theory).
This powerful tool provides precise analytic information about the rational
maps obtained by perturbing a parabolic cycle, and underlies Shishikura's
proof that the boundary of the Mandelbrot set has Hausdorff dimension two.
Other applications appear in [Douady-Buff-Devaney-Sentenac], [Jellouli2] and
[Tan] in this volume. Refinements are included in the Appendix.
2 Techniques in complex dynamics

A wide range of modern methods of complex analysis and dynamics can be
seen at work in the articles of this volume.
The most classical techniques, successfully used by Fatou and Julia, are the
Poincare metric, Schwarz Lemma and Mantel's theorem on normal families.
One can see these methods applied in [Petersen-Ryd], [Shishikura-Tan] and
the appendix to [Hai'ssinsky].
The Riemann mapping theorem, together with the Caratheodory theory, is
a central tool for understanding the topology of the Julia set and its com-
plementary components. The rays coming from the Riemann representation
often meet (land) at same points and cut the Julia set. They are preserved
under iteration and transfer easily to the parameter spaces. Using external
rays, one obtains a type of Markov partition called a Yoccoz puzzle. These
methods appear in [Hai'ssinsky], [Milnor], [Roesch] and [Tan].
Quasiconformal deformations, the Beltrami equation, the Ahlfors-Bers the-
ory and its refined version G. David's theorem. Starting with Sullivan's work,
the integrability of measurable complex structures has found remarkable ap-
plications to complex dynamics. For example, it is essential to the Douady-
Hubbard theory of polynomial-like maps, Mandelbrot-like families, which in
turn forms the foundations of the theory of renormalization and the univer-
sality of the Mandelbrot set. These methods appear in [Douady], [McMullen],
[Douady-Buff-Devaney-Sentenac] and [Hai'ssinsky].
Holomorphic motions. This tool for studying the variation of dynamics in
families plays a crucial role in [Roesch]. See also [McMullen], [Douady-Buff-
Devaney-Sentenac] and [Hai'ssinsky].
Grotzsch inequality about moduli of annuli. These conformal invariants
play an essential role in work on local connectivity of Julia sets and the
Mandelbrot set. See [Milnor] and [Roesch].
Yoccoz inequality. This application of the Grotzsch inequality is discussed
Introduction xi
in [Tan], Appendix D and in [Haissinsky].

Parabolic implosion. This important modern theory is developed in de-
tail in [Shishikura2], and is applied in [Tan], [Jellouli2] and [Douady-Buff-
Devaney-Sentenac].
Expansion on subsets of the Julia set away from recurrent critical orbits.
This is the content of Mane's result and is the theme of [Shishikura-Tan]. See
[Yin] for an example of application.
Jakobson's theorem. Collet-Eckmann maps. This is one of the key results
in real dynamical systems having a complex flavor, and is exposed here in
[Luzzatto].
Circle homeomorphisms, distortion of cross-ratios. They are important on
their own right and have important applications in complex dynamics. Both
features are exhibited in [Petersen].
Transferring results on the dynamical planes to the parameter space. This
is the most common way to get information about the parameter space. See
[McMullen], [Douady-Buff-Devaney-Sentenac], [Haissinsky], [Roesch], [Tan],
[Jelloulil] and [Luzzatto].
Thurston's theory, creating rational maps from combinatorial information.
One example of this is mating of polynomials. And the fact that mating is
an adequate way to describe topologically the dynamics is reported here in
[Shishikural].
Acknowledgements. This is above all a collective work. It is a privilege
for me, the editor, to be able to thank all the contributors, who participated
with enthusiasm, energy, rigour and dedication. Special thanks go to C.
McMullen, who provided valuable help at every stage of the preparation, and
A. Manning, R. Oudkerk and S. van Strien for their linguistic, technical and
moral support.
Tan Lei, Department of Mathematics, University of Warwick, Coventry
CV4 7AL, United Kingdom
e-mail: tanlei@maths.Warwick.ac.uk
Preface
John Hubbard
Holomorphic dynamics is a subject with an ancient history: Fatou, Julia,

Schroeder, Koenigs, Bottcher, Lattes, which then went into hibernation for
about 60 years, and came back to explosive life in the 1980's.
This rebirth is in part due to the introduction of a new theoretical tool:
Sullivan's use of quasi-conformal mappings allowed him to prove Fatou's no-
wandering domains conjecture, thus solving the main problem Fatou had left
open.
But it is also due to a genuinely new phenomenon: the use of computers
as an experimental mathematical tool. Until the advent of the computer,
the notion that there might be an "experimental component" to mathemat-
ics was completely alien. Several early computer experiments showed great
promise: the Fermi-Past a-U lam experiment, the number-theoretic computa-
tions of Birch and Swinnerton-Dyer, and Lorenz's experiment in theoretical
meteorology stand out. But the unwieldiness of mainframes prevented their
widespread use.
The microcomputer and improved computer graphics changed that: now a
mathematical field was behaving like a field of physics, with brisk interactions
between experiment and theory.
I mention computer graphics because faster and cheaper computers alone
would not have had the same impact; without pictures, the information pour-
ing out of mathematical computations would have remained hidden in a flood
of numbers, difficult if not impossible to interpret. For people who doubt this,
I have a story to relate. Lars Ahlfors, then in his seventies, told me in 1984
that in his youth, his adviser Lindelof had made him read the memoirs of
Fatou and Julia, the prize essays from the Academie des Sciences in Paris.
Ahlfors told me that they struck him at the time as "the pits of complex anal-
ysis": he only understood what Fatou and Julia had in mind when he saw
the pictures Mandelbrot and 1 were producing. If Ahlfors, the creator of one
of the main tools in the subject and the inspirer of Sullivan's no-wandering
domains theorem, needed pictures to come to terms with the subject, what
can one say of lesser mortals?
In this preface, I will mainly describe the events from 1976 to 1982, as I
saw them.
The first pictures
For me at least, holomorphic dynamics started as an experiment. Dur-
ing the academic year 1976-77, I was teaching DEUG (first and second year
calculus) at the University of Paris at Orsay. At the time it was clear that
willy-nilly, applied mathematics would never be the same again after com-
Xlll
xiv J. Hubbard
puters. So I tried to introduce some computing into the curriculum.

This was not so easy. For one thing, I was no computer whiz: at the time
I was a complex analyst with a strong topological bent, and no knowledge of
dynamical systems whatsoever. For another, the students had no access to
computers, and I had to resort to programmable calculators. Casting around
for a topic sufficiently simple to fit into the 100 program steps and eight
memories of these primitive machines, but still sufficiently rich to interest
the students, I chose Newton's method for solving equations (among several
others).
This was fairly easy to program. But when a student asked me how to
choose an initial guess, I couldn't answer. It took me some time to discover
that no one knew, and even longer to understand that the question really
meant: what do the basins of the roots look like?
As I discovered later, I was far from the first person to wonder about this.
Cayley had asked about it explicitly in the 1880's, and Fatou and Julia had
explored some cases around 1920. But now we could effectively answer the
question: computers could draw the basins. And they did: the math depart-
ment at Orsay owned a rather unpleasant computer called a mini-6, which
spent much of the spring of 1977 making such computations, and printing
the results on a character printer, with X's, 0's and l's to designate points
of different basins. Michel Fiollet wrote the programs, and I am extremely
grateful to him, as 1 could never have mastered that machine myself.
In any case, Adrien Douady and I poured over these pictures, and even-
tually got a glimpse of how to understand some of them, more particularly
Newton's method for z3 — 1 and z3 - z. At least, we understood their topology,
and possibly the fact that we concentrated on the topology of the Julia sets
has influenced the whole subject; other people looking at the same pictures
might have focussed on other things, like Hausdorff dimension, or complex
analytic features.
Also, I went for help to the IHES (Institut des Hautes Etudes Scientifiques),
down the road but viewed by many at Orsay as alien, possibly hostile terri-
tory. There dynamical systems were a big topic: Dennis Sullivan and Rene
Thorn were in residence and Michael Shub, Sheldon Newhouse, and John
Guckenheimer were visiting that spring.
I learned from Sullivan about Fatou and Julia, and especially about the
fact that an attracting cycle must attract a critical point, and more generally
that the behavior of the critical points dominates the whole dynamical pic-
ture. This suggested how to make parameter-space pictures, and the mini-6
made many of these also. (These pictures are among the great-grand parents
of the present volume, and whether the authors know it or not, they appear
in the genealogical tree of most if not all the papers.) But having the pictures
was no panacea: they looked chaotic to us, and we had no clear idea how to
Preface xv
analyze them.
The year 1981-82: an ode to the cafe of Paris

At the end of 1977,1 went to Cornell, where I thought and lectured about
the results we had found. In particular, Mandelbrot saw the pictures that
the mini-6 had produced, and correctly calling them "rather poor quality,"
invited me to give a lecture at Yorktown Heights in 1978, saying that he had
often thought about the "Fatou-Julia fractals," although he had never made
pictures of them. But nothing much got proved until the next visit to France,
for the academic year 1981-82.
By then, many things had subtly changed. Douady had understood that
it was wiser to iterate polynomials before iterating Newton's method, as they
are considerably simpler. His sister Veronique Gautheron had written pro-
grams to investigate the dynamics of polynomials. Computers had improved;
Veronique used a machine, now long defunct, called a Goupil (later a small
HP), but for me the arrival of the Apple II was decisive. Mandelbrot had
access to the IBM computer facilities of Yorktown Heights; he had produced
much better pictures of the Mandelbrot set than we had, and had published
a paper about it. Feigenbaum had performed his numerical experiments, and
the physicists were interested in the iteration of polynomials, more particu-
larly renormalization theory.
Perhaps it was an illusion, but it seemed to me that holomorphic dynam-
ics was in the air. Milnor and Thurston had long studied interval maps and
were beginning to consider polynomials in the complex. In the Soviet Union
(behind the iron curtain, very much in existence at the time) Lyubich and
Eremenko, who were at the time just names we were vaguely aware of, were
also starting to study holomorphic dynamics. In Brazil, Paolo Sad had pro-
duced a paper (hand-written) on the density of hyperbolic dynamics. The
paper was wrong, and the result is still the main unproved question of the
theory. (At the time 1 did not appreciate the importance of the result Sad
had claimed, but I clearly recall coming back to our apartment on the Rue
Pascal and hearing from my wife that Sullivan had telephoned all the way
from Brazil with the message that Sad's paper "was coming apart at the
seams.") But Sad's techniques led to one of the most important tools of
the subject, the Mane-Sad-Sullivan A-lemma and holomorphic motions. In
Japan, Ushiki had been an early advocate of computer graphics. His brilliant
student Shishikura was beginning to take an interest in the field. In any case,
the stage was set for a fertile year, and indeed 1981-82 was simply wonderful:
I will describe three episodes, all of which occur in cafes, and all of which are
somehow connected with computers.
• The connectivity of the Mandelbrot set. Mandelbrot had sent

us a copy of his paper, in which he announced the appearance of islands off
xvi J- Hubbard
the mainland of the Mandelbrot set M. Incidentally, these islands were in

fact not there in the published paper: apparently the printer had taken them
for dirt on the originals and erased them. (At that time, a printer was a
human being, not a machine.) Mandelbrot had penciled them in, more or
less randomly, in the copy we had.
One afternoon, Douady and I had been looking at this picture, and won-
dering what happened to the image of the critical point by a high iterate of
the polynomial z2 + c as c takes a walk around an island. This was difficult
to imagine, and we had started to suspect that there should be filaments of
M connecting the islands to the mainland. Overnight, Douady thought that
such filaments could be detected as barriers: something had to happen along
them, and found that is should be the arguments of the rays landing at 0.
In any case, Douady called the following morning, inviting me to join him
at a cafe Le Dauphin on the Rue de Buci. He had realized that what we had
discovered was that the Mandelbrot set was connected: over a croissant, he
wrote the statement, c i-> <^c(Pc(0)) is an isomorphism C \ M -» C \ B.
Sullivan flew from the United States to hear Douady speak on our proof
in the analytic dynamics seminar at Orsay that week. Sibony was also in
the audience, as were Kahane and many others. The following week, Sibony
announced he had an independent proof that the map proposed by Douady
was a bijection.
Not long after, I made a list of all the quadratic polynomials whose critical
point is periodic of period 5. Then I asked the computer (an Apple II, with
a pen plotter attached) to draw all the Julia sets. Today this would be
virtually instantaneous; at the time it took several hours. Then I looked
really carefully at the drawings, trying to see what made them different from
each other. After I marked the orbit of the critical point, in each case a tree
was staring me in the face. A bit of reflection soon told me some necessary
conditions a tree with marked points must satisfy in order to be a possible
tree for a polynomial.
In a day or so I had drawn all the trees that could be drawn corresponding
to the critical point being periodic of period 6. The fact that these did indeed
correspond to the appropriate polynomials was strong evidence that the de-
scription was right. Not long afterwards Douady came up with the algorithm
for external angles in terms of trees. The complex kneading sequence was
born. These trees (now called Hubbard trees) together with external rays
have now become a central tool of combinatorics and classifications.
• Matings. One night in the spring of 1982,1 set the computer (the same
Apple II, now equipped with a dot-matrix printer), to drawing Julia sets of
rational functions of degree 2, running through a list of parameter values
where the two critical points were periodic. Douady actively disapproved of
this activity, thinking that we should focus on quadratic polynomials until
Preface xvii
they were better understood.

The next morning I had a pile of perhaps 40 such drawings (on those folded
sheets with holes along the sides typical of dot-matrix printers), most of which
looked like junk. But several evidently had some structure. I collected these,
and met Douady at the local cafe (this time the cafe des Ursulines, near Ecole
Normale, whose owner at the time was very welcoming to mathematicians).
He looked at one of them, and after a while drew in two trees connecting the
orbits of the critical point. One was the tree of the rabbit, the other the tree
of the polynomial with the critical point periodic of period 4, with external
angles 3/15 and 4/15. It was immediately clear that the picture really did
represent this object.
This suggested many experiments to confirm the existence of matings,
which we carried out; soon we came up with the mating conjecture: two
quadratic polynomials can be mated unless they belong to conjugate limbs
of the Mandelbrot set. But we had to wait for Thurston's theorem and the
work of Silvio Levy, Mary Rees, Mitsu Shishikura, and Tan Lei, to see the
mating conjecture proved for post-critically finite polynomials. As far as I
know, there is still no reasonable mating conjecture in degree 3.
• Polynomial-like mappings. One of the great events of that year
was Sullivan's proof of the non-existence of wandering domains for rational
functions. Douady and I were both present at his first lecture on the subject,
and immediately saw the power of his methods: invariant Beltrami forms and
the Ahlfors-Bers theorem.
I had written my thesis (under Douady's direction) about Teichmiiller
spaces, and was quite familiar with these techniques, but had never thought
of applying them to dynamics; Sullivan knew better. He had studied Ahlfors's
work on Kleinian groups, and had realized that the same techniques could
be used in holomorphic dynamics. In Sullivan's view, there is a dictionary
relating Kleinian groups to holomorphic dynamics; the no-wandering domains
theorem showed the power of this program. Trying to understand further
parts of this dictionary has been an important motivating force for a lot of
the research in the field, and more particularly Curt McMullen's.
In short order Douady and I used quasi-conformal mappings to show that
the multiplier map gives a uniformization of the hyperbolic components of
the interior of M, something we had conjectured but couldn't prove.
Soon thereafter, in some cafe in the north of Paris, Douady was ruminating
about polynomial-like mappings. Of course, he didn't yet have a precise
definition, but he had seen that if an analytic mapping mapped a disc D c C
to C so that the boundary of D maps outside of D, winding around it several
times, then many of the proofs about polynomials would still go through.
Thinking over what he had said, I saw that evening that if one could
construct an appropriate invariant Beltrami form, then the polynomial-like
xviii J. Hubbard
mapping would be conjugate to a polynomial. This time 1 called Douady

and asked him to meet me at the Cafe du Luxembourg, and presented my
argument. My proof of the existence of the invariant Beltrami form was
shaky, as I was requiring extra unnecessary conditions, but Adrien soon saw
that if we got rid of these, the invariant Beltrami form was easy to construct.
The straightening theorem was born.
The following year, back in Cornell, I was making parameter-space pictures
of Newton's method for cubic polynomials. I saw Mandelbrot sets appearing
on the screen, which wasn't really surprising, but was simply amazed to see a
dyadic tree appearing around it. Looking carefully at this tree, I found that it
reflected the digits of external angles of points in the Mandelbrot set written
in base 2. Seeing that these angles were made by God, and not by man, was
an extraordinary realization to me. I started my Harvard colloquium lecture
two days later by saying I was changing its subject completely, because 1 had
made an amazing discovery two days earlier. That colloquium was one of the
best lectures I delivered in my life.
Quasi-conformal mappings have remained one of the central tools in holo-
morphic dynamics, eventually becoming a field in its own right called quasi-
conformal surgery, and providing an amazing flexibility in chopping up and
reassembling the very rigid objects of the field. Applications are too numer-
ous to list, but Shishikura's sharp bound on the number of nonrepelling cycles
of a rational function, and his construction of Herman rings from Siegel discs,
stand out as early monuments to the power of this method.
That spring, I went to the United States for a couple of weeks and gave
a number of lectures about these results, at Columbia and Cornell, and the
SUNY graduate school. I had never met John Milnor, though I had practically
been raised on his books. But he came to the Columbia lecture, and about a
week later, back in France, I received a letter from him pointing out that the
proof of the connectivity of the Mandelbrot set also proved that the external
argument of real polynomials in the boundary of M is a monotone function
of the parameter, and that this settled the entropy conjecture of Metropolis,
Stein and Stein from the 1950's.
Another important event of that year for Douady and me was a Seminaire
Bovirbaki lecture in which Malgrange presented work of Ecalle on indifferent
fixed points. The theory of Ecalle cylinders and the parabolic implosion grew
out of that lecture. These results, and many others from that year, such as
the existence of Julia sets that are not locally connected, and the connection
of polynomial-like mappings with renormalization, are clearly part of the
genealogy of all the papers of this volume; I would call them the grandparents
of the present volume.
Preface xix
The Orsay notes

In the fall of 1982,1 returned to Cornell. The situation was then as follows:
a huge amount had been discovered, and largely proved. In particular, we
had proved that if the Mandelbrot set is locally connected, then hyperbolic
quadratic polynomials are open and dense, and we had formulated the MLC
conjecture (the Mandelbrot set is locally connected). We had also constructed
the combinatorial model M of the Mandelbrot set and the canonical mapping
M —> M, and proved that if MLC holds, then this map is a homeomorphism.
We had complete proofs of the combinatorial description of Julia sets
of strictly preperiodic polynomials and polynomials with attracting cycles,
although there were gaps for polynomials with parabolic cycles, which led to
gaps in understanding the landing of rays at roots of components.
But our proofs were handwritten, mainly understandable only to Douady
and me; next to nothing had been published. The only publications I can
think of from that year were two Notes aux Comptes-Rendus, one by Douady
and me on the connectivity on the Mandelbrot set and trees, and one by
Sullivan on the no-wandering domains theorem.
Getting this material organized and written was essentially Douady's work:
he gave a course at ENS and Orsay the following year, delivering each week
the same lecture in both places and then writing it, using the notes of of
Pierrette Sentenac, Marguerite Flexor, Regine Douady and Letizia Herault.
Tan Lei and Lavaurs were in the audience, and become the first generation
of students in this field. On my side, Ben Wittner and Janet Head were also
first generation students in the field at Cornell.
I was not present for most of the writing of the Orsay notes, and am always
amazed at the extraordinary new inventions present in the notes. The tour
de valse was in our hand-written notes (where it was called the lemme du
coup de fouet; this name was considered offensive by Pierrette Sentenac), but
the arrivee au bon port and the local connectivity of the Julia set of z2 + z
are among Douady's inventions of that year. Others contributed to the notes,
including Lavaurs, who provided some of the central ideas of the parabolic
implosion and Tan Lei with her resemblance between the Mandelbrot set near
a preperiodic point and the Julia set for that point.
Other publications followed: the Mane-Sad-Sullivan paper in 1983, and
the polynomial-like mappings paper by Douady and me, in 1985. Together, I
would describe these and the Orsay notes as the parents of the papers of the
present volume.
Of course, the present volume has another parent: Bill Thurston's topolog-
ical characterization of rational functions. Sullivan had linked holomorphic
dynamics with quasi-conformal mappings, and Thurston invented another
great new technique by linking holomorphic dynamics to Teichmiiller the-
ory. For a postcritically finite polynomial, there are 3 ways of encoding the
xx J. Hubbard
dynamics: my trees, the external arguments of the critical values, and the
Thurston class. Only the last extends to rational functions, and with this
theorem it became clear that post-critically finite branched mappings provide
the right way to encode the combinatorics of rational mappings.
The work after this is no longer history, but current events, and I leave an
introduction to the present volume to Tan Lei.
Department of Mathematics, Cornell University, Ithaca, NY 14853-7901,

U.S.A. and Departement de Mathematiques, Universite de Aix-Marseille I,
13331 Marseille Cedex 3, Prance.
email: jhh8@cornell.edu and John.Hubbard@cmi.univ-mrs.fr
The Mandelbrot set is universal
Curtis T. McMullen*
Abstract
We show small Mandelbrot sets are dense in the bifurcation locus
for any holomorphic family of rational maps.
1 Introduction
Fix an integer d > 2, and let pc{z) = zd + c. The generalized Mandelbrot set
Md C C is defined as the set of c such that the Julia set J(pc) is connected.
Equivalently, c e Md iff p"(0) does not tend to infinity as n -» oo. The
traditional Mandelbrot set is the quadratic version M.2-
A holomorphic family of rational maps over X is a holomorphic map
f:X xC-»C
where X is a complex manifold and C is the Riemann sjphere. For each t E X
the family / specializes to a rational map /t : C —> C, denoted ft{z). For
convenience we will require that X is connected and that deg(/ t ) > 2 for all
t.
The bifurcation locus B(f) C X is defined equivalently as the set of t such
that:
1. The number of attracting cycles of ft is not locally constant;
2. The period of the attracting cycles of ft is locally unbounded; or

3. The Julia set J(ft) does not move continuously (in the Hausdorff topol-
ogy) over any neighborhood of t.
It is known that B(f) is a closed, nowhere dense subset of X; its complement
X - B{f) is also called the J-stable set [MSS], [Mc2, §4.1].
As a prime example, pc(z) = zd + c is a holomorphic family parameterized
by c 6 C, and its bifurcation locus is dMj. See Figure 1.
In this paper we show that every bifurcation set contains a copy of the
boundary of the Mandelbrot set or its degree d generalization. The Man-
delbrot sets Md are thus universal; they are initial objects in the category
of bifurcations, providing a lower bound on the complexity of B(f) for all
families / ( .
For simplicity we first treat the case X = A = {t : \t\ < 1}.
'Research partially supported by the NSF. 1991 Mathematics Subject Classification:
Primary 58F23, Secondary 30D05.
1
C. McMullen
Figure 1. Mandelbrot sets of degrees 2, 3 and 4.
Theorem 1.1 For any holomorphic family of rational maps over the unit
disk, the bifurcation locus B(f) C A is either empty or contains the quasi-
conformal image of dMd for some d. "
The proof (§4) shows that B(f) contains copies of dMd with arbitrarily
small quasiconformal distortion, and controls the degrees d that arise. For
example we can always find a copy of dMd with d < 2 2 d e g ^^~ 2 , and generically
B(f) contains a copy of dM2 (see Corollary 4.4). Since the Theorem is local
we have:
Corollary 1.2 Small Mandelbrot sets are dense in B(f).
There is also a statement in the dynamical plane:
Theorem 1.3 Let f be a holomorphic family of rational maps with bifurca-

tions. Then there is a d>2 such that for any c € Md and m > 0, the family
contains a polynomial-like map ff:U-+V hybrid conjugate to zd + c with
mod(t/ -V)> m.
Corollary 1.4 / / / has bifurcations then for any e > 0 there exists a t such
that ft(z) has a superattracting basin which is a (1 + e)-quasidisk.
Proof. The family contains a polynomial-like map / f : U -> V hybrid

conjugate to po(z) = zd, a map whose superattracting basin is a round disk.
Since mod(y — U) can be made arbitrarily large, the conjugacy can be made
nearly conformal, and thus ft has a superattracting basin which is a (1 -I- e)-
quasidisk. •
For applications to Hausdorff dimension we recall:
Theorem 1.5 (Shishikura) For any d > 2, the Hausdorff dimension of

dMd is two. Moreover H. dim(J(p c )) = 2 for a dense Gs of c e dMd.
This result is stated for d — 2 in [Shi2] and [Shil] but the argument generalizes
to d > 2. Quasiconformal maps preserve sets of full dimension [GV], so from
Theorems 1.1 and 1.3 we obtain:
Corollary 1.6 For any family of rational maps f over A, the bifurcation set
B{f) is empty or has Hausdorff dimension two.
Corollary 1.7 / / / has bifurcations, then H. dim(J(/ t )) — 2 for a dense set

ofteB(f).1
For higher-dimensional families one has (§5):
Corollary 1.8 For any holomorphic family of rational maps over a complex
manifold X, either B{f) = 0 or H. dim(B(/)) = H. dim(X) = 2 dime X.
Similar results on Hausdorff dimension were obtained by Tan Lei, under a

technical hypothesis on the family / [Tan].
A family of rational maps / is algebraic if its parameter space X is a
quasi-projective variety (such as C n ) and the coefficients of ft(z) are rational
functions of t. For example, pc{z) = zd + c is an algebraic family over X — C.
Such families almost always contain bifurcations [Mel]:
Theorem 1.9 For any algebraic family of rational maps, either
1. The family is trivial (ft and fs are conformally conjugate for all t,s £
X); or
2. The family is affine (every ft is critically finite and double covered by a

torus endomorphism); or
3. The family has bifurcations (B(f) ^ 0).
Corollary 1.10 With rare exceptions, any algebraic family of rational maps
exhibits small Mandelbrot sets in its parameter space.
'This set of t can be improved to a dense G$ using Shishikura's idea of hyperbolic

dimension.
4 C. McMullen
This Corollary was our original motivation for proving Theorem 1.1.
As another application, for t e C^ 1 let
and let
Cd = {t : J(ft) is connected}
denote the connectedness locus. Then we have:
Corollary 1.11 (Tan Lei) The boundary of the connectedness locus has full
dimension; that is, H. dim(dCd) = H. dim(Cd) = 2d — 2.
Proof. Consider the algebraic family ga(z) = zd + azd~l, which for a ^ 0

has all but one critical point fixed under ga. By Theorem 1.9, this family has
bifurcations at some a S G Then there is a neighborhood U of (a, 0,... ,0) 6
Cd~2 such that for t e U all critical points of ft save one lie in an attracting
or superattracting basin. If t £ B(f) D U, then the remaining critical point
has a bounded forward orbit under ft, but under a small perturbation tends
to infinity. It follows that B(f) D U = dCd n U ^ 0, and thus dim{dCd) >
dim B(f) = 2d-2. •
Remark. Rees has shown that the bifurcation locus has positive measure in
the space of all rational maps of degree d [Rees]; it would be interesting to
known general conditions on a family / such that B(f) has positive measure
in the parameter space X.
Acknowledgements. I would like to thank Tan Lei for sharing her re-
sults which prompted the writing of this note. Special cases of Theorem 1.1
were developed independently by Douady and Hubbard [DH, pp.332-336] and
Eckmann and Epstein [EE].
2 Families of rational maps

In this section we begin a more formal study of maps with bifurcations.
Definitions. A local bifurcation is a holomorphic family of rational maps
ft{z) over the unit disk A, such that 0 € B(f).
The following natural operations can be performed on / to construct new
local bifurcations:
1. Coordinate change: replace /( by mt o ft o m^1, where m : A x C - + C

is a holomorphic family of Mobius transformations.
2. Iteration: replace ft(z) by /"(z) for a fixed n > 1.
3. 5ase change: replace / t (z) by ./>(*) (z), where <>

/ : A -> A is a noncon-
stant holomorphic map with </>(0) € B(f).
The first two operations leave the bifurcation locus unchanged, while the last
transforms B{f) to <j>-\B{f)).
Marked critical points. We will also consider pairs (/, c) consisting of
a local bifurcation and a marked critical point; this means c : A —> C is
holomorphic and f[{ct) = 0. The operations above also apply to (/, c); a
coordinate change replaces cj with mt(ct) and a base change replaces Q with
Misiurewicz points. A marked critical point c of / is active if its forward

orbit
(/t "(ct) : n = 1,2,3,...)
fails to form a normal family of functions of t on any neighborhood of t = 0

in A. A parameter t is a Misiurewicz point for (/, c) if the forward orbit of ct
under ft lands on a repelling periodic cycle. If t = 0 is a Misiurewicz point,
then either c is active or Ct is preperiodic for all t.
Proposition 2.1 / / c is an active critical point, then (/, c) has a sequence

of distinct Misiurewicz points tn —• 0.
Proof. This is a traditional normal families argument. Choose any 3 distinct

repelling periodic points {ao, bo, CQ} for /o, and let {at, bt, c t } be holomorphic
functions parameterizing the corresponding periodic points of ft for t near
zero. Since (/™(ct)) is not a normal family, it cannot avoid these three points,
and any parameter t where /"(c t ) meets at, bt or ct is a Misiurewicz point. •
Ramification. Next we discuss the existence of univalent inverse branches

for a single rational map F(z). Let d = deg(F, z) denote the local degree of
F at z 6 C; we have d > 1 iff z is a critical point of multiplicity (d — 1).
We say y is an unramified preimage of z if for some n > 0, Fn(y) = z and
deg(Fn,y) — 1. We say z is unramified if it has infinitely many unramified
preimages. In this case its unramified preimages accumulate on the full Julia
set J(F).
Proposition 2.2 / / z has 5 distinct unramified preimages then it has in-

finitely many.
6 C. McMullea
Proof. Let E be the set of all unramified preimages of z, and let C be the
critical points of F. Then F~l(E) C E U C, so if \E\ is finite then
d\E\= Yl l + mult(/',z) < \F~l{E)\ + 2d-2< \E\ + 4d-4
and therefore \E\ < 4. •
Corollary 2.3 Let (/, 0) be a local bifurcation with marked critical point.
Then the set of t such that ct is ramified for ft is either discrete or the whole
disk.
Proof. By the previous Proposition, the ramified parameters are denned by

a finite number of analytic equations in t. •
Proposition 2.4 After a suitable base change, any local bifurcation f can be
provided with an active marked critical point c such that CQ is unramified for
/o.
Remark. It is possible that all the active critical points are ramified at
t = 0. The base change in the Proposition will generally not preserve the
central fiber / 0 .
Proof. The set C = {(t, z) 6 A x C : f[(z) = 0} is an analytic variety with
a proper finite projection to A. By Puiseux series, after a base change of the
form <j>{t) = etn all the critical points of / can be marked by holomorphic
functions [c],... , c™}. Since t = 0 is in the bifurcation set, by [Mc2, Thm.
4.2], there is an i such that {f"(c\)) is not a normal family at t = 0. That is,
c* is an active critical point.
Next we show c1 can be chosen so that for generic t it is disjoint from the
forward orbits of all other critical points. If not, there is a o3 and n > 1 such
that f"(c{) — c\ for all t. Then cP is also active and we may replace c' with cJ.
If the replacement process were to cycle, then c* would be a periodic critical
point, which is impossible because it is active. Thus we eventually achieve
a c* which is generically disjoint from the forward orbits of the other critical
points.
In particular, there is a t such that c\ is unramified for ft. By Corollary 2.3,
the set R C A of parameters where c\ is ramified is discrete. By Proposition
2.1, there are Misiurewicz points tn for (/,c*) with tn —> 0. Choose n such
that tn & R, and make a base change moving tn to zero; then c* is active, and
CQ is unramified for /Q. •
Misiurewicz bifurcations. Let (/, c) be a local bifurcation with a marked

critical point. We say (/, c) is a Misiurewicz bifurcation of degree d if
Ml. /o(co) is a repelling fixed-point of / 0 ;
M2. Co is unramified for /o;
M3. ft(ct) is not a fixed-point of / ( , for some t; and
M4. deg(/ t , ct) = d for all t sufficiently small.
Proposition 2.5 For any local bifurcation (f, c) with c active and CQ unram-
ified, there is a base change and an n > 0 such that (/", c) is a Misiurewicz
bifurcation.
R e m a r k . The delicate point is condition (M4). The danger is that for every
Misiurewicz parameter t, the forward orbit of Q might accidentally collide
with another critical point before reaching the periodic cycle. We must avoid
these collisions to make the degree of / " at ct locally constant.
Proof. There are Misiurewicz points tn -> 0 for (/, c), and ct is unramified
for all t near 0, so after a base change and replacing / with / " we can arrange
that (/,c) satisfies conditions (Ml), (M2) and (M3).
We can also arrange that deg(/ t ,c ( ) = d for all t ^ 0. However (M4)
may fail because deg(/ t , ct) may jump up at t = 0. This jump would occur if
another critical of /( coincides with ct at t = 0.
To rule this out, we make a further perturbation of fQ. Let at locally
parameterize the repelling fixed-point of /[ such that /o(co) = aa. Choose a
neighborhood U of a0 such that for t small, ft is linearizable on U and U is
disjoint from the critical points of ft. (This is possible since /g(ao) ^ 0 )
Let bt 6 U — {at} be a parameterized repelling periodic point close to at.
Then bt has preimages in U accumulating on at. Choose s near 0 such that
fs{cs) hits one of these preimages (such an s exists by the argument principle
and (M3)). For this special parameter, cs first maps close to as, then remains
in U until it finally lands on bs. Since there are no critical points in U. we
have deg(/j, c) = d for all i > 0.
Making a base change moving s to t — 0, we find that (fn,c) satisfies
(M1-M4) for n a suitable multiple of the period of bs. •
3 The Misiurewicz cascade

In this section we show that when a Misiurewicz point bifurcates, it produces
a cascade of polynomial-like maps.
8 C. McMullen
Definitions. A polynomial-like map g : U —> V is a proper, holomorphic

map between simply-connected regions with U compact and U C V C C
[DH]. Its filled Julia set is denned by
it is the set of points that never escape from U under forward iteration.
Any polynomial such as pc{z) = zd + c can be restricted to a polynomial-
like map pc : U —> V of degree d with the same filled Julia set. Moreover
small analytic perturbations of pc : U -»• V are also polynomial-like.
A degree d Misiurewicz bifurcation (/, c) gives rise to polynomial-like maps
/ " : Bo -> Bn, by the following mechanism. For small t, a small ball Bo about
the critical point ct maps to a small ball Bi close to, but not containing, the
fixed-point of ft. The iterates B; = fl(B) then remain near the fixed-point
for a long time, ultimately expanding by a large factor. Finally for suitable t,
as Bi escapes from the fixed-point it maps back over Bo, resulting in a degree
d map / " : Bo -» Bn D Bo. Since most of the images (Bi) lie in the region
where ft behaves linearly, the first-return map / " : Bo ^ Bn behaves like a
polynomial of degree d.
This scenario leads to a cascade of families of polynomial-like maps, in-
dexed by the return time n. Here is a precise statement.
Theorem 3.1 Let (/, c) be a degree d Misiurewicz bifurcation, and fix R > 0.
Then for all n ^> 0, there is a coordinate change depending on n such that
ct — 0 and
whenever \z\, |£| < R. Here t = tn(l + 7n£)> tn and j n are nonzero, and j n ,
tn and en tend to zero as n —> oo.
The constants in O(-) above depend on / and R but not on n.

The proof yields more explicit information. Let Ao = /o(/o(co)) be the
multiplier of the fixed-point on which c0 lands, and let r be the multiplicity
of intersection of the graph of ct and the graph of this fixed-point at t = 0.
Then for t = tn, the critical point ct is periodic with period n, and we have:
tn ~ CXon/r, (3.1)
1
7n = CV'"" ', and (3.2)
n/r il/(<| 1)
en = n(|Ao|- +|A o r - )) (3-3)
for certain constants C, C depending on / . Due to the choice of roots, there
are r possibilities for tn and (d— 1) for 7,,; the Theorem is valid for all choices.
Finally for £ fixed and t — tn(l + 7 n £), the map / " is polynomial-like near Q
for all n » 0, and in the original ^-coordinate its filled Julia set satisfies
Notation. We adopt the usual conventions: tin — O(bn), a,, >c bn, an ~ 6n
and n » 0 mean \an\ < C\bn\, (1/C)|6 n | < \an\ < C\bn\, an/bn -» 1 and
n > N, where C and N are implicit constants.
Proof. We will make several constructions that work on a small neighbor-
hood of t — 0. First, let at parameterize the repelling fixed-point of ft such
that aQ = /o(c 0 ). Let A( = //(o t ) be its multiplier. There is a holomorphically
varying coordinate chart u — 4>t(z) defined near z = a« such that
(j>toft°<t>tl(u) = Atu (3.4)
for u near 0. We call u = <pt(z) the linearizing coordinate; note that u — 0 at

a*.
We next arrange that u = 1 is an unramified preimage of cf. Since CQ
is unramified by (M2), its unramified preimages accumulate on a0. Let b0
be one such preimage, with /Q(Ô) = co a n d &o in the domain of <f>Q. Then
bQ prolongs to a holomorphic function bt with f[{bt) = ct. Replacing <f>t by
<Pt(z)I'<t>i{bt). we can assume u = <j>t{bt) = 1.
For small t, the composition /f o <f>[1 is univalent near u = 1. By apply-
ing a coordinate change z •-> m t (2), where m t is a Mobius transformation
depending on t, we can arrange that c( = 0 and that
~lf) (3.5)
on B(l.f).
Since deg(/ ( , 0) = d for £ near 0 by (M4), we have
4>toft(z) = Y,
d
(3.7)
with ^d(O) / 0. Here A0(t) - ft(Q) is the u-coordinate of the critical value.
By (M3), ct is not pre-fixed for all t, so there is an r > 0 such that
A0(t) = fB{t) (3.8)
where B{0) # 0.
Next for n > 0 we choose tn such that
(3.9)
10 C. McMullen
z-coordinate plane •u-coordinate plane
o• • • o
at = 0 ^ ^ ^ bt = \
u i-> XtU
Figure 2. Visiting the repelling fixed-point
More precisely, for t = tn we will arrange that ct maps first close to at, then
lands after n iterates on bt, and thus returns in p further iterates to ct; see
Figure 2. In the u-coordinate system, ft is linear and bt = 1, so the equation
f?+l(ct) = bt becomes
= 1 when £ = tn. (3.10)
By the argument principle, for n ^> 0 this equation has a solution tn close to
any root of the approximation Agf £(0) = 1 obtained from (3.8). Moreover
tn
(verifying (3.1)), and tn satisfies (3.9) because fl{bt) = ct. (There are actually
be r solutions for tn for a given n; any one of the r solutions will do.)
We now turn to the estimate of fl+n+p(z) for (t,z) near (tn,0). We will
assume throughout that t = tn + s and that:
z\ and \s/tn\ are 0{ATnl{d-l)) (3.11)
where A = |A0| > 1. (To see this is the correct scale at which to work,
suppose diam(.B) x diam/ t 1 + n + p (5), where B is a ball centered at z — 0.
Then d i a m / t ( 5 ) x (diamB) d , and / " is expanding by a factor of about
A n , while ft is univalent, so we get d i a m 5 x A"(diam5) d , or diamB x
1+ +p n r
A -n/(d-i) similarly |/ f " (0)| x A {s/tn)t n x (s/tn) = 0 ( d i a m 5 ) when s
is as above.)
It is also convenient to set
so that we may assert:
z andt areO{A~n). (3.12)

By (3.11) the n iterates of ft(z) lie within the domain of linearization, so

by (3.7) we have
<t>t o fl+n(z) =K >nA0(d)zd(l + O(\z\ + \t\)).
The first term is approximately 1. Indeed, A" = A"n(l + O(ns)), so by (3.8)

we have
rf
rf O(ns)
by (3.10). Similarly, A? = AJ(1 + O(<)), so
rf O(ns) =
A%A0(d)zd + r -
using (3.11) and (3.12). The expression above as a whole is O(A~ n ^ d ~^), so
composing with the univalent map / f o ^ 1 introduces (by (3.5)) an additional
error of size O(A" 2n ^ d ~ 1) ), which is already accounted for in the O(-) above.
Thus the expression above also represents ft rn+p(z).
Finally we make a linear change of coordinates of the form z >-> anz,
conjugating the expression above to
Setting an = (A5^ to normalize the coefficient of zd, we have

an\ x A"/^" 1 ' and thus:
with * = tn(l + 7,,f), 7 n and en as in (3.2) and (3.3). Notice that if \z\ and
|f | are bounded by R in the expression above, then (3.11) is satisfied in our
original coordinates. Reindexing n, we obtain the Theorem. •
12 C. McMullen
4 Small Mandelbrot sets

We now show the Misiurewicz cascade leads to small Mandelbrot sets in pa-
rameter space. From this we deduce Theorems 1.1 and 1.3 of the Introduction.
Hybrid conjugacy. Let 51,52 be polynomial-like maps of the same degree.
A hybrid conjugacy is a quasiconformal map <j> between neighborhoods of
K(gi) and K(g2) such that <f> o g1 = g2 o 0 and d(j>\K{g\) = 0. We say gx
and 52 a r e hybrid equivalent if such a conjugacy exists. By a basic result
of Douady and Hubbard, every polynomial-like map g of degree d is hybrid
equivalent to a polynomial of degree d, unique up to affine conjugacy if K(g)
is connected [DH, Theorem 1].
Theorem 4.1 Let (/, c) be a degree d Misiurewicz bifurcation. Then the pa-
rameter space A contains quasiconformal copies M.% °f the degree d Mandel-
brot set Md, converging to the origin, with dM^ contained in the bifurcation
locus B(f).
More precisely, for all n^$> 0 there are homeomorphisms
K : Md -> Mnd C A
such that:
1. / " is hybrid equivalent to zd + f whenever t = (/>„(£);
2. d{0,Mnd)~\M\-nlr;
3. di&m{Mnd)/d{0,Mnd) x lAol""^- 1 ';
4- <j)n extends to a quasiconformal map of the plane with dilatation bounded

by 1 + O(en); and
5- ^ ° <t>n(0 = ? + O(en), where ^ ( 0 = *»(1 + 7»0-
The notation is from (3.1) - (3.3).

We begin by recapitulating some ideas from [DH]. Let A(R) = {z : \z\ <
R}, and let
be a holomorphic family of mappings defined for (£, z) G A(R) x A(R), where

R > 10 and ^(0) = 0. Let M C A(R) be the set of f such that the forward
orbit <?£(0) remains in A(R) for all n > 0.
Lemma 4.2 There is a S > 0 such that if sup |/i(f, z)| = e < 5 then there is
a homeomorphism
<f>: Md ->• M
such that for all£ 6 Md, g^Q is hybrid equivalent to zd + £
and 4> extends to a 1 + O(c)-quasiconformal map of the plane.
Proof. Let p ? (z) = zd + f. Since /? > 10 we have M d C A(fl) and K{p{) C
A(i?) for all £ € M d ; indeed these sets have capacity one, so their diameters
are bounded by 4 [Ah]. In addition, for £ € Md the map p ? : U —> A(/Z)
is polynomial-like, where £/ = p^"1(A(/?)). By continuity, when sup \h\ is
sufficiently small, M is compact and g^ is polynomial-like for all ^ £ A1.
By results of Douady and Hubbard, we can also choose 6 small enough
that \h\ < 5 implies there is a homeomorphism
<j>: Md - > M
such that #0(0 is hybrid equivalent to zd + f [DH, Prop. 21].
Now assume |/i| < e < <$. For t E A let Mt denote the parameters where
the critical point remains bounded for the family
9tt = zd+i + t5-h{i,z\

and define <j>t : Md -4 Ait as above. Then <j>t is a family of injections, with
(j>o(z) = z, and (j>t(£) is a holomorphic function of t for every ^. (For example
this is clear at £ € dMd because Misiurewicz points are dense in dMd; for the
general case see [DH, Prop. 22].)
By a theorem of Slodkowski [SI] (cf. [Dou], [BR]), <j>t{z) prolongs to a holo-
morphic motion of the entire plane, and its complex dilatation (it — dcfit/dfa
gives a holomorphic map of the unit disk into the unit ball in L°°(C). By
the Schwarz lemma, \\fit\\oo < \t\; since cf> = 0 £ ^, we obtain a quasiconformal
extension of <j> with dilatation 1 + O(e). The bound on \(j>(£) - £| similarly
results by applying the Schwarz Lemma to the map A —¥ C - {0,1, oo} given
by t >-> <f>t(O> o n c e three points have been normalized to remain fixed during
the motion. •
Proof of Theorem 4.1. Fix R > 10. For all n » 0, Theorem 3.1 provides
a family of rational maps of the form
defined for (f,z) € A{R) x A(R), where t = ^ ( £ ) =J_ B (1 + 7 n 0 - The

preceding Lemma gives homeomorphisms 4>n : Md —>• Ai^ C A(/J) for all
n » 0. Setting <j>n = ij)n o <j>nj the Theorem results from the Lemma and the
bounds (3.1) - (3.3). I
14 C. McMullen
Example. The quadratic family (/, c) = (z2 + t - 2,0) is a Misiurewicz

bifurcation of degree d — 2, with Ao = 2 and r = 1. Thus M2 contains small
copies M\ of itself near c = - 2 , with rf(A^J, - 2 ) x 4"" and diam .M? x 16" n .
Consequences. Assembling the preceding results, we may now prove the
Theorems stated in the Introduction. Here is a more precise form of Theorem
1.1:
Theorem 4.3 Let f be a holomorphic family of rational maps over the unit
disk with bifurcations. Then there is a nonempty list of degrees
2de
DC {2,3,... j2 ^)-2}
such that for any e > 0 and d € D, B(f) contains the image of dMd under a
(1 + e)-quasiconformal map.
If the critical points of f are marked {c\,... ,c™} such that
(i < N) <=> c* is active and c\ is unramified for some t,
then we may take
D = {infsupdeg(/*,c') : i < N}.

1
k
Proof. Let BQ = B(f). After a base change we can assume that / is a local
bifurcation with critical points marked as above. By Proposition 2.4, there is
at least one active, unramified critical point, so N > 1. For any i < N, we can
make a base change so c\ is active and unramified; then by Proposition 2.5,
a further base change makes (/ n ,c : ) into a degree d Misiurewicz bifurcation.
Let di = inf t sup t deg(/ t *,cj). We claim d = d{ < 22i*^ft)~2. Indeed,
deg(/ t ", c\) assumes its minimum outside a discrete set, and it is equal to d
near t = 0, so di > d. On the other hand, cj, lands on a repelling periodic cycle,
so deg(/0'c, Cg) = d for all k > n, and therefore di < d. Finally d is largest if c1
hits all the other critical points of / before reaching the repelling cycle; in this
case d = (pi + l)(p2+l) • • • (pm+l) for some partitionpi+pz+- • -+pm = 2d—2.
The product is maximized by the partition 1 + lH h l , s o d < 22deg(/t)"2.
By Theorem 4.1, the bifurcation locus B(f) contains almost conformal
copies dMj of dMd accumulating at t - 0, with diam(A^S) < d(0,M%).
Letting <f>: A -4 A denote the composition of all the base-changes occurring
so far, we have B(f) — 4>~l{Bo). Then <t> is univalent and nearly linear on
M^ for n » 0, so (f>(dM2) C Bo is a (1 + e)-quasiconformal copy of dMd. •
Let Rat d be the space of all rational maps of degree d; it is a Zariski-open

subset of P 2 d + 1 . We now make precise the statement that a generic family
contains a copy of the standard Mandelbrot set.
Corollary 4.4 There is a countably union of proper subvarieties R C Ratj

such that for any local bifurcation, either ft & R for all t, or B(f) contains a
copy of dM%.
Proof. On a finite branched cover X of Ratd, the critical points of / 6
can be marked { c ^ / ) , . . . ,<?d-2{f)}. Clearly deg(/n,a{f)) = 2 outside a
proper subvariety Vrni; of X. Let R be the union of the images of these
varieties in Ratd, and apply the preceding argument to see D — {2} if some
ft t R. •
Proof of Theorem 1.3. The proof follows the same lines as that of Theorem
4.3; to get mod(V-U) large one takes R large in Theorem 3.1. Thus Theorem
1.3 also holds for all d e D. •
5 Hausdorff dimension
In this section we prove Corollary 1.8: for any holomorphic family / of rational
maps over a complex manifold X, we have H. dim(5(/)) = H.dim(X) if
Recall that the Hausdorff dimension of a metric space X is the infimum of

the set of 5 > 0 such that there exists coverings X = (J X{ with ]T(diam Xi)s
arbitrarily small.
Lemma 5.1 Let Y be a metric space, X a subset ofYx [0,1]. Then
H. dim(X) > 1 + inf H. dim(Xt)
where Xt = {y : (y,t) e X}.
Here Y x [0,1] is given the product metric.
Proof. Fix 6 with 6 + 1 > H. dim(X). For any n there is a covering X c
[)B{yuri) x U with |/i| = n and £>f +1 < 4" n . Note that '
Xtc{JB(yi,r,)
and
Let En be the set of t where the integrand exceeds 2" n ; then ^2m(En) <
53 2~" < oo. Thus almost every t belongs to at most finitely many En, so
almost every Xt admits infinitely many coverings with ^rf < 2~n -> 0.
Therefore 5 > inf H. dim(A'e), and the Theorem follows. •
16 C. McMullen
The Lemma above is related to the product formula
H. dim(X x Y) > H. dim(X) + H. dim(F);
see [Fal, Ch. 5] and references therein.

Proof of Corollary 1.8. Suppose B(f) ^ 0. Then there is a t0 e B(f) and
a locally parameterized periodic point a(t) of period n such that a(t) changes
from attracting to repelling near tQ [MSS], [Mc2, §4.1]. More formally this
means the multiplier X(t) = (fn)'(a(t)) is not locally constant and |A(to)| = 1-
Choosing local coordinates we can reduce to the case X = A" and t0 = 0.
Let A s = A x {s} for s € A"" 1 . For coordinates in general position, X(t)
is nonconstant on A o . Shrinking the A"" 1 factor, we can also assume a(t)
changes from attracting to repelling in the family / | A S for all s. Then
and H.dim5(/|A S ) = 2 by Corollary 1.6. Applying the Lemma above to

B{f) C A x A""1 we find
H.dim(£(/)) > (2n - 2) +inf H.dim£(/) 5 = 2n = H.dim(X).
References
[Ah] L. Ahlfors. Conformal Invariants: Topics in Geometric Function The-
ory. McGraw-Hill Book Co., 1973.
[BR] L. Bers and H. L. Royden. Holomorphic families of injections. Ada

Math. 157(1986), 259-286.
[Dou] A. Douady. Prolongement de mouvements holomorphes (d'apres

Slodkowski et autres). In Seminaire Bourbaki, 1993/94, pages 7-20.
Asterisque, volume 227, 1995.
[DH] A. Douady and J. Hubbard. On the dynamics of polynomial-like map-

pings. Ann. Sci. Ec. Norm. Sup. 18(1985), 287-344.
[EE] J.-P. Eckmann and H. Epstein. Scaling of Mandelbrot sets generated

by critical point preperiodicity. Comm. Math. Phys. 101(1985), 283-
289.
[Fal] K. J. Falconer. The Geometry of Fractal Sets. Cambridge University
Press, 1986.
[GV] F. W. Gehring and J. Vaisala. Hausdorff dimension and quasiconfor-

mal mappings. J. London Math. Soc. 6(1973), 504-512.
[MSS] R. Mane, P. Sad, and D. Sullivan. On the dynamics of rational maps.

Ann. Sci. Ec. Norm. Sup. 16(1983), 193-217.
[Mel] C. McMullen. Families of rational maps and iterative root-finding

algorithms. Annals of Math. 125(1987), 467-493.
[Mc2] C. McMullen. Complex Dynamics and Renormalization, volume 135

of Annals of Math. Studies. Princeton University Press, 1994.
[Rees] M. Rees. Positive measure sets of ergodic rational maps. Ann. scient.
Ec. Norm. Sup. 19(1986), 383-407.
[Shil] M. Shishikura. The boundary of the Mandelbrot set has Hausdorff

dimension two. In Complex Analytic Methods in Dynamical Systems
(Rio de Janeiro, 1992), pages 389-406. Asterisque, volume 222, 1994.
[Shi2] M. Shishikura. The Hausdorff dimension of the boundary of the Man-

delbrot set and Julia sets. Annals of Math. 147(1998), 225-267.
[SI] Z. Slodkowski. Holomorphic motions and polynomial hulls. Proc.

Amer. Math. Soc. 111(1991), 347-355.
[Tan] Tan L. Hausdorff dimension of subsets of the parameter space for

families of rational maps. Nonlinearity, 11(1998), 233-246.
MATHEMATICS DEPARTMENT, HARVARD UNIVERSITY, 1 OXFORD S T ,

CAMBRIDGE, MA 02138-2901, U.S.A.
e-mail: ctm@abel.MATH.HARVARD.EDU
Baby Mandelbrot sets are born in
cauliflowers
Adrien Douady
with the participation of
Xavier Buff, Robert L. Devaney and Pierrette Sentenac
Abstract
We show that at the neighborhood of a cusp point Co # 1/4 of the
Mandelbrot set M, there is a sequence (Afn) of small quasi-conformal
copies of M in M, tending to CQ. Moreover, for each n, the copy
Mn is encaged in a set Fn which is homeomorphic to the Julia set
,/< of z2 + \ + e for e > 0 small, in fact in a nested sequence of sets
(Tn,m)m which are homeomorphic to the preimage of Jt by z t-¥ z2"',
and accumulate to Mn. We prove that the distance from Mn to CQ
tends to 0 like 1/n2 and its diameter, with or without the decorations
U m r n , m , tends toO like 1/n3.
1 The phenomena in the Mandelbrot set

When one looks closely at the Mandelbrot set M at the neighborhood of the
cusp point Co of a primitive hyperbolic component, one can observe a sequence
(Mn) of small copies of M in M, tending to CQ. Each copy Mn is encaged in
a set Tn which resembles the Julia set J( of z2 + \ + c for e > 0 small, in fact
in a nested sequence of sets (r n > m ) m which resemble the preimage of Jt by
z— i >• z2<n, and accumulate to Mn (Fig. 1).
The distance from Mn to Co tends to 0 like 1/n2 and its diameter, with or
without the decorations [Jm F n i m , tends to 0 like 1/n3.
In this paper, we shall first give a model T for this figure, then we shall
describe a more general situation in which T appears quasi-conformally in a
set Mp defined in the parameter space. Then we have to prove on one hand
that T does appear in Mp (section 10), and on the other that MF appears
in the Mandelbrot set in the neighborhood of CQ (section 11).
A precise statement is given in Theorem 1, and in Theorem 2 for the
more general situation. The proof of Theorem 2 will be given in sections 10.
It relies on concepts and results concerning polynomial-like mappings which
can be found in [DH Pol-like], with a complement by Lyubich [Ly], and on
the description of parabolic implosion [La], [D/AMS], [Sh], [TL], [DSZ]. We
shall also use Slodkowski's theorem on extension of holomorphic motions [SI],
[D/Sl]. Starting from these results, the main difficulty is in keeping track of
notations.
19
20 A. Douady, X. Buff, R. Devaney and P. Sentenac
7 8 9
Figure 1: The phenomena in the Mandelbrot set.
But first of all, we have to specify what we mean by "to appear quasi-
conformally" (section 2).
2 Conventions
We denote by Ts the translation by s, i.e. we set Ts(z) = 2 + 5.
We denote by 5# the circle of radius R in C, and by AR^ the annulus
{z\ R' < \z\ < R}.
We denote by Kc the filled Julia set and by Jc the Julia set for z2 + c, and
by M the Mandelbrot set.
The Julia set Ji is called the cauliflower; a Julia set Ji+e with e > 0 small
is called an imploded cauliflower. The Julia set Ji+e (Fig. 2) does not tend
to Ji when e -» 0. It does not actually have a limit, but there is a family of
compact sets (Ji )^ G T (Julia-Lavaurs sets) such that, if a sequence (J± +€n )
has a limit with en > 0 tending to 0, this limit is one of the sets Jia. The
compact set J i depends continuously on a for a € T [D/AMS].
Figure 2: A sequence of imploded cauliflowers.

Given a compact set K in C, connected, full and containing more than
one point, we denote by (j>K the isomorphism C \ K —> C \ D such that
\im.((pK{z)/z) is real > 0.
Given two compact sets X and Y in C, we say that X appears quasi-
conformally in Y if there is a quasi-conformal homeomorphism ip : C —> C
such that <p(X) C F and </>(<W0 C dY (the last condition is there in order to
exclude the case <p(X) C K, that one could always realize as soon as K ^ 0).
3 The model
Consider an imploded cauliflower Ji+( and choose R > 1 so that J±+t is
contained in the annulus A^ i . Denote by r o (e) the set Ji+( enlarged by
the factor R3^2, so that it is now contained in AR? #. Denote by Fm(e) the
inverse image of r o (e) by the map z >-> z2™ ; these sets are contained in
disjoint annuli and their union is the set of points whose orbit under Z H Z 2
has a (unique) point in r o (e).
We set (Fig. 3)
Up to a quasi-conformal homeomorphism of C, the set E(e) does not de-

pend on the choice of R provided it satisfies the requirement. It is possible to
choose R satisfying the requirement for all values of e which are sufficiently
small. Indeed the set L = \JaeT </ii<r is a compact set which does not contain
0 andd R suchh that
0, tht L CCA i ill do.
will d
We set
T - {0} U | J hn(E(en))
n>no
s4
w
* ,i
s
)
(3
Figure 3: The model H(e).
where en = ^ , and ftn(z) = ^2 + -^3-z, the constant C being chosen small

enough to ensure that the sets hn(E(en)) are contained in disjoint discs. Once
again, up to a quasi homeomorphism of C, the set T does not depend on the
choice of n 0 , R and C, provided they satisfy the requirements.
We shall prove
O
THEOREM 1.- If CQ is the root point of a primitive component of M, the set

T appears quasi-conformally in M, with 0 (in T) corresponding to CQ (in M).
We shall first describe a more general situation in which the set T appears
quasi-conformally in some set MF defined in the parameter space. In section
11 we shall show by a renormalization argument that this situation occurs in
that of Theorem 1.
4 A more general situation

Let A be an open set in C and Ao S A, and let f — {f\ : U'x —> U\)\e\ be an
analytic family of quadratic-like maps, i.e. of polynomial-like maps of degree
2. We suppose that f\0 has a fixed point ao with derivative 1 ; this implies
that fx0 is hybrid equivalent to z2 + | , i.e. to z + z2. We are interested only
with values of A which are close to AQ.
Let g\ : V\ -* U\ be an analytic isomorphism, depending analytically

on A, with Vx relatively compact in Ux \ U'x (Fig.4). We suppose that the
open sets Ux, U'x, Vx are Jordan domains with a C 1 boundary undergoing a
holomorphic motion.
Figure 4: The more general situation.
For each A e A, the map f\ has two fixed points a'x and a'x ; we may
not be able to decide which is which, in a continuous way, but the function
A i-> 5(X) = (a'x - a") 2 is well defined and holomorphic on A . We suppose
that the function 8 vanishes at Ao with a non zero derivative.
We denote by UJX the critical point of f\, we set an(X) = /"(WA) and
We denote by FA the map U'x U VA —>• Ux which induces fx on U'x and gx

on VA. We define K(FX) as the set of points z such that Fx(z) is defined and
belongs to U'x U Vx for all n. It is a full compact set in C.
We denote by MF the set of values of A for which u)X £ K{FX). It is a
closed set in A. Note that A 6 Mp does not imply that K(FX) is connected; in
fact it is never connected, since U'x n K(F\) and VA n K(F\) are disjoint^ not
empty as they contain K(f\) and gx1{K(fx)) respectively, and cover K(FX).
Theorem 1 will be a consequence of the more general:
THEOREM 2.- In this situation, the set T appears quasi-conformally in Mp.
We can normalize the situation in the following way: we suppose Ao = 0,
a0 = 0, and
Indeed, under the hypotheses made above, one can reduce fx(z) to this form
by a change of variables of the form (A, z) M- (u(X),A(X).z + B(X)), with
u(Ao) = 0, u'(Ao) / 0.
5 Fatou coordinates
Let / : [ / ' — > ( / be a holomorphic map and fi a connected open set mUUU'.
We denote by fl/f the quotient of f2 by the equivalence relation identifying
z with f(z) whenever z and f(z) are denned and belong to Q.
A map 4> : fi —>• C is called a Fatou coordinate if it satisfies the following
conditions:
• 0 is an analytic isomorphism of fi onto its image;
• <j> conjugates / to T\ on ft, i.e. <f>(f(z)) = <j>[z) + 1 when z and /(z) are
in S2 ;
• <j> induces a bijection of il/f onto C / Z .
If 0 is a Fatou coordinate on fi, fi// has naturally a quotient analytic-
structure and this bijection is an analytic isomorphism.
Given fi, a Fatou coordinate on Q,, if it exists, is unique up to an additive
constant. This follows from the fact that a C-analytic automorphism of C / Z
preserving the generator of H1(C/Z) is necessarily a translation.
If ^ : U[ U U\ —»• U'2 U U2 is a quasi-conformal conjugacy between / i : C/J —>
{/i and / 2 : U'2 —> (/2, and S7j is the domain of a Fatou coordinate for / 1 :
then fi2 = * ( ^ i ) is the domain of a Fatou coordinate for / 2 . Indeed Q2//2 is
quasi-conformally homeomorphic to C/Z, thus analytically isomorphic to it.
In view of the straightening theorem for polynomial-like mappings and
analytic families of polynomial-like mappings, many results that are known
for the family of quadratic polynomials Pt : z H-> Z + z2 + e (i.e. z > 4 z 2 + | + f
up to a change of origin) can be transported to our situation.
For Po : z H-» Z + z2, one can define two Fatou coordinates <3>po : S7£o -> C
and $po : f2po -t C such that Qpg (resp. fipo ) contains the disc Df =
{z I \z - r\ < r} (resp. D~ = {z \ \z + r\ < r} ) at right (resp. at left) of 0
for r > 0 small enough (in fact one can take r = \ ) . A maximal domain for
<t>p0 is B' U B"U ]topo, 0[, where B' and B" are the two main squares of the
parabolic chessboard, i.e. the connected components of K(P0) \ (jPo~ n (R)
which intersect D~ for r > 0 small, and wp0 = — | is the critical point of
o
Po- As a holomorphic function, <J>po extends to all of K(P0), but is no longer

injective there.
A maximal domain for <5>po is C \ R_. Indeed there is a branch h of P^
denned on C \ R_ and which maps it into itself; for any z in C \ R__ the
point hn{z) tends to 0 tangentially to R + when n —>• c», so one can extend
$ P 0 from £>+ to C \ R_ by $p o (z) = $po(hn(z)) + n for n big enough.
In our situation, the map / 0 is quasi-conformally conjugate to Po. We can
choose a Fatou coordinate <^ whose domain f^ contains D~ (for r > 0 small
enough) and the points an = /^(wo) for n > 1.
PROPOSITION 1.- There is a Fatou coordinate <f>^ for f0 whose domain QQ
contains VQ.
Proof : Remind the way one constructs a quasi-conformal conjugacy be-
tween /o and Po in [DH Pol-like]. By the Caratheodory theorem, the map / 0
extends to a continuous map U'a —>• UQ which is a local diffeomorphism except
at <x>0. First we choose a R > 1, and C 1 diffeomorphisms rj : dU0 —> S^ =
{z | \z\ = R} and rf : dU'o ->• SXR, where R' = y/R, so that J?(/ 0 (Z)) = (?/(z)) 2
for z € 9t/g. Second we fill (77, 77') into an arbitrary C 1 diffeomorphism T/J
between Uo \ f/o a n d t^ e closed annulus ylfl.R'. At that moment one can
arrange so that ip~l{[-R, —R1]) n VQ = 0. Later ] - 00, —1[ becomes the ex-
ternal ray of the cauliflower of argument | , i.e. itself. So Vo is contained in
*" X (C \ R_) where * is the quasi-conformal conjugacy between / 0 and Po
constructed using tp. •
We choose r 0 > 0 small and 90 e]0,n[, and we define the sector 5 as the
set of values of A such that |A| < r0 and \Arg{\)\ < 9Q. We set S* = 5 \ {0}.
If r 0 has been chosen small enough we have 5 C A, and for A £ 5 there
exist ([D/AMS],[DSZ]) Fatou coordinates fo : 9,^ -> C and 0+ : fi+ -^ C
satisfying the following conditions:
• QQ (resp. fi^" ) contains a disc £>+ (resp. D~ ) ;
• Vo C 9+ ;
• on(0) = /on(wo) e ô forn > 1 ;
• the fixed points a'x, a'l of /A belong to the boundaries dfl^, dil^ ;
• the set fi* = {(A, z) | A 6 S z e fij} is open in S x C ;
• for A G 5* the intersection fi^ nf2^ is the domain of a Fatou coordinate;
• (normalization) (j>^(an(X)) = n, cf>\(6A) = 0.
Even though being the domain of a Fatou coordinate is a pretty restrictive
condition for an open set, the conditions above do not determine the domains
ilf. We can arrange so that fi* is relatively compact in U\, bounded by a
C 1 Jordan curve which depends continuously on A. We could on the opposite
arrange so that fij = fi^ for \ € S*.
Restricting 5 if necessary, we have V\ C fi^ f° r all A G 5.
6 The phase
With the conditions above, the function (X,z) 1—>• 4>^(z) is continuous on fi*,
and C-analytic on fi* D (5* x C) ([Sh],appendix).
For A 6 5 ' , the functions <f>~^ and <j>^ both induce a Fatou coordinate on
fl^ D fi^, so <j>~i — 4>^ is a constant with a value r(A) € C. This defines a
function r : 5* —> C called the Zi/W phase, which is holomorphic and tends
to 00 when A tends to 0 ; we have [D/AMS], [Sh], [DSZ] :
Reducing slightly 60 and r0 if necessary, the map r induces a C-analytic

isomorphism A i-» A of S* onto a closed set S which contains a set of the form
{Z\\Z\>Rg, \Arg(-Z)\<ff0}.
We have ^ ~ 7 | ; ^ ~ f ? ; ^ x ~ -ff~ w f l e n ^ tends to 0, or equivalently
when A = r(A) tends to oo in S.
Extending fa to a function on K(f0) by fa(z) = fa(fo(z)) ~n> w e define
the Lavaurs map La for a 6 C ([La], [D/AMS], [DSZ]) by
the domain of definition of La being (fa ) 1(T-a(fa(£l£)). On each compact

set of this domain, / " tends to La uniformly when A tends to 0 and n tends
to oo in such a way that r(A) + n tends to a.
7 A sequence of Mandelbrot-like families

We set Vx = (j>x(Vx) for A € S, and Wn = {A 6 5 | A + n £ Vx} where
A = T~1(X). We set Wn = T~l(Wn). The sets Wn are disjoint. Indeed, for
A € Wn we have an(X) 6 Vx. More precisely Wn is the set of values of A such
that a,i(X) e fi^ U fi^ for 1 < « < n and an(X) 6 Vx.
We choose centers, i. e. base points in all these sets: LJX for Ux, bx for
Vx, 0 = <j>x(b\) for Vx, Xn = —n for Wn, and An = r " 1 ^ , , ) for Wn, so that
When n tends to infinity, the open set Tn(Wn) tends to VQ, with conver-
gence of the boundaries. In particular diam(Wn) tends to So = diam(V0). It
follows that
An ~ —2 and diam(Wn) ~ —5— .
n n6
For A 6 Wn, we denote by V"A' the connected component of fx"(Vx) which
contains LJX, and we set Gx = 5A ° / " : Vx' —> Ux (note that n is determined
by A since the sets Wn are disjoint).
For a e Tn(Wn), we set Ga,n = Gx with A = T~l(a - n). For o e Vo, we
denote by Vâ the connected component of L~1(Vo) which contains o>o, and
define Ga : Vo>0. -> Uo by Ga = gQoLa.
PROPOSITION 2.- a) Take n big enough. Then for X e Wn, the map Gx :
V^ —> Ux is quadratic-like, and the family (Gx)xewn is Mandelbrot-like.
b) For a 6 Vo, the map Ga : V§a —>• Uo is quadratic-like, and the family
{Ga)aqv0 is Mandelbrot-like.
c) When n tends to infinity, the family (Gf<T,n)ffgTn(wrn) tends to (Ga)aeyo,
uniformly on each compact in its domain of definition in Tn(Wn) x C.
Proof : a,l) : For i < n, denote by V^ the component of /A"1 04) which
contains at. I f l Vx'i+l is holomorphic and
proper with no critical point, thus is an isomorphism. The map fx : Vx —> Vx x
is a ramified covering of degree 2. It follows that / " : Vx —> Vx is a ramified
covering of degree 2, and Gx = gx o /£ : V{ -+ £/* is quadratic-like.
a, 2) : For n big enough, when A ranges on a curve which follows closely the
boundary dWn, the point A follows dWn, the point an(X) = A + n ranges on
a curve close to dV0, thus an(X) on a curve close to dV0 , so finally GX(LJX) =
gx{anW) ranges on a curve in Uo \ U'o, which makes 1 turn around 0.
b,l) : Define Vâl as the connected component of L~}_i(Vo) which contains
ai(0) = /o(w o ). Then LCT_i : Vg'î —> VQ is an isomorphism, while fo : Vla —»
Vo)(r>1 is a ramified covering of degree 2. So (?„ = j 0 ° I j : VQO —> t/o is
quadratic-like.
b,2) : When a follows dVo, the point La(u)o) = {(po)1^) follows dVo, and
Ga{oJo) ranges on a curve in UQ\UQ, making 1 turn around 0.
c) : This follows from the fact that / " , where A = T~1(CT - n), tends to
La when n tends to oo. •
8 About Mandelbrot-like families

In this section, we consider a Mandelbrot-like family h = (h\)xew, provided
with a horizontally analytic tubing © ([DH Pol-like]). More precisely we
assume:
• W is a Jordan domain in C, with C 1 boundary.
• For each A € W, the map ©A is a quasi-conformal embedding of AR2tR
in C ; for each Z £ A&R the point © A ( ^ ) depends holomorphically on A.
We set Cx = Qx(S]p) and C'x = ex(SR) ; we denote by Ux (resp U'x ) the
Jordan domain bounded by Cx (resp. C'x ).
• hx is a quadratic-like mapping U'x —> Ux ; the map h : (A, z) M- (A, hx{z))
is C-analytic and proper W ->• 14, where U is the open set {(A,z) | A €
W and z G Ux} and 14' is defined similarly ; we denote by UJX the critical
point of hx.
• We have 0 A (^ 2 ) = hx{Qx{Z)) for Z e SR. _ _
• The map h extends continuously to a map W —>• U, and 0 : (A, Z) H->
(A, ©A(-^)) extends continuously to a map W x AR2 R —>• U, such that ©A is
injective on AR2:R for A € dW.
The map \ i-^- uix extends continuously to W.
• For A e dW, we have hx(u)X) e Cx.
• When A ranges over dW making 1 turn, the vector hx(u>x) — UJX makes
1 turn around 0.
We denote by Mh the connectedness locus of the family h:
M h = {A e W | K(hx) connected} = {A | wA e K{hx)} .
It is proved in [DH Pol-like] that Mj, is homeomorphic to the Mandelbrot

set M. More precisely, for any A G M h , there is a unique c = x(A) e M such
that hx is hybrid-equivalent to z i-> z2 + c, and the map x s o defined is a
homeomorphism of M h onto M.
Using the tubing 6 , we can extend % to a homeomorphism x© of W onto
M/M = {c e C | ^M(C) < 2Log(R)}, where £ M is the potential function
defined by M.
The map xe is defined in the following way:
We first define foonW. For X eW and z 6 Ux\K(hx), there is a A; such

that zk =hkx{z)£UxÛ'x = ex(AR2tR), then we set Qx(z) = jgLog\exl(zk)\.
If there are two values for k, they give the same value for Qx(z). We set
Gx(z) = 0 for zEK(hx).
We denote by Hx the set of points z such that Qx{z) < Gx{u)X) : if A 6 Mh,
we have Hx = K(hx), and if A € W \ Mh the set Hx is limited by an 8-shaped
curve.
We define <j>x : Ux - Hx -> Afi2iW, where Log(px) = Gx(ux), by the
conditions :
• 0A = G^1 on t/A \ U'x ;
• Â(fcA(z)) = (<M*))2 for zeU'x\Hx;
• (j>x is continuous on Ux\ Hx.
For A e W \ Mh, we have hx(uix) E Ux\ Hx, and Xe(A) is defined by
For A e M h , we set xe(A) = x(A).
It is proved in [DH Pol-like, IV-4] that the map xe defined this way is a
homeomorphism of W onto an open set in C, which is easily seen to be WM
under the hypotheses made here, and that xe is quasi-conformal on W \ Mh
for any W relatively compact in W (Prop. 20).
Lyubich has proved ([Ly]) that there exists a quasi-conformal mapping
X of a neighborhood W* of M h in W onto a neighborhood W^ of M in
C extending x- I* t n e i i follows from Rickmann's lemma ([DH Pol-like I-
5, Lemma 2],[Ri]) that xe is quasi-conformal on W for any W relatively
compact in W, as conjectured in [DH Pol-like].
Let now F be an arbitrary compact set in AR2>R, and denote by F m the

preimage of F by Z H* Z2™.
Baby Mandelbrot sets axe born in cauliflowers 29
PROPOSITION 3.- We have
xe'irj(rm)) = {A g w | h™+1(wx) e e A (r)} .
Proof: Setting c = xe(A) anM{c) = 0c(c) 6 Tm ; <£c(Q™+1(0)) 6 T ; /i™ + 1 K) 6 6 A ( r ) . •
COROLLARY.- We have
MhU{\ | (3/c) /»*(a,A) e
9 Appearance of E(en)
Let us come back now to the situation of sections 4 to 7. We shall apply the
above Corollary to the family Gn = (G\ : Vx —> U\)\£wni or equivalently to
the family {G^n)aeyQ.
We fix R > 1. For each n, we shall choose a tubing 0 n = {&x)xewn for
G n , with 0 A : AR2R —> U\ \ Vx' quasi-conformal.
We first take a tubing 0f for f denned on A' x Agi ^ with A' a neighbor-
hood of 0 in A. Then the straightening theorem provides a homeomorphism
ipx '• J(fx) —> Jt(x) induced by a map which is AV-quasi-conformal with a Kt
independent of A for A in a neighborhood A" of 0. In particular for the center
An of Wn denned in section 7, we get ip\n : J(f\n) —> JCn with some tn. This
€n is determined by Tp(en) = — n, where Tp is the lifted phase for the family
of polynomials (Pt), taking ip$b$ as a base point on the right side. It follows
that -~ir ~ -n and cn ~ ^ j .
We denote by ro(e n ) the set Jtn enlarged by the factor R* as in section
3, and by r m (e n ) the preimage of r o (e n ) by Z >-» Z2™, agreeing with the
notations of section 3 and of Prop. 3.
We can then choose for each n big enough a homeomorphism 0 " : AR2>R —>
Z7An N KA'n so that
• 0 ° is X-quasi-conformal with a K independent of n ;
• lOn(/C/ ) — JXnY'-'nK^)) '•Or |zi| — It ,
• e°n{Z) = il>£{Rri.Z) for Z € r o (e B ).
When A ranges in Wn, the set J(f\) remains a Cantor set contained in
U'x N VA and undergoes a holomorphic motion. The sets dU\ and dVx' also
undergo a holomorphic motion. By Slodkowski's theorem, one can find a
holomorphic motion (tA) of U\n \ Vx' which induces the above mentioned
ones on dU\n, dV\n and J{fxn). Setting Q\ = ixoQ^, w e obtain a horizontally

analytic tubing 0 n for Gn such that 6A(r o (e n )) = J ( / A ) for each A in Wn.
PROPOSITION 4.- The set E(en) appears quasi-conformally in MFC\Wn.

Proof : By [DH Pol-like] completed by [Ly], the map xe n i s a quasi-
conformal homeomorphism of Wn onto WM- Taking F = ro(e n ), the set
M U Urn^M^™) is precisely the one denoted by E(en) in section 2, so the
Corollary of Prop. 3 tells that the image of E(e n ) by Xel i s the set
Xn = M G n U {A e Wn I (3k) Gkx(ux) 6 J(fx)} .
It is clear that X n c M f n W n .
The points £ € E(en) which are either Misiurewicz points in M or parame-
ter values for which R~* •(^'MCO) 2 '" IS a (repelling) periodic point in Jln form a
dense set in the boundary d=.(en). For such a point f, the point A = x«~nl(0 i s
a parameter value such that, for some k, the point G*(k/;0 is a repelling point
either under G\ or under fx ; so £ € 9M F . Therefore X i
10 Proof of Theorem 2
We have seen that \n — Xen '• Wn -> WAf is a quasi-conformal map which
induces a homeomorphism Xn -> H(en) with en ~ ^ j .
Choose C > 0 such that C.WM is containd in the open disc D±, and define
Xn : Wn -> T_n(C.iyM) by \ n = hn o X n o T" 1 , where /i n (Z) = C.Z - n, and
X : Un>n0 Wn ~> U n > no T_ n (C.W M ) by X|vvn = Xn-
Set ro(e) = — ^ for e e 5*, and 77 = T0 O r" 1 : S ->• 5 0 = r o (5*) =
{Z I |Z| > -z= \Arg(-Z)\ < &}. Remind that ?7(Z)/Z tends to 1 when Z
tends to 00 in 5 ; reducing #0 if necessary we get that rj (Z) also tends to 1.
LEMMA.- The map x can be extended to a quasi-conformal homeomorphism

V> : 5 -4 So which induces 7/ on dS.
Proof : The map Tn o Xn o 7 1 n : Tn{Wn) -> C.W M has a limit Xoo :
Vo —> CWn when n tends to infinity, obtained by applying the straightening
theorem for Mandelbrot-like families to the family {Ga)a^yo, and multiplying
by C.
The map Xoo extends continuously to closures, inducing on the boundaries
the C'-diffeomorphism
C4'M O OO 1 O g0 o r - 1 : dV0 -> aK0 -»• 3t/ 0 -+ 5^ 2 -

We denote by 7 ^ the inverse diffeomorphism d(C.WM) -> dV0.

For each n, the map Tn o x" 1 o T_n also extends continuously to closures,
inducing on the boundaries a diffeomorphism 7 n : 8{C.WM) —+ d(Tn(Wn)),
and 7 n tends to 700 in the C 1 topology on d(C.W M ) when n —> 00.
Let £° be the vertical line defined by Re(Z) = -\, which separates CWM
from T-{(C.WMYLB.n& let 5° be the vertical strip bounded by £° and Ti(£°),
so that 5° 3 C.WM.
Let £ be a C 1 curve which coincides with £° outside of a compact set, and
separates the closures of Vo and T_i(VJ>), and let B be the domain limited by
£ and T\ (£), so that B contains the closure of V^. Let 0 be a diffeomorphism
£° —*• £ which coincides with the identity outside a compact set.
One can find a diffeomorphism v^ : Bo \ C.WM -» 5 \ Vo which coincides
with 700 on d(C.WM), with /? and T\ o /? o T_j on 9B 0 , and with the identity
outside a compact set. This map v^ is quasi-conformal.
There is a no such that, for n > no, the closure of Tn(Wn) is contained
in B, and one can find a C'-diffeomorphism vn : Bo \ C W ^ - > B \ ^ ( W ^ )
which coincides with 7 n on 9 ( C W M ), with /? and Ti o /? o T_i on 9B 0 , and
with the identity outside a compact set independent of n. These maps vn can
be chosen so that they are A'-quasi-conformal with a K independent of n.
Then the maps (T_n o vn o T n ) n > no and (x~') n >n 0 match up to a quasi-
conformal homeomorphism v : H° -> W, where H° and H are the regions on
the left of T_ no+ i(£°) and Tl n o + i(£) respectively. This map coincides with
the identity outside a horizontal strip.
One can easily extend v to a quasi-conformal homeomorphism v : C —> C
inducing x^ 1 a l s o f° r n < no- We can then modify v on a neighborhood of
dSo, keeping it quasi-conformal, so that it agrees with r?"1 on dS0- Then the
inverse map r/> satisfies the required properties:
• ij> induces a quasi-conformal homeomorphism 5 -> So ;
• V;ivvn = Xn for each n ;
• ty coincides with 77 on OS. •
Now r0 o ip o T is a quasi-conformal homeomorphism of S* onto itself which
coincides with the identity on the boundary, we extend it to a quasi-conformal
homeomorphism ip : C —> C by taking the identity on C \ S*.
The map tp~l induces a homeomorphism of
r = {o}ur-1(Ur-»(=(e»)))
n>no
onto {0} U Un>no ^ri, so T' appears quasi-conformally in MF-

The set T defined in section 3 differs from T' by two slight modifications:
• the sequence (en) which is equivalent to ^j is replaced by ff exactly ;
• the map r^ 1 is replaced on T^n(WM) by its affine tangent "map at - n .
32 A. Douady, X. Buff, R. Devaney and P. Sentenac.
These do not affect the quasi-conformal type, in other words there is a

quasi-conformal homeomorphism of C onto itself which maps T onto T'.
This finishes the proof of Theorem 2.
11 Proof of Theorem 1
Let Co be the cusp point of a primitive hyperbolic component of the interior
of the Mandelbrot set M, and denote by Q\ the map 2 H z 2 + c0 + A. Let A
o
be the component of K^ which contains 0, and k be the period of A, so that

QQ induces a proper map A -> A of degree 2. We suppose CQ ^ j . There is
a parabolic point a 0 6 dA of period k, with (<2§)'(ao) = 1.
One can find a sequence (6n)ngz such that :
(BO1) : 6 n+1 = Q 0 ( V ) f o r a l l n ;
(BO2) : bN = 0 for some N ;
(BO3) : bn i A for n < N ;
(BO4) : b-mk -» Qo when m —> oo.
Indeed, there is a Fatou coordinate </>Q0 for QQ defined on a domain QQO
with
• aoe dil^0 ;
• D(a0, r) \ A C QQ0 for some r > 0 ;
• Ko&Qo) = C~ = {Z\ Re(Z) < 0}.
Let b' be the preimage of 0 under QQ which is not in Qo~'(A). Since
a0 e J{Qo), there is a 6 € £>(a o ,r) and an N such that Qo~A(b) = 6', thus
Q0N(6) = 0. Then b e ft+0, and we can define b_mk € fi+0 by (^$0(6_mfc) =
0go(6) — m. We define bn by 6n = Qo(6_mjt) if n = — mfc + ^ with v > 0 (this
is independent of the way we write n as — rnk + v ). Then (6 n ) ne z satisfies
(BO1) to (BO4).
Note that, for a sequence obtained by this construction, we have 6_mt €
QQO for m > 0. But more generally, for an arbitrary sequence (6 n ) n e Z satis-
fying (BO1) to (BO4) and an arbitrary fi+ which is the domain of a Fatou
coordinate with range containing C~, there is a mo such that 6..m* G f2T for
?n > m 0 .
Remark : We shall actually prove a result which is stronger that the one
stated in Theorem 1. We shall prove that, for any sequence (6n) satisfying
(BO1) to (BO4), there is a v and a map C —> C which makes T appear
quasi-conformally in M, and which for each m maps the center hm(Q) of the
m-th copy of M in T to the point b_mk+v.
In this situation, we have (see Fig. 6):

LEMMA.- One can find Jordan domains UQ, U'O, VQ with C 1 boundary such
that :
(JD1) :AcU^cUlcU0;
(JD2) : QQ induces^ proper map f0 : UQ-> Uo of degree 2 ;
(JD3) : VOCUO\ U'Q ;
(JD4) : be e Vo for some £ = -mk < N, and Q$~l induces an isomor-
phism <7o : VQ —»• Uo.
Proof : One can find Jordan domains U and U' with Cl boundary such
that A c t / ' and that QQ induces a quadratic-like mapping U' —>• U, and
one can choose U contained in an arbitrary small neighborhood of A. The
sets Qo(A) for 0 < i < k are disjoint (due to the fact that CQ is the root of
a primitive hyperbolic component), and one can choose U so that the sets
Ql(U) are disjoint.
For n < N, denote by Vn the connected component of Qo^N~n\U) which
contains bn. Then the sets Vn are disjoint, for each n the set Vn is a Jordan
domain and QQ~U induces a homeomorphism Vn —> U. For the Poincare
metric of C \ {Jo<i<kQo(^-)> ^ n e diameter of Vn is bounded, since there is
a branch of QQ1 defined on the universal covering of this set and mapping
it into itself ; so the euclidean diameter of V-mk tends to 0 when m -> oo.
There is an mi > mo such that V-mk C U for m > mi.
Now the domains U, U' and V-mk satisfy (JD1), (JD2) and (JD4) for any
m> mi.
SUBLEMMA.- Let P C C+ = {Z | Re{Z) > 0} be a topological closed disc
such that P fl 2m.P = 0 for m ^ 0. Then, given s0 > 0, one can find a
C 1 -curve C such that :
(API) : C C E = {Z | 0 < Re(Z) < s0} ;
(AP2) : T2vi(C) = C ;
(APS) : \.C n C = 0 ;
(AP4) : (Vm eZ) CD 2m.P = 0.
Proof of sublemma : Take fi > supx+yiep^-. Let J be a C^-arc joining
s0 - fj.soi to s 0 + fJ,soi in {Z = x + yi | |y| < /ix} and avoiding (Jmez 2 m-i3-
Set J m = 2~ m .J. Take si < s 0 such that 2/xsi < | . Consider the segments
J^ = [2" m (s 0 + fj,soi) si + 2/xsji], J" = [si + 2^sxi sx + (2TT - 2fj.Si)i],
J™ = [Sl + (2TT - 2/iSj)i 2?ri + 2- m (s 0 - /j,soi)]. Then, for m big enough, J m
is contained in {Z \ 0 < Re(Z) < inf(2si,s0)}, and
is a piecewise C'-curve satisfying (API) to (AP4). One can smooth it and

obtain a C'-curve C still satisfying (API) to (AP4). •
End of proof of the lemma : Proceeding like in section 8 and taking
logarithm, one can find a C^diffeomorphism h of the universal covering of
y = 2n - fix
y =:.
y — 2n - 2fix
y = -2/xtf
Figure 5: The proof of the sublemma.
U \ K(fo) onto a vertical strip S = {Z | 0 < Re(Z) < So} which conjugates
/ to Z i-)- 2.Z and the deck transformation to T27ri. Take P = h(V) with V
a lift of V-mik. Note that P is a topological closed disc, with P n 2 m .P = 0
for m ^ 0.
By the sublemma, one can find a C1-curve C in £, invariant under T27ri,
avoiding | j 2 ~ m . P , and such that | . C lies between i.R and C. There is a
unique m 2 such that the set 2~m2.P lies in the strip between C and \.C. Set
^ = — (rm + m2)fc. The projections of h~l(C) and ft"1^^) are C 1 Jordan
curves which bound Jordan domains Uo and (7Q satisfying(JDl) and (JD2),
and taking Va = VtC\ Qo^'îUo) the conditions (JD3) and (JD4) also are
satisfied. •
We can now finish the proof of Theorem 1. Take for instance U\ = Uo
for every A e C . For A close to 0, let U'x (resp. V\ ) be the Jordan domain
bounded by the component of Q^k(dU\) (resp. Q^t(dU\)) close to dU'o (resp.
to dV0 ) (see Fig. 6).
There is a simply connected neighborhood A of 0 such that dU'x and dV\
undergo a holomorphic motion when A ranges in A, and that the inclusions
Ux
<?A0 = Q3
Figure 6: The mappings F\o and F\.
U'x C Ux, Vx C Ux \ U'x hold for A g A. Then Q\ and QA induce maps

fx '• U'x —> C/\ and <?A : Vx —>• £/\, and we are in the situation described in
section 4. The condition <5'(0) ^ 0 follows from [DH/O] Expose XII (ABP),
Cor. 2 du Th. 1, or [TL].
By Theorem 2, the set T appears quasi-conformally in Mp. But Mp
appears in M by A — i >• CQ + A. Indeed, if A € Mp, the critical point u)\ = 0
has an orbit under F\ which stays in £/A, SO <5"(0) does not tend to oo and
c0 + A 6 M. If A € dMp, the sequence of functions (A' >-> Fx,(uJx'))n is not
normal at A, the same holds for (Q"'(0) a n d c0 + A 6 9M.
This finishes the proof of Theorem 1.
Final remarks : To say that T appears quasi-conformally in M is a rather

weak statement as we don't give a bound for the dilation ratio. Computer
pictures seem to indicate that, if the radius R is chosen in an appropriate
way, this dilation ratio is not too big. But we don't have an estimate.
The statement of quasi-conformal appearance does not imply the estimates
stated in section 1 on the diameter of the sets Mn and their distance to cQ.
These could be derived from the computations made. We leave that to the
courageous reader.
The drawings are done by X. Buff using algorithms and programs by Chris-
tian Henriksen.
References.
[D/AMS] A. DOUADY, Does a Julia set depend continuously on the Polyno-
mial ?, Proceedings of Symposia in Applied Math., 49, 91-139, (1994).
[DH Pol-like] A. DoUADY & J. HUBBARD On the dynamics of polynomial-

like mappings, Ann. Sci. ENS Paris vol 18 (1985), 287-343.
[DSZ] A. DOUADY, P . SENTENAC & M. ZINSMEISTER, Implosion parabolique
et dimension de Hausdorff, C.R.A.S. t.325, Serie I, 765-772, (1997).
[D/Sl] A. DOUADY, Prolongement de mouvements holomorphes [d'apres Siod-
kowski et autres, Seminaire Bourbaki 775, (1993).
[La] P . LAVAURS, Systemes dynamiques holomorphes, Explosion de points
periodiques paraboliques, These de doctorat, Universite Paris-Sud in Orsay,
(1989).
[Ly] M. LYUBICH, Feigenbaum-Coullet-Tresser Universality and Milnor's
Hairiness Conjecture, Preprint IHES/M/96/61, to appear in Annals of Math.
[Sh] M. SHISHIKURA, Bifurcation of parabolic fixed points, in this volume.
[SI] Z. SLODKOWSKI, Extensions of holomorphic motions, Prepublication
IHES/M/92/96, (1993).
[TL] TAN LEI, Similarity between the Mandelbrot set and Julia sets, Comm.
Math. Phys., 134, 587-617, (1990).
Adrien Douady, Departement de Mathematique, Universite de Paris-Sud,
Bat. 425, 91405 Orsay, France,
e-mail: Adrien.Douady@ens.fr
Xavier Buff, Universite Paul Sabatier, Laboratoire Emile Picard, UFR MIG,
118 route de Narbonne, 31062 Toulouse Cedex, France,
e-mail: buff@topo.math.u-psud.fr
Robert Devaney, Department of Mathematics, Boston University, Boston MA
02215, U.S.A.
e-mail: bob@math.bu.edu
Pierrette Sentenac, Departement de Mathematique, Universite de Paris-Sud,
Bat. 425, 91405 Orsay, France.
e-mail: sentenac@topo.math.u-psud.fr
Modulation dans l'ensemble de
Mandelbrot
Peter Hai'ssinsky
July 20, 1999
Abstract
We prove A.Douady's and J.H. Hubbard's Tuning Theorem from
the sketch given by A. Douady. This result explains why the Mandel-
brot set contains "little copies" of itself, and where they can be found.
We also give a topological model of the tuned Julia sets, and some
further corollaries.
Le but de cet article est de montrer le Theoreme de Modulation
de A. Douady et J.H. Hubbard, ainsi que de corollaires immediats. Ce
resultat explique la presence de petites copies de l'ensemble de Man-
delbrot dans lui-meme. Nous donnons aussi un modele topologique de
l'ensemble de Julia d'un polynome module, ainsi que quelques corol-
laires.
Nous commengons cette introduction par des rappels afin de situer le con-
texte et pouvoir ensuite enoncer les resultats principaux.
Rappels
On se restreint a la famille quadratique, bien que de nombreux resultats
restent vrais pour des polynomes de degre plus eleve. La particularity des
polynomes est l'existence a l'infini d'un point super-attractif, de bassin con-
nexe, et dont le bord est leur ensemble de Julia.
On considere la famille de polynomes
Z (-»• Z2 + C,
_. Kc = {z G C, Q™(z) ~h °°) l'ensemble de Julia rempli, et

Jc = dKc l'ensemble de Julia de Qc;
Ce sont des compacts totalement invariants, et Kc est de plus plein.
Nous allons enumerer les principaux resultats dont on aura besoin. Les
demonstrations se trouvent dans [9], sauf mention differente.
Proposition 0.1. (I, Chap.3, Prop.l)
J soit 0 G Kc et Kc est connexe,
De deux choses I'une. , .. -, , „
sott Kc est un Cantor.
37
38 Peter Haissinsky
Pour tout c dans C, on peut definir la eoordonnee de Bottcher <f>c au

voisinage de l'infini qui conjugue Qc a z >—> z2, <pc tangente a l'identite a
rinfini. Si Kc est connexe, alors <f>c : C \ K, -—> C \ D cst une representation
conforrne de C \ Kc.
On definit le potentiel de Qc par Gc — log|<pc|. On prolonge G, a C comine
suit.:
- pour : e C \ A',., on utilise la relation Gc{Qc{z)) — 2Gr(z).
- pour ~ G /\'f, on prolonge par continuite en posant Gc(z) = 0.
On definit les equipotentiellcs par G~'(?/), r\ > 0; si /; > G f (0) alors c'est
un cerele topologique. Le rayon externe 1ZC(9), ou 0 est compte en nombre
de tours, est 0,7 '(r/e'2"^). pour r/ assez grand. On le prolonge tant qu'on ne
"bute" pas sur un point precritique a; de Qc, i.e. Arg<pc(w) = ff. Pour tout z
de TZC{Q)- on dit que 2 a pour argument 0. Si lim,,_>o 0^l{r}e2lTc(>) existe alors
H,.{9) aboutit.
Exemples.
: *
*****
9*
Figure 1. Ensembles de Julia quadratiques.

Si A'r est, connexe et localement connexe, on peut definir ie lacet de
Caratheodory par l'application j c : T -> 3A'O fonction continue et surjec-
five, qui a 0 associe le point ou TZC(O) aboutit.
Proposition 0.2. (I, Chap.8, 2, Prop.2)
\f(c.O) € C x Q/Z. TZ,:(8) aboutit ou bute sur un point precritique de Qc.
Soit M = {v e C, <fi/ que Kc soit connexe} = {c G C, 0 G A",}
l'enscmble de Mandelbrot. Qet ensemble est compact connexe et plein (I.
Chap.8, Theoreme 1).
L'application
&(z)
Modulation dans 1'ensemble de Mandelbrot 39
est holomorphe par rapport a (c,z), partout ou 0C est definie.

Cette application induit l'uniformisante
$ M : C \ M -> C \ D ,
c H-> </>c(c)
de potentiel GM = log|<E>A/|-
On definit les equipotentielles et les rayons externes de M de la meme fagon.
o
Pour k > 1 fixe, {c & C tq 3 A:-cyde attractif pour Qc} CM. et chaque
composante est appelee composante hyperbolique de M de periode k.
Remarque. Vc 6 C, Qc admet au plus un cycle non repulsif (car degQ c = 2).
Pour toute composante hyperbolique W. on a la representation conforme
pw : W —> D, qui a c associe le multiplicateur du cycle attractif de Qc (II,
Chap.19, Th.l). L'application pw se prolonge continument au bord; on note
/%,'(0) le centre de W et p^'(l) sa ratine.
Definitions.
V{) — {centres des composantes hyperboliques} = {c e C , 0 soit periodique) ;
T>\ = {racines des composantes hyperboliques} = {points paraboliques} ;
£>2 = {points de Misiurewicz} = {c g C, 0 soit strictement preperiodique}.
Proposition 0.3. (II, Chap.13, Th.l et [19]) V0 e Q/Z, KM{6) aboutit a
c e dM;
- c € T>\ si 9 est a denominateur impair,
- c € T>2 sinon.
Theoreme 0.4. (Relations entres arguments dynamiques et argu-
ments externes)
- (I, Chap.8, I, Cor.2): Pourc&M, Argcc—ArgMc.
- (II, Chap.IS, Th.l, Complement, Chap. 18, Prop.l): Pour c € X>o, il existe
1ZM{0-) el TIM{8+) aboutissant a la racine de la composante de c. Us aboutis-
o
sent dans le plan dynamique a la racine de la composante de Kc contenant c,

en etant adjacents a cette composante.
- (II, Chap. 13, Th.l, Complement et [21]): Pource Vx, il existe TlM{9^) et
11 M{8+) aboutissant a c. Us aboutissent dans le plan dynamique a la racine
o
de la composante de Kc contenant c, qui est un point parabolique, en etant

adjacents a cette composante.
- (I. Chap.8, III, Th.2): Pour c G T>2, il existe 0,, ..., 9V tels que leurs
rayons externes aboutissent a c dans M. 11$ y aboxitissent aussi dans le plan
dynamique, et ce sont les seuls.
Je suggere aussi au lecteur de se reporter a l'approche de J. Milnor ([17]).
Applications a allure polynomiale, mouvements holomorphes et fig-
uration rigide. Nous decrivons ici les outils analytiques dont nous aurons
besoin.
40 Peter Haissinsky
Definitions.
> line application a allure polynomiale de degre d > 2 est la donnee d'un
triplet ([/', U, g) ou U' et U sont des disques topologiques tels que U' CC U
et tels qu'on ait g : U' —¥ U holomorphe propre de degre d.
> K(g) — ng~n(U') est l'cnsemble de Julia reinpli;
> J(g) = <9/C, est l'ensemble de Julia.
Definition. Deux applications a allure polynomiale ([/', £/, / ) et ( V , V, g)
sont hybridement equivalentes s'il existe une application quasiconforme ip :
[/ -> V7 qui conjugue / a g et telle que dtp = 0 pp. sur A"(/). On note
/ ~M> -9-
Pour la definition et les proprietes des applications quasiconformes on peut
se reporter a [1].
Independarninent des propositions suivantes, on peut montrer que de nom-
breuses proprietes des ensembles de Julia (remplis) de polynomes se trans-
posent a ce cadre. Us sont d'ailleurs etroitement lies par le theoreme suivant:
Theoreme 0.5. (Theoreme de Redressement, [10], /, Th.l)
Soit (V, U, g), une application a allure polynomiale de degre d > 2. // existe
un polynorne P de degre d tel que P ~hyb 9-
De plus, si K(g) est connexe, alors P est unique a conjugaison affine pres.
En particulier, si g est quadratique avec un ensemble de Julia connexe, alors
il existe un unique c G M tel que Qc ~hyb </•
Definition (famille analytique). Soit A une variete complexe et soit g =

(gx : U'x —> CA)AeA une famille d'applications a allure polynomiale de degre
d > 2. On pose U = {(A,z), z G V\}, W = {(A,z), z € U'x) et on definit
g(X, z) — g\(z). Alors g est une famille analytique si les conditions suivantes
sont satisfait.es:
(i) W et U sont homeomorphes a A x D au-dessus de A,
(ii) la projection de 14' dans A est propre, ou la fermeture de W est prise dans
W,
(iii) l'application g : 14' -^tU est C-analytique et propre.
Remarque. Si le bord des domaines U\,\ € A, sont parametres par un
mouvement holomorphe (voir ci-dessous) alors les conditions (i) et (ii) ci-
dessus sont satisfaites.
Theoreme 0.6. (Dependance par rapport aux parametres, [10]; //,

Th.2, Prop. 14; IV, Prop.21)
Soit (.9A)A€A une famille analytique d'applications a allure quadratique.
(i) II existe x '• A —>• C continue telle que VA € A, gx ~hyb QxW-
(ii) On suppose que A est homeomorphe a D. On definit Mg = {A £
A, K(gx) connexe}.
Si x '• Mg -> x{Mg)(C M) est propre et x(Mg) est connexe alors x est
Modulation dans I'ensemblc de Mandelbrot 41
mi revetement rarnijie. De plus, si Mrj CC A. alors \ : Mg —> M est un

revetement ranrific de dearc d fini que Von peut calculer en comptant le norn-
hre de tours lie fi$u.'x) — U\ (u>\ : point critique de ij$ quand mi pareourt une
fins un lu.ret en.tourant Mfr
Definitions, (a) Sous les hypotheses de (ii), si M9 est compact dans A et le

degrc de \ cst d = 1, alors \ est un homeomorphisme et on dit que la f'amille
est uumdelhrote.sqm'.
(b) Si \ : My ~~> M \ {l/'i} o s t " n homeomorphisme alors g est dite semi-
lnnudclbrot.osque.
C o m p l e m e n t . M. Lyubich a dcmontre que x es>t quasiconforme lorscjue la
f'amille <>st inandelbrotestjue ([15]). i.e. il existe une extension quasiconforme
an plan de \ .
Rappelons <|u'un mouvement holomorphe d'un ensemble A' C C au-dessus
d'une \ariete complexe A est une application /, : A x .V -> C qui verifie les 3
conditions suivantos:
(i) il existe Ao 6 A tel <juc pour tout x G A', <(Ao,x) = x;
(ii) a .r G -V fixe. A i-> i.(X.x) est holomorphe:
(iii) a A 6 A fixe, x >-> i.(\,x) est injective.
Lc> resultat le plus fort, sur les mouvements holomorphes est un theorcme
de Z.Slodkowski ([20]):
Theoreme 0.7. Un mouvement holomorphe au-dessus du disque unite se

proloni/e en un mouvement holomorphe de Unite la sphere. De plus, a A fixe,
I'application, z ^> i.(X.z) est quasiconforme.
Enfin. nous generalisons la notion d'application a allure polvnorniale:

Definition. Soit / tine fraction rationnelle et soit. P un polynome ayant son
ensemble de Julia rempli K{P) connexe. On dit que P figure de nmniere
vigido dans f s'il existe un voisinage U de K(P), une application continue
injective ^ : U —> C tels que:
(a) pour tout z G K(P), v> ° P(z) - I ° ^{z);
(1)) 1'application y? G H',^ pour tout;; < 2, i.e. y admet des derivees ]>artielles
an sens des distributions V integrables;
(e) 0^ = 0 ])p. sur K(P).
Pour ]>lus de precisions sur les figurations rigides. on peut se reporter a
[11],
Resultats principaux
La premiere partie de cet article consiste a demontrer le theorerne ci-dessous
scion l'esciuisse de [6].
42 Peter Haissinsky
Dans la suite, W designe une composante hyperbolique de centre cQ et de

ratine c.\. On dit que W est primitive si la fleur du point parabolique de c\
n'est formee que d'un petale attractif.
Theoreme 1. de Modulation (A. Douady & J.H. Hubbard)

(a) Si W est une composante primitive de M de periode k, on peut trouver
an voisinage A de W. et deux families d'ouverts (£/,', Lr,.)ceA tels que, (gc =
QrU'l- '• U'c ~^ U,)cqA soit une famille mandelbrotesque.
(b) Si W n'est pas primitive, on peut settlement construire A, les U'c et les Uc
de facon que A soit un voisinage de W \ {c-\}, et que \ induise une famille
sem,i-mandelbrotesque. De plus, Q1/4 figure de rnaniere rigide dans QC1.
Figure 2. M et un zoom.
On note Mco = Xcol(M) obtenu par ce theoreme.

Complement (M.Lyubich, [15]). Si W est, primitive, la petite copie de
M est quasiconforme a M.
La demonstration consiste a construire des ouverts U'. et. Uc afin que les Q*
soient des applications a allure quadratique. Ensuite, on applique la theorie
des applications a allure polyiiomiale qui est bien adaptee si W est primitive.
Dans le cas non primitif, on n'obtient pas Mg CC A: une etude plus detaillee
pst necessaire.
Dans la seconde partie, nous etudions l'operation de modulation. A c0 €
VQ fixe, on note
Co ± • : M -¥ MCo
c->X^(c).
Nous montreroiis le theoreme suivant (voir §6 pour un enonce precis):

Theoreme 2. L 'operation de modulation agit sur les ensembles de Julia rem-

plis comme suit: Kcoxc est homeomorphe a KCQ dans lequel on a remplace les
composantes connexes de I'interieur par des "copies" de Kc.
En appendice, nous etudions la dependance du point d'aboutissement d'un
rayon en fonction du parametre c.
Notes. Cet article est tire de [11]. Les representations d'ensembles de Julia,
ainsi que de l'ensemble de Mandelbrot ont ete faites a l'aide d'un logiciel de
C.T. McMullen.
Remerciements. Je tiens a remercier Tan Lei pour son encouragement et
pour tout le temps qu'elle a accorde a cet article. Ses nombreux commentaires
ont graridement clarifie cette exposition, et je lui dois aussi des simplifications
de demonstration.
Partie I: Theoreme de modulation

1 Premiere construction
Soit W une composante de periode k, de centre CQ et de racine c\. On suppose
que Co / 0 et c\ ^ 1/4. Dans ces conditions, il existe deux rayons externes
Âf(î) et TZM{02), 0i < 92, qui aboutissent a cx (Figure 2). D'apres le
Theoreme 0.4, TlCo{9\) et, TZCo(92) aboutissent a yu qui est fc-periodique, et qui
O
est la racine de la composante connexe U\ de Kco contenant cQ (Figure 1). Soit

7] > 0. On appelle A l'ouvert delimite par UÔx) U {yl }WR(]t,{92) U (G(: = ri)
et contenant U\. On note A' la composante connexe de Q~*(A) contenant,
U\, qui est bien definie car U\ est /c-periodique.
Proposition 1.1. A' est delimite par quatre rayons externes d'angle 9y, 92,
d\ et 9'2 et par Vequipotentielle de niveau rf = r)/2k. De plus, A' C A et
Qk0 : A' —> A est propre de degre deux, et sa valeur critique est Co-
DEMONSTRATION. On pose IJX = Ql^l(Ui); tant que 1 • Ul+l est conjugue a l'identite sur D, car U{ ne contient pas de point
critique. En revanche, Qca\uk=v0 • Uo ~> U\ est conjugue a z ^ z 2 dans D, qui
se prolonge an bord. Le point (/i a done deux preimages, y[ d'argument interne
1/2 et lui-meme. Comme Qk0 est univalente au voisinage de y[, il existe 9[ >
9', tels que leurs rayons aboutissent en u! et tels que < nkJ -1 , , !'
[ 2K92 = 92 (mod 1).
On definit A\ selon l'enonce. On a:
U\ C Qka(A\) ==> A' - A\ et Q^ : A1 -> A est propre car e'est un
polynome.
44 Peter Haissinsky
a un
Comme A est connexe, Q£0U' degre qui est le meme que Qk0\un i-e-
2.
De plus, le point CQ n'a qu'un seul antecedent par Q^\A '• c'est done la
valeur critique. •
Proposition 1.2. II existe un ouvert Ai, voisinage deW\ {c{\, tel que pour
tout c G Ai, 7tc{0s) (resp. 1ZC(9'S)), s = 1,2, aboutissent a un mime point
yi(c) k-periodique (resp. y[(c), preimage de j/i(c) par Qkv), ce qui permet de
construire Qkc : A'(c) —» A[c) propre de degre deux. De plus, pour tout c G Ax,
le point yi(c) est repulsif et c est la valeur critique de Qkc.
DEMONSTRATION. On considere l'ouvert Ax delimite par les rayons TZM(&I)

et HM{Oz), et l'equipotentielle 77 contenant W (voir Figure 2). Par le Theoreme
A.3 (voir Appendice), il existe une fonction holomorphe c i-> 2/1 (c) telle que,
pour tout c G Ai, on a yi(c) = 7c (î) = 7c(#2) et j/i(c) est repulsif.
Les rayons d'angle B\ et 0'2 aboutissent pour tout c s Ai car leurs or-
bitcs n'entrent plus jamais dans ]6\, ^ [ (c/Appendice, Cor. A.5 et Prop. A.6).
D'apres la Proposition A.I, il existe aussi une fonction holomorphe c i-> y[(c)
telle que, pour tout c G A1; on a y[(c) = Jc{9[) = 7c(^)-
Soit^(c), cG A1; l'ouvert delimite par 7^c(î)u{yi(c)}UT2.c(^2)U{Gc = r/}
et contenant c. On note A'(c) la composante connexe de <5~fc(^4(c)) ayant
;</i(c) et ?/i(c) dans sa fermeture. Soit A[(c) l'ouvert defini comme Ax dans
Kco (voir Figure 3). Qkc(dA$c)) = dA{c), done ^ ( c ) = A'{c).
Pour tout c G Ax, Qkc : A'(c) -> A(c) est holomorphe propre, done admet
un degre (A(c) connexe) que Ton definit par d(c) = -^- fg
f):
F i g u r e 2: A t .
c f h o l o n i o r p h o et a v a l o u r e n t i e r e . d o n e c o n s t a n t e s u r A| e o n n e x e : d(<'
</(r n ) — 2. O n n u m t r o d(v m o m e quo c cst la valour c r i t i q u o .
Q1
.lusqu'ici. on a defini une famille {A'(c) —'-> A(c))ri.,\, holomorphc propre
de dej^re deux. Ses elements ne sont pas a allure polynoiniale car A'(c) n'ost
pas rolativement compact dans A(c). II est, done necessaire de changer un
pen leurs definitions. Pour cela. nous allons distinguer le eas primit.if du c;\.s
non pritnitif.
46 Peter Hai'ssinskv
e 4: A.
2 W primitive
D'apres le Theoreme 0.4, TZC] (#i) et TZCx (9->) aboutissent au meme point y\(c$
parabolique de vraie periode k; on construit de la meme fagon .4'(f'i) et A(ci).
Conime c\ est la raeine de W primitive. y$c$ n'a qu'un seul petale repulsif
dont l'axe est contenu dans C \ A'(c-i). Nous allons utiliser les resultats de
21] obtenus par implosion parabolique. Nous utilisorons d'abord un enonce
sur le plan des parametres, puis son pendant dynamique.
Theoreme 2.1. (Implosion parabolique - plan des parametres)

Avec les notations precedentes, pour tout s > 0. il exi.ste 9\ (z\9\ — £,9\\ et
0> &}9o< 9> + t[ tc.ls (juf
Soit 5 > 0 fixe assez petit que Ton determinera ulterieurement. On definit
A l'ouvert borne du plan des parametres borde par les rayons 1ZM(9I) et
T^-hi{(h)- par c-. ainsi que par l'equipotentielle (GM = if).
D'a]>res le Theoreme A.3, pour tout, c £ A, les rayons dynamiques 7Z.C{9\)
et 7£r(02) aboutissent a un meme point, y-2(c), qui est preperiodique repulsif
et (jui depend C-analytiquemont de c.
Potir c € A. on peut alors definir l'ouvert Uc contenant A(c) et borde par
les rayons TZc(0-[) et 1Zr(02), par 2/2(0). ainsi que Tequipotentielle (Gc — rj); et
on note aussi U'c la composante connexe de Q^k(Uc) contenant A'(c).
Si e est assez petit, alors 2/2(cj) appartient au petale repulsif de iji(c\), et
son image reciproque par la branche de Q~~k qui envoie y\ (r,) sur y\ (c-t) envoie
y-i{c\) a l'interieur de A(c\). On choisit e ainsi. Ceci implique notamment que

(ry,. = Q*|f/' : t-^ —> f/c)cgA e s t une famille d'applicatioiis a allure quadratique.
Notons'w = {{c.z), c G A, z G f/ 0 }, W = {(c,z), c G A, z G ^ } et
.9:W'->W, (c,z)^(r,5c(z)).
Coinme les domaines C/r bougent holomorphiquement, on en deduit que
les conditions (i) et (ii) de la definition d'une famille analytique sont verifiees.
De plus, fy est propre car Qkc{z) est polynomiale.
D'apres le Theoreme 0.6, on peut definir \ '• A —> C telle que gc ~^f,
o
Q\(c); f l e plus, x est continue sur A et analytique sur A/Co ou A/Co = {c G

A,K(gc) cormexe}.
Montrons que MCQ est un compact de A:
> Mco est ferme dans A car MCo — n n > 0 {c G A, gn(c, c) G U'}.
> Afin de montrer que MCo CC A, on va montrer que pour tout c assez
proche de dA. K{gc) est un Cantor.
Pour ceci, on s'appuie sur un resultat dynamique de [21]:
Theoreme 2.2. (Implosion parabolique - plan dynamique)
Soit A = A — Ai ; pour tout c G A, il existe n > 0 tel que Q"fc(c) soit separe
de c par nc{0,)\J {y2(c)} U nc(02).
On en deduit que pour tout c G A, K(gc) est totalement discontinu. Quant
au cas c $ A proche de <9A, ceci decoule du fait que dA C (C — M) U {c} et
A'(.9(:) C Ke.
Done, Afro CC A.
Remarque, x(^'i) = 1/4.
En effet, soit c G P1-'; il existe <pc tel que gc ~hyb Qx(c) et il existe za(c) G f/^ fixe
attractif tel que g'c{za(c)) — pw(c). Soit U un voisinage relativement compact
dans le bassin immediat de za(c) : (pc : U —>• <pc(U) est biholomorphe. Done,
<^c(^(f)) est un point fixe de Q x(c) et Q'x(r)(tPc(za{c))) = .^(z a (c)).
Par suite, on a
X • W -4 VI-'Q
et coinme \ est continue, x( c i) — 1/4-

Par construction, c G Lrc pour tout c G A: il s'agit done de la valeur
critique de gc (tout point d'image c est precritique pour Qc, et grc est quadra-
tique). Notons UJC = g~l{c) le point critique de gc. D'apres le Theoreme 0.6,
X : A/f.o —> A/ a un degre d egal au nombre de tours de gc{<jJc) — u)c — c — LJC
quand c parcourt une fois un lacet qui tourne (une fois) autour de MCo.
Puisquc 9c est quadratique, on peut appliquer le theoreme des fonctions
implicit.es pour montrer que c M- UJC est holomorphe (point critique simple).
Posons h(c) = c — UJC. Cette fonction est done holomorphe et ne s'annule
pas au voisinage de dA (car Mco est compact dans A). Pour montrer que
48 Peter Hmssinsky
d=l, nous pouvons done utiliser le principe de l'argument pour montrer que
(I est le nombre des racines de h(c) = 0 compte avec multiplicite. Nous alloris
proeoder en deux etapes: d'abord montrer que la seule solution dans A est
Co et ensuite qu'il s'agit d'une raeine simple.
Suit <•' line raeine; ceci signifie que l'application f/f- fixe le point critique.
En comptant ses points fixes, on en deduit; que Tun est c' et l'autre est le
point repulsif an bord de son bassin : il coincide avec 7,'(î) et 7^(6*2). Or ces
conditions earactorisent r.o. done c! — CQ. Le fait que c-n est une raeine simple
a ete inoiitre par Gleason (voir [9], II, Chap. 19. Lemine 2).
Par suite, d = 1 et la famille est mandelbrotosque.
3 W non primitive
Nous alloris changer la definition des A'(c) et A(c) en deux temps.
On fixe ="<) > 0 et on definit dans le plan des paramotres:
- LM(0\) ••= {<• e A2, «r fl;U ( r ) = Bi+eoCi^c)},
- Lu(02) := {r e A,. «r.9A/(r) = 62 - e0G\,(c)},
- A c A|. 1'ouvert contcnant \V et delimite par Lm(9l) et Lm(02).
Lemme 3.1. Soit 0 < e < £Q ; pour tout c dans A,
L(0,) = {z, argc(z) = 0, + sG?(z), 0 < Gc(z) < r,}

lC A[C))
L{0,) = {z. argc(z) = B2 - eG*(z), 0 < G,.(z) < r,}
sovi him d<>.fi.ni.es jusyu'c.n iji{c).
DEMONSTRATION. Si <• e A n M, il n'y a pas de probleme.

Soit done r € A \ M: L(B\) est mal definie si et seulement si il existe une
preimage *i de r dont l'argument t est. dans ]#i.fl2 + £T]2[ et le potentiel s est
plus grand quo \J'-j1 •
Dans le plan dynamique de QCo, il revient au meme de eonsiderer ZQ =
o,T()' o <E>A/(f) et d'etudier l'existence d'une preimage U>Q = </>,"."' (f>s+2"rt) de ZQ
p a r Q,.n telle (jue 0 < n < rf et 8X < t < 0] + es2.
Soit il' = {z = (p;:J (c°+'2"'t), 0 < .9 < 7;' et 6i < t < Oi + es2}; il suffit d e
montrer (|ue Torbite d'un point de Q,' ne tombe jamais dans le domaine de
definition do ;,,. a savoir B '=f {^(r"" 1 " 2 "'), 0 < ,s < //, et 6», + £,,.s2 < t <
« 2 -::.s-'}.
On a il' n D = 0. L'application de premier retour est dorinee par QkCo
car il' C A'(r). On note Q = Q^(Q') = {z = fcBl{e^21*1). 0 < s < 7/ et
ti\ < t < (9, + f.s 2 /2*}; etant donne que il n A'(c) C $2' et il n B = 9, L(6i)
es( bien definie. II en est de meme pour L{02). •
On a choisi e de telle sorte que L(9S) ne contienne pas la valeur critique.

Etant dans A(c), L(9S) admet deux preimages lisses par Qk, l'une L'(0S)
aboutissant a y\{c), l'autre L{9[) a y[(c).
Soit z € L'\ex) : 2kArgcz = 0a ± s22kG2(z).
>Arycz > 0-\ + sG'2(z) et L'(0i) est au-dessus de L{9{).
>Ar(jcz < 9-2 - eG'2(z) et L'(02) est au-dessous de L{02).
O n n o t e A2(c) l ' o u v e r t c o n t e n a n t y[(c) l i m i t e p a r L(8S), s =1,2 e t {Gc =
k
7 / }. ot A'2{c) = Q- {Ai{c) n A'{c)) limite par L{9's), L'(9S), s = 1, 2 et {Gc =
•;/'} (voir Figure 5).
On a A'2{c) C A2(c). Maintenant, Vc G A, 3A2(c) et dA2{c) n'ont que
//i(f) en commun et Qk : A'2(c) —> A2(c) est holomorphe propre de degre
deux.
(Gc = V)
L(82)
Figure 5: A'2(c) C A2(c).
Second temps
Le point ij\(c) etant repulsif, on peut rajouter un petit disque linearisable
de inaniere a construire une application a allure polynomiale gc. Cependant,
il nous faut choisir ces disques plus soigneusement pour avoir une famille
analytique.
Lemme 3.2. Pour tout c G A, il existe un secteur invariant Ai(c) issu de
yi(<). disjoint de A(c), et une application continue i : A x Ai(co) —> C telle
que, pour tout c € A, L(C, A I ( C 0 ) ) = A^c).
DEMONSTRATION. Notons p(c) = (<2*)'(2/i(c)). Pour tout c G A, nous
considerons la coordonnee linearisante de j/i(c) pour Qkc, que Ton prolonge
an plan: on a done une application holomorphe nc : C —s- C, univalente
an voisinage de 0, telle que K C (0) = ?yi(c) et Kc(p(c) • z) — Qkr o KC(Z) pour
tout .; G C. Nous normalisons KC par K'C(0) = 1. Par suite, KC depend
C-analytiquement de c.
50 Peter Haiusinsky
Not.ons F,. la eomposante connexe de K~1CR,.(0I) U TZAfh)) nontenant 0 et

qni (>st invariante par la multiplication par p(c). On transports dans ce plan
les restrictions aux rayons des applications (f>~lo(prn afin d'avoir un mouvement
holoinorphe j : A x Fru —> C. Chaque F c est un quasi-disque.
Nous voudrions prolonger j en un mouvement holomorphe global tout en
gardant la conjugaison avec les multiplications par p(c). Ceci semblc ctre
un probleme difficile si on ne veut pas retrecir A. En revanche, en utilisant
['extension barveentrique de A. Douady et C..J. Earle (voir [8]), on peut trou-
ver un prolongement j : A x C -> C continue, qui est quasiconforrne a c fixe,
et q u i s a t i s f a i t p(c) • j { c . z) = j ( c . p ( c o ) • z) p o u r t o u t (c, z ) .
On considers maintenant S(CQ) la eomposante connexe de C — r c o ne
eontenant pas au voisinage de 1'origine K~1(A(CO)). II existe A > 1 et une
a])plication conforme h : S(cQ) --» E, oil E ost un seeteur droit, tels que
h(p(cn) • z) - Xh(z). On note alors S(r) = hrl(\z\ - r) C SCo pour r > 0.
Comme les domaines linearisables de y\(c) bougent continument, il existe
une application continue c »-» r(c) > 0 tclle que KC soit, univalente sur le
secteur Sec(c) borne, issu de l'origine, et borde par r(c)Uj(c, 6(r(c))). Notons
Aj(c) = K,.(Sec(c,)). On pose
i(cz) = Kroj(c,(hoKcJ[(z • r(c)/r(co)).
Nous prolongeons / par <j>c ' o(pro sur (A(CQ) \-4•>(<;'())). On a bien A,(r) CC
A|(c). Ceci permet done de definir des domaines U'(. et Uc qui varient con-
tinument. et d'obtenir une famille analytique d'applications a allure quadra-
ti(|iie dans A (c/le cas primitif). Par le Theoreme 0.6, nous obtenons A/,o,
lieu de connexite de la famille, et une application \ : -^,0 ~^ ^ continue.
L e m m e 3.3. Pour c € Mco, on a limc_.>Cl \;(c) = 1/4.
R e m a r q u e . Par la construction de A, l'adherence dans C de M,.o est c'i.
DEMONSTRATION. On commence par montrer que x\\v = P\\\^°P\v- I' s'ensuit
que le lemme est vrai sur W, et meme sur W. On considere une suite (cn) de
M,-n de limite c\. Par un corollaire des inegalites de Yoccoz (voir e.g. [14]),
l('s x{('n) appartiennent. a des membres de diametres tendant vers zero, i.e.
ils out meme limites que les racines de ces membres. a savoir 1/4. •
Par suite, ,\ est propre et son extension \ : MCo U {<••]} —> M est un
revetement ramifie de degre fini (Theoreme 0.6). D'aprcs la remarque ci-
dessus, on peut considerer un lacet base a la racine c.[ qui fait une fois le tour
de M,.o dans A U {ci}. En adaptant la demonstration du Theoreme 0.6 dans
le cas MCu CC A, on montre que le degre de \ est egal au nombre de tours
de (c — u-v) quand c parcourt une fois ce lacet. Comme dans le cas primitif,
la foncfion c H-> C — ui,. est holomorphe et c.o est la seule racine comptee avec
Modulation dans I'ensemble de Mandelbrot 51
multiplicite. Par consequent. \ es*' u n homeomorphisme. On en deduit que

(ft:)re.\ e s t u n e famille semi-mandelbrotesque.
Bien qu'on ue puisse pas definir d'application a allure polynomiale a la
vacine c\ (comme dans le cas primitif), on a cependant la
Proposition 3.4. Le, polynome Qi/i figure de manicre rigide dans QCI.
o
DEMONSTRATION. D'apres [12], les composantes connexes de AV, f>ont des

(juasidisques. On note b\ celle qui contient c.\.
Soit if : U\ —> D une application conforme telle que < , ,'
[ ip(uCl) = 0.
Cette application conjugue Q^ a
z2 + 3
B: z2'
Comme b\ est un quasidisque, <p se prolonge quasiconforinement a C. et.

il ue reste plus qu'a montrer que Qx/A figure dans B.
Pour cela. nous procedons comme dans la demonstration de la Prop. 2.1
de[12].
Notons So un secteur invariant issu de 1/2 a l'exterieur du chou-fleur, que
Ton decoupe en domaines fondamentaux Dn, et soit To un secteur issu de 1
conespondant a SQ. On note Sn l'image reciproque de SQ issu de Q^"(l/2).
Sur line reunion de quadrilateres, on construit une application quasiconforme
(.• par pull-back. Sur les secteurs restant. on prolonge par des arcs de eereles.
Pour montrer que w est ACL (voir [1]), il suffit de considerer une suite
d'applications ((.'„.) de H-"1-" uniformement bornee qui converge uniformement
vers v. Cette suite est en fait donnee par la construction de U1. Pour montrer
qif elle est uniformement bornee, il suffit done de calculer la norme L" de dz4h
a l'exterieur du chou-fleur.
La restriction de xj> a un domaine Dn a une eccentricite
A',,, x log-
Pour rondure, montrons:
L e m m e 3.5. On a v; € II' 1 '" pour tout v e]l, 2[.
DEMONSTRATION. On note P = Q\/.\.
f C.
52 Peter Haissinskv
1/rc
z2 + 1/4
Figure 6: Estimation des distances.
Or
Notons a €]1,2[ la dimension de Hausdorff de JP et soit p g]l, a+ °_ 2 [-

Par le theoreme de Koebe,
r:t par l'inegalite de Holder, cette somme est bornee par
D'apres [18], comme pv > 1 et ^~", > a, les deux sommes sont finies. De
plus,
\
< Jaci>)
On trouve
<CC.
Par suite, l'application est dans Wl<v et Q 1/4 figure de maniere rigide dans
B.
Remarque. Reciproquement, ip l
6 W 1 ' 2 car liîlz^ x X] r) ' L(D ) J00 ',-1
J2 l / n 3 < oo .
Partie II: Operation de modulation

Soient CQ £ T>0 de periode k > 1, et W sa composante hyperbolique. On
considere \Co : MCo —> M. Si W n'est pas primitive, on prolonge \CQ par
continuite en posant Xco(ci) = 1/4-
On definit alors
c0 M ->• M c o
Figure 3. Deux ensembles de Julia de la meme copie.

Dans cette partie, nous allons analyser l'operation de modulation afin
d'obtenir des liens de caractere topologique ou/et combinatoire entre les
differents ensembles de Julia. Le theoreme principal est le suivant (voir §6
pour un enonce precis):
Theoreme 2 L 'operation de modulation agit sur les ensembles de Julia rem-
plis cornrne suit: Kco±c est homeomorphe a KCo dans lequel on a remplace les
composantes connexes de I'interieur par des "copies" de Kc.
Figure 4. Modulateur et zoom d'un polynome module.

54 Peter Haissinsky
Nous commencerons tout d'abord par etudier la combinatoire pour en ex-

traire des invariants de Mco. Avant de demontrer le theoreme, nous constru-
irons un modele abstrait de modulation de compacts. Ensuite, nous definirons
un homeomorphisme entre ce modele et l'ensemble de Julia du polynome
module. Une fois le theoreme demontre, on enoncera quelques corollaires
immediats.
Le but est de comprendre la combinatoire de l'ensemble de Julia module
a partir de celles de l'ensemble de depart et de CQ. A cette fin, nous allons
etudier des puzzles associes a la copie MCo isolant les petits ensembles de
Julia.
4 Puzzle associe a une copie MCQ

La copie Mco de M differe essentiellement de M en ce que les rayons dyadiques
qui aboutissaient de maniere unique a un point de M, ne sont plus les seuls
qui aboutissent dans la copie. Plus precisement:
Definition. Soit T>2 = {c € A/Co, 3n > 1, Qnc(c) = yi(c)}.
Get ensemble comprend exactement les parametres de Mco qui representent
les points d'aboutissement des rayons dyadiques dans M.
Commenc,ons par une proposition preliminaire qui va nous permettre de
definir les puzzles:
Proposition 4.1. II existe une suite decroissante d'ouverts connexes An, n >
1, tels que :
(i) Aoo n M = M co \ {t>2 U { C l }), ou A ^ = n A n ;
(ii) Vzo G Q^l{yi), H existe une fonction holomorphe z : An —>• C holomorphe
telle que Q'(l(z(c)) = ]ji(c) avec Z(CQ) = ZQ ; de plus, si t G Q/Z est I'angle
d'un rayon qui aboutit a ZQ, alors le rayon correspondant pour Qc aboutit a
DEMONSTRATION. NOUS allons proceder par recurrence sur n.

> D'abord, on considere Ao delimite par 72.^/(0 1 )U{CI}U7Â/(^2) 6t contenant
<-Q. D'apres la Proposition 1.2, on sait que l'application de Ao, z/i : c — i >• yi{c)
est bien definie et verifie la propriete (ii). On note Ai la composante connexe
de Qjok(A) (c/Prop. 1.1) qui contient c0.
Assertion 0-A. Soit t G JîQîVi))' 0 < * < ^; •' aôrs ^c(*) aboutit pour
tout, c, G Ao.
En efFet, pour tout m > 1, 2mt ^]î,^[- Done les rayons correspondant
aboutissent en vertu du Corollaire A.5.
Par la Proposition A.I, on en deduit la propriete (ii) pour ces points.
Assertion 0-B. Les rayons 1ZÎ{9\) efRf,i{0'2) aboutissent a un meme point,

(•, G t>>.
D'apros la Proposition 0.3, 1ZM(0\) aboutit a c^ G T>2- car B\ est strietoment
proporiodiqne. Dans h'c.,. TZC.,(O\) aboutit a r2 (voir Theoieme 0.4). Or.
c-2fcAo. done 7lr.,(0[) et 7Z(..,{B'2) aboutissent a y[(c->) = r-2: on en deduit que
7v,,,(fl.',) aboutit aussi a r_>.
Ceei nous permet de do.finir A, C An delimite par RM(8\)\J{C2}UHM(Q'.,).
c> On suppose cine Ton a un ouvcrt A,, borde par les rayons preimages de 0\
et O-j par un it.en'1 au plus fcn-ie.me. Ces rayons aboutissent a des points de
T>-2- On eonsidcre aussi .4,, la composante connexe de QrBnk(A) qui contient
A s s e r t i o n n - A . Soil t € ^ ~ ' ( Q ^ ( y i ) ) , Q 1, 2r"t n'appartient pas aux intervalles dont les rayons
sont dans A,, par construction. Done les rayons correspondant aboutissent en
veitu du Corollaire A.5.
Par la Proposition A.I, on en deduit la propriete (ii) pour ces points.
Assertion n-B. Les rayons qui aboutissent a un point, c,,î € (Q7n ilji)^
A,,) ahoutissant dans le plan des pararnetres a un mime point cn+i € T>2-
La deinonstration est la meme que 0-B.
On pent done definir A n+ i C An l'ouvert borde par ees rayons et contenant
Co-
Si <• e A-X, n .1/. eela signifie d'une part que son orbife par Q,. reste bornee.
et d'autro part que c G Q1}k(A(c)) = g~n(A(c)) pour tout n > 1. et done
c G h'dj,•) : ee tjiii implique c G MCo. •
Definitions. En suivant l'idee des puzzles de Branner-Hubbard et ^'oecoz
(voir [2]. [14j). on considere la suite de partitions definies eomine suit, pour
<• e A,c :
- on fixe // > 0 et on pose Po = U,,<n<fc(7l(.(2"(?1)u{Q<"(y1)}U^,.(2"6',))U(a =
;/). dont la partition deh'nit: le puzzle de niveau 0. note VQ : la fermeture des
coniposantes eonnexes bornees du eomplementaire sont les pieces de Po. On
definit par recurrence F,, = Q~[(Tnî), qni induit le puzzle V,, de niveau //.
()( envoie une piece de "P,, sur une piece de V,,-\.
- un bout ,r = (A"o D A"i D ...) est une suite decroissante de pieces telle q\w
\fn > 0. A",, G V,,. Qr agit sur les bouts. On note SP 1'ensenible des bouts. On
nnmit £p de la distance ultrametrique du conime suit : si ./• = (A"o D A'I 3 ...)
<'f .'/ = (i"0 ^ i'i 3 •••)- o n 110t(> ;/ = sup{n G MU {oc\. A",, = V,,} et
</„(.»•.//)- 1/2".
- on definit enfin 1'impression du bout ,r par I(.r) = OX,,, compact connexe.
On niiiiiit 1'enseinble des impressions de la metrique de Hausdorh" dH sur les
compacts non vides do C.
56 Peter Haissinsky
A tout z E Kc qui n'est pas dans la grande orbite de yi(c), on associe le

bout N(z) = (A'o D A'x D ...) tel que Vn > 0, z £ A'n. Si 2 est une preimage
de i)i, on definit N(z) comme l'ensemble des bouts dont l'impression contient
z. N est done une application multivalued surjective. Elle est continue en ce
sens que si zk tend vers 2 alors tout point d'accumulation de N(zk) appartient
a N(z). De plus, on a 2 G / o N(z).
La Proposition 4.1 entraine le resultat suivant:
- Vn > 0,
in : A n x U 0 <p< n r p (c 0 ) -»• C
(c,z) ^4>~l 0 ^ ( 2 )
est un mouvement holomorphe de point base CQ tel que /<,i(rp(co)) = Tp(c).
- ces mouvements holomorphes induisent un homeomorphisme ic : £V(co) —>
£p(c) (e'est en fait une isometrie) conjugant QCo a Qc.
Une adaptation de la demonstration du Theoreme 5.7 b) de [14] conduit
a (voir aussi [16]):
Proposition 4.2. Soit c € Mco \ T>2, alors pour tout bout x de Qc, ou Men
I(x) est un point, ou bien I(x) est homeomorphe a KX(, (cj.
Remarque. Si c G T>2, alors l'impression de tout bout est un point. D'une

part, les rayons aboutissant aux points precritiques vont morceler les petites
copies du Julia; d'autre part, comme le polynome est sous-hyperbolique, il
existe une metrique telle que Qc soit fortement dilatante, et par consequent
o
les pieces ont un diametre qui decroit exponentiellement (7^= 0).
Corollaire 4.3. Pour tout e dans Mco, l'ensemble de Julia Jc est localement
c.onnexe en tout, point qui n'appartient pas a une petite copie.
On verifie aussi que l'application I : (£p,du) —» (Comp(C),dfi) est semi-

continue superieurement et continue aux bouts dont l'impression est un sin-
gleton. Elle est de plus injective.
Corollaire 4.4. Si c € Mco \ V2, on note Sc — {z S Kc, tel que {2} soit
l'impression d'un bout) ; alors il existe un homeomorphisme de Sco sur Sc qui
conjugue Q,.a et Qc.
5 Modulation de compacts du plan

Soient K C C un compact connexe, localement connexe, plein, d'interieur non
vide et L c C compact connexe, plein. On se propose de definir KL comme
etant A" auquel on a remplace les composantes connexes de son interieur par
des copies de L (Figure 7).
L Kr
Figure 7: Modulation de compacts.
5.1 Introduction
o
D'apres [9, 7], si A' est compact, connexe, localement connexe et plein, K a
o
an plus un nombre denombrable de composantes connexes {ft}j 6 j^ = ô(K);

chacune est un domaine de Jordan, et l'intersection des fermetures de deux
d'entre elles contient au plus un point. Pour chaque £/*, on choisit un centre
.c,. On eonsidere aussi des uniformisantes <pi : D —>• £/,; telles que ip(0) = Xi
et y'(0) > 0. On appelle rayon interne de Ui tout segment ip(re2lwt), ou
t £ E / Z est fixe et r parcourt [0,1]. Ceci nous permet de definir des arcs
lvglcuumtaircs:
> On definit sur K la relation d'equivalence fermee ~ engendree par:
y si x — y ou s'il existe U GK tel que x, y G U.
Comme A' est localement connexe, pour tout e > 0, il n'y a qu'un nombre fini
de classes de diametre superieur a e. On en deduit que son espace quotient
A* est compact et metrisable (voir [3], Chap. 1). On note g : K —> K la
projection, qui est continue et propre. Les fibres sont ou bien des points, ou
o
bien des reunions fermees et connexes de fermetures de composantes de A';

en particulier, les fibres sont connexes.
R e m a r q u e . II se peut que K soit un singleton, comme dans le cas du "lapin"
de A. Douady.
> Pour tout couple (z, w) de K, il existe un unique chemin de K qui relie
leurs projections. On definit alors l'arc reglementaire ]z, W[K comme le releve
de cet arc qui joint z a w, homeomorphe a ]0,1[, qui traverse l'interieur de K
de maniere minimale le long des rayons internes.
O
o Pour chaque U € TTQ(K), on definit TTU : K —¥ U continue surjective par:

- si :;• € U, 7T(/(.r) = :;;,
T7 f )/ £ \X, XTAK , s
- Ninon, il e x i s t e u n u n i q u e y e U t q < , r . ' _ ft e t o n p o s e TTU(X) = y.
ir est localement constante sur K\U.
58 Peter Haissinsky
On definit 1'application continue
ip : K -» K x nZ7»
z - » (30), (7^(2:))).
(^ est injective: soient x ^ y,et done ]x, y[K^ 0 ;

- si ]x, y[Kn K= 0 alors g{x) / p(y);
O
- si ]x, y[Knno(K) ^ 0, alors il existe U telle que ]x, y[i<nU / 0, done nu(x) /
TTu(y)- ^ _
Par suite, on peut plonger K dans K x Ilt/j a l'aide de ip:
^(/v) = K = {(C,^) G ^ x 1 1 ^ , 6 = TTi(g-l(O) si g{Ui) + £, et sinon
3£ tq g(Uk) = £ et &_= 7r^(a;fc) pour z ^ A:}.
Remarque. Si a; G t/flV, alors on peut aussi bien choisir xu que Xy comme
centre.
Ce modele de K fournit une fac,on de definir Ki dans K x IlLj tel que les
Li soient des copies de L.
5.2 Lemme et definition

Pour chaque composante U de l'interieur de K, on considere ju : T —> dU
l'extension de % au bord, qui est un homeomorphisme.
Soient L c C u n compact connexe plein, et JL son lacet de Caratheodory,
bien defini pp. (d'apres P. Fatou).
Pour ehaque composante U, on suppose que 1'application
o 7- 1 (z)
est bien definie.
Lemme 5.1. et Definition. On pose KL = {(C,^) £ K x

C, ei sinon 3k tq g(Uk) = £ et & = ^ ( 0 ; *
A1 A'/, esi compact et on Vappelle "K module par L".
DEMONSTRATION. K X riL, etant un espace metrique compact, il suffit de

verifier que Ki est ferme par les suites: soit £™ = (t;n, (^™)i>o), n > 0, une
suite de KL convergeant vers |°° = (f°°, (^°°)j>o)-
On definit Q. :=• î>o{g{xi)}, qui correspond aux classes d'equivalence
ayant plus d'un element.
A chaque n > 0, on associe un point yn 6 A" comme suit: ou bien fn ^ fi,
et on pose yn = g^1^"); ou bien £" e fi, alors il existe kn tel que g{xkn) = £,n
et pour tout ?! 7^ fc,,, ^f = i>iTTt(xkn), et dans ce cas on pose yn = xkn.
Modulation dans l'enscmble de Mandelbrot 59
On extrait une sous-suite (y1H.) convergeant vers /73c G A'. En particulier,

i" - .'/(</») -» fj(Voc) = ' ° C -
On eonsidere les deux cas suivants:
premier cas: g(yx) £ 52. Comme g est continue, la suite yn est convergente.
Pour ehaque /. il existe un voisinage connexe de yx disjoint de V,. Par suite,
pour it assez grand, n,(yn) = ni(yx) et £,°° = v',TTt{yx.}.
Done £"• 6 A"/..
second cas: g(ijoo) € 52.
On not*1 $2(7/00) lensemble des f/, qui contiennent y^, dans leur adherence.
On choisit U 6 il(yx) vorifiant la propriete suivante si elle est satisf'aite: il
existe une sous-suite (ynj) telle que Un]ynj,yx{^ 0 et yn} —> yx.
Si I") ^ ['. alors TTi(yx) = 7r,(xr) ear y ^ € U et g(yx) — fj(-i'v)-
Soit £•", ^ ^(17,0) : pour A: 2s> 1, ynt appartient a un voisinage connexe de yx
disjoint de U, et done 7r,(?/ n J = î{yx) et on en deduit que £?° = •i:>,nl(y!X>) —
r,-,(.»•,).
Soit (.", G ^^(yoc) \ {(/} ; [)our j ^ 1, on a aussi TT,(//„;) = ^,((7^.) et done
On en deduit que A't est compact.
Remarque. A'/y est bien le modele que Ton cherehait en ce sens qu'on pout
definir les injections suivantes:
ZD : K\ (U,>of7,) KL
(voir Figure 7)
zu, : L —> A"/,
•T'G'L/
et K, - (Urc7,:(I)) U ~(A" \ (U ! > o f/ i ).
6 Theoreme 2
Enonce. 5oii Co 6 2?o \ {0} de periode k et soit c G i l / ; on note c =
Co _L r. A",,,. A',., A'c les ensembles de Julia rernplis de CQ, C, C respechvement.
Alois K; est homeomorphe a KI:OKC-
Nous allons definir un homeomorphisme en utilisant notre travail sur les
bouts. Nous reinarquons que, pour tout x € £Co, g est constante sur l{x); ceci
nous permet de definir Q : £Co —> K,o par Q(x) = g(I(x)). Cette application
est continue car dn(I(x),Xn) —> 0.
60 Peter Haissinsky
On definit alors g : A^ —>• Kcg par g(z) = QOLON(Z). Cette application est
bien definie car pour tout x G N(z), z 6 I(x) et done Q(L(I{X))) = g(i(z)).
De plus, elle est surjective et continue d'apres les proprietes de continuite de
G, i. et TV.
DEMONSTRATION DU T H E O R E M E 2. On numerate les copies de Kc dans K-c
fie maniere coherente avec les Ui.
On definit (p : A^ —> KCOKC par:
- si x i (uA'c(i)), i = g(x), et £ = ipjTj{g-lg{x));
- si x G Kc(i), S, = g{x), & = x, et pour j ^ i, ^ = ^JiTJ(xt).
Montrons que (p est un homeomorphisme.
> Continuite: il s'agit d'une adaptation de la demonstration du Lemme 5.1.
On considere une suite (xn) de Ki convergente. On lui associe une suite (yn)
de KCo en s'appuyant sur la Proposition 4.1, puis on precede au meme type
de discussion, selon la position d'un point d'accumulation de (yn).
t> Surjectivite : soit ^ = (f, ^ ) £ KCQKC ',
• si £ ^ 0, g est injective et g~l(Z,) repond a la question.
• si £ = ,g([/,), 3A: tel que g(Uk) = £ et ^ = ipiTri(xk),i ^ A;; d'ou ^ ( ^ ) = |
avec Cfc G /<f:(fc)-
> Injectivite : soient x, y tels que ip(x) = (p{y) = (£, ^,);
• si ^ ^ fi alors x, y G A'c \ ( 0 ^ ( 1 ) ) ou g est injective, d'ou x = y.
• sinon, il existe k tel que g(xk) = £, et pour tout i 7^ k, ^ = 4>iiii{xk). Du
coup, .T et ?/ appartiennent a un meme Kc{k). Par consequent, x = (p{x)k —
7 CoroUaires
Dans ce paragraphe, nous deduisons des coroUaires du Theoreme 2.
Corollaire 7.1. Soient CQ G POI c £ Do U P 2 ; on noie c = t'o J_ c, et Hco,

Hc, Hf les arbres de Hubbard de CQ, C, C respectivement. H^ s'obtient en
o
remplagant les composantes connexes de HCo par Hc en recollant par (3, j3'.
Les points marques sont alors determines comme suit:
j ^ 0, xi+i = Xi dans Hc(j + 1),

- si Xi G Hc(j) est defini,
j - 0, xi+i - Qc(xi) dans Hc(l).
Corollaire 7.2. Soient eg G T>Q, C G T>2; on note c = Co _L c ; c a des argu-
ments externes ti, ... , tv dans M. Les arguments externes de c dans M sont
les arguments dans Kco des points de dU\ d'arguments internes tj, ... , tv.
DEMONSTRATION. Soit 7^M(<) aboutissant a c, alors TZ£(t) aboutit aussi a c

(c G V2).
Or. t G 7, rl (r) — 7,^'(7t/i (7'f1 (""))) P a r le Theoreme 2. On conclut du fait que
Corollaire 7.3. Soient eQ S £>0 \ {0}, c £ Vx \ {1/4}; on note c = c.o ± c;

c a deux arguments externes t\ et t-i dans M. Les arguments externes de c
dans M sont les arguments dans KCo des points de dU\ d'arguments internes
/, et to.
DEMONSTRATION. On suppose que TZ\i{t) aboutit a c, alors 1Z,-(t) aboutit a
Or. t £ 7,-Tl(yi(c)) = 7(^l(7f'i(7^1(?7i('?))))- On conclut par le Theoreme 0.4.
< 0-> €
Definition. Soient qu'on decompose en base deux:
t eT
0i = 0.u\...u],, -i = 1,2 et
alors (6»,jy2) ± t :=
Theoreme 7.4. Soient Co 6 "Po, <" &T>\U T>2',on note c — CQ ± c; soit t un

argument externe de e dans M, alors il lui correspond un argument externe i
de c dans M, a savoir f. = (61, #2) -L t.
Lemrae 7.5. Soit z £ dU\ d'argument interne t; 7£(.o(#i),7£Co(#2) adjacents

a U\ en tjt. alors ArgCnz = (^1,^2) J- t.
La preuve du lemrne est inclusc dans la demonstration de la Proposition

1, Section 7 de [5]. Le theoreme est alors un corollaire immediat de ce qui
precede.
Remarque. Etant donne qu'en base deux, les nombres dyadiques ont deux
ecritures. on obtient deux valeurs de i. Ceci est du au fait que M s'etend
au-dela de Mco en e (presence de filament(s)).
A Aboutissement de rayons rationnels

Dans cet appendice, nous montrons que les rayons rationnels aboutissent a
des ])oints qui dependent holomorphiquement du parametre dans de "grands"
domaines du plan.
62 Peter Haissinsky
Proposition A.I. Solent 0 G Q/Z et A C C un ouvert connexe tels que,

pour tout c G A, TZC(B) aboutit; alors c \—> -)c(9) est. holomorphe et 7C(0) est
preperiodique repulsif.
S'il existe 0' G Q/Z tel que 1ZC(0') aboutit pour tout c G A et, pour un
paramet.re c G A, on a ~}c(Q) — 7r($'). alors on a egalite sur tout A.
DEMONSTRATION. On considere les applications hn(c) = d>^(e]/'n^2"e).
Cette suite (hn) converge simplement vers ^r(0). II suffit done de rnontrer
que cette suite est uniformement convcrgente sur tout compact.
Soit jj > 0; pour c € A n {GM < 77}, 0" 1 : {|z| > e7*} ->• C est conforme
et tangente a l'identite a l'infini, done ^'({1^1 > f>''}) ^ {l^l > 4.f;r'} par le
theoreme du quart de Koebe. Par consequent, pour n ^> 1, |/i n (c)| < 4,e7'.
D'apres le theoreme de Montel, (hn) est une famille normale, done sa limite.
7f:(^); est une fonction holomorphe de c.
Coinme 6 est rationnel, il existe ( > 0 et k > 1 (ininimaux) tels que
fA ' On considere l'application suivante:
p : A ->• C
C
<iui est holomorphe. De plus. Qec{yc(Q)) appartient a l'ensemble de Julia Jc,

done \p(c)\ > 1. D'apres le Lemme A.2 ci-dessous, p est ouverte, et on en
deduit <jue \p(c)\ > 1.
Par hypothese, on a 7,-(^) = Ji(O'), et ees points sont. solutions de 1'equation
F(c. z) := Qi+k(~) ~ Qi(~) = 0. Comme
f |(c,v(«) = (Q'+fc)'(
(car Ql(^',.(8)) est repulsif), ces points coincident aussi sur un voisinage de
r par l'unieite du theoreme des fonctions implicites; done sur tout A par
prolongement analytique. •
L e m m e A . 2 . Soit X = {(c,z) € C 2 , Qkc{z) = z] : l'application p : X -> C

(I.ejinie par p(c, ~) — (Qj)'(z) est ouverte.
DEMONSTRATION. A' est une courbe algebrique de C 2 , done ses composantes

sont non bornees. De plus, 1'application a : X —> A", definie par o(c,z) =
(c,Qc(z)) est periodique et propre.
Or p(c,z) = 2kIl0<j<iCo-J{c, z). Done, quand (c,z) G A' -> 00, p(c, z) aussi
et cette application n'est pas constante, done ouverte. •
Une application de la Proposition A.I est le theoreme suivant.
T h e o r e m e A . 3 . Solent () < 9X < Q2 < ••• < ^ < 1 £ Q ; on suppose que
leur rayons Tl,u{(),), i = 1,...,;/, aboutissent a un meme point c et que ce sont
les seuls. Soit A la composante connexe de C N ( 7 J A / ( # I ) U 7 2 M ( # J / ) U { C } ) qui ne

contient pas I'origine. Alors, pour tout c € A, les rayons dynamiques %c($i)
aboutissent tous au meme point £(c) et I'application c — i > £(c) est holomorphe.
Dans sa demonstration, nous utiliserons la Proposition 0.2 sous la forme:
Corollaire A.4. Soit 0 G Q/Z et soit c £ M d'argument t; si I'orbite posi-
tive de 9 ne contient pas t, alors 7Z.c{0) aboutit.
Nous aurons aussi besoin de la proposition suivante:
Proposition A.5. ([9], I, Chap.4, Prop.4) Soit CQ G T>OWD2- Ce point
est une extremite de son arbre de Hubbard.
DEMONSTRATION DU THEOREME A.3. Si c G M, comme Kc est connexe, la
grande orbite de c est dans Kc et tout rayon rationnel aboutit (Proposition
0.2). Supposons done c £ M et notons t £]0i,6u[ son argument externe.
D'apres la Proposition 0.3, ou bien c € T>2, ou bien c G V\. Distinguons ces
deux cas.
Cas parabolique (c G £>i)- On peut extraire une demonstration de ce cas
des travaux de J. Milnor (sections 2 a 5 de [17]), ou il utilise aussi la Prop. A.I
(voir Theoreme 3.1 de [17] ou Theoreme D.8 de [21]). Nous preferons presenter
une approche differente. D'apres le Theoreme 1 de [19], les rayons 7£a(#,),i =
1, ... ,f, aboutissent au point parabolique a qui attire c.
Soient 0 < 0~ < 0+ < 1 les arguments des rayons qui aboutissent a Q en
etant adjacents a la composante de Fatou qui contient c. D'apres le Theoreme
D.2 de [21], les rayons ^ M ( ^ ± ) aboutissent a c, et leurs arguments font done
partie des 0{.
On peut definir A± la composante connexe de C \ ( 7 £ M ( 0 ~ ) U { C } U 7 £ M ( 0 + ) )
qui ne contient pas I'origine. Montrons le Theoreme A.3 avec A± a la place
de A et les 9± a la place des 0,-. D'apres le Theoreme principal de [13], I
(voir aussi [9], II, Chap. 18, Prop.l, [17] Cor. 4.3 ou [21], Th.D.3.b)), c est
la racine d'une composante de centre CQ G X>O D A. Pour ces pararnetres,
les ensembles de Julia sont homeomorphes et les rayons 7?.co(0~) et 'R,Co(0+)
separent I'origine de CQ. Comme la valeur critique est une extremite de son
arbre de Hubbard (Prop. A.5), les orbites des 0* par doublement de Tangle
n'entrent jamais dans l'intervalle ]$~,9+[." Le Corollaire A.4 montre done
que ces rayons aboutissent pour tout c G A. De surcroit, comme les rayons
aboutissent a un meme point pour le centre Co, on conclut par la Proposition
A.I que les rayons aboutissent a un meme point pour tout c G A avec une
dependance holomorphe. En travaillant avec A ± , on peut montrer qu'il n'y a
pas d'autres rayons que les TIMI^) qui aboutissent a c (Cor. D.6 de [21] ou
Cor. 5.4 de [17])), et done que v = 2 et A = A±. Ceci termine la demonstration
dans le cas parabolique.
64 Peter Hai'ssinsky
Cas de Misiurewicz (c G V2)- D'apres le Theoreme 1 de [19], les rayons

aboutissent dans le plan dynamique de c a c. Ce point r etant strictement
preperiodique et Textremite de son arbre de Hubbard (Prop. A.5). on en
deduit que les orbites positives des 6( n'atteignent jamais t. Par suite, le
Corollaire A.4 s'applique et les rayons 7lc(9i), i = 1, ...v, aboutissent pour
tout c € A. Tons ces rayons sont preperiodiques et aboutissent en un point
pre-repulsif. Par le Corollaire A.4 et l'unicite du theoreme des fonetions im-
plicitos. ils aboutissent aussi dans un voisinage eonnexe A de r. Par suite,
on a 7,-(#i) = ••• = l'c(^v) pour tout e G A U A (e.onnexe) par la Proposition
A.I. •
Remarque. Dans le cas paraboliquc, on montre meme un peu plus: tous les
rayons dynamiques qui aboutissent aux points paraboliques de c continuent
a aboutir a un meme jjoint dans tout A.
References
[1] LA'. Ahlfors, Lectures on quasiconform.al mappings. Van Nostrand,
1966.
[2] B. Branner & J.H. Hubbard, The iteration of cubic polynomial, part II:
Patterns and parapatterns, Acta Math., vol. 169, pp. 229-325, 1992.
[3] R.Daverman, Decompositions of manifolds, Pure and applied mathe-

matics, 124, Academic Press, Inc., Orlando. Fla., 1986.
[4] G.David. Solutions de Vequation de Deltrami avec. ||/'||oo = 1- Ann.

Acad. Sri. Fenn. Ser. A 1 Math., 13. pp. 25-70, 1988.
[5] A. Douady, Algorithms for computing angles in the Mandelbrot set, in

Chaotic Dynamics and Fractals, Atlanta 1985, Notes Rep. Math. Sci.
Engrg., vol.2, pp. 155-168, 1986.
[6] A. Douady. Applications de la chirurgie holomorphe. Proc. Int. Congr.

Math.. Berkeley, pp. 724-738, 1986.
[7] A. Douady. Descriptions of compact sets of C, Topological methods in

modern mathematics (Stony Brook, 1991). pp.429-465, 1993.
[8] A. Douady <k C.J. Earle, Conformally natural extension of homeomor-

phisms of the circle, Acta Math., vol. 157, pp.23-48, 1986.
[9] A. Douady & J.H. Hubbard, Etude dynamique des polynomcs complexes
I & II, Pub. math. d'Orsay 84-02 et 85-05, 1984/85.
[10] A.Douady & J.H. Hubbard, On the dynamics of polynomial-like map-

pings, Ann. Sci. ENS Paris, vol. 18, pp. 287-343, 1985.
[11] P. Hai'ssinsky, Applications de la chirurgie holomorphe, notamment aux

points paraboliques, These de l'Universite de Paris-Sud, 1998.
[12] P. Hai'ssinsky, Parabolic points and [i-conformal maps, prepublication

do TENS de Lyon, 1998.
[13] P. Hai'ssinsky, Deformation J-equivalents de polynomes geometrique-

ment finis - Pincements de polynomes, prepublication de TENS de Lyon,
1998.
[14] J.H. Hubbard, Local connectivity of Julia sets and bifurcation loci: three
theorems of J. C. Yoccoz, Topological methods in modern mathematics
(Stony Brook, 1991), pp.467-511, 1993.
[15] M.Lyubich, Feigenbaum-Coullet-Tresser universality and Milnor's

h,airi,ness conjecture, preprint IHES, 1996, a paraitre.
[Hi] .1. Milnor, Local connectivity of Julia sets: expository lectures, dans ce
volume.
[17] J. Milnor, Periodic orbits, external rays and the Mandelbrot set: an
expository account, preprint Stony Brook, Asterisque, a paraitre.
[18] C.T. McMullen, Hausdorff dimension and conformal dynamics I & II,
preprint, 1997.
[19] C. Petersen & G. Ryd, Convergence of rational rays in parameter spaces,

dans ce volume.
[20] Z. Slodkowski, Holomorphic motions and polynomial hulls, Proc. AMS,

vol.111, pp. 347-355, 1991.
[21] TanL., Local properties of the Mandelbrot set at parabolic points, dans
ce volume.
Peter Hai'ssinsky, Department of Mathematics, University of Jyvaskyla,

PL 35. Fin-40351 Jyvaskyla, Finland,
e-mail: phaissin@math.jyu.fi
Local connectivity of Julia sets:
expository lectures
John Milnor
Abstract
An exposition of results of Yoccoz, Branner, Hubbard and Douady
concerning polynomial Julia sets. The contents are as follows:
1. Local connectivity of quadratic Julia sets (following Yoccoz) . 68
2. Polynomials for which all but one of the critical orbits escape
(following Branner and Hubbard) 87
3. An infinitely renormalizable non locally connected Julia set
(following Douady and Hubbard) 100
Appendix: Length-area-modulus inequalities 107
References 113
Introduction
The following notes provide an introduction to work of Branner, Hubbard,
Douady, and Yoccoz on the geometry of polynomial Julia sets. They are an
expanded version of lectures given in Stony Brook in Spring 1992.
Section 1 describes unpublished work by J.-C. Yoccoz on local connectivity
of quadratic Julia sets. (Compare [Hu3].) It presents only the "easy" part of
his theory, in the sense that it considers only non-renormalizable polynomials,
and makes no effort to describe the more difficult arguments which are needed
to deal with local connectivity in parameter space. It is based on second hand
sources (Hubbard [Hul] together with lectures by Branner and Douady), and
uses the language of Branner and Hubbard, The presentation is quite different
from that of Yoccoz. (Compare Problem 1-e at the end of §1.)
Section 2 describes the analogous arguments used by Branner and Hub-
bard [BH2] to study higher degree polynomials for which all but one of the
critical orbits escape to infinity. In this case, the associated Julia set J is
never locally connected. The basic problem is rather to decide when J is
totally disconnected. This Branner-Hubbard work came before Yoccoz, and
its technical details are not as difficult. However, in these notes their work is
presented simply as another application of the same geometric ideas.
Section 3 complements the Yoccoz results by describing a family of exam-
ples, due to Douady and Hubbard, showing that an infinitely renormalizable
quadratic polynomial may have non-locally-connected Julia set. An Appendix
describes needed tools from complex analysis, including the Grotzseh inequal-
ity.
67
68 J. Milnor
We will assume that the reader is familiar with the basic properties of Julia
sets and the Mandelbrot set. (For general background, see for example [Be],
[Br2], [CG], [Dl], [D2], [EL], [LI], or [St]. A brief survey of the Mandelbrot
set is given in §3.) In particular, we will make use of external rays for a
polynomial Julia set J(f) C C . (Compare [DH1], [DH2], [D5], [GM], [M2],
[M3].)
1. Local Connectivity of Quadratic Julia Sets

This section will prove the following.
Theorem 1.1 (Yoccoz). / / fc(z) = z2 + c is a quadratic poly-

nomial such that:
(1) the Julia set J(fc) is connected,
(2) both fixed points are repelling, and
(3) fc is not simply renormalizable,
then J(fc) is locally connected.
(For the definition of renormalizability, see 1.4, as well as §3.) In terms

of the familiar parameter space picture for the family of quadratic maps
fc{z) = z2 + c, condition (1) says that the parameter value c belongs to
the Mandelbrot set M , while (2) says that c does not belong to the closure
of the central region bounded by the cardioid, and (3) says that c does not
belong to any one of the many small copies of M which are scattered densely
around the boundary of M . (Compare Figure 1.)
Remarks: The proof gives much more, since it effectively describes the
Julia set by a new kind of symbolic dynamics. With a little more work, the
Yoccoz method can also deal with the finitely renormalizable case. Since the
case of a map with attracting or parabolic fixed point had been understood
much earlier [DH2], we see that conditions (2) and (3) can be actually replaced
by the following weaker pair of conditions:
(2;) / has no irrationally indifferent periodic points (Cremer or Siegel
points), and
(3') / is not infinitely renormalizable.
These modified conditions (1), (2') and (3') are all essential. In fact:
Condition (1): For c £ M the Julia set J(fc) is a Cantor set, which
is certainly not locally connected.
Condition (2'): Sullivan and Douady showed that a polynomial Julia
set with a Cremer point is never locally connected. (Compare [Su], [Dl],
Local connectivity of Julia sets 69
Figure 1. Boundary of the Mandelbrot set M , and a detail of the

one-third limb showing several small copies of M . (Compare §3.)
[M2]. For a more explicit description of non local connectivity, see [S01-2].)
Similarly, Herman has constructed quadratic polynomials with a Siegel disk
having no critical point on its boundary. The corresponding Julia set cannot
be locally connected. (See [He,§17.1], [D1,II,5], [D4].)
Condition (3'): In §3, following unpublished work by Douady and Hub-
bard, we will describe infinitely renormalizable polynomials for which J is
not locally connected.
Thus the sharpened version of Theorem 1.1 comes fairly close to deciding
exactly which quadratic polynomial Julia sets are locally connected. Yoc-
coz has also proved a corresponding result in parameter space: For c in
the Mandelbrot set M , if the polynomial fc(z) — z2 + c is not infinitely
renormalizable, then M is locally connected at c . For proofs of this more
difficult result, see [Hu3], [K].
The Yoccoz puzzle. The proof of Theorem 1.1 begins as follows. If a
connected quadratic Julia set has two repelling fixed points, then one fixed
point (to be called /?) is the landing point of the zero external ray, and the
other (called a) is the landing point of a cycle of q external rays, where
q > 2 . (Compare [M2], [Pel].) Let 0 = Co >-> cx >->••• be the critical
orbit. The Yoccoz puzzle of depth zero consists of q non-overlapping closed
topological disks Po(co), P0{c\), • • • , Po{cq-i) , called "puzzle pieces", which
are obtained by cutting the region G < 1 along the q external rays landing
at a . Here G is the Green's function (or canonical potential function)
which vanishes only on the filled Julia set, and satisfies G(/(.z)) = 2G{z) ;
while Po(ci) denotes the unique puzzle piece of depth zero which contains
70 J. Milnor
2/7
^-I
1/7
yvi
1/14 \
\ \
4/7 \ \ /'""I
yJ
/ /
\ 9/liV Ic ) \P1(-C2)
___Ju—-^
\ P 0 ( 0 ) 11/14
Figure 2. Julia set for f(z) = z2 + i, showing the Yoccoz puzzles

of depth zero and one. Here q = Z. The 1/7, 2/7 and 4/7 rays
land at the fixed point a .
the post-critical point c, = / o t (0) .

Inductive Construction: If P d , . . . , PJ are the puzzle pieces of
depth d , then the connected components of the sets / - 1 (PJ; ) are the puzzle
pieces P^x of depth d + 1 . As an example, there are always 2q — 1
pieces of depth one. These consist of q pieces Pi(c 0 ), -Pi(ci), . . . , Pi(cq_i)
which touch at the fixed point a , together with q — 1 additional pieces
•Pi(-ci), . . . , Pi(—Cq-\) which touch at the pre-image —a .
Since the a-fixed point is used to construct this puzzle, the points in its
grand orbit will have to be dealt with by a special argument. (See 1.9.) It
will be convenient to introduce the abbreviated notation
Jo = (J/~»
for the set of all points in the Julia set which belong to this excluded grand
orbit. For any z € J \ «70 , it is easy to see that there is a unique sequence
of puzzle pieces
Po(z) D P i ( z ) D P2(z) D •••
which contain z .
De
Main Problem. Let z 6 J \ Jo any point which belongs to the Julia
set ,/(/) but not to the countable set \J f~n{a) C J{f) , and let Pd(z) be
the unique puzzle piece of depth d which contains this point z . Does the
intersection f\dPd{z) of these puzzle pieces consist of the single point zl
The associated annuli. Following Branner and Hubbard, consider the
puzzle pieces Pd{z) D Pd+\{z) of two consecutive depths containing some
given z € J \ Jo . With luck, the smaller puzzle piece Pd+i(z) will be
contained in the interior of Pd{z) . In this case the difference set
Ad{z) = Interior(P d (z)) \ Pd+l(z)
will be an annulus, whose modulus mod Ad(z) is a well defined positive
real number. (Compare the Appendix.) However, we must also allow for
the possibility that the boundary of Pd(z) intersects Pd+\ . In that case,
we describe Ad(z) as a "degenerate annulus", and define modAd(z) to be
zero.1 As an example, in Figure 2 the annulus AQ(-Ci) has positive modulus,
but the annulus ^4o(0) around the critical point is degenerate.
With these definitions we can reformulate the Main Problem as follows:
Modified Main Problem (Branner and Hubbard). Given a point
z 6 J \ Jo , is the sum
y^j mod Ad(z)
d
infinite? If so, using the Grotzsch inequality, it is not difficult to prove that
the intersection f] Pd{z) must consist of the single point z . (See A.7 in the
Appendix.)
The sequence of puzzle pieces
Po(0) D Fi(0) D P2(0) D •••
containing the critical point will play a special role. For this sequence to be
well defined, we need the following.
Standing Hypothesis 1.2. We assume that 0 ^ Jo , so that

the orbit of the critical point does not eventually land on the fixed
point a .
This is no real restriction. For in the contrary case where 0 e JQ so that

the critical orbit does end at a , we are in the post-critically finite case, and
local connectivity can be established by more classical methods. (Compare
[DH2] or [M2].)
'It will be convenient to refer to Ad(z) loosely as an "annulus" even when it may be
degenerate, and hence consist of one or more simply connected regions. In fact it might be
clearer to work with the closed ring Ad{z) = Pd{z) \ Interior(Pd+i (z)) which has interior
Ad(z) , and which has the homotopy type of a circle even in the degenerate case.
72 J. Milnoi
We can compare the sequence {Pd(z)} for an arbitrary point z € J \ Jo

with the sequence {P<j(0)} as follows. Define the semi-critical depth S(z) to
fie t/ie largest integer d > 0 /or w/iic/i Pd(;z) = Pd(0) , taking S(z) to be
+00 i/ P d (z) = Pd{0) for all d, and setting S(z) = - 1 if Pd(z) ^ Pd(0)
for all d . Thus S(z) is large whenever z is close to the critical point.
To each point z0 6 J \ Jo we can associate the orbit ZQ •->• z\ i-> z2 •->
• • •, and the sequence of numbers S(zo), S(zi), S(z 2 ), • • • • (The sequence
associated with the critical orbit 0 = CQ I—> c\ H-> • • • will be of particular
interest.) Following Branner and Hubbard, this sequence can be displayed
graphically by introducing the associated tableau. (Compare Figures 4, 6, 9.)
Definition. The tableau associated with an orbit ZQ I->- Z\ H-> Z^ \-± • • •

in Jo is an array with one column associated with each Zj and one row
associated with each depth in the Yoccoz puzzle. We will draw a solid vertical
line at depth d in the j-th column whenever d < S(z;) , and a double vertical
line when d = S(ZJ) . Such single or double vertical line segments correspond
to puzzle pieces Pd{zi) which coincide with the critical puzzle piece Pd(0) ,
and a long chain of such vertical segments indicates an orbit point Zj which
is very close to the critical point. Diagonal arrows of length m, pointing
north-east, correspond to iterates f°m , mapping some puzzle piece Pd{zi)
by a branched covering map onto Pd-m{zi+m) .
Given any orbit z0 )->• zx i-> z2 •-> • • • in J o , note that / carries any
puzzle piece Pd(zi) of depth d > 0 onto the puzzle piece Pdî{zi+{) . This
map Pd(zi) —> Pd-i(zi+i) is either a conformal isomorphism or a two-fold
branched covering according as Pd{zi) does or does not contain the critical
point. Now consider the restriction of / to the annulus
Ad{Zi) = Intenoi(Pd{Zi)) \ Pd+1(Zi) .
We can distinguish three cases, as follows.
Critical Case. If d < S(ZJ) so that 0 e Pd+i(zi) (indicated by a

vertical line segment in the tableau), then Ad(zi) coincides with Ad(0) , and
will be called a critical annulus. In the non-degenerate case, note that Ad(zi)
surrounds the critical point 0, and that / maps the annulus Ad(zi) onto
Ad-i(zi+i) by a 2-fold covering map. It follows easily that
modAd(zi) = ± modAd-i(zi+i) .
Off-Critical Case. If d > Sfe) so that 0 g Pd{zi) (indicated by a blank
in the tableau), then / maps Ad(zi) biholomorphically onto ^4d_i(2;i+1) and
the moduli are equal. In this case, we will say that the annulus Ad(zi) is
off-critical.
Figure 3. Critical, semi-critical and off-critical annuli (left); and a

"child" (right).
Semi-Critical Case. Finally, if d — S(zi) so that 0 e Ad(zi) (corre-

sponding to a double line in the tableau), we will say that Ad(zi) is semi-
critical. Here the topology is more complicated, since / maps the annulus
Ad(zi) onto the entire puzzle piece / 3 d _ 1 (z i+1 ) . However, it is easy to check
that Ad(zi) is non-degenerate if and only if the annulus Ad_i(zi+i) is non-
degenerate. For an inequality between moduli, see Problem 1-b.
As examples, in the schematic picture Figure 3 (left), the annulus Ao{z)
is critical, while ^1(2) is semi-critical, and ^2(2) is off-critical. In Figure
2, the annuli Ao{-c\) and i4 0 (-c 2 ) are both semi-critical.
Definition. The critical annulus Ad+k{0) in the Yoccoz puzzle will be
called a child of the critical annulus Ad(0) if and only if Ad+k(Q) is an
unramified two-fold covering of Ad(0) under the map fok . (See Figure 3
(right), and Figure 4. Our terminology is based on [Hul], but with several
modifications.)
Note that the modulus of such a child is always exactly half the modulus
of the parent. In a tableau, starting with any critical annulus, represented by
a single vertical segment, we can locate the parent (necessarily unique if it
exists), by drawing a line to the north-east and extending until it hits another
vertical segment. Our strategy for solving the Modified Main Problem can
now be summarized as follows:
Find a critical annulus of positive modulus, and prove that it has so

many descendents, children and grandchildren and so on, that the
modulus sum is infinite.
Branner and Hubbard summarize the tableau properties which are needed
to carry out this argument in three rules, as follows.
74 J. Milnor
Z
Z
O z
l Z
2 ZZ i
3 ZT.A
4 Z Zr
5 Z Zr
6 ' 7Z-7 Zfi
12 Z 13 Z 14
A2
A3
A4
A?
A
8
A9
A
10
All'
Figure 4. An example: the tableau associated with the orbit of the

point z0 = 1 for the map f(z)=z2 — 1.6. Here we see by following
the top diagonal arrow that the critical annulus ^3(2:3) = ^ ( O )
is an unramified two-fold covering of the critical annulus A0(z6) =
A0{0) . In particular, the critical annulus at depth 3 is a child of the
critical annulus at depth 0 . Similarly, the critical annuli at depths
4 , 6, 8, 10 are children of the critical annulus of depth 1 .
First tableau rule: Every column of a tableau is either all critical, or all
off-critical, or has exactly one semi-critical depth S(z;) , and is critical above
and off-critical below.
Now let us compare the tableau of the critical orbit
0 = Co i—> c\ \—> c-i \—>•
with the tableau of some given orbit z0 i-> z\ i-> • • • in Jo . (The case
zQ = c0 is not excluded.) If the tableau of {ZJ} is critical or semi-critical
at depth d in column m , draw a line "north-east" from this critical or
semi-critical annulus, as indicated by the dotted line on the right, and draw
a corresponding line north-east from depth d of column zero in the critical
tableau.
Second tableau rule: Everything strictly above the diagonal line on the left
must be copied above the diagonal line on the right.
m m+k
anything copied
d-k d-k
here here .1
Figure 5. Illustration for the second and third tableau rules, with the
critical tableau on the left, and the tableau for z0 t-» z\ i-> • • • on
the right.
The proofs of these two rules are easily supplied. D

Now suppose that the critical annulus of depth d is a child of the critical
annulus of depth d — k , as indicated in the figure, and suppose that the
tableau of {ZJ} is semi-critical at depth d of column m .
Third tableau rule: Following the diagonal arrow from this semi-critical
annulus of depth d in the tableau of {ZJ} , we must reach a semi-critical
annulus at depth d — k , as illustrated.
Proof. According to the hypothesis, fok maps Ad(0) onto Ad_fc(0) ,
where the point zm is an element of this annulus Ad(Q). Therefore f°k{zm) =
Zm+k must belong to Ad-k(0) . O
Definition. We will say that a critical annulus Ad(0) is excellent if it
contains no post-critical points, or equivalently if there is no semi-critical
annulus in the d-th row of the critical tableau.
In Figure 6, note that the critical annuli at depths 1 , 3 , 4 , 6 , 7 , 8 , 9 , . . .
are excellent. Each one'has exactly two children, which are also excellent. On
the other hand, 0, 2, 5, 10, 18 (Fibonacci numbers minus three) are not
excellent. Each of these has only one child; however this child is excellent.
Definition: We will say that the critical tableau is recurrent if the num-
bers S(cjt) with k > 0 are unbounded, so that there are columns which go
arbitrarily far down. It is periodic if S(cfc) = oo for some k > 0 , so that the
76 J. Mi/nor
0 1 2 3 4 5 6 7 9 10 11 12 13 14 15
:i
n
n n
71
Figure 6. The "Fibonacci" critical tableau. (Compare [BH2].) Here

the closest recurrences of the critical orbit come after a Fibonacci
number of iterations:
1, 2 , 3 , 5 , 8 , 13, 2 1 , 34, 5 5 , ....
If un is the n-th Fibonacci number, then the semi-critical annulus

for column number un occurs at depth un+i - 3 . The diagonal
dotted lines illustrate the genealogies
0 <- 3 <- 8 ^ "•
\
with arrows from child to parent. Since A3(0) is the only child of
A0(0) , it follows from 1.3 that every descendent of A3{Q) has at
least two children.
16 5 80
Figure 7. Graph of the unimodal map x \-t x2 -1.8705286 • • • which realizes the
Fibonacci tableau. (Compare [ML].) The first eight points on the critical orbit
are marked. The critical orbit closure is a rather thin Cantor set, which is plotted
underneath the graph. (Compare Problem 1-f.)
Figure 8. The Yoccoz puzzle at depths zero through five for this quadratic
Fibonacci map, drawn to the same scale. Note that the critical pieces are the
biggest ones at any specified depth.
78 J. Milnor
2k 3k
d+k
Figure 9. Finding Children. The successive dotted lines illustrate

Lemma 1.3 (a), (b) and (d).
corresponding column goes all the way down. In this case, it follows easily
that Pd{ci) = Pd(ci+k) for all d and i. (Compare Lemma 1.5 below.)
Lemma 1.3. / / the critical tableau is recurrent but not periodic,

then:
(a) Every critical annulus has at least one child.
(b) Every excellent critical annulus has at least two children.
(c) Every child of an excellent parent is excellent.
(d) Every only child is excellent.
Proof of (a). Start at depth d of column zero in the critical tableau,

and march to the right until we first meet another critical annulus ( = solid
vertical line), say at column k . (Figure 9.) Now marching diagonally south-
west, we find the first child at depth d + k .
Proof of (b). Suppose that the annulus Ad(0) is excellent. Choose k as
in (a). By hypothesis, the fc-th column cannot be critical all the way down.
Hence it must be semi-critical at some depth d! > d . Starting at column k
and depth a" , proceed diagonally right (north-east). By the tableau rules,
column 2k must be semi-critical at depth d' — k . Similarly column 3fc
must be semi-critical at depth d' — 2k , and so on, until we again reach depth
d . Furthermore, as we follow this diagonal, we do not meet any other critical

or semi-critical annuli. In particular, at the point where we reach depth d,
there cannot be a critical or semi-critical annulus. (The hypothesis that d is
excellent comes in at this point. Compare Figure 9.) Now let us again march
to the right at depth d until we reach a critical annulus, say at column rn .
Again turning 135° and proceeding diagonally south-west, we cannot meet
any critical or semi-critical annulus until we are back at column 0 . In this
way we prove that the critical annulus of depth m + d is also a child of d .
Proof of (c). This is clear, since if f(Ad+k(0)) = Ad(0) where Ad+k{0)
contains a post-critical point, then Ad(0) does also.
Proof of (d), by contradiction. Consider a child d' which is not
excellent, and let dt = n' - k be the parent. Then for some k' > k the
&'-th column has semi-critical annulus at depth d! . (The case k' — k is
illustrated in Figure 9.) Following the diagonal up from column k' and
depth d! , we must meet a semi-critical annulus at depth dx by the third
tableau rule. (Thus the parent is not excellent.) Now proceed to the right at
depth di until we meet a critical annulus, say at column m . Then it follows
as above that d\ + m is a second child; hence d' was not an only child. •
Definition 1.4. (Compare [DH3], [Mc2].) A quadratic polynomial / is
renormalizable if there exists a closed topological disk A o around the critical
point, and an integer p > 2 (called the renormalization period), with the
following four properties:
(1) A o should be centrally symmetric about the critical point so

that / carries A o onto the image Ai = /(A o ) by a ramified
2-fold covering map.
(2) / should induce conformal isomorphisms
A, -=->• A 2 - ^ • • • -=-» A p ,
C
where A^ = / '(A O ) . In particular, the critical point should be
disjoint from Ai U • • • U A p _i .
(3) However, the disk A p should contain A o in its interior.
(4) Finally, the entire orbit of the critical point under the map
fop should be contained in the original disk A o .
(Compare the discussion of "polynomial-like mappings" in §2, and the dis-

cussion of "tuning" in §3.) If these conditions are satisfied, then the small
filled Julia set Kp = K(f°p\Ao) can be defined as the compact connected
set consisting of points whose orbits remain in Ao under all iterations of
f°p . This is a proper subset of the filled Julia set K(f) of the original map
80 J. Milnor
Figure 10. An example of simple renormalization: the Julia set for

f(z) — z2 — 1.75 , with renormalization period p = 3 . (The pa-
rameter value c = -1.75 is the root point of a small copy of the
Mandelbrot set.) The filled Julia set K(f°s\Ao) is a topological disk
bounded by a "cauliflower".
f . Following McMullen [Mc2], the map / is called simply renormalizable of

period p if also:
(5) Each intersection foi{Kp) n foj(Kp) with 0 < i < j < p

is either vacuous, or consists of a single point which does not
disconnect either of these two sets.
We will be interested only in the simply renormalizable case. See Figure

10 for a simply renormalizable example, and Figure 11 for a map which is
renormalizable but not simply renormalizable. (Compare [RS].)
We can now state two basic lemmas.
L e m m a 1.5. If the critical tableau associated with f is periodic,

then f is simply renormalizable. More precisely, if Pd(cp) =
•fd(O) for all depths d , with p minimal, and therefore P<i(ci+p) =
Pd(ci) for all i and d by the Second Tableau Rule, then f is
simply renormalizable of period p .
L e m m a 1.6. If the critical orbit lies completely within the union

) u Pda) u ••• u Pifo-i)
of those puzzle pieces of depth one which touch the fixed point a ,
then f is simply renormalizable of period q . (See Figure 2.)
The following construction will be needed to prove Lemmas 1.5 and 1.6.
Recall that each of the puzzle pieces P 0 (CJ) of depth zero consists of points in
,2
F/gure I I . An example of crossed renormalization: The Julia set
for f(z) = z2 + .41964338 + .60629073 i is shown on the left,
with a schematic diagram of its renormalization on the right. The
associated small Julia set K(fo2\&0) is a simple arc from c2 to c4 .
Our discussion will always exclude such examples.
Figure 12 Left: A Puzzle Piece and Thickened Piece of Level 0

Right: Thickened Pieces of Level 0 and 1 for z z2 +
82 J. Milnor
some closed subset of the filled Julia set K(f) , together with points outside
of K{f) which have potential G and external angle t satisfying inequalities
of the form
0 < G < 1, U < t < t'i .
(Here 0 0 so small that every external ray
whose angle differs from tt or t't by at most rj intersects this disk. Now let
P0(ci) consist of the disk De(a) , together with the region bounded by:
(1) the segment U — rj < t < t't + rj of the equipotential G — 1 ,
(2) the external ray segments of angle ^ — r\ and tfi + rj which extend from
this equipotential G = 1 to their first intersection with De(a) , and
(3) an arc of the boundary of De(a) .
Evidently this thickened puzzle piece PO(CJ) contains the original Po(ci) .
We now construct thickened puzzle pieces of greater depth by the usual in-
ductive procedure: If PJ is a thickened puzzle piece of depth d , then each
component of / ^ ( P j j ) is a thickened puzzle piece of depth d + 1 .
The virtue of these thickened pieces is the following statement, which is
easily proved by induction: / / a puzzle piece p£* contains PJ*^ , then the
corresponding thickened piece P^ contains PJ +1 in its interior. In other
words, this construction replaces all of our annuli by non-degenerate annuli.
We will only make use of thickened pieces which are small enough to
satisfy the following additional restriction: If Pd(z) contains the critical
point, then the interior of Pd(z) must already contain this critical point. If
1.2 is satisfied, then this requirement is easily satisfied for any bounded values
of the depth d , and this will suffice for the applications (Lemmas 1.5 through
1.8). Note however that we cannot expect this condition to be satisfied for
unbounded values of d . In fact this would be impossible whenever the fixed
point a is an accumulation point for the critical orbit.
Proof of Lemma 1.5. Choose d > p so large that the critical an-
nulus Ad(0) is a child of Ad_p(0) , and let A o = Pd(0) be the thick-
ened critical puzzle piece of depth d. Then A p = /° P (A O ) is equal to
Pd-p(cp) = Pd-p{0) , which contains AO in its interior. The hypothesis guar-
antees that the successive points Cp, c 2p , c 3p , . . . all belong to A o . Since
p is minimal, the various puzzle pieces Pd_j(cj) with 0 < i < p cannot over-
lap, so any intersection point must belong to the boundary of these puzzle
pieces, and therefore cannot disconnect the small filled Julia set. Therefore
/ is simply renormalizable. •
Proof of Lemma 1.6. (suggested by M. Lyubich). Let Ao = Pi(0) .
Then it is easy to check that the successive images A, = / o i (A 0 ) are disjoint
from the critical point for 0 < i < q , and that Aq contains Ao in its
interior. (Compare Figure 2.) We will prove inductively that cqi € Px(0) C
A o for every i. It certainly follows from this inductive hypothesis that
Cqi+i 6 P 0 (ci) n K(f) c Pi(ci) , and similarly that cqi+j e P0{CJ) H K(f) c
Pi(cj) for 0 < j < q. In particular, c gi+9 _ 1 must belong to Pi(cg_i) ,
hence c, i + , 6 /(Pi(c,_i)) = P0(0) . By the hypothesis of 1.6, the point
cqi+q is known to lie in one of the puzzle pieces Pj (ch) which lie around a .
Evidently it can only lie in P x (0) , as required. •
Corollary 1.7. If f is not simply renormalizable, then there

exists a critical annulus of positive modulus.
Proof. According to 1.6, the critical orbit must visit one of the puzzle
pieces Pi(-ci) , . . . , Pi(—c?_i) which surround the pre fixed point —a .
Suppose for example that cd € Pi(—Cj) . It is easy to check that the corre-
sponding annulus A0(c,i) — Ao(-Ci) has positive modulus. (Compare Figure
2.) Pulling this annulus back inductively along the critical orbit, it follows
that Aj(0) also has positive modulus. •
As another another application of thickened puzzle pieces, we can study
orbits which stay away from the critical point.
L e m m a 1.8. Let z0 H-» ZX I-> • • • be an orbit in J \ Jo which

is disjoint from some critical puzzle piece P/v(0) . Then the in-
tersection P) Pd(zo) of the puzzle pieces containing z® reduces to
the single point ZQ .
Proof. Let Ui = Interior(P^_ 1 ) , where the various puzzle pieces of depth
N — 1 have been numbered so that the critical value cx belongs to Pff°\ .
Each puzzle piece of depth > N is contained in some unique P^_! , and we
give it the Poincare metric distj(a;, y) associated with the corresponding Ui.
For i > 0 , note that there are exactly two branches of f~l on Ui, call them
<7i and g2 , and that each of these maps gK on Ui is univalent and carries
Ui onto a proper subset of some Uj . Thus each gK is Poincare distance
decreasing. Since every puzzle piece of depth N is a compact subset of the
associated Ut (we need the thickening at this point), it follows that
<listj(gK(x), gK(y)) < Xdisti(x,y)
throughout any such puzzle piece, provided that i > 0 , where A < 1 is a
uniform constant. Let 6 be the maximum of the Poincare diameters of the
puzzle pieces of depth iV . Then for any h > 0 , since none of the puzzle
pieces
-> PN+h-l(zi) - > • • • - > • P/V+iOzh-l) -)• Pff(zh)
84 J. Milnor
(except possibly the first) can be contained in the critical value puzzle piece
P/!jO}l , it follows that the Poincare diameter of PN+h(z0) is at most XhS ,
which tends to zero as h -> oo . D
In order to deal with the possibility that f°d{zo) = a , so that the puzzle
piece Pd{z0) is not uniquely defined, we will need the following.
Definition. For any point z in the Julia set let Pd(z) be the union of
the puzzle pieces of depth d containing z. (In most cases, Pd{z) is equal
to the unique depth d puzzle piece Pd{z) which contains z . However, if
fod(z) = a then P^(z) is a union of q distinct puzzle pieces.)
We can now state and prove the principal result of this section.
Theorem 1.9. Suppose as usual that f is quadratic, with con-

nected Julia set, with both fixed points repelling, and not sim-
ply renorrnalizable. Suppose further that the critical orbit is dis-
joint from the fixed point a , Then for any z € J(f) we have
Proof. Since we assume that / is not simply renormalizable, we know

from Corollary 1.7 that there exists some critical annulus Am(0) which has
positive modulus. We will first prove that f\Pd(0) — {0} , then prove that
f] Pd(z) = {z} for any z 6 J(f) which is not an iterated pre-image of a ,
and finally prove the corresponding result when fon(z) = a .
Critically Recurrent Case. Suppose that the critical orbit is recurrent,
so that 1.3 applies. First consider the puzzle pieces Pd(0) around the critical
point. If the non-degenerate annulus Am(0) has at least 2k descendents
in the k-th generation for each k , then each of these contributes exactly
mod>l m (0)/2* : to the sum ^2dmodAd(0) . Hence this sum is infinite, as
required. On the other hand, if there are fewer descendents in some genera-
tion, then one of them must be an only child, hence excellent by Lemma 1.3d.
Using Lemma 1.3b and 1.3c, we again see that J2dmod Ad(0) is infinite.
Therefore in either case the intersection p| Pd(0) reduces to the single point
0.
Now consider a point z0 ^ 0 of the Julia set. We assume that the orbit
Zo >-> Z\ —
i > • • • is disjoint from a, so that the puzzle pieces Pd{zo) are
well defined. If the orbit of z0 does not accumulate at zero, then we have
f)Pd(zo) = {z0} by Lemma 1.8. Suppose, on the other hand, that the origin
is an accumulation point of {zn} . In other words, suppose that the tableau of
the point z^ has critical annuli reaching down to all depths. For each depth
d, let us start at column zero (corresponding to the point ZQ itself) and
advance to the right until we first hit a critical annulus at column n , then
proceed diagonally back down until we reach column zero at depth n + d.
It follows from this construction that the annulus An+d(z0) is conformally

isomorphic to ^4^(0) . Furthermore, distinct values of d must correspond to
distinct values of n + d. Thus the sum ^modA^zo) is also infinite,
hence f)Pd(zo) = {zo} , as required.
Critically Non-recurrent Case. Now suppose that the critical orbit is
not recurrent. Then the critical value /(0) satisfies the hypothesis of Lemma
1.8. Hence f]Pd(f(0)) = {/(0)} , and it follows easily that f l ^ 0 ) = {0} .
Next consider a point z0 ^ 0 . Again we may assume that the orbit {z,j}
accumulates at zero, since otherwise the conclusion would follow from Lemma
1.8. Again 1.7 tells us that there exists one depth m such that Am(0) has
positive modulus. The corresponding depth m for the tableau of ZQ must
have infinitely many columns k which are critical. For each of these, let
us proceed diagonally back down in the tableau of ZQ , ignoring whatever
critical or semi-critical annuli we may meet, until we reach column zero at
depth m + k . Each time we meet a critical or semi-critical annulus, we lose
up to half of the modulus. (Problem 1-b below.) However, the hypothesis
that the critical orbit is non-recurrent guarantees that such losses will only
occur a bounded number of times. Thus ^modAd(zo) has infinitely many
summands which are bounded away from zero. Hence this sum is infinite, and
we have proved that f)Pd(zo) = {z0} in this case also.
Iterated Pre-images of a. If some forward image zn = / on (zo) is
equal to the fixed point a , then the above arguments do not make sense
as stated. In this case, there are q distinct puzzle pieces P$ of depth n
which have z0 as common boundary point. Each of these is contained in
a unique sequence of nested puzzle pieces Pn D P^'+i D • • • which have
z0 as common boundary point. Assertion: For each one of these q nested
sequences the intersection f]d Py reduces to the single point z0 . In fact,
the proof of Lemma 1.8 applies equally well to this situation. Evidently the
statement that fld^dlô) = {ô} follows immediately.
Thus we have proved Theorem 1.9, r\d^diz) = iz} i m a ^ cases. The-
orem 1.1, as stated at the beginning of this section, is a straightforward
consequence. (Compare Problem 1-a below.) •
For further developments on the puzzle technique, see for example [F],
[L2], [L3], [L4], [Pe3].
Here are some problems for the reader.

Problem 1-a. Local Connectivity. Prove that the intersection of J(f)
with each puzzle piece is connected. Conclude that J(f) is locally connected
at z whenever f]Pd{z) = {z} .
Problem 1-b. Semi-Critical Annuli. Show that a semi-critical annulus
Ad(zi) of depth d > 0 is non-degenerate if and only if Ad_i(zi+i) is non-
86 J. Milnor
degenerate. In the non-degenerate case, show that Ad(zi) is the union of

(1) a ramified two-fold covering of Ad-i{zi+i) , and
(2) a conformal copy of Pd{zi+i) .
Splitting Ad-i{zi+i) by a carefully chosen simple closed curve through the
critical value, use the Grotzsch inequality to prove that m o d Ad(zi) >
Problem 1-c. Non-degenerate Annuli. Show that an annulus Ao (z)

of depth zero is non-degenerate if and only if it is semi-critical, and show that
Ad(zm) is non-degenerate if and only if Ao{zd+m) is non-degenerate.
Problem 1-d. Illustrating the Tableau Rules. Suppose that we try
to modify the Fibonacci tableau of Figure 6 by changing just one column.
For example, suppose that we place the semi-critical annulus for column 5
at depth 3 or 4 or at depth > 6 , instead of at depth 5 , without changing
columns 0 through 4. Show that the tableau rules would then force column
8 to end already at the semi-critical depth 0 or 1 or 2 respectively.
Problem 1-e. The Yoccoz r-function. For each depth d > 0 ,
consider the sequence of puzzle pieces f°l(Pd(0)) = Pd~i(ci) with 0 < i < d .
If one of these pieces is equal to the critical puzzle piece Pd_;(0) , then choose
the smallest such i and set r(d) = d — i. Otherwise, if there is no such
i, set r(d) = - 1 . Thus - 1 < r{d) < d for all d > 0 . Show that
r(d + 1) < r{d) + 1 in all cases, and show that the annulus Ad{0) is a
child of ^ n (0) if and only if n = r(d) with r(d + 1) = -r(d) + 1 > 0 .
Problem 1-f. Persistent Recurrence. Show that the critical tableau is
recurrent if and only if the function r(d) is unbounded, and that it is periodic
if and only if r ( d + l ) = r(d)+l for large d . By definition, the critical orbit is
persistently recurrent if the tableau is non-periodic with limôo r(d) = +oo .
Show that the Fibonacci tableau is persistently recurrent. In the persistently
recurrent case, show that the critical orbit is bounded away from the fixed
point /3 . (Otherwise, for any depth d we could choose a post-critical point
c-n G Pd{P) , then choose the smallest k with 0 € Pd+k{cn-k) , and conclude
that r(d+ k) = 0 .) Show that the critical orbit is bounded away from every
iterated pre-image of /3, and hence that its closure is nowhere dense in the
Julia set. Similarly show that the critical orbit is bounded away from the
fixed point a ; and use this fact to show that the critical orbit closure is a
Cantor set. More generally, show that the critical orbit is bounded away from
any periodic or pre-periodic point. (For further information, see [L2].)
Problem 1-g. The Critical Orbit is Generically Dense. It is con-
venient to say that a property of certain points in a compact set is generically
true if it is true throughout a countable intersection of dense relatively open
subsets. For a generic point c 6 dM in the boundary of the Mandelbrot set,
show that the critical orbit for the map fc is everywhere dense in the Julia
set. (Proof outline, suggested by McMullen and by J. Kwapisz: Let M(p,e)

be the set of c 6 dM such that the critical orbit for fc comes within e of
every periodic orbit of period p . Using Montel's Theorem, show that this
set is dense for p > 2 . It is clearly open. Therefore a generic c belongs
to the intersection of the M(p, e) , which implies that the critical orbit is
dense in J .) As one corollary, it follows that a generic fc is not persistently
recurrent, and also there are no excellent annuli in the generic case. If fc
is renormalizable of period k , show that the union of the small Julia sets
f°'(Jk) is a proper subset of the Julia set J(fc) • Conclude that fc is not
renormalizable in the generic case.
2. Polynomials For Which All But One

of the Critical Orbits Escape
This section will describe easier applications of ideas from §1. Its object
is to describe the Branner-Hubbard theory of polynomials for which only one
critical point has bounded orbit. However, we begin with the simpler case
where no critical point has bounded orbit. The following is well known.
Theorem 2.1. Let f be a polynomial of degree n>2 for which

all critical orbits escape to infinity. Then the filled Julia set K
(in this case equal to the Julia set J — dK ) is a Cantor set of
measure zero2. Furthermore, the dynamical system (J, f\j) is
homeomorphic to the one-sided shift on n symbols3.
Proof. Let uj\, . . . , wn_i be the (not necessarily distinct) critical points
of / , and let G : C —>• K+ be the canonical potential function ( = Green's
function), which satisfies G(f(z)) = nG(z) and vanishes precisely on the
filled Julia set K . Thus G(w;) > 0 by hypothesis. The critical points of G
in C \ K are the points uii and also all of their preimages under iterates
of / . Hence the critical values of G (other than zero) are the numbers of
the form G{u)i)/nk with k > 0 .
The puzzle of / is constructed as follows. Choose a number 7 , not of
the form G(oji)/nk , so that 0 < 7 < Min{G(/(wj))} . Then the region
Xo = G - 1 ( 0 , 7] contains no critical values of / , and is bounded by smooth
curves. Similarly, each locus
Xd = rd(X0) = G-Ô.T/n"]
2
Shishikura has shown that the Hausdorff dimension of this Cantor set can be arbitrarily
close to two.
3
Both Przytycki [Pr] and Makienko [Ma] have proved the sharper result that any rational
Julia set which is totally disconnected and contains no critical point must be isomorphic
to a one-sided shift.
J. Milnor
Figure 13. Nested puzzle pieces, and the annulus Ad(z) .
is bounded by smooth curves. Note that the complementary region C \ Xd

cannot have any bounded component, since the harmonic function G cannot
have a local maximum. It follows that the locus Xd is the disjoint union
of a finite number of closed topological disks. By definition, each of these
closed disks will be called a puzzle piece Pd of depth d . Since these puzzle
pieces contain no critical value of f , it follows easily that f maps each Pd
of depth d > 0 by a conformal isomorphism onto some puzzle piece f(Pd)
of depth d — 1 .
If Pd D Pd+i a r e nested puzzle pieces of depths d and d + 1 , then the
set
Ad = Interior(P d ) \ Pd+\
is a well defined annulus of strictly positive modulus. (Figure 13.) We will call
such an Ad an annulus of depth d in the puzzle. Evidently f°d maps each
annulus of depth d by a conformal isomorphism onto an annulus of depth
zero. Hence the moduli of all annuli constructed in this way are uniformly
bounded away from zero.
For any point z in the filled Julia set we can form the nested sequence of
puzzle pieces Po (z) D P\ (z) D • • • , all containing z . Since the moduli of the
annuli Interior Pd(z) \ Pd+\{z) are bounded away from zero, it follows that
the intersection f] Pd{z) reduces to the single point z . Since each boundary
circle dPd is disjoint from K , this implies that K is totally disconnected.
The proof that J has measure zero will be based on the McMullen in-
equality
area(P d + i) <
1+ 4TT m o d (Ad) '
with Ad = Interior(P d ) \ Pd+\ as above. (Compare the Appendix.) However,
to apply this inequality in a useful way, we will need to construct annuli
somewhat differently. Choose a number e > 0 which is small enough so that
V
Figure 14. The "thin annulus" A*{Pd) C Ad C Pd .
there are no critical values of G within the closed interval [7 — 6 , 7 ] . Then

each connected component of G~ ! [(7 — e)/nd, j/nd] is a n annulus, and there
is one such "thin annulus"
A {fi) — rd n u
within each connected component Pd of G l [0, j/nd] . The construction is

such that:
(a) these annuli A*{Pd) have modulus uniformly bounded away

from zero, say moAA*{Pd) > e > 0 , and
(b) every puzzle piece Pd+i which is contained in Pd is actu-

ally contained in the smaller disk Pd \ A*{Pd) (the bounded
component of C \ A*{Pd) ).
Thus the McMullen inequality takes the following sharper form. For each
fixed puzzle piece Pd ,
ST <P \ <? area(Pd) area(Pd)
*h*< l + ^mod^P,)) " l + 4»e '
to be summed over those puzzle pieces of depth d + 1 which are contained
in the given Pd . It now follows inductively that the sum of the areas of all
puzzle pieces of depth d satisfies
l
a r e a ( P 0 ) / ( l + 47re)d .
Since this tends to zero as d —> 00 , and since J c\JPd , it follows that J
has area zero.
90 J. Milnor
To prove that (J , f\j) is isomorphic to the one-sided d-shift, we proceed

as follows. We will construct a closed subset F c C , consisting of paths
leading out to infinity, such that:
(1) F contains all critical values of / ,
(2) /(F) C T , and
(3) the complement U = C N F is simply-connected, and contains
the Julia set.
The construction will be based on the following. By a loxodrome in C \ K
will be meant a smooth curve which cuts the family of equipotential curves
at some constant angle. If we fix this angle, note that there is a unique
loxodrome through any point of C \ K which is not a critical point of G .
The required F will be a locally finite union of loxodromes, all with the same
fixed angle, each starting a some postcritical point of / and extending out
to infinity. If we choose the angle badly, then such a loxodrome may run into
some critical point of G . However, for all but countably many choices of
angle this problem will not occur. There remains the possibility that some
postcritical point z is itself a precritical point of / , and hence a critical
point of G . In this case, there will be finitely many distinct loxodromes of
given angle starting at z , and we can simply choose any one of them, making
compatible choices along the orbit of z . The required properties (1), (2), (3)
are now easily verified.
Since /(F) C F , it follows that / " ' ( [ / ) C (/ . Furthermore, since U is
simply connected, and contains no critical values of / , it follows that every
one of the d branches of / " ' near a point of U extends uniquely to a
holomorphic map <ft : U —> U . The images
...,gn(U) C . V
are disjoint open sets covering the Julia set. We will show that the intersec-
tions
Jt = Jn9i(U)
are disjoint compact sets which form the required Bernoulli partition J =
J\ U • • • U Jn of the Julia set. That is: "*
For each sequence of integers ia , ii, . . . between 1 and n there
exists one and only one orbit ZQ t—> z\ i—• z^ t—t • • • in the Julia
set with Zk 6 Jik for every k .
In fact zQ can be described as the intersection of the nested sequence of

sets ffio o 5it ° • • • ° 9ik(J) • To prove this statement, we use the Poincare
metric on U . Since each & restricted to the compact set J c U shrinks
Poincare distances by a factor bounded away from one, it follows that each
such intersection f\5«o ° 9h ° ''' ° P H ( - 0 consists of a single point. This

proves 2.1. •
Maps with exactly one bounded critical orbit

This will be an exposition of results due to Branner and Hubbard [BH3].
We now suppose that exactly one of the n — 1 critical points of / , counted
with multiplicity, has bounded orbit, while the orbits of the remaining n - 2
critical points escape to infinity. (Thus we exclude examples such as 2 4
z3 + i, for which a double critical point has bounded orbit; however, a double
critical point escaping to infinity is fine.) Furthermore, we assume that n >
3 , so that at least one critical orbit does escape. Then, according to Fatou
and Julia, the Julia set is disconnected, with uncountably many connected
components. Let
CQ H Ci H+ c 2 H
be the unique bounded critical orbit.
As in the proof of Theorem 2.1, choose a number 7 > 0 which is not a
critical value of G , and so that the region G~1(Q, 7] contains no critical
value of / . Again we define the puzzle pieces of depth d to be the connected
components Pd of the locus
\JPd = G'l[o,j/nd] .
Thus each point z € K determines a nested sequence PQ (Z) D PI (Z) D • • •,
and the central problem is to decide whether or not f]d Pd{z) = {z} . Again
we look at the intermediate annuli
Ad{z) = InteriorPd(z) \ Pd+\{z) .
As in the Yoccoz proof, such an annulus is said to be semi-critical, critical,
or off-critical according as the critical point c0 belongs to the annulus itself,
or to the bounded or the unbounded component of its complement. (For this
purpose, we ignore the other critical points, whose orbits escape to infinity.)
This Branner-Hubbard puzzle is easier to deal with than the Yoccoz puzzle
for three reasons:
(a) All of the annuli Ad(z) are non-degenerate, with strictly positive
modulus.
(b) The various puzzle pieces of depth d are pairwise disjoint.
(c) For each z € K , the intersection f]dPd(z) of the puzzle pieces con-
taining z has an immediate topological description: It is equal to the con-
nected component of the filled Julia set K which contains the given point
z . For this intersection is clearly connected, and no larger subset can be
connected since each boundary dPd(z) is disjoint from K .
The tableau of an orbit z0 i-> z, M- • • • is defined just as in §1, and can
be described as a record of exactly which of the annuli Ad(zi) are critical or
92 J. Milnor
Figure 15. An example: Graph of the function f(x) = x(x — c o ) 2 / ( l — c 0 ) 2 on

the unit interval I = [0,1], with critical points CQ = .76 and c o /3 . There is
just one bounded critical orbit c0 t-» 0 i-> 0 >-> • • •. Since the iterated pre-images
of any point of J are dense in J , and since there are three distinct branches
of f~l mapping I into itself, it follows that the Julia set J is completely
contained in the real interval I . This Julia set is plotted underneath the graph.
Since CQ and 0 evidently belong to distinct connected components of J , it
follows from Theorem 2.2 that J is totally disconnected.
semi-critical or off-critical. First suppose that the tableau of the critical orbit
is not periodic. We continue to assume that the degree is n > 3 and that
exactly one of the n-l critical points has bounded orbit.
Theorem 2.2. / / the critical tableau is not periodic, or equiva-

lently if the post-critical points cx, c2 , . -. are all disjoint from
the critical component f]d Pd(c<)) , then for every point z0 of the
filled Julia set K the sum ^ d m o d ^ d ( z 0 ) is infinite, hence
f\dPd{zo) = {20} • It follows that J - K is a totally discon-
nected set of area zero.
Thus J is again homeomorphic to a Cantor set. However, in this case

( J , f\j) is not isomorphic to a shift, or even a sub-shift. For there are critical
points in J , hence / is not locally one-to-one on J .
The proof of this theorem is quite similar to the proof of the Yoccoz
theorem. However there are simplifications, leading to the sharper result
which is stated: £ m o d Ad{z) = 0 0 for all z 6 K. In particular, 1.8 is
replaced by the following statement, which is true whether or not the critical
tableau is periodic. Consider an orbit z0 •-> zx t-> • • • .
L e m m a 2.3. Suppose that the points z\, z-i, ... are all disjoint
from some neighborhood PN(CQ) of the critical point CQ . Then
the annuli Ad(zo) have modulus uniformly bounded away from
zero, hence ]>^d
Proof. Each annulus Ad{zi) of depth d > 0 has modulus at least half
of the modulus of Ad-\{zi+\) . In fact, if i > 0 and d > N then these two
annuli are conformally isomorphic. Thus
where the right side is bounded away from zero since there are only finitely
many annuli of depth zero. •
Proof of 2.2. If the critical tableau is recurrent, then the proof pro-
ceeds exactly as is the Yoccoz argument: Every critical annulus either has
at least 2" descendents in the n-th generation for every n , or else has a
descendent with this property. Since all annuli are non-degenerate, it follows
that £} mod.4 d (co) = oo . On the other hand, if the critical tableau is non-
recurrent, then it follows from 2.3 that JZ modAd(co) = oo . In particular, it
follows that the collection of puzzle pieces {Pd(c0)} forms a fundamental sys-
tem of neighborhoods of the critical point c0 . The corresponding statements
for any iterated pre-image of co follow immediately.
Now consider a point z0 € K , with orbit zo >-+ Z\ >-*••• which never
meets the critical point c0 . If this orbit does not accumulate at CQ , then
the statement that Yldm°d Ad(zo) = oo follows from 2.3. On the other
hand, if this orbit does accumulate at c0 , then since ^ m o d A d ( c o ) — oo ,
an easy tableau argument shows that ^2d m o d Ad(zo) = oo also. Thus
(~]d Pd{z) = {z} in all cases.
Since each boundary dPd{z) is disjoint from the filled Julia set K , it
follows that K is totally disconnected, and hence that J = K . To prove
that this set has measure zero, we proceed as in the proof of Theorem 2.1.
Choose e > 0 so that the interval [7 — e, 7] contains no critical values
of the function G : C -> R+ . Then each puzzle piece Pd(z) contains
a unique component A*(Pd(z)) of the set G~1[(7 — f)/nd, 7/n d ] . These
annuli A*(Pd(z)) c Ad(z) are also non-degenerate, and the proof above
shows equally well that ^ d modj4*(P<i(z)) = 00 . Hence, just as in the
proof of 2.1, for each fixed puzzle piece Pd , the total area of the puzzle
pieces Pd+\ of depth d 4-1 which are contained in Pd satisfies
94 J- Milnor
Define the ratio /i(P<j) by the formula

area/ 3,* = n{Pd) ^
Pd+lCPd
so that this McMullen inequality takes the form l+47rmod^l*(Pd) < v{Pd) •
Then, substituting this formula inductively, we can write
area(Px) = •••
-i) area(P d) ,
PoD-DPd
where the left hand expression is to be summed over all puzzle pieces of
depth zero, the next over all pairs Po D Pi , and so on. (If PQ D • • • D Pd ,
note that P o , . . . , Pd-i are uniquely determined by Pd .) Let % be the
minimum value of the product fi(Po) • • • /J-(Pd-i) as P o , . . . , Pd~\ varies
over all sequences of nested puzzle pieces P o D Pj D • • • D P d _! . Then we
see from this last equality that
]Parea(Po) >
to be summed over all puzzle pieces of depth zero or d respectively. Thus,
if we can prove that rjd —t oo as d —> oo , then it will follow that
area(J) < ^ a r e a ( P d ) < ^ a r e a ( P 0 ) / % -> 0 ,
hence area(J) = 0 as required.
Clearly 1 < rji < r/2 < • • • • If these numbers tended to a finite limit
L < oo , then for each d we could find puzzle pieces Po(d) D P\{d) D • • • D
Pd-\{d) so that fi(P0(d)) • • •/j,(Pd-i{d)) < L. Hence we could choose a
puzzle piece P o which occurs infinity often as Po(d) , then choose Pi c P o
which occurs infinitely often as Pi(d) , and so on. In this way, we could find
a sequence
Po D Pi D P 2 D •••
with fi(P0) • --iiiPd-i) <L<oo for every d . Since l+47rmod (A* {Pi)) <
fi{Pi) , this would imply that
< L < oo ,
contradicting our statement that ^2 mod Ad{z) — oo for all z g K . This
completes the proof of 2.2. •
On the other hand, if the critical tableau is periodic, then we will prove
the following.
Figure 16. Example for Theorem 2.4: Graph of the map f(x) =
x(2x — l)(5x - 4) on the unit interval. In this case the connected
interval [0, | ] is contained in the filled Julia set K . The orbit of
the critical point | escapes to — oo .
Figure 17. Julia set for this map, drawn to the same scale, showing
the puzzle pieces of depth zero and one. Each non-trivial component
of J is homeomorphic to a quadratic Julia set (the "basilica").
96 J. Milnor
Figure 18. Julia set for z \-y z3 + az2 + 1, with a = -1.10692 + .63601 i,
showing the puzzle pieces of depth zero and one. Each non-trivial component of
J is homeomorphic to the Julia set for the quadratic map 2 H z2 + i .
T h e o r e m 2.4. Still assuming that just one critical orbit CQ >—»

Ci H-> • • • is bounded, if the associated critical tableau is periodic
of period p > 1 , so that Pd(co) = Pd(cp) for all depths d, then
the connected component of the filled Julia set K = K(f) which
contains CQ is non-trivial, that is, consists of more than one point.
In fact, a component of K is non-trivial if and only if it contains
some iterated pre-image of CQ .
Thus there there are countably many non-trivial components of K . These

countably many components are everywhere dense in K , since the iterated
pre-images of any point of the Julia set are dense in the Julia set. Since
a disconnected Julia set necessarily has uncountably many components, it
follows that there are uncountably many single points components. (The Julia
set of a rational function may have uncountably many non-trivial components.
Compare [Mel]. However, in the polynomial case no such example is known.)
m m+p m+3
*
Figure 19. Tableau for an orbit z0 H> Z\ H-> • • • which intersects

every critical puzzle piece. Here the stars stand for entries copied
from the periodic critical tableau.
Proof of Theorem 2.4. If Pd{cQ) = Pd{cp) for all d , then it follows

from the tableau rules that Pd{c{) = Pd{cp+i) for all i and d . Hence the
entire orbit
C0 I-)- C2p
p
of c0 under f° is contained in the critical component p] d Pd(co) C K(f) .
This intersection certainly has more than one point. For either it contains
Co cp , or else CQ is a superattracting point, in which case some entire
neighborhood of c0 belongs to f]dPd{co) •
It follows easily that every pre-critical point in K(f) also belongs to a
non-trivial connected component. For if the orbit z0 i-> z\ >->•••• intersects
the critical component C\dPd(co) , then we can choose the smallest I > 0
for which zt belongs to this critical component. It follows easily that foi
maps the component (~)d Pd(z0) containing z0 homeomorphically onto this
critical component.
Now consider an orbit z0 H-> ZX I-» • • • in K(f) which is disjoint from this
critical component. This means that the tableau of this orbit has no columns
which are completely critical. The proof is now divided into two cases:
Case 1. Suppose that there exists a fixed puzzle piece Pd(co) which
is disjoint from this orbit {zn} . Then according to Lemma 2.3 we have
^2dmodAd(zQ) = oo , hence C\dPd{zo) = {zo} •
98 J. Milnoi
Figure 20. A polynomial-like mapping (g, A) with K = K(g, A)

totally disconnected.
Case 2. If the orbit {zn} intersects every critical puzzle piece, then we
use a tableau argument as follows. Choose a depth TV so that the periodic
critical tableau has no semi-critical annuli at depths > N . By hypothesis,
there are infinitely many pairs (d, m) , with d > N , so that the d-th. row
of the tableau for ZQ is semi-critical in column m and off-critical in earlier
columns. (Compare Figure 19.) Using the tableau rules, we can then compute
the tableau in column m + i and depth d — i for 0 < i < d — N . In fact the
entries in column m + jp and depth d — jp are semi-critical, and the others
are off-critical. It now follows that the annulus Aa+m+i(zo) is conformally
isomorphic to an annulus of depth N . Hence its modulus is bounded away
from zero. It follows that £ ] t mod A((z0) — oo , which completes the proof
of Theorem 2.4. •
We can understand this proof better by introducing the following concepts,

which are due to Douady and Hubbard [DH3].
Definition. By a polynomial-like map is meant a pair (g, A) where A C
C is a closed topological disk and g is a continuous mapping, holomorphic
on the interior of A , which carries A onto a closed topological disk g(A)
which contains A in its interior, such that g maps boundary points of A
to boundary points of g (A) .
The degree S > 1 of such a polynomial-like mapping is a well defined
topological invariant. Note that almost every point of g(A) has precisely 6
pre-images in A . The filled Julia set K(g,A) is defined to be the compact
set consisting of all z0 E A such that the entire orbit zQ •->• zx i-> • • • of z0
under g is defined and is contained in A .
L e m m a 2 . 5 . Such a polynomial-like map of degree 6 has 6-1

critical points, counted with multiplicity, in the interior of A .
The filled Julia set K(g, A) is connected if and only if it contains
all of these 6—1 critical points.
Proof. Consider the nested sequence of compact sets

g(A) D A D <T'(A) D g~2(A) D •••
with intersection K(g,A). First suppose that the boundary dA contains no
post-critical points, that is points g°k(uii) with k > 0 where w; is a critical
point of g . Then clearly each g~k(A) is a compact set bounded by one or
more closed curves. In fact, each g~k(A) is either a closed topological disk or
a finite union of closed topological disks, each of which maps onto the entire
disk A under g°k . To see this, note that for each component B of g~k(A)
the image gok(B) is compact, and that gok maps the boundary dB into
dA and maps the interior of B onto an open subset of the interior of A .
Since the interior of A is connected, this implies that g°k(B) = A . If some
such component B were not simply-connected, then some component B' of
C \ g fc(A) would be bounded. A similar argument would then show that
g°k must map B' onto the complementary disk C \ Interior(A) , which is
impossible.
Applying the Riemann-Hurwitz formula to the ramified covering A —>
g(A) , we see that the number of critical points of g in the interior of A ,
counted with multiplicity, is equal to nx(g{A)) ~ x ( A ) — n-1 — 1. (Here x
is the Euler characteristic.) Similarly, applying this formula to g~l(A) -» A ,
we see that the number of critical points in g~1{A) is equal to nx{g~l{A)-
l
X(A) . Thus all of the n - 1 critical points are contained in g~ {A) if and
only if x{g~l(A)) = 1 , so that <?~'(A) consists of a single topological disk.
Similarly, it follows by induction that all of the critical points are contained
in <7~*(A) if and only if g~k(A) is a topological disk. If every g~~k(A) is a
disk, then it follows that the set K = f]g~k(A) is connected. On the other
hand, if some ^""^(A) consists of two or more disks, then each one of these
disks must contain a point of K , which is therefore disconnected.
To complete the proof, we must allow for the possibility that dA may
contain some post-critical point of g . Clearly there can be at most d — I
post-critical points in the annular region <7(A) \ A . Hence we can choose
a disk A] C g{A) whose boundary avoids these post-critical points. If A]
contains A in its interior, and also contains all critical values g(u>i) in its
interior, then it is easy to check that the pair (g , g'1 (Ai)) is a polynomial-
like map of the same degree, and with the same filled Julia set, but with no
post-critical points in the disk boundary. The proof then proceeds as above.
D
100 J. Milnor
Remark. Douady and Hubbard prove much sharper statements: If (g, A)

is polynomial-like of degree 6 > 2 , with K(g, A) connected, then there ex-
ists a polynomial map ip of degree <5 so that if> on some neighborhood of
K (ip) is quasi-conformally conjugate to g on a neighborhood of K(g, A) .
Furthermore, this quasi-conformal conjugacy can be chosen so as to satisfy
the Cauchy-Riemann equations (in an appropriate sense) on the compact set
K(tp) . The polynomial map ip is then uniquely determined up to affine con-
jugacy. In the case S = 2 , one has the further statement that ip depends
continuously on (g, A) .
Now let us return to the situation of Theorem 2.4.
Lemma 2.6. / / the critical tableau is periodic of period p > 1 ,

then for any critical puzzle piece Pd(co) with d sufficiently large,
the pair (/ op , Pd(co)) is polynomial-like of degree two. Further-
more, the critical orbit
c0 >-» cp 1-4 c 2 p •-*
under f°p is completely contained in Pd{co) , so that the filled

Julia set K(f°p, Pr{co)) is connected. In fact K(f°p , Pd(co))
is equal to the intersection of the critical puzzle pieces f]k /\(co) ,
and hence is precisely equal to the connected component of K(f)
which contains CQ .
This is proved by a straightforward tableau argument. Details will be left

to the reader. D
In this way, Branner and Hubbard show that each non-trivial component of
K(f) is homeomorphic to an appropriate quadratic Julia set. As examples,
Figures 17 and 18 illustrate the case p = d = 1 .
3. An Infinitely Renormalizable
Non Locally Connected Julia Set
This section describes an unpublished example of Douady and Hubbard. It
begins with an outline, with few proofs, of results from Douady and Hubbard
[DH1], [DH2], [DH3], [D3]. (See also [Ml], [D5], [Sch], [M3].)
Hyperbolic components in the Mandelbrot set
Let fc(z) = z2 + c. By definition, the Mandelbrot set M is the compact
set consisting of all parameter values c € C such that
the Julia set J(fc) is connected <=> 0 has bounded orbit .
However, it is often convenient to identify M with the corresponding set of

polynomials fc. A map fc e M is hyperbolic (on its Julia set) if and only
if it has a necessarily unique attracting periodic orbit. The hyperbolic maps
in M form an open subset of the plane, and each connected component H
in this open set is called a hyperbolic component in M . Let p be the period
of the attracting orbit, and let Ap = Ap(/C) 6 D be its multiplier. The basic
facts about hyperbolic components in M are as follows:
(1) Any two maps in the same hyperbolic component H C M

have attracting orbits of the same period p . We call p — PH the
period of H .
(2) Each hyperbolic component H is conformally isomorphic to

the open unit disk D under the correspondence fc i-> Ap(/C) .
In fact this correspondence extends uniquely to a homeomorphism
between the closure H and the closed unit disk D .
In particular, each H has a unique center point CH which maps to Ap(c//) =

0 , and each boundary dH contains a unique root point r# € dH which
maps to Xp(rH) = 1 . If the map fc has a superattracting periodic orbit, then
evidently c is the center point for one and only one hyperbolic component.
(3) Similarly, if fr has a parabolic periodic orbit, then r is the

root point for one and only one hyperbolic component H . If the
period of this orbit is p and the multiplier is Xp = exp(2irim/p')
then the period of H is pp'.
By definition, the principal hyperbolic component Hy is the set of fc

having an attracting fixed point. If the multiplier is Aj € D then a brief
computation shows that c = Xx(2 - Ai)/4 , hence
Xi(c) = 1-vT^lc,
taking that branch of the square root which lies in the right half plane. The
boundary dHy is the cardioid, consisting of all points which have the form
c = ei9{2 - eifl)/4 , so that X^ (c) = eiB .
Satellites. Given any hyperbolic component H of period p > 1 and
any root of unity e2mm^p' ^ 1 , it follows from (2) that there is a unique
point r e dH with Ap(r) = e 2irim/p ' . According to (3), this r is the root
point of a new hyperbolic component H' of period pp'. We say that H'
is a satellite, which is attached to H at internal angle m/p'. In the special
case where H is the principal hyperbolic component HQ , we will use the
notation H' — H(m/p') for this satellite, and in general we will use the
notation H' = H * H(m/p') . (Compare the discussion of "tuning" below.)
102 J. Milnor
Figure 21. The period two component H(l/2) in the Mandelbrot

set, with an arrow pointing to its satellite #(1/2) * H(l/7) .
As an example, taking m/p' = 1/2 , the point r = - 3 / 4 is the root point of

the period two component #(1/2) , consisting of all c with \c + 1| < 1/4 .
External rays. We consider not only external rays for the Julia set in
the z-plane ( = dynamic plane), but also external rays for the Mandelbrot set
in the oplane ( = parameter plane). Let H c M be a hyperbolic component
of period p > 1. Then exactly two external rays in C \ M land at the root
point TH • Let 0 < a < & < 1 be their angles. These angles have period
exactly p under doubling, and hence can be expressed as fractions of the
form n/(2 p - 1) . As an example, for the satellite component H(l/p) we
have a = 1/(2P - 1) and b = 2/(2 p - 1) .
For any c € i / U { Y / / } the corresponding external rays Ra(c) and Rt,(c)
in C \ J{fc) land a t a periodic point which can be described as the "root
point" of the Fatou component containing c . See Figure 22 for a schematic
picture of the way these external rays are arranged in the parameter plane
and in the dynamic plane. In particular
0 < a/2 < b/2 < a < b < 1 + a/2 < (1+ b)/2 < 1 ,
where the a/2-ray and the 6/2-ray always land at distinct points: exactly
one of these two rays is periodic.
Renormalization and Tuning. Recall that a map fc 6 M is simply
renormalizable if there is an integer p > 2 called the renormalization period,
and a closed topological disk A in the z-plane, centrally symmetric about
the origin, so that:
(1) f°p~l maps the image /(A) by a conformal isomorphism

onto a disk / o p (A) which contains A in its interior,
cardiojd
Figure 22. Schematic picture of external rays for

a hyperbolic component H C M in the c-plane,
and a corresponding picture in the dynamic plane.
(2) the entire orbit of 0 under f°p is contained in A , and

(3) the images fot(Kp) and f°i(Kp) do not disconnect each otheT
for 0 < i < j < p. Here Kp denotes the small filled Julia set
consisting of points of A whose orbits under fop remain in A
forever.
The following facts are surely true (compare [Ml], [Mc2], [D5], [M3]), al-
though I am not aware of any published proof.
The set of all fc 6 M which are simply renormalizable of period
p consists of a finite number of small copies of M (sometimes
with the root point deleted). Each of these small copies contains
a unique hyperbolic component of period p . Conversely, each
104 J. Milnor
Figure 23. Julia set for the root point of H{\/2)
for the root point of #(1/7) (here the

middle ray a = 6/2 has not been labeled),
and for the root point of #(1/2) * #(1/7)

hyperbolic component H of period p > 2 determines a small

copy of M which can be described as the image of an associated
mapping c H H * c , which embeds M homeomorphically onto a
proper subset H * M C M . The elements of this form (with the
possible exception of the root point rf[ ) , are precisely the elements
of M which are simply renormalizable of period p .
This embedding c i-> H*c maps each hyperbolic component H' of
period p' > 1 conformally onto a hyperbolic component H*H' of
period pp' in such a way that the multiplier Appi(H*c) is equal to
Ap'(c) . In particular, this embedding maps center points to cen-
ter points and root points to root points. It carries the principal
hyperbolic component Hy onto H itself, and maps the satel-
lite H{m/p') of Hv onto the satellite H * H(m/p') of H .
This * -product operation between hyperbolic components is asso-
ciative, and has the principal hyperbolic component Hy as two-
sided identity element. Thus it makes the collection of hyperbolic
components into a monoid (âssociative semigroup with identity)
which operates as a semigroup of embeddings of M into itself.
This monoid is free non-commutative.
Intuitively, the Julia set for H*c' is obtained from the Julia set of a map
fc belonging to the hyperbolic component H by replacing each bounded
component of C x J(fc) by a copy of J(fc') and defining the dynamics
appropriately so that only the copy which is pasted in place of the critical
component is mapped non-homeomorphically. The result is described as fc
tuned by f^ .
In terms of external angles, let a < b be the two external angles for the
root point of H . These angles have periodic binary expansions of the form
.ai • • • op Oj • • • a p • • • and .b\ • • • bp b\ • • • bp • • • , both with period exactly equal
to p . Assertion: If the point d € dM is the landing point of a ray with
angle i € R / Z , then the image H * d is the landing point of a ray whose
binary expansion can be obtained by inserting the p-tuple a^^- • -ap in place
of each zero in the binary expansion of t and the p-tuple 6i • • • bp in place
of each one. (See [D3].)
The example
Consider a sequence of integers 1 < p t < p 2 < • • • •
Theorem 3.1. / / this sequence diverges to infinity sufficiently

rapidly, then the sequence of subsets H(l/pi) * H(I/P2) * • • • *
H(l/pk)*M intersects in a single point u) € M with the property
that the Julia set J(fw) is not locally connected.
106 J. Milnor
The proof begins as follows. It will be convenient to use the abbreviation

r(k) for the root point r/f(i/pt) on the cardioid. Start with any Pi > 1 and
consider the embedding c H-> H(l/pi) * c from M into itself, which carries
the period 1 root point n? = 1/4 to the root point r(l) = r//(i/ pi ) . Since
this embedding is continuous, we can place H(l/pi) *c as close as we like to
r(l) by choosing c close to 1/4 . In particular, if P2 is sufficiently large,
then the entire l/p2-limb will be close to 1/4 by the Yoccoz inequality,
hence every point c e H(1/pi) * H(I/P2) * M will certainly be close to r(l) .
Under such a small perturbation, note that the parabolic fixed point of / r (i)
splits up into a repelling fixed point for fc, together with a nearby period
Pi orbit. Thus, by choosing P2 large, we can place this entire period p t
orbit into an c-neighborhood of the parabolic fixed point of fr^) .
Similarly, choosing p 3 even larger, for any c 6 H{pi)*H(p2) * H(p3) *M

we can guarantee that the parabolic period pi orbit for H{l/p\) * r(2) is
replaced by a nearby period pi p2 orbit. In particular, we can guarantee that
this new orbit lies within the e + e/2 neighborhood of the original parabolic
fixed point. Continuing inductively, we can guarantee that all of the new
orbits which are constructed lie within the 2e neighborhood of the original
fixed point. Furthermore, we can easily guarantee that these successive copies
of M have diameter shrinking to zero.
Thus, if v is the unique point in the intersection, we know that the
Julia set J(fw) contains orbits of period 1 ,Pi, P1P2, • • • , all lying within
the 2e neighborhood of the original parabolic fixed point. These various
periodic points can be described as the landing points of external rays of
angles ak < bk where 0 < 01 < 02 < • • • < b2 < 6] . Furthermore, it
is easy to check that the difference 6^ — ak tends rapidly to zero, so that
lim/t-joo ak = limjb-Kjo it . On the other hand, as noted in the preceding
section, the external rays of angle ak/2 and bk/2 land at different points,
which are negatives of each other. Each such landing point is either close
to the original parabolic fixed point, or close to its negative. Thus we have
a sequence of angles ak/2 and bk/2 tending to a common limit, yet the
landing points of the corresponding external rays for J(fJ) do not tend to
a common limit. According to Caratheodory, this implies that J{fJ) is not
locally connected. D
More generally, if Hi, H2, •. • is any sequence of hyperbolic compo-

nents which are not centered along the real axis, then it is again true that
limbec ak = linifc-too bk . Let w 6 M be any element of the intersection
f]k Hi * H2 * • • • * Hk * M . Whenever the landing points in J{fw) of the rays
of angle ak/2 fail to converge to the critical point, it follows that J(fw) is
not locally connected. In particular, this is true whenever the sequence of
components Hk converges sufficiently rapidly to the root point r<? = 1/4 .
This seems to be the only known obstruction to local connectivity in the

infinitely renormalizable case. Thus whenever the landing points of the ak/2
rays do converge to the critical point, we can ask whether J(fw) is in fact
locally connected.
For further papers related to local connectivity of Julia sets, see for exam-
pie [Ki], [Pe2], [PM], [S01], [S02], [Z].
Appendix: Length-Area-Modulus Inequalities

The most basic length-area inequality is the following. Let I2 C C be
the open unit square consisting of all z = x + iy with 0 < x < 1 and
0 < y < 1 . By a conformal metric on I2 we mean a metric of the form
ds = p(z) \dz\
where z >->• p(z) > 0 is any strictly positive continuous real valued function
on the open square. In terms of such a metric, the length of a smooth curve
7 : (a, 6) -> I2 is defined to be the integral
LP(l) = /
and the area of a region U C I2 is defined to be
area p ([/) — p(x + iy)2dx dy .
In the special case of the Euclidean metric ds = \dz\, with p(z) identically
equal to 1, the subscript p will be omitted.
Theorem A.I. / / area /) (/ 2 ) (the integral over the entire square)

is finite, then for Lebesgue almost every y £ (0,1) the length
Lp{ly) of the horizontal line -yy : 1i-> (t, y) at height y is finite.
Furthermore, there exists y so that
Lp{ly)2 < area p (/ 2 ) . (1)
In fact, the set consisting of all y £ (0,1) for which this inequality
is satisfied has positive Lebesgue measure.
Remark 1. Evidently this inequality is best posible. For in the case of

the Euclidean metric ds = \dz\ we have
L( 7 y ) 2 = area(/ 2 ) = 1 .
Remark 2. It is essential here that we use a square, rather than a rect-
angle. If we consider instead a rectangle R with base Aa; and height Ay ,
108 J. Milnor
then the corresponding inequality would be

. , „ Ax
(2)
for a set of y with positive measure.
Proof of A.I. We use the Schwarz inequality
which says (after taking a square root) that the inner product of any two
vectors in the Euclidean vector space of square integrable real functions on
an interval is less than or equal to the product of their norms. We may as
well consider the more general case of a rectangle R = (0, Ax) x (0, Ay) .
a
Taking f(x) = 1 and g(x) = p(x, y) for some fixed y , we obtain
Ax \ 2 /-Ax
p{x, y)dx) < Ax p(x, y)2 dx ,

/ Jo
or in other words 2
L,( 7y ) < Ax f(•Ax p(x,y)2dx,
Jo
for each constant height y. Integrating this inequality over the interval
0 < y < Ay and then dividing by Ay , we get
A
AT
2
^ ) d <Ay { -~
A - )^ ) • (3)
In other words, the average over all y in the interval (0, Ay) of Lp(jy)2
is less than or equal to ^ areap(^) . Further details of the proof are
straightforward. •
Now let us form a cylinder C of circumference Ax and height Ay by
gluing the left and right edges of our rectangle together. More precisely,
let C by the quotient space which is obtained from the infinitely wide strip
0 < y < Ay in the 2-plane by identifying each point z = x + iy with its
translate z + Ax . Define the modulus mod(C) of such a cylinder to be the
ratio Ay/Ax of height to circumference. By the winding number of a closed
curve 7 in C we mean the integer
W = -r— <j> dx .
Ax J1
Theorem A.2 (Length-Area Inequality for Cylinders). For

any conformal metric p(z)\dz\ on the cylinder C there exists
some simple closed curve 7 with winding number +1 whose
length Lp(j) = § p(z)\dz\ satisfies the inequality

Lp{j)2 < aretLp(A)/mad(A) . (4)
Furthermore, this result is best possible: If we use the Euclidean
metric \dz\ then
L( 7 ) 2 > area(^)/mod(yl) (5)
for every such curve 7 .
Proof. Just as in the proof of A.I, we find a horizontal curve 7 y with
On the other hand in the Euclidean case, for any closed curve 7 of winding
number one we have
= J> \dz\ > <f> dx - Ax ,
J-, Jy
hence L{y)2 > (Ax) 2 = area(C)/mod(C) . D
Definitions. A Riemann surface A is said to be an annulus if it is
conformally isomorphic to some cylinder. An embedded annulus A C C
is said to be essentially embedded if it contains a curve which has winding
number one around C .
Here is an important consequence of Theorem A.2.
C o r o l l a r y A . 3 ( A n A r e a - M o d u l u s I n e q u a l i t y ) . Let A c C
be an essentially embedded annulus in the cylinder C , and suppose
that A is conformally isomorphic to a cylinder C . Then
In particular:
mod(C') < mod(C) . (7)
Proof. Let ( H z be the embedding of C onto A C C . The Euclidean

metric \dz\ on C , restricted to A , pulls back to some conformal metric
P(C)MCI o n C , where p(() — \dz/dQ. According to A.2, there exists a
curve 7' with winding number 1 about C whose length satisfies
L p ( 7 ') 2 < area,(C')/mod(C') .
This length coincides with the Euclidean length L(j) of the corresponding
curve 7 in A CC , and area p (C) is equal to the Euclidean area area(>l) ,
so we can write this inequality as
2
< area(.4)/mod(C) .
110 J. Milnor
But according to (5) we have

area(C)/mod(C) < L( 7 ) 2 .
Combining these two inequalities, we obtain
area(C)/mod(C) < area(^)/mod(C) ,
which is equivalent to the required inequality (6). •
Corollary A.4. The modulus of a cylinder is a well defined con-

formal invariant.
Proof. If C is conformally isomorphic to C then (7) asserts that
mod(C') < mod(C) , and similarly mod(C) < mod(C') . •
It follows that the modulus of an annulus A can be denned as the modulus
of any conformally isomorphic cylinder. Furthermore, if A is essentially
embedded in some other annulus A', then mod(vl) < mod(yl') .
Corollary A.5 (Grotzsch Inequality). Suppose that A' C A
and A" C A are two disjoint annuli, each essentially embedded
in A . Then
mod{A') + mod(,4") < mod(i4) .
Proof. We may assume that A is a cylinder C . According to (6) we
have
area(
mod(A') < 5 , ) mod(C) , mod^") < area(
£ ) mod(C) .
v ; K
~ area(C) ' v ; - ^ ^ ^ *• >
where all areas are Euclidean. Using the inequality
area(A') + area(^") < area(C) ,
the conclusion follows. •
Up to this point, we have only considered cylinders or annuli of finite mod-
ulus. If we take an arbitrary Riemann surface with free cyclic fundamental
group, then it is always conformally isomorphic to some cylinder, providing
that we allow also the possibility of a one-sided infinite or two-sided infinite
cylinder (that is, either the upper half-plane or the full complex plane modulo
the identification z = z + 1 ) . By definition, a Riemann surface conformally
isomorphic to such an infinite cylinder will be called an annulus of infinite
modulus.
Now consider the following situation. Let U C C be a bounded simply
connected open set, and let K C U be a compact subset such that the
difference A = U \ K is an annulus (which may have finite or infinite
modulus).
Corollary A.6. Suppose that K C U as described above. Then

K reduces to a single point if and only if the annulus A — U\K
has infinite modulus. Furthermore, the diameter of K is bounded
by the inequality
^ l ^ § (8)
Proof. According to A.2, there exists a curve with winding number one
about A whose length satisfies L2 < area(/l)/mod(j4) . Since K is en-
closed within this curve, it follows easily that diam(K) < L/2 , and the
inequality (8) follows. Conversely, if K is a single point then using (7) we
see easily that mod(yl) = oo . D
Corollary A.7. (Branner-Hubbard). Let Kx D K2 D K3 D

• • • be compact subsets of C with each Kn+i contained in the
interior of Kn . Suppose further that each interior K% is simply
connected, and that each difference An = A"° \ Kn+\ is an an-
nulus. If 535° mod(An) is infinite, then the intersection f] Kn
reduces to a single point.
Proof. It follows inductively from the Grotzsch inequality that the mod-
ulus mod(K° \ Kn) tends to infinity as n -> oo . Hence by (7) and A.7
the intersection of the Kn is a point. D
Corollary A.8 (McMullen Inequality). Again suppose that
K C U C C , and that A = U \ K is an annulus. Then
area(iC) < area(t/)/e 4jrmod ^ ) . (9)
Proof. (Compare [BH, II].) We will first prove the weaker inequality
< 7*^^ . (10)

~ l + 47rmod(^)
The isoperimetric inequality asserts that the area enclosed by a plane curve
of length L is at most L2/(4TT) , with equality if and only if the curve is a
round circle. (See for example [CR].) Combining this with the proof of A.6,
we see that
,.,, . L2 aiea{A)
area(tf) < — < ^
4TT 4
Writing this inequality as 4xmod{A) <axea.(A)/aiea{K) and adding +1
to both sides we obtain the completely equivalent inequality l+47rmod(yl) <
area((7)/area(/C) . This in turn is equivalent to (10).
112 J. Milnor
To sharpen this inequality, we proceed as follows. Cut the annulus A

up into n concentric annuli Ai, each of modulus equal to xnod(A)/n .
Let Ki be the bounded component of the complement of Ai, and assume
that these annuli are nested so that Ai U Ki — K°+l with K\ — K . Let
K°+1 =Al)K = U. Then
area.(Ki+i)/areai(Ki) > 1 + 47rmod(yl)/n
by (10), hence
area(t/)/area(/O > (1 + 47rmod( J 4)/n) n ,
where the right hand side converges to e4)r mod ( A ) as n —> oo . •
As a final illustration of modulus-area inequalities, consider a flat torus
T = C/A . Here A C C is to be a 2-dimensional lattice, that is an additive
subgroup of the complex numbers, spanned by two elements Ai and A2
where Ai/A2 £ R . Let A C T be an embedded annulus.
By the "winding number" of A in T we will mean the lattice element
w E A which is constructed as follows. Under the universal covering map
C -> T , the central curve of A lifts to a curve segment which joins some
point ZQ 6 C to a translate ZQ + w by the required lattice element. We say
that A cT is an essentially embedded annulus if w / 0 .
Corollary A.9 (Bers Inequality). If the annulus A is embed-
ded in the flat torus T = C/A with winding number w 6 A ,
then
«dM)<=j=p. (U)
Roughly speaking, if A winds many times around the torus, so that |u;|
is large, then A must be very skinny. A slightly sharper version of this
inequality is given in Problem A-3 below.
Proof. Choose a cylinder C which is conformally isomorphic to A . The
Euclidean metric \dz\ on A C T corresponds to some metric p(C)|d£| on
C , with
a r e a ^ C ) = aiea(A) .
By A.2 we can choose a curve 7' of winding number one on C , or a corre-
sponding curve 7 on A C T , with
< area p (C)
pK1
' ~ mod(C') mod(^l) - mod{A) "
Now if we lift 7 to the universal covering space C then it will join some point
ZQ to z0 + w . Hence its Euclidean length £(7) must satisfy £(7) > \w\ .
Thus
H2 S
which is equivalent to the required inequality (11). •
Concluding P r o b l e m s :
P r o b l e m A - l . In the situation of Theorem A.I, show that more than
half of the horizontal curves yy have length Lp(jy) < y/2aieap(P) . (Here
"more than half" is to be interpreted in the sense of Lebesgue measure.)
P r o b l e m A-2. Show that the converse to A.7 is false: A nest K\ D
K2 D K3 D • • • of compact subsets of C may intersect in a point even if
Y^° mod(K° \ Kn+\) is finite. (For example, do this by showing that a
closed disk D of radius 1/2 can be embedded in the open unit disk D so
that the complementary annulus A = D \ D has modulus arbitrarily close
to zero.)
Problem A-3 (Sharper Bers Inequality). If the flat torus T = C/A
contains several disjoint annuli At, all with the same "winding number"
w & A , show that
^mod(Aj) < area(T)/H2 .
If two essentially embedded annuli are disjoint, show that they necessarily
have the same winding number.
References
[A] L. Ahlfors, "Conformal Invariants", McGraw-Hill 1973.
[Be] A. Beardon, "Iteration of Rational Functions", Grad. Texts Math. 132,
Springer 1991.
[Bl] P. Blanchard, Disconnected Julia sets, pp. 181-201 of "Chaotic Dynamics
and Fractals", ed. Barnsley and Demko, Academic Press 1986.
[BDK] P. Blanchard, R. Devaney and L. Keen, The dynamics of complex
polynomials and automorphisms of the shift, Inv. Math 104 (1991) 545-
580.
[Brl] B. Branner, The parameter space for complex cubic polynomials, pp.
169-179 of "Chaotic Dynamics and Fractals", ed. Barnsley and Demko,
Academic Press 1986.
[Br2] B. Branner, The Mandelbrot set, pp. 75-105 of "Chaos and Fractals",
edit. Devaney and Keen, Proc. Symp. Applied Math. 39, Amer. Math.
Soc. 1989.
[BH1] B. Branner and J. H. Hubbard, The iteration of cubic polynomials,
Part I: the global topology of parameter space, Acta Math. 160 (1988),
143-206;
114 J. Milnor
[BH2] B. Branner and J. H. Hubbard, The iteration of cubic polynomials,

Part II: patterns and parapatterns, Acta Math. 169 (1992), 229-325.
[CG] L. Carleson and T. Gamelin, "Complex Dynamics", Springer 1993.
[CR] R. Courant and H. Robbins, "What is Mathematics?", Oxford U. Press
1941.
[Dl] A. Douady, Systemes dynamiques holomorphes, Seminar Bourbaki, 35e
annee 1982-83, n° 599; Asterisque 105-106 (1983) 39-63.
[D2] A. Douady, Julia sets and the Mandelbrot set, pp. 161-173 of "The
Beauty of Fractals", edit. Peitgen and Richter, Springer 1986.
[D3] A. Douady, Algorithms for computing angles in the Mandelbrot set,
pp. 155-168 of "Chaotic Dynamics and Fractals", ed. Barnsley & Demko,
Acad. Press 1986.
[D4] A. Douady, Disques de Siegel et anneaux de Hermann, Seminar Bourbaki,
39e annee 1986-87, n° 677; Asterisque 152-153 (1987-88) 151-172.
[D5] A. Douady, Chirurgie sur les applications holomorphes, Proc. Int. Cong.
Math. Berkeley, A.M.S. (1987) 724-738.
[DH1] A. Douady and J. H. Hubbard, Iteration des polynomes quadratiques
complexes, C. R. Acad. Sci. Paris 294 (1982) 123-126.
[DH2] A. Douady and J. H. Hubbard, Etude dynamique des polynomes com-
plexes I & II, Publ. Math. Orsay (1984-85).
[DH3] A. Douady and J. H. Hubbard, On the dynamics of polynomial-like
mappings, Ann. Sci. Ec. Norm. Sup. (Paris) 18 (1985), 287-343.
[EL] A. Eremenko and M. Lyubich, The dynamics of analytic transformations,
Leningr. Math. J. 1 (1990) 563-634.
[F] D. Faught, Local connectivity in a family of cubic polynomials, thesis,
Cornell 1992.
[GM] L. Goldberg and J. Milnor, Fixed point portraits of polynomial maps
II, Ann. Sci. Ec. Norm. Sup. 26 (1993), 51-98.
[He] M. Herman, Recent results and some open questions on Siegel's lineariza-
tion theorem of germs of complex analytic diffeomorphisms of Cn near a
fixed point, pp. 138-198 of Proc 8 th Int. Cong. Math. Phys., World Sci.
1986.
[Hul] J. H. Hubbard, Puzzles and quadratic tableaux (according to Yoccoz),
preprint 1990.
[Hu2] J. H. Hubbard, Parapuzzles and local connectivity in the parameter
plane (according to Yoccoz), preprint 1990.
[Hu3] J. H. Hubbard, Local connectivity of Julia sets and bifurcation loci:
three theorems of J.-C. Yoccoz, pp. 467-511 of "Topological Methods in
Modern Mathematics", ed. Goldberg and Phillips, Publish or Perish 1993.
[Ka] J. Kahn, Holomorphic Removability of Julia Sets, Stony Brook I.M.S.
Preprint 1998#11.
[Ki] J. Kiwi, Non-accessible Critical Points of Cremer Polynomials, Stony
Brook I.M.S. Preprint 1995#2.

[LI] M. Lyubich, The dynamics of rational transforms: the topological pic-
ture, Russian Math. Surveys 41:4 (1986), 43-117.
[L2] M. Lyubich, Milnor's attractors, persistent recurrence and renormal-
ization, pp. 513-541 of "Topological Methods in Modern Mathematics",
ed. Goldberg and Phillips, Publish or Perish 1993.
[L3] M. Lyubich, Renormalization ideas in conformal dynamics. Current
developments in mathematics, 1995 (Cambridge, MA), 155-190, Internat.
Press, Cambridge, MA, 1994.
[L4] M. Lyubich, Dynamics of quadratic polynomials. I, II. Acta Math. 178
(1997) 185-247, 247-297.
[LM] M. Lyubich and J. Milnor, The Fibonacci unimodal map, J. Amer.
Math. Soc. 6 (1993), 425-457.
[Ma], P. Makienko, Totally disconnected Julia sets, MSRI preprint 042-95.
[Mel] C. McMullen, Automorphisms of rational maps, pp. 31-60 of "Holo-
morphic Functions and Moduli", edit. Drasin et al., Springer 1988.
[Mc2] C. McMullen, "Complex Dynamics and Renormalization," Ann. Math.
Studies 135, Princeton U. Press, 1994
[Mc3] C. McMullen, "Renormalization and 3-manifolds which fiber over the
circle", Ann. Math. Studies 142, Princeton U. Press, 1996.
[Ml] J. Milnor, Self-similarity and hairiness in the Mandelbrot set, pp. 211-
257 of "Computers in Geometry and Topology", edit. Tangora, Lect. Notes
Pure Appl. Math. 114, Dekker 1989
[M2] J. Milnor, "Dynamics in One Complex Variable, Introductory Lectures",
Vieweg 1999, to appear.
[M3] J. Milnor, Periodic orbits, external rays and the Mandelbrot set: an
expository account, Stony Brook I.M.S. Preprint 1999#3; Asterisque to
appear.
[Pel] C. L. Petersen, On the Pommerenke-Levin-Yoccoz inequality, Ergodic
Theory Dynamical Systems 13 (1993) 785-806.
[Pe2] C. L. Petersen, Local connectivity of some Julia sets containing a circle
with an irrational rotation. Acta Math. 177 (1996), 163-224.
[Pe3] C. L. Petersen, Puzzles and Siegel disks, "Progress in Holomorphic
Dynamics", pp. 50-85, Pitman Res. Notes Math. Ser., 387, Longman,
Harlow, 1998.
[PM] R. Perez-Marco, Fixed points and circle maps, Acta Math. 179 (1997),
243-294.
[Pr] F. Przytycki,.Iterations of rational functions: which hyperbolic compo-
nents contain polynomials? Fund. Math. 149 (1996), 95-118.
[RS] J. Riedl and D. Schleicher, On Crossed Renormalization of Quadratic
Polynomials, "Problems on Complex Dynamical Systems" (Proceedings of
conference on holomorphic dynamics), Kyoto 1997.
116 J. Milnor
[Sch] D. Schleicher, Rational parameter rays of the Mandelbrot set, Stony

Brook I.M.S. Preprint 1997#13.
[Sh] M. Shishikura, The Hausdorff Dimension of the Boundary of the Man-
delbrot Set and Julia Sets, Ann. of Math. 147 (1998), 225-267.
[S0l] D. E. K. S0rensen, Accumulation theorems for quadratic polynomials,
Ergodic Theory Dynam. Systems 16 (1996), 555-590.
[S02] D. E. K. S0rensen, Describing quadratic Cremer point polynomials by
parabolic perturbations. Ergodic Theory Dynam. Systems 18 (1998), 739-
758.
[St] N. Steinmetz, "Rational Iteration: Complex Analytic Dynamical Sys-
tems", de Gruyter 1993.
[Su] D. Sullivan, Conformal dynamical systems, pp. 725-752 of "Geometric
Dynamics", edit. Palis, Lecture Notes Math. 1007, Springer 1983.
[Z] S. Zakeri, Biaccessibility in Quadratic Julia Sets II: The Siegel and Cremer
Cases, Stony Brook I.M.S. Preprint 1998#1.
Institute for Mathematical Sciences,

State University of New York at
Stony Brook NY 11794-3660, U.S.A.
e-mail: jack@math.sunysb.edu
Holomorphic motions and puzzles
(following M. Shishikura)
Pascale Roesch*
Abstract
We prove that the Mandelbrot set is locally connected at Yoccoz
parameters, following an idea of Shishikura.
The aim of this article is to present a short proof, due to M. Shishikura

(unpublished), of the following theorem first obtained by J.-C. Yoccoz:
Main Theorem (Yoccoz). The Mandelbrot set M is locally connected at

parameters c for vihich fc(z) — z2 + c is a non-renormalizable map whose
fixed points are repelling and whose critical point is recurrent.
For parameters c satisfying the above conditions, another theorem of Yoc-

coz (see the paper by J. Milnor in this volume) asserts that the Julia set Jc
of fc is locally connected, and in particular exhibits a fundamental system
of neighborhoods Pn{c) of c—in the dynamical plane of fc—whose closures
have connected intersections with Jc. We derive the Main Theorem from this
dynamical result in the following way. Given a prescribed parameter cQ 6 M,
we use the Douady-Hubbard conformal representation of C \ M to construct
neighborhoods Vn of c0—in the parameter plane—whose closures have con-
nected intersections with M. Then, to show that {CQ} — CWn, we consider
the annuli An = Vn \ Vn+\ and prove that the sum of their moduli is infinite.
To get this estimate, we compare the modulus of An with the modulus of
the "corresponding" annulus An(c0) = Pn(c0) \ Pn+i(c-o)- In M. Shishikura's
proof, this key comparison relies on quasiconformality properties of holomor-
phic motions and the observation that, for c € Vn, the dynamical piece Pn(c)
moves holomorphically.
Section 1 briefly recalls the construction of Yoccoz puzzles and some per-
tinent results which will be needed later (for proofs, the reader is referred
to Milnor's article [M2]). Section 2 deals with parallel constructions in the
parameter plane; it provides a para-puzzle "corresponding" to the dynamical
Yoccoz puzzle (see also [B, H] for similar constructions). Section 3 completes
the proof of the Main Theorem, and in particular explains how to compare
the modulus of a parameter annulus with the modulus of a critical annulus
in the dynamical plane.
'Research partially supported by the ESF.
117
118 P. Roesch
I am very grateful to M. Shishikura for explaining this proof that I adapted

in [R] to study local connectivity in the parameter plane of cubic Newton
maps.
1 Puzzles in the dynamical planes

1.1 External rays
We consider the family of quadratic polynomials fc(z) = z2 +c parameterized
by c e C. By definition, the filled Julia set Kc of fc is the complement of the
open set
5 c (oo) = {z 6 C | f:(z) > oo}
n—yoo
which is the basin of infinity. The potential function
GC:C—>R+, *.—> lim2-log + |/ e B (z)|>
(where log + :r = sup{0,loga;}) is a proper harmonic function whose zero set

is Kc and whose critical values are bounded by GC(Q). Hence, the open set
Uc = {z | Gc(z) > G c (0)} is a punctured disc and the classical Bottcher
Theorem provides a unique conformal representation
(f>c'Uc—> {z | \z\ >rc), where logr c = G c (0),
which conjugates fc to z i-> z2 and satisfies Gc =

The level curve {z E C | Gc{z) = v}, v > 0, is called the equipotential of
level v and is denoted by Ec(v). The external ray of angle t e'R/Z, denoted
by Rc(t), is the gradient line of Gc stemming from infinity with the angle t
(measured via the Bottcher coordinate <j>c). When Kc is connected, i.e., when
c belongs to the Mandelbrot set, <j>c is a conformal representation from 5 c (oo)
to C \ ID and external rays are just the inverse images by tf>c of the rays in
C\D.
An external ray is an analytic curve that either hits a critical point of Gc
or accumulates on the Julia set Jc. Moreover, if t is rational and Rc(t) hits
no critical point, then Rc(t) lands at (i.e. converges to) some point of Jc that
is a repelling or parabolic (eventually) periodic point [DH1]. Conversely, for
c € M, each repelling periodic point is the landing point of at least one (and
at most finitely many) external ray(s) (see [Ml]).
1.2 Wakes of the Mandelbrot set

We now briefly describe the regions in the parameter plane where the fixed
points of fc are both repelling and one of them is the landing point of a
Holomorphic motions and puzzles 119
prescribed cycle of external rays. These conditions define parameters for

which the Yoccoz puzzle can be constructed.
Theorem 1.1 [DH1]. The map
fCvAf — > C \ D
\
is a conformal representation—where 4>c is the Bottcher coordinate on Uc 9 c.
Thus, the conformal homeomorphisms <f>c and <j>M have the same image,
C \ D , and they map c—with different meanings—to the same point. In
particular, every point w e C x I ) has the form w — <f>c(c) for c = ^(ui).
As in the dynamical plane, the external ray of angle t S R/Z in the
parameter plane is the set
and the equipotential of level v > 0 is the closed curve

£(v) = {c€C\\og\*M(c)\=v}.
One easily checks that the ray 11(0) is exactly the half-line ]l/4, +oo[. For
c £ C \ 7£(0), the ray Rc(0) lands at a repelling fixed point of fc usually
denoted by /3C. The second fixed point ac of fc is attracting if and only if the
parameter c lies inside the open region % C M bounded by the central car-
dioid. The multiplier f'c(ac) defines a conformal representation from H to the
unit disc D which extends to a homeomorphism between the boundary curves.
We denote by 7: R/Z ~ dB —> dH the inverse of the latter homeomorphism.
The following two theorems (see [DHL M2]) describe the regions of C \ %
over which the external rays landing at ac have fixed angles.
Theorem 1.2 [DH1].
a) Every external ray TZ(t) of rational angle t converges to some parameter
in M.
b) A parameter c 6 M is the landing point of some external ray H(p/q)
with (p, q) = 1 and q even if and only if the critical point of fc is eventually
periodic. Moreover, for such a parameter c, the external ray H(r/s) converges
to c if and only if, in the dynamical plane of fc, the external ray Rc(r/s) lands
at the critical value c.
c) For any rational p/q with (p, q) — 1 and q > 1, the parameter f(p/q) € dH
is the landing point of exactly two external rays TZ(t") and TZ(t+), where the
angles tx = i* {p/q) belong to the same orbit under multiplication by two and
satisfy 0 < V < t*~ < 1.
120 P- Roesch
The proper arc TZ{t~)un{t+)U^{p/q) divides the plane into two connected
components. We denote by W p /, the component which does not contain the
domain H.
Theorem 1.3 [GM]. For any parameter c in W p /,, there are exactly q ex-
ternal rays that land at the fixed point ac, namely the rays RC(2H+), where
t+ = t+(p/q) and i = 0 , . . . ,q - 1. Moreover, if we cut out the plane along
these rays, the critical value c lies in the region bounded by Rc{t~) and Rc{t+).
1.3 The Yoccoz puzzle

Finally, we define the Yoccoz puzzle (see [M2]) for every parameter^; belonging
to a wake Wp/g with (p, q) = 1 and q > 1 (every point of M \ H belongs to
such a Wp/q).
Definition 1.4. Let c € Wp/q, t+ = t+{p/q) and Xc = {z € C | Ge{z) < 1}.

The Yoccoz puzzle of fc is given by the following sequence of graphs:
/oc = dXc U (Xc fl y fle(2it+)))
Icn = / r C o ) for all n > 1.
The puzzle pieces of depth 0 are the q connected components of Xc \ /§.

Each of them contains exactly one point /*(0) with 0 1 are
the connected components of f~n{Xc) \ Icn = f~n(Xc \ /^). The puzzle
piece containing a given point z will be denoted by f£(z). The puzzle pieces
containing the critical value c will be simply denoted by P£,..., P£,
For c € Wp/q, a theorem of Yoccoz [M2] asserts that, if fc is non-renorm-

alizable and combinatorially recurrent, then its Julia set is locally connected,
and in particular (X°=i ?n ~ ic)-
The map fc is combinatorially recurrent if, for every n, there exists m > 0
such that /™(0) belongs to JF£(O). If the critical point 0 is recurrent then fc is
obviously combinatorially recurrent. Conversely, if fc is non-renormalizable
and combinatorially recurrent then it follows from the theorem of Yoccoz that
the point 0 is recurrent.
Milnor's paper [M2] gives further information about the puzzle pieces P^
that will be needed in Section 3. In particular, it provides a sequence of
annuli A^ having the following properties:
Acn. — f£. \ Pnt+i surrounds the critical value c of fc and is non-

degenerate—meaning that PP D Pf...;
• /"• n° induces an unramified covering map from Acn^ onto AcnQ; and
the sum X ^ > o m ° d ^ n IS

infinite, where modj4£ denotes the modulus
oiAi.
2 Puzzles in the parameter plane

As in the previous section, we will use the following conventions: objects in
the parameter plane will be denoted by calligraphic capital letters (such as
V, TV), while those in dynamical planes will be denoted by roman capitals
(such as P, R). Symbols associated with objects from the dynamical plane
of fc will also carry a subscript or superscript c.
2.1 Yoccoz para-puzzles

Since the combinatorics of the Yoccoz puzzle of fc depends on the wake W p /,
that contains c, we will construct one para-puzzle for each W p / g , using the
same angles as in the dynamical plane and replacing the Bottcher coordinate
by the Douady-Hubbard conformal representation. Fix two relatively prime
integers p, q with q > 1.
Definition 2.1. Let t+ = t+(p/q) and Xn = {c e Wp/q | log|^ M (c)| < 2~n},
n > 0. The Yoccoz para-puzzle is given by the following sequence of graphs
defined in the closure of W p /,:
where Jn = j i e R/Z | 2nt 6 {t+,... ,29'H+}\. The para-puzzle pieces

of depth n are the connected components of the Xn \ Xn.. A general para-
puzzle piece of depth n will be denoted by Vn and the one containing a given
parameter c by Vn(c).
Remark 2.2. For any parameter c in W p /,, Theorem 1.1 shows that /§ con-
sists of converging rays and equipotentials. The graph /f has the same prop-
erty provided the critical value c does not lie on /§. By induction, if c does
not belong to I^_1 U • • • U /£ then /£ U • • • U 1$ is a union of converging rays
and closed equipotentials. We will now characterize the set of parameters c
for which this occurs.
Lemma 2.3. A parameter c € W p /, belongs to Xn if and only if the critical
value c of fc lies on /£. As a consequence, if c belongs to a para-puzzle piece Vn
of depth n, then c is outside of 1% U • • • U /Q.
122 P. Rcesch
Figure 1: Shape of dynamical and parameter pieces of depth < 3 for W1/3.
Proof. This is clear for n — 0. Now recall that if c does not belong to Kc—or
equivalently if c does not belong to M—the Bottcher coordinate <f>c is denned
on a disc Uc containing the critical value. For parameters c in Wp/g \ M, the
equality $M(C) = <fic{c) implies that the argument of <f>c(c) belongs to [t~, t+]
(Theorem 1.3). Hence, by definition of graphs and para-graphs, a parameter
c £ Wp/q is in Xn \ M if and only if the critical value c is in /£ \ Kc. Then
Theoreml.2-b) completes the proof for c G M. Indeed, the rays of Xn \ I o
land at parameters c such that /"(c) = ac, and each of these parameters c
is the landing point of q external rays corresponding to the q dynamical rays
landing at the critical value c (c is a preimage of ac). The second part of the
statement follows readily. •
parameter plane
model plane
dynamical plane of fc
holomorphic motion
Figure 2: Relations between dynamical, parameter and reference planes.
Remarks 2.4.
a) Let n be the minimal index such that In contains the critical value. Then
the graphs / £ , . . . , / £ do not contain the critical point, and therefore consist
of rays and equipotentials. This holds in particular when the parameter c
belongs to Vn-\-
b) All para-puzzle pieces are simply-connected since each graph Xn is made
connected by the equipotential it involves.
c) The unique para-puzzle piece VQ of depth 0 is bounded by lZ(t+), ~R-{t~)
and £(1). In the dynamical plane of fc for c € Vo, the curves Rc(t+), Rc(t~)
and Ec(l) are exactly the rays and equipotential that bound the piece Poc
containing the critical value (Theorem 1.3). Thus, using the canonical pa-
rameterizations of rays and equipotentials, we get a natural homeomorphism
from dV0 to dP§ for any CELVQ.
2.2 Relations between puzzles and para-puzzles

We want now to generalize Remark 2.4-c) to deeper graphs, i.e., to construct
a "canonical" homeomorphism from dVn to dP£° for any n > 0 and any
parameter Co e Vn. Canonical here means the homeomorphism extends to a
quasiconformal map from Tn to Pf*. The construction proceeds by induction
and relies on the fact that, when c moves within a para-puzzle piece Vn, the
graph 7£+1 moves holomorphically.
Recall that a holomorphic motion of a given subset Z C C parameterized
by a complex manifold C is a map C x Z ->• C x C of the form (c, z) •-> (c, hc(z))
which is holomorphic on each slice C x {z}, z € Z, injective on each slice
{c} x Z, c £ C, and satisfies /ico(2) = z for every z 6 Z and a given CQ £ C.
L e m m a 2.5. Let Co be a parameter in a para-puzzle piece Vn of depth n > 0.

There exists a holomorphic motion
5UC/I i/ia< /= + 1 = /i£(C +1 ) /or even/ c e P B .

Moreover, ifn > 1, Men /^ coincides with hn_x on Vn x ( / J T I / ^ , ) (where
Vn-\ = 'Pn-i(co) D ^ n ) and for every c eVn the diagram
/CO ^
I*
JC
n n
1^
is commutative.
124 P. Roesch
Proof. Let Vn = {z £ C | log + |z| > 2~( n + 1 ) }. Since (j>^{z) is holomorphic in

the two variables (c, z) and is defined for (c, z) e Vn x Vn, the map
(<=,*)
is a holomorphic motion of <j>c^(Vn) (note that this set contains the equipo-
tential of /^_j). Moreover, for c 6 Vn, the critical point of fc remains outside
of /^ + 1 (Lemma 2.3), so that fc admits local inverses on I^+l. Therefore,
hn extends uniquely over Vn x (/^°+1 \ K^) to a holomorphic motion—
still denoted by hn—satisfying fc o hcn = hcn o fCo. One easily checks that
Kco) = /£ + 1 \ Kc. The final extension
U Vn X
is given by the A-Lemma of Mane-Sad-Sullivan (see [D]) which shows that any
holomorphic motion of a given subset in C extends to a holomorphic motion
of the closure. Since the extended map hcn is continuous, hcn{I^+x) = /£ + 1 .
The others desired properties are satisfied by construction. •
pco
parameter plane
dynamical plane of CQ
Figure 3: Holomorphic motion of over Vn.

Notations 1. From now on, we fix a parameter CQ 6 M and we suppose that

c0 lies in some para-puzzle piece Vn at each depth n. For every c in Vn, we
denote by D^+l the puzzle piece of depth n + 1 in the dynamical plane of fc
which is bounded by h^(dP^+1), where hn is the holomorphic motion given
by Lemma 2.5.
L e m m a 2.6. There exists a unique homeomorphism
H:
n>0 n>0
coincides with ((/>£„) ' o $ M outside of M. Moreover, H maps dVn onto

and is given, on Vn fMn+\, by the inverse holomorphic motion of the
critical value: H{c) = (h^)~l(c).
Proof. We will prove the following assertion ^P(fc) by induction on k > 0:
The map {<t>co)"~1 ° $ M > restricted to
(dV0 u {Vo n l i ) u • • • u (pfc_! n J/t)) x M,
extends to a homeomorphism H from dVo U (Po H l i ) U • • • U {Vk

to ^Pp00 U (P™ n /=°) U • • • U (Pj*^n /^°) which takes dTk onto dP?.
The step k = 0 follows from remark 2.4-c) and we now assume that ty(k) is
satisfied. _
We first show that (j>ô$M induces a homeomorphism from ('PfcnXfc+i)\M
onto (P^nl^+1) N J o . I n fact > fca ° ^ M : c x M ~+ c x J«> maps the point
whose Douady-Hubbard coordinate is re2"1 to the point whose Bottcher co-
ordinate is the same. Moreover, by induction, the homeomorphism H takes
dVk onto dff. Therefore, the image of ( ^ n l k + 1 ) \ M under <j>^ o <E>M is
exactly (PJ n I?+l) \ J^.
Note that on (Vk n l f c + i ) \ M, the map H satisfies
H{c) = 0" 1 o* M (c) - ^ ( ^ ( c ) ) = (^)- 1 (c).
Thus we extend H by setting H{c) = (/ifc)""'(c) f° r c e P* n l w . To see

that this extension is still injective, we have to show that, if two distinct
external rays Tl(ti),TZ(t2) C I* + i land at points cx,c2 € Vk with the same
image under H, then C] = c2. Since cx,c2 belong to Vk with A; > 1, they
are Misiurewicz parameters (i.e., parameters for which the critical point is
eventually periodic). This follows readily from Theorem 1.2-b) and the fact
that t\,t2 have even denominators since they both differ from t±. Then
Theorem 1.2-b) implies that the two dynamical rays Rci(ti), Rîh) land at
c\ and c2 respectively. Since hck' (Rco(tj)) = RCj(tj) for j € {1,2}, the rays
126 P. Roesch
and i ? ^ ^ ) both land at H{c\) — H(c2). In turn, this implies that

RCl(ti) and RCi(t2) land at cx. By Theorem 1.2-b), fc(t2) lands at cu and
therefore c\ = c2.
Finally, we have to show that H maps dVk+\ onto dP^x- Since H is
a homeomorphism from &Pk U (P^ n I A + i) to d.F£° U fp£° n /£%), it maps
dVk+i to 9Qfc+i for some piece Qk+i of depth k + 1. On the other hand, for
c € 'Pjt+i, the critical value c of / c is in Dk+l because it moves continuously
with c, belongs to Dk\l for c = c0 and does not meet Ik+l (Lemma 2.3).
Hence, if c 6 dVk+i lies inside some equipotential segment of dVk+x, then the
critical value c belongs to dDck+v Therefore, H(c) = {hk)~l{c) e d-Pfc+j and
lies inside some equipotential segment of dP^+y Since PkXi is the only piece
of depth k + 1 whose boundary contains this point, Qk+i = P^Xi- ^
Corollary 2.7. The para-puzzle piece Vk+i containing Co is the set of pa-
rameters c for which the two puzzle pieces F^ + ] and Dck+l coincide. Thus c
belongs to Vk \ Vk+v if and only if the critical value c lies in Dck \ Dck+l.
Proof. As in Lemma 2.6, for c € Vk+i, the critical value c of fc is in D%+1.

Conversely, suppose c belongs to Vk \ T'fc+i- Let ct, t € [0,1], be a contin-
uous path in Vk joining c0 to C] = c, crossing 5Pfc+i at exactly one point c^,
and avoiding 2/t+i D M (such a path can be constructed using equipotentials
and rays). Then for t < to, the critical value / Ct (0) = c% belongs to Dck+V
Moreover, since the parameter path Cj crosses Ik+i at t — to, the critical
value /c,(0) gets out of Dck*+l when t passes over t0. Then Lemma 2.6 insures
that, for t > t0, the critical value does not cross dDck'+l again (because the
path does not cross Vk+i). Hence cx = c lies outside of Dk+l. D
Lemma 2.8. If D^+1 C £>£> then Vn+1 C Vn.

Proof. The corollary 2.7 shows that the homeomorphism H of Lemma 2.6
maps dPn+l n Vn into dD°?+v If ££>+1 C D^°, then H~l(dD^+l) is compact,
so dVn+\ H 'Pn is also compact (since it is closed in Vn). On the other hand,
dVn+\ is connected, and it is not entirely included in dVn because of the piece
of equipotential it contains. Therefore Vn+\ C Vn. D
3 Proof of the Main Theorem

3.1 Extending the holomorphic motions
Let Co € M be a parameter satisfying the hypotheses of the Main Theorem
i.e., for which / ^ ( z ) = z2 + Co is a non-renormalizable map whose fixed points
are repelling and whose critical point is recurrent. Since CQ does not belong
to the central cardioid, a well-known result (see for example [GM]) asserts
that Co lies in some wake Wp/q with (p, q) — 1 and q > 1; moreover, Co is
not a Misiurewicz parameter. Thus c0 belongs to a complete sequence of
pieces Vn, n > 0, of the para-puzzle constructed in Section 2. As in the
dynamical planes, we consider the sequence of (possibly degenerate) annuli
An = Vn\ Vn+l. We also denote by Acn° the annulus P™ \ P*+x which
surrounds the critical value of / ^ .
The key ingredient for constructing a quasiconformal homeomorphism
An -> A% is the following theorem of Z. Slodkowski [S, D]:
Theorem 3.1 (Slodkowski). Every holomorphic motion H of Z C C par-

ameterized by the open unit disc D and inducing the inclusion map on {0} x Z
extends to a holomorphic motion / i : i x C - » D x C . Moreover, for every
parameter c € D, the map hc: C —> C is a quasiconformal homeomorphism
whose dilatation constant Kc is bounded by j^rb-
Applying this theorem, we can extend each hn to a holomorphic motion

hn of C parameterized by Vn. Then, as in Lemma 2.6, we consider the map
An -> An° which assigns to each parameter c the point (hn)~l(c). This
map is a quasiconformal homeomorphism (see the proof of Proposition 3.2
below). However, without further care, its dilatation will depend on n. We
will solve this problem by restricting to the sequence of non-degenerate annuli
,4£° described at the end of Section 1, whose modulus sum is infinite and for
which / n '~"° induces an unramified covering map from A^ to A^.
3.2 Comparing the moduli

The Main Theorem is now a straightforward consequence of the following:
Proposition 3.2. There exists a constant K > 1 such that, for every integer
n 6 {nt, i > 0},
^ m o d ^ < mod A < KmodA* .

K
Proof. Let n 6 {n;, i > 0} and dn be the degree of the covering map
yn~n0. £co _+ J4«_ To g e t t n e required inequality, it suffices to construct a
quasiconformal homeomorphism Hn: -4 n ~* A^ whose dilatation is bounded
independent of n.
For every c g Vn and m € {no, n}, let A^, denote the annulus D^ \ D^+x •
Then the map /"~ n ° induces an unramified covering map of degree dn from
An to Ano. Indeed, Lemma 2.5 and the definition of D^, D^+l show that the
128 P. Roesch
diagram
is commutative, so / " "° maps Acn onto A£o. Lemma 2.3 ensures that no crit-
ical point of /"~ n ° can enter Acn while c remains in Vn, so the map /"~ n ° | ^
has no ramification points and its degree is dn.
Now Slodkowski's Theorem gives a holomorphic motion
which extends /&„„ and consists of /iTc-quasiconformal homeomorphisms h^0.

For every c € Vn the homeomorphism /i£0: Aô —>• J4£0 lifts—via the holo-
morphic covering maps /"0~"° and /"~"°—to a homeomorphism
which is again ifc-quasiconformal, coincides with hcn on dP£°+l, and depends

holomorphically on c—because f£-~no o h^ = h^ o /^~"°. Hence the map
is a holomorphic motion.
We now recall that, according to Corollary 2.7 and Lemma 2.6, a parame-
ter c is in An U dVn+i if and only if the critical value c belongs to Acn U dD^+1,
or equivalently if and only if (ft^)~1(c) lies in ^ U ^ P ^ j . Let /fn be the map
On 9 P n + i , the map Hn coincides with H and hence is a homeomorphism onto

dP?+l (Lemma 2.6).
Assertion. The map Hn is a quasiconformal homeomorphism with dilatation
Kn = sup{Xc, ceVn} < swp{Kc, c E Vno+1} = K < +00.

We prove this assertion below assuming that all maps are sufficiently
smooth. The general case can be derived by approximation arguments, us-
ing compactness of quasiconformal homeomorphisms (see [DH2, Lemma IV.3]
and [LV, Theorem 2.2 in Section VI.2]). Note that the proof of Theorem 3.1
(in [D]) yields an explicit approximation of class C°° for the extended holo-
morphic motion.
Consider the Beltrami coefficient fi = dHn/dHn. Differentiating the rela-
tion hcn o Hn(c) = c, we get the identity
dhcn{Hn{c)) dHn(c) + dhl{Hn{c)) dHn{c) = 0
which shows that
dK{Hn{c))
The Ahlfors-Bers Theorem then yields a A"n-quasiconformal homeomorphism
<p: An —> <f>{An) satisfying d<f> — ndcj>. By a straightforward calculation of
d(Hn o 0~'), the map 4>n = Hn o <f>~1 is holomorphic. One can easily see that
Hn is a proper map, so $„ is a holomorphic ramified covering map from (j>{An)
to An°0. By Lemma 2.8 and the Riemann-Hurwitz formula, $„ has no ramifi-
cation points, so its degree is given by the degree of H on the inner boundary
component of An. Using the fact that Hn induces a homeomorphism from
dVn+i to dP^v we conclude that $„ is a conformal homeomorphism. The
result follows. •
In fact the above proof gives the following more precise result:
Corollary 3.3. For each n € {ni, i > 0}, there exists a number Kn > 1
such that
-T7- mod A^ < mod An <
and Kn —> 1 as n tends to infinity.

Proof. Since the sum Xî>o m o ^^S *s infinite, Proposition 3.2 shows that
^ i > 0 mod.4 n , is also infinite, so that the intersection of the para-puzzle pieces
Vn reduces to {c 0 }. On the other hand, the preceding proof bounds the
dilatation of Hn by Kn = sup{Kc, c € Vn}- To see that the sequence Kn
tends to 1, choose a conformal representation \ : ^n 0 ~~> ®> which maps CQ
to 0. Then Slodkowski's Theorem 3.1 asserts that Kc < ^ j ^ [ . This implies
the result since the pieces Vn shrink to CQ. D
To conclude this paper, we characterize the parameters rays 1Z(t) which
converge to a given Yoccoz parameter c.
Corollary 3.4. Let c € M be a parameter satisfying the hypotheses of the
Main Theorem. The angles t 6 K/Z for which the parameter ray H(t) con-
verges to c are exactly those for which the dynamical ray Rc(t) converges to
the critical value c of fc.
130 P. Roesch
Proof. Observe that if a ray Tl(t) or Rc(t) enters some (para-) puzzle piece of
depth n then all its points of potential less than 2~ n lie in that piece. Now
Lemma 2.6 shows that lZ{t) enters Vn(c) if and only if Rc{t) enters the piece
P£ containing the critical value of fc. Since f]Vn(c) = f]P£ = {c}, the result
follows. •
References
[A] AHLFORS L. — Lectures on Quasiconformal Mappings, Wadsworth &
Brook/Cole, Advanced Books & Software, Monterey 1987.
[B] BRANNER B. — Puzzles and para-puzzles of quadratic and cubic poly-

nomials, Proc. of Symp. in Appl. Math. 49 (1994), 31-69.
[D] DoUADY A. — Prolongement de mouvements holomorphes [d'apres

Slodkowski et autres], Seminaire Bourbaki n° 775, Asterisque 227
(1995), 7-20.
[DH1] DOUADY A., HUBBARD J. H. — Etude dynamique des polynomes com-

plexes I & II, Publ. math. d'Orsay 1984.
[DH2] DOUADY A., HUBBARD J. H. — On the dynamics of polynomial-like

mappings, Ann. Sci. Ec. Norm. Sup. 18 (1985), 287-343.
[GM] GOLDBERG L. Ft., MlLNOR J. — Fixed points of polynomial maps, part

II: fixed point portraits, Ann. Sci. Ec. Norm. Sup. 26 (1993), 51-98.
[H] HUBBARD J. H. — Local connectivity of Julia sets and bifurcation

loci: three theorems of J.-C. Yoccoz, in Topological Methods in Modern
Mathematics, 467-511, Goldberg and Phillips eds, Publish or Perish
1993.
[LV] LEHTO O., VlRTANEN K. I. — Quasiconformal Mappings in the Plane,

Springer- Verlag, 1973.
[Ml] MlLNOR J. — Dynamics in One Complex Variable: Introductory Lec-

tures, IMS Preprint, SUNY Stony Brook 1990.
[M2] MlLNOR J. — Local connectivity of Julia Sets: Expository lectures, in

this volume.
[R] ROESCH P. — Topologie locale des methodes de Newton cubiques,

these, ENS de Lyon 1997.
[S] SLODKOWSKI Z. — Extensions of holomorphic motions, Ann. Scuola

Norm. Sup. Cl. Sci. 22 (1995), 185-210.
Pascale ROESCH
Universite des Sciences et Technologies de Lille
UFR de Mathematiques
59655 Villeneuve d'Ascq CEDEX, FRANCE
email:roesch@agat. univ-lillel.fr
Local properties of the Mandelbrot
set at parabolic points
Tan Lei
Abstract
We formulate the technique of parabolic implosion into an easy-to-
use result: Orbit correspondence, and apply it to show that for CQ a
primitive parabolic point, the Mandelbrot set M outside the wake of
Co is locally connected at CQ. This, combined with known results inside
the wake, shows that M is locally connected at CQ. The appendices
contain sketches of relative results and their proofs.
1 Introduction
Denote by Qc the map z 1-4 z2 + c. The Mandelbrot set M is defined to be
the set of c such that limn..>oo Q"(0) /> 00.
We say that CQ is a primitive (resp. non-primitive) parabolic point if QCo
has a periodic point of multiplier 1 (resp. e2m6 with S e Q \ Z ) .
For precise definition of the following notations, see §3.2. Denote by tpM :
C \ D ->• C \ M the conformal isomorphism fixing 00 and tangent to the
identity there. If two M-rays RM{9±) land at the same point c0 we denote
by wake{6±) the component of C s RM(0+) U RM{9~) U {CO} which does not
contain 0. Denote by R^ (9) the external ray of QCQ of angle 9.
Theorem 1.1 Let co ^ 1/4 be a primitive parabolic point. More precisely

Qco has a k-periodic point of multiplier 1, with k > 1.
a) (landing property) For the two angles 9* for which it C o (^ i ) land at a com-
mon parabolic point z$ and are adjacent to the Fatou component containing
c0, the M-rays RM^) land at c 0 .
b) (parametrisation) A germ of RM(9±) can be written as \Jn>no cJ([0,1]) for
some no > 0 satisfying: Cn injective, c^ 4l (0) = c*(l), and, forT(z) = z1 and
some C, C > 0,
V W °T*o^(c±(i)) = <£_,(*) and g < |c*(t) - c | < ^ . (1)
c) (transversality) The local solution of Q*{z) — z = 0 can be written as

{c.Q+v2,g±{v)) withg± holomorphic inv, <?±(0) = za, g'±(0) ^ 0. Moreover the
multiplier map X±(v) = (Qô+v2)'(g±(v)) satisfies A±(0) = 1 and A'±(0) ^ 0.
d) (local connectivity) The set M \ wake{9*~) is locally connected at c0.
133
134 Tan Lei
e) (no other rays involved) For any other ray RM(O outside of the wake, the
impression of its corresponding prime end does not contain c0 (in particular
the ray does not land at CQ) and M N ( w o t e ( ^ ) U {co}) is connected.
The last statement says that M has no ghost umbilical cords at c0, in
the terminology of D. S0rensen. We will formulate techniques of parabolic
implosion into a parabolic orbit correspondence result and then apply it to
construct a sequence of puzzle pieces Pn in C x wake{9±), with c0 on the
boundary, such that tpM o Tk o ifiM(dPn) = dPn-\ and diam(P n ) >c 1/n2.
Together with known results inside the wake, see [Hu, Mi, S0] and Ap-
pendix D, we can conclude that M is locally connected at every primitive
parabolic points.
A similar statement for non-primitive parabolic points is included in [DH2],
[Mi] and [S0]. In Appendix A and D we give a short account of these results,
together with interesting consequences.
Our treatment follows very closely the original approach of Douady and
Hubbard ([DH2]), who invented the theory of parabolic implosion while prov-
ing landing properties of M-rays.
J. Hubbard in [Hu, Theorem 14.6] (see also Appendix E in the present
paper) sketched a proof of the local connectivity of M at a primitive parabolic
point, using the theory of Mandelbrot-like families together with the fact that
M is locally connected at the cusp c = 1/4. This proof is less elementary than
the one in the present paper, but has an advantage that it turns the local
connectivity property at a cusp into a semi-direct consequence of the existence
of sub-Mandelbrot-like families (semi-direct means that extra combinatorial
information is needed), and thus can be adopted to more general settings.
We show in Appendix B that the similar statement for repelling orbit
correspondence can be applied to show that M is locally connected at Misi-
urewicz points (i.e. c values such that Qc(0) is strictly preperiodic).
For a more combinatorial approach of landing properties and local con-
nectivity of M at various points, see [Scl] and [Sc2].
Acknowledgements. Discussions with A. Douady and M. Shishikura have
been very fruitful. J. Milnor has helped to largely improve both the global
structure and local details.
2 Orbit correspondence in parabolic implo-

sion
We will use the terminology and results in [Si].
Definition. By a Fatou coordinate, associated with a rational map / on
a region U, we will mean a univalent map $ : U —> C which satisfies the
following condition for every z € U,
f(z) e U <=> $(z) + 1 6 *(£/) ^ *(/(«)) = *(*) + 1 ,
and moreover satisfies the convexity condition that whenever both w and
w + n belong to $(U) the intermediate points w + 1, w + 2 , ... , w+ (n— 1)
must also belong belong to $(£/). Note that if 4> exists / is also univalent in
U.
The set up. Fix r > 0 small. Let A r C C be a bounded connected open set
with 0 on the boundary, U an open neighbourhood of 0 and fs(z) be a family
of holomorphic maps on U satisfying:
1. (s, z) i-> fs(z) is well defined, continuous on A r x U and holomorphic in
z, with/ s (z) = \{s)z + O(z2);
2. (s, z) i-+ fs(z) is holomorphic on A r x U;
3. /0"(0) ^ 0 ; _
4. Set o(s) - (X(s) - l)/(27ri). Then a is continuous on A r with cr(O) = 0,
and a maps AT univalently onto {z | \z\ < r, | argz| < TT/4}.
R e m a r k . In practice fs{z) and a(s) are often defined in a larger sector
neighbourhood of s = 0 than A r . But we are only interested in the sub-
sector of parameters which is mapped by cr(s) bijectively onto {2 | \z\ <
r, I argz| < TT/4}. Here the width TT/4 is not really relevant, it can be any
angle in the interval (O,TT).
Theorem 2.1 (Douady-Lavaurs-Shishikura): If r is sufficiently small,

then there exist the fallowings (see Figure 1):
For s = 0 there are open Jordan domains fi-,o and tt+$ which are respec-
tively an attracting and a repelling petal for fo, satisfying £}.-# n fi+io = {0},
with associated Fatou coordinates $>±tQ : fl^fl ~^ C.
For s £ A r \ {0} there is a Jordan domain 51S containing two fixed points
for fs on its boundary and two Fatou coordinates $ ± s : fls —> C with
(2)
Taking Do to be fi-Ûfi+.o, both the closure fis and the complement C\Qs
depend continuously on s E A r , using the Hausdorff metric for the space
of compact subsets of the Riemann sphere, and the correspondence (s, z) >->•
<t>±s(z) is continuous in both variables for either choice of sign, wherever it
is defined. Moreover $±}S(z) is holomorphic in s for s € A r .
Proof. Let a{s) £ C be defined so that X(s) = e 2 " a < s ' and |3?(a(s))| < 1/2.
This theorem was proved in [SI], Proposition 3.2.2 and Appendix, with -4^
in place of -4-r, and with | arg(a(s))| < 7r/4 + e in place of s E A r . For s E A r
136 Tan Lei
we have | arga(s)| ss | arg<r(s)| < | . Using the Taylor series expansion of ex

one can see easily that the difference -^h? — -^H is continuous on A r . Since the
Fatou coordinates are defined up to addition of constants (relative to 2), one
can add the above difference to 3>_5 in Shishikura's normalisation without
alerting the continuous dependence on (s,z). Therefore (2) holds together
with the rest part of the theorem. •
Using this result, we can restate and prove the "Tour de Valse" theorem
of Douady and Sentenac ([DH2], expose n°XI) as follows.
Figure 1 : Parabolic orbit correspondence
Proposition 2.2 (Parabolic orbit correspondence). Let X Cfi_,oand

Y C fi+]o be compact sets. If r is sufficiently small, then given continuous
functions x : A r x [0,1] —> X and y : A r x [0,1] —> Y, holomorphic in s for
s G A r , the equations
f?{x{s,t))=y{s,t) and + 71- (3)
have at least one common solution s — sn(t) in A r , which depends continu-

ously on t, provided that n is sufficiently large. Moreover
1
lim 71 - r-rr = Co(t) and lim 1 - — 0 as n -> 00 (4)
n-too O(Sn{t)) no-(sn(t))
where Co(t) is finite and continuous on t, and the convergences are uniform
on t (see Figure 1).
Before giving the proof we give a quick application showing that the Julia
sets move discontinuously:
Application I. Let fs(z) — (l + s)z + z2. For A r a sector neighbourhood of 0

containing the direction of iR the family satisfies all the required conditions.
The Julia set J(/o) of fQ is a cauliflower (see the left picture in Figure 1 of
[Si], in this volume). Let x be any point in fL.,0-
So x £ J{fo)- We will show that there is a sequence sn € A r converging to

0 such that x 6 J{fsJ f° r all n - It follows that for any converging subsequence
J(/ Sn J, lim J(/ Sn J ?M/o).
In order to find sn, we define at first an appropriate function y(s). On
J(/o) there is a repelling periodic cycle of period two. Let y(0) be a preimage
of this cycle such that j/(0) E fi+,o- For any s near 0, there is a unique
y(s) which is holomorphic on s and is mapped by some iterates of fs onto a
repelling periodic cycle of period two. Clearly y(s) € </(/,).
Now we can apply Proposition 2.2 for x(s) = x to conclude that there is
a sequence sn € A r , sn -> 0, such that f"n(x) = y(s n ). As a consequence
x e «/(/ s J for every large n.
Proof of Proposition 2.2. First note that we only need to find a solution for the
right equation in (3) , since combining it with (2) we get $+<s(x(s,t)) + n =
<&+iS{y(s,t)). By definition of Fatou coordinates we have $+)S(f"(x(s,t))) =
$ + i S (x(s,t))+n = $4 iS (y(s,£)). Since $ + i S isunivalent, we have f?(x(s,t)) =
y(s, <). This is the left equation in (3) .
To simplify the situation, assume at first x(s, t) = x(s) and y(s, t) = y(s)
(i.e. they are independent oft). Set x(s) = $_,,(ar(s)) and y(s) = $ +)S (j/(s)).
Then x(s) and y(s) are continuous in A r . For n large, we are seeking for
solutions of the following:
i(s) + n- -^rr = y{s) . (5)
Let x(s) — y(s) = Co + -H(s), with Co = x(0) — y(0), H(s) continuous in

A r and H(0) = 0. We may assume r is small so that |//(S)IA7 ^ V^-
Set F(s) = C0 + n - ^ y = i - ^ with zn = 1/(CO + n), and G(s) =
F(s) + H{s). Then (5) becomes G(s) = 0. We will at first find a solution
for F(s) — 0 and then use it to approximate a solution of G — 0.
For n large enough,
zn £ D{zn, |z n | 2 ) C {z | \z\ < r, | arg(z)| < j} .
Let D — o~x{D(zn, \zn\2)) C A r . By univalence of a, the set D is a closed

topological disc. Moreover there is a unique s'n £ D such that cr(s'n) = zn, i.e.
F{s'n) = 0. Now
1 _ |g(s) - zn
\F(s) \dD = >\ and
zn a{s) dD zn(j(s)
Therefore \F-G\ao < \F\dD- By Rouche's theorem we conclude that there is

a unique sn G D such that G{sn) = 0. (Here is another argument showing the
138 Tan Lei
existence of sn: We have |arg(F(s))-arg(G(s))| s G a D < TT/4. Since arg(F)| a D

is of degree one so is arg(G)|ao- Now G must have at least one zero in D for
otherwise arg(G(s)) would extend to a continuous, homotopically non-trivial
map from D to R/27rZ. This is impossible.)
Therefore sn is a common solution of the equations in (5) and (3) .
Furthermore, by definition of Co and (5) ,
lim n ?—- = Co .
This is part of (4) , the other part of (4) easily follows.

Finally assume that x(s,t) and y(s,t) do depend on t. Set as above
x{s,t) = $->s(x(s,t)) and y(s,t) = $+iS(y(s,t)). Then x(s,t) -_y(s,t) -
C0{t) + H{s, t), with C0(t) = x(0, t) - j/(0, t), H(s, t) continuous on Ar x [0,1]
and H(Q, t) = 0. By uniform continuity of H(s, t) we may choose r small such
that \H(s,t)\ < 1/4 on A r x [0,1]. The same argument above would find a
unique sn(t) as zero of G{s, t). This implies that sn(t) is continuous in t. •
3 Proof of Theorem 1.1

In this section let (co, z0) be such that ZQ is a fc-periodic point (k > 1) of
QCo with multiplier 1 and it is on the boundary of the Fatou component of
QCo containing c0. We would like to find an appropriate family fs(z) and two
appropriate functions x(s,t) and y(s,t) in order to apply Proposition 2.2.
3.1 The family fs relative to w

We will at first define holomorphic functions g±(v),a±(v) for v = sjc - c0,
define a holomorphic map I(c), choose a vector w with certain properties and
then define a family of maps fs relative to w satisfying the set up of §2.
Lemma 3.1 Set F{c,z) = Qkc(z) - z. Then (QhJ"(zo) ? 0 and F{c,z) = 0

has two local solutions which can be written as (co + v2,g±{v)) with g±{v)
holomorphic.
Proof. By assumption Q^ is the first return map about z0 with derivative 1

at z0. It follows that each parabolic basin attached to the point z0 is fixed
by Q^. In other words, different parabolic basins attached to zo belong to
different orbits under the iteration of Q^. On the other hand, Q^ has only
one critical point so there can be at most one periodic cycle of parabolic
basins. It follows that there is exactly one such basin attached to ZQ. As a
consequence (<9^,)"(zo) ¥^ 0 (by local theory of Fatou flowers).
Now consider F(c, z). We have

dF d2F
F{c0, Z0) = 0, — = 0, #0.
By Weierstrass preparation theorem the local solutions of F(c, z) = 0 coincide

with the solution of
{z - zo)2 + (c - co) • A{c - co) • {z - zo) + (c - co) • B{c - c0) = 0 (6)

where A(*),B(*) are holomorphic functions. Moreover the analytic set {c |
F(c, z) has multiple solutions} can not have dimension one, so it has dimen-
sion zero, hence is equal to the singleton {co} in a neighbourhood of Co.
Now replace c — Co by v2 in (6) , we get
{z - zo)2 + v2A(v2) • {z - zo) + v2B{v2) = 0 .

As the discriminant is not constantly 0, it is in the form v2mH(v2) with
H(0) ^ 0. Therefore s/H(v2) is a locally well defined holomorphic function.
So the solutions of (6) (and of F(c, z) — 0) are the two holomorphic functions
9±(v) = -v2A{v2)±vmyjH{v2). ( We will show later that <4(0) / 0 therefore
771=1.) •
Set
a±(v) =
Lemma 3.2 For c = Co + v2 set I(c) — a+{v)a^{v). The correspondence

c •-> I(c) is holomorphic throughout a neighbourhood of CQ with I{co) = 0.
Furthermore, given a direction w such that I(co + hw)/\I(co + hw)\ —> —1
as h \ 0, and a small number r > 0, there is an open connected set A r
containing {hw, h > 0 small} and with 0 on the boundary, such that a+
maps a choice A of \fK~r univalently onto {z \ \z\ < r, \ arg(z)| < 7r/4}, and
maps {h^/w, h > 0 small} onto an arc tangent to R + at 0 (see Figure 2).
Proof. Since z — Q^+îz) = 0 has only two solutions g±{v), we have (by
residue theorem)
I _ 2iri 2TTZ
/
= +
,z-,oMQkco+v2(z)-z z
(Qkco+v>)'(g+(v)) - l (Qk0+v>)'(9-(v)) - 1
1 1
a+{v) a4v) '
But the left hand side is holomorphic on v and has finite value at v = 0. So
lim — — H — is finite. (7)
v->o a+(v) <y-\v)
140 Tan Lei
Set a±(v) = r ± (v)e i l ? ± w . Then (7) gives 77-(u) - r)+(v) -> TT as u -» 0

(exchange <7+ and #_ if necessary).
Now we turn to I(c). As a symmetric function of the roots of (6) , I(c)
is a function of the coefficients of (6) . So c H-> I(C) is holomorphic in a
neighbourhood of CQ. Clearly /(c 0 ) = 0. It has then a power series expansion
I(c0 + s) = as" H with v > 1. Thus there are v distinct directions w such
that /(co + hw)/\I(co + hw)\ ->• - 1 as /i \ 0.
Figure 2 : Finding «/,. A and A r , in the case a'+(0) ^ 0
Fix a choice of w and a choice of ^Jw. Let v = v(h) — h • ^Jw. Then

I(c0 + v2) = cT+(v)a_(v) and ^ = e^+W+'-W). So 77+(u) + 77_(v) ->• TT as
/i\0.
Combining the two equations on r)±(v) we conclude that, as v = hy/w,
lim/,^0 V+(v) exists and is equal to 0. In other words, a+{v) maps {h^/w, h >
0 small} onto an arc tangent to K+ at 0. (Similarly <r_(v) maps it onto an
arc tangent to R~ at 0.)
As a holomorphic function fixing 0, <?+{v) = b(v)p in a neighbourhood of
0, with b univalent, 6(0) = 0 and p > 0 some integer. There is therefore a
(almost) sector neighbourhood A of 0 containing the direction of ^Jw which
is mapped by cr+(v) univalently onto {z | \z\ < r, | arg(z)| < TT/4}. Since A
has an opening angle close to ir/(4p) < TT/4, the map v2 is univalent on A,
whose image will be the required set A r . •
Definition of fs relative to w. Fix a choice of w such that /(co+/iw)/|/(co+
hw)\ —> — 1 as h \ 0. Lemma 3.2 provides two sets A and A r . For v € A and
s = v2 e A r , set g(s) = g+(v) and define f,(z) to be Qkco+S(z + g(s)) - g(s).
Corollary 3.3 The family fs satisfies the hypothesis in the set up of §2 with
(s,a(s)) parametrised by (v2,a+(v)).
Proof. The conditions 1 and 2 about regularities of fs(z) are easily verified,
with X(s) = f's(0) = (Qkco+J(g(s)) = (Q^+s)'(g+(v)). So o{s) := ^ 1 =
<J+(v). For condition 3, we have /Q'(0) = (Q£,)"(zo) / 0 (Lemma 3.1). Con-
dition 4 follows from Lemma 3.2. •
3.2 The angles 9±, r]^ and the functions x(s,t), y(s,t)
Here we need to understand more about the dynamics of Qco and of Qc for
nearby c. We will at first recall the classic theory of Riemann representations,
external rays, state two results about the dynamics of Qc<), and then use
perturbations to construct x(s,t) and y(s,t).
We consider a Model plane as a complex plane together with the action
of q(= Qo) : z i-> z 2. This plane is used to give (external) coordinates in
the dynamical plane of Qc as well as in the parameter plane in the following
sense:
For c 6 C define <pc from a subset of the model plane onto a subset of
the dynamical plane of Qc such that <pc conjugates q to Qc, fixes oo and is
tangent to the identity there (this is often called the Bottcher coordinates).
It is known that the maximal domain of definition of </?c is C x Xc, with Xc
a bounded star-like compact set containing the closed unit disc. Moreover, if
c e M, the set Xc coincides with the closed unit disc, and if c ^ M, there is
zc £ C \ Xc such that <fc(zc) — c.
Define ifM from {|z| > 1} (a subset of the model plane) onto C \ M (a
subset of the parameter plane) by <PM(ZC) = c. Due to a result of Douady-
Hubbard, and independently Sibony, we know that y>M is a well defined con-
formal isomorphism, fixing oo and tangent to the identity there.
The rays RM{0), Rc(9) are defined to be ^ M ({e' i+2 ' rifl | 0 < /x}) and
i
Pc({etl+2n9 I 0 < //}) respectively. Such a ray is said to land at a point z if
the limit, as ji \ 0, exists and is equal to z. We say also that z has 9 as
external angle.
The great importance of these rays are:
1. they are preserved by the dynamics;
2. when two or more rays land at the same point, they provide a clean cut
of the plane, and, together with equipotentials, provide puzzle pieces which
are candidates of local connected neighbourhoods;
3. they often depend analytically on the parameter.
Concerning our polynomial Qco, we have the following known result:
Lemma 3.4 (Douady-Hubbard) The Julia set of Q^ is locally connected. The

convex hull of the critical orbit forms a topologically finite tree (Hubbard tree)
with the critical value CQ as an extremity. The point ZQ, as the unique parabolic
142 Tan Lei
point on the boundary of the Fatou component U containing c0, has two ex-
ternal angles adjacent to U. These angles are denoted by 9± (with 9± € (0,1)
and 9~ < 8+). The period of them by angle doubling is equal to per(zo).
For a proof, see [DH2], pages 51-52. From this one derives easily
Lemma 3.5 In the same setting, there is a sequence of prerepelling (i.e.

eventually periodic repelling) points zn in the Hubbard tree such that zn —¥ ZQ,
zn has two external angles rj^ with r]~ /• 9", 77+ \ , 9+, as n —> 00, and, for
k the period of z0, <3^(zn) = zn_u 2k9+ = 6+ (mod 1) and 2k9~ - 8~ (mod
I)-
Proof. In the segment [a, CQ] of the Hubbard tree (where a is the fixed point of
Qco with a non-zero rotation number), choose a point x so that in [x, z®] \ {ZQ}
there is no branching points of the tree (this is possible since there are only
finitely many branching points), and no points in the basin of attraction of
the parabolic orbit (this is possible as we have assumed that c0 is primitive).
By expansion property on the Julia set, there is a minimal N such that Q^
restricted to [x, z0] is no more injective. This implies that there is z' 6 [x, z0]
such that Qi~x{z') = 0 and Q^(z') = CQ. By minimality of N, Q?o restricted
to [z',z0] is a homeomorphism onto its image. As Q^{zo) is a point in the
parabolic orbit, distinct from ZQ, and is on the Hubbard tree, by the choice
of x we have [Q^(2o))Zo] 3 [x, z0] D [z',z0]- Therefore
o] D {Z',Z0} .
As Qô reverses the orientation, there is a point Z\ E [z',z0] fixed by Q^, i.e.
z\ is a periodic point. It must be repelling as it is not in the parabolic orbit.
It must have exactly two external rays, as it is in the Hubbard tree, and is not
a branching or extremal point. By expansion again the local inverse branch
g of Q*o mapping ZQ to z0 maps [zx, z0] into itself. Denote by zn the orbit of
Z\ under g. They satisfy the desired property of the Lemma. See Figure 3. •
Figure 3 : The segment [Q^(z0), c0] in the Hubbard tree of Q^

A proof of a similar statement in the case that 0 is preperiodic for Qc can
be found in [T], Lemma 2.3(2), and in the case 0 is periodic for Qc can be
found in [Sc2], Lemma 5.1.
In order to define an appropriate y(s, t), we define at first ro(t) for t € [0,1]
in the log-model plane as follows (we have chosen deliberately log e2m9+ to
be 2m'9+ - 2m for the purpose to give a similar picture as in the dynamical

plane and in the parameter plane, Figures 4, 5 and 6): Fix A > 0 small and
ri\ large, the map t >-» Fo(t) should be piecewise linear and continuous on
[0,1/2) and (1/2,1], with
Fo(0) = A/2* + 2md+ - 2TTI , F 0 (l/6) ft + 2m'0+ - 2-Ki

r o (i/3) = A + 271^77^ — 27TI , Fo(2 ) 271-177+ - 2 ? n ,
F o (2/3)
F o (5/6) - r o (i) = A/2* +
r o ([| ,1])
2-KiO-
Todf, ID
c
To([il )
A/, k A
2
mi ID
ID
2xi6 + - 2 r r i
r0([o .Jl)
Figure 4 : The curve Fo (t) and its direction of parametrisation
We then define y(O,t) = <pco(er°^). Although F o has a discontinuity at

t = 1/2, by the choice of 77^, the map y(0, t) is continuous everywhere and
parametrises a Jordan arc.
The function y(0, t) is stable in a neighbourhood of CQ (cf. [DH1], last
chapter), in other words, y(s,t) = <pco+s<p^l{y(0,t)) = ^ co+s (e r °W) is well
defined and is an analytic continuation of y(O,t). One may choose jx small
enough and r^ large enough so that y(s, t) belongs to a compact set Y in Vt+fi
forall(s,t) € A ^ x [0,1].
Define x(s,t) = x(s) to be simply <2co+s(co + «), where I is the minimal
integer such that x(0) € Cl-fi- We may choose a small enough neighbourhood
of co so that x(s) belongs to a compact set X in X7_o.
We now define Fn(t) for n > 0 inductively as follows: choose 9 e (9~, 0 4 ).
Let r be the map from [9 - 27ri, 9 ) to its self mapping z to 2z (mod 2 T « Z ) .
144 Tan Lei
Then 2fti9~ and 2TTI0+ - 2iri are fixed points of rk. Set
and define r n ([0,1/2)) (resp. r n ((l/2,l])) to be the lift by r~fc of the arc
r n _i([0,1/2)) (resp. EV-i((l/2,1])) with the above initial condition. Note
that
r n (-) = rvi(o), rB(-) = rB_1(i),
lim rn(t) = 2iriri+i+n - 2ni and lim Tn(t) = 27rw7~1+n .
Moreover
U r "([°' «^ =
^ + 27ri9+ - 27ri | 0 < /x < A}
n>0
r
U
n>0
»([f> !]) =
Note also that for q : z H4 Z2 in the model plane we have qk(eTn^) =
()
c0 + s
Figure 5 : The dynamical plane of QCo+s
3.3 Application II of Proposition 2.2 for a fixed w

Fix a choice of c0 + w along which I(c)/\I(c)\ -> —1, as in Lemma 3.2.
We can now apply Proposition 2.2 to the family fs relative to to and the
functions x(s) - g(s) and y(s,t) — g(s) defined in the previous sub-sections.
There exists then no > 0 and, for every n > n0, t 6 [0,1], a point sn(t) € A r
such that fsn{t)(x(sn(t)) - g{sn(t))) = y{sn(t),t) - g(sn(t)). In other words,
Lemma 3.6 We have Cn(t) := c0 + sn(t) — <PM(eril'+n^) for n > n0 and

te [0,1].
Proof. Fix n > no so that sn(t) exists.

We claim at first that x(sn(t)) = VcO+Sn(t)(ernit)) for t e [0,1] (cf. Figure
5). Set Q = Qc+s^t) and <p = ^+Sn{t). Define Vj{t) = Qk^n-^(x(sn(t))) for
j = 0, • • • , n. We want to show yj(t) = <p o exp(Fj(2)) for t E [0,1] and j E
{ 0 , l , - - ' ,n},i.e. that the following diagram commutes for j = 0,1, ••• ,n—1:
Qk
Since Q* is univalent where the Fatou coordinates are defined and maps
both j/i(l/6) and 2/o(0) to yo{l/6), we must have yi(l/6) = yo{O) = <p o
exp(F0(0)) = <poexp(ri(l/6)). So the diagram commutes for t = 1/6, j = 0.
The rest follows easily by lifting and by induction.
Let cn(t) = Co + sn(t). Recall that x(s) — Q^ +S (c o + s). We want to show
now cn(t) = Vc,n(t){eTkl+n(t)) (cf. Figure 5).
One can choose c= co + s close enough to CQ SO that fi5 (given by Theorem
2.1) is disjoint from c, Qc{c), • • • , <3c'~'(c)- This last set is the set of critical
values of Qf. As a consequence, any arc 7 in Q.s from Q\l(x{s)) to x(s)
would lift by Q~kl to an arc starting from x(s) and ending at a point w(c)
which is independent of the choice of 7. We claim now w(c) = c: On one
hand, w(c0) = c0, which is an easy consequence of the pulling back of the
Fatou coordinates $_:o- On the other hand, w(c) is continuous on c, and is
an element of Q~hl(Qf(c)). As a consequence, w(c) = c provided that c is
close enough to c0.
We can show that w(cn(t)) = </3cn(t)(er*'+"'t') using the lifting argument as
above. Therefore cn(t) = y C n (o( e r t ' + n ( t ) )-
Hence cn{t) = ¥> M (e r " +n(0 ) for t e [0,1]. •
3.4 Conclusions
Lemma 3.7 (= Theorem 1.1.a)) The M-ray of angle 0* lands at Co and is
tangent to CQ + w at CQ, for each direction CQ + w along which I(c)/\I(c)\
converges to —1,
Proof. Note that for Cn(t) defined in Lemma 3.6,
7 + := I J e n ([0,1/6]) = ^ M (exp({/i + 2m6+ \ 0 no
146 Tan Lei
is a germ of RM{0+). By Lemma 3.2 and Corollary 3.3, the arc a({hw, h >
0 small}) is tangent to R + at 0. On the other hand, since i/er(s n (t)) -> 1 uni-
formly (by (4) of Proposition 2.2), the arc CT(T_CO7+) = a(U n > 0 «n([0,1/6])
is also tangent to K+ at 0 (where T l ^ denotes the translation z *-+ z — Co).
Therefore CT(T_C07+) and CT({/MU, h > 0 small}) are tangent. So RM{8+)
lands at c0 and is tangent to c0 + u; at Co.
Similarly RM{0~) lands at c0 and is tangent to CQ + w. •
Corollary 3.8 (= Theorem 1.1.c)) The functions g±(v) defined in Lemma

3.1 and their multiplier functions X±(v) satisfy (<7±)'(0) ^ 0 and A'±(0) ^ 0.
Proof. Relying on the fact that <fM is a conformal isomorphism we know that
there is only one M-ray of angle 9+. Combining with Lemma 3.7 we conclude
that there is a unique w such that J(c)/|/(c)| converges to —1 along c0 + w.
It follows that I(c) is univalent.
If o±{v) were holomorphic on c, the function I(c) = o+{v)o-{v) could not
be univalent (since cr±(0) = 0). Therefore none of er±(ti) is holomorphic on c.
It follows that none of g±(v) can be holomorphic on c.
Since I(c) is univalent, I(CQ + v2) as a function of v has a double zero at
0. It follows that each cr±(v) has a simple zero at 0, i.e. o±(0) ^ 0. It follows
easily that A'±(0) / 0, as \±(v) = 1 + 2ma±(v). Finally
A'±(0) = f (
v=0
->fc\/
)'(*) — (co + u2) + ^(
(ceo) u=0
± (0)
so ^ ( 0 ) ± 0.
Corollary 3.9 There exist constants C, C" > 0 independent of n,t such that
C'/n2 < \cn(t) — co| < C/n 2 , /or oH t £ [0,1] and alln>no.
Proof. The curve (s,<r(s)) is parametrised by (v2,a+(v)). By the left limit in

(4) of Proposition 2.2 and the fact that C 0 (i) is continuous on t, we know
that (?{sn{t))/^ -» 1 uniformly. Now a(sn(t)) = a+(vn(t)), where vn(t) =
\/sn(t) € A, and a+ is univalent by the above corollary, so a+(vn(t))/(b •
vn(t)) -> 1 (for some 6 ^ 0 ) uniformly. Therefore b • vn(t)/^ —> 1 uniformly.
Hence
Cn(t) - Cp _ Vn(t)2
l/{bnf l/(bn)2
uniformly. In particular the constants C, C" required in the Corollary exist.
Proof of Theorem l.l.b), d) and e).

b) To get c+([0,1]) and c~([0,1]) for n > no we just need to reparametrise
c n ([0,1/6]) and c n ([5/6,1]) respectively. The rest follows from Lemma 3.6
and Corollary 3.9.
d) By construction and the fact that RM^) land at c0, for n > n 0 , the
set
c
- ( [ ^ ] ) u U <v»([o,^])u |J Ulf.iDufo}
m>n m>n
bounds a Jordan domain (called a puzzle piece, Figure 6). Let us denote
it by Pn. We have diam(P n ) < C/n2 due to the above corollary. Moreover
dPn fl M has only two points CQ and c Tl (l/2). So M D Pn is connected (this is
a general fact, cf. for example Lemma 2.1 of [T]). Therefore M \ wake^)
is locally connected at Co and M \ (wake(6±) U {c0}) is connected.
e) Given any RM{0 outside the wake, there is n such that £ > 77+ or
C < Vn- S° the corresponding impression is contained in (M \ wake(r/^)) U
{c n (l/2)}, hence is disjoint from CQ. •
Co
RM{0+)
Figure 6 : The puzzle piece Pn
4 Application III of Proposition 2.2

Here is a qualitative version and the most common form of parabolic implo-
sion:
Corollary 4.1 In the set up of §2, for any x e fi-,o, y E fi+,o, there are
sn -> 0 and kn —> oo such that f^ converges uniformly on compact sets of
£l-fi to a holomorphic map g with g(x) = y.
148 Tan Lei
Proof. Apply Proposition 2.2 to x(s, t) = x and y(s, t) = y, we find a sequence

sn —> 0 such that /" n (x) = y. We will see that /"Jn-, 0 converges uniformly
on compact sets to a function g satisfying g(x) = y and g o f0 = f0 o g.
Recall from the proof of Theorem 2.1 that for a(s) defined as A(s) = e2ma^s)
with a(s) close to 0, the function ^ - jnyrr = -^ - ^TJT is continuous in
A r , in particular has a finite value at s = 0.
In [S2], 3.4, Shishikura proved that for any sequence sn —> 0 such that
-r—, = kn + /?„ with kn 6 N, kn -» +oo and /?„->• /5 € C, the sequence
/*n"lfJ-,o converges uniformly to a limit function.
Now assume that sn is a solution of (5) . We have
= n + x{sn) - y(sn) + ( )
with x(sn) -y{sn) -> f (0) - y(0) as n -> oo. So the hypothesis of Shishikura's
theorem is satisfied with kn — n. Therefore /"^ \Q_ O converges uniformly to a
limit function g. Clearly g(x) = y, g is holomorphic and g o fQ = fQ o g. •
A Non-primitive case
Assume now that Q^ is a quadratic polynomial with a ^-periodic point of
multiplier e2mp/q with q > 1. In this case a similar theory of parabolic implo-
sion can be applied to show:
Theorem A . I (Douady-Hubbard, [DH2])

a) (landing property) For the two angles 6* for which Rco(9±) land at a com-
mon parabolic point ZQ and are adjacent to the Fatou component containing
Co, the M-rays RM(0*) land at CQ.
b) (parametrisation) A germ of RM(9±) can be written as (J n > n o c*([0,1]) for
some n0 > 0 satisfying: c,, injective, c*+1(0) = c j ( l ) , and, forT(z) = z2 and
some Ci,C2 > 0,
±(t)) = c*_l(t) and ^ < |c*(<) - co\ < ^ . (8)

n n
c) (transversality) The local solution of Q^(z) — z = 0 can be written as
(c, h(c)) with h holomorphic, h(co) = ZQ and /i'(co) ^ 0. Moreover the multi-
plier function A(c) = (Qkc)'(h{c)) satisfies A(co) = e 27rip ^ and A'(c0) / 0.
The inequalities in (1) and (8) can be used to prove a result of the
degree of tangency. See Appendix C.
B Misiurewicz case
A parallel theory for repelling periodic points can be sketched as follows:
Theorem B.I Let fs(z) = \(s)z + 0{z2) be a holomorphic family of maps

such that |A(0)| > 1. Then for s in a neighbourhood of 0, there are neigh-
bourhood Us of 0 and linearising coordinates $ s such that for any z, n with
fi(z) € Us, j = 0,1, • • • , n, we have
and 4>s(z) depends holomorphically on (s, z).
Proposition B.2 (Repelling orbit correspondence). In the same set-

ting, let x(s), y(s) be two holomorphic functions with x(0) = 0, x(s) £ 0
and ?/(0) € Uo \ {0}- Then there is a sequence sn —> 0 for n > N such that
f?n(x{sn)) = y{sn) and fln{x{sn)) E USn for j = 0,1, • • • ,n. Moreover, if
A'(0) 7^ 0, there are constants C\,Ci such that ,A&.n < \sn\ < , A S , n .
Proof. Set x(s) = $s(x(s)), y(s) = <$>s(y(s)). They are holomorphic functions
with x(0) = 0, i(s) =£ 0 and y(0) ^ 0. Given n, we seek for solution of
X(s)nx(s) - y(s) .
With the help of power series expansions one can carry out a similar proof as
the one of Proposition 2.2. •
One can find extensive applications of Proposition B.2 in [DH3] and [Me].
We present here an application to the Mandelbrot set (Parts 1) and 2) are
due to Douady and Hubbard, Parts 3) and 4) are due to Tan Lei):
Theorem B.3 Let c0 be a Misiurewicz point, more precisely there are k > 0,
/ > 0 minimal such that Q^l(0) = <5^(0) and I > 0. Then
1) The angles of M-rays landing at CQ coincide with the angles of Qco-rays
landing at C,Q
3) M is locally connected at CQ.

4) The impression of the prime end of any other M-ray does not contain CQ.
Details of proofs can be found in [T].
C degree
The following illustrates a quick application of the inequalities in Theorems
1.1, A.I, although the statement is not in its sharpest form.
150 Tan Lei
Theorem C.I For Co ^ 1/4 a primitive parabolic point, there is a unique

hyperbolic component W attached to CQ, and for the two rays RM^*) landing
at Co, we have that the degree of tangency at Co of RM{9+) and RM(0~) is 2 +
(i.e. is larger than or equal to 2, but smaller than 2 + n for any fi > 0), as
well as the degree of tangency at CQ of RM{01) and dM, i = + or —. For CQ
a non-primitive parabolic point, there are exactly two hyperbolic components
Wo and W\ attached to CQ and the two rays RM(6±) landing at c$ separate
them, moreover the degree of tangency at CQ of RM(6l) and dWj is 2 + , i = +
or —, j — 0,1, as well as the degree of tangency at CQ of dWo and dW\.
Proof. For the existence of such hyperbolic components see Corollary D.4
below.
Here we only prove the part about degree of tangency, which is due to Tan
Lei.
Assume at first CQ is non-primitive. We need an inequality between hyper-
bolic metric and Euclidean metric: Let U be a simply connected domain in
C. Then for x, y £ U, we have
\y-x\ <e2du{x^-d(x,dU) , (9)

where du denotes the hyperbolic metric of U. For a proof, see Appendix A
of [ST].
Let Wo, W\ be the two hyperbolic components attached to c0. We are going
to show for n large enough, and c* = c*(0) (in the notation of Theorem A.I),
\c+ - CrT+1| < B • d(c+, Wo U Wx), |c- - c-+1\ < B • d(c-,W0 U Wx) ,
(10)
where B is a constant.
To prove this estimate, we need the relationship between c* in (8) and
the help of hyperbolic metric. We will only work on the + case. The other
one is similar. Take a simply connected domain V in the model plane so
that T~kq{V) has a component V contained in V and that the two sets U =
<PM{V), and U' = <PM{V) both contain c+ for n> N. Therefore, for Vn the
component of T-kqn(V) with e2nie+ on the boundary, and Un = ^ M ( K I ) C U,
Now apply (9) we get
\<+N ~C + , I < eM"<<+"-<+"-M> • d(4+N, dU)<B- d(c:+N, dU) .
In particular, since (Wo U Wi) n U = 0,

for any large n. This is (10) .

Assume by contradiction that the 0 + -ray is tangent to dWo with a degree
of tangency at least 2 + fi, n > 0. Then
d{<£, Wo U Wj) < d{c+, Wo) < C • | 4 - c o | 2+ " . (11)
So
k>n k>n
k>n k>n
where B', B" are constants, the first and the last inequalities is due to (8) ,
the second one is trivial, the third one is due to (10) and the fourth one is
due to the assumption (11) . Therefore
B" ~ k"
k>n k>n
This is a contradiction to the fact that ^2 -j^ converges.

This shows that the rays have degree 2 + tangency to each dW0 and dW\.
A similar calculation using (1) would show that in the primitive case, the
two rays RM^) are tangent to each other at Co with degree 2 + tangency.B
D A short account of relative results

(One can find in [Mi], [Scl] and [Sc2] radically different approaches of results
in this appendix).
Here we give a short exposition of a proof for
T h e o r e m D . I For c Misiurewicz or on the boundary of a hyperbolic compo-

nent, M is locally connected at c.
The part concerning Misiurewicz points is already established in Theorem

B.3. We will only deal with the remaining cases.
We will sketch a proof of a stronger result (Theorem D.3) and show how
to deduce from it the above result. We will also give a list of important
corollaries.
For Co / 1/4 a parabolic point, define Ang(co) to be the set of two angles
such that the corresponding dynamical rays for QCo land at the parabolic
point ZQ on the boundary of the Fatou component containing c. and adjacent
to this Fatou component. For eg = 1/4 we set by convention Ang(c0) = {0,1}.
152 Tan Lei
Our starting point is the following combination of Theorems 1.1.a and

A.I.a, proved by parabolic implosion technique in the case CQ ^ 1/4 (in the
case c0 = 1/4 it can be checked by hand):
Theorem D.2 Let CQ be a parabolic point. Denote by 6i: the two angles in
Ang(co). Then the M-rays RM(9±) land at c$.
Now define wake(co) to be the connected component not containing 0 of
Cx(i?M(0+)u/?M(r)u{Co}).
Given any hyperbolic component W of M it is quite easy to see (with
the help to quasi-conformal surgery) that the multiplier p(c) of an attracting
periodic point z(c) for each c £ W realizes a conformal homeomorphism from
W onto the unit disc D, and extends to a homeomorphism of the closure (see
[DH2], pp.108-109). Denote by p~ l (l) the root of W. The parabolic points
on dW are exactly p~l(e2iriQ).
Theorem D.3 Let CQ be a parabolic point with Ang(co) — {0~,Q+}.

a) There is a holomorphic function /3CO(C) defined on wake(c.o) such that
/3C0(c) is the common landing point of Rc(0+) and of Rc(8~).
b) There is a hyperbolic component W{CQ) of M with root CQ and W(co) C
wake(ca). Moreover for the common landing point z0 of 11^(0*),
lira /3[c) = z0 .
cewake{cc),c—>co
c) Structure of sub-wakes ofW(cg) (no ghost limbs):
MDwake(co) \ (J wake(c') = W{c0) . (12)

co}, parabolic
d) M is locally connected at c for any c € dW(co) which is non-parabolic;

MOwake(c0) is locally connected ate®; andM\wake{d) is locally connected
at c' for any parabolic d € dW(co) \
Proof.
a) can be found in [Mi], Theorem 3.1. It is proved by an elementary
argument. A sketch is included in Theorem D.8 below.
b) is due to a parabolic perturbation argument (in the non-imploded di-
rection). See [Mi], section 4.
c) We will follow essentially the route of D. S0rensen ([S0]).
At first we apply a) and b) to every parabolic point on the boundary of
W(c0). Combining with the tuning algorithm (cf. for example [Hal, §7] and
[S0]), we obtain the following: for any parabolic c' € dW(ca) x {c0} and any
c e wake(c'), the common landing point /3c>(c) of Rc(9±(c')) has a constant
and non-trivial combinatorial rotation number p{d). Moreover, if we choose
a converging non-constant sequence c'n of such parabolic points, p(c'n) —• oo.

This allows us to apply the powerful Yoccoz inequality (see for example [Hu]
and [S0]), to show that
diam(M D wake(c'n) ) - > 0 asB4oo. (13)
sub-wake
RM(0+)
Figure 7 : The sub-wake structure of W(CQ)
Two more ingredients are needed to end the proof:

A. The parabolic points are dense in dW(co), moreover, a consequence of the
tuning algorithm gives £e'eaw(« oM co}, parabolic 9+(c>) ~ r (c') = 0+ ~ e~<
B. Every point of dM is accumulated by points in the hyperbolic components
(by a normal family argument, cf. Branner [Br]).
Denote by M' the difference of the left set in (12) to the right set. Assume
it is not empty. Let c € dM' \ W(co). Part A says that there is no room left
for M' to contain hyperbolic components. Therefore by Part B the point c
must be the limit of some cn € wake(c'n), for some non-constant converging
parabolic sequence c'n € dW{c$) x {co}. Applying (13) , we get c 6 dW(c0).
A contradiction.
d) This is a trivial topological consequence of c): Let c 6 dW(c0) be non-
parabolic. Choose dn, c'^ two sequences of parabolic points on dW{co) tending
to c from opposite sides, with 0+(c'n) < 0~(c'^). Then the short arc in dW(co)
connecting c'n and c^ together with the two M-rays RM{0^_(c'n)), RM{0"{c'^))
bounds a simply connected region Un. Let Mn = M D Un. Then {M n } is
a nested sequence of connected neighbourhoods in dM of c. As f]n Mn is
disjoint from any of the sub-wakes, and f]n Mn c M (as M is closed) we
conclude from (12) that C\nMn = {c}.
154 Tan Lei
For c = CQ, the points c'n,d^ can be chosen such that 0^(c'n) /• 6+ and
6-(cj[) \ 0~~, and (/„ the union of two simply connected regions bounded by
the short arc in dW(cQ) connecting c'n and cj( and the four M-rays RM(9+(dn)),
RM(0'(d^)) a n d RM^)- The rest follows.
Similar argument can be carried out for c = d 6 dW(cQ) parabolic. •
Proof of Theorem D.I. Let W be a hyperbolic component with root c0.
Assume at first that c0 is primitive. Theorem D.3.d) applied to wake(c0)
shows that M n wake(c0) is locally connected at Co- Theorem 1.1.d shows
that M \ wake(co) is locally connected at CQ. AS a consequence M is locally
connected at CQ.
Another consequence of this is that W = W(CQ), for W^co) given by
Theorem D.3.b). In other words W(CQ) is the unique hyperbolic component
with root Co. We claim that this is also true whether Co is primitive or not.
Proof. We proceed by induction on the period of W. For period 1, there is
a unique hyperbolic component which is the central cardioid. Its root 1/4
is primitive and the assertion is true. Assume it is true for all hyperbolic
components of period up to k. Assume W is of period k + 1. We want
to show W = Vl^Co). If Co is primitive we are done. Assume Co is non-
primitive. Then by a perturbation argument as in [Mi], section 4 (see also
[Ha2]), one can show that there is a hyperbolic component W\ with root cj,
with Co G dWy \ {c\} and with period less than or equal to k. By induction
Wx = W(ci). HW T£ W(CQ), Theorem D.3.c) applied to wake(co) shows that
Wr\wake(co) — 0, and applied to wake(c\) shows that Wn(M ^wake(c0)) —
0. This is a contradiction. •
Coming back to the proof of Theorem D.I, we may already conclude that
M is locally connected at any c 6 dW which is non-parabolic, by applying
Theorem D.3.d) to W = W{co).
We may assume now that Co is non-primitive. By the above argument
c0 € dW(c\) for some different parabolic point c\. Theorem D.3.d) applied
to wake(co) shows that M D wake(co) is locally connected at c0; and then
applied to wake{c{) shows that M \ wake(co) is locally connected at Co- As
a consequence M is locally connected at Co.
For any d € dW parabolic, it is the root of a hyperbolic component W
by Theorem D.3.b). We can then repeat the same argument above with W
replaced by W, and conclude that M is locally connected at d. •
Let us restate a few important corollaries in the proof:
Corollary D.4 Every parabolic point is the root of a unique hyperbolic com-
ponent. Moreover CQ is on the boundary of precisely two (resp. one) hyperbolic
components if co is non-primitive (resp. primitive).
Corollary D.5 For CQ a parabolic point M \ {c0} has exactly two connected
components.
Corollary D.6 For co a parabolic point and 0 any angle not in Ang(co). The
impression of the prime end corresponding to RM(6) does not contain CQ. In
particular Co is the landing point of precisely two M-rays, i.e. those with angle
in Ang(co).
Note that the same result as Corollary D.6 for Co a Misiurewicz point is
stated in Theorem B.3.
The following important result comes as a corollary: For any 0 e T, define
L(9) to be the landing point of RM{9) if it exists. For any c a Misiurewicz
point, define Ang(c) to be the set of angles such that the corresponding
dynamical ray for Qc lands at c.
Corollary D.7 (Douady-Hubbard-landing-theorem)
1. The map L is well defined on the set of rationals with even denominators
and maps it onto the set of Misiurewicz parameters. Moreover for any c
Misiurewicz L"'(c) = Ang(c), and Ang(c) consist of finitely many rationals
with even denominators.
2. The map L is well defined on the set of rationals with odd denominators
and maps it onto the set of parabolic parameters. Moreover for any c parabolic
Proof. By an elementary argument it is quite easy to see that L is well defined

on the set of rationals and maps it into the set of Misiurewicz and parabolic
parameters (cf. [PR] in this volume or [GM]).
Dynamical properties tell us that Ang(c) consists of non-empty but finitely
many rationals with even denominators for c Misiurewicz; and of two rationals
with odd denominators for c parabolic.
Our theorem B.3 establishes L~l{c) = Ang(c) for c Misiurewicz, and
Corollary D.6 establishes L~l(c) = Ang(c) for c parabolic. •
Remark. One can find a more combinatorial approach of Corollary D.7 in
[Mi] and [Scl]. One hidden difficulty in proving Corollary D.7.1 is that for
9 a rational with even denominator, L(8) must be a Misiurewicz point, or
equivalently can not be a parabolic point. Our proof, as well as the original
proof of Douady-Hubbard, relies on results about parabolic points. However,
there are ways to avoid this, for example one can use Thurston's algorithm
to construct a parameter c such that 9 € Ang(c) (a rather involved method),
or one can apply the pretty elementary method given by Petersen and Ryd
(cf. [PR], in this volume).
The following is a refinement of Theorem D.3.a) and gives a dynamical
description of wake(co):
T h e o r e m D.8 Let CQ be a parabolic point with Ang(co) = {9~,9+}. Then
wake(c0) coincides with the set of c € C such that Rc{8~) and Rc(9+) both
land, and land at the same repelling periodic point.
156 Tan Lei
Proof. The main ideas are the following: At first this is true for c € C \ M,
by a combinatorial study conducted by Milnor ([Mi], Lemmas 2.6, 2.9 and
2.11). An argument of P. Hai'ssinsky ([Hal], Proposition A.I) shows that the
landing points z+(c) of i? c (^ ± ) are holomorphic in c, on any connected open
set where they are both defined, and moreover on this connected open set
either z+(c) = 2-(c) or z+(c) ^ z-(c) for any c. We can then conclude easily
by combining these two arguments. For details, see [Mi], Theorem 3.1. •
The following function is as important as the function Pco(c) denned in
Theorem D.3.a).
Theorem D.9 Let W be a hyperbolic component with root CQ. Let a co (c) ê
an attracting periodic point for c £ W depending holomorphically on c. Then
a^c) extends to a holomorphic map on the entire wake(co), is a periodic
point of Qc with constant period, repelling for c € waA;e(co) v W, and co-
incides, in each sub-wake of W, i.e. wake(c'), c' 6 dW parabolic, with the
holomorphic function Pc'(c) (which is the landing point of Rc(9±(c'))).
Proof. This is basically a corollary of Theorem D.3. By the property of the
multiplier function p(c) and the implicit function theorem, a^c) extends
holomorphically to a neighbourhood of W N {CO} in wake(co). Now fix d G
dW \ {c0} parabolic. Let W be the unique hyperbolic component with root
d (Corollary D.4). Theorem D.3.b) applied to W = W{d) says aCo(c) =
Pc'{c) on W. Since PC'{c) is defined and holomorphic on the entire sub-wake
wake(d), so is ^ ( c ) , and they coincide throughout the subwake.
We need now the help of a third holomorphic function £(c) which is denned
to be a co (ci) for some c t € wake(d) \ M (with d as above), and to be its
analytic continuation throughovit C \ (M L>RM(0)) following the holomorphic
motion of the Cantor Julia set. Clearly cvco(c) = £(c) on a open set. So aCo(c)
extends to C \ (M U RM(Q)) and coincides with £(c) there. In particular
ao(c) extends to wake(co) \ M.
Finally by (12) every point c G wake(co) is either outside of M, or in
one of the sub-wakes, or in W. Therefore the theorem. •
Generally speaking aCo(c), as a periodic point of Qc, can be analytically
continued along any path in C \ {finitely many parabolic points}. But we
may loose tract of the rays landing at a^c), it may become non-repelling, and
we do not know a priori whether these parabolic points present real branch
points or not. Theorem D.9 answers these questions at least in wa/ce(0 ± (c o )).
Remark. In case W is the central cardioid of M, the proof of D.3, D.9 can be
made more elementary and thus lead to a simpler proof of local connectivity
of M at c0 = 1/4. See [GM], Appendix C.
Our final corollary is the following Lavaurs continuity result ([DH2], Chap-
ter XVII: Une propriete de continuite), which is a stronger version of Theorem
D.3.b):
Theorem D.10 (Lavaurs) For 9 a rational angle, denote byj$(c) the landing
point of Rc{9) if it exists. Then 70(c) is continuous o n C \ Uit>i RMQ11®)-
Here the rays RM{V) do not include their landing points. This in particular
says that 7e(c) is continuous at the landing points of these rays.
Proof. We will only treat the case 0 has odd denominator. The other case is
a much easier exercise.
On the open set
C N | J RM(2*9)
79(c) is well denned (Douady-Hubbard) and holomorphic ([Hal], appendix).

So we just need to show that fg(c) is continuous at L{6) = Co, the landing
point of RM(8) (as this will imply that 72i«(c) for I > 1 is continuous at L(9)
and as a consequence j$(c) is continuous at L(2m9) for m > 1).
Let U be a neighbourhood of CQ such that U \ RM{9), U PI wake(c0) and
U \ wake(co) are all simply connected. Set j(c) = j$(c). Denote as before,
by ZQ the parabolic point of Qco on the boundary of the Fatou component of
Co-
Case 1. The point CQ is primitive. In this case the other angle ff in
Ang(co) is not in the form 2l9. So -y{c) is well defined in U x RM(&)- But
y(c) - Pcc(c) f° r c £ Unwake(co) (Theorem D.3.a)) and /3C0(c) - gi{v), i = +
or - , where v = y/c - Co and ^±(^) is defined in Section 3.1 (they are the two
local periodic points perturbed from ZQ). AS U \ RM{9) is simply connected,
y(c) = gi{y/c - Co) on U \ RM{9) and is continuous at Co (as gi{\/c — c0) is).
Case 2. The point Co is non-primitive. This is the case where the two angles
9,9' in Ang(co) are in the same orbit by angle doubling and 7(c) is denned on
U x {RM{0) U RM{9')), and is holomorphic on U \ {RM{9) U RM{9') U {d})
([Hal], appendix).
In U n wake(c0) we have 7(0) = 0co(c) and 7(00) = z0 (Theorem D.3.a)
and b)). So 7 is continuous at CQ on (U C\ wake(co)) U {CQ}.
In [/ N tuofce(co), let Mî the hyperbolic component containing Co on the
boundary. As c e U HWy tending to Co radically, the same perturbative
argument of Milnor ([Mi], section 4) shows that j(c) is repelling for Qc and
tends to z0. Let z(c) be this repelling periodic point. It has analytic contin-
uation along any path in U \ {co} (shrinking U if necessary) and is periodic
of constant period for Qc- It is therefore a well defined holomorphic function
in U N wake(co) as this last set is simply connected. As 7(0) is holomorphic,
we have y(c) = z(c) on (U \ wake(c0)) U {co} and continuous at c0, as z(c)
is. •
158 Tan Lei
E Local connectivity via Mandelbrot-like

families (following Hubbard)
Here we sketch a proof of the local connectivity of M at a primitive parabolic
point using the theory of Mandelbrot-like families and the fact that M is lo-
cally connected at 1/4. We follow essentially the route indicated by Hubbard
([Hu]).
Let c0 be a primitive parabolic point, more precisely Q,^ has a k periodic
point of multiplier 1. In [Hal] in this volume, it is shown that there are
open discs l/0, U\ containing c0, open discs Bc (resp. B'c) whose boundary
undergo a holomorphic motion, such that f = {Q* : B'c —> Bc, c € UQ}
forms a Mandelbrot-like family, with U\ C UQ, U\ is the set of c such that the
critical point of Qkc in B'c does not escape B'c after one iteration, U\ n M is
connected (it is in fact a puzzle piece in the sense of Yoccoz), and, for Mf the
connectedness locus and \ '• Mf -* M the straightening map, Mf C M D UQ
and x is a homeomorphism with x(co) — 1/4.
Let Un to be the set of c £ Uo such that the critical point of Q* in B'c
does not escape 5^ after n iterations. Then Un is again a puzzle piece and
n»fn = Mf.
Denote by W the unique hyperbolic component with c$ as root (Corollary
D.4). Choose dn, d'n two sequences of parabolic points on dW converging to c0
such that 6*{c'n) < 9+ and 0~(c'^) > 9~. Denote by Vn the simply connected
domain containing RM(0+) bounded by the short arc in dW connecting dn
and d'n and the two rays R.M(0+(dn)) and i?M(0~(c^)) (here we need Theorem
A.I for the landing of these rays). Denote by Mn the set x{Vn n Mf).
Set Ln = Vn n Un. Then Ln c L n _i and Ln D dM is a connected neigh-
bourhood (in dM) of Co. We now show that f)nLn =
c0 G f| Ln = (f|L n ) n Mf = f|(LB n Mf) C P ( K n Mf) =
where the first equality is due to Q^ Ln C f]n Un = Mf, the second is trivial,
the set inclusion is because Ln C Vn, the next equality is by definition of Mn
and the fact that x is a homeomorphism, the following is due to the facts
that limn-xx, x( c n) = limn->oo x{c'n) = 1/4 ( ^ X i s continuous) and that M
is locally connected at 1/4 (see [GM] Appendix C for an elementary proof),
and the final inequality is by properties of x-
References
[Br] B. Branner, The Mandelbrot set, Proc. Symp. Appl. Math., AMS 39
(1989), 75-105.
[DH1] A. Douady & J. Hubbard, Etudes dynamique des polynomes complexes,

I, Pub. Math. d'Orsay, 84-02.
[DH2] A. Douady & J. Hubbard, Etudes dynamique des polynomes complexes,

II, Pub. Math. d'Orsay, 85-05.
[DH3] A. Douady & J. Hubbard, On the dynamics of polynomial-like map-

pings, Ann. Sci. ENS, vol. 18, pp. 287-343, 1985.
[GM] L. Goldberg & J. Milnor, Fixed points of polynomial maps, part II,
Ann. Sci. Ens. 26 (1993), 51-98.
[Hal] P. Hai'ssinsky, Modulation dans I 'ensemble de Mandelbort, in this vol-

ume.
[Ha2] P. Haissinsky, Deformation J-equivalente de polynomes geometrique-

ment finis - Pincements de polynomes, ENS Lyon, preprint no. 235,
1998.
[Hu] Hubbard, Local connectivity of Julia sets and bifurcation loci, in Topo-
logical Methods in Modern Mathematics, ed. by L. Goldberg and A.
Phillips, Publish or Perish, Inc. 1993.
[Me] C. McMullen, The Mandelbrot set is universal, in this volume.
[Mi] J. Milnor, Periodic Orbits, External Rays and the Mandelbrot set: An
expository account, Stony Brook IMS preprint #1999/3, Asterisque,
to appear.
[PR] C. Petersen & G. Ryd, Convergence of rational rays in parameter

spaces, in this volume.
[SI] M. Shishikura, Bifurcation of parabolic fixed points, in this volume.
[S2] M. Shishikura, The boundary of the Mandelbrot set has Hausdorff di-
mension two, S.M.F. Asterisque 222, 1994.
[Sci] D. Schleicher, Rational external rays of the Mandelbrot set, Asterisque,

to appear.
[Sc2] D. Schleicher, On fibers and local connectivity of Mandelbrot and multi-

brot sets, Stony Brook IMS preprint #1998/13.
160 Tan Lei
[S0] D. S0rensen, Complex Dynamical Systems: Rays and Non-Local Con-

nectivity, Ph.D. thesis, DTU Denmark, 1994.
[ST] M. Shishikura & Tan L., An alternative proof of Mane's theorem on

non-expanding Julia sets, in this volume.
[T] Tan L, Voisinages connexes des points de Misiurewicz, Ann. Inst.

Fourier, Grenoble, 42, 4, 707-735, 1992.
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF WARWICK, COVEN-

TRY CV4 7AL, UNITED KINGDOM
e-mail: tanlei@maths.warwick.ac.uk
Convergence of rational rays
in parameter spaces
Carsten Lunde Petersen and Gustav Ryd*
Abstract
We give an elementary proof of the landing Theorem for rational
external rays of the Mandelbrot set and related connectedness loci for
the one-parameter families of polynomials {Pc{z) = zd + c} c e C , d > 2.
The proof is quite general and apphes with marginal changes to nu-
merous other families with one free critical point.
1 The setting
Fix an integer d > 2 and consider the family Pc(z) = Pc — zd + c, c 6 C, of
rnonic polynomials of degree d with a degree (d - 1) critical point at the
origin. For each c let Jc denote the Julia set for Pc and define the domain of
attraction to infinity
Ac = Ac(oo) = {ze C|P"(z) —> oo, as n -» oo}.
Let 4>c : fic —> C N D(i?) be a Bottcher coordinate for Pc at oo normalised

by being tangent to the identity at oo. Then 4>~x has a unique maximal radial
extension to a subset (C \ D) c of C \ D. This radial extension terminates
at a point w with \w\ > 1 if and only if (j)~x extends continuously to w and
4>cl(w) is a (pre)critical point for Pc. Thus (C \ B ) C = C \ D, if and only
if c £ Ac. Let Ac = ^ ' ( ( C \ D) c ), then <j>c : Ac —> (C \ D) c is biholomor-
phic and depends holomorphically on the parameter c. That is if z^ € A^
then there exists neighbourhoods U of ZQ and V of c0 such that the map
(2, c) H-> (j>c{z) is analytic on U x V.
Define Md - {c e C|c ^ Ac} and let $ d : C \ Md —> C \ D denote the
unique Riemann map tangent to the identity at infinity, then $d(c) = <f>c(c).
Given 0 e T = R/Z and c G C : The dynamical_ (external) ray Rc(9) of J c is
the analytic arc ^ 1 ( { e x p ( s -I- i2z6)\s > 0} n (C \ D ) C ) . It starts at 00 and
either ends at a precritical point z0 e Un>oFc~n(c) or ends by accumulating
on some subset of the Julia set Jc. The parameter (external) ray R.Md{0) of
Mi is the analytic arc $j'({exp(s -I- I2TTO)\S > 0}). A ray is called a rational
ray if 9 is rational, i.e. 9 € Q/Z. A ray R is said to land or converge, if the
•Research undertaken at Dept. of Mathematics, KTH, 100 44 Stockholm, Sweden.
161
162 Carsten Lunde Petersen and Gustav Ryd
accumulation set R \ R is a singleton subset of J (if it is a dynamical ray)

or Md (if it is a parameter ray).
On the boundary of Md we distinguish two particular and different types of
parameters, parabolic parameters and Misiurewicz parameters. A parameter
c0 is called parabolic, if P^ has a parabolic cycle, i.e., there exists a positive
integer q and a periodic point z0 such that P^(z0) = z0 and (F%Y(zo) = 1. A
parameter c0 is called a Misiurewicz parameter, if the orbit of c0 is finite and
ends in a repelling periodic orbit, i.e. there exist positive integers l,q such
that P ^ ( c o ) = P4(co) and |(P«)'(P4(co))| > 1.
Figure 1: The Mandelbrot set with a periodic (under angle doubling) and a
strictly preperiodic ray.
In this paper we give an elementary proof of the landing Theorem for

rational external rays of Mdi for any integer d > 2:
Theorem 1 (Douady-Hubbard) Given 0 e T with dl6 = dl+q0 mod 1, for

some minimal integers I > 0 and q>l. Then the parameter ray RMA
on a parameter CQ e dMd. Furthermore,
1. if I > 0 then CQ is a Misiurewicz parameter, P^(co) = -^^(co) and the

dynamic ray RCQ{0) land at CQ.
2. ifl = 0 then CQ is a parabolic parameter, the (unique) parabolic orbit

for Pco has one cycle of immediate attracted basins. This cycle contains
both 0 and c and its exact period is q. Moreover the dynamic ray Rc0 (0)
lands on the parabolic point to which CQ is attracted under iteration by

pq
r
c0'
Figure 2: The Julia set for a quadratic polynomial with a period-2 ray landing
at a parabolic fixed point.
This Theorem is well-known, at least for d = 2, since [DH]. The original
proof is however rather involved. There are several other proofs, such as
[HS], which uses iteration on Teichmiiller spaces, [Mil], [Sch] and [K]. In [CG]
this theorem is claimed, but there seems to be a part missing. This paper
has evolved from [Pet] and [Ryd], which both were extending Theorem 1
to different settings using the same techniques, but formulated in terms of
Caratheodory Kernel convergence and hyperbolic distance on the one side
and harmonic functions on the other side. Here we have chosen to formulate
the proof in terms of normal families.
The idea behind our proof for landing of the parameter ray R = RMd(0)
with dl+q0 = dl9 mod 1 is the following simple one: Let Co E M* be an ac-
cumulation point of R. If Co E JCo then ' > 0 a n d P£9(co) — P^fa) a n ( l the
dynamical ray RCo(0) lands at CQ. Thus CQ is a Misiurewicz parameter, by
the snail-Lemma, [DH] (alternatively the repellingness of P1CQ{CQ) is proved
directly, also in [DH] by showing Pco is expanding in the orbifold metric). If
c0 E KCQ \ JCo, then / = 0 and Co is contained in a immediate basin of ex-
act period q for a parabolic point. Thus CQ is a parabolic parameter. The
Douady-Hubbard landing Theorem then easily follows.
Our proof differs from the original proof in [DH] mainly in our direct
approach to convergence and uses only elementary analytical means. Our
164 Caxsten Lunde Petersen and Gustav Ryd
Figure 3: The Julia set for a quadratic Misiurewicz polynomial with a prepe-
riodic ray landing at the critical value Co.
proof has easy generalisations (beyond the plenty of the original proof) to
many other settings.
2 The central tool

We will frequently use gc : C —> R+, the Green's function for Jc with pole
at co. It is the sub-harmonic function, which is harmonic on Ac, coincides
with log |0c(-z)| on Ac and which is identically zero on C \ Ac. Similarly we
use GMA '• C —y M+, the Green's function for M,* with pole at oo. It is the
sub-harmonic function, which coincides with log \&Md\ o n C \ M& and which
is 0 on M^. It satisfies the relation GMd{c) = gc{c) for all c 6 C. Note that
an external ray is a gradient line for the appropriate Green's function.
Our central tool is a sequences \j)n of holomorphic maps, which relates the
dynamics in the filled Julia set, KCo, with the dynamics in Aco. The maps ipn
are obtained as limit maps of sequences of inverse branches of iterates of Pc
with c in the appropriate parameter ray. The underlying idea is the same as
the idea behind parabolic implosion.
Given a parameter c € C, an angle 77 6 T, and a potential level p > gc{c)
we define 'sectors' V(c, r/,p) = {z\p2/3 < gc{z)} s Rc(v + | ) and V'(c, 77,p) as
the sub-sector of V(c, rj,p), which maps univalently onto V(c,r],pdq) by Pq
and which intersects the ray Rc{ri). Given 0 e T with 77 = dl6 = d'+q6 mod 1,
for some minimal / > 0 and q > 1. Let c0 6 dMd be any limit point of RM4{9)
and let {c,j} gN C RMA^) ^ e a sequence converging to c0. Define
Rj = (RCl (9) n {z\gC] (z) > gC] (CJ)}) U {cx>} (2.1)
Then the sequence {Rj}eN is a sequence of compact sets in C. Passing to

a subsequence, if necessary, we can suppose this sequence converges to some
compact subset R C C containing Co and oo. Then R D Ac — Rc<,(9) U {oo},
because the Bottcher coordinates <f>c depends continuously on the parameter
c on any set of the form {z|(7c(z) > p}, where p > gc(c)/d. Moreover c0 e R
and P'+q{R) C P'(R).
Notation 2.1 During the rest of this article we use without further notice
the symbols 9, I, q, rj, Co, Cj, R and Rj for j e N in the sense introduced in
the above paragraph.
Proposition 2.2 Let ZQ 6 RdK^ be arbitrary. There exist w = wzo in

Rciri), a 'sector'V = V20 = V ^ . r / . ^ M ) ^d{ij)n = ipn<20 : V —> tfc)n>0,
a sequence of holomorphic maps such that
1. For all n > 0 : ^n{w) = Pc"0(20) and Vn+i = P^ o xpn.
2. For alln> I : xjjn o Pc« = F « o i on V - V'(c o ,77, P c H ) aiJd in

particuiar
3. One of the following mutually exclusive cases occurs
(a) The ifin are univalent with i/>n(V) Q (Ka x Jco) and for all n, m > 0
such that either n tne extensions still en-
joying the above properties.
Notation 2.3 We shall refer to the above items as Property 1, Property 2,

Property 3a and Property 3b respectively.
Proof : We prove the Proposition only in the extreme case z0 = c0, and
leave the analogous general case to the reader.
We can suppose GMd{cj) = 9CJ(CJ) < d~l. For each j let Nj be given by
1 < dN> • gC}{cj) < dq and PCj'{cj) 6 RCl{v)- By relative compactness, pass-
ing to a subsequence if necessary, we have wf.=P^J (CJ) —> w e Rcoiv) with
1 < 9coH < d".
166 Caxsten Lunde Petersen and Gustav Ryd
Define V — V^^rj^g^w)) and note that V is a simply connected do-

main. We can assume that V does not contain any critical values for PCjJ,
discarding possibly finitely many initial values of j . Also we can assume that
Rj fl {z\gCj{z) > gCj(wj)} C V for all j . For each j define ^0,Cj : V —> C as
the unique branch of PCj J , which maps Wj onto Cj (see Figure 4). Moreover
for each n with 0 < n < Nj define ^n,Cj • V —> C by ^n>Cj. = Pcn o ÔJCJ. .
-N
Figure 4: Image domains ipnyCj(VL) viewed in the Bottcher coordinate at

infinity (we have taken L ^ gCo(w)dq and M to be a multiple of q).
The maps t/>n,Cj are univalent and satisfy for each fixed j :
Vn,0<n<Nr. ^ K ) = I*(c,-) and ^ n+1}C . = PCj. o ^ n (2.2)

Vn, / < n < iV,- - ^ : ^ oF| = /^ o ^c.. (2.3)
V71,171 with 0 < n, m < A/} — M and either n < / or \n — m\ £ qL
j i (2.4)
Where M = M(0) is any integer which satisfies
— < min{distT(rfn0, dTO) |0 < n < m < / + q}.

d
Note that iV^ —> 00 as 7 —>• 00. Moreover for any L > d9 and any n G N the
restrictions of -0njCj to y L = V H {z\gco(z) < L} are uniformly bounded.
We shall obtain the maps ipn as locally uniform limits of the maps ijjn)Cj as
j -» 00. By Montel's Theorem there exists a holomorphic map ^0 ' V —> C
such that, extracting a subsequence if necessary, the maps ipo>Cj converge lo-
cally uniformly to ip0. Define for each n € N a holomorphic map rpn : V —>• C
by tpn = P " o ip0. Then ipnc. ^ t ipn, because Pcn converge locally uniformly
to P ^ as j -> oo. By the Hurwitz Theorem the maps ipn are either univa-
lent or constant. Moreover ipn(V) C Km or equivalently gco = 0 on ipn(V),
because gC} —>• 0 uniformly on t/>o)Cj {V1) for each fixed L.
By locally uniform convergence the properties (2.2) and (2.3) also holds
for the ipn (with P^,), so Properties 1 and 2 follows for ipn. Similarly (2.4)
holds for ipn and ipm if not both ipn and ipm are constants. Moreover if
,/co n tpn(V) ^ 0, then tpn is constant, by the openness of non-constant holo-
morphic maps. By (2.2) if one ipn is constant they all are, so the Property 3
follows.
Suppose z0 ^ Vo,co(Vco) a n d l e t {zj e Rj)jeN ^ e a sequence converging
to z0. Then the ratio gCj(zj)/gCj(cj) diverges to infinity, since if not, then
z0 € ô,co(^co) by the locally uniform convergence of ^0,Cj to ipQiC0. But then
for any p > 0 there exists N € N, such that V? > iV the maps Vn,*, are de-
fined and univalent on the sector K(co,??, p). Thus the domain of ipn%Zo can
be taken to include V = V(CQ, r), 0). q.e.d.
We obtain some immediate corollaries.
Corollary 2.4 For z0 6 R D ATCo we have
&• Vn € N the map ipniZ0 is constant.
Proof : The map Vi = i>i,z0'•ô —^ ^«o ^s either univalent or constant by

Property 3. Moreover xpt{w) = P4( 2 o)' a n d ^( P co( w )) = ^ ' ( ^ o ) by Prop-
erty 1 and Property 2 respectively. As Pîw) ^ w the Corollary follows.
q.e.d.
Corollary 2.5 Let z 6 ( i ? n / ( J \ Jco and suppose Plco{z)

Pc'0(2) belongs to a q-periodic immediate basin of some parabolic point a.
Proof : The map ipt = ipi,z '• Vz —> K^ is univalent by Property 3 and
Corollary 2.4. Hence the image ipt(Vz) is contained in some Q-periodic Fatou
component U. A periodic Fatou component is either a hyperbolic, a parabolic
or a rotation domain. The Fatou domain U can not be hyperbolic, because
hyperbolicity is an open condition. It can not be a rotation domain either,
because any point on ip^R^^)) is non recurrent. q.e.d.
Corollary 2.6 For z0 € R n K^ we have
Proof : If z0 6 J co , then T/>O 20 is constant, because it can not be open. Thus

!%>{*>) = P£"(zo) by Corollary 2.4.
Next assume P^(z0) = P£9(zo) and ZQ $ JCQ. Then P^{zo) is a q peri-
odic Siegel point, because the alternative, attracting, is an open condition.
However the set of centres of Siegel disks is discrete and R is a contin-
uum. Thus we get a contradiction with Corollary 2.5. More detailed let
U denote the Siegel disk with centre Pc'o(zo) and let z' 6 it be a point with
Plco[z') € U \ {Pc'0(z0)}- Then P^{z') # P^{z'), because a Siegel disk con-
tains precisely one periodic point. But then U is a parabolic domain and not
a Siegel domain, by Corollary 2.5, hence P'0{z0) ^ P^q(zQ). This contradic-
tion completes the proof. q.e.d.
For w 6 Rc{v) with gc(c) < d • gc{u>) define a real analytic parametrisation
Vc<w : R+ — • Rc(r]) by TeiV1{t) = 4>;l(exp{gc{w) • d"1 + I2TTT/)) of the sub-arc
of the dynamic ray RC(T]) from w to oo. The map Tcw satisfies
Moreover Vcw has a unique real analytic extension to the interval ] — t o >°c[,
where t0 — -^—j log(gc(c)/(d • gc(w))). In particular for c € Md the parametri-
sation i\u, extends to VC:W : R —> Rc(v), a parametrisation of RC(TJ). In the
light of Corollary 2.6 we define arcs:
Definition 2.7 For ZQ e {R n K^) \ J^ let F 2o : R+ —> (KCo \ J^) be the

real analytic arc given by Vzo = ip0tZO o rco>Wto. Then for all t > 0 (t e R if
By Corollary 2.5 the arc P^{rzo) is contained in a g-periodic immediate basin

of a parabolic point a'Zo and F i 0 (i) —>• aZB, where a^0 = P c ' 0 (a 20 ) = P£< {<*,<>).
(-+OO
3 The Douady-Hubbard Landing Theorem

Proposition 3.1 If CQ 6 JC(t, then I > 0 and CQ is a Misiurewicz parameter
with PCO(CQ) = P^q(c0). Moreover the dynamic ray /?co(0) iands on c 0 .
Proof : By Corollary 2.6 we have P^(co) = P£q(c0). If / = 0 then c0 is
q-periodic, but then also the critical point 0 is ^-periodic, because it is the
only preimage of c0. This contradicts however the fact that CQ e dMd, thus
l>0.
Let us next prove that the dynamical ray Rco(0) C R lands on co. however
this follows from Corollary 2.6, because RîP) OJ^CRDJco and the set of
points z0 with P1C0{ZQ) = P£q(zo) is discrete. Hence it follows from the snail
Lemma, [DH, Exp. VIII, p. 71, le. P], that P^fo) is repelling or parabolic,
however it can not be parabolic, because the unique critical point is prepe-
riodic. Alternatively by [DH, Exp. V.4 p. 40-42, le. P] f*,(co) is repelling.
q.e.d.
Proposition 3.2 Suppose CQ 6 U C K^, where U is some Fatou component.

Then c0 is a parabolic parameter, U is an immediate parabolic basin of exact
period q and the preperiod I = 0.
Proof : Define Ut = PlC0(U), then Ui is a g-periodic immediate parabolic
basin, by Corollary 2.6 and Corollary 2.5. But then so is U, because the
cycle of such immediate basins should contain a critical point and value. Let
the (exact) period of U be p. For n > 0 define 7 n = /^J(r c o ). Then j p 6 U.
Let $ : U —• C be a Fatou coordinate for P^ on U normalised by $(c 0 ) = 0
and $(i*jj,(c<))) = 1, i.e. $ o / ^ = l + $ . The map $ has a univalent inverse
branch
* : EL, - {z = x + iy\y > - 1 } —> tt
with tf(0) = co and with f£(ft) C fi.
There exists n > 0 such that 7pn)7p(n+i) Q Q> because 7m([0,1]) is com-
pact and f%(jm(t)) = jm(t + 1) = 7m+,(<) for all m > / and t > 0. Define
7Pn = $(7pn) and 7P(n+i)) = $(7P(n+i)), so that j ^ t + 1) = %„(*) + g/p and
similarly for 7p(n+i)- The map z -> exp(2irizp/q) maps these two arcs onto
two simple closed curves, which are rotations (around the origin) of each
other. But the later intersect, being rotations of each other, and so do j p n
and 7 p ( n+ i). Since 7 pn and 7p(n+i) are contained in Q, they also intersect. But
then p is a multiple of q by Property 3a and thus p = q.
Next we prove that / = 0. Choose n with j n q C Q so that 7( n + m ), C fi
for all m > 0. Then 7, n = $(7n<7) satisfies the hypothesis of the isotopy
Lemma 3.3 below. Let F n be the isotopy, whose existence is granted by
Lemma 3.3. Increasing n, if necessary, we can assume F n (R + x [0,1]) C H_i,
because 7,(n+i) = 1 + 7«« and f n (« + 1, s) = 1 + fn(t, s). Define an isotopy
F n : R,. x [0,1] —• Q by Tn = * o f. Then the image r B (K+ x [0,1]) does
not contain any of the critical values for P^. Thus there exists a unique lift
F o of F n to Fc"9 with ro(O,0) = CQ. But then by construction
ro(n,0) = r o ( n , l ) = 7 o ( n ) = 7 n , ( 0 ) .
which contradicts Property 3a, when I > 0. q.e.d.
Note that the above isotopy argument is already given in [DH, Exp. XII.2
p. 35-36 2e. P]
Lemma 3.3 Let ji : R+ —^ C be an arc with 7i(0) = n and such that for
all t > 0 7i(t + 1) = 7i(t) + 1. Then there exists r : R+. x [0,1] —• C, an
i s o t o p y of arcs w i t h F ( t , 1) = ^ ( t ) , T ( t , 0 ) = n + t : = -yo{t), F ( 0 , [ 0 , 1 ] ) = n
and for allt>0,se [0,1] : T{t + 1, s) = 1 + F(£, s).
Proof : Extend 71 to an arc 71 : R+ —• C by use of the periodicity
71 (t + 1) = 7i(£) + 1. For some r > 0 sufficiently big the vertical translate
7i + ir is disjoint from the arc R, the real axis, and those two arcs bound
a 'horizontal strip B. Let h > 0 be such that there exists a biholomorphic
map <j>: {z = x + M/|0 < y < h) —> B with <j>(z + 1) = 1 + <£(z). Then ^ ex-
tends to a homeomorphism of the closed sets (the boundaries being analytic
arcs) and we can suppose <f> normalised by ^(0) = n. Let w = <f>~x{ir), then
a isotopy F with the required properties is obtained as
{ (1 - 4s)(n + t) + 4s • <j>(t)
(j>{t + 2(s - \)w) - <£(2(s - \)w) + n if \ < s < §

ifO<s<^
71 (4(s - f )t + 4(1 - 3) • ^l{<t>{t + w)-

ir)) if f < s < 1
q.e.d.
The following Lemma implies that even in the periodic case the set R n Jc<>
is a singleton.
Lemma 3.4 Suppose 9 is q periodic and let U be a Fatou component for P^

with Ru = U n R ^ 0. If Ru connects two distinct boundary points a ^ fi of
U, i.e. a,fi& Ru- Then the restriction P^ : U —> U has two distinct critical
points in U.
Proof : The Fatou component U is a g-periodic immediate basin for a
parabolic periodic point, which we can assume to be a, by Corollary 2.5 and
the remark following Definition 2.7. ^
Let <j>: U —> D be a Riemann map, say mapping a to 1. Let L — 4>(Ru)
and define P as the conjugate Blaschke product P = <f> o P^ o <j>~1. Then L
connects the parabolic point 1 for P with a repelling fixed point /? 7^ 1. The
conclusion of the Lemma is then equivalent to: P has at least two distinct
critical points in D. To reduce indexing define for x € L : z = z(x) = <j>~1 (x).
Choose a round disk w around /? such that P is univalent on ui and
w C P(u). We can choose a possibly smaller round disk u ' c w around (3
such tha£ for any x euj'nL: z(x) £ F ^ and <f>(rz([0,1])) C w. The first
because F^, only intersects J^ in one point. And the second, because for
any z € (R D K^) \ (JCo U Fc,,) the map ipo,z extends univalently to the sec-
tor V — V(co,r),0) and î(F 2 ([0,1])) is contained in the compact subset
{w € i?co(?7)|l < g^w) < d2"} of V. Hence by the Koebe Distortion Theo-
rem for univalent maps the diameter of </>(Fz([0,1])) tends to 0 as the distance
of x to S 1 = 83 tends to zero.
Let x 6 a/ D L be arbitrary and let 7 = $ o Tz. Then 7 : K —> 3 is an
analytic arc with 7(0) — x and P o 7(s) = j(s + 1). But then j(s) -»• /? as
s —> 00, because P univalently covers w by P{UJ).
Let f/i, t/2 C D be the two complimentary components ofJD> \ 7. Consider
say U\. It contains a unique connected component U{ of P~l(U\), because
the dynamics of P on 7 is diffeomorphically conjugate to translation by 1.
Moreover the degree of the restriction P : U[ —> Ui is at least 2, because
U\ n S 1 is covered at least twice by U[ C\ S 1 . Thus U\ contains a critical point
for P. And similarly so do f/2. q.e.d.
Remark 3.5 Consider the family Qc(z) = z4 - 2z3 + z2 + z + c, c £ C of

quartic polynomials. For each polynomial Qc the real axis contains one crit-
ical point u) 6] - j , 0 [ with critical value vc. For Qo the critical value v0 is
contained in the same interval and is attracted to the parabolic fixed point
0. T i e two other critical points are complex conjugate and are attracted to
the parabolic fixed point 1. Moreover the external ray of argument 0 equals
[1,00]. For any sequence {cn}n€t) C]0,00[ converging to 0 the Hausdorff limit
R of the segments of external rays [wn,oo] C E equals [i>o,oo]. In this case
the segment ]0,1[ is like the arc Ry in the proof of the above Lemma.
Proposition 3.6 Suppose CQ E U C K,^, where U is some Fatou component.

Then the dynamic ray R^ (6) lands on the parabolic point on the boundary
ofU.
P r o o f : It was proved in Proposition 3.2 that c0 is a parabolic parameter, U

is an immediate parabolic basin of exact period q and 0 is g-periodic (1 = 0).
Let a denote the parabolic point to which Co is iterated by P^. Then a 6 R
by closeness and forward invariance of R. We claim that RC\ Jco = {a}.
The set R n Jco is a finite collection of g-periodic repelling or parabolic
points, by Corollary 2.6 and analytic continuation. If this set was to contain
any point other than a, then there should be a Fatou component U' such
that Rf)U' connects two different points of R n J^. But no such component
exist by Lemma 3.4 and the fact that P^ has only one critical point. q.e.d.
P r o o f of T h e o r e m 1: The cluster set Cl = RMd{e) v % i W Q dMd is a

continuum and is thus connected. Let CQ € Cl. By Propositions 3.1 and 3.2,
Co has either parabolic or Misiurewicz dynamics. Since the period q and the
preperiod I is fixed, the parameter c$ varies in a finite set and is thus unique,
which proves convergence. The dichotomy follows from Proposition 3.1 and
Proposition 3.2. Finally the statements about dynamic rays is contained
in Proposition 3.1 for the Misiurewicz case I ^ 0, and follows by combining
Proposition 3.2 and Proposition 3.6 for the parabolic case I = 0. q.e.d.
References
[CG] L. Carleson and T. Gamelin. Complex Dynamics. Springer-Verlag,
1993.
[DH] A. Douady and J.H. Hubbard. Etude Dynamique des Polynomes Com-
plexes. Premiere Partie, Publications Mathematique d'Orsay 84-02,
1984 and Deuxieme Partie Publications Mathematique d'Orsay, 85-04,
1985.
[HS] J. H. Hubbard and D Schleicher. The Spider Algorithm. In Proceedings

of Symposia in Applied Mathematics, volume 49, 1994.
[K] J. Kiwi. Rational rays and critical portraits of complex polynomials.

1997, IMS Preprint 1997/15 SUNY at Stony Brook.
[Mil] J. Milnor. Periodic orbits, external rays and the Mandelbrot set; an
expository account. Preprint, 1995.
[Pet] C. L. Petersen. Convergence of (pre)periodic rays in parameter spaces.

Manuscript, 1997.
[Ryd] G. Ryd. Iterations of one parameter families of complex polynomials.

PhD thesis, Department of Mathematics, KTH Stockholm, Sweden,
1997.
[Sch] D. Schleicher. Rational Parameter Rays of the Mandelbrot Set. Novem-

ber 1997, IMS Preprint 1997/13 SUNY Stony Brook.
CL. P E T E R S E N , I M F U F A , ROSKILDE UNIVERSITY, P O S T B O X 260,

DK-4000 ROSKILDE, DENMARK, e-mail: lunde@ruc.dk
GUSTAV R Y D , RESEARCH & T R A D E , 103 95 STOCKHOLM, SWEDEN
e-mail: gustav@rat.se
Bounded recurrence of critical points
and Jakobson's Theorem
Stefano Luzzatto
Abstract
We prove that the set of quadratic interval maps which satisfy a
strong bounded recurrence condition on the orbit of the critical point
has positive Lebesgue measure. This implies classical Theorems of
Jakobson and Benedicks-Carleson.
1 Introduction
1.1 Statement of results
In this paper we will be concerned with the dynamics of maps belonging to
the quadratic family
fa{x) = x2-a
for a £ fi£ = [2 - e, 2], e > 0 and x G / = [-2,2]. Let CQ = co(a) = / o (0)

denote the critical value of /„ and for i > 0, q = q(a) = /*(co). Notice that
this differs slightly from the usual convention of letting Cy denote the critical
value.
Definition (Bounded recurrence of the critical point). Fox 5 > 0, a >

0 and a 6 R we say that fa satisfies the bounded recurrence condition
{BR)n = (BR)n(5,a) if for all A: € [ 1 , 4
0<i<k
We say that / satisfies (BR) if it satisfies (BR)n for all n and we let
9.* = n*e{a, 6) = {a e Slt : /„ satisfies (BR)(S, a)}
Theorem. For every a > 0 there exists S > 0 suc/i £/m£
1ST*
|—— - > 1 a s e - > 0.
êre | • | denotes Lebesgue measure.
173
174 Stefano Luzzatto
The proof also shows that the complement of Q* in Qe contains an open and
dense set so that in particular fi* is a nowhere dense set of positive Lebesgue
measure. In section 4.1 we shall show that for small e and any a < log \/2, all
parameters in Q*e have an exponentially growing derivative along the critical
orbit. By standard results this implies that the corresponding maps /„ admit
an absolutely continuous invariant probability measure /i. In fact quite minor
modifications of the constructions and estimates given here can be used to
obtain the same conclusion. In any case we have as a Corollary the following
classical
Theorem (Jakobson, 1981). The parameter value 2 is a (one-sided) full

Lebesgue density point of the parameters for which the corresponding maps fa
admit an absolutely continuous invariant probability measure.
This result is considered a milestone in the theory of dynamical systems,

both for the intrinsic interest of the statement as well as for the ideas and
techniques introduced. It has also motivated a large amount of deep and
difficult research, although the full potential of the concepts and methods
involved is still far from being fully grasped or realized. The main purpose
of this paper is to take advantage of the substantial progress which has been
made through the application and generalizations of the original ideas to
more complicated examples, and to revisit the arguments in their original
and simplest setting.
The formulation of the main result in terms of condition (*) is new in this
setting and is motivated by several factors. Amongst these are the fact that
it appears naturally in the study of maps with infinite derivatives, where the
same condition is imposed on the recurrence near the critical points and near
the singularities; it is extremely simple to state and it is not hard to develop an
intuitive feeling for it; a simple application of the Ergodic Theorem actually
shows that it is satisfied quite generally, i.e. typical points satisfy (*) in
relation to their recurrence with respect to any other given typical point,
for any maps admitting an absolutely continuous invariant measure with a
density in LP,p > 1.
The proof is divided into 5 sections, in which the different kinds of ar-
guments required are developed: e.g. the combinatorial structure, the prob-
abilistic argument, the distortion estimates. Throughout the proof I have
tried to identify the exact assumptions used at each step and to bring these
to the reader's attention. In connection with this I have also tried to show
how most of the arguments apply in much more general situations than the
one considered here and to give some indication as to how they have been or
could be generalized.
Bounded recurrence of critical points 175
Acknowledgements. These notes would never have been written without

the enthusiastic support of the participants in the Non-Hyperbolic Dynamics
Seminar at Warwick during the spring term 1997 : James Gallagher, Oleg
Koslovsky, Lambros Lambrou, Hans Henrik Rugh, Sebastian van Strien and
Tan Lei. Special thanks also to Hans Henrik Rugh, Samual Senti and War-
wick Tucker for reading preliminary versions of this paper and making useful
suggestions, to Sebastian van Strien for many discussions, his encouragement
and his excellent advice, and to Jacob Palis and Marcelo Viana for their con-
tinuing friendship and support and expert guidance which has led me to the
problems and techniques developed in this paper. This work was supported
by a grant of the British Engineering and Physical Sciences Research Council
(EPSRC) No. GR/K86329
1.2 Idea of the proof

We fix e > 0 and restrict our attention to the parameter interval ft = ftE =
[2 - e , 2] (from now on we shall usually omit the subscript e). We begin with a
combinatorial construction leading to the definition of the set ft*. The details
are slightly technical but the basic idea is easy to explain. Fix a partition I
of the critical neighbourhood A = (-8, S) which is fine enough so that two
points which are close enough together and always fall in the same element
of the partition can be said to have essentially the same recurrence. We then
construct a sequence of partitions V^^V^,... of ft so that points in the
same element of the partition of order n "stay together" up to time n; let
ft'n' denote the union of elements which satisfy a certain condition which will
be shown to imply {BR)n.
It then remains to show that ft* = n n ft' n ' has positive measure and
|ft*|/e -> 1 as e —• 0. For this it is sufficient to show (i) that the mea-
sure Iflt"-1' \ ft(n'| of parameters which need to be excluded at each stage is
exponentially small in n and (ii) that no exclusions have to be made before
some time TV —> oo as e —>• 0. The second claim follows immediately from
the fact that the critical point is non-recurrent for the parameter value a = 2
while the first is non-trivial and constitutes the heart of the proof. In proba-
bilistic language it corresponds to saying that the conditional probability of
not satisfying {BR)n given that (B.R)n_i is satisfied, is exponentially small
in n.
Before going into the details let's consider a particularly simple example
which illustrates some of the mechanisms involved. Let T : I —> / be an
expanding interval map with a Markov partition V. For simplicity we assume
that T(LJ) = I for all elements w € V. Then by standard results there is a
sequence V\,Vi,... of refinements of V such that T" maps each u>n 6 Vn to
/ with uniformly bounded distortion. Now fix an arbitrary point eel and
consider the set T* = n n r ( n ) where
r< B) = {x e I : \fn{x) -c\> 8e-an)
for some a > 0 and 5 > 0. I claim that for arbitrarily small a and sufficiently
small 6, |F*| > 0. Indeed, it follows from the bounded distortion property that
the proportion of each interval u>n which falls into An = (c — 6e~an, c + 8e~~an)
at time n is of the order of 5e~an. Thus the same is true of the entire interval
and the total measure of points which fall into An at time n for some n, is
bounded by some constant multiple of ^2 5e~an which can be made arbitrarily
small by taking 5 small.
This simple argument can be interpreted in the following way. The dis-
tortion bounds (which in this case follow from the expansivity of the map)
and the Markov property imply that small intervals wn grow sufficiently large
sufficiently fast and in a sufficiently regular way that their distribution in the
original interval / at time n is essentially random. Therefore the probability
of falling into An is proportional to the size of An and the fact that this tends
to zero exponentially fast in n then yields the conclusion. Notice that the
construction shows that F* must be nowhere dense.
Our condition (BR)n is very similar to the one given above although some-
what stronger since it takes into account the past orbit of points. Since we are
interested in the recurrence of a single (critical) point for different parameters
we shall consider the image in the dynamical interval / of the iterates of the
critical point for different parameter values. It is easy to see that for any
n > 0 the images fa(c) for an interval of parameter values in form an interval
in /. To simplify the terminology we shall usually refer to these sets as iter-
ates of an interval of parameter values. We shall show below that intervals
of good parameters are growing in size at exponential speeds and that our
construction of the partitions in parameter space yields uniform distortion es-
timates. The main difficulty here compared to the situation in the toy model
described above is the absence of a Markov structure and even of a uniform
lower bound on the images of partition elements. To overcome this problem
we develop a statistical argument to show (and make precise the notion) that
most partition elements have uniformly large images and therefore can be
considered to have a sufficiently random distribution. More specifically we
prove two key technical estimates which link the size and number of elements
in P' n ) to their recurrence: elements with strong recurrence are small and not
too many. Therefore the average recurrence of the critical points of maps in
fif""1' up to time n is very low and it follows by a large deviation argument
that for exponentially most parameters it will be sufficiently low so that they
will belong toft<n>(i.e. the ratio Ift^ 1 ) - n^\/\n('n-1'>\ -> 0 exponentially
fast).
In section 2 we give the complete combinatorial definition of the sets Q^
and their partitions p(nK In section 3 we assume two technical lemmas (the
ones mentioned above) and carry out the large deviation argument. In sec-
tion 4 we prove three fundamental lemmas which constitute the basis of the
proof. These deal respectively with expansion estimates outside the critical
neighbourhood, parameter dependence, and the notion of binding. In section
5 we prove the bounded distortion property for elements of V^ and show
that all points of fiM actually satisfy condition (BR)n. Finally, in section 6
we prove the technical lemmas used in the large deviations argument.
1.3 Technical remarks

Remarks about specific aspects of the proof will be interspersed throughout
the text. I will limit myself here to some remarks of a more general nature.
Persistence of hyperbolicity. On one level, Jakobson's Theorem as well

as the Theorem given here and the many other generalizations which will be
mentioned below are about the general question of persistence of hyperbolicity.
While uniform hyperbolicity is essentially an open condition, this is not the
case for the weaker forms of hyperbolicity which occur in systems with critical
points or (in higher dimensions) homoclinic and heteroclinic tangencies. How-
ever, these Theorems do say that in certain cases non-uniform hyperbolicity
is persistent in a strong measure-theoretical sense. In the present setting the
proof makes it quite clear that the uniform hyperbolicity in most of the phase
space, in fact outside the critical neighbourhood A, which moreover persists
under perturbations, is crucial. However the fact that these estimates do not
hold everywhere in a persistent way means that parameter choices have to
be made and is reflected in the fact that Q* contains no open sets. On the
other hand the fact that we can take |A| -> 0 without affecting the uniform
expansion estimates is reflected in the fact that 2 is a full Lebesgue density
point of fi*.
Connected to this point is that of characterizing dynamically the Lebesgue
density points of all parameters for which the corresponding maps have ab-
solutely continuous invariant probability measures (actp's); an interesting
statement would be something like: a* is a (possibly one-sided) full Lebesgue
density point of parameters for which an acip exists (such measures always
reflect a certain degree of hyperbolicity, see next paragraph) if and only if it
satisfies certain dynamical conditions, e.g. it has an acip itself. Some very
partial results are known (see references below). Notice that maps without
acip's can be partial density points of maps with acip's (i.e. the ratio of the
measure of parameters for which an acip exists in an e neighbourhood of such
a map tend to a number strictly between 0 and 1 as e tends to 0. For example
it follows from the hyperbolicity of the renormalization operator that this is
true for infinitely renormalizable maps in the quadratic family. There are no
known examples of maps which are in the closure of the set of maps with
acip's but are zero Lebesgue density points of this set, or even for which such
density is not defined. Such examples would be very interesting.
It is very likely that questions related to the degree of persistence of hyper-
bolicity should be naturally formulated and connected to questions related to
the the very notion of hyperbolicity and to the degrees of hyperbolicity which
can occur, even though this notion is not yet well deinfed ! All maps in the
quadratic family (and even in much more general families) with an absolutely
continuous invariant probability measure fi are non-uniformly hyperbolic with
respect to /x: there exists A > 0 such that log|(/")'(x)| 1 / n —> A for //-almost
every x, i.e. the derivative is growing asymptotically at a (uniform) exponen-
tial rate. However it is clear that not all such maps exhibit the same degree
of hyperbolicity: for example the rate of decay of correlations, which reflects
some fairly intrinsic properties of the map, can be exponential in certain cases
and subexponential in others [NS98, BLvS99].
Construction of the invariant measure. The construction and the es-

timates given here can be used for parameters in 0* to prove the existence
of an absolutely continuous invariant probability measure. Indeed they easily
give rise to an induced map with very strong hyperbolicity properties which
in fact implies exponential decay of correlations for Holder continuous observ-
ables. Indeed, a version of Lemma 4.3 shows immediately that there exists
a simple partition V of the phase space, a function P : V —• N and an in-
duced uniformly expanding map F with uniformly bounded distortion given
by F\UJ — fp(-u)\uj. This map however does not satisfy infwep |P(w)| > 0, i.e.
elements of V do not grow to a fixed size under F. Nevertheless Proposition
3.1 implies that F can be iterated and V refined to give a new induced map
F, a refinement V of V and a function P : V -> N with F\(w) - fp\(w) such
that all OJ e V grow to fixed size under F. Moreover P is summable and
has an exponentially decaying tail: \{x : P(x) > n } | is exponentially small in
n. The existence of an acip fi for / then follows by classical arguments. By
recent results of Lai-Sang Young [You98], the exponentially decaying tail of
N immediately implies that if /x is mixing it also satisfies strong stochastic
properties such as exponential decay of correlation and the central limit the-
orem. See [BLvS99] for a generalization to a larger class of maps satisfying
some very weak conditions on the critical orbits.
Comparison with other approaches. All proofs and generalizations of

Jakobson's Theorem must essentially rely on the construction of some se-
quence of partitions with some combinatorial information which keeps track
of the certain aspects of the itinerary of partition elements. The partition
needs to be constructed in such a way as to guarantee uniform distortion

bounds and sufficiently strong analytic estimates on the size of partition ele-
ments in terms of their combinatorics. These ingredients are then all plugged
into a statistical argument to obtain the necessary estimates on the condi-
tional probabilities of satisfying the required conditions.
There are two main reasons for which this essential similarity in the vari-
ous proofs is not always easily recognizable. One is that different approaches
differ in the specifics and the order in which these constructions and estimates
are carried out. In this paper for example, the partitions are constructed fol-
lowing [BC91] by fixing a partition in dynamical space near the critical point
and essentially pulling it back via a special family of maps from parameter to
dynamical space. In other approaches, such as recent notes of Yoccoz [Yoc],
the partition is constructed topologically using preimages of a fixed point.
Both constructions do the job but require different approaches to the distor-
tion bounds and to the way the combinatorial information is defined. This
can also lead to an (apparently) different structure to the overall argument
as definitions and estimates need to be introduced at different stages.
Probably the most confusing aspect of trying to compare different proof
is related to the choice of the (essentially geometric) conditions which are
needed as an intermediate step in the construction. The existence of an acip
is a relatively abstract notion and therefore a more concrete condition is in-
troduced and is then shown on the one side to occur with positive probability
in parameter space and on the other to imply the existence of an acip. De-
pending on what kind of condition is chosen, the proportion of work required
to prove its persistence in parameter space and the fact that it implies the
existence of an acip can vary considerably. In this paper a bounded recur-
rence condition on the critical orbit is used which I would say is about half
way between the two: the proof of the existence of an acip given condition
(*) involves the construction of an induced Markov map and the estimates
involved are almost identical to those required in the proof of the persistence
of condition (*) in parameter space. Yoccoz, on the other hand, defines a pri-
ori a set of parameters which by definition have induced Markov maps. The
existence of an acip is then almost automatic and most of the work goes into
proving the persistence of such structures in parameter space. Interestingly
this involves also defining a condition on the critical orbits in terms of their
itineraries with respect to the given partition, which is strongly reminiscent
of our condition (*).
1.4 Historical Remarks

The existence of quadratic maps with absolutely continuous invariant prob-
abilities was first noticed by Ulam and von Neumann [UvN45] for the case
a = 2; they used the very special fact that / 2 is a Chebyshev polynomial

smoothly conjugate to the angle-doubling map on the circle to obtain an ex-
plicit formula for the density of the invariant measure (i. The extensive work
carried out during the 50's, 60's and 70's on uniformly expanding maps, see
[Ren57, Wil79, LY73, Bow79] formed the basis for the more sophisticated
techniques required to treat more general maps with critical points. The ex-
istence of an acip is now known to follow from some very weak conditions on
the critical orbit, see [NvS88] and [BLvS99]. Jakobson [Jak81] was the first to
show the existence of acip's could be typical in some sense, even for systems
with critical points. Benedicks and Carleson [BC85] gave an alternative proof
based on controlling the recurrence of the critical point and other proofs in-
clude [Ryc88, BC91, Tsu93a, Yoc] and the one given in the present paper.
The proof in [Tsu93a] is particularly remarkable for its conciseness as it takes
full advantage of the special properties of maps close to the top quadratic
map. It has recently been proved that in the special case of the quadratic
family almost all parameters have either an attracting periodic orbit or admit
an absolutely continuous invariant probability measure [GS97, LMN97].
There are many interesting generalizations of Jakobson's Theorem. These
are usually limited to proving the persistence of some geometric and/or ana-
lytic condition (analogous to the bounded recurrence condition given above)
but do not go into the details of the ergodic properties of systems satisfying
such conditions. In general however, the methods used can be adapted to
show that invariant measures with certain hyperbolicity properties exist. In
[dMvS93] a generalization is proved to show that under a mild transversality
condition, maps satisfying a non-recurrence condition on the critical orbit
are full Lebesgue density points of maps with acip's. In [TTY92] a sketch
of the proof is given for one-parameter families of maps with finitely many
critical points. In [Tsu93b] it is shown that under similar transversality as-
sumptions as in [dMvS93] maps with a finite number of critical points and
satisfying some exponential growth and bounded recurrence conditions along
forward and backward critical orbits are Lebesgue density points of similar
maps. Such transversality conditions have been shown to be satisfied for
all maps with exponentially growing derivatives along the critical orbit in
the quadratic family [Tsu] and for maps with non-recurrent critical points in
rather general polynomial families [vS]. In [Ree86] a version of Jakobson's
Theorem is proved in the context of rational maps; in [Rov93] "contracting
Lorenz maps" are studied in which the map has a critical point coinciding
with a discontinuity; in [Thu98]families of maps with completely degenerate
("flat") critical points are studied; in [LVb, LT] maps in which both singu-
larities with infinite derivatives and critical points coexist and in [PRV98] a
family of maps with an infinite number of critical points is analysed.
Of particular interest are generalizations to higher dimensional situations.
The ideas and techniques required for such a major breakthrough were first
introduced in [BC91] in the context of the Henon family of maps. These
were generalized in [MV93, Via93] to quadratic-like maps in the context of
hornoclinic bifurcations. In [DRV96, Cos98] similar techniques were applied
to study the dynamics near saddle-node bifurcation points. In [Via97] a class
of systems with many expanding directions is shown to be persistent under
perturbations. In [PR97] and [PRV] a class offlowswith an equilibrium point
with complex eigenvalues is considered and in [LVa] a generalized model of
the Lorenz equations is developed to include parameters for which the return
map contains folds as well as an unbounded derivative. These results do not
study ergodic properties of the systems but rather concentrate on proving the
persistence of certain geometrical/analytical structures which embody some
degree of hyperbolicity. The ergodic properties in some cases have also been
studied, see [BY93, BY98, BVb, BVa] for the Henon maps, [Alv97, AV] for
the systems studied in [Via97] and [HL] for generalized Lorenz-like attractors
of [LVa].
2 The combinatorial structure

For the rest of the paper we fix constants
log\/2>A>a><5>i>£>0
in the order given, where the symbol ~> means that the constant on the right
is chosen sufficiently small after the constant on the left has been fixed. To
simplify the notation we shall also introduce the constants
A = A - 2a and (3 = a/A.
The constant A gives the expansion rate outside a critical neighbourhood. It
can be chosen arbitrarily close to log y/2 and we shall take advantage of this
fact to simplify certain estimates. The constants a and 5 are those used in
the definition of condition (*) and in the formulation of the results. Notice
that taking a smaller strengthens the statement of the Theorem. The con-
stant i. is an incidental constant used in the construction to define a critical
neighbourhood A + = (S\ <5') which strictly contains the critical neighbour-
hood A = (-6,6). The fact that s is chosen last is no problem as our Main
Theorem is an asymptotic statement about what happens as e —> 0 once the
other constants have been fixed.
2.1 Inductive assumptions

Let fi(0) = Q = fi£ and Vm = {fi(0)} denote the trivial partition of Q.. Given
n > 1 suppose that for each k < n — 1 there exists a set fi(fc' C fi and a
182 Stefa.no Luzzatto
partition V^ of Q(k) into intervals such that each w 6 "P(fc) has an associated
itinerary constituted by the following information (for the moment this is
given as abstract data, the geometrical meaning of this data will become
clear in the next section). A sequence
0 = Vo < Vi < ••• <Vs < k, s — s(u) > 0
of escape times. Between any two escape times rjî and rji (and between ns
and k) there is a sequence
7]i-i < V\ < • • • < l/t < T)i t = t(u, i) > 0
of essential return times (or essential returns) and between any two essential
returns Vj_\ and t/j (and between ut and r/;) there is a sequence
Pj-x < Hi < • • • < /xu < Vj u - U{LJ, i, j) > 0
of inessential return times (or inessential returns). Following every essential

(resp. inessential) return there is a time interval
Wj + hPj] (resp. [fij + l^pj] )
with pj > 0 called the binding period associated to the return time Uj (resp.
fij). A binding period cannot contain any essential or inessential return times
although it may contain a sequence of bound return times (or bound returns).
Associated to each essential, inessential and bound return time is a positive
integer r called the return depth (which is roughly the (norm of the) logarithm
of the distance from the critical point) . We define the functions
Tl(k) : n{k) -> N and
which associate to each a € Jlf*' respectively the total sum of all return depths
and the total sum of essential return depths only of the element w € V^
containing a in its itinerary up to and including time k. By construction
both functions are constant on elements of V^.
We also assume that the following two analytic properties are satisfied.
Bounded recurrence: Each a 6 fi'fc) satisfies (BR)k-

Bounded distortion: For every u> € V^, the map
Cj : u> —> ujj = {cj(a) : a £ u>}
is a diffeomorphism with uniformly bounded distortion V j < v + p + 1

where v is the last essential or inessential return of w before time k and
p is the binding period. More precisely, there exists a constant D > 0

independent of A; and w 6 V^ such that
Moreover if u+p+1 < k the same statement holds for all i/+p+1 < j <
k + 1 replacing w by any subinterval u>' C o> which satisfies w^ C A"1".
2.2 Definition of fi^ and V{n)

For r € N let
Then
A+ = {0}U (J Jr and A = {0} U \J Ir.
\T\>TS+-\ \r\>n-\
where
r\5 = logJ" 1 and rs+ = tlog^" 1

and we can suppose without loss of generality that rs,rg+ € N. Finally
subdivide each Ir into r 2 subintervals of equal length. This defines partitions
X,I+ of A + with I = I + | A - An interval belonging to either one of these
partitions is of the form IT%m with m € [l,r 2 ] Let / £ m and I^m denote the
elements of I + adjacent to / r>m and let Ir<m = / ' m U / r i m U / £ m . If / r , m happens
to be one of the extreme subintervals of I+ then let / ' m or /£ m , depending
on whether IT<m is a left or right extreme, denote the intervals (-2(5', -S1] or
[6L, 2(5') respectively.
Remark 1. The reason for the subdivision of each IT into r 2 subintervals will
become clear in the last part of the argument in the proof of the distortion
estimates, see (12) and Sublemma 5.2.2. It is related to the fact that the
distortion of a single iterate on each Ir is of the order of 1 (which is "not
summable") whereas the distortion of a single iterate of / r , m is of the order
of 1/r2 (which is "summable").
We now define a refinement V^ of V^1]. Let u 6 V{n~l) and un =
(a), a e w} denote its "image" at time n. We distinguish different cases.
Non-chopping times. We say that n is a non-chopping time for u> if one

(or more) of the following situations occur:
2. n belongs to the binding period associated to some (essential or inessen-

tial) return time v < n of LJ;
3. ujn n A + 7^ 0 but wn does not intersect more than two elements of the
partition I + .
In all three cases we let u> 6 V^n\ If wn n (A Pi I±u) ^ 0 in cases 6) or c) we
say that n is a bound return time or an inessential return time respectively
for UJ € V^ and we define the corresponding return depth by
r = m a x { | r | :wnnlr ^0}.
Notice that we are allowing the possibility of intervals having inessential or
bound returns not actually intersecting A, but just the partition elements of
I+ adjacent to (but outside) A. This is for minor technical reasons related
to the estimates in Section 5.1.
Chopping times. In all remaining cases, i.e. if wn n A + jt 0 and wn

intersects at least three elements of I + , we say that n is a chopping time for
u £ j>(n-i). \Ye define a natural subdivision
w = a / u | J w(r'm)U<y.
(r,m)
so that each Wn fully contains a unique element of I+ (though possibly

extending to intersect adjacent elements) and u)ln and w£ are components of
ujn \ (A + U un) with \u)ln\ > 5l and \w?\ > 5L. If the connected components of
u)n \ (A + L)u)n) fail to satisfy the above condition on their length we just glue
them to the adjacent interval of the form Wn • By definition we let each
of the resulting subintervals of u be elements of V^n\ The intervals wl, uip
and a/ r ' m ' with \r\ < r^ are called escape components and are said to have an
escape at time n. All other intervals are said to have an essential return at
time n and the corresponding values of |r| are the associated essential return
depths. We make the same remark here as above: partition elements I±TS do
not belong to A but we still say that the associated intervals u/ ± r i ' m ) have a
return rather than an escape.
This defines the partition V{n) of Q^-V. In particular the functions S{n)
and 7£(n) are defined on fi'""1' and constant on elements of V^nK Thus we
can define
= {we t(n) : £(n)(w) < an/10}
and
•p(n) _ J>W\Q{n)
Notice that an itinerary of escape times, return times, etc. is well defined for
all elements of P' n ) up to time n. In Section 5 we shall prove that the two
analytic assumptions on the recurrence and the distortion are also satisfied:
Proposition 2.1. The inductive assumptions are satisfied for O'n)

In particular all parameters in Sl^ satisfy (BR)n. Let
0* - p| n(n).
n>0
2.3 Renormalization properties of the combinatorics

We give an alternative combinatorial description of the partition elements de-
fined above. This description is crucial to the probabilistic argument which
will be applied in the next section and highlights some remarkable renor-
malization properties of the construction. To each w 6 V^ is associated a
sequence 0 = r}0 < r]i < • • • < r]s < n, s — S(LJ) > 0 of escape times and a
corresponding sequence of escaping components
w C u>M C ••• C w^ 0 ' with u){m) C fi(T|<) and w(r)i) G P ( r ) i ) .
Letting
uj{r)i) =u for all s + 1 < i < n
gives a well defined parameter interval a/1"5 associated to a; 6 'P ( n ) for each
0 >
and let
denote the natural partition of Q^ into intervals of the form u^K Notice
that
fi(n-l) = Q(n) C . . . C QW
and
since the number s of escape times is always strictly less than n and therefore
in particular a/"*' = u> for all w £ VM. For a given w = u^ £ Q (i) , 0 (w) = {a/ = < > + l ) € Q<i+1> : u/ C w}
denote all the elements of Q' I+1) which are contained in u> and Q(*+1)(o>) the
corresponding partition. Then we define
A£ ( i ) :<2 ( i + 1 ) (a;)-+N where A£{i)(a) = £^+l)(a) - £M{a)
gives the total sum of all essential return depths associated to the itinerary of
the element u' £ Q'*+1)(w) containing a, between the escape at time rji and
the escape at time ?7i+i. Finally we let
Qii+1)(uj, R) = {u1 £ Q(i+1> :u'Cu, A5 (i) (o/) = R}.
3 The probabilistic argument

Modulo the proof of Proposition 2.1 which will be given in section 5, it remains
to prove that
|ir| > o
and that 2 is a Lebesgue density point of fi*. As mentioned above our main
strategy is to estimate the average value of the recurrence of the critical point
for parameters in Q'"" 1 ' taking into account the orbit up to time n. This is
the heart of the proof, where the combinatorial structure and the analytical
estimates really need to work together.
A relatively easy argument can be used to show that intervals in V^ are
exponentially small in terms of their recurrence, i.e. in terms of the total
sum of their return depths. However this in itself does not imply that the
total measure of points belonging to intervals with strong recurrence is small,
we also need to estimate the cardinality of the set of intervals in V^ with
a given recurrence. The easiest way to do this is to use the sequence of
return depths as a way of indexing the elements of 'P' n ) and to use standard
combinatorial techniques to estimate the possible number of sequences whose
terms add up to a given total. Unfortunately (unboundedly) many intervals
may have the same sequence of returns depths and therefore we need to divide
the combinatorial structure into blocks for which an indexing with bounded
multiplicity can be obtained. This is precisely the purpose of the construction

of section 2.3. These blocks have in an intrinsic interest as they contain certain
scale-invariant properties of the construction. Recall that 0 = a/A <C 1.
Proposition 3.1. For all i < n — 1, w G Q(l) and R > 0 we have
53 W\ < eim-l)R\w\.
The statement in the this Proposition may be considered the key estimate
in the entire paper. Starting with a "large" interval (the notion of escape
time has the purpose precisely of fixing the concept of large in this context)
we can construct a subdivision into subintervals which are again large after
a certain history of returns to the critical neighbourhood, quantified through
the total return depth function R,. The tail of R on one escaping component,
i.e. the set of points belonging to intervals which have not yet achieved large
scale after a total number of returns with return depths R, is exponentially
small in R.
Proposition 3.1 will be proved in Section 6. For the moment we assume
this estimate and complete the proof of our main Theorem. In section 3.1
we show how the estimate of Proposition 3.1 which applies to a single block
leads to a bound for the overall average recurrence up to time n. In section
3.2 we complete the proof with the large deviations argument.
3.1 Average recurrence

Lemma 3.1.
Proof. The equality follows immediately by the definitions. For the inequality,
let 0 rs u'6Q( i+1 )(u,R) V R>rs )
and therefore
5] e^'V/Vl < (1 + e-"/3)|a;| < e^'M. (1)
Since £<"> = A£(0) + • • • + A,?*11"1) and is constant on elements of

we can write XLe0<"> ef(n)f^^2|ci)| as
Notice the nested nature of the expression. Applying (1) repeatedly gives
3.2 Large deviations

The definition of f2'"' gives
and therefore using Chebyshev's inequality (A/ |{x G / : / ( x ) > M}\ < Jr f)
and Lemma 3.1 we have
in*"-1)
which implies
Thus using the fact that for e > 0 small there exists a N > 0 such that
QU) = Q(J+I) f o r a u j < N ( this follows easily from the fact that c,-(2) =
-2,Vj > 2) we get
j | 1
\ i=N / V i=N
Moreover it is easy to see that N(s) —> +OQ as e —> 0. Therefore
|fi'|/|fi £ | -> 1 as e -> 0
which implies that 2 is a one sided Lebesgue density point of fi*.

4 Fundamental tools
We state and prove three preliminary Lemmas which are not difficult but
whose importance cannot be overestimated. They provide the foundations
for the estimates of Sections 5 and 6 and are also of independent interest.
Generalizations of each of these results have been successfully developed and
applied.
The first is an estimate on the expansion outside a given critical neigh-
bourhood A = (—6,6). As a corollary we will show that all parameters in
f2* satisfy the Collet-Eckmann condition. The second is a basic result about
the dependence of iterates of the critical points on the parameter. The third
formalizes the intuitive idea that when an iterate of a critical point lands in
A it then retraces part of its initial orbit. Therefore information on the past
is useful to study the future.
4.1 Expansion outside A

Lemma 4.1. For all 0 < A < log 2 and 6 > 0 small, there exist constants
e > 0 and Cs > 0 such that for any a e fi£, / = / „ , n > 1 and x e I such
that when x, fx,..., fn~xx $. A we have
. (UE1)
/ / moreover fnx £ A + we have
\(fn)'x\ > eXn (UE2)
Proof. Consider the map h : [0,1] —> [-2,2] given by h(9) = 2 cos ir9 and
define a family of maps on [0.1] by
9a = h~l ofaoh.
Notice that for a = 2, h defines a conjugacy between fi = x2 — 2 and the
angle doubling map g2{9) = 29 mod 1. Indeed using the standard equality
cos2ip — 1 = cos2ip we have
f2(h{9)) = (2cos7T0)2 - 2 = 2cos2?r0 = h(g2{9)).
Now by the chain rule
I(/T)'(*)l =\h'(gna(h~1(x))) • (9:)'(h-\x)) • (h-iyix)]

Using the fact h'(8) = sinTrfl is close to zero only near the extreme points it
is easy to see that
tCs iir-l(x)i(-S,S)
for some constant Cs > 0 depending on 6. Since Ks^'WI = 2 n for any 9 this
clearly implies the statement with A = log 2 for a = 2. For a < 2 an explicit
calculation shows that
\9'a(e)-92(0)\<Cs\2-a\<eC's
for some (large) C's and therefore
I G O T ) I > (2 - eC'sY = e10^2-^ > eX

for any given A < log 2 and 5 > 0 as long as e is small enough. •
Remark 2. Estimates analogous to (UE1) with A depending on 5 hold in very
great generality, e.g. for any C2 unimodal map without periodic attractors
[Mis81, Man85]. The remarkable aspects of the statement given here are 1)
that the constant A can be chosen independent of 6, and 2) that the constant
Cs can be dropped if fn(x) is closer to the criticalpoint than x is. The proof
given here takes full advantage of the special nature of the parameter value
a = 2, namely that it is smoothly conjugate to a uniformly expanding map. A
similar result holds for perturbation of more general maps with non-recurrent
critical orbits [dMvS93] or even for maps with certain exponential growth
(Collet-Eckmann) conditions on the forward and backward orbits of the crit-
ical points [Tsu93b], however it is harder there to obtain good estimates on
the dependence of constants such as Cs,C's and e in terms of 6. Sharp es-
timates (or indeed any estimates at all) for A and Cg in terms of S in more
general situations would be extremely interesting. These issues are related to
the questions mentioned above about the characterization of maps which are
Lebesgue density points of maps with acip's.
Recall that A, a and S are assumed to have been fixed.
Corollary 4.1.1. For all e > 0 sufficiently small, a € Oe and k > 1 such
that fa satisfies (BR)k we have
Proof. Let N be a large constant such that Ci(a) $ A and (fl)'(co{a)) > 3 1 for
alii < N and all a e £le- N can be chosen arbitrarily large if e is chosen small,
this follows from the fact that for a = 2 the critical point is in the preimage
of a fixed point with derivative equal to 4. Let 0<N<fi<---<i/s<k

denote all the iterates such that cVj 6 A. By that chain rule we have, for any
fixed o 6 f2e and / = / o ,
(/* +1)'(co) = (fN)'(co) • (r-N)'(cN) •. • (/•*-*-') / (c,._ 1 ) • ( / * + 1 - " ) ' ( c J -

Moreover each term in the product, except for the first one, can be further
split up into a piece of orbit outside A for which (UE2) holds, and the last
iterate inside A. Thus rearranging the terms be separating iterates outside
A and inside A gives
and using the definition of (BR)k yields
Finally notice that that whenever cVj 6 A we have logical""1 > logtS"1 and
therefore condition (BR)k also implies in particular a bound on the number
of returns:
slogtT 1 < ^2 l o g ^ l " 1 < ait.
Substituting this into the last display above completes the proof. •
4.2 Parameter dependence

Lemma 4.2. For all a e Qe and all k > 1 such that fa satisfies (BR)k,
Proof. Let F : $7e x / -» / be the function of two variables defined inductively

by
F(a, x) = fa(x) and Fk(a, x) = F(a, f^x).
Then, for x — Co,
c'k(a) = daFk(a, co) = daF(a, ft'co) = - 1 + /i(cfc_,)«4_1(a).

Iterating this argument gives
and dividing both sides by (/*)'(<*) = /:(c fc _i)/^(c fc _ 2 ). . ./ifa)£(<*) gives
c 1
fc(°) _ i ,
fc
(/ )'()
Now choose iV so that
and e > 0 sufficiently small so that for all a E ile,
(fine,) < -(3.5) i V i < N. (3)

Then, by (2),
The upper bound is obtained in the same way. •

Remark 3. A similar remark applies here as in the previous section. The proof
is relatively straightforward in this case where we take advantage of (3) which
follows from the fact that the critical point is preperiodic repelling. More-
over, in this case transversality of the family has a clear geometric meaning:
a preperiodic repelling critical point is a codimension one condition in the
space of maps and the statement in the Lemma is saying that the quadratic
family crosses this subspace transversally and with positive speed. Similar
statements are usually assumed in more general versions of Jakobson's Theo-
rem; a geometric interpretation is not obvious in such cases but one can work
with the analytic condition that the expression in (2) is bounded away from
0. Tsujii [Tsu] has recently shown that the quadratic family does indeed sat-
isfy such a transversality condition at Collet-Eckmann parameters and van
Strien [vS] has proved the same for rather more general polynomial families
at Misiurewicz parameters (non-recurrent critical points).
The results just mentioned certainly have an intrinsic interest, although
in this context it is not clear why any form of transversality at all should
be required. Transversality conditions, in the theory of local bifurcations for
example, are often imposed to guarantee that one moves away from a bad
situation with positive speed; the opposite is true here: we begin with a map
with an acip and want to show that this phenomenon persists. Probably some
smoothness condition (maybe even a very weak one) on the dependence of
the maps with respect to the parameter is all that is really needed. In fact we
shall not use the statement of Lemma 4.2 directly but rather an immediate
consequence stated in the following Corollary. The rest of the proof could
most likely be adapted to work even if the constants 4 and 1/4 below were
replaced by some other constants possibly even depending on e, which would
probably be necessary if the transversality conditions were not satisfied.
Corollary 4.2.1. Let ui C fi and suppose that all a £ w satisfy (BR)k-

Then for all 1 • Wj given by tp(ci{a)) = c, o c^x(a). By the

mean value theorem there exists ( e w such that |WJ|/|WJ| = |y'(&)|- Then by
the chain rule and Proposition 4.2 we get the desired statement. •
4.3 Binding
Lemma 4.3. Let 0 < k <n — 1, u € V^ and suppose that k is an essential
return time for ui with return depth r. Then there exists an integer
p<2r
such that
> |wfc|W with 0 = a/A < 1/8.
Proof. For a € u let 7 = 7(0) be the interval [cfc(a),0] (or [0, ck(a)]) and for
j > 0, let 7j = fl+i(j). We define the binding period of ck(a) as
p{a, k) = sup{m 6 N : |T,| < e'2aj V j < m - 1}
and the binding period of u> as
p = p(uj, k) = min{p(a, k) : a € w}
Intuitively p(a, k) is essentially constant on u but strictly speaking there can

be some small variation. An additional argument is required at the end of
the proof to deal with this problem.
For the next two sublemmas we fix a parameter a g o ; and let p = min{p —
l,k}. We shall show below that p < i but for the moment we want to make
sure that our bounded recurrence condition holds during the time interval
under consideration.
194 Stefano Luzz&tto
Sublemma 4.3.1 (Distortion during the binding period). For all y0

and z0 in 7o and 0 <i <p+l we have
: = Da
Proof. By the chain rule
(f)'{yo) 7=0
By the definition of / , /'(z) = 2|z| V z € / and /"(z) = 2. Then by the mean

value theorem, \f'{zj) - / ' ( % ) | < |7j|sup ze/ /"(z) = 2|7,-|. Then, using the
fact that log(l + x) < x,
and thus it is sufficient to show that
j=0
2 7
Since |7j | < e" "- this follows immediately if we show that
which in turn follows if |cj| > ea* since

\Vj\ > h\ - |7j e'aj j - e-2 - e~aj ) > - e~a).
Let N be the smallest integer such that e~aN < S (i.e. N is the earliest time
at which condition (BR) allows a return to A; notice that N ~ logcS"1/0)-
Then for j > N we have clearly have \CJ\ > e~ai by condition (BR)j and for
j < N we have the basic estimate \CJ\ > 1 — e4^ (since 4 is the maximum
expansion rate). Therefore we just choose e small enough (recall that it can
be chosen arbitrarily small after fixing a and 6). •
Sublemma 4.3.2 (Expansion at the end of the binding period). Let
x — Cjt(a). We have an estimate for the length of the binding period:
1 2 1
logl^r ^p^-loglxl" <fc;
and two alternative estimates for the growth of the derivative at the end of
the binding period:
\(f"+1)'(x)\ > and |(/ p+1 )'(x)| >

Notice that the derivatives are calculated at x — Ck(a), thus there is an

overall expansion during the binding period which more than compensates
the small derivative at the return.
Proof. By assumption (BR)p is satisfied and therefore |(/i)'(co)| > eXj for all
j < p. By the mean value theorem, the definition of binding and the bounded
distortion estimate above,
e
d. |7p| _ \{f
Taking logarithms and rearranging yields
- 1 +21ogD
Q
H
~ \ + 2a
(4)
<l(log\x\-l+]ogDa) + l:
A <"
L
as long as \x\ < 28 is sufficiently small. Thus p < k and in particular
p = p+1. Similarly, using the fact that \f'{z)\ < 4, Vz and that |7P| > e'2ap,
we have
and therefore p > (21og|a;|"1 - logD 0 )/(2a + log4) > loglxl"1. This com-
pletes the estimates on the size of p. To get the derivative estimates, notice
that by the mean value theorem there exists £0 € 7o such that |(/p)'£o| =
|7p|/|7o|. Thus by Sublemma 4.3.1, the fact that \f'(x)\ = 2\x\ and |-yo| = x2
and the chain rule, we have
The two required estimates for the expansion are now easily obtained by
substituting a bound for p in terms of x in one case and a bound for x in
terms of p in the other. More precisely (4) gives e- 2 a p > e-*°\oe\x\-i/\ >
|a;|4a/A and therefore \{fp+l)'{x)\ > g g > Ixl*" 1 . Similarly (4) also gives
log jxl" 1 > Ap/4 i.e. jar]-1 > e ^ 4 and therefore |(/ p+1 )'(a;)| > ^ ^
> e(A-2Q)(p+l)/5 Q
We can now complete the proof of the Lemma. Suppose first of all that
p(a) is constant on oi. Then by Sublemma 4.3.2 and using the fact that
|ct(o)| > e- 2 r by assumption, we have |(/ p+1 )'(c fc (o))| > \ck(a)\50-1 > e^~l>
and thus by Corollary 4.2.1 |w fc+p+ i| > e ^ - 1 ) ^ ^ ! . The result then follows
from the assumption that k is an essential return for UJ implying that e~r/r2 <
|w/t| < 2e~ r . If p(a) is not constant on ui then for all a 6 u we have
|(/?)'(cfc+1(«))l ~ l(/aP)'(co(«))l ~ I(/Op)'(co(a))| ~ \(fj(ck+1(a))\.

Here the symbol ~ is used to indicate that the quantities on either side are
the same up to a multiplicative constant which does not depend on a, u or k.
The middle ~ follows from the inductive hypotheses and the other two from
Sublemma 4.3.1. Therefore
and the rest of the argument proceeds as above. •
5 Distortion and recurrence bounds

In this section we prove the general inductive step in the definition of fi(n)
and p(nK Notice that the combinatorial properties follow by construction
thus we only need to prove the two analytic conditions: bounded distortion
and bounded recurrence. We begin with the latter.
5.1 Condition (BR)n

Lemma 5.1. All points in fi'n) satisfy (BR)n.
Recall that functions TZ^ and E^ are defined on Q.^ and constant on
elements of V^ for all k < n and most importantly we have S W(a) < Aafc/10
for all k <n and all a 6 fi'n). We claim that it is sufficient to prove that for
all a 6 fi(n) and 0 < k < n we have
(5)
by virtue of the following
Sublemma 5.1.1. Let a 6 ft(n) and suppose that ft<*>(a) < ak/2 for all
k <n. Then a satisfies (BR)n.
Proof. Notice first of all that by construction cv{a) e A only if a has some
non-zero return depth associated to a return at time v. In other words all
the returns to A which contribute to the sum in the definition of condition
(BR) are accounted for by considering all the return times and return depths
associated to a. By construction, if v < n is a return for u> and the associated
return depth is r, then all points in u satisfy |cj,(a)| > e~2r (in fact they
satisfy ^"a] ~ e~T but the weaker estimate is sufficient for our purposes).
Therefore \og\c,,(a)\~x < 2r and
D
Equation 5 says that the at least one fifth of the total sum of return depths
associated to the orbit of the critical point of a given parameter a is formed
by essential return depths. Thus all we need to do is prove that the inessential
and bound returns occurring between two essential returns have a total sum
of return depths which can be bounded in terms of the return depths of the
preceding essential return. We divide the argument into two parts; first we
consider the essential returns and then the bound returns.
Sublemma 5.1.2. Let UJ 6 V^\ v 6 [1, n] and suppose that all points in w
satisfy {BR)V and that v is an essential return for u>. Set v := fio and let
v < n\ < • • • < fiu be a maximal sequence of inessential returns occurring
after time v and before any subsequent chopping time. Let ro, r\,...,ru be the
corresponding inessential return depths. Then
2
r=0
Proof. The idea of the proof is simple. Firstly, there are no chopping times
between time fio and time fiu and the intervals are growing exponentially,
in fact the growth can be estimates in terms of the return depths using the
binding period estimates. Secondly, large intervals cannot fall too close to
the critical point or else they would give rise to a chopping time. Thus
we can bound the return depths at each return in terms of the size of the
interval. More formally, let p0,... ,pu be the lengths of the binding periods
corresponding to the returns /io, • • •, / v By the bounded recurrence condition
and Sublemma 4.3.2 we have for each o 6 u> and i — 0 , . . . , u,
By Lemma 4.1 we also have
and therefore combining these two estimates and using the chain rule gives
198 Stefa.no Luzzatto
Thus by Corollary 4.2.1,
Now since |io is an essential return we have |w w | > e~r°/r% > 8e~' 1+a)r ° for
some a > 0 which can be chosen small and so
e -(l+a)ro
which immediately implies
t=0 *• A "
Sublemma 5.1.3. Leiw € P ^ ' , /i 6 [l>n] and suppose that /x is an essential

or inessential return for LJ with return depth p. Let p > 0 be the length of the
binding period following time \i and let p, < Q\ < • • • < £„ < fi+p be the bound
return times of w, i.e. those times for which u>^ D A ^ 0, and pi,... ,pv the
corresponding bound return depths: p, = max{|r| : w^ n lT ^ 0}. Then
t=l
Proof. For simplicity we fix a 6 LJ which has the deepest sequence of bound
returns. By the definition of binding the points c^+â) is close to Cj_i(a) for
all j < p and thus, intuitively, a bound return £ for c^ occurs exactly when
Co (a) has a return at time Q — \i + 1. We assume for the moment that this
is indeed the case. The main idea is that by the definition of binding period
C(. and C£_M+i are very close (of the order of e-2QCi-^+i) a n ( j by condition
(Blfyb-n+i, Icfc-p+il is of the order of e ~ a ^~' i + 1 and therefore the return
depth Pi corresponding to the bound return at time Q is almost the same as
the associated return depth rt corresponding to the return at time Q - fi+ 1.
Thus the sum of the bound return depths is uniformly almost the same as the
sum of all the return depths associated to co(a) up to time p. By condition
{BR)P and the estimates on the length of p this is of the order of ap ~ ar.
The details are left to the reader.
It can happen however that £ - fj,+1 is not a return time but c^^+i falls
very close to A so that c^ falls inside A and gives rise to a bound return time.
This situation has no real relevance to the dynamics and could to all effects
and purposes be ignored by modifying slightly the definitions. To be formally
correct however there are at least a couple of ways to deal with it. One way is
to change slightly the definition of condition (BR)n so that not all returns to
A are counted but those to a neighbourhood which shrinks slowly from A to

A minus the two extreme subintervals of the partition. In this way we could
guarantee that the above situation does not occur. Alternatively, notice that
the bound return depths corresponding to such situations are necessarily very
small, essentially r{. The techniques used for the binding estimates can be
applied here to show that there is a binding period of length ~ r& following
time £ — n + 1 during which this situation cannot happen, and is therefore
very rare. Once again the details are straightforward and are left to the
reader. •
5.2 Bounded distortion

In this section we shall prove a strong version of the distortion bounds.
Lemma 5.2. Let u e V{n). Then
1/r
< e > = D
for all a, b e w and all k < vq + pq + 1 where uq is the last essential or

inessential return of LJ before time n and pq is the length of the corresponding
binding period. Ifn>vq+pq+l then the same statement holds for all k <n
restricted to any subinterval Q C OJ such that \Qk\ C A + .
Notice that the distortion bound improves as S tends to zero. This is only
true under the additional assumptions stated that k < uq+pq+l or restricting
t o i i C u such that \Qk\ C A + . Without any assumptions the result still holds
for any k < n with a constant D& independent of n but depending on 5.
By Lemma 4.2 it is sufficient to show that
= D
i.e. critical orbits with the same combinatorics satisfy the same derivative
estimates. By standard arguments (see also the proof of Lemma 4.3), (6)
follows if we can show that
where Dj ^
Let 0 < vi < • • • < vq < k be the essential and inessential returns of u up
to time k. By construction there is a unique element IPi,mi in I+ associated
to each V{. Let p; be the length of the binding period associated to i^. For
notational convenience define UQ and p0 so that u0 + p0 + 1 = 0. We suppose

first that k < vq + pq + 1. Then write
"q+Pq 9 - 1 "i+l+Pi+l
E^ = E E *•
j=0 i=0 Ci
Sublemma 5.2.1. For each i — 0 , . . . , q — 1 we have
Proof. Write
"i + l+P.+ I "i+l"l l/.+ i+Pi+i
E ^ = E A+A>,+1+ E ^, (7)
I'.+Pi+I I/J+Pi+I "i+1+1
We shall estimate each of the three terms on the right hand side in separate
arguments.
For the first notice that since î/ i+l Q lpitmi condition (UE2) implies that
|(/"' +1 --')'c i+1 _ i; '(a)| > ex(-Vi+l~j) and therefore by Corollary 4.2.1 we have
\u>j\ < 4e~A("i+1~:')|w[/,+1|. Moreover Ic^^-jla)] > SL and so letting A =
we have
E»>o e~Xi
" f e- A ^K +1 |r t <4AK +1 |^'. (8)
The last inequality follows from the fact that 5~l < 6'1 < epi+l. For the
second term we immediately have
A^^KJe*" (9)
since DVi+l is the supremum of \u)Vx+l \/\cVi+l \ over all points cUi+l in w,,JH and
|ci/,+1| > 2e" Pi+1 by definition.
To estimate the third term we argue as follows. By Corollary 4.2.1 and
the bounded distortion during binding periods we have for all j € [vi+\ +
M < 8£>aK+1|sup|(/r^')'(c,i+1)(a)|. (10)

Now fix a G a; and let jj,j > 0 be as in section 4.3. By the definition of
binding periods we have |TJ| < e - 2 Q 0 ~" i + l ) and \jo\ > e~2pi+i and therefore
Moreover \f'a{cVi+1{a))\ < 2e~« +1 and therefore
Substituting in (10) gives
V Dj < 1 6 Z ? a K + 1 | e ^ ' V 6 " / """, < 16D*|w v ...|e ft+1 . (11)
Using the fact that A depends only on A and DQ is large when a is small,
equations (8)-(ll) complete the proof. •
Sublemma 5.2.1 gives immediately
f,+p, q-l l/jii+Pi+i 9-1
E i=E E ^<ÊKIe"-
D
(12)
j=0 t=0 Vi+Pi + 1 t=0
Since |wVi| < e~px/pf we have almost obtained our estimate except for the
fact that we have no control over the multiplicity with which each pi occurs.
To overcome this problem we show that essentially only the last return with
a given return depth needs to be considered in the sum (12) since the lengths
of the preceeding ones form a decreasing geometric sequence. First of all
subdivide the sum on the right hand side into partial sums corresponding to
return times with the same return depth:
i=0 r>r^ + i-Pi=r
Sublemma 5.2.2. For any r >rg+,
E
i:pi=r
Proof. Let (ij; = i/ijt j = 1 , . . . , m be the subsequence of returns with return

depths equal to r. By construction \iv^m | < 5e" r /r 2 . Using the binding period
estimates and (UE2) we have for all o e w and j = 1,..., m - 1,
Therefore by Corollary 4.2.1, |w w | > 5'( 1 - 5 «|w w+1 |/4 and
j=o
•
Substituting the estimate of Sublemma 5.2.2 into (12) and (13) we get
r TT
>> s+
This completes the proof of the Lemma for k < uq + p, + 1 . If k > uq +pq + 1
we consider the additional terms Dj restricting ourselves to some subinterval
u C w with Qk C A + . Clearly the preceding estimates are unaffected by this
restriction. Using (UE2) we have
and therefore using also the fact that |/'(c,(a))| > <5' since LJJ n A + = 0 we
get
6 Combinatorial and metric estimates

The aim of this section is to prove Proposition 3.1: for alH < n - 1, LJ €
and R > 0 we have
We use a counting argument to show that the cardinality of Q('+1\LJ,R) is

bounded above by e^R and an analytic argument to show that the size of each
u)' 6 Q(i+1)(w,i?) is bounded above by eW~VR\u\.
6.1 The counting argument

Lemma 6.1. For allO 0, we have
We begin with a standard abstract counting argument. Let SR denote the

the set of all sequences of pairs of integers ( i n , mi), (±r 2 ,m 2 ),.. .,(±r,,m,)
with s > 1, |n| + • • • + |r,| = R, \r,\ > rt and m, € [1, r]] for all j = 1,..., s.
Sublemma 6.1.1.
#5s<e».
Proof. We estimate first of all the cardinality of the set of possible sequences
rx,..., rs of positive integers Tj > r{ with rx-\ 1- r, = R for a fixed s > 1.
Notice that the number of of possible sequences as prescribed corresponds
exactly to the number of ways of partitioning R objects into s disjoint non-
empty subsets and can be bounded above by the number of ways of choosing
s balls from a row of R + s balls since this will determine a partition of
the remaining R balls into at most s disjoint subsets (the fact that each of
these subsets must contain at least r$ elements will be used in the calculation
below). Using Stirling's approximation formula for factorials k\ e [1,1 +
l/4k]V2irkkke-h we have
(R + s)\ {R + s)R+*
~ Rlsl ~ RRs« ~
To estimate the second term we use the fact that srs < R and that log(l+:r) <
x for all x > 0 and we get
To estimate the first term we write first of all

\R
R + *Y =
s/R -I R
Since s/R < l/tj and x I - > l a s a : - + 0 both terms inside the parenthesis
tend to 1 as 8 —• 0 and it is sufficient to choose a sufficiently small 6 to get
Next we need to allow each Tj to occur with multiplicity r? and this gives a
bound
r < 0R/4
3 —
Also, each r^ can have a positive or negative sign. This gives an additional
multiplicative factor of 2 s and thus a bound of
e (iUog2)/r Je /Jft/4
Finally, keeping in mind that for fixed R we have s < R/rs we can sum over
all possible sequence lengths s to get
E
s<R/rs
6
_ e\og{R/r,)efiR/3
The statement in Lemma 6.1 now follows from the above and the following es-
timate about the actual number of partition elements which can share exactly
the same sequence.
Sublemma 6.1.2. For u € Q'*"1' and any sequence ( r i , m i ) , . . . , (r s ,m s )

with s > 1, \TJ\ > r,5, rrij € [1, r|] there exists at most rf elements CJ € Q^(u>)
with a maximal sequence of essential returns 7?i_i < v\ < • • • < vs < rji such
that wVj n lri>mj ^ 0 for j = 1 , . . . , s.
Proof. The proof is an immediate consequence of the construction described

above. Thefirsttime, after time r/î, that w intersects A + in a chopping time,
every subinterval which arises out of the 'chopping' has either an escape time,
in which case the sequence above is empty and the lemma does not apply or
an essential return time in which case a unique pair of integers rx and m\ are
associated to it. Thus no two elements created up to this time can share the
same sequence. Fixing one of these subintervals which has an essential return
we consider higher iterates until the next time that it intersects A + . At this
time it is further subdivided into subintervals. Those which have essential
returns at this time ail have another uniquely defined pair of integers r 2 and
m.2 associated to them. However multiplicity can occur for those which have
escape times: all the subintervals which fall in A + \ A have an escape at this
time, and therefore belong to Q ^ and we do not consider further iterates,
but all share the same first (and only) term of the associated sequence of
return depths. However the number of such subintervals can be estimated by
the number of elements of the partition Z + | A + \ A p l u s a * most two elements
which can escape by falling outside A + . The number of such intervals is then
< 2(rg+ — rs)rj + 2 < rf. In the case of the intervals which have two or more
returns we repeat the same reasoning to get the result. D
6.2 The metric estimates

Lemma 6.2. For all 0 0 and w G Q (i+1) (w, R)
we have
Proof. Prom the construction of u> there is a nested sequence of intervals
UJ d U) *' d • • ' (Z UJ l' (Z (x? ° ~ U)
such that each u/"^ has an essential return at time Vj (intuitively UJW is
created as a consequence of the intersection of a/"-*-1' with A at time Uj).
Write
Id; I |<>
Letting r 0 = r,$ we have

Sublemma 6.2.1. For j = 0 , . . . , s - 1 toe /ia«e
<
Proof. We consider first the cases j — 1 , . . . , s - 1. The bounded distortion

property holds up to time Vj + pj + 1 and so
To estimate the numerator notice that by definition uy^+i C A + and there-

fore by (UE2) and Corollary 4.2.1 | w ^ ° l > c* ( "' +1 - |/J " w " 1) |wiji.t 1> l/4 and
therefore
To estimate the denominator, observe that by construction w^ has an es-

sential return at time Vj, and therefore by Lemma 4.3 |i<VJ+p.+1| > |w^J |8/3 >
For j = 0, recall that u/" o) = w 6 Q (i " 15 which means that w(l/o) has an
escape at time 77^1 = i/0. Suppose first that u)i[ C A + . The the bounded
distortion holds up to time i>\ and we have
\Jj
and we just need to show that |o;^ o) | > De 90rs. There are two cases to
consider according to the position of wffl. If u>^ n A ^ 0 then / r>m C
wlo°' S / r i m for some rs+ > r > rs and we can repeat the binding arguments
of section 4.3 and together with (UE2) and Corollary 4.2.1 these give
w (£i)| > |wf; o+po " 1) |/4 > |w^ o ) | w /4 > e~80r/4r1613 > De~90rs.
On the other hand if u)%o) D A + = 0 then by construction |w^ o) | > 5l and

using (UE2) again and the fact that the critical point is quadratic
as long as i < 4/?. This completes the proof in the case ufy' C A + .
If uTuo is not completely contained in A + we cannot apply the bounded
distortion property to the whole of a/"o) up to time vx. However, restricting
ourselves to a maximal subinterval w^ 0 ' C w'"o) with Gr^f1 C A + we do have
I I l l I
Therefore it is sufficient to show that |w^ o) | > De~90n. This is clearly true
since by hypothesis \u>i'/°''\ intersects A (since it contains la^ 1 '! which is con-
tained in A) and also extends all the way to the boundary of A + (since we
have assumed that |w^ o) | is not properly contained in A + ) . Therefore
•
The statement in the Lemma now follows immediately. Substituting into (14)
and using the fact that |w|/|w("^| < 1 since <D-C UJ^3'1 gives
M
The result follows since rs > r$. D
References
[Alv97] J. Alves. SRB measures for nonhyperbolic systems xuith multidi-
mensional expansion. PhD thesis, IMPA, Rio de Janeiro, 1997.
[AV] J. Alves and M. Viana. Statistical properties of some multidimen-
sional attractors. In preparation.
[BC85] M. Benedicks and L. Carleson. On iterations of 1 — ax2 on (—1,1).
Annals of Math., 122:1-25, 1985.
[BC91] M. Benedicks and L. Carleson. The dynamics of the Henon map.

Annals of Math., 133:73-169, 1991.
[BLvS99] H. Bruin, S. Luzzatto, and S. van Strien. Decay of correlations in

one dimensional dynamics, preprint 1999.
http://www.maths.Warwick.ac.uk/ luzzatto/papers.html.
[Bow79] R. Bowen. Invariant measures for Markov maps of the interval.

Comm. Math. Phys., 69:1-17, 1979.
[BVa] M. Benedicks and M. Viana. Random perturbations and statistical

properties of some Henon-like maps. In preparation.
[BVb] M. Benedicks and M. Viana. Solution of the basin problem for

certain non-uniformly hyperbolic attractors. In preparation.
[BY93] M. Benedicks and L.-S. Young. SBR-measures for certain Henon

maps. Invent. Math., 112:541-576, 1993.
[BY98] M. Benedicks and L.-S. Young. Markov extensions and decay of

correlations for certain Henon maps, to appear in Asterisque, 1999.
[CE83] P. Collet and J.-P. Eckmann. Positive Lyapunov exponents and

absolute continuity for maps of the interval. Ergod. Th. & Dynam.
Sys., 3:13-46, 1983.
[Cos98] M. J. Costa. Global strange attractors after collision of horseshoes

with periodic sinks. PhD thesis, IMPA, Rio de Janeiro, 1998.
[dMvS93] W. de Melo and S. van Strien. One-dimensional dynamics. Springer

Verlag, Berlin, 1993.
[DRV96] L. J. Diaz, J. Rocha, and M. Viana. Strange attractors in saddle-

node cycles: prevalence and globality. Invent. Math., 125:37-74,
1996.
[GS97] J. Graczyk and G. Swiatek. Generic hyperbolicity in the logistic

family. Annals of Math., 146:1-52, 1997.
[HL] M. Holland and S. Luzzatto. Invariant measures and decay of

correlations for systems with criticalities and singularities. Work
in progress, 1999
[Jak81] M.V. Jakobson. Absolutely continuous invariant measures for one-

parameter families of one-dimensional maps. Comm. Math. Phys.,
81:39-88, 1981.
[JN96] M.V. Jakobson and S. Newhouse A two dimensional version of a

folklore theorem. Amer. Math. Soc. Transl. (2), 171: 89-105, 1996.
[LMN97] M. Lyubich, M. Martens, and T. Nowicki. Almost every real

quadratic map is either regular or stochastic. Preprint Stony
Brook, 1997.
[LT] S. Luzzatto and W. Tucker. Non-uniformly expanding dynamics

in maps with criticalities and singularities, preprint
http://www.maths.Warwick.ac.uk/ luzzatto/papers.html.
[LVa] S. Luzzatto and M. Viana. Lorenz-like attractors without invariant

foliations. In preparation.
[LVb] S. Luzzatto and M. Viana. Positive Lyapunov exponents for

Lorenz-like maps with criticalities. to appear in Asterisque.
[LY73] A. Lasota and J.A. Yorke. On the existence of invariant measures

for piecewise monotonic transformations. Trans. Amer. Math. Soc,
186:481-488, 1973.
[Man85] R. Mane. Hyperbolicity, sinks and measure in one dimensional

dynamics. Commun. Math. Phys., 100:495-524, 1985.
[Mis81] M. Misiurewicz. Absolutely continuous invariant measures for cer-

tain maps of the interval. Publ. Math. IHES, 53:17-51, 1981.
[MV93] L. Mora and M. Viana. Abundance of strange attractors. Ada

Math., 171:1-71, 1993.
[NS98] T. Nowicki and D. Sands Non-uniform hyperbolicity and universal

bounds for S-unimodal maps. Invent. Math., 132:633-680, 1998
[NvS88] T. Nowicki and S. van Strien. Absolutely continuous measures

under a summability condition. Invent. Math., 93:619-635, 1988.
[PR97] A. Pumarino and A. Rodriguez. Persistence and coexistence of

strange attractors in homoclinic saddle-focus connections, volume
1658 of Lect. Notes in Math. Springer Verlag, Berlin, 1997.
[PRV] M. J. Pacifico, A. Rovella, and M. Viana. Persistence of global

spiraling attractors. In preparation.
[PRV98] M. J. Pacifico, A. Rovella, and M. Viana. Infinite-modal maps with

global chaotic behaviour. Annals of Math., pages 1-44, 1998.
[Ree86] M. Rees. Positive measure sets of ergodic rational maps. Ann. Sci.
Ecole Norm. SupA0 Serie, 19:383-407, 1986.
[Ren57] A. Renyi. Representations for real numbers and their ergodic prop-
erties. Acta Math. Acad. Sci. Hungar., 8:477-493, 1957.
[Rov93] A. Rovella. The dynamics of perturbations of the contracting

Lorenz attractor. Bull. Braz. Math. Soc, 24:233-259, 1993.
[Ryc88] M. Rychlik. Another proof of Jakobson's theorem and related re-

sults. Erg. Th. & Dyn. Sys., 8:93-109, 1988.
[Thu98] H. Thunberg. Positive Lyapunov exponents for maps with flat

critical points, to appear in Erg. Th. & Dyn. Sys., 1998.
[Tsu] M. Tsujii. A simple proof of the monotonicity conjecture in the

quadratic family. Erg. Th. & Dyn. Sys. to appear.
[Tsu93a] M. Tsujii. A proof of Benedicks-Carleson -Jakobson Theorem.

Tokyo J. Math., 16:295-310, 1993.
[Tsu93b] M. Tsujii. Positive Lypaunov exponents in families of one-

dimensional dynamical systems. Invent. Math., 111:113-137, 1993.
[TTY92] P. Thieullen, C. Tresser, and L.-S. Young. Exposant de Lyapunov

positif dans des families a un parametre d'applications unimodales.
C. R. Acad. Sci. Paris, 315, Serie 1:69-72, 1992.
[UvN45] S. Ulam and von Neumann. Bull. AMS, 1945.
[Ves87] A.P. Veselov. Integrable mappings and Lie algebras. Soviet Math.
Dokl., 35:211-213, 1987.
[Ves91] A.P. Veselov. Integrable maps. Russian Math. Surveys, 46(5):1-51,

1991.
[Via93] M. Viana. Strange attractors in higher dimensions. Bull. Braz.
Math. Soc, 24:13-62, 1993.
[Via97] M. Viana. Multidimensional nonhyperbolic attractors. Publ. Math.

JHES, 85:69-96, 1997.
[vS] S. van Strien. Generic unfoldings of rational maps. Work in

progress.
[Wil79] R.F. Williams. The structure of the Lorenz attractor. Publ. Math.
IHES, 50:321-347, 1979.
[Yoc] J.-C. Yoccoz. Weakly Hyperbolic Dynamics. Birkhauser, Basel, in

preparation.
[You98] L.-S. Young. Statistical properties of dynamical systems with some

hyperbolicity. Ann. of Math 147:585-650, 1998.
M A T H S D E P T , WARWICK UNIVERSITY, COVENTRY CV4 7AL, UK

E-mail: luzzattoQmaths. Warwick .ac.uk
Url: http://www.maths.Warwick.ac.uk/"luzzatto
The Herman-Swiatek Theorems
with applications
Carsten Lunde Petersen
Abstract
This article first presents a proof of the following Theorem of Her-
man, [Her] : An analytic circle homeomorphism / with a critical point
and irrational rotation number 0 is quasi-symmetrically conjugate to
the corresponding rigid rotation, if and only if 0 is of constant type.
The proof is presented so as to clearly exhibit, that it only uses the
hypotheses on Df through the fact that / satisfies the iWiatek Cross-
ratio distortion inequality [Swi], and that there is at least one critical
point ao on the circle, where / contracts (quasisymmetrically) like a
power xv with v>\.
Secondly the Theorem is applied, using an idea of Herman, to prove
the existence of quadratic Siegel polynomials, Pg(z) = et2w6z + z2, with
9 of non-constant type, for which the Julia set is locally connected.
1 Introduction
In his survey at Seminaire Bourbaki, February 1987, A. Douady [Dou] pre-
sented a conjectural procedure to produce synthetically a quadratic polyno-
mial Pg with a Siegel disk, whose rotation number is 9 and whose boundary
is a quasi-circle containing the critical point. The idea was to use quasi-
conformal surgery on a certain cubic Blaschke product fg, whose restriction
to S 1 is a real analytic circle homeomorphism with rotation number 8 and with
a critical point, an inflection point, (see also page 12). However in order to
perform the surgery, the restriction of fg to S 1 has to be quasi-symmetrically
conjugate to the rigid rotation with rotation number 6. This lead Douady to
pose the question: Does there exist any irrational 8 for which the conjugation
on S 1 of fg to the corresponding rigid rotation is quasi-symmetric? Herman,
who attended the seminar, responded positively only a few days later. Using
Swiatek's newly developed Cross-ratio distortion inequality [Swi] (see also
Theorem 1.3), he proved the following result, which in the folklore is known
as the Herman-Swiatek Theorem:
Theorem 1.1 ([Her]) Let f be an analytic circle homeomorphism with ir-

rational rotation number 8. Then f is c-quasi-symmetrically conjugate to the
rigid rotation (of angle 8), if and only if 8 is of constant type. Moreover the
constant c> 1 depends only on Df and the constant type bound N for 8.
211
212 Carsten Lunde Petersen
I shall give a proof of this Theorem with the analytic assumption replaced
by a technical hypothesis to be presented below. The proof is an edited
version of Yoccoz proof, in his unpublished manuscript [Yoc, Section 3].
We shall identify the circle with T = R/Z and give T the induced orien-
tation. Moreover we shall not distinguish a circle map / and lifts of / to R.
Let / be an orientation preserving circle homeomorphism. Call a quadruple
(a, b, c, d) admissible i f a < 6 < c < d < a + l i n R and define the cross-ratio
and the cross-ratio distortion
D(a,b,c,d,f) =
We shall be working under the following technical assumptions on /
Hypothesis 1 T i e circle homeomorphism f satisBes

a. Tiere exists a constant C > 1 suci tiat for any family (ai} bit ci; di)iEl of
admissible quadruples with max(#{i|:r 6 ]a.;,dj[}) < m :
b. There exist constants u, C > 1 and (at least) one 'critical point' a0 6 T,
such that for all x, y € R with \x — ao\ < \y - ao\
x a
~°
f(y) ~ /(ao) y
Note that for any circle homeomorphism / and for any t € R the circle homeo-
morphism ft = f + t satisfies Hypothesis 1 if and only if / does and that with
the same constants. Though not explicitly formulated the following slightly
stronger version of Theorem 1.1 is implicit in both [Her] and [Yoc].
Theorem 1.2 Let f be a circle homeomorphism with irrational rotation

number 8 and which satisfies Hypothesis 1. Then f is c-quasi-symmetrically
conjugate to the rigid rotation (of angle 8), if and only if 8 is of constant
type. Moreover the constant c > 1 depends only on the constant type bound
N for 8 and the bounds C, v, C > 1 in Hypothesis 1.
The connection between Theorem 1.1 and Theorem 1.2, that is the con-
nection between Hypothesis 1 and conditions on the derivative Df is made
through the following Theorem of Swiatek:
T i e Herman-Swiatek Theorems with applications 213
Theorem 1.3 ([Swi]) Let f be a C3 orientation preserving circle homeo-

morphism and suppose f has finitely many critical points OQ, aly... an such
that there exist neighbourhoods Wi C T of'<n for each i and such that :
1. / has negative Schwarzian derivative Sf on Wi \ a ; .
2. There exist constants At > 0 and U e N such that
V i e W , : At\x - a;!2'1 < f'(x) < 2Ai\x - atf1'
3. The variation of log Df onT^ U?=0Wi is bounded by p > 0.
Then / satisfies Hypothesis 1 with C, u, C given by the Ai,lit and p.
Proof : For Hypothesis l.a. this is the Cross-Ratio Inequality of Swiatek

[Swi, p.112]. Hypothesis l.b. follows by integration of 2. q.e.d.
Here Sf denotes the Schwarzian derivative of / and is given by
Sf = D2\ogDf-l-{D\ogDff
Any real analytic circle homeomorphism with a critical point is easily seen
to satisfy the hypotheses and thus the conclusion of Theorem 1.3. Thus
Theorem 1.1 follows by combining Theorem 1.2 and Theorem 1.3.
The fact that c in Theorem 1.1 depends essentially only on the constant
type bound N for 0, enables the proof of a result for quadratic polynomials,
with Siegel disks of certain non constant type rotation numbers:
For \8\ = 1 let eg be the complex number for which the quadratic polyno-
mial Qce(z) = z2 + eg has a fixed point with multiplier exp(i27T0). Herman
has proved the following Theorem :
Theorem 1.4 (Herman, unpublished) There exist irrational 0 of non-

constant type, such that the boundary of the Siegel disk for the quadratic
polynomial QCe is a Jordan curve containing the critical point.
Here I take his proof one step further and prove:
Theorem 1.5 There exist irrational 9 of non-constant type, such that the
Julia set of the quadratic polynomial QCs is locally connected.
2 Notation and prerequisites

Following the established conventions of the subject, we shall freely use the
letter c to denote any constant, which depends only on Df or as here the
constants C,u,C > 1 of Hypothesis 1. In particular the constants do not
depend on the irrational rotation number 6 of / . That is given any / satisfy-
ing Hypothesis 1 and any t ) 6 T such that /,, = / 4- rj has irrational rotation
number, the constants c for fv only depend on / .
Notation 2.1 We shall use freely the following notations:
1. f will denote a critical circle map with a (not necessarily unique) critical
point a0 and irrational rotation number 9.
2. The irrational 9 £ [0,1] has (the unique) continued fraction expansion

with integer coefficients bn and with convergents ^ for n e N:
0=
b2 +
where pn = bnpn^ + p n _ 2 and qn = bnqn^ + qn_2.
3. For each i € Z deSne at = /~*(ao), (note that the critical orbit is

indexed time reversely).
4. For n > 0 we define
"*•(*)], nodd,
[f~q"(x),x}, n even.
i<qn}, K = Kn^, {a,J
We recall that the irrational number 9 is of constant type with bound N,

if and only if the sequence {bn} of coefficients is bounded by N, and more
generally of constant type, if the coefficients are bounded by some unspecified
N. Moreover 9 is of constant type, if and only if, it is Diophantine of exponent
2, where in general 9 is Diophantine of exponent p > 2, if and only if there

exists C > 0 such that
V<7> l , V p € Z : \9--\ > —.
The most basic property of circle homeomorphisms is captured by the

Poincare semi-conjugation Theorem:
Theorem 2.2 (Poincare semi-conjugation Theorem) Let f be any ori-

entation preserving circle homeomorphism with irrational rotation number 9.
Then f is semi-conjugate to the rigid rotation R$(z) = z + 0 with angle 9,
i.e. there exists a continuous map <f>: T —> T such that <j> ° / = Re ° 4>-
The Poincare semi-conjugation Theorem implies that the combinatorial
orbit structure of any orientation preserving circle homeomorphism is the
same as that for the corresponding rigid rotation, in particular:
Remark 2.3 The circle homeomorphism f has the following properties:

1. For even n (respectively for odd n) the point of K^ following (respec-
tively preceding) a point a* for 0 1 all the intervals In(fj[x)) for 0 < j < qn+\
and In+iiPix)) for 0 < j < qn have mutually disjoint interiors.
3. For any x € T
T= U Jn(f(x))= U W(Z))U (J W/ J (*))

0<j<qn+i+qn 0<j<qn + i 0<j<qn
3 Theorem 1.2 and the ' a priori ' bounds

In this section we prove Theorem 1.2. We shall assume throughout this
section that / is a circle homeomorphism with irrational rotation number 9
and which satisfies Hypothesis 1. Moreover the letters ao, C, v, C shall always
denote respectively the critical point and the constants of Hypothesis 1. We
begin with a fundamental, but elementary consequence of Hypothesis l.a.:
Lemma 3.1 For any admissible quadruple (a, b, c, d) and any N € N one has
[fN(a), fN(b), fN(c), fN(d)} < Cm • [a, b, c, d],

216 Caxsten Lunde Peteisen
where m € N is tie covering number

max#{j|0
m= m a x # { j | 0 < j < N and x 6 }fj(a),fj{d)[}.
Proof : By Hypothesis l.a. we have
q.e.d.
Let us fix n > 0, say even to fix the ideas. Set q = qn, m(x) = mn(x),
I(x) = In{x) for x £ T (see Notation 2.1) and define for x e T
mn(f"(x))
(a;)) + mn{x) m n (/«(a:)) + mn(x)'

Lemma 3.2 There exists a constant c > 1 such that Vz 6 T and for all
z € /(*)
c'1 • mm{m(z),m{f"(z))} < m{x) < m(z) + m{fq{z)).
Proof : The right-hand inequality is a useful but trivial observation (see
the figure). Regarding the left hand inequality, let us prove that, there exists
a constant cL > 1 such that
Vx e T m(x) > c;1 • min{m(/«(a;))) m(/- ? (:r))}. (1)
Hz) z f(z)
I 1—I—I—I—
H lq
f (x) f (x) f"(x) x f(x)
Figure 1: Relative positions of the relevant points.
Given any x € T write min{m(/'(x)),m(/"" 9 (x))} = r(x) • m(x), we shall

prove r{x) is bounded independent of x. We can assume r(x) > 1 so that
m(x) < rnin{m(/«(x)),m(/- ? ( l ) )}, which implies A(x) > \. We estimate
(x)) = rnjfîx)) m{x)
mlT I - 4 - Tf\ \ T Q\ T\\ 1 —i~ T I T * !

The intervals f3Qf~3q(x),x[), 0<j<q cover any point y € T at most 3

times. Thus by Lemma 3.1
hence r(x) < 4 • C 3 - 1.

Let c = 4 • C 3 , so that the constant c\ — c - 1 applies in (1), by arbitrari-
ness of x 6 T. Let x e T and z € /(x) be arbitrary. Then by (1) either
771(2) < m(f q{x)) + m(x) < (ci + 1) • m(x) = c • m(x) or

m{f(z)) < m(f(x)) + m(x) < (c, + 1) • m(x) = c • m(x).
q.e.d.
Proposition 3.3 There exists a constant c > 1 such that for aiJ x 6 T and
2 € /(x) one has
c" 1 • m(x) < 771(0) < c • m(x).
P r o o f : By the above Lemma 3.2 it suffices to prove that there exists C\ > 1
such that for all z € T
cf1 • 771(2) < Tn{fq(z)) < cj • m(z) (2)
Let y € T be a point where the continuous function z •-» m(z) attains its
minimum, so that A(y) > 1/4, and let y, = f^{y) for all j e Z. Given any
z e T let j be the minimal non negative integer with yj E In{f2q{z))- Then
0 < j < qn+i + qn by Remark 2.3.3. Thus by Lemma 3.1 there exists c2 > 1
such that Cj ' < A(j/j), because the intervals ]yi+2q, yi-q[, 0 c^x increasing c2 if necessary. Writing
out the definitions of A(%), A(yj+q) and A(y J+ 2,) we obtain for c3 = c2 + 1
that:
j-) < c3m{yj+q)
[(/i+9) < c3m(yj+2q)

• m{yj+2q) < c3m(yj+q)
So that
c^lm{yj+q) < m{yj),m{yj+2q) < c3m{yj+q) (3)
sj+q
1
Let Ci be the constant from Lemma 3.2 then

cll mm{m{yj+2q), m{yj+q)} < m{z) < m(yj+2q) + m(yj+q) (4)
n{Vj), ™{yj+g)} < m{f{z)) < m{yj) + m{yj+q) (5)
Combining the estimates (3) with the pair of estimates (4) and (5) we obtain
(2), and the Proposition follows. q.e.d.
Corollary 3.4 The circle homeomorphism f is minimal, i.e every orbit is

dense and any Poinc&re semi-conjugacy is a homeomorphism and true conju-
gacy.
Proof : Suppose to the contrary that / is a Denjoy counter example with
invariant Cantor set L. Let z be a left endpoint of some (wandering) interval
/ of the complement of L. Then
lim m2n{z) = 0,
n>oo
which contradicts Proposition 3.3. Here | / | denotes the length of the interval
/. q.e.d.
Theorem 3.5 There exists a constant c > 1 such that

Vx G Kn : mn_i(a:) < c • mn(x).
Proof : Tofixthe ideas let us suppose n is even and define for i € Z
and
_
Then by Proposition 3.3 there exists cj > 1 such that
(6)
-t-
Figure 3: Relative positions of the points denning Si.
Thus it suffices to prove that there exists c > 1 such that
c~l < Ef for 0 < i < qn. (7)
Moreover by Lemma 3.1 it suffices to prove (7) only in the case i = 0, because
the intervals ]a.i+qn, ai+iqn_l [, 0 1 such that
So < c2 • E_, n . (8)
We can suppose uo < 1, as we want to bound «o away from 0. Thus by

Hypothesis l.b.
S-i < u - i < C" • uvQ.
Invoking Lemma 3.1 again (from E_ t to £_,„) we get
S - f c < C • C3 • ul (9)
As we can suppose uo < 1, we obtain from (6), (8) and (9)
— < L o < c2 • O • G • uQ
ZC\
This shows that both u0 and Eo are bounded from below and completes the
proof. q.e.d.
The above Theorem and/or the following Corollary is often in the lit-
erature referred to as d-priori real bounds for critical circle maps. They
are used as a bootstrap to control the geometry and topology of holomorphic
maps, which restricts to critical circle homeomorphisms. See for example [dF],
[dFdM], [dPaWdM], [GS], [Pet] and [Yam]. The papers [GS] and [dFdM] also
contain proofs of the d-priori real bounds.
220 Caisten Lunde Petersen
Corollary 3.6 {h-priori real bounds) Tiere exists a constant c > 1 such
that
V n > 2 : |o0 - a,,,, | < c • |a,n - ao|.
Now Theorem 1.2 follows as a Corollary of Theorem 3.5. We need however

an initial Corollary first:
Corollary 3.7 Suppose 0 is of constant type. Then there exists a constant

c > 1 depending only on C, v, C and on the bound N on the coefficients bn
so that
m
Vz6T,Vn€N: " ( ^ < c.
mn+i(z)
Proof : Let z € T be arbitrary. Given n € N there exists x € Kn such
that either z € / n _i(i) or z 6 / n_i(/"9 "-l (2;)) by Remark 2.3.3. For this x
there exists k < N + 1 such that z e In{f±kqn{x))- Let C! be the constant
from Proposition 3.3 and let c2 be the constant from Proposition 3.5 then we
obtain
. C C
— 2 1
mn{z) mn(x)
q.e.d.
Forn € N let Mn = d(Z, qn9) = \z- Rgn(z)l where the last equality holds
for any z € T exactly because Rg is a rigid rotation. Then for any n € N
bn+1Mn < M n _, < (1 + bn+l)Mn.
Corollary 3.8 Suppose 6 is of constant type. Then there exists a constant

c > 1 depending only on C, u, C and on the bound N on the coefficients bn
so that f is c-quasi-symmetrically conjugate to the rigid rotation Rg.
Proof : By Corollary 3.4 there exists a homeomorphism h : T —> T con-
jugating the rigid rotation Rg to / . Let z £ T and 0 < 6 < | be arbitrary.
Choose n € N such that Mn+l < S < Mn. Let us suppose n is even tofixthe
z-8 z z+8
Req-(z) R?'(z) Rl'(z) Rln'(z)

ideas (n odd being similar) then
mn+1(h(z)) < \h{[z,z + S\)\ <mnW>«{h{z))) (10)

(h(z))) < \h{[z - S,z])\ < mn{h(z)). (11)
Let Ci be the constant from Proposition 3.3 and let c2 be the constant from
Proposition 3.7 then we obtain
I < \h([z,z + S})\ < (?c
q.e.d.
Corollary 3.9 Suppose 6 is of non-constant type. Then f is not quasi-

symmetrically conjugate to the rigid rotation.
Proof : It suffices to prove that a homeomorphism h : T —>• T conjugating
/ to the rigid rotation Rg is not quasi symmetric. To this end choose a
subsequence {bnt}>gN diverging to oo. Then
IM4.*-iM)l _ Mnt-> ^ l vn
which contradicts that h is quasi symmetric, because
|/n»-i(flo)l ^ Tnnk-i{ag) y .
|/ n t _ 2 (o 0 )| m nt _ 2 (ao) ""
where c is the constant from Corollary 3.6. q.e.d.
Theorem 1.2 is a collection of Corollary 3.8 and Corollary 3.9
4 Local connectivity of Julia sets

Let /o : C —> C denote the Blaschke product
fo(z) = zl
1-32
For each irrational 9 e]0,1[ let X$ be the unique unimodular constant for
which the restriction fg = \g • f0 : S 1 —> S 1 is real analytic, has 1 as critical
point and has irrational rotation number 6. Thus Theorem 1.1 applies. Let
Jfs denote the Julia set of fg and let Jg C J/s denote the boundary of the
immediate attracted basin A#(oo) for oo. It was proved in [Pet] that
Theorem 4.1 For every irrational 9 the subset Jg and the full Julia set J;s
are locally connected.
We shall prove the following Theorem which combined with Theorem 4.1
implies Theorem 1.5:
Theorem 4.2 There exists irrational 9 of non-constant type for which the
Julia set J^ for QCs is homeomorphic to Jg.
The proof of this Theorem is based on the similar result for 6 of constant
type:
Theorem 4.3 (Douady, Ghys, Herman, Hubbard and Shishikura)

Suppose 9 is of constant type with bound N. Then there exists a K(N)-quasi
conforms! homeomorphism <pg : C —> C, which is a holomorphic conjugacy
between fg and P^ on the immediate attracted basins of oo and which maps
Jg (homeomorphically) onto Jc$.
The idea of the proof of Theorem 4.2 is to extract a uniform limit of the
maps <j>g of Theorem 4.3 for a sequence {#„} of irrationals of constant type
converging to an irrational of non-constant type.
To facilitate the proof of Theorem 4.2 we will first outline the proof of
Theorem 4.3:
Suppose the number 6 is of constant type. Let h : S 1 —> S 1 denote the con-
jugacy, between fg and the rigid rotation Rg. As 9 is of constant type, the
Herman-Swiatek Theorem, Theorem 1.2 states that the map h is c-quasi-
symmetric, with a constant c > 1, which only depends on the bound N on
the coefficients of the continued fraction expansion for 9. Let ipg : D —>• D
be a A'-quasi-conformal extension of h (with a K > 1 which only depends on
c and hence on N) and define Fg : C —> C by
1 Ipg l O Rg O t/jg if Z € D .
Let <70 denote the standard almost complex structure on C, and let og denote
the Fg invariant almost complex structure given by
{ V(o)()
((rl>oF?)*(<r0)(z)
crQ(z)
HFn(z)eB
otherwise.
Finally let <f>g : C —• C be the integrating map, the quasi-conformal home-

omorphism for which <rg — <^(<r0), normalized so that the conjugate map
T i e Herman-Swiatek Theorems with applications 223
fa o Fg o (f>gl equals the quadratic polynomial Qg(z) = z2 + eg, with an indif-

ferent fixed point of multiplier el2ir*.
The conjugacy h : S 1 —> S 1 between fg and Rg, normalised by h(l) = 1
depends continuously on 9 in the C°-topology (uniform convergence), because
fg and Rg depend continuously on 9 in the C°-topology.
Thus by choosing the quasi-conformal extension ipg in some canonical way,
say by using the Ahlfors-Beurling extension, [LV, Th. 6.3] or the Douady-
Earle extension, [DE] we can suppose that also the quasi-conformal extensions
xjjf depend continuously on 9 in the C°-topology. To fix the ideas we choose
to use say the Douady-Earle extension for every 9.
Lemma 4.4 Let (9n) be a sequence of irrationals converging to some irra-

tional 90. Suppose the 0n are of constant type with a uniform constant type
bound N. Then 9Q is of constant type with bound N and the two sequences
((j>gn) and (j>gl converge to (j>g0 respectively <j>^ in the C°-topology.
Proof: The coefficients of 6n converge to those of 6Q so that 60 is of constant
type with bound N. We shall prove that <j>gn converges uniformly (C°) to <f>g0,
from which the Lemma follows.
There exists K(N) > 1, such that each map 4>sn is /f-quasi conformal.
Thus extracting a subsequence, if necessary we can assume the sequence con-
verge C° to a K quasi-conformal homeomorphism $0. We shall prove that
<l>0 = fa- (12)
As (Fgn) converges to Fg0 in the C°-topology, and {Qc^n) converges to Q^ in
the C°-topology, the map <p0 conjugates Fg0 to Qg0 and the restriction of <f>0
to the immediate attracted basin of oo is biholomorphic. Thus (12) holds on
the closure of the immediate attracted basin of oo, by uniqueness of the holo-
morphic conjugacy. Moreover (12) also holds on the 'Siegel-disk' D, because
the Douady-Earle extension depends continuously on the boundary data hg.
Finally it holds on the grand orbit of D, because both (f>o and <j>g0 conjugates
dynamics. Thus any (C°) limit function of the <j>gn equals cj>g0. Combining
this with the precompactness of the sequence (<j>Sn) and the compactness of
C completes the proof. q.e.d.
Proof of Theorem 4.2: Let (en) be a summable sequence of strictly

positive numbers. Let (dn) be any unbounded sequence of natural numbers
and let 0i be any irrational of constant type. Given a natural number k\ let
for each n > 1 the irrational number 9\in be obtained from #i by replacing the
(n + fci)-th coefficient bn+ki in the continued fraction expansion of 9\ with dx.
Then the sequence (#i,n) converges to 9X. Hence by Lemma 4.4 there exists
n such that
Let 62 = #i,n and let fc2 = k\ + n. Then we may restart the process using
02, k2, ^2 and e2 t 0 obtain an integer fc3 > k2 and an irrational number 83 of
constant type, with bk2 = d\, bk3 = d? and with all other coefficients of the
continued fraction expansion of 93 equal to those of 0x, such that <fo2 and <f>e3
are c2 close in the Co topology. Proceeding recursively we find a sequence
(0n) of irrational numbers of constant type converging to some irrational of
non constant type 60 and a Cauchy sequence (in the C°-norm) of (quasi-
conformal) homeomorphisms ((j>$n), such that the sequence of inverse maps is
also a Cauchy sequence. Let <f>0 denote the limit of the sequence {<f>en), then
(j>o is a homeomorphism conjugating Fg0 to Q^. In particular JCs = <f>o(j0o).
q.e.d.
References
[dF] E. de Faria. Assymptotic rigidity of scaling ratios for critical circle
mappings. Ergod. Th. & Dyn. Sys. 17(1998), 227-260.
[dFdM] E. de Faria and W. de Melo. Rigidity critical circle mappings
I. preprint IMS#1997/16, Institute for Mathematical Sciences,
Stony Brook, November 1997.
[dFaWdM] E. de Faria an W. de Melo. Rigidity critical circle mappings
II. preprint IMS#1997/17, Institute for Mathematical Sciences,
Stony Brook, November 1997.
[Dou] A. Douady. Disques de Siegel et anneaux de Herman. Semi-
naire Bourbaki, Volume 1986-87, expose no. 677, Asterisque
153(1988), 151-172.
[DE] A. Douady and C. J. Earle. Conformally natural extension of
homeomorphisms of the circle. Ada Math. 157(1986), 25-48.
[GS] J. Graczyk and G. Swiatek. Critical circle maps near bifurcation.
Comm. Math. Phys. 176(1996), 227-260.
[Her] M. R. Herman. Conjugaison quasi symetrique des homeo-
morphismes analytiques du cercle a des rotations. Version tres
tres preliminaire, 19 pages manuscript, 1987.
[LV] O. Lehto and K. I. Virtanen. Quasiconformal mappings in the
plane. Springer Verlag, 1973.
[Pet] C .L. Petersen. Local connectivity of some Julia sets containing a
circle with an irrational rotation. Ada Math. 177(1996), 163-224.
[Swi] G. Swiatek. Rational Rotation Numbers for Maps of the Circle.

Comrn. Math. Phys. 119(1988), 109-128.
[Yam] M. Yampolsky. Complex bounds for critical circle maps. Ergod.

Th. & Dyn. Sys. 19(1999), 227-257.
[Yoc] J.-C. Yoccoz. Structure des orbites des homeomorphismes analy-

tiques possedant un point critique. 30 pages manuscript, 1987.
IMFUFA, ROSKILDE UNIVERSITY, POSTBOX 260, DK-4000 ROSKILDE,

DENMARK, e-mail: lunde@ruc.dk
PERTURBATION D'UNE FONCTION LINEARISABLE
HABIB JELLOULI
Resume :
Dans cette note, on demontre dans un premier temps que si P(z) —
e2niaz + z2 est linearisable en 0 et si (E2L)neN e s t la suite des reduites de
a dans le developpement en fraction continue, alors pour n assez grand les
qn petales en 0 du polynome quadratique parabolique Pn(z) = e2fft«" z +
z2, sont proches du bord du disque de Siegel de P. Ensuite on considere
f(z) — e2irtaz + O(z2) une fonction holomorphe linearisable en 0 et (fe)eei
une famille R-analytique de fonctions holomorphes sur le disque de Siegel
maximal A de / en 0, avec fa = f, fg(O) = 0 et /g(0) = e2*0 pour
tout 8 E I. On demontre que si a est un nombre diophantien alors on
peut construire pour k 6 N, une famille R-analytique {ipk,e)eei de fonctions
holomorphes tel que tfk,e conjugue fg a sa partie lineaire avec un terme
d'erreur O{{6 - a)k) uniformement sur tout compact de A. Enfin on donne
une reduction d'une conjecture de Douady et on utilise la famille (vfc,e)ee/
pour donner quelques proprieties sur la conjecture reduite.
Abstract :
In this note, we first prove that if P(z) = e27rtQz + z2 is linearizable at
0 and if {pn/<ln)neN is the sequence of the approximants of a in the devel-
opment in continued fraction then, for n sufficiently large, the qn petals at
0 of the parabolic polynomial Pn{z) = e2vt(-Pr>^qn^z + z2 come close to the
boundary of the Siegel disc of P. Later, we consider a linearizable holomor-
phic function /(z) = e2iriaz + O{z2) at 0 and (fg)eei an R-analytic family of
holomorphic functions on the maximal Siegel disc A of / at 0, with fa = / ,
fg(O) = 0 and f'g{0) = e2irie for any 9 € /. We show that if a is diophan-
tine, then we can construct for k € N, a family of R-analytic holomorphic
functions, say (fk,e)eei such that <pk,e conjugates fg to its linear part with
an error term O((0 — a)k) uniformly on any compact of A. Finally we give
a reduction of a conjecture by Douady, and we use the family (fk,e)eei t o
give some properties on the reduced conjecture.
227
228 Habib Jellouli
0. Introduction
Considerons f(z) = e2*iaz + O(z2) une fonction holomorphe linearisable
en 0 telle que a 6 I - Q et designons par A son disque de Siegel maximal
en 0. Alors on sait qu'il existe un isomorphisme ip de A sur le disque
unite D = {z € C/\z\ < 1} avec <p(0) = 0 et conjuguant / a la rotation
Ra(z) = e2"iaz (i.e. <p(f(z)) = Ra(<p(z)) pour tout z <E A). On appelle
sous-disque de Siegel pour / en 0 tout sous-ensemble A' de A de la forme
A' = tp'1^) ou r e]0,1] et D r = {z € C tel que \z\ < r}.
Soit maintenant / un intervalle ouvert de R contenant a et (/e)eg/ une
famille R-analytique de fonctions holomorphes sur A, verifiant : fa = / ;
/ e (0) = 0 et /^(0) = e27rie pour tout 6 e I. Pour U C A et (0,b) € / x N,
b
on note U(0,b) = f] fg\U) = {zeU tel que fe{z) E U pour 1 < i < b}.
i=0
Enfin dans tout ce qui suit on note Rg{z) = e2nt9z et (9n = ^-)neN
designe la suite des reduites de a dans le developpement de a en fraction
continue.
Theoreme 1. Soit Ai C A2 deux sous-disques de Siegel distincts pour f en
0 et (6n)ngN une suite d'entiers naturels positifs verifiant bn = o(qnqn+y).
i) si K est un compact dans Aj, alors pour n assez grand on a:
ii) si K est un compact dans A2 - Ai, alors pour n assez grand on a :

Kc(A2-A1)(9n,bn).
Un cas interessant se presente lorsqu'on etudie la famille des polynomes

quadratiques P$(z) — e2ni6z + z2. A ce propos on sait que chaque polynome
quadratique Pgn (z) = e2rrt«n z+z2 admet qn bassins immediats paraboliques
attaches a 0. Comme consequence du Theoreme 1, on a le corollaire suivant:
Corollaire 1. Si A' C A est un sous-disque de Siegel pour Pa(z) =
e2*taz + z2, alors il existe un rang no 6 N tel que pour tout n > no, chaque
bassin immediat parabolique B attache a 0 pour Pgn(z) = e27rt«» z+z2 verifie
B n (A - A 7 ) / <t>.
On dit que a satisfait a une condition diophantienne d'ordre /3 > 0

s'il existe une constante c > 0 telle qu'on ait \a - ^ | > -5^3 pour tout
^ € Q. On dit que a est un nombre diophantien s'il satisfait a une condition
diophantienne d'ordre (3 pour un certain (3 > 0.
II est bien connu depuis Siegel que si a est un nombre diophantien, alors
tout germe holomorphe f(z) = e2ntaz + O(z2) est linearisable en 0.
Perturbation d'une fonction Hnearisable 229
Theoreme 2. Supposons que a est un nombre diophantien et soit k un

entier naturel positif donne. Alors il existe un intervalle ouvert J C I con-
tenant a et une famille K-analytique (fk,e)eeJ de fonctions holomorphes
sur leurs domaines de definitions notes Dom(ipk,e) et verifiant les trois pro-
prietes suivantes :
i) <Pk,a = V;0 e Dom(tpKe); ipk,e(O) = 0 et ¥>'M(0) = ¥>'(0) pour tout 0 e J.
ii) si K est un compact de A alors pour 0 assez voisin de a, on a:
K C Dom(ipkfi)
iii) ifkfi ° fe — R& ° <Pk,e + O((9 — a)k) uniformement sur tout compact de
A.
Ainsi ipkfi conjugue fg a sa partie lineaire avec un terme d'erreur O((6 —
a)k) uniformement par rapport a 8 sur tout compact de A. (fk,e)eeJ es*
appelee famille de coordonnees suivies a l'ordre k £ N associee a la famille
(fe)eei-
Corollaire 2. Supposons que a est un nombre diophantien et soit k un en-
tier naturel positif donne. Si (bn)n€® est une suite d'elements de N verifiant
K = o(qnqn+i) o-lorsipkfinofin = RinO(pkfn+O((6n-ct)k~~:l) uniformement
sur tout compact de A et uniformement par rapport a j avec 0 < j <bn.
Notons Pe(z) = e27ri6z + z2 ; K(Pe) = l'ensemble de Julia rempli de P9.
Le paragraphe 4 de cet article est une tentative pour aborder la conjecture
suivante de Douady :
Conjecture. Si a est un nombre diophantien et (9n = pn/qn)neN est la
suite des reduites de a, alors on a :
Jjm^nus^PflJn K (Pa)) = mes(i: (Pa))
(oii mes designe la mesure de Lebesgue).

Le theoreme suivant permet de faire une reduction de cette conjecture :
Theoreme 3. Si Pa{z) = e2niaz+z2 est Hnearisable en 0 et si (a n ) n 6 N est
une suite de nombres reels verifiant lim ctn = ex, alors les deux proprietes
n—>+oo
suivantes sont equivalentes :
i) lim mes(K(Pan)D K (Pa)) = mes(K (Pa))
o
ii) il existe une composante connexe Q de K (Pa) verifiant:
lim mes(K(Pa ) n fi) = mes(Q).

230 Habib Jellouli
II en resulte que si on a lim mes(K(Pgn) D A') = mes(A') pour

n—•+00
tout sous-disque de Siegel A' = y)~1(Dr) avec r e]0,1[, alors on aura
lim mes(K(Pg ) D A) = mes(A) et par suite la conjecture sera vraie.
Notre tentative consiste a conjuguer P$n par la coordonnee suivie <pk,en sur
A' (la famille (<Pk,en)neN e s t donnee par le theoreme 2) puis en passant
dans le revetement universel I = { w e C/Re w < 0} de B£ = {z € C/0 <
\z\ < r} au moyen de II(w) = re^, on aboutit a une fonction holomorphe,
injective, Gn : H -> C de la forme Gn(w) = w + \new + hn(ui).
La fonction Gn semble posseder des proprietes tres interessantes ; quel-
quesunes sont demontrees au paragraphe 4.
1. Preliminaires
1.1. Developpement en fraction continue

Si x est un nombre reel on note [x] sa partie entiere et {x} = x — [x] sa
partie decimale. Soit a 6 l - Q , On pose a0 = [a] e Z, ct-o = {a} E [0,1[;
on definit par induction, pour n £ ff, un entier an et un reel an par :
an = et an = \ \ .
On pose p_2 = 9-1 = 0, p_i = q-2 = 1, et on definit inductivement

pour n e N des entiers pn et qn par les relations pn = anpn-i + pn-2 et
9n = an<ln-i + Qn-2- Les rationnels ^-, pour n € N s'appellent les reduites
du nombre irrationnel a. On a alors les proprietes suivantes :
pnAqn =l , pour n > 1,

n+1
pnqn+i -Pn+iqn = ( - l ) , pourn>0,
|a —I < , pour n > 0,
<7n 99
9n+i > qn > I , pour n > 2 .

\ 2 /
1.2. Disque de Siegel maximal
Soit f(z) = e2vlotz + O(z2) une fonction holomorphe definie dans un
voisinage de 0 telle que a e R - Q.
Definition. / est dite linearisable on 0 s'il existe un domaine simplement
connexe U de C tels que 0 G U et / est un isomorphisme de U sur lui-meme.
Dans ces conditions U est appele un disque de Siegel pour / en 0.
Perturbation d'une fonction linea.risa.ble 231
Lemme. Si f est linearisable en 0 alors I'ensemble des disques de Siegel

pour f en 0 est ordonne par inclusion.
En effet : Si U et V sont deux disques de Siegel pour / en 0 tel que
U <£V, alors il existe un point z0 6 U avec ZQ fi V. Soit ip un isomorphisme
de U sur le disque unite D tel que ip(0) = 0. II en resulte d'apres le lemme
de Schwarz que ip(f(z)) = e2r:%aij){z) pour tout z eM.
Considerons la courbe de Jordan L image reciproque par ip du cercle de
centre 0 et passant par ip(z0). On en deduit que L est invariante par / et
que l'orbite directe du point ZQ par / est dense dans L (car a 6 K — Q).
Par suite V D L = <f>, ce qui donne V C U car 0 6 V.
Definition. Si / est linearisable en 0 alors d'apres le lemme precedent,
l'union A de tous les disques de Siegel pour / en 0 est aussi un disque de
Siegel pour / en 0, appele disque de Siegel maximal pour / en 0.
Remarque. Chaque isomorphisme tp de A sur le disque unite D avec
ip{0) — 0, conjugue / a la rotation Ra(z) = e2viocz.
ip est appele une application linearisante de / en 0. Si <p est une autre
application linearisante de / en 0 alors il existe A 6 C avec |A| = 1 et
ip — \ip.
2. Demonstration du theoreme 1
2.1. Demonstration du theoreme 1

On peut supposer sans perte de generalite que A2Â. Soit r±, r^ 6]0,1[
tel que A» = tp-1^) ou D n = { z £ C/\z\ < n} pour i 6 {1,2}. Posons
5 = ini{\x - y\,x 6 9A2 et y € dA} • 8 > 0 et il existe e > 0 tel que
\fe{z) — f(z)\ < f P o u r t o u t ^ G A2 et pour tout 6 verifiant \0 — a\ < s.
Du fait que /(A 2 ) = A2, on en deduit que fe{z) E A pour tout z 6 A2 et
\9-a\ <e.
Ainsi la fonction Fg(z) = (cp o fg o <p~x)(z) est definie sur Dr2 pour tout
|# — a\ < e, R-analytique en 9 et holomorphe en z. II est facile de voir qu'on
peut ecrire F6(z) = e2niez + (9 - a)z2He{z) pour \9 - a\ < e et z e D r2 .
Posons c = swp{\z2Hg(z)\, z € Dr2 et \9 — a\ < e}. Comme lim 9n = a;
n—•+oo
alors il existe no £ N tel que \9n — a\ < e pour tout n > no, et en utilisant
la propriete \9n — a\ < -—-—, on obtient :
? (1)
qnQn+l QnQn+1
pour tout z € Dr2 e* pour tout n > no-

232 Habib Jellouli
i) Si K est un compact de Ai alors il existe r[ e]0, ri[ tel que p(K) C

Par hypothese, bn = o(qnqn+i)> done il existe n\ > UQ tel que :
< pour n > ny.

QnQn+l C
On en deduit d'apres (1) que pour n > m et \z\ < r[ + j ri b ~ r ' (avec
0 < j < bn - 1), on a : \F9n(z)\ < r[ + (j + 1 ) ^ .
II en resulte que si z £ D r ' alors Fjj (z) € D ri pour 0 < j < bn et
n > ni, d'ou K C Ai(0 n , &n) pour n > nj,.
ii) Si K est un compact de A 2 - Ai, alors il existe r " , ^ ' €]r 1 ) r 2 [ avec:
r'l < rl etip(K)c Dr» - Dr».
Comme bn = o(qnqn+i) alors il existe n2 > no tel que
bn .
!L — ,— ) pour n>n2.
— < min V C C /
On en deduit d'apres la double inegalite (1) que pour n > n 2 et
-" " r 2 - r'i

rl jr < \z\ < r2 + j—7—- (avec 0 < j < bn - 1)
"n
on a:
Ainsi si z € Dr^' - Br» on aura F$n (z) e Dr2 - D r i pour 0 < j < bn et
n > n 2 , d'ou if C (A2 - Ai)(0 n ,6 n ) pour n > n 2 .
Ceci acheve la demonstration du theoreme 1.
2.2. Demonstration du corollaire 1
Soit A' C A i m sous-disque de Siegel pour le polynome quadratique
linearisable Pa(z) = e2iriaz + z2. Designons par Bn le bassin imm^diat
parabolique attache a 0 pour Pgn(z) = e " ^ z + z ^ e t qui contient le point
critique de P$n.
Le point xn — —e Vliv. (verifiant Pgn(xn) — 0) se trouve sur le bord
de Bn et pour n assez grand xn ^ A (car x = —e2nta £ A). Fixons une
courbe de Jordan invariante L pour Pa dans (A — A ). On en de'duit qu'il
existe no £ N tel que Bn n L ^ <f> pour n > no. D'autre part, d'apres le
theoreme 1, ii), en prenant bn = qn, Vn € N, alors il existe n\ > no tel que
L C (A - A )(9n,qn) pour n > nj. Or on sait que les qn bassins immediats
parabouques attaches a 0 pour Pgn sont donnes par Bn = Pg (Bn ) avec
0 < j < qn - 1. En conclusion B^ n (A - A') # <f> pour n > ni et

0 < j < qn — 1. Ainsi le corollaire 1 se trouve demontre.
2.3. Complements
1) Si Ai C A est un sous-disque de Siegel pour / en 0 et (6n)neN une suite
d'entiers naturels positifs verifiant bn — o(qnqn+\), alors il existe un rang
no e N tel que pour n > no, toute composante connexe de Ai(0n,bn) est
simplement connexe.
En effet : d'apres le theoreme 1, il existe n 0 € N tel que pour n > n0,
on a Ai c A(6n,bn). On pose Dr = </?(Ai) = {z e C/\z\ < r}.
Ainsi pour n > no et 0 < k < bn, les fonctions gn). = ipo f£ otp~1 sont
definies et holomorphes sur D r .
Soit C une composante connexe de Ai(dn,bn) pour n > no, et soit F
un lacet de Jordan dans C. Notons W la composante connexe bornee de
C - T et M = (p(W) C D r . Ainsi pour 0 < k < bn et z 6 dM = ip(T) on a
|j7n,fc(2)| < r- II e n resulte d'apres le principe du maximum que pour tout
z & M et pour tout k 6 {0,1,2,..., bn}, on a |<7nifc(z)| < r. Ceci exprime
que W C Ai(0 n , bn) et par suite C est simplement connexe.
2) II est facile de voir d'apres le theoreme 1 qu'au sens de la distance de
Hausdorff entre compacts de C, d(Ai(9n,bn)) (resp. 9((A2 — Ai)(6n,bn)))
tend vers 9Ai (resp. c?Ai U CJA2) lorsque n tend vers +00.
3. Demonstration du theoreme 2
3.1. Preliminaire
Si Cl est un domaine de C, alors on note par H(Q,) l'ensemble de toutes
les applications holomorphes sur fi et a valeurs dans C.
Considerons deux domaines U, V de C et ip : U —> V une application
holomorphe telle que (p'(z) ^ 0 pour tout z £U.
On definit l'operateur lineaire ip* de H(V) dans %{U) par :
V9 ^ r P ° u r 9 e H(V).
Et lorsque • V et xp : V —> W sont holomorphes et verifiant <p'(z) 7^ 0
pour z e U et tp'(^) 7^ 0 pour £, &V, alors (ip o ip)* = <p* o ip*. Si de plus (p
et ^ sont des isomorphismes, alors (ip o ip)t = ip* o tpt.
234 Habib Jellouli
ii) Soit un diagramme commutatif
I'
Si ip,tl> et h sont holomorphes et verifiant <p'{z) ^ 0 pour z eU, xj>'(z) ^
0 pour z e V et h'(z) ^ 0 pour z € U, alors on a aussi un diagramme
commutatif pour les operateurs associes :
En particulier si #0 £ %(^0 est un point fixe pour ip* (i.e. V*<7o = go), alors
/o = h*go 6 ?^(W) est un point fixe pour ip*.
3.2. Lemme. Soit a un nombre diophantien et A € C. Considerons
fa{z) — e2*laz + O(z2) une fonction holomorphe dans un voisinage de 0 et
designons par A son disque de Siegel maximal en 0.
Alors, pour toute fonction holomorphe G : A —> C avec G(0) — G'(0) =
0, il existe une seule fonction holomorphe g : A —> C verifiant g(0) = 0,
g'(O) = Xetg-f*g = G.
Demonstration : Soit D = { . z e C / | , z | < l } une application
linearisante de / en 0. On en deduit, d'apres la propriete ii) ci-dessus, que
le diagramme suivant est commutatif
H(A) - ^ ^ H(A)
On pose H = <ptG 6 H{H). On a if (0) = H'(0) = 0. Ainsi on se ramene
a montrer qu'il existe une seule fonction holomorphe h : D —)• C telle que
h(0) = 0, h'(0) = A et h - R*ah = if. Ecrivons Jf(z) = 5 3 a n z n , ft,(z) =
n>2
+oo
^ cn2:n. Done l'egalite h — R^h = H est equivalente a l'egalite ^ c n (l -
n>0 n=0
+00
anz ou encore
n=2
co = 0
ci quelconque
c
n = -L_e2J°yn-i)a pour n > 2 .
Soit m n l'entier relatif verifiant \(n - l)a — mn\ < \ pour n > 2. On a
1 _ e 2f*(n-l)a| _ | e iri(n-l)a _ e-7r*(n-l)ai _ 2| sin(7r(n — l ) a ) |
= 2|sin(7r[(n-l)a-mn])| .
En utilisant 1'inegaUte de Jordan l\x\ < \ sinx| pour |x| < f, on obtient
|1 — e 27r »(n-l)"| > 4|( n _ l ) a _ TOn| — 4( n _ I)|Q, _ _ 2— I .
Or a est un nombre diophantien, done il existe /3 > 0; il existe M > 0 tel

que | a — p/g| > -$p pour tout p/g € Q. Par consequent, on aura
d'ou
-i)«| - 4M
Ainsi la fonction /i(0) = N cn2;n, ou c\ arbitraire et Cn = a^?n-i)a pour
n > 2, est holomorphe sur D et verifie h — R*ah = H.

En conclusion, il existe une seule fonction holomorphe h S 'H(P) verifiant
h(0) = 0, h'(0) = A (ainsi ci = A) et h — R^h = H. Par suite, le lemme se
trouve demontre en considerant g = (p*h 6 H(A).
3.3. Dans ce qui suit, "on se place dans les hypotheses et notations du
theoreme 2: a etant done un nombre diophantien, f(z) = e2viaz+O(z2) une
fonction holomorphe dans un voisinage de 0, A son disque de Siegel maximal
en 0; >p designe une application linearisante de / en 0, et enfin (/e)ee/ est
une famille R-analytique de fonctions holomorphes sur A verifiant fa = f,
fe(0) = 0 et # ( 0 ) = e27rie pour 6&I.
236 Habib Jellouli
3.3.1. Proposition. Pour tout k € N, il existe un intervalle ouvert J C I

contenant a et une famille R-analytique {gkfi)eeJ de fonctions holomorphes
sur A verifiant gk,e(0) = 0, g'k6(0) = 1 et gkfi - f*egkfi = O{{9 - a)k)
uniformement sur tout compact de A.
D e m o n s t r a t i o n : O n &ipof = Raoipet R*a (idu) = ido- Done
ga = V*(idD) verifie ga-f*ga = 0. (2)
La fonction fgga est definie et holomorphe sur l'ensemble A* (9) = {z 6 A

tel que f$(z) 6 A et f'g{z) ^ 0}. Remarquons que A*(a) = A et si K
est un compact de A, alors K C A* (9) pour 6 assez voisin de a. On
en deduit d'apres (2) que ga - f%ga — O((6 — a)) uniformement sur tout
compact de A et par suite la proposition est vraie pour k = 0 et k = 1 avec
9o,9 = gi,e — ga = $•• Nous allons continuer la preuve par recurrence sur
keK
Supposons qu'on a gk,g - fggk,s = (9 ~ a)kGkfi ou (gk,e)eeJ est E-
analytique, holomorphe sur A, gkye{0) = 0, g'k9(0) — 1 et Gkie : A* (9) ->• C,
holomorphe sur A*(^), R-analytique en 9 £ J.
En particulier, Gk,a : A —> C verifie Gk,a(0) = G'k a(0) = 0; par suite,
d'apres le lemme 3.2 applique a G — —Gk>a, il existe une seule fonction
holomorphe hk}O : A -> C tel que /ifc,a(0) = 0, h'ka(0) = 1 et hk%a-f^hkta =
—Gkâ. D'oii Gkfi + (hk}a — fghkâ) = O{{9 — a)) uniformement sur tout
compact de A. Posons
+ g
U ( )kK (3)
pour 9 dans un petit intervalle ouvert J' C J contenant a. On a
,<? - fe9k+i,e = 1 + re_a\k \9k,e - fe9k,e + (9- a)k(hk,a -
= O{{9 — a)k+1) uniformement sur tout compact de A .

D'autre part gk+i,e(0) = 0 et g'k+ltg(0) = 1.
Ceci acheve la demonstration de la proposition.
3.3.2. Corollaire. Soit k € N et {gk,a)ee.J I® famille R-analytique de
fonctions holomorphes sur A donnee par la proposition 3.3.1. Si A' est
un sous-disque de Siegel distinct de A, alors il existe un intervalle ouvert
Perturbation d'une fonction linearisabie 237
J' C J contenant a tel que, pour tout 9 E J', il existe une seule fonction
<Pk,g : A ' —> C, holomorphe sur A ' et verifiant:
i) V e ( O ) = 0, <pfk e(0) = V ' ( 0 ) , Vkt9{z) # 0 pour z e A ' - { 0 } ;
" ) 5fc.e(2)<p'fc,e(2) = Vfc.eC'Z) V«ur tout z e A ' .
£>e p/ws, /a famille (ipk,9)eej' est R-analytique.
La demonstration de ce corollaire necessite le lemme preparatif ci-dessous
dont l'enonce est dans un cadre plus general et dont on aura besoin plus
tard dans la demonstration du theoreme 2.
3.3.3. L e m m e . Soient U un domaine simplement connexe de C, ZQ S U
et A £ C*. Si h : 14 —t C est une fonction holomorphe avec h(zo) = 0,
h'(z0) = 1 et h{z) ^ 0 pour z € U - {zo}» aXors il existe une seule fonction
i/> : U —> C holomorphe sur 14 et verifiant:
i) tl;{zQ) = 0, V'(zo) = A, tp(z) ? 0 pour zeU- {z0};
ii) h(z)ip'(z) = ^>(z) pour tout z eU.
De plus, si h varie dans une famille M.-analytique (hg)eî, alors la
famille {ipg)eel associee est aussi fit-analytique.
Demonstration :
— 14 etant simplement connexe, par suite la fonction
est definie, holomorphe sur W et on verifie facilement que la fonction ip

definie sur U par tp(z) — X(z — Zo)eH^ est solution du lemme 3.3.3.
— Unicite: supposons que rp : 14 —> C est solution du lemme. On pose
h(z) = (z - zo)hi(z) et tp(z) = {z - zo)ipi(z) done hj,(z) ^ 0 et V'i(2;) 7^ 0
pour tout z £14. O n aura
h(z)tp'(z) — tp(z) pour tout z £ U <=>
= (z — zo)V>i ( 2 ) pour tout z al4
pour tout z e u - {zo\
pour tout zEU

tpi(z) (z ~ zo)h(z)
t^- = H'(z) pour tout z e U
' - Q pourtoutzeW
238 Habib Jellouli
II existe c e C* = C - {0} tel que
ipi(z) = ceH<-^ pour tout z 6 U •&•

H
ip(z) =c(z- zo)e ^ pour tout z eU .
Mais ip'{zo) = A donne c = A, et par suite ip = tp.

— Si (hg)g£i est une famille R-analytique de fonctions holomorphes sur U
avec hg(z0) = 0, h'g(z0) = 1 et he(z) ^ 0 pour z eU- {z0}, alors pour 9 e /
iM*) = A(z - ô)exp( f l i ^ L - M l ^ ) pour tout z eU .
II est clair a partir de cette formule que la famille (ipg)eeI est aussi E-
analytique.
Demonstration du corollaire 3.3.2. D'apres la formule (3) dans la
demonstration de la proposition 3.3.1, on a:
9k,a = 9k-l,a = = 9l,a = 90,a = 9a = — •
tp(z)
On pose S = min .. . > 0.
z€dA' lfi'(z)
Comme lim gk e = 9k a = 9a uniformement sur tout compact de A,
alors il existe un intervalle ouvert J' C J contenant a tel que
\9k,e(z) -9c(z)\<S< \ga{z)\ pour z e dA' et 9 E J' .
On en deduit, d'apres le theoreme de Rouche, que pour tout 9 e J', on a

9k,e(z) ¥=0 pouv z e A' - {0}.
Maintenant, il suffit d'appliquer le lemme 3.3.3 a U = A', z0 — 0,
A = y'(0) et h$ = gkfi pour 9 G J'. La fonction fk,e : A' —> C est done
donnee par
Nous sommes maintenant en mesure de dormer la demonstration du

theoreme 2 enonce dans l'introduction.
3.4. Demonstration du theoreme 2.
— Soit A' c A' deux sous-disques de Siegel distincts de A. D'apres le
corollaire 3.3.2, il existe un intervalle ouvert J' C I (resp. J' C / ) contenant
a et une famille (fk,e)eeJ' (resp. (<fik,9)eeji) v£rifiant les proprietes du
corollaire 3.3.2.
Perturbation d'une fonction linearisabJe 239
Comme A' C A', alors d'apres l'unicite, on aura ipk,e/A' = <Pk,9 pour
6 e J' fl </'. En d'autres termes, pour tout compact K de A, on a K C
Dom(y>fcie) pour 6 assez voisin de a. Ceci demontre ii).
— D'apres le corollaire 3.3.2, la famille (<pk,o)eeJ' verifie 0 € Dom((pk,e),
Vfc,e(0) = 0 et ip'k9(0) = ¥>'(0) pour 0 € J'. D'autre part, on a gk,a = ga =
$-. Ceci donne gk,a{z)v' {%) = v(z) pour tout z e A . II en resulte d'apres
l'unicite dans le corollaire 3.3.2 que <pk,a = V- Ceci demontre i).
— Pour achever la demonstration du theoreme 2, il reste a verifier iii).
Soit K un compact de A. Considerons deux sous-disques de Siegel A',
A" verifiant K c A' C A" C A. Alors, il existe e > 0 tel que, pour tout
6 e]a - e,a + e[c J', on a:
a) A" c Dom(vfe,e);
b) A' C {Z e A" tels que fg(z) e A" et f'e{z) ± 0};
c) 9kAz) ± ° P° ur zeA"- {0}.
Ainsi on a, pour 6 e]a — e,a + e[, gk,e{z)ip'kg(z) = tfik,9(z) pour z € A";
en particulier fffc.e(/«(*))Vfc.eCM*)) = fk,e(Mz)) P o u r z e ^''• Done,
d'apres b) ci-dessus, on obtient
9k,e(fe{z)) , Ugofg\\z) = Ueofe)(z) pour ze A' ;
d ou
' \Je9kfij{z) • ytpk,e o / e j (z) = ^fc,s o /ej(^) pour z 6 A'. Or la
fonction z i-> (fggk,9)(z) est holomorphe sur A', et verifie (/e5fe,e)(0) = 0,
(fe9kfi\ (0) = 1 et (fggk,e){z)^0 pour z e A ' - { 0 } . II en resulte, d'apres

le lemme 3.3.3, que Ton a pour z 6 A'
D'autre part
fzt-(fS9kM)»_ [U-9kAZ),r . fZ 9M
Jo Z(fS9k,e)(t) Jo ZQkAZ) Jo 9*At)(fS9k,e)(t)
D'apres la proposition 3.3.1, on peut ecrire fffc,e(O — (f$9k,d)(0 = (8 ~
«)fcGjb,fl(O» o u G M holomorphe en ^ G A'(fl)' = {£ e A / / 9 ( O € A et
^ et R-analytique en 0. On pose
Jo ?5fc,fl(0 Jo 9kAZ)(fe9k,e){0
240 Eabib Jellouh
On aura
(vk,e o fa) (z) = e2™V(0)*eF*.«W • eV-°)k«».»(>) .
(<Pk,e°fe)(z) = ( ^ o V J M ) ( z ) + (eC-fl')t**.»W - l ) . ( i ? e o ^ M ) ( z ) (4)
ef 1
et on a lim = 1, done il existe (i > 0 tel que \t\ < n implique
l
\e - 1| < 2|t|. Les fonctions Rg o ipk e et Hkg sont uniformement bornees
sur le compact K, par suite il existe e' > 0 (e' < e), il existe une constante
M > 0 tels que |(iiflO¥>M)(z)| < M, | H M ( z ) | < M et \{0-a)kHk,e(z)\ < /i
pour z 6 if et 0 e]a-e',a + e'[. On en deduit, d'apres (4), lfk,e°fe)(z) =
f /?g o yjfcje j (2) + C((0 — a)k) uniformement sur K. Ainsi, le theoreme 2 se
trouve demontre.
3.5. Demonstration du corollaire 2. Soit K un compact de A. Con-
siderons un sous-disque de Siegel A' avec K C A' C A.
D'apres le theoreme 2, il existe e > 0 et M > 0 tels que, pour 8 g
]a - e,a + e[, on a (pky9 o / e = Re o yjfe e + (0 - a)kHktg, ou |iffci9(z)| < M
pour 2 € A .
D'autre part, d'apres le theoreme 1 (point i)), il existe n 0 6 N tel que
K C A'(9n,bn) et \8n — a\ < e pour n > n0. On en deduit que, pour
0 < j < bn et n > no, on a sur if les egalites suivantes:
<Pk,en o f6n = i? S n o ^fc,,9n + (8n - a)kHkfin
<pk,en o / | n = i ? | n o <pkfin + (9n - a)k \R6n o HktSri + Hktgn o f6n
(0n - a)k
Du fait que \Hk,gn(z)\ < M pour 2 6 A et f6 (z) € A' pour 0 < j < bn et
z € K, il en resulte qu'on a
1 Y, u 4 r ° ° ^*,9n o /ir i } ) («)i <jM<bnM. (5)

Or, par hypothese, bn — o(qnqn+i); done ilexiste A > 0 telque bn < Xqnqn+i
pour n e N. Maintenant, d'apres le preliminaire 1.1, on a qnqn+i < l^n -
a| - 1 . Ainsi on aura
(z)\ < XM\°n ~ a\~x pour z 6 K et n > n 0 .
1
Soit done \<Pk,en ° fg — Rg ° Vfe,en I < \M\9n — a\k pour z 6 K, n > n0
et 0 < j < bn. Ceci acheve la demonstration du corollaire 2.
3.6. Remarque. L'hypothese bn = qn pour tout n G N permet d'ameliorer
leeerement
&
la conclusion du corollaire 2. En effet, on a \6n — a\ < —-— <
> I n i qnqn+i
-y, done bn = qn < \9n - a^1!2. Par suite, d'apres l'inegalite (5) ci-dessus,
on obtient <fik,en ° fg = Rg ° Vfc,^ + C((^n — a) 2) uniformement sur
tout compact de A et uniformement par rapport a j avec 0 < j < qn.
4. Remarques sur la Conjecture
Rappelons la conjecture de Douady deja enoncee dans l'introduction :

Conjecture. Si a est un nombre diophantien et si (0n = pn/Qn)n&n est
o o
la suite des reduites de a, alors lirrin^+oo mes(K(Pgn)n K (Pa)) = mes(K
(P a )) ou on a note Pg(z) = e2ni9z + z2.
Dans ce qui suit, on demontre le theoreme 3 (qui permet de reduire la
conjecture ci-dessus), et on donne des motivations pour aborder la conjec-
ture reduite.
4.1. Demonstration du theoreme 3
Commengons par etablir les deux lemmes preparatifs ci-dessous et dont
on aura besoin encore plus tard.
4.1.1. Lemme. Soient U un domaine simplement connexe de C (autre
que C), LJ0 € U et f : U -> C holomorphe et injective sur U. Si K est un
compact dans U, alors il existe une constante fj, > 1 (independante de f) tel
que ^|/'(wo)| < |/'(w)| < n\f'(u)0)\ pour ui G K.
Demonstration : Considerons un isomorphisme tj) du disque unite D =
{z G C/|z| < 1} sur U tel que tp(O) = UJ0.
Ainsi les deux fonctions ip et g = / o tp sont holomorphes, injectives sur
D, par suite d'apres le theoreme de distorsion de Koebe ([Po], p. 21), on
obtient pour z £ l :
+|z|)3 - v(o) • - a - \z\r (1+\z\r - V(o) •

242 Habib Jellouli
On en deduit que pour z 6 D :
Or tp~1(K) est un compact dans D, done il existe r € [0,1[ tel que tp~1(K) C
D,. = {ze C/|z| < r}. D'ou on aura :
^ I/'MI < (izf) V'MI pour w e A\
II suffit de prendre fi = ( j ^ ) et le lemme se trouve demontre.

4.1.2. Lemme. Soient U,V deux domaines simplements connexes de C,
tjj un isomorphisme de U sur V et (t(>n) une suite de fonctions holomorphes
sur U, qui converge vers ip uniformement sur tout compact de U.
Si K est un compact dans V et U\ est un domaine simplement connexe
delimite par un lacet de Jordan dans U sans points doubles, verifiant K C
ip(Ui) ; alors il existe un rang no 6 N tel que pour n > no, on a :
i) ijjn est injective sur U-y;
ii) KdPniUl).
Demonstration : Considerons deux domaines simplement connexes U2,
U3 de U avec U\ C Ui ; Ui C Uz ; U3 C U et U3 delimite par un lacet de
Jordan sans points doubles. On pose
($1 = mf{\i>(z) - i>(z')\;z eUi et z' e U - U2}

S2 = ini{\i/)(z) - u\; z e dU3 et w € ^(W2)}
63 = iai{\ip(z) -w\;ze dUx e t w e i f } .
Alors il existe un rang n0 E N tel que pour tout n > n0, on a :
\rl>n{*)-1>{z)\<j pourzeWi (6)
|^n(z)-^)|<y pom ze6U3 (7)
\^n(z) - ^{z)\ no, on a :
\ipn(z) - ip(z)\ < \ip{z) - w\ pour z e dUz et w €

\$n{z) - tp(z)\ < \\j)(z) - UJ\ pour z e dUx et w € K .
On en deduit d'apres le theoreme de Rouche et le fait que tp est un isomor-

phisme que pour n > n 0 , on a :
pour tout w e ip{U2), il existe un seul z € U3 tel que ipn(z) = w , (9)
pour tout u> 6 K, il existe un seul z 6 U\ tel que tpn(z) = w . (10)

En particulier (10) implique que K C tjJn{Ui) pour n > no, d'ou ii) est
demontre.
Pour n > no, ipn est injective sur Wi, en effet :
Soit zj_,z2 e Wi tel que ipn(zi) = ^(22) = w e ipn(Ui)- D'apres (6),
on a if>n(Ui) C ip(lh), et on a Wi C W3 par hypothese. Ainsi on obtient
21,22 € W3, w e ip(U2) avec Vv»(zi) = $n{z2) = w. II en resulte de (9) que
z\ = z ii).
* Montrons que ii) =$> i) :
Designons par A le disque de Siegel maximal pour Pa en 0 et soit fi une
o
composante connexe de K (Pa)-
On sait qu'il existe k € N tel que la restriction ga de P% a f2 est un
isormorphisme de Q sur A. Posons w0 = g^1^) 6 0. Soit A' un sous-
disque de Siegel arbitraire dans A avec A ' / A et posons fl' = g~1(A') C ft.
Considerons deux sous-disques de Siegel auxiliaires A", A'" de A tel
que A' C A" C A'" C A et posons Q" = gzH&"), 0'" = ff-Â"').
La suite de fonctions (P«n )ngN converge vers ga uniformement sur tout
compact de Q. On en deduit d'apres le lemme 4.1.2 q'uil existe un rang
no € N tel que pour n > no, on a :
P*n est injective sur Q!"\ (11)
A'c.P* n (ft")C A . (12)

D'apres (11), on peut appliquer le lemme 4.1.1 a P*n sur Cl"', par suite il
existe une constante fi > 1 (independante de n) tel que :
Or lim (P^ )'(vo) = (f^)'(ô),

n
par suite il existe un rang ni > no tel
n—>-f*oo
que pour n > ni, on a :
< 2|(Pafc)'(u;o)| (14)
244 Habib Jellouli
on
En posant fn = 2/i|(P*)'(wo)|, /fc = ^ ( i ^ ' M I , obtient d'apres (13)
et (14) que pour n > ni, on a :
M2 < K ^ J ' H I ni :
M >es(((C - K(PaJ) nfi")< mes(PQfcn((C - K(PaJ) n fi")
< ^mes((C - if(Pa n )) D 0") , (16)
d'autre part P ^ ( ( C - /JC(PQn)) n «") = (C - K{PaJ) n P* n (n") car
P - ^ C - i f ( P Q J ) = C - i f ( P a J = Pan(C - K(PaJ).
On en deduit d'apres (12) et (16) que pour n > rii, on a:
mes((C - K(PaJ) n A') < nlmes((C - K(PaJ) n fi) (17)

mes((C - -ftT(PaJ) D fi') < M2"2mes((C - A-(P a J) n A) (18)
d'ou de (17) et (18) on tire que pour qu'une composante connexe Q, de
K (Pa) verifie lim mes(ii:(P a ) n f l ) = mes(12) il faut et il suffit que A
n—^+oo
verifie lim mes (K(Pa) n A) = mes (A).
n—>+oo
Ainsi ii) implique qu'on a lim mes (K(Pan)C\A) = mes (A) et par suite
n—>+oo
o
pour toute composante connexe O,- de K (Pa), on a : lim uo.es(K(Pan) (1
o
Ecrivons K (Pa) = [j £lj, on a :
lim mes(i<:(P a Jn.fi:(P Q ))= lim V mes (K(PaJ D fi,-). (19)
Remarquons que mes(K(Pan) DClj) < mes(f2j) pour tout j , n 6 N avec de

plus la serie y^mes(fij) converge (vers mes (if (Pa))).
Par consequent on peut permuter les symboles " lim " et "V^" dans
n—>+oo ~—J
(19) et on obtient :
lim mes(K(Pan)n K (Pa)) = Vmes(Oj) = m e s ^ (Pa))
d'oii ii) => i) et le theoreme se trouve demontre.

4.2. Motivation de la conjecture reduite.
Comme on l'a deja mentionne dans l'introduction, le travail qui va suivre
est une tentative pour aborder la conjecture reduite (equivalente a la conjec-
ture de Douady d'apres le theoreme 3), c'est-a-dire : lim raes(K(Pen)n
n—y+oo
A) = mes(A) ?
Le modele obtenu pour realiser cette etude est une fonction holomorphe,
injective Gn : H ->• C (ou H = {w G C/Re u) < 0}), de la forme Gn(w) =
uj+\ne"+hn(uj) (qui en un sens on pretend qu'elle possede un comportement
dynamique comparable a celui de la fonction w i-> u) + \new)-
Commengons par demontrer deux lemmes, pour se faire la main :
4.2.1. Lemme. Soit P(z) = e27™? z + a2z2 avec 0 ^ a2 6 C, p et q etant
deux entiers premiers entre eux. Alors I'iteree glfeme de P est de la forme
pi(z) = z + aq+lzi+1 + O{zi+2) avec aq+1 / 0.
Demonstration : Voir [Jl], §2, Lemme 2.
+OO
4.2.2. Lemme. Soit f(z) = z + aq+iz q+1

+ \_, a zk
k avec a
q+i ¥" 0> e*
k=q+2
+ OO
f(z) = fj,\z+ 2_] (J-kzk avec /iiÔ, deux fonctions analytiques au voisinage
k-2
de 0. Alors dans un petit voisinage de 0 la fonction F — ip o / o ip"1 est de
la forme :
Demonstration : II s'agit d'un calcul facil en partant de l'identite F o ip =
4.2.3. Notations et changement de coordonnees.

Soient a un nombre diophantien, (8n = pn/qn)nen â suite des reduites
de a dans le developpement en fraction continue, Pg{z) = e2?Tiez + z2, k > 2
un entier naturel donne et ((pk,e)eeJ qu'on notera tout simplement (<ê)eeJ
la famille R-analytique donnee par le theoreme 2 (avec a € J) et associee
246 Habib Jellouli
a la famille (Pe)eeR- A designe toujours le disque de Siegel maximal de Pa

en 0, D = {z e C/\z\ < 1} une application linearisante de Pa en
0.
Fixons dans tout ce qui suit r e]0,1[ et considerons deux reels auxiliaires
ri, r 2 verifiant 0 < r < r\ < r 2 < 1.
On pose Aj = p " 1 ^ . ) c A avec Brj = {z € C/\z\ < r,} pour j €
{1,2}. Comme P a : A —> A et ip : A —>• D sont deux isomorphismes avec
lim Pgn = P a et (d'apres le theoreme 2) lim <pa = tp uniformement
w
n-»+oo n—>+oo «
sur tout compact de A, alors il en resulte d'apres les theoremes 1,2 et le
lemme 4.1.2, qu'il existe un rang n 0 € N tel que les proprietes suivantes
sont vraies pour n > no :
A1cA2(0n,<?n); (20)
A 2 cDom( V ( ! J; (21)
et sont
Pdn PQ injectives sur A 2 ; (22)
DrC^JAx). (23)
Ainsi pour \z\ < r, on a y>^"1(z) 6 Ai est bien defini, Pg ((ps~1(z)) € A 2
pour 0 < j < qn et les proprietes (21) et (22) montrent que les applications
z I—v [ipe o P3Q o ipe~l){z) sont holomorphes et injectives sur D r (pour 0 <
3 < Qn)-
On en deduit d'apres la remarque 3.6, qu'il existe une constante M > 0
et un entier ni € N tel que pour \z\ < r et n > ni, on a :
6n n < M\6n - a\k~2 . (24)
On note Fgn (z) = (ips °Pgn o<p^"1)(z) pour \z\ < r et n > no. Remarquons
que d'apres le lemme 4.2.1, on peut ecrire :
Pfc (z) = z + aqn+1z^+1 + O{zqn+2) avec aQn+1 # 0.

Et en utilisant le lemme 4.2.2, on obtient pour \z\ < r :
1o(zqn+2) (25)
•
Soit TVo > max(n o ,n 1 ) tel que \0n — a\k 2 < ^£_ pour n > iV0.
Introduisons une nouvelle coordonnee u> € H = {u G C/Re w < 0} par
^ = IIn(w) = re?" et definissons la fonction Gn(w) pour w € H et n > iV0
par :
>^ / \ . 1 1 •* t? n v y » T-F / \ /o^\
Perturbation d'une fonction iinearisabie 247
Dans cette formule log designe la branche principale du logarithme (i.e.

-7T < Im log(.) < 7r).
4.2.4. Proprtetes.
Pour n > No, on a les proprietes suivantes :
(•Pi) Gn : HI -¥ C est definie, holomorphe sur HI et verifiant :
n n o Gn = F9n o n n
(P2) Gn est injective sur i et on a :
Gn(u) + 2ftiqn) — G n (w) + 2niqn pour tout w e H
(•P3) Gn(ui) = u> + O((9n - a)*" 1 ) uniformement sur M.

(P 4 ) Gn(u) - w ~ Anew pour R e w - ^ - 0 0 ; ou An = qnaqn+i(-^j^)qn• De
plus on a : 0 < |A n | < 2.
(P5) Gn{w + 2ni) = Gn{u)) + 2m + O((6n - a)2k~2) uniformement sur tout
demi-plan H = {w€ C/Re w < -£} avec I > 0.
Demonstration :
): D'apres l'inegalite (24), en prenant j = qn, on obtient :
\F9n(z) - z\ < M\9n-a\k~2 pour|z|<r etn>7V0.
Par suite d'apres le principe du maximum, on aura :
2
-1|<—\0n-a\ <\ pour |z| < r et n > AT0. (27)
z r
Ainsi
avec z = II n (w) 6 D r , done Gn est definie et holomorphe sur H. De plus,
(n n o Gn)(w) = re^f1 = ret: . ^M = F9n (z) = (F$n o nn)(w) .

z
): du fait que n n (w + 2iriqn) = n n (o)), on en deduit immediatement que

d ( w + 2iriqn) = Gn(ui) + 2Triqn pour tout u> G H. Montrons que G n est
injective sur HI : Soit a>i,w2 € H tel que Gn{u>i) = Gn{u)2).
II en resulte de la propriete {P{) que Fgn(Un(uJi)) = Fe n (n n (w 2 )),
mais Iln (a>i), n n (w2) 6 D r et F$n est injective sur D r , done on doit avoir
n n (wi) = II n(w2) et il existe m € Z tel que w\ = u)<i + 2/nimqn. Ceci donne
248 Habib Jellouli
Gn(wi) = Gn(u2 + 2nimqn) ou encore Gn(uii) = Gn(u2) + 2nimqn, d'ou

m — 0 et oj\ = w2.
(P 3 ): Tout d'abord on a : | log(l + z)\ < 2\z\ pour \z\ < ^, en effet
(-1)™- 1 — pour|z|<l;
m.
m
m=l
+00 +00
done I log(l + z)\ < ^ z m

\\ < \ \ ^2 ( i ) m = 2\z\ pour \z\ < \.
z
m=\ m=0
e
D'autre part, Gn(uj) = w + qnlog (l + ( - 1J J avec z — II n(w)
et d'apres (27), on a : |— 1| < 5 pour \z\ < r. Ceci entraine que :
\Gn(w) -u\< 2qn\Fen^Z' - 1| < —qn\6n - a|fc~2 (d'apres (27))

z r
or qn < \6n — a\~2 et par suite on obtient :
2M
\Gn(w) — u>\ < \9n — a\ pour u 6 H et n > No ,
r
e'est-a-dire Gn(u) = u + O((9n — a)*" 1 ) uniformement sur H.
(Pi): D'apres (25), on a : 6n

^' - 1- "fntv? z^ Pour z
~+ °-
Comme Gn(u>) — w = qn^>g (1 + (— 1 j J avec z = Un(uj), alors

on aura pour Re w -> - 0 0 :
r t, A ,, n (Fe^z)
II est clair que |An| > 0 car a 9 n + 1 ^ 0. Montrons que |An| < 2 :
On pose S = {/i: D —>• C, holomorphe, injective avec fe'(0) — 1 = /i(0) =
0}. Un resultat sur la theorie des fonctions univalentes (voir [Go] page 494)
perniet de dire que si h(z) = z + s m + 1 z m + 1 + sm+2zm+2 +... est un element
de 5, alors on a |s m +i| < ^ e t l' m egalite est optimale. Maintenant d'apres
(25) l'application :
est un element de la classe S et par suite on obtient :
ou encore |An| < 2.

(Ps): * Etape 1. Considerons Fgn(z) = ((pg oPJj^1 oip~*)(z) pour |z| < r
et n > No. D'apres l'inegalite (24), en prenant j = qn-i, on obtient :
\FOK{Z) - e2™3j^T"z\ < M\en - a\k~2 .
Compte tenu de l'egalite qnpn-i -Pn<7n-i = ( ~ l ) n (voir preliminaires 1.1),

alors on a :
qn
( 1
et on aura \F6n(z) - e2™ "- z\ < M\6n - a\k~2 pour \z\ < r et n > No.
Par consequent, d'apres le principe du maximum, on a :
1| < |6>a|fc~2 ^ pour \z\ < r et n > NQ.
Ainsi la fonction G n : H —> C donnee par :
avec z
Gn(u)) = LJ+ (-l)n+12ni + qn log ( *2»J+i ) = n n(w)
est definie et holomorphe sur EL De plus on voit facilement qu'elle verifie

l'egalite II n o Gn = Fgn o II n .
* Etape 2. Soit £ > 0 et Bfe = {w € C/Re w < - ^ } .
Posons e n = ^jr\6n — a ^ " 1 et soit Ni £ N tel que iVi > iVo et en < |
pour tout n> N±.
Ceci entraine que si n > TV!, on a : G n (E^) C El et Gn(Et) C H.
Montrons que pour n > Ni, on a : G n o G n = G n o G n sur E^. On a :
n n o (G n o Gn) = ( n n o G n ) o G n = F S n o ( n n o Gn) = (F6n oFeJo Hn
de meme II n o (G n o Gn) = (-Pen ° -fen) ° n n , or Fgn o i*gn = Fgn o Fgn car

Pa" oP?"" 1 = P, 9 "^ 1 oP«». Done II n o (G n oG n ) = II n o (G n o G n ) et on en
250 Habib Jellouli
deduit qu'il existe m £ Z tel que ( G n o Gn)(uj) = (Gn o Gn){u)) + 2Trimqn

pour tout w g B j . O n peut ecrire pour tout u £ H :
2nimqn = [Gn(Gn(w)) - Gn(w)} + [Gn{u>) - Gn{w)] + [Gn(u) - G n (G n (w))].

(28)
D'autre part on a :
lim Gn(uj) — IJJ = 0 et lim Gn(u>) — w = (—l)n~*~127ri ,

Re u—¥—oo R^ u—I — oo
maintenant par passage a la limite quand Rew —>• —oo dans (28), on obtient
m = 0 et par suite Gn o Gn = Gn o G n sur Bl^.
* Etape 3. Posons
W = qn log (*M) et sn(U) = Qn log
OXL z = Un(uj) = re^ pour u € H et n > Aî.

En d'autres termes : Gn(w) = w + Hn{w) et Gn(w) = w+ (-l) n + 1 27ri +
Hn{w). D'apres la demonstration de la propriete (P 3 ), on a :
|-Hn(w)|<e n pour w 6 H et n > Ni (29)
un meme raisonnement donne :
\Hn{w)\ < en pour w 6 H et n > iVi. (30)
On en deduit d'apres la formule de Cauchy que pour w € H,
< ^ • (32)
* Etape 4. Pour w 6 Hi et n > Ni, on a:
G n (G n (w)) = Gn(w) + (-l) n + 1 27ri + Hn(Gn(u))

= w + Hn(w) + (-l)n+12m + Hn{w)
ou on a pose ^n(u>) = Hn{u + Hn(ui)) — Hn{w).

En utilisant la formule suivante
~ r1 _
Hn(w + a) - Hn(w) = o / H'n(w + at)dt
Jo
on obtient : _ __
\i>n(w)\<\Hn(0j)\- SUp |<(0|.
I€-W|<£n
On en deduit d'apres (29), (32) et le fait que e n < | , qu'on a :

2e2
\4>n(u)\ < —f- pour w eMe etn>Nx. (33)
D'autre part, on a:
Gn(Gn(w)) = Gn(u>) + Hn(Gn(u>))

= u + (-l)n+12iri + Hn(uj) + Hn(w + (-
ou on a pose
iMw) = Hn(u + (-l) n+1 27ri + Hn(w)) - Hn{w + (-
Un raisonnement analogue utilisant (30) et (31), montre :
2e2
IV'nHI < -y pour w e Me et n > Nx. (34)
Maintenant en vertu de 1'identite Gn o Gn = Gn o G n , on obtient :
Hn(u)+(-l)n+12m)-Hn(uj) = %j}n{w)-i>n{w) pour w € Me etn>Ni.
Par suite d'apres (33) et (34), on a \Hn(u + (-l)n+12wi) - Hn(w)\ < *&
pour w € Me et n > Ni. On en deduit que \Hn(w + 2TTZ) - Hn(u)\ < ^f*
pour w E Me et n > Ny. Ceci exprime qu'on a Gn(w + 27ri) = Gn(a;) + 27ri +
O((9n - a)2k~2) uniformement sur Mf.
4.2.5. Quelques questions
Soit A' C A un sous-disque de Siegel de Pa en 0 et notons par A' son
symetrique par rapport au point critique a> = -le2via de Pa. Designons
par ^l(Pgn) la composante connexe de K(Pgn) - {0}, contenant le point
critique wn = -±e2nie" de Pgn.
Question 1. Existe-t-il un rang no € N tel que :
(K(P6J - n ( P O ) H A' = <j> pour tout n > n 0 ?
Au vu de la definition de la fonction Gn, alors on peut montrer qu'une

reponse positive a la question 1, entraine une reponse positive a la question
suivante :
252 Habib Jellouli
Question 2. Existe-t-il un rang iVo 6 N tel que pour tout n > No, la
fonction Gn n'a pas de points fixes dans H ?
Remarque. On pretend que les cinq proprietes de 4.2.4 et une reponse
positive a la question 2 ci-dessus sont suifisantes pour poursuivre cette etude
et aboutir a une reponse positive a la conjecture de Douady enoncee dans
l'introduction.
Remerciements.
Je remercie le Professeur A. Douady qui m'a propose ce sujet, ainsi que
le Professeur J.H. Hubbard pour ses remarques tres utiles.
Bibliographie
[Go] G.M. GOLUZIN, Geometric theory of functions of a complex variable,
vol. 26, Translations of Mathematical monographs, A.M.S. 1969.
[Jl] H. Jellouli, Indice holomorphe et multiplicateur, dans ce volume.
[Po] C. POMMERENKE, Univalent functions, Vandenhoeck and Ruprecht,
Gottingen, 1975.
Institut preparatoire aux etudes d'ingenieur de Nabeul, Departement de

Mathematiques et d'Informatique, Merazka, 8000 Nabeul, Tunisie.
INDICE HOLOMORPHE ET MULTIPLICATEUR
HABIB JELLOULI
Resume : Soit P\0{z) = Xoz + z2 ou Ao = e2™*^, (p,q) £ Z x N*, p et

q etant premiers entre eux et z £ C. II est bien connu qu'au voisinage du
point fixe parabolique 0, l'iteree g-eme de P\ o e s t analytiquement conjuguee
a une fonction de la forme z ^ z + zq+1 + Az2q+1 + O(z2q+2). Le nombre
complexe A est un invariant de conjugaison analytique.
On considere des perturbations de P\Q dans la famille des polynomes
quadratiques P\{z) = Xz + z2 ou A est dans un petit disque D(\Q,r) centre
en Ao et de rayon r.
Pour r assez petit, le multiplicateur p(X) du cycle de longueur q le plus
proche de 0, definit une fonction holomorphe en A 6 D(X0, r). On utilise la
notion d'indice holomorphe pour donner les valeurs des derivees premiere
et seconde du multiplicateur en Ao, et par des proprietes dynamiques des
P\, on donne un encadrement de la partie reelle du nombre complexe A.
Enfin on utilise ces resultats dans l'etude d'explosion des cylindres de Fatou-
Ecalle.
Abstract : Let PXo(z) = Xoz + z2 where Ao = e2irip'q, (p,q) E 1 x W,

(p> q) = 1 and z € C. It is well known that the g-th iterate P%o of P\o is
analytically conjugate to a holomorphic function of the form z H-> z + zq+1 +
Az2q+1 + O(z2q+2) in a neighbourhood of 0, where A is a complex number
invariant under analytic conjugation.
We consider perturbations of P\o in the quadratic polynomials P\ (z) —
Xz + z2 with A 6 B(A0,r), r small.
For r sufficiently small, the multiplier p(X) of the g-periodic cycle close
to 0, is holomorphic as a function of A 6 £>(Ao,r). We use the holomorphic
index to calculate the values of the first and second derivatives of p(X) at
Ao. Furthermore we bound the real part of A from above and below, using
the dynamical properties of P\. Finally we use these results in the study of
the explosion of Fatou-Ecalle cylinders.
253
254 Habib Jellouli
1. Notion d'indice holomorphe
Soit U un domaine de C, z0 6 U et / : U -> C une fonction holomor-

phe pour laquelle ZQ est un point fixe (f{zo) — ZQ). On definit I'indice
holomorphe (voir [M]) de / au point ZQ et on note i(f, ZQ) comme etant le
residu
. . . s i r i
ou le cercle d'integration \z — zo\ = e est oriente positivement et e choisi

assez petit verifiant f(z) — z ^ 0 pour 0 < \z — ZQ\ < £.
Le lemme suivant dont on aura besoin exprime essentiellement que
I'indice holomorphe est un invariant de conjugaison analytique
Lemme 1.
1) Si ZQ est un point fixe simple de f, alors on a
2) Si ZQ est un point fixe de f et F = ho f o h~x ou h est holomorphe,

injective au voisinage de zo, alors on a
i(f,zo)=i(F,h(zo)) .
Preuve : 1) Par hypothese, f(z) = z0 + ax(z — z0) + a2(z — zo)2 +

avec ai ^ 1. Ceci donne le developpement en serie de Laurent
+
1 11 °°
v +
77-7 = 7; r,z - zo)v +J2bk(z-z0)k ;
- f(z) (1 - ai)(z - zo) ~
par suite
2) ler cas : /'(z 0 )^l.On a: F(h(z0)) = h(z0), F'(h(z0))

D'apres 1), on obtient
i{f Zo) = i{Fh{))

' T^
2e cas : f'(z0) = 1. Designons par m > 2 l'ordre de multiplicity de z0
comme point fixe de / . Soit s > 0 tel que f(z) — z ^ 0 et f'(z) ^ 1 pour
0 < | z - z o | < e.
Posons
c= inf \f(z) -z\>0 ,
\z-Z0\=£
ft(z) = f(z) + t , Ft = ho ft oh"1 pour \t\ < (t < % ;
ceci donne \ft(z) — z\ > -| pour \z — ZQ\ = e.

D'apres le principe de l'argument, le nombre N(t) des points fixes de ft
a l'interieur du cercle j£ — {z/\z — ZQ\ = e} est donne par
J_ / iMzldz .
2m J7e ft(z) - z
D'apres cette formule, N(t) est une fonction continue et est a valeurs dans
N, done constante. Compte tenu du fait que N(0) = m, alors N(t) = m
pour \t\ < fi. Designons par Pi(t),P2(t),...,/3m(t) les points fixes de ft
avec &(<)) = /32(0) = ... = ^ ( 0 ) = z0, et on a pj(t) ± z0, //(^(t)) =
f'(f3j(t)) ^ 1 pour 0 < |«| < JU et 1 < j < m.
Par suite, d'apres le ler cas, on a i(ft,(3j{t)) = i(Ft,h(Pj(t))), ce qui
donne
soit
z - ft(z) 2ni fhhm) z - Ft{z)
Comme les contours d'integration j£ et /i(7e) sontfixes,il est clair que les
deux membres de (1) sont continus en t pour |t| < fj. et par suite, par passage
a la limite, quand t tend vers 0 dans (1), on aura i(f, ZQ) = i(F, h(zo)).
2. Polynomes quadratiques et multiplicateur
On considere la famille (PA)ASC de polynomes quadratiques definis sur

Cpar P\(z) = Xz + z2.
On note P " la n-eme iteree de P\. On appelle cycle periodique, de
longueur m 6 N* pour PA, une suite de points distincts {zo, z i , . . . , zm-i\
telle que zj = P\{ZJ-I) et z0 = P A (z m _i).
Ainsi, P™(ZJ) = Zj pour 0 < j < m—1, et la meme valeur est obtenue en
evaluant la derivee (PjJ")' en chaque point Zj du cycle (en efiFet (P™)'(z0) =
r i t o 1 P'M) = (P?)'(zj) Pour 0 < j < m - 1).
Cette valeur commune p est appelee le multiplicateur (ou encore la valeur
propre) de ce cycle.
256 Eabib Jellouli
Un cycle de multiplicateur p est dit attractif (resp. repulsif; resp. in-

different; resp. indifferent parabolique; resp. superattractif) si \p\ < 1 (resp.
\p\ > 1; \p\ = 1; p = e2iria avec a e Q; p = 0).
Dans toute la suite, on fixe Ao = e2'Kip/q ou (p, q) e Z x N* avec p et g
premiers entre eux.
Lemme 2. L'ordre de multiplicite de 0 comme point fixe du polynome Pqo
est{q+l).
Preuve : P\o (z) — e2™p/qz+z2. Soit m l'ordre de multiplicite du point fixe
0 de Pqg (2 < m < 2«); done on a : Pqo(z) = z + amzm + am+1zm+1 + ...
avec am^0.
II s'agit de montrer que m = q + 1. En vertu de l'identite Pqo o Pô =
P\ o ° P\o) on obtient
z2) + am(e2™p/iz + z2)m + O(zm+1)

+ amzm + ...) + {z + amz
En egalant les termes en zm, on aura
ame2*imp/q = e2nip'qam , ce qui donne e 2«("»-i)p/« = i .
par suite, il existe qi G N* tel que m = q^q + 1.

Montrons que q\ — 1. On a: P\0{z) = z(l + aqiq+izqiq + . . . ) avec
a
qiq+i 7^0- {z £ C/a g i g + iZ 9 1 9 G K_} definit qiq demi-droites appelees axes
d'attraction de Pqo.
Pour chaque axe d'attraction L, l'ensemble BL = {z G C tel que la suite
P™q(z) tend vers 0 suivant la direction L} est un ouvert, et il existe une
composante connexe Ai de Bi qui contient, pour chaque z G Bi, les termes
P™q(z) pour n assez grand. A^ est appele un bassin parabolique attache a
0 (car 0 G Aj_) et ces bassins sont en nombre gig. De plus P\0(AL) — Aî,
oil L' = X0L = {Xnz,z E L\; d'ou P? (Ar) = AL et les gig bassins se
repartissent en gi cycles. On sait que chaque cycle doit contenir au moins
un point critique de P\o. Or P\o admet un seul point critique et par suite on
a un seul cycle de bassins paraboliques attaches a 0, autrement dit gi = 1.
Soit done m — q + 1, et ceci acheve la preuve du lemme 2.
Lemme 3. // existe e > 0 et r > 0 tels que, pour tout A G C avec 0 <
| A — Ao | < T, le polynome Pq admet exactement q points fixes dans le disque
pointe 0 < \z\ < e; de plus ces q points forment un cycle de longueur q pour
Preuve : P\0(z) = Xoz + z2 ou Ao = e2v%p^q. Remarquons que 0 est un

point fixe simple de P*o pour 1 < k < (q — 1), tandis que, d'apres le lemme 2,
0 est un point fixe de PjJo de multiplicite (q + 1). Comme les zeros d'une
fonction holomorphe (non identiquement nulle) sont isoles, alors il existe
e e]0,1[ tel que P f o ( z ) - z ^ 0 pour 0 < |z| < 2e et 1 < k < q. On pose
-ye = {z € C/\z\ = e} , 72e= {z e C/\z\ = 2e} ,

c = ^f {\P^(z) - z avec z€jeUj2e}>0 .
II existe r e]0,1—e[ tel que \P£(z)-z\ > | pour |A—Ao| < r, z £ 7£U72e
et 1 < k < q. D'apres le principe de l'argument, le nombre JVfc)£(A) (resp.
-/Vfc,2e(A)) des points fixes de P* dans le disque \z\ < e (resp. \z\ < 2e) pour
| A — Ao| < r et 1 < k < q est donne par
D'apres ces formules, il est clair que les fonctions Nk,£(x) e t -Wfc,2e(A)
sont continues en A et a valeurs dans N par suite constantes. Ainsi Nk,e(X) =
Nk,s(xo) et Nk,2i:(x) = Nk,2e(Xo) pour |A — Ao| < r et 1 < k < q. Soit done
%AX) = Nk,2e(X) = 1 Sll<k<q-1
On en deduit que 0 est le seul point fixe de P* (1 < k < (q - 1))

dans le disque \z\ < 2e, tandis que P^, pour 0 < |A — Ao| < r, admet 0
comme point fixe simple et q points fixes dans le disque pointe 0 < \z\ < e.
Remarquons que P ' , pour |A — Ao| < r n'admet pas de points fixes dans la
couronne e < \z\ < 2e (car iV9>e(A) = Nqi2e{X) — q+ 1). D'autre part, on a
< e => \P$z)\ < 2s; en effet
= \Xz + z2\ < |A| \z\ + \z\2 < (1 + |A - Ao|)|z| + \z 2
< (1 + r)e + e2 < (2 - e)e + e2 =2e (car r < 1 - e) .

Ainsi, si ZQ designe l'un des points fixes de P% avec 0 < \ZQ\ < e, on aura
Zj = P3X{ZQ) pour 0 < j < q — 1, tous des points fixes de PjJ distincts deux
a deux dans le disque pointe 0 < \z\ < e; par suite les Zj, 0 < j < q — 1,
representent les q points fixes de P% pour 0 < |A — Ao| < r et forment bien
un cycle de longueur q.
Notation : Pour 0 < |A - Ao| < r, on note par p(X) le multiplicateur du
cycle periodique de longueur q pour PA dans le disque pointe 0 < \z\ < e
(donne par le lemme 3).
258 Habib Jellouli
Lemme 4. p : A H-» p(A) est une fonction holomorphe sur le disque |A —

Ao| < r; de plus p(Xo) = 1.
Preuve : Pour 0 < |A — Ao| < r, on designe par /3i(A),/? 2 (A),... ,/?q(A) les
q points distincts du cycle de longueur q dans le disque pointe 0 < \z\ < e
donne par le lemme 3. On suppose f3j($ — P$l3j-i(X)) pour 2 < j < q.
On construit le polynome
Qx(z) = (z- px{X)){z - /32(A)) (z- (3q(X))
q 1
= z + bq-xiX)!?- + + &i(A)z + 60(A)
ou, pour 1 < k < q, bq_k($ = ( - l ) f c J 1 j 2 J k g
Pjk(X) sont les fonctions symetriques elementaires des points /3j(X)(l <
3 < q)- On considere les q polynomes symetriques de Newton en 0j(X)
(1 < j < q) definis par
'fc(A) = E ( ^ ( A ) ) f c , l<k<q .
3=1
On sait que chaque fonction symetrique elementaire s'exprime comme

un polynome en les polynomes symetriques de Newton
D'autre part, d'apres le theoreme du residu, on a pour 0 < |A - Ao| < r

etl<k<q
) l ,
dz .
Par suite, A (-* <Tfe(A) est holomorphe sur le disque |A — Ao| < r avec
<rfe(Ao) = 0. On en deduit que A \-¥ 6g_fc(A) est holomorphe sur le disque
|A - Ao| < r avec bq-k(X0) = 0 et Q\0(z) — zq. On a
p(A) = (i^WA)) = flni0iW) = Il(A + 2/3i(A)) = {-2YQx{-\).

3=1 3=1
II en resulte que A i-» p(X) est definie et holomorphe sur le disque |A —

Ao| < r avec p(X0) = ( - 2 ) « Q A o ( - ^ ) = Ag = 1.
Remarque : II est naturel de penser a p(X0) = 1 comme etant (P^ o )'(0)
ou dans ce cas le cycle de longueur q est forme de points tous egaux a 0.
3. Derivees du multiplicateur et l'indice holomorphe

Dans tout ce paragraphe, on conserve les notations utilisees en 2.
Indice holomorphe et multiplica,teur 259
q + q
Proposition 1. p'{X0) = - | - et p"(X0) = Â ~ ^ ou A =
X A
Preuve : A i->- p(X) etant holomorphe sur |A — Ao| < r avec p(Xo) = 1, done
il existe 0 < r' < r verifiant p(A) / 1 pour 0 < |A - Ao| < r1. Pour e donne
par le lemme 3 et |A — Ao| < r', on pose
II est clair que A H^ r(A) est holomorphe dans le disque |A - Ao| < r'
avec T(A 0 ) = i{Pxo,0) = A; par suite r(A) = A + 0{X - AO). D'autre part,
d'apres le theoreme du residu, on a pour 0 < |A - AQ| < r'
T(A) =
+ T-T (d'apres le lemme 1) .

1-A9 l-p(X)
Ceci donne
P{X) = X
- 1 ;A«-I)T(A)
= 1 - q | T - (A - Ao) + ^ " ^ ( A - A0)2 + O((A - A 0 ) 3 )|
fl - ( | - ( A - Ao) + O((X - A o ) 2 )) (yl + O(X - Ao))|
A — Ao) H r-r-o—(A — AQ) + O((A — Ao)
= 1 - g ( A - Ao) + g +
^ 1 }
( A - Ao)2 + O((A - Ao)3) .
D'ou la proposition.
Proposition 2. Etant donne P\0{z) = e27rip/qz + z2, alors il existe une
fonction analytique ip injective au voisinage de 0 avec <p(0) = 0, tel que
260 Habib Jellouli
FXo=cpo Plo o tp'1 est de la forme FXo (z) = z + zq+1 + Az2q+1 + O(z2q+2).
Le nombre complexe A est un invariant de conjugaison analytique verifiant
Preuve : D'apres le lemme 2, on peut ecrire P% (z) = z + azq+1 +... avec

aÔ. Dans ces conditions, il est bien connu ([Be]) qu'il existe une fonction
analytique ip injective au voisinage de 0, <p(Q) = 0, tel que F\o = v°P\ oip'1
prend la forme FAo(z) = z + zq+1 + Az2i+1 + O(z2i+2).
Un calcul simple donne A = i(J*\ o ,0); done, d'apres le lemme 1, on a
aussi A = i(P%o,0) qui est un invariant de conjugaison analytique.
La formule de l'indice holomorphe (voir [M]) pour une fraction ra-
tionnelle / : C ->• C, non identiquement egale a l'identite, exprime que
j
f(z)=z, zee
Le cas d'un polynome Q, oo est un point fixe superattractif, done d'apres
le lemme 1, on a i(Q, oo) = 1 et on aura
Q(z)=z, zee
Dans notre cas, P$z) = Xz + z2 et l'iteree g-eme P^ admet 2q points
fixes (non necessairement distincts).
Conservons les notations du 2) et designons, pour 0 < |A — Ao| < r,
/3i(A),^2(A),... ,Pq(X) les q points distincts du cycle de longueur q dans
le disque pointe 0 < \z\ < e donne par le lemme 3, et posons /?o(A) — 0,
/3q+i($,... ,/32«-i(A) les autres points fixes de P% dans C. Ainsi on a
2q — 1
3=0
Soit
—-—h TTT (d'apres la preuve de la proposition 1) .
Rappelons que limA-jAo r ( ^ ) = A. D'autre part, on sait que le parametre

Ao = e2wip/q se trouve sur le bord de I'ensemble de Mandelbrot M; par suite
on peut choisir une suite (An)ngN* de nombres complexes avec An ^ M Vn 6

N* et linin-n-oo An = Ao. II en resulte que tout point fixe du polynome P%
est repulsif; autrement dit \(P\n)'(Pj{K))\ > 1 pour 0 < j < 2q - 1. Or la
fonction u> — i >• j ^ transforme {w e C/|w| > 1} sur {w €. C/Re a* < -|}; il
en resulte, d'apres le lemme 1, que
ReWP^/Ân))) < \ ,VneFf et V 0 < j < 2« - 1 ;
et on a
On en deduit que
2 < W ( T ( A B ) ) <
et, par passage a la limite quand n —> +oo, on obtient
Ceci acheve la preuve de la Proposition 2.
4. Application a l'etude d'explosion des cylindres
Pour des proprietes generales sur les phenomenes de bifurcation para-

bolique et d'explosion des cylindres de Fatou-Ecalle, on pourra consulter les
references suivantes [Br], [DH], [L], [SI] et [S2].
Rappels et Notations.
Considerons le polynome quadratique complexe P\o (z) — XQZ + z2 ou
Ao = e2nip/q. L'origine z = 0 est un point fixe parabolique et on a ex-
actement q axes d'attraction et q axes de repulsion dans un petit voisinage
de 0 ou Pqo et Pô9 sont bien definis. On peut trouver q domaines fij
(k e Z/qZ) contenant chacun un axe attractif et stable par Pqo et q do-
maines Ux0 k (k E Z/qZ) contenant chacun un axe repulsif et stable par
Pôq (le choix de fiô i restant bien sur arbitraire avec Qô suivant Cl^g ) .
L'espace quotient Cô (resp. Cô ) de fl^ (resp. £1^ ) sous l'action
de P%o (resp. P^ 7 ) est un cynndre topologique conformement isomorphe a
C* comme surface de Riemann et appele cylindre de Fatou-Ecalle.
262 Habib Jellouli
On notera St (resp. S7 ) la zone limitee par dClt (resp. d£l7 )

n0I
inclus et son image par PjJo (resp. P^) i incluse. Ces zones sont appelees
des regions fondamentales.
Considerons maintenant des perturbations de P\a dans la famille des
polynomes quadratiques P\{z) = Xz + z2 ou A = Aoe27™" assujetti a la
condition ajÔet | arga| < j .
Sous ces hypotheses et pour A assez voisin de Ao, on a un cycle de
longueur q pour PA note (/3i(A), /^(A),..., /39(A)) donne par le lemme 3 et
on peut trouver 2q regions fondamentales s£k et S7k (k E Z/qZ) dont le
bord de chacune est forme de deux courbes (une image de l'autre par P^)
joignant 0 et le point (ik{\) du cycle de longueur q.
Les cylindres quotients CXk = S£k/P% et C^k — S7k/P7~q sont encore
conformement isomorphes a C* comme surface de Riemann (en convenant
d'envoyer /3fe(A) en 0 et 0 en oo).
Partons d'un point de S~£k, en iterant P% un grand nombre de fois, on
arrive a S7k et par suite les cylindres C^k et C7k sont desormais identifies
par Faction de P%. Cette identification induit un isomorphisme G\ de C*
sur C* de la forme G\{z) = G\{1) • z. Par construction des cylindres, un
voisinage de /3fc(A) dans S7k est envoye par iterations de P% sur un voisinage
de /3fc(A) dans S^k (ce phenomene reste vrai pour A = Ao), ceci induit une
application holomorphe bijective F\ d'un voisinage de zero dans C* vers
un voisinage de zero dans C* prolongeable par continuity par -F\(0) = 0 et
verifiant F'x(0)=£0.
Dans ([DH],II) page 75, on a le developpement limite suivant :
= exp (^J) = exp
ou log designe la branche principale du logarithme, p(A) designe le multi-

plicateur du cycle (A(A), • . . , /9q(A)) de longueur q donne par le lemme 3 et
K est une constante (dependant de Ao).
Notre resultat dans cette section montre que la constante K s'exprime
en fonction du nombre complexe A = i(P% ,0) (indice holomorphe de P*
en zero).
Proposition 3. Pour A = Aoe2™a assez voisin de Ao, assujetti a la condi-
tion a ^ 0 et | arga| < j , on a le developpement suivant :
Indice holomorphe et rnultiplicateur 263
Preuve : D'apres ([DH], II) page 75, on a :
on peut ecrire :
log(p(A)) = log(p(X0)) H T T T ( A - Ao)

p(Xo)
p"(X0)p(X0)-(p'(X0)r 2 2
H (A — Ao) + o[(X — Ao) )
comme P(XQ) = 1, on aura :
log(p(A)) = p'(X0)(X - Ao) 11 + ° ,A — ( A - Ao) + o(X - Ao) I
d'ou:
. P(x0)(p(x0)y
log(p(A)) p'(Ao)(A-A o ) L1 2p'(A0) (A - Ao) + o(A - Ao)
p'(Ao)(A-Ao)
'(A))2 / +
i( P"(XO) \
22
2Aop'(A0) V ( ' ( A0)))) 2
V(p'(A / W
'
Tenant compte de la proposition 1, on a :
Ceci acheve la preuve de la proposition 3.

Remarque. : Le developpement limite donne par la proposition 3 nous
permet de retrouver l'inegalite Re(A) < 3~ deja demontree dans la propo-
sition 2, en effet : prenons A = Aoe27r8a avec a > 0 assez petit, on aura:
(2)
264 Habib Jellouli
d'autre part on a:
or l'origine z = 0 est un point fixe indifferent pour P\{z) = Xz + z2 car

|A| = 1, par suite le multiplicateur p(X) du cycle de longueur q donne par le
lemme 3 doit verifier |p(A)| > 1 (voir [Bl]).
Ainsi en utilisant (2) et (3) et par passage a la limite quand a tend vers
0 + , on obtient
d'ou
Remerciements :
Je tiens a remercier les Professeurs A. Douady, J.H. Hubbard, J. Milnor
et M. Shishikura pour leurs aides et encouragements.
Bibliographie
[Be] A.F. BEARDON, Iteration of rational maps, Springer-Verlag, New-
York, 1991.
[Bl] P. BLANCHARD, Complex analytic dynamics on the Riemann sphere,
Bull. Amer. Math. Soc. 11 (1984), 85-141.
[Br] B. BRANNER, Fatou coordinates and Ecalle cylinders, manuscrit 1993.
[DH] A. DOUADY et J.H. HUBBARD, Etudes dynamiques des polynomes
complexes I, II, Publications Mathematiques d'Orsay, 1984.
[L] P. LAVAURS, Systemes dynamiques holomorphes : explosion de points
periodiques paraboliques, these de doctorat de l'Universite de Paris-Sud,
Orsay, France, 1989.
[M] J. MILNOR, Dynamics in one complex variable, Introductory lectures,
Preprint of IMS, SUNY at Stony Brook, 1990/5.
[SI] M. SHISHIKURA, On the parabolic bifurcation of holomorphic maps,
Proc. of NogoyaConf., Sept. 1990.
[S2] —, The Hausdorff dimension of the boundary of the Mandelbrot set and
Julia sets, Annals of Math. 147(1998), 225-267.
Institut preparatoire aux etudes d'ingenieur de Nabeul, Departement de

Mathematiques et d'Informatique, Merazka, 8000 Nabeul, Tunisie.
An alternative proof of Mane's
theorem on non-expanding Julia sets
Mitsuhiro Shishikura and Tan Lei
Abstract
We give a proof of the following theorem of Mane: A forward invari-
ant compact set in the Julia set of a rational map is either expanding,
contains parabolic points or critical points, or intersects the ui-limit
set of a recurrent critical point. Moreover the boundary of a Siegel
disc is contained in the oj-limit set of some recurrent critical point. We
establish also a semi-local version of the result.
1 Statements
A classical theorem of Fatou says that for a given rational map / , if no
critical points are in its Julia set / ( / ) , then / is uniformly expanding on a
neighborhood of J(f). Such a map is called hyperbolic. A local version of
this says that if a; € J(f) is not in the closure of the critical orbits, there is
a neighborhood U of x such that the diameters of the components of f~nU
shrink to zero as n —> oo.
The next best category of rational maps are.those with no parabolic orbits
but with each critical point in the Julia set having a finite orbit. They are
called sub-hyperbolic maps. Such maps expand also uniformly on a neighbor-
hood of the Julia set, but with respect to a Riemannian metric with certain
singularities.
If we now allow parabolic orbits, but still require critical points in the
Julia set to have a finite orbit, we are into geometrically finite maps. These
maps have also some sort of expansion on the Julia set, but only on sector
neighborhoods of the parabolic points. See for example [TY].
A critical point in the Julia set having a finite orbit is a particular case of
being non-recurrent, in the following sense: Denote by w(c), the w-limit set of
c, to be {z € C | there exists nk ->• oo such that z = \imfnh(c)}. We say that
c is recurrent if c € OJ(C). Mane has a result that establishes expansion prop-
erties (or contraction properties of / " * ) , at points or on compact invariant
subsets of the Julia set away from the parabolic points and the us-limit sets of
recurrent critical points. This is an important result with many applications,
for example combined with conditions on the critical orbits (such as Collet-
Eckmann conditions), one may obtain metrical or geometrical information
about the Julia set of more general type of rational maps.
In this paper we present an alternative proof of Mane's result, along with
the first application by Mane, which is about the boundary of Siegel discs.
265
266 M. Shishikura and Tan Lei
We then apply it to a semi-local setting, following ideas of H. Kriete (see

[Kr]). There are many further applications in the literature, see for example
[Yin], in this volume, [CJY] and [PR].
Mane's result can be stated in the form of the following three theorems:
Denote by / a given rational map of degree at least two, by J(f) its Julia
set (which may or may not be C), and by N an integer depending only on /
(to be made precise in §2).
Theorem 1.1. (point version) If a point x G J(f) is not a parabolic periodic
point and is not contained in the u)-limit set of a recurrent critical point, then
for all e > 0 there exists a neighborhood U of x such that, for each n > 0 and
each connected component V' of f~n(U),
(a) the spherical diameter of V is < e and deg(/" : V —> U) < N.
(b) For all £j > 0 there exists n0 > 0 such that if n > no, the spherical
diameter ofV is <E\.
Theorem 1.2. (compact set version) Let A C «/(/) be a compact invariant
set (i.e /(A) C A) not containing critical points or parabolic periodic points.
If A is disjoint from the u -limit set of every recurrent critical point, then it is
expanding (see §5 for the precise definition). In particular a Cremer periodic
point must be contained in u(c) for some recurrent critical point c.
Theorem 1.3. (an application) Let F be either the orbit of a Cremer periodic
point, the boundary of a Siegel disc, or a connected component of the boundary
of a Herman ring. There exists a recurrent critical point c such that UJ(C) D F.
A recent application of these results by Yin ([Yi]) shows that if J(f)
contains no recurrent critical points, then it is shallow (a notion introduced
by C. McMullen), and therefore has Hausdorff dimension < 2 (a result first
obtained by Urbanski).
Section 2 presents a proof of Theorem 1.1 due to M. Shishikura. It is
edited by Tan Lei based on a hand-written manuscript of M. Shishikura, and
is presented in a form that is also valid for the case J(f) = C, and for the
generalization in Section 4. Appendix E is written by M. Shishikura. The
rest is written by Tan Lei. Section 3 contains proofs of Theorems 1.2 and 1.3
following Mane. Section 4 translates the above three theorems into a semi-
local version, and gives a brief description of an interesting application of
H. Kriete. The appendices contain some classical estimates on the Poincare
metric and alternative proofs of some of the lemmas.
Our proof of Theorem 1.1 is in spirit the same as Mane's original one, but
differs in presentation. It emphasizes on the use of Poincare metric, and gives
a direct argument rather than by contradiction. It also provides a bit of more
quantitative information.
An alternative proof of Mane's Theorem 267
The authors would like to thank C. McMullen and A. Manning for helpful
discussions and K. Astala for allowing us to include his elementary proof of
one of our lemmas.
2 Proof of Theorem 1.1

For z S C , denote by Dr(z) the open disc centred at z of radius r, and by D
the unit disc. We state first a general result:
Lemma 2.1. For any 0 < r < 1 and any positive integer p, there exists a
constant C(p, r) > 0 such that for any holomorphic proper map g : V —>• O of
degree < p, with V simply connected, each connected component of g~1(Dr(0))
has diameter < C(p,r) with respect to the Poincare metric on V. Moreover
, 0 C(p, r) = 0 .
Proof. (For a more elementary proof due to K. Astala, see Appendix D.) Let
A{ be the (finitely many) maximal concentric open annuli in D surrounding
DT(0) such that Ai does not contain critical values of g. For E C V a
component of ^ ( D ^ O ) ) and each i, there is A\ c V such that A\ surrounds
E and g : A\ —>• A{ is a covering. So
mod A\ = -—r——- mod Ai > - mod Ai and

deg(ffUj V
mod ( K £ ) > y mod A\ > - V mod At = - ° S ( 1 / r )
The rest is classical (cf. Appendix B). •

Next we define a universal constant.
Definition of No: There exist zx,--- , zNoî 6 D such that {§ < \z\ < 1} C
XT1 *
(XT
Let / be the given rational map of degree at least two.
Definition of O0 and J': Set ft0 = C if J(f) = C. Otherwise, by Sul-
livan's classification of the dynamics in the Fatou set we can find a non-
empty compact set L as a disjoint union of finitely many closed Jordan
domains and closed sub-annuli in the Herman rings and their preimages
such that f(L) c L, L contains all critical points in the Fatou set and
L n J(f) = {parabolic periodic points}. Set fi0 = C \ L. In both cases
we have f~l(Q0) C_Oo and £7o_n J(f) = J(f) \ {parabolic periodic points}.
In case_ J(f) = C set J' = C. In case J(f) ^ C, i.e. T^ ^_C, we may then
assume fijCC (so it is bounded). We set K' = {z, fn(z) € flQ for all n > 0}
and J' = dK'. Then f~l{J') C J' and the Hausdorff distance between J'
and <9/~"(f2o) tends to 0 as n ->• oo, with respect to the Euclidean metric
(in fact both K' and J' are totally invariant but we will not need these
properties). Notice that in this case J(f) C J', moreover L can be chosen so
that </(/) 7^ J' if and only if some Siegel disc or Herman ring of / contains
postcritical points.
In both cases f20 D J' contains no parabolic periodic points. The critical
points in X70 are all contained in J' and are of the following three types:
I. c is a recurrent critical point, i.e. c 6 to(c).
II. c is a non-recurrent critical point with c £ uj(d) for some recurrent critical
point d.
III. c is a non-recurrent critical point not contained in the w-limit set of any
recurrent critical point.
Definition of c i( N and C o : Denote by c\, • • • ,cv the critical points of type
III, deg(/, c) the local degree of / at c, and by N a positive integer greater
than or equal to Y[^=ideg(f,Ci) ( t m s makes sense even when there are no
critical points of type III) 1 . Set Co = N • No • C{N, 2/3), where C(-, •) is the
constant given by Lemma 2.1.
Denote by d(z, E) the Euclidean distance between a point z € C and a
closed subset E of C, and by di&m. w{W) the diameter of W with respect to
the Poincare metric of W (assuming that W is a hyperbolic Riemann surface
and W C W).
L e m m a 2.2. (1) Let x 6 0,0(1 J1. There exists fii with x G fix and satisfying:
{ fti C flo *s open and hyperbolic, not necessarily connected,

/"Hfii) C fix, dni(f(Ci)J(Cj)) > Co, if f{a) ? f(Cj)
rff!i(ci, / n (c*)) > Co, for i = 1, • • • , v and n > 1 ,
where d^1 is the Poincare metric on fii.

(2). Given J7X satisfying (*) above, for all C > 0 and ei > 0, there exists
n0 such that if n > n0 then for all W C W C /~"(Qi) such that W is
simply connected, W'flJ' ^ 0 and diamw(W) < C, the set W has spherical
diameter < £\.
Proof. (For a different proof relating directly the Poincare metric to the
spherical metric, see Appendix E.)
(1). For each i = 1, • • • ,v, there exists a repelling periodic point tt ar-
bitrarily close to Ci (and therefore with f(ti) close to /(c;)) such that the
orbit of ti does not contain x. We choose these points so that for Zo =
(Ji Orbit(ii), the set C \ Za is a hyperbolic surface, d-^ (f (ci)> f (cj)) > Co
l
Note that TV can be chosen depending only on the degree d of / , for example = d?d
if f(ci) ± f(cj) and dj, (a, fn{a)) > CQ for i = 1, • • • , v and n > 1. Now
set Qi = Oo \ ZQ C C \ Zo. It satisfies all the conditions.
(2) With the help of a Mobius transformation we may assume 0, oo ^ fij.
We will need the following inequality. L e t a e W e W c f t c C \ { 0 , o o }
with W a simply connected domain and diamvp(W') < C; we have
(**j diamgphericaiVK' < C • inf < d(a, dQ), T— >

I \a )
with C" a constant depending only on C. See Appendix A for a proof.
Case J' ^ C. We may then assume J' c 15^(0) for some R > 0. For W, W
as in (2), choose a eW C)J' we get from (**),
diamspherica,W < C • d(a,9/-"(fi 1 )) .
Since the Hausdorff distance (with respect to the Euclidean metric) between
J' and df~n(Qi) tends to 0 as n -> oo, there is no such that for n > no and
every a e J' we have d(a, 9/" n (fii)) < £i/C".
Case J' = J(/) = C. Let S = C'/ev For a pair (W, W) as in (2), if there
is a eW with |a| > 5, then by the inequality (**) the spherical diameter of
W is less than e\. It remains to consider a pair W',W with W C Ds{0)-
Choose Z C C \ f2i a periodic orbit (which is surely non-exceptional). Now
J(f) n Ds(0) is covered by finitely many discs Dr(dj) with r — e1/(2C") and
aj e J(/)n£>s(0). By properties of the Julia set, for each j , there is N(j) such
that fn{Dr{aj)) D Z for n > N(j). Hence D r (aj)n/- n (Z) ^ 0 forn > AT(j).
Now let n0 = maxj N(j). Fix n > n0. Then for any a 6 J(f) n £>s(0), there
is j such that a e Dr(a,j). So d(a, 9/"n(f2i)) < d(a,f-n{Z)) < 2r = e^/C.
Lemma 2.3. For any fix satisfying (*) above, let Uo C C be a round disc
such that £/„ C fii N Uc recurrent critical poinMc) and diamnif/0 < Co.
Then, for every n > 0,
deg(n). for every D$(z) C UQ with 0 < s < d(z,dUo)/2, and every connected
component V of f~n(Ds(z)), V' is simply connected and deg(/ n : V —>•
Dt{z)) < N;
diam(n). for every Dr(w) C Uo with 0 < r < d(w,dU0)/2, and every
connected component V of f~n(Dr(w)), diamo^ < Co.
Proof. We claim at first that f~n(U0) for any n > 0 contains no critical
points other than Cj, • • • , cv. First note that f(ui(c)) — CJ(C) for a recurrent
critical point c. So for Q2 = fij \ |JC recurrent critical point w(c)> w e h a v e
f~l(£l2) C fi2- In particular f~n(U0) C fi2 for all n > 0. Secondly recall that
the critical points in fli are in three different types, I, II and III, and those
of types I and II are contained in (J c recurrent critical point w ( c ) ( n o t i c e t h a t
for c a recurrent critical point, w(c) coincides with the closure of the orbit
{/ n (c), n > 0}). Therefore fi2, hence f~n(U0), contains only critical points
of type III, i.e. cx, • • • ,cv.
Now let us prove the assertion by induction on n. For n = 0, it is obvious.
Suppose it is true up to n — 1. We will prove deg(n) and diam(n) in two
distinct steps.
Step 1. deg(n) follows essentially from diam(O), • • •, diam(n-l).
Let Ds{z) be as in the Lemma. Let V be a component of f~n{Ds{z)).
j n
Then Vj = f (V) are components of f-( ~rt(Ds(z)) (j = 0,1, • • • , n ) . So, by
the hypothesis of induction, for j = 1, • • • , n, the set Vj is simply connected
and diamn^V}) < CQ.
Now V is also a component of /"^(Vi). Since diam ni (Vi) < Co, V\
contains at most one critical value in {/(CJ), i = 1, • • • , v\ (by Condition (*)),
and contains no critical values of other types of critical points. Therefore V
is simply connected and deg(/ : Vj —>• V^+i) = deg(/, q) if Q e V^- and = 1
otherwise.
Fix i € {1, • • • , v}. If Q 6 V} for some maximal j between 1 and n, then
fm(ci) £ Vj for m > 1, by the induction assumption diamn^V,-) < Co and
Condition (*). Hence Ci £ Vj for 0 < / < j . This means that Q appears at
most once in Vo = V, V\, • • •, Vn = Ds(z). Since no critical points other than
c\, • • • ,cv occur in these components, we have
n-l
deg(/ n : V -> D.(z)) = J ] deg(/ : Vj -»• Vj+1) < [ ] deg(/, *) < N .
j=0 i
Thus deg(n) is proved for any Ds(z) as in the Lemma.

Step 2. Given Dr(w) as in the Lemma, diam(n) for Dr(w) follows from
deg(n) applied to several Ds(z)'s.
• Let V be a component of f~n(Dr(w)). Apply deg(n) to Dr(w) we know
that V is simply connected and deg(/" : V —> DT(w)) < N.
Now choose z\, • • • , Zf\ro-i € Dr(w) such that
ATo-l
Dr{w) \ D$r(w) C Q
We have, for i = 1, • • • , No — 1,
d ( z i , 9C/ 0 ) > d ( w , dU0) - d { w , z{) > d ( w , dU0) -r>2r-r=r.

So we can apply deg(n) to Drfa) and V a component of f~n(Dz(zi)) to
conclude that V is simply connected and deg(/ n : V -> Dz.(zi)) < N.
Since ^r = \-\r, by Lemma 2.1 each connected component of f~n{Dir{zi))

in V has diameter < C(N, f) with respect to dv>, and hence with respect to
(in, since V C f~n{ili) C fii.
Similarly each connected component of /~"(£)2r(u;)) has diameter at most
C(N, §) with respect to dni.
Now Dr(w) is the union of the following NQ open connected sets: DiT{w)
and Dir(zi) C\DT(w) for i = 1, • • • , JV0 — 1. Since / " : V -> Dr(w) is a proper
map of degree < iV, the preimage by / " " of each of the above No sets has at
most N connected components in V, and V is covered by these components.
Therefore V is covered by at most TV • No sets of diameter < C(N, | ) . Hence
diam ni V < N • No • C(N, §) = Co since V is connected.
This completes the induction. •
Proof of Theorem 1.1. Let x £ HQ d J' not contained in the w-limit set of a
recurrent critical point. By Lemma 2.2(1), there is ilx satisfying (*) and
x 6 {Hi n J') \ (J w(c) .

c recurrent critical point
We may assume 0,oo £ il\. Fix e > 0. We will prove Theorem 1.1 for all
x1 G (fi, n J') \ |JC recurrent critical point w ( c ) w i t h bounds depending only
on Hi.
First note that there exists 5 > 0, if V C V C Hi with V simply connected
and diamv(^') < 6 then the spherical diameter of V is less than e (see
Appendix A). Take 0 < p < 1 such that C{N, p) < 5 (where C{N, p) is given
by Lemma 2.1).
Let x' eilinf\ Uc recurrent critical point w ( c )' L e t ô be a round disc
such that x' eUQC fy s Q c recurrent critical point w ( c ) a n d diamniô < Co
(as in Lemma 2.3).
Let r > 0 be such that r < |d(x', 9C/0). Define [/ = D^x') C Dr(x').
Fix n > 0. Let V^ be a connected component of f~n{DT{x')) and V be
a connected component of f~n{U) in V. Then by Lemma 2.3 both V and
V are simply connected and deg(/ n : V ->• Dr{x')) < N. So diamyV"' <
C{N,p) < S (Lemma 2.1). Hence V has spherical diameter < e. This
proves (a). For (b), let ei > 0 and C = C{N,p). Let n0 be the integer
given in Lemma 2.2(2) (which depends only on Hi) and assume n > n0.
For V and I" as above (relative to f~n) we have V C V C /~"(fii), V is
simply connected, V n J' / 0 (since x' 6 J' n [/ and Z" 1 ^') C J') and
diamv(V') < C{N,p) = C. So we can apply Lemma 2.2(2) to W = V and
W' — V to conclude that the spherical diameter of V' is less than si.
This ends the proof of Theorem 1.1. •
3 Proofs of Theorems 1.2 and 1.3

Proof of Theorem 1.2. We say that a compact set A is expanding if there is N
such that for any n > N, minÂ ||(/ n )'(2)|| > 1, where || • || is with respect
to the spherical metric.
Assume that A is a forward invariant compact set not containing critical
points and parabolic points. We may assume 0, oo ^ A and thus work with
Euclidean metric rather than spherical metric.
Assume by contradiction that A is not expanding. In other words, there
are nk —>• oo, zk G A, such that \{fni°)'{zk)\ < 1. We will show that any
accumulation point of {fnk {zkfykeN ^s *n ^ n e w-limit set of some recurrent
critical point.
Assume not. We may assume the entire sequence fnk(zk) converges to a
point x £ A. Then x satisfies the conditions of Theorem 1.1. Take e > 0
such that d(A, c) > 2e for every critical point c. Let U be a round disc
centred at x associated to e given by Theorem 1.1. For k large, fnk(zk) is
in U. Let Vk be the component of f~nk(U) containing zk. By Theorem 1.1,
diam(/J(Vfc)) < e for 0 < j < nk. Since p(zk) £ A for all j , we have p(Vk)
does not contain critical points of / . Therefore fnk : Vk —v U is a bijection.
Let ipk : U —t Vk be the inverse. Then the family {(pk, k € N} is normal
and any limit function tp must be constant, since diamVfc -> 0 as k —> oo
(Theorem 1.1). This contradicts the fact that
\<p'(x)\ = lim Wk{fnk{zk))\ = lim } > 1.

k-+oo kôo |(/"<=)'{z k )\
m
Proof of Theorem 1.3. The orbit of a Cremer periodic point is in J', invariant
and non-expanding. So by Theorem 1.2 it intersects ui(c) for some recurrent
critical point c. It is therefore contained in co(c).
Now let F denotes either the boundary of a Siegel disc or a component of
the boundary of a Herman ring. We want to show that there is a recurrent
critical point c such that F C OJ{C). The proof consists of two steps.
c
*• U c recurrent critical point w ( c )-
Since there are only finitely many (recurrent) critical points, the right
hand set is closed. Note also that there are only finitely many critical points
and parabolic periodic points.
Assume the assertion is not true. Then there is x € F satisfying the
conditions of Theorem 1.1. There is therefore a connected open neighborhood
V of x such that components of f~nq(V) intersecting F have diameter tending
to 0. On the other hand, consider the "harmonic measure" \x of F, namely that
induced by the boundary map of the conformal linearization map tp : A —> B,
with B the Siegel disc with boundary F or respectively the Herman ring with
one boundary component F, and with A = O, or respectively an annulus.

It is known that /z is non-atomic with support the whole F, fq preserves /z
and is ergodic with respect to it. From the properties of V and f~nq{V)
we know that V n F has positive harmonic measure and f~nq(V) D F has
harmonic measure tending to 0 as n tends to oo. This is a contradiction since
/ ' preserves the harmonic measure.
2. There is one recurrent critical point c such that FCw(c).
For any recurrent critical point c, we have /(F n w(c)) C (F n w(c)). As
f : F -> F is ergodic with respect to /i, the set F n w(c) has either 0 or full
harmonic measure. But there are only finitely many such critical points. By
1 there is c recurrent such that F Pi w(c) has full harmonic measure. Now
since Fnw(c) is closed and the support of the harmonic measure is the whole
set F, we must have F flw(c) = F. •
4 A semi-local version of the above results

Although we stated and proved Theorems 1.1, 1.2 and 1.3 for a rational
map as a global dynamical systems of C, our proof actually works, with
only minor modification, for restrictions of a rational map in a sub-domain
of C, considered as a semi-local dynamical system, together with semi-local
recurrent critical points and semi-local invariant compact sets. We may then
conclude that only semi-local recurrent critical points are relevant to semi-
local invariant subsets.
More precisely, let F : C —> C be a rational map. Let Q' c Q be two
open subsets of C with Q' ^ Q and Q # C, such that_F| Q ; : Q' ->• Q
is a holomorphic proper map. Consider / — F\QT : Q' —> Q as a dynamical
system that is only defined on Q'. (This is a bit like polynomial-like mappings,
although it is not required that Q' is relatively compact in Q, and no quasi-
conformal technique is needed).
Now critical points of F in Q' are also critical points of / . We say that
a critical point c escapes if there is k > 0 such that c, /(c), • • • , fk(c) 6 Q
but fk(c) iQ'. We call {c, /(c), • • • , fk(c)} the /-orbit of c. Although such
critical points may very well be recurrent under F, they will not play relevant
roles to the /-invariant sets in Q'.
We modify the definition of fi0 as follows: Define L for the rational map F
as before. Let V be the union of L and the /-orbits of escaping critical points.
Set fi0 = Q \ L'. Since f(L D Q') C F(L) C L, we have f(U D Q') C L'.
Therefore/-^fto) CSV _
Now we can define K' = {z £ Q'Jn(z) 6 f-l{n0) for all n > 0} and
J' = dK'. Then again
(A) / - 1 (J') C J' and
(B) the Hausdorff distance between «/' and <9/~"(fio) tends to 0, as n -> oo,
with respect to the Euclidean metric.
We can now restate Theorems 1.1, 1.2 and 1.3 for / provided we make
the following changes: In the statements of the theorems replace "x € J{fT
by "x € Qo H J'" > "recurrent critical point" by "/-recurrent critical point",
"A C J(fT by "A C nonJ'", and finally T " by T C fionJ' with / n ( r ) C fi0
for all n".
For the proof of these theorems, we make the following changes: In the
proof of Lemma 2.2(1) choose the points ti whose /-orbits escape, and set
ZQ to be the union of these /-orbits. Since there are such kind of points
arbitrarily close to each Cj, Condition (*) is easily achieved. In the proof of
Lemma 2.2(2), the study of the case J' = C can be deleted. Notice that the
only properties of J' needed is (B) (in the proof of Lemma 2.2(2)) and (A)
(at the very end of §2). This completes the proof of Theorems 1.1, 1.2 and
1.3 in this generalized setting.
We now describe briefly an application by H. Kriete (see [Kr]),-of this
generalized version of Mane's result: Let F be a polynomial. Then Goldberg-
Milnor's Fixed Point Portrait (cf. [GM]) provides regions in C bounded by
periodic external rays that separate Cremer points and Siegel discs. Apply
Theorem 1.3 to F restricted to each of these regions we obtain a critical
point which is recurrent within the region and whose orbit accumulates to
the Cremer point or Siegel boundary of the region.
A The inequality
Recall that d(a, E) denotes the Euclidean distance between a point a e C
and a closed subset E of C
Lemma A . I . Let W 7^ C be an open and simply connected proper subset of

C. Leta,be W. Then
\b-a\< (c 2 ^ 8 -*) - 1) • d{a, dW) .
Proof. Let F : D -4 W be a conformal map with F(0) = a. Set r = \F~l(b)\.

Then
J. — 7*
Denote by h(z) — (F(z) — a)/F'(0) the normalized univalent map. By Koebe

1/4 and distortion theorems we have
± and
Combining these we get
I6 l*
The right hand side is equal to ( e 2 ^ " ' 6 ' - 1) • d(a, dW). •
Lemma A.2. Let a e W C W C fi C C \ {0, oo} with W open and simply
connected. Assume di&mwiW) < C. Then
diam spherica ]W < C inf < d(a, dQ), —-

I a
where C = 2C"(e 2C — 1) with C" a universal constant.
Proof. Let C" be the Lipschitz constant between the spherical metric and the
Euclidean metric. We have
erica iW' < C"diamEuC|ideanW < 2C" sup6eVi/, |6 - a\
< 2C"snpbeW,{e2dâ'» - l)d{a,dW)
2C
< 2C"(e - l)d{a,dW) = C'd{a,dQ.) .
For a suitable choice of the spherical metric, the map H(z) = 1/z is an
isometry. Apply the above inequality to H(a) C H(W) C H{W) C
we get
diam s p h e r i c a l ^' = diam.phericaiif (VK') < C'd(H(a),dH(n)) < £- ,
where the last inequality is due to the facts that H(a) — -a and 0 £ H(fl).B
From this one deduces easily that for any e > 0, there is S such that if
C < 5 then diamsphericaiW < £• This fact was needed at the end of §2 for
B Control between modulus and diameter

Lemma B . I . For 6 6 (0,oo), there are two strictly decreasing continuous
functions u(6) and v(5) with
lim u(5) = lim v(5) — +oo; lim u(6) = lim v(5) = 0 ,
(5-»0+ (5->0+ i5-»+oo <5-++oo
satisfying the following property: Given any pair (V, E) with 7 c C a hyper-
bolic open simply connected subset (i.e. # ( C \ V) > 2) and E C V compact
so that V \ E is an open annulus, then for m the modulus of V N E and 6
the diameter of E with respect to the Poincare metric on V, we have
u(6) <m< v(S) .
Proof. We may assume V = D, and E contains 0 and s with s € (0,1) and

rfD(0, s) = S. We have
log- = 5 therefore s = s(d~) = —&—- .
Note that E C £>s(0) so A D D \ A,(0). As a consequence
m > mod (D s D s (0)) = i - log - = -!- log ^ j =: «(<5) .
To get v(6), we use the solution of Grotzsch extremal problem (cf. [All],
Chapter III) that m < mod (D \ [0, s(S)}) =: u((5). •
In the case of Lemma 2.1, we have v(S) > m > (log(l/r))/(27rp) so
The function C(p,r) as a function of r is strictly increasing and continuous,

with limr_>o+ C(p,r) = 0 and limr_>i- C(p,r) — +oc.
C Bounded distortion
We use Lemma 2.1 to recover a distortion estimate in [CJY] (although we
don't need it in our paper). In the same setting as Lemma 2.1, denote by
Bv(z,s) the Poincare disc in V centred at z with Poincare radius s. Define
By$(w,t) similarly. Then there is a constant C\ depending only on ,s and p
such that, for any z £ V,
BB(g(z),Cl)Cg(Bv(z,s)) c BB(g(z),s) .
Proof. The right hand inclusion is due to Schwarz Lemma. As for the left
hand side, we may assume g(z) = 0. Let £?p(0, t) be the largest disc contained
in g(Bv(z, s)). Let E be the connected component of g~l(Bv(0, t)) containing
z. Obviously E n dBy{z,s) ^ 0. Therefore the Poincare diameter 5 of E
satisfies S > s. By Lemma 2.1 and Lemma B.I, s < 6 < C(p,r(t)) =: Cp(t),
where
Clearly Cp(t) is a continuous strictly increasing function of t. As a conse-

quence t > {Cp)"l(s) =: Ci(p,s).
D An elementary proof of Lemma 2.1

This proof is due to K. Astala, who kindly allowed us to include it here.
In the setting of Lemma 2.1 we may assume V — D. Therefore g : D -> D
is a Blaschke product of degree at most p, more precisely
— diZ
with fliSD and £ < p.

Let z' € D such that |ff(z')l ^ r - Then there is at least one i such that
z— < r1/p .
1 -
As Pi(z) := i_^*z is an isometry for the Poincare metric, we have c?p(z', a;) —
=:
dj)(Pi(z'), 0) < log }tfi/p ^ - As a consequence z' £ |J- = 1 S p ( a i : M).
Therefore g~ {Dr) C ULi-^D( a »'-^)- ^ o w a n ^ c o n n e c t e d component of
l
x
g~ (Dr) would have diameter with respect to dp at most p x 2M —: C'(p, r).
From the definition of M one can see clearly that C'(p, r) —>• 0 and r —> 0.
E Alternative proof of Lemma 2.2

The Poincare metric A(z)|dz| of C \ {0,1} has an estimate
logA(z) = log , / l + 0(1) as z -> 0 .
See [A12, p.18, (1-24)] (or also [McM, pl3, Theorem 2.3]). An easy calculation
shows that for any C > 0 and e > 0, there exists 8 > 0 such that if y 6 C
and 0 < \y\ < 6 then
Using affme transformations sending 0, 1, oo to 1, w, oo, one can show that

for any 0 < r 0 and e > 0, there exists 5 > 0 such that if
r < \w\ < R, dsPhericai(iu,2/) < <5 and i / 6 C \ { 0 , l,w}, then
diamspherical{z| rfCv{0,w}(2:'^) ^ C} < e-

The same argument can be used with {0, oo} replaced by {0,1} or {l,oo}.
Therefore we conclude that for any C > 0 and e > 0, there exists 5 > 0 such
that ifrfsphericai(w,y) < 6 and y € C \ {0,1, w} then
dCMOlw](z,y) < C} < e .
278 M. Shishikuia and Tan Lei
Now let us define fii. Let t be a repelling periodic point which is not in
the orbits of x and q's. Let Zo = U%Lof~n(t), where M is to be determined
later and Qi = f20 — Zo- By the choice of t, we know that x and the orbits of
Cj's belong to fii. We may suppose that 0,1, oo G Zo.
To prove (1), set C = Co and
e = m i n K p h e r i c a i l f t e ) , / ^ ) ) ! / ^ ) ^ /( C i )}U
Kphericalfci, /"(Cj))|i =l,...,V, n > 1} ,

and we obtain 6 satisfying the above. Since the inverse orbit of t is dense in
the Julia set, we can choose large M so that for any z 6 J(f) there exists
an element w G Zo = l)%Lof~n(t) with daphericetx(z,w) < 6. In particular, if
y = f(ci), we Zo and dsphericai(w,2/) < 6 then
diamsphericai{z| dni(z,y) < Co} < diamsphericai{z| dCvRi,™}^'^) < Co} < £ •
Therefore rfsphericai(/(ci),/(cj)) > Co if /(CJ) ^ /(CJ). The same argument

applies tO d S pherical(ci,/ n (Cj)).
To prove (2), set C = C and e = t\ and we obtain 6 satisfying the above
property. Choose n 0 so that every point of the Julia set is within spherical
distance 5 from f~n°(Z0). Then for any y G f n fi0 C J ( / ) , there exists
w e Zo such that £4phericai(iu,2/) < <5 and hence
diamsphericai{,z| df-nini)(z,y) < C}
< diamsphericai{^| dC^{o,i,w}(z>y) <C} <ei .

It follows that if n > n0, W C /""(î) and dia,mf-n{ni)(W) < C then W
has a spherical diameter less tahn ei. •
Remark. [McM, p.39, Theorem 3.6] uses a similar argument.
References
[All] L.V. Ahlfors, Lectures on Quasiconformal Mappings, Wadsworth,
1987.
[A12] L.V. Ahlfors, Conformal Invariants, McGraw-Hill, 1973.

[CJY] L. C. Carleson, P. W. Jones and J.-C. Yoccoz, Julia and John, Bol.
Soc. Bras. Mat. vol. 25, p. 1-30 (1994).
[GM] L Goldberg and J. Milnor, Fixed points of polynomial maps, part II,
Ann. Sci. Ec. Norm. Sup. vol. 26 (1993), p. 51-98.
[Ma] R. Mane, On a theorem of Fatou, Bol. Soc. Bras. Mat. vol. 24, p.
1-11 (1993).
[Kr] H. Kriete, Recurrence and periodic points, manuscript, February

1998.
[McM] C. T. McMullen, Complex Dynamics and Renormalization, Princeton

University Press, 1994.
[PR] F. Przytycki and S. Rohde, Porosity of Collet-Eckmann Julia sets,

Fund Math. 155 (1998), pp. 189-199.
[TY] Tan Lei and Yin Yongcheng, Local connectivity of the Julia set for
geometrically finite rational maps, Science in China (Series A), vol.
39, number 1 (1996), pp. 39-47.
[Yi] Yin Yongcheng, Geometry and dimension of Julia sets, in this vol-
ume.
Mitsuhiro Shishikura, Hiroshima University, Department of Mathematics,

Kagamiyama, Higashi-Hiroshima 739-8526, Japan,
email : mitsu@math.sci.hiroshima-u.ac.jp
Tan Lei, Department of Mathematics, University of Warwick, Coventry
CV4 7AL, United Kingdom.
e-mail: tanlei@maths.warwick.ac.uk .
GEOMETRY AND DIMENSION OF JULIA SETS
Y I N YONGCHENG
ABSTRACT. We show that the Julia set of a critically non-recurrent rational

map is shallow, in the sense of McMullen (except the case that the Julia set
is the whole sphere). This recovers, as a corollary, a theorem of M. Urbanski
with a different approach.
1991 Mathematical Subject Classification: 30D05, 58F08
1. INTRODUCTION
Let / : C -> C be a rational map with degree deg / > 2. A point z is
said to be a periodic point if fk (z) = z for some k > 1. The minimal of such
k is called the period. For a periodic point ZQ, denote the multiplier of ZQ
by A = (fk)'(z0). The periodic point z0 is either attracting, indifferent or
repelling according to | A | < 1, |A| = l o r | A | > l . In the indifferent case,
we say zo is parabolic if A is the root of unity.
The Julia set, denote by J(f), is the closure of all repelling periodic
points. Its complement is called Fatou set, denote by F(f). A connected
component of F(f) is called a Fatou component. D. Sullivan proved that
each Fatou component U is preperiodic, this means there exist integers
m > 1, n > 0 so that
and every periodic Fatou component is either an attracting basin, a par-

abolic basin, a Siegel disk or a Herman ring. For the classical results of
complex dynamics, see [Be], [Mi] and [CG].
Definition 1. A rational map /:C—»C is called critically non-recurrent if
each critical point c £ •/(/) is non-recurrent, i.e. c ^ w(c).
Denote B{x, e) a ball of venter x and radius e in the spherical metric on
C, D{x, e) a disk of venter x and radius e in the Euclidean metric in C.
The following definition is given by CMcMullen, [Me].
The research is partially supported by NSF of China (No.19531060) and Morningside
Centre of Mathematics, Chinese Academy of Sciences.
281
282 Yin Yongcheng
Definition 2. A compact subset X c C i s shallow if there exists 0 < k < 1

such that any ball B(z, r), z 6 X, r > 0, contains a smaller ball B(y, kr) C
C\X.
It's easy to prove that if X is shallow, then its box dimension BD{X) < 2
and its Hausdorff dimension HD{X) < BD{X) < 2.
The main result of this note is the following statement.
Main Theorem. For a critically non-recurrent rational map f : C —> C,
the Julia set J(f) is shallow or </(/) = C.
A directly consequence is
Corollary. (Urbanski, [Ur]j For a critically non-recurrent rational map
f :C-*C,HD(J(f))<2 orJ(f) = C.
Remark. C. McMullen, F. Przytycki and S. Rohde proved the same the-
orems for quadratic polynomials with Siegel disk of bounded type rotation
numbers and Collet-Eckmann rational maps without parabolic points, [Me],
[PR],
Notations. 1. d is the spherical metric on C.
2. The distortion Dist(X, z) of a compact set X about z ^ X is
max d(y, z)l min d(y, z).
y€X yex
3. Two positive numbers A and B are K-commensurable or simply
commensurable if K~l < A/B < K for some K > 1 independent of A and
B, Ax B means A and B are commensurable.
Acknowledgement. The work was done during the program on complex dy-
namics hold at Morningside Centre of Mathematics in 1998. Thanks to Prof.
Yang Lo and Prof. Li Zhong organised this program and to all members in
the seminar on complex dynamics at Beijing.
2. BACKGROUND MATERIALS
2.1 The distortion theorem. The distortion theorem of univalant maps is

a powerful tool in the study of complex dynamics. The following distortion
theorem of the version for p-valent maps is well-known (see [CJY] or [ST]
appendix C).
Lemma 2.1. Let Ui and XJ<i be simply connected domains in C and g :
Ui —^ Ui be a proper holomorphic map of deg g <p. then
{w\pU2(w,g(z0)) < C""1} C gdzlpu^zô) < r}) C {w\pU2{w,g(z0)) < r}
for any ZQ € U\, where pux and pu2 are hyperbolic metrics ofU\ and U2, C
is a constant depending only on p and r.
Denote p the hyperbolic metric in D(0,1) and Bp(0,r) the hyperbolic
ball of centre at 0 and radius r.
Geometry and dimension of Julia, sets 283
Lemma 2.2. Let h : D(0,1) —> D(Q, 1) be o proper holomorphic map of

degree at most p and h(0) = 0. Then there exists a constant R* depending
only on p and R such that D(0,R*) C h(Bp(0,R)) and the component of
h~1(D(0,R*)) containing 0 is a subset of Bp(0,R).
Proof. Assume deg h = p' <p. Let 0 = z0, z\, • • • , zpi-\ be the preimages
of h(z0) = 0 and hj : D(0,1) -> D(0,1) be the univalent map so that
hj(Zj) = 0 , 0 < j < p' - 1. Then
P'-I
h{z) = e» ft hi(z)
3=0
for some 6 e K.
There are at most p' many points of {ZJ \ 0 < j < p' — 1} in the closed
hyperbolic ball Bp(0,R). We can find a hyperbolic circle 7 of centre at 0 in
Bp(0,R) such that
for z £ 7 and 0 < j < p' - 1. Then
p(hj(z),Q) =
Change it to the Euclidean metric,

g2(p'-l) _ 1
d(M*),0)> — = Cv
e 2(p'-l) -|- I
Then
P'-I
d(fc(z),0) = Y[ d(fc,-(r),O) > C / = R*(p',R).
j=o
and i?* = min{i?*(p', i?) : 1 < p' < p} is the constant which we need.
QED
Prom lemma 2.2, we have a useful corollary in spherical metric.
Corollary 2.3. Let U 9 ZQ be a simply connected domain and g : U —>
B(WQ, 26) be a proper holomorphic map of degg < p, WQ = g(zo). If g maps
B(zo,r) c U into B(WQ,6), then there exists a constant K depending only
on p so that B(wo,r*) c g{B(zo,r)) and the component of g~l(B(w0,r*))
containing z0 is a subset of B(zo,r), where r* > diamg(B(zo,r))/K.
Proof. Assume (dg(B(zo,r))) D {dB(wo,6)) ^ 0 at first. Let U' be the
component of g~1(B(u>o,6)) containing ZQ. Then mod (U \ U') > ^ ^ .
284 Yin Yongcheng
For any two points x,y E dB(zo,r), pu(x,z0) and pu(y,zo) are C(p)-
commensurable with 1 for a constant C(p) depending only on p. Take
R= min pu(x,z0). By lemma 2.2, there exists r* = 26R* such that
x£dB(zo,r)
B(wo,r*) c g(B(zo,r)) and the component of g~1 (B(WQ, r*)) containing ZQ
is a subset of B(zo,r). Denote K = 1/R*. Then r* > dia,mg{B(zo,r))/K.
In the case (dg(B(z0, r)))n(dB(w0,6)) = 0, take 8' = d(w0, dg(B{z0, r)))
< 5. We replace U by the component of g~1(B(w0,26')) containing z0. Use
the same method as above, we get the constant K.
QED
2.2 Mane's theorem In [Ma](see also [ST]), R. Mane proved a beautiful
theorem. It is a key lemma in the proof of our main result.
Lemma 2.4. (Mane's theorem) Let f : C -> C be a rational map. If a
point x 6 J(f) is not a parabolic periodic point and is not contained in the
at -limit set of a recurrent critical point, then for all e > 0 there exists a
neighbourhood U of x such that
(a) For all n > 0, every component of f~n(U) has diameter < e;
(b) There exists p > 0 such that for all n > 0 and every component V
ofrn(U),
degree(fn : V -)• U) < p.
2.3 Dynamics near the parabolic point. Let / 0 : C -»• C be a ratio-

nal map with a non-degenerate parabolic fixed point 0, i.e. /o(0) = 1 and
/o(0) ^ 0. Choose a neighbourhood Uo 3 0 so that / 0 maps Uo homeomor-
phically onto .a neighbourhood UQ3 0.
Lemma 2.5. (Fatou coordinates) Take/o, Uo andU^ as above. Then there
exist simply connected domains D± compactly contained in Uo n U'Q, whose
union forms a punctured neighbourhood ofO, satisfying
f±(D±) CD±U {0} and Dn>0 ft(D±) = {0}.
Moreover, there exist univalent maps $± : D± -> C such that

(1) *±(/o(*))
3 3
(2) Range ($+) D {C | - - T < arg(C - Ro) < -n] and Range (<2>_) D
{C I -7T < arg(C + Ro) < -TT} for some Ro > 0.
The domains D+ and D- in lemma 2.5 are called attracting petal and
repelling petal respectively.
Set Vo* = {C|ReC < -Ro}, Vo = {ClReC < -Ro, |Imfl >
{C | - Ro - 1 < ReC < -flo, |ImC| < i?o}, % = O0 - 3, j > 1 and
Vo* = i l ^ V y ) , y 0 = S I 1 fa), fij = $I^(i2j), j > 0. Prom lemma 2.5,
Vo is contained in the Fatou set F(f0) = C \ J(/o) and f0 : fij+i -» Clj is
homeomorphic for j > 0.
3. PROOF OF THE MAIN THEOREM
Let / : C —> C be a rational map. In this section, we always assume

J(f) ± C.
For x e J(f), r > 0, define
h{x, r) = sup{£ | there is B(y, £) C F(f) D B(x, r)}.
Then /i : J(f) x R+ -> R+ is continuous.

Since int(J(/)) = <f>, hence
for any r > 0.

Proof of the main theorem. First of all, we suppose / is a critically non-
recurrent rational map having no parabolic periodic points. From lemma
2.4, there are So > 0 and p < oo so that for any x 6 J(f), n > 0 and any
component V of f~n(B(x, 2S0)), V is simply connected and
deg(/ n : V ->• B(x, 25O)) < p.
For any z0 € </(/) and 0 < r < So, look at the forward images Bm =
fm(B(z0,r)), zm=fm{zo), m>0.
Let no be the smallest integer so that diam5 n o + i > (V Then Zoô <
dianxBno < 80 for some 0 < l0 < 1. By corollary 2.3, there are a constant
K depending only on p and a ball B(zno,ro) C Bno with ro > diam5 n o /i^
such that the component of f~n°(B(zno,ro)) containing z0 is a subset of
B(zo,r). There exists B{yo,-h(rQ)) c B(zno,r0) C\ F(f). Let £»0 be a
component of f~n°{B(yo,-h(ro))) contained in B(zQ,r), y 6 Dor\f~n°{yo).

Then Dist(9£>o, y) < M for some M depending only on p and dianxDo ^ r.
Therefore, there exists 0 < k < 1 which does not depend on zo and r so
that ^(y, kr) C Do C B(z0, r) D F ( / ) . The Julia set J ( / ) is shallow.
Now let / be a non-recurrent rational map having parabolic periodic
points. For simplicity, we suppose / has only one non-degenerate parabolic
fixed point 0.
286 Yin Yongcheng
Take Qj(j > 0) as in section 2. The set Xo = J(f) \ (U,->ifij U {0}) is

a compact subset of J(f). By lemma 2.4, there exist Si > 0 and p! < oo
so that for all x 6 Xo, n > 0, 0 < (5 < #i and any component V of
f~n(B(x, 26)), V is simply connected and
deg(/ n : V-+B(x,26))<Pl.
Let _K"i be the constant depending only on pi in corollary 2.3. Suppose D

contains a ball B(z,e) with e > diamD/Ki. Then f{D) contains a ball
B(f(z),e*) such that the component of f~1(B(f(z),e*)) containing z is a
subset of B(z,e), where e* > diaxaf {D) / KQ for a constant K$ depending
only on K\.
Choose N large enough so that N > 128K0RQ and d(Q6nn) > 4K0dia.mD,n
for n > TV. Take 52 = ridia.mQN for a suitable 77 so that
B(x, 2S2) C Vo* = V0U (UjyoQj)
for any x £ UjL^j.

Now we prove that the Julia set </(/) is shallow.
For any z0 6 «/(/) and 0 < r < 60 = min(£i,<S2), denote Bm =
fm(B(z0,r)), zm = fm(z0), m > 0. Let n 0 be the smallest integer such
that diam.Bno+i > #o- Then diam5 m < Jo for m < no and diamB no x 8Q.
If Zno G Xo, then the same argument as in no parabolic periodic points
case shows there exists B(y, kir) C B(zo,r) D F(f) for some 0 <fci< 1.
Suppose zno 6 (Lij>iflj) U {0} and mo > 0 is the smallest integer such
that Zj 6 (l>j>itij) U {0} for all ra0 < j < n 0 . Then J5mo contains a
ball W = B(zTOo,e*) with e* > diamBmo/ATo such that the component of
f~m°(W) containing ZQ is a subset of B(ZQ,T).
We claim that there exists a ball B(yQ,rQ) c 5 m o fl F ( / ) such that
f~m°(B(yo,rQ)) has a component contained in B(zo,r) for some ro x
diam5 m o .
(1) Ifz m o = 0, take r 0 = - e * .
(2) If zno e njo for 1 < jo < N, then Bno c 5(zno,<J0) andB(z no ,26 0 ) C
y0*. Let B' be a component of f-^n°-m^(B(zno,2S0)) containing Bmo.
Then / n »- m o : B' -> 5(z n o ,2J 0 ) is univalent and mod (J5' \ Bmo) >
— log 2. By distortion theorem of univalent map, Dist(9Wri, zn) x 1 and
Z7T
there is 5(y3, rg) C Wi n F ( / ) for some rj x 50, where Wx = fn°-m°(W).
Hence there is B(yo, r0) C W D -F(/) for some r 0 x diamB mo .
(3) If zno E dj0 for jo > N, zmo e Qjo + (n 0 - m 0 ) = Clio.
(3a) When diam5 m o > 2Jftr0diamfij0, d(zmo,dW) > 2diamfiio, take
(3b) When diam5 mo < 2Kodi&m£lio, D = B(zmo,2dmmBmo) C V*.

Let D — $_(£>), W — $-(W), Wi = $_(Wi) and zmo = $_(2:mo). Then
Jmo) < 16Dist(9W,zmo) < 16iô-
If d(5mo, dW) > 4i?0, then there is a ball 5 = B(y0, \d{zmo, dW)) C Vb
such that $I X (5) = S contains a ball B(yo,ro) C W n F o C f f n f (/) for
r 0 xdiamB m o .
If d ( i m o , 9 # ) < 4i?0, then W C 5(zmo,64A:oi?o) and.Wi = W +
(n0 - m0) C B(zno,6AK0Ro), zno = zmo + (n0 - m0) = $_(z no ). Since
^n0 e fij-o, Jo > N > 128KORO, we have D' = B(zno,128KoRo) C V*,
D' = QZ^D') C V , B no C £>' and mod (£>''\ 5 n o ) > ^-Iog2. Hence
2TT
Dist (9Wi, zno) < 16Dist (dWi, zno) = 16Dist (dW , zmo) < 162KO.
Let D" be the component of /-( n o-«o)( J p') containing Bmo. Then
jno-mo . £)/' _). £>' j s univalent. By distortion theorem, there is a ball
#(yo.ro) C Ty n F(/) for some r 0 x diamB mo .
The proof of our claim is completed.
Now we come back to the proof of the main theorem.
If mo = 0, we are done. If mo > 1, then Bmo-i n f~1(B(yo,ro))
contains a ball with radius r'Q x diamJ3mo_i. The same argument as in no
parabolic points case shows that there is B(y, k2r) C -B(zo. r)nF(/) for some
0 < fe2 < 1- Take k = min(&i,fc2).Then for any z0 G J(f), r > 0, B(z0, r)
contains a ball B(y, kr) C F(f). This means J(f) is shallow.
QED
REFERENCES
[Be] A.Beardon, Iteration of Rational Functions, Springer-Verlag, 1991.
[CG] L.Carleson and T.Gamelin, Complex Dynamics, Spring-Verlag, 1993.
[CJY] L. Carleson, P.W. Jones and J.-C. Yoccoz, Julia and John, Bol. Soc. Bras. Mat.
25, N . I (1994), 1-30.
[Ma] R.Mane, On a theorem of Fatou, Bol. Soc. Bras. Mat. 24 N . I (1993), 1-11.
[Me] C.McMullen, Self-similarity of Siegel disks and the Hausdorff dimension of Julia
sets, Acta Math. 180 (1998), 247-292.
[Mi] J.Milnor, Dynamics in one complex variable, Stony Brook Preprint 5 (1990).
[PR] F.Przytycki and S.Rohde, Porosity of Collet-Eckmann Julia sets, Fund Math. 155
(1998), 189-199.
[ST] M.Shishikura and Tan Lei, An alternative proof of Mane's theorem on non-
expanding Julia sets, in this volume.
[Ur] M.Urbanski, Rational functions with no recurrent critical points, Ergod. Th. and
Dynam. Sys. 14 (1994), 391-414.
INSTITUTE OF MATHEMATICS, ZHEJIANG UNIVERSITY, HANGZHOU, 310027,

P.R.CHINA
E-mail address: yinQmath.zju.edu.cn
On a theorem of M. Rees for matings
of polynomials
Mitsuhiro Shishikura
Abstract
We show the following theorem of M. Rees: if the formal mating of
two postcritically finite polynomials is (weakly) equivalent to a rational
map, then the topological mating is conjugate to the rational map.
The dynamics of complex polynomials has been extensively studied and many
results are obtained. For example, a polynomial of Thurston's type admits
a description of the dynamics in terms of external angles, etc (see [DH1]).
However, much less is known for rational maps in general (see [R] for degree
two case). The mating is a way to construct a rational map or a branched
covering of S2 from a pair of polynomials with connected Julia sets. When
this is possible, we can understand the dynamics of the rational map in terms
of that of the polynomials. The notion of the mating was also introduced by
Douady and Hubbard [D].
There are two notions of the mating —the formal mating and the topo-
logical mating. In fact there is another, the degenerate mating, which is a
slightly modified version of the formal mating. To define the formal mating,
we take two copies of C each with the circle at infinity and make a topological
sphere by gluing the copies together along the circles at infinity. Then define
a mapping on the sphere using one polynomial on one copy and the other
polynomial on the other copy. To define the topological mating, take the
disjoint union of the filled-in Julia sets of the polynomials and identify the
point with external angle 9 for one polynomial with the point with external
angle —9 for the other polynomial. Then there is a naturally induced map
on the quotient.
Now the question is when they are "equivalent" to a rational map (mata-
bility). For the case of degree two, there are several works [L], [W], [R] and
the final result on the combinatorial matability is obtained by Tan Lei [T].
As for the relationship between the two matings, M. Rees [R] proved:
/ / the formal mating of two hyperbolic polynomials of Thurston's type is
Thurston equivalent to a rational map, then the topological mating is conjugate
to the rational map. ([R] §0.15).
The main theorem (§1.7) of this paper claims that this is also the case even
if the polynomials are merely of Thurston's type and the formal mating can
be replaced by the degenerate mating. This generalization has an interesting
consequence. If we start with two polynomials of Thurston's type which have
empty filled-in Julia sets (consequently with Lebesgue measure zero), when
289
290 M. Shishikura
they are "matable", by gluing these Julia sets (which are "dendrites"), we
obtain not only a topological 2-sphere but also a complex structure on it such
that the induced dynamics is analytic with respect to this structure. For
example, this is the case when z2 + i is mated with itself.
The main idea of the proof is similar to [R]. However, the critical points in
the Julia set cause a lot of complications. The reader should notice that many
of our arguments are trivial in hyperbolic case. M. Rees says she has also a
proof for this generalization, so the author does not insist on the priority for
it. This paper is aimed simply to provide a proof available for the sake of
future development of the theory of matings.
In §1, we define the matings and state the main theorem. In §2, when two
polynomials are matable, we construct a semi-conjugacy h from the formal
mating to the rational map. In §3, we study the properties of h. In §4, we
conclude that two points are mapped by h to the same point if and only if
they are ray-equivalent. Here we use a generalization of Douady's lemma
(Remark 4.3(3)) instead of Lemma 3.7 in [R]. Together with this, the use of
the map G (Lemma 4.5) is the new idea, which made it possible to prove the
theorem in non-hyperbolic case.
I would like to thank A. Douady, J. H. Hubbard, Mary Rees and Tan Lei
for helpful discussions, especially Tan Lei introduced me the subject of the
mating, told me the importance of the theorem, finally encouraged me to
include this paper in this volume and helped the arrangement for retyping
the paper. I would like to thank also IHES and Akizuki project (founded by
Toyosaburo Taniguchi) for supporting my visit to IHES in the year 1989/90,
during which this paper was written. The original version of this paper had
been circulated as IHES preprint M/90/48 (1990).
1 Matings of Polynomials and Main Theorem

Let fi, f2 be monic complex polynomials of degree d (> 2). We consider only
the case where fi, f2 are of Thurston's type, i.e., all the critical points are
periodic.
We are concerned with two notions of the matings of / i and f2 —the
formal mating and the topological mating. The degenerate mating is a slightly
modified version of the formal one.
1.1 Formal mating. We add to the complex plane C "the circle at infinity"
which is symbolically denoted by { oo • e27ri* | 6 € K/Z }. Define
C = C U { oo • ,2m9
e
On a theorem of M. Rees for matings of polynomials 291
which is homeomorphic to a closed disc with the natural topology. Extend

the polynomials f}r. C -> C to C by setting
£(00 • e2**) = 00 • e2*lde.
Let Ci, C2 be two copies of C, and from now on we consider that fj acts
on Cj, i.e. /.,•: Cj —> Cj (J = 1,2). Let ~ be the equivalence relation in
Q LJ Q (disjoint union) which identifies 00 • e2jrie in Q and 00 • e'2irxS in
Q for 9 e R/Z. Then the quotient by ~ is topologically a two dimensional
sphere. So denote
In this paper, S2 is a notation specially reserved for this quotient. Finally,

the formal mating /1 iL / 2 : S2 -» S 2 is defined by
F = /, JL h =
which is well defined. This defines an orientation preserving branched cov-

ering of S2 onto itself. The equator Sl of S2 is the quotient image of the
circles at infinity of Q and Q . Here, S1 is also a notation reserved for the
equator. However, there is a natural correspondence between the equator
and the standard unit circle by 00 • e2*te e Q f ) e2rn9, and the restriction of
the formal mating, F: S1 -4 S1 corresponds to the map z »-» zd on the unit
circle.
Remark. Since we added the circle to C, we have to destroy the conformal
structure in order to introduce a differentiable structure on S2. However, we
can keep the conformal structure on a neighbourhood of the filled-in Julia sets
Kf; ( c C,-). Thus we can talk about conformal mappings in int K^ LJ int K/2.
We are interested in the relationship between F = /i_li / 2 and rational
maps. Since we have lost the conformality at least near the equator, it is too
optimistic to expect that F is conjugate to a rational map. So we introduce
a weaker notion of equivalence than the conjugacy.
1.2 Thurston equivalence. Let / : S2 —• S2 be an orientation preserving

branched covering. Define
Qf = {critical points of / } and
n>0
We say / is of Thurston's type, if # P ; < 00. Two such maps /, g are defined
to be Thurston equivalent if and only if there exist orientation preserving
292 M. Shishikura
homeomorphisms 6,9': S2 -> S2 such that 9 = & on Pf, 9(Pf) = Pg, 9 and
0' are isotopic relative to Pf, and the following diagram commutes:
s2 —- * S2
'I 1-
S2 - !Us".
Definition, / i and / 2 are strongly matable if /i-LL / 2 is Thurston equivalent to
a rational map.
Remark. Thurston gave a topological criterion for a branched covering to be
equivalent to a rational map (see [DH2]). Namely, if a branched covering /
is not equivalent to a rational map, then there exists a so-called Thurston
obstruction, which consists of disjoint simple closed curves in S 2 \ Pj with
certain homotopical growth property under f~n. This criterion is in general
difficult to check. However, in case of degree two, there are extensive works
[Ft], [T] using Levy cycle criterion which is proved to be equivalent to the
Thurston's. For higher degree, it is known [ST] that Levy cycle criterion is
not equivalent to Thurston's.
In 1.4, we will see how to reduce F when a "removable obstruction" exists.
1.3 Ray-equivalence. Let Kj be the filled-in Julia set of fj (j = 1,2). By

Douady and Hubbard [DH1], there exists a conformal mapping ipK, '• C\Kj -»
C \ D, where D is the open unit disc, such that:
ipKj(z)/z -> 1 as z —> oo;
<p~Kl extends continuously to the boundary 9 0 (and gives the parameterisa-

tion off dKj).
d )
Moreover ipj,1 also extends continuously to the circle at infinity by
<p-K) | 5 1 = i d s i .
The external ray for /,- of angle 9 6 R/Z is
and the closed external ray of angle 9 is its closure in S2, i.e.,
i f = /#> U W-K){e2^)} U {ex, • e™ € C,-}.

We define the ray equivalence ~ on S2 as the equivalence relation gener-

ray
ated by
x ~ y if x and y are in the same closed external ray.

ray
Hence if x and y are on two external rays which touch each other either at the
equator S1 or at the Julia sets dK\ U dK2, then x and y are ray-equivalent,
in other words, two point are ray-equivalent if and only if they are connected
by a path consisting of closed external rays.
Remark. If an external ray in a ray-equivalence class is periodic (as a set),
then all other rays in this class are also periodic with the same period. It
follows that if a ray-equivalence class contains a preperiodic point, then this
class contains only finitely many external rays.
1.4 Degenerate mating. F': S2' -> S2'. Let
V = {ray equivalence class containing at least two points of Qp U Pp},

Y = {£ | £ is a ray equivalence class intersecting Qp U Pp and
there exist m, n > 0 r/ e Y' such that / m (£) = /"(??)}.
If Y = 0, we define S2' = S2 and F' = F. If Y ? 0 and all £ e Y are

trees (i.e., contain no non-trivial loop), then we define the degenerate mating
F' = S2' -> S2' of /i and / 2 as follows: In S 2 , shrink each ( € F to a point
(see the remark above), then this makes a new topological sphere, which is
denoted by S2'. Let TT: S2' —> S 2 ' be the natural projection.
There exists a continuous map F: S2' —> S 2 ' such that Fow = TTOF and it
is not a branched covering only in a neighbourhood of n;(F~l(Y) \ Y). Each
component oiix{F"x{Y)\Y) is a tree and its image by F is a point. So we can
modify F in a small neighbourhood of TT(F~1(Y) N Y) to obtain a branched
covering F': S2' -> S 2 '. This is well-defined up to Thurston equivalence.
If Y 7^ 0 and Y contains £ which is not a tree, then we do not define the
degenerate mating.
When Y ^ 0 and all £ € Y are trees, take disjoint closed disc neighbour-
hoods U$ for £ € Y. Then { dU^ | £ 6 K } gives a Thurston's obstruction for
F. The construction of F' means that this obstruction is removable in the
sense that we can reduce it so that the new branched covering is defined.
1.5 Matability.
Definition. Two monic polynomials are matable if and only if the degenerate
mating exists and is Thurston equivalent to a rational map.
294 M. Shishikura
1.6 Topological mating. Using the ray-equivalence, we can define the quo-
tient ~space~S2~/âind the topological mating F* = [/i±L/ 2 ]*: S21 ~—>
ray ray
S2/ ~ as the induced map from F. Since every ray lands on a point in dKtl
ray
or dKf2, we have
S7~=A' A UA7 2 /~,
ray " ' ray
where the point in dKf{ with external angle 6 is identified with the point in
8Kh with external angle -6 (0 € K/Z).
In general. S21 ~ is not always homeomorphic to the sphere. It is not
ray
known if the quotient is always a Hausdorff space. In [ST], we gave an example
for which S2/ ~ is a sphere but f\ and fa are not matable.
ray
1.7 Main theorem. Our main result in this paper is:

Main theorem. Let / i , /2 be monic polynomials of Thurston's type of degree
d (> 2). If the degenerate mating F' = (/i-LL $2)' is Thurston equivalent
to a rational map R: C —> C, then there exists a continuous semi-conjugacy
h: S2 -^-C such that:
(i) the following diagram commutes:
c —5->. c,
where F = f\ J l /2 is the formal mating;
(ii) h is a uniform limit of orientation preserving homeomorphisms;
(iii) h is conformal in intK;l LJ int Kf2 onto C \ JR and h~l(C N JR) =
int A^/, LJ int X/ 2 ;
(iv) for x, y G 5 2 , /i(x) = /i(y) if and only if x and y are ray-equivalent.
In other words, the topological mating F*: S2/ ~—> S2/ ~ is conjugate
ray ray
to R by h: S2/ ~->- C induced by h.
ray
Remark. If both f\ and / 2 are hyperbolic, this result was obtained by M. Rees
(Lamination Map Conjugacy Theorem [R] §0.15). However, she stated the
theorem in terms of laminations.
Corollary. Under the hypothesis of the theorem, the ray-equivalence is closed.
Moreover, II\E'- E -> JR induces a semi-conjugacy between zd\si and R\jR-
This semi-conjugacy is a uniform limit of topological embeddings.
For further properties of h, see Proposition 3.1.
2 Construction of the semi-conjugacy

In this section, we show the existence of semi-conjugacy from the mating to
a rational map, when the polynomials are matable.
2.1. Theorem. Suppose that the degenerate mating F' = (/i-lL /2)' of poly-
nomials /i and / 2 is Thurston equivalent to a rational map R: C —• C. Then
there exists a continuous mapping h: S2 —> C, satisfying (i), (ii) and (iii) of
Main Theorem 1.7.
Proof. The main idea of the proof is to use the lifts of the equivalence maps
9,9' and to show their convergence using the expanding metric for R.
To make the idea clearer, we first deal with the case F' = F, i.e. the case
where there is no degenerate Levy cycle for F. We shall see later the case
F' ^ F, i.e. the case where F' is really degenerate.
Assuming that F' — F is Thurston equivalent to R, we shall construct
sequences 9n, Hn (n — 0 , 1 , . . . ) and a neighbourhood V in S2 of super-
attracting periodic orbits of F with F(V) C V (V = 0 if there is no super-
attracting orbit), satisfying the following conditions:
(a) 9n: S2 -> C is an orientation preserving homeomorphism;
(b) 6noF = Ro0n+x-
(c) 0n = 9n+i on PF U F"n(V), 9n{PF) = PR and 9n is conformal in F" n (V');
(dl) Hn: S2 x [0,1] -» C is continuous;
(d2) //„(•, t) is a homeomorphism for any t e [0,1];
(d3) Hn(;0) = 6n,Hn{;l) = 6n+l\

(d4) Hn(x,t) = 9n(x) for xePFU F~n{V), t € [0,1].
First of all, by the definition of Thurston equivalence, we have 60 — 9, 9X = 9'
and Ho satisfying (a)-(d4) except the conditions on F~n(V) in (c) and (d4).
When F has a super-attracting orbit, we need to modify 90, 9\ and Ho as
follows so that these conditions are satisfied.
Let { C; I 3 = 0 , 1 , . . . ,p - 1} be a super-attracting orbit of F (Q —
F>(Co) and Co = Fp(Co)), and denote <j = 90(Q, kj = deg 0 F and k =
kok{ ...kp-y. For each Cj (j = 0 , 1 , . . . , p - 1), there exists a local coordinate
function uij from a neighbourhood Wj of Q into C such that tUj(Cj) = 0 and
RP: Wj —> (wj)k in this coordinate. A Dehn twist D^ around £'• of angle t is
a homeomorphism of C defined by
Dlj): Wj-> eitv^ • Wj in Wj

296 M. Shishikura
and
D[tj) = id in C \ Wjy
where v is a continuous function such that v(r) = 1 (0 < r < rj), v(r) = 0
(r > r 2 ) with 0 < r, < r 2 and {\WJ\ < r 2 } C Wj.
For two homeomorphisms 9,8': S2 —> C, we write 0 ~ 9' is there is an
isotopy H: S2 x [0,1] -> C such that #(•, 0) = 0, //(•, 1) = 0' and H{x, t) =
0(:r) for x e PF and t E [0,1]; we write 0 ss 0' if in addition there is a
neighbourhood V of {(,-} such that H(x,t) = 9{x) for x e V and t g [0,1].
Since {(,} and {Q} are super-attracting and deg^ F = deg(, R, there
exists a local conformal conjugacy between F and R near these orbits. Hence
there exists a homeomorphism ff0: S2 -> C such that 0J, ~ 90, O^ ° F = Ro 0'Q
near Q and 0J, is conformal near Q. Since F and R are branched coverings
branching only over Pp and P# respectively, we can lift the isotopy between
do and 0'0 to an isotopy starting from 6\. So we obtain a homeomorphism
9\: S2 -> C such that 9'0 o F = R o 0[ and 9[ ~ ^ ~ 6>0 ~ <?o- H e n c e
Ro6'0 = Roffi near C r It follows that
for some tj 6 •pZ. On the other hand, for each j and s € K, there is a
homeomorphism D : C ->• C such that £>P+1) o /? = /? o D and D % £ ) ^ ,
where £)« ' = Z?i° and « is defined similarly as above.
Let ^ = Dll] O Dil' o • • • o Z?^:' 1 o 9'0, w h e r e 3 are determined below.
Then there exists a homeomorphism 6\: S2 -> C such that 6\ ~ 9\ and
0* o F = /? o 0*. By the above remark, we have
#* ~ n ( 0 ) o n(1) o • • • o n(p-1) o «'
where s p = s0. Hence 0J « 9* if Sj = tj + Sj+i/fcj (j = 0 , 1 , . . . ,p - 1). And

this equation has a solution (in T^ZYyZ) since fc = koki... A;p_i > 1. Using
this solution (SJ), we redefine new 0O (resp. 8i) to be 02 (resp. 9\) and new
i/ 0 to be the isotopy fixing PF and a neighbourhood V of {Q}. If there are
several super-attracting orbits, this procedure can be done for those orbits.
Finally we can take V so that F(V) C V. Thus for new 90,9i and Ho, the
conditions (a)-(d4) are all satisfied.
Now we define 0n, Hn inductively. Suppose we have 0 n _i, 9n and Hn^x
satisfying the conditions. Considering F and R as branched coverings, we
can lift Hn..i by F and R to obtain Hn: S2 x [0,1] -> C such that: # „ is
continuous; Hn(-, 0) = 0 n ; and
(e) Hn.l(F(x),t) = R(Hn(x,t)).
Then define 9n+\ — Hn(-, 1). It is easy to check the conditions for 0n+l, Hn.
Thus we have constructed the sequences 9n, Hn.
To show the convergence of 9n, we need the following:

2.2. Theorem (Douady-Hubbard [DH1]). Let R: C -> C be a rational map
of Thurston's type and U(C C) an open set whose closure does not contain
super-attracting orbits. There exist A > 1 and a Riemannian metric ||-|| on
U with a finite number of singularities such that
if z, R(z) e U and v 6 TZC, then ||T2i?(i;)|| > A||u||;
and at the singularity, ||-|| has the form u(z)|dz|/|z|" with 0 < /3 < 1 and
u{z) > 0 continuous. This metric defines the same topology on U as the
standard one. Moreover, the metric d defined by ||-|| satisfies
(spherical distance of x and y) < K • d(x, y) if x, y 6 U .
Even though this theorem says that R is infinitesimallv expanding on JR,

R is not locally expanding, even not locally injective whenever fi/j D JR ^ 0.
So we use the following notion.
2.3. Definition. Let U: [0,1] -> U \ PR be an arc. The homotopic length of
n is
/i-length(a) = inf{length(c*') | a' is a rectifiable arc in U \ PR,

homotopic (in U \ PR) to a fixing end points},
where length(a') is with respect to ||-||.

2.4. Lemma, (i) d(a(0),a(l)) < /i-length(a).
(ii) Let U' be a set such that U' C U. For any e > 0, there exists S > 0 such
that if a is an arc in U' \ PR with diam(a) < 8 then /i-length(a) < e,
where diam(-) is the diameter with respect to the spherical metric.
(iii) If a is an arc such that a, R(a) c U \ PR, then h-length(R(a)) >

A • /i-length(a).
The proof is left to the reader.
In Theorem 2.2, we may take U so that U' = C^60(V) C U. For x G S 2 \ (Vn
P F ), let us consider the arc Hn{x, •): [0,1] -> C. If x e S2 \ (F-n(V) U Pf),
then the image of this arc is contained in C^(0n(F-n(V))UPR) C C \ (9n(V)U
PR) = U' \ PR. SO we can define the homotopic length for Hn. By (e) and
298 M. Shishikura
Lemma 2.4(ii), if a; € S2\(F-n{V)UPF), then F(x) 6 S2\(F^n-^{V)L)PF)

and
/i-length(i/ n (z,-)) < j(
On the other hand,
C= sup
because of the uniform continuity of i/ 0 and Lemma 2.4(ii). Using the con-
dition (c), we have
(fl B (i),fl n+1 (x))= sup d{Hn{x,Q),Hn(x,l))
< sup /i-length(# n(a;, •))

2
This shows the uniform convergence of 6n with respect to the spherical metric
of C. Let h = limn-+oo 9n- Then h must satisfy h o F = i? o h by (b). Other
properties of /i are also easy to check, using the fact that
n>0
B
C \ JR = (J /2- (0o(V)) = | J
n>0 n>0
Thus the theorem is proved in the case F' — F.

Let us consider the case F' ^ F. Let TT: S2 -4 S2' be the collapsing
map, and F: S2' —> 5 2 ' such that TX o F — F o -K (see Definition 1.4). By
the assumption, there exist the equivalence maps 8, ff: S2' -4 C such that
0 o F' = R o ff etc. Define
QF,={xeS2'\irl(x)?lpt},
Note that QF consists of trees and QF> C Qf U PF<, QR C QR U PR.

We construct the sequences #„, Hn and V as before with the modified
conditions (a')~(d4'):
(a') 0n: S2 -+ C is continuous, flnIS2 \ F~n{QF): S2 \ F~n(QF) -> C \

R~n(QR) is orientation preserving homeomorphism, and for every z e
/J" n (Q«), ^ ' W is a tree;
(b') = (b); (c') = (c) with PF, PR replaced by PFUQF, PRUQR, respectively;
(dl') = (dl);
(d2') Hn(x, t) $ PR for x$PFUQF,te [0,1];

(d3') = (d3); (d4;) = (d4) with PF replaced by PF U QF.
First, we may assume that 9 = & on PFi U QF< and the isotopy between
9 and 9' fixes QFi as well as PFi, by lifting 9, 9' and the isotopy by F' and
R and replacing them by the lifts. We can modify them again, as before, so
that 9 and 9' coincide and are conformal in a neighbourhood V = 7r(V) of
the super-attracting orbits of F', and the isotopy fixes V.
By the definition of F', there exists a continuous map rj — S2' -> S2'
such that F = F' o T]\ r) is homotopic to the identity by a homotopy which
differs from the identity only in a small neighbourhood of F~l(QF') \ Qp.
Now define 90 = 9 o TT and 9\ — 9' o rj o n, and construct the homotopy Ho
from (isotopy between 9 and 9')oir and 0'o(homotopy between r/ and id)o7r.
Then it is easy to see that 90, 9y, Ho satisfy the conditions (a')-(d4'). By
the lifting argument, we can construct sequences 9n, Hn, since the condition
(d2') is enough to guarantee the possibility of lifting. Then using (b') and the
homotopy Hn, one can check (a') for 9n+i-
Note that we can take Ho with the additional condition:
H0(x,t) i90{V)loxxiV.
Hence for all n > 0,
Hn(x,t) i R~n(90(V)) for x i F~n(V).
Now the same argument as before applies to show that 9n converges.

Finally it is easy to see that 0 n 's can be approximated by orientation
preserving homeomorphisms (by (a')). •
3 Properties of the semi-conjugacy

Here are some properties of h.
3.1. Proposition. Let F, R and h be as in Theorem 2.1. Then we have:
(i) h is surjective; for any z € C, h~x(z) is closed, connected and non-

separating (i.e., its complement is connected).
(ii) If x and y in S2 are ray-equivalent, then h{x) = h(y).

300 M. Shishikura
(iii) F: h~l(z) -> h~1(R(z)) is a map of degree degzR. More precisely, for
zeC,xeh-l{R(z)),
^2 degy F = deg. J?,

i( 2 )
where degy F, deg2 /? are the local degrees of F, R at y, z, respectively.

In particular, if z is non critical for R, then F|ft-i(2j is injective.
Proof, (i) These properties follow from the fact that h is a uniform limit of
homeomorphisms. The surjectivity and the closeness of h~l(z) are easy to
see.
Suppose that h~1(z) is not connected, i.e., there are open sets U\ U2 in S2
such that/i" 1 (2) C UiUU2, UXC\U2 = 0 and Uinh~l{z) ^%,U2Dh~x(z) ^ 0.
Then h{S2 \ U\ U U2) is a compact set not containing z. Take a connected
neighbourhood of z such that V D ^S2 \U\U U2) = 0. Since /i is a uniform
limit of homeomorphisms, there exists a homeomorphism h': S2 —>• C such
that
d(h, h!) < min{distance (2, dV), distance (V, h(S2 \ ( / , U i/2))}.
Then it follows that h'{S2 \ Ux U I/a) n V = 0, hence h'"l(V) Cl/iU f/2l and
/i'(/i-'(z)) C V hence ^ n /i'-'(K) ^ 0, (72 n ^"'(V) ^ 0. This contradicts
with the fact that /i'~'(V) is connected.
Suppose thatft"'1(2) is separating, i.e., there exist non empty sets Vi and
V2 in S2 such that Vx n Va = 0 and S2 \ ft"1^) = V, U V2. For any w 6
C \ {z}, either /i"'(u») C V\ or /i~'(u)) C V2, since h~l{w) is connected.
Define W{ = {w € C x {z} | /i"1 (u;) c V^} (i = 1,2). Then ^ n W2 = 0 and
jy, u W2 — C s {z}. By the continuity of h, the set function w -> h~x(w)
is "upper semi-continuous", hence Wi are open. Since Vi were non-empty, so
are Wi. This contradicts the fact that C \ {z} is connected,
(ii) Fix r0 > 1, 6 E R/Z, j — 1 or 2, and consider the arc
a = ifi^dre2"9 | r0 < r < 00}).
Then Fn(a) = ^{{re2*""9 | r f < r < 00}) and
diamF n (a) -^ 0 (in S2) as n -> 00.
By the uniform continuity of h, we have
diamRn{h{a)) = diam/i(F"(a)) -^ 0.
Note that Rn(h(a)) C JR.

If (Un^"(M a ))) nPR^$, there is an m such that Rm(h(a)) contains a

repelling periodic point. Since Rm(h(a)) is connected and diamRJ1 (h(a)) —> 0
(n ->• oo), /?m(/j(a)) must be one point. Then so is h(a), since h(a) is
connected and h(a) C R~m{Rm(h(a))).
If (U n -^"(^( Q ))) n -PR — $> t n e n w e c a n u s e t n e homotopic length (Defi-
nition 2.3). By Lemma 2.4, we have
h-length{h(a)) < {^)n(h-length(Rn{h(a))) -» 0.

A
Hence in any case we have /i(^^1(roe2'r**)) = h(y]l(oo • e2m6)). By the con-

tinuity of tpx1 and h, we conclude that /i({v?tf!(re2'rtfl) I 1 < r < °°}) is
one point. Then the definition of the ray-equivalence implies that each ray-
equivalence class is mapped to one point by h.
(iii) Define the function
N{z,x)=
for z £ C and x € S2. It is easy to see that this function is lower semi-
continuous. On the other hand, for fixed z, N(z,x) is upper semi-continuous
as a function of x 6 h~i(R(z)}, since
N(z,x)=d-
Since h~l(R(z)) is connected and N is integer valued, N(z,x) is constant

on this set and it is denoted by N(z). For any z e C, N(z) > 1, by the
surjectivity of h. If R(z) is not a critical value, we have N{z) = 1, since
Ylz'eR-URiz)) •^( 2 ') = ^ an( ^ # ^ " 1 ( ^ ( z ) ) = ^- When w is a critical value of
R, by approximating w by non critical values, one can show that
N(z) > deg2 /? for 0 € i?" 1 (w).
Again using YlzeR-l(w) N(z) = d, we obtain N(z) = degz R for any z £ C . D
4 Proof of (iv) of Main Theorem

In this section, we prove the property (iv) of the semi-conjugacy h. The key
of the proof is:
4.1. Proposition. Let / i , f2, R and h be as in Theorem 2.1. Then #h~1(z)nSi
is bounded for z E C, where S 1 is the equator of S2.
Assuming this result, we can now complete:
302 M. Shishikura
Proof of Main Theorem. By Proposition 3.1(ii), h~l{z) consists of ray-equiva-

lence classes. By the above proposition, there are only finite number of ray-
equivalence classes in h~l(z) and each of them contains a finite number of
external rays. Then it follows that each ray-equivalence class is closed. So
h~x(z) can contain only one ray equivalence class, otherwise h~~l{z) would
be disconnected. In other words h(x) = h(y) if and only if x and y are ray
equivalent. •
Proof of Proposition 4.1.

4.2. Lemma. Let Xn be compact metric spaces and gn: Xn —> Xn+i maps
(n = 0 , 1 , 2 , . . . ) . Suppose that there exist 0 < /x < 1, 6 > 0 and an integer
N > 0 such that
(i) f o r n > 0 , x,y £ X n ,
if d(gn{x),gn(y)) < 6, then d(x,y) < fJ-d{gn(x),gn(y));
(ii) each Xn can be covered by N sets of diameter < 6.
Then#Xn<N (n = 0 , l , 2 , . . . ) .
4.3. Remark. (1) The condition (i) implies that gn are injective.
(2) In our application, Xn are always subsets of S1 and the metric is the one
which is induced from R/Z (hence the total length of S 1 is 1). Then the
condition (ii) is trivial, in fact one can take an integer N greater than 1/6.
(3) Lemma 4.2 is a generalisation of Douady's lemma which says:
Let X be a compact metric space and g: X —> X a continuous locally expand-
ing map. If g is injective on X, then X is finite.
To prove this from Lemma 4.2, put Xn — X and gn = g, and check the
condition (i) from the local expanding property and injectivity of g. (See the
proof of Lemma 4.4).
Proof of Lemma 4.2. Note that, by condition (i), if U C Xi+i is a set with
diam(£/) < S then U' = g^{U) has:
diam U' < fi diam U < fiS.
By (ii), Xn is covered by N sets U\,...,UN with diamUi < 6. Then Xn-\

will be covered by ./V sets U[,...U'N with diam U[ < /J, diam Ui < /J.6, where
U- = gZl{Ui). Repeating this procedure, one obtains N sets [/*,..., U^ with
diam U* < nn8 which cover Xo. Since / x n ( 5 - > 0 a s n - » o o , Xo cannot contain
more than iV points. A similar argument applies to Xn. D
4.4. Lemma. Let V be any open neighbourhood in C of QR D JR. Then there
exists 6 > 0 such that
(i1) For z e C x V, x, y 6 h'^z) n S\ if d(F{x),F(y)) < 6, then d{x,y) =
Ld(F(x),F(y)).
On a theorem ofM. Rees for matings of polynomials 303
Proof. If not, one can choose xn, yn 6 Sl such that
h(xn) = h(yn) e C \ V, d(F(xn), F(yn)) -> 0 and d(zn, yn) > —.
(Note that if d(x, y) < ^ then d{F(x), F(y)) = d • d(x, y}.)

Extracting a convergent subsequence, we may assume that xn -* x, yn -> y.
Then we have h(x) = h(y) £ QR D JR, d(F{x),F(y)) = 0 and d(x,y) >
^ > 0. This contradicts with the fact that F\h-i^x)) is injective. (Proposi-
tion 3.1 (hi)). D
If R is hyperbolic, we may take V = 0. In this case, applying Lemma 4.2
to Xn — h~l(Rn(z)) n Sl and gn — F\xn, one obtains Proposition 4.1.
Now we consider non hyperbolic case. The problem is the behaviour of
h~1(z) in the neighbourhood of QR n JR. For z e ^ U PR, there are only
finite number of external rays in h~1(z), by Remark 4.3(3).
4.5. Lemma. Let w € QR D JR. Then there exists a neighbourhood U of
h~x(u}) in S2, a homeomorphism G: U ->• G(U) C S 2 , and a neighbourhood
V of UJ in C satisfying:
If z e V, then h-1{z) C U and there is u; = G(z) 6 C such that
G: h~l{z) -¥ h~1(w) and it is injective; and
(i") There exists 8 > 0 such that for z e V, x,y 6 /i"^ 2 ) n 5" w i t f l
d(C(x),G(y)) < S, we have d(z,y) < ±d(G{x),G(y)).
Proof. Let I > 0 be the integer such that a = ${<*}) is the first periodic point
in the orbit of w. Then h"1(a) is periodic as a set under F. Let p > 0 be the
period of external rays in h~1(a), so that all the external rays in h~x(a) are
fixed under Fp. We construct G so that it satisfies the commutative diagram
(S2,h-
In the above diagram we consider Fl as a branched covering from a simply
connected neighbourhood of h~l(w) to a simply connected neighbourhood of
h~l(a). Then we can lift F9 by Fl to obtain G, since F"1 is homotopic to
identity in this neighbourhood by a homotopy fixing the points in h~l (a)C\PR.
Now we verify the properties of G. By continuity of h there exists a
neighbourhood V of u such that if z 6 V then h~x(z) C f/. By the above
diagram, we have G{h-X(z)) c F-'(F i + p (/i- l ( w )) = h-^R-^Rîcu))) which
is a disjoint union of h~l(zi)'s (zi e R~'(Rl+p(u)))). Since /i-1(2:) is connected,
h^1(z) must be contained in one of h~1(zi))s. Let w = G(z) = Z{. Note that
F p | s i in a neighbourhood of h^(a) n 5 l is locally an expanding map with
304 M. Shishikura
factor dp, then so is G|si in a neighbourhood of /i -1 (w) fl S1. So one obtains

(ii'), restricting the neighbourhood if necessary. •
Let QR PI JR = {ui,..., u>k}- Denote the U, G, V, G, 5 in Lemma 4.5 for
w = uii by Ui,Gi,Vi,Gi,6i. We may take VJ's disjoint. Let V = (Jj V{ and
denote by <50 the 5 in Lemma 4.4. Finally let /x = 1/rf, 5 = mino<i<K- (5i and
an integer TV > 1/5.
Now, take any z e JR and set z0 = 2. Define Xn,gn,zn+1 inductively as
follows:
Xn=h-1(zn)nS1;
if zn <£ V, then gn = F\Xn and zn+i = R{zn);
ii zne Vi, then gn = Gi\Xn and zn+i =
Then {X n ,y n }, together with n,S,N as above, satisfies the assumptions (i),
(ii) of Lemma 4.2. Hence
#h-*{z)nSl=#X0<N
and this bound does not depend on z. Thus Proposition 4.1 is proved. •
References
[DHl] A. Douady and J. H. Hubbard, Etudes dynamiques des polynomes
complexes, I and II, avec la collaboration de P. Lavaurs, Tan Lei et
P. Sentenac, Publication Mathematique d'Orsay, 84-02 (1984), 85-04
(1985).
[DH2] A. Douady and J. H. Hubbard, A proof of Thurston's topological char-

acterization of rational functions. Ada Math., 171 (1993), 263-297.
[D] A. Douady, Systemes dynamiques holomorphes, (Bourbaki seminar,
Vol. 1982/83) Asterisque, 105-106 (1983), 39-63.
[L] S. Levy, Critically finite rational maps, Ph. D. Thesis, Princeton Uni-
versity, 1985.
[R] M. Rees, A partial description of parameter space of rational maps of
degree two: Part I. Ada Math., 168 (1992), 11-87.
[ST] M. Shishikura and Tan Lei, A family of cubic rational maps and mating
of cubic polynomials, preprint, to appear.
[T] Tan Lei, Matings of quadratic polynomials, Erg. Th. & Dyn. Sys., 12
(1992), 589-620.
Oil a theorem of M. Rees for matings of polynomials 305
[W] B. Wittner. On the bifurcation loci of rational maps of degree two.

Ph. D. Thesis, Cornell University, 1986.
Hiroshima University, Department of Mathematics, Kagamiyama, Higashi-

Hiroshima 739-8526, JAPAN
Le theoreme d'integrabilite des
structures presque complexes
Adrien Douady
d'apres des notes de X. Buff
Abstract
We give a proof of the theorem of Morrey-Bojarski-Ahlfors-Bers in
the L2 framework, using Fourier transform. We prove successively the
following results:
Th. 1 : For ft 6 L ^ ^ C ) with ||/i||i,°° < 1 , there is a unique
$ = $^ in HîC) which is a solution of the Beltrami equation defined
by ft , such that $(z) — z tends to 0 at infinity. Moreover, if ft depends
analytically on a parameter A , then $^ depends analytically on A .
Th. 2 : If ft is C°° on a neighborhood of z0 with \ft(zo)\ < 1, the
Beltrami equation has local solutions which provide local C°° charts.
Th. 3 : If ft <= C^pCC) with \\ft\\Lco < 1, the map ^ is a C°°
diffeomorphism of C onto itself.
Th. 4 : For n in L ^ ^ C ) with \\P\\L°° < 1, the map $„ is a
homeomorphism of C onto itself.
Th. 5 : For fi in L°°(C) with ||/i||i» < 1, there exists a unique
$ = $Ji in /f/0C(C) which is a solution of the Beltrami equation defined
by ft, and a homeomorphism of C onto itselffixing0 and 1. Moreover, if
ft depends analytically on a parameter A, then ${, depends analytically
on A.
Nous donnons une demonstration du Theoreme de Morrey-Bojar-
ski-Ahlfors-Bers dans le cadre L2, en utilisant la transformee de Fou-
rier. Pour 1'enchainement des resultats, voir le paragraphe 3 de l'article.
1 Motivation
Soient U et V deux ouverts de C et (j>: U —> V un diffeomorphisme C°° (au
sens reel) preservant l'orientation, ce que nous ecrivons <j> € Diff+(U;V).
x l
Pour x 6 U, notons S^{x) l'ellipse Tx(j)~ (S ), consideree a homothetie reelle
pres, et posons n<f,{x) = §|/§f (i) • On a |/i^,(a;)| < 1. Les conditions suivantes
sont equivalentes :
(1) L'application <j) est conforme, i.e. holomorphe ;
(2) £$ est le champ de cercles ;
(3) M* = 0.
La donnee du champ d'ellipses S^ sur U (toujours a homothetie reelle
pres en chaque point) et celle de la fonction / ^ : ! / - > D sont equivalentes :
307
308 A. Douady, d'apres des notes de X. Buff
l'ellipticite (rapport des longueurs des axes) K de £4,(1) et k = \n<t>{x)\ sont

lies par K = | ^ | , i.e. k — | ^ j , et Arg(fi) — -28 ou 9 est l'argument du
petit axe de £#(x).
On dit que (j> est un difFeomorphisme K-quasi-conforme si le champ £$
est a ellipticite bornee par K, i.e. si H/x^l^ < ~^. Nous verrons com-
ment definir la quasi-conformite pour des applications qui ne sont pas des
diffeomorphismes.
La question suivante est naturelle : Etant donne un ouvert U de C et un
champ d'ellipses £ sur U, peut-on trouver un diffeomorphisme 0 : U —> V tel
que £ = £<t> ?
Question equivalente : Etant donne une fonction /i : U —>• D, peut on
trouver un diffeomorphisme (j> : U —> V tel que \i = n^, ?
Si on veut pour <f> un difFeomorphisme C°°, une condition necessaire est
bien sur que \i soit C°°. En termes de £, cette condition s'exprime moins bien.
Nous verrons (Th. 2 et Th.3) que cette condition est suffisante localement,
et aussi globalement si U = C et JJ. a support compact.
Mais pour les besoins de la chirurgie holomorphe comme pour ceux de
l'etude des groupes kleiniens, on est amene a considerer V equation de Beltrami
pour des fonctions n qui ne sont pas continues. On cherche alors des solutions
<j) qui ne sont pas C 1 , et les derivees partielles ff et | | doivent etre prises au
sens des distributions.
Une condition naturelle a imposer a <> / est alors d'etre dans l'espace de
Sobolev i// oc . C'est en tous cas une bonne presentation de la theorie. Nous
traitons le cas fj. E £^mP avec ||/i|k~ < 1) nous reservant de nous debarrasser
de l'hypothese [i a support compact a la fin. Les resultats que nous presentons
sont dus a Morrey, Bojarski et Ahlfors-Bers. La demonstration que nous
donnons utilise seulement des techniques L2 et la transformee de Fourier.
2 Quelques espaces fonctionnels

Soit U un ouvert de C. On note C{U) (resp. Cl{U), C°°{U)) l'espace vectoriel
des fonctions continues sur U a valeurs dans C. On note CCOfnp([/) l'espace des
fonctions a support compact dans U, et on definit de meme Clom-, C'^mp{U).
L'espace L2(U) est le complete de Ccomp(C/) pour la norme IS definie par
(ll/lb)2 = //l/l 2 = / / I/WI2 dxdy

ou z = x + iy. C'est un espace de Hilbert complexe.
Nous preferons cette definition a celle qui consiste a voir les elements
de L2(U) comme des fonctions de carre integrable definies a une fonction
negligeable pres. En effet, suivant une morale heritee de Laurent Schwartz,
nous rechignons a evaluer une fonction en un point lorsqu'elle n'est pas con-
tinue. II nous arrivera cependant de parler des elements de L2{U) comme de
"fonctions".
On designe par L2comv(U) I'espace des fonctions de L2(U) a support com-
pact dans if, et par L2OC(U) I'espace des fonctions sur U qui sont localement
dans L2. On defiuit de rneme LX{U), Llmmp{U), Llloc(U). L'epace L}JU) est
le plus grand des espaces de fonctions sur U que nous considerons.
Pour L°°(U) il faut etre un peu plus soigneux : la boule fermee de rayon
r dans L°°(U) est la fermeture dans Lloc(U) de l'ensemble des / e C comp (t/)
tels que ||/||L<» < r (la topologie sur LjÛ) etant celle induite par le produit
des Ll{U') pour U' relativement compact dans U).
On ecrit L2 pour L 2 (C), etc.
Pour / € Ll[oc(U) et (/ n ) une suite d'elements de Ljoc(U), on dit que fn
tend vers / au sens des distributions sur U si
Pour f et g dans Ljoc(U), on dit que g = -£ au sens des distributions si
tdh
Ox '
La fonction g est uniquement determinee (comme element de Llloc{U)) par
cette condition.
On definit I'espace de Sobolev Hloc(U) comme I'espace vectoriel des fonc-
tions / € Ll>c(U) admettant des derivees partielles J£ et | ^ au sens des
distributions dans L2loc(U) . Pour / e HjÛ), on definit f{ et §£ par les
formules habituelles f£ = \{%- - i%-), % = U^- + i | f ) . Si U = 0, f est
J
oz Z ^ ox oy'' uz I ^ ox oy' oz '
holomorphe (lemme de Weyl).
Une fonction de Hloc(U) n'est pas necessairement continue, et c'est bien
dommage car cela faciliterait la demonstration du Th. 4. Par exemple la
fonction z H-> log | log \z\\ appartient a /// 0C (C).
3 Enonces
Theoreme 1.- Soit fi € ^Smpi^) avec IIMII/-00 < 1- Alors il existe une
fonction $ = $ , , £ ///^(C) et une seule qui soit solution de I'equation de
Beltrami
definie par /x, et telle que $(z) — z tende vers 0 quand z —> oo.
De p/us, si /i depend analytiquement (resp. continument) d'un parametre
X, la fonction $^ depend analytiquement (resp. continument) de X.
Theoreme 2.- Si \i est C°° au voisinage de z0 avec |M(ZO)| < 1, I'equation de

Beltrami definie par /x admet une solution locale qui est un diffeomorphisme
C°° d'un voisinage de ZQ sur un ouvert de C.
Theoreme 3.- Si /x 6 C^ n p (C) avec ||/X||L°° < 1, I'application $,, est un

diffeomorphisme C°° de C sur lui-meme.
Theoreme 4.- Pour /x G L^ m p (C) auec ||/i||z,°° < 1, I'application $ M est un

homeomorphisme de C sur lui-meme.
Theoreme 5.- Pour zz G L°°(C) avec \\H\\L<*> < I, U existe un $ = $J, G

et un seuZ qw soit solution de I'equation de Beltrami
et qui soit un homeomorphisme C —>• C avec $(0) = 0 et $(1) = 1.

De plus, si /j, depend analytiquement (resp. continument) d'un parametre
X, la fonction $J, depend analytiquement (resp. continument) de X.
4 Rappels sur la transformation de Fourier

On ecrit L2 pour L2(C), etc.
Pour / 6 Lx, on definit la transformed de Fourier f de f par la formule:
fiO = fit + iv) = fff(x + iy)e~2i dxdy.
On a alors f £ Co, i.e. / est continue et tend vers 0 a l'infini,
7^9 = f - g , II/IU-<
La derniere egalite permet de definir la transformation de Fourier /
comme isometrie de L 2. Pour f et g dans L2 , on a
Fg = f * g , /(*) = / ( - * ) et H/IU-<
Pour / 6 Llloc telle que / soit definie, et que §£ G L2 et ^ e L 2 , on a
et %-=
d
d'ou
On dit qu'une fonction / € L°° est a decroissance rapide si / 6 O(l/\£\k)

pour tout k 6 N. On appelle espace de Schwartz et on note S l'espace
vectoriel des fonctions C°° a decroissance rapide et dont toutes les derivees
sont a decroissance rapide.
L'espace de Schwartz est stable par multiplication, par convolution et par
transformed de Fourier: on a / € S si et seulement si / € S. Par ailleurs, si
/ est C°° a support compact, alors / est a decroissance rapide, et si / € Llloc
est a decroissance rapide, alors / est C°°.
Definition : On appelle transformation de Hilbert-Bending I'isometrie

2
£ : L —> 1? definie par
Si / € Iw P > la fonction F = ^ * / appartient a L2loc et verifie | f = /

au sens des distributions, et F(z) 6 O{\) a l'infini. C'est la seule solution de
^ — f qui tende vers 0 a l'infini. On a alors ^ = £(/), la fonction £(/) est
holomorphe en dehors du support de / , et dans O(^) a 1'infini.
Remarque : II n'y a pas de fagon naturelle de choisir pour tout / G I? une
solution F de §f = / .
5 Dans le cadre L2
Demonstration du Theoreme 1 : L'equation de Beltrami s'ecrit, en posant
$(2) = z + <p{z) et / = dip/dz,
L'operateur JJL- : h i-> n • h de L2 dans lui meme est de norme k < 1 et £ est
une isometrie, done 1 - fi-C : L2 —> L2 est inversible et si on note /2 l'^lement
/z de L2, on obtient
/ = (1-M-£)"V ,
soit
/ = pi + fi-C(fl) + fi-C{n-C(p)) + ....
Ceci demontre l'existence et l'unicite de la solution / 6 L2. A partir de / , on
obtient <p par </? = ^ * / et $ par $(z) — z + <p(z). Alors <p est holomorphe
au voisinage de l'infini et tend vers 0 a l'infini.
Si (i,\ depend de fac.on C-analytique (resp. continue) d'un parametre A 6 A

(variete C-analytique) (resp. espace topologique), l'operateur iiy depend
analytiquement (resp continument) de A, done aussi (1 — fiyC)~l. H en est
de meme de f?\, done aussi de /A, <f\ et finalement 4>^. cqfd.
6 Un cas particulier
Une partie du Theoreme 3 est que $, ou ce qui revient au meme tp, est
C°°. Ceci pourrait se deduire du theoreme d'hypo-ellipticite des operateurs
differentiels elliptiques, dont un cas particulier est le suivant :
Theoreme. Si sur un ouvert U de C on a | | + u | ^ = v avec u et v dans
C°° et ||u||L=c < 1, dors cpeC00.
Mais il semble difficile de deduire le Th. 3 de ce theoreme. C'est pourquoi
nous choisissons de redemontrer le theoreme d'hypo-ellipticite dans la situa-
tion qui nous interesse, avec une demonstration qui donne pour le meme prix
le Th. 2, a partir duquel nous obtiendrons le Th. 3.
Pour le moment on va demontrer la proposition suivante.
Proposition 1.- Supposons fj, 6 C^p.
a) si ||/2||Li = I < 1, alors $„ est C°° ;
b) si ||/i||£,i = £ < 1/3, alors $M est un diffeomorphisme.
Demonstration : On a / = //•£(/) + fj.. Par consequent on peut ecrire
/ = A* (£
On sait que fi est a decroissance rapide et on va montrer que / est a decroi-
ssance rapide.
Lemme. Soit h : R + —> R + une fonction continue, tendant vers 0 mais qui
ne soit pas a decroissance rapide. Alors il existe un entier k € N et une suite
(xn) avec xn —>• oo, telle que x!^h(xn) —> oo, et
h(x)
sup —— ->• 1.
IJ. r n\xn)
Demonstration : Soit k un entier tel que h(x) & O(l/xk). Pour tout n e N,
est un ferme non vide de R ; notons xn un point ou la fonction h atteint son

maximum sur An. Pour tout x G R, on a alors
Kx) - ~ - ~^Kxn) si x & An, et h(x) < h(xn) si x e An.

Le theoreme d'mte'grabilite des structures presque complexes 313
Par consequent
hlx)
1 < sup - ^
|l-*nl<v
cqfd.
Reprenons la demonstration de la Partie a) de la Prop. 1. Supposons que
/ n'est pas a decroissance rapide. On va appliquer le lemme a
h{x) = sup |/(C)|.

ICI=*
On trouve done une suite (Q) telle que xn — |£ n | satisfasse aux conditions du
lemme avec h(xn) = |/(Cn)|- Ecrivons
/ = An * iyj + (A - An) * I J / 1 + A,
ou fin — A • XD(O,V^T)- Alors, ||AnlU' < &, de sorte que
An * ( £ / ) (Cn) < ^ SUp
Par ailleurs, pout tout k' € N on a
et
HF I
Done
avec £ < 1 et /(Cn) qui n'est pas dans O(l/xn), d'ou une contradiction.
Ceci montre que / est a decroissance rapide, et g aussi puisque ces deux
fonctions ont meme module. Par suite f et g sont C°°, done aussi ip et 4>.
Passons maintenant a la demonstration de 6). On suppose que ||A||n <
1/3, et de
on deduit ll/ILi < HAIU. • ll/IUi + HAIIii. Done

En posant g — d<j)/dz, et done g — £/£ • / , on a
d'ou
v < \\g\W < 2-
En tout point z € C, on a done
6t
dz - 2 = 2 dz
II en r&sulte que $ est un diffeomorphisme local. II est propre car $(z) =
2 + o(l). Done e'est un diffeomorphisme. cqfd.
7 Existence de solutions locales, cas C 'CO
Le Theoreme 2 est une consquence immediate de la proposition suivante :

Proposition 2. Soit fi 6 C°°(C) et ZQ € C. On suppose que \^{zo)\ < 1.
i4/ors i/ eitste un diffeomorphisme * : C -> C tei gue
9z ~ ^ ' dz
au voisinage de ZQ.
Demonstration : Quitte a faire une translation, on peut supposer que z0 = 0.
1) On va d'abord traiter le cas ou n(z0) = 0. On choisit une fonction
h 6 C^p telle que ft = 1 au voisinage de 0. On definit alors
hJz)=h[-
de sorte que
,f =e'|K) e.
On verifie alors que
dh£ aft aft aft

= £ et -^— •77-
,1 ar\ (77/
Posons n£ = ft£ • /x.

L e m m e . - On a ||/xc||x,i —• 0 quand e -> 0.
Le theoreme d'intSgrabilite des structures presque complexes 315
Demonstration : On peut supposer /i 6 C~ mp . Alors, /x € 5 et de /x(0) = 0,

on deduit
JJ
Sous-lemme.- Soit w 6 S telle que JJw(x,y) dxdy — 0. Alors on peut
trouver u et v dans S telles que
du dv
dx dy
Demonstration : Soit h € C ^ n p ( R ; R ) telle que fRh = 1. Posons
p(x) — / w(x,y)dy
et ecrivons w = w^ + W2 avec
w,(a;, y) = p(x)h(y) et w;2(x, y) = w(x, y) - p{x)h(y) .
On a
p(x) dx = 0,
/ €R
done
(Vj/ 6 R) / tuija;, y) da; = 0, et (Var € R) / w2{x, y)dy = 0.

JxeR JyeR
Posons 'X I* + OO
/
Wi(t,y)dt= - / wi(t,y)
et
•y= - 0 0 Jj
y+oo
/
W2(x, t)dt = — W2(x,t)dt.
— — oo */ y
Alors on a du/dx = wx, dv/dy = w2 et u et v sont dans 5. cqfd.
Reprenons la demonstration du lemme. Appliquant le sous-lemme a ft, on
peut ecrire
dv \ dhe dhe
d£ dr]
Par consequent
<
dh dh
<e + dr]
cqfd.
Nous reprenons maintenant la demonstration de la proposition 2.

Pour e assez petit, ||/}£||z,i < 1/3. D'apres la proposition l.b), il existe un
diffeomorphisme *J/ : C -» C tel que
O2 C72 O2
au voisinage de z0-
2) On s'interesse maintenant au cas ou /x(zo) — avec |/zo| < 1.
Soit A : C -> C telle que
En posant ^ = ty o yl, l'equation pour VP est une equation de Beltrami
avec ft e C°° et /i(0) = 0. cqfd.
8 Proprietes topologiques des solutions

locales
Le Theoreme 2 admet comme consequence :
Proposition 3.- Soit fj, 6 C°° et z0 tel que \n(zo)\ < 1, e< soit cj> e Hloc{U)
une solution de l'equation de Beltrami
sur un voisinage U de z0. Alors

a) <j> est C°° au voisinage de ZQ,
b) ou bien (f> est une constante au voisinage de ZQ, OU bien il existe un
v > 0 tel que <f> soit en ZQ un revetement ramifie de degre v,
c) si v — 1, 0 est un diffeomorphisme au voisinage de z0.
Demonstration : D'apres la proposition 2, on peut trouver un diffeomor-
phisme * : C —> C tel que
Le theoreme d'integrabilite des structures prcsque complexes 317
au voisinage de z0. Posons V = #(£/), z\ — *(zo) et• 4> = <f>o vf 1 , de sorte

que <fi = (poty. Alors 0 6 #/ oc , et 50/dz = 0 au voisinage de z,. Par suite,
<j> est holomorphe au voisinage de z\ d'apres le lemme de Weyl, done verifie
les proprietes a), b) et c) (au voisinage de Z\) et il en est de meme de tp (au
voisinage de z0). cqfd.
9 Demonstration du theoreme 3
Sous les hypotheses du theoreme 3, l'application <J> : C -> C est C°° et propre.
Son degre topologique est 1 car $(-z) — z + o(l). Si $ etait constante sur un
on vert, non vide U, on aurait U ^ C, done 6^ aurait un point frontiere z0 et
on aurait une contradiction avec la proposition 3.b. Done $ n'est constante
sur aucun ouvert non vide et en chaque point z £ C elle a un degre local
vz > 0.
Pour tout ID 6 C, le degre total au dessus de w est egal au degre topolo-
gique, i.e.
E * = !•
ze*-'(u>)
Par suite $"'(w) ne contient qu'un seul point et le degre local de $ en ce
point est 1. Autrement dit, $ est bijective et est un diffeomorphisme local.
C'est done un diffeomorphisme. Ceci demontre le Theoreme 3.
10 Inegalites de Grotzsch
Proposition 4.- Soient A\, A2 et A3 des anneaux complexes, <p\ : Ay —> A3
et y>2 : A2 —> A2 des plongements C°° non homotopiquement triviaux et
disjoints. On suppose que ip\ est K\-quasi-conforrne et que ip2 est K2-quasi-
conforrne. Alors
mod(.43) > —-mod(J4i) + — mod(,42) •

A A
Demonstration : Voir [Ah, Theorems 3 et 4]. cqfd.

Corollaire 1.- Soient Ri et /? 2 tels que R2 > Ri > 0, et soit <f> un plongement
C°° de DR2 dans C. On suppose que 4> est holomorphe sur £)#, avec 0(0) = 0
et 4>'(Q) = 1, et K-quasiconforme sur DR2 \ D^. Alors 4>{DR2) contient DR3
avec,
R3 = \R^KR\'XIK.
Demonstration : Posons U — 4>(DR2) et soit <f>u '• U —* DR4 un isomorphisms

holomorphe tel que <fiu{0) = 0 et ^ ( O ) = 1 ; posons $4 = <t>u°4>- P ° u r r petit,
soit h(r) = r ( l + o(l)) tel que (p(DT) D Ai( r ). Alors 04 induit un plongement
de DR2 \ Dr dans DRA \ A>(r), holomorphe sur DR1 \ DT et if-quasi-conforme
sur DR2 \ DRl. D'apres la proposition 4, on a
^ log ^ y = mod(D,u \ Dh{T)) > mod(DRl \ Dr) + ^mod(DR2 \ DR
1 Rt 1 i?2
= —log + T;log—.
27r r K R\
D'ou
log(/?4) + £ > (l - -^ J log fli + -^ log i?2,
avec un e arbitrairement petit, done qu'on peut supprimer. Autrement dit,
D'apres le theoreme du 1/4 de Koebe, U contient un disque de rayon R3 avec

R3 = /ti/4. cqfd.
Corollaire 2.- Plagons nous sous les hypotheses du theoreme 3 avec ||//||j,°o =
k < 1, et supp(/i) c DR, et posons K = (1 + A;)/(l - k). Alors
a) Pour Z\ et 22 dans C auec |z2 ~~ zi| < r> on a
*(z,)| < 4r 1 / K (2i? + r ) 1 " 1 ^ ,
6,) on a $(D R ) C D 2 H , et pour tout z eC, on a |$(z) - z| < 3i?.

Demonstration : a) : Pour l^l < R + r, on applique le corollaire 1 a C defini par
Pour |zi I > R + r, on remarque que pour t fixe avec |t| < r, la fonction z H->
$(x; + i) - $(2) est holomorphe sur C\DR+r, bornee par 4r]/K(2R + ry-l'K
sur le cercle de rayon R + r, et tend vers 0 a l'infini. L'inegalite o) decoule
alors du principe du maximum.
b) : La premiere assertion est une variante du Theoreme du | de Koebe,
en utilisant le fait que $ est univalente sur C - D R et est sous la forme
$(2) = 2 + 6i2 - 1 + • • • a 1'infini (voir [Po, page 8, foramle (8)]). II en resulte
que |$(2) — z\ < 3R sur D/j . La fonction 2 i-> $(2) — 2 est holomorphe sur
C — DR et tend vers 0 a l'infini; comme elle est majoree par 3R sur le cercle
de rayon R, elle Test aussi sur C — DR. cqfd.
11 Demonstration du theoreme 4
Soit (i e •i'Smp a support dans DR et verifiant ||/i||z/» = k < 1. Choisissons
une fonction 77 > 0 dans C ^ p telle que JJ77 = 1, et pour e > 0 posons
7/E(,z) = p-r7(J), de sorte que Jfr)£ — 1. Soit (en) une suite tendant vers 0, et
posons /i,, = r]en * /i. On a alors
• Mn G C comp ;
• supp(/i n ) € DR pour n assez grand;
• IKIU~ < * ;
• fj,n —> /j, d a n s L2.
Pour chaque n, en appliquant le theoreme 3, on trouve un diffeornorphisme

$ n de C sur lui-meme tel que
Les $ n forment une famille equi-continue d'apres le corollaire 2 de la proposi-

tion 4, et d'apres le theoreme d'Ascoli, on peut extraire un suite qui converge
uniformement sur tout compact vers une application continue $ : C —> C.
Les $ ^ ' satisfont aux memes conditions, avec le meme k, puisqu'ils sont
/C-quasi-conformes avec le meme K que les $ n , mais avec R' = 3R. Us for-
ment done egalement une famille equicontinue, et il en resulte que $ est un
homeomorphisme.
Pour demontrer le Theoreme 4, il suffit de montrer que 4> = $^. Posons
On a
\\9n\\l? =
Done quitte a extraire encore une sous-suite on peut supposer que (/„) et (gn)
convergent faiblement vers des fonctions f et g dans L 2. Posons ip = li
de sorte que <j>(z) = z + <f>(z). II reste a demontrer le lemme suivant.
Lemme.
a) f — dip/dz (au sens des distributions) ;
b) g = d(p/dz (au sens des distributions) ;
c) f = M • 9 + A«-
Demonstration :
a) : Soit h e C^nv On a
/ / / • * -
320 A. Dou&dy, d'apres des notes de X. Buff
b) : idem avec d/dz au lieu de d/dz.

c) : On a
m - Hn9n = Mi? - 9n) + (f* ~ A»n)ffn-
Dans cette somme, n(g — gn) tend vers 0 faiblement dans I? car pour tout
h e L2,
/ M<7 - gn)= I {hn){g - gn)
tend vers 0. De plus, on a \\(i - /in||x,2 -» 0 et ||ffn||x,2 est borne, done
||(/X-Mn)0n||j,i < l l M - ^ I I ^ I I S n l l ^ -> 0.
Par suite fingn tend vers fig et /„ = fingn + fin tend vers fig + fi au sens des
distributions. D'autre part, / „ tend vers / faiblement dans L2, done aussi au
sens des distributions. Par suite f = fi • g + fi. cqfd.
Ceci acheve la demonstration du Theoreme 4.
12 Un scrupule
Si fi depend continument (resp. analytiquement) d'un parametre A e A,
nous avons vu que <£>A est un homeomorphisme C —> C et que <p\ depend
continument (resp. analytiquement) de A pour la topologie L2. Mais la
dependance est-elle continue (resp. analytique) pour la topologie de la con-
vergence uniforme sur tout compact? La reponse est O U I dans les deux
cas:
a) : Cas continu. Pour tout Ao il existe un voisinage A' de Ao, un k < 1 et
un R tels que, pour A 6 A', fi\ soit de norme L°° bornee par & et a support
dans D/j (c'est du moins l'hypothese qu'il faut faire pour que ga marche...).
Alors les $A forment une famille equicontinue, done relativement compacte
pour la topologie de la convegence uniforme sur tout compact. Et sur une telle
famille, toutes les topologies raisonnables coincident, en tous cas la topologie
L2OC et la topologie de la convergence uniformle sur les compacts.
b) Cas analytique. On peut utiliser le lemme suivant :
Lemme,- Soient E\ et E% deux espaces de Frechet, L : E\ —> Ei une ap-
plication lineaire continue injective. Soient A une variete C-analytique et
F\ : A —>• Ei une application. Posons F-i = ioF\. On suppose Fi continue et
i<2 C-analytique. Alors Fj est C-analytique.
Demonstration : On peut supposer que A est un voisinage de 0 dans C"
contenant ( D i ) n . Pour k = (k\, ...,kn) posons
ck = /"••• f e-2i*<h't>F1(e2i*^+-+t»))dtl...dtn .
Le theoreme d'integrabiiite des structures presque complexes 321
Coinme F2 est analytique, on a F2(z) = XêN" ^{ck)zk pour \z\ < 1, et comme
i est injective, on a F\{z) = Ylkew ckzk- cqfd.
13 Solutions locales, cas L°°

Proposition 5.- Soit p G L^j, avec ||^L°°|| < 1, considerons $ = $ M et soit
U un ouvert de C.
a) Si f est une fonction holomorphe sur $>(U), on a :
En particulier, / o $ g H}0C{U), et est solution sur U de ['equation de Beltrami

definie par \i.
b) Reciproquement, si est de la forme f o ^ , ou f est holomorphe
sur
Demonstration : a) : C'est imrnediat si ft £ C^ m p , puisqu'alors $ est C°°.
Sinon, considerons les <J>n definis dans le paragraphe 11. Pour tout ouvert
V relativement compact dans U, on a $n(V) C $(U) pour n assez grand,
^ -> | f dans L2(V), / ' o $ „ - + / ' o $ uniformement sur V. Done
</••*•>•£-</•••>•£•
Or (/' = *„) •*? £
Demonstration analogue pour la deuxieme formule. Les autres assertions
de (a) en decoulent immediatement.
b) : Ici encore le resultat est evident pour fi 6 C^mp, et nous allons ap-
procher $ par une suite (<!>„) comme au paragraphe 11, mais la demonstration
est plus delicate. II convient de considerer, plutot que des derivees partielles
| f et § les formes differentielles d<j> = | f dz et ~d<f> = %dz.
La norme L2 des formes differentielles de degre 1 est invariante par les ap-
plications conformes et quasi-invariante par les applications quasi-conformes:
Si <j>: \\\ -> W2 est un diffeomorphisme conforme (resp. if-quasi-conforme),
pour toute forme differentielle a de degre 1 sur W2, on a ||</)*a||jr,2 = ||a||i2
(resp. ±\\cl>'a\\L2 <\\a\\L, < K\\Pa\\L2).
Soit <f> € HfÛ) une solution sur U de l'equation de Beltrami definie par
fi, et notons / la fonction <f> o $~' sur $(U). Soit V un ouvert relativement
compact dans U. On veut montrer que / est holomorphe sur $(V). Pour n
assez grand, on a §{V) C $n{U), posons alors fn = <f>° ^ ' sur
Un calcul donne
soit, en tenant compte de | | = /x|^,
Lemme.- Soit (u,,) une suite dans L'^mp avec ||un||L°° < rn independant de
n et ||un||x,2 —> 0, et soit g £ L2. Alors \\ung\\i2 —t 0.
Demonstration : Soit e > 0, et soit gx e Cc^np telle que ||ff-ffi ||x,2 < ^ ; . Pour
tout n, on a ||un(s-#i)IU 2 < f, e t p o u r n assez grand, on a \\un\\L2 <
d'ou ||u n 5i|| L 2 < f et ||u n 5|| t 2 < e. cqfd.
Fin de la demonstration de la Prop. 5 : On a ||/i - /xn||/,2 -> 0, | ^ € L 2

independant de n, et H ^ j ^ F l l i 0 0 - f^F = T^fc- ^ e n r 6 s u l t e e n vertu du
lemme que ||$*9/ n ||L* ->• 0, done | |^/n.| |i.2(V) -*• 0 .
D'autre part /„ — <j> o $ ~ ' -> / uniformement sur V, done df = 0 sur
V au sens des distributions, et / est holomorphe d'apres le lemme de Weyl.
cqfd.
Corollaire.- Soient U un ouvert de C et fj, e L°°(U) avec ||/i||z,°° < I. II

existe alors une solution $ € if/O(,(C/) de Vequation de Beltrami definie par /x
5«r £/ gui est un homeomorphisme de U sur un ouvert V de C.
5i satisfait a ces conditions, toute (p € HjîJJ) qui est solution de la
meme equation de Beltrami est de la forme F o 0, oil F est une fonction
holomorphe sur V.
Demonstration : Prolongeons fi par 0 sur C — U. Alors la restriction de $^
a U repond a la question. Une fonction $ solution de l'equation de Beltrami
qui est injective sur U est necessairement de la forme G o <£M ou G est un
isomorphisme de $p{U) sur un ouvert V. cqfd.
14 Demonstration du Theoreme 5
Soit /x 6 L°°(C) avec ||/X||L°° < k < I. Pour R > 1, definissons /xR par
fj,R = /j, sur D/e et /xR = 0 sur C - DR. Posons <J>R = 3 " ^ , et soit <l>Jj
Le theoreme d'intSgrabilite des structures presque complexes 323
l'homeomorphisme de C de la forme Ao<£>R, OU A : C —> C est une application

affine ajustee de fagon que $5? l&isse fixes 0 et 1.
Pour i?' > R, posons %1>R,K = §R, o (fc^)" 1 , de sorte que $R, = ipR,R< o$R.
D'apres la Prop. 5, tpw est holomorphe, et evidemment univalente, sur
Proposition 6.- a^ : L'application ipR,R' a w"e limite ipR quand R' oo

c'est une fonction holomorphe et univalente sur UR.
b) ^R^RI tend vers I'identite quand R et R! tendent vers I'infini. .
Lemme.- // existe une constante KP telle que, pour tout R > 2, pour toute
fonction <f> holomorphe et univalente sur DR, laissant fixes 0 et 1, et tout z
tel que \z\ < j , on ait
—h 1—12
\<l>(z)-z\<
R
Demonstration : D'apres un theoreme de Koebe, l'ensemble 5 des C

holomorphes univalentes telles que <^(0) = 0 et f'{0) = 1 est compact (pour
la convergence uniforme sur les compacts de D). Par suite, il existe une
constante K, telle que, pour toute tp € S et tout z tel que \z\ < | , on ait
\<p(z)-z\ <K\Z\2.
Pour R > 1, notons SR (resp. SR) l'ensemble des • C, holomor-
phes univalentes et telles que </?(0) = 0 et <p'(0) = 1 (resp. fixant 0 et 1). Pour
4> e SR et \z\ < f, on a \<j>(z) - z\ < ^ . En posant <ffl = ^ y , de sorte que
^ 6 5 1 , on obtient
1 K\Z\
z\ <
K\Z\
-1 < l + HBl)
R R +
R '
Or \<(>(1) — 1| < ^, et \-T7y: — 1| < ^ pour une certaine constante A/. On
obtient alors
avec /c11 = 2/c + M. cqfd.

Demonstration de la Prop. 6 : Demontrons d'abord l'assertion (6). On a
mod{UR - [0,1]) > ^ m o d ( D R - [0,1]) > ±- .

K is.
Par suite, par le Theoreme du \ de Koebe, UR D DR avec R = \R~*. La
fonction V;R,R' e s t holomorphe sur ce disque et on a lipRîz) — z\ < K ^ des
que R > 2\z\, d'ou (b).
Demontrons maintenant (a). Pour R < R! < R", on a V-*R,R" = IPR\R" °

il>ntR'- Fixons R ; si on prend une suite de valeurs de R' tendant vers oo, les
V;/?,R' forment une suite de Cauchy pour la norme de la convergence unifonne
de tout compact de UR. Cette suite a une limite ipR, qui ne depend pas de la
suite de valeurs de R' choisie. D'ou (o). cqfd.
Demonstration du Theoreme 5 : Pour R < R', on a ft = TI>R> o yiRRi sur

UR, d'ou ipR o <$>R = ipRi o $^, sur D/j. II existe done une fonction $ = $J, €
Hloc(C) induisant ipn o $ ^ sur D# pour tout R > 2. Cette fonction est un
homeomorphisme C —> C, solutioin de l'equation de Beltrami definie par //
laissant fixe 0 et 1.
Montrons l'unicite. Si $ 6 ///^(C) est une autre solution de l'equation
de Beltrami definie par /z, la fonction ^ est de la forme F o $ J ou F est
holornorphe sur C d'apres le Cor. de la Prop. 5. Si de plus ^ est injective et
laisse fixes 0 et 1, il en est de meme de F, done F est l'identite et \t = 4>^.
Enfin si [i depend d'un parametre A € A, pour tout sous-ensemble A' de
A tel que (3A; < 1)(VA € A')||/iA|U°° < k, ^XtR tend vers ${,x uniformement
sur A' x A pour tout compact A de C. Si fix depend continument (resp. C-
analytiquement) de A, Q>* R depend continument (resp. C-analytiquement)
de A pour tout R, et il en est de meme de & cqfd.
References
[Ah] L. V. Ahlfors, Lectures on Quasiconformal mappings, Wadsworth k
Brooks/Cole, 1987.
[Po] Ch. Pommerenke, Boundary Behaviour of Conformal Maps, Springer-
Verlag, 1992.
Adrien Douady, Departement de Mathematique, Universite de Paris-Sud,

Bat, 425, 91405 Orsay, France,
e-mail: Adrien.Douady@ens.fr
Xavier Buff, Universite Paul Sabatier, Laboratoire Emile Picard, UFR MIG,
118 route de Narbonne, 31062 Toulouse Cedex, France,
e-mail : buff@topo.math.u-psud.fr
BIFURCATION OF PARABOLIC FIXED POINTS
MlTSUHIRO SHISHIKURA
Abstract : We present here the foundation of parabolic implosion theory.

We show that, in the bifurcation of parabolicfixedpoints with one attracting
petal, Fatou coordinates exist and depend continuously on the parameter.
We study also properties of Ecalle cylinders, transition maps and return
maps.
§1. INTRODUCTION
This paper is devoted to the study of the bifurcation of parabolic peri-

odic points for holomorphic maps of one complex variable. Local behavior of
parabolic periodic points was already studied by Fatou ([Fa]) and Leau ([Le])
(see also Camacho, [Ca]). Fatou constructed a partial conjugacy between
parabolic dynamics and a translation, which we call a "Fatou coordinate"
(see §2.2).
However once we perturb maps which have parabolic periodic points,
very complicated but interesting bifurcation phenomena can occur via an
interaction between local and global dynamics.
For example, the Julia set can vary discontinuously (Douady, [Do]).
The aim of this paper is to present a tool to study such a bifurcation, both
qualitatively and quantitatively.
The study of the bifurcation of parabolic periodic points was begun by
Douady and Hubbard in their Orsay lecture note [DH]. They introduced
the "Ecalle cylinders" for parabolic periodic points and for their perturba-
tion, and applied it to the study of the external rays of the Mandelbrot
set. The theory of Ecalle cylinders was then developed by Lavaurs in his
thesis [La], and he applied this technique to prove many results-such as the
"limit shape" of the Mandelbrot set at c = | , the limits liminfn_+oo Jcn and
liinsup,^^ KCn for cn —> | , the non-local-connectivity of the cubic Man-
delbrot set (the connectedness locus for cubic polynomials), etc. Recently
the author proved that the boundary of the Mandelbrot set has Hausdorff
dimension two, and in the proof, the Ecalle cylinders were also an essential
tool (see [Sh]).
Typeset by
325
326 Mitsuhiro Shishikura
The difficulty in the study of the bifurcation of parabolic periodic points

comes from the fact that after a small perturbation, the parabolic periodic
point bifurcates into two or more periodic orbits and the behavior of the
orbits near the parabolic periodic point changes drastically according to the
parameter. The incoming and outgoing Ecalle cylinders are defined to be
the quotients of fundamental regions for the dynamics near the parabolic
periodic point. (See §2 for precise definition.) The fundamental regions
can be taken so that they depend continuously on the parameter, so the
correspondence between the original dynamical plane and the cylinders is
"continuous" in a suitable sense. The orbits near the parabolic periodic
point are traced by the transition map from the incoming cylinder to the
outgoing cylinder, and this is the part which represents the drastic change
according to the parameter. Through the incoming (resp. outgoing) cylin-
der, orbits come in to (resp. go out from) the neighborhood of the parabolic
periodic point, and the transition map connects the incoming orbits to the
outgoing orbits.
If we consider orbits which return many times to either of the fundamen-
tal regions, it is natural to study the first return map, i.e., the map which
assigns to a point in the fundamental region its first orbit point which comes
back to that region. By the quotient, there is an induced first return map
on the Ecalle cylinder. So one can study iteration of the first return map
instead of the original map.
Moreover it turns out that if the perturbed map has a new fixed point
which is parabolic or irrationally indifferent, then the corresponding return
map also has parabolic or irrationally indifferent fixed point at the end of
the cylinder. So we have a chance to iterate this procedure. In fact, iterating
this procedure means that we are tracing parameters such that the angle
of multiplier of the fixed points correspond to successive continued fraction
expansion of the limit. It is also experimentally known that such a limit
has quite "thick" Julia set and the Mandelbrot set also look "thick" at this
parameter see for example, [Ma], [Mi]). This observation can be justified
to some extent, for example, we can use this technique to prove that the
boundary of the Mandelbrot set has Hausdorff dimension two.
This note is organized as follows: In section 2, we study parabolic fixed
points locally and construct Fatou coordinates, Ecalle cylinders and the
transition map. Bifurcation of such fixed points are studied locally in section
3, and the return map is defined. The extension and global properties these
objects are discussed in section 4.
This paper is a revision of Chapters 2-4 of: The parabolic bifurcation
of rational maps, 19° Coloquio Brasileiro de Matematica, IMPA, 1993. The
author would like to thank the editor, Tan Lei, for her help and encourage-
ment in writing this paper, and R. Oudkerk, for various comments and for
producing the drawings from our sketches.
FIGURE 1. The Julia sets of z2 + c for c = \ and c = 0.25393 + 0.00046i
§2. LOCAL THEORY ON PARABOLIC FIXED POINTS
§2.1. Preliminary coordinate change and petals.

2.1.1. In this note, we are concerned with parabolic periodic points of
analytic maps. For simplicity, we only treat the case where the period is
one (i.e. fixed point) and the multiplier is one. Moreover by a coordinate
change we may assume that the fixed point is at z = 0. Therefore we are
interested in the class of maps satisfying
/o(0) = 0 and = 1.
We need a further assumption on the non-degeneracy of the fixed point,

that is
fo(O) * 0.
If this is the case, by another coordinate change, we may assume that
fg(O) = 1, hence
fo(z) = z + z 2 + ....
So define
/o is defined and analytic in a neighborhood of 0, 1

"{* /o(0) = 0, /i(0) = 1, and /£(0) = 1 )'
2.1.2. For an / 0 6 To, define /g by conjugating by w = I(z) = — :

z
Then /Q is defined in a neighborhood of w = oo, f0*(oo) — oo, and it is easy

to see that
fo(w) = w + l-\ 1- 0{—z) as w ->• oo,
w w
where a is a constant. So /Q is, as the first approximation, close to the
translation T(w) = w + 1. Since each orbit of T lies on a horizontal line,
and the images of horizontal lines by z = I~1(w) are circles tangent to the
real axis at z — 0, each orbit of /o also lies on a curve which "looks like" a
circle tangent to the real axis at z = 0.
FIGURE 2. Orbits of / 0 and /0*
2.1.3. Let, for a large b > 0, fi~ = {z e C| Re(2;) > 6 - |Im(z)|} and
f2+ = {z € C| Re(z) < -b + | Im(z)|}. It is easy to see from the form of /£
that if b is sufficiently large, then on fi~ U Q+, /Q is defined and injective,
and satisfies
\fS(z)-(z+l)\<-.
Then it follows that
/o (?F) c f i - U {oo} and /0*(fi+ U {oo}) D fi+.
If z Gfi~,then (/o) n (z) are defined for all n > 0 and converge to oo as
n -> oo. Conversely if there is a neighborhood of ZQ on which the iterates
(/o) n (z) are defined for all n > 0 and converge to oo as n -> oo, then
(/o)"(zo) e fi~ for some n.
2.1.4. Now define £2_ = J - ^ f i - ) , 0+ = /- 1 (f2 + ). We call ft_ an attract-

ing petal, and £2+ a repelling petal.
Remark. In this section, we use the subscript ± (fi±, $±, S±, etc.) to
denote objects in the original coordinate (in which /o is defined) and the
superscript ± (fi ± , $ ± , 5 ± , etc.) to denote corresponding objects in the
coordinate w = — 1/z (in which /£ is defined). The difference (for example,
between S± and 5^) seems to be subtle, however this allows us to avoid
more complicated notations and to save symbols. Anyway they represent
the "same" objects in different coordinates.
They satisfy:
ft_ and ft+ are simply connected domains bounded by Jordan curves
passing through 0;
ft_Uf2+U{0} is a neighborhood of 0, on which /o is defined and injective;
fo(TT) C ft- U {0} and / 0 (fi + U {0}) D Q^.
2.1.5. For an analytic function / which has a parabolic fixed point £ , the
parabolic basin of is
B { z has a neighborhood on which fn (n = 1,2,...) 1

are defined and fn -» ( uniformly as n -» oo J
Note that £ itself is not in the parabolic basin.
Lemma 2.1.6. For fo E To, the parabolic basin of z = 0 can be expressed
by
n>0
Proof. It immediately follows from the fact on £l~~ in 2.1.3. •
FIGURE 3. ft+, 5+
§2.2. Fatou coordinates.

As we have seen in 2.1, the conjugate /Q of /o € !FQ looks like the trans-
lation T(w) = w + 1. In fact, we see here that there are partial conjugacies
between / 0 and T.
Proposition 2.2.1. Let f0 e Fo and /£, SI* etc. be as in 2.1. There
exist conformal mappings (analytic diffeomorphisms) ~ : Q~ -> C and
+ : fi+ —> C satisfying
*-(/ 0 *(tu)) = *"(t«) + 1 and
$ ± (tu) = w — alogw; + c±,o + o(l)

as tu tend to oo within sectors {z\ Re(z) > b — fc|Im(z)|} ( c £J~) /or $~
ana1 {z| Re(z) < — 6 + k\ lm(z)\} ( c fi+) /or $ + , w/iere c±)0 are consfanfs
and k is any constant between 0 and 1. These $ ^ are unigue up fo additive
constant.
The proof will be given in §2.5 and §2.6.
2.2.2. Remark. As a convention, the branches of logarithm in the above
expression are chosen so that they coincide in the upper component of Q~ D
Q + and differ by 2?ri in the lower. Note here that if a is not zero, then $~
and $ + cannot coincide on the both components of Q~ D fi+, because of
the difference of the branches of the logarithm. In this case, f0 cannot be
conjugated to T in any neighborhood of 0.
Definition. The map $_ = $~ o / : Q_ -> C is called the incoming Fatou
coordinate of /o and $ + = $ + o / the outgoing Fatou coordinate of / 0 .
2.2.4. It can be easily seen that for any k with 0 < k < 1, there is a
large b' > 0 such that the image $~(fl~) (resp. $ + (fi + )) contains a set
{z\ Re{z) > 6'-Jfc|Im(z)|} (resp. {z\ Re(z) < -b'+ k\lm(z)\}) and on this
set there is a well-defined inverse branch of <fr~ (resp.
§2.3. The Ecalle cylinders.

2.3.1. Let ZQ and z$ be points satisfying Re^g" < —6 — 2 and b < ^
where b is the constant used to define il±. Define the vertical lines l~ =
{ZQ + iy\y 6 R} and £+ = {z£ + iy\y € R}. Then /${£') lies to the right
of l~, both contained in Q.~ and they bound a closed region S~ ( c fi~),
which looks like a vertical strip. Similarly fo{(+) and £+ bound a closed
region S+ (C fl+). It is easy to see that every orbit of /Q in fi~ (resp.

lands on S~ (resp. S+) exactly once, except when the orbit lands on one of
the boundary lines for S~ (resp. 5"), in which case it lands on the other
boundary line as well.
Let S- = I~1(S~) and S+ = I~1(S+). Their shape look like a croissant
with the tips at the fixed point z = 0. They are called fundamental regions
for the dynamics of fo\s±'
2.3.2. Definition. (Ecalle cylinders) Construct the quotient of 5_ by
identifying z € rl{l~) and fo(z) E / " H / o ^ " ) ) :
C_ = S_/~, where z ~ fo(z).
Obviously C_ = 5 _ / ^ is topologically a cylinder and can be naturally

identified with the quotient £2_/^, where z ~ fo{z) if both z and fo(z)
belong to fi_. Therefore C_ has a structure of a Riemann surface, since /o
is conformal. Similarly C+ = S+/ ~ = Q+/ ~ is defined. We call C_ the
incoming Ecalle cylinder and C+ the outgoing Ecalle cylinder.
Let 7T : C —> C* = C s {0} be the universal covering map n(z) = e 2irtJ! .
Lemma 2.3.3. The maps TC O $_ and TT o $ + induces isomorphisms
¥_ : C_ = S_/^ -> C* anrf ¥+ : C_ = S+/^ -> C*.
Proof. It follows from Lemma 2.5.4 in §2.5. •
§2.4. Transition map.

It follows from 2.2.4 that the map £/0 = $_ o ($+)~ x = $~ o ( $ + ) " 1
can be defined in a region {w\ \ Im(tu)| > b" + 2| Reiu|} for large b" > 0. By
its definition it satisfies
whenever the both sides are defined. Using the relation, one can extend £f0
to the region {w \ lmw\ > 770} for some 770 > 0. The relation implies that
£fo = n o £fo o TT~i : {0 < |iu| < e~2nri0} U {e27rr"> < \w\ < 00} -> C*
is well-defined, where n(w) = e2vtw. The asymptotic behavior of $* implies

that £f0 (w)—w tend to a limit as Imw —¥ 00 or — 00. (The limits are different
for hnw —>• 00 and —00.) Hence £f0 extends analytically to 0 and 00, so
that £fo{0) = 0, £/o(oo) = 00, £'fo(0) ± 0, and £^(00) ^ 0.
Definition. We call £f0 the transition map or Ecalle transformation. Since

the Fatou coordinates $>± are determined up to additive constants, the
transition map is defined up to pre- and post-composition of linear maps of
the form z i-> az.
Normalization. We normalize $~ and $ + by adding constants so that
<&~(w) - 4>+(tu) -> 0 when w e fi~ D fi+ and I m w - ^ o o .
Then automatically
$~(u>) - $ + (w) -> 27ria when w 6 fi~ n f i + and I m w - ^ -oo.
Under this normalization, it is easy to see that the transition map satisfies
Note that we can still add the same constant to $~ and +, so there is
one more degree of freedom to normalize.
Comparing with the definitions in 2.3, we can interpret the transition
map as a map from the outgoing Ecalle cylinder C+ to the incoming Ecalle
cylinder, denned only in a neighborhood of the ends of C+.
§2.5. The existence of the Fatou coordinates.

2.5.1. In this section, we prove the existence of the Fatou coordinates in a
more generalized setting. This form is necessary for the application to the
perturbed maps. To state the continuity of the construction, it is convenient
to define the following topology.
Definition. (Compact-open topology together with the domain of defini-
tion) In what follows, a map / is always associated with its domain of
definition T>{f) (an open set of C or C/Z etc.); that is, two maps are con-
sidered as distinct, if they have distinct domains, even if one is an extension
of the other. A neighborhood of an analytic map / : V{f) -> C is a set
containing
{.9 : V{g) -> C| g is analytic, V{g) D K and sup d(g(z), f(z)) < e},
z<EK
where K is a compact set in T>(f), e > 0 and d(-, •) is the spherical met-
ric. (If the map is to some other space, then d should be replaced by an
appropriate metric.) A sequence {/„} converges to / in this topology if
and only if for any compact set K C 1>(f) there exists an n 0 such that
K C V(fn) (n > n 0 ) and fn\n -> ftc uniformly on K as n -> oo. The sys-
tem of these neighborhoods defines "the compact-open topology together
with the domain of definition", which is unfortunately not Hausdorff, since
an extension of / is contained in any neighborhood of / .
For bi, b2 6 C with Rebt < Reb2, define
QihM) = {« € C| Re(2-&!) > - | I m ( z - 6 i ) | , R e ( z - 6 2 ) < | I m ( z - 6 2 ) | }.
If bx = —oo (resp. b2 = oo), the condition involving &i (resp. 62) should be
removed.
Proposition 2.5.2. Let F be o holomorphic function defined in Q with
Q - Q(bi,b2) and Reb2 > Re&i + 2 (here bx or b2 may be - 0 0 or ooj.
Suppose
\F(z) - {z + l)\ < ±, and \F'(z) - 1| < | for z 6 Q.
Then
(0) F is univalent on Q.
(1) Let ZQ £ Q be a point such that Re 61 < Rezo < Re 62 - 5/4. Denote by
S the closed region (a strip) bounded by the two curves £ — {ZQ + iy\y € R}
and F(£). Then for any z £ Q, there exists a unique n € Z such that Fn(z)
is defined and belongs to S — F(f).
(ii) There exists a univalent function $ : Q —> C satisfying
whenever both sides are defined. Moreover $ is unique up to addition of a

constant.
(iii) Fix a point z0 e Q. If we normalize $ by $(z 0 ) = 0, then the corre-
spondence F 1—> $ is continuous with respect to the compact-open topology.
Proof. (0) and (i) are easy and left to the reader.
(ii) Let z0 e Q be as in (i). Define hi : {z\ 0 < Re z < 1} -> Q by
hi{x + iy) = (1 - x)(zo + iy) + xF(z0 + iy), for 0 < x < 1, y € R.

Then we have
•£J = *"(*» + iy) - (*b + ty), -j1 = ixF'(z0 + iy) + i(l - 1).
Hence
1
{F(z0) - (z0 + iy + 1)} + x(F'(za + iy) - 1)| < -
4
Therefore \^tl^g^ < 1/3 and hi is a quasi-conformal mapping onto

the strip 5, and satisfies hi1(F(z)) = h^lz) + 1 for z € £. Let <T0 be
the standard conformal structure of C, and take the pull-back a = h\o~o
on {z\ 0 < Rez < 1}. Then extend a to C by a = (T n )V on {z| - n <
Rez < - n + 1}, where T(z) = z + 1 . By Ahlfors-Bers measurable mapping
theorem [A], there exists a unique quasi-conformal mapping h2 : C —> C
such that hô-Q — a a n d h2(0) — 0, h2{\) = 1. By the definition of a, T
preserves a. Hence h2 oT o h2l preserves the standard conformal structure
a0, i-e. conformal, therefore it must be an affine function. Since T has
no fixed point in C, so does h2 o T o h^1, hence it is a translation. Using
h2oToh2~l{G) = 1, we have ^ o T o / i " 1 ~T.
Now define $ by $ = h2 o h^1 on 5 , and extend to the whole Q using
the relation &(F(z)) = $(z) + 1. Then $ is well-defined by (i), continuous
and homeomorphic by the above relations on h^1 and h2. Moreover $ is
analytic outside the orbit of £, then analytic in the whole Q by Morera's
theorem. Thus we have obtained the desired univalent function $.
If $ ' is another such function, then $"(z) — <£' o $ - 1 commutes with T
at least in $(Q). Hence $"(z) — z extends to C as a periodic function, then
$ " extends to C as a holomorphic function commuting with T. Similarly
$ " - 1 also has this property, therefore <p must be an affine function. However
an affine function commuting with the translation T is also a translation by
a constant. Hence the assertion follows.
(iii) Let us consider F and Fo defined in the same Q and satisfying the
condition of the Proposition. As in (ii), we can construct hi, h2, $ for
F and hit0, h2i0, $o for Fo. It is easy to see that on any compact set of
{z|0 < Rez < 1}, ^-/f^- -> ^f-l^f' as F -)• Fo. Hence aF = h\aQ ->
°~Fo = Î.OÔ and h2 —>• h2$ on any compact set as F -> F o . It follows from
the definition of the extension of $ that $ —> $o as F —> Fo. •
2.5.3. A strip 5 as in Proposition 2.5.2 (ii), is called a fundamental region
for the map F\Q. The quotient space
C = S/~, where l 9 z ~ F(z) e F{£)
= Q/~, where z^F(z)\izeQC\F-1{Q)
is topologically a cylinder which is called the Ecalle cylinder. Moreover, C
has a natural structure of a Riemann surface, using F near £ as a coordinate
patching.
Lemma 2.5.4. Let F, Q, S be as above. Then TT o $ induces an isomor-
phism
Proof. It is easy to see from the construction that $ is a covering map.

Moreover it induces an isomorphism from the fundamental groups, since
the generator of the fundamental group of C corresponds to F and that of
C is T. D
2.5.6. It is easy to see that Proposition 2.5.2 and Lemma 2.5.4 imply Propo-
sition 2.2.1 (the existence part) and Lemma 2.3.3 respectively.
§2.6. An estimate for the Fatou coordinate.

2.6.1. Let F{z) be an analytic function such that v(z) — F(z) — z is close
to 1 (for example, as in Proposition 2.5.2). Note that if v = 1 then F(z) =
z + v(z) = z + 1 is the time one map of the constant flow ^f = 1- So it
is natural idea to compare the orbit of F with that of the flow ^| = v(z).
This flow can be solved by the integration
rz rfc
(0
If we regard £ in this formula as a complex function, the correspondence
z n t gives a new coordinate in which theflowis actually the constant flow.
So according to the above idea, the Fatou coordinate must have a similar
behavior as t(z). The following statement justifies this observation.
Proposition 2.6.2. Suppose that $ and v are holomorphic functions in a
region U satisfying:
$ is univalent in U, \v(z) — 1| < 1/4 for z E U and
$(z -I- v(z)) = $(z) + 1, if z,z + v(z) e U.
(i) There exist universal constants Ri,Ci,C2 > 0 such that ifU — {z \z —
zo\ < R} for R>Ri, then
(ii) Suppose U = {z 6 C*| 0X < argz < 92} (02 < #i + 2TT) and \v'(z)\ <
C/\z\1+v (z 6 U) for some C,u>0. For z0eU and 6[, 6'2 with 0x < 0\ <
9'2 < 92, there exists R2,C3 > 0 and ( e C such that
for z satisfying 9[ < argz < 9'2, dist(z,C \ U) > R\. Moreover C3 depend
only on 9i,9[.
See [Y] for a similar estimate.
Proof, (i) We may suppose z0 = 0. We take R so that R » 1. It follows
from Koebe's distortion theorem [P] that if \z\ < R — 2, then
viz >(z + v(z)) - $(z
since $ is univalent in {C| |C - A < 2 }. Hence C < |$'(z)| < C" if

|z| < .R - 2. We have |$"(z)| < C/.R if |z| < ii/2. (In fact, by Cauchy's
formula, $"(z) can be expressed in terms of an integral of $'(C)/(C ~ z)2
over the contour {Q\ |( - z\ = i?/3}, then use the above estimate.) Using
Taylor's formula
a) = at)dt,
we obtain |1 - &(z)v(z)\ < C/R if \z\ < R/2 - 5/4 (this implies \z\ < R/2
and \z + v(z)\ < R/2). Again by Cauchy's formula, |(1 - &{z)v{z))'\ =
\<!>"(z)v(z) + &(z)v'(z)\ < C/R2 if \z\ < R/A. It also follows from Cauchy's
formula that \v'(z)\ < C/R and \v"(z)\ < C/R2 for \z\ < R/2. Therefore
v'(z)\ < \v'(0)\ + C/R2 for \z\ < 5/4 Hence we have |$"(z)| < C(l/R2 +
v'(0)\) for \z\ < 5/4. By the above Taylor's formula again, we have |1 -
&(0)v(0)\ < C(l/R2 + \v'(0)\) < C/R. So we obtain the desired inequality.
(ii) Note that in the sector 9[ < argz < 9'2 we have dist(z,C \ U) >
for some constant d > 0. So the result in (i) applies to the region
{ w — z\ < Ci\z\}, and gives
1
v(z) l+u
Integrating this formula along a path in the smaller sector, we obtain the
inequality. •
2.6.3. Proof of the estimate in Proposition 2.2.1. Let v(z) — /Q (z) - z =
1 + a/z + O(l/z 2 ), $ = $* and U = {z e C*| \ < argz < ^ } . Since
\v'(z)\ < C/\z\2, we can apply Proposition 2.6.2 and obtain an estimate in
a smaller sector:
Since l/v(z) = 1 - a/z + O(l/z2), we have

<3>-(z) = z - alogz + const + O(l/z).
a
§3. LOCAL THEORY OF THE PARABOLIC BIFURCATION
§3.1. The phenomenon. Let / 0 € To be as in §2, i.e.
As we have seen in §2.1.2, the orbits of / 0 are contained in curves which

"look like" circles tangent to the real axis at z = 0. In this section, we study
how such orbits bifurcate after perturbation.
First the parabolic fixed point 2 = 0 has multiplicity two as a fixed point,
i.e., as the solution of the fixed point equation fo(z) — z — 0. Therefore
in general, it bifurcates into two fixed points after perturbation. Their
multipliers are close to one, and depending on the perturbation, they can
be attracting, repelling, or indifferent.
/„ s V
//"V
4&
Case 1 *— _ Case 2
\
FIGURE 4. Orbits for f0 and its perturbation
Two examples of perturbed orbits are depicted in the Figure. One no-
tices a big difference for orbits starting from a point on the left of the origin,
for example, on the negative real axis. For /o, such an orbit simply con-
verges to the origin; in Case 1, the orbit converges to an attracting fixed
point which bifurcated from the origin; however in Case 2, the orbit can go
through an "opening" between the two fixed points and leave the neighbor-
hood. Such a new type of orbits can give rise to a drastic change of the
global dynamics (such as the inflation of the Julia set). However it takes
an extremely long time for these orbits to go through the opening, so we
need to consider a large number of iterates of the map in order to see this
phenomenon. And the effect of the perturbation might accumulate during
the iteration, so it becomes difficult to control.
For the perturbed system as Case 1, there is not a drastic change in
the global dynamics, so we concentrate on Case 2. The distinction between
Case 1 and Case 2 can be given by the multiplier A of a bifurcated fixed
point. Roughly speaking, Case 1 corresponds to A which is closer to the
real axis than to the unit circle (or arg(A - 1) e [-f ,-f] U [^,^f]) and
Case 2 corresponds to A closer to the unit circle than to the real axis (or
arg(A - 1) € [7r/4,37r/4] U [5TT/4, 7TT/4]). (There are two fixed points and
two multipliers A, A' to consider. However by the theory of holomorphic
indices (see [Mi]) we know that
1 1
A-l A'-l
is continuous hence bounded, whereas A's are close to 1. So we can take
either one in the above discussion.)
We analyse such a bifurcation as in Case 2 using the Fatou coordinates
and the Ecalle cylinders. These objects for /o are already constructed in
§2, using "fundamental regions" S_ and S+. An important fact (which
we prove later) is that such fundamental regions continue to exist after a
perturbation corresponding to Case 2. The difference is that now the tips of
the croissant shape are at two distinct fixed points. So we can also construct
Ecalle cylinders and Fatou coordinates.
This setting allows us to trace the orbits which go through the open-
ing between the fixed points. For this purpose, it is essential to know how
an orbit coming in through the incoming fundamental region S-j goes out
through the outgoing fundamental region S+j. In fact, we will see that this
correspondence (incoming point in S-j) •-> (outgoing point in S+j) induces
a translation between the two Fatou coordinates and an isomorphism be-
tween the Ecalle cylinders. The crucial point is that this correspondence
depends sensitively on the perturbation (or the parameter), however the de-
pendence can be controlled in terms of A (or more precisely 1/logA where
log 1 = 0 ) . In fact, in a suitable coordinate with a normalization, the above
correspondence can be written a s z ^ z - 2ni/ log A. (See Proposition 3.2.2
(iii).)
In Case 2, points near the tips of the croissants S+j have orbits which
return many times to S±j. So the (first) return map can be defined on
S+j, and its iteration can be used to analyse the long term behavior of
those orbits. Such a return map induces a (well-defined, continuous and
analytic) map 7£/ on the Ecalle cylinder C+,/. Studying the dynamics of
11 f, we are able to obtain more information on the Julia sets.
In this section, we state these facts more precisely and give the proof.
FIGURE 5. Fundamental domains S±j
§3.2. Statements.
3.2.1. Let fo(z) = z + z2 H € To be as in §2. If / is an analytic mapping
close to /o, then it has twofixedpoints near 0 counted with the multiplicity.
By changing the coordinate by a translation near the identity, we can shift
one of them to the origin, so / reduces to the form
Note here that there are in general two choices- which fixed point to put
at the origin. In fact, for a parametrized family of maps containing /o, it
is in general impossible to reduce them to the above form by coordinate
changes depending continuously on the parameter. However this problem
can be avoided either by taking a branched double cover of the parameter
space, or restricting the family to the part Case 2 in the discussion in 3.1
(for example maps belonging to T\ defined below).
Anyway we may suppose that the map under consideration belongs to
_. _ J - / i s defined and analytic in a neighborhood of 0, 1
V /(o) = o. /'(o) / o /•
For / 6 J-, we express the derivative /'(0) as
/'(0) = exp(27ria(/))
where a(f) € C and — | < Rea(/) < | . Our study is focused on the maps
in the following class:
?i = {feT\ |arga(/)|<7r/4}.
Let <r(f) be the other fixed point of / near the origin (ff(f) = 0 if the
fixed point is degenerate). Since o~(f) is a solution of f(z) — z — 0, with
/'(0) = exp(27ria(/)) and /"(0) ~ /£(0) - 1, it is a solution of (e 27ria -
l)z + z(l + O{a) + O{z)) = 0. So(T(/) = (-27ria + O(a 2 ))(l + O(a) + O(z)),
and the O(z) term tends of 0 as a -> 0. Therefore a(f) = O(a) and
a(f) = -2jrta(/)(l + o(l)) as / -> /o.
Let us denote the objects defined for /o 6 .Fo m §2 with an additional

subscript "0", i.e., <3>±, $*, S±, etc are denoted by <3>±,o> $?> £±,0) etc.
Proposition 3.2.2. Le< /o € TQ. There exists a neighborhood Afo of /o fin
the sense of the topology defined in 2.5.1) such that if f 6 A/"o H T\, then
there exist closed Jordan domains S-j, S+j and analytic functions $ - , / ,
$ + / satisfying the following:
(i) S±j is bounded by an arc £± and its image f(£±) such that l± joins two
fixed points {0,<r(/)}, £±nf(e±) = {0,<r(/)} and S-JHS+J = {O,cr(/)};
(ii) $± / is defined, analytic and injective in a neighborhood of S'± * =
{ ( ) }
(Hi) if z 6 S'_ t then there is an n > 1 such that fn{z) 6 S'+ t and for the
smallest such n,
(iv) when f £ T\ tends to /o, S±j —> 5±,o U {0} in the Hausdorff metric,
and ^ ± , / —> ^±,o in the topology defined in 2.5.1.
The proof will be given later in §3.4.
It is often easier to work with (pf — (^+j)~1 and for applications, the
following is more powerful.
Proposition 3.2.3. Let ip0 be the inverse function ( $ + | 0 ) ~ 1 defined in
Qo = {w: |arg(-tu-6')l<2jr/3}
for large £0 > 0 (see 2.2.4). There exists a neighborhood A/o of f0 such
that if f e_A/o l~> Fi, then there exist large £o>»?o > 0, and analytic maps
<pf : Qf -> C and %j : {w 6 C : | lmw\ > r/0} -> C, where
Qf = {weC: |arg(-w-^ 0 )|<27r/3 and \ arg(w + ^ - - Co)I < 2TT/3 },

satisfying the following:

(i) ff(Qf) C V{f); Ifw,w + le Qf, then
<pf(w + l) - fotpf(w);
ipf(w) -» 0 when w e Qf and Imw -> +00; yf{w) -> <x(f) when w 6 Qf
and Imw —> —00;
f«J / / |w| > 770,
TZf(w) — it) ienrfs io —l/a(f) when Imw —> 00, £0 a constant when Imw —>
finj //u; € Q/ satisfies \w\ > TJQ and w' = TZf(w) + n € Qf for some n € Z,
fn{<ps{w)) = V / (w') i/n > 0, f-n(<pf{w')) = ^(ti;) i/n < 0;
fwj w/ien / € ^1 and / —> /o,
^5/ -> y)0 0"^ ^ / + - 7 7 \ ~* ^/o •
Moreover the convergence of itf + -^fy; is uniform with respect to the Eu-
clidean metric.
The proof will be given in §3.5.
3.2.4. The induced return map Ilf on the Ecalle cylinder.
By Proposition 3.2.3 (ii), Hf induces a well-defined mapping IZf on C*
via ir(w) = e2niw:
Kf = notifon-1:{0< \z\ < e~2nr">} U {e27r"° < \z\ <oo}->C*.
Moreover it extends to 0 and 00 analytically, and satisfies: 7£/(0) = 0,

TZf(oo) = 00, the fixed point 2 = 0 has multiplier
27ri
and 00 has also a non-zero multiplier.

Suppose that w,w' € Q/, \w\ > 770 and %f(n(w)) — TT(W'). Then by
definition, Hf(w) + n = w' for an integer n, and therefore fn{y>f(w)) =
<Pf{w') (n > 0), or f~n(<Pf{w')) = ipf{w) (n < 0). Note that
tif(w) -(w- —rjr) ->• €fo(w) - w,
hence this expression is bounded in {|Imw| > J/o} (taking j/o larger if nec-
essary). On the other hand, R e l / a ( / ) —> +00 when T\ 3 f -> /o- There-
fore for f E. Ti sufficiently close to /o, we have Re7i/(to) < Rera and
2TT/3 < arg(7J/(w) — w) < 4TT/3. SO in this case, the above integer n is
positive (or even large) if Re w and Re w' stay bounded. This is also the
case if
I arg(u/ + — - - w)\ < 2?r/3.
If n is large, we can interpret that one iteration of 72./ corresponds to a large

number of forward iteration of / via ipf. So we call 72/ the induced return
map and this map is denned on (a subset of) C*, which is identified with
the "outgoing Ecalle cylinder" C+j = S+j/ ~ . Together with these facts,
Proposition 3.2.3 (iii) implies
(ill') If U, U' are domains contained in Q/ such that
TZf is defined on ir(U) and Kf{n(U)) C n{U') for some m > 1,
<Pf\u, T|[/' are injective, and
I arg(w/ + 2^/) -w)\< 2TT/3 for w e U, w' e U',
then there exists an n > m such that
fn = tpf o {-K\U')~1 oiVfowo (ipflu)'1 on <pf(U).
Prom Proposition 3.2.3 (iv), we have:

(iv') e^KKilf -» £fo when / 6 7V0 n Tx and / -s- / 0 .
It follows that if {/„} is a sequence in Mo n JF\ such that /„ -> f0 and
l/ r t (/n) — kn-* -0 (constant) as n -> 00, where kn are integers, then there
exists a limit
Here the condition on a can be rewritten in the form:
It is sometimes convenient to consider the induced map on C/Z instead

of C*. Define the quotient map ?ri : C —• C/Z by TTI(W) = w mod Z, and
the identifying map 7r2 : C/Z -> C* by 7r2(w mod Z) = exp(2iriw). Then
7r is decomposed as 7r2 O TTI. Note that TT2 sends the upper (resp. lower)
end of C/Z to z — 0 (resp. oo). Let Hf = Kiojif o ?rfx = TT^1 O 7?./ O ?r2.
The dynamics of 7£/ near the upper end of C/Z is close to the translation
by - I / a mod Z, and near the lower end it is close to a translation by a
constant.
3.2.5. To study the quantitative aspect, it is often more convenient to work
with <p*j = / o ipf. Let us define £>(»?) = {w 6 C| \w — ir?| < |»?|/4} and
£>'(JJ) = {w € C| |w - iT/l < |r?|/2}, for rj <E R
As a corollary of Proposition 2.2.1 and Proposition 3.2.3 (iv), we have
the following:
For large r) > 0, there exists a neighborhood Afi(n)(c M>) of /oi depending
on r), such that if / € -^Vi(^)nJ"i, then y>/ is defined and injective on D(±n),
the image ip*f(D(±r))) is contained in D'(±r]) and | < |((/>J)'| < 2 on D(±T)).
3.2.6. Remark. For / e f with 3?r/4 < a r g a ( / ) < 5TT/4, instead of being
in T\, we can obtain a similar result as Propositions 3.2.2 and 3.2.3 with
the following changes:
the lower end of C/Z corresponds to the fixed point 0;
in the formulae and definitions, ^4TT should be replaced by -^TTJ, except
for the multiplier in 3.2.4, Tl'{{oo) = exp(-27ri^y);
in Proposition 3.2.3 (i), y>f{w) —> a(f) (w € Q/ and Imio —• +oo),
<Pf(w) —t 0 (w € Qf a n d Imw —> —oo).
§3.3. A preliminary coordinate change- unwrapping coordinate.

We would like to prove the statements in §3.2 using Propositions 2.5.2
and 2.6.2, in the same frame work as the Fatou coordinates for the unper-
turbed system /o (i.e., Proposition 2.2.1 and Lemma 2.3.3). For this pur-
pose, we introduce a new coordinate w (unwrapping coordinate) in which
the map looks like a translation. For this purpose, we do not need the
assumption on the argument of a(f).
3.3.1. Recall that 0 and a = <r(/) are the only fixed points of / in a small
neighborhood of 0, if / € T is sufficiently close to /o. So
u(z) = „,(,) = M f ^
is analytic and non-zero in the neighborhood. In particular, for / = / 0 ,
fr(/0) = 0, Uf0(z) = (fo(z) — z)/z2 = 1 + O(z). It is easy to see, for example
by using an integral representation, that the correspondences / •-> <r(/) and
/ M- Uf(z) are continuous (where we suppose that Uf is defined is a fixed
344 Mitsuhiro Shishikura.
small neighborhood V of zero). Moreover we may suppose that cr(f) £ V

and 1/2 < \uf(z)\ < 2 on V.
Recall that
«r(/) = -27rta(/)(l+o(l)) a s / ^ / 0 .
3.3.2. Suppose that / € T is as in 3.3.1 and a(f) ^ 0 (i.e. a(f) ± 0).

Let us introduce a new coordinate w G C by
where a = a(f) and a = a(f). Also define the translation T/(w) by
Tf(w) =W jjr.
Here are the important properties of ry for / € T sufficiently close to

/o with a(f) / 0: _
(i) The map T / : C - » C \ {0,O"(/)}, W *-¥ z = Tf(w) is a universal covering,
whose covering transformation group is generated by the translation Tf\
Tf(w) —>• 0 as Imaw -> +00, and Tf(w) ->• er as Imam —> -00.
(ii) Let TO(W) = I~1(w) = -l/w. Then
sup dc(Tf(w),T0(w)) ->0 as / - > / 0 ,

tu: I Reau;|<3/4
where d^-(-, •) is the spherical metric on C Hence on any compact set, Tf(w)
converges to TQ(W) in the spherical metric when / —> /o.
(iii) For any e > 0, there exists Ro > 0 independent of / such that if
w 6 C \ | J T?DRo, where £)fio = {«/ | |u»'| < Ro},
then |r/(«))| < e.

Proof, (i) is obvious. For (ii), consider the function
1 1
27ri
l-e- C ~ 27riC"
It is easy to see that this function is analytic in {£ : | Re£| < k} for any 0 <
k < 1 (for example k = 3/4), and bounded on the boundary of this region
(including the limit Im£ -> ±oc). Hence by the maximum value principle,
it is bounded on the region as well. This implies that |T/ - 2^ferol < C|<xJ
Tf'DR
FIGURE 6. Ff(w) and Tf in the unwrapping coordinate
therefore the assertion holds, since 2-iria 1.

(iii) It follows from (ii) that if RQ is large, Tf(dDR0) is contained in a small
disk around 0. Since Tf is a covering map, the image of £>#0, Tf(Dn0) must
contain the exterior of the small disk. Then the image of C \
cannot go to the exterior. So the assertions are proved. •
3.3.3. Now we define the lift of / to the ^-coordinate. First note that if
/ G T sufficiently close to / 0 and Ro > 0 is large,
\<T(f)uf(z)/(l + zuf(z))\<C\<r\<l/2
for w € C \ UngzT^Dfi,, and z = rf(w) (by 3.3.1 and 3.3.2 (iii)).

Define Ft : V(Ff) -»• C, where V{F}) = C \ Ung
- « * ' - • • / ( - )
where we use the branch of logarithm with — n < Imlog(-) < n.

Ii f e T sufficiently close to / 0 , a(f) ^ 0 and Ro > 0 is large, then Ff
has the following properties:
(i) foTf = TfoFf and Tf o Ff = Ff o T/;
(ii) For it; G (
4'
(iii) Ff(w) = w + 1 + O(l/w2) a s l m a w - ^ +00;

(iv) Denote Fo = fo — TQ1 o f0 o T0, then
sup \Ff{w) - F0(w)\ ^ 0 when /
Proof, (i) Using e-2iri011u = 1 - a/z (z = T/(W)) and /(z) = z + z(z - o > ,
we have:
exp(-27rtaF/M) =
g
- _
TV O F,(ttf) = ; . °^Z) r. r = /(*) = / O Tf(w).

f n /v
' z + z{z - a)u - (z - a)(l + zu) JK ' '
The other relation is obvious.
(ii) and (iv) Since \cru(z)/(l+zu(z))\ = O(a), we can estimate the logarithm
by Taylor expansion, and obtain:
sup Ff{w) _L ™(») ) I = 0(a).
As i ^ is periodic with period ^, it is enough to give the estimates in {w :

I Re aw I < 1/2} \ DR,,. Prom the facts that <r(/) -+ 0, u/(z) -> ufo{z) =
ifo(z) — z)/z2 uniformly on V and ry —>• ro uniformly on {tu : |Reaiy| <
1/2}, as / -» /o, we conclude that when / -> /o
27ria(/)(l } ^|
TQ(W) T0(W)2 + T0(w)(f0(To(w)) - T0(w))

and the convergence is uniform on {w : |Reaw| < 1/2} \ DR0. SO (iv) is
proved. This also proves (ii), since FQ(W) — w + 1 + O(l/w) and one can
take a larger RQ if necessary.
(iii) Since e2"iaW = /'(0) = 1 - < T ( / ) U / ( 0 ) , it is easy to see that Ff(w)-(w+
1) —• 0 as Imcrai -> +oo. Then by the periodicity of Ff, the decay must be
exponential. (Use the Fourier expansion and estimate the coefficient.) •
§3.4. Proof of Proposition 3.2.2.

3.4.1. Let TV be a neighborhood of /o in T such that all the statements

in §3.3 hold for / € T with a(f) ^ 0. Prom now on, we assume that
/ 6 Nnfx, therefore a ( / ) ^ 0 and |arg(a(/))| < TT/4. For ft > 0, let
«7 = Q(ft, -ft + ^ y ) and fi+ = Q(ft - - l y , -ft) = T / ^ ) ,
where Q(-, •) is the region defined in 2.5.1. If ft is large enough then f2^ c
V{Ff) for all / 6 M D .Fi. Moreover we may assume that for the same
ft >0,
and Uj = Q(-oo,-ft),
where fij stand for fi* defined for f0 in §2.1.3.
Let us fix two points tu* 6 fi* such that Re w~ > ft and Re w+ < -ft —
5/4. Then we may assume w± e fi* for all / e M"D J i and Re £ > 2ft + 2,
taking JV smaller if necessary.
By our assumption, \Ff(w) - (w + 1)| < -| and |i^(w) - 1| < \ on
fit, hence Proposition 2.5.2 can be applied. By (i), there exist fundamental
regions 5^ (c Ctf) bounded by I* = {w± + iy\y € R} and F^). By
(ii), there exist univalent analytic maps $J : Qj —> C and $^ : fl^ -¥ C
satisfying
and $*(«;*) - fcjftu;*). By (iii), *J= -* *± as / -> / 0 .

Let S±j, £± be the images of Sf, ^ ± by T/, with two points 0 and <r(/)
added. Then it is easy to check the assertion (i) in Proposition 3.2.2. For
example, to prove SLj D S+j = {0,CT(/)}, it is enough to show that
sjn{J
and this will be the case if | Re ^TTJ | is large.
Since T/ is injective on a neighborhood of 5*, {rf\s±)~l extends to a
neighborhood of S'± f. Then define
It is easy to check (ii) and (iv) in Proposition 3.2.2. So we only need to

check (iii). Before doing this, we prove the following Key lemma.
Lemma 3.4.2.
$toT/(«))-$7(w)
7 + — r ^ O , when f -> f0.
' a{f)
Proof. The function $J" o Tf is defined on QJ = l y 1 ^ ) and satisfies

the same functional equation as 3>J. It follows from the uniqueness in
Proposition 2.5.2 (ii) that there exists a constant d\ = di(f) such that
$}oT/ = $- + di.
By 3.3.3 (iii), Ff(w) = w + 1 + O(l/w2) as Imaw -> oo, hence
Jw~ Ff(O-C
converges to a constant, when w tend to oo satisfying TT/4 < arg(u> — w0) <
3TT/4 . It follows from Proposition 2.6.2 (ii) that there exist constants c±(f)
such that
when w tend to oo within a sector of the form {w | 6[ < &rg(w - w0) <
Q'2} CQ*, where TT/4 < 0[ < 9'2 < 3n/4 and w0 € C.
Thus we can express the constant d\(f) as
dx(f)= Um(ôTf(w)-^J(w))
Now define £f(w) = $ 7 o ( $ t ) - 1 on $t(f27 n Q^"), which contains at

least the vertical strip {w | |Reio| < 1, |Imw| > %} for a large r)0 > 0.
This T/O c a n be chosen uniformly for / € Af ft ^ i . It satisfies the functional
equation £f(w + l) = £f(w) + l whenever the both sides are defined. By this
relation, £f can be extended to {w |Imw| > 770}. Prom the asymptotic
behavior of <£^, we have
£f(w) = w + c_(/) — c+(/) + o(l) when Imw -» 00.
Similarly, £f(w) - w tends to a constant as Imiu -> -00.

Let TJQ < 7}i < T}2- Consider the integral of £f(w) — w along the rectan-
gular contour with vertices irfi, 1 + ir\i, 1 + ir?2, ir]2- By Cauchy's theorem,
the total integral is zero, and the integrals on the vertical sides cancel by
the periodicity of £f (w) - w. Hence, we have
0= / {£f{w) - w)dw - / {£f{w) - w)dw,
where the integrals are along segments. Letting T]2 -+ oo, we obtain
= / (£f(w) w)dw.
Jim
By Proposition 2.5.2 (iii), we have $ 7 —• $ ^ and 4>^ -* 4>J as / -> / 0 ,

therefore if —• 5/ 0 uniformly on [tr/i, 1 + i7?i]. Then
c_ (/) - c+ (/) = / (£/(«>) - w)dti; -)• / (Sfo (w) - w)dw = 0,

Jitj! Jit)!
because of the normalization for £f0 (§2.4). So the assertion follows, since
$;or / -$7 + ^ y = c+(/)-c_(/). •
3.4.3. Normalization. Now we change the normalization for <3>7 and re-
place it by $7 + c+(/) - c_(/). According to Lemma 3.4.2, this replace-
ment depends continuously on / . So it does not affect other statements of
Proposition 3.2.2. After this normalization, we have:
<bjoTf{w) = *"(«;)- - - i - = Tf(w)ot>J(w) ;
£f(w) = w + o(l) as Imto —>• oo.

Let us prove 3.2.2 (iii). Note that T/(57) and Sf are both vertical
strips contained in fiy", and Tf(Sj) is on the left side of St. So it is easy
to see that for any point w e Tf(Sj), its forward orbit under Ff passes
through 5^" before it goes out of Q^. Let n be the smallest integer such
that Ff{w) e Sf. Then
n = $7(jyH^)) - ^ 7 7 + n.
(/j
This would imply (iii), since $-,/(.z) = ^^(rT" 1 ^)) and $ + i / (/ n (z)) =
4>j(F?(w)) by the definition of $ ± > / . D
350 Mitsuhiro Shishikura.
§3.5. Proof of Proposition 3.2.3.

Lemma 3.5.1. There exist a neighborhood MQ (C M) of /o and a large
£o > 0 (independent of f) such that i f / e J V o f l f i , then
Qf = {weC: | arg(-w - £o)| < 2TT/3 and | arg(w + — r - £0)l < 2TT/3 },
(defined in 3.2.3) is contained in 3>f (fit).

Froo/. Let us prove that <J>t maps the left boundary of fit to the left of Qf
and the right boundary of fit to the right of Qf.
By the proof of 3.3.3, we may assume that \Ff(w) - (w + 1)| < e < 1/4
on fi^ for small e > 0, taking JVo smaller and £o larger. Furthermore,
Proposition 2.6.2 (i) implies that if w € Q^ and dist(z,d£lf~) > Rlt then
w) - l/(Ff(w) - w)\ < C2/dist(z,d£lp.
Hence
< C2/dist{z, dStf) + e/(l - e).
So taking £ small and ^2 > £1 > 0 large, we can guarantee that
I arg($+)'(«,)| < it/12 on 0Q(6 - I/a, -fc)
Integrating ($^")'(io) along dQ(& - l/«, —^2)1 w e conclude that
I a r g ( * ; H - $ ; ( - & ) ) I <TT/3
on the right boundary of Q(& - I/a, -£2), and
2TT/3 < arg(*£(tu) - *^(fa - I/a)) < 4TT/3 .
Shrinking Afo, we may assume that $t(-^2) is uniformly bounded for

/ e J V o n f i . Therefore if we take £0 (for Qf) large enough, $^"(-£2) is on
the right of Qf for all / e A^nJ17!. Similarly $J (£2) is on the left of TJx(Qf)
for all / e A/"o DT\. Applying Tf and using the relation $ + o T / = T / o $ ~
(see the normalization in 3.4.3), we conclude that $^(£2-a) — T/o#7(£ 2 )
is on the left of Qf.
Together with the result on the angle, this implies that $ ^ maps the
left (resp. right) boundary of Q(£2 - I/a, - & ) to the left (resp. right) of
Qf. Therefore $^(Q(£2 - 1/", -&)) covers Qf, and so does $J(ttf). •
Thus the inverse map ( ^ t ) " 1 is defined on Qf. So define
<Pf = Tfo ($^ )~x on Qf.
It is easy to check that it satisfies Proposition 3.2.3 (i).

Let nf = Sf - £^y. Since E/ = $f~ o ( $ + ) - 1 and $^" o T/ = 7/ o $ J ,
there are other expressions of 7£/:
tif = Tfo$-o (fcj)"1 = *J o 7) o (*^) -1 .

As we have seen in the proof of Lemma 3.4.2, £f hence Ttf are defined in
{w : | Imui| > T]0} for some rj0 > 0. It is easy to see that
-R.f(z+l)=iif(z) + l;
itf(z) — z tends to —l/a(/) when Im^r —> ex), to a constant when Imz —¥
—oo.
To prove (iii), suppose that w, w' = TZf(w) + n G Qf for some n G Z. If
n>0,
= T / o C ^ ) - 1 ^ ; o Tf o (*+)"1(i«) + n)
Similarly we have / n((pf(w')) = <ff(w) for n < 0.

Finally (iv) follows from the continuity of $f etc.
Thus Proposition 3.2.3 is proved. •
§4. GLOBAL BIFURCATION
§4.1. Global dynamics and extended Fatou coordinates. When the

dynamics is defined globally, i.e., in the case of rational maps or entire func-
tions, the bifurcation of parabolic fixed points can have global effects, which
are more interesting and important. To study such a global phenomenon,
we need to know the connection between the global orbits and the orbits
near the fixed point. One way to treat this is to extend the Fatou coordi-
nates and the transition map to a larger region, or if possible to the maximal
domain of extension,
Such an extension can be determined from the functional equations

which the Fatou coordinates etc. satisfy, for example, $_ o /£ = $_ + 1 =
T o <£_. The only problem is to control the domain of definition in order
to obtain a single-valued function. For example, for the unperturbed map
fo(z)=z + z2-\ , it is easier to extend $_ than <£+, and in fact we will
extend <p0 = ($+)~ 1 instead of $+.
Another importance of the extension is that the maximal extension often
has good properties- such as being a branched covering, belonging to a
compact class of maps. See §5, for example. Then we can extract more
information on the Fatou coordinates and the transition map.
For a perturbed map f(z) = z + O(e2*lct)z + O{z2), it is more difficult to
choose the domain to extend. And we cannot decide the maximal extension.
In this case, if we try to extend them just by analytic continuation, we will
soon encounter multi-valuedness of the function. However it is still possible
to choose an appropriate domain so that the continuity of Fatou coordinates
etc with respect to the map / is guaranteed. Since the domain of definition
of the Fatou coordinate varies depending on the map / , we need to use the
"compact open topology together with the domain of definition" (defined
2.5.1), in order to express the continuity.
After extending the domains of definition, we can formulate a global
version of Proposition 3.2.3. We will use this form in the application.
§4.2. The extension for the unperturbed map / 0 6 TQ.

4.2.1. Let /o € T$, i.e. /o is defined and analytic in a neighborhood of 0
and satisfies
/o(0) = 0, /£(0) = 1 and fg(O) = 1.
We denote the domain of definition of /o by 2?(/o), which we need to take
into the consideration. We will also use other notations defined in §§2 and
3, such as $±,o, £f0, Q±,o, etc.
4.2.2. Extension of <p0.
The map <p0 — (^+,o)~1 can be extended analytically to a region X>(y?o)
which satisfies the following:
(i) V(ip0) contains Q o = {w 6 C| 7r/3 < arg(w + £0) < W 3 } and
{w\ \lmw\ > rjo} for large £ 0 ,% > 0.
(ii) If w, w + 1 6 V(tp0), then
ipo(w+ 1) = f0oip0(w).
Moreover in the above formula, one side is defined if the other side is defined.
More precisely ,w + le V(ip0) if and only if w e V(<p0) and ipo(w) € P(/o).
(iii) If /o is a rational map or an entire function, then T>(<p0) = C.
Proof, (i) If follows from the definition and the estimate in Proposition 2.2.1
that Qo is contained in $o,-(^+) on which (p0 was originally defined. Then
we take (ii) as the definition of T>(y>o), in other words, w E. T>((po) if and only
if there is an integer n > 0 such that w - n € Qo and fl ((fo(w - n)) 6 Z>(/o)
for j' = 0 , . . . , n — 1. In this case, we define ipo(w) = foifoiw ~ n))- This is
well-defined, i.e., the definition does not depend on the choice of n because
of the relation on po in Qo.
It follows also from the estimate in Proposition 2.2.1 that
{w\ | lmw\ > 7} + 2| Rew\} C $+,o(0_ n ft+) = * J ( f i " D fi+)
for large 77 > 0. Hence if |to| > 7/ + 1 then there is an integer m such that
u> + m € $+,o(^-)- Therefore by 2.1.3, fo(fo('w + m) is defined for all
n > 0. It follows from the definition that w € T>(<p0). Thus (i) is proved.
(iii) is a consequence of (ii), since we can iterate rational maps and entire
functions anyway. •
4.2.3. Extension of $_,o- Let B be the basin for the fixed point 0, i.e.,
z 6 B if and only if there is a neighborhood of z on which / " (n = 1,2,...)
are defined and converge to 0 uniformly.
Then $ ^ 0 can be extended to B, which contains fi_. If z e B, then
fQ(z) e B and
$ - , o o / 0 ( z ) = $-,o(2) + l.
Proof. If z € S, then /£(z) € ft_ for some n > 0, by Lemma 2.1.6. So let
$_ i0 (z) — $-,o(fo(z)) ~ n- Then it is easy to see that this is well-defined,
and satisfies the functional equation. •
For simplicity of notation, let us write $0 — $-,()•
4.2.4. Extension of £fo. Let B = (PQ1(B), then T(B) = B and B contains
{w I I Imw| > T70} for some rj0 > 0.
Now define £f0 : B -> C by
£fo = $ o ° Vo-
This definition coincides with the original definition in 2.4. It satisfies
£fo(w + 1) = 4 H + 1foru; € 5.
We always assume that f/0 is normalized, i.e.,
£fo(w) ~ w
+ °(1) as Irntu-^ 00.
The induced map £/0 = ?r o Sf0 on"1 : n(B) —> C* is well-defined.

Moreover it extends to 0 and 00 analytically by £/0 (0) = 0 and £/0 (00) = 00,
and £'fo(0) ? 0, £'fo{oz) ± 0. So V{£fo) = *{B) U {0,oo}.
4.2.5 Inverse orbits and <po.

There is a natural interpretation of (p0 in terms of inverse orbits.
Definition. For a mapping / , a sequence of points {ZJ}JLQ is called an
inverse orbit (of z0) for / , if Zj e T>(f) and f(zj) = Zj_i for j > 1.
Lemma 4.2.6. For w € T>(ipo), let Zj = (fO{w - j ) (j = 0 , 1 , . . . ) . Then
{2j}|i 0 is an inverse orbit for f0 converging to 0. This gives a one to one
correspondence between 2?(v?o) and the set of inverse orbits converging to 0,
except the orbit Zj — 0 (j — 0,1,...).
Moreover if Zj (j > 1) are not critical points, then ip'0(w) / 0.
Proof. If w € V{ip0), then w-j£ V{<p0) (j = 0 , 1 , . . . ) and ipo{w - j) ->• 0
by (4.1.2), so the first statement is obvious. Let {ZJ} be an inverse orbit
converging to 0 and suppose Zj / 0 for some j . For large j , say for j > j 0 ,
Zj belongs to fi_ U fi+ (see (4.1.1)). But f$ tends to 0 uniformly on fi_,
so there exists jo such that Zj € fi+ for j > jo- Let w = (<PO\Q0)~1(ZJ) + i
(j > jo)- 1^ i s easy to see that w does not depend on j > jo and corresponds
to the inverse sequence {ZJ}. If w and w' give the same inverse sequence,
take j > 0 such that u> — j , w' — j £ Qoi then we have in = w' by the
injectivity of ipo on Qo.
The last statement follows from the facts that <po{w) = fo ° Vo(w — n)
for w € X>(y>o) and that ^ 0 ^ 0 in Q o . D
4.2.7. Grand orbits and $ 0 - There is a similar interpretation for $ 0 :

$ 0 (z) = $0(-z') if and only if there is an n > 0 such that f£{z) = /^(z');
= $0(2') {mod Z) if and only if there are n,m > 0 such that
O
The equivalence class for the latter one is called a grand orbit.
§4.3. Extension of ipf and Hf for perturbed maps.
4.3.1. Extension of iff. Let / 6 j V 0 n J i a s i n §3. Then <pf : Qf ->• C
was denned. We extend it to a region T>(<pf) defined by:
eC ™ - " e g / f o r an integer n > 0 and
The extension of cpi is automatically determined by <pf(w) — fn(<pf(w — n))

for w — n 6 V(tpf) with n in the definition. Obviously Qf C T>{<ff) and if
l £ > ( ) , then <ff{w) 6 D(/) and
+ l) = foipf{w).
Continuity. When / e jVo n f0 and / -> /o, we have iff -> y>0
respect to the compact open topology together with the domain of definition.
Proof. Let K be any compact set contained in T>((p0). There exists an n > 0
such that T~n(K) = {w-n\w 6 K} C Qo. By definition of 2%o), /#(#) C
V(f0) for j - 0 , 1 , . . . , n. If / is sufficiently close to / 0 , then T~n(K) is
contained in Qf and /•*(#) C Z>(/) for j = 0 , 1 , . . . , n. This implies that
K C T>(ipf). Moreover fn((pf(w - n)) -> fo(<po(w - n)), since Proposition
3.2.3 already guarantees the continuity on Qf. V / | Q / -> Volco- ^
Remark. We extend y>/ only to the right of Q/. It is possible to extend to

the left to some extent, but there is a danger of the multivaluedness.
4.3.2. Extension of $ / = $J o TJ1. For / 0 , we had <b0 = % : B -> C

It is not easy to define its counterpart for / . Let
5} = {w G C : £0 < Rew <
where £o is given in Proposition 3.2.3. Then define a region
there exists an integer n such that

f fn[z) e Tf(S}) and 0 < n ^
= [z 6 V(f)
Note that 77 is injective on S*f. We define $ / : T>(®f) -> C by:
where n is the integer appeared in the definition of
Claim. <§>f is well defined and satisfies
+ 1 i / z , / ( z ) e !?(*/).
Moreover <ff -4 y>0, as f E AfoDTo and f -> /o-
Froo/. First note that if 0 < n < m < ^ ( 3 ^ ) } and fn{z),fm{z) €

J
Tf(S*j), then for all j with n < j < m, / (z) 6 T/(5p. This can be proved
by looking at ^ - " ( T / I S ; ) ^ / " ^ ) ) -
This implies the well-definedness and the functional equation. The con-
tinuity can be proved as in 4.3.1. •
4.3.3. The return map Hf. We redefine the return map Hf as follows.
First define:
tif(w) = Qf((pf(w)) - -^j- on Vx.
If w,w + 1 e £>i, then
nf(w +1) = nf(w) +1.
Finally define
V(nf) = {w + n\w G P i , |*/(v>/M) - t«| < R e ( ^ 7 7 ) ) and
and
Then we have:
Claim. 7t/ is well-defined and satisfies
7lf(w + l) = Hf(w) + l forw
/£ coincides with the definition in §5. Moreover TZQ + l/a(f) —* 5^0, as

/ 6 J V 0 n f 0 and f -* f0.
Proof. This follows from 4.3.1 and 4.3.2. •
§4.4. A global version of Proposition 3.2.3.

Now we can state a global version of Proposition 3.2.3 with extended
domain of definitions.
Proposition 4.4.1. Let f0 6 TQ. There exists a neighborhood Af0 of f0
such that if f G Afo H T\, then there exist large ôô > 0, and analytic
maps iff : V(ipf) ~> C and fif : V{Hf) -> C satisfying the following:
(i) T>((ff) contains the set
Qf = {weC: | arg(-w - £o)| < 2TT/3 and | arg(to + -^j- - ô)| < 2TT/3 },
and <Pf(V(<pf)) C P ( / ) ;
Ifw,w + l e T>(cpf), then
= fo<pf(w);
<Pf(w) —> 0 when w 6 Q/ an<f Imw —• +oo; ff(w) —> <r(/) w/ien it) € (2/
and Irnto —> — oo.
(ii) V(TZf) contains {w € C : | Imu>| > 770} and is invariant under T(w) =
w+ 1; Ifw eV{Hf), then
TZf(w) — w tends to —l/a(f) when Imio —» oo, to a constant when Imw —>
—oo.
(in) Ifw € V(V,f)nT>((pf) and w' = Hf(w) + n € V{<pf) for some n 6 Z,
fn((pf(w)) = <pf{w') ifn > 0, /~ n (<^/(w')) = V/(t") ifn < 0.
Moreover if | arg(to' + 2a77) ~ w )l ^ 27r/3, t/ien n > 0.

fivj w/ien f E Ti and / —> /o,
Proof. This follows from Proposition 3.2.3, 3.2.4, 4.3.1-4.3.3. Note that the
condition |$/(y>/(to)) —tu| < Re( 2a (/)) w a s added in the definition ofD(TZf)
in order to guarantee the last statement of (iii). •
§4.5. Branched covering and transition map.

4.5.1. In this section, we study the property of the maximal extension of
the transition map for analytic maps which are branched coverings on some
domain.
Definition. An analytic mapping f : U -> V, where U and V are domains
of C, is called a branched covering, if for any z e V, there is a neighborhood
W of z such that for any connected component D of f~l{W), f\o is proper
and f : D — f~1(z) -» W \ {z} is a covering in ordinary sense.
Definition. (Immediate parabolic basin) Let ( b e a parabolic fixed point
of an analytic function / . A connected open set B C V{f) is called an
immediate parabolic basin of Q for / , if B is a connected component of the
parabolic basin of C for / (see (4.1.2)) such that f{B) = B and f : B -> B
is proper (hence a branched covering of finite degree).
For a rational map, a parabolic periodic point always has an immediate
parabolic basin and it coincides with the ordinary definition. However a
parabolic fixed point for an analytic map in general may not have any
immediate parabolic basin.
Lemma 4.5.2. Suppose that f is defined and analytic in a neighborhood of

0 and / ' ( O ) = 1 , / M ( O ) = 0 ( 1 < k < q ) , / ( « + 1 ) ( O ) ^ 0 (q > 1 ) , i.e. q + 1
is the order of f(z) — z at 0. Then f has at most q immediate parabolic
basins of 0, each of which contains at least one critical value of f, except
for a parabolic Mobius transformation. Moreover if an immediate parabolic
basin contains only one critical point (or critical value), then it is simply
connected.
Proof. This is a well-known argument for rational maps. Let us see it briefly.
By a local analysis at 0 (see for example or [Mi]), it can be shown that
there exist q disjoint simply connected open sets V^\...,V^ ( c T>(f))
("attracting petals") such that 0 € dV®; f(V{i)) c V « U {0}; / is
injective on V ^ ; a point z € T>(f) belongs to the parabolic basin of 0 if
and only if fn(z) e UjVW for some n > 0; and the orbit spaces VW/~
(where z ~ f(z) if z, f(z) 6 V^), are isomorphic to C.
Let B be an immediate basin of 0, and 5" the parabolic basin. By the
above, B n UjV(i) f <j> and UjV(i) C Bl. Since B is a component of 5 s , B
contains one of VW 's. So there are at most q immediate basins.
Define Vn(i) (n = 0,1,...) inductively, as follows: set V0(i) = V&; if
/(V n w ) C Vn(<), V ^ is the component of f-l(V^) containing VnW, which
exists and / ( V ^ ) = Vn c V^^. It is easy to see that for each i, Un>0Vn(i)
W
is a component of BK NOW suppose B = Un>0Vn is an immediate parabolic

basin, hence by definition, / : B —>• B is a branched covering. If B contains
no critical value or no critical point, then / : V^^ -> Vn , hence / " :
Vn —> V^ are covering maps, therefore they induces a covering map B ->
VW/~- Moreover vji*' hence B are simply connected. Therefore B must
be isomorphic to C and / is a Mobius transformation.
Similarly, if B contains only one critical point or critical value, then one
can show inductively that Vn are simply connected, hence so is B. •
4.5.3. We suppose now that /o e T is a function satisfying /Q(0) = 1 and

/o'(0) = 1. So we have ip0 and £/0 as in §§2 and 4. Let us denote go = £/ 0 .
Suppose that /o has an immediate parabolic basin B of 0. The above
argument applies to V^ — fi_ defined in §2.1, hence B is unique and
contains fi_. Let 5 ' be the (whole) parabolic basin of 0, and define B =
H and & = <pg ^B"). By 4.2.4, {w\ \ lmw\ > T?0} C BK Since for any
there is n > 0 such that Tn(w) e B, we have
{w\ | Imw| > T?O} C B.
Denote by Bu (resp. by Bl) the component of B containing {w\ Imw >

(resp. {w\ Ixaw < —rj}). (The superscript "u" stands for upper, and
Bifurcation of parabolicfixedpoints 359
" f for lower.) In general, Bu and Bl may coincide. Obviously TBU = Bu,
TBl = Bl. Then define Bu = *(!?•), Bl = * ( #
Definition. Let .Fo* be the set of functions / e T such that /'(0) =
1, /"(0) = 1 and / has an immediate parabolic basin which contains only
one critical point of / . Then this basin is automatically simply connected
by Lemma 4.5.2.
Proposition 4.5.4. Let f0 6 ?o and B, Bu, Be as above. Then gQ : Bu U

Bl - > C is a branched covering of infinite degree, ramified only over one
point v € C*.
The sets Bu, Bl, Bu U {0}, Bl U {oo} are simply connected, and J5" D
l
B = <p.
Proof. Let us first show that $o : B -> C is a branched covering. In fact,
$o|n- is injective, hence $ 0 : /o""(^-) <~> .B -+ T~ n $o(ft-) is a branched
covering (n > 0). So it follows that $o on B is a branched covering.
Now let us show that <po • Bu U Bl -> B is a branched covering. Let z
be a point in B. Take simply connected neighborhoods U, U' of z such that
U C U' and U' contains at most One forward orbit of the critical point. Let
z' be a point in y?g 1{U) D (Bu U Bl). Let U'n be the component of / 0 ~ n (t/')
containing y>o{z' - n) (n = 0,1,...). Then for some m > Q, U'n (n > TO)
do not contain the critical point. Hence there exist inverse branches fn~rji :
(/£, -• f/^,+fc of/Q. The family { / Q ^ ^ } is normal, since it avoids at least
three values (0 and the orbit of the critical point). Moreover it converges
to 0 uniformly on compact sets, since it does so near ipo{z' — m). Hence
there exists an n > m such that Un — foin+ (Um) C fi+ = ô(Qo)- Let
V = Tn o (volaJ-HUn)- Then z1 6 V and ^ 0 |v = (/ o n |yJ ° (Volfio) ° T~n.
So ipa : V -^ U \s & branched covering with at most one critical point, since
U contains at most one critical orbit. This shows that each component of
y>o 1(U) is either unramified or ramified over a common point in U, hence
tpo : Bu U Be -> B is a branched covering.
Therefore £fo = $ 0 o <p0 : Bu U Bl -> C and 5 / o : Bu U B^ ->• C* are
branched coverings. Moreover it is easy to see from the above that £f0 is
ramified only over v = n o o(c), where c is the unique critical point.
Now let U be acomponent of /(^"1(/o(fi_)) contained in B, different from
Q.-. Then we have fo(U) D t/ = (j> (n > 1). Hence for any component V of
<PQ1{U), n\v is injective (by the functional equation for ipo)- On the other
hand, n o $ 0 : t/ -> C* is infinite to one, since n o $ 0 |[/ = n o $0|n_ ° /o|i/
and 7T o $o|n^ is infinite to one. Hence go is of infinite degree.
Let us show the simple connectivity of Bu (or Bl). If 7 is a closed curve
in Bu, 7' = T~n-y c Qo for some n > 0. Let W be a region bounded
by </>O(T')> not containing 0. Since both the immediate basin B and fl+ =
Vo(Qo) are simply connected and do not contain 0, we have W C 5 n f i + . It
follows that 7' is trivial in Bu and so is 7. Therefore Bu is simply connected.
Then Bu U {0} (or Be U {0}) is also simply connected, since the fun-
damental group of Bu is generated by a curve around 0, which is trivial in
£"U{0}.
Finally, let us show that Bu and Bl are different components of B.
Suppose Bu = Bl. Then Bu = Bl = C, since Bu and Be are invariant
under T, simply connected, and contain half planes. Therefore B contains
Q- and £1+ = </?o(Qo), and the union is a punctured neighborhood of 0. But
B is simply connected punctured neighborhood of 0, so B = C \ {0}. Hence
/o is a parabolic Mobius transformation, since it is analytic on C and has
no periodic point in B. Then /o has no critical point and this contradicts
the assumption. Thus we have Bu n Bl - <f>, hence Bu D Be = <f>. O
By the normalization in 2.4, we have g'0(0) = 1. So 0 is again a parabolic

fixed point of go-
Lemma 4.5.5. Let /o € To and g0 = £/ 0 . Then g'0'(0) ^ 0 and go has a
simply connected immediate parabolic basin which contains only one critical
point. In other words, go belongs to To after a linear scaling of the coordi-
nate. Moreover gô{v) (n = 0,1,...) are defined and go(v) —>• 0 (n —> 00),
where v is the unique critical value of go-
Proof. Obviously 30 ^ 0. So there exist attracting petals V^ (i — 1 , . . . , q)
for go as in the proof of Lemma 4.5.2, where q + 1 = ord(go(z) — z). Since
go : Bu U Bl —> C* is a branched covering, we can also construct Vn
and go : V^^ -> Vn is a branched covering with at most one critical
point. Then Vn are simply connected and degpolyio (n — 0,1,...) are
eventually constant. It follows that Bi = UjVJt* are simply connected and
#0 : 5 j —> Bi is a branched covering with at most one critical point. Hence
Bi are immediate parabolic basins. By Lemma 4.5.2, we have q = 1, i.e.,
<7o (0) # 0. The rest follows easily. •
APPENDIX : HOLOMORPHIC DEPENDENCE

OF THE FATOU COORDINATES (BY TAN LEI)
We show here that the Fatou coordinates depend analytically on the

parameter, with a proof given by A. Douady.
Proposition A.I. In the setting of Proposition 3.2.2 (resp. 2.5.2), assume
in addition that fe(z) (resp. Fs(z)) is a family of holomorphic maps for
s 6 A U {0} C Cfc, with A C Cfc on open connected set containing 0 in
the boundary, such that fs (resp. Fs(z)) is continuous for s € A U {0}
and analytic in s € A, then the Fatou coordinates $±,/, := $±, s (resp.

$ s := <f>Fs) depend analytically on s for s € A.
Proof. We will at first deal with the situation in Proposition 2.5.2. In its
proof we replace F by Fs and hi(w) by hi<a(w). The definition of /i1]S indi-
cates that it is continuous in (s,w), quasi-conformal in w and holomorphic
in s. Therefore the complex structure as = h\ s<To is holomorphic in s.
By Ahlfors-Bers theorem, h2tS(w), the solution of as (normalized as in the
proof of Proposition 2.5.2), is again continuous in (s,w), quasi-conformal in
w and holomorphic in s. Now define 4>(s, z) = $s(z) = h2<s ° h^l(z), as in
the proof of Proposition 2.5.2. It is continuous in (s, z) and holomorphic in
z and satisfies the other properties of Proposition 2.5.2. It remains to show
that $(s, z) is holomorphic in (s,z).
Fix p = (s, 2) a base point. Relative to it we define w = h~i(z') and three
holomorphic functions P(z) = <J>(s~, z), Q(s) = hitS(w) and H(s) — /i2,s(u>).
Note that (s, z) for z close to z and (a, Q(s)) for s close to ? are two local
complex varieties in the space of (s,z), both passing through p, and being
transversal there. Furthermore P(z) is the value of <3> along the leaf {s, z),
and, as $(s, Q(s)) = H(s), the function H(s) is the value of $ along the
leaf (s,Q(s)).
(The following argument applies to maps in a more general setting. It
says roughly that if a continuous map from (an open set of) Cfc+1 to C is
complex differentiable along two transversal continuous foliations of complex
varieties, then the map is in fact holomorphic. It mimics strongly the proof
for the case where the map is differentiable in each variable.)
Assume at first s varies in a open set of C (i.e. k = 1).
Formal differential. Define at first a formal differential <5<J>^ as follows:
it is the C-linear map from C2 to C (having the form (s, z) •-> as + bz) which
maps the vector (0,1) to P'(z), and the vector (l,Q'(s)) to H'(s). Indeed,
this defines the linear map (i.e. the values of a and b) uniquely, as the
vectors (0,1) and (l,Q'(s)) are transversal.
Continuity. Now we let the base point p vary. By continuity of hii8(w)
and /i2,s(w) we conclude that w is continuous in p, and the three functions
P(z), Q(s) and H(s) all depend continuously in p.
By Cauchy formula for holomorphic functions, the derivatives of these
three holomorphic functions also depend continuously on p. Therefore
P'{z)i Q'(s) ajid H'(s) are all continuous in p. It follows that the for-
mal differential 6<frp (i.e. the values a and b) depends also continuously on
P-
Integration along the leaves. Let p and p 0 be in C2 and e apart
(with e small). We may connect p to po by an arc i(t), t € [0,1] which goes
parallel to the z-axis for t e [0,1/2], and then follows the curve (s, hi<s(wi))
for some w\. We use dot-derivative to denote derivative relative to t. We
claim that
r
$(p) - $(po) = / <S$7(t) 7(0 dt .
Jo
Proof. Transversality implies that fQ \\j(t)\\dt < Ce where C < 00. Let
Pi =7(1/2).
Assume t 6 [0,1/2]. We have 7(0 = (so, z(t)) and <&(y(t)) = PPo(z(t)),
with dot-derivative equal to Ppo(z(t)) • z(t) — S^^^j(t). Here we use the
trivial fact that Pp?{z) = P^t){z). So $(pi) - *(p 0 ) = / Q 1/2 <S*7(t)7(t) dt.
Now let t vary in [1/2,1]. As 7(0 follows the curve (s, hifB{w{)) with wi
unchanged, we have the same functions Q(s) and H(s) along the curve. So
7(0 = (s(t),Q(s(t))), and $(7(0) = #(*(*)) w i t l x dot-derivative equal to
//'(s(O)-s(O = <5*7(t)7(0. here we use the fact that j(t) = (1, Q'(s(t)))-»(0»
and 5$7(t) maps (1, Q'(s(t)) to H'(s(t)). So $(p)-$(pi) = J x / 2 <5$7(t)-y(0 dt.
The formal differential is equal to the differential. For p, po as
above,
(p - Po)| =
<f Jo
Letting e ^ O w e see that $ is complex differentiate at p 0 with differ-
ential d$Po — 8$Po, and with d# Po depending continuously on po.
Assume now that s = (si, • • • , Sfc) varies in an open set of Cfc. The same
argument as above shows that $ is holomorphic in each variable Sj. As $
is continuous in (s,z), it is in fact holomorphic.
Finally we come to the setting of Proposition 3.2.2. For fs a family
of maps the approximate Fatou coordinates for fa will provide a holomor-
phic family Fs(z) satisfying the conditions of Propositions 2.5.2 and A.I.
Therefore the Fatou coordinates $±,s(z) are holomorphic for s € A.
REFERENCES
[A] Ahlfors, L., Lectures on quasiconformal mappings, Van Nostrand, 1966.

[Ca] Camacho, C , On the local structure of conformal mappings and holo-
morphic vector fields, Asterisque 59-60 (1978) 83-94
[Do] Douady, A., Does a Julia set depend continuously on the polynomial?
Proceedings of Symposia in Applied Mathematics, Vol. 49, 1994.
[DH] Douady, A. and Hubbard, J. H., Etude dynamique des polynomes
complexes, Publ. Math. d'Orsay, ler partie, 84-02; 2me partie, 85-04.
[Fa] Fatou P., Sur les equations fonctionnelles, Bull. Soc. Math. France 47
(1919) 161-271 and 48 (1920) 33-94, 208-314.
[La] Lavaurs, P., Systemes dynamiques holomorphes: explosion de points
periodiques paraboliques, These de doctorat de l'Universite de Paris-Sud,
Orsay, France, 1989.
[Le] Leau, L., Etude sur les equations fonctionnelles a une ou plusieures
variables, Ann. Fac. Sci. Toulouse 11 (1897).
[Ma] Mandelbrot, B., On the dynamics of Iterated maps V: Conjecture that
the boundary of the M-set has a fractal dimension equal to 2, p. 235-238,
Chaos, Fractals and Dynamics, Eds. Fischer and Smith, Marcel Dekker,
1985.
[Mi] Milnor, J., Dynamics in one complex variables: Introductory lectures,
Preprint SUNY Stony Brook, Institute for Mathematical Sciences, 1990.
[P] Pommerenke, Ch., Univalent functions, Vandenhoeck & Ruprecht, Got-
tingen, 1975.
[Sh] Shishikura, M., The Hausdorff dimension of the boundary of the Man-
delbrot set and Julia sets, Annals of Math. 147(1998), 225-267.
Hiroshima University, Department of Mathematics, Kagamiyama, Higashi-

Hiroshima 739-8526, JAPAN
References
[Intro] Tan L., Introduction.
[Hubbard] J. Hubbard, Preface.
[McMullen] C. McMullen, The Mandelbrot set is universal.
[Douady-Buff-Devaney-Sentenac] A. Douady, with the participation of X.

Buff, R. Devaney and P. Sentenac, Baby Mandelbrot sets are born in
the cauliflowers.
[Hai'ssinsky] P. Haissinsky, Modulation dans I'ensemble de Mandelbrot.
[Milnor] J. Milnor, Local connectivity of Julia sets: Expository lectures.
[Roesch] P. Roesch, Holomorphic motions and puzzles (following M. Shishi-

kura)
[Tan] Tan L., Local properties of the Mandelbrot set at parabolic points.
[Petersen-Ryd] C. Petersen and G. Ryd, Convergence of rational rays in pa-

rameter spaces.
[Luzzatto] S. Luzzatto, Bounded recurrence of critical points and Jakobson's

Theorem.
[Petersen] C. Petersen, The Herman-Swiatek theorems with applications.
[Jelloulil] H. Jellouli, Perturbations d'une fonction linearisable.
[Jellouli2] H. Jellouli, Indice holomorphe et multiplicateur.
[Shishikura-Tan] M. Shishikura and Tan L., An alternative proof of Mane's

theorem on non-expanding Julia sets.
[Yin] Yin Y.-C, Geometry and dimension of Julia sets.
[Shishikural] M. Shishikura, On a theorem of Mary Rees for the matings of

polynomials.
[Douady] A. Douady, d'apres des notes de X. Buff, Le the'oreme d'integra-

bilite des structures presque complexes.
[Shishikura2] M. Shishikura, Bifurcation of parabolic fixed points.
365

(London Mathematical Society Lecture Note Series 274) Tan Lei - The Mandelbrot Set, Theme and Variations-Cambridge University Press (2000)

Uploaded by

Document Information

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

(London Mathematical Society Lecture Note Series 274) Tan Lei - The Mandelbrot Set, Theme and Variations-Cambridge University Press (2000)

Uploaded by

Copyright:

Available Formats

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES

Managing Editor: Professor N.J. Hitchin, Mathematical Institute,

46 p-adic Analysis: a short course on recent work, N. KOBLITZ

The Mandelbrot Set,

Cambridge University Press

Published in the United States of America by

© Cambridge University Press 2000

This publication is in copyright. Subject to statutory exception

First published 2000

Cambridge University Press has no responsibility for the persistence or

An alternative proof of Marie's theorem on non-expanding Julia sets 265

Complex dynamical systems is a fascinating field. It has both the visual

M = {c £ C : f™(z) remains bounded as n —> oo.}.

Nevertheless M exhibits a rich geometric and combinatorial structure, with

1 A detailed description of the contents of

The first results in this direction were obtained by Douady-Hubbard using

[Douady] gives a new proof of Ahlfors-Bers theorem on integrability of

2 Techniques in complex dynamics

in [Tan], Appendix D and in [Haissinsky].

Holomorphic dynamics is a subject with an ancient history: Fatou, Julia,

puters. So I tried to introduce some computing into the curriculum.

The year 1981-82: an ode to the cafe of Paris

• The connectivity of the Mandelbrot set. Mandelbrot had sent

the mainland of the Mandelbrot set M. Incidentally, these islands were in

they were better understood.

mapping would be conjugate to a polynomial. This time 1 called Douady

The Orsay notes

Department of Mathematics, Cornell University, Ithaca, NY 14853-7901,

2. The period of the attracting cycles of ft is locally unbounded; or

Figure 1. Mandelbrot sets of degrees 2, 3 and 4.

Corollary 1.2 Small Mandelbrot sets are dense in B(f).

There is also a statement in the dynamical plane:

Theorem 1.3 Let f be a holomorphic family of rational maps with bifurca-

Proof. The family contains a polynomial-like map / f : U -> V hybrid

For applications to Hausdorff dimension we recall:

Theorem 1.5 (Shishikura) For any d > 2, the Hausdorff dimension of

Corollary 1.7 / / / has bifurcations, then H. dim(J(/ t )) — 2 for a dense set

For higher-dimensional families one has (§5):

Similar results on Hausdorff dimension were obtained by Tan Lei, under a

Theorem 1.9 For any algebraic family of rational maps, either

2. The family is affine (every ft is critically finite and double covered by a

3. The family has bifurcations (B(f) ^ 0).

'This set of t can be improved to a dense G$ using Shishikura's idea of hyperbolic

Proof. Consider the algebraic family ga(z) = zd + azd~l, which for a ^ 0

2 Families of rational maps

1. Coordinate change: replace /( by mt o ft o m^1, where m : A x C - + C

2. Iteration: replace ft(z) by /"(z) for a fixed n > 1.

3. 5ase change: replace / t (z) by ./>(*) (z), where <>

Misiurewicz points. A marked critical point c of / is active if its forward

(/t "(ct) : n = 1,2,3,...)

fails to form a normal family of functions of t on any neighborhood of t = 0

Proposition 2.1 / / c is an active critical point, then (/, c) has a sequence

Proof. This is a traditional normal families argument. Choose any 3 distinct

Ramification. Next we discuss the existence of univalent inverse branches

Proposition 2.2 / / z has 5 distinct unramified preimages then it has in-

d\E\= Yl l + mult(/',z) < \F~l{E)\ + 2d-2< \E\ + 4d-4

and therefore \E\ < 4. •

Proof. By the previous Proposition, the ramified parameters are denned by

Misiurewicz bifurcations. Let (/, c) be a local bifurcation with a marked

Ml. /o(co) is a repelling fixed-point of / 0 ;

M2. Co is unramified for /o;

M3. ft(ct) is not a fixed-point of / ( , for some t; and

M4. deg(/ t , ct) = d for all t sufficiently small.

3 The Misiurewicz cascade