Annals of Mathematics

High Dimension Diffeomorphisms Displaying Infinitely Many Periodic Attractors

Author(s): J. Palis and M. Viana
Source: Annals of Mathematics, Second Series, Vol. 140, No. 1 (Jul., 1994), pp. 207-250
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/2118546 .
Accessed: 09/07/2013 10:14

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to Annals of
Mathematics.

http://www.jstor.org

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 Annals of Mathematics,140 (1994), 207-250

High dimension diffeomorphisms

displaying infinitelymany
periodic attractors
By J. PALIS and M. VIANA

Introduction
We extendto higherdimensionsa remarkabletwo-dimensional resultof
Newhouse provedin the seventies: many (residual subsets of an open set)
smoothdiffeomorphisms near one exhibitinga homoclinictangencyhave in-
finitelymany coexistingsinks. The saddle associated with the homoclinic
tangencyis takento be sectionallydissipative.In thisgeneralsettingwe have
to circumvent a major difficulty:theusual lack of(transversal)differentiability
of invariantdynamicfoliationsof codimensionhigherthan one. Also, in gen-
eral thereis no global smoothcentermanifoldand so the problemcannotbe
reducedin thiswayto the two-dimensional case. Insteadthe questionis solved
by producinghyperbolic sets whose foliationsare "essentially"differentiable
(a la Whitney)witha largefractaldimension(thickness).
The recentdevelopmentof (dissipative)dynamicshas been muchinflu-
enced by the discoveryof some strikingbifurcatingphenomenasuch as the
Lorenz-likeattractors,the Henon-likeattractors,the Feigenbaumand Coullet-
Tressercascades of period doublingbifurcationsand Newhouse's infinitely
manycoexistingsinks(attractingperiodicorbits).
This last phenomenon, althoughmorethantwentyyearsold, has remained
essentiallya topicin the studyofsurfacediffeomorphisms. That is, the known
examplesin higherdimensionswereobtainedby takinga two-dimensional pro-
totypeand "multiplying" it by a stronglycontractingdiffeomorphism. In the
presentpaper we are able to extend the originalresultto otherdimensions
(even to infinitedimension)in its fullstrength,showingin generalthe abun-
dance (fromthe topologicalpoint of view) of the diffeomorphisms displaying
infinitelymanycoexistingsinks. Indeed,we provethe following result:

MAIN THEOREM. Near any smoothdiffeomorphism exhibitinga homo-

clinictangencyassociatedto a sectionallydissipativesaddle, thereis a residual
subsetofan openset ofdiffeomorphisms such thateach ofits elementsdisplays
manycoexistingsinks.
infinitely

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
208 J. PALIS AND M. VIANA

We point out that our methodsalso implya one-parameterversionof

this theorem:A genericunfoldingof such a (quadratic) homoclinictangency
yieldsresidualsubsetsof intervalsin the parameterline whose corresponding
diffeomorphisms exhibitinfinitelymany sinks. For surfacediffeomorphisms
this versionhas been obtainedin [Rob].
"Smooth" in the statementabove means of class C2 and "near" means
closenessin the C2 sense. Recall that a homoclinictangencyis just a tangency
betweenthe stable and unstable manifoldsof a saddle periodicpoint. The
saddle is called (codimension-one)sectionallyor stronglydissipativeif it has
just one expandingeigenvalue(positiveLyapunovexponent)and the product
of any twoeigenvalueshas normless than one; i.e., any contracting eigenvalue
is strongerthantheexpandingone. We observethatin orderto get abundance
ofsinkswhenunfoldinga homoclinictangency,we assume that the associated
saddle is sectionallydissipativeforotherwisewe are bound to obtain periodic
orbitsofsmallerindex (saddles,sources);see [Rom]. On the otherhand,when
we genericallyunfoldsuch a quadratichomoclinictangencyin any dimension,
we obtain a ratherstrikinglist of dynamicalphenomena:
* Henon-likestrangeattractors[MV], [V], based on the remarkablework
of Benedicks-Carleson[BC];
* cascades of period doublingbifurcationsof sinks[YA];
* residual subsets of open sets of intervalsin the parameterline whose
elementsexhibitinfinitely many coexistingsinks [N2], [N3], [Rob], and the
presentpaper.
Thus, sincehomoclinictangenciesare a commonbifurcating dynamicfea-
ture, evidenceof the existenceof diffeomorphisms with infinitely
many co-
existing sinks is rather abundant. It is quite possible, however,that in a
parametrizedformin termsof Lebesgue measurein the parameterspace this
is a rare phenomenon(measurezero). This is a subject of muchinterestand
we referthe readerto [PTI fora discussionon thisand otherrelatedquestions.
As pointedout above, a basic difficulty in extendingNewhouse'soriginal
resultfromtwo to higherdimensionsis the usual lack of differentiability of
invariantfoliations(in our case, unstable foliations)of codimensiongreater
than one, whichare in generaljust H6ldercontinuous.This has been a major
obstaclein makingsenseof(transversal)fractaldimensionsofinvariantCantor
sets. In particular,thisis the case of thickness,a conceptused by Newhouseto
showthat manypairs of Cantorsets in theline necessarilyintersecteach other
unless one lies in a gap of the other;see Section 1. These intersections cor-
respondto homoclinictangenciesthat in turngeneratesinkswhen unfolded.
Also, the questiontreatedin thispaper has a semi-globalcharacter,involving
the wholeorbitof tangency.In particular,it cannotbe solvedby just assum-

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 209

ing that we can locally linearizethe map near the associated saddle, e.g., by
supposingthat it has nonresonanteigenvalues.Even more,we cannot apply
the standardprocedureof "reducingdimensions"throughprojectionalong the
strongcontractingdirectionsonto a centralmanifold,whichin factdoes not
exist in general.The keyidea introducedhereto circumvent these difficulties
is to obtain a hyperbolicset whose higher-codimension unstable foliationis
indeed "differentiable"(or "intrinsically
differentiable")ifwe avoid the strong
contractingdirections.This idea may be usefulin othersimilardynamicsit-
uations involvingfractaldimensionsof hyperbolicsets or, more directly,the
of theirinvariantfoliations.
differentiability
The formaldefinitions of intrinsicdifferentiability
and thicknessforthe
unstablefoliationare givenin Section1 and developedin Sections2 through4.
Specially,in Section2 we briefly describesomebasic propertiesofintrinsicdif-
ferentiability.A new relevantconditionon the homoclinictangencyis made
explicitin Section3: Assumption(3.2) in Proposition3.2 yieldsthe construc-
tion of hyperbolicsets whose unstable foliationis intrinsicallydifferentiable,
sincewe can then "avoid"thestrongcontracting directionsas suggestedabove.
In Section 1 we also presenta sketchof the whole proofof our main result.
It followsmorecloselya new proofof Newhouse'stwo-dimensional resultpre-
sentedin [PT] than the originalpapers. For the sake of simplifying the argu-
we in
ment, show Section 5 that we may assume that the saddle associated
to the homoclinictangencyhas a unique least contractingeigenvalue.In Sec-
tion 6 we constructthickinvariantCantor sets that appear when we unfold
the homoclinictangencyand whoseelementshave stable and unstablemani-
foldsthat transversally intersectthose of the associated saddle. This is done
througha renormalization technique.Finally,in Section7 we finishthe proof
of the theoremby just assemblingtogetherthe factsestablishedin the previ-
ous sections.We also brieflyindicatehow the parametrizedversionwe stated
above followsfromthese arguments.

1. Main ingredients and sketch of the proof

We begin by recallinga fewnotionsand facts,mostlyfrom[N1], [N3],
whichplay a centralrole in the argumentsbelow.
Let K C R be a Cantorset and K be its convexhull. A presentationof
K is an orderingU = (Un)n of its gaps, i.e., of the connectedcomponentsof
K\K. For each n and u E aUn let Tr(K,U,u) = length(C)/length(Un), where
C is the connectedcomponent of K \ (Ul U ... U Un) that containsu. The
thicknessof K is definedas
Tr(K) = sup infr(K, U, u)
U U

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
210 J. PALIS AND M. VIANA

wherethe supremumis takenoverall presentations ofK and the infimum

over
all points u E aUn as above. The local thicknessof K at x E K is

Tr(K, x) = lim(sup{T(L): L c K n [x -aF, x + a] a Cantor set}).

Recall that a basic set fora diffeomorphism is a compact,invariant,tran-

sitive, hyperbolicset, with a dense subset of periodic orbits,which is the
maximal invariantset in a neighbourhoodof it. A basic set is persistent:
Any Ck-smallperturbationof the map, k > 1, yieldsa unique basic set near
the initialone, called the "analytic"or "smooth"continuationof it. Let now
A be a (nontrivial)basic set of a C2 diffeomorphism p: M -- M, whose
stable foliationis of codimensionone, i.e., such that dimWS(x) = m - 1,
m = dimM, forall x E A. Let z E WS(A) and 0: [-a,a] -- M be a C1
embeddingtransverseto WS(A) at z = O(O). The local stablethicknessof A
at z is rS(A,z) = T(O-l(WS(A)), O). This is independentof the choice of @,
as a consequenceofthe factthat (underthe codimension-one assumption)the
holonomymaps (i.e., the projectionsalongthe leaves) ofthe stable foliationof
A can be extendedto C1 maps. Actually,this smoothnessof the holonomyof
WS(A), togetherwiththe transitivity of WI A, also impliesthat rS(A,z) has
the same value foreveryz E WS(A). We denoteby rS(A) this constantvalue
and call it the (local) stable thicknessof A. This is a strictlypositivefinite
numberand depends continuouslyon the diffeomorphism, in the sense that
if A, denotes the analyticcontinuationof A fora diffeomorphism b which
is C2-close to p, then rs(Afp) is close to rS(A). Local unstable thicknesses
rU(A,z) and rU(A) are defined in a similar way, when WU(A) has codimen-
sion one. In particular,both the stable thicknessand the unstablethickness
are well-defined if M is a surface.
Now we outlinethe proofof our main result. We start by recallingthe
main ideas in the proofofNewhouse'stheorem([Ni], [N2],[N3]),as presented
in [PT], and then describethe key ingredientsinvolvedin extendingthese
argumentsfromtwo to higherdimensions.This extensionwill be carriedout
in the forthcoming sections.
Let us considerfirstthe case of a homoclinictangency(on a surface)
involvinga thickhorseshoe.By this we mean that the homoclinictangency
is associated to a periodicpoint p belongingto some basic set A such that
Tu(A) rs(A) > 1. A crucial fact here is the gap lemma: if K1, K2 are Cantor
sets in the real line such that r(Ki) .r(K2) > 1 and K1 is not entirelycontained
in a gap of K2 nor vice-versa,then K1 n K2 7&0. By consideringthe line
e of tangenciesbetweenthe stable and the unstable foliationsof the basic
set A and applyingthe lemma to e n WU(A) and e n WS(A), one concludes
KV(whose closurecontains
that thereexistsa C2-openset of diffeomorphisms

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 211

the diffeomorphism p exhibitingthe tangency)such that every b E JKhas

tangenciesbetweenleaves of the stable and unstablefoliationsof the analytic
continuationA, of A. In particular,a dense subset of diffeomorphisms E
JKexhibitshomoclinictangenciesassociated to periodicpoints in AP (e.g.,
to the analytic continuationof p). We assume p to be dissipativeat the
saddle-pointinvolvedin the tangency,i.e., I detDfk(p) I < 1, k =period ofp,
and then the same holds for b E JK. Recall that, underthis dissipativeness
assumption,a small perturbationofa diffeomorphism exhibitinga homoclinic
tangencyyieldsa periodicsinkand that,moreover, a sinkis persistentthrough
small perturbationsof the map. Thus, foreach positiveintegern, the subset
of KVconsistingof elementswith n sinks is open and dense. Then, by a
standard Baire categoryargument,the elementsof a residual subset of JK
displayinfinitely manycoexistingsinks.
In orderto proveNewhouse'stheoremin thegeneral2-dimensional case we
just assume the existenceofa homoclinictangency,not necessarilyassociated
to a thickCantorset. Then we proceedas followsto constructan open set JK
as above. We continueto take the diffeomorphism p to be dissipativeat the
saddle-pointp. Also, we maysupposethatp belongsto a nontrivialhyperbolic
basic set A1; if this is not the case then we just perturb p so as to create
transversehomoclinicorbitstogetherwitha new orbitoftangencyassociated
to p. We then consider rU(Al) = rU(Al,p) > 0 and furtherperturbations of
the diffeomorphism are to be takensmallenoughso thatA1 and rU(Al) persist
essentiallyunchanged. The crucial step in the argumentconsistsin showing
that any small unfoldingof the tangencyproducesnew horseshoesA2 with
arbitrarilylarge stable thickness,in particularsatisfyingrS(A2) TUr(Al) > 1.
In [PT] such A2 are constructedvia a renormalization procedureas follows.
Let (W)E(_,6) =
po p, be an arc of diffeomorphismsgenericallyunfolding
the tangency.Then one findsparametervalues su -n 0 and small domains
QnC M convergingto thetangency(see Figure1), suchthat p (Qn) nQn#0
and, up to appropriaten-dependentresealingof Qn,the returnmaps -Wln I
Qn converge in the C2 topology to the endomorphism q(x, y) = (1 - 2x2,x)
as n -x oc (convergenceholds in the Ck topology,any k > 1, if the arc
is taken to be C?). From the fact that @ is conjugate to b(x,y) = (1 -
2lxi,x) one concludesthat it has invarianthyperbolicsets with arbitrarily
large thickness. Then we take A2 to be the analytic continuationfor Wn
Qn of some of these 4-hyperbolicsets. One checksthat leaves of Wu(Al)
and WS(Al) have some transverseintersections with, respectively, leaves of
to
WS(A2) and Wu(A2). Finally,a heteroclinictangencyassociated periodic
points P1 E A1, P2 E A2, may be created, again by a small perturbation
(thus with a negligibleeffecton A2 and rS(A2)). This is done throughan
auxiliarysaddle P that also originatesfromthe 1-dimensional endomorphism:

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
212 J. PALIS AND M. VIANA

P\ I

FIGURE 1

It exhibitsa homoclinictangencyand its stable and unstablemanifoldshave

points of transverseintersectionwith the dual invariantmanifoldsof points
both in A1 and A2. Thus, arguingas in the particularcase above, thereexists
an open set in Diff2(M) withpersistentheteroclinictangenciesinvolvingA1
and A2 and so also withpersistenthomoclinictangenciesassociated to A1 or
A2.
Now we discussin moredetail than in the introduction the main difficul-
ties in extendingthis resultto higherdimensions.We also presentin a more
formalway the ideas to overcomethem. As we said before,some of these
ofinvariantfoliationsor the definition
ideas (likethe intrinsicdifferentiability
and invarianceof local thickness)are quite generaland do not requirethe
assumptionsof codimension1 or of sectionaldissipativeness.
It is by now classic that the unfoldingof a homoclinictangencyof an
m-dimensionaldiffeomorphism leads to the creationof nontrivialbasic sets
([Sm]). It is easy to checkthat we can get these basic sets togetherwithnew
homoclinictangenciesoutside them. Therefore,we may again assume that
p is part of a nontrivialbasic set A1. However,it is not clear what Tru(Al)
should be taken to mean now, since transversesectionsto Wu(Al) are no
longerlines. Another,perhaps even more serious difficulty arises fromthe
factthat, Wu(Al) havingcodimensionbiggerthan 1, the projectionsalong its
leaves mayhave a bad metricbehaviour:In generaltheyare not Lipschitzbut
just H6ldercontinuous.
In orderto bypass these difficultieswe proceedas follows.The firststep
is to show that the unfoldingof a homoclinictangencyyieldsthe formation,

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 213

forarbitrarilysmall values A2of A, of otherhomoclinictangenciesassociated

to periodicpointsPAof Lo, such that:
(I) DW4(pA),? =period ofPA,has a unique weakestcontractingeigenvalue
which,as a consequence,is a real number.
Hence, we may assume rightfromthe startthat Dfk(p), k =period ofp,
satisfies(I). What we do then is to constructA1 embeddedin the manifold
M in an infinitesimally2-dimensionalway,transverselyto the strongestcon-
tractingdirectionsof Dfk(p). By this we mean that A1 is takensuch that at
everyx E Al, the intrinsictangentspace

ITx A1 = span v: thereis (xn)n E AN so that x?, -+ x and

- X _ 4
(for simplicitywe considerhere M = Rtm)is 2-dimensionaland, like the
plane generatedby the unstable and the weakeststable directions,is (uni-
formly)transverseto the codimension-two plane generatedby the strongest
contractingdirections. By abus de langage,we say that the angle (mean-
ing its trigonometric tangent) betweenthe above 2-planesis bounded from
above. For such a A1 we provethat the projectionmaps ir: So n Wu(Al) --
ZE n Wu(Al) along leaves of Wu(Al) are intrinsicallydifferentiable,Zo, E1
beingtransversalsectionsto leaves of Wu(Al). That is, thereexistsa contin-
uous map Zo n Wu(A1) 3 x | ) 07r(x) E ?(Rtm-l,JRml-) such that forevery
x E Zo n Wu(Al), we have
7r(x)- 7r(z)l- 07rllx) (X - z) - 0 as z - xwith z E
Lon Wu(Al),
lix zl-
whereZl is identified witha diskin Rjm-l. Here and in whatfollowsC(RiP,Rq)
denotesthe space of linearmaps fromIRPto Rq. Then we definethe local un-
stable thicknessof Al at x E Al by Tru(Al,x)= Tr(*(Wu(Ai)n Zo),*(x)),
whereSo is a transversalsectionto Wu(Al) at x and Tr:Wu(Al) n Zo , R
is any intrinsically map such that 0& I lTx(Wu(A1) n Zo) is a
differentiable
bijection.Intrinsicdifferentiability
of theunstablefoliationallows us to check
thatthis definitiondoes not dependon the choices of Z0 and Trand to prove
that Tru(Al,x) is strictlypositiveand independentof x E A1. Also it varies
continuouslywiththe diffeomorphism o E Diff2(M). We then show that by
arbitrarilysmall perturbations(essentiallynot affecting Al and Tru(Al)) one
obtains a hyperbolicbasic set A2 with codimension-one stable foliationand
large stable thickness,namely TU(Al)rS(A2) > 1. This is done by a natu-
ral extensionto higherdimensionsof the renormalization schemementioned
above. We also check that Al and A2 may be taken to be heteroclinically
related (existenceof mutual transverseintersections betweenleaves of their
stable and unstablefoliations)and, moreover, to exhibita tangencyq between

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
214 J. PALIS AND M. VIANA

W'(pl) and WS(p2), P1 E A1, P2 E A2 periodicpoints. We recall that the

stable foliationof A2, being of codimensionone, admits an extensionto a C1
foliationPFs(A2)definedin a neighbourhoodof A2 (whichwe may assume to
containq). An implicitfunctionargumentshowsthat thereexists an intrin-
sically differentiablemap 7rW: Wu (A1) n WI (pi) -? M such that 7ri(P1) = q
and each 7ri(x), x E Wu(A1) n Wls~(p1), is a pointoftangencybetweenleaves
of Wu(A1) and PFs(A2)(the imageof 7ri is the set oftangencies,see Figure2).

'(Al) ~ WsA2

WS~~~~~~~pi)
~~W(^

FIGURE 2

We let Wr2be the projectionalong the leaves of fFs(A2)onto WjU~c(p2),

which
we identifywithan intervalin DRi.
Then the theoremfollowsby an application
ofthe gap lemmato the Cantorsets Wr2? wil(W11(A1)flWlS~(pi)) and WS(A2)fl
W1jOW(P2) A

2. Intrinsically smooth maps

In thissectionwe describea notionof "intrinsic"differentiabilityoffunc-
tions on compactsubsetsof jim, closelyrelatedto Whitney[W], and we list
some of its basic properties.The main resultis Proposition2.10, on intrinsic
differentiability
of invariantsectionsof (contracting)bundlemorphisms.
Let X c jtm be a compactset and p: X -> In be continuous.We say
that p is intrinsically
C1 on X ifthereexistsa continuousmap Ap: X x X -

?(Itm, li) suchthat

f(x) - ~(z) = Ap(x, z) (x- z) forall x,z E X.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 215

Such a AW (whichis, in general,farfromunique) is called an intrinsicderiva-

C1+7'on X ifit admitssome-y-Hblder
tiveof W.We say that Wis intrinsically
continuousintrinsicderivative.
Example 2.1. Let (: U -) R', U an open rectanglein Rjm be a C1 map.
Then WI X is intrinsically C1 on X, foreverycompactX C U: One maytake
Ap(x, z) = (Aifj(x, z))i,j where,denotingy(i) (Xi, Xi Zi+ ,Zm) -

A/'i j(X, Z) = j ,
a~o ifxi-zi
ifXi=zi. (Y
If, moreover,Wis C2 then this AWis Lipschitzcontinuousand so Wis intrin-
sicallyC1+7'on X forevery0 < -y< 1.
Example 2.2. Let A: j2 ,R 2 be given by W(x,y) = (ax, py) with
1 < a < p and let X C JR2be definedas follows.Take K to be the standard
middle-thirdCantor set and let f: K K be the unique continuousmap
-

satisfyingf(0) = 0, f(1) = 1 and f(#3ir?1) = f(#ir) ? 1 forall p, q E N.

C1 functionon K with Af(k, k) = 0 forall k E K. We
This is an intrinsically
take X = graph(f) and then

is (xz) = f 0; x = (k,f(k)),z = (e,f(e)),

denoting
is an intrinsic forWon X. Observethat IJAW(x,x)ll= ar forall
derivative
x E X; comparewiththe previousconstruction.
This second example illustratesthe main point in the proofthat Al,
C1 unstable foliation(Proposi-
as constructedin Section 3, has intrinsically
tion3.4); one uses the geometryofthe domainto obtainan intrinsicderivative
with the smallestnormof the derivativeof the map "restrictedto factors".
The proofof the following C1 (or C1+'Y)maps is im-
propertiesof intrinsically
mediate.
LEMMA 2.3 (Chain Rule). Let Wp:X - ] and 'b: Y
R' -) RP be intrin-
sicallyC1 (resp. C1+'Y)with so(X) C Y. Then Vbo p is intrinsically
C1 (resp.
C1+7') and one may take A(b o Wp)(x,z)= AV)(/(x),Wp(z)) Ap(x,z). .

LEMMA 2.4 (Uniformlimits). Let a sequence Wk: X -l IRi of intrinsi-

cally C1 maps convergeuniformly to Wp:X -+ R' and admitintrinsicderiva-
tives AWkconverging uniformly to '1: X x X -? ?(Rm',R). Then W is in-
trinsicallyC1 and one may take Aso= (D.
LEMMA 2.5 (Restrictions).Let Wp:Xl x X2 -+W RW and for xl E X1
and X2 E X2 definemaps (xli: X2 -+ R' and WX2: X1 -+ R' by Wxi(X2) =
9(Xl,X2) = WX2(X,).

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
216 J. PALIS AND M. VIANA

(a) If (p is intrinsically C1 then the same holds for every epX1 and pX2,

with AW.1(X2,Z2) * V2 = AW((X1,x2),(X1,z2)) * (0,V2) and ApX2(Xi,Zl) V =

AW( (X1 ,X2), (Z1 ,X2)) * (Vl ,?)

(b) If all (PWxand W`2 admit intrinsic derivatives Anxj and AWX2 which,
moreover,vary continuouslywithx1 and X2 then (W is intrinsicallyC1 with
- (VlV2) = AOZ2 (X1,Zl) *V, + AWX (X2,Z2) *V2.
AW((X1iX2),(Z1,Z2))

Remark2.6. The following simpleremarkwill be used in the forthcom-

ing sections. Let (: X -- Rn be Lipschitzcontinuousand U C X x X be
such that {IIx - zII: (x, z) E U} is bounded away fromzero. Then thereis
a Lipschitz continuous map A: U -t l(RJ, Rn) such that W(x) - p(z) =
L(x, z)(x-z) forevery(x, z) E U.

is a local property.We
Let us also observethat intrinsicdifferentiability
say that (: X -? ]Rnis locallyintrinsically
C1 (resp.C1+^) ifeveryx E X has
a neighbourhoodVxC X such that WI Vx is intrinsically C1 (resp. C1+').

LEMMA 2.7. W: X - C1 if and onlyif thereexistsa

in is intrinsically
continuous map &p: X , ?(RImn) such that

lim ((x) ( z& (x(x-Z) = 0 for everyx E X.

C1 if and onlyif it is locallyintrinsically

In particular,(Wis intrinsically C1.

Proof. The "onlyif" affirmativeis provedbyfixingan intrinsicderivative

AW of and takingOW(x) AW(x, x). For the proofof the
( = "if" part we take
OWas in the and
statement define0: X x X -, ?(R', IRn)by

O(x,z) (x-z) = 9(x)-9(z)-09(x) (x-z) and

O(x,z) v = 0 wheneverv (x-z)=0.
Then O(x,z) dependscontinuously on (x, z) E X x X and Ad(x, z) = &W(x)+
O(x,z) is an intrinsicderivativeforW. The second part of the lemma is an
immediateconsequenceof the firstone. El

Remark2.8. Noticethat,fora sequence (Wk)k converging uniformlyto

(p: X -- Rn, having&O0kconverging uniformly to some (D: X -t l(Ri, Rn)
to assure that Wis intrinsically
is not sufficient C1.

Remark 2.9. The same constructionpermitsus to checkthat Wis in-

trinsicallyC1+^ if and only if &Was above can be foundwhichis -y-H6lder
continuousand, moreover,satisfies

IIW(x) - p(z) - 09(x)(x - Z)II ? C lix - ZII1+'

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 217

forsome C > 0 and everyx, z E X. In particular,by Whitney[W] (see also

C1+^ function,-y> 0, admits a C1+^ extensionto an
[St]), any intrinsically
open neighbourhoodof its domain. However,in the presentpaper we make
no use of such extension,except forthe simpleone-dimensionalsituationin
Lemma 4.4 wherean explicitconstruction is provided.
Proposition3.4 willbe provedby meansofthe following generalresulton
contractingbundlemorphisms, whichis a versionforintrinsicdifferentiability
of Theorem6.1 in [HP]. Let X0 C X be compact subsets of R' and B =
BR(0) be a compactball in R'. Let f: Xo -) X be a homeomorphism and
F: X0 x B - X x B be a continuous map of the formF(x, v) = (f(x), Fx(v)).
Assumethat,forsome c < 1 and everyx E Xo, Fx has Lipschitzconstant< c,
so that thereis a unique continuoussection &: X -+ B satisfying3(f(x)) =
Fx(&(x)) for all x E X0 (F-invariance).

PROPOSITION 2.10.
(a) Suppose that F and f1 are intrinsically
C1 (on theirdomains) and
thereis a < 1 such that

(2.1) -*|Af-1(x,z)
IIAFy(v,w)IH < a,

for any x,z E X, y E Xo and v,w E B. Then & is intrinsically

C1 on X.
(b) Suppose that F and f-1 are intrinsically
C1+7'(on theirdomains) and
thereis b < 1 such that
(2.2) IIAFy(vw)II jAf-l(xz) 1l+7 < b .

for any x,z E X, y E Xo and v,w E B. Then a is intrinsically

C1+ on X.
Proof. Let S (resp.S) be thespace ofcontinuousmaps a: X -+ B (resp.
Z: X x X -) I(Rm, Rn)), endowed with the sup-norm. Define F: S x S
S x S by F(a, Z) = (a', Z') with
(2.3) a'(x) = Fy(a(T))
(2.4) E'(x ,z) = AF(Tx, z) A f 1(x, z)
+ AFX(a (T)C(Z)) a (X,Z) O Af (x, z)
where,forsimplicity, we denote x = f-1 (x), = f-1 (z). By construction
(recall also Lemma 2.5), if E is an intrinsicderivativefora then Z' is an
intrinsicderivativefora'. Observenow that F is a fibercontraction:
(i) Given (C1, E1), (C2,Z2) E S x S and denoting (a'J,'Z) = F(ai, Ei),
1 <i <2, and f=f1(x), wehave

aIJ- or = sup IIF(1 (X)) - FX(C2(X))II CSUPI1 (X-) (x) 11

x x
-
< CII1 a211 (C < 1);

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
218 J. PALIS AND M. VIANA

(ii) Given a E S and El, E2 E S and denoting (a', Ei) = F(a, si), 1 < i < 2,
and x = (x) = f-1(z), we have

1Ii
1E - E'l| < sup(IIAFy(a(T),a(Q,))Ij (x,~)- E2(X ,Z)II-*|Af1(Xiz) Z)
x
< a JjEl -E211 (a < 1).

It follows([HP]) that F has a unique fixedpoint (c5,Z) and Fn(a, ) -+

(t,) as n -) +?0, forevery (a, Z) E S x S. Clearly a = 3&. Choose

now ao E S intrinsically C1 and Z0 an intrinsicderivativeof ao. Writing
Y'(ao, Eo)(an, En) we have that En is an intrinsic derivative of an for
=

everyn. Since (an)n -+ & and (En)n -+ Z, it followsfromLemma 2.4 that Z

is an intrinsicderivativefor&. This proves(a). Now we showthat underthe
assumptionsof (b) E is -y-H6ldercontinuous.Fix C > 0 largeenoughso that
AF,) x E Xo, AFV, v E B, and Af-1 are all (C, -y)-H6lder continuous.Let
X1, X2, z E X. Subtractingthe equalities

E(xiz) '=(xi, z)+AFj&(Ti), -'(xi, z)

(recall (2.4)) we obtain

IIE(Xl,Z) - E(x2,z)II < CIIxl -T2IV IIAf-1II +? AF*II . Cjxl - X2II'

+ CllCJ(Tl)- CJ(T2)jj^'jjtj*- JAf-1 1

+ ||AF*|jjjt(Tl,_z) - t(y2i,z~)J * IlAf 111

? IIAFII * 111 * C|X1 - X2II1',

where 11AF*11= supvxIzI|AFV(x,z)II and IIAF*II = supxIvIwjAFx(v,w)II.

we get
continuous
Notingthat& is IIEII-Lipschitz

IIE(xl, Z) - E(X2, Z)II ? Clijxl - X2II^'

? (IIAFII11IAf1II)JII(f lXl,f1Z) - (f fz)II
X2,f

withCi = (CIlAf-11)1+^'+ CIIAF*11

+ (CIIAf-1I IIIItjj)l++ CIIAF*I11I1iI.By
recurrenceand usingthe factthat

IIAF*IIIIAf1IIIfYl -f-1Y211 ?< IIAF*1IIlAf-111+'Ilyl- Y211'

< blyl - Y211I

we find,foreveryk > 1
k-1
IE(xl, Z) - (X2, Z)II < C(Z b)Ixil- X2II'
0
+ (IIAF*II f kZ) _tyf-kX2,if kZ)II.
fkXi,
ljAP11II)kIIZy

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 219

On the otherhand

IILF*II IIAf-II < max{IIAF*II, IIAF*II IIAf-1I1"1 < max{c,b} < 1

and so passingto the limitas k -- +?0 leads to

oo

IIE(x1, Z) - Z(X2, Z)II ? C1( b')llxl - x2II'.

0

In the same way one provesthat E is -y-HMlder

continuousin the secondvari-
able. El

Remark 2.11. It is an interestingquestion whetherthe conclusionof

the propositionstillholds underthe slightlymoregeneral(and morenatural)
assumptions

IIAFf-1(x)(vw)II IIAf1(x,x)II < a < 1, forall x E X,v,w E B

lIAf-1(x,x)ll1+< b < 1, forall x E X,v,w E B.
IIAFf-l(x)(vw)lI
We close withanothersimpleexamplewhichwill be of futureuse.

Example 2.12. It is a wellknownfactthat the map X: u - (I ? u)-1 is

definedand smoothon {u E 2(WRm,Tm)): IIuII < 1}. In what followswe let AX
be some (fixed)intrinsicderivativeforX on, say, {u: Ilull < 1/2}; observethat
we musthave IIAX(0,0)11= IIDX(O)II= 1. Let R > 0 be fixedand B = BR(O)
denotethe closed R-ball in (W(Rn,in). Let also H E ?(1R+f, Rm+f) be fixed
and writeH = (a d)" a E ?(RmRm), b E .C(RnJRm),c E (lRmRn),
d E /(RnRn). < 1/(2R), so thatthe graphtransform
Suppose I1bll1a-111
inducedby H
r: h ~-+ (c + dh) (a + bh)-1 = (ca-1 + dha-1) X(bha-1)
is well defined on B. For hl,h2 E B and g E L (Rnm,Jn), we let

AF(h1, h2) 9 = (dga-1) X(bhla-1)

+ (ca-1 + dh2a-1) AX(bhla-1, bh2a-1) (bga-1).
Then, clearly, AF is an intrinsicderivativefor F on B. Observe that
AF(h1,h2) is (uniformly) Ila-111is close to zero.
close to g -+ dga-1 if llbll

3. Intrinsically smooth foliations of basic sets

Let qo be a transversehomoclinicpointof some hyperbolicfixed(or pe-
riodic) point p of a C2 diffeomorphism p: M -+ M. We assume qo V Wss(p)
conditionto be stated in
and anothermild (open and dense) transversality

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
220 J. PALIS AND M. VIANA

(3.2) below. Then our goal is to prove that there exists a hyperbolic basic set
A1 containing p and qo and whose unstable foliationis intrinsicallyC1. For the
sake of simplicitywe restrictto the case when p is C2-linearizable on a neigh-
bourhood U of p. Apart from the corresponding nonresonance assumptions
no other conditions on the eigenvalues of Do(p) (or the dimensions of Wu(p),
Ws(p)) are required forthis construction. Moreover, a dual result holds when
q0 VtWuu(p).

Example 3.1. Let f: IA3 -I R3 be a C' diffeomorphismsuch that for

(x, y,z) close enough to the origin, f(x, y, z) = (ax, Ay,Oz) with a > 1 >
A > 0 > 0 and A > a6. Suppose moreover that p = (0, 0, 0) has a transverse
homoclinic point of the form qo = (0,0,6), 6 small. Let us parametrize a
segment -yof Wu(p) near qo by [-F, F] 3 x F-* (x, y(x), z(x)) and assume
that y'(0) $& 0. For n > 1 sufficientlylarge fn(-y) intersects {x = al}, a
small, in the point qn = (a, Ay(aa-'), Onz(aa-')). On the other hand
n
n= f (q0) = (0,0,on6) and so

IlIn- (al0,0)~ consyA() -

c
J|qn- (0,O0) y(ag
Therefore,the projectionfrom{x = 0} to {x = al} along WU(p) is not even
Lipschitzon {p} U {qn: n > 0}. This illustratesthe main obstructionfor
an invariantfoliationof a basic set (with codimensionbiggerthan 1) to be
smooth.

Startingour constructionof A1, let us denote by al,... , ac, Al, , As,

u + s = m, the eigenvaluesof Do(p), with Iua| > ...> ?aul > 1 > JA11 >
_. > IAe We define 1 < w < s by IAlI = -= IAw > IAw+1I > ...> A
and let Es = Ew e EBs be the invariantsplittingsuch that Do(p) I Ew has
eigenvaluesA1,-, Aw and Do(p) B has eigenvaluesAw+l,..., A8. We
E5s
choose C2 linearizing coordinates I( .u ,(,,1 ...) s on U and, clearly, we
may assume that

(Al) Wuoc(p) C {Ci ==

s= = 0} and Wls(p)
. C {61 = . = u = 0};
(A2) EW = {J0} x Rw x {08-W} and the strong stable manifold (tangent to
ESS at p) satisfies Wss(p) C {= = ... = (w = 0}.

Up to a convenientchoice of riemannianmetricwe have, for a = au ,

A= A11 = JAwl and 0==Aw+11,

(Bi) (a -6) J|vJJ

< 1|D(p(p)vJJ, forall v E Eu
(B2) (A - c) JlvJJ
< IJDo(p)vll< (A + c) JlvJJ, forall v E Ew
(B3) JJDo(p)vll< (6 + c) JvJJ, forall v E EsB

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 221

where6 > 0 is fixedsmall enoughso that 0 + 26 < A - 26 < A + 2, < r- 26.

(In the case w = s, i.e., if all contractingeigenvalueshave the same norm,
ESs = {O}, W5s(p) = {p} and we leave 0 undefined).Fix points q E W1s(p)
and r = -pN(q) E Wiu,(p) in the orbitof qo. Take
V = V6 = 11](6lv - u
'Wl < 6} X 1Il((l' (0s1l ' PI

where 6 > 0 is small and p > 0 is fixedin such a way that {p, q} C int(V) C
V C U. Let n = n(6) be minimumsuch that r E int(On(V)). (We suppose 6
convenientlyadjustedso that (9n+N(V) cuts V in twocylindersas in Figure3.)

n ?N(V)

/ .............~~~~~~~..................... ....................... ......../

Ws(P)

FIGURE 3

We let A = flkez ~(n+N)k(V) and then take simply Al = Un+N p?(A). The
next propositionexpressesthe crucialgeometricpropertyof A: the intrinsic
tangentspace to Wju,(A) at every x E Wju,(A) is containedin a (u + w)-
dimensionalsubspace whose angle to RU+wx {08W} bounded
is uniformly
fromabove. We denoteby (vu1,V5s)the componentsof a vectorv E R' with
respectto the splittingRm = Ru+w x R-.

PROPOSITION 3.2. There are C > 0 and -y> 0, independentof the n

chosenin the definitionofA, and thereexistssome (C,y)-Holder continuous
map A: WuJo(A)x WiuJ(A) -- I(Ru+wR8-w) such that

IA(x,z)H< C and (x - z)s, = A(x, z) (x - z)uw for all x, z

. E Wuc(A).
Proof. Let W = Wu(A) nv, V = f n+N and V1,V2 be the twoconnected
components(cylinders)of +(V) n v. For k > 0 let Uk C W x W be the set
of pairs (x, z) such that

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
222 J. PALIS AND M. VIANA

(i) +-iY(x), +-i(z) belongto the same Vjj, ji = 1,2, for0 < i < k;
(ii) V/-k(X),4r-k(z) belongto differentVj/'s.
Observethat W = Uk>oUk is densein W x W: in fact(x,z) E W x W \ W
if and only if x and z belong to the same local unstable leaf. We construct
A on each Uk, by recurrenceon k, and then it extendsfromW to W x W
by uniformity. The definitionof A I Uo is ratherarbitrary.Clearly,thereis
K > 1 (independentof n) such that (see Figure4)

I(x -z)ss < K and |(x-z)u11H > K-1 forall x E Vi, z E Vj, i j.

EU

.... . .. . . . . . . . . . . .. . . . . . . . . . . . . . .. ..
... . ... ..

______ ____--
~~~~~V2
V1
Ew

ESs

FIGURE 4

large (dependingonlyon K)
by Remark2.6, forCo > 0 sufficiently
Therefore,
there are Co-Lipschitz continuous maps AO: Uo -) L(lRu+w, Rsw-) with norm
< Co and satisfying

(x-z)ss = Ao(x,z). (x-z)uw forall (x, z) E Uo.

We take A I Uo to be any such map. Let now (x, z) E Uk, k > 1, and denote
x ='F1(x), z = V/-1
(z). Observethat (x, z) E Uki1. We write

1Pn = Tno? L E (Ru+u,JRu+u T E L(s-}R8-v

with(recallabove) JIL-1 <? (A - )-1 and IITII< (? + E). Moreover,we choose

as in Example 2.1 a Lipschitzcontinuousintrinsicderivative

A,,=
(z11 q$11uw
Lssq$1

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 223

of q$= on a neighbourhoodof {p, r}. Then AVb(r,() = Lq((#nTj, pone)o (pn

definesan intrinsicderivativeforVband so, denotingy = (nxf pnz); we have

(x -
z) = AuwqOuw(y)* Ln -
-)uw + Ass4ouw(Y) T
Tn -Z)ss

= (Auw/quw(Y) + Assquw(y) TLA(x) L-) .L2 ( )uw

and analogously

(x-z)ss = (Auw0ss(Y) + Ass0ss(Y) T A(T,,) L-) LL -

We wantto define
(3.1) A(x, z) = (A uw ss(y) + Ass ss(y)TnA zL- n)
0 (Auwq0uw(y) + AssouW(y)TnA(xT z)L-n)-i

= A(x, z) (x - z)uw. In order to show that

so that, automatically, (x -z)ss
(3.1) makessense we need thefollowinggeneric(open and dense) hypothesis:
(3.2) Duw4uw(r)is an isomorphism.
Here D denotestheusual derivativeand (3.2) meansthatunstable/weak-stable
directionsare not sent to strong-stabledirectionsby (pN. Notice now that
if > 0 is small (and then n is large) V1 and V2 are small and so y =
6
(-p-N(X),(pN-(Z)) must be close to either (r,r) or (p,p). This means that
(up to reducing 6) we may assume Auwquw(y) to be close to Duwquw(r) =
Azuwquw(r,r)or Duwquw$(p)= Auwquw(pp). On the otherhand, if n is large
we also have

IIAss uw(y)
TA(x, z)L II < constC0(6X )

small. It followsfrom(3.2) that (Auwouw(Y)+ A ~ssuw(y)TnA(xz)L-n) is

invertibleand the normof its inverseis at most
ai = 1 + max{|j(Duwquw(p)) 1jj, ||(Duw4uw(r))11}.

Moreover,the same argumentalso showsthat if n is taken large enoughthe

norm of (Auwqss(y) + Ass~ss(y)TnA(x, z)L-n) is at most

a2 = 1 + max{jjDuw~ss(p)|I| IjDuw4ss(r)1}.
This provesthat (3.1) is indeed definedand gives IIA(x,z)HI< Co, as long as
we take fromthe beginningCo > aja2. Successiverepetitionofthisprocedure
extendsthe definitionof A to W = Uk>oUk (keepingIIAII< Co) and now the
proofwill be completeifwe showthat such A is (C, y)-Hdldercontinuousfor
some -y> 0 and C > Co. Let (x, z) E Uk and (x,z) E Uk with k < k. Denote
Xi = 0-'(x), i(x) and zi = V)-/(z), for i > 0. If k < k then Xk and Xk
i;,i =
V1's and so IIA(xk,Zk) -AQk, Zk)II < 2CO < 2COKIIXk- 4kII
belongto different

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
224 J. PALIS AND M. VIANA

If, on the contrary,k = k then both (Xk, Zk) and (k, Zk) belong to Uo and
so, by construction,IIA(xk,Zk) - AQk, Zk)I < CoIIXk- k1. Now we proceed
by induction:Let 0 < i < k and assume that

IIA(xi+i, zi+i) - A(Xi+l, zi+,) 11< Cllxi+l -.'Xi+lII"

forsome 0 <y < 1 and some large C > 0. Clearly,

Hxi - Jijj > (|IDS 1|(n?N)Hxi~i-?iH.

On the otherhand,expressingA(xi, zi) and A(xj, zi) in termsof A(xi+,, zi+,),
A(?i+l, zi+1) and thensubtractingwe get
IIA(xi, zi) - A(5x , zi)II < a3 11Xi-xiI

E )n A(xi+1,
+ a4 (0+ zi+l ) - A (,i+ 1,zi+l ) II

with a3, a4 dependingonlyon W=(N. Taking C > 0 sufficiently large with

respectto a3 and diam(U) we obtain a3Ijxi -NiI < (C/2)llxi - diI7. On the
otherhand

a4 (0 +E)IIA(xi+,zi++) - A(Qi+lzi+l)1

(0 + E . IID7-11rF)
? Ca4IIDW-1IlyN I xi - IF,
C
K x-xi

as long as we fix -y < log((A - E)/( + e))/ log JIDW-1 11(and take n large
enough). Hence IIA(xi,zi) - A(i, zi)JI< Cjlxi -xljjjy and by inductionthis
givesIIA(x,z) -A(z, z)II < CIIx-XII1. By symmetry, Holdercontinuityinthe
second variablealso follows. O
Remark 3.3. It is also clear fromthe argumentthat A(p,p) = 0, i.e.,
ITpWUc(Al) C Eu e Ew.
PROPOSITION 3.4. For A1 as abovethemap F: WU(Al) 3 x -4 TXWU(X)
C1 on compactparts of Wu(Al).
is intrinsically

Proof. Given any compact K C Wu(Al) thereis k > 0 such that K C

Un+N k+i(W) where,as before,W = WU(A) n V. Note that F(Wk+i(X)) =
Dpk+i(x) . F(x) forall x E WU(Al). Therefore,the propositionwill followif
we provethat the restrictionof F to W is intrinsically
C1 and we proceedto
do this. The argumentgoes as follows.Let
DuVu ) and DDs?u}
DV)= DsVb ob=

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 225

denote the expressionof the derivativesof / = Pn+N and X = 0N in the

splittingRm = Ru x R8. Note that forx E W

D+(x) = Db(Wnx). (Un 0)

II < (a - E) and IISiI<
R8) satisfyingIIU-I
with U E 12(Ru, IRu), S E 12(1R8,
(A + e). We use Proposition2.10 with f = 'b, Xo = W, X = f(W), B a
closeddiskofradiusR = IIDuOs
IIII(Duu)-II + 1 in C(Ru,IR') lRuXS and
F: Xo x B -* inducedby D+: givenx E Xo
X x B beingthe graphtransform
and h E B
(3.3) Fx(h) = (Du/)s(x) + DsVts(x) h) * (DuVu(x) + DsVu(x) *h)-1.
Observethat, as long as n is takensufficiently
large (by choosing6 small),

IDsVbu(x)h. (Du4,u(x))-1i < R IIDsQuIII(Duou)i11(A + )n(c - n <K 1

and, analogously,IIDOP(x) h*(Dubu(x))1 11 < 1. ThereforeFx(h) is wellde-

< IIDu,,s(x)
finedby (3.3) and it belongsto B: IIFx(h)II (Du/ou(x))-1II+1< R.
Clearly,F is intrinsicallyC1 (it extendsto a C1 map on a neighbourhoodof
X0 x B) and the combinationof the constructionsin Lemma 2.5 and Ex-
ample 2.12 yieldsan intrinsicderivative
-f(X1,X2) 0
AF((xi, hi), (X2,h2)) =
AFhi (Xi, X2) AFX2(hi, h2))
forsome C1 > 0 (dependingonlyon ),
satisfying,

(hi Ih2)11< Cl (A
11AFy ) forall y e Xo and all h1,h2 E B.

C1 but choosingAf 1 in such a way

It is also clear that f-1 is intrinsically
that the hypothesisof Proposition2.10 holds is somewhatdelicate. Using
Proposition3.2 we may writeforx, z E X
f1 (x) - f i(z)
= ((f-(x) - -
fi(Z))uw, A(fi-(x), f-i(z))(f l(x) f-1(z))uw)
= (L-n(i-i (X) --1 (Z))U, A(fi- (x), f-i (z))L-n(O-i (X) - ?-1
Therefore,fixingAO-' an intrinsicderivativefor /-1,
/ ~~L-n 08
Af -1 (X, Z) = * AO-1 (XI Z)
( A(f-i(X), f (z)) L
L-n
is an intrinsicderivativeforf1. Recall moreoverthat IIL-II < (A -
Thus, forsome C2 > 0 dependingonlyon Co and , IIAf-1(x,z)II < C2(A -

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
226 J. PALIS AND M. VIANA

0', forall x, z E X. In thisway,havingfixed? > 0 and 6 > 0 smallenough,

we have

(3.4) JAl/Fy(hl, 11
h2)11 ||fZl (xz) || < ClC2 ( ) as < 2

forall y E Xo, x, z E X and h1,h2 E B. This provesthat the invariant

section 3f of F is intrinsicallyC1 and the propositionfollowssince clearly
TxWU(X)= graph(&(x)) foreveryx E W. F

PROPOSITION3.5. Let x E Wu(Al), S0 and El be (small) C1 sections

transverseto Wu(x) and ir: 0onwu(Al) - Eifnwu(Al) denotetheprojection
C1.
along the leaves of Wu(Al). Then ir is intrinsically

Proof. Fix coordinates (ok) = (?1,.... I 7,(1, .. ,() such that So C

{T/= (0, 0,.. ,0)}, E1 C {rT= (1, 0, . . .,0 )} and each leafintersecting
Y0 in a
point (0, z) can be writtenas the graphof a function( = G(r; z). Then

graph (a(r,; z)) = T(?,,G(,;z))WU(?h

G(r; z))

and so, by the previous proposition, (r, G(r; z)) - (7; z) is an intrinsi-
cally C1 map. For -E < t < 1 + E we defineg(t; z) = G(ti; z), where
(1,0, .. . I,0), and then

H: (t,g(t;z)) dg
dt(t; =_9_ (to;
z) a Z)

C1. Notice that ir(0,z) = (a, G(i;z)) = (g, (1; z)). We fix
is intrinsically
an intrinsicderivativeAH of H and define,for each t, AHt(Ci, (2) v =
AH((t,I1), (t, 2)) *(0,v) (recallLemma 2.5). For (0, zi), (0, Z2) E EY n Wu(Al)
we denote by F(t; z1, Z2) E C(RS, RS) the unique solution of the initial value
problem:
d (t; z1, Z2) = AHt(9(t; Zi), g(t; Z2)) F(t; Zl, Z2)

LF(0; Zl, Z2) = id.

Then (Zi, Z2) 1 F(t; z1, Z2) is continuousand we claim that it is an intrin-
sic derivative for z ~-? g(t; z). In fact, putting 0(t) = g(t; zl) - g(t; Z2)-

{
F(t; zl, Z2) (Z1 - Z2), we get

d(t) = AHt(g(t; Z1), g(t; Z2)) *a(t)

9(0) =O.
Hence 0 is identicallyzero, whichprovesthe claim. In particular(Zl, Z2)
F(1;z1, Z2) is an intrinsicderivativefor z "-? g(1; z) and so ir is intrinsically
C1, as we stated. F

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 227

Remark 3.6. In particular ir is bi-Lipschitz continuous and so it preserves

metric invariants such as Hausdorffdimension or limit capacity.

Remark 3.7. If one assumes that (o is of class C3 then stronger conclu-

sions follow from the arguments above: Using Proposition 2.10(b) one gets
that Y: Wu(Al) 3 x - TXWU(X) is intrinsicallyC1+'Yforsome By> 0 and so
the same holds for the holonomy maps ir: o n Wu(Al) -* El Wu(Al) of
Wu(Al). On the other hand, this last conclusion will also be obtained (in the
case w = 1) in Proposition 4.3, by a differentapproach which applies also if
(0 is just C2.

4. Thickness in higher dimensions

Let Al be as constructed in the previous section. Here we assume that w =

1; i.e., Do(p) has a unique (necessarily real) weakest contracting eigenvalue
A = Al. Then we consider ir: A1nfWs (p) -* R to be an arbitraryintrinsically
C1 map such that ker(Air(p,p)) does not contain ITp(A1 n Ws,(p)) = EW
(i.e., Air(p,p) I Ew is bijective) and we show that Tr(r(A fl WlO(p)), 7r(p)) is
independent of the choice of ir. We call

TU(Al,p) = T-(r(A1 nfWs,(p)), ir(p)), any ir as above,

the local unstable thickness of Al at p and prove that it is strictly positive
and varies continuously with the diffeomorphism:A C2-small perturbation of
p yields a small variation of Tru(Alp).
We keep the notation of Section 3. Let 1rw: A1 n Ws,(p) -* R be the re-
strictionto A, nWs (p) C {0u} x R8 of the projection (l..... . .l uI... I) ,
Ci We claim that 1rwis a homeomorphism onto its image Kw and more-
over rW-1 is intrinsicallyC1+'Y on KW. In fact, by Proposition 3.2, for every
x, z E A1 n Wls(p) we have

(x - z)ss = A(x, z) .
(0U, 7rw(x) - 7rw(z))

and so (for C > hAilas in the proposition)

1hrw(x)- 7rw(z)I < lix - Zii ? (C + 1) 1hrw(x)- 7rw(Z)I

This proves that 1rwis invertibleand 7r-1is Lipschitz continuous. Moreover,

-
Ar-(s, t): R 3)v (s), 1rW-l(t)) (0u, v)) E R'
(v, A(7rw-1
defines a H6lder continuous intrinsicderivative for 7rw-l.
Let now ir: A1ln Wls(p) -* R be an intrinsically C1 map such that
the kernel of Air(p, p) does not contain ITp(Ai n Wls(p)) = graph(A(p, p)
{0u} x R). Then ir o 7r,-l is intrinsically C1 with A(ir 0 r1rl)(0, 0) = Air(pp)

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
228 J. PALIS AND M. VIANA

AHw1(O.,O)#7 0. Since ir(Ai nlW1 (p)) = (7ra 7rl)(Kw), we conclude that

r(ir(Aln W,(p)), 7r(p)) = ir(Kw, 0), as a consequenceof the following
simple
result.

LEMMA 4.1. Let K c R be a Cantor set, y E K and 9: K -- IR be an

intrinsically 5$0. Then r(g(K),g(y)) = -r(K,y).
C1 map withLAg(y,y)

Proof. Fix 6 > 0 small and take E > 0 such that

(4.1) jAg(xl,z1)I < (1 + 6) |9g(X2,z2)1 forall x1,x2,z1, Z2 E [Y-e, Y +l].
Clearly,g is a homeomorphism on [y- e, y + E] if E is small enough. Hence,
given Cantorset c [y e, y + e], U = (Un)n is a presentationof L if and
a L -
only if g(U) = (g(Un))n is a presentationof g(L). From (4.1) we get that for
everyu E nUn,n > 1,

T(g(L), g(U), g(u)) < (1 + 6)T(L, U, u).

Therefore,r(g(L)) < (1 + 6)T(L) foreverysuch L and so r(g(K),g(y)) <
(1 + 6)r(K, y). Since 6 is arbitrarywe concludethat Tr(g(K),g(y)) < r(K, y)
and the reverseinequalityis provedin the same way. E

For futureuse, let us state also the next result,whichfollowsfromthe

previouspropertiesby well-known arguments:Part (a) is a directconsequence
of Proposition3.5; forpart (b) one also uses the transitivity
of o ItAl.

PROPOSITION 4.2.
(a) Let q E Wu(p), E be a Cl sectiontransverseto WU(p) at thepointq and
ir: WU(Al) n E - R be an intrinsically C1 map such that ITq (Wu(Al) n E)
is not containedin ker(AQir(q,q)).Then r(ir(Wu(Aj) n E),ir(q)) = ru(Al,p).
(b) More generally,given z E WU(Al), E a transversesectionto WU(Al) at
z and 7r: Wu(Al)nE -+ R submersionwithITz (Wu(Aj)nE) ? ker(A7r(z,z)),
then T(7r(Wu(Al)n i),ir(z)) is equal to ru(Al,p).

Our strategy to prove that TU(Al,p) = r(Kw,O) is positive is to write

Kw as an invariantset of some C1+ expandingmap, so that the reasoning
in Proposition6 of [N3] (in the language of [PT, Ch. IV]) can be applied to
it. For the constructionof this map we proceed as follows. Let W be the
connectedcomponentof WS(p) n v containingp. By construction, forevery
x E A, v the connectedcomponentof Wu(x) n v containingx intersects
_

W in a unique point, whichwe denote by 7rs(x). Observe also that every

Wf= {z: 4ji(z) E V for0 < j < f and 0'(z) E W}, f > 1, has exactly2e
components,denotedby Wje, 1 < j < 2'. For each f > 1 we definea map
V
gj Kw Kw by 9 = arm7rsoo ri i .

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 229

Clearly, gi is intrinsicallyC1 and it followsfromthe next result that it is even

intrinsicallyCl+7.

PROPOSITION 4.3. There are C > 0 and ayE (0,1) such that for every
e > 1 and 1 < j < 2t the map 7rsI (Wje n Al): Wj, n Al -- W admits a
(C,0y)-Hdlder continuous intrinsic derivative.

For the sake of clearness we postpone the proof of this result to the end
of the section. Recall also Remark 3.7.
Observe that ge maps KW onto Kw in a 2e to 1 way: For each 1 <
i < 2t and Kiye= 7rw('ot/(Wj,e n Al)) the map ge I Kje: Kjie - Kw is a
homeomorphism. It is also easy to check that the Kj,e, 1 < j < 2t,have their
convex hulls Kje two by two disjoint. Moreover, for > 1 large enough every e
(ge I Kjie) is an expansion:
(4.2) I zAge(tr)Ij > 2 (say) for every t,r E Kite.

e
We now fix > 1 so that (4.2) holds and denote g = gi and Kj = K3,e. At
this stage the fact that Kw has strictlypositive local thickness can be proved
just by translating, in a more or less straightforwardway, the arguments in
[N3], [PT, Ch. IV] into our language of intrinsicdifferentiability.
Alternatively,
one can show that g can in fact be extended to a C1+7 map G defined on a
neighbourhood of Kw, which is easy to prove in this one-dimensional setting.

LEMMA 4.4. Let K C R be a Cantor set and g: K -- R be intrinsically

C1 (resp. intrinsicallyC1+7). Then there exists G: ]R -* R a C1 (resp. C1+7)
extension of g.

Proof. Fix 0: [0,1] -* [0,1] a C' function such that 0(x) = 0 and
0(1 - x) = 1 whenever 0 < x < 1/3. Let also ao = infK, bo= sup K. We set:
G(x) = g(x) if x e K;

G(x) = (g(a) + Ag(a,a)(x-a)) (i -? ))

+ (g(b) + zg(b, b)(x- b))0 (xa)

if x belongs to a gap (a, b) of K; G(x) = g(bo) + LAg(bo,bo)(x - bo) if x > bo

and G(x) = g(ao) + Ag(ao, ao)(x - ao) if x < ao. We also define H: 1R-O R
by putting: H(x) = Ag(x, x) if x E K;

H(x) = Ag(a,a) (i-E (t )) + g(b,b)0 (x )

+ 0i (x -) ( bgqj,a) - Ag(b,b) + (x -) (Ag(b,b) - Ag(a, a)))

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
230 J. PALIS AND M. VIANA

if x E (a, b), a gap of K; H(x) = Ag(bo,bo) if x > bo and H(x) = Ag(ao, ao)
if x < ao. It is a simpleexerciseto checkthat H is continuous(Cr if Ag is
C7) and G'(x) = H(x) foreveryx E IR. E

This means that Kw is a dynamicallydefinedCantor set, in the same

sense as in [PT, Ch. IV] and our claim that r(KW, 0) > 0 follows.
Let us nowobservethat (Kw, G), as we constructedit,variescontinuously
withthe diffeomorphism W. In orderto explain this affirmative
we let (p be a
diffeomorphismC2-close to (Wand denote by Kw((p), Kj(np), Kj( p) and G(np)
the objects obtainedby performing the above construction for(p (note that we
may,and do, take f((p) e (pis closeenoughto p). Then it is straightforward
= if
to checkthat (Kw(np),G(np)) is close to (Kw, G) in the sense that

(a) corresponding endpointsof Kj(np) and Kj are close;

(b) G(np)is Cl-closeto G and theirderivativeshave nearbyHolderconstants
(see also the proofof Proposition4.3 below).

It follows,see again [PT, Ch. IV], that r(Kw(p), 0) is close to r(KW,0) and
thisprovesthat the local unstablethicknessru(Al, p) is a continuousfunction
of the diffeomorphism, as we claimed.
We close this sectionby presentingthe proofof Proposition4.3.

Proof. We keep the notationfromthe proofof Proposition3.2. More-

over, we use C and 7yas genericnotationforconstants(C > 0 large, 0 <
7y< 1) dependingonly on the diffeomorphism W. Let S = {S(x): x E A1}
where S(x) denotes the connectedcomponentof WS(x) n V containingx.
Note that W4,j E S for every 1 < j < 2e, f > 1. For S,S E S let
7r(S, ; .): S n A1 --+S n A1 be the projectionalong the leaves of Wu(Al),
inside V. By Proposition3.5 7r(S,S; ) is C-Lipschitzcontinuous. On the
other hand, as observedbefore,{IIx - zIl: (x, z) E (S x S) n UO,S E S}
is bounded away fromzero. Hence (see Remark 2.6) thereis a C-Lipschitz
continuousmap L(S, S;,): ((S n Al) x (S n Al)) n U0 -O L (Rs, Rs) such that

(4.3) 7r(S,S; x) - 7r(S, S; z) = A(S, S; x, z) (x - z)

forevery(x, z) E ((SnA1) x (SnA1))nU0. Hereand in whatfollowswe identify

each S E S with its image under the projection (4l, ... - , ,, Cm+i.... I ') ''4
((u+l,... , Cm). We writethe expressionof z\(S, 5; x, z) with respectto the
splitting IRs = JR' x RS-l

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 231

and thendefine
Aw = Aw +Aw? s A(x, z) 0
A (S,S;xz) = I (SS;xz),
A s= Asw+Ass A(x,z) O.
whereA(x, z) is as givenbyProposition3.2. Then A(S, 5;.,.) is (C, 'y)-Hdlder
continuousand satisfies(4.3). Moreover,Aw is uniformly boundedaway from
zero and infinity,as a consequenceof that same proposition. We also need
some information on how A (S, 5; , ) varieswith S, S E S and this is easy to
get. Let us denote dist(S, 5) = SUPXEsnA 7rI(S,S; x)-x lI. It is easy to see
11
that A\may be takensuch that IML(S, 5; x,z) - idlI< Cdist(S, 5) and so

(4.4) (S, S; XIZ) -1| < C dist(S, S)

Aiw

for every x,z E S n A1. Note that if S E S then b(S) C S' for some
S' E S. We let ALb(.,) be a Lipschitzcontinuousintrinsicderivativefori
and denoteby zLb(S;., ) its restrictionto (S n A1) x (S n A1). Throughthe
above identification
we maythinkof A+b(S;., ) as takingvalues in C(RS,RS).
Proceedingas beforefor A we constructa new, (C, -y)-Hdldercontinuous,
intrinsicderivativeAb(S;., ) for b 1(S n A1) havingthe form

AOb(S; x, z) = ( 0) (S; x, z)

withrespectto the splittingRs =

-R x Rl1. In the same way we obtain a
continuousintrinsicderivativefor /-1I (s(S) n Al):
(C, -y)-Hblder

AO-'(S;XI, Z)= (dW> 0) (S;X',Z').

Observe that Proposition3.2 impliesthat oibwis uniformly bounded away

fromzero and infinity;
moreover,independentlyof all the choices,we have

Aw-' (S; O(x), A(z)) Aow(S; x,z) = 1 foreveryx, z E S n A1.

We also pointout that this construction
yields

(4.5) Abw(S; r(S, S; x), r(S, S; z))-Abw(S; x, z) < C dist(S, Sj)

forevery x, z E S n A1. Now we fix So, So E S and constructan intrinsic

derivativeA7r(So,Sgo;,) for 7r(So,So; ) in the followingway. For (x,z) E
((So n A1) x (So n A1)) n UO we set simplyA7zr(So, So; x, z) = A (So, go; x, z).
Let now (x, z) E ((So n A1) x (So n A1)) n Uk, i.e., k > 1 is the minimum
integersuch that (0-k(X),,O-k(Z)) E Uo. We denote x~= 7r(SoSo;x), Xj =
-i (x), ?Ij = -ii(x) and analogously for z, zj and ij, j > 0. We also let

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
232 J. PALIS AND M. VIANA

Sj = S(xj) = S(zj), Sj = S(j) = S(ij), 0 < j < k. Then we define

Ai\r(So,So; X,Z) =
-\Ok(Sk; Xk,Zk) *'(Sk, Sk; Xk,Zk) *AO k(Sk; X,Z).
We are left to check that with this definition A7r(So, So;,) is (C, -y)-Hblder
continuous: we consider x, z,w E So f A1 and prove that

(4.6) zAir(So,So;x , z) -z7r(So, So; x, w) < C Iz - w 1

with (xz) E Uk, (x,w) E Ue and (z,w) E Urm k,f,m > 0. Clearly, j | 4
dist(Sj, Sj) is exponentially decreasing and i zIj -wj is exponentially
increasing, at least while j < m. It follows that, if Iz - wl is small enough
depending on D = sup{dist(S', S"): S', S" E S}, then there exists s < m
such that

(4.7) w > dist(S., Se).

1zs-Ws
Note that for the purpose of proving (4.6) it is no restrictionto assume that
Iz - wI is small with respect to D and we do so fromnow on. We fixs minimum
satisfying (4.7) and observe that there is A < 1 (depending only on p) such
that

(4.8) dist(Sj, Sj) < CA3 fz - wIt forj > s,

(4.9) Izi-wiI < CASIO -z-wl for0<j<s.
This is an easy consequence of the definitionof s and the exponential variation
of dist(Sj, Sj) and Izj - wjI mentioned above. Since the 7r(Si,Se; ) admit a
uniformLipschitz constant, we also get that

(4.10) izj-ibjl < CAS-j Iz-wl for0<j<s.

Now we take r = min{k, X,s4. By construction

Tr(So, So; x, z) =
(7:i ) (So, So; x, z)

with Azrw(So, So; x, z) = AO' (Sr; rv r) *Axw(Sr,Sr; r,Zr) z).

* -(S,; Xv
and analogously for (x, w). Then

So; X,Z)
Azrrw(So, Arw(Sr,Sr; Xr,Zr)
Airw(So,So; x, w) - Xr, r)
rW(rw(Srr;

R(x, z, w) = J7J (Si; xi ii) Aw(Si; xi, WO

1 AIbw(Si; Xivjfi)* AIw(Si; xi, Zi)

From (4.9), (4.10) and the uniform Holder continuity and boundedness of
AOW~i,
( ), AOWb(Si;v) it follows that

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 233

We claim that

/.'irw(Sr, Sr; Xri

rw(SrvSrvrr) Zr)_ 1 < C1Z-W
(4.12)

Note that this implies (4.6) in an easy way. In fact, from (4.11), (4.12)
we get IArww(So,So;x,z) -A7rw(So,So;x,w)I < CIz - wI and then, since
ALrs(So,So;x,y) = A ip) A7rw(So,So;xy), for y = z or w, the same holds
forA7r,. Therefore,the proposition will followonce we have proved this claim.
Suppose firstr = k; then, because r < s < m, we also have r = 1. It follows
that
Azrw(Sr,Sr; Xr,Y) = Aw(Sr, Sr; Xr, Y), Y = Zr or Wr,
and so (4.12) is a direct consequence of (4.9) and the fact that Aw is Holder
continuous and bounded away from zero and infinity.Let now r = s. Once
more by definition,

z\7rw(Sr,Sr; Xr, Zr) = zw(Sk, Sk; Xk,Zk) f i ,(Si . X Z )-

Note that dist(S,, 5j) < or-J dist(Sr, Sr) for r < j < m and so, using (4.5),

ft A?w (Si; xi, Zi)

On the other hand, by (4.4) and (4.8), IAw (Sk, Sk; Xk,Zk) -1 1 < C IZ - Wl
and this proves that IJ 7rw(Srr; xrzr) - 11 < Clz - wIj. In this same way
one shows that IA7rw(Sr,Sr; Xr, Wr)- <? Clz - wI' and then (4.12) follows
immediately. El

5. Unique least contracting eigenvalue

Let ((p be a C2 one-parameter family of diffeomorphismswhich at

p = 0 goes through a homoclinic tangency associated to a hyperbolic fixed (or
periodic) point p of po. We prove here that under a few generic assumptions
to be stated below such a family exhibits, at fib-- 0, homoclinic tangencies
associated to periodic points Pj -- p such that

(I) Dp2. (Pj), j = period ofPj, has a unique weakest contractingeigenvalue

which, therefore, a real number.
is
A dual statement, for expanding eigenvalues, can be obtained in the same
way. No assumptions on the dimensions of Wu(p) or WS(p) are required in
this section.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
234 J. PALIS AND M. VIANA

Observefirstthat,generically,
eitherDpo(p) satisfiesthe propertyin (I)
or else
(II) Dpo(p) has exactlytwo weakestcontractingeigenvaluesand these are
complexconjugatenumbers.
In the firstcase thereis nothingto prove,so we assume fromnow on that
(II) holds. Notationanalogous to that of Section 3 is used; this means that
w = 2, A1 = Aeio, A2= Ae-i with A > 1A31and XER \ {kir: k E Z}. For
the sake of simplicitywe assume that thereare C2 tindependent coordinates
('i I ... the
I, u1, .... , ) linearizing Ad, forIuI small, on a neighbourhoodU
ofthe analyticcontinuation pHofp. Moreover,we maytake thesecoordinates
to satisfyconditions(Al)-(B3) of Section3. Here we may even assume that
DHo,(pj) I Ew is conformalwith respectto the euclidean metricinduced by
the coordinates 1,(2. We also suppose that the tangencyis quadratic and
the family((p unfoldsit generically.Then we may take, say for itt> 0,
points qH E W',(p,), rHE Wlu(p,) dependingcontinuouslyon ,tt,such that
(r,) = qH forsome fixed N > 1, ro,q0 belong to the orbit of tangency
N

and r,, qH are points of transverseintersectionof Wu(pa), Ws(pa) forevery

/u> 0. Recall, moreover, that Wu(p,) and Ws(p,) also have,fora sequenceof
parameter values itt= ,uj -- 0, points of tangential intersection. For each fixed
,u= ,uj and everysufficientlylargen > 1 we maytake,as in Section3, a neigh-
bourhood V = V(j, n) of {pp, q,} such that A(j, n) = nkeZ M<jn+N)(v) is a
n+invariant hyperbolicset and Wn+N I A(j, n) is conjugateto the 2-shift.
Then, givenany periodicpoint P E A(j, n), one may findparametervalues fit
arbitrarilyclose to ,uj forwhichAp has homoclinictangenciesassociated to
(the analyticcontinuationof) P. We considerP = P(j, n) to be the unique
Wn+N-fixed point in A(j, n) \ {pi,}. Clearly,the orbitof pj passes arbitrarily
close to pHifj and n are sufficiently in orderto concludeour
large. Therefore,
argumentit is now sufficient to show:

PROPOSITION 5.1. Givenj > 1 (large) thereexistarbitrarilylargevalues

of n = nj for which Dpn+N(p) has a unique weakest contractingeigenvalue.

we continueto denotePu= guj

Proof. Forthesake ofnotationalsimplicity
=
and P P(j, n). We makeuse ofthe followingelementaryfactwhoseproofwe
omit.

LEMMA 5.2. Let P 1 Q1II

GL (RP), Q E GL (Rq) satisfyIIP-1I < 1. Let
(
E

L= ) EGL(RPxIRq) and denote

B
L1 =(A B_) and L )= A( ( n>>

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 235

Assume that A E GL (IRP) (and so also D- E GL (Rq)). Then thereare no >

onlyon (11P-1l1
1 and Co > 0, depending IIQIIIILII,11L-1ll,
IIA-llll(D-)-'11),
such thatfor everyn > no thereexistlinearmaps f: IRP- Rq",g: Rq -+ RP
satisfying
(a) lfII < Co and llll < Co(1lP-11 IllQll);
(b) graph(f), graph(g) are Ln-invariant.
Moreover,given 6 > 0 thereis ni = ni(6) > no such thatfor n > nl
(c) Il(1rfLn7r71)p-n - All < 6 and - D-11 6, where
IlQn(irgL-11jr1)
7rf: graph(f) -IRP and lrg: graph(g) -Rq are the canonicalprojections.

Let then
AUU Auw Aus /A1 0 0\
DON= | AW Aww Aws)| D9p(p)= |O Aw 0
Asu Asw Ass \ 0 0 Asv
be the expressionsof DWN and DWp,(p)with respectto the splittingRm =
RI X R2 x Rs-2 = Eu x EW x E5s. We also denote

(A1 wuA ww
) (Auu uw Aus
~~~~~AS-
A-w As-
A- A
~~"KAWU A1WW/ and DWPiN yA-
SW
511~~~~S As

and assume the genericproperty(cf. (3.2))

(5.1) AL=o(ro), and so also A-(pU= 0, qo), is an isomorphism.
Then a firstapplicationof the lemma yields,forj and n sufficiently
large,
linear maps fuw: RU xJ2 whose graphsare
JRS-2 fSs: Rs-2 ,- RU x R22,

invariantunder DHn+N(p) and satisfy

(i) Dgn+N(p) I graph(fuw)is conjugateto K1 (j pAn) forsome iso-

morphismKi E GL(Ru x R2) whichis close to XsopN(z)) and so also to
Ap=o(ro);
(ii) Dp(fl+N)(n ) Igraph(fss)is conjugateto A-nK2, some K2 E GL(IR'-2)
also to A-(pU= 0, qo).
close to A-((P) and therefore
Observe,on the otherhand, that Auu(ri) is an isomorphism, since Wu(p,)
and Ws(p,) intersecttransversely at ri. Then, by using the lemma again,
we obtain linearmaps fu: JRU--1R2, fW: R2 ,- RU such that the pre-images
7r1 (graph(fu)),7r-1(graph(fw))oftheirgraphsunderthe canonicalprojection
7ruw:graph(fuw)-- JRUx JR2are Dpn+N(p)-invariant subspaces satisfying
(iii) Dn+N (p) I7rU-
(graph(fu)) is conjugateto K3An forsome K3 E GL(Ru)
close to Auu(rp);

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
236 J. PALIS AND M. VIANA

(iv) Dp-(n+N)(p) I 7r-(graph(f,)) is conjugateto A- K0 forsomeKo

GL(]R2) close to Aw(qj);
GAuuv... vALw denoting the entries of A-1). Altogether,and recalling proper-
ties (B1)-(B3) of Section3, we see thisimpliesthat forj > 1 sufficiently
large
thereis c > 1 such that forn > j the isomorphismDW(n+N) has exactly
(a) u eigenvalueswithabsolutevalue > c-1 (a - )n;
(b) 2 eigenvalues,say A1,A2I with I1lI,IA21 E IC-' (A- )n, c(A + )n];
(c) (s - 2) eigenvalues with absolute value < c(9 + ?)n.
This reducesthe completionof our argumentto showingthat thereexist ar-
bitrarilylargevalues ofn forwhich IA1 5 1A21. Note firstthat Auu(t = 0, ro)
is not an isomorphism, since Wu(p),WS(p) are tangentat ro. We assume this
tangencyto be quasi-transversal, meaningthat dimker(Auu(11( = 0, ro)) = 1.
Since ker(Aw) = Awu(ker(Auu))and ker(Auu)= A- (ker(A- )), it follows
that ker(A-w(p = O,qo)) is also 1-dimensional.We fix {v1, V2} to be a ba-
sis of JR2 7r1(graph(fw)), orthogonalwithrespectto the euclideanmetric
induced by the coordinates 1,(2 and such that v2 E ker(A (,u = 0,qo)).
Referring alwaysto this metric,we denoteby C(v, a) the closed cone in ]R2of
amplitudea > 0 aroundthe directionof a (nonzero)vectorv. By construc-
tion Ko(C(v,,87r/18)) C C(iowr/18),with io = Aw(p = O,qo) vi, if Ko is
sufficiently near A- ((,t= 0,qo), i.e., ifj and n are large enough(recall (iv)).
Also, since Aw is conformal,A-n(Cpbo,7/18)) = C(vnT,/18), i = Awno. n

On the otherhand, since the eigenvaluesof Aw are not real, thereexist in-
finitelymany values of n forwhich fn = Awn i0 E C(vi,67r/18) and so
C(vn,7r/18)C C(v,, 77r/18).This meansthat forsuchvalues ofn, A-nK0 has
a strictlyinvariantcone A-nKo(C(v1,87r/18))C C(v,, 77r/18)and so its two
eigenvalues(A1)-' and (A2)-1 musthave different norms.This completesour
proof. ?

Remark5.3. It is also clear from(a)-(c) that dimW*(P) = dimW*(p),

* = u or s, and D(pn+N(p) is sectionallydissipativeif Dpo (p) is.

6. Renormalization and thick basic sets

Let Wobe a C2 diffeomorphism with a (nondegenerate)homoclinictan-
gency associated to a fixed (or periodic) point p. We suppose here that
dimWu(p) = 1 and Dpo(p) is sectionallydissipative,i.e., the product of
any two of its eigenvalueshas normless than one. The main goal is to show
that the unfoldingof such a tangencyby a genericone-parameterfamilyof
diffeomorphisms ((p yieldsthe formation, forarbitrarilysmall values of ,,
of hyperbolicbasic sets A2 = A2(,U) having(codimensionone stable foliation

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 237

and) arbitrarilylarge stable thicknessrS(A2). For futureuse we also check

that

(a) A1 and A2 are heteroclinically related; i.e., there exist some mutual
betweentheirstable and unstableleaves;
transverseintersections
(,3) thereexist periodicpoints P1 E A1, P2 E A2 such that Wu(p1) has a
nontransverseintersectionwith Ws(P2).

Actually,forthe proofof (a), (i3) we consideronlythe case when Dnpo(p)has

a unique least contractingeigenvalue. We suppose in additionthat, besides
the homoclinictangency,the point p also has transversehomoclinicorbits
and, moreover,these involvethe same separatrixof Wu(p) (resp. the same
connectedcomponentof Ws(p) \ W's(p)) as the tangency.Since any diffeo-
morphismwitha homoclinictangencymaybe approximatedby anothersatis-
fyingtheseconditions,theyrepresentno restriction to the purposeof proving
our main result.
For the constructionofA2 we firstdeduce a higher-dimensional versionof
the renormalizationschemein [TY], [PT, Ch. III]. The basic idea is to show
that, under a fewgenericassumptions,the unfoldingof (p) containsthe
unfoldingof familiesofnearlyquadraticdiffeomorphisms. More precisely,one
of iteratesof (p, to appropriatedomains close to the
shows that restrictions
tangencyhave the form

(x, Y) F-4 (X2+ v,xA) + e(v, x, Y) forsome A E Rlm-l and E C2-small,

chosencoordinates(x, Y) E IR x
whenwrittenin conveniently 'm-l and pa-
rameter v.
Let us describethis constructionin moredetail. We assume once more
that the ',oH,,u small, admit C2 ,u-dependent
linearizingcoordinates((, Z) E
R X Rm-l on a neighbourhoodof p. We fixthese coordinatesin such a way
that Wc(p~l)
{o C = O} and Wjc(p~l) c {Z = O}. The assumptionon the
eigenvaluesof Dnpo(p)means that we may choose the normin Rm to be such
that

(6.1) upI. IISpI< 1 (foreverysmall,u)

where a,, is the expandingeigenvalueof Dnop(p,) and S, = DnpoI Es(p,).
Let qo = (0, Qo) E Wls(p) and ro = (po,O) E Wiu,(p) belong to the orbitof
tangency, say ro = p-N(qo), N > 1. Then for (1u,,, Z) close to (0, po,O) we
may write N(, Z) as

(vt + a2Z+ b(( - Po)2 + bl(( - po)I'+ b2ft2

+ h(p, -poZ),
Qo + Vi + Al (( - po) + A2Z + H(p, - po, Z))

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
238 J. PALIS AND M. VIANA

wherewe have v,b,bilb2 E R, a2 E ?(Rm-1, R), VA1 E ?(R,I Rm-), A2 E

?(R~m-l,Im-l) and

(6.2) Dh = 0, DH = O.&(h = 0,Oh= 0&,h = 0 at (0,00 ).

We also assumethat the homoclinictangencyis quadraticand that it is gener-

ically unfoldedby the family((p, whichcorrespondsto having b 54 0 and
v 540. Firstwe introducen-dependentreparametrizations

(6.3) v = ?n(,u) = VU2n,/+ (a2J2nSnQo- po_O).

to checkthat for n > 1 suf-

Let K > 0 (large) be fixed. It is not difficult
ficientlylarge ?n maps some small intervalIn, close to zero in the ,u-space,
diffeomorphically onto [-K, K]. DenotingOn= (OnI In)-1, we now introduce
(n, ,u)-dependentcoordinates(x,Y) definedby

(6.4) ((, Z) = fn,(x, Y) = (cu/l2nx C-ny + Qo + VIa)

+ pour,-n

and take In: [-K, K]m+l -4 R x M to be given by

'J'(P, x,Y) = (,u,(, Z), ,u= On(P), ((, Z) = ,(X Y).

Let us also denote 4 : IR x M --I R x M, (/,, r7)= (/,, Pp(71)). Now a direct
calculationyieldsthe expressionof 4pn+Nin the coordinates(P, x,Y):

(n )(fl xi Y)
TO
(,b62 + bifxinu>
(- + b2u2t2 + v + a2uySy(Y + Vu>) + o2nh(,u Xn,

A161 + A2 (Sln Y + UnSSn, (Qo + VP)) + unH(p, Xn, Yn))

where ,u = On(V), Xn = Ynn = S1n(UnY + Qo + V/,). Observe that

IAI,IXnI< constIolcI-nand IIYnlI
< constIISIInand,inviewof(6.1), (6.2),this
implies that u2nh(, XnX,n) and unH(t, XnYn) -- 0 as n -x oc, uniformly
on (v, x, Y) e [-K, K]m+l. Using also (6.3) one concludesthat the sequence
(l4ln+NlJn)(f, > Y) converges uniformly to

x(P,6x, Y) = (v, bx2 + bi POx + b2Po + v, Aj.x)

V V2/

when n -x oc. Moreover,essentiallythe same argumentapplies to the deriva-

tivesand we concludein thiswaythat ( @ X in the C2
-4 topology.
Finally,we introduce

~b,,+ bipo b p2 bb2p bipo y1Pb

+P YPY+.b1P
2v -4v2 v2 xbx+ 2v 2vb

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 239

and then,immediately, the expressionof 4pn+N withrespectto (v,x, Y) con-

verges(in the C2 topology)to
A
(6.5) X(v,x, Y) = (v,x2 + v,Ax),
I E whereA =
n-1.
~~~b
Remark 6.1. As a consequence,the genericunfoldingof a homoclinic
tangencyassociated to a sectionallydissipativesaddle yieldsthe formation of
sinksclose to the orbitof tangency.In fact, xv(x,y) = (x2 + v,Ax) has at-
tractingperiodicorbits,say forv close to zero. The presenceofsuchorbitsis a
persistentphenomenonundersmallperturbations.It followsthat forarbitrar-
ily smallvalues of , (,a = On(v),v close to zero) ,o has an attractingperiodic
orbitcontainedin a (constlpl)-neighbourhood of the orbitof tangency.
Now the constructionof A2 proceeds in the same way as in the two-
dimensionalcase; we sketchthe main points and referthe reader to [PT,
Ch. VI] formore details. The crucial fact here is the existenceforthe map
x l4 x2-2, and so also forX-2: (X, Y) l-4 (x2 -2, Ax), ofinvariantexpanding
Cantor sets Kj with thickness r(Kj) -4 +oo as j +oo. Moreover, these Kj
are transitiveand have a dense subsetof periodicorbits. It followsthat each
Kj has, forn large and ,a = On(v), v close to -2, an analyticcontinuation
as a hyperbolicbasic set Kj(n, /,) of (0n,1,0o +N o n-b). In particular,the

Kj (n, ,a) have codimension-1 stable foliationand stable thicknessrS(Kj (n, ,t))
close to r(Kj) > 1. Then, we just take A2 = n,(Kj (n,tt)) withj and n
large and ,A= On(V), v close to -2.
Let, moreover,P = (2,2A) and Q be any other (hyperbolic)periodic
point of X-2. We denote by P(v) and P(n, ,u) the analyticcontinuationsof
P for,respectively, xi',and ( , o (,0n+N o bn- ), n large, ,A= On(v), v close
to -2. We also introduceanalogousnotationsforQ. Now, it is easy to check
that
* WS(Q(v)) has transverseintersectionswith Wu(P(v)) forall v close
enoughto -2;
* Wu (Q(v)) intersectsWS(P(v)) ifand onlyifv < -2 and thisintersection
is transverseif v < -2.
Combiningthis with the fact that (compact parts of) stable and unstable
manifoldsof hyperbolicperiodicpointsvarycontinuously withthe maps, one
concludesthat foreach largen thereare ,t-valuesAn = On(f7n),V1nclose to -2,
forwhich(see Figure5)
* P(n, ftn)and Q(n, ftn)are heteroclinically
related;
* Wu (Q(n, ftn))also has nontransverse
intersectionswith WS(P(n, ftn))
.
This applies, in particular,when Q is a periodic point in any of the Kj.
Therefore,thesecommentsreducethe proofof properties(a), (,B)to checking

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
240 J. PALIS AND M. VIANA

W'(Q (n, p))

(6.6) pis hteroclincallyreatedtoP(n, ,a) 1 (

t \ . /~~~~~~~Ws
P(n, p))

Q(n, pl) /Ws (Qn, pl))

W'(P(n,p1))

FIGURE 5

that

(6.6) relatedto P(n,/u) =

P-,is heteroclinically n,p(P(n,p))
forarbitrarily largen and ,A= On(v),v in a fixedneighbourhoodJ of -2.
The basic strategyto provethis is the same as in two dimensions.Let
us firstgive a briefoutlineof it. We show that thereexist compactdomains
au1(n,p) C Wu1(P(n,,u)), ES(n, ) C WS(P(n, u)) 7 C W1u(p), Es c WS(p)
such that (see Figure 6)

Oru(n, 6
Es(n>p) /\U(P)

\ \'-s o~W'(P(n, p)) \ cu

Wu (P(n, p)) p ,
\Es

FIGURE 6

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 241

* P(n, ,) E a1u(n,[) nfl s(n, m,);

* the pointof tangencyqo belongsto 49au n a&S;
* auc(n,M) -*, au and Es(nf,I) -+ Es as n -* oo (and so t -* 0).
Moreover,our assumptionon the existenceof transversehomoclinicorbits
associated to p, together with the construction of Es, a1u, to be described
below, assure that Wu(p) (resp. Ws(p)) accumulateson au (resp. Es) "from
the appropriate side" so that it eventually intersectsEs (resp. a.U) transversely.
Hence, for n large Wu (p,1) (resp. WS(p,1)) cuts ES(n, ft) (resp. au"(n,p)) and
the affirmative
(6.6) above follows.
On the otherhand, this strategyrequiresconsiderablymorecare in the
presenthigher-dimensional setting,especiallyin the constructionof Es(n,iA),
Es. We concentrateon this, the argumentfor au1(n,,u), a.u being analogous
(and simpler). As mentionedbefore,we may suppose that p,1has a unique
weakestcontractingeigenvalue,whichwe denoteby AP. We also assume once
morethe transversality condition(5.1):

(6.7) ,=o(ro) is an isomorphism.

First we deduce the followingresulton existenceof invariantsplittingsin a
neighbourhood ofthe tangency.As before,we denote Es' = {O} x {O} x IIm-2,
EUW= lRx lRx {Om-2}. We also let 0. = 100ol,A = JAol,9 = fjDpo(p) I EssJJand
fix? > 0 such that 0 + 2E < A - 2E. Moreover,we continueto denoteby U a
neighbourhoodofp whereC2 linearizingcoordinatesforthe p,1 are defined.

PROPOSITION 6.2. Thereexistneighbourhoods B of qo and I of 0 E R

and constantsCo > 0 and nO> 1 such thatfor n > no and a E I thereis a
Holder continuoussplittingTBM = En EDnss satisfying
(a) dim El (z) = 2 and dim E',, (z) = m -2 for z E B;
(b) Esns admitsan integralfoliationRS,1
(c) Don+?N(z) *E* (z) = E,,(,o+N(z)) for * = uw or ss and for every
z E B nl(jn+N)(B) such that ,o' (z) E U for 0 < i < n.
(d) For everyz as in (c) we have
-
n
1) lv < ||Dson+N(z) V|| < Co(a + E) l|VII if v E El (z)
and J|D n+N(z) v|| < Co(O + E)n l|VII if v E Enss(z).

Proof. The construction ofthe subbundlesE', Esns is done by a stand-

ard fixedpointargumentanalogousto that in [HP, Theorems6.1, 6.2] or Sec-
tion 2 in the presentpaper. Thus we just presentan outlineof it, leavingthe
details to the reader. We let y1ube a segmentin Wiu,(p) containingp and
ro in its interior
and we take B and I to be small neighbourhoodsof qo and
0 E R, respectively,accordingto certainconditionsto be stated in the sequel.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
242 J. PALIS AND M. VIANA

For simplicitywe choose B to have the formB11x B5, B11a neighbourhoodof

o E lR _ Eu and B, a neighbourhoodof Qo in Rm-1 l Es. Then we denote
OUB = 'Bu x B, and &SB = Bu x OBR.Since we are assuming the tangency to
be nondegeneratewe may fix -yu, I, Bu and B5 in such a way that, forMtE I,
large,
N(pyu) does not intersectOSB (see Figure 7). Then, forn sufficiently

l pn+N (B) lll

FIGURE 7

we have
(6.8) epn+N(OuBn xn') fL B = 0 (actually0uB n X, = 0)
(6.9) <#n+N(OsB n Xn,) n OSB = 0
whereXn,, denotestheset ofpointsz suchthat p(z) E U forevery0 < i < n.
Let us describe firstthe constructionof En. We let E1 be the constant
(parallel) 2-dimensionalvectorbundleon OSB givenby E1 (z) = DpN(ro) EUw.
By (6.7) this is transversalto ESs and so, forn largeand any z E OSB n xn,/,
Dn(z) . El(z) is close to EUw. It follows that E2(r7) = DHn+N(z) El(z),
= definesa two-dimensional
(n+N) (I) bundleon Hn+N (OSB n XnH,,) n B,
withE2(r7)uniformly close to DfON(ro) .Euw ifB and I are smalland n is large.
Then, recallalso (6.9), we maytake a two-dimensional bundleF on B ofclass
C1 coincidingwith E1, E2 on theirdomainsand such that E(z) is uniformly
close to DpZ(ro) Euw; givenany 6 > 0 then,up to choosingB and I small
.

and n largeenough,we may assume that angle(E(z), D9po(ro) Euw) < 6 for .

we fix 0 < 6 < angle(DfpN(ro) Euw,EBs). Then

everyz E B. In what follows~~~~~~~~~~~~ .

we defineX = X(6, F) to be the space of continuous2-dimensionalbundles

E such that
* angle(E(r7), D~pN(ro) Euw) < 26 forevery 7 E B and
.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 243

* (n) = E(n) forevery r E B \ 9,+N (Bn xbaXn).

We also introducethe graph-transform T: X -* X givenby

f (E)(n7) = D~p'+N(Z)
=
IL+~z *-E(z),
E() zZ== ~I (n+N)(n

if r E fT+N(B n Xn,) and T(E)(r) = E(n) otherwise. Observe that by

constructioneveryE E X is transversalto ESS. It followsin a fairlyeasy way
that,at least forn large,T is well-defined (T(X) C X) and a Co-contraction.
Hence T has a unique fixedpoint,whichwe call En. Moreover,the same
type of hyperbolicity argumentas in [HP] or Section 2 shows that El is
Holdercontinuous.Thus El satisfies(a), and (c). On the otherhand, (d) is
a directconsequenceofthe transversality ofEl to ESS and the factthat most
iterationsare done insidethelinearizingneighbourhood U. A similarargument
allows us to obtain Ess. We start by taking El(r7) = ESS, for 7rE &UB,
and then proceed as forEn, only this time takingnegativeiterates: we set
E2(z) = D N(rEl((r7), r= ( (z), forz E i
As before,we extend E1, E2 to a C1 (m - 2)-dimensionalbundleF on B and
thenwe defineX to be the space of continuous(m - 2)-dimensionalbundles
E on B such that
* angle(E(z), ESs) < 26 foreveryz E B and
* E(z) = E(z) forevery z E B \ (9p(n+N)(B) n Xn
The graph-transformT: X -* X is defined by

T(E)(z) = D~p-(n+N) (n) . E(n), 7j = fTn+N (z)

if z yp(n+N)(B) Xn,,
l 4 and T(E)(z) = E(z) otherwise. Then, as in the
previouscase, we concludethat T has a unique fixedpoint,whichis a H6lder
continuousinvariantsubbundle,and we take this to be Es. Moreover,Ens
can be integratedto a n+N-invariant foliation nsThis can be seen as fol-
lows. First,whenextendingE1, E2 above we maytake E to coincidewiththe
tangentbundleof a C1 foliationF of B by (m - 2)-dimensionalsubmanifolds.
Now considerthe space X of foliationsF of B by (m - 2)-submanifolds of
class C1 such that the tangentspaces to leaves of F varycontinuouslywith
the pointand satisfy
* angle(TF(z), E') < 26 forz E B and
* Tz.F(z) = E(z) if z E B \ ( (n+N)(B) n Xn
T: X -* X givenby
We also introducethe graph-transform
T(.F)(z) = connectedcomponentof p n+N)(F(pfl+N(z))) n B containing
z,

if z E (jn+N)(B) n Xnm and t(F)(z) = F(z) otherwise. The fact that

negativeiterationsexpand the leaves of foliationsF E X assures that T is

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
244 J. PALIS AND M. VIANA

well defined,at least ifn is largeenough. Moreover,the same calculationsas

beforeshow that T is a contraction,withrespectto the C0-distancebetween
tangentbundles. It followsthat T has a unique fixedpoint _,nss and then,
clearly,we musthave TzgY4(z) = E"i,,(z) foreveryz E B. L

con-
Remark6.3. For futureuse let us also state explicitlythe following
sequencesof the renormalization techniquesabove:
(a) Given 6 > 0 we have angle(Eu(z), Do (ro) .Euw) ? 6 forevery(u, z) E
I x B, as long as B and I are small enoughand n is sufficientlylarge.
(b) Given B and I small and 6 > 0 thenforlargen, angle(E' (z), ESs) < 6
forevery(L, z) e I x B.

Keeping the same notationas before,we take B, I and 6 to be fixed

(small) and n to be large,dependingon p. In particular,we suppose 9n(J) C
I. We observe that for * = uw or ss and all the values of ,p and n under
consideration, we must have E* (P(n, [L)) = graph(f*), recall (i) and (ii)

to K
in Section5. Thus,Dpn+N E (P(n, p)) is conjugate ( ) for
some K E GL(1R2)closeto ALo(ro). We let u(n, ,u),A(n,p/)be itseigenvalues,
ja(n,f)l > jA(n,,4)I.The convergence
to (6.5) impliesthatU(n,A) is closeto
4 (if n is large and A = 9n(v), v close to -2). Hence, A(n,,4) has the same
sign as (detAu=o(ro)(Aouo)n). We claim that forthe purposeof provingour
main theoremit is no restriction to suppose that

(6.10) A(n,f) > 0 forftE an(J) and n arbitrarily

large.

This can be justifiedas follows. If (Aouo) > 0 and detA=o(ro) > 0

then thereis nothingto prove. If (Aouo) < 0 we just restrictto the n-values
havingthe appropriateparityso that detA=o(ro) (Aouo)n> 0. Suppose .

now (Aouo) > 0 and detA1,0 (ro) < 0. We note firstthat, by the arguments
developedpreviouslyin this section,thereare p-values f4n = On(vin),Vn close

to -2, forwhich P(n, fln)has homoclinictangencies.On the otherhand, in

the presentsituationwe have A(n,fln)< 0 < u(n, ^n). This means that the
sPn have homoclinictangenciesfallingin the secondofthe previoustwocases.
Moreover,it is clearthat ifthe theoremholdsforeach sOn then,since 4n-? 0,
it also holdsfor po. Therefore,we may,withoutany loss ofgenerality, assume
that (6.10) holds.
Finally,we prove(6.6) (forthe values ofn as in (6.10)). For simplicitywe
use to representequalityup to a (multiplicative)constantnot depending
on n or 6 and const to denote such a constant. Distances and angles refer
alwaysto the euclideanmetricassociatedto the coordinates((, Z) above. We
begin by introducingthe foliationsP, = ({( = const}) and u = (ON({Z =

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 245

const})) definedon B. We suppose B smallenoughso that the leavesofthese

outsidea hypersurface
two foliationsintersecttransversely
L = AdN(L,), L = {( , Z): a(fH Z) E Es}
Due to the quadraticnatureof the tangencywe even have
(6.11) angle(T,/(z), Tz(z)) dist(z,L,1) forz E B.
A fairlysimple calculation shows that LIXmay be writtenas a graph =
g(f, Z) and, moreover, p-N(P(r,-I f)) is at a distance -n of LX,in
{( > g(py,Z)} or{( < g(, Z)} dependingon whetherb is positiveor negative.
It followsthat P(n, -) is placed at a distance d IoKln of L,, to the side
determinedby the vector (0, A) (recall (6.5)). Now we let S be the (d/2)-
neighbourhoodof P(n, f) in its stable manifold.For n sufficiently
large
(6.12) angle(TzS,ES) < b la-n forz E S.
This is a direct consequenceof the fact that the (local) stable manifoldof
P(n, lt) = 0L l(P(n, ii)) converges to a hyperplane {x = const} as n -* oc,
togetherwith the formof the coordinatechanges (6.4). Note also that, by
construction,
dist(z,LI,) > constla/l-n forz E S.
We take EW to be the line fieldon S givenby Ew(z) = E' (z) n TzS. Re-
mark6.3(a) and (6.12) implythat EW is almost colinearwith (0, A): if n is
large then angle(EW(z),(0, A)) < 26 forz E S. We orientEW in such a way
that EW. (0, A) > 0 and foreach 71oE FO = WsS(P(n,At)) n s we define
Wo(rqo)to be the positivetrajectoryof ?jounder Ew. Now we considerthe
set r D FO of points ?q= f_(n+Nq)(70) with no E Fo and H0-(i+?N)(71o)E U for
0 < i < n. We point out that r n B containsa neighbourhoodof P(n, pi)
in W5s(P(n,p)) with fixedradius diam(B). This may be seen fromthe
followingremarks: (i) (p-N(ro) containsa neighbourhoodof V ( ))
in Wss( rN(P(n,,u))) with radius -%K lu- and, by (6.7), it is transversal
to EUw at everypoint; (ii) the expansionduringthe n (negative) iterations
inside U is > constlA,11Th ? jojn. For each i > 1 and rq E r, we define
Wi(7/) = tn+N) (Wil((47+NW))). Since Ew is (+)-invariant and A(n,At)
is positive,we have Wj(rq)D Wi-,(rq) foreveryi > 1 and rqE r. We claim
that, as long as they stay inside B, the Wi(,q) remainnearlycolinearwith
(0 A):
(6.13) angle(TzWi(r), (0, A)) < 26 forz E Wi(rI), i > 1, 71 E r
In order to prove this we note first that TzWi(ri) C E1'(z) and thus, by
Remark 6.3(a), angle(TzWj(rq),DfpN(ro)E9w) < 6. Therefore,(6.13) will

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
246 J. PALIS AND M. VIANA

followifwe prove

(6.14) (r), E) < 6 |a,,

angle(TzWi forz E Wi(r), i > 1, EF.

We do this in an inductiveway. Let i > 1 and suppose that (6.14) holds for
every T(Wi1(rqo), ro E Fo C F and ( E Wi_1(rio)(note that for i = 1
this is contained in (6.12)). Let us also point out that, due to the way
we have orientedEW, the Wj(rio), j = 0,1,... ,i - 1, grow in the direc-
tion opposite to L,. Thus, we have dist((,L,) > constlIo~-n for every
( E Wi_1(rio),rioE Io. In view of (6.11) and the inductionhypothesis
this impliesangle(T(Wi_1(ro),TF/u(()) > constlujl-n. Then, angle(D~p-N
TWi_1(,qo),Eu) > constla/j-n and so angle(D(p +N) T(Wii(r10),EU) ?
constIA,1jn. This means that angle(TzWi(rq),Es) < IA/,n< l for
z = 9jn+N)(,) 71 = 7In+N)(tq0), and so the proofof (6.13), (6.14) is com-
plete. In particular,foreach rqE r n B the curve Ui>1Wi(rq)containsa
segmentW(rq) nearlycolinearwith (0, A) and connectingrjto the boundary
of B. We just let Es(n, M) = U17ErnBW(T1)C WS(P(n, p)) and then (6.14),
Remark6.3(b) and the factthat P(n, fi) -* qo implythat these Es(n, ,) con-
verge,in the C1 sense,to a domain E' C WS(p) containingqo in its boundary.
to checkthat,givenany segmentof Wu(p) intersect-
Finally,it is not difficult
ing transversely the connectedcomponentof WS(p) \WSs(p) that containsqo,
its positiveiterateseventuallyintersectEs transversely.

7. Proof of the main result

Finally,we explainhowtheideas and resultsin theprevioussectionsfitto-
getherto proveour main theorem.We startwitha generalC2 diffeomorphism
(o witha homoclinictangencyassociated to a sectionallydissipativesaddle p.
We showhowto obtain,aftera certainnumberofC2-smallperturbations of(,
a new diffeomorphism containedin the closureof an open set K C Diff2(M)
exhibitingpersistenthomoclinictangencies. For the sake of clearnesswe di-
vide the proofintofoursteps. In ordernot to overloadthe notationwe denote
all perturbeddiffeomorphisms by p; on the otherhand we state carefullythe
propertiesobtainedaftereach perturbation.

Step 1. Up to a firstperturbation, we may suppose that (o satisfiesthe

(generic) assumptionsof Section 5: The homoclinictangencyis quadratic
(and quasi-transversal);the saddle point p is C2-linearizableand has either
1 (real) or 2 (complex) weakest contractingeigenvalues;finally,the trans-
versalitycondition(5.1) holds. Then, proceedingas in Section5, we obtain a
diffeomorphism o arbitrarilyclose to p, exhibitinghomoclinictangenciesasso-
ciated to a periodicsaddle p which(besides beingalso sectionallydissipative,

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 247

recall Remark5.3), has a unique least contractingeigenvalue. In the sequel

we stilldenoteby (o and p such o and P, respectively.

Step 2. We may again suppose that the saddle p is C2-linearizableand

the homoclinictangencyis quadraticand satisfies(5.1). Also, we mayassume
that the pointoftangencyis not containedin the strongstable manifoldofp.
Then fordiffeomorphisms arbitrarilyclose to (p,p has transversehomoclinic
intersections togetherwith a new homoclinictangency.Note that, by conti-
nuity,these transverseintersections are also outside WSs(p) and theysatisfy
(3.2). Hence, by Section 3, p belongs to a nontrivialbasic set Al with in-
trinsicallyC1 unstablefoliation.Moreover,the resultsof Section 4 apply to
Al since,afterStep 1, we have a unique weak-stabledirection.In particular,
Tru(Al) is strictlypositiveand remainsboundedaway fromzero underfurther
(small) perturbations.

Step 3. The newtransverseand nontransverse homoclinicorbitsare con-

structed(Step 2) in such a way that theyinvolvethe same connectedcom-
ponentof Ws(p) \ Wss(p) and of Wu(p) \ {p}. It is also clear that we may
continueto suppose that the saddle p is C2-linearizable.Moreover,we may
assume that thisnew homoclinictangencyis again quadraticand satisfiesthe
genericcondition(6.7): Otherwisewe just replace soby some nearbydiffeo-
morphismforwhichthis holds. Then, by genericallyunfoldingthis tangency
as in Section6, we obtain arbitrarily of s0exhibitingbasic
small perturbations
sets Al and A2 such that: rU(Al)rs(A2) > 1 and

* the leaves of Wu(Al) (resp. Ws(Al)) have transverseintersections

with
those of WS(A2) (resp. WU(A2));
* thereare periodicpointsP1 E Al, P2 E A2 suchthat Wu(pl) and WS(p2)
also have a pointq of nontransverseintersection(i.e., a tangency).

Step 4. Let U2 be a neighbourhoodof A2 such that WS(A2) admitsan

extensionto a C1 foliationR2 definedon U2. By C1 we mean here that the
tangentspaces to the leaves TFP(z) varyin a C1 fashionwiththe point z.
Clearly,we maytake q to belongto U2 and thenan implicitfunctionargument
allows us to definethe "line"oftangenciesbetweenthe leaves of WU(Al) and
P near q. This is based on the following:

LEMMA 7.1 (Implicitfunction).Let X C iRm be compactand I C R11

be a compactinterval.Let F: X x I -J R be intrinsically
C1 and (xoto) E
X x int(I) be such that

(7.1) F(xoto) = 0 and AFx,(to,to) $ O.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
248 J. PALIS AND M. VIANA

Then thereexist V C X a compactneighbourhood of xo and a unique intrin-

sicallyC1 map f: V -* I such thatf(xo) = to and F(x,f(x)) = 0 for every
xEV.

Proof. Let a > 0, Vo a neighbourhood of xo and 6 > 0 (small) be such

that IAFx(s, t)1 > a forevery x E Vo and s, t E [to-6, to+6]. We take V C Vo
so that IF(xto)I <? a6/2 for every x E V. Then, for x E V, F(xto - 6).
F(x, to + 6) < 0 and so there is f (x) E [to-6, to + 6] such that F(x, f (x)) = 0.
Moreover, f (x) is unique since AFx is never zero on [to- 6, to + 6]. The same
kind of argument shows that the function f: V -* I defined in this way is
continuous. Finally, Af(x, z) = -AFf(x)(x, z)/AFz(f(x), f(z)) (where we
use the same notation as in Lemma 2.5) is an intrinsicderivative for f. D

In order to apply the lemma we firstfix U C U2 a small neighbourhood of q

and &, a C' vector fieldon U orthogonal to the leaves of F2. By Proposition 3.5,
WU(Al) n U contains an intrinsicallyC1 diffeomorphicimage Y of X x I where
I is a compact interval and X is a small compact neighbourhood of P1 in
Al n Wls(pl). We let (u be some intrinsicallyC1 vector field on Y tangent
to the leaves of WU(Al) and then we define F(y) = (u(y) -s(y). Hypothesis
(7.1) in the lemma corresponds to having a quadratic tangency at q and, up
to considering an additional perturbation of p,we may assume this to be the
case. As a conclusion we get that there exist V, a compact neighbourhood
of P1 in Al n Ws,(pl) and irl: V1 -* WU(Al) n U an intrinsicallyC1 map
such that each rl(x), x E V1, is a point of tangency between WU(x) and
some leaf of 2. We also introduce 7r2: U -* Wlu c(p2), the projection along
the leaves of 2 onto Wluc(p2) which we identify with an interval in R11(via
some C1 diffeomorphism). If necessary, we perform a last perturbation of
,o so that AZrl (pl,pl) ITp1(Al n Ws, (pi)) is not tangent to the stable leaf
42(q). Then (7r2o i1) is intrinsicallyC1 and i\(7r2o 7r1)(pl,pl) is bijective.
We let K1 = (7r2o i1n)(Vi) and K2 be a small compact neighbourhood of P2 in
A2 n Wulo,(P2)and then

* K1 and K2 intersecteach other at P2;

* T(Kip2) oT(K2, P2) = rU(Al) Tr(A2) > 1.

Now from the gap lemma and the continuous variation of the thickness, we
get that for a whole C2-open set K of perturbations of o the corresponding
sets K1, K2 intersecteach other, which corresponds to heteroclinic tangencies
involving Al and A2. Since Al and A2 are heteroclinically related, it follows
in an easy way that given any periodic point P in Al U A2 (e.g., j = p) a
dense subset of elements of K exhibits homoclinic tangencies associated to
(the analytic continuation of) P. As explained in Section 1, this implies that

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
 HIGH DIMENSION DIFFEOMORPHISMS 249

residuallyin K the maps have infinitelymanycoexistingsinks. This completes

our proof.
On the otherhand, the parametrizedversionof the theoremnow follows
by the fact that, given a genericone-parameterfamilyof diffeomorphisms
passingthrougha homoclinictangency,thepreviousargumentscan be carried
out makinguse only of perturbations along theparameter:At each step the
new, perturbed,diffeomorphism is taken belongingto the initialfamily.We
just observea fewpointsin this direction.First of all, the perturbationde-
scribed in Section 5 is of that kind and so, as before,we may assume that
the periodicpointassociatedto the homoclinictangencyhas a uniqueweakest
contracting eigenvalue.Of course,in performing sucha perturbationalongthe
parameterline we also have to assure that all the conditionsconcerningthe
initialhomoclinictangencyare still valid forthe new one. Throughanother
small perturbationwe obtain a basic set Al as in Section3, witha homoclinic
tangencyassociated to a periodicpoint pi E A1 satisfyingall the previous
transversality conditions.Then, again as before,an additionalsmall pertur-
bation inside the parametrizedfamilyyields a basic set A2 as in Section 6:
It has large stable thickness(the productwith the unstablethicknessof A1
is greaterthan 1) and, forsome periodicpoint p2 E A2, p1 and P2 are het-
eroclinicallyrelated (mutual transverseintersections betweentheirinvariant
manifolds)and moreoverWu(pi), WS(p2) also exhibitan orbitof tangency.
The last part of the proofis thensimilarto that of the main theorem.

INSTITUTO DE MATEMATICA PURA E APLICADA (IMPA), RIO DE JANEIRO, BRAZIL

UNIVERSIDADE DO PORTO, PORTUGAL

REFERENCES

[BC] M. BENEDICKS and L. CARLESON, On iterationsof 1-ax2 on (-1, 1), Ann. ofMath. 122
(1985), 1-24.
[HP] M. HIRSCH and C. PUGH, Stable manifoldsand hyperbolicsets,Proc. A.M.S. Symp.Pure
Math., 14(1970), 133-163.
[MV] L. MORA and M. VIANA, Abundanceofstrangeattractors,Acta Math. 171(1993), 1-71.
[Ni] S. NEWHOUSE, Non-densityof Axiom A(a) on S2, Proc. A.M.S., Symp. Pure Math. 14
(1970), 191-202.
[N2] , Diffeomorphismswithinfinitelymanysinks,Topology13(1974), 9-18.
[N3] , The abundance of wild hyperbolicsets and nonsmoothstable sets fordiffeo-
morphisms,Publ. Math. I.H.E.S. 50(1979), 101-151.
[PT] J. PALIS and F. TAKENS, Hyperbolicity and Sensitive-ChaoticDynamics at Homoclinic
Bifurcations, CambridgeUniversity Press, 1992.
[Rob] C. ROBINSON, Bifurcationto infinitelymanysinks,Comm. Math. Phys. 90(1983), 433-
459.
[Rom] N. ROMERO, Persistenceofhomoclinictangenciesin higherdimensions,to appear Ergod.
Th. and Dynam. Sys.

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions
250 J. PALIS AND M. VIANA

[Sm] S. SMALE, Diffeomorphisms withmanyperiodicpoints,Differential and Combinatorial

Topology, PrincetonMath'l Series,PrincetonUniv. Press (1965), 63-80.
[St] E. STEIN, Singular Integralsand Differentiability
Propertiesof Functions,Princeton
Math'l Series,PrincetonUniversityPress, 1970.
[TY] L. TEDESCHINI-LALLI and J.A. YORKE, How oftendo simple dynamicalprocesseshave
manycoexistingsinks?,Comm. Math. Phys. 106(1986), 635-657.
infinitely
[V] M. VIANA, Strangeattractorsin higherdimensions,Bull. Braz. Math. Soc. 24(1993),
13-62.
[W] H. WHITNEY, Analyticextensionsofdifferentiable functionsdefinedin closedsets,Trans.
A.M.S. 36(1934), 63-89.
[YA] J.A.YORKE and K.T. ALLIGOOD, Cascades ofperiod-doubling bifurcations:
a prerequisite
forhorseshoes,Bull. A.M.S. 9(1983), 319-322.

(Received November10, 1992)

This content downloaded from 147.65.4.59 on Tue, 9 Jul 2013 10:14:33 AM

All use subject to JSTOR Terms and Conditions