Partial hyperbolicity and robust transitivity - Project Euclid

Acta Math., 183 (1999), 1-43 (~) 1999 by Institut Mittag-Leffler. All rights reserved

Partial hyperbolicity and robust transitivity by LORENZO J. DiAZ

ENRIQUE R. PUJALS

PUC-Rio Rio de Janeiro, Brazil

Univ. Federal do Rio de Janeiro Rio de Janeiro, Brazil

and

RAiJL URES Univ. de la Rep(tblica Montevideo, Uruguay

1. I n t r o d u c t i o n Throughout this paper M denotes a three-dimensional boundaryless compact manifold and Diff(M) the space of gl-diffeomorphisms defined on M endowed with the usual Cl-topology. A ~-invariant set A is transitive if A=w(x) for some xEA. Here w(x) is the

forward limit set of x (the accumulation points of the positive orbit of x). The maximal invariant set of ~ in an open set U, denoted by A~(U), is the set of points whose whole orbit is contained in U, i.e. A ~ ( U ) = ~ i e z ~i(U). The set A~(U) is robustly transitive if Ar is transitive for every diffeomorphism r CLclose to ~. A diffeomorphism ~EDiff(M) is transitive if M=w(x) for some xEM, i.e. if A ~ ( M ) = M is transitive. Analogously, ~ is robustly transitive if every r gLclose to also is transitive, i.e. if A ~ ( M ) = M is robustly transitive. In this paper we focus our attention on forms of hyperbolicity (uniform, partial and strong partial) of a maximal invariant set A~(U) derived from its robust transitivity. Observe that U can be equal to M, and then ~ is robustly transitive. On one hand, in dimension one there do not exist robustly transitive diffeomorphisms: the diffeomorphisms with finitely many hyperbolic periodic points (Morse~ Smale) are open and dense in Diff(S1). On the other hand, for two-dimensional diffeomorphisms, every robustly transitive set A~(U) is a basic set (i.e. A~(U) is hyperbolic, transitive, and the periodic points of ~ are dense in A~(U)). In particular, every robustly transitive surface diffeomorphism is Anosov and the unique surface which supports such diffeomorphisms is the torus T 2. These assertions follow from [M3] and [M4]. In dimension bigger than or equal to three, besides Anosov (hyperbolic) diffeomorphisms there are robustly transitive diffeomorphisms of nonhyperbolic type. As far as we know, three types of such diffeomorphisms have been constructed: skew products, This work is partially supported by CNPq, PRONEX-Dynamical Systems and FAPERJ (Brazil), and CSIC and CONICYT (Uruguay).

L.J. DIAZ, E.R. PUJALS AND R. URES

Derived from Anosov, and deformations of the time-T map X~ of the flow of a transitive Anosov vector field X. Before describing these examples let us recall that the standard Derived from Anosov

(DA) diffeomorphisms, defined on the two-torus T 2, are obtained via saddle-node bifurcations of Anosov systems: the unfolding of the bifurcation leads to structurally stable maps (the DA-diffeomorphisms) whose nonwandering set is a source and a nontrivial hyperbolic attractor, see [Sm] and [W]. This two-dimensional construction can be carried to higher dimensions to get DA-diffeomorphisms which are robustly nonhyperbolic and transitive, see [M1] and [C]. Chronologically, the first examples of nonhyperbolic robustly transitive diffeomorphisms were skew products. Such diffeomorphisms were constructed in the four-dimensional torus T 4 = T 2• 2 by perturbing the product of a DA-diffeomorphism and an Anosov one, see [Sh]. Nowadays we also know that one can perturb the product of any diffeomorphism O having a hyperbolic transitive attractor Ar and the identity Id on any compact manifold to get G Cl-close to 9 x Id with a robustly nonhyperbolic transitive attractor Ac. Moreover, A t = A t ( U )

for some neighbourhood U, and A t ( U ) is robustly

transitive. In particular, if 9 is Anosov (i.e. A r transitive, see [BD1].

then the perturbation G is robustly

All robustly transitive diffeomorphisms mentioned above (skew products and DAmaps) are nonisotopic to the identity, but there also are robustly nonhyperbolic transitive diffeomorphisms isotopic to the identity: Given any transitive Anosov vector field X let X~ be the flow of X at time T. Then one can perturb X~ to obtain a robustly transitive diffeomorphism, see [BD1]. In dimension bigger than or equal to three, besides the constructions above, one can also obtain robustly nonhyperbolic transitive sets (of semilocal nature) via cycles containing periodic points of different indices (dimension of the stable manifold), see [Dill, [Di2] and [DR]. The nonhyperbolic transitive sets A~ (U) quoted above always contain periodic points with different indices and coincide with the closure of their transverse homoclinic points (i.e. the transverse intersections between the invariant manifolds of a periodic point). The previous examples fit into the category which we call strong partially hyperbolic (see the definition below): there is a D~-invariant partially hyperbolic splitting of Th~,(u)M= E 8| ~ with three nontrivial bundles, where E 8 and E ~ are hyperbolic directions (contracting and expanding, respectively) and E c is a nonhyperbolic central direction. On the other hand, recently, see [B] and [BV], there have been constructed examples of robustly transitive diffeomorphisms which do not admit three nontrivial invariant bundles (i.e. either E s or E ~ above is trivial).

PARTIAL HYPERBOLICITY AND ROBUST TRANSITIVITY

It is important to mention t h a t in this paper we are only concerned with transitive sets which are locally maximal. Notice t h a t one can also define transitive sets (in a robust way) as follows: given a hyperbolic saddle P of ~, for every r close to ~ define Er as the closure of the transverse homoclinic points of PC (Pc is the continuation of P). Such sets are transitive, but in general they fail to be locally maximal: in some cases sinks or sources accumulate to EC, see [BD2]. Even more, they do not admit any nontrivial Dr

splitting, see [BD2] and the constructions in [DU]. Our goal here is to characterize the forms of possible hyperbolicity for a maxi-

mal invariant set A~(U) which is robustly transitive.

We prove that, in the case of

three-dimensional compact manifolds, the robustly transitive sets A~(U) are generically

partially hyperbolic. Now let us state precisely our results. We begin by giving some basic definitions. Let be a diffeomorphism and A a ~-invariant set. A splitting T A M = E |

is dominated if

E and F are D~-invariant and there are constants m > 0 and K < 1 such that D m -1 II(D~m)lE~ll.ll((x~ ) )IF~(~>II 0 and 0 < A < I such that for every n > 0 one has

IlDx~n(v)[[ ~C~llvll (resp. IID~-~(v)ll 0 and 0 < A< 1, such that

< CAnllv*]] .llvcl], IIO o ( >

-n(vD II9IIO

n(v )ll

C nllvUll. IIv ll,

for all n > 0 , viEEix, i=s,c,u. A partially hyperbolic set A~ expands (resp. contracts) volume in the central bundle if E c is volume-expanding (resp. -contracting). By a volume-expanding bundle F of A~, we mean a Dqo-invariant bundle F such that there are constants C > 0 and a > l that

such

IJaCF(x)(~k)l >Ca k for all x c A ~ , k~> 1, where JaCF(x) ~ denotes the Jacobian of ~ in the bundle F at x. We say that a Dqoinvariant bundle F is volume-contracting if it is volume-expanding for ~-1. As we have mentioned, a partially hyperbolic set can also be hyperbolic. Here, to avoid misunderstandings, we adopt the following convention: the partially hyperbolic sets we consider are genuinely partially hyperbolic, meaning that their central directions are nontrivial and nonhyperbolic. Given y)ET(U) the set A~,(U) has robustly real eigenvalues if there is a gl-neigh bourhood b/~, of ~ such that for every r eigenvalues of D p r n

are

and every periodic point PEAr

all the

real (n is the period of P). Consider the subset P(U) of T(U) of

diffeomorphisms ~ such that A~, (U) has robustly real eigenvalues and hyperbolic points of different indices (i.e. A~,(U) is not uniformly hyperbolic). When U=M we let 7)='P(M). THEOREM C. The set A~(U) is strong partially hyperbolic for all qocP(U)Ncl(U). In the case of transitive diffeomorphisms we have a stronger version of the previous result:

L . J . D I A Z 1 E . R . P U J A L S A N D R. U R E S

COROLLARY D. Let ~E7). Then ~ is strong partially hyperbolic. The existence of periodic points with complex (nonreal) eigenvalues prevents the existence of a splitting having three nontrivial directions (recall that we are considering three-manifolds).

Theorem C means that the existence of such points with complex

eigenvalues is the unique obstruction for the strong partial hyperbolicity.

The next

theorem says that if the nonhyperbolic set A~(U) has complex eigenvalues then it satisfies a stronger form of partial hyperbolicity: the central bundle is either volume-expanding or -contracting. Let ~ c T ( U ) . The set A~(U) has complex eigenvalues if there is some periodic point P E A ~ ( U ) such that Dpy~ n has two eigenvalues with the same modulus (n is the period of P). We denote by ];(U) (resp.)2) the subset of T(U) (resp. 7-) of diffeomorphisms such that A~(U) (resp. ~) is not uniformly hyperbolic and has complex eigenvalues. THEOREM E. Let ~ be a diffeomorphism in A(U) that can be approximated by

diffeomorphisms in ~)(U). Then the central bundle of TA~(u)M is two-dimensional and volume-expanding/contracting: if TA~(u)M=ES@E cu then E cu is volume-expanding, and if T A ~ ( u ) M = E ~ @ E ~ then. E c~ is volume-contracting. In the case of a transitive diffeomorphism Theorem E can be read as COROLLARY F. Let ~ET- be a diffeomorphism which can be approximated by diffeo-

morphisms in ~;. Then ~ is partially hyperbolic and volume-expanding/contracting in the central bundle. Note that since Av(U) is not uniformly hyperbolic it contains points of indices one and two. Our proof shows that all periodic points with complex eigenvalues have the same index. Finally, the following corollary gives the connection between the absence of strong partial and uniform hyperbolicity and the approximation by homoclinic tangencies. Recall that a hyperbolic periodic point P has a homoclinic tangency at x if the invariant manifolds of P have a nontransverse intersection at x. COROLLARY G. Let ~C~4(U) be such that A~(U) is neither strong partially hyper-

bolic nor uniformly hyperbolic. Then ~ can be approximated by some r with a homoclinic tangency associated to some hyperbolic periodic point in Ar Let us observe that in dimension bigger than two the existence of homoclinic tangencies does not lead to creation of sinks or sources, and thus homoclinic tangencies are not an obstruction for transitivity. We remark that Corollary G can be formulated in the case of transitive diffeomorphisms.

PARTIAL HYPERBOLICITY AND ROBUST TRANSITIVITY

In view of the results above, let us summarize the different types of robustly transitive sets A~(U) in three-manifolds. For that let P(Q) denote the set of periodic points of Q in A~(U). We also consider the subsets PR(Q) (resp. Pc(Q)) of P(Q) of points having only real eigenvalues of different moduli (resp. having two eigenvalues of the same modulus, this case including periodic points with eigenvalues of multiplicity bigger than one and, obviously, periodic points with complex (nonreal) eigenvalues). (1) Suppose that A~(U) is hyperbolic. Then Pc(Q) is robustly empty if and only if for every r Cl-close to Q the set Ar has a hyperbolic splitting with three nontrivial directions. (2) Suppose that A~(U) is robustly nonhyperbolic. Then 9 A~(U) contains (robustly) points of indices one and two, 9 A~(U) is (robustly) nonstrong partially hyperbolic if and only if Q can be approximated by diffeomorphisms r with P c ( r 1 6 2 9 A~(U) is robustly nonstrong partially hyperbolic if and only if Q can be approximated by a diffeomorphism r with a homoclinic tangency (associated to some point

of P(r As we have mentioned, the unique surface which supports robustly transitive diffeomorphisms is the two-torus. This means that Cat least for surfaces) the existence of such transitive diffeomorphisms gives some topological information about the surface. For higher dimensions we would like to know if it is possible to deduce some topological information about the ambient manifold M from the existence of robustly transitive diffeomorphisms. In the case of three-manifolds, we study the connection between the existence of transitive diffeomorphisms in M and the growth of the fundamental group of M. As an application of Theorem B, we obtain an obstruction for the existence of robustly transitive diffeomorphisms on manifolds with finite fundamental group. The formulation of this obstruction depends on the integrability of the central bundle: note that, to the best of our knowledge, it is an open question whether the central bundle is necessarily integrable, even in the simplest case of three-manifolds. Let E~(Q)| i=s or u, be a partially hyperbolic splitting of M for QcDiff(M), where Ec(Q) has dimension two. The splitting is dynamically coherent if there exists a foliation 9re(Q) tangent to E~(Q). Notice that, by the hyperbolicity, Es(Q) or E"(Q) (according to the case) is integrable, and then one can define the stable/unstable foliation ~i(Q), tangent to Ei(Q), i=s, u. THEOREM H. Let M be a three-dimensional boundaryless compact manifold. Suppose that M supports a robustly transitive diffeomorphism having a dynamically coherent splitting. Then the fundamental group ~h(M) is infinite.

L.J. DIAZ, E.R. PUJALS AND R. URES

Let us say a few words about the organization of this paper. The main step of our constructions is the following preliminary result: THEOREM 1.1. There is a residual subset T~(U) of T(U) such that for every

~ET~(U) the set A~(U) has a partially hyperbolic splitting TA~(u)M=Ei(~)| i=s or u, where Ei(~) is one-dimensional and uniformly hyperbolic. We give an outline of the proof of this theorem in w In w we introduce the different types of perturbations that we use in this paper (perturbation of the derivative and creation of cycles). Theorem 1.1 is proved in w which is the main and the longest section of this paper. This section is divided in three parts: estimates on the eigenvalues (w

angular estimates of the bundles (w

splittings (w Theorem 1.1.

and construction of uniformly dominated

Finally, in w we prove the theorems in this introduction by using

Acknowledgments. The authors are grateful to Ch. Bonatti, J. Palis, M. Sambarino, M. Shub, and M. Viana for many useful and encouraging conversations. The authors acknowledge the warm hospitality of IMPA (Rio de Janeiro, Brazil), Departamento de Matem~tica of PUC-Rio (Rio de Janeiro, Brazil) and IMERL (Montevideo, Uruguay) while preparing this paper. Finally, we also thank the referee's suggestions for improving the presentation of this paper.

2. O u t l i n e o f t h e p r o o f of T h e o r e m 1.1 To explain the main ideas and difficulties of the proof of Theorem 1.1 (actually, the key result in this paper) let us begin by saying a few words about a stronger two-dimensional version of our result. From now on fix the open set U and denote by P ( ~ ) the set of periodic points of ~ in U, and by PR(~) the subset of P ( ~ ) of periodic points having all eigenvalues real and different in modulus. THEOREM ([M3]). Every Cl-robustly transitive set A~(U) of a surface diffeomorphism ~ is a basic set (hyperbolic, locally maximal, and with dense periodic points). Let us assume that A~(U) is infinite, otherwise, as we have mentioned in the introduction, the result is immediate. To prove the result it is enough to see that P ( ~ ) is robustly hyperbolic, or equivalently (due to the fact that we are in dimension two) that the number of sinks and sources is finite and constant in a neighbourhood of ~. From the transitivity and since we are assuming that A~(U) is infinite, in our case this number is zero. Arguing by contradiction, if P ( ~ ) is not hyperbolic then one gets an elementary

PARTIAL HYPEI~BOLICITY

AND I~OBUST TRANSITIVITY

bifurcation of some periodic point (saddle-node, flip or Hopf). In dimension two, such bifurcations lead to the creation of new sinks or sources, contradicting the fact that the number of sinks and sources is locally constant. We also observe that in dimension two homoclinic tangencies generically lead to the creation of sinks or sources, see [PV]. Thus, in the case of surface diffeomorphisms, such bifurcations are also forbidden. In higher dimensions the examples quoted in the introduction show that a robustly transitive set A~(U) can be nonhyperbolic and its periodic points can bifurcate. Moreover, one can also have homoclinic tangencies. Actually, the main difficulty in the proof of Theorem 1.1 arises from the fact that in dimension three the list of forbidden bifur-

expansive/dissipative homoclinic tangencies (i.e. homoclinic tangencies associated to periodic

cations of points in A~(U) is rather limited: Hopf bifurcations and sectionally

points such that the modulus of the product of any pair of eigenvalues is bigger/less than one).

Let us observe that, for example, sectionally dissipative homoclinic tangencies

imply the creation of sinks, see [PV], and thus they are forbidden in our context. The proof of Theorem 1.1 is by contradiction: assuming that A~(U) is not partially hyperbolic we create either a sink or a source in U. Since A~(U) is infinite this contradicts its robust transitivity. Let us now be much more precise and sketch some key ideas and ingredients of our proofl An important difficulty in the proof is to find a suitable candidate for the role of D~-invariant splitting over A§

For that we first restrict our attention to the

diffeomorphisms ~) such that PR(~) is dense in A~(U), and prove that such diffeomorphisms are generic in T(U) (see Lemma 4.2). For points PcPR(~) there is a splitting TpM=E~,OE~,OE~ with three nontrivial directions (E~ is the eigenspace associated to the eigenvalue Ai(P), where ]As(P)[~no

Since the periods tn tend to infinity and every pn is hyperbolic, for each n there is a neighbourhood Vn (r of r such that

P)C(r

for

(4.4)

We claim that there are diffeomorphisms (,~-~r and points P~EP~'n(~n) such that

Clearly, the lemma follows from the claim: applying Lemma 3.1 to gn we obtain a new I sequence ~--~r such that and o z (, E S [ c t "~ c u So it remains to prove the claim. We argue by contradiction. Assume that the claim is false. Then for a fixed n0 (big) we get # > 0 such that IAc(Pr r

(1+#) k,

where k is the period of PC,

for all r close to r and every PC c ( P ~ ) l ' n ~ 1 6 2 This assertion follows from Lemma 3.1 by arguing as in the proof of Lemma 4.5. Now, recall that ]A,(Pn,r t~ and IA~(Pn, Cn)l > (1 +6)t~ (see Lemma 4.6). By Lemma II.9 in [M3] (which, in a few words, asserts that robust hyperbolicity of periodic points implies that the angles between the stable and unstable bundles are bounded away from zero) there is 3,>0 such that a(E~(r

E pr ~u

>7

for all r close to r and every PcE(P~)I'n~

PARTIAL HYPERBOLICITY AND ROBUST TRANSITIVITY

21

Taking n>l/~/one has P ~ n ( r for all %bclose to r contradicting (4.4). This ends the proofs of the claim and of the lemma. [] The next step in the proof of the proposition is to get a saddle-node periodic point R such that a(E~({), E~({)) is small. We use the following claim whose proof we postpone until the end of this section. CLAIM 4.12. Suppose that r and that Pn and Qn are the periodic points of Cn in (4.3), Pn and Qn with index 2. Then there are sequences of diffeomorphisms (r r and r162 and of points RnEPR(r of period kn, such that (1) max{a(E~{~(r E ~ ( r a(E~(r E~.(r ~ 0 such that ao(E*(B),Er

is defined for every B close to 7)(r for every 13 close to 7)(r

Proof. We argue by contradiction. Suppose that for every C > 0 there is B close to 7)(r such that ao(E*(B),E~(B))C for every family of

PARTIAL

HYPERBOLICITY

periodic linear maps B &close to T)(r and that the first possibility holds:

AND ROBUST

29

TRANSITIVITY

In the sequel let us assume that we have fixed r

ao(ES(B), Ec~(B)) > C

for all B &close to 7:)(r

(4.6)

Our goal is to prove that if (4.6) holds and E s (r E c~ (r is (hyperbolically) dominated on the period (see Lemma 4.24 below) then the splitting is uniformly dominated. Let us begin with the following two-dimensional lemma about perturbations of linear maps. LEMMA 4.20. Consider sequences of (2 x 2)-diagonal matrices {(A~,m)~_l }-~>o,

'

bi,m '

such that (a) I-Ii~l Ai,m=A-~ =Id:~ for every m, (b) there is a constant c>0 such that c-l 0 there is rno such that for every m ) r n 0 there are families of triangular matrices (Ai,m)i=z m satisfying (i) [[Ai,m-Ai,m[[~0 to be determined later. Arguing inductively and m m bearing in mind that l-L=1 ai,m = 1 and I-Ii=1 bi,m =-t-1, a straightforward calculation gives

(;)

m

(10)

30

L . J . D I A Z , E . R . P U J A L S A N D R. U R E S

where

*m ~-- E

ai,m 6,,~6j,m

j=l

"

bi,m = "i=j+l

' j=l

= 5m

aj,m

"i=j+l

~,m . .

aj,~ j=l

and m

j,m= 1-I b',m i=j+l

ai,m

We choose (inductively) the 6j,m such t h a t

,m Similarly, we have

m

j-1

j=l

--

(~m~j,rn i--

1

j=l

'

ai,m z

1

=Sine

bj,m

j=l

and m

ej,,.

^ Cj,m =

"1-1" ai, m H bi,m" i=j+l

T h a t is, l~3,ml = 1~3,m1-1. Notice t h a t , by construction,

-

-

aj,m

>0

and

aj,m

'

'

bj,rn

'

>0.

Consider now the sums

j=l

Since the lay,m[ and

j=l

bj,m

[bj,m[ are bounded, we have t h a t these sums cannot be b o u n d e d

simultaneously. Suppose, for instance, t h a t the first one is not bounded. T h e n there is m such t h a t Sr~ >

2x/e.

Observe t h a t ~,,~=6,~Sm. T h u s taking

5mC(x/Sm, 89 we have

PARTIAL HYPERBOLICITY

31

AND ROBUST TRANSITIVITY

Clearly, by the definition of -dj,m, J=:

(;) (10)

J=:

(0)( ) 4-1

'

By the definition of 5m one gets II_~j,,~-Aj,,~II < : . Now it is enough to take Zij,m=fij,,~. This completes the proof when Sm is not bounded. Finally, if Sm is bounded then Sm is not bounded. Then the proof of the lemma follows as before by considering lower triangular matrices instead of the upper triangular ones. [] We have the following reformulations of the previous lemma and of Proposition 4.7 in terms of cone fields whose proofs are immediate. LEMMA 4.21. Let ( (Ai,m)~m:)m>~O be families of linear maps satisfying the hypothe-

ses of Lemma 4.20. Consider any pair of cones Ch and Cv around i=(1, 0) and j = ( 0 , 1), respectively, such that i~C v and j~C h. Then the perturbations (Ai,m) in Lemma 4.20 can be taken satisfying (a) either ]-Lm=l.di,m(Ch) CC v, (b) or H~=:Ai,~(C")CC h. LEMMA 4.22. There are ~-,6>0 such that for every family of linear maps 13 b-close t o / ) ( r it holds ES(BP) ~ C~(ECU(Bp))

for every PEPR(r

Here C~(F) denotes the cone field of size ~- around F.

Denote by P~(r

= { P 6 PR(r

of period m ~>n}.

PROPOSITION 4.23. Assume that hypothesis (4.6) holds. Then there is no such that the D@-invariant splitting (E~(@)OE~"(r162 is uniformly dominated for every CeTZ(U) close to r

Proof. We first prove the dominance in the period for periodic points: LEMMA 4.24. Under the hypothesis of Proposition 4.23 there are A, 0 < A < I , and no such that

IIDxCnlE~(r II'NDr162162

< An

for every x 6 P~t~

of period n ~ no.

Proof. We argue by contradiction and suppose that the lemma is false. Then there are sequences of points Pm6PR(r Pm of period n,,, and of increasing numbers (k,~)---*l- such that [[D p~r ~~[Ep..(r .

~,~.,,(p~)(r E..~

> (kin) n'~

for all m.

(4.7)

32

L.J. DIAZ, E.R. PUJALS AND R.. URES

On the other hand, from Lemma 4.6,

IlDPmr II(Dp,,,r

~< (l-a)"= I1~'11, (V u ) II4 (l+a)--"~ IIv"II,

(4.s) V u E E~,~m(pm).

Now, if km is close enough to 1 one has ( 1 - ~ ) " ~ < k~ ~. Therefore, there is wEE~,,~(pm) , [Iv~I]= Ilwll = 1, such that

IIDp~ r

II(DP~ r

(4.9)

(w)II = (k~) -~.

Let

T,,~ = ( k,~ ) 1/ "'' ,

7,~ ---+1 as m ---* oo .

We now perturb the derivative of D e along the orbit Pm, r

..., r

in the direction Dp,.,,r

plying the action of the derivative Dr162

modifying the derivative in the direction spanned by Dp.,~r

multi(without

by the factor ~-m. In

this way we obtain families 13m-*:D(r such that

IlB,~,,,-,(p,,)

... BP,,(vS)ll 9IIB~

.-. B~,2,,_,(p,,) (w)II = 1.

(4.1o)

Take the unit vector w'EECU(Bp,,,) in the direction B-1

....

u1

) (w)

(

and write

BP.,(v*)l].

p,~ = IIBr

(4.11)

For each 0 ~0 satisfying Lemma 4.22 we consider the cone fields (see Lemma 4.21) g'(Bp,~) = [g~-(E~(Bp,~))] ~

and

gh(Bp,~) = [C~-(E~(Bp~))],

(A c denotes the complement of A) and the matrices (Ti,m). Suppose, for instance, that we can perturb the (T~,m) according to case (a) in Lemma 4.21. By perturbing B (which is close to :D(r along the segment of C-orbit {I'm, r .-., r we obtain g close to B such that (1) CCj(p~)Ej=Ej+I, (2) E~(Ccj(p,~))=E~(Br for every j, (3) Cr where Ti,m is the perturbation of Ti,m in Lemma 4.21. Therefore, (4) l l j = 0

E j = l l i = l - L i , m = m.

~r

On one hand, by Lemma 4.22 and by definition,

Es(Cr

C [C~-(EC~'(Cr162

On the other hand, let C : C r

~ = C~'(Cr

for all 0 ~ C > 0 for all xEA~,(U), see (5.2).

PARTIAL

HYPERBOLICITY

AND

I:tOBUST

TRANSITIVITY

39

Therefore, ol(EU(~n),ECS(~n))---~O implies that a(E~(qo,0, E~(~n))--*0. We now have two possibilities:

I)~(P~, w=)l (1) IAc(Pn,~On)l § ])t=(Pn, ~on)] is uniformly bounded from above.

(2) I~c(P~, ~n)l

In the first case, using Fact 4.13 and Lemma 3.1, we get r

close to ~n such that

I~ (P~, Cn)l = I)~ (Pn, Ca)l, which contradicts (5.3). Finally, if Is is uniformly bounded, we apply Lemma 3.1 to perturb (/9n through the orbit of Pn to get Cn with IA~(Pmr162 contradicting (5.3).

5.4. P r o o f of Corollary D

The corollary follows applying the arguments in the proof of Theorem B to the splittings E~ | E c~ and E~ O E cs.

5.5. P r o o f s of T h e o r e m E a n d C o r o l l a r y G We divide the proof of the theorem into two parts: approximation by homoclinic tangencies (which will imply the corollary) and expansion/contraction of the volume. 5.5.1. Approximation by homoclinic tangencies. Proof of Corollary G. By hypotheses there are r close to ~, and P and Q c H y p P ( r of indices 2 and 1, respectively, such that Q has an expanding complex eigenvalue (recall Remark 4.15). The next lemma immediately implies the corollary. LEMMA 5.4. Let r be as above. Then given any RCHyp P(r

of index two there is r close to r with a homoclinic tangency associated to Re, where Re is the continuation of R f o r e . Proof. By Lemma 3.3 there is ~ close to r such that (1) WS(R~, ~) and W~(Q~, ~) have a nontrivial transverse intersection, and (2) W~(R~, ~) and WS(Q~, ~) have a point of quasitransverse intersection. Since Q~ has an expanding complex eigenvalue, W~(R~, ~) spirals around W~(Q~, ~). Finally, since WU(R~, ~) and WS(Q~, ~) are quasitransverse we can perturb ~ to obtain with a homoclinic tangency associated to Re. 5.5.2. The bundle E c~ is volume-expanding. follows immediately from the proposition below.

[]

Proof of Theorem E. The theorem

40

L.J. DIAZ, E.R. PUJALS AND R. URES

PROPOSITION 5.5. Let qoEA(U). Then we have: (1) if TAr(v)=ES(~)@EC~'(qo), then E c ( ~ ) = E ~ ( W ) is volume-expanding, (2) if TAr(v)=E~'(~)@E~S(qo), then E~(qo)=E~S(~) is volume-contracting.

Remark 5.6. This proposition does not hold (in general) if WET'(U). To see this observe that in such a case there are two possible choices for the central bundle, either

EC=E ~s or E~=E c~. Now just take qo having a fixed point P of index two with IA~(P))%(P)I 0 arbitrarily small and m~ big, we get r close to r such that Cm~ (y)= y and m e --1

1 me

E

l~162