Nonlinearity

PAPER Related content

- Dominated splitting and zero volume
Dominated splittings for flows with singularities Vitor Araujo and Mário Bessa

- Shades of hyperbolicity for Hamiltonians

To cite this article: Vitor Araujo et al 2013 Nonlinearity 26 2391 Mário Bessa, Jorge Rocha and Maria
Joana Torres

- Pressures for geodesic flows of rank one

manifolds
Katrin Gelfert and Barbara Schapira
View the article online for updates and enhancements.

Recent citations
- Dominated splitting for exterior powers and
singular hyperbolicity
Vitor Araujo and Luciana Salgado

This content was downloaded from IP address 131.170.21.110 on 22/12/2017 at 20:59

IOP PUBLISHING NONLINEARITY
Nonlinearity 26 (2013) 2391–2407 doi:10.1088/0951-7715/26/8/2391

Dominated splittings for flows with singularities

Vitor Araujo1 , Alexander Arbieto2 and Luciana Salgado3
1 Instituto de Matemática, Universidade Federal da Bahia, Av. Adhemar de Barros, S/N , Ondina,

40170-110-Salvador-BA-Brazil
2 Universidade Federal do Rio de Janeiro, Instituto de Matemática, PO Box 68530, 21945-970

Rio de Janeiro, Brazil

3 Instituto de Matemática Pura e Aplicada-Estrada Dona Castorina, 110, Jardim Botânico,

22460-320 Rio de Janeiro, Brazil

E-mail: [email protected], [email protected] and [email protected]

Received 24 October 2012, in final form 2 July 2013

Published 18 July 2013
Online at stacks.iop.org/Non/26/2391

Recommended by R de la Llave

Abstract
We obtain sufficient conditions for an invariant splitting over a compact
invariant subset of a C 1 flow Xt to be dominated. In particular, we reduce
the requirements to obtain sectional hyperbolicity and hyperbolicity.

Mathematics Subject Classification: 37D30, 37D25

1. Introduction

The theory of hyperbolic dynamical systems is one of the main paradigms in dynamics.
Developed in the 1960s and 1970s after the work of Smale, Sinai, Ruelle, Bowen [9, 10, 26, 27],
among many others, this theory deals with compact invariant sets
for diffeomorphisms and
flows of closed finite-dimensional manifolds with hyperbolic splitting of the tangent space.
That is, if X is a vector field and Xt is the generated flow then we say that an invariant and
compact set
is hyperbolic if there exists a continuous splitting of the tangent bundle over

, T
M = E s ⊕ E X ⊕ E u , where E X is the direction of the vector field, the subbundles are
invariant under the derivative DXt of the flow Xt
DXt · Ex∗ = EX∗ t (x) , x ∈
, t ∈ R, ∗ = s, X, u;
E s is uniformly contracted by DXt and E u is uniformly expanded: there are K, λ > 0 so that
DXt |Exs 
Ke−λt , DX−t |Exu 
Ke−λt , x ∈
, t ∈ R. (1.1)
Very strong properties can be deduced from the existence of such a structure; see for
instance [13, 25].
Weaker notions of hyperbolicity, such as the notions of dominated splitting, partial
hyperbolicity, volume hyperbolicity and singular or sectional hyperbolicity (for singular flows),

0951-7715/13/082391+17$33.00 © 2013 IOP Publishing Ltd & London Mathematical Society Printed in the UK & the USA 2391
2392 V Araujo et al

have been proposed to try to enlarge the scope of this theory to classes of systems beyond the
uniformly hyperbolic ones; see [1] and [5] for singular or sectional hyperbolicity. However,
the existence of dominated splittings is the weaker one.
Many researchers have studied the relations of the existence of dominated splittings with
other dynamical phenomena, mostly in the discrete time case, such as robust transitivity,
homoclinic tangencies and heteroclinic cycles, and also the possible extension of this notion
to endomorphisms; see for instance [5, 6, 15, 19–21, 33].
However, the notion of dominated splittings deserves attention. Several authors used this
notion for linear Poincaré flow, see [11, 16, 17], and this is useful in the absence of singularities,
as in [12]. We remark that this flow is defined only in the set of regular points. However, in
the singular case, it is a non-trivial question to obtain a dominated splitting for the derivative
of the flow, and thus allowing singularities. Indeed, it is difficult to obtain it from a dominated
splitting for linear Poincaré flow. See [14], for an attempt to solve this, in the context of
robustly transitive sets, using the extended linear Poincaré flow.
Definition 1.1. A dominated splitting over a compact invariant set
of X is a continuous
DXt -invariant splitting T
M = E ⊕ F with Ex = {0}, Fx = {0} for every x ∈
and such
that there are positive constants K, λ satisfying
DXt |Ex  · DX−t |FXt (x)  < Ke−λt , for all x ∈
, and all t > 0. (1.2)
The purpose of this article is to study the existence of dominated splittings for flows
with singularities. On one hand, we study this question in the presence of some sectional
hyperbolicity, as this theory was built to understand flows like the Lorenz attractor, which
are the prototype of non-hyperbolic dynamics with singularities, but with robust dynamical
properties. On the other hand, we present several examples to clarify the role of the condition
on the singularities.
This article is organized as follows. In section 2, we present the definitions and auxiliary
results needed in the proof of theorem A, and some examples showing that the domination
condition on the singularities is necessary. In section 3, we present applications of our main
result and more examples implying that these results are not valid for diffeomorphisms. In
section 4 theorem B is proved assuming theorem A. In section 5, we prove theorem A and
theorem 2.4. Theorem C and corollary 2 are proved in section 6 using the previously proved
results. Finally, in section 7, we present the examples stated in propositions 1 and 2.

2. Statements of the results

Let M be a connected compact finite n-dimensional manifold, n  3, with or without a

boundary. We consider a vector field X, such that X is inwardly transverse to the boundary
∂M, if ∂M = ∅. The flow generated by X is denoted by {Xt }.
An invariant set
for the flow of X is a subset of M which satisfies Xt (
) =
for
all t ∈ R. The maximal invariant set of the flow is M(X) := ∩t
0 Xt (M), which is clearly a
compact invariant set.
A singularity for the vector field X is a point σ ∈ M such that X(σ ) = 0.  The set
formed by singularities is denoted by Sing(X). We say that a singularity is hyperbolic if the
eigenvalues of the derivative DX(σ ) of the vector field at the singularity σ have a non-zero
real part.
A compact invariant set
is said to be partially hyperbolic if it exhibits a dominated
splitting T
M = E ⊕ F such that subbundle E is uniformly contracted, i.e. there exist C > 0
and λ > 0 such that DXt |Ex 
Ce−λ for t  0. In this case F is the central subbundle
of
.
Dominated splittings 2393

We say that a DXt -invariant subbundle F ⊂ T
M is a sectionally expanding subbundle

if dim Fx  2 is constant for x ∈
and there are positive constants C, λ such that for every
x ∈
and every two-dimensional linear subspace Lx ⊂ Fx one has
| det(DXt |Lx )| > Ceλt , for all t > 0. (2.1)
Definition 2.1 ([18, definition 2.7]). A sectional-hyperbolic set is a partially hyperbolic set
whose singularities are hyperbolic and whose central subbundle is sectionally expanding.
This is a generalization of hyperbolicity for compact invariant sets with singularities
accumulated by regular orbits, since the Lorenz attractor and higher dimensional examples are
sectional-hyperbolic [1, 7, 18, 30] and because of the following result the proof of which can
be found in [1, 5, 19].
Lemma 2.2 (Hyperbolic lemma). Every compact invariant subset of a sectional-hyperbolic
set without singularities is a hyperbolic set.
The main result of this paper is the following.
Theorem A. Let
be a compact invariant set of X such that every singularity in this set
is hyperbolic. Suppose that there exists a continuous DXt -invariant splitting of the tangent
bundle of
, T
M = E ⊕ F , where E is uniformly contracted, F is sectionally expanding
and for some constants C, λ > 0 we have
DXt |Eσ  · DX−t |Fσ  < Ce−λt for all σ ∈
∩ Sing(X) and t  0. (2.2)
Then T
M = E ⊕ F is a dominated splitting.
As a corollary, we obtain sufficient conditions to obtain hyperbolicity in the non-singular
case using the hyperbolic lemma.
Corollary 1. Let
be a compact invariant set without singularities for the vector field X.
Suppose that there exists a continuous DXt -invariant decomposition of the tangent bundle of

, T
M = E ⊕ F , where E contracts and F is sectionally expanding. Then
is a hyperbolic
set.
We note that in definition 2.1 domination is required. As a consequence of theorem A, the
domination assumption is only necessary at the singularities, so that we obtain the following
equivalent definition of sectional-hyperbolicity.
Definition 2.3. A compact invariant set
⊂ M is a sectional-hyperbolic set for X if all
singularities in
are hyperbolic, there exists a continuous DXt -invariant splitting of the
tangent bundle on T
M = E ⊕ F with constants C, λ > 0 such that for every x ∈
and
every t > 0 we have
(1) DXt |Ex 
Ce−λt ;
(2) | det DXt |Lx | > C −1 eλt , for every two-dimensional linear subspace Lx ⊂ Fx ;
(3) DXt |Eσ  · DX−t |Fσ  < Ce−λt for all σ ∈
∩ Sing(X).

2.1. Hyperbolicity versus sectional-expansion

We present here some motivation for theorem A. We observe that, for a hyperbolic set, both
splittings (E s ⊕ E X ) ⊕ E u and E s ⊕ (E X ⊕ E u ) are dominated. Moreover, it is easy to see
that a hyperbolic set
has no singularities accumulated by regular orbits within
. Indeed,
any singularity σ accumulated by regular orbits on a hyperbolic set
would be a discontinuity
point for the hyperbolic splitting, due to the absence of the flow direction at σ .
2394 V Araujo et al

Our next result shows that we cannot have a hyperbolic splitting without the flow direction,
unless we restrict ourselves to finitely many singularities, all of them isolated on the non-
wandering set.
Theorem 2.4. Let
be a compact invariant set of X. Suppose that there exists a continuous
DXt -invariant decomposition of the tangent bundle T
M = E ⊕ F over
and constants
C, λ > 0, such that for every x ∈
and all t > 0
DXt |Ex 
Ce−λt and DX−t |FXt (x) 
Ce−λt .
Then
consists of finitely many hyperbolic singularities.
We cannot replace the assumptions on theorem 2.4 either by sectional-expansion or by
sectional-contraction, as the following examples show.
Example 1. Consider a Lorenz-like singularity σ for a C 1 flow {Xt }t∈R on a 3-manifold M,
that is, σ is a hyperbolic singularity of saddle-type such that the eigenvalues of DX(σ ) are
real and satisfy λ2 < λ3 < 0 < −λ3 < λ1 and λ1 + λ2 > 0.
Let Ei be the eigenspace associated to the eigenvalue λi , i = 1, 2, 3, and set E = E3
and F = E1 ⊕ E2 . Then the decomposition is trivially continuous, not dominated (F admits
vectors more sharply contracted than those of E) but E uniformly contracts the lengths of the
vectors and F uniformly expands the area, that is, F is sectionally expanded.
Example 2. Consider a hyperbolic saddle singularity, σ , for a C 1 flow {Xt }t∈R on a 4-manifold
M such that the eigenvalues of DX(σ ) are real and satisfy
λ2 < λ3 < 0 < λ4 < λ1 , λ1 + λ3 > 0 and λ2 + λ4 < 0.
Let Ei be the eigenspace associated to the eigenvalue λi , i = 1, 2, 3, 4, and set F = E2 ⊕ E4
and E = E1 ⊕ E3 . Then the decomposition is trivially continuous, not dominated as before,
E is uniformly area contracting, since DX−t |E expands area for t > 0; and F is uniformly
area expanding. In other words, F is sectionally expanded and E is sectionally contracted.
In both examples above we have sectional-expansion and sectional-contraction along the
subbundles of a continuous splitting but the splitting is not dominated. These examples involve
a trivial invariant set: an equilibrium point. But there are examples with compact invariant
sets having singularities accumulated by regular orbits and also with a dense regular orbit; see
section 7.
This suggests that theorem 2.4 might be generalized if we assume domination at the
singularities of
together with sectional-expansion along F and uniform contraction along
E over regular orbits. This is precisely the content of theorem A.

3. Applications and other examples

To present the next result characterizing sectional-hyperbolic sets through sectional Lyapunov
exponents, we first have to present more definitions.
We say that a probability measure µ is X-invariant if µ(Xt (U )) = µ(U ), for every
measurable subset U ⊂ M and every t ∈ R. Given a compact X-invariant set
we say that
a subset Y ⊂ M is a total probability subset of
if µ(Y ) = 1 for every X-invariant measure
µ supported in
. We note that each singularity σ of
belongs to Y since µ = δσ is a
X-invariant probability measure.
Let A : E × R → E be a Borel measurable map given by a collection of linear bijections
At (x) : Ex → EXt (x) , x ∈ M, t ∈ R,
Dominated splittings 2395

where M is the base space (we assume it is a manifold) of the finite-dimensional vector bundle
E, satisfying the cocycle property
A0 (x) = I d, At+s (x) = At (Xs (x)) ◦ As (x), x ∈ M, t, s ∈ R,
with {Xt }t∈R a smooth flow over M. We note that for each fixed t > 0 the map At : E →
E, vx ∈ Ex → At (x) · vx ∈ EXt (x) is an automorphism of the vector bundle E.
The natural example of a linear multiplicative cocycle over a smooth flow Xt on a manifold
is the derivative cocycle At (x) = DXt (x) on the tangent bundle T M of a finite-dimensional
compact manifold M.
A concrete instance of this structure we will use is the derivative cocycle At (x) = DXt |
Ex restricted to a continuous DXt -invariant subbundle E of the tangent bundle T M.
According to the multiplicative ergodic theorem of Oseledets [3, 4], since in this setting
we have that sup−1t1 log At (x) is a bounded function of x ∈
, there exists a subset R
of
with total probability such that for every x ∈ R there exists a splitting
Ex = E1 (x) ⊕ · · · ⊕ Es(x) (x) (3.1)
which is DXt -invariant and the following limits, known as the Lyapunov exponents at x, exist
1
λi (x) = lim log DXt (x) · v,
t→±∞ t

for every v ∈ Ei (x)\{0}, i = 1, . . . , s(x). We order these numbers as λ1 (x) < . . . < λs(x) (x).
One of these subbundles is given by the flow direction (at non-singular points of the flow) and
the corresponding Lyapunov exponent is zero for almost every point.
The functions s and λi are measurable and invariant under the flow, i.e., s(Xt (x)) = s(x)
and λi (Xt (x)) = λi (x) for all x ∈ R and t ∈ R. The splitting (3.1) also depends measurably
on the base point x ∈ R. If F is a measurable subbundle of the tangent bundle then by ‘the
Lyapunov exponents of E’ we mean the Lyapunov exponents of the non-zero vectors in F .
Given a vector space E, we denote by ∧2 E the second exterior power of E, defined as
follows. For a basis v1 , . . . , vn of E, then ∧2 E is generated by {vi ∧ vj }i=j . Any linear
transformation A : E → F induces a transformation ∧2 A : ∧2 E → ∧2 F . Moreover, vi ∧ vj
can be viewed as the 2-plane generated by vi and vj if i = j ; see for instance [3] for more
information.
In [2], a notion of sectional Lyapunov exponents was defined and a characterization of
sectional-hyperbolicity was obtained based on this notion.
Definition 3.1 ([2, definition 2.2]). Given a compact invariant subset
of X with a DXt -
invariant splitting T
M = E ⊕ F , the sectional Lyapunov exponents of x along F are the
limits
1
lim log  ∧2 DXt (x) ·
v
t→+∞ t

whenever they exists, where
v ∈ ∧2 Fx − {0}.

As explained in [2], if {λi (x)}s(x)
i=1 are the Lyapunov exponents, then the sectional Lyapunov
exponents at a point x ∈ R are {λi (x) + λj (x)}1i<j s(x) and they represent the asymptotic
sectional-expansion in the F direction.
As an application of theorem A, we have the following result, which is an extension of
the main result in [2] assuming continuity of the splitting, instead of a dominated splitting.
Theorem B. Let
be a compact invariant set for a X such that every singularity σ ∈
is
hyperbolic. Suppose that there is a continuous DXt -invariant splitting T
M = E ⊕ F such
that Tσ M = Eσ ⊕ Fσ is dominated, for every singularity σ ∈
.
2396 V Araujo et al

If the Lyapunov exponents in the E direction are negative and the sectional Lyapunov
exponents in the F direction are positive on a set of total probability within
, then the
splitting is dominated and
is a sectional-hyperbolic set.
We now apply the main result to the setting of weakly dissipative three-dimensional flows.
We recall that an open subset U is a trapping region if Xt (U ) ⊂ U, t > 0 and a compact
invariant subset
is attracting if it is the maximal invariant subset
(U ) = ∩t
0 Xt (U ) inside
the trapping region U . We say that a compact invariant subset
for the flow of the vector field
X is weakly dissipative if div(X)(x)
0 for all x ∈
, that is, the flow near
infinitesimally
does not expand in volume.
We also say that a hyperbolic singularity σ of X is C 1 -linearizable if there exists a
diffeomorphism h : Vσ → B from an open neighbourhood of σ ∈ M to an open neighbourhood
B of the origin in R3 such that h(Xt (x)) = etA ·h(x), Xs (x) ∈ Vσ for |s|
t and A = DX(σ ).
For this it is enough that the spectrum of DX(σ ) satisfies a finite number of non-resonance
conditions; see e.g. [28]. Hence this is a C r -generic condition for all sufficiently large r  1.
Theorem C. Let X be a C 1 vector field on a three-dimensional manifold M admitting a
trapping region U whose singularities (if any) are hyperbolic and C 1 linearizable. Let us
assume that the compact invariant subset
=
(U ) is weakly dissipative and endowed with
a one-dimensional continuous field F of asymptotically backward contracting directions, that
is, x ∈
→ Fx is continuous and for each x ∈
, Fx is a one-dimensional subspace of Tx M,
and also
1
lim inf log DX−t | Fx  < 0 for all x ∈
. (3.2)
t→+∞ t

If at each singularity σ ∈ U there exists a complementary DXt -invariant direction Eσ such

that Eσ ⊕ Fσ = Tσ M is a dominated splitting, then
is a hyperbolic set (in particular,
has
no singularities).
Since we use the attracting and dominated splitting assumption only to prove the non-
existence of singularities in
, as a consequence of the proof we obtain the following.
Corollary 2. Let X be a C 1 vector field on a three-dimensional manifold M admitting compact
invariant subset
, without singularities, which is weakly dissipative and endowed with a one-
dimensional continuous field F of asymptotically backward contracting directions.
Then
is a hyperbolic set.
These results are extremely weak versions of the following conjecture of Viana, presented
in [31].
Conjecture 1. If an attracting set
(U ) of smooth map/flow has a non-zero Lyapunov exponent
at Lebesgue almost every point of its isolating neighbourhood U , that is
1
lim sup log DXt  > 0 for Lebesgue almost every x ∈ U, (3.3)
t→+∞ t
then
(U ) has a physical measure: there exists an invariant probability measure µ supported
in
(U ) such that for all continuous functions ϕ : U → R
 
1 t
lim ϕ(Xt (x)) dt = ϕ dµ for Lebesgue almost every x ∈ U.
t→+∞ t 0

Indeed, if U is a trapping region and
=
(U ) satisfies the assumptions of theorem C, then

is a hyperbolic attracting set, for which it is well known that (3.3) is true and there exists
some physical measure; see e.g. [5, 9, 10].
Dominated splittings 2397

We now provide several examples which clarify the role of the assumption of domination
at the singularities in theorem A and the relations between dominated splittings for
diffeomorphisms and flows.
Proposition 1. There exist the following examples of vector fields and flows.
(1) A vector field with an invariant and compact set containing singularities accumulated by
regular orbits inside the set, with a non-dominated and continuous splitting of the tangent
space, but satisfying uniform contraction and sectional-expansion.
(2) A vector field with an invariant and compact set with a continuous and invariant splitting
E ⊕ F such that E is uniformly contracted, F is area expanding but it is not dominated.
(3) A suspension flow whose base map has a dominated splitting but the flow does not admit
any dominated splitting.
Item (3) above is based on a diffeomorphism described in [8] and similar to another
suggested by Pujals in [5, example B.12].
An important remark is that our results are not valid for diffeomorphisms, as the following
result shows.
Proposition 2. There exists a transitive hyperbolic set for a diffeomorphism with a non-
dominated splitting which is sectionally-expanding and sectionally contracting.

4. Proof of theorem B

We follows the lines of [2]. The following proposition, whose proof can be found in [2], is the
main auxiliary result in the proof.
We fix a compact Xt -invariant subset
. We say that a family of functions {ft :
→ R}t∈R
is subadditive if for every x ∈ M and t, s ∈ R we have that ft+s (x)
fs (x) + ft (Xs (x)).
The subadittive ergodic theorem (see e.g. [32]) shows that the function f (x) = lim inf ft (x)
t
t→+∞
coincides with f
(x) = lim 1 ft (x) in a set of total probability in
.
t→+∞ t
Proposition 3. Let {t → ft :
→ R}t∈R be a continuous family of continuous function
which is subadditive and suppose that f (x) < 0 in a set of total probability. Then there
exist constants C > 0 and λ < 0 such that for every x ∈
and every t > 0 we have
exp(ft (x))
C exp(λt/2).

Proof. See [2, proposition 3.4]. 

Remark 1. In [2] only
= M(X) was considered to conclude that {Xt }t∈R is a sectional-
Anosov flow. However all statements are valid for a compact invariant subset of M.

Proof of theorem B. Define the following families of continuous functions on

φt (x) = log DXt |Ex  and ψt (x) = log  ∧2 DX−t |FXt (x) , (x, t) ∈
× R.
Both families φt , ψt :
→ R are easily seen to be subadditive. From proposition 3 applied
to φt (x) and the hypothesis on the Lyapunov exponents, there are constants C > 0 and γ < 0
such that exp(φt (x)) = DXt |Ex 
C exp(γ t) for all t > 0 showing that E is a contractive
subbundle.
Analogously, the hypothesis on the sectional Lyapunov exponents and proposition 3
applied to the function ψt (x) provides constants D > 0 and η < 0 for which  ∧2
DX−t |FXt (x) 
Deηt , so F is a sectionally expanding subbundle. Now theorem A ensures that
the splitting E ⊕ F is a dominated splitting. 
2398 V Araujo et al

5. Proof of theorem A and theorem 2.4

We need an auxiliary lemma for which the following definition is necessary. Given a pair
 the angle  (Ex , Fx ) between the two
of subspaces Ex , Fx of Tx M such that Ex ∩ Fx = {0},
subspaces is defined by
1
sin  (Ex , Fx ) := , x ∈
,
π(Ex )
where π(Ex ) : Ex ⊕ Fx → Ex is the projection onto Ex parallel to Fx defined in the vector
space Ex ⊕ Fx .
We say that an invariant splitting E ⊕ F = T
M of the tangent bundle over a invariant
subset
has angle uniformly bounded away from zero if the dimensions of the fibers Ex , Fx
are constant for all x in
and there exists θ0 > 0 such that sin  (Ex , Fx )  θ0 for each x in
.
The next lemma specifies the subbundle which contains the flow direction.
Lemma 5.1. Let
be a compact invariant set for X. Given an invariant splitting E ⊕ F of
T
M with the angle bounded away from zero over
, such that E is uniformly contracted,
then the flow direction is contained in the F subbundle, for all x ∈
.

Proof. We denote by π(Ex ) : Tx M → Ex the projection on Ex parallel to Fx at Tx M, and

likewise π(Fx ) : Tx M → Fx is the projection on Fx parallel to Ex . We note that for x ∈

X(x) = π(Ex ) · X(x) + π(Fx ) · X(x)
and for t ∈ R, by linearity of DXt and DXt -invariance of the splitting E ⊕ F
DXt · X(x) = DXt · π(Ex ) · X(x) + DXt · π(Fx ) · X(x)
= π(EXt (x) ) · DXt · X(x) + π(FXt (x) ) · DXt · X(x)
Let z be a limit point of the negative orbit of x. That is, we assume without loss of
generality since
is compact, that there is a strictly increasing sequence tn → +∞ such
that lim xn := lim X−tn (x) = z. Then z ∈
and, if π(Ex ) · X(x) is not the zero vector,
n→+∞ n→+∞
we get
lim DX−tn · X(x) = lim X(xn ) = X(z), but also
n→+∞ n→+∞

lim DX−tn · π(Ex ) · X(x)  lim c−1 eλtn π(Ex ) · X(x) = +∞. (5.1)
n→+∞ n→+∞

This is possible only if the angle between Exn and Fxn tends to zero when n → +∞.
Indeed, using the Riemannian metric on Ty M, the angle α(y) = α(Ey , Fy ) between Ey
and Fy is related to the norm of π(Ey ) as follows: π(Ey ) = 1/ sin(α(y)). Therefore
DX−tn · π(Ex ) · X(x) = π(Exn ) · DX−tn · X(x)
1

· DX−tn · X(x)
sin(α(xn ))
1
= · X(xn )
sin(α(xn ))
for all n  1. Hence, if the sequence DX−tn · π(Ex ) · X(x) in the left hand side of (5.1) is
unbounded, then lim α(X−tn (x)) = 0.
n→+∞
However, since the splitting E ⊕F has an angle bounded away from zero over the compact

, we have obtained a contradiction. This contradiction shows that π(Ex ) · X(x) is always
the zero vector and so X(x) ∈ Fx for all x ∈
. 
Dominated splittings 2399

The proof of theorem 2.4 follows from lemma 5.1.

Proof of theorem 2.4. We assume, arguing by contradiction, that there exists a point
x ∈
\ Sing(X), that is, x is a regular point: X(x) = 0. Hence lemma 5.1 ensures that
X(x) ∈ Fx . But the same lemma 5.1 applied to the reversed flow X−t generated by the field
−X on the same invariant set
shows that X(x) ∈ Ex . Thus X(x) ∈ Ex ∩ Fx = {0}.  This
contradiction ensures that
⊂ Sing(X). But our assumptions on the splitting show that each
σ ∈
is a hyperbolic singularity: the eigenvalues of DX |Eσ are negative and the eigenvalues
of DX |Fσ are positive. It is well known that hyperbolic singularities σ are isolated in the
ambient manifold M; see e.g. [22]. We conclude that the compact
is a finite set of hyperbolic
singularities. 

Before proving theorem A we need some extra notions and definitions.

We recall that, for a invertible continuous linear operator L, the minimal norm (or co-
norm) is equal to m(L) := L−1 −1 . Hence condition (1.2) in definition 1.1 is equivalent
to
DXt |Ex  < Ke−λt · m(DXt |Fx ) for all x ∈
and for all t > 0. (5.2)
The multiplicative ergodic theorem of oseledets [3, 4] provides the following properties
of Lyapunov exponents and Lyapunov subspaces, in addition to the existence of Lyapunov
exponents on the total probability subset R of
, as presented in section 3.  For any pair of
disjoint subsets
 I, J ⊂ {1, . . . , s(x)}, the angle between the bundles E I (x) = i∈I Ei (x) and
EJ (x) = j ∈J Ej (x) decreases at most subexponentially along the orbit of x, that is
1
lim log sin  (EI (Xt (x)), EJ (Xt (x))) = 0, x ∈ R;
t→±∞ t
which implies, in particular, that for any pair i, j ∈ {1, . . . , s(x)} with i = j and vi ∈
 vj ∈ Ej (x) \ {0}
Ei (x) \ {0}, 
1
lim log | det(DXt | span{vi , vj })| = λi (x) + λj (x).
t→±∞ t

Proof of theorem A. Let x ∈ R ⊂
be a regular point of the flow of X. From

lemma 5.1 we know that X(x) ∈ Fx . Since F is a DXt -invariant subbundle, considering
 multiplicative cocycle At (z) = DXt | Fz for z ∈
, t ∈ R, there exists a splitting
the linear
Fx = s(x)j =1 Fj (x) of Fx into a direct sum of Lyapunov subspaces. One of these subspaces is
E generated by X(x) = 0,
X  which we rename F1 (x) = ExX in what follows. We also have the
corresponding Lyapunov exponents λFj (x), j = 1, . . . , s(x).
 we consider span{X(x), v}. From the
Fixing i = 2, . . . , s(x) and v ∈ Fi (x) \ {0},
assumption of sectional expansion of area together with the subexponential control on angles
at regular points we obtain
1
0 < λ
lim inf log | det(DXt | span{X(x), v})| = λF1 (x) + λFi (x) = λFi (x).
t→+∞ t
Hence λFi (x)  λ > 0 for all i = 2, . . . , s(x), x ∈ R and X(x) = 0. 
If we consider the cocycle Bt (z) = DXt | Ez , x ∈
, t ∈ R, then we find the splitting
r(x)
Ex = i=1 Ej (x) into Lyapunov subspaces and the corresponding Lyapunov exponents
i (x), i = 1, . . . , r(x). It is easy to see that the assumption of uniform contraction along E
λE
implies that λEi (x)
−λ < 0 for all x ∈ R and 1
i
r(x).
2400 V Araujo et al

DXt |Ex 
Then φt (x) = log m(DX t |Fx )
is a subadditive family of continuous functions satisfying
φt (x) 1 1
φ(x) = lim inf
lim inf log DXt |Ex  − lim sup log m(DXt |Fx )
t→+∞ t n→+∞ t n→+∞ t

= max{λE
i (x), 1
i
r(x)} − min{λi (x), 1
i
s(x)}
−λ − 0 = −λ
F


for all x ∈ R such that X(x) = 0.
For x = σ for some singularity σ ∈
we have φ(σ )
−λ as a direct consequence of
the assumption of domination at the singularities.
We have shown that φ(x) < 0 for all x ∈ R. Applying proposition 3 we see that there
exists C > 0 such that exp φt (x)
Ce−λt/2 for all x ∈
and t > 0, which gives us the
condition (5.2). This concludes the proof of theorem A. 
The proof of corollary 1 follows from theorem A and the hyperbolic lemma 2.2.

6. Proof of theorem C and corollary 2

We start by showing that the one-dimensional asymptotically backward contracting subbundle

F over
is in fact a uniformly expanding subbundle. Indeed, we note that ft (x) = log DX−t |
Fx  is a subadditive family of continuous functions satisfying f (x) < 0 for all x ∈
.
In particular, f (x) < 0 in a set of total probability. Hence by proposition 3 we obtain
ft (x)
Ce−tλ/2 for all t > 0 and some constants λ, C > 0. (This trivially implies
DXt | FXt (x)   etλ/2 /C, x ∈
, t  0 and so DXT | Fx  > 2 for all x ∈
and
T  2 log(2C)/λ.)
Now we use the assumption that
is an attracting set whose possible singularities σ are
hyperbolic, C 1 linearizable and admit a complementary DXt -invariant direction Eσ such that
Eσ ⊕ Fσ = Tσ M is a dominated splitting.

Lemma 6.1. The attracting set
does not contain singularities.

Proof. By the C 1 -linearization assumption and the stable/unstable manifold theorems, there
exists a corresponding Xt -invariant (weak- or strong-)unstable one-dimensional manifold Wσu
contained in
(since
is attracting) such that Fσ = Tσ Wσu and X−t (z) −→ σ for each
t→+∞
z ∈ Wσu .

Claim 1. For all z ∈ Wσu we have X(z) ∈ Fz .

We note that this is trivially true for z = σ . Let us assume, by contradiction, that
X(z) ∈/ Fz for some z ∈ Wσu \ {σ } (and thus for all such z by DXt -invariance). By assumption
there exists a diffeomorphism h : Vσ → B from an open neighbourhood of σ ∈ M to an open
neighbourhood B of the origin in R3 such that h(Xt (x)) = etA · h(x), Xs (x) ∈ Vσ for |s|
t
and A = DX(σ ). By a linear change of coordinates we can assume without loss of generality
that Dh(σ ) · Fσ = R × 02 and Dh(σ ) · Eσ = 0 × R2 . Let v : Vσ ∩
→ R3 denote the
continuous unitary vector field such that span(vx ) = Dh(x) · Fx , x ∈ Vσ ∩
.
We are assuming that vz has a non-zero component along the Eσ ≈ 0 × R2 direction
for z ∈ Wσu \ {σ }. Since X−t (z) → σ the domination of the splitting Eσ ⊕ Fσ ensures
t→+∞
that  (etA · vz , Eσ ) −→ 0. However the invariance of F ensures that FX−t (z) −→ Fσ and
t→−∞ t→−∞
since h is a C 1 conjugation we also have  (etA · vz , Fσ ) −→ 0. This contradiction proves the
t→−∞
claim.
Dominated splittings 2401

Using the claim together with the previous estimate on expansion along F , we have that
∞ > sup X(x)  X(XkT (z)) = DXkT · X(z)  2k X(z), k  1
x∈

which is a contradiction and completes the proof of the lemma. 
From now we use only the assumptions of corollary 2, that is, that
is a weakly dissipative
compact invariant subset without singularities having a continuous field of asymptotically
backward contracting directions.
In this setting, the linear Poincaré flow is well defined over
. If we denote by
Ox : Tx M → Nx the orthogonal projection from Tx M to Nx = {v ∈ Tx M :< v, X(x) >= 0}
the orthogonal complement of the field direction at x ∈
, then the linear Poincaré flow is
given by Pxt = OXt (x) ◦ DXt : Tx M → NXt (x) , x ∈
, t ∈ R.
 is invariant, from the splitting Tx M = Ex ⊕ NX
X
We note that since the flow direction
X(X−t (x))
X(x)

we can write DXt = and since | det DX−t |  1 by the weak dissipative
0 P −t
assumption, we get | det(P −t )|  m0 = minz∈
(X(x)/X(X−t (x))) > 0.
Claim 2. The subbundle F̃ of the normal bundle given by {F̃x = Ox (Fx )}x∈
is P t -invariant
and uniformly backward contracting.
Indeed, the continuity of the subbundle F ensures that  (Fx , X(x)) is bounded away from
zero, and so there exists κ > 0 such that for every v ∈ F̃x we have |β|
κv such that
βX(x) + v = u ∈ Fx . This ensures that
Px−t · v = OX−t (x) (DX−t · (u − βX(x)))
OX−t (x) (DX−t · u − βX(X−t (x)))

DX−t · u
Ce−tλ/2 v + βX(x)
Ce−tλ/2 (v + κvX(x))
 

C 1 + κ sup X(z) e−tλ/2 v = C̃e−tλ/2 v.
z∈

This proves claim 2.
This is enough to guarantee the existence of a complementary P −t -invariant and expanding
subbundle through a graph transform technique. To this end, let us consider the one-
dimensional subbundle Gx of Nx orthogonal  to F̃x . Wecan then write Nx = F̃x ⊕ Gx and so,
a (x) bt (x)
since F̃ is P t -invariant, we get Px−t =
t
, where at (x) : F̃x → F̃X−t (x) , ct (x) :
0 ct (x)
Gx → GX−t (x) and bt (x) : Gx → FX−t (x) . We know that |at (x)|
C̃e−tλ/2 and since
0 < m0
| det(P −t )| = |at (x)| · |ct (x)| we obtain |ct (x)|  m0 etλ/2 /C̃, x ∈
, t > 0. We
fix T > 0 such that m0 eT λ/2 /C̃  2.
Let us now consider the family of continuous one-dimensional complementary subspaces
of F̃x inside Nx given by the graph of a linear map L : Gx → F̃x , i.e., we consider the vector
space G of all L = {Lx : Gx → F̃x : x ∈
} with the norm L1 − L2  = supx∈
L1x − L2x .
We have that G with this norm is a Banach space.
It is easy to see that the map P −t transforms the graph of Lx into the graph of
L̂X−t (x) = (Px−t · Lx + πX−t (x) · P −t ) ◦ [(I − πX−t (x) ) · P −t ]−1 , where πz : Nz → F̃z is
the orthogonal projection from Nz into F̃z at z ∈
. In this way we get a map P −T : G → G
which we claim is a contraction with respect to the norm given above. Indeed, for any pair
Li = {Lix }x∈
, i = 1, 2 we get
L̂1X−T (x) − L̂2X−T (x)  = Px−t · (L1x − L2x ) ◦ [(I − πX−t (x) ) · P −t ]−1 
1

|aT (x)| · L1x − L2x /|ct (x)|
L1x − L2x 
2
2402 V Araujo et al

Figure 1. A double homoclinic saddle connection expanding area.

for each x ∈
. We thus have a P −T fixed point L0 defining a P −T -invariant subbundle

Ẽ = {Ẽx = {(u, L0x (u)) : u ∈ Gx }x∈
of the normal bundle. Since P −T commutes with each
P s for all s ∈ R, we see that Ẽ is P s -invariant for all s ∈ R.
Claim 3. The subbundle E = Ẽ ⊕ E X is DXt -invariant.
Indeed, for any v = u + βX(x) with u ∈ Ẽx , β ∈ R, we have DXt · v = DXt · u + βX(Xt (x))
and DXt · u = Pxt · u + γ X(Xt (x)) for some γ ∈ R, thus DXt · v = Pxt · u + (β + γ )X(Xt (x))
with Pxt · u ∈ ẼXt (x , hence DXt · v ∈ EXt (x) and the claim is proved.
At this point we have a continuoussplitting Tx M = Ex ⊕ Fx , x ∈
. Using this
At (x) 0
invariant splitting we can write DX−t = and the continuity of the splitting
0 Bt (x)
ensures that  (Ex , Fx ) is bounded away from zero. This ensures that 1
| det DX−t | =
| det At (x)| · |Bt (x)| · sin  (Ex , Fx ) and since Bt (x) = DX−t | Fx we have |Bt (x)|
Ce−tλ/2 ,
hence | det At (x)| grows exponentially fast.
This shows that
has a continuous invariant splitting E ⊕F with F uniformly contracting
and E area expanding for the backward flow. In a three-dimensional manifold this shows that
E is sectionally expanding for −X. From corollary 1 we have that
is a hyperbolic set for
−X, hence
is a hyperbolic set and the proof of theorem C and corollary 2 is complete.

7. Examples

This section is devoted to presenting some examples.

7.1. Proof of proposition 1

Here we present examples stated in proposition 1.
Item (1). Consider a vector field on the plane R2 with a double saddle homoclinic connection
which expands volume and multiply this vector field by a contraction along the vertical
direction; see figure 1. For the construction of the planar vector field, see e.g. [23,
chapter 4, section 9].
The compact invariant set W , for the vector field X, formed by the pair of homoclinic
connections together with the singularity s admits a splitting E ⊕ F , where E is
the vertical direction and F is the plane direction. This splitting is continuous, E is
uniformly contracting and F is uniformly area expanding.
Dominated splittings 2403

1
W1

W2

− s
λ
1 s λ−
2
2
s 3 s
4 4
λ+
s
λ+1 1 2
W3

W4
−1
S1

Figure 2. A sketch of Bowen’s example flow.

Indeed, since we assume that the singularity s expands area, i.e., | det DXt (x)|Lx | 
ceλt for all t > 0 and all 2-plane Lx ⊂ Fx , for some fixed constants c, λ > 0, we can
fix a neighbourhood U of s such that
| det DXt (x)|Lx |  ceλt , ∀0
t
1, ∀x ∈ U
and consequently
| det DXt (x)|Lx |  ceλt , ∀t > 0
whenever the positive orbit of x is contained in U . Now we observe that there exists
T > 0 such that for all x ∈ W \U we have Xt (x) ∈ U for all t ∈ R satisfying |t| > T .
Hence we can choose k = inf{c−1 e−λt | det DXt (x)|Lx | : x ∈ W \U, 0
t
T } > 0
and write | det DXt (x)|Lx |  k · ceλt for 0
t
T ; and for t > T
| det DXt (x)|Lx | = | det DXt−T (XT (x))|LXT (x) | · | det DXT (x)|Lx |  ceλ(t−T ) · k · ceλT
= kc2 eλt .
So we can find a constant κ > 0 so that | det DXt (x)|Lx |  κeλt for all x ∈ W and
every t  0.
To obtain a non-dominated splitting, just choose the contraction rate along E to be
weaker than the contraction rate of the singularity s.
Example 3. Considering, in the previous item (1),
as the union of the saddle s with
only one homoclinic connection, we have T
M = E ⊕ F a non-dominated continuous
splitting over a compact invariant and isolated set containing a singularity of the flow.
Moreover,
is the α-limit set of all points of the plane R2 × {0} in the region bounded by
the curve
except the sink; see [23, chapter 4, section 9].
Item (2). Consider the flow known as the ‘Bowen example’; see e.g. [29] and figure 2. We have
chosen to reverse time with respect to the flow studied in [29] so that the heteroclinic
connection W = {s1 , s2 } ∪ W1 ∪ W2 ∪ W3 ∪ W4 is now a repeller. The past orbit
under this flow φt of every point z ∈ S 1 × [−1, 1] = M not in W accumulates on
either side of the heteroclinic connection, as suggested in the figure, if we impose
the condition λ− − + +
1 λ2 < λ1 λ2 on the eigenvalues of the saddle fixed points s1 and s2
(for more specifics on this see [29] and references therein). In this setting, the saddle
singularities s1 , s2 are area expanding: | det Dφt (si )| grows exponentially with t > 0,
i = 1, 2.
2404 V Araujo et al

s2 s3
s4 s1

S
Figure 3. The Bowen flow embedded in a sphere.

We can now embed this system in the 2-sphere putting two sinks at the ‘north and
south poles’ of S2 ; see figure 3. Let us denote the corresponding vector field on S2
by X.
Let Y be a vector field in S1 corresponding to the ‘north-south’ system: the
corresponding flow has only two fixed points N and S, N is a source and S a sink.
We can choose the absolute value of the eigenvalues of Y at N, S to be between the
absolute value of the eigenvalues of s1 , s2 .
In this way, for the flow {Zt }t∈R associated to the vector field X × Y on S2 × S1 , the
singularities {si × S}i=1,2 of the invariant set
:= W × {S} are hyperbolic saddles,
and the splitting T
(S2 × S1 ) = E ⊕ F given by
E(p,S) = {0} × TS S1 and F(p,S) = Tp S2 × {0} for p ∈ W,
is continuous, DZt -invariant, E is uniformly contracted and F is area expanding (by
an argument similar to the previous example), but it is not dominated.
Example 3 and items (1) and (2) of proposition 1 show that theorem A is false without
the dominating condition on the splitting E ⊕ F at the set of singularities within
.
Item (3). Now we present a suspension flow whose base map has a dominated splitting but the
flow does not admit any dominated splitting.
Let f : T4 × T4 be the diffeomorphism described in [8] which admits a continuous
dominated splitting E cs ⊕E cu on T4 , but does not admit any hyperbolic (uniformly contracting
or expanding) subbundle. There are hyperbolic fixed points of f satisfying (see figure 4):
• dim E u (p) = 2 = dim E s (p) and there exists no invariant one-dimensional subbundle of
E u (p);
• dim E u (
p ) = 2 = dim E s (

p ) and there exists no invariant one-dimensional subbundle of
s
E (
p );
• dim E s (
q ) = 3 and dim E u (q) = 3.
Hence, the suspension flow of f with constant roof function 1 does not admit any dominated
splitting. In fact, the natural invariant splitting E cs ⊕ E X ⊕ E cu is the continuous invariant
Dominated splittings 2405

q
s p

Figure 4. Saddles with real and complex eigenvalues.

splitting over T4 × [0, 1] with bundles of least dimension, and is not dominated since at the
point p the flow direction E X (p) dominates the E cs (p) = E s (p) direction, but at the point q
this domination is impossible.
This completes the proof of proposition 1.

7.2. Proof of proposition 2

We present first some auxiliary notions.
Let f : M → M be a C 1 -diffeomorphism and  ⊂ M a compact f -invariant set, that is,
f () = . We say that  is transitive if there exists x ∈  with dense orbit: the closure of
{f n (x) : n  1} equals . The invariant set  is topologically mixing if, for each pair U, V
of non-empty open sets of , there exists N = N (U, V ) ∈ Z+ such that U ∩ f n (V ) = ∅ for
all n > N .
We say that  is an attracting set if there exists a neighbourhood U of  in M such that
∩n
0 f n (U ) = ; in this case we say that U is an isolating neighbourhood for . An attractor
is a transitive attracting set and a repeller is an attractor for the inverse diffeomorphism f −1 .
It is well known that hyperbolic attractors or repellers can be decomposed into finitely many
compact subsets which are permuted and each of these pieces is topologically mixing for a
power of the original map; see e.g. [25].
Now, we present the example stated in proposition 2. Consider a hyperbolic attractor
(a Plykin attractor, see [24]) , with an isolating neighbourhood U , defined on the two-
dimensional sphere S2 for a diffeomorphism f : S2 → S2 , with a splitting T S2 = E ⊕ F
satisfying Df |E 
λs and Df |F   λu where 0 < λs < 1 < λu and λs · λu < 1.
Now we consider the following diffeomorphism
g = f1 × f2 : M → M, with M = S2 × S2 ,
where f1 := f 2 and f2 := f −1 . We note that 1 := ∩n
0 f1n (U ) is a hyperbolic attractor
with respect to f1 whose contraction and expansion rates along its hyperbolic splitting
T1 S2 = E1 ⊕ F1 are λ2s , λ2u , respectively. Likewise 2 := ∩n0 f2n (U ) is a hyperbolic
repeller with respect to f2 whose contraction and expansion rates along its hyperbolic splitting
T2 S2 = E2 ⊕ F2 are λ−1 −1
u , λs , respectively. This implies in particular that Df1 contracts
area near 1 , that is, λs · λu < 1 and that Df2 expands area near 2 , that is, (λs · λu )−1 > 1.
2 2

Moreover, the set

 := 1 × 2
is a hyperbolic set for g, whose hyperbolic splitting is given by (E1 × E2 ) ⊕ (F1 × F2 ).
Now we note, on the one hand, that the splitting T M = E ⊕ F with E = E1 and
F = F1 ⊕ E2 ⊕ F2 is not dominated, since the contraction rate along E2 is stronger than
2406 V Araujo et al

the contraction rate along E1 , but E is uniformly contracted and F is uniformly sectionally-
expanded, since the Lyapunov exponents 2 log λu , − log λu , − log λs are such that each pair
has positive sum.
On the other hand, the decomposition T M = T1 S2 ×T2 S2 = E ⊕F where E = E1 ⊕F1
and F = E2 ⊕ F2 is continuous, since each subbundle is continuous; is also Dg-invariant;
but it is not a dominated decomposition. In fact, we have the contraction/expansion rates
0 < λ2s < 1 < λ2u associated to E and the corresponding rates λ−1 −1
u , λs associated to F ,
−1 −1
satisfying the relation λs < λu < 1 < λs < λu (since 0 < λs · λu < 1 and 0 < λs < 1 < λu
2 2

imply that λ2s < λ−1 u and λs

−1
< λ2u ). Therefore the splitting E ⊕ F cannot be dominated.
In addition, as noted above, both Dg −1 |E and Dg |F expand area, so that E is sectionally
contracted and F is sectionally expanded.
Finally, since both f1 |1 and f2 |2 are topologically mixing, then g | is transitive; see
e.g. [32].
This example shows that there are transitive hyperbolic sets with a non-dominated,
although continuous, splitting satisfying the sectional-expansion and uniform contraction or
sectional-contraction conditions. This completes the proof of proposition 2.

Acknowledgments

This is a part of the PhD thesis of LS at Instituto de Matemática-Universidade Federal do Rio de

Janeiro under a CNPq (Brazil) scholarship and she is now supported by an INCTMat-CAPES
post-doctoral scholarship at IMPA.
The authors thank C Morales and L Dı́az for many observations that helped improve the
statements of the results.
Authors were partially supported by CNPq, PRONEX-Dyn.Syst., FAPERJ and FAPESB.

References

[1] Araújo V and Pacifico M J 2010 Three-dimensional Flows (Results in Mathematics and Related Areas, 3rd Series,
A Series of Modern Surveys in Mathematics) (Heidelberg: Springer)
[2] Arbieto A 2010 Sectional lyapunov exponents Proc. Am. Math. Soc. 138 3171–8
[3] Arnold L 1998 Random Dynamical Systems (Berlin: Springer)
[4] Barreira L and Pesin Y 2007 Dynamics of systems with nonzero Lyapunov exponents Nonuniform Hyperbolicity
(Encyclopedia of Mathematics and its Applications vol 115) (Cambridge: Cambridge University Press)
[5] Bonatti C, Dı́az L J and Viana M 2005 A global geometric and probabilistic perspective, mathematical physics,
III Dynamics Beyond Uniform Hyperbolicity (Encyclopaedia of Mathematical Sciences vol 102) (Berlin:
Springer)
[6] Bonatti C, Gourmelon N and Vivier T 2006 Perturbations of the derivative along periodic orbits Ergod. Theory
Dyn. Syst. 26 1307–37
[7] Bonatti C, Pumariño A and Viana M 1997 Lorenz attractors with arbitrary expanding dimension C. R. Acad.
Sci. Paris Sér. I Math. 325 883–8
[8] Bonatti C and Viana M 2000 SRB measures for partially hyperbolic systems whose central direction is mostly
contracting Israel J. Math. 115 157–93
[9] Bowen R 1975 Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms (Lecture Notes in
Mathematics vol 470) (Berlin: Springer)
[10] Bowen R and Ruelle D 1975 The ergodic theory of Axiom A flows Invent. Math. 29 181–202
[11] Doering C I 1987 Persistently transitive vector fields on three-dimensional manifolds Dynamical Systems and
Bifurcation Theory (Pitman Research Notes in Mathematics vol 160) ed M I Camacho, M J Pacifico and
F Takens (London: Longman) pp 59–89
[12] Gan S and Wen L 2006 Nonsingular star flows satisfy axiom A and the no-cycle condition Invent. Math.
164 279–315
[13] Katok A and Hasselblatt B 1995 Introduction to the Modern Theory of Dynamical Systems (Encyclopeadia of
Applied Mathematics vol 54) (Cambridge: Cambridge University Press)
Dominated splittings 2407

[14] Li M, Gan S and Wen L 2005 Robustly transitive singular sets via approach of an extended linear Poincaré flow
Discrete Contin. Dyn. Syst. 13 239–69
[15] Liang C and Liu G 2008 Dominated splitting versus small angles Acta Math. Sin. (Engl. Ser.) 24 1163–74
[16] Liao S T 1980 On the stability conjecture Chin. Ann. Math. 1 9–30
[17] Mañé R 1982 An ergodic closing lemma Ann. Math. 116 503–40
[18] Metzger R and Morales C 2008 Sectional-hyperbolic systems Ergod. Theory Dyn. Syst. 28 1587–97
[19] Morales C A, Pacifico M J and Pujals E R 2004 Robust transitive singular sets for 3-flows are partially hyperbolic
attractors or repellers Ann. Math. (2) 160 375–432
[20] Moryiasu K 1991 The extension of dominated splittings for C 1 -regular maps Tokyo J. Math. 14 425–37
[21] Newhouse S 2004 Cone-fields, domination, and hyperbolicity Modern Dynamical Systems and Applications
(Cambridge: Cambridge University Press) pp 419–32
[22] Palis J and de Melo W 1982 Geometric Theory of Dynamical Systems (Berlin: Springer)
[23] Perko L 2001 Differential Equations and Dynamical Systems (Texts in Applied Mathematics vol 7) 3rd edn
(New York: Springer)
[24] Plykin R V 1974 Sources and currents of A-diffeomorphisms of surfaces Math. USSR Sbr. 94 243–64
[25] Robinson C 2004 An Introduction to Dynamical Systems: Continuous and Discrete (Upper Saddle River, NJ:
Pearson/Prentice Hall)
[26] Sinai Y 1968 Markov partitions and C-diffeomorphisms Func. Anal. Appl. 2 64–89
[27] Smale S 1967 Differentiable dynamical systems Bull. Am. Math. Soc. 73 747–817
[28] Sternberg E 1958 On the structure of local homeomorphisms of euclidean n-space: II Am. J. Math. 80 623–31
[29] Takens F 1995 Heteroclinic attractors: time averages and moduli of topological conjugacy Bull. Braz. Math.
Soc. 25 107–20
[30] Tucker W 1999 The Lorenz attractor exists C. R. Acad. Sci. Paris 328 (Série I) 1197–202
[31] Viana M 1998 Dynamics: a probabilistic and geometric perspective Proc. Int. Congress of Mathematicians
(Berlin, Germany, 1998) vol 1, pp 557–78
[32] Walters P 1982 An Introduction to Ergodic Theory (Berlin: Springer)
[33] Wen L 2002 Homoclinic tangencies and dominated splittings Nonlinearity 15 1445–69