Randomness in hyperbolic dynamics
Jamerson Bezerra
Rio de Janeiro July 20, 2020 Instituto Nacional de Matematica´ Pura e Aplicada
Jamerson Bezerra
RANDOMNESS IN HYPERBOLIC DYNAMICS
Thesis presented to the Post-graduate Program in Mathe- matics at Instituto Nacional de Matematica´ Pura e Aplicada as partial fulfillment of the requirements for the degree of Doctor in Philosophy in Mathematics.
Advisor: Marcelo Viana Co-advisor: Carlos Gustavo Moreira
Rio de Janeiro 2020 “The first rule of discovery is to have brains and good luck. The second rule of discovery is to sit tight and wait till you get a bright idea.” George Polya´ ACKNOWLEDGMENTS
I would like to express my gratitude to my advisor, Professor Marcelo Viana, for the enlightening conversations and for his support throughout the entire process of construction of this work. I would also to thank my co-advisor, Professor Carlos Gustavo Moreira (Gugu), for his persistence in convince me to work with the second part of this thesis and for accept me as his student in the last months of my Ph.D. period. Without him this work would not be possible. I would like to reinforce that Viana and Gugu’s support and presence during the pandemic was essential to the conclusion of this work and for that I am really thankful. I am particular grateful to, my collaborator and friend, Professor Mauricio Poletti, for accept to work with me, receive me at Orsay where we were able to finish the first part of this thesis, and for be one of the responsible in my choice of dynamical systems as my work area. I would like to offer my special thanks to Sandoel Vieira, Thomas´ Spier, Hugo Araujo,´ Sergio Ibarra and Ivan Passoni for them availability and patience to hear me and for all the important commentaries and observations about the work with- out which this thesis would not be the same. I wish to acknowledge the help of Gregory Cosac, Fernando Lenarduzzi and Marlon Flores. In uncountable moments I disturb them, among other things, with corrections of English, review of my projects and works and without think twice they were always available to help me with these tasks. For that I have no words to say but thank you. I would like to specially thanks my friends that help me in my academic jour- ney for the countless advices and inspirations. Just to mention some of them: Alan Anderson, Andres´ Chavez,´ Argenis Garcia, Caio Doria,´ Davi Lima, Eduardo San- tana, Felipe Gonc¸alves, Luis Fernando, Mateus Sousa, Paulo Ribenboim, Pedro Gaspar,´ among others. I admire you and you are example for me. I would like to thanks my awesome wife, M´ıriam Pessoa.ˆ She gave me all the support that a needed to walk through this path. Her patience to deal with me can not be found in another human been. She gave me countless meaningful advices in numerous situations and for this reason she is as important as me for this thesis. There is not a way to compile a set words in such a way to express my gratitude to her. I only can say thank you my dear. I would like to say thank you to my mother and my father to raise me, in very adverse circumstances, prioritizing education at every step of the path. They always believed that the unique way to have better opportunities in the life is through education and so gave me all the necessary conditions that bring me until here. Finally, I wish to thanks all the friends that I made during this journey. Beyond all the above mentioned I would like to mention: Marcos, Gisele and Roberto, Clea and Maxwell, Thais and Alcides, Eduardo Garcez, Eduardo dos Santos, Vi- tor, Ermerson and many others. Thank you all.
4 ABSTRACT
This thesis is a contribution to the widely studied theory of hyperbolic dynamics. The work is structured in two independent parts which are intrinsically related: Lyapunov exponents for linear cocycles (model for non-uniformly hyperbolic dy- namics) and geometric properties of horseshoes (uniformly hyperbolic dynam- ics). In the first part, we address the problem of continuity and simplicity of the Lyapunov spectrum for random product k-tuples of quasi periodic cocycles. In dimension two, we prove that for any r ≥ 1 there exists a C0 open and Cr dense set of k-tuples of Cr quasi periodic cocycles whose random product is a continuity point of the Lyapunov exponents with positive value. Restricting to the k-tuples of Schrodinger¨ cocycles, the same results holds. In higher dimensions, we prove that among the Cr, r ≥ 1, k-tuples of quasi periodic cocycles with one coordinate being a diagonal cocycle, there exists a Cr dense and C1 open set such that the ran- dom product defined by cocycles in this set has simple Lyapunov spectrum and is a C0 continuity of the Lyapunov exponents. In the second part of this work, using the Erdos¨ probabilistic method, we prove that for typical Cr horseshoes with s- splitting, one dimensional weak stable bundle and upper stable dimension smaller than one there exists a subhorseshoe with almost the same upper stable dimension which is contained in a C1+ locally invariant submanifold tangent to the center unstable direction.
Key-words: Lyapunov exponents. Quasi-periodic cocycles. Higher dimen- sional horseshoes. Reduction of the dimension. CONTENTS
Acknowledgments 3
Abstract 5
I Random product of quasi-periodic cocycles 9
1 Introduction 10
2 Definitions and Statements 12
3 Holonomies 15
4 Proof of the results 18 4.1 Proof of Theorem A ...... 18 4.2 Proof of Theorem B ...... 21 4.3 Proof of Theorem C ...... 23 4.4 Proof of Proposition I.10 ...... 26
II Embedding small horseshoes in lower dimensions 32
5 Introduction 33 5.1 History and motivation ...... 33 5.2 Results ...... 36 5.3 Work in progress ...... 37
6 Horseshoes 39 6.1 Dominated splitting and partially hyperbolic sets ...... 39 6.2 Horseshoes ...... 41 6.3 Invariant manifolds ...... 44 6.4 Markov partitions ...... 46
6 6.5 Symbolic dynamics ...... 47 6.6 Symbolic dynamics notation ...... 48
7 Metric interactions between horseshoes and its symbolic models 53 7.1 Stable diameters ...... 54 7.2 Upper dimensions ...... 56
8 Horseshoes with s-splitting and one dimension weak stable bundle 60 8.1 Bounded distortion estimates ...... 61 8.2 Applications of bounded distortion ...... 64 8.3 Counting pieces ...... 68 8.4 Extractions ...... 72
9 Random perturbations and projections 76 9.1 Random perturbations ...... 77 9.2 Translating randomly in the weak stable directions ...... 78 9.3 Nonrecurrence and transversality condition ...... 81 9.4 Marstrand like argument ...... 85
10 Proof of the main theorem 89
7 LIST OF FIGURES
3.1 Stable holonomy ...... 16 3.2 Linear stable holonomy ...... 16
3.3 Holonomy Ht ...... 16
4.1 Weak twisting condition...... 19 4.2 Example of algebraic set. Cartan umbrella given by the equation z(x2 − y2) − x3 = 0...... 26 1 1 4.3 Transversality between V0 and A(S ) = {Ht; t ∈ S }...... 28
0 5.1 Λg embedded in S ...... 37
6.1 Smale horseshoe...... 41 6.2 Homoclinic intersection in dimension 2...... 42 6.3 s-splitting...... 42 6.4 Non s-splitting...... 43 6.5 Markov Partition with small diameter...... 47 6.6 From stable cylinder to unstable cylinder...... 52
7.1 Piece P(a) with diameter larger than ζ...... 54 7.2 Stable diameter ...... 55
9.1 Random perturbations ...... 79 9.2 Projection of the boxes ...... 85
10.1 Picking boxes with disjoint projections...... 89
8 Part I
Random product of quasi-periodic cocycles
9 1 INTRODUCTION
When studying the Lyapunov exponents of linear cocycles two main questions appears frequently: Are the exponents continuous with respect to the cocycle and how frequently do we have the maximum number of different exponents? We say that the Lyapunov spectrum of a cocycle is simple if it has the maxi- mum number of different exponents (this is the case when all Oseledets spaces are one dimensional). If we deal with two dimensional cocycles, this means our ma- trices take values in SL2(R), much more is known than in the higher dimensional case, when it takes values in GLd(R), d > 2. It was proved by Bochi [8] that in the C0 topology generically two dimen- sional cocycles either have uniform hyperbolicity or have only one exponent. An- other result, by Avila [4], says that in the smooth topology there is a dense set of Cr-cocycles for 0 ≤ r ≤ ∞ with simple spectrum. In particular these two results imply that in the C0 topology the exponents do not behave continuously, more- over, the only continuity points are the hyperbolic and cocycles with only one exponent. This motivates the following question: Are there open and dense sets of cocy- cles with simple spectrum, or, in other words, is the simplicity of the spectrum a generic property in smooth topology? For two dimensional cocycles over bases with some hyperbolic behaviour this was proved in many scenarios (see [48], [1], [42]). For two dimensional smooth quasi-periodic cocycles, (Cr topology, 0 < r ≤ ∞) Wang and You proved in [50] that the set with simple spectrum is not open, in particular the exponents are not continuous with respect to the cocycle. For higher dimensional cocycles the problem of simplicity becomes more deli- cate, this problem goes back to the works of Guivarc’h-Raugi [24] and Gol’dsheid- Margulis [22], where they work with random product of matrices. By random product of matrices we mean the cocycle generated by composing randomly a set of matrices accordingly to a probability in the group of matrices. They prove sim- plicity of the Lyapunov spectrum in this scenario with some generic conditions on the support of the probability measure.
10 For more general higher dimensional cocycles over hyperbolic maps, if some bunching conditions are assumed, simplicity was proved to be generic in the Cr topology, for 0 < r < ∞ ([10], [2], [6]). For quasi-periodic cocycles very few is known in higher dimension about sim- plicity, see for example [18] where they find some strong conditions to have sim- plicity of the biggest Lyapunov exponents. In this work we study the Lyapunov exponents of cocycles over dynamics that have both behaviors, a random part and a isometric (quasi periodic) part. Specifically we deal with random product of quasi-periodic cocycles by this we mean the cocycle generated by taking a probability measure in the set of quasi- periodic cocycles and iterating randomly according to this probability. We prove that, for two dimensional cocycles, there exists an open and dense set of cocycles that are points of continuity for the Lyapunov exponents and have simple spectrum, even in the C0 topology, and in higher dimension, with one of them taking values in the diagonal group, we find open and dense sets with simple spectrum. Observe that in contrast to the non-random case, where, in the C0 topology, generically we have one Lyapunov exponent outside of uniform hyperbolicity, in the random product scenario we get generically positive Lyapunov exponents also in the C0 topology. This change of behavior when we ad some randomness on the dynamics was already exploited in many cases, see for example [47], [27], [7].
11 2 DEFINITIONS AND STATEMENTS
Given an invertible measurable map f : M → M and a measurable application d d A : M → GLd(R), we define the linear cocycle as the map F : M × R → M × R given by
F(x,v) = ( f (x),A(x)v).
Usually we denote the linear cocycle by the pair ( f ,A) and, sometimes, when the map f is fixed (and there is no ambiguity) we denote just by A. Its iterates are given by Fn(x,v) = ( f n(x),An(x)v), where A( f n−1(x))···A(x), if n ≥ 1 An(x) = Id, if n = 0 A( f n(x))−1 ···A( f −1(x))−1 if n ≤ 1 In the case that f preserves a probability measure µ which is ergodic and ± log||A 1|| is µ-integrable, Oseledets theorem [37] says that there exists k ∈ N, real d 1 k numbers λ1(A) > ... > λk(A) and a decomposition of R = E (x)⊕···⊕E (x) by measurable subspaces Ei(x) such that for µ-a.e. x ∈ M we have
• A(x)Ei(x) = Ei( f (x)) for all i, and
n i • lim ||A (x)v|| = λi(A), for all v ∈ E (x). |n|→∞
The numbers λ1(A),··· ,λk(A) are called the Lyapunov exponents associated with the cocycle ( f ,A) and the set formed by them is called Lyapunov spectrum, we say that A has simple spectrum if k = d. When M = S1 and f is a rotation of angle θ ∈ (0,1] we say that the cocycle ( f ,A) is a quasi-periodic cocycle and usually write (θ,A). k Fix θi ∈ (0,1] for all i ∈ Ik = {0,...,k} and ν = ∑i=0 νiδi a probability measure 1 Z ˆ on Ik. We will denote by Leb, the Lebesgue measure on S . Let X = Ik , X = X × S1 and consider the invertible (locally constant) skew product fˆ : Xˆ → Xˆ , given by ˆ f ((xn)n∈Z,t) = ((xn+1)n∈Z,t + θx0 )
12 1 and observe that the measure µˆ = νZ ×Leb is fˆ-invariant. We also write fi : S → 1 f (t) = t + f n = f ◦···◦ f f −n = f −1 ◦···◦ f −1 n ∈ S , i θi, x xn−1 x0 and x x−n x1 , for all N and x ∈ X.
From now on we will fix θ0 irrational. In particular, we have that ( fˆ, µˆ ) is an ergodic system. 1 For measurable maps Ai : S → GLd(R), i ∈ Ik, we define the random product ˆ ˆ ˆ ˆ of the quasi periodic cocycles (θi,Ai)i∈Ik as the cocycle ( f ,A), where A : X → ˆ GLd(R) is given by A(x,t) = Ax0 (t). r 1 k+1 Note that each vector (A0,...,Ak) ∈ (C (S ,GLd(R))) defines a random product ( fˆ,Aˆ). We abuse of the notation and denote by Aˆ the above map and the
point (A0,...,Ak) which define Aˆ. Let 0 ≤ r ≤ ∞ (or r = ω for the analytic case), by Cs topology in the product r 1 k+1 space (C (S ,GLd(R))) , for s ∈ [0,r] (or s ∈ [0,∞] ∪ {ω} in the case when r = ω) we mean the topology given by the distance
d s (Aˆ,Bˆ) = maxd s 1 (A ,B ), C C (S ,GLd(R)) i i i∈Ik
r 1 k+1 where Aˆ,Bˆ ∈ (C ( ,GL ( ))) and d s 1 denote the distance that gen- S d R C (S ,GLd(R)) s r 1 erates the C topology in C (S ,GLd(R)). r 1 k+1 s We say that the Aˆ ∈ (C (S ,GLd(R))) is a C -continuity point for the Lya- r 1 punov exponents if for all sequences Aˆk ∈ C (S ,GLd(R)) converging to Aˆ in the s C -topology we have that the Lyapunov exponents of Aˆk converge to the Lyapunov exponents of Aˆ. These cocycles where already introduced in [5, section 5.2], to show that co- cycles over some partially hyperbolic dynamics are not continuous in general. If our cocycle A takes values in SL2(R) recall that we have λ1(A) = −λ2(A), in particular we have simple spectrum if and only if we have one positive expo- nent. If the cocycle takes values in the space of two by two matrices with positive determinant, we can reduce to SL2(R) just dividing by the square root of the de- terminant.
Theorem A. For r ∈ [0,∞] ∩ {ω}, there exists a C0 open and Cr dense subset r 1 k+1 of (C (S ,SL2(R))) , such that the random product defined by cocycles in this set has positive Lyapunov exponent and is a C0-continuity point for the Lyapunov exponents.
Let ϕ : S1 → R be a continuous function. The Schrodinger cocycle associated 1 1 1 to the function ϕ is defined as ( f ,Aϕ ), where f : S → S and Aϕ : S → SL2(R) is given by
13 ! ϕ(x) −1 Aϕ (x) = . 1 0
In the literature it is common to use the function ϕ(x) = E −u(x), with E ∈ R and u : S1 → R a continuous function. The reason for that notation is the relation of 2 2 the Schrodinger cocycles with the Schrodinger operator Hu,x : l (Z) → l (Z), n (Hu,x(z))n = zn+1 + zn−1 + u( f (x))zn, given by the eigenvalue equation
Hu,x(z) = E · z. For a detailed survey on this topic see [17]. 1 When f is a rotation of angle θ ∈ (0,1] in S , Schrodinger cocycles, ( f ,Aϕ ), are quasi periodic cocycles which we will denote just by (θ,Aϕ ). Theorem B. For r ∈ [0,∞] ∪ {ω}, there exists a C0 open and Cr dense subset r k+ of (C (S1,R)) 1, for 0 ≤ r ≤ ∞, such that the random product defined by the Schrodinger cocycles associated with the functions in this subset has positive Lya- punov exponent and is a C0-continuity point for the Lyapunov exponents.
In order to state the result in higher dimensions let Dd(R) be the subgroup of diagonal matrices in GLd(R). Theorem C. For d > 2 and r ∈ [0,∞] ∪ {ω}, there exists a Cr dense subset of r 1 r 1 k C (S ,Dd(R)) × (C (S ,GLd(R))) such that the random product defined by co- cycles in this set has simple Lyapunov spectrum and is a C0-continuity point of the Lyapunov exponents. If r ∈ [1,∞] ∪ {ω} this set is also C1 open. Observe that, since the Cr dense set in the Theorem C is formed by cocycles with simple Lyapunov spectrum and C0-continuity points of the Lyapunov expo- nents, we have a C0 open and Cr dense set with simple Lyapunov spectrum for each r ∈ [0,∞] ∪ {ω}. 1 1 1 Remark 1. We say that two cocycles A,B ∈ C(S ,GLd(R)) over f : S → S , are r r 1 C -cohomologous if there exists C ∈ C (S ,GLd(R)) such that A(t) = C−1( f (t))B(t)C(t).
Cohomologous cocycles have the same Lyapunov exponents. As a consequence,
Theorem C, is also valid for cocycles such that A0 belongs to the set of cocycles r C -cohomologous to cocycles taking values in Dd(R).
14 3 HOLONOMIES
Given x ∈ X, we define its stable set as
s W (x) = {y ∈ X, such that for some k ≥ 0, yi = xi for i ≥ k},
and the unstable set as
u W (x) = {y ∈ X, such that for some k ≤ 0, yi = xi for i ≤ k}.
We write x ∼s y if x and y are in the same stable set and similarly we write x ∼u y if x and y are in the same unstable set. s s 1 1 If x ∼ y we define the stable holonomy from x to y, hx,y : S → S , as
s n −1 n n0 −1 n0 hx,y = lim ( fy ) ◦ fx = ( fy ) ◦ fx . n→∞
where n0 is the smallest integer such that xi = yi for all i ≥ n0. Analogously, we u u 1 1 define for x ∼ y the unstable holonomy from x to y, hx,y : S → S , as
u n −1 n n0 −1 n0 hx,y = lim ( fy ) ◦ fx = ( fy ) ◦ fx . n→−∞
where n0 is the biggest integer such that xi = yi for all i ≤ n0.. Now considerx ˆ,yˆ ∈ Xˆ ,x ˆ = (x,t) andy ˆ = (y,t0). We writex ˆ ∼s yˆ if x ∼s y and 0 s u u 0 u t = hx,y(t) and we writex ˆ ∼ yˆ if x ∼ y and t = hx,y(t). Define the Linear Stable Holonomies associated with the random product ( fˆ,Aˆ) as the family of linear maps
s,Aˆ d d ˆ s {Hxˆ,yˆ : R → R ;x ˆ,yˆ ∈ X,xˆ ∼ yˆ},
given by
ˆ s,A n 0 −1 n n0 0 −1 n0 H 0 = lim (Aˆ (y,t )) Aˆ (x,t) = (Aˆ (y,t )) Aˆ (x,t). (x,t)(y,t ) n→∞ where n0 ≥ 1, is such that xi = yi for i ≥ n0. Analogously we define the Linear Unstable Holonomies
u,Aˆ d d ˆ u {Hxˆ,yˆ : R → R ;x ˆ,yˆ ∈ X,xˆ ∼ yˆ}.
15 Figure 3.1: Stable holonomy Figure 3.2: Linear stable holonomy
∗ ∗,Aˆ We use the notations ∼ or Hxˆ,yˆ , meaning that the sentence remains true for any ∗ ∈ {s,u}. Observe that for fixedx ˆ,yˆ ∈ Xˆ withx ˆ ∼∗ yˆ the map
ˆ 0 1 k+1 ∗,Aˆ A ∈ (C (S ,GLd(R))) 7→ Hxˆ,yˆ varies continuously in the C0 topology. Now we will fix some notations that will be essentials in the following sec- tions.
From now on, p ∈ X will be the fix point of the shift map defined by pi = 0 for all i ∈ Z, z ∈ X will the homoclinic intersection point defined by z0 = 1 and 0 u zi = 0 for all i 6= 0. Let t = hp,z(t) and define
s u Aˆ s,Aˆ u,Aˆ h = hz,p ◦ hp,z and Ht = H(z,t0)(p,h(t)) ◦ H(p,t)(z,t0). (3.1)
Figure 3.3: Holonomy Ht
16 1 ˆ Aˆ Observe that for any fixed t ∈ S the map A 7→ Ht varies continuously in the C0 topology. We omit the index Aˆ when the random product that we are dealing with is clear by the context.
17 4 PROOF OF THE RESULTS
4.1 Proof of Theorem A
In this section all cocycles takes values in SL2(R), in particular, for any map A, we have at most 2 exponents λ+(A) and λ−(A). So, in this section and in the next the term Lyapunov exponents refers to λ+ and this is not a restriction since λ+(A) = −λ−(A). ˆ Recall that the system ( f , µ) is ergodic, since we are assuming that θ0 ∈ R\Q. ˆ ˆ 0 1 k+1 Let ( f ,A) be the random product of (A0,...,Ak) ∈ (C (S ,SL2(R))) .
Definition I.1 (Weakly pinching). We say that the cocycle ( fˆ,Aˆ) is weakly pinch-
ing if the cocycle (θ0,A0) has λ+(A0) > 0 with respect to the Lebesgue measure on S1.
1 By Oseledets theorem, if λ+(A0) > 0, there exists a measurable map S 3 1 1 t 7→ (e+(t),e−(t)) ∈ RP × RP where e+(t) is the direction of the Oseledets decomposition corresponding to λ+(A0) and e−(t) the direction corresponding to λ−(A0). Take p, z, h and Ht as defined in the end of Section 3.
Definition I.2 (weakly twisting). We say that a weakly pinching cocycle ( fˆ,Aˆ) is weakly twisting if
Ht({e+(t),e−(t)}) ∩ {e+(h(t)),e−(h(t))} = /0
for a positive measure subset of t ∈ S1
Observe that both conditions, weakly pinching and weakly twisting, only de-
pends on (θ0,A0) and (θ1,A1). The pinching and twisting conditions defined here (and in section 4.3) are gen- eralizations of the ones introduced in [2] for cocycles over hyperbolic dynamics. We will call the cocycle ( fˆ,Aˆ) weakly simple, if it is both weakly pinching and weakly twisting.
18 Figure 4.1: Weak twisting condition.
Define the projective cocycle ˆ ˆ 1 ˆ 1 ˆ ˆ PFAˆ : X × RP → X × RP , (xˆ,[v]) 7→ ( f (xˆ),[A(xˆ)v]) ˆ and let m be an PFAˆ-invariant measure that projects on µ. Using Rokhlin’s Disin- tegration Theorem, we can find a measurable mapx ˆ 7→ mxˆ such that Z 1 m = mxˆdµ and mxˆ({xˆ} × RP ) = 1. Xˆ We say that m is u-invariant if there is a total measure set X0 ⊂ Xˆ such that for 0 u u everyx ˆ,yˆ ∈ X withx ˆ ∼ yˆ, we have (Hxˆ,yˆ)∗mxˆ = myˆ. Analogously, we say that m is s-invariant if the same is true changing unstable by stable holonomies. Finally, m is su-invariant if it is both s an u invariant. By [36, Proposition 3.9], weakly simple implies that the projective cocycle,
PFAˆ, do not admit any su-invariant measure. We have the following criteria. ˆ r 1 k+1 Proposition I.3. Let r ∈ [0,∞]∪{ω} and take A ∈ (C (S ,SL2(R))) . If the ran- dom product ( fˆ,Aˆ) is weakly simple then, there exists a C0 neighborhood of Aˆ in r 1 k+1 (C (S ,SL2(R))) , such that any random product of cocycles in this neighbor- hood, has postive Lyapunov exponent and is a C0-continuity point of the Lyapunov exponent. Proof. The invariance principle of [3] (see [42, Theorem 6.2] for a version that
fits into our settings) says that if λ+(Aˆ) = 0 then any m, PFˆ invariant measure, is su-invariant. Since, ( fˆ,Aˆ) is weakly simple, we conclude that λ+(Aˆ) > 0. Take Aˆ → Aˆ such that λ (Aˆ ) λ (Aˆ). Consider m , PFˆ invariant measure k + k 9 + k Aˆk that projects to µˆ , u-invariant such that
Z Aˆk(xˆ)v λ (Aˆ ) = log dm . + k kvk k
19 Up to taking a sub-sequence we can assume that
∗ mk * m and λ+(Aˆk) → a < λ+(Aˆ), ˆ where m is an PFAˆ-invariant measure that projects to µ. + As λ+(Aˆ) > 0, otherwise it is a continuity point, we have that m = αm + − ∗ R βm , with α + β = 1 where m = δe∗(xˆ)dµ(xˆ) and e∗(xˆ) is the Oseldets sub- space corresponding to λ∗(Aˆ) for ∗ = + or −. So,
Z Aˆ(xˆ)v log dm = αλ (Aˆ) + βλ (Aˆ). kvk + −
By assumption we have that αλ+(Aˆ)+βλ−(Aˆ) < λ+(Aˆ), which implies that α < 1. Recall that the holonomies varies continuously in the C0 topology, then for ∗ ∗,Aˆk ∗,Aˆ everyx ˆ ∼ yˆ with ∗ = s or u, we have that Hxˆ,yˆ converges uniformly to Hxˆ,yˆ , so we can apply [42, Theorem A.1] to conclude that m is u-invariant. Observe that − − 1 + − the measure m is s-invariant and m = β (m−αm ), then m is also u-invariant. Analogously we conclude that m+ is su-invariant. Since ( fˆ,Aˆ) does not admit su-invariant measures we conclude that Aˆ is a continuity point of the Lyapunov exponent λ+. ˆ r 1 k+1 0 Assume now that there exists a sequence Ak ∈ (C (S ,SL2(R))) , of C - discontinuity points of the Lyapunov exponent, converging to Aˆ in the C0 topol- ogy. Repeting the above argument we can see that for each k ∈ N we can find a PF -invariant measure, that projects to µˆ that is su-invariant. Passing to a subse- Aˆk quence we can find m such that mk converges to it, again by [42, Theorem A.1] ˆ m is an su-invariant measure for PFAˆ. This contradicts the fact that the random product ( fˆ,Aˆ) does admit any su-invariant measure. 0 ˆ r 1 Therefore we can find a C neighborhood of A ∈ (C (S ,SL2(R))) such that the random product defined by the cocycle in this neighborhood is a C0 continuity point of the Lyapunov exponent.
To conclude the proof of theorem A, we just need to prove that the weakly simple random products are Cr dense.
r r 1 2 Proposition I.4. For r ∈ [0,∞]∪{ω}, there exists a C dense set of C (S ,SL2(R)) such that for any (A0,A1) in this set, the random product of (θi,Ai)i∈Ik is weakly r 1 k−1 simple for all (A2,...,Ak) ∈ (C (S ,SL2(R))) . r Proof. By [4] we can find a C -dense set of A0 such that λ+(A0) > 0. In other words we have a Cr dense set weakly pinching.
20 −1 1 Now observe that in our setting Ht = A0(h(t)) A1(t). Take K ⊂ S with 1 Leb(K) > 2 such that t 7→ (e+(t),e−(t)) is continuous in K. Thus, since h pre- serves the Lebesgue measure we have that Leb(K ∩ h−1(K)) > 0. Take t ∈ K ∩ h−1(K) such that t is a density point for the Lesbegue measure. If
Ht{e+(t),e−(t)} ∩ {e+(h(t)),e−(h(t))} = /0,
then, by continuity, there exists a neighborhood of t in K ∩ h−1(K), containing a set with positive measure satisfying the same property which implies that the random product is twisting.
If Ht{(e+(t),e−(t)}∩{e+(h(t)),e−(h(t)} 6= /0we change A1 by A˜1 = A1 ◦Rθ , for θ small.
This implies the holonomy of (A0,A˜1) is given by H˜t = HtRθ . Then we can take θ arbitrarily small such that
H˜t ({e+(t),e−(t)}) ∩ {e+(h(t)),e−(h(t))} = ∅.
Consequently, (θi,Ai) is weakly twisting concluding the proof.
4.2 Proof of Theorem B
In the case of Schrodinger¨ cocycles the perturbation to get weakly simple cocy-
cles is more delicate because we can only perturb ϕi, i ∈ Ik. So to conclude the Theorem B, using I.3, we just need to prove the following Proposition.
r Proposition I.5. For r ∈ [0,∞] ∪ {ω}, there exists a dense subset of (C (S1,R))2 such that, for any (ϕ0,ϕ1) on this subset, the random product of the Schrodinger r 1 k−1 cocyles (θi,Aϕi )i∈Ik , is weakly simple, for any (ϕ2,...,ϕk) ∈ (C (S ,R)) . Proof. In [4], it is proved that for any r ∈ [0,∞] there is a dense subset of maps 1 r 1 ϕ : S → R in C (S ,R) such that the cocycle (θ0,Aϕ ) has positive Lyapunov exponent, which in our case is equivalent to say that for a dense subset of ϕ0 ∈ r 1 C (S ,R), the random product of (θi,Aϕi ), with i = 0,...,k, is weakly pinching. r 1 k+1 Take a point (ϕ0,...,ϕk) ∈ (C (S ,R)) such that the random product of
(θi,Aϕ )i∈Ik is weakly pinching and assume, without loss of generality, that ϕ0 does not vanish identically (this can be made since we have density of weakly
pinching). Consider p,z,h and Ht as in Section 3 and observe that ! −1 1 0 Ht = (Aϕ0 (h(t))) Aϕ1 (t) = , (4.1) ϕ0(h(t)) − ϕ1(t) 1
21 and s u h(t) = hz,p ◦ hp,z(t) = t + (θ1 − θ0) is a rotation and, in particular, preserves the Lebesgue measure. Observe that the matrix Ht preserve the vertical axis, i.e. e2 = (0,1) is the unique fixed point of the action of the matrix Ht in the projective space when ϕ1(t) 6= ϕ0(h(t)). r With a small change of ϕ1 in the C topology we can assume that ϕ0(h(t)) 6= 1 ϕ1(t) for Leb-a.e. t ∈ S (for example adding a suitable constant to ϕ1). Hence, e2 is the unique direction in the projective space which is invariant by Ht. Consider the following set
1 L2 = {t ∈ S ;e2 ∈ {e+(t),e−(t)}},
where e+ and e− are the Oseledets subspaces associated to the cocycle (θ0,Aϕ0 ) (which we know that has positive Lyapunov exponent). 1 We claim that Leb(S \L2) > 0. Indeed, otherwise, we have that e2 is a Os- 1 eledets subspace of the cocycle (θ0,Aϕ0 ) for Leb-a.e. t ∈ S . Then, we have the following possibilities:
1 1. e2 = e+(t) = e+( fθ0 (t)), for some t ∈ S : In this case we have,
e2 = e+( fθ0 (t)) = Aϕ0 (t)e+(t) = (−1,0),
which is a contradiction.
1 2. e2 = e+(t) = e−( fθ0 (t)), for Leb-a.e. t ∈ S : We have,
e = e ( f 2 (t)) = A ( f (t))A (t)e (t) = (− (t),1), 2 + θ0 ϕ0 θ0 ϕ0 + ϕ0
1 for Leb-a.e. t ∈ S . This shows that ϕ0 vanishes identically, a contradiction. 1 Therefore, let δ = Leb(S \L2) ∈ (0,1) (the case δ = 1 is simpler and follows 1 analogously). By Lusin’s theorem, there exists Γ ⊂ S , such that the functions e+ 2−δ and e− are simultaneously continuous in Γ and Leb(Γ) > 2 . Observe that the condition on the measure of Γ and the fact that h preserves the Lebesgue measure shows that,
Leb(Γ ∩ h−1(Γ)) > 1 − δ.
22 Then, −1 1 Leb(Γ ∩ h (Γ) ∩ (S \L2)) > 0. −1 1 Let t ∈ Γ ∩ h (Γ) ∩ (S \L2) be a density point for the Lebesgue measure. Since e2 ∈/ {e+(t),e−(t)} we have e2 ∈/ Ht({e+(t),e−(t)}). Assume that
Ht({e+(t),e−(t)}) ∩ {e+(h(t)),e−(h(t))} 6= ∅.
Then, changing ϕ1 in a small neighborhood of t and using (4.1) we can make
Ht({e+(t),e−(t)}) ∩ {e+(h(t)),e−(h(t))} = ∅.
Since t is a density point, Ht, e+ and e− are continuous in t, we have that the above property is preserved for a positive measure neighborhood of t. r 1 So we conclude that, fixed ϕ0 ∈ C (S ,R) such that the cocycle (θ0,Aϕ0 ) r 1 has positive Lyapunov exponent, there exists a dense subset of ϕ1 ∈ C (S ,R) such that the random product of (θi,ϕi) is weakly twisting, for any ϕ2,...,ϕk ∈ r C (S1,R) and so, weakly simple.
4.3 Proof of Theorem C
From now on our cocycles take values on GLd(R), for d > 2. As always, let p,z,h and Ht be as in section 3.
Definition I.6 (Pinching). We say that the random product ( fˆ,Aˆ) is Pinching if the
Lyapunov exponents λ1(p),··· ,λl(p) of the cocycle ( f0,A0) satisfies that l = d and for any 1 ≤ j ≤ d − 1 and Leb-a.e. t ∈ S1 the sums
λi1 (p) + ··· + λi j (p), (4.2)
for all sequences 1 ≤ i1 < ... < i j ≤ d, are distinct.
Take I and J subsets of {1,··· ,d} with the same cardinality and consider the map PI,J : GLd(R) → R defined as the determinant of the matrix obtained taking the minor associated with the lines designated by the elements in I and columns designated by elements of J.
Definition I.7 (Twisting). We say that the cocycle ( fˆ,Aˆ) is Twisting if for any I and J as above Z |log|PI,J(Ht)||dt < ∞. S1
23 Remark 2. The definition of twisting given above is adapted to suit our purposes. In general, a cocycle ( fˆ,Aˆ) is said twisting if the growth rate of all minors of
H ˆn are sub exponential, i.e., fxˆ (t)
lim log PI,J(H ˆn ) = 0, n→∞ fxˆ (t) for all I,J. This, basically, says that the angle between the Oseledets subspaces goes to zero at most sub exponentially.
We say that the random product ( fˆ,Aˆ) is Simple if it is both Pinching and Twisting. Observe that even if d = 2 being simple is a stronger condition than being weakly simple. In order to prove Theorem C we need the following result which is a version of the main result of [43].
Theorem I.8. If the cocycle ( fˆ,Aˆ) is simple, then the Lyapunov spectrum is sim- ple and it is a continuity point with respect to the C0 topology of the Lyapunov exponents.
In [43] the result is stated for Holder¨ cocycles with some more general dy- namics, but as mentioned in [43, section 4.1] we only need to have well defined holonomies that varies continuously with respect to the cocycle. 1 1 Observe A0 : S → Dd(R) is defined by d functions a1,...,ad : S → R such that (A0(t))i,i = ai(t) and (A0(t))i, j = 0 for i 6= j. Then, by Birkhoff’s ergodic theorem, the Lyapunov spectrum of ( f0,A0) is the set Z log(ai)dLeb .
1 So, after suitable choose of bi > 0, we can definea ˜i : S → R given bya ˜i(t) = biai(t) such that the diagonal cocycle A˜0 defined usinga ˜ has the property 4.2. Moreover, this is a C0 open condition. r Hence after a C small perturbation of A0 we can assume that the random product ( fˆ,Aˆ) of (A0,A1,...,Ak) is always pinching for any (A1,...,Ak) in r 1 k C (S ,GLd(R)) , and r ∈ [0,∞] ∪ {ω}. If ( fˆ,Aˆ) is twisting, then, by Theorem I.8 we have that ( fˆ,Aˆ) has simple Lya- punov spectrum and is a C0-continuity point of all Lyapunov exponents. In partic- r 1 k+1 0 ular, any (B0,...,Bk) ∈ (C (S ,GLd(R))) which is C -close to (A0,...,Ak) has also simple Lyapunov spectrum. So, to conclude the proof of the Theorem C it is enough to prove the following theorem:
24 Theorem I.9. Let d > 2, r ∈ [0,∞] ∪ {ω}. Then the set of maps A1 in r 1 ˆ ˆ C (S ,GLd(R)), such that the random product ( f ,A) of (θi,Ai)i∈Ik is twisting, is Cr dense. Moreover for r ∈ [1,∞] ∪ {ω} this set is also C1 open.
r 1 Proof. Assume first that r ∈ [1,∞]. Consider A1 ∈ C (S ,GLd(R)) and denote by ( fˆ,Aˆ) the random product of (θi,Ai). Note that, for p,z ∈ X as in Section 3,
s −1 u hz,p(t) = f0 ◦ f1 and hp,z = I.
s u −1 u So, h = hz,p ◦ hp,z = f0 ◦ f1. Moreover, since t = hp,z(t) and observing that
s −1 u H(z,t)(p,h(t)) = A0(h(t)) A1(t) and H(p,t)(z,t) = I, we get that
s u −1 r 1 Ht = H(z,t)(p,h(t)) ◦ H(p,t)(z,t) = A0(h(t)) A1(t) ∈ C (S ,GLd(R)).
Therefore, to see that a random product ( fˆ,Aˆ) is twisting, we must show that for any I,J ⊂ {1,...,d} with same cardinality the map PI,J : GLd(R) → R satisfies
Z −1 log PI,J(A0(h(t)) A1(t)) dt < ∞. (4.3) S1 The next proposition (which will be proved in Section 4.4) is the technical −1 tool in the proof that property (4.3) above holds for A0(h(t)) A1(t) in a dense r 1 subset of C (S ,GLd(R)) and we will state it for more general polynomial maps (polynomials with the variable being the coordinates of the matrix in GLd(R)), than PI,J.
Proposition I.10. Let P : GLd(R) → R be a non-constant polynomial map, d ≥ 1 and r ∈ [1,∞] ∪ {ω}. Then, the set
r 1 1 1 AP = A ∈ C (S ,GLd(R));log|P ◦ A| ∈ L (S ,m) ,
r 1 is open and dense subset of C (S ,GLd(R)). Using the Proposition I.10 and the continuity of the invertible map r 1 r 1 φA0 : C (S ,GLd(R)) → C (S ,GLd(R)) given by
−1 φA0 (A)(t) = A0(h(t)) A(t),
25 r 1 ˆ ˆ we can see that the set of A1 ∈C (S ,GLd(R)) such that the random product ( f ,A) is twisting, that is,
−1 1 1 PI,J ◦ (A0(h(t)) A(t)) ∈ L (S ,GLd(R)),
for all maps PI,J (there are only finite of them) is open and dense. Hence, we conclude the theorem for r ∈ [1,∞] ∪ {ω}. 1 1 Assume now that r ∈ [0,1). Then, we can approximate A1 by B1 ∈ C (S , r GLd(R)) in the C topology and, after that, using the Proposition I.10 again, find 1 r D1 close to B1 in the C topology (and then close to A1 in the C topology) such
that the random product of (θi,Di)i∈Ik is twisting, where D j = A j for all j 6= 1. This concludes the result for any r ∈ [0,∞] ∪ {ω}.
4.4 Proof of Proposition I.10
First we will give a brief review about basic real algebraic geometry. A subset l V ⊂ R is said to be an Algebraic set if there exist finitely many polynomials f1,..., fm ∈ R[X1,...,Xl] such that
m \ V = [ fi = 0]. i=1
Figure 4.2: Example of algebraic set. Cartan umbrella given by the equation z(x2 − y2) − x3 = 0.
An algebraic set V is said irreducible if, whenever V = F1 ∪ F2, with Fi alge- braic sets, then V = F1 or V = F2. It is known, see [9, Theorem 2.8.3,page 50], that every algebraic set V can be written as the union of irreducible algebraic sets
26 V1,...,Vp such that Vi * ∪ j6=iVj. The sets Vi are called the irreducible components of V.
For a point x0 ∈ V, we define the Zariski tangent space of V at the point x0 as the linear space m n o T ZarV = \ x ∈ l f (x ) · x = , x0 R ;∇ i 0 0 i=1
where ∇ fi denotes the gradient vector of the polynomial fi. We say that a point x0 ∈ V is a regular point if
(T ZarV) = (T ZarV) x ∈ V , dim x0 min dim x ;
and x0 is a singular point of V if it is not regular. If V0 is an irreducible algebraic set we define the dimension of V0 as the num- Zar ber dim(Tx V) for any regular point x ∈V. For general algebraic sets V we define the dimension as
dimV = max{dim(Vi);i = 1,..., p},
where Vi are the irreducible components of V. It is important also to reinforce that, when V is an irreducible algebraic set and x ∈ V is a regular point of V, there exists a neighborhood of x in V which ∞ Zar is a C manifold and in this case Tx V is in fact the tangent space TxV of this neighborhood at x, see [9, page 66]. Let Sing(V) be the set of singular points of V (sometimes we will call this set singular part of V and its complement as regular part of V). It is also known, see [9, Proposition 3.3.14, page 69], that Sing(V) is an algebraic subset of V and
dim(Sing(V)) < dimV. (4.4)
That will be the crucial property in the proof of the Proposition I.10. For more details about algebraic and semi algebraic sets, see [9]. Consider P : GLd(R) → R a non-constant polynomial map. Fix r ∈ [1,∞] (we will deal with the case r = ω later) and consider the following algebraic set,
V 0 = [P = 0].
Observe that V 0 has dimension at most d2 − 1 (as defined above) and then its regular part is a regular submanifold of GLd(R) of dimension equal to dimension of V 0. The idea to prove the Theorem 4.4 is the following: if the closed curve
t 7−→ Ht
27 does not intersects the algebraic set V0, then there is nothing to prove (the log of a positive function in a compact is integrable). But, in general that is not the case, intersection between Ht and V0 can be large. If that is the case we can use a version of Thom transversality theorem I.11 (V0 is not a regular manifold) to perturb Ht suitably and guarantee that the intersection is transversal (which is enough to get integrability that we are pursuing) as described in the Figure 4.3.
1 1 Figure 4.3: Transversality between V0 and A(S ) = {Ht; t ∈ S }.
Let V 1 = Sing(V 0) ⊂ V 0, be the singular part of V 0.
Theorem I.11 (Thom transversality theorem). Fix r ∈ [1,∞]. Let M be a manifold and N ⊂ M be a submanifold. Then, the set
r 1 {A ∈ C (S ;M); A t N}, is dense. If N is closed then the above set is also open.
Remark 3. A particular case of this theorem that we will use is when codimension r of N is large than 1. In this case, the unique way to a map A ∈ C (S1;M) to be transversal to N is
1 A(S ) ∩ N = ∅.
For a complete proof and details around this result, see [23, Corollary 4.12, page 56]. As a corollary of Theorem I.11 in our context, we have
r 1 1 1 Lemma I.12. Take A ∈ C (S ;GLd(R)) and assume that d(A(S ),V ) > 0. Then, r 1 for all U ⊂ C (S ;GLd(R)), neighborhood of A, there exists V ⊂ U , open, and a > 0 such that for all B ∈ V we have:
28 •d (B(S1),V 1) > a > 0
0 •B t V .
Observe that as a corollary of this lemma and compacity of S1, we have
1 0 #B(S ) ∩V < ∞.
In fact, we can assume that the intersection above has the same cardinality for every B ∈ V . 1 Let A and V as in Lemma I.12 and consider B ∈ V . Set {t1,...,tk} ⊂ S such 1 0 that {B(t1),...,B(tk)} = B(S ) ∩V . By transversality we have that,
0 0 (P ◦ B) (ti) = ∇P(B(ti)) · B (ti) 6= 0,∀i = 1,...,k.
1 Set s1 = 2 min{|ti −tl|;i 6= l} > 0. By Taylor’s formula we have
0 P ◦ B(t) = (P ◦ B) (ti)(t −ti) + oi(|t −ti|).
Consider s2 > 0 such that if |t −ti| < s2 for some i = 1,...,k, then
0 oi(|t −ti|) 1 0 (P ◦ B) (ti) − ≥ (P ◦ B) (ti) 6= 0. |t −ti| 2
In particular, there exists C1 > 0 such that if |t −ti| < s2, for some i = 1,...,k, we have 0 oi(|t −ti|) |log (P ◦ B) (ti) − | ≤ C1. |t −ti| 1 1 Define s = 2 min{s1,s2} and consider the intervals Ii = (ti −s,ti +s) ⊂ S . So, Z Z Z |log|P ◦ B(t)||dt = |log|P ◦ B(t)||dt + |log|P ◦ B(t)||dt 1 1 k k S S \ ∪i=1 Ii ∪i=1Ii = (I) + (II).
1 k By compacity of S \∪i=1 Ii, and the fact that the function P◦B does not vanish in this set, there exists another constant C2 > 0 such that
|log|(P ◦ B(t))|| ≤ C2,
1 k for every t ∈ S \ ∪i=1 Ii. This gives Z 1 k (I) = |log|P ◦ B(t)||dt ≤ C2m( \ ∪i= Ik) 1 k S 1 S \ ∪i=1 Ii
29 Note that by, Taylor’s Formula in each Ii, we have
0 oi(|t −ti|) log|P ◦ B(t)| = log (P ◦ B) (ti) + + log|t −ti|. |t −ti| Hence, by the choice of s, and the fact that the logarithm funtion is integrable at the origin, we conclude that
Z k Z (II) = k |log|P ◦ B(t)||dt = ∑ |log|P ◦ B(t)||dt ∪i=1Ik i=1 Ii k Z k Z 0 oi(|t −ti|) ≤ ∑ |log (P ◦ B) (ti) + |dt + ∑ |log|t −ti||dt i=1 Ii |t −ti| i=1 Ii k ≤ ∑ m(Ii)C1 +C < ∞. i=1
Therefore, for each B ∈ V we have that log|P ◦ B| ∈ L1(S1;m). To conclude the proof of the main lemma what is left to show is that the fol- lowing set
r 1 1 1 A = {A ∈ C (S ;GLd(R); d(A(S ),V ) > 0}, (4.5)
r 1 is dense in C (S ;GLd(R)). This is a consequence, as we will see, of the fact that 1 the V is an algebraic variety of codimension greater or equal than one in GLd(R). Consider the following chain of algebraic sets:
m m−1 1 0 V V ... V V ,
where V i = Sing(V i−1). We know that this is a finite chain since V i is an algebraic variety of dimension strictly less than the dimension of V i−1. We can assume m that V is a regular submanifold of GLd(R) (the chain stops in this moment). Moreover, by definition, we have that each V i is a closed subset of V i−1 and i−1 i V \V is a regular submanifold of GLd(R) of codimension large than 1. m Using the remark 3 with M = GLd(R) and N = V which is a regular subman- ifold of GLd(R), we have that the set
r 1 1 m Am = {A ∈ C (S ;GLd(R)); A(S ) ∩V = ∅},
r 1 is dense in C (S ;GLd(R)). More generally, using the same argument, now with i i+1 M = GLd(R) and N = V \V , we will obtain that the set
r 1 1 i i+1 Ai = {A ∈ C (S ;GLd(R)); A(S ) ∩V \V = ∅},
30 r 1 is dense in C (S ;GLd(R)), for every i = 1,...,m − 1 (note that we are using strongly that the codimension is large enough). Hence, the set
m \ A = Ai, i=1
r 1 is (open) and dense in C (S ;GLd(R)). Observing that this intersection is the set A defined above in 4.5, we conclude the proof of the main lemma in the case r ∈ [1,∞]. Assume that r = ω. In this case we have that P ◦ A : S1 → R (P is a poly- nomial and in particular a analytic function) is an analytic function for any A ∈ ω 1 C (S ,GLd(R)). In this case, either P ◦ A is constant equal to zero or, has only finitely many zeros and those zeros have finite order in the sense that if we con- 1 sider a neighborhood of a zero ti ∈ S and write
mi P ◦ A(t) = (t −ti) gi(t),
where mi ∈ N, gi 6= 0 on this neighborhood. Since the function logt is m-integrable on [0,1], we conclude that
1 1 log|P ◦ A| ∈ L (S ,m),
ω 1 for all A ∈ C (S ,GLd(R)) such that P◦A is not zero. Assume then that P◦A ≡ 0. Since, the polynomial P is non constant, the interior of [P = 0] is empty. So, for any A ∈ [P = 0] and for any ε > 0, there exist B ∈ GLd(R), ||B|| = 1 and a positive 1 δ < ε such that A+δB ∈/ [P = 0]. In particular, taking A = A(t0), for some t0 ∈ S ω 1 and defining A˜ ∈ C (S ,GLd(R)) given by
A˜(t) = A(t) + δB,
we have that A˜ is δ-Cω close to A and P ◦ A˜ does not vanish identically. Hence, we fall in the previous case and
1 1 log P ◦ A˜ ∈ L (S ,m).
This concludes the proof of the Proposition I.10.
31 Part II
Embedding small horseshoes in lower dimensions
32 5 INTRODUCTION
5.1 History and motivation
In his celebrated work on celestial mechanics, [41], Poincare,´ already mentioned that an indication of the complexity of the three bodies problem is given by the understanding of the dynamics of, what he called, bi-asymptotic solutions, which is, nowadays, known as transversal homoclinic intersections associated to a saddle type periodic point. By analyzing carefully this geometric phenomena, S. Smale, in the sixties, introduce the concept of horseshoe: a model of non trivial (not given by a finite amount of periodic orbits), chaotic, persistent, higher dimensional dynamics, which has a continuous pair of laminations globally defined and it is conjugated to a symbolic model. The conjugation with the symbolic model allow us to fully understand the dy- namical behavior inside of the horseshoe (entropy, growth rate of periodic orbits, transitive orbits, among others). One of the simplest ways to add an extra difficulty is assuming that for some point of the horseshoe, for instance periodic, the stable and unstable manifolds in- tersect each other, not transversely (which would imply that this point were inside a new horseshoe) but tangentially (lack of hyperbolicity). That is the scope of the work of Newhouse in [34]. He proved that close to any given surface horseshoe, Λ, exhibiting a quadratic homoclinic tangency, associated with a periodic point p, and thickness larger than one there exists an open set U with persistence of tangencies. For a precise statement and the definition of thickness see [39]. In the case of surface horseshoes the stable and unstable laminations can be extended to regular (C1+) foliations. Using this regularity Newhouse was able s s to move the regular Cantor sets K (p) = Wloc(p) ∩ Λ (stable Cantor at p) and u u K (p) = Wloc(p) ∩ Λ (unstable Cantor at p) along these foliations to get regular Cantor sets on the same line (line of tangencies). When we unfold this tangency these new Cantor sets move through the line and its intersection are equivalent to new tangencies. This observation was crucial to reduce the problem to the study
33 of intersections of Cantor sets on the real line which he handled using what is currently called Newhouse Gap lemma (the thickness bound is used here). Later, in [33], Newhouse proved that if the initial tangency is associated with a dissipative periodic point, then there exists a residual set of diffeomorphisms
R∞ ⊂ U with infinitely many sinks. The existence of these sinks generically is now called Newhouse phenomena. This gives an indication of how complicated the dynamics can be if we leave the region that defines the horseshoe. Newhouse, Palis and Takens, in [35], initiate an attempt to fully understand the typical behavior of the dynamics f in a region U that contains the horseshoe Λ and a region V that contains the quadratic homoclinic tangency associated with a
periodic point p ∈ Λ. They proved that if dimH(Λ) < 1, then for any smooth one
parameter family (gt)|t|≤t0 unfolding a homoclinic tangency, that is
• g0 contains a tangency in V associated with pg0 (the hyperbolic continuation of p);
• (gt)−t0 we have that the set of parameters t such that Λ˜ t, the maximal invariant in U ∪ V, is hyperbolic has total density at 0. In other words, for any smooth small family unfolding a homoclinic tangency at V, (gt), we have that Λ˜ g is hyperbolic with total density at the first parameter. In order to prove this result Newhouse, Palis and Takens made use again of regularity of the invariant foliations to reduce the problem to study translations (arithmetic difference) of regular Cantor sets in the tangent line and whose the sum of its Hausdorff dimensions are smaller than one. In particular, the arithmetic difference between these Cantor sets has zero Lebesgue measure which indicates the density property attained by this result. The Newhouse, Palis and Takens result, above described, led the theory to a natural question: what happens if dimH(Λ) > 1. Aiming to answer this question Palis conjectured, in [38], that for typical pairs of Cantor sets K1 and K2 with dimH(K1)+dimH(K2) > 1 the arithmetic difference K1 − K2 contains an interval. Moreira introduced, in [31], the notion of stable intersection between the pairs k of C regular Cantor sets (K1,K2) which means that there exists a neighborhood of the pair (K1,K2) such that the intersection happens for any pair of regular Cantor sets in this neighborhood. Moreira and Yoccoz, in [32], proved that there exists an k open and dense set U of pairs of C regular Cantor sets such that if (K1,K2) ∈ U , then there exists t ∈ R such that (K1,K2 +t) has stable intersection. In their work, Moreira and Yoccoz, introduced new perturbations techniques and used tools from 34 the probability theory to address the problem. These ideas and techniques was the fuel necessary to understand the typical dynamical behavior in the region U ∪V, described above, when we unfold a quadratic tangency associated with a periodic point in a fat surface horseshoe Λ. More precisely, in [30], Moreira and Yoccoz proved that given a surface horseshoe Λ for a diffeomorphism f with dimH(Λ) > 1 and a quadratic homoclinic tangency associated with a periodic point p ∈ Λ, then for an open and dense set U0, g0 ∈ U0, in a neighborhood of f and for all smooth one parameter family passing through g0 we have that the set of t such that gt has tangency robustly has positive inferior density at 0. Among others, the lack of regularity of the invariant foliations is one of the most fundamental difficulties when we try to push the above results from surfaces to higher dimensional horseshoes. It is not hard to build 3 dimensional horse- shoes given by a transversal homoclinic intersection associated to a periodic point in which the unstable lamination is not even Lipschitz. It is enough, for example, to choose the eigenvalues of the periodic point suitably and assume that the ho- moclinic intersection occurs exactly over the strong stable manifold (not a generic condition). See [40]. Focusing on generalizing Newhouse results, Palis and Viana, in [40], dealt with these difficulties proving that typical transversal homoclinic intersections are inside of horseshoes with intrinsically C1+ unstable (or stable) laminations 1 which is enough to recover many of the geometric properties of the horseshoe (for example the thickness can be defined and we can preserve regularity of regular Cantor sets by holonomies). Using this weak notion of regularity, Palis and Viana, were able to prove that near any smooth diffeomorphism exhibiting a tangency associated with a sectional dissipative saddle, there exists a residual subset of an open set such that each of its elements displays infinitely many sinks. Another problem in the study of higher dimension horseshoes is the bad be- havior of the Hausdorff dimension of the horseshoe with respect to the diffeomor- phism. In [13], Bonatti, Diaz and Viana build an example of a horseshoe Λ, for a diffeomorphism f , in dimension 3, with dimH(Λ) < 1 which can be approxi- mated by diffeomorphisms g such that dimH(Λg) > 1 (Λg is the hyperbolic con- tinuation of Λ, see Section 6.2). The idea is again to consider a horseshoe Λ, with 1Intrinsic derivatives (also known as Whitney’s derivatives) is a notion that can be defined, differently of the traditional derivatives, on compact sets. Whitney proved that when a function is C1 intrinsic, there exists an open set such that the function can be extended to this open set, but in general this extension does not give any information for the point of view of the original dynamics. 35 dimH(Λ) < 1, given by a transversal homoclinic intersection happening exactly, again, over the strong stable leaf of the periodic point (not generic). When this intersection is unfolded it emerges horseshoes with Hausdorff dimension larger than one. In [29], Moreira, Palis and Viana, propose a script to generalize the New- house, Palis and Takens result and Moreira and Yoccoz result to higher dimen- sional horseshoes. They suggest to use a different notion of dimension, the upper stable dimension, inspired by the concept of affine dimension introduced by Fal- coner in [20], which seems to suit better from a dynamics point of view. The upper stable dimension, d¯s(Λ), which measures the size of the hyperbolic set inside the stable leaves, is always upper semi continuous with respect to the diffeomorphism (under extra hypothesis is continuous, see Chapter 8) and is an s ¯ upper bound for stable dimension at any x ∈ Λ, that is, dimH(Wloc(x)∩Λ) ≤ ds(Λ) which motivates the upper in the name of d¯s(Λ). See Section 7.2 to a precise definition of the upper stable dimension. In [28], Matheus, Moreira and Silva used the upper stable dimension to study horseshoes with sharp s-splitting (see Section 6.1). They proved that typical C∞ large (stable dimension larger than one for any point in the horseshoe) horseshoes contains blenders of codimension one. For more details about blender see [12]. Among other things, the present work has the intention to continue the formal- ization of the ideas described in the script proposed by Moreira, Palis and Viana. In the next section, we describe our results. 5.2 Results The examples mentioned in the previous section gives an indication that strong stable leaves passing through many points of a given horseshoe is an obstruction to good geometric properties of the horseshoe. The main result of this thesis is a first attempt to deal with this kind of issue. Theorem A (Main Theorem). Let M be a compact manifold, dim(M) ≥ 2, and r r > 1. Consider Λ ⊂ M be a horseshoe for f ∈ Diff (M) with a s-splitting, TΛM = ss w u w E ⊕ E ⊕ E . Assume that dim(E ) = 1 and d¯s(Λ) < 1. Then, given ε > 0 r r 0 there exists g ∈ Diff (M),C close to f , and a subhorseshoe Λg of the hyperbolic continuation of Λ for g, Λg, such that ¯ 0 ¯ 1. ds(Λg) ≥ ds(Λg) − ε; 36 0 2. For every x ∈ Λg we have ss 0 W (x) ∩ Λg = {x}. Theorem A has the following geometric property which is a direct conse- quence of the characterization obtained by Bonatti and Crovisier in [11]. Theorem B (Dimension reduction). There exists a locally g-invariant, C1+, sub- 0 0 w manifold S of M which contains Λg and such that for every x ∈ Λg,TxS = E (x)⊕ Eu(x). 0 Figure 5.1: Λg embedded in S 5.3 Work in progress We want to point out that, in typical terms, the hypothesis of the Theorem A r are not particular in the sense that for a typical C horseshoe Λ with d¯s(Λ) < 1 there exists a subhorseshoe with almost the same upper stable dimension which has s-splitting. This is attained first guarateeing the existence of a periodic point 37 with simple spectrum (typical among the Cr horseshoe) and later building random perturbations (see Section 9.1) to suitably find cone fields which will led us to a dominated splitting inside the stable bundle. An interesting consequence of the Theorem A is the regularity (C1+) of the 0 unstable laminations of the subhorseshoe Λg meaning that this lamination can be extended to a C1+ foliations (similarly to what happens in the case of surface 0 horseshoes). The reason is that since the strong stable leaves only intersect Λg in a single point (Ess is a normal direction), we can ignore the strong stable direction 0 and so guarantee that bunching conditions are satisfied for Λg. Using, therefore, the theory developed by Hirsch, Pugh and Shub [25] we will be able to obtain the aimed regularity. Adding the information provided above we will achieve the following result: Proposal 1 (In progress). For typical Cr diffeomorphism f with horseshoe Λ with dimH(Λ) > 1 we can build subhorseshoes Λ1 and Λ2 which are invariant by some iterate f k satisfying that: 1. Λ1 has a globally defined strong stable lamination and regular unstable holonomies; 2. Λ2 has a globally defined strong unstable lamination and regular stable holonomies; 3. It holds max < 1 and d¯s(Λ1) + d¯u(Λ2) > 1. d¯s(Λ1), d¯u(Λ2) 4. For each p ∈ Λ and p ∈ Λ the projections πss| s , along the 1 1 2 2 Λ1∩Wloc(p1) strong stable leaves, and πuu| u , along the strong unstable leaves, Λ2∩Wloc(p2) are injective and its inverse has a C1+ extension. 38 6 HORSESHOES 6.1 Dominated splitting and Partially hyperbolic sets Let Λ ⊆ M be a compact manifold. A splitting of TΛM is a decomposition TΛM = E1 ⊕ ··· ⊕ Ek, in continuous linear fiber subbundles Ei ⊆ TM over Λ, where the dimension of Ei(x) is independent of x ∈ Λ, for all 1 ≤ i ≤ k. Let f ∈ Diffr(M), r ≥ 1, and assume that Λ ⊆ M is a compact f -invariant set. We say that a splitting TΛM = E1 ⊕ ··· ⊕ Ek is a dominated splitting for Λ if it satisfies • Invariance: d f (x)Ei(x) = Ei( f (x)), for every x ∈ Λ and for 1 ≤ i ≤ k; • Domination: There exist C > 0 and σ < 1 such that n n n d f (x)|Ei(x) ≤ Cσ m(d f (x)|Ei+1(x)) for all n ≥ 1 and all 1 ≤ i ≤ k − 1. Λ is said to be a partially hyperbolic set for f if admits a dominated splitting TΛM = E1 ⊕···⊕Ek (partially hyperbolic splitting) such that E1 is uniformly con- tracted and Ek is uniformly expanded, that is, there exist C > 0 and λ ∈ (0,1) such that n n −n n d f (x)|E1(x) ≤ Cλ d f (x)|Ek(x) ≤ Cλ . c s u In this case we call the bundles E2 ⊕ ··· ⊕ Ek−1 = E , E1 = E and Ek = E , re- spectively, the center, stable and unstable bundles for f over Λ. When the partially hyperbolic set Λ has trivial center bundle, Ec = {0} we say that Λ is a (uniformly) hyperbolic set for f ∈ Diffr(M). ss w u When Λ is hyperbolic with a (nontrivial) splitting TΛM = E ⊕E ⊕E , with Es = Ess ⊕ Ew uniformly contracted and Eu uniformly expanded, the bundles Ess 39 and Ew are called respectively strong stable bundle and weak stable bundle and ss w u we say that Λ has a s-splitting TΛM = E ⊕E ⊕E . Analogously, we can define a u-splitting for Λ. Observe that this notion requires that the dimension of the manifold M must to be at least 3. A riemannian metric on M is said to be adapted to a partially hyperbolic set s c u Λ, for f , if the partially hyperbolic splitting TΛM = E ⊕ E ⊕ E satisfies that: • d f (x)|Es(x) < 1, −1 • d f (x)|Eu(x) < 1 and • d f (x)|Es(x) ≤ m(d f (x)|Ec(x)) ≤ d f (x)|Ec(x) ≤ m(d f (x)|Eu(x)). Basically, the existence of such a metric on the manifold M indicates that the contraction, expansion and the domination in the partially hyperbolic splitting of TΛM can be see in the first iteration of the dynamics f . It is known that given a partially hyperbolic set for f ∈ Diffr(M) in a com- pact manifold M with riemannian metric g, there exists a riemannian metricg ˜, equivalent to g, which is adapted to Λ. For more details see ([16], page 17). s c u We say that a splitting, TΛM = E ⊕ E ⊕ E , of a partially hyperbolic set Λ (with an adapted metric) is a sharp splitting if there exist positive real numbers λ sc and λ cu such that sc sup d f (x)|Es(x) < λ < inf m(d f (x)|Ec(x)) x∈Λ x∈Λ cu ≤ sup d f (x)|Ec(x) < λ < m(d f (x)|Eu(x)). x∈Λ Similarly we can define sharp dominated splitting when we have a dominated splitting instead of a partially hyperbolic splitting of the set Λ. ss w u In the case that Λ is hyperbolic with a s-splitting, TΛM = E ⊕ E ⊕ E , we can interpret Λ as a partially hyperbolic set with nontrivial center bundle Ec = Ew. We say that Λ has a sharp s-splitting if the (partially hyperbolic) splitting TΛM = Ess ⊕ Ew ⊕ Eu is a sharp splitting. In a similar way we define a sharp u-splitting for Λ. All the concepts introduced in this section are robust which means that if, Λ ⊆ M is a compact set satisfying the P property, where P stands to either 1. Λ has a dominated splitting or 2. Λ has a partially hyperbolic splitting or 40 3. Λ has a hyperbolic splitting, then there exist an open neighborhood U ( f ) ⊂ Diff1(M) of f ∈ Diffr(M) and an open neighborhood U of Λ such that if g ∈ U ( f ) and K ⊂ U is a g-invariant set, then K satisfies the property P (see [16], page 16). 6.2 Horseshoes Let Λ ⊆ M be a (uniformly) hyperbolic set for f ∈ Diffr(M), r ≥ 1, with an adapted metric. We will assume throughout this entire work that our metric is adapted, unless we make explicit the contrary. Λ is said to be isolated if there exists open neighborhood, U ⊆ M, of Λ such that \ Λ = f j(U). j∈Z Definition II.1. A compact set Λ ⊆ M is called a horseshoe for a Cr diffeomor- phism, f : M → M, if it is an infinite, isolated, topologically mixing and totally disconnected hyperbolic set. We say that Λ is a Cr horseshoe if Λ is a horseshoe for some f ∈ Diffr(M). Example II.2. Let f : R2 → R2 which maps the unit square Q ⊂ R2 in the ”horse- shoe shape” as described in the figure 6.1. The maximal invariant inside the square Q is called Smale horseshoe. Figure 6.1: Smale horseshoe. Example II.3. In the same spirit of the dynamics of the Smale horseshoe we can define horseshoes by homoclinic intersections as described in the figure 6.2. 41 Figure 6.2: Homoclinic intersection in dimension 2. Example II.4. In the case that the periodic point, with transversal homoclinic intersection, has simple spectrum we can build a horseshoe with s-splitting. Figure 6.3: s-splitting. Example II.5. The case of complex eigenvalues led to the a horseshoe without s-splitting. Remark 4. If Λ is a horseshoe, then Per( f |Λ) = Λ. That can be proved using the shadowing property of hyperbolic sets (or the conjugation with the shift map (see Section 6.5). Remark 5. The isolated condition for a hyperbolic set Λ is equivalent to the Local product structure which means that the intersections of stable and unstable manifolds of points of Λ lies in Λ (see Section 6.3). 42 Figure 6.4: Non s-splitting. Let Λ be a horseshoe for a f ∈ Diffr(M) and consider U ⊆ M be a small open neighborhood of Λ. As mentioned in the end of the Section 6.1, choosing U small enough, we have that for any g ∈ Diff1(M) close to f in the C1 topology, if K ⊆ U is a g-invariant set, then K is hyperbolic. Since Λ is a horseshoe, a stronger conclusion holds. There exists a neighborhood U ( f ) in Diff1(M) and a 0 continuous map φ : g ∈ U ( f ) 7−→ φg ∈ C (Λ;M) satisfying that 1. There exists K > 0 such that dC0 (incΛ,φg) ≤ KdC0 ( f ,g) (6.1) 0 where incΛ ∈ C (Λ;M) is the inclusion map; 2. The following diagram f Λ / Λ φg φg Λg g / Λg commutes. See ([46], page 101). The sets Λg = ϕg(Λ) are called hyperbolic (or analytical) continuations of Λ by g. They satisfies that \ n Λg = g (U), n∈Z r r for any g ∈ U ( f ). In particular, Λg are C horseshoes for all g ∈ U ( f )∩Diff (M). 43 6.3 Invariant manifolds In what concerns the study of regularity of invariant foliations by diffeomorphisms there are three different (although intrinsically related) classes of problems. 1. Existence, uniqueness and regularity of the leaves; 2. Regularity of the foliations/lamination (transversal regularity or regularity by holonomies); 3. Regularity of the leaves with respect to the hyperbolic continuation of the leaf (regularity with respect to the diffeomorphism). In this section we will discuss and try to bring references about each of these problems. Consider Λ ⊂ M a horseshoe for a diffeomorphism f ∈ Diffr(M) with s- ss w u splitting, TΛM = E ⊕ E ⊕ E . s The stable manifold theorem states that there exists a continuous map Φ f : r Es Λ → Emb (Ddim ;M) such that s s Φ f (x) = Wf ,ε (x) for some ε > 0. This implies that there exists a unique stable lamination s s F f = {Wf (x);x ∈ Λ}, s of Λ by stable manifolds such that each leave Wf (x) is an immersed submanifold of M of class Cr. There also exists a continuous map 1 s r dimEs g ∈ U( f ) ⊂ Diff (M) 7−→ Φg ◦ φg : Λ → Emb (D ;M) The same can be said about the unstable lamination. So, s u r (i) Wf ,Wf always exist, are unique and have C leaves; s (ii) The maps g ∈ U( f ) 7−→ Wg,ε (φg(x)) are continuous for all x ∈ Λ. See for example [39]. We have similar results for the strong stable foliation/lamination F ss (the one whose leaves are tangent to the strong stable space Ess). Define ss n n −εn n W (x) = y ∈ M; d( f (x), f (y)) ≤ Ce m(d f (x)|Ew(x)) . 44 This definition does not depends on the choice of (small) ε > 0. This lamination for x ∈ Λ is also unique (if the splitting of TΛM is fixed which our case is). (see [16]). In [25] is proved that for every x ∈ Λ, W ss(x) is uniquely defined, invariant by r ss f and is a C -immersed submanifold of M tangent to E . For η > 0, if Dη (x) ⊂ W ss(x) is a disc of length η in W ss(x) it depends continuously on f in the Cr- topology and x ∈ Λ. So, (1) W ss always exists, is unique and have Cr leaves. ss (2) The map g 7−→ Wg,η vary continuously. At this moment we wonder if there exists a weak stable lamination tangent to Ew. If we assume that Λ has a sharp s-splitting we can use the theorem in ([39], page 158) to prove that we have a weak unstable manifold tangent to Ew ⊕ Es defined for any x ∈ Λ and intersect this manifold with W s to guarantee the existence of W w. But, that was only possible since we have a sharp separation between the bundles Ess and Ew ⊕ Eu. In general, the existence of the weak stable manifold can fail and even when exists the unicity can be not assured. See [39]. Until now the discussion is pretty much satisfactory to address the classes of problems mentioned in the beginning of section in a continuous fashion. What can we say about higher regularities? If r > 1, then the bundles Es,Eu,Ess and Ew are always Holder¨ continuous ([16], page 44). Remark 6. If f ∈ Diff1(M), then in general the bundles Es and Eu are only con- tinuous. It is possible to build examples of 2-dimensional Anosov diffeomorphisms which are C1 but the bundle are only continuous using Franks lemma [21]. In [44] they proved that W s,W ss and W u are also Holder¨ continuous. In [45] they give a better explanation of the fact that inside of a leaf of the stable lamina- tion, W s(p), the lamination W ss can be extended to a regular C1+ foliation. This last fact will repeatedly be used throughout this work. r Let { fµ }µ∈X be a C family of diffeomorphisms. For µ0 ∈ X, let Λµ0 be a horseshoe for fµ0 and denote by Λµ the hyperbolic continuation of Λµ0 for any µ ∈ X (the family { fµ } is taken in such a way that the hyperbolic continuation is well defined). Take x0 ∈ Λµ0 and let xµ = φ fµ (x0) ∈ Λµ be the hyperbolic 45 continuation of x0. The map µ 7−→ xµ , is Cr (See Appendix 1 in [39]). This fact will be useful in the subsequent sections. Summarizing the discussion so far: s u ss 1. There are uniquely defined laminations (continuous), W f , W f and W f such that the leaves are of class Cr which is the regularity of f ; 2. These laminations are Holder¨ continuous at Λ, W ss vary in a C1+ fashion inside of any leaf of the stable lamination W s. 3. The existence of a weak stable lamination, W w, tangent to Ew can fail and even in the cases in which such a lamination exists there is no reason to be unique (in general it is not); ∗ 4. The maps g ∈ U ( f ) 7−→ Wg,ε (φg(x)) are only continuous (actually, we have r dimE∗ the continuity of the map g ∈ U ( f ) 7−→ Φg ◦ φg : Λ → Emb (D ;M) ∗ dimE∗ where Wg,ε (φg(x)) = Φg(φg(x))(D ). r In any dimension, we have that the map g ∈ U ( f ) ⊂ Diff (M) 7−→ φg(x) is of class Cr for any x ∈ Λ. Here U ( f ) is endowed with the Cr topology. Remark 7. More can be said in the two dimensional case where we have a better understanding of the regularity of these laminations (which can be extended to C1+ foliations in a neighborhood of Λ) and hence of the geometric properties of surface horseshoes. Remark 8. Another possible (future) discussion is regarding the regularity of the holonomies. The Holder¨ regularity (transversal regularity) of the foliations described in this section stands for the Holder¨ regularity of the holonomies as well. For a more general discussion about the different types of regularity of general foliations see [44]. 6.4 Markov partitions Let Λ be a horseshoe for f ∈ Diffr(M). A subset P of Λ is said to be a rectangle if for any x,y ∈ P, [x,y]ε0 ∈ P, where {[x,y] } = W s (x) ∩W u (y). ε0 ε0 ε0 We say that a rectangle P is proper if P = IntΛ P. 46 s Let P be a rectangle with diam(P) ≤ ε0. For any x ∈ P, denote by W (x;P) the W s (x) ∩ P set ε0 . A Markov partition for a horseshoe Λ, for f ∈ Diffr(M), is a finite partition P = {P(a);a ∈ A} formed by proper rectangles of Λ satisfying: (i) IntΛ P(a) ∩ IntΛ P(b) = ∅ for all a,b ∈ A; −1 (ii) For any a,b ∈ A and x ∈ Λ such that x ∈ IntΛ P(a)∩ f (IntΛ P(b)) we have f (W s(x;P(a))) ⊂ W s( f (x);P(b)) f (W u(x;P(a))) ⊃ W u( f (x);P(b)). The set A is called the alphabet associated to the Markov partition P. The el- ements of P will frequently be called by either boxes or pieces of the Markov partition. It was proved by Bowen, in [15], that given a horseshoe Λ we can find Markov partitions with arbitrarily small diameters. We will use these information in the subsequent sections without make any mention. Figure 6.5: Markov Partition with small diameter. 6.5 Symbolic dynamics Given a horseshoe Λ for f ∈ Diffr(M) with Markov partition P = {P(a);a ∈ A}, we say that a pair (a,b) ∈ A2 is admissible if f (P(a)) ∩ P(b) 6= ∅. Define n o Z Σ = θ ∈ A ; (θi,θi+1) is admissible for everyi ∈ Z . Σ is a compact subset of AZ with the product topology and σ : Σ → Σ (the shift map) is a homeomorphism. 47 In [15], it was also prove that there exists a homeomorphism i f : Σ → Λ which make the following diagram σ Σ / Σ i f i f Λ / Λ f commutes. The pair (σ,Σ) is called symbolic representation of the horseshoe dynamics ( f |Λ,Λ). For each a ∈ A consider an small open neighborhood U(a) of P(a) such that U(a) ∩U(b) = ∅, for all a,b ∈ A (this is possible because |A| < ∞ and the piece P(a) are two by two disjoints and so have positive distance). Set U = ∪a∈AU(a) and let U ( f ) ⊂ Diffr(M) be an open neighborhood of f such that for any g ∈ U ( f ) the analytical continuation of Λ by g, Λg, is contained in U. The set Pg = Pg(a) := U(a) ∩ Λg;a ∈ A is a Markov partition for Λg and there exists ig : Σ → Λg homeomorphism such that the following diagram σ Σ / Σ ig ig Λg g / Λg commutes for all g ∈ U ( f ). For each element x ∈ Λg, there exist θ = (θi)i∈Z ∈ Σ such that \ − j {x} = g (U(θi)). j∈Z 6.6 Symbolic dynamics notation Now we introduce a notation which will be used throughout this work. • Positive infinite sequences or (local) stable manifolds: n o + + + N0 + + Σ = θ = (θn )n≥0 ∈ A ; θi θi+1 is admissible ; • Negative infinite sequences or (local) unstable manifolds: n o − − − N0 − − Σ = θ = (θ−n)n≥0 ∈ A ; θ−iθ−i+1 is admissible ; 48 • Positive finite words or stable cylinders: + W = {a = a0a1 ...an; n ∈ N0, ai ∈ A and aiai+1 is admissible}; • Negative finite words or unstable cylinders: − W = {a = a−n ...a−1a0; n ∈ N0, a−i ∈ A and a−ia−i+1 is admissible}; • (Bi) finite words or boxes: W = {a = a−n ...a0 ...am; n,m ∈ N0, ai ∈ A and aiai+1 is admissible}; We see all the above introduced sets Σ+,Σ−,W+,W−,W as subsets of 2Σ, meaning that any element θ (or a) in this sets is a subset of Σ. For example an element a = a−n ...a0 ...am ∈ W is the subset of elements θ ∈ Σ such that θi = ai for every −n ≤ i ≤ m. We also note that W+ ⊂ W and W− ⊂ W, once that if + − a = a−n ...a0 ...am ∈ W with n = 0 we have a ∈ W and if m = 0, a ∈ W . + − If a = a0 ...am ∈ W and b = b−k ...b0 ∈ W with a0 = b0, we use the nota- tion b ∩ a = b−k ...b0a1 ...am ∈ W, to the actual intersection between the subsets a and b of Σ. Using this notation we can always write any element a ∈ W as a = a− ∩a+, with a− ∈ W− and a+ ∈ W+. We set + : Σ → Σ+ and − : Σ → Σ− the natural projections +(θ) = θ + − (θ) = θ −. So, giving any θ ∈ Σ, there exist θ + ∈ Σ+ and θ − ∈ Σ− such that θ = θ − ∩ θ + (this is an abuse of the language since, θ is an element of Σ and θ − ∩ θ + is a subset of Σ which is equal to {θ}). Observe that the shift map naturally acts in these sets: If a ∈ W, for any j ∈ Z, σ j(a) = θ ∈ Σ;σ − j(θ) ∈ a and if θ ± ∈ Σ±, then σ j(θ ±) = θ ∈ Σ;σ − j(θ) ∈ θ ± . It is worthwhile to mention that when a = a−n ...a0 ...am ∈ W and −n ≤ j ≤ m, we still see the set σ j(a) as an element of W that, although has the same symbols as a, is a different element in W (the position of the 0-th coordinated 49 changes). In other words, if, without loss of generality, j ≥ 0, σ j(a) = b = b−n− j ...b0 ...bm− j ∈ W, where bi− j = ai for −n ≤ i ≤ m. For any a = a−n ...a0 ...am ∈ W, |a| = m + n + 1 denote the size of the word a. Consider a = a−n ...a0 ...am ∈ W and b = b−k ...b0 ...bl ∈ W finite words 2 such that the pair (am,b−k) ∈ A is admissible. Define the right concatenation of the word a by b as the word a · b = c = c−n ...c0 ...cm+|b|W R with ( ai if − n ≤ i ≤ m ci = bi−m−1−k if m + 1 ≤ i ≤ m + |b| We define the left concatenation of the word b by the word a as the word a · b = d = d−k−|a| ...d0 ...dl ∈ W, L with ( bi if − k ≤ i ≤ l di = ai+k+1+m if − k − |a| ≤ i ≤ −k − 1 Observe that these operations are not commutative, i.e., in general, we do not have a · b 6= b · a and a · b 6= b · a. R R L L We denote by a·n the right concatenation of the word a ∈ W by a, n-times. R Analogously we define ·nb the left concatenation of the word bW by b n-times. L Since W+ ⊂ W the definition of right concatenation applies to this finite words but in this case is more natural to denote a · b only by ab ∈ W+. The same R holds to the left concatenation between negative finite words, c,d ∈ W−, which will be denoted by cd = c · d ∈ W−. In this case for a ∈ W+,b ∈ W− we write L a·n = an and ·nb = bn. Remember that this concatenations are only possible if the R L admissible properties are satisfied by the words a and b. + Given a = a0 ...am,b = b0 ...bl ∈ W with a0 = b0 we define the maximum common word in a and b as the word + a ∨ b = a0 ...ak ∈ W , where a j = b j for all 0 ≤ j ≤ k and ak+1 6= bk+1. Similarly, given c = c−n ...c0,d − = d−k ...d0 ∈ W , with d0 = c0 we define the minimum common words in c and d as the word − c ∧ d = d−m ...d0 ∈ W , 50 where ci = di for all −m ≤ i ≤ 0 and c−m−1 6= d−m−1. Note these define commu- tative operations: a ∨ b = b ∨ a and c ∧ d = d ∧ c. This above introduced symbolic notations and operations turns out to be really useful when we are dealing with horseshoes. A flavor of such applications can be seen when we consider the following set of notations. Let f ∈ Diffr(M) with a horseshoes Λ which is conjugated to Σ. As described in the end of the Section 6.2 there exists a neighborhood U ( f ) ⊆ Diff1(M) such that the hyperbolic continuation of Λ for any g ∈ U ( f ), Λg is also conjugated to Σ with continuous conjugation ig : Σ → Λg. For a = a−n ...a0 ...am ∈ W and g ∈ U ( f ), write n \ − j Pg(a) = g (Pg(a j)) ⊂ Λg. j=−m This set is called a piece (or box) associated to a in the horseshoes Λ. When + a ∈ W the set Pg(a) is denoted by Hg(a) and is called stable cylinder and when − a ∈ W , the set Pg(b) is denoted by Vg(a) and is called unstable cylinder. Set, for θ + ∈ Σ+ the local stable manifold associated with θ + as s + \ − j + Wg,loc(θ ) = g (U(θ j )), j∈N and for θ − ∈ Σ−, the local unstable manifold associated with θ − as u − \ j − Wg,loc(θ ) = g (U(θ j )). j∈N These are open subsets of some leaf in the stable and unstable lamination of Λ respectively. Note that, by the conjugation property, if a ∈ W−, we have −|a| −|a| g (Vg(a)) = Hg(σ (a)). 51 Figure 6.6: From stable cylinder to unstable cylinder. Another important observation is that the restriction of ig to a, ig|a : a → Pg(a) ± s/u ± is also a homeomorphism (the same holds for θ and Wg,loc(θ ) instead of a and Pg(a)). + − The homeomorphisms, i|∗, ∗ ∈ {a,θ ,θ }, allow us to identify boxes, unsta- ble/stable cylinders and local stable/unstable manifolds with its symbolic repre- sentations. 52 7 METRIC INTERACTIONS BETWEEN HORSESHOES AND ITS SYMBOLIC MODELS We start this section describing the metrics defined on Σ, Σ− and Σ+ which will be used in all the metric consideration that we will make in this work (remember the notation introduced at Section 6.6). Define, for θ +,θ˜ + ∈ Σ+ + + −|θ +∨θ˜ +| d+(θ ,θ˜ ) = 2 , and for θ −,θ˜ − ∈ Σ− − − −|θ −∧θ˜ −| d−(θ ,θ˜ ) = 2 . Taking θ,θ˜ ∈ Σ and writing θ = θ − ∩ θ + and θ˜ = θ˜ − ∩ θ˜ + with θ −,θ˜ − ∈ Σ− and θ +,θ˜ + ∈ Σ+ set − − + + d(θ,θ) = max d−(θ ,θ˜ ), d+(θ ,θ˜ ) . Using the standard metric introduced above we can prove (see [26], page 598) that the continuous conjugation i f : Σ → Λ actually is a bi-Holder¨ map (but in general not Lipschitz) with uniformly Holder¨ exponent and Holder¨ constant in a neighborhood U ( f ) ⊂ Diffr(M), this will be useful to transcribe metric properties from symbolic dynamics to the horseshoes which can be exemplified by the next proposition whose the proof will be skipped (use the fact that the conjugation i f is Holder¨ continuous to pass the problem to symbolic dynamics). Proposition II.6. Let P be a Markov partition for Λ with alphabet A. For any g ∈ U ( f ), if diam(Pg(a)) ≥ ζ > 0 for some a ∈ W, then 1/δ dist(Λ\Pg(a);Pg(a)) & ζ 0 − −1 −1 where δ0 = [i f ][i f 1] and [i f ],[i f ] ∈ (0,1] are the Holder¨ exponents of i f and i f respectively. 53 Figure 7.1: Piece P(a) with diameter larger than ζ. 7.1 Stable diameters Let Λ be a horseshoe for f ∈ Diffr(M) and P be a Markov partition with alphabet A for Λ. For each b ∈ W−, define |b| Ds( f ;b) = sup d f (x)|Es(x) . −|b| x∈ f (Vf (b)) Ds is called the stable diameter along Vf (b). When the diffeomorphism f is fixed we will only write Ds(b) instead of Ds( f ;b). Set −1 −1 τ = sup dg (x)| s and τˆ = sup dg(x)| s . g Eg(x) g Eg(x) x∈Λ x∈Λ For any g ∈ U ( f ). Note that 0 < τg ≤ τˆg < 1 and that if U ( f ) is chose small enough, there exists τ,τˆ ∈ (0,1) such that 0 < τ ≤ τg ≤ τˆg ≤ τˆ < 1, (7.1) for every g ∈ U ( f ). Proposition II.7. 1. For any b ∈ W−, |b| |b| τ f ≤ Ds(b) ≤ τˆf ; 2. If b,c ∈ W− we have Ds(bc) ≤ Ds(b)Ds(c). 54 Proof. Note that for every x ∈ Λ and for any n ∈ N n n d f (x)|Es(x) ≤ τˆf and −n −n d f (x)|Es(x) ≤ τ f . So, n n n n τ f ≤ m(d f (x))|Es(x) ≤ d f (x)|Es(x) ≤ τˆf , for every x ∈ Λ and n ∈ N. Taking n = |b| we conclude the item 1. −|bc| −|b| |b| Observe that if x ∈ f (Vf (bc)), then x ∈ f (Vf (b)) and f (x) −|c| ∈ f (Vf (c)) and |bc| |c| |b| |b| d f (x)|Es(x) ≤ d f ( f (x))|Es( f |b|(x)) · d f (x)|Es(x) , which implies |bc| d f (x)|Es(x) ≤ Ds(b) · Ds(c), −|bc| for every x ∈ f (Vf (bc)). So, Ds(bc) ≤ Ds(b) · Ds(c). For b ∈ W− set Dˆ (b) = W s (y) ∩V (b) , s sup diam ε0 f y∈Λ∩Vf (b) where the diameter used is defined using the metric in the manifold M. Figure 7.2: Stable diameter 55 − −|b| s Consider b ∈ W and x1,x2 ∈ f (Vf (b)) such that x2 ∈ W (x1) and d( f |b|(x ),d f |b|(x )) = W s (y) ∩V (a), 1 2 diam ε0 f |b| where y ∈ f (x1). By mean value theorem,