Professional Documents
Culture Documents
com
Pedro Duarte
Silvius Klein
Lyapunov Exponents
of Linear Cocycles
Continuity via Large Deviations · Volume 3
www.Ebook777.com
Free ebooks ==> www.Ebook777.com
Volume 3
Series editors
Henk Broer, Groningen, The Netherlands
Boris Hasselblatt, Medford, USA
www.Ebook777.com
The “Atlantis Studies in Dynamical Systems” publishes monographs in the area of
dynamical systems, written by leading experts in the field and useful for both
students and researchers. Books with a theoretical nature will be published
alongside books emphasizing applications.
Lyapunov Exponents
of Linear Cocycles
Continuity via Large Deviations
Free ebooks ==> www.Ebook777.com
www.Ebook777.com
In memory of João Santos Guerreiro and
Ricardo Mañé, professors whose friendship
and intelligence I miss
Pedro Duarte
vii
viii Preface
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Prologue. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Main Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Continuity Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Large Deviations Type Estimates . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Summary of Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 Estimates on Grassmann Manifolds . . . . . . . . . . . . . . . . . . . . . . . 23
2.1 Grassmann Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Projective Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.2 Exterior Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.3 Grassmann Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.1.4 Flag Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Singular Value Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.1 Singular Value Decomposition . . . . . . . . . . . . . . . . . . . 32
2.2.2 Gaps and Most Expanding Directions . . . . . . . . . . . . . . 35
2.2.3 Angles and Expansion . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Lipschitz Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3.1 Projective Action. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.3.2 Operations on Flag manifolds . . . . . . . . . . . . . . . . . . . . 50
2.3.3 Dependence on the Linear Map. . . . . . . . . . . . . . . . . . . 57
2.4 Avalanche Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.4.1 Contractive Shadowing . . . . . . . . . . . . . . . . . . . . . . . . 64
2.4.2 Statement and Proof of the AP . . . . . . . . . . . . . . . . . . . 68
2.4.3 Consequences of the AP. . . . . . . . . . . . . . . . . . . . . . . . 73
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
ix
Free ebooks ==> www.Ebook777.com
x Contents
www.Ebook777.com
Contents xi
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Acronyms
xiii
Chapter 1
Introduction
1.1 Prologue
We are about to offer the reader a distilled version of the entire monograph. All the
main actors will be briefly introduced, along with their principal roles.
We begin with the avalanche principle, first formulated for SL(2, R) matrices in
[14]. The AP is a deterministic statement relating the norm growth of a long chain
of matrices to the corresponding product of matrix norms.
In order to state our general version of this principle, let us introduce some
notations. Given a matrix g ∈ Mat(m, R), we denote its singular values by
s1 (g) ≥ s2 (g) ≥ · · · ≥ sm (g) ≥ 0 . Moreover, let gr(g) = ss21 (g) (g)
≥ 1 represent
the gap ratio between its first two singular values. When gr(g) > 1, the first singu-
lar value s1 (g) is simple, so the corresponding most expanding (singular) direction
v(g) ∈ P(Rm ) is well defined.
Given a sequence of matrices g0 , g1 , . . . , gn−1 ∈ Mat(m, R), we define its expan-
sion rift as
gn−1 . . . g1 g0
ρ(g0 , g1 , . . . , gn−1 ) := .
gn−1 . . . g1 g0
The angle between two matrices g, g ∈ Mat(m, R), denoted by α(g, g ), is the
sine of the angle between the g-image of v(g) and the orthogonal complement of
v(g ). Hence α(g, g ) > ε means that the g-image of the most expanding direction
of g is reasonably well aligned with the most expanding direction of g . Moreover,
if gr(g) and gr(g ) are large enough, then α(g, g ) ≈ ρ(g, g ) = gg g
g
.
The avalanche principle can be formulated as follows. Let 0 < κ ε2 . Given a
sequence of matrices g0 , g1 , . . . , gn−1 ∈ Mat(m, R), if for all indices i
gi gi−1
(angles) ≥ε
gi gi−1
(gaps) gr(gi ) ≥ κ −1 ,
or equivalently
1
L k (A) = lim log sk (A(n) (x)).
n→∞ n
The maximal Lyapunov exponent L 1 (A) corresponds to the largest singular value
growth rate of the iterates of A.
Many of the arguments in this monograph involve an inductive procedure, where
we take a large block A(n) (x), corresponding to a time scale n, and divide it up into
smaller blocks of approximately the same lengths. We transfer measurements from
the smaller blocks to the larger ones and repeat this procedure (Fig. 1.1).
This inductive procedure is based on the AP, where the individual matrices gi
correspond to the smaller blocks and the product matrix gn−1 . . . g1 g0 corresponds
to the larger block.
We illustrate this approach with a sketch of a new proof of the Oseledets Multi-
plicative Ergodic Theorem.
To start off, we show that if L 1 (A) > L 2 (A), then the most expanding direction
v(A(n) (x)) of the iterates A(n) (x) of the cocycle is well defined and it converges for
almost every phase x. To do that, it is enough to prove that except for a set of phases
of arbitrarily small measure, and for n n 0 1, we have
1 A(n+m) (x)
lim log = L 1 (A) − L 1 (A) = 0,
n,m→+∞ n + m A(n) (x) A(m) (T n x)
1
lim log gr(A(n) (x)) = L 1 (A) − L 2 (A) > 0 .
n→+∞ n
This means that asymptotically, the gap dominates the angle almost surely. Hence
for any large enough finite time scale, this will happen outside a set B of phases with
arbitrarily small measure.
Using Birkhoff’s ergodic theorem, we have that for long enough orbits, the iterates
T m x of a base point x visit the set X \ B frequently, hence these visits can be chosen
to be well distributed throughout the orbit. We refer to such times m as avalanche
times of the base point x.
Therefore, if we consider a long block of length n 1, we may divide it up into
smaller blocks, bounded between consecutive avalanche times, with lengths of order
n 0 n μ(B).
Two consecutive blocks of this kind have the form A(m −m) (T m x) and A(m−m )
(T m x), where m , m, m are consecutive avalanche times (Fig. 1.2). By construction,
the gap and angle assumptions of the AP hold uniformly for all such consecutive
blocks. The AP is then applicable, and (1.2) implies
By taking exterior powers we can extend the notion of most expanding direction of
A and introduce the most expanding flag of A as the measurable filtration
V2 (x) ⊂ V3 (x) ⊂ Rm ,
where for i = 2, 3, the subspace Vi (x) has dimension i and it contains the i most
expanding directions of A. The orthogonal complements V3 (x)⊥ ⊂ V2 (x)⊥ ⊂ Rm
of these subspaces contain the least expanding directions of the cocycle A. They
determine another filtration which is (T, A)-invariant and which turns out to be
exactly the Oseledets filtration.
More could be said about both Lyapunov exponents and the Oseledets filtration
if instead of the almost sure convergence of n1 logA(n) (x) to L 1 (A) provided by
Kingman’s ergodic theorem, a stronger, quantitative type of convergence in measure
were available. Enter the large deviation type estimates, i.e. estimates of the form
1
μ {x ∈ X : logA(n) (x) − L 1 (A) > ε} < ι(n, ε) (1.4)
n
Mixing MPDS are always ergodic, but the converse is not true in general.
The d-torus Td = (R/Z)d with its normalized Haar measure μ on the σ -field F
of Borel sets determines a probability space (Td , F, μ). We mention a few classes
of MPDS on the torus.
Example 1.1 (toral translations) Given ω ∈ Rd , the translation map T : Td → Td ,
T x := x + ω (mod 1), preserves the Haar measure μ. This MPDS is ergodic iff
the components of ω are rationally independent. Toral translations are never mixing.
Example 1.2 (toral endomorphisms) Given a matrix M ∈ GL(d, Z), the endomor-
phism T : Td → Td , T x := M x (mod 1), preserves the Haar measure μ. The
endomorphism T is ergodic iff the spectrum of M does not contain any root of
unity. Ergodic toral automorphism are always mixing.
The compositions of a toral endomorphisms with a translation is called an affine
endomorphism. This provides another class of MPDS on the torus. See [32] for the
characterization of the ergodic properties of affine endomorphisms.
Let Σ be a compact metric space and consider the space of sequences X =
Σ Z . The (two-sided) shift is the homeomorphism T : X → X defined by T x :=
{xn+1 }n∈Z for x = {xn }n∈Z .
6 1 Introduction
Example 1.3 (Bernoulli shifts) Given a probability measure μ ∈ Prob(Σ), the shift
T : X → X preserves the product probability measure μZ . The MPDS (X, μZ , T )
is called a Bernoulli shift. Bernoulli shifts are ergodic and mixing.
Let Σ be a compact metric space and let F be its Borel σ -field. A Markov or
stochastic kernel on Σ is a function K : Σ × F → [0, 1] such that
(1) for every x ∈ Σ, K (x, ·) ∈ Prob(Σ) is a probability measure in Σ,
(2) for every A ∈ F, the function x → K (x, A) is F-measurable.
The iterated Markov kernels are defined recursively, setting
(a) K 1 = K ,
(b) K n+1 (x, A) = Σ K n (y, A) K (x, dy), for all n ≥ 1.
Each power K n is itself a Markov kernel on Σ. A probability measure μ on Σ is
called K -stationary if for all A ∈ F , μ(A) = K (x, A) μ(d x).
Given a stochastic kernel K on Σ and a K -stationary probability measure μ, there
exists a unique probability P on the sequence space X = Σ Z such that the stochastic
process {en : X → Σ}n≥0 , en (x) := xn , has initial distribution μ and transition
kernel K , i.e., for all x ∈ Σ and A ∈ F,
1. P[ e0 ∈ A ] = μ(A),
2. P[ en ∈ A | en−1 = x ] = K (x, A).
Example 1.4 (Markov shifts) The two-sided shift T : X → X preserves the proba-
bility measure P. The MPDS (X, P, T ) is called a Markov shift.
When Σ is finite, the Markov shift is ergodic iff for all x, y ∈ Σ there exists
m = m(x, y) ∈ N such that K m (x, y) > 0, and it is mixing iff there exists m ∈ N
such that for all x, y ∈ Σ, K m (x, y) > 0.
Definition 1.3 We say that (K , μ) is strongly mixing if there are constants C < ∞
and 0 < ρ < 1 such that for every f ∈ L ∞ (Σ), all x ∈ Σ and n ∈ N,
f (y) K n (x, dy) − f (y) μ(dy) ≤ C ρ n f ∞ .
Σ Σ
Theorem 1.1 Given an ergodic MPDS (X, μ, T ), for any observable ϕ and for μ
almost every point x ∈ X ,
1
n−1
lim ϕ(T j x) = ϕ dμ .
n→+∞ n X
j=0
Let Gr(Rm ) denote the set of all linear subspaces of Rm . This space has a natural
topology (see Chap. 2). It is a compact Riemannian manifold, called the Grass-
mannian of the Euclidean space Rm . For each 0 ≤ k ≤ m we define
These subsets are precisely the connected components of the Grassmannian Gr(Rm ).
A subset B ⊆ X × Rm is called a measurable bundle over X when there exists a
measurable function E : X → Gr(Rm ) such that
B = { (x, v) ∈ X × Rm : v ∈ E(x) } .
Definition 1.4 We call linear cocycle on B over (X, μ, T ) any measurable map
FA : B → B defined by a measurable family of linear maps A(x) : B(x) →
B(T x) through the expression FA (x, v) := (T x, A(x)v).
We call the linear maps A(x) the fiber action and the transformation T the base
dynamics of the linear cocycle FA . We identify FA with the pair (T, A), or simply
with A when T is fixed. When B = X × Rm is a trivial bundle, a linear cocycle
on B over (X, μ, T ) is determined by a measurable function A : X → Mat(m, R).
Except for the more technical purposes in Chap. 4, we will always consider linear
cocycles on trivial bundles.
A cocycle A is said to be μ-integrable if
log+ A(x) μ(d x) < +∞ .
X
Definition 1.6 Let Σ be a compact metric space and let μ ∈ Prob(Σ) be a proba-
bility measure on Σ.
A random Bernoulli cocycle is any cocycle A : X → Mat(m, R) over the
Bernoulli shift (X, μZ , T ), where X = Σ Z is the space of sequences in Σ, and
the function A depends only on the first coordinate x0 .
www.Ebook777.com
1.2 The Main Concepts 9
1
lim logA(n) (x) v = λ j .
n→+∞ n
The second MET of Oseledets [23] gives a more precise description of the fiber
asymptotic behavior of a linear cocycle over an invertible ergodic MPDS.
Theorem 1.4 (MET I) Let (X, μ, T ) be an invertible and ergodic MPDS, and let
FA : B → B be a μ-integrable linear cocycle. Then there exist numbers λ1 > λ2 >
· · · > λk+1 ≥ −∞ and a family of measurable sub-bundles B j ⊂ B, 1 ≤ j ≤ k +1,
such that for μ-almost every x ∈ X ,
(a) B(x) = ⊕k+1 j=1 B j (x),
(b) each sub-bundle B j is A-invariant,
1
(c) for every v ∈ B j (x) \ {0}, lim logA(n) (x) v = λ j ,
n→±∞ n
1
(d) lim logsin min (⊕ j≤l B j (T n x), ⊕ j>l B j (T n x)) = 0, for any l = 2, . . . , k.
n→±∞ n
Hence, by Kingman’s ergodic theorem the following limit exists for μ-a.e. x ∈ X ,
1
L 1 (A) := lim logA(n) (x).
n→+∞ n
This limit is exactly the first Lyapunov exponent λ1 in both versions of the MET.
Using exterior powers of the cocycle A, the remaining Lyapunov exponents can
be expressed in terms of similar limits. Again by Kingman’s ergodic theorem the
following limits exist for all 1 ≤ j ≤ m, and μ-a.e. x ∈ X ,
10 1 Introduction
1
j
1
Λ j (A) = lim log∧ j A(n) (x) = lim log sk (A(n) (x)) .
n→+∞ n n→+∞ n
k=1
where ∧ j g denotes the jth exterior power of g ∈ Mat(m, R). The Lyapunov expo-
nents of the cocycle A can then be characterized by
1
L j (A) = Λ j (A) − Λ j−1 (A) = lim log s j (A(n) (x)) ,
n→+∞ n
with the convention that Λ0 (A) = 0. The Lyapunov spectrum of a cocycle A is the
sequence of its Lyapunov exponents:
Definition 1.10 We say that Ξ and A are compatible if for every integer N ∈ N, for
every set F ∈ F N (A) and for every ε > 0, there is an observable ξ ∈ Ξ such that:
1 F ≤ ξ and ξ dμ ≤ μ(F) + ε . (1.5)
X
Definition 1.11 Let A ∈ Cm and 1 ≤ k ≤ m such that L k (A) > L k+1 (A).
We define E k− (A) : X → Gr m−k (Rm ) as the measurable component of the Ose-
ledets filtration of A corresponding to the Lyapunov exponents ≤ L k+1 (A). If T is
invertible, we define E k+ (A) : X → Gr k (Rm ) to be the direct sum of the components
of the Oseledets decomposition corresponding to Lyapunov exponents ≥ L k (A).
cocycles. We will not study these types of problems in this monograph, but we will
discuss them briefly in the last chapter.
In probability theory and harmonic analysis there are several inequalities describing
the deviation of a function from its mean.
The most basic result of this kind is Chebyshev’s inequality. We formulate it in
its exponential form. For any t, λ > 0 and for any random variable X ,
P[ X − E(X ) ≥ λ ] ≤ e−λ t E[et |X −E(X )| ] . (1.6)
where the sup is taken over all intervals I ⊂ T and f I = |I1| I f . There is a
universal constant c > 0 such that if f BMO < +∞, then for all λ > 0,
x ∈ T : f − f T ≥ λ ≤ e−c λ/ f BMO . (1.7)
Let X 0 , X 1 , X 2 , . . . be a real valued random process and denote by Sn = n−1
j=0 X j
the corresponding sum process. Tail events of this process correspond to the deviation
of its averages n1 Sn from their means E( n1 Sn ).
There are several types of large deviation inequalities describing tail events, such
as Chernoff bounds (see [28]), which we formulate for a random i.i.d. process {X n }:
1
P[ Sn − μ ≥ λ ] < C max{e−(c λ /σ ) n , e−(λ/K ) n }
2 2
(1.8)
n
for some universal constants C < ∞ and c > 0, where μ = E(X 0 ), σ 2 = var(X 0 )
and K = X 0 ∞ .
The asymptotic behavior of tail events forms the subject of the theory of large
deviations (see [25]). A classical result in this theory is the following theorem due
to H. Cramér.
Theorem 1.5 If the random process {X n } is i.i.d. with mean μ = E(X 0 ) and finite
moment generating function M(t) := E[et X 0 ] < +∞ for all t > 0, then
1 1
lim log P
Sn − μ > ε = −I (ε)
n→+∞ n n
We now give a general formulation of the large deviation principle (see [25]).
Given an increasing sequence of integers {rn } and a lower semi-continuous function
I : R → [0, +∞), we say that the random process {X n } satisfies a large deviation
principle with normalizing sequence {rn } and rate function I , if for any closed set
F ⊂ R,
1 1
lim sup log P Sn ∈ F ≤ − inf I (x) ,
n→+∞ r n n x∈F
We note that the large deviation principle holds under the assumptions of The-
orem 1.5 with rn = n and the rate function specified in that theorem. Other large
deviation principles, including for Markov processes, can be found in [30].
Given a dynamical system (X, μ, T ), any observable ξ : X → R determines
the random process X n = ξ ◦ T n . Let ξ = X ξ dμ be the mean of this random
process, i.e., the space average of the observable, and let Sn ξ := n−1 j=0 ξ ◦ T be
j
the corresponding sum process, i.e., the usual Birkhoff sums. There are many results
regarding large deviations for dynamical systems (see for instance [17, 21, 26, 33]).
Given a linear cocycle A : X → Mat(m, R) over the dynamical system (X, μ, T ),
we say that A satisfies a large deviation principle with rate function I (ε) if
1 1
lim log μ logA(n) − L 1 (A) > ε = −I (ε) .
n→+∞ n n
Many large deviation principles, as well as other limit theorems, have been devel-
oped both in the context of random Bernoulli cocycles [8, 20, 29], and that of random
Markov cocycles [7].
The large deviation principle for additive processes (Birkhoff sums of some given
observable), and respectively for sub-additive processes (products of matrix valued
random processes), are asymptotic results. Our study of continuity properties of
Lyapunov exponents of linear cocycles does not require asymptotic statements, but
only good upper bounds on the measure of tail events, for the random processes given
by the base and fiber dynamics. We call these bounds large deviation type (LDT)
estimates.
To describe these LDT estimates we introduce the following formalism. From
now on, ε, ι : (0, ∞) → (0, ∞) will represent functions that describe respectively,
the size of the deviation from the mean and the measure of the deviation set.
Definition 1.12 Let 1 < p ≤ ∞ and let E and I be two families of functions. A set
P = P( p) of LDT parameters (with constant p) is a collection of triplets in N × E × I
such that
(1) The set E consists of non-increasing functions ε : (0, ∞) → (0, ∞).
14 1 Introduction
Example 1.5 The set E consists either of constant functions ε(t) ≡ ε, with 0 < ε <
1, or else of powers ε(t) ≡ C t −a for some C, a > 0.
The set I consists of functions of one of the following kinds:
(a) exponentials ι(t) ≡ M e−c t ,
(b) sub-exponentials ι(t) ≡ M e−c t
b
We finish remarking that the known large deviation principles, in the context
of random cocycles (see [7, 8, 15]), do not provide uniform fiber LDT estimates.
Hence they can not be used directly in our scheme for proving continuity of Lyapunov
exponents. However, the spectral theory approach used to prove them is enough to
derive uniform fiber LDT estimates.
Proving base and fiber (uniform) LDT estimates for quasi-periodic cocycles uses
harmonic analysis and potential theory tools, along with the arithmetic properties of
the torus translation.
(b) If T : X → X is invertible then for some α = α(A) > 0 and for all B1 , B2 ∈ V,
This theorem is proved in Chaps. 3 and 4. Item (a) and the general continuity
statement are proved in Theorem 3.1. Item (b) follows from Theorems 4.7 and 4.8.
We present two applications of the ACT, to random and to quasi-periodic cocycles.
For each of these models we specify the space of cocycles C = {(Cm , dist)}m∈N and
the sets of deviation functions E and I.
We begin with the random Markov application of the ACT, studied in Chap. 5.
Let (X = Σ Z , P, T ) be a Markov shift, where Σ is a compact metric space,
and the probability P on X is determined by a Markov kernel K on Σ and by a
K -stationary probability measure μ.
Define the space B∞ m (K ) of measurable functions A : Σ × Σ → GL(m, R) such
that A and A−1 are uniformly bounded, i.e., A∞ < +∞ and A−1 ∞ < +∞. We
endow this space with the metric d∞ (A, B) = A − B∞ .
Each A ∈ B∞ m (K ) determines the random Markov cocyle  : X → GL(m, R),
defined by Â(x) := A(x0 , x1 ). We will make the notational identification A = Â.
Theorem 1.7 If (K , μ) is strongly mixing then all conclusions of Theorem 1.6 hold
on the space of measurable cocycles C = {(Im∞ (K ), d∞ )}m∈N with a Hölder mod-
ulus of continuity.
1.5 Summary of Results 17
The main ingredient in this application is the following fiber uniform LDT estimate
of exponential type (see Theorem 5.3).
0 < b < 1.
t
k · ω ≥ for all k ∈ Zd \ {0} ,
|k|
d+δ0
where for any real number x we write x := mink∈Z x − k .
We formulate the continuity statement, which follows from Theorem 6.1.
18 1 Introduction
The continuity of Lyapunov exponents for spaces of cocycles with low regularity
was studied by Arbieto and Bochi [1], Bessa and Vilarinho [4] and by others. Since
the results in this monograph are closer to the high regularity regime, we will focus
our review on these types of models.
There is a large amount of work dedicated to the case of analytic, quasi-periodic
cocycles. Classic results on the subject are due to Goldstein and Schlag [14] and
to Bourgain and Jitomirskaya [10]. These results refer to Schrödinger cocycles, as
defined in Sect. 7.2, and continuity is understood relative to the energy parameter
and/or the frequency. In [10], the authors prove joint continuity in the energy E and
the frequency ω, at all points (E, ω) with ω irrational.
In [14], for the one frequency case, assuming a strong Diophantine condition on
the frequency, the authors prove a sharp fiber LDT estimate and establish the AP for
SL(2, R) matrices. Based on these ingredients, they develop an inductive procedure
that leads to Hölder continuity of the (top) Lyapunov exponent as a function of
the energy E, under the assumption of a positive lower bound on the Lyapunov
exponent. A similar approach is applied to the multifrequency Diophantine torus
translation case, leading to weak-Hölder continuity of the Lyapunov exponent, the
weaker modulus of continuity being due to a weaker version of the fiber LDT estimate
available in this case.
Extensions of the ideas and results in [14] to other related models were obtained
in [11, 18, 19].
Bourgain proved in [9] joint continuity in energy and frequency for the multifre-
quency torus translation model.
A higher dimensional version of the AP, along with a higher dimensional version
of the result in [14], were obtained in [27] for Schrödinger-like cocycles, under the
restrictive assumption that all Lyapunov exponents are simple. It was also indicated
1.6 Literature Review 19
in [27] that this method is in some sense modular, a statement that motivated in part
our current work.
With motivations that are both intrinsic and related to mathematical physics prob-
lems (e.g. spectral properties of Jacobi-type operators), the study of continuity prop-
erties of the Lyapunov exponents has been extended from Schrödinger cocycles to
more general ones, including higher dimensional cocycles and/or cocycles with sin-
gularities. Each extension comes with significant technical challenges, requiring new
methods.
C. Marx and S. Jitomirskaya proved joint continuity in energy and frequency
(one frequency case) for Mat(2, C)-valued analytic cocycles (see [16] and references
therein). Using a different approach, A. Ávila, S. Jitomirskaya, C. Sadel extended
this result to multidimensional (i.e. Mat(m, C)-valued) analytic cocycles (see [2]).
We note that both results mentioned above ([2, 16]) treating one frequency torus
translations, rely crucially on the convexity of the top Lyapunov exponent of the
complexified cocycle as a function of the imaginary variable, by firstly establishing
continuity away from the torus. This approach immediately breaks down in the
multifrequency case.
Our work in [12] presents a geometric, conceptual approach to the AP, which
allows us to generalize it to higher dimensions, namely to blocks of GL(m, R) matri-
ces, and further (see Chap. 2), to any blocks of nonzero matrices in Mat(m, R). We
use this general AP in [12] to prove Hölder (or weak-Hölder for multifrequency
translations) continuity of the Lyapunov exponents of GL(m, R)-valued analytic
cocycles in a neighborhood of a cocycle with simple Lyapunov exponents. More-
over, continuity of all Lyapunov exponents (but without a modulus of continuity)
holds everywhere, regardless of the multiplicity of the Lyapunov exponents.
Our next goal was to handle cocycles with singularities (i.e. not necessarily
GL(m, R)-valued), which, as explained in Chap. 6, is especially delicate in the mul-
tifrequency case. While unlike in [2, 16], we do require Diophantine translations,
and the translation frequency is fixed, our method applies equally to translations on
the one or the higher dimensional torus.
At the other end of the type of ergodic behavior of the base dynamics—the random
case, continuity results for linear coccycles over Bernoulli shifts in the generic case
go back to Furstenberg and Kifer [13].
Le Page proved in [24] Hölder continuity of the top Lyapunov exponent for a
one-parameter family of cocycles over the Bernoulli shift, under irreducibility and
contraction assumptions, which are assumed to hold uniformly throughout this fam-
ily. We are not aware of any generalization of this theorem of Le Page to irreducible
cocycles over strongly mixing Markov shifts.
Compared with [24], our result provides continuity of all Lyapunov exponents,
regardless of the gaps in the Lyapunov spectrum and it holds in the space of all
irreducible cocycles, not just for one-parameter families. It is also more general
since we consider cocycles over mixing Markov shifts, and not just over the Bernoulli
shifts. Moreover, we assume that our cocycles are locally constant, i.e. they depend
on a finite number of coordinates, and not just on one coordinate.
20 1 Introduction
Bocker-Neto and Viana [6] proved continuity of the Lyapunov exponents for
two-dimensional cocycles over Bernoulli shifts without any irreducibility assump-
tions. This result does not provide a modulus of continuity. A higher dimensional
version of this result was announced by A. Ávila, A. Eskin and M. Viana (see the
monograph [31]). An extension of results from [6] to a particular type of cocycles
over Markov systems (particular in the sense that the cocycle still depends on one
coordinate, as in the Bernoulli case) was obtained in [22]. Other related results were
recently obtained in [3].
We note, for the interested reader, that a general one-stop reference for continuity
results for random cocycles is Viana’s monograph [31].
References
1. A. Arbieto, J. Bochi, L p -generic cocycles have one-point Lyapunov spectrum. Stoch. Dyn.
3(1), 73–81 (2003). MR 1971187 (2004a:37063)
2. A. Ávila, S. Jitomirskaya, C. Sadel, Complex one-frequency cocycles. J. Eur. Math. Soc.
(JEMS) 16(9), 1915–1935 (2014). MR 3273312
3. L. Backes, A.W. Brown, C. Butler, Continuity of Lyapunov exponents for cocycles with invari-
ant holonomies, preprint (2015), 1–34
4. M. Bessa, H. Vilarinho, Fine properties of L p -cocycles which allow abundance of simple and
trivial spectrum. J. Differ. Equ. 256(7), 2337–2367 (2014). MR 3160445
5. J. Bochi, Genericity of zero Lyapunov exponents. Ergodic Theor. Dynam. Syst. 22 (2002)(6),
1667–1696. MR 1944399 (2003m:37035)
6. C. Bocker-Neto, M. Viana, Continuity of Lyapunov exponents for random 2d matrices, preprint,
to appear in Ergodic Theory and Dynamical Systems (2010), 1–38
7. P. Bougerol, Théorèmes limite pour les systèmes linéaires à coefficients markoviens. Probab.
Theor. Relat. Fields 78(2), 193–221 (1988). MR 945109 (89i:60122)
8. P. Bougerol, J. Lacroix, Products of random matrices with applications to Schrödinger oper-
ators, in Progress in Probability and Statistics, vol. 8 (Birkhäuser Boston Inc, Boston, MA,
1985). MR 886674 (88f:60013)
9. J. Bourgain, Positivity and continuity of the Lyapounov exponent for shifts on Td with arbitrary
frequency vector and real analytic potential. J. Anal. Math. 96, 313–355 (2005). MR 2177191
(2006i:47064)
10. J. Bourgain, S. Jitomirskaya, Continuity of the Lyapunov exponent for quasiperiodic operators
with analytic potential. J. Statist. Phys. 108(5–6), 1203–1218 (2002). Dedicated to David Ruelle
and Yasha Sinai on the occasion of their 65th birthdays. MR 1933451 (2004c:47073)
11. J. Bourgain, M. Goldstein, W. Schlag, Anderson localization for Schrödinger operators on
Z with potentials given by the skew-shift. Comm. Math. Phys. 220(3), 583–621 (2001). MR
1843776 (2002g:81026)
12. P. Duarte, S. Klein, Continuity of the Lyapunov exponents for quasiperiodic cocycles. Comm.
Math. Phys. 332(3), 1113–1166 (2014). MR 3262622
13. H. Furstenberg, Y. Kifer, Random matrix products and measures on projective spaces. Isr. J.
Math. 46(1–2), 12–32 (1983). MR 727020 (85i:22010)
14. M. Goldstein, W. Schlag, Hölder continuity of the integrated density of states for quasi-periodic
Schrödinger equations and averages of shifts of subharmonic functions. Ann. Math. (2) 154(1),
155–203 (2001). MR 1847592 (2002h:82055)
15. H. Hennion, L. Hervé, Limit Theorems for Markov chains and Stochastic Properties of Dynam-
ical Systems by Quasi-Compactness, vol. 1766, Lecture notes in mathematics (Springer, Berlin,
2001)
References 21
16. S. Jitomirskaya, C.A. Marx, Analytic quasi-perodic cocycles with singularities and the Lya-
punov exponent of extended Harper’s model. Comm. Math. Phys. 316(1), 237–267 (2012).
MR 2989459
17. Y. Kifer, Large deviations in dynamical systems and stochastic processes. Trans. Am. Math.
Soc. 321(2), 505–524 (1990). MR 1025756 (91e:60091)
18. S. Klein, Anderson localization for the discrete one-dimensional quasi-periodic Schrödinger
operator with potential defined by a Gevrey-class function. J. Funct. Anal. 218(2), 255–292
(2005). MR 2108112 (2005m:82070)
19. S. Klein, Localization for quasiperiodic Schrödinger operators with multivariable Gevrey
potential functions. J. Spectr. Theor. 4, 1–53 (2014)
20. É. Le Page, Théorèmes limites pour les produits de matrices aléatoires, Probability measures
on groups (Oberwolfach, 1981), Lecture notes in mathematics, vol. 928 (Springer, Berlin-New
York, 1982), pp. 258–303. MR 669072 (84d:60012)
21. A.O. Lopes, Entropy and large deviation, Nonlinearity 3(2), 527–546 (1990). MR 1054587
(91m:58092)
22. E. Malheiro, M. Viana, Lyapunov exponents of linear cocycles over Markov shifts, preprint
(2014), 1–25
23. V.I. Oseledec, A multiplicative ergodic theorem. Characteristic Ljapunov, exponents of dynam-
ical systems. Trudy Moskov. Mat. Obšč. 19, 179–210 (1968). MR 0240280 (39 #1629)
24. É. Le Page, Régularité du plus grand exposant caractéristique des produits de matrices aléa-
toires indépendantes et applications. Annales de l’institut Henri Poincaré (B) Probabilités et
Statistiques 25, no. 2, 109–142 (1989) (fre)
25. F. Rassoul-Agha, T. Seppäläinen, A course on large deviations with an introduction to Gibbs
measures. Graduate Studies in Mathematics, vol. 162 (American Mathematical Society, Prov-
idence, RI, 2015). MR 3309619
26. L. Rey-Bellet, L.-S. Young, Large deviations in non-uniformly hyperbolic dynamical systems.
Ergodic Theor. Dynam. Syst. 28(2), 587–612 (2008). MR 2408394 (2009c:37029)
27. W. Schlag, Regularity and convergence rates for the Lyapunov exponents of linear cocycles. J.
Mod. Dyn. 7(4), 619–637 (2013). MR 3177775
28. T. Tao, Topics in random matrix theory. Graduate Studies in Mathematics, vol. 132 (American
Mathematical Society, Providence, RI, 2012). MR 2906465 (2012k:60023)
29. V.N. Tutubalin, Limit theorems for a product of random matrices. Teor. Verojatnost. i Primenen.
10, 19–32 (1965). MR 0175169 (30 #5354)
30. S.R.S. Varadhan, Large deviations and applications. École d’Été de Probabilités de Saint-Flour
XV-XVII, 1985–87, Lecture Notes in Mathematics, vol. 1362 (Springer, Berlin, 1988), pp. 1–49.
MR 983371 (89m:60068)
31. M. Viana, Lectures on Lyapunov exponents. Cambridge Studies in Advanced Mathematics
(Cambridge University Press, 2014)
32. P. Walters, An introduction to ergodic theory. Graduate Texts in Mathematics, vol. 79 (Springer,
New York, 1982). MR 648108 (84e:28017)
33. L-S. Young, Large deviations in dynamical systems. Trans. Am. Math. Soc. 318(2), 525–543
(1990). MR 975689 (90g:58069)
Chapter 2
Estimates on Grassmann Manifolds
Abstract The main result of this chapter, called the Avalanche Principle (AP),
relates the expansion of a long product of matrices with the product of expansions
of the individual matrices. This principle was introduced by M. Goldstein and
W. Schlag in the context of SL(2, C) matrices. Besides extending the AP to matrices
of arbitrary dimension and possibly non-invertible, the geometric approach we use
here provides a relation between the most expanding (singular) directions of such a
long product of matrices and the corresponding singular directions of the first and last
matrices in the product. The AP along with other estimates on the action of matrices
on Grassmann manifolds will play a fundamental role in the next chapters, when we
establish the continuity the LE and of the Oseledets decomposition.
The projective space is the simplest compact model to study the action of a linear
map. Given an n-dimensional Euclidean space V , consider the equivalence relation
defined on V \ {0} by u ≡ v if and only if u = λ v for some λ = 0. For v ∈ V \ {0},
the set v̂ := {λ v : λ ∈ R \ {0}} is the equivalence class of the vector v relative to
this relation. The projective space of V is the quotient P(V ) := {v̂ : v ∈ V \ {0}} of
V \ {0} by this equivalence relation. It is a compact topological space when endowed
with the quotient topology.
The second metric, d, corresponds to the Euclidean distance. More precisely, given
u, v ∈ S(V ),
d(û, v̂) := min{u − v, u + v} (2.2)
measures the smallest chord of the arcs between u and v and between u and −v. The
third metric, δ, measures the sine of the arc between two points on the sphere. More
precisely, given u, v ∈ S(V ),
u ∧ v
δ(û, v̂) := = sin(∠(u, v)). (2.3)
u v
The fact that δ is a metric on P(V ) follows from the sine addition law, which implies
that sin(θ + θ ) ≤ sin θ + sin θ , for all θ, θ ∈ [0, π2 ].
These three distances are equivalent. For all û, v̂ ∈ P(V ),
δ(û, v̂) = sin ρ(û, v̂) and d(û, v̂) = chord ρ(û, v̂). (2.4)
The inequalities
2θ π
≤ sin θ ≤ chord θ = 2 sin(θ/2) ≤ θ ∀0 ≤ θ ≤
π 2
imply that
2
ρ(û, v̂) ≤ δ(û, v̂) ≤ d(û, v̂) ≤ ρ(û, v̂). (2.5)
π
Because of (2.4), these three metrics determine the same group of isometries on the
projective space.
www.Ebook777.com
2.1 Grassmann Geometry 25
v1 ∧ · · · ∧ vk ∧ w1 ∧ · · · ∧ wk = v1 ∧ · · · ∧ vk + w1 ∧ · · · ∧ wk .
Let Λnk be the set of all k-subsets I = {i1 , . . . , ik } ⊂ {1, . . . , n}, with i1 < · · · < ik ,
and order it lexicographically. Given a basis {e1 , . . . , en } of V , define for each I ∈ Λnk ,
the kth exterior product eI = ei1 ∧ · · · ∧ eik .Then the ordered family {eI : I ∈ Λnk }
n
is a basis of ∧k V . In particular dim ∧k V = k .
The exterior algebra ∧∗ V inherits an Euclidean structure from V . More pre-
cisely, there is a unique inner product on ∧∗ V such that for any orthonormal basis
26 2 Estimates on Grassmann Manifolds
Interestingly, the norm of the simple k-vector v1 ∧· · ·∧vk is equal to the k-dimensional
volume of the parallelepiped generated by its factors vj . More precisely,
where Volk stands for the k-dimensional Hausdorff measure. To explain this fact first
notice that if the vectors v1 , . . . , vk are pairwise orthogonal then
v1 ∧ · · · ∧ vk v1 vk
= ∧ ··· ∧ =1
v1 · · · vk v1 vk
because the vectors {vj /vj : j = 1, . . . , k} are orthonormal. This shows that v1 ∧
· · · ∧ vk = v1 · · · vk and establishes (2.6) in this case. In general we use the
Gram-Schmidt orthogonalization method, defining recursively
j−1
vj , v
v1 = v1 and vj = vj − i
vi for j = 2, . . . , k.
i=1
vi 2
At each step, when we replace vj by vj , both wedge products and k-volumes are
preserved. Hence v1 ∧ · · · ∧ vk = v1 ∧ · · · ∧ vk and
Formula (2.6) also implies that for any simple vectors e = e1 ∧ · · · ∧ er and
f = f1 ∧ · · · ∧ fs in V ,
e ∧ f ≤ e f . (2.7)
operator, ∗ : ∧∗ V → ∧∗ V defined by
The Hodge star operator maps ∧k V isomorphically, and isometrically, onto ∧n−k V ,
for all 0 ≤ k ≤ n. Geometrically it corresponds to the orthogonal complement
operation on linear subspaces, i.e., for any simple k-vector,
The geometric meaning of the ∨-operation reduces by duality to that of the sum
∧-operation and the complement ∗-operation.
Any linear map g : V → V induces a linear map ∧k g : ∧k V → ∧k V , called the
kth exterior power of g, such that for all v1 , . . . , vk ∈ V ,
This construction is functorial in the sense that for all linear maps g, g : V → V ,
The considerations in Sect. 2.1.2 show that the Plücker embedding satisfies the fol-
lowing relations:
Proposition 2.1 Given E ∈ Gr k (V ), E ∈ Gr k (V ), consider unit vectors v ∈ Ψ (E)
and v ∈ Ψ (E ).
(a) If (E, E ) ∈ Gr k,k (∩) then ψ(E ∩ E ) = v
∨ v .
(b) If (E, E ) ∈ Gr k,k (+) then ψ(E + E ) = v
∧ v .
A duality between sums and intersections stems from these facts.
Proof The first statement is obvious because if E ∩ (Kg) = {0} then K(g|E ) = {0}. If
E +(Rg) = V then, since Kg∗ = (Rg)⊥ , we have E ⊥ ∩(Kg∗ ) = E ⊥ ∩(Rg)⊥ = (E +
Rg)⊥ = {0}. Hence by 1, the linear map g∗ |E ⊥ : E ⊥ → g∗ (E ⊥ ) is an isomorphism.
It is now enough to remark that g∗ (E ⊥ ) = g−1 (E)⊥ . In fact, the inclusion g∗ (E ⊥ ) ⊂
www.Ebook777.com
30 2 Estimates on Grassmann Manifolds
Hence both g∗ (E ⊥ ) and g−1 (E)⊥ have dimension equal to dim(E ⊥ ), and the equality
follows.
We warn the reader that the notation ϕg is used for both the projective and the
Grassmannian actions of g ∈ L (V ).
Similarly, given k ≥ 0 such that k + dim(Rg) ≥ n = dim V , the pull-back by g
is the map ϕg−1 : Gr k (g−1 ) ⊂ Gr k (V ) → Gr k (V ), E → g−1 E, where:
Definition 2.5 The domain is defined by
Gr k (g−1 ) := {E ∈ Gr k (V ) : E + (Rg) = V }.
From the proof of Proposition 2.1 we obtain a duality between push-forwards and
pull-backs which can be expressed as follows.
Proposition 2.3 Given g ∈ L (V ) and k ≥ 0 such that k + dim(Rg) ≥ n = dim V ,
we have Gr k (g−1 ) = Gr n−k (g∗ )⊥ and for all E ∈ Gr k (g−1 ),
(g−1 E)⊥ = g∗ (E ⊥ ).
In Sect. 2.3 we derive a modulus of Lipschitz continuity, w.r.t. the metric δ, for
the sum and intersection operations.
the length of the flag F, and the length of the signature τ . Let F(V ) be the set of all
flags in V , and define Fτ (V ) to be the space of flags with a given signature τ . Two
special cases of flag spaces are the projective space P(V ) = Fτ (V ), when τ = (1),
and the Grassmannian Gr k (V ) = Fτ (V ), when τ = (k).
The general linear group GL(V ) acts naturally on F(V ). Given g ∈ GL(V )
the action of g on Fτ (V ) is given by the map ϕg : Fτ (V ) → Fτ (V ), ϕg F =
(gF1 , . . . , gFk ). The special orthogonal subgroup SO(V ) ⊂ GL(V ) acts transitively
on Fτ (V ). Hence, all flag manifolds Fτ (V ) are compact homogeneous spaces. Each
of them is a compact connected Riemannian manifold where the group SO(V ) acts
by isometries. Since Fτ (V ) ⊂ Gr τ1 (V ) × Gr τ2 (V ) × · · · × Gr τk (V ), the product
distances
τ ⊥ := (n − τk , . . . , n − τ1 ).
Similarly, given a signature τ such that τi + dim(Rg) ≥ n for all i, the pull-
back by g on flags is the map ϕg−1 : Fτ (g−1 ) ⊂ Fτ (V ) → Fτ (V ), ϕg−1 F :=
(g−1 F1 , . . . , g−1 Fk ), where:
Definition 2.8 The domain of ϕg−1 is defined by
Fτ (g−1 ) := {F ∈ Fτ (V ) : F1 + (Rg) = V }.
Singular value geometry refers here to the geometry of the singular value decom-
position (SVD) of a linear endomorphism g : V → V on some Euclidean space V .
It also refers to some geometric properties of the action of g on Grassmannians and
flag manifolds related to the singular value decomposition of g.
denote the sorted singular √ values of g.√The adjoint g∗ has the same singular values
as g because the operators g∗ g and g g∗ are conjugate.
The largest singular value, s1 (g), is the square root of the maximum value of Qg
over the unit sphere, i.e., s1 (g) = maxv=1 g v = g is the operator norm of g.
Likewise, the least singular value, sn (g), is the square root of the minimum value
of Qg over the unit sphere, i.e., sn (g) = minv=1 g v. This number, also denoted
by m(g), is called the least expansion of g. If g is invertible then m(g) = g−1 −1 ,
while otherwise m(g) = 0.
2.2 Singular Value Geometry 33
Therefore, by the considerations at the end of Sect. 2.1.2, the families of k-vectors
{vI = vi1 ∧· · ·∧vik : I ∈ Λnk } and {vI∗ = vi∗1 ∧· · ·∧vi∗k : I ∈ Λnk } form two singular bases
for ∧k g and ∧k g∗ , respectively, while the products sI = si1 . . . sik are the singular
values of both ∧k g and ∧k g∗ .
Proposition 2.6 For any 1 ≤ k ≤ dim V , ∧k g = s1 (g) . . . sk (g).
Proof The maximum product sI is attained when I = {1, . . . , k} ∈ Λnk . Hence
∧k g = s1 . . . sk .
The volume expansion factor of a linear map g : V → V between two Euclidean
spaces V and V is defined by
det + (g) := det(g∗ g).
On the third step we have used the fact that ∧n V has dimension 1.
Because of this property the volume expansion factor behaves multiplicatively.
Proposition 2.8 Given Euclidean spaces V , V and V , if g : V → V is an
isomorphism and g : V → V any linear map then
www.Ebook777.com
2.2 Singular Value Geometry 35
Proof The matrix A ∈ Mat(n, R) with entries aij =
gvi , vj represents the linear
map g in the given orthonormal bases. Consider the isometries U : Rn → V and
U : Rn → V respectively defined by Uei = vi and U ei = vi for all i = 1, . . . , n.
Then g = U AU ∗ and
sk (g)
gr k (g) := ≥ 1.
sk+1 (g)
∧k g2
gr k (g) = = gr 1 (∧k g).
∧k−1 g ∧k+1 g
Proof The first equality follows from Proposition 2.6. The two first singular values
of ∧k g are s1 (∧k g) = s1 (g) . . . sk−1 (g)sk (g) and s2 (∧k g) = s1 (g) . . . sk−1 (g)sk+1 (g).
Hence
36 2 Estimates on Grassmann Manifolds
s1 (∧k g) sk (g)
gr 1 (∧k g) = = = gr k (g).
s2 (∧k g) sk+1 (g)
Given g ∈ L (V ), if gr(g) > 1 then the singular value s1 (g) = g is simple.
Definition 2.14 In this case we denote by v(g) ∈ P(V ) the associated singular
direction, and refer to it as the g-most expanding direction.
By definition we have
ϕg v(g) = v(g∗ ). (2.16)
The subspace vk (g) is the direct sum of all singular directions associated with the
singular values s1 (g), . . . , sk (g). We have
The subspace vk (g) is the direct sum of all singular directions associated with the
singular values sn−k+1 (g), . . . , sn (g). Again we have
ϕg F = g F := (g F1 , . . . , g Fk ).
Proposition 2.12 Given g ∈ L (V ) such that gr τ (g) > 1, the push-forward induces
a map ϕg : Fτ (g) → Fτ (g∗ ) such that ϕg vτ (g) = vτ (g∗ ).
Hence
vτ ⊥ (g) = (vn−τk (g), . . . , vn−τ1 (g)) = (vτ1 (g), . . . , vτk (g))⊥ = vτ (g)⊥ .
Proposition 2.14 Given g ∈ L (V ) such that gr τ ⊥ (g) > 1, the pull-back induces a
map ϕg−1 : Fτ−1 (g) → Fτ−1 (g∗ ) such that ϕg−1 vτ (g) = vτ (g∗ ).
38 2 Estimates on Grassmann Manifolds
We end this section proving that the orthogonal complement involution conjugates
the push-forward action by g ∈ L (V ) with the pull-back action by the adjoint map g∗ .
Proposition 2.15 Given g ∈ L (V ) such that gr τ ⊥ (g) > 1, the action of ϕg−1 on
Fτ (V ) is conjugated to the action of ϕg∗ on Fτ ⊥ (V ) by the orthogonal complement
involution. More precisely, we have Fτ−1 (g) = Fτ ⊥ (g∗ )⊥ and Fτ−1 (g∗ ) = Fτ ⊥ (g)⊥ ,
and the following diagram commutes
ϕg∗
Fτ ⊥ (g∗ ) −−−−→ Fτ ⊥ (g)
⏐ ⏐
⏐ ⏐⊥
·⊥ · .
Fτ−1 (g) −−−−→ Fτ−1 (g∗ )
ϕg−1
Proof To see that Fτ−1 (g) = Fτ ⊥ (g∗ )⊥ , notice that the following equivalences hold:
F ∈ Fτ−1 (g) ⇔ F1 + Rg = V
⇔ F1⊥ ∩ Kg∗ = {0} ⇔ F ⊥ ∈ Fτ ⊥ (g∗ ).
Exchanging the roles of g and g∗ we obtain the relation Fτ−1 (g∗ ) = Fτ ⊥ (g)⊥ .
Finally, notice that it is enough to prove the diagram’s commutativity at the Grass-
mannian level. For that use Proposition 2.3.
Throughout this section let p̂, q̂ ∈ P(V ), and p ∈ p̂, q ∈ q̂ denote representative
vectors. The projective distance δ(p̂, q̂) was defined by
p, q2 p ∧ q
δ(p̂, q̂) := 1− = = sin ρ(p̂, q̂).
p2 q2 p q
We also define the minimum distance between any two subspaces E, F ∈ Gr(V ),
Proof Item (a) follows because orthogonal projections are self-adjoint operators.
Given w ∈ P⊥ , we have πu (w) = πv (w) = 0, which implies w ∈ K(πu − πv ).
Hence P⊥ ⊂ K(πu − πv ). Since u and v are non-collinear, πu − πv has rank 2. Thus
K(πu − πv ) = P⊥ , which proves (b).
For (c) we may assume that V = R2 and consider u = (u1 , u2 ), v = (v1 , v2 ), with
u1 + u22 = v12 + v22 = 1. The projections πu and πv are represented by the matrices
2
u12 u1 u2 v12 v1 v2
U= and V =
u1 u2 u22 v1 v2 v22
where α = v1 v2 −u1 u2 and β = v12 −u12 = −(v22 −u22 ). This proves that the restriction
of πv − πu to the plane P is anti-conformal. The similarity factor of this map is
πv − πu = πv (u) − u = πv⊥ (u) =
sin ∠(u, v)
Finally, since u − v, u v ⊥ v,
www.Ebook777.com
40 2 Estimates on Grassmann Manifolds
To finish (b) we still have to prove that πE − πF = πE,F ⊥ . Restricting our
attention to the subspace V0 = (E ∩ (E ∩ F)⊥ ) ⊕ (F ∩ (E ∩ F)⊥ ), because πE − πF
vanishes on V0⊥ we can assume that V = E ⊕ F. In particular dim V = 2k. Consider
the orthogonal invariant decomposition of Lemma 2.3. It is enough to check that the
relation πE − πF = πE,F ⊥ holds on each plane Pj . Therefore we may as well
assume that k = 1. Notice that over the subspace E we have πE − πF = πE,F ⊥ . Since
the linear map πE − πF is anti-conformal, the norm πE − πF is attained along E,
which implies that πE − πF = πE,F ⊥ . This proves item (b).
Since πE,F is an orthogonal projection all its singular values are in the range [0, 1].
Hence, for any unit vector u ∈ E, πE,F (u) ≥ m(πE,F ) ≥ det + (πE,F ). Thus
2.2 Singular Value Geometry 41
Item (c) follows taking the maximum over all unit vectors u ∈ E.
The following complementary quantity to the distance δ(p̂, q̂) plays a special role
in the sequel.
Definition 2.23 The α-angle between p̂ and q̂ is defined to be
|
p, q|
α(p̂, q̂) := = cos ρ(p̂, q̂).
p q
In order to give a geometric meaning to this angle we define the projective orthog-
onal hyperplane of p̂ ∈ P(V ) as
The number α(p̂, q̂) is the sine of the minimum angle between p̂ and Σ(q̂). As in
Definition (2.19), given a subspace F ⊂ V we write
α(p̂, q̂) = sin ρmin (p̂, Σ(q̂)) = δmin (p̂, Σ(q̂)) (2.20)
α(p̂, q̂) = 0 ⇔ δ(p̂, q̂) = 1 ⇔ p ⊥ q. (2.21)
As before, the number α(E, F) equals the sine of the minimum angle between E
and Σ(F).
Proposition 2.18 For any E, F ∈ Gr k (V ),
Next we characterize the angle α(E, F). Consider the notation of Definition 2.22.
Definition 2.27 We say that two τ -flags F, G ∈ Fτ (V ) are orthogonal, and we write
F ⊥ G, if Fj ⊥ Gj for some j = 1, . . . , k.
ατ (F, G) = 0 ⇔ G ⊥ F.
As in the previous cases, the number ατ (F, G) equals the sine of the minimum
angle between F and Σ(G).
Proposition 2.20 For any F, G ∈ Fτ (V ),
gn−1 . . . g1 g0
ρ(g0 , g1 , . . . , gn−1 ) := ∈ [1, +∞) .
gn−1 . . . g1 g0
The key concept of this section is that of angle between linear maps. The quantity
α(g, g ), for instance, is the sine of the angle between ϕg (v(g)) = v(g∗ ) and Σ(v(g )).
As we will see, this angle is a lower bound on the expansion rift of two linear maps
g and g .
44 2 Estimates on Grassmann Manifolds
Proof Let us write α = α(ŵ, v(g)) and σ = σ (g). Take a unit vector v ∈ v(g)
√ such
that ∠(v, w) is non obtuse. Then w = α v + u with u ⊥ v and u = 1 − α 2 .
∗
Choosing√a unit vector v √∈ v(g∗ ), we have gw = α g v∗ + gu with gu ⊥ v∗ and
gu ≤ √1 − α s2 (g) = 1 − α 2 σ g. We define the number 0 ≤ κ ≤ σ so that
2
and also
gw2 = α 2 g2 + gu2 = g2 α 2 + (1 − α 2 )κ 2
= g2 α 2 ⊕ κ 2 ≤ g2 α 2 ⊕ σ 2 ,
Next proposition relates the expansion rift ρ(g, g ) with the angle α(g, g ) and the
upper angle β(g, g ).
Proposition 2.23 Given g, g ∈ L (V ) with a (1)-gap pattern,
g g
α(g, g ) ≤ ≤ β(g, g )
g g
Proof Let α := α(g, g ) = α(v(g∗ ), v(g )) and take unit vectors v ∈ v(g), v∗ ∈ v(g∗ )
and v ∈ v(g ) such that
v∗ , v = α > 0 and g v = g v∗ .
Since ϕg (v(g)) = v(g∗ ), w = gg vv is a unit vector in ŵ = v(g∗ ). Hence, applying
Proposition 2.22(a) to g and ŵ, we get
g g v g g
α(g, g ) g = α(ŵ, v(g )) g ≤ ≤ ,
g v g
g w2
= a2 + (1 − a2 ) κ 2 = a2 ⊕ κ 2 .
g2
46 2 Estimates on Grassmann Manifolds
and √
gw a ∗ 1 − a2 g u
=√ v +√ .
g w a2 ⊕ κ 2 a2 ⊕ κ 2 g
√
The vector v can be written as v = α v∗ + w with w ⊥ v∗ and w = 1 − α2 .
Set now b := α(ϕg (ŵ), v(g )). Then
√
gw
αa 1 − a2
g u, v
b =
, v
≤ √ + √
g w a2 ⊕ κ 2 a2 ⊕ κ 2 g
√
αa κ 1 − a
g u
2
≤√ + √
, w
a2 ⊕ κ 2 a ⊕κ
2 2 g u
√
αa κ 1−a 2
≤√ + √ w
a2 ⊕ κ 2 a ⊕κ
2 2
√ √ √
αa κ 1 − a2 1 − α 2 α2 ⊕ κ 2
=√ + √ ≤ √ .
a2 ⊕ κ 2 a2 ⊕ κ 2 a2 ⊕ κ 2
We use Lemma 2.21 (7) on the last inequality. Finally, by Proposition 2.22(a) applied
to g ∈ L (V ) and the unit vector gw/gw ∈ ϕg (ŵ),
g g w ≤ g b2 ⊕ σ (g )2 g w
≤ g g b2 ⊕ σ (g )2 a2 ⊕ κ 2
≤ g g κ 2 ⊕ α 2 ⊕ σ (g )2 ≤ β(g, g ) g g,
where on the two last inequalities use items (6) and (5) of Lemma 2.21.
∧k (g g)
αk (g, g ) ≤ ≤ βk (g, g )
∧k g ∧k g
Proof Apply Proposition 2.23 to the composition (∧k g ) (∧k g). Notice that by Def-
inition 2.15, the Plücker embedding satisfies Ψ (vk (g)) = v(∧k g). Hence
αk (g, g ) = α(vk (g∗ ), vk (g )) =
v(∧k g), v(∧k g )
= α(∧k g, ∧k g ).
The next results show how close the bounds α(g, g ) and β(g, g ) are to each other
and to the rift ρ(g, g ).
2.2 Singular Value Geometry 47
n−1
gn−1 . . . g1 g0 n−1
(i)
α(g , gi ) ≤ ≤ β(g(i) , gi )
i=1
gn−1 . . . g1 g 0 i=1
In this section we will derive some inequalities describing quantities such as the
contracting behavior of a linear endomorphism on the projective space, the Lipschitz
dependence of a projective action on the acting linear endomorphism, the continuity
of most expanding directions as functions of a linear map, and the Lipschitz modu-
lus of continuity for sum and intersection operations on flag manifolds. Except for
Propositions 2.28 and 2.29, the content of this section will be only used in Chaps. 4
and 5.
p q 1 1
− ≤ max{ , } p − q.
p q p q
p q u v p q
− = − ≤ u − v = −
p q u v q q
= q−1 p − q = max{p−1 , q−1 } p − q.
q −
p, q p
vp (q) := .
q −
p, q p
Proof Let p ∈ p̂ and q ∈ q̂ be unit vectors such that θ = ∠(p, q) ∈ [0, π2 ]. We can
write q = (cos θ ) p + (sin θ ) vp (q). Hence
and
g p ∧ g q gp ∧ gvp (q)
δ(ϕg (p̂), ϕg (q̂)) = = (sin θ ) .
g p g q g p g q
where the superscript emphasizes the distance in matter. Given a projective map
f : X ⊂ P(V ) → P(V ), we denote by Lipd (f ) the least Lipschitz constant of f with
respect to the distance d. Next proposition refers to the projective metrics δ and ρ
defined in Sect. 2.1.1.
www.Ebook777.com
50 2 Estimates on Grassmann Manifolds
Proof Item (1) of this proposition follows from Proposition 2.22(b), because
δ(ŵ, v(g)) < r implies α(ŵ, v(g)) = 1 − δ(ŵ, v(g))2 ≥ 1 − r2.
Item (2) reduces to (1), because we have δ(û, v̂) = sin ρ(û, v̂), which implies that
B(ρ) (v̂, a) = B(δ) (v̂, sin a).
To prove (3), take unit vectors v ∈ v(g) and v∗ ∈ v(g∗ ) such that g v = g v∗ .
Because v is a g-most expanding vector, πv⊥∗ ◦ g = g ◦ πv⊥ ≤ s2 (g) ≤ κ g.
Given x̂ such that δ(x̂, v(g)) < r, and a unit vector x ∈ x̂, by Proposition 2.22(a)
g 1 1
≤ ≤√ .
gx α(x̂, v(g)) 1 − r2
σ (g) κr
δ(ϕg (x̂), v(g∗ )) ≤ δ(x̂, v(g)) ≤ √
α(x̂, v(g)) 1 − r2
1 1 # ⊥ $
(Dϕg )x v = πv⊥∗ (g v) + π" ⊥
ϕg (x) − πv∗ (g v).
gx gx
As before let V be a finite n-dimensional Euclidean space. Recall that the Grassmann
manifold Gr k (V ) identifies through the Plücker embedding with a submanifold of
P(∧k V ). Up to a sign, E ∈ Gr k (V ) is identified with the unit k-vector e = e1 ∧· · ·∧ek
associated to any orthonormal basis {e1 , . . . , ek } of E. Recall that the Grassmann
distance (2.10) on Gr k (V ) can be characterized by
Definition 2.31 Given E, F ∈ Gr(V ), we say that E and F are (∩) transversal if
E + F = V . Analogously, we say that E and F are (+) transversal if E ∩ F = {0}.
θ+ (E, F) := e ∧ f ,
θ∩ (E, F) := e⊥ ∧ f ⊥ .
Since the chosen unit vectors are unique up to a sign, these quantities are well-defined.
Next proposition establishes a Lispchitz modulus of continuity for the sum and
intersection operations on Grassmannians in terms of the previous quantities.
e∧f e ∧ f
d(E + F, E + F ) = −
e ∧ f e ∧ f
≤ K e ∧ f − e ∧ f
≤ K (e ∧ (f − f ) + (e − e ) ∧ f )
≤ K (e − e + f − f )
www.Ebook777.com
52 2 Estimates on Grassmann Manifolds
The measurement on the (∩) transversality admits the following lower bound in
terms of the angle in Definition 2.24.
θ∩ (E, F) ≥ αr (E, E ∩ F + F ⊥ ).
θ∩ (E , E ⊥ ) = det + (πE ,E )
=
∧r πE,E (v1 ∧ · · · ∧ vr ), v1 ∧ · · · ∧ vr
=
πE (v1 ) ∧ · · · ∧ πE (vr ), v1 ∧ · · · ∧ vr
=
v1 ∧ · · · ∧ vr , v ∧ · · · ∧ v
= αr (E, E ).
1 r
Hence
θ∩ (E, F) = e ∧ f ≥ e0 ∧ f0 − e ∧ f − e0 ∧ f0
≥ θ∩ (E0 , F0 ) − e ∧ (f − f0 ) − (e − e0 ) ∧ f0
≥ θ∩ (E0 , F0 ) − f − f0 − e − e0
≥ θ∩ (E0 , F0 ) − d(F, F0 ) − d(E, E0 ).
Proof Since πF ⊥ ,E ⊥ is an orthogonal projection, all its singular values are in [0, 1].
Thus, because det + (πF ⊥ ,E ⊥ ) is the product of all singular values, while m(∧i πF ⊥ ,E ⊥ )
is the product of the i smallest singular values, we have
Hence
Thus
Proof Exercise.
Let Dτ (V ) denote the space of all τ -decompositions, which we endow with the
following metric
dτ (E· , E· ) = max dτi −τi−1 (Ei , Ei ),
1≤i≤k+1
⊥
Notice that dim Fi = τi and dim Fk−i+1 = τk−i+1 = dim V −τi , i.e., the subspaces
Fi and Fk−i+1 have complementary dimensions. We will refer to this quantity as the
transversality measurement between the flags F and F .
In the next proposition we complete F and F to full flags of length k + 1 setting
Fk+1 = Fk+1 = V . Set also τ0 = 0 and τk+1 = dim V .
Proposition 2.36 If θ (F, F ) > 0 then the following is a direct sum decomposition
in the space Dτ (V ),
%
k+1
V = Fi ∩ Fk−i+2 ,
i=1
with dim(Fi ∩ Fk−i+2 ) = τi − τi−1 for all 1 ≤ i ≤ k + 1.
Proof Since the subspaces Fi and Fk−i+1 have complementary dimensions, the rela-
tion θ∩ (Fi , Fk−i+1 ) > 0 implies that
V = Fi ⊕ Fk−i+1 . (2.22)
By Lemma 2.5, θ∩ (Fi , Fk−i+2 ) ≥ θ∩ (Fi , Fk−i+1 ) > 0. Therefore Fi + Fk−i+2 =V
and
⊥
dim(Fi ∩ Fk−i+2 ) = τi + τk−i+2 − dim V
= τi + (dim V − τi−1 ) − dim V = τi − τi−1 .
Given two linear maps g0 , g1 ∈ L (V ) with τ -gap ratios such that ατ (g0 , g1 ) > 0,
they determine a τ -decomposition of V as intersection of the image by ϕg0 of the most
expanding τ -flag for g0 with the least expanding τ ⊥ -flag for g1 (see Definitions 2.18
and 2.21). The corresponding intersection transversality measurement is bounded
from below by the angle ατ (g0 , g1 ).
In particular, if ατ (g0 , g1 ) > 0 then the flags vτ (g0∗ ) and vτ ⊥ (g1 ) determine the
decomposition vτ (g0∗ ) vτ ⊥ (g1 ) ∈ Dτ (V ).
θ∩ (Fi , Fk−i+1 ) = θ∩ (vτi (g0∗ ), vτi (g1 )⊥ ) = ατi (vτi (g0∗ ), vτi (g1 )) = ατi (g0 , g1 ),
g − g
drel (g, g ) := .
max{g, g }
Notice that this relative distance is not a metric. It does not satisfy the triangle
inequality. We introduce it just to lighten the notation.
16
d(v(g1 ), v(g2 )) ≤ drel (g1 , g2 ).
1 − κ2
Proof Let g ∈ Lκ and λ > 0. The singular√ values (resp. singular vectors) of g are
the eigenvalues (resp. eigenvectors) of g∗ g. Hence sj (λ g) = λ sj (g), for all j. We
also have v(λg) = v(g) and gr(λ g) = gr(g).
Consider the subspace Lκ (1) := {g ∈ Lκ : g = 1}. The projection g →
g/g takes Lκ to Lκ (1). It also satisfies v(g/g) = v(g) and
g1 g2
− ≤ 2 drel (g1 , g2 ).
g1 g2
Hence we can focus our attention on the restricted map v : Lκ (1) → P(V ).
Let Lκ+ (1) denote the subspace of g ∈ Lκ (1) such that g = g∗ ≥ 0, i.e., g is
positive semi-definite.
Given g ∈ Lκ (1), we have g∗ g = 1 = g, gr(g∗ g) = gr(g)2 and v(g∗ g) =
v(g). Also, for all g1 , g2 ∈ Lκ (1),
Hence, the mapping g → g∗ g takes Lκ (1) to Lκ+2 (1) and has Lispschitz constant
2. Therefore, it is enough to prove that the restricted map v : Lκ+2 (1) → P(V ) has
(locally) Lipschitz constant 4 (1 − κ 2 )−1 .
Let δ0 be a small positive number and take 0 < ε0 δ40 . The size of δ0 will be
fixed throughout the rest of the proof according to necessity. Take h1 , h2 ∈ Lκ+2 (1)
such that h1 − h2 < ε0 and set p̂0 := v(h1 ). By Proposition 2.29 we have
⎛ ⎞
κ 2
δ
ϕh1 B(p̂0 , δ0 ) ⊂ B ⎝p̂0 ,
0 ⎠
⊂ B(p̂0 , δ0 ),
1 − δ02
where all balls refer to the projective sine-metric δ defined in (2.3). The second
inclusion holds if δ0 is chosen small enough. Take any p̂ ∈ B(p̂0 , δ0 ) and choose unit
vectors p ∈ p̂ and p0 ∈ p̂0 such that
p, p0 > 0. Then p =
p, p0 p0 + w, with
w ∈ p⊥ ⊥
0 , h1 (p0 ) = p0 and h1 (w) ∈ p0 . Hence
ϕh2 B(p̂0 , δ0 ) ⊂ B(p̂0 , δ0 ).
By Proposition
√ 2 2.29 we know that T1 = ϕh1 |B(p̂0 ,δ0 ) has Lispchitz constant κ =
δ + 1−δ
κ 2 0 1−δ2 0 ≈ κ 2 , and assuming δ0 is small enough we have 1−κ1
≤ 1−κ 2 . Notice that
2
0
although the Lispchitz constant in this proposition refers to the Riemannian metric
ρ, since the ratio Lipδ (T1 )/Lipρ (T1 ) approaches 1 as δ0 tends to 0, we can assume
that Lipδ (T1 ) ≤ κ . Thus, by Lemma 2.7 below applied to T1 and T2 = ϕh2 |B(p̂0 ,δ0 ) ,
we have d(T1 , T2 ) ≤ 2 h1 − h2 and
1 4
d(v(h1 ), v(h2 )) ≤ d(T1 , T2 ) ≤ h1 − h2 .
1 − κ 1 − κ2
Free ebooks ==> www.Ebook777.com
2.3 Lipschitz Estimates 59
1
d(x1∗ , x2∗ ) ≤ d(T1 , T2 ),
1−κ
Proof
i
≤ g1 j−1 g2 i−j g1 vj − g2 vj
j=1
Ll,κ := {g ∈ L (V ) : gr l (g) ≥ κ −1 },
www.Ebook777.com
60 2 Estimates on Grassmann Manifolds
and define
l max{1, g1 , g2 }l−1
Cl (g1 , g2 ) := .
max{∧l g1 , ∧l g2 }
16
d(vl (g1 ), vl (g2 )) ≤ Cl (g1 , g2 ) g1 − g2 .
1 − κ2
Proposition 2.41 Given 0 < κ < 21 and integers 1 ≤ k < k + r ≤ dim V , there
exists δ0 > 0 such that for all g ∈ L (V ) and E ∈ Gr k (V ), if
(a) σk (g) < κ and σk+r (g) < κ,
(b) δ(E, vk (g)) < δ0
then
(1) σr (g|E ⊥ ) ≤ 2 κ,
20r
(2) δ vr (g|E ⊥ ), vk+r (g) ∩ E ⊥ ≤ δ(E, vk (g)).
1 − 4 κ2
Proof Consider the compact space
By Corollary 2.2, with Cr (hE , h) = r, we get a bound on δ(vr (hE ), vr (h)). The
second distance is zero because vr (h) = vk+r (g) ∩ vk (g)⊥ . Finally we use item (2)
of Proposition 2.30 to derive a bound on the third distance. Notice that although
the conclusion of Proposition 2.30 is stated in terms of the distance d, the ratio
between the metrics d and δ is very close to 1 when δ0 is small. Finally notice that
vk (g) ⊂ vk+r (g) implies θ∩ (vk+r (g), vk (g)⊥ ) = 1.
Lemma 2.9 Given g1 , g2 ∈ L (V ), p̂ ∈ P(g1 ) ∩ P(g2 ) and any unit vector p ∈ p̂,
1 1
d(ϕg1 (p̂), ϕg2 (p̂)) ≤ max{ , } g1 − g2 .
g1 p g2 p
The final four lemmas of this section apply to invertible linear maps in GL(V ).
They express the continuity of the map g → ϕg with values in the space of Lipschitz
or Hölder continuous maps on the projective space. These facts will be needed in
Chap. 5.
δ(ϕg1 (p̂), ϕg1 (q̂)) δ(ϕg2 (p̂), ϕg2 (q̂))
g1 p ∧ g1 vp (q) g2 p ∧ g2 vp (q)
−
=
−
δ(p̂, q̂) δ(p̂, q̂) g1 pg1 q g2 pg2 q
g1 p ∧ g1 vp (q) − g2 p ∧ g2 vp (q)
≤
g1 pg1 q
1 1
+
−
g2 p ∧ g2 vp (q)
g1 pg1 q g2 pg2 q
≤ g1−1 2 g1 p ∧ (g1 vp (q) − g2 vp (q)) + g1−1 2 (g1 p − g2 p) ∧ g2 vp (q)
+ g−1 2 g−1 2 (g1 p
g1 q − g2 q
+ g2 q
g1 p − g2 p
) g2 2
1 2
−1 2
≤ g1 (g1 + g2 ) g1 − g2
+ g2 2 g1−1 2 g2−1 2 (g1 + g2 ) g1 − g2
= (g1−1 2 + g2 2 g1−1 2 g2−1 2 ) (g1 + g2 ) g1 − g2 .
Proof Given p̂ = q̂ in P(V ) consider unit vectors p ∈ p̂, q ∈ q̂ and set v = vp (q). We
have p = q = v = 1 and
p, v = 0. This last relation implies p ∧ v = 1.
Hence
gp ∧ gv = (∧2 g)(p ∧ v) ≥ (∧2 g)−1 −1 ≥ g−1 −2 .
Analogously
gp ∧ gv = (∧2 g)(p ∧ v) ≤ ∧2 g ≤ g2 .
We also have
g−1 −2 ≤ g p g q ≤ g2 .
www.Ebook777.com
2.3 Lipschitz Estimates 63
where C1 (g1 , g2 ) = α max{g1 g1−1 , g2 g2−1 }2(1−α) C(g1 , g2 ), and C(g1 , g2 )
stands for the constant in Lemma 2.10.
δ(ϕg1 p̂,ϕg1 q̂) δ(ϕg2 p̂,ϕg2 q̂)
Proof Setting Δ1 := δ(p̂,q̂)
and Δ2 := δ(p̂,q̂)
, from Lemmas 2.10 and 2.11
we get
α
Δ − Δα
≤ α max{Δα−1 , Δα−1 }
Δ1 − Δ2
1 2 1 2
≤ α max{g1 g1−1 , g2 g2−1 }2(1−α)
Δ1 − Δ2
≤ α max{g1 g1−1 , g2 g2−1 }2(1−α) C(g1 , g2 ) g1 − g2 .
or, equivalently,
gn−1 . . . g1 g0 g1 . . . gn−2 ε
1.
g1 g0 . . . gn−1 gn−2
The AP was introduced by Goldstein and Schlag [6, Proposition 2.2] as a tech-
nique to obtain Höder continuity of the integrated density of states for quasi-periodic
64 2 Estimates on Grassmann Manifolds
Here we prove a shadowing lemma saying that under some conditions, a loose pseudo-
orbit of a chain of contracting maps is shadowed by a true orbit of the mapping
sequence. In particular, a closed pseudo-orbit is shadowed by a periodic orbit of the
mapping chain.
Given a metric space (X, d), denote the closed ε-ball around x ∈ X by
Lemma 2.14 (shadowing lemma) Consider ε > 0 and 0 < δ < κ < 1 such that
δ/(1 − κ) < ε < 1/2.
Given a family {(Xj , dj )}0≤j≤n of compact metric spaces with diameter 1, a chain
of continuous mappings {gj : Xj0 → Xj+1 }0≤j≤n−1 defined on open sets Xj0 ⊂ Xj , and
a sequence of points xj ∈ Xj , assume that for every 0 ≤ j ≤ n − 1:
2.4 Avalanche Principle 65
Proof The proof’s inductive scheme is better understood with the help of Fig. 2.1
(see also Fig. 2.2), where we set zji := (gj−1 ◦ · · · ◦ gi+1 ◦ gi )(xi ) for i ≤ j ≤ n, with
the convention that this composition is the identity when i = j. Of course we have
to prove that all points zji are well-defined.
The boxed expressions represent upper bounds on the distance between the points
respectively above and below the box. The ith row represents the orbit of xi ∈ Xi by
the chain of mappings {gj }j≥i . All points in the jth column belong to the space Xj .
To explain the last upper bound at the bottom of each column, first notice that
zii = xi . By (a), zii−1 = gi−1 (xi−1 ) is well-defined, and by (c), zii−1 ∈ Xi0 (2 ε) ⊂ Xi0 (ε).
Likewise zi−1i−2
∈ Xi−10
(ε), and zii−2 = gi−1 (gi−2 (xi−2 )) is well-defined. Then by (d)
we have
d(zii−1 , zii−2 ) = d(gi−1 (xi−1 ), gi−1 (gi−2 (xi−2 ))) ≤ δ. (2.25)
X0 X1 X2 X3 ... Xn−1 Xn
g g g g gn−2 gn−1
z00 −→
0
z01 −→
1
z02 −→
2
z03 −→
3
. . . −→ z0n−1 −→ z0n
δ κδ κ δ κ δ
n−3 n−2
g g g3 gn−2 gn−1
z11 −→
1
z12 −→
2
z13 −→ . . . −→ z1n−1 −→ z1n
δ κ n−4δ κ n−3δ
g2 g3 gn−2 gn−1
z22 −→ z23 −→ . . . −→ z2n−1 −→ z2n
κ n−5
δ κ n−4δ
g3 gn−2 gn−1
z33 −→ . . . −→ z3n−1 −→ z3n
.. .. ..
. . .
gn−2 gn−1
n−2
z n−2 −→ zn−2
n−1 −→ zn−2
n
δ
gn−1
zn−1
n−1 −→ zn−1
n
znn
Fig. 2.1 Family of orbits for the chain of mappings {gj : Xj0 → Xj+1 }j
66 2 Estimates on Grassmann Manifolds
All other bounds are obtained applying (b) inductively. More precisely, we prove by
induction in the column index j that
(i) all points zji in the jth column are well-defined and belong to Xj0 (ε),
(ii) distances between consecutive points in the column j are bounded by the expres-
sions in Fig. 2.1, i.e., for all 1 ≤ i ≤ j − 1,
The initial inductive steps, j = 0, 1, 2, follow from (a), (c) and (2.25). Assume
now that the points zji in jth column satisfy (i) and (ii). Then their images zj+1
i
= gj (zji )
are well-defined. By (b) we have for all 1 ≤ i ≤ j − 1,
Together with (2.25) this proves (ii) for the column j + 1. To prove (i) consider any
1 ≤ i ≤ j. By (c) and the triangle inequality,
j j
i
d(zj+1 , ∂Xj+1
0
) ≥ d(zj+1 , ∂Xj+1
0
) − d(zj+1
i
, zj+1 )
j
≥ d(gj (xj ), ∂Xj+1
0
) − d(zj+1 , zj+1 )
l−1 l
l=i+1
j
δ
≥ 2ε − κ j−l δ ≥ 2 ε − ≥ ε.
1−κ
l=i+1
This proves (i) for the column j + 1, and concludes the induction.
Conclusion (1) follows from (b) and the following claim, to be proved by induction
in i.
For every i = 0, 1, . . . , n − 1, g(i) (B(x0 , ε)) ⊂ Xi0 (ε), where g(i) = gi−1 ◦ · · · ◦ g0 .
Consider first the case i = 0. Given x ∈ B(x0 , ε),
d(g0 (x), ∂X10 ) ≥ d(g0 (x0 ), ∂X10 ) − d(g0 (x0 ), g0 (x)) ≥ 2 ε − d(g0 (x0 ), g0 (x))
≥ 2 ε − κ d(x0 , x) ≥ 2 ε − κ ε > ε
By (b), g(i) acts as a κ i contraction on B(x0 , ε) and g(i) (B(x0 , ε)) ⊂ Xi0 (ε). Thus for
every x ∈ B(x0 , ε),
n−1
n−1
δ
(n)
d(gn−1 (xn−1 ), g (x0 )) = d(znn−1 , zn0 ) ≤ d(znl , znl−1 ) ≤ κ n−l−1 δ ≤ .
1−κ
l=1 l=1
Hence
Thus g(n) (x) ∈ B(x0 , ε), which proves that g(n) (B(x0 , ε)) ⊂ B(x0 , ε).
In the statement and proof of the AP we will use the notation introduced in Sect. 2.2.3.
Given a chain of linear mappings {gj : Vj → Vj+1 }0≤j≤n−1 we denote the composition
of the first i maps by g(i) := gi−1 . . . g1 g0 . Throughout this chapter, a b will stand
for a ≤ C b for some absolute constant C.
Theorem 2.1 (Avalanche Principle) There exists a constant c > 0 such that given
0 < ε < 1, 0 < κ ≤ c ε2 and a chain of linear mappings {gj : Vj → Vj+1 }0≤j≤n−1
between Euclidean spaces Vj , if
(a) σ (gi ) ≤ κ, for 0 ≤ i ≤ n − 1, and
(b) α(gi−1 , gi ) ≥ ε, for 1 ≤ i ≤ n − 1,
then
(1) d(v(g(n) ), v(g0 )) κ ε−1 ,
∗
(2) d(v(g(n)∗ ), v(gn−1 )) κ ε−1 ,
# $n−1
(3) σ (g(n) ) κ κ (4+2ε2
ε)
,
n−2
n−1
κ
(4)
logg(n) + loggi − loggi gi−1
n 2 .
i=1 i=1
ε
Remark 2.3 (On the assumptions) Assumption (a) says that the (first) gap ratio of
each gj is large, gr(gj ) ≥ κ −1 . Given (a), by Propositions 2.23 and 2.24, assumption
(b) is equivalent to a condition on the rift, ρ(gj−1 , gj ) ε for all j = 1, . . . , n − 1.
Remark 2.4 (On the conclusions) Conclusions (1) and (2) say that the most expand-
ing direction v(g(n) ) of the product g(n) , and its image ϕg(n) v(g(n) ), are respectively
κ/ε-close to the most expanding direction v(g0 ) of g0 , and to the image ϕgn−1 v(gn−1 )
of the most expanding direction of gn−1 . Conclusion (3) says that the composition
map g(n) has a large gap ratio. Finally, conclusion (4) is equivalent to
Free ebooks ==> www.Ebook777.com
2.4 Avalanche Principle 69
for some universal constant C > 0. These inequalities describe the asymptotic almost
multiplicative behavior of the rifts
C κ/ε2
ρ(g0 , g1 , . . . , gn−1 ) ρ(g0 , g1 ) ρ(g1 , g2 ) . . . ρ(gn−2 , gn−1 ).
Proof The strategy of the proof is to look at the contracting action of linear mappings
gj on the projective space.
For each j = 0, 1, . . . , n consider the compact metric space Xj = P(Vj ) with the
normalized Riemannian distance, d(û, v̂) = π2 ρ(û, v̂). The reader should be warned
of the notational similarity between this projective metric and the one defined in (2.2).
We do not refer to the metric (2.2) in this proof. However, the distance in the statement
of the AP can be understood as any of the four equivalent projective distances δ, d,
ρ or d. For 0 ≤ j < n define
The domain of the projective map ϕgj : P(gj ) ⊂ Xj → Xj+1 clearly contains the
open set Xj0 . Analogously, the domain of ϕgj−1 ∗ : P(gj∗ ) ⊂ Xj → Xj−1 contains Yj0 .
We will apply Lemma 2.14 to chains of projective maps formed by the mappings
ϕgj : Xj0 → Xj+1 and their adjoints ϕgj−1
∗ : Yj0 → Xj−1 .
Take positive numbers ε and κ such that 0 < κ ε2 , let r := 1 − ε2 /4, and
define the following input parameters for the application of Lemma 2.14,
1
εsh := arcsin ε,
π
√
r + 1 − r2 4κ
κsh := κ ,
1 − r2 ε2
κr 2κ
δsh := √ .
1−r 2 ε
A simple calculation shows that there exists 0 < c < 1 such that for any 0 < ε < 1
δsh
and 0 < κ ≤ c ε2 , the pre-conditions 0 < δsh < κsh < 1 and 1−κ sh
< εsh < 1/2 of
the shadowing lemma are satisfied.
Define xj = v(gj ) and xj∗ = v(gj−1
∗
). This lemma is going to be applied to the
following chains of maps and sequences of points
www.Ebook777.com
70 2 Estimates on Grassmann Manifolds
from which we will infer the conclusions (1) and (2). Let us check now that assump-
tions (a)–(d) of Lemma 2.14 hold in both cases (A) and (B).
By definition ∂Xj0 := {v̂ ∈ Xj : α(v̂, xj ) = 0 } = { v̂ ∈ Xj : v̂ ⊥ xj }. Hence,
if v̂ ∈ ∂Xj0 then d(xj , v̂) = 1, which proves that d(xj , ∂Xj0 ) = 1. Analogously,
∂Yj0 = {v̂ ∈ Xj : v̂ ⊥ xj∗ } and d(xj∗ , ∂Yj0 ) = 1. Therefore assumption (a) holds.
By definition of Xj0 (ε),
π
v̂ ∈ Xj0 (ε) ⇔ d(v̂, ∂Xj0 ) ≥ ε ⇔ ρ(v̂, ∂Xj0 ) ≥ ε
# π $2
⇔ δ(v̂, ∂Xj ) = α(v̂, xj ) ≥ sin
0
ε
#π $ 2
⇔ δ(v̂, xj ) ≤ cos ε .
2
#π $
1 ε2
cos εsh = cos arcsin ε ≤ 1 − = r,
2 2 4
we have Xj0 (εsh ) ⊂ B(δ) (xj , r) and Yj0 (εsh ) ⊂ B(δ) (xj∗ , r), and assumption (b) holds
by Proposition 2.29 (3).
By the gap assumption,
Therefore
2 2
d(ϕgj (xj ), ∂Xj+1
0
)= arcsin δ(ϕgj (xj ), ∂Xj+1
0
)= arcsin α(ϕgj (xj ), xj+1 )
π π
2
≥ arcsin ε = 2 εsh .
π
Similarly, by the gap assumption,
∗ ∗ ∗ ∗ ∗
α(ϕgj−1
∗ (x ), x
j j−1 ) = α(v(gj−1 ), v(gj−1 )) = α(gj+1 , gj ) = α(gj , gj+1 ) ≥ ε,
∗ 2
∗ (x ), ∂Y
j−1 ) ≥ arcsin ε = 2 εsh .
0
d(ϕgj−1 j
π
Free ebooks ==> www.Ebook777.com
2.4 Avalanche Principle 71
This proves that (c) of the shadowing lemma holds. Notice that in both cases (A) and
(B), the assumption (c) holds trivially for the middle points, because ϕgn−1 (xn−1 ) =
xn∗ ∈ Yn0 (2 εsh ) and ϕg0∗ (x1∗ ) = x0 ∈ X00 (2 εsh ).
It was proved above that Xj0 (εsh ) ⊂ B(δ) (xj , r) and Yj0 (εsh ) ⊂ B(δ) (xj∗ , r). By (2.5)
we have d(û, v̂) ≤ δ(û, v̂). Thus by Proposition 2.29 (1),
ϕgj (Xj0 (εsh )) ⊂ B(δ) (xj∗ , δsh ) ⊂ B(d) (xj∗ , δsh ) with xj∗ = ϕgj (xj ),
and analogously,
(δ)
∗ (Y (ε )) ⊂ B
ϕgj−1 j
0
sh (xj−1 , δsh ) ⊂ B(d) (xj−1 , δsh ) with xj−1 = ϕgj−1 ∗
∗ (x ).
j
δsh κ
δsh .
(1 − κsh ) (1 − κsh
2n
) ε
s2 (g)
(Dϕg )v(g) = = σ (g).
g
On the other hand, by (1) the distance from v(g(n) ) to v(g0 ) is of order κ ε−1 ε
and
Lip(ϕg0 |B(v(g0 ),κ ε−1 ) ) (Dϕg0 )v(g0 ) = σ (g0 ) ≤ κ.
Therefore n−1
(n) κ (4 + 2 ε)
σ (g ) κ (κsh ) n−1
≤κ ,
ε2
www.Ebook777.com
72 2 Estimates on Grassmann Manifolds
Before proving (4), notice that applying (3) to the chain of linear maps g0 , . . . , gi−1
we get that g(i) := gi−1 . . . g0 has a first gap ratio for all i = 1, . . . , n.
We claim that
(i)
α(g , gi ) − α(gi−1 , gi )
κ ε−1 . (2.27)
δsh
d(v(g(i)∗ ), v(gi−1
∗
)) ≤ κ ε−1 .
(1 − κsh )(1 − κsh
2i
)
(i)
α(g , gi ) − α(gi−1 , gi )
=
α(v(g(i)∗ ), v(gi )) − α(v(g∗ ), v(gi ))
i−1
≤ d(v(g(i)∗ ), v(gi−1
∗
)) κ ε−1 .
For any i, the logarithm of any ratio between the four factors α(g(i) , gi ), β(g(i) , gi ),
α(gi−1 , gi ) and β(gi−1 , gi ) is of order κ ε−2 . In fact, by (2.27)
(i)
log α(g , gi )
1
α(g(i) , gi ) − α(gi−1 , gi )
≤ κ ε−2 .
α(gi−1 , gi ) ε
By hypothesis (a), σ (gi ) ≤ κ. From conclusion (3) we also have σ (g(i) ) < κ,
provided we make the constant c small enough. Hence by Lemma 2.4,
(i)
log β(gi−1 , gi )
κ and
log β(g , gi )
κ .
2 2
α(gi−1 , gi ) ε2 α(g(i) , gi ) ε2
Since κ 2 ε−2 κ ε−2 , the logarithms of the other ratios between the factors above
are all κ ε−2 . Thus, for some universal constant C > 0, each of these ratios is
−2 −2
inside the interval [e−C κ ε , eC κ ε ].
Finally, applying Proposition 2.25 to the rifts ρ(g0 , . . . , gn−1 ), ρ(g0 , g1 ), ρ(g1 , g2 ),
etc., we have
−2
n−1
α(g(i) , gi ) ρ(g0 , . . . , gn−1 )
n−1
β(g(i) , gi ) −2
e−n C κ ε ≤ ≤ n−1 ≤ ≤ en C κ ε ,
i=1
β(gi−1 , gi ) i=1 ρ(g ,
i−1 i g ) i=1
α(gi−1 , gi )
Proposition 2.42 There exists c > 0 such that given 0 < ε < 1, 0 < κ ≤ c ε2 and
g0 , g1 , . . . , gn−1 ∈ Mat(m, R), if
1
(gaps) gr(gi ) > for all 0 ≤ i ≤ n − 1
κ
gi gi−1
(angles) >ε for all 1 ≤ i ≤ n − 1
gi gi−1
then
( )
max d(v(g(n)∗ ), v(gn−1
∗
)), d(v(g(n) ), v(g0 )) κ ε−1
n−2
n−1
κ
(n)
logg + loggi − loggi gi−1
n 2 .
ε
i=1 i=1
Proof Consider the constant c > 0 in Theorem 2.1, let c := c (1 − 2 c2 ) and assume
0 < κ ≤ c ε 2 .
Assumption (gaps) here is equivalent to assumption (a) of Theorem 2.1. By Propo-
sition 2.24, the assumption (angles) here implies
2 κ2
α(gi−1 , gi ) ≥ ρ(gi−1 , gi ) 1 −
ρ(gi−1 , gi )2
2 κ2
≥ε 1− ≥ ε 1 − 2 c2 ε2 =: ε ,
ε 2
Given a chain of linear maps {gj : Vj → Vj+1 }0≤j≤n−1 between Euclidean spaces Vj ,
and integers 0 ≤ i < j ≤ n we define
With this notation the following relation holds for 0 ≤ i < k < j ≤ n,
Next proposition states, in a quantified way, that the most expanding directions
v(gn,i) ) ∈ P(Vi ) are almost invariant under the adjoints of the chain mappings.
74 2 Estimates on Grassmann Manifolds
κ κ (4 + 2 ε) n−i
d(ϕgi∗ v(g(n,i+1) ), v(g(n,i) )) ( ) .
ε ε2
Proof Consider κ, ε, κsh and εsh as in Theorem 2.1. From the proof of item (3) of the
∗
AP, applied to the chain of mappings gn−1 , . . . , gi∗ , we conclude that the composition
(n,i) ∗ ∗ ∗
g = gi ◦ · · · ◦ gn−1 is a (κsh ) -Lipschitz contraction on the ball B(v(gn−1
n−i
), εsh ).
(n,i+1)∗ ∗
On the other hand, by (2) of the AP we have d( v(g , v(gn−1 ) ) κ ε−1 and
∗
d( v(gn−1 ), v(g(n,i)∗ ) κ ε−1 . Since κ ε−1 ε εsh , both projective points
(n,i)∗ ∗
v(g ) and v(g(n,i+1)∗ ) belong to the ball B(v(gn−1 ), εsh ). Thus,
Proposition 2.44 Let c > 0 be the universal constant in Theorem 2.1. Given num-
bers 0 < ε < 1 and 0 < κ < c ε2 , and given two chains of matrices g0 , . . . , gn−1
and g0 , . . . , gn−1
in Mat(m, R), both satisfying the assumptions of the AP for the
given parameters κ and ε, if drel (gi , gi ) < δ for all i = 0, 1, . . . , n − 1, then
κ
(a) d( v(gn−1 . . . g0 ), v(gn−1 . . . g0 ) ) + 8 δ,
ε
gn−1 . . . g0
κ δ
(b) log n + .
gn−1 . . . g0 ε2 ε
Proof Item (a) follows from conclusion (1) of Theorem 2.1, and Proposition 2.40,
d( v(gn−1 . . . g0 ), v(gn−1 . . . g0 ) ) ≤ d( v(gn−1 . . . g0 ), v(g0 ) )
+ d(v(g0 ), v(g0 )) + d( v(g0 ), v(gn−1
. . . g0 ) )
κ 16 δ κ
2 + + 8 δ.
ε 1−κ 2 ε
Free ebooks ==> www.Ebook777.com
2.4 Avalanche Principle 75
which implies
gi 1
≤ .
gi 1−δ
Since the two chains of matrices satisfy the assumptions of the AP we have
log gi gi−1
δ .
gi gi−1 ε
log gn−1 . . . g0
≤
log gn−1 . . . g0 g1 . . . gn−2
gn−1 . . . g0 g1 g0 . . . gn−1 gn−2
g g . . . gn−1
gn−2
+
log 1 0
gn−1 . . . g0 g1 . . . gn−2
n−2
n−1
+
log gi
+
log gi gi−1
i=1
gi i=1
gi gi−1
www.Ebook777.com
76 2 Estimates on Grassmann Manifolds
κ δ
2 n 2 + (n − 2) δ + (n − 1)
ε ε
κ δ
n + ,
ε2 ε
pτj (g)
πτ,j (g) = ,
pτj−1 (g)
and
pτj (g) = πτ,1 (g) . . . πτ,j (g).
Proposition 2.45 (Flag AP) Let c > 0 be the universal constant in Theorem 2.1.
Given numbers 0 < ε < 1, 0 < κ ≤ c ε2 and a chain of matrices gj ∈ Mat(m, R),
with j = 0, 1, . . . , n − 1, if
(a) στ (gi ) ≤ κ, for 0 ≤ i ≤ n − 1, and
(b) ατ (gi−1 , gi ) ≥ ε, for 1 ≤ i ≤ n − 1,
then
∗
(1) d(vτ (g(n)∗ ), vτ (gn−1 )) κ ε−1
(n) −1
(2) d(vτ (g ), #vτ (g0 )) $ κε
n
(3) στ (g(n) ) ≤ κ (4+2
ε2
ε)
n−2
n−1
κ
log π(g(n) ) + log π(gi ) − log π(gi gi−1 )
n 2 .
i=1 i=1
ε
Proof For each j = 1, . . . , k, consider the chain of matrices ∧τj g0 , ∧τj g1 , . . . , ∧τj
gn−1 . Assumptions (a) and (b) here imply the corresponding assumptions of
2.4 Avalanche Principle 77
Theorem 2.1 for all these chains of exterior power matrices. Hence, by (1) of the AP
∗
Thus, taking the maximum in j we get d(vτ (g(n)∗ ), vτ (gn−1 )) κ ε−1 , which proves
(1). Conclusion (2) follows in the same way.
Similarly, from (3) of Theorem 2.1, we infer the corresponding conclusion here
n
(n) (n) (n) κ (4 + 2 ε)
στ (g ) = max στj (g ) = max σ (∧τj g ) ≤ .
1≤j≤k 1≤j≤k ε2
has complex dimension one and determines a point in P(Cm ), denoted by v(g) and
referred to as the g-most expanding direction.
Given points v̂, û ∈ P(Cm ), we set
v, u
α(v̂, û) := where v ∈ v̂, u ∈ û. (2.28)
v u
78 2 Estimates on Grassmann Manifolds
Given g, g ∈ Mat(m, C), both with (first) gap ratios, we define the angle between
g and g to be
α(g, g ) := α(v(g∗ ), v(g )).
With these definitions, the real version of the AP leads in a straightforward manner
to a slightly weaker complex version, stated and proved below. However, adapting
the original proof to the complex case, replacing each real concept by its complex
analog, would lead to the same stronger estimates as in Theorem 2.1.
Proposition 2.46 (Complex AP) Let c > 0 be the universal constant in Theorem 2.1.
Given numbers 0 < ε < 1, 0 < κ ≤ c ε4 and a chain of matrices gj ∈ Mat(m, C),
with j = 0, 1, . . . , n − 1, if
(a) σ (gi ) ≤ κ, for 0 ≤ i ≤ n − 1, and
(b) α(gi−1 , gi ) ≥ ε, for 1 ≤ i ≤ n − 1,
then
∗
(1) d(v(g(n)∗ ), v(gn−1 )) κ ε−2
(n)
(2) d(v(g ), v(g# 0 )) 2 $κ ε−2
n
(3) σ (g(n) ) ≤ κ (4+2
ε4
ε )
n−2
n−1
κ
(4)
logg(n) + loggi − loggi gi−1
n 4 .
i=1 i=1
ε
s3 (gR ) s2 (g)
σ(2) (gR ) = = = σ (g). (2.29)
s1 (gR ) s1 (g)
The g-most expanding direction v(g) ∈ P(Cm ) is a complex line which we can
identify with the real 2-plane v(2) (gR ). This identification, v(g) ≡ v(2) (gR ), comes
from a natural isometric embedding P(Cm ) → Gr 2 (R2m ).
Consider two points v̂, û ∈ P(Cm ) and take unit vectors v ∈ v̂ and u ∈ û. Denote
by U, V ⊂ Cm the complex lines spanned by these vectors, which are planes in
Gr 2 (R2m ). Consider the complex orthogonal projection onto the complex line V ,
πu,v : U → V , defined by πu,v (x) :=
x, v v. By (2.28) we have α(v̂, û) = πu,v .
∗
On the other hand, since the adjoints πu,v : V → U of πu,v both as a complex and as
Free ebooks ==> www.Ebook777.com
2.4 Avalanche Principle 79
a real linear maps coincide, it follows that πu,v = πU,V is the restriction to U of the
(real) orthogonal projection onto the 2-plane V . Thus, by Proposition 2.19(b),
α2 (U, V ) = ∗ π ) = det (π ∗ π ) = π 2 = α(v̂, û)2 .
det R (πu,v u,v C u,v u,v u,v
In particular,
α(2) (gR , (g )R ) = α(2) (v((gR )∗ ), v((g )R )) = α(v(g∗ ), v(g ))2 = α(g, g )2 . (2.30)
Take κ, ε > 0 such that κ < c ε4 , 0 < ε < 1, and consider a chain of matrices
gj ∈ Mat(m, C), j = 0, 1, . . . , n − 1 satisfying the assumptions (a) and (b) of the
complex AP. By (2.29) and (2.30), the assumptions (a) and (b) of Proposition 2.45
hold for the chain of real matrices gjR ∈ Mat(2m, R), j = 0, 1, . . . , n − 1, with
parameters κ and ε2 , and with τ = (2). Therefore conclusions (1)–(4) of the complex
AP follow from the corresponding conclusions of Proposition 2.45. In conclusion
(4) we use the (2)-singular value product π(g) := g2 = ∧2 gR .
References
1. A. Ávila, S. Jitomirskaya, C. Sadel, Complex one-frequency cocycles. J. Eur. Math. Soc. (JEMS)
16(9), 1915–1935 (2014). MR 3273312
2. M. Barnabei, A. Brini, G.-C. Rota, On the exterior calculus of invariant theory. J. Algebra 96(1),
120–160 (1985). MR 808845 (87j:05002)
3. J. Bourgain, Positivity and continuity of the Lyapounov exponent for shifts on Td with arbitrary
frequency vector and real analytic potential. J. Anal. Math. 96, 313–355 (2005). MR 2177191
(2006i:47064)
4. J. Bourgain, S. Jitomirskaya, Continuity of the Lyapunov exponent for quasiperiodic operators
with analytic potential. J. Statist. Phys. 108(5–6), 1203–1218 (2002), Dedicated to David Ruelle
and Yasha Sinai on the occasion of their 65th birthdays. MR 1933451 (2004c:47073)
5. P. Duarte, S. Klein, Continuity of the lyapunov exponents for quasiperiodic cocycles. Comm.
Math. Phys. 332(3), 1113–1166 (2014). MR 3262622
6. M. Goldstein, W. Schlag, Hölder continuity of the integrated density of states for quasi-periodic
Schrödinger equations and averages of shifts of subharmonic functions. Ann. Math. 154(2)(1),
155–203 (2001). MR 1847592 (2002h:82055)
7. W. Schlag, Regularity and convergence rates for the Lyapunov exponents of linear cocycles. J.
Mod. Dyn. 7(4), 619–637 (2013). MR 3177775
8. S. Sternberg, Lectures on Differential Geometry (Prentice-Hall Inc, Englewood Cliffs, 1964).
MR 0193578 (33 #1797)
www.Ebook777.com
Chapter 3
Abstract Continuity of Lyapunov Exponents
X × Rm (x, v) → (T x, A(x)v) ∈ X × Rm ,
dist(B, A) ≥ B − A L ∞ .
is locally Lipschitz.
www.Ebook777.com
3.1 Definitions, the Abstract Setup and Statement 83
logA± L p ≤ C < ∞,
holds for all cocycles A ∈ C , and if we endow Cm with the distance given by
dist∞ (A, B) := A − B L ∞ ,
Definition 3.4 We say that Ξ and A are compatible if for every integer N ∈ N, for
every set F ∈ F N (A) and for every ε > 0, there is an observable ξ ∈ Ξ such that:
1 F ≤ ξ and ξ dμ ≤ μ(F) + ε. (3.6)
X
As mentioned in the introduction, the main tools in our results are some appropriate
large deviations type (LDT) estimates for the given dynamical systems (meaning the
base and the fiber dynamics). An LDT estimate for the base dynamics says that given
an observable ξ : X → R, we have
1 n−1
μ {x ∈ X : ξ(T x) −
j
ξ dμ > ε} < ι(n, ε),
n j=0 X
where ε = o(1) and ι(n, ε) → 0 (as n → ∞) represent, respectively, the size of the
deviation from the mean and the measure of the deviation set. The above inequality
should hold for all integers n ≥ n 0 (ξ, ε).
In classical probabilities, when dealing with i.i.d. random variables, large devia-
tions are precise asymptotic statements, and the measure of the deviation set decays
exponentially. For our purposes, and for the given dynamical systems, we need
slightly different types of estimates (not as precise, but for all iterates of the system
and satisfying some uniformity properties). Moreover, in some of our applications
(e.g. to certain types of quasi-periodic cocycles), the available decay of the measure
of the deviation set is not exponential in the number of iterates, but slower than
exponential. This is the motivation behind the following formalism.
Fix a constant 1 < p ≤ ∞. From now on, ε, ι : (0, ∞) → (0, ∞) will represent
functions that describe respectively, the size of the deviation from the mean and the
measure of the deviation set. We assume that the deviation size functions ε(t) are non-
increasing. We assume that the deviation set measure functions ι(t) are continuous
and strictly decreasing to 0, as t → ∞, at least like a power and at most like an
exponential, in other words we assume that:
1
log t log t as t → ∞.
ι(t)
Free ebooks ==> www.Ebook777.com
3.1 Definitions, the Abstract Setup and Statement 85
p−1
Denote by φι (t) the inverse of the map t → ψι (t) := t ι(t)− 2 p−1 . We then also
assume that the increasing function φι (t) does not grow too fast, or more precisely
that:
φι (2t)
lim < 2.
t→∞ φι (t)
Let E and I be some spaces of functions, with E containing deviation size functions
ε(t) and I containing deviation set measure functions ι(t). We assume that I is a
convex cone, i.e., the functions a · ι(t) and ι1 (t) + ι2 (t) belong to I for any a > 0 and
ι1 , ι2 ∈ I. Let P = P( p) be a set of triplets p = (n 0 , ε, ι), where n 0 ∈ N, ε ∈ E and
ι ∈ I. An element p ∈ P is called an LDT parameter. Our set of LDT parameters
P should satisfy the condition: for all ε > 0 there is p = p(ε) = (n 0 , ε, ι) ∈ P
such that εn 0 ≤ ε, so P contains LDT parameters with arbitrarily small deviation
size functions.
We now define the base and fiber LDT estimates, which are relative to given
spaces of deviation functions E, I and set of parameters P.
Definition 3.5 An observable ξ : X → R satisfies a base-LDT estimate if for every
ε > 0 there is p = p(ξ, ε) ∈ P, p = (n 0 , ε, ι), such that for all n ≥ n 0 we have
εn ≤ ε and
1 n−1
μ {x ∈ X : ξ(T j x) − ξ dμ > εn } < ιn . (3.7)
n j=0 X
1
μ {x ∈ X : logB (n) (x) − L (n)
1 (B) > εn } < ιn . (3.9)
n
www.Ebook777.com
86 3 Abstract Continuity of Lyapunov Exponents
The proof of the abstract continuity theorem (ACT) of the Lyapunov exponents
will be finalized in Sects. 3.5 and 3.6. In Sect. 3.2 we prove that the upper semicon-
tinuity of the maximal Lyapunov exponent holds uniformly in cocycle and phase,
for a large set of phases. While this result is interesting in itself, in our scheme it
ensures that in an inductive procedure based on the AP, the gap condition holds. The
inductive procedure, which is a type of multiscale analysis leading to the proof of
continuity of the Lyapunov exponents, is described in Sect. 3.3 (the base step) and
Sect. 3.4 (the inductive step).
We note that the use of the nearly upper semicontinuity of the maximal Lyapunov
exponent result in Sect. 3.2 represents a major point of difference between our induc-
tive procedure and the one employed by Goldstein and Schlag in [4] or by Schlag
in [7]. It is also what allows us to treat, within this scheme, random models (see
Chap. 5) and (in a future work) identically singular quasi-periodic models.
time argument used by Katznelson and Weiss [6] in their proofs of the Birkhoff’s and
Kingman’s ergodic theorems. However, the results in [3, 5] require unique ergodicity
of the system, a property that Bernoulli and Markov shifts do not satisfy. By replacing
unique ergodicity with a weaker property—namely that a base-LDT holds for a large
enough set of observables, which we show later to hold for Markov shifts—we obtain
a (weaker) version of the uniform upper semicontinuity in [5], one which holds for
a large enough set of phases.
1
logB (n) (x) ≤ L 1 (A) + ε (3.10)
n
holds for all x outside of a set of measure < ιn .
Up to a zero measure set, the exceptional set depends only on A, ε.
(ii) Assume that L 1 (A) = −∞.
For every t < ∞, there are δ = δ(A, t) > 0, n 0 = n 0 (A, t) ∈ N and ι = ι(A, t) ∈ I,
such that if B ∈ Cm with dist(B, A) < δ, and if n ≥ n 0 , then the upper bound
1
logB (n) (x) ≤ −t (3.11)
n
holds for all x outside of a set of measure < ιn .
Up to a zero measure set, the exceptional set depends only on A, t.
Proof Throughout this proof, C will stand for a positive, finite, large enough constant
that depends only on the cocycle A, and which may change slightly from one estimate
to another.
If B ∈ C is at some small distance from A, then it will be close enough to A
in the L ∞ distance as well, hence we will assume that for μ a.e. x ∈ X we have
B(x) < C.
Moreover, in the case (i), when L 1 (A) is finite, since we also assume A to be
1
bounded, we may choose the constant
L C such that for all n ≥ 1 we have
1 (n)
n logA 1 < C and hence also L 1 (A) < C.
L
The proofs for each of the two cases are similar, but the argument will differ in
some parts. We first present the case L 1 (A) > −∞ in detail, then indicate how to
modify the argument for the case L 1 (A) = −∞.
88 3 Abstract Continuity of Lyapunov Exponents
1
lim logA(n) (x) = L 1 (A) for μ a.e. x,
n→∞ n
hence the number
1
n(x) := min{n ≥ 1 : logA(n) (x) < L 1 (A) + ε} (3.12)
n
is defined for μ a.e. x ∈ X .
For every integer N , let
N
1
U N := {x : n(x) ≤ N } = {x : logA(n) (x) < L 1 (A) + ε}.
n=1
n
Next we will bound from above logB (n) (x) by logA(n) (x) + o(1) for all
cocycles B with dist(B, A) < δ where δ will be chosen later, for all 1 ≤ n ≤ N and
for a large set of phases x ∈ X .
Since A is L 1 -bounded, logA(n) ∈ L 1 (X, μ), so A(n) (x) = 0 for μ-a.e. x ∈ X .
Moreover, if B ∈ Cm with dist(B, A) < δ (where δ 1 is chosen below), we
have B(x) − A(x) < δ and B(x) < C for μ-a.e. x ∈ X .
Then for x outside a null set and for 1 ≤ n ≤ N , we have:
B (n) (x)
logB (n) (x) − logA(n) (x) = log
A(n) (x)
B (n) (x) − A(n) (x) B (n) (x) − A(n) (x)
≤ log[ (n)
+ 1] ≤
A (x) A(n) (x)
1 1
≤ nC n−1 δ · (n)
≤ N C N −1 δ · (n)
.
A (x) A (x)
Hence
1
logB (n) (x) ≤ logA(n) (x) + δ N C N −1 (3.14)
A(n) (x)
V ∈ F N (A), and we will show that V has measure at most ε.
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 89
1
logA(n) (x) > log 1/t ,
n n
hence
N
1 log 1/t
V ⊂ {x : logA(n) (x) > }.
n=1
n n
1 log 1/t Cn CN ε
μ {x : logA(n) (x) > }< ≤ = .
n n log 1/t log 1/t N
Therefore,
μ(V ) < N ε/N = ε,
n(x1 ) if x1 ∈ O
x1 = x n1 =
1 if x1 ∈
/O
n(x2 ) if x2 ∈ O
x2 = T n1 x1 n2 =
1 if x2 ∈
/O
...
n(xk+1 ) if xk+1 ∈ O
xk+1 = T xknk
n k+1 =
1 if xk+1 ∈
/ O.
Free ebooks ==> www.Ebook777.com
90 3 Abstract Continuity of Lyapunov Exponents
n 1 + · · · + n p ≤ n < n 1 + · · · + n p + n p+1 ,
p
logB (n) (x) = bn (x) = bn 1 +···+n p +m (x) ≤ bn k (xk ) + bm (x p+1 ). (3.16)
k=1
We estimate each term separately. Each estimate is valid for x outside of a null
set.
For the last term we use the trivial bound:
• Or xk ∈
/ O, so n k = 1, in which case bn k (xk ) = logB(xk ) ≤ C.
Therefore,
www.Ebook777.com
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 91
We conclude:
We add up (3.17) and (3.18) for all 1 ≤ k ≤ p, and then use (3.16) to get:
p
logB (n) (x) ≤ (n 1 + · · · + n p ) (L 1 (A) + 2ε) + 2C 1O (xk ) + C N
k=1
n−1
≤ n (L 1 (A) + 2ε) + 2C 1O (T j x) + C N .
j=0
Divide both sides by n to conclude that for μ-a.e. x ∈ X and for all n ≥ n 0 ,
1
n−1
1 CN
logB (n) (x) ≤ L 1 (A) + 2ε + 2C 1 (T j x) + . (3.19)
n n j=0 O n
By the choice of n we have CnN < ε, so all that is left is to estimate the Birkhoff
average above. We use the compatibility condition. Since O ∈ F N (A), there is an
observable ξ = ξ(A, ε) ∈ Ξ such that 1O ≤ ξ and X ξ dμ < μ(O ) + ε < 3ε.
Then, applying the base-LDT to ξ , there is p = p(ξ, ε) = p(A, ε) ∈ P, p =
(n 0 , ε, ι), such that for n ≥ n 0 we have εn ≤ ε and
1 1
n−1 n−1
1O (T j x) ≤ ξ(T j x) < ξ dμ + εn < 4ε,
n j=0 n j=0 X
1
n(x) := min{n ≥ 1 : logA(n) (x) < −2t}. (3.20)
n
The sets U N are defined as before. Fix N = N (A, t), then U = U(A, t) = U N so
that μ(U ) < 1/t. Furthermore, U ∈ F N (A) and if x ∈ U then 1 ≤ n(x) ≤ N and
−2N t
Let 0 < δ < t Ne C N −1 , and let B ∈ Cm with dist(B, A) < δ, so B(x)− A(x) < δ
for μ-a.e. x ∈ X . Then clearly, for any 1 ≤ m ≤ N and for μ-a.e. x ∈ X we have:
e−2N t e−2mt
B (m) (x) − A(m) (x) < m C m−1 δ ≤ N C N −1 δ < ≤ . (3.22)
t t
For these phases x and number of iterates m, there are two cases.
Case 1. A(m) (x) is extremely small, i.e.
In this case, using (3.22) we get B (m) (x) < e−2t m (1 + 1/t), so
If m = n(x), then using (3.24) in the second case and using directly (3.23) in the
first case, we conclude that for μ-a.e. x ∈ U we have
logB (n(x)) (x) ≤ −2t n(x) + 1/t ≤ (−2t + 1/t) n(x), (3.25)
Remark 3.1 Note that since our cocycles are in L ∞ , Proposition 3.1 above also
implies the upper semicontinuity of the top Lyapunov exponent as a function of
the cocycle. If L 1 (A) = −∞, this in particular implies the continuity of L 1 at A;
moreover, since L 1 (A) ≥ L 2 (A) ≥ · · · ≥ L m (A), this also implies the continuity
at A of each Lyapunov exponent. Therefore, from now on, we may assume that
L 1 (A) > −∞.
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 93
The main application of Proposition 3.1 is the following lemma, which we use
repeatedly throughout the inductive argument. It gives us a lower bound on the gap
between the first two singular values of the iterates of a cocycle, thus ensuring the
gap condition in the Avalanche Principle.
Throughout this chapter, if A ∈ C is such that L 1 (A) > L 2 (A) ≥ −∞, then
κ(A) denotes the gap between the first two LE, i.e. κ(A) := L 1 (A) − L 2 (A) > 0
when L 2 (A) > −∞, while if L 2 (A) = −∞ then κ(A) is a fixed, large enough finite
constant.
We are now under the assumptions of the abstract continuity Theorem 3.1.
Lemma 3.1 Let A ∈ Cm be a cocycle for which L 1 (A) > L 2 (A) and let ε > 0.
There are δ0 = δ0 (A, ε) > 0, n 0 = n 0 (A, ε) ∈ N and ι = ι(A, ε) ∈ I such that for
all B ∈ Cm with dist(B, A) < δ0 and for all n ≥ n 0 , if
(n)
L (B) − L (n) (A) < θ, (3.26)
1 1
1
log gr(B (n) (x)) > κ(A) − 2θ − 3ε. (3.27)
n
Moreover,
L (n) (n)
1 (B) − L 2 (B) > (κ(A) − 2θ − 3ε) (1 − ιn ). (3.28)
Proof Fix ε > 0. If L 2 (A) = −∞, let t = t (A) := −2L 1 (A) + κ(A).
Since L 1 (A) > L 2 (A), the cocycle A satisfies a uniform fiber-LDT with a para-
meter p = p(A, ε) ∈ P and in a neighborhood around A of size δ(A, ε) > 0.
The compatibility condition holds for all cocycles in C , hence also for ∧2 A. Note
that L 1 (∧2 A) = L 1 (A) + L 2 (A), hence L 1 (∧2 A) > −∞ iff L 2 (A) > −∞.
The nearly uniform upper semicontinuity of the top LE (Proposition 3.1) can then
be applied to ∧2 A, and it gives parameters δ > 0, ι ∈ I, n 0 ∈ N that define the range
of validity of (3.10) and (3.11) respectively. These parameters depend on A and ε
when L 2 (A) > −∞ and only on A when L 2 (A) = −∞.
Pick δ = δ(A, ε) > 0, n 0 = n 0 (A, ε) ∈ N, ι = ι(A, ε) ∈ I such that both
the uniform fiber-LDT and Proposition 3.1 apply for all cocycles B ∈ Cm with
94 3 Abstract Continuity of Lyapunov Exponents
dist(B, A) < δ, for all n ≥ n 0 and for all x outside a set of measure < ιn . Fix such
B, n, x.
For any matrix g ∈ Mat(m, R) we have
s1 (g) g2
gr(g) = = ∈ [1, ∞]. (3.29)
s2 (g) ∧2 g
1 1 1
log gr(B (n) (x)) = 2 logB (n) (x) − log∧2 B (n) (x). (3.30)
n n n
The uniform fiber-LDT gives a lower bound on the first term on the right hand
side of (3.30):
1
logB (n) (x) > L (n) (n)
1 (B) − εn > L 1 (B) − ε.
n
Moreover, from assumption (3.26) we have
L (n) (n)
1 (B) > L 1 (A) − θ ≥ L 1 (A) − θ,
hence
1
logB (n) (x) > L 1 (A) − θ − ε. (3.31)
n
Proposition 3.1 applied to ∧2 A will give an upper bound on n1 log∧2 B (n) (x).
If L 2 (A) > −∞, so L 1 (∧2 A) > −∞, from part (i) of Proposition 3.1 we get
1
log∧2 B (n) (x) < L 1 (∧2 A) + ε = L 1 (A) + L 2 (A) + ε. (3.32)
n
Combine (3.30)–(3.32) to conclude that for all chosen B, n, x we have:
1
log gr(B (n) (x)) > κ(A) − 2θ − 3ε,
n
which proves (3.27). Integrating in x we derive (3.28).
Now if L 2 (A) = −∞, so L 1 (∧2 A) = −∞, use part (ii) of Proposition 3.1 to get
1
log∧2 B (n) (x) < −t = 2L 1 (A) − κ(A). (3.33)
n
Combine (3.30)–(3.33) and get (3.27) in this case as well. Then (3.28) follows as
above.
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 95
For the rest of this chapter, we are given an ergodic MPDS (X, F, μ, T ), a space
of measurable cocycles C , a set of observables Ξ and a set of LDT parameters P with
corresponding spaces of deviation functions E and I. We assume the compatibility
condition in Definition 3.4 between Ξ and any cocycle A ∈ C , the base-LDT for any
observable ξ ∈ Ξ , the uniform L p -boundedness condition (put p = 2 to simplify
notations) on any cocycle A ∈ C with L 1 (A) > −∞ and the uniform fiber-LDT for
any cocycle A ∈ C with L 1 (A) > L 2 (A). These LDT estimates hold for parameters
p ∈ P.
We show that the finite scale Lyapunov exponents have a continuous behavior if
the scale is fixed. We are not able to prove actual continuity of these finite scale
quantities, unless we make some restrictions on the space of cocycles. However, this
continuous behavior at finite scale is sufficient for our purposes, as the inductive
procedure described in the next section leads to the actual continuity of the limit
quantities (the LE) as the scale goes to infinity.
Proposition 3.2 (finite scale uniform continuity) Let A ∈ Cm be a cocycle for which
L 1 (A) > L 2 (A). There are δ0 = δ0 (A) > 0, n 01 = n 01 (A), C1 = C1 (A) > 0 and
ι = ι(A) ∈ I such that for any two cocycles B1 , B2 ∈ Cm with dist(Bi , A) ≤ δ0
where i = 1, 2, if n ≥ n 01 and dist(B1 , B2 ) < e−C1 n , then
(n)
L (B1 ) − L (n) (B2 ) < ι1/2 . (3.34)
1 1 n
Proof Let ε0 := κ(A)/10 > 0. Since L 1 (A) > L 2 (A), the uniform fiber-LDT and
Lemma 3.1 hold for A, ε0 . Choose parameters p = p(A) ∈ P, p = (n 0 , ε, ι) and
δ0 = δ0 (A) > 0 such that εn 0 ≤ ε0 and Lemma 3.1 and the fiber-LDT hold for all
cocycles B ∈ Cm with dist(B, A) ≤ δ0 and for all n ≥ n 0 .
Let C0 = C0 (A) > 0 such that for all such cocycles B we have B L ∞ ≤ eC0
and for all n ≥ 1,
(n) 1
L (B) ≤ logB (n)
(x)
2 ≤ C0 .
1
n L
1
logBi(n) (x) > L (n)
1 (Bi ) − εn ≥ −C 0 − ε0 . (3.35)
n
96 3 Abstract Continuity of Lyapunov Exponents
1 (n)
logB (n) (x) − 1 logB (n) (x) = 1 log B1 (x)
1 2 (n)
n n n B2 (x)
1 B1(n) (x) − B2(n) (x)
≤
n min{B1(n) (x), B2(n) (x)}
1 (C0 +ε0 ) n
n−1
(n− j−1) ( j)
≤ e B2 (T j+1 x) B1 (T j x) − B2 (T j x) B1 (x)
n j=0
Integrating in x we conclude:
1
logB (n) (x) − 1 logB (n) (x) μ(d x) < e−n (C1 −2C0 −ε0 ) . (3.37)
1 2
B
n
n n
By Cauchy-Schwarz we have
1
logB (n) (x) − 1 logB (n) (x) μ(d x)
1 2
Bn n n
1 1
≤ logB1(n) (x) 2 · μ(Bn )1/2 + logB2(n) (x) 2 · μ(Bn )1/2 ,
n L n L
hence
1
logB (n) (x) − 1 logB (n) (x) μ(d x) C0 ι1/2 . (3.38)
1 2 n
Bn n n
(n) 1
L (B1 ) − L (n) (B2 ) ≤ logB (n) (x) − 1 logB (n) (x) μ(d x) < ι1/2 ,
1 1 1 2 n
X n n
In this section we derive the main technical result used to prove our continuity
theorem, an inductive tool based on the avalanche principle (2.42), the uniform fiber-
LDT in Definition 3.7 and the nearly uniform upper semicontinuity Proposition 3.1.
All estimates involving two consecutive scales n 0 , n 1 of the inductive procedure
will carry errors of order at most nn01 . We begin with a simple lemma which shows
that we may always assume that n 1 is a multiple of n 0 , otherwise an extra error term
of the same order is accrued.
This implies
r r
L (n 1) (n n 0 )
1 (A) − L 1 (A) ≤ [L (r ) (A) − L 1(n n 0 ) (A)] ≤ 2C ,
n1 1 n1
This implies
q (q) 1
L ((n+1) n0 )
(A) − L (n 1)
1 (A) ≤ [L (A) − L (n 1)
1 (A)] ≤ 2C ,
1
(n + 1) n 0 1 (n + 1)
Then
B (m 2 +m 1 ) (x)
> e−(m 1 +m 2 )(η+2εn ) (3.40)
B (m 2 ) (T m 1 x) B (m 1 ) (x)
Proof Applying (one side inequality in) the fiber-LDT to the cocycle B at scale
m 2 + m 1 , for all x outside a set of measure < ιm 2 +m 1 < ιn , we have:
1
logB (m 2 +m 1 ) (x) > L (m 2 +m 1 )
(B) − εm 2 +m 1 ≥ L (m 2 +m 1 )
(B) − εn ,
m2 + m1 1 1
hence (m 1 +m 2 )
B (m 2 +m 1 ) (x) > e(m 1 +m 2 ) L 1 (B)−(m 2 +m 1 ) εn
. (3.41)
Applying (the other side inequality in) the fiber-LDT to the cocycle B at scales
m 2 , m 1 , for all x outside a set of measure < ιm 2 + ιm 1 < 2ιn , we have:
1
logB (m 1 ) (x) < L (m
1
1)
(B) + εm 1 < L (m
1
1)
(B) + εn ,
m1
1
logB (m 2 ) (T m 1 x) < L 1(m 2 ) (B) + εm 2 < L 1(m 2 ) (B) + εn .
m2
Thus
(m 1 )
B (m 1 ) (x) < em 1 L 1 (B)+m 1 εn
, (3.42)
(m )
B (m 2 ) (T m 1 x) < e m 2 L 1 2 (B)+m 2 εn
. (3.43)
B (m 2 +m 1 ) (x)
B (m 2 ) (T m 1 x) B (m 1 ) (x)
(m 2 +m 1 ) (m 1 ) (m 2 +m 1 ) (m 2 )
> em 1 (L 1 (B)−L 1 (B))+m 2 (L 1 (B)−L 1 (B))−2(m 2 +m 1 ) εn
Remark 3.3 We note that all is needed in the proof of (3.40) is the availability of
the fiber-LDT estimate precisely at scales m 1 , m 2 and m 1 + m 2 and not at all scales
n ≥ n 0 . This is of course irrelevant here, but it may be helpful in other contexts,
when the (full) fiber-LDT estimate is not available a-priori.
Proposition 3.3 (inductive step procedure) Let A ∈ Cm be a measurable cocycle
such that L 1 (A) > L 2 (A). Fix 0 < ε < κ(A)/20.
There are C = C(A) > 0, δ = δ(A, ε) > 0, n 00 = n 00 (A, ε) ∈ N, ι = ι(A, ε) ∈
I such that for any n 0 ≥ n 00 , if the inequalities
(a) L (n 0) (2n 0 )
1 (B) − L 1 (B) < η0 (3.44)
(n 0 )
(b) L (B) − L 0 (A) < θ0
1
(n )
1 (3.45)
hold for a cocycle B ∈ Cm with dist(B, A) < δ, and if the positive numbers η0 , θ0
satisfy
2θ0 + 4η0 < κ(A) − 12ε, (3.46)
we have:
(n 1 )
L (B) + L (n 0 ) (B) − 2L (2n 0 ) (B) < C n 0 . (3.48)
1 1 1
n1
Furthermore,
(a++) L (n 1) (2n 1 )
1 (B) − L 1 (B) < η1 (3.49)
(n 1 )
(b++) L (B) − L 1 (A) < θ1 ,
1
(n )
1 (3.50)
where
n0
θ1 = θ0 + 4η0 + C , (3.51)
n1
n0
η1 = C . (3.52)
n1
Proof Since L 1 (A) > L 2 (A), the cocycle A satisfies a uniform fiber-LDT. Moreover,
Lemma 3.1 also applies.
Pick δ = δ(A, ε) > 0, n 00 = n 00 (A, ε) ∈ N, ι = ι(A, ε) ∈ I such that for any
n ≥ n 00 , both the uniform fiber-LDT and Lemma 3.1 apply for all cocycles B ∈ Cm
with dist(B, A) < δ, for all n ≥ n 00 and for all x outside a set of measure < ιn .
Let n 0 ≥ n 00 and assume (3.44) and (3.45) hold for cocycles B ∈ Cm with
dist(B, A) < δ. Assume moreover that the uniform L 2 -bound in Definition 3.3 (with
p = 2) applies to all such cocycles.
100 3 Abstract Continuity of Lyapunov Exponents
B (2n 0 ) (x)
> e−n 0 (2η0 +4εn0 ) ≥ e−n 0 (2η0 +4ε) =: εap (3.53)
B (n 0 ) (T n 0 x) B (n 0 ) (x)
1
gr(B (n 0 ) (x)) > en 0 (κ(A)−2θ0 −3ε) =: , (3.54)
κap
which will ensure that the gaps condition in the avalanche principle also holds.
Let Bn 0 be the union of the exceptional sets in (3.53) and (3.54). To simplify
notations, replace the deviation set measure function ι by 4 ι, so we may assume
μ(Bn 0 ) < ιn 0 (we will tacitly do this throughout the paper).
−1/3
Let n 1 be an integer such that n 1+ 0 ≤ n 1 ≤ n 0 · ιn 0 . Since ι(t) decreases at least
like t −c (as t → ∞) for some c > 0, and since ι depends on ε and A, n 00 might
need to be chosen larger, depending on ε and A so that if n 0 ≥ n 00 then the integer
−1/3
0 , n 0 · ιn 0 ] is large enough.
interval [n 1+
Moreover, due to Lemma 3.2 we may assume that n 1 = n · n 0 for some n ∈ N.
To see this, note that once (3.48) is proven for scales that are multiples of n 0 , in
particular for the scales n 1 = n n 0 and n 1 = (n + 1) n 0 , then using (3.39) we derive
(3.48) for any scale n 1 such that n n 0 ≤ n 1 ≤ (n + 1) n 0 . Furthermore, (3.49) and
(3.50) will be derived directly from (3.48).
For every 0 ≤ i ≤ n − 1 define
gi = gi (x) := B (n 0 ) (T i n 0 x).
1
gr(gi ) > for all 0 ≤ i ≤ n − 1,
κap
gi gi−1
> εap for all 1 ≤ i ≤ n − 1.
gi gi−1
n−2
n−1
κap
logg (n) + loggi − loggi gi−1 n · 2 .
i=1 i=1
εap
3.4 The Inductive Step Procedure 101
κap
Note that = e−n 0 (κ(A)−4η0 −2θ0 −11ε) < e−ε n 0 .
εap
2
Since the deviation set measure functions ι ∈ I decay at most exponentially fast,
κ
we may assume that e−εt ≤ ι(t) for t ≥ n 00 . Hence we have ε2ap < ιn 0 .
ap
n−2
logB (n 1 ) (x) + logB (n 0 ) (T in 0 x)
i=1
n−1
− log B (2n 0 ) (T (i−1)n 0 x) n ιn 0 . (3.55)
i=1
1 1 1
n−2
logB (n 1 ) (x) + logB (n 0 ) (T in 0 x)
n1 n i=1 n 0
2 1
n−1
− log B (2n 0 ) (T (i−1)n 0 x) ιn 0 .
n i=1 2n 0
Denote
by f (x) the function on the left hand side of the estimate above, so
f (x) ιn for all x ∈
/ B̄n 0 . Clearly
0
n − 2 (n 0 ) 2(n − 1) (2n 0 )
f (x) μ(d x) = L (n 1)
1 (B) + L 1 (B) − L 1 (B).
X n n
The term on the left hand side of the above inequality can be written in the form
(n 1 )
L (B) + L (n 0 ) (B) − 2L (2n 0 ) (B) − 2 [L (n 0 ) (B) − L (2n 0 ) (B)] ,
1 1 1 1 1
n
hence we conclude:
(n 1 )
L (B) + L (n 0 ) (B) − 2L (2n 0 ) (B)
1 1 1
n0 2 n0
<C + [L (n 0)
(B) − L (2n 0)
(B)] C . (3.56)
n1 n 1 1
n1
Clearly the same argument leading to (3.56) will hold for 2n 1 instead of n 1 , which
via the triangle inequality proves (3.49), that is, the conclusion (a++).
We can rewrite (3.56) in the form
(n 1 )
L (B) − L (n 0 ) (B) + 2[L (n 0 ) (B) − L (2n 0 ) (B)] < C n 0 . (3.57)
1 1 1 1
n1
which establishes (3.50), that is, the conclusion (b++) of the proposition.
We are now ready to prove our abstract continuity result for Lyapunov exponents of
linear cocycles.
Theorem 3.2 Let A ∈ Cm be a measurable cocycle for which L 1 (A) > L 2 (A).
Then the map Cm B → L 1 (B) is continuous at A and the map Cm B →
L 1 (B) − L 2 (B) is lower semicontinuous at A.
L (n) (2n)
1 (A) − L 1 (A) < ε. (3.58)
We will apply the inductive step Proposition 3.3 repeatedly. We first choose the
relevant parameters (which will depend on A and ε) so that both the inductive step
Proposition 3.3 and the finite scale continuity Proposition 3.2 apply. The latter will
ensure that the assumptions (3.44)–(3.46) of the inductive step Proposition 3.3 are
satisfied for a large enough scale n 0 = n 0 (A, ε), so we can start running the inductive
argument with that scale.
Let ι ∈ I be the sum of the corresponding deviation measure functions in the
inductive step Proposition 3.3 and the finite scale continuity Proposition 3.2.
Let δ0 be less than the size of the neighborhood of A ∈ Cm from the inductive
step Proposition 3.3 and from the finite scale continuity Proposition 3.2 respectively.
Let C1 be the constant in the finite scale continuity Proposition 3.2 and let C be the
constant in the inductive step Proposition 3.3.
Finally, let the scale n 0 be greater than the thresholds n 00 from the inductive step
Proposition 3.3, n 01 from the finite scale continuity Proposition 3.2 and n 02 from
(3.58) above. Moreover, assume n 0 to be large enough so that e−C1 2n 0 < δ0 , ιn 0 < ε
1/2
−1/3 −0−
and n ιn
0+
for n ≥ n 0 and C n 0 ε.
Let δ := e−C1 2n 0 and let B ∈ Cm with dist(B, A) < δ.
Since δ = e−C1 2n 0 < e−C1 n 0 , we can apply the finite scale continuity Proposi-
tion 3.2 (with B2 = B and B1 = A) at scales 2n 0 and n 0 and get:
(n 0 )
L (B) − L (n 0 ) (A) < ι1/2 =: θ0 < ε, (3.59)
1 1 n0
(2n 0 )
L (2n 0 ) 1/2
(B) − L 1 (A) < ι2n 0 < ι1/2
1 n 0 = θ0 . (3.60)
The inequality (3.59) also implies the assumption (3.26) in Lemma 3.1 for n = n 0 ,
hence by (3.28) we have
L 1(n 0 ) (B) − L (n 0)
2 (B) > (κ(A) − 2θ0 − 3ε) · (1 − ιn 0 ). (3.62)
The inequalities (3.61) and (3.59) imply, respectively, (3.44) and (3.45), that is,
the assumptions (a) and (b) in the inductive step Proposition 3.3.
Moreover, 2θ0 + 4η0 < 2ε + 12ε = 14ε < κ(A) − 12ε, so the condition (3.46)
between parameters is also satisfied.
We can apply the inductive step Proposition 3.3 and conclude that for n 1 n 1+0
we have:
L (n 1) (2n 1 )
1 (B) − L 1 (B) < η1 , (3.63)
(n 1 )
L (B) − L 1 (A) < θ1 ,
(n )
(3.64)
1 1
104 3 Abstract Continuity of Lyapunov Exponents
where
n0
θ1 = θ0 + 4η0 + C , (3.65)
n1
n0
η1 = C . (3.66)
n1
Again, the inequality (3.64) implies the assumption (3.26) in Lemma 3.1 for
n = n 1 , hence by (3.28) we have
L (n 1) (n 1 )
1 (B) − L 2 (B) > (κ(A) − 2θ1 − 3ε) · (1 − ιn 1 ). (3.67)
Furthermore, (3.63) and (3.64) ensure that the assumptions (a) and (b) of the
inductive step Proposition 3.3 hold at scale n 1 .
n0
Moreover, 2θ1 + 4η1 = (2θ0 + 8η0 ) + 6C < (2ε + 24ε) + ε < κ(A) − 12ε,
n1
hence the inductive step Proposition 3.3 applies again, and for n 2 n 1+
1 we have:
L (n 2) (2n 2 )
1 (B) − L 1 (B) < η2 , (3.68)
(n 2 )
L (B) − L (n 2 ) (A) < θ2 , (3.69)
1 1
where
n1
θ2 = θ1 + 4η1 + C , (3.70)
n2
n1
η2 = C . (3.71)
n2
Note that
n0 n1
2θ2 + 4η2 = (2θ0 + 8η0 ) + [10C + 6C ] < (2ε + 24ε) + ε < κ(A) − 12ε.
n1 n2
It is now clear how we continue this procedure. Going from step k to step k + 1,
we choose a scale n k+1 n 1+
k and from Lemma 3.1 we get
L (n k) (n k )
1 (B) − L 2 (B) > (κ(A) − 2θk − 3ε) · (1 − ιn k ), (3.72)
where nk
ηk+1 = C (3.75)
n k+1
3.5 General Continuity Theorem 105
and
nk
θk+1 = θk + 4ηk + C
n k+1
k−1
ni nk
= (θ0 + 4η0 ) + 5C +C
i=0
n i+1 n k+1
∞
ni
< (θ0 + 4η0 ) + 5C
i=0
n i+1
Hence
θk+1 < 23 ε. (3.76)
Moreover
k−1
ni nk
2θk+1 + 4ηk+1 = (2θ0 + 8η0 ) + 10C + 6C
i=0
n i+1 n k+1
∞
ni
< (2θ0 + 8η0 ) + 10C
i=0
n i+1
< (2θ0 + 8η0 ) + 20C n 0−0− < (2ε + 24ε) + 20ε = 46ε,
so 2θk+1 +4ηk+1 < κ(A)−12ε, ensuring that the inductive process runs indefinitely.
Now take the limit as k → ∞ in (3.74), and using (3.76) we have
L 1 (B) − L 1 (A) ≤ 23ε,
which proves the lower semicontinuity at A of the gap between the first two LE.
Note that estimate (3.74) in the proof of Theorem 3.2 says that if the cocycle B
is close enough to A, then
(n)
L (B) − L (n) (A) ε (3.77)
1 1
Indeed, it is enough to first ensure that the base step of the inductive procedure,
i.e. that the estimates
(n 0 )
L (B) − L (n 0 ) (A) < ι1/2 =: θ0 < ε
1 1 n0
(2n 0 )
L (2n 0 ) 1/2
(B) − L 1 (A) < ι2n 0 < ι1/2
1 n 0 = θ0
hold not just for a single scale n 0 , but for a whole (finite) interval of scales N0 =
[n 00 , en 00 ] =: [n − +
0 , n 0 ], where n 00 is greater than the applicability threshold of various
estimates (e.g. uniform fiber-LDT, finite scale continuity etc.).
Let ψ(t) := t 1+ and define inductively the intervals of scales N1 = [ψ(n − 0 ),
ψ(n + 0 )] =: [n −
1 , n +
1 ], N k+1 = [ψ(n −
k ), ψ(n +
k )] =: [n −
k+1 , n +
k+1 ] for all k ≥ 0.
It follows that if n = n 1 ∈ N1 , then n = n 1 ψ(n 0 ) = n 1+ 0 for some n 0 ∈ N0 ,
and so (3.63) and (3.64) hold for all n 1 ∈ N1 .
Continuing inductively, for every k ≥ 1, if n = n k+1 ∈ Nk+1 , then there is
n k ∈ Nk such that n = n k+1 ψ(n k ) = n 1+ k and then (3.73) and (3.74) hold as well.
The intervals N0 and N1 overlap because
n− −
1 = ψ(n 0 ) = n 00 < e
1+ n 00
= n+
0.
Then since ψ is increasing and Nk+1 ψ(Nk ), the intervals Nk and Nk+1 will
overlap for all k ≥ 0.
Therefore, if n ≥ n − 1 then n ∈ Nk+1 for some k ≥ 0 and so (3.77) holds. This
means, moreover, that we may apply Lemma 3.1 at all such scales and conclude that
for all x outside a set of measure < ιn ,
1
log gr(B (n) (x)) > κ(A) − 5ε.
n
We conclude that the following uniform, finite scale statement holds.
Lemma 3.4 Given a cocycle A ∈ Cm with L 1 (A) > L 2 (A) and 0 < ε < κ(A)/100,
there are δ = δ(A, ε) > 0, n 0 = n 0 (A, ε) ∈ N and ι = ι(A, ε) such that for all
n ≥ n 0 and for all B ∈ Cm with dist(B, A) < δ we have:
(n)
L (B) − L (n) (A) < ε, (3.78)
1 1
1
log gr(B (n) (x)) > κ(A) − 5ε, (3.79)
n
for all x outside a set of measure < ιn .
Corollary 3.1 For all m ≥ 1, and for all 1 ≤ k ≤ m, the Lyapunov exponents
L k : Cm → [−∞, ∞) are continuous functions.
Proof Let A ∈ C be a measurable cocycle. If L 1 (A) > L 2 (A), then we can conclude,
from Theorem 3.2 above that L 1 is continuous at A and that L 1 − L 2 is lower
semicontinuous at A.
3.5 General Continuity Theorem 107
is continuous everywhere.
It is then a simple exercise (see Lemma 6.1 and Theorem 6.2 in [2] for its solution)
to see that this is all that is needed to conclude continuity of each individual Lyapunov
exponent, irrespective of any gap pattern.
The following proposition, which is also interesting in itself, will be the main ingredi-
ent in obtaining the modulus of continuity of the top Lyapunov exponent. It gives the
rate of convergence of the finite scale exponents L (n)
1 (B) to the top Lyapunov expo-
nent L 1 (B) and it gives an estimate on the proximity of these finite scale exponents
at different scales.
These estimates are uniform in a neighborhood of a cocycle A ∈ Cm for which
L 1 (A) > L 2 (A), and they depend on a deviation measure function ι = ι(A) ∈ I,
which will be fixed in the beginning. Define the map ψ(t) = ψι (t) := t · [ι(t)]−1/3 ,
and let φ = φι be its inverse. Moreover, for every integer n ∈ N, denote n++ :=
−1/3
ψ(n) = n ιn and n-- := φ(n), so (n++)-- n.
These estimates will be obtained by applying repeatedly the inductive step Propo-
sition 3.3. In order to obtain the sharpest possible estimate, when going from one
scale to the next, we will make the greatest possible jump, which is why we have
−1/3
defined the “next scale” n++ above as n ιn .
φ(n)
L (n)
1 (B) − L 1 (B) < C ≤ ι1/3
n-- (3.80)
n
(n++) n
L (B) + L (n) (2n)
1 1 (B) − 2L 1 (B) < C ≤ ι1/3
n . (3.81)
n++
Proof To prove (3.80) it is enough to show, under similar constraints on B and n,
and for some function ι = ι(A) ∈ I, that
φ(n)
L (n) (2n)
1 (B) − L 1 (B) < C . (3.82)
n
This would imply, for all k ≥ 0,
φ(2k n)
L 1(2 (B) − L 1(2
k k+1
n) n)
(B) < . (3.83)
2k n
Since we assume that for all deviation measure functions ι ∈ I, the corresponding
map φ = φι satisfies
φι (2t)
lim < 2,
t→∞ φι (t)
there is 0 < r < 1 such that for large enough n we have φ(2n) ≤ 2r φ(n).
Hence for all k we have:
φ(2k n) φ(n)
k
≤ rk .
2 n n
We can then sum up for k from 0 to ∞ in (3.83) and derive (3.80).
To prove (3.82) and (3.81) we follow the same procedure, based on the inductive
step Proposition 3.3, used in the proof of the general continuity Theorem 3.2 and of its
extension Lemma 3.4, but with some modifications. We will work again with intervals
of scales instead of individual scales. We fix ε0 := κ(A)/100 so all subsequent
parameters, including the deviation measure function ι, will be fixed and dependent
only upon the cocycle A. Let ι ∈ I, δ0 > 0, C1 > 0 and C > 0 be as in the beginning
of the proof of Theorem 3.2.
Pick n − −
0 ∈ N large enough that for n ≥ n 0 the inductive step Proposition 3.3 and
the finite scale continuity Proposition 3.2 apply, and that L (n) (2n)
1 (A) − L 1 (A) < ε0 .
n−
Assume also n − −C1 2 e 0
< δ0 , ιn − < ε0 , C(n − −0− 1/2
0 to be large enough that e 0) < ε0 ,
0
−1/3
n 0+ ιn for n ≥ n −
0 and since ι decays at most exponentially, we may also assume
− −1/2 n−
that n 0 ιn − < e .
0
0
− +
Now set n + − +
0 := e , N0 := [n 0 , n 0 ] and δ := e
n0 −C1 2n 0
.
The assumptions above ensure that for all cocycles B ∈ Cm with dist(B, A) < δ,
+
and for all n 0 ∈ N0 , since δ = e−C1 2n 0 ≤ e−C1 2n 0 < e−C1 n 0 , the finite scale
continuity Proposition 3.2 applies at scales 2n 0 , n 0 . This implies, as in the proof of
Theorem 3.2, the assumptions in the inductive step Proposition 3.3 for every n 0 ∈ N0 .
3.6 Modulus of Continuity 109
Let n − − + + − +
1 := ψ(n 0 ), n 1 := ψ(n 0 ) and N1 := [n 1 , n 1 ] ψ(N0 ).
We may assume that for every n 1 ∈ N1 there is n 0 ∈ N0 such that n 1 =
−1/3 −1/3
n 0 ιn 0 n 0 ιn 0 (= ψ(n 0 )) (this is because by Lemma 3.2, the estimates involv-
ing scales n 1 ∈ N1 which are not divisible by n 0 will only carry an additional error
of order nn01 ).
We apply the inductive step Proposition 3.3 and obtain:
n0 φ(n 1 )
L (n 1) (2n 1 )
1 (B) − L 1 (B) < C C ,
n1 n1
(n 1 )
L (B) + L (n 0 ) (B) − 2L (2n 0 ) (B) < C n 0 Cι1/3 .
1 1 1 n0
n1
The following theorem shows that locally near any cocycle A ∈ Cm for which
L 1 (A) > L 2 (A), the top Lyapunov exponent has a modulus of continuity given by a
map that depends explicitly on a deviation measure function ι, hence on the strength
of the large deviation type estimates satisfied by the dynamical system.
More generally, if for some 1 ≤ k ≤ m the cocycle A has the Lyapunov spectrum
gap L k (A) > L k+1 (A), then the map Λk := L 1 + · · · + L k satisfies
Λk (B1 ) − Λk (B2 ) ≤ ω(dist(B1 , B2 )). (3.85)
speed of convergence Proposition 3.4 apply with deviation measure function ι for all
cocycles B ∈ Cm with dist(B, A) < δ0 and for all n ≥ n 00 .
Let C1 = C1 (A) > 0 be the constant from Proposition 3.2.
Set δ := min{δ0 , 21 e−C1 4n 00 }.
Let Bi ∈ Cm be measurable cocycles with dist(Bi , A) < δ (i = 1, 2) and put
dist(B1 , B2 ) =: h (< 2δ ≤ e−C1 4n 00 ).
Set n := 2C1 1 log(1/ h) ∈ N.
Then e−C1 4n ≤ h ≤ e−C1 2n , so dist(B1 , B2 ) = h ≤ e−C1 2n and n ≥ n 00 .
All of this preparation shows that we can apply the finite scale uniform continuity
Proposition 3.2 to B1 , B2 at scales n and 2n and get:
(n)
L (B1 ) − L (n) (B2 ) < ι1/2 < ι1/3 , (3.86)
1 1 n n
(2n)
L (B1 ) − L (2n) (B2 ) < ι1/2 < ι1/2 < ι1/3 . (3.87)
1 1 2n n n
Since dist(Bi , A) < δ ≤ δ0 and n ≥ n 00 , we can also apply the uniform speed of
convergence Proposition 3.4 to Bi (i = 1, 2) at scale n and have:
L (n++)
1 (Bi ) − L 1 (Bi ) < ιn1/3 , (3.88)
(n++)
L (Bi ) + L (n) (2n)
1 (Bi ) − 2L 1 (Bi ) < ιn .
1/3
1 (3.89)
The more general assertion of the theorem follows by simply taking exterior
powers. Indeed, the cocycle ∧k A has the property
References
In his 1968 paper [12] in the Transactions of the Moscow Mathematical Society,
V. Oseledets proved his now famous Multiplicative Ergodic Theorem. Assuming the
integrability of the cocycle, this theorem proves the existence of a measurable and
(T , A)-invariant filtration of the fiber
and the existence of a sequence λ1 > λ2 > · · · > λk ≥ −∞, such that for μ-a.e.
phase x ∈ X and for every vector v ∈ Fj (x) \ Fj+1 (x),
1
lim logA(n) (x) v = λj .
n→+∞ n
We also mention the proofs of Gol’dsheid and Margulis in [5] (for a detailed
presentation of this proof see [1]), Mañé (see his monograph [10]), Walters [19] as
well as variants of these proofs by Viana (see his recent monograph [17]) or Bochi
(see the lecture notes [3] on his web page).
Many extensions of this theorem are available, including those of Gol’dsheid and
Margulis in [5], Kaı̆manovich in [7] or Ruelle in [15, 16].
In this chapter we give a new proof of the multiplicative ergodic theorem, which
is based upon the AP. More precisely, we use the estimate in the AP on the dis-
tance between the most expanding direction of a product of matrices and the most
expanding direction of the first term in the product.
We assume the base dynamics to be invertible. However, the existence of the
Oseledets filtration for non-invertible base dynamics can be reduced to the invertible
case by a natural extension construction (see Sect. 1.3 in [13]).
The Oseledets decomposition is usually obtained under the assumption that both
the base dynamics and the fiber action are invertible. Our pooof does not require
invertibility of the fiber action.
We construct the Oseledets filtration as the μ-a.e. limit as n → ∞ of filtrations
corresponding to the singular value decomposition of the iterates A(n) (x) of the cocy-
cle A. The convergence of these (finite scale) filtrations follows from our extension of
the AP concerning estimates on the distance between most expanding singular direc-
tions of products of matrices. The assumptions of the AP are ensured by Kingman’s
ergodic theorem, which provides μ-a.e. convergence to the Lyapunov exponents
of certain quantities related to the iterates A(n) (x) of the cocycle. Our proof of the
Oseledets theorem uses a concept of avalanche times when the assumptions of the
AP are satisfied. It was pointed to us that this notion was related to that of good times,
introduced recently by Gouëzel and Karlsson in [6]. We note that the AP allows us to
derive quantitative estimates that seem to have no conceptual correspondence in [6].
Finally, we mention that our method of proving Oseledets theorem may also be used
to establish (4.1).
If a quantitative version of the convergence in Kingman’s ergodic theorem is
available, that is, if our system satisfies fiber large deviation type (LDT) estimates,
4.1 Introduction and Statements 115
then we establish a rate of convergence of the finite scale filtrations to the Oseledets
filtration.
Moreover, if the LDT is uniform in the cocycle, we derive continuity of the
Oseledets filtration as a function of the cocycle, in an appropriate average sense.
The argument is again inductive and based upon the AP, whose assumptions are
shown to hold off of small sets of phases related to the exceptional sets in the LDT
estimates.
We construct the subspaces of the Oseledets decomposition of the cocycle A as
intersections between components of the orthogonal complements of the filtration
of A and components of the filtration of the adjoint cocycle.
The continuity of the Oseledets decomposition (under the same assumption of
having uniform LDT estimates) is derived using a similar scheme as the one employed
for the continuity of the filtration. However, this needs to be combined with a careful
analysis of the Lipschitz behavior of the intersection of vector subspaces, which we
obtained in Chap. 2.
A precise formulation of the continuity of the Oseledets filtration and decompo-
sition as functions of the cocycle requires some preparation.
We introduce (see the preamble to Sect. 4.3) a general topological space of mea-
surable cocycles. We then allow perturbations of a given cocycle within the whole
space.
We define spaces of measurable filtrations and decompositions and endow them
with appropriate topologies (see Sect. 4.3.2).
In the case of higher dimensional (i.e. Mat(m, R)-valued, with m > 2) cocy-
cles, as we perturb a given cocycle, the dimensions of the corresponding subspaces
of its Oseledets filtration or decomposition may change. We define some natural
projections / restrictions of these filtrations / decompositions, which will allow us to
formulate and to prove stronger continuity results. In Sect. 4.3.1 we establish the con-
tinuity of the most expanding direction, in Sect. 4.3.3 that of the Oseledets filtration,
and finally in Sect. 4.3.4 we obtain the continuity of the Oseledets decomposition.
We note that as with the Lyapunov exponents, our continuity results are quanti-
tative.
To give an idea of these continuity results, we formulate here a simplified, partic-
ular version of our results in Sect. 4.3.
Let (X, μ, T ) be an ergodic dynamical system with T invertible.
Let Cm be a space of measurable cocycles A : X → Mat(m, R), endowed with a
distance (dist) at least as fine as the L ∞ -distance.
We make the following assumptions:
i. The base dynamics satisfies an LDT estimate for a rich enough (relative to Cm )
set of observables.
ii. Every cocycle A ∈ Cm satisfies a uniform (relative to dist) integrability condition.
iii. Every cocycle A ∈ Cm with L1 (A) > L2 (A) satisfies a fiber LDT which is uniform
in a neighborhood of A.
If A ∈ Cm is such that L1 (A) > L2 (A), then its Oseledets decomposition contains
a one dimensional subspace E1 (A)(x) corresponding to the maximal Lyapunov expo-
116 4 The Oseledets Filtration and Decomposition
nent L1 (A). This defines (after identifying one dimensional subspaces with points in
the projective space P(Rm )) a measurable function E1 (A) : X → P(Rm ).
By the continuity of the Lyapunov exponents established in Chap. 3, if A ∈ Cm
is such that L1 (A) > L2 (A), then for any nearby cocycle B we have L1 (B) > L2 (B).
Hence E1 (B) is well defined as well, and we will prove the following.
Theorem 4.1 With the settings and assumptions described above, if A ∈ Cm with
L1 (A) > L2 (A), then locally near A the map
μ {x ∈ X : d(E1 (B1 )(x), E1 (B2 )(x)) > dist(B1 , B2 )α } < ω(dist(B1 , B2 )),
This result (and the more general ones in Sect. 4.3) are applicable to both random
(i.i.d. or Markov) irreducible cocycles and to quasi-periodic cocycles, since LDT
estimates will be established for these models (see Chaps. 5 and 6).
Continuity of the Oseledets decomposition for GL(2, C)-valued random i.i.d.
cocycles was obtained by Bocker-Neto and Viana in [4]. Their result is not quan-
titative but it requires no generic assumptions (such as irreducibility) on the space
of cocycles. Another related result was recently obtained in [2]. A different type of
continuity property, namely stability of the Lyapunov exponents and of the Oseledets
decomposition under random perturbations of a fixed cocycle, was studied in [9, 11].
We formulate the ergodic theorems of Birkhoff and Kingman, then define the LE of a
linear cocycle over a measurable bundle. We obtain a new proof of the multiplicative
ergodic theorem of Oseledets using the AP.
ture τ ⊥ = (m − τk , . . . , m − τ1 ).
A τ -decomposition of Rm is a family E· = {Ej }1≤j≤k+1 of vector subspaces such
that Rm = ⊕k+1 j=1 Ej , and dim Ej = τj − τj−1 for all j = 1, . . . , k + 1. We denote by
Dτ (Rm ) the space of all τ -decompositions of the Euclidean space Rm .
Given two flags F ∈ Fτ (Rm ) and F ∈ Fτ ⊥ (Rm ), of complementary signatures,
the quantity θ (F, F ) measures the transversality between each subspace Fj in F
and the corresponding subspace Fk−j+1 in F . When θ (F, F ) > 0 all these pairs
(Fj , Fk−j+1 ) of subspaces have a transversal intersection and the following family of
subspaces F F = {Fj ∩ Fk−j+2 }1≤j≤k+1 is a τ -decomposition (see Proposition 2.36
in Chap. 2).
Table 4.1 summarizes these notations.
The proofs of Birkhoff’s pointwise ergodic theorem and Kingman’s ergodic theorem
can be found in most monographs covering topics in ergodic theory (see for instance
[17, 18]). It is also worth mentioning in this context the simple proofs by Katznelson
and Weiss [8]. The method in [8] is based on a stopping time argument, an instance
of which will appear in our proof of the MET in Sect. 4.2.3, and was also used in
Chap. 3 to establish a type of uniform upper semicontinuity of the maximal LE.
Theorem 4.2 (Birkhoff’s ergodic theorem) Let (X, μ, T ) be an ergodic dynamical
system, and let ξ ∈ L 1 (X, μ) be an observable. Then
1
n−1
ξ(T j x) → ξ(x)μ(dx) for μ a.e. x ∈ X.
n j=0 X
Lemma 4.1 (Fekete’s Subadditive Lemma) Given a sub-additive sequence {an }n≥0
the following limit exists
an an
lim = inf ∈ [−∞, +∞).
n→∞ n n≥1 n
B = { (x, v) : x ∈ X, v ∈ E(x) }.
We denote by B(x) the fiber over the base point x and note that as a set, it coincides
with the subspace E(x).
4.2 The Ergodic Theorems 119
1
L1 (A) := lim logA(n) (x).
n→∞ n
The number L1 (A) is called the first Lyapunov exponent of A.
Proof The sequence of functions fn (x) = logA(n) (x) satisfies the sub-additivity
property (4.2) and f1+ = log+ A ∈ L 1 (X, μ). The conclusion follows by
Theorem 4.3.
Proposition 4.2 If A is a μ-integrable cocycle then the following limit exists for any
1 ≤ i ≤ m and μ-a.e. x ∈ X,
1
Li (A) := lim log si (A(n) (x)). (4.3)
n→∞ n
The number Li (A) ∈ [−∞, +∞) is called the ith Lyapunov exponent of A.
Moreover, for all 2 ≤ i ≤ m,
log∧i A ≤ i logA,
the integrability condition X log+ A dμ < +∞ for A implies that all cocycles ∧i A
are also μ-integrable. Because ∧m+1 A(x) ≡ 0 we have L1 (∧m+1 A) = −∞. Let k
be the first integer 1 ≤ j ≤ m + 1 such that L1 (∧j A) = −∞. Then L1 (∧k−1 A) >
L1 (∧k A) = −∞. By Proposition 2.6 we have for 1 ≤ i ≤ k,
Notice that ∧i−1 A(n) (x) is eventually non-zero because L1 (∧i−1 A) > −∞. Hence,
taking logarithms and applying Kingman’s theorem, the limit (4.3) exists and the
relation (4.4) holds. Notice that for i = k we get
120 4 The Oseledets Filtration and Decomposition
For k ≤ i ≤ m, since si (A(n) (x)) ≤ sk (A(n) (x)), by comparison we infer that Li (A) =
−∞ as well.
Corollary 4.1 If X log+ A(x) dμ(x) < +∞ then for μ-a.e. x ∈ X, and 1 ≤ i ≤ m,
1
lim log∧i A(n) (x) = L1 (A) + · · · + Li (A).
n→∞ n
Proof Apply Proposition 4.2, using (4.4) inductively.
1
λA (x, v) := lim sup logA(n) (x)v,
n→+∞ n
1
λ−
A (x, v) := lim inf logA(n) (x)v.
n→+∞ n
Note that λA (x, 0) = −∞. These functions satisfy the following properties:
Proof Item (a) follows from the inequality A(n) (x) v ≤ A(n) (x) v. Item (b) is
a straightforward consequence of the definition. Item (c) follows from the inequality
logA(n) (x)(v + v ) ≤ log A(n) (x)v + A(n) (x)v
≤ log 2 max{A(n) (x)v, A(n) (x)v }
= log 2 + max{logA(n) (x)v, logA(n) (x)v }.
4.2 The Ergodic Theorems 121
and the fact that lim supn→+∞ n1 logA(n) (x)v < limn→+∞ n1 logA(n) (x)v implies
that the ratio A(n) (x)v /A(n) (x)v converges geometrically to 0. Finally, item (e)
follows from the identity A(n) (x)v = A(n−1) (Tx)(A(x)v).
Given a real number λ ∈ R, the set
is a linear subspace of E(x), because of items (b) and (c) of the previous proposition.
This family of subspaces determines a finite filtration (flag)
{0} Fλ1 (x) Fλ2 (x) · · · Fλk (x) Fλk+1 (x) = E(x)
which by item (e) is invariant in the sense that A(x) Fλ (x) ⊆ Fλ (Tx), for all x ∈ X. The
multiplicative ergodic theorem (MET) gives a precise description of this filtration
and its relation with the Lyapunov exponents.
Assume that A is μ-integrable and L1 (A) > L2 (A). The following proposition is
about the existence of a measurable function v∞ (A) : X → P(Rm ) with the most
expanding direction of the cocycle A. For each n ∈ N we define a partial function
v(A(n) (x)) if gr(A(n) (x)) > 1
v(n) (A)(x) :=
undefined otherwise.
Definition 4.3 Let (Y , d) be a metric space. We say that a sequence of partial func-
tions fn : Dn ⊆ X → Y is μ almost everywhere Cauchy if given ε > 0 there exists a
set B ∈ A with μ(B) < ε and n0 ∈ N such that for all n ≥ n0 , the function fn (x) is
well-defined on X \ B, i.e., X \ B ⊆ Dn , and the sequence {fn (x)}n≥n0 is Cauchy for
every x ∈ / B.
Proposition 4.4 Let A be a μ-integrable cocycle such that L1 (A) > L2 (A). The
sequence of (partial) functions v(n) (A) from X to P(Rm ) is μ almost everywhere
Cauchy. In particular, it converges μ almost everywhere to a (total) measurable
function v(∞) (A) : X → P(Rm ). Moreover, for μ-a.e. x ∈ X,
1
lim sup log d(v(n) (A)(x), v(∞) (A)(x)) ≤ L2 (A) − L1 (A) < 0.
n→+∞ n
(a) m0 = n0 ,
(b)
k = n for some
m
k ≥ 1, and
(c)
mi − 2 mi−1
< ε mi for all i ≥ 1.
Proof Choose k ≥ 1 such that 2k ≤ n/n0 < 2k+1 , and define θ = 1k log2 (n/n0 ) − 1,
so that 0 ≤ θ < 1k . The sequence mi := n0 2(1+θ)i satisfies (a) and (b). From
mi mi + 1
− 1 ≤ n0 2(1+θ)(i−1) − 1 < mi−1 ≤ n0 2(1+θ)(i−1) < 1+θ (i ≥ 1)
21+θ 2
we obtain, multiplying by 2/mi ,
mi − 2mi−1
2mi−1
1
1 2mi−1
log 2 2
=
1 − ≤ 1− θ
+
θ − ≤ + .
mi mi 2 2 mi k mi
1 − 1
≤ 1 − 1 = 2 − 1 ≤ log 2 .
1/k
2θ 21/k 21/k k
log 2
Item (c) follows choosing r = 2l where l ∈ N is such that l
+ 1
2l+1
< ε.
Lemma 4.3 Given ε > 0 small enough and a measurable set Ω ⊂ X such that
μ(Ω) > 1 − ε/4 there is a measurable subset Ω0 ⊆ Ω with μ(Ω0 ) > 1 − ε and
there are integers n0 ≥ r such that
(a) For each x ∈ Ω0 there is a ε-doubling sequence {mi }i≥0 satisfying m0 = n0 and
T mi x ∈ Ω for all i ≥ 0;
(b) For all x ∈ Ω0 and n ≥ r n0 , there is an ε-doubling sequence {mi }i≥0 satisfying
m0 = n0 , mk = n for some k ≥ 1, and T mi x ∈ Ω for all 0 ≤ i < k.
1 ε
lim #{ 0 ≤ j ≤ m − 1 : T j x ∈
/ Ω } = μ(X \ Ω) < . (4.5)
m→∞ m 4
Given a phase x, if we denote by m(x) the first integer such that the inequality
1 ε
#{ 0 ≤ j ≤ m − 1 : T j x ∈
/ Ω}<
m 4
holds for all m ≥ m(x), then by (4.5), m(x) is defined for μ-a.e. x ∈ X.
For every integer n, let Un := {x ∈ Ω : m(x) ≤ n}. Since Un ⊂ Un+1 and ∪n Un
has full (relative) measure in Ω, there is n0 = n0 (ε) such that μ(Ω \ Un0 ) < ε/2.
4.2 The Ergodic Theorems 123
mi − m
< ε mi /6. Setting m = m0 = n0 and m = mk = n, the sequence {m }i≥0
i 0 k i
satisfies (b), and a simple calculation shows that it is ε-doubling.
The next proposition says that for any given ε > 0 there is a measurable set of
phases Ω0 with μ(Ω0 ) > 1 − ε, such that if x ∈ Ω0 then there exists an ε-doubling
sequence of avalanche times, that is, times where the assumptions of the AP hold.
Proposition 4.5 Let A be a μ-integrable cocycle such that L1 (A) > L2 (A).
Given 0 < κ < L1 (A) − L2 (A) and 0 < ε κ, there exist integers n0 ≥ r ≥ 1
and a measurable set Ω0 ⊂ X with μ(Ω0 ) > 1 − ε such that for any x ∈ Ω0 and
for any n ≥ r n0 there exists an ε-doubling sequence {m0 , . . . , mk } satisfying the
following properties: m0 = n0 , mk = n and for all 0 ≤ i < k,
(1) gr(A(mi ) (x)) ≥ emi (κ−2ε) and gr(A(mi+1 −mi ) (T mi x)) ≥ emi (κ−2ε)(1−ε)/(1+ε) ,
A(mi+1 ) (x)
(2) ≥ e−5mi ε .
A(mi+1 −mi ) (T mi x) A(mi ) (x)
Moreover, for each x ∈ Ω0 there exists an ε-doubling sequence {mi }i≥0 with m0 = n0
such that (1) and (2) hold for all i ≥ 0.
1
lim logA(n) (x) = L1 (A),
n→∞ n
1
lim log∧2 A(n) (x) = L1 (A) + L2 (A) < 2L1 (A) − κ.
n→∞ n
Take 0 < ε κ small. For any n0 ∈ N consider the measurable set Ωn0 (ε) of
phases x ∈ X such that for all n ≥ n20 we have
124 4 The Oseledets Filtration and Decomposition
en (L1 (A)−ε) ≤ A(n) (x) ≤ en (L1 (A)+ε) and ∧2 A(n) (x) ≤ en (2L1 (A)−κ) . (4.7)
We assume that n0 is chosen large enough so that also μ(X \ Ωn0 (ε)) < ε/2.
Setting Ω := Ωn0 (ε), by Lemma 4.3, there exist integers r and n0 ≥ max{n0 , r},
and a measurable subset Ω0 ⊂ Ω such that for all x ∈ Ω0 and n ≥ r n0 , there is an
ε-doubling sequence {mi }0≤i≤k satisfying m0 = n0 , mk = n, and T mi x ∈ Ωn0 (ε) for
all 0 ≤ i < k.
Item (1) follows from the fact that if x ∈ Ωn0 (ε) then by (4.7)
A(n) (x)2 n0
gr(A(n) (x)) = ≥ en (κ−2ε) for all n ≥ .
∧2 A(n) (x) 2
Applying the estimate above with n := mi yields the first inequality in item
(1), while the second follows by putting n := mi+1 − mi . Note that the ε-doubling
condition implies that mi+1 − mi > 1−ε m ≥ 21 n0 .
1+ε i
For item (2) we use again (4.7). Since x, T mi x ∈ Ωn0 (ε),
A(mi+1 ) (x)
≥ e−2 ε mi+1 ≥ e−5 ε mi .
A(mi ) (x) A(mi+1 −mi ) (T mi x)
Proof (of Proposition 4.4) Given 0 < ε κ < L1 (A) − L2 (A), consider the
integers n0 ≥ r ≥ 1 and the measurable set Ω0 ⊆ X provided by Proposition 4.5.
Given x ∈ Ω0 and n ≥ n0 r by this proposition there is a ε-doubling sequence
{m0 , m1 , . . . , mk } such that m0 = n0 , mk = n and both gap and angle conditions (1)
and (2) hold for all 0 ≤ i < k. We apply the avalanche principle to the sequence of
two matrices g0 = A(mi ) (x) and g1 = A(mi+1 −mi ) (T mi x), with g1 g0 = A(mi+1 ) (x).
The key parameters in this application of the AP are
k−1
k−1
e−mi θ ≤ e−(3/2) n0 θ
e−n0 θ .
i
≤
i=0 i=0
Taking n0 large enough, the bound e−n0 θ becomes arbitrarily small. This proves
that the sequence {v(A(n) (x))}n≥n0 is Cauchy. Moreover, passing to the limit as n →
+∞,
d(v(A(n0 ) (x)), v(∞) (A)(x)) e−n0 θ .
Therefore, as n = n0 is arbitrary,
1
lim sup log d(v(n) (A)(x)), v(∞) (A)(x)) ≤ −θ.
n→+∞ n
Finally, since ε > 0 can be taken arbitrarily small, and κ can be taken arbitrarily
close to L1 (A) − L2 (A), we conclude that
1
lim sup log d(v(n) (A)(x), v(∞) (A)(x)) ≤ L2 (A) − L1 (A).
n→+∞ n
Next proposition asserts the convergence of this sequence of partial functions. The
limit function vk(∞) (A) is called the most expanding k-plane of the cocycle A.
Proposition 4.6 If Lk (A) > Lk+1 (A) then the sequence of partial functions vk(n) (A)
from X to Gr k (Rm ) is almost everywhere Cauchy. In particular, it converges μ almost
everywhere to a (total) measurable function vk(∞) (A) : X → Gr k (Rm ). Moreover, for
μ-a.e. x ∈ X,
1
lim sup log d(vk(n) (A)(x), vk(∞) (A)(x)) ≤ Lk+1 (A) − Lk (A) < 0.
n→+∞ n
126 4 The Oseledets Filtration and Decomposition
1
lim log∧j A(n) (x) = L1 (∧j A)
n→+∞ n
1
lim sup log d vj(n) (A)(x), vj(∞) (A)(x) ≤ Lj+1 (A) − Lj (A).
n→+∞ n
Proposition 4.7 The set of μ-regular points of a cocycle has full μ-measure.
Proof By Proposition 2.11, gr 1 (∧j A(n) (x)) = gr j (A(n) (x)). By Definition 2.15 we
have Ψ vj (A(n) (x)) = v(∧j A(n) (x)), where Ψ stands for the Plücker embedding. To
finish apply Corollary 4.1 and Proposition 4.6.
We recall Definition 2.11 from Sect. 2.2.
Given a linear map A : V → V between Euclidean spaces V and V of dimension
m, we call singular basis of A any orthonormal basis {vj }1≤j≤m of V consisting of
singular vectors vj of A such that A vj = sj (A) for all j = 1, . . . , m. Notice that for
every 1 ≤ k ≤ m, the unit k-vector v1 ∧ · · · ∧ vk is a most expanding vector of ∧k A.
Proposition 4.8 Consider a μ-integrable cocycle A, and a μ-regular point x ∈ X.
If γ = L1 (A) = · · · = Lk (A) > Lk+1 (A) then for any v ∈ vk(∞) (A)(x) \ {0}, we
have
1
lim logA(n) (x) v = γ .
n→+∞ n
In particular, λA (x, v) = λ−
A (x, v) = γ .
Proof Consider a singular basis {v1,n , . . . , vm,n } for the linear map A(n) (x). Let
{v1 , . . . , vk } ⊂ vk(∞) (A)(x) be an orthonormal family obtained as limit of the sequence
{v1,ns , . . . , vk,ns }, for some subsequence of integers ns .
Let wn = v1,n ∧ . . . ∧ vk,n and w = v1 ∧ . . . ∧ vk .
Because x is μ-regular, wˆn → ŵ. Thus possibly changing the sign of v1,n , we may
assume that wn → w as n → +∞. Therefore,
∧k A(n) (x) w
∧k A(n) (x) w − ∧k A(n) (x) wn
−1 =
∧k A(n) (x) wn ∧k A(n) (x)
∧k A (x) w − ∧k A(n) (x) wn
(n)
≤ ≤ w − wn → 0.
∧k A(n) (x)
k
∧k A(n) (x) w ≤ A(n) (x) vj ,
j=1
4.2 The Ergodic Theorems 127
we have
1
k γ = L1 (A) + · · · + Lk (A) = lim log∧k A(n) (x)
n n→∞
1 1
= lim log∧k A(n) (x) wn = lim log∧k A(n) (x) w
n→∞ n n→∞ n
1 1
k k
≤ lim inf logA(n) (x) vj ≤ lim sup logA(n) (x) vj
n→∞ n n→∞ n
j=1 j=1
1
k
≤ lim logA(n) (x) = k γ .
n→∞ n
j=1
1 1
Consider now ci,n := logA(n) (x) − logA(n) (x) vi which satisfies
n n
k
k 1
k
(n)
0 ≤ ci,n ≤ cj,n = logA (x) − logA(n) (x) vj .
j=1
n n j=1
1
λA (x, vj ) = lim logA(n) (x) vj = γ .
n→+∞ n
Now, given v ∈ vk(∞) (A)(x) \ {0}, assume, by contradiction, that there exists a
sequence ns → +∞ such that
1
lim logA(ns ) (x) v < γ . (4.8)
n→+∞ ns
k
(ns )
A (x) v =
2
v, vj,ns 2 A(ns ) (x) vj,ns 2
j=1
k
= v, vj,ns 2 sj (A(ns ) (x))2 ≥ v, vj,ns 2 sj (A(ns ) (x))2 .
j=1
1 1
lim logA(ns ) (x)v ≥ lim log sj (A(ns ) (x)) = Lj (A) = γ ,
s→+∞ ns s→+∞ ns
128 4 The Oseledets Filtration and Decomposition
1
λA (x, v) = lim logA(n) (x) v = γ ,
n→+∞ n
Proof Take 0 < ε κ := L1 (A) − L2 (A), and consider the measurable set Ω0 and
the order n0 ∈ N provided by Proposition 4.5. For x ∈ Ω0 let {mi }i be an ε-doubling
sequence of avalanche times. Then for all i ≥ 0,
A(mi+1 ) (x)
α A(mi ) (x), A(mi+1 −mi ) (T mi x) ≥ e−5 mi ε .
A(mi ) (x) A(mi+1 −mi ) (T mi x)
d(v(mi ) (A∗ )(x), v(∞) (A∗ )(x)) and d(v(mi+1 −mi ) (A)(x), v(∞) (A)(x))
Since Ωi has measure μ(Ωi ) > 1 − ε with arbitrary ε, this proves the lemma.
Let us abbreviate v̂(x) := v(∞) (A)(x), v̂n (x) := v(n) (A)(x), v̂∗ (x) := v(∞) (A∗ )(x)
and v̂n∗ (x) := v(n) (A∗ )(x). Moreover, let us respectively denote by v(x), vn (x), v∗ (x)
and vn∗ (x) unit measurable sections of v̂(x), v̂n (x), v̂∗ and v̂n∗ (x).
Lemma 4.5 Assume L1 (A) > L2 (A) and let v : X → Rm be a unit measurable
section of v(∞) (A). Then A(x)∗ v(Tx) = 0 for μ-almost every x ∈ X.
Proof By Lemma 4.4, α(v̂(Tx), v̂∗ (Tx)) > 0 for μ-a.e. x ∈ X. By Proposition 4.4
applied to the adjoint cocycle A∗ , for μ-a.e. x ∈ X and all large enough n ≥ 1,
Finally, since
From now on, given a matrix A(x) and a projective or Grassmannian point v̂ we
will abbreviate ϕA(x) v̂ and ϕA(x)−1 v̂ writing respectively A(x) v̂ and A(x)−1 v̂. With this
notation, from (2.16) we obtain
A(n) (x)v(A(n) (x)) = v(A(n) (x)∗ ) and A(n) (x)∗ v(A(n) (x)∗ ) = v(A(n) (x)). (4.9)
The following proposition establishes the invariance of the most expanding sub-
bundles vk(∞) (A).
130 4 The Oseledets Filtration and Decomposition
Proof By Proposition 2.15, (b) reduces to (a). Working with exterior powers we can
reduce (a) to the case k = 1, i.e., (a) reduces to the identity A(x)∗ v̂(Tx) = v̂(x).
By Proposition 4.4 and (4.9),
v̂(x) ≈ v̂n (x) = v(A(n) (x)) = A(n) (x)∗ v(A(n) (x)∗ ) = A(n) (x)∗ v̂n∗ (T n x)
and analogously
Hence
converges to 0 as n → +∞.
On the last occurrence of ≈ we apply Proposition 4.4.
1
n−1
1 1
f (T n x) = f (x) − (f − f ◦ T )(T j x),
n n n j=0
1 1
n−1 n−1
f (T x) ≥
j
fp (T j x) ≥ L,
n j=0 n j=0
Proof Because the cocycles A and A∗ play symmetric roles, it is enough proving the
μ-integrability of the function logA v∗ .
Applying Proposition 4.10 to A∗ , we see that A(x) v∗ (x) = ±v∗ (Tx). From this
invariance relation, for μ-a.e. x ∈ X,
n−1
logA(n) (x) v∗ (x) = logA(T j x) v∗ (T j x).
j=0
132 4 The Oseledets Filtration and Decomposition
and hence
1 1 1 1
logA(n) (x) + log α v̂∗ (x), v̂n (x) ≤ logA(n) (x)v∗ (x) ≤ logA(n) (x).
n n n n
By Proposition 4.1, n1 logA(n) (x) converges to L1 (A) almost surely. By Lemma 4.4,
α v̂∗ (x), v̂(x) > 0, and hence n1 log α(v̂∗ (x), v̂n (x)) converges to zero.
Thus, for μ-almost every x ∈ X,
1
n−1
1
lim logA(T j x) v∗ (T j x) = lim logA(n) (x)v∗ (x) = L1 (A).
n→∞ n n→∞ n
j=0
The function logA(x)v∗ (x) is bounded from above by the μ-integrable func-
tion log+ A(x). Hence, h(x) := log+ A(x) − logA(x)v∗ (x) is a non-negative
measurable
function whose Birkhoff averages converge μ-almost everywhere to
log+ A dμ − L1 (A). By Lemma 4.7 it follows that h ∈ L 1 (X, μ), which implies
that logA v∗ ∈ L 1 (X, μ).
Thus, by Birkhoff’s theorem, X logA(x) v∗ dμ = L1 (A).
Proposition 4.12 Assume L1 (A) > L2 (A). Then for μ-a.e. x ∈ X,
1
(a) lim log α v(∞) (A∗ )(T n x), v(∞) (A)(T n x) = 0.
n→+∞ n
1
(b) lim sup log α A(n) (x)v(∞) (A)(x), v(∞) (A)(T n x) = 0.
n→+∞ n
Proof Take unit measurable sections v, v∗ : X → P(Rm ) of v(∞) (A) and v(∞) (A∗ ),
respectively, and as before let us write v̂(x) := v(∞) (A)(x) and v̂∗ (x) := v(∞) (A∗ )(x).
Consider the function f (x) := log α(v̂∗ (x), v̂(x)). By Lemma 4.6, for (a) it is
enough to prove that f − f ◦ T ∈ L 1 (μ).
By Proposition 4.10 we have
By Proposition 4.11, logA v∗ ∈ L 1 (X, μ), and logA∗ (v ◦ T ) ∈ L 1 (X, μ). Hence
by Lemma 4.6 this implies (a).
4.2 The Ergodic Theorems 133
As before, we use the notation v̂n and v̂n∗ for the sub-bundles v(n) (A) and v(n) (A∗ ),
respectively. Since A(n) (x) v̂n (x) = v̂n∗ (T n x), by Proposition 2.35 we have
α(A(n) (x) v̂(x), v̂(T n x)) ≥ α(v̂∗ (T n x), v̂(T n x)) − d(v̂∗ (T n x), v̂n∗ (T n x))
− d(A(n) (x) v̂n (x), A(n) (x) v̂(x)).
Now take 0 < κ < L1 (A) − L2 (A) and 0 < ε κ arbitrary small. By item
(a), for all large enough n we have α(v̂∗ (T n x), v̂(T n x)) ≥ e−nε . Because as n grows,
A(n) (x) has a large gap ratio, it acts as a strong contraction in a neighborhood of v̂n∗ (x).
Hence by Proposition 4.4 for all n large enough
d(A(n) (x) v̂n (x), A(n) (x) v̂(x)) d(v̂n (x), v̂(x)) ≤ e−n (κ−ε) .
We can not guarantee that the second distance d(v̂∗ (T n x), v̂n∗ (T n x)) converges to 0
μ-almost everywhere, but since v̂n∗ converges almost surely to v̂∗ , with the speed
provided by Proposition 4.4, for μ-a.e. x ∈ X there is a sequence of times {ni }i such
that
d(v̂∗ (T ni x), v̂n∗i (T ni x)) ≤ e−ni (κ−ε) ∀ i.
λ− −
∧k+r A (x, vk ∧ vr ) = λ∧k A (x, vk ) + λ∧r A (x, vr ).
Moreover, if λ− −
∧r A (x, vr ) = λ∧r A (x, vr ) then λ∧k+r A (x, vk ∧ vr ) = λ∧k+r A (x, vk ∧ vr ).
Proof Because ∧k [vk(∞) (A)] has dimension one, vk is a limit point of the sequence
of most expanding vectors for ∧k A(n) (x). Hence, by Proposition 4.8 we have
λ−
∧k A (x, vk ) = λ∧k A (x, vk ).
134 4 The Oseledets Filtration and Decomposition
λ∧k A (x, vk ) + λ− −
∧r A (x, vr ) ≤ λ∧k+r A (x, vk ∧ vr ).
By Proposition 4.10 we have ∧r A(n) (x) vr ∈ ∧r [vk(∞) (A)(T n x)⊥ ]. Hence by Propo-
sition 2.34 (b)
∧k A(n) (x)vk ∧r A(n) (x) vr ≤ αn (x)−1 ∧k+r A(n) (x) (vk ∧ vr ),
where
αn (x) := αk A(n) (x)vk(∞) (A)(x), vk(∞) (A)(T n x) .
Lemma 4.8 Let {u1 , . . . , um } be a limit singular basis of E(x) at some μ-regular
point x ∈ X. Then for all i = 1, . . . , m,
λ−
∧i A (x, u1 ∧ . . . ∧ ui ) = λ∧i A (x, u1 ∧ . . . ∧ ui ) = L1 (∧i A).
4.2 The Ergodic Theorems 135
Proof Let {u1,n , . . . , um,n } be a singular basis of A(n) (x), and {u1 , . . . , um } a corre-
sponding limit singular basis for the cocycle A. Choose k such that
Proposition 4.15 Consider a cocycle A such that Lk (A) > Lk+1 (A). Then
Li A|v⊥k = Li+k (A) for any 1 ≤ i ≤ m − k,
where A|v⊥k stands for the restriction of A to the invariant bundle vk(∞) (A)⊥ .
Therefore, the conclusion follows subtracting these identities for consecutive indexes
i and i − 1.
Let us prove (4.10). This identity is reduces to a pair of inequalities. We will use
Propositions 4.13 and 4.14 to establish each of these inequalities.
Fix a μ-regular point x ∈ X, and consider a limit singular basis {u1 , . . . , um } of
the fiber E(x). By Lemma 4.8,
On the last step we use that uk+1 ∧ · · · ∧ uk+i is a non zero vector in the fiber of the
bundle ∧i [vk(∞) (A)⊥ ].
For the converse inequality, choose an orthonormal basis {u1 , . . . , uk } of vk(∞)
(A)(x) and extend it with a limit singular basis {uk+1 , . . . , um } for the cocycle A|v⊥k .
By Lemma 4.8 applied to the cocycle A|v⊥k we get
λ−
∧i A (x, uk+1 ∧ · · · ∧ um ) = λ∧i A (x, uk+1 ∧ · · · ∧ um ) = L1 (∧i A|v⊥
k
).
136 4 The Oseledets Filtration and Decomposition
In particular, for every non-zero vector v in the fiber over x of this sub-bundle,
1
λ−
A (x, v) = λA (x, v) = lim logA(n) (x) v = Lk+1 (A).
n→+∞ n
Proof The stated relation is a simple application of Proposition 2.41 to the matrices
g = A(n) (x). Notice that for a generic point x ∈ X these matrices have exponentially
large gap ratios gr k (A(n) (x)) and gr k+r (A(n) (x)). By this proposition
(n)
δ vr (A(n) (x)|v⊥k ), vk+r (A)(x) ∩ vk(∞) (A)(x)⊥ δ(vk(n) (A)(x), vk(∞) (A)(x) )
converges to zero. Hence the proposition follows by taking the limit as n tends to
+∞. The last statement is a consequence of Proposition 4.8.
Given a signature τ = (τ1 , . . . , τk ), we define a sequence of partial functions
vτ(n) (A) on X taking values on Fτ (Rm ),
vτ (A(n) (x)) if gr τ (A(n) (x)) > 1
vτ(n) (A)(x) :=
undefined otherwise
whose converge is established below. The almost sure limit of this sequence of
functions, denoted by vτ(∞) (A), will be called the most expanding τ -flag of A.
We say that the Lyapunov spectrum of a cocycle A has a τ -gap pattern when
If moreover
L (A) = L+1 (A), for all ∈
/ {τ1 , . . . , τk },
we will say that the Lyapunov spectrum of A has exact gap pattern τ . In this case
we write λj (A) := Lτj (A), for j = 1, . . . , k + 1. These numbers span the complete
Lyapunov spectrum of A without repetitions,
λ1 (A) > λ2 (A) > · · · > λk (A) > λk+1 (A) ≥ −∞.
Proposition 4.17 If the Lyapunov spectrum of A has a τ -gap pattern, then the
sequence of partial functions vτ(n) (A) from X to Fτ (Rm ) is almost everywhere Cauchy.
In particular, it converges μ almost everywhere to a (total) measurable function
vτ(∞) (A) : X → Fτ (Rm ). Moreover, for μ-a.e. x ∈ X,
1
lim sup log d(vτ(n) (A)(x), vτ(∞) (A)(x)) ≤ −gapτ (A) < 0.
n→+∞ n
We are now able to state and to prove the Oseledets Multiplicative Ergodic The-
orem, which has two versions, one on the existence of the Oseledets filtration, and
the other on the existence of the Oseledets decomposition.
Proof Assume the cocycle A has a Lyapunov spectrum with exact gap pattern τ =
(τ1 , . . . , τk−1 ), where 0 = τ0 < τ1 < . . . < τk−1 < τk = dim E, and E = E(x)
denotes the fiber of B. Set by convention vτ(∞) 0
(A) = {0} and vτ(∞)
k
(A) = E(x).
(∞) ⊥
Define Fj (x) := vτj−1 (A)(x) for j = 1, . . . , k + 1, so that dim Fj (x) = dim E −
τj−1 . This implies (b).
The invariance (a) follows from Proposition 4.10.
(∞)
To shorten notation let us respectively write vk and v⊥ k instead of vk (A)(x) and
vk(∞) (A)(x)⊥ . Given v ∈ Fj \Fj+1 = v⊥ ⊥
τj−1 \vτj , consider the orthogonal decomposition
138 4 The Oseledets Filtration and Decomposition
v = u + v , with u ∈ v⊥ ⊥
τj−1 ∩ vτj , u = 0, and v ∈ vτj . By Proposition 4.16 the non-zero
vector u is in the fiber of
vτ(∞)
j −τj−1
A| ⊥
vτ = vτ(∞)
j−1
(A)⊥ ∩ vτ(∞)
j
(A),
j−1
and
λ−
A (x, u) = λA (x, u) = Lτj−1 +1 (A) = Lτj (A) = λj (A).
λ−
A (x, v) = λA (x, v) = λA (x, u + v ) = λA (x, u) = λj (A).
Recall that Rg and Kg denote respectively the range and the kernel of any given a
linear map g : V → V .
Proof We make use of three functorial properties of exterior powers which can be
easily checked:
(1) ∧k Rg = R∧k g , (2) ∧k (g|E ) = ∧k g|∧k E and (3) (∧k g)−1 = ∧k (g−1 ).
Thus ∧k (Kg⊥ ) = ∧k (Rg∗ ) = R∧k g∗ = K∧⊥k g , and
When the cocycle takes invertible values, i.e., A : X → GL(m, R), the backward
iterates A(−n) (x) correspond to forward iterates by the inverse cocycle (T −1 , A−1 ).
Proof Assume that A has a Lyapunov spectrum with exact gap pattern τ =
(τ1 , . . . , τk ), where 0 = τ0 < τ1 < . . . < τk < τk+1 = dim E, and E = E(x)
denotes the fiber of B. Set by convention vτ(∞)0
(A) = {0} and vτ(∞)
k+1
(A) = E(x).
(∞) ∗ (∞) ⊥
Define Ej (x) := vτj (A )(x) ∩ vτj−1 (A)(x) for j = 1, . . . , k + 1.
By Proposition 4.10, both sub-bundles vτ(∞)j
(A∗ ) and vτ(∞)
j−1
(A)⊥ are A-invariant,
and hence the same is true about the intersection. This proves (b).
For (a) consider the flag valued measurable functions
Thus, in view of Proposition 2.36 it is now enough to see that θ (v∗τ (x), v⊥
τ (x)) > 0
for μ-a.e. x ∈ X. But by Definition 2.34 and Lemma 2.6,
The final positivity follows from Lemma 4.4. This proves (a), or in other words that
{Ej (x)}1≤j≤k+1 is a direct sum decomposition of E(x) with dim Ej (x) = τj − τj−1 , for
all j = 1, . . . , k + 1.
140 4 The Oseledets Filtration and Decomposition
n−1
logA(n) (x) v∗ (x) = logA(T j x) v∗ (T j x).
j=0
By Proposition 4.11, the function logA v∗ is μ-integrable with logA v∗ dμ =
L1 (A). Therefore, by Birkhoff’s ergodic theorem we have
1
lim logA(n) (x) v∗ (x) = L1 (A).
n→+∞ n
n−1
(n) −n ∗ −n
logA (T x) v (T x) = logA(T −j x) v∗ (T −j x)
j=0
we get
A(n) (T −n x)+ v∗ (x) = A(n) (T −n x) v∗ (T −n x)−1 v∗ (T −n x),
so that
logA(−n) (x) v∗ (x) = − logA(n) (T −n x) v∗ (T −n x).
Thus
1
lim logA(−n) (x) v∗ (x) = −L1 (A).
n→∞ n
Next consider the case j = 1 and r = τ1 > 1. In this case τ0 = 0 and the
intersection sub-bundle Ej is the r-dimensional A-invariant bundle vr(∞) (A∗ ).
Given a unit vector v1 in vr(∞) (A∗ ), include it in some orthonormal basis {v1 , . . . , vr }
of vr(∞) (A∗ ) and take the unit r-vector w = v1 ∧ · · · ∧ vr . Applying the previous case
to the cocycle ∧r A and w ∈ v(∞) (∧r A), we conclude that
1
lim log∧r A(n) (x) (v1 ∧ . . . ∧ vr ) = L1 (∧r A) = r L1 (A).
n→±∞ n
Free ebooks ==> www.Ebook777.com
4.2 The Ergodic Theorems 141
r
log∧r A(n) (x)(v1 ∧ · · · ∧ vr ) ≤ logA(n) (x)vi ≤ r logA(n) (x),
i=1
and since both upper and lower bounds of this sum converge to r L1 (A), as n → ±∞,
we conclude that for all i = 1, . . . , r, and in particular for i = 1,
1
lim logA(n) (x)vi = L1 (A).
n→±∞ n
Ej = vτ(∞)
j
(A∗ ) ∩ vτ(∞)
j−1
(A)⊥ = vτ(∞)
j −τj−1
(A|v⊥τ ).
j−1
1 1
lim logA(n) (x) v = lim log(A|Bj )(n) (x) v
n→±∞ n n→±∞ n
= Lτj −τj−1 (A|Bj ) = Lτj (A) = λj (A).
sin min (⊕j≤l Ej (x), ⊕j>l Ej (x))
= ∠m in(v(∞) (A∗ ), v(∞) (A)⊥ )
τl τl
≥ ατl (vτ(∞)
l
(A∗ ), vτ(∞)
l
(A)) = α(v (∞)
(∧τl A ), v(∞) (∧τl A)).
∗
Given an ergodic system satisfying base and uniform fiber LDT estimates, we prove
that the Oseledets filtration and decomposition vary continuously with the cocycle in
an L 1 sense. We begin by proving the continuity of the most expanding direction, as
it contains the main ingredients of our argument. We then define the space of mea-
surable filtrations and endow it with an appropriate topology. Using the construction
of the Oseledets filtration and decomposition in Sect. 4.2.3, we deduce the continuity
www.Ebook777.com
142 4 The Oseledets Filtration and Decomposition
of these two quantities from that of the most expanding direction. This is obtained
via some Grassmann geometrical considerations established in Chap. 2.
Throughout this section we will be under the assumptions of the ACT (Theo-
rem 3.1 in Chap.3). That is, given an ergodic MPDS (X, μ, T ), a space of measur-
able cocycles C , a set of observables Ξ , a set of LDT parameters P, we assume the
following:
1. Ξ is compatible with every cocycle A ∈ C .
2. Every observable ξ ∈ Ξ satisfies a base-LDT estimate: for every ε > 0 there is
p = p(ξ, ε) ∈ P, p = (n0 , ε, ι), such that for all n ≥ n0 we have εn ≤ ε and
1 n−1
μ {x ∈ X :
ξ(T j x) − ξ dμ
> εn } < ιn . (4.11)
n j=0 X
μ {x ∈ X :
logB(n) (x) − L1(n) (B)
> εn } < ιn . (4.12)
n
As before, εn := ε(n), ιn := ι(n), where ε, ι : (0, ∞) → (0, ∞) are such that
the deviation size functions ε(t) are non-increasing, while the deviation set measure
functions ι(t) are continuous and strictly decreasing to 0, as t → ∞, at least like a
power and at most like an exponential. The latter constraint is just for convenience,
so we can write, whenever needed, e−c0 n + ιn ιn . Moreover, it will be convenient
in this section to assume a stronger decay at infinity of ι(t), which holds in all our
applications. Hence we will assume that
We employ the Lipschitz estimates on Grassmann manifolds and the avalanche prin-
ciple derived in Chap. 2.
Recall from Sect. 4.2.3 that the most expanding direction of the nth iterate of a
cocycle A ∈ C defines a partial function
4.3 Abstract Continuity Theorem of the Oseledets Filtration 143
(n) v(A(n) (x)) if gr(A(n) (x)) > 1
v (A)(x) :=
undefined otherwise.
where the quantity under the integral sign refers to the distance between points in
the projective space P(Rm ).
Clearly all the functions v(n) (A) are in L 1 (X, P(Rm )), and by dominated conver-
gence we have that as n → ∞,
We will prove that if L1 (A) > L2 (A), then locally near A, the map B → v(∞) (B)
is continuous with a modulus of continuity depending on the LDT parameter.
We do so by deriving a quantitative version of the convergence in (4.13), which
moreover is somewhat uniform in phase and cocycle. This more precise convergence
comes as a consequence of the availability of the LDT estimates for our system, as the
exceptional sets of phases in the domain of applicability of the avalanche principle
can be precisely (and uniformly in the cocycle) measured.
Fix a cocycle A ∈ C such that L1 (A) > L2 (A). Let κ(A) := L1 (A) − L2 (A) > 0 if
L2 (A) > −∞ or else let κ(A) be a large enough constant. Fix ε0 := κ(A)/100. What
follows is a bookkeeping of various exceptional sets related to notions from Chap. 3.
They will eventually define and measure the exceptional sets in the domain of applica-
bility of the avalanche principle (AP) in Proposition 2.42, for certain sequences of
iterates of a cocycle B in a small neighborhood of A. They depend on A and ε0 , hence
only on A.
Pick for the rest of this subsection δ = δ(A) > 0, n0 = n0 (A) ∈ N, ι = ι(A) ∈ I
such that, by Lemma 3.4 in Chap. 3 we have: for all B ∈ Cm with dist(B, A) < δ,
1
gr(B(n) (x)) > en κ(A)/2 =: (> 1) (4.14)
κn
holds for all n ≥ n0 and for all x outside a set of measure < ιn , and
(n)
L (B) − L (m) (A)
< κ(A)/50 (4.15)
1 1
As before, (4.14) will ensure the gap condition in the AP, while (4.15) via
Lemma 3.3 will ensure the angle condition.
Fix a cocycle B with dist(B, A) < δ. We will define, for all scales n ≥ n0 , the
exceptional sets outside which the AP can be applied for various block lengths and
configurations of block components.
The exceptional set in the nearly uniform upper semicontinuity of the maximal
Lyapunov exponent (Proposition 3.1 in Chap. 3) depends on A and ε0 , hence only on
n (A). Its measure is μ [Bn (A)] < ιn .
A, and we denote it by Busc usc
Let
1
Bldt
logB(n) (x) − L (n) (B)
> ε0 }
n (B) := {x ∈ X : 1
n
g
A simple inspection of the proof of Lemma 3.1 in Chap. 3 shows that Bn (B) is
g
contained in the exceptional set in (4.14), and its measure satisfies μ [Bn (A)] ιn .
(n)
Note also that (4.14) ensures that v(B (x)) is defined (since there is a gap between
the two largest singular values).
Moreover, a simple inspection of the proof of Lemma 3.3 in Chap. 3, combined
with (4.15), shows that for 2n ≥ m1 , m2 ≥ n ≥ n0 the bound
Note from (4.14) and (4.16) that κε2n = e−n κ(A)/10 < ιn 1, hence the condition
n
on κ and ε from the AP is satisfied.
The bound on the distance between most expanding directions in the conclusion
of the AP is κn
= e−3 n κ(A)/10 < ιn . (4.17)
εn
When using the AP, we will always apply (4.16) to configurations for which n is
fixed and m1 = n while n ≤ m2 ≤ 2n. This motivates defining
−n ldt
Ban (B) := [Bldt
m (B) ∪ T Bm (B)].
n≤m≤3n
Let
n (B) := Bn (B) ∪ Bn (B),
Bga g a
ga ga
so μ [Bn (B)] n ιn , and if x ∈
/ Bn (B), both the gap and the angle conditions hold
for appropriate block components at scale n.
−1/3
Let n0 ≥ n0 and 2n0 ≤ n1 ≤ ιn0 . If we define
Bap
n0 (B) := T −in0 Bga
n0 (B), (4.18)
−1/3
0≤i<n0−1 ιn0
−1/3
then μ [Bn0 (B)] n0−1 ιn0 n0 ιn0 < ιn0 and if x ∈
ap 1/2 ap
/ Bn0 (B), then the AP can be
applied to a block of length n1 whose n components have lengths n0 , except for the
last, whose length is the remaining integer m satisfying n0 ≤ m ≤ 2n0 .
Now define for all n ≥ n0 the nested, decreasing sequence of exceptional sets
Bn (B) :=
ap
Bk (B). (4.19)
k≥n
Clearly
μ [Bn (B)] ≤
ap 1/2
μ [Bk (B)] ιk ι1/2
n
k≥n k≥n
−1/3
/ Bn (B) then for any scales n0 , n1 such that n0 ≥ n and 2n0 ≤ n1 ≤ ιn0 ,
and if x ∈
ap
/ Bn0 (B), the AP can be applied to a block of length n1 whose components
since x ∈
have lengths n0 .
Remark 4.2 Let us summarize the accounting above. Given a cocycle A ∈ Cm with
L1 (A) > L2 (A), there are parameters δ, n0 , ι depending only on A so that for any
cocycle B ∈ Cm with dist(B, A) < δ and for any scale k ≥ n0 , there are exceptional
ap 1/2
sets Bk (B) and Bk (B) of measure < ιk such that
−1/3 ap
1. If n0 ≥ n0 , 2n0 ≤ n1 ≤ ιn0 and if x ∈ / Bn0 (B), then the AP with parameters
κn0 , εn0 can be applied to a block B(n1 ) (x) of length n1 whose components have
lengths n0 , except possibly for the last, whose length is between n0 and 2n0 .
2. If n ≥ n0 and if x ∈ / Bn (B), then for any scales n0 , n1 such that n0 ≥ n and
−1/3
2n0 ≤ n1 ≤ ιn0 , the AP with parameters κn0 , εn0 can be applied to a block
B(n1 ) (x) of length n1 whose components have lengths n0 .
Proof Consider the block B(n1 ) (x) and break it down into n−1 many blocks of length
n0 each, and a remaining block of length m with n0 ≤ m < 2n0 . In other words,
write n1 = (n − 1) n0 + m, for some n0 ≤ m < 2n0 and define
gi = gi (x) := B(n0 ) (T i n0 x)
for 0 ≤ i ≤ n − 2, and
Then
g(n) = gn−1 . . . g1 g0 = B(n1 ) (x),
for 1 ≤ i ≤ n − 2, while
κn0
d(v(B(m) (x)), v(B(n0 ) (x))) . (4.21)
εn0
κmj
d(v(B(mj+1 ) (x)), v(B(mj ) (x))) < . (4.22)
εmj
4.3 Abstract Continuity Theorem of the Oseledets Filtration 147
−1/3
Moreover, since ιn0 ≥ m0 ≥ n01+c ≥ 2n0 , and since x ∈ / Bn0 (B), hence x ∈
/
ap
Bn0 (B),
Lemma 4.10 is applicable also at scales m0 , n0 , and we have
κn0
d(v(B(m0 ) (x)), v(B(n0 ) (x))) < . (4.23)
εn0
κn0 κmj
k
κn
≤ + 0.
εn0 j=0
εmj εn0
The last inequality holds because by (4.17) we have κεnn = e−3 n κ(A)/10 , hence the
series above converges rapidly, and so its sum is comparable with the first term.
Moreover,
dist(v(n) (B), v(∞) (B)) < ι1/2
n . (4.25)
Proof The estimate (4.24) follows directly by taking the limit in Lemma 4.11 with
n0 = n and m → ∞. The second estimate follows by integration in x.
Moreover,
dist(v(n) (B1 ), v(n) (B2 )) < ιn . (4.27)
148 4 The Oseledets Filtration and Decomposition
Proof To prove (4.26) we will use the Lipschitz continuity of the most expanding
singular direction in Proposition 2.40 from Chap. 2.
Put gi = gi (x) := Bi(n) (x), i = 1, 2.
g g
Let x ∈ / Bn (B1 ) ∪ Bn (B2 ), which is a set of measure ιn . We show that the
assumptions of Proposition 2.40 hold for all such x.
Firstly note that by (4.14) we have gr(gi ) = gr(Bi(n) (x)) > κ1n > 1, so in particular
v(Bi(n) (x)) are well defined.
Moreover, the fiber-LDT estimate applies to Bi and we have
1 κ(A) κ(A)
logBi(n) (x) > L1(n) (Bi ) − ε0 > L1(n) (A) − − > −C0 ,
n 50 100
where we used (4.15) in the estimate above, and C0 = C0 (A) < ∞.
Then
gi = Bi(n) (x) > e−C0 n .
Since for every x, Bi (x) < C(A) < ∞, by possibly increasing C0 , we may also
assume that
gi = Bi(n) (x) < eC0 n .
g1 − g2 = B1(n) (x) − B2(n) (x) ≤ n eC0 (n−1) dist(B1 , B2 ) < e−(C1 −2C0 ) n .
16
d(v(g1 ), v(g2 )) ≤ drel (g1 , g2 ) e−C2 n .
1 − κn2
We are now ready to formulate and to prove the continuity of the most expanding
direction.
Theorem 4.6 Let A ∈ Cm with L1 (A) > L2 (A). There are δ > 0, ι ∈ I, c > 0, α > 0,
all depending only on A, such that for any cocycles B1 , B2 ∈ Cm with dist(Bi , A) < δ,
where i = 1, 2, we have:
μ {x ∈ X : d(v(∞) (B1 )(x), v(∞) (B2 )(x)) > dist(B1 , B2 )α } < ωι (dist(B1 , B2 )),
4.3 Abstract Continuity Theorem of the Oseledets Filtration 149
Proof Fix any C2 > 0 and let C1 be the constant in Proposition 4.19.
Put dist(B1 , B2 ) =: h and choose n ∈ N such that h e−C1 n . Since h ≤ 2δ and
n 1/C1 log 1/h, by taking δ small enough we may assume that n ≥ n0 .
Apply Proposition 4.19 to get that for x outside a set of measure < ιn ,
1/2
Combine (4.29) and (4.30) to conclude that for x outside a set of measure ιn
and for c0 < min{C2 , 3κ(A)/10}, we have
meaning that B will still have a τ = (2, 3) gap pattern. However, this might not be
its exact gap pattern, as we could have L1 (B) > L2 (B), leading to a finer gap pattern,
say τ = (1, 2, 3). If τ were the exact gap pattern of B, then its Oseledets filtration
would be a τ ⊥ = (m − 1, m − 2, m − 3)-flag
Now F(A)(x) and F τ (B)(x) are both τ ⊥ -flags, and we may define a distance
between them component-wise (as points in the same Grassmann manifold). The
distance between the measurable filtrations F(A) and F τ (B) as functions on X will
be the space average of the pointwise distances.
Furthermore, the Oseledets decomposition E· (A) of the cocycle A with exact
τ = (2, 3) gap pattern, consists of a 2-dimensional subspace E1 (A)(x) corresponding
to L1 (A) = L2 (A), a one dimensional subspace E2 (A)(x) corresponding to L3 (A), and
an m −3-dimensional subspace E3 (A)(x) corresponding to the remaining (and equal)
Lyapunov exponents.
If a small perturbation B of A has (as above) the finer τ = (1, 2, 3) gap pattern,
then its Oseledets decomposition will consist of subspaces E1 (B)(x) (one dimen-
sional, corresponding to L1 (B)), E2 (B)(x) (one dimensional, corresponding to L2 (B)),
E3 (B)(x) (one dimensional, corresponding to L3 (B)) and the subspace E4 (B)(x) (m−3
dimensional, corresponding to the remaining Lyapunov exponents).
In order to compare the Oseledets decompositions of A and B, we would have to
“patch up” the first two Oseledets subspaces for B. In other words, we will consider
the natural restriction E·τ (B) of the Oseledets decomposition E· (B), consisting of the
subspaces E1 (B) ⊕ E2 (B), E3 (B), E4 (B).
Free ebooks ==> www.Ebook777.com
4.3 Abstract Continuity Theorem of the Oseledets Filtration 151
We make the obvious observation that for two dimensional (i.e. Mat(2, R)-valued)
cocycles, or for cocycles of any dimension with simple Lyapunov spectrum, these
projection/restriction of the filtration/decomposition are not needed.
Let us now formally define the space of measurable filtrations.
Given two signatures τ = (τ1 , . . . , τk ) and τ = (τ1 , . . . , τk ), we say that τ
refines τ , and write τ ≥ τ , if {τ1 , . . . , τk } ⊇ {τ1 , . . . , τk }.
Given τ ≥ τ , there is a natural projection ρτ,τ : Fτ (Rm ) → Fτ (Rm ), defined by
www.Ebook777.com
152 4 The Oseledets Filtration and Decomposition
Finally, we endow the space F(X, Rm ) of all measurable filtrations of Rm with the
topology determined by the following neighborhood bases,
Theorem 4.7 Let A ∈ Cm be a cocycle with a τ gap pattern. Then locally near A,
the map
Cm B → F τ (B) ∈ F⊃τ (X, Rm )
154 4 The Oseledets Filtration and Decomposition
is continuous with a modulus of continuity ω(h) := [ι (c log h1 )]1/2 for some con-
stant c = c(A) > 0 and for some deviation measure function ι = ι(A) from the
corresponding set of LDT parameters.
In fact, a stronger pointwise estimate holds:
μ {x ∈ X : d(F τ (B1 )(x), F τ (B2 )(x)) > dist(B1 , B2 )α } < ω(dist(B1 , B2 )),
Proof Since A has a τ = (τ1 , . . . , τk ) gap pattern, Lτj (A) > Lτj−1 (A) for all indices
j, so L1 (∧τj A) > L2 (∧τj A). We may then apply the continuity of the most expanding
direction in Theorem 4.6 to ∧τj A and obtain that
and
vτ(∞)
j
(A)(x) = Ψ −1 (v(∞) (∧τj A)(x)).
the first two assertions follow from the continuity of the most expanding direction
and the fact that the Plücker embedding Ψ and the orthogonal complement ⊥ are
isometries. The third assertion is an immediate consequence of Proposition 4.20.
By the continuity of the Lyapunov exponents, any nearby cocycle B has the same
or a finer gap pattern τ ≥ τ . Let E·τ (B)(x) denote the τ -restriction of E· (B)(x) to
the space of decompositions with signature τ . Clearly we have
where
vτ (B(n) (x)) = vτ1 (B(n) (x)), . . . , vτk (B(n) (x))
= Ψ −1 (v(∧τ1 B(n) (x))), . . . , Ψ −1 (v(∧τk B(n) (x))) .
Consider the exceptional sets defined in Sect. 4.3.1 for each dimension τj , that is,
define
Bn (B) := Bn (∧τj B).
1≤j≤k
Redefine κ(A) := min κ(∧τj A), which subsequently determine κn and εn as in (4.14)
1≤j≤k
and (4.16).
Since A has a τ = (τ1 , . . . , τk ) gap pattern, the estimates on the most expanding
direction, namely Remark 4.2, Propositions 4.18 and 4.19 are applicable to ∧τj B,
1 ≤ j ≤ k. We summarize the relevant results in the following remark.
Remark 4.3 There are parameters δ, n0 and ι, depending only on A, such that the
following hold for all cocycles B with dist(B, A) < δ and for all scales n ≥ n0 .
156 4 The Oseledets Filtration and Decomposition
4. The partial functions vτ(n) (B) satisfy the following finite scale uniform continuity
property. Given C2 > 0, there is C1 = C1 (A, C2 ) < ∞ such that for any cocycles
Bi ∈ Cm with dist(Bi , A) < δ for i = 1, 2, if dist(B1 , B2 ) < e−C1 n , then for x
outside a set of measure < ιn we have:
dτ vτ(n) (B1 )(x), vτ(n) (B2 )(x) < e−C2 n . (4.34)
Proof The statements in item 1 above follow from (4.14) and (4.16) applied to ∧τj B
for 1 ≤ j ≤ k.
Each component of the flag vτ (B(n) (x)) converges, for μ-a.e. x ∈ X, by Proposi-
tion 4.4 and the fact that B has the τ gap pattern.
The rate of convergence in item 3 is a consequence of Proposition 4.18 applied in
each component of the flag vτ (B(n) (x)), that is, applied to the exterior powers ∧τj B
for 1 ≤ j ≤ k. The same argument holds for item 4.
Remark 4.4 Since A has the τ gap pattern, so does A∗ . Therefore, by possibly dou-
bling the size of the exceptional set, we may assume that the rate of convergence
(4.33) holds for both B and B∗ . The same applies to the finite scale continuity (4.34).
if
gr τ (B(n) (x)) > 1 and θ vτ(n) (B∗ )(x), vτ(n) (B)(x)⊥ > 0,
otherwise it is undefined.
4.3 Abstract Continuity Theorem of the Oseledets Filtration 157
/ Bn (B).
Clearly this map is well defined for all x ∈
We begin by establishing a lower bound on the transversality measurement for
the flags defining this finite scale decomposition.
Proof This lower bound follows easily from Proposition 2.39 in Chap. 2 and the
second inequality in item 1 of Remark 4.3.
θ vτ(n) (B∗ )(x), vτ(n) (B)(x)⊥ = θ vτ (B∗ (n) (x)), vτ (B(n) (x))⊥
∗
= θ vτ (B(n) (T −n x) ), vτ (B(n) (x))⊥
≥ ατ (B(n) (T −n x), B(n) (x)) ≥ εn .
Next we establish the convergence to E·τ (B) of the finite scale decomposition
introduced above.
κn
d E·(n) (B)(x), E·τ (B)(x) < 2 . (4.36)
εn
Proof Fix the phase x and the scale n. For simplicity of notation let
With these notations we have E·τ (B)(x) = F F and E·(n) (B)(x) = F0 F0 .
By Proposition 2.38 in Chap. 2, we have
d E·(n) (B)(x), E·τ (B)(x) = d(F0 F0 , F F )
1 1
≤ max , (dτ (F0 , F) + dτ ⊥ (F0 , F )). (4.37)
θ (F0 , F0 ) θ (F, F )
while applying (4.33) to B and using the fact the the orthogonal complement ⊥ is an
isometry, we get:
dτ ⊥ (F0 , F ) = dτ ⊥ vτ(n) (B)(x)⊥ , vτ(∞) (B)(x)⊥
κ
= dτ vτ(n) (B)(x), vτ(∞) (B)(x) <
n
. (4.39)
εn
and by Lemma 2.37 in Chap. 2 combined with (4.38) and (4.39) we have:
Remark 4.5 The proposition above shows in particular that the partially defined
finite scale decompositions E·(n) (B)(x) converge for μ-a.e. x ∈ X to the τ -restriction
E·τ (B)(x) of the Oseledets decomposition of B.
Proposition 4.23 (finite scale continuity) There are constants C1 = C1 (A) < ∞
and C3 = C3 (A) > 0 such that for any cocycles Bi ∈ Cm with dist(Bi , A) < δ for
i = 1, 2, if dist(B1 , B2 ) < e−C1 n , then for x outside a set of measure < ιn and n ≥ n0
we have:
d E·(n) (B1 )(x), E·(n) (B2 )(x) < e−C3 n . (4.42)
Proof Let C2 > κ(A)/2. We apply item 4 of Remark 4.3. There is C1 = C1 (A) such
that for any cocycles Bi ∈ Cm with dist(Bi , A) < δ for i = 1, 2, there is a set of
phases of measure < ιn such that outside of that set, (4.34) holds for both B1 , B2 and
B1∗ , B2∗ .
Fix such a phase x, and to simplify notations, for i = 1, 2 let
We are now ready to formulate the ACT for the Oseledets decomposition.
Theorem 4.8 Let A ∈ Cm be a cocycle with a τ gap pattern. Then locally near A,
the map
Cm B → E·τ (B) ∈ Dτ (X, Rm )
is continuous with a modulus of continuity ω(h) := [ι (c log h1 )]1/2 for some con-
stant c = c(A) > 0 and for some deviation measure function ι = ι(A) from the
corresponding set of LDT parameters.
In fact, a stronger pointwise estimate holds:
μ {x ∈ X : d(E·τ (B1 )(x), E·τ (B2 )(x)) > dist(B1 , B2 )α } < ω(dist(B1 , B2 )),
Proof The first two assertions are derived from the speed of convergence in Proposi-
tion 4.22 and the finite scale continuity in Proposition 4.23 in exactly the same way
we derived the continuity of the most expanding direction in Theorem 4.6.
The third assertion is an immediate consequence of Proposition 4.21.
160 4 The Oseledets Filtration and Decomposition
References
Abstract In this chapter we prove the continuity of all Lyapunov exponents, as well
as the continuity of the Oseledets decomposition for a class of irreducible cocycles
over strongly mixing Markov shifts. Moreover, gaps in the Lyapunov spectrum lead
to a Hölder modulus of continuity for these quantities. This result is an application
of the abstract continuity theorems obtained in previous chapters, and generalizes a
theorem of E. Le Page on the Hölder continuity of the maximal LE for one-parameter
families of strongly irreducible and contracting cocycles over a Bernoulli shift.
C(E 0 , . . . , E m ) := { x ∈ X + : x j ∈ E j , for 0 ≤ j ≤ m },
Let (L ∞ (Σ),
·
∞ ) denote the Banach algebra
of complex
bounded F-measurable
functions with the sup norm
f
∞ = supx∈Σ f (x). The following concept corre-
sponds to condition (A1) in [1].
Definition 5.6 We say that a Markov system (K , μ) is strongly mixing if there are
constants C > 0 and 0 < ρ < 1 such that for every f ∈ L ∞ (Σ), all x ∈ Σ and
n ∈ N,
164 5 Large Deviations for Random Cocycles
f (y) K n (x, dy) − f (y) μ(dy) ≤ C ρ n
f
∞ .
Σ Σ
Proposition 5.1 If the Markov system (K , μ) is strongly mixing then the Markov
shift (X, Pμ , T ) is a mixing dynamical system.
n+
p−1
Ex0 [ f (en , . . . , en+ p )] = ··· f (xn , . . . , xn+ p ) K n (x0 , d xn ) K (x j , d x j+1 )
Σ Σ j=n
n+ p−1
··· f (xn , . . . , xn+ p ) μ(d xn ) K (x j , d x j+1 ) = Eμ ( f ).
Σ Σ j=n
Hence
converges to
−1
Eμ ( f ) ··· g(x−q , . . . , x−1 ) μ(d x−q ) K (x j , d x j+1 ) = Eμ ( f ) Eμ (g).
Σ Σ j=−q
The mixing property of the shift (X, Pμ , T ) follows applying the previous argument
to the indicator functions of any two cylinders, because the σ -algebra of cylinders
generates the Borel σ -field of X .
Examples of strongly mixing Markov systems arise naturally from Markov ker-
nels satisfying the Doeblin condition (see [3]). We say that K satisfies the Doeblin
condition if there is a positive finite measure ρ on (Σ, F) and some ε > 0 such that
5.1 Introduction and Statements 165
K (x, E) ≥ 1 − ε ⇒ ρ(E) ≥ ε.
Given E ∈ F, define
Proposition 5.2 Let K be a Markov kernel on (Σ, F). If K satisfies the Doeblin
condition then there are sets Σ1 , . . . , Σm in F and probability measures ν1 , . . . , νm
on Σ such that for all i, j = 1, . . . , m,
1. Σi ∩ Σ j = ∅ when i = j,
2. Σi is K -forward invariant, i.e., K (x, Σi ) = 1 for x ∈ Σi ,
3. νi is K -stationary and ergodic with νi (Σ j ) = δi j ,
4. lim n→+∞ K n (x, Σ1 ∪ · · · ∪ Σm ) = 1, with geometric uniform speed of conver-
gence, for all x ∈ Σ,
5. ν(Σ1 ∪ · · · ∪ Σm ) = 1, for every K -stationary probability ν.
Moreover, for every 1 ≤ i ≤ m there is an integer pi ∈ N and measurable sets
Σi,1 , . . . , Σi, pi ∈ F such that
1. {Σi,1 , . . . , Σi, pi } is a partition of Σi ,
2. K (x, Σi, j+1 ) = 1 for x ∈ Σi, j and 1 ≤ j ≤ pi , with Σi, pi +1 = Σi,1 ,
3. the Markov system (Σi, j , K pi ) is strongly mixing for all 1 ≤ j ≤ pi .
FA (x, v) := (T x, A(x) v) ,
where we identify A with the function A : X → GL(m, R), A(x) := A(x0 , x1 ), for
x = (xn )n∈Z ∈ X . These will be referred to as random Markov cocycles.
The cocycle FA is determined by the data (K , μ, A), and identified by the function
A whenever the Markov system (K , μ) is fixed.
Definition 5.9 Let Gr(Rm ) denote the Grassmann manifold of the Euclidean space
Rm . An F-measurable section V : Σ → Gr(Rm ) is called A-invariant when
In general E may fail to be a Borel set, but it is an analytic set in the sense of
descriptive set theory (see [8, Definition 14.1 and Exercise 14.3]). By [8, Theorem
21.10] this set is universally measurable, and in particular it is measurable w.r.t. μ.
Hence, because of the strong mixing property,
5.1 Introduction and Statements 167
μ(Σ \ E) = lim Es0 [1Σ\E (en )] = lim Ps0 { x ∈ Ω : en (x) ∈ Σ \ E } = lim Ps0 (∅) = 0 ,
n→∞ n→∞ n→∞
which proves that for μ-a.e. s ∈ Σ there exists a sequence x ∈ Ω such that x0 = s0
and xn = s for some n ∈ N.
Then Vk (s) = Ak (xn−1 , xn ) . . . Ak (x1 , x2 ) Ak (x0 , x1 ) Vk (s0 ), which implies that
the sequence Vk (s) converges to A(xn−1 , xn ) . . . A(x1 , x2 ) A(x0 , x1 ) V0 when k →
∞. Thus, Vk (s) converges for μ-a.e. s ∈ Σ, and the limit function V (s) =
limk→∞ Vk (s) is a measurable and proper A-invariant section, with the same dimen-
sion as the sections Vk . This proves that the cocycle A is reducible.
For the reader’s convenience we briefly recall some definitions and notations
regarding the Lyapunov exponents, Oseledets filtrations and decompositions of a
cocycle A in any space of cocycles Cm .
The ergodic theorem of Kingman allows us to define the Lyapunov exponents
L j (A) with 1 ≤ j ≤ m as L j (A) := Λ j (A) − Λ j−1 (A) where
1
Λ j (A) := lim log
∧ j A(x)
for μ-a.e. x ∈ X.
n→∞ n
These theorems are proved in Sect. 5.4.1. They are applications of Theorem 3.1
in Chap. 3, and Theorem 4.7 and 4.8 in Chap. 4. The main ingredients in these appli-
cations are two theorems on base and fiber uniform LDT estimates of exponential
type that we now formulate.
We begin with the base LDT theorem. Consider the metric d : X × X → [0, 1]
The last set, Hα (X ), is the space of Hölder continuous functions with exponent α
w.r.t. the distance d on X . In fact it follows easily from the definition that
f (x) − f (x )
vα ( f ) = sup .
x=x x )α
d(x,
vk ( f g) ≤
f
∞ vk (g) +
g
∞ vk ( f ),
vα ( f g) ≤
f
∞ vα (g) +
g
∞ vα ( f ).
f g
α ≤
f
α
g
α ,
and clearly
1
α =
1
∞ + vα (1) = 1 + 0 = 1. The proof that (Hα (X ),
·
α ) is a
lattice and a Banach space is left as an exercise.
Theorem 5.2 Let (K , μ) be a strongly mixing Markov system. For any 0 < α ≤ 1
and ξ ∈ Hα (X − ) there exist C = C(ξ ) < ∞, k = k(ξ ) > 0 and ε0 = ε0 (ξ ) > 0
such that for all 0 < ε < ε0 , x ∈ Σ and n ∈ N,
⎡ ⎤
1 n−1
Pμ ⎣ ξ ◦ T j − Eμ (ξ ) > ε ⎦ ≤ C e−k ε n .
2
n j=0
The fiber LDT theorem, proved in Sect. 5.3.2, has the following statement.
n−1
n−1
Sn (ξ )(x) := ξ̂ ◦ T j (x) = ξ(x j ).
j=0 j=0
Definition 5.14 We say that ξ satisfies LDT estimates of exponential type if there
exist positive constants C, k and ε0 such that for all n ∈ N, 0 < ε < ε0 and x ∈ Σ,
1
P+ y ∈ X : Sn (ξ )(y) − Eμ (ξ ) > ε
+
≤ C e−n k ε .
2
x
n
is called the second characteristic function of η, also known as the cumulant gener-
ating function of η (see [11]).
Proof For (1) notice that the assumptions imply that the parametric integral Ex (e z η )
and its formal derivative Ex (e z η η) are well-defined continuous functions on the disk
|z| < a. Since c(η, x, 0) = log Ex (1) = log 1 = 0, (2) follows. Property (3) holds
because dcdt
(η, x, 0) = Ex (η 1)/Ex (1) = Ex (η). The convexity (5) follows by Hölder
inequality, with conjugate exponents p = 1/s and q = 1/(1 − s), where 0 < s < 1.
In fact, for all t1 , t2 ∈ R,
s t2 η 1−s
c(η, x, s t1 + (1 − s) t2 ) = log Ex [ et1 η e ]
t1 η s
t2 η 1−s
≤ log Ex [e ] Ex [e ]
= s c(η, x, t1 ) + (1 − s) c(η, x, t2 ).
Definition 5.16 We call limit cumulant generating function of the process {Sn (ξ )}n≥0
any function c(ξ, ·) : Da (0) → C such that there exist a constant C > 0 and a
numeric sequence {δn }n≥0 for which the following properties hold:
(1) c n (ξ, ·) is well defined and analytic
on Da (0), for all n ∈ N,
(2) n c(ξ, z) − cn (ξ, x, z) ≤ C z + δn , for all n ∈ N, z ∈ Da (0) and x ∈ Σ,
(3) limn→+∞ δn = 0.
www.Ebook777.com
172 5 Large Deviations for Random Cocycles
Before discussing why they exist, let us draw some conclusions from the existence
of limit cumulant generating functions.
Proof The function c(ξ, z) is analytic on Da (0) because it is the uniform limit of the
sequence of analytic functions n1 cn (ξ, x, z). This proves (1).
Item (2) follows directly from Proposition 5.5(2).
Consider now the sequence of analytic functions
ĉn (ξ, z) := cn (ξ, x, z) dμ(x).
Σ
Then
d ĉn
(ξ, 0) = Ex [Sn (ξ )] dμ(x) = Eμ (ξ̂ ).
dt Σ
The next proposition relates the existence of a limit cumulant generating function
for the process {Sn (ξ )}n≥0 with LDT estimates of exponential type for ξ .
Proof Let us abbreviate c(t) = c(ξ, t). We can assume that c (0) = Eμ (ξ̂ ) = 0.
Otherwise we would work with ξ = ξ − Eμ (ξ ) 1, for which Eμ (ξ̂ ) = 0. Notice that
the normalized process {Sn (ξ )}n≥0 admits the limit cumulant generating function
c(ξ , t) = c(t) − t Eμ (ξ ) = c(t) − t c (0).
5.1 Introduction and Statements 173
Since h > c (0), we can choose 0 < t0 < a such that for all t ∈ (−t0 , t0 ),
h t2
0 ≤ c(t) < .
2
By Definition 5.16, for all t ∈ (−t0 , t0 ),
C0 n h t 2
Ex [et Sn (ξ ) ] = ecn (ξ,x,t) ≤ en c(t)+C |t|+δn ≤ e 2 ,
2
C t0 +supn≥0 δn
where C0 := 2e . Thus, by Chebyshev’s inequality (see (1.6)) we have
that for all t < t0
C0 −n tε− h 2t
2
P+
x [ Sn (ξ ) > nε ] ≤ e
−tnε
Ex [et Sn (ξ ) ] ≤ e .
2
ε
Given 0 < ε < ε0 := h t0 , pick t = h
∈ (0, t0 ). This choice of t minimizes the
2
− tε− h 2t
function g(t) = e . For this value of t we obtain
C 0 − ε2 n
P+
x [ Sn (ξ ) > nε ] ≤ e 2h .
2
We can derive the same conclusion for −ξ , because c(ξ, −t) is a limit cumulant
generating function of the process {Sn (−ξ )}n≥0 ,
1 ε2
P+
x [ Sn (ξ ) < −nε ] = Px [ Sn (−ξ ) > nε ] ≤ C 0 e− 2 h n .
2
Thus, for all x ∈ Σ, 0 < ε < ε0 and n ∈ N,
ε2
P+
x [ |Sn (ξ )| > nε ] ≤ C 0 e
− 2h n
.
Remark 5.2 To obtain a sharp upper bound on the rate function for the large devia-
tions of the process Sn (ξ ) we should have used the Legendre transform of the convex
function c(t)−t c (0). Here because we do not care about sharp estimates, but mainly
to avoid dealing with the degenerate case where c(t) is not strictly convex, we have
replaced c(t) − t c (0) with its upper bound h2t on the small neighborhood (−t0 , t0 ),
2
Corollary 5.1 Assume there is continuous map c : X → H(Da (0)) such that
(a) for each (K , μ, ξ ) ∈ X, the function c(ξ, z) := c(K , μ, ξ )(z) is a limit cumulant
generating function of the process {Sn (ξ )}n≥0 on Da (0),
(b) the parameters C and δn in Definition 5.16 can be chosen uniformly in X.
Then
(1) For each (K , μ, ξ ) ∈ X there exists a neighborhood V in X such that V satisfies
uniform LDT estimates of exponential type.
2
(2) If there exists h > 0 such that dtd 2 c(ξ, 0) < h for all (K , μ, ξ ) ∈ X then X
satisfies uniform LDT estimates of exponential type.
Proof Given (K 0 , μ0 , ξ0 ) ∈ X, let c0 (t) := c(K 0 , μ0 , ξ0 )(t), and take h > c0 (0).
By continuity of c : X → H(Da (0)) there exist a neighborhood V of (K 0 , μ0 , ξ0 ) in
X and t0 > 0 such that for any (K , μ, ξ ) ∈ V, the function c(ξ, z) := c(K , μ, ξ )(z)
satisfies for all t ∈ (−t0 , t0 ),
dc h t2
c(ξ, t) − t (ξ, 0) < .
dt 2
The argument used to prove Proposition 5.7 shows that V satisfies uniform LDT
estimates of exponential type.
The strategy to meet the assumptions of Corollary 5.1, i.e., to prove the existence
of a limit cumulant generating function for the process {Sn (ξ )}n≥0 , is a spectral
method that we describe now.
Define a family of Laplace-Markov operators
(Q t f )(x) = (Q K ,ξ,t f )(x) := f (y) et ξ(y) K (x, dy),
Σ
on some appropriate Banach space B, embedded in L ∞ (Σ, F), and containing the
constant functions. Notice that by definition (Q t 1)(x) = Ex [et ξ̂ ]. Hence, iterating
this relation we obtain the following formula for the moment generating function of
Sn (ξ ): for all x ∈ Σ and n ∈ N,
Ex [et Sn (ξ ) ] = (Q nt 1)(x).
From this relation we infer that c(t) = log λ(t) is a limit cumulant generating function
for the process Sn (ξ ). Therefore, by Proposition 5.7, ξ satisfies LDT estimates of
exponential type.
To obtain uniform LDT estimates, through Corollary 5.1, we assume some
weak continuous dependence of the family of operators t → Q K ,ξ,t on the
observed Markov system (K , μ, ξ ), which implies that the eigenvalue function
λ(t) ∈ H(Da (0)) also depends continuously on (K , μ, ξ ).
We mention briefly some of the origins of this subject. One is the proof by H.
Furstenberg and H. Kesten of a law of large numbers for random i.i.d. products
of matrices [4], which was later abstracted by Furstenberg to a seminal theory on
random products in semisimple Lie groups [5]. In this context, a first central limit
theorem was proved by Tutubalin in [16].
Since its origin, the scope of Furstenberg’s theory has been greatly extended by
many contributions (see for instance [6, 14]).
Another source is a central limit theorem of S.V. Nagaev for stationary Markov
chains (see [12]). In his approach Nagaev uses the spectral properties of a quasi-
compact Markov operator acting on some space of bounded measurable functions.
This method was used by E. Le Page to obtain more general central limit theorems,
as well as a large deviation principle for random i.i.d. products of matrices [10]. Later
P. Bougerol extended Le Page’s approach, proving similar results for Markov type
random products of matrices (see [1]).
The book of Bougerol and Lacroix [2], on random i.i.d. products of matrices, is an
excellent introduction on the subject in [1, 10]. More recently, the book of Hennion
and Hervé [7] describes a powerful abstract setting where the method of Nagaev
can be applied to derive limit theorems. It contains several applications, including to
dynamical systems and linear cocycles, that illustrate the method.
Let B be a Banach space and let L (B) denote the Banach algebra of bounded linear
operators T : B → B. Given T ∈ L (B), we denote its spectrum by σ (T ), and its
spectral radius by
ρ(T ) = lim
T n
1/n = inf
T n
1/n .
n→+∞ n≥0
(Q K )n f − f, μ 1
∞ ≤ C ρ n
f
∞ .
Defining
H0 = { f ∈ L ∞ (Σ) : f, μ = 0 },
c 1 = lim (c 1 + (Q K )n h) = lim (Q K )n f ≥ 0,
n→+∞ n→+∞
(Q K )n f − f, μ 1
∞ =
(Q K )n [ f − f, μ 1]
∞
≤ C ρ n
f − f, μ 1
∞
≤ 2 C ρ n
f
∞ .
∞,
(B3) Bα is a lattice, i.e., if f ∈ Bα then f , f ∈ Bα ,
(B4) Bα is a Banach algebra with unity 1 ∈ Bα and vα (1) = 0.
Assume also that this family is a scale of normed spaces in the sense that (see [9])
for all 0 ≤ α0 < α1 < α2 ≤ 1
(B5) Bα2 ⊂ Bα1 ⊂ Bα0 ,
(B6) vα0 ( f ) ≤ vα1 ( f ) ≤ vα2 ( f ), for all f ∈ Bα2 ,
α2 −α1 α1 −α0
(B7) vα1 ( f ) ≤ vα0 ( f ) α2 −α0 vα2 ( f ) α2 −α0 , for all f ∈ Bα2 .
The next proposition shows that an example of a scale of Banach algebras sat-
isfying (B1)–(B7) are the spaces of α-Hölder continuous functions on (Σ, d). The
norms on these spaces are defined as follows: for all α ∈ (0, 1] and f ∈ L ∞ (Σ), let
f (x) − f (y)
f
α := vα ( f ) +
f
∞ , with vα ( f ) := sup .
x,y∈Σ d(x, y)α
x= y
satisfies (B1)–(B7).
vα ( f g) ≤
f
∞ vα (g) +
g
∞ vα ( f ),
that holds for all f, g ∈ L ∞ (Σ). The monotonicity properties (B5) and (B6) are
straightforward to check. The function α → log vα ( f ) is convex. In fact, given
α1 , α2 , s ∈ [0, 1],
| f (x) − f (y)|s+(1−s)
log vs α1 +(1−s) α2 ( f ) = log sup s α +(1−s) α2
x= y d(x, y) 1
s 1−s
| f (x) − f (y)| | f (x) − f (y)|
≤ log sup sup
x= y d(x, y)α1 x= y d(x, y)α2
= s log vα1 ( f ) + (1 − s) log vα2 ( f ).
We now adopt a second set of assumptions that rule the action of the Markov and
Laplace-Markov operators associated to observed Markov systems (K , μ, ξ ) ∈ X
on the Banach algebras Bα .
Assume there exists an interval [α1 , α0 ] ⊂ (0, 1] with α1 < α20 such that for all
α ∈ [α1 , α0 ] the following properties hold:
(A1) (K , μ, −ξ ) ∈ X whenever (K , μ, ξ ) ∈ X.
(A2) The Markov operators Q K : Bα → Bα are uniformly quasi-compact and
simple. More precisely, there exist constants C > 0 and 0 < σ < 1 such that
for all (K , μ, ξ ) ∈ X and f ∈ Bα ,
Q nK f − f, μ1
α ≤ C σ n
f
α .
Q K ,z ξ ( f ξ i ) ∈ Bα and
Q K ,z ξ ( f ξ i )
α ≤ M
f
α .
Q K 1 ,z ξ1 f − Q K 2 ,z ξ2 f
∞ ≤ M
f
α dist((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 ))θ .
The interval [α1 , α0 ] will be called the range of the scale of Banach algebras. In
the fiber LDT theorem we will need to take α0 small enough to have contraction in
(A2), but at the same time we need α1 bounded away from 0 to have uniformity in this
contraction. The need for the condition α1 < α20 is explained below (see Remark 5.3).
180 5 Large Deviations for Random Cocycles
The positive constants C, σ , M, b and θ above will be called the setting constants.
Examples of contexts satisfying all assumptions (B1)–(B7) and (A1)–(A4) are
provided by the applications in Sects. 5.3.1 and 5.3.2.
The symmetry assumption (A1) allows us to reduce deviations below average to
deviations above average, thus shortening the arguments.
(A2) is the main assumption: all Markov operators Q K : Bα → Bα are quasi-
compact and simple, uniformly in (K , μ, ξ ) ∈ X. This will imply that, by possibly
decreasing b, all Laplace-Markov operators Q K ,z ξ : B α → Bα are also quasi-
compact and simple, uniformly in (K , μ, ξ ) ∈ X and z < b.
(A3) is a regularity assumption. The
operators Q K ,z ξ act continuously on Bα ,
uniformly in (K , μ, ξ ) ∈ X and z < b. Moreover, it implies that Db z →
Q K ,z ξ ∈ L (Bα ), is an analytic function.
Finally, (A4) implies that the function (K , μ, ξ ) → λ K ,ξ (z) is uniformly Hölder
continuous. Here λ K ,ξ (z) denotes the maximal eigenvalue of Q K ,z ξ .
These facts follow from the propositions stated and proved in the rest of this
subsection.
Hypothesis (A3) implies that Q K ,zξ ∈ L (Bα ), for all z ∈ Db . In particular the
function Q K ,∗ξ : Db → L (Bα ), z → Q K ,zξ , is well-defined, for every (K , μ, ξ ) ∈
X. The next proposition establishes its analyticity.
This is the first order Taylor remainder formula for h(z) = eb z at z = z 0 . To shorten
notation we write Q z for Q K ,z ξ . Replacing b by ξ(y), multiplying by f (y) K (x, dy)
and integrating over Σ we get
Q z f − Q z0 f z
z−ζ
− Q z0 ( f ξ ) = Qζ ( f ξ 2) dζ.
z − z0 z0 z − z0
Q z − Q z0
lim = Q z0 (ξ ·).
z→z 0 z − z0
1 + 2σ
Γ 0 = { w ∈ C : w = },
3
1−σ
Γ1 = { w ∈ C : w − 1 = },
3
and Rz (w) = R K ,zξ stands for the resolvent of Q K ,zξ ,
−1
Rz (w) := w I − Q K ,zξ .
∞
T0−1
C
T −1
≤
T0−1
n+1
T − T0
n = ≤ .
n=0
1−
T0−1
T − T0
1−Cε
Q 0
=
L 0 + N0
≤ 1 + C σ .
We now go through the given operators, one at a time. Assume 0 < b0 < b is
small and take z ∈ Db0 . For Q K ,z ξ , item (1) follows from assumption (A3), taking
C0 := M, while (2) follows from (A3) and Proposition 5.11 with the same constant.
For the operator Rz (w), we have
∞
Q n0
R0 (w) = w−1 (I − w−1 Q 0 )−1 = w−1
n=0
wn
∞ ∞ ∞
P0 N0n P0 N0n
= w−1 + w −1
= + .
n=0
w n
n=0
w n w − 1 n=0 wn+1
5.2 An Abstract Setting 183
/ int(Γ0 ) ∪ int(Γ1 ) implies w − 1 ≥ 1−σ
Notice also that w ∈ 3
and w ≥ 1+2σ
3
, and
hence
∞
n
P0
C σ
R0 (w)
≤ +
w − 1 w w
n=0
∞ n
3 3C 3σ 3 + 3C
≤ + = =: C1 .
1−σ 1 + 2σ n=0 1 + 2σ 1−σ
Proof (of Proposition 5.12) By Lemma 5.2 for all |z| < b and w ∈ / int(Γ0 ) ∪ int(Γ1 ),
the operator norm
Rz (w)
is uniformly bounded. This implies that the spectrum Σz
of Q K ,z ξ is contained in int(Γ0 ) ∪ int(Γ1 ), and hence we can write Σz = Σz0 ∪ Σz1
with Σzi ⊂ int(Γi ), for i = 0, 1. By the spectral theory of bounded operators on
Banach spaces (see [15, Chap. XI]) if we denote by Hz and E z the subspaces of Bα ,
respectively associated to the spectral components Σz0 and Σz1 , then for all z ∈ Db0 ,
with b0 > 0 small enough,
(a) the operators Q z , Pz , L z and Nz commute,
(b) L z f = Q z f ∈ E z , for all f ∈ E z ,
(c) Nz f = Q z f ∈ Hz , for all f ∈ Hz ,
(d) Q z = L z + Nz ,
(e) Bα = E z ⊕ Hz ,
(f) Pz is the projection to E z parallel to Hz .
For z = 0, the condition (A2) implies that the operator Q 0 |Bα is quasi-compact and
simple, with spectrum Σ00 ⊂ Dσ and Σ01 = {1}. Since 1 is a simple eigenvalue, E 0 =
1 is the space of constant functions. The operator Q 0 leaves invariant the subspace
of functions with zero average and acts on it as a contraction with spectral radius
≤ σ . Hence we must have H 0 = { f ∈ Bα : f dμ = 0 }. Thus for all f ∈ Bα ,
P0 f = ( f dμ) 1 and N0 f = Q 0 f − ( f dμ) 1. Since 1 is a simple eigenvalue
of Q 0 , a continuity argument implies that Σz1 is a singleton, i.e., Σz1 = {λ(z)}, for
all z ∈ Db . It follows easily that dim(E z ) = 1, and λ(z) = L z 1, μ/Pz 1, μ. By
perturbation theory, and Proposition 5.11, the function λ : Db0 → C is analytic.
Hence, to finish the proof of Proposition 5.12, it is now enough to establish items
(6)–(10).
Take 0 < b0 < b according to Lemma 5.2. Fixing a reference probability measure
μ0 on Σ, we can write, for all z ∈ Db ,
184 5 Large Deviations for Random Cocycles
L K ,z ξ 1, μ0
λ K ,μ,ξ (z) = . (5.6)
PK ,z ξ 1, μ0
PK ,z ξ 1, μ0 ≥ 1 −
PK ,z ξ 1 − PK ,0 1
α ≥ 1 − C0 b0 .
λ K ,μ,ξ (z) − 1 ≤ L K ,z ξ 1, μ0 − L K ,0 1, μ0
PK ,z ξ 1, μ0 PK ,0 1, μ0
L K ,z ξ 1 − L K ,0 1, μ0 C0 PK ,z ξ 1 − PK ,0 1, μ0
≤ +
1 − C0 b0 (1 − C0 b0 )2
C0 b0 C02 b0
≤ + = O(b0 ).
1 − C0 b0 (1 − C0 b0 )2
Thus, given ε >0 we can make b0 > 0 small enough so that for all (K , μ, ξ ) ∈ X,
and all z ∈ Db0 , λ K ,μ,ξ (z) − 1 < ε. This implies (6).
To prove (7), choose p ∈ N such that C σ p ≤ (σ + 2ε ) p , and make b0 > 0 small
enough so that
p ε
p C0 b0 < (σ + ε) p − (σ + ) p = O(ε).
2
We then have
p p
Nzp
≤
N0
+
Nzp − N0
p−1 p
≤ C σ p + p C0
Nz − N0
≤ C σ p + p C0 b0
ε
≤ C σ p + (σ + ε) p − (σ + ) p < (σ + ε) p .
2
This follows from (5.7), Lemma 5.2, and the algebraic relation
R K 1 ,zξ1 (w) − R K 2 ,zξ2 (w) = −R K 1 ,zξ1 (w) ◦ (Q K 1 ,zξ1 − Q K 2 ,zξ2 ) ◦ R K 2 ,zξ2 (w).
Remark 5.3 The condition α1 < α20 and the assumption (A4) are only needed to
prove item (10) of Proposition 5.12.
Theorem 5.4 Given (K 0 , μ0 , ξ0 ) ∈ X and h > (c K 0 ,ξ0 ) (0), there exist a neighbor-
hood V of (K 0 , μ0 , ξ0 ) ∈ X , C > 0 and ε0 > 0 such that for all (K , μ, ξ ) ∈ V,
0 < ε < ε0 , x ∈ Σ and n ∈ N,
+ 1
ε2
Px
Sn (ξ ) − Eμ (ξ ) ≥ ε ≤ C e− 2 h n . (5.10)
n
Remark 5.4 Averaging in x, w.r.t. μ we get for all 0 < ε < ε0 , (K , μ, ξ ) ∈ V and
n ∈ N,
186 5 Large Deviations for Random Cocycles
1 ε2
P+ Sn (ξ ) − Eμ (ξ ) ≥ ε ≤ C e− 2h n .
μ
n
Proof In fact,
n
ξ(x j )
n−1
((Q K ,zξ )n 1)(x0 ) = ez j=1 K (x j , d x j+1 ) = Ex0 e z Sn (ξ ) .
Σn j=0
The next proposition shows that c K ,ξ (z), defined in (5.9), is a limit cumulant
generating function of the process {Sn (ξ )}n≥0 . Moreover it says that the parameters
C and δn in Definition 5.16 can be chosen uniformly in X.
Thus
log Ex et Sn (ξ ) − n log λ K ,ξ (z) = log Ex et Sn (ξ ) − log λ K ,ξ (z)n
t S (ξ )
Ex e n − λ K ,ξ (z)n
≤
min{λ K ,ξ (z)n , Ex et Sn (ξ ) }
5.2 An Abstract Setting 187
Proof (of Theorem 5.4) Combine Proposition 5.13 with Corollary 5.1.
Here we prove the base-LDT and uniform fiber-LDT estimates for irreducible cocy-
cles over mixing Markov shifts. These results follow from the abstract Theorem 5.4.
To deduce Theorem 5.2 from Theorem 5.4 we specify the data (Bα ,
·
α ) and X,
and check the validity of the assumptions (B1)–(B7) and (A1)–(A4).
Consider a strongly mixing Markov system (K , μ) on the compact metric space Σ.
−
Let X − = Σ Z0 be the space of sequences in Σ indexed in the set Z−0 of non-positive
integers. Since Z−
0 is countable, the product X −
is a compact metrizable topological
space. We denote by F its Borel σ -field. The kernel K on Σ induces another Markov
on X − defined by
kernel K
( ..., x−1 ,x0 ) :=
K δ( ..., x−1 ,x0 ,x1 ) K (x0 , d x1 ).
Σ
Let P−μ denote the Kolmogorov extension of (K , μ), which is also the unique
-stationary measure. Theorem 5.4 will be applied to the Markov system ( K
K , P−μ ).
−
Consider the spaces Hα (X ) introduced in Definition (5.11). Its functions can be
regarded as measurable functions on X − .
Proof This follows essentially from Proposition 5.10. The metric space (X − , d) has
diameter 1 but it is not compact (see Remark 5.1). Hence, formally, this proposition is
not a direct consequence of Proposition 5.10. Properties (B1), (B3) and (B4) follow
from Proposition 5.4. For α = 0, the seminorm v0 measures the variation of f . Hence
H0 (X ) = L ∞ (X ), while the norm
·
0 is equivalent to
·
∞ . This proves (B2). The
remaining properties, (B5)–(B7), can be proved as in Proposition 5.10.
188 5 Large Deviations for Random Cocycles
Fix 0 < α0 ≤ 1 and 0 < L < +∞ and consider the space X of observed Markov
, P−
systems ( K − −
μ , ξ ) over the fixed Markov system ( K , Pμ ), with ξ ∈ Hα0 (X ) and
−
ξ
α0 ≤ L. This space is identified with a subspace of Hα0 (X ) and endowed with
the corresponding norm distance.
Proof We shall write Q = Q K. The first inequality follows from (c) of Proposi-
tion 5.8. For the second, notice that if k ≥ 1 then vk (Q n f ) ≤ vk+n ( f ). Indeed, for
x ) ≤ 2−k with k ≥ 1, we have
x = (xn )n≤0 and x = (xn )n≤0 in X − such that d(x,
x0 = x0 . Thus
n
(Q f )( . . . , x−1 , x0 ) − (Q n f )( . . . , x , x )
−1 0
n−1
≤ f ( . . . , x 0 , x 1 , . . . , x n ) − f ( . . . , x , x 1 , . . . , x n ) K (x j , d x j+1 )
0
Σn j=0
n−1
≤ vk+n ( f ) K (x j , d x j+1 ) = vk+n ( f ),
Σ n j=0
The next proposition shows that X satisfies (A2) with range [α1 , α] for any given
0 < α1 ≤ α. The setting constants C > 0 and 0 < σ < 1 depend on the number α1 .
5.3 The Proof of LDT Estimates 189
Proposition 5.17 If (K , μ) is strongly mixing, then given 0 < α1 < α0 there are
constants C > 0 and 0 < σ < 1 such that for all α1 ≤ α ≤ α0 , Q K : Hα (X − ) →
Hα (X − ) is quasi-compact and simple with spectral constants C and σ , i.e., for all
f ∈ Hα (X − ),
(Q K )n f − f, P−
μ 1
α ≤ C σ
f
α .
n
Note that if Fk− is the sub σ -field of F− generated by the cylinders in the coordinates
−
x−k+1 , . . . , x−1 , x0 , we have f k = E− − −
μ ( f |Fk ), and in particular Eμ ( f k ) = Eμ ( f ),
for all k ∈ N. By definition of f k ,
Q n ( f − f k )
∞ ≤
f − f k
∞ ≤ vk ( f ) ≤ 2−αk vα ( f ). (5.12)
Q n h
∞ ≤ C ρ n
h
∞ . (5.13)
Q n ( f k − E−
μ ( f ) 1)
∞ =
Q
n−k
h
∞ ≤ C ρ n−k
h
∞ (5.14)
≤Cρ n−k
Q ( f k −
k
E−
μ( f ) 1)
∞
≤ C ρ n−k
f k − E−
μ ( f ) 1
∞ ≤ 2C ρ
n−k
f
∞ .
α1 √
Setting σ = max{2− 2 , ρ} we have 0 < σ < 1. From the inequalities (5.12)
and (5.14), with k = n/2, we have
Q n f − E− −
μ ( f ) 1
∞ ≤
Q ( f − f k )
∞ +
Q ( f k − Eμ ( f ) 1)
∞
n n
≤ 2−α 2 vα ( f ) + 2Cρ 2
f
∞
n n
≤ σ n vα ( f ) + 2Cσ n
f
∞ .
vα (Q n f − E− −
μ ( f ) 1) = vα (Q ( f − Eμ ( f ) 1))
n
≤ max{
Q n f − E−
μ ( f ) 1
∞ , 2
−n α
vα ( f ) }
≤ max{ σ n vα ( f ) + 2Cσ n
f
∞ , σ 2n vα ( f ) }
= σ n va ( f ) + 2Cσ n
f
∞ .
Q n f − E−
μ ( f ) 1
α ≤ 4Cσ
f
α ,
n
Proof (of Proposition 5.15) Property (A1) is obvious. Proposition 5.17 proves (A2).
By Proposition 5.16 the operator Q K : Hα (X − ) → Hα (X − ) is bounded with norm
Q K
≤ 2. Since Hα (X − ) is a Banach algebra, given 0 < α ≤ α0 and ξ ∈
X ≡ Hα0 (X − ) ⊂ Hα (X − ), the multiplication operator Dξ : Hα (X − ) → Hα (X − ),
Dξ f := ξ f , is uniformly bounded for ξ ∈ X. Thus, because Q K ,ξ = Q K ◦ Deξ , the
map Q K ,∗ : Hα (X − ) → L (Hα (X − )), ξ → Q K ,ξ , is analytic. These considerations
imply (A3). A simple computation, using that (Hα (X − ),
·
α ) is a Banach algebra,
shows that for all f ∈ Hα (X − ) and ξ1 , ξ2 ∈ X,
Q K ,ξ1 f − Q K ,ξ2 f
α ≤ 2 e L
ξ1 − ξ2
α
f
α ,
In this section we use Theorem 5.4 to establish the fiber LDT Theorem 5.3. First we
specify the data (Bα ,
·
α ) and the metric space X. Then we check that assumptions
(B1)–(B7) and (A1)–(A4) hold.
Consider the space B∞ m (K ) of random cocycles over a Markov system (K , μ).
For each cocycle A ∈ B∞ m (K ) we define a Markov kernel on Σ × Σ × P(R ) by
m
K A (x, y, p̂) := δ(y,z,A(y,z) p̂) K (y, dz). (5.15)
Σ
We will see that (c.f. Corollary 5.2), under the assumptions of Theorem 5.3, this
kernel admits a unique K A -stationary probability measure μ A in Σ × Σ × P(Rm ).
For each A ∈ B∞ m (K ) consider the observable ξ A : Σ × Σ × P(R ) → R
m
www.Ebook777.com
192 5 Large Deviations for Random Cocycles
X := { (K A , μ A , ±ξ A ) : A ∈ B∞
m (K ), A irreducible, L 1 (A) > L 2 (A) }.
Next we define the scale of Banach algebras. Recall the following projective
distance (see (2.3))
p ∧ q
δ( p̂, q̂) := ,
p
f
α := vα ( f ) +
f
∞ , (5.17)
f (x, y, p̂) − f (x, y, q̂)
vα ( f ) := sup . (5.18)
x,y,∈Σ δ( p̂, q̂)α
p̂=q̂
For simplicity of notation from now on we will simply write p for both the vector
and the corresponding projective point.
Consider the normed space Hα (Σ × Σ × P(Rm )) of all functions f ∈ L ∞ (Σ ×
Σ × P(Rm )) such that vα ( f ) < +∞, endowed with the norm (5.17).
vα ( f g) ≤
f
∞ vα (g) +
g
∞ vα ( f ), f, g ∈ L ∞ (Σ).
The monotonicity properties (B5) and (B6) are straightforward to check. The
assumption (B7) follows from the convexity of the function α → log vα ( f ), whose
proof is analogous to that of Proposition 5.10.
Given A ∈ B∞ ∞
m (K ), consider the linear transformation Q A : L (Σ × Σ ×
P(Rm )) → L ∞ (Σ × Σ × P(Rm )) defined by
(Q A f )(x, y, p) := f (y, z, A(y, z) p) K (y, dz). (5.19)
Σ
(a)
A(±n)
∞ ≤ max{
A
∞ ,
A−1
∞ }n .
(b)
A(n) − B (n)
∞ ≤ n max{
A
∞ ,
B
∞ }n−1
A − B
∞ .
n−1
(n) (n)
A −B = (A( j) ◦ T n− j )(A ◦ T n−1− j − B ◦ T n−1− j ) B (n−1− j) .
j=0
vα (Q nA f ) ≤ καn (A) vα ( f ).
Hence
Ex f (en , A(n) p) − f (en , A(n) q)
vα (Q nA f) = sup
x∈Σ, p=q δ( p, q)α
Ex f (en , A(n) p) − f (en , A(n) q)
≤ sup
x∈Σ, p=q δ( p, q)α
! α "
δ(A(n) p, A(n) q)
≤ vα ( f ) sup Ex
x∈Σ, p=q δ( p, q)
= vα ( f ) καn (A).
In particular,
lim κ n (A)1/n = inf{ καn (A)1/n : n ∈ N }.
n→+∞ α
Proof We write as before Mn = A(n) . Recall that given M ∈ GL(m, R), the quantity
(M) := max{log
M
, log
M −1
} is sub-multiplicative, in the sense that for any
matrices M1 , M2 ∈ GL(m, R), (M1 M2 ) ≤ (M1 ) + (M2 ). By Lemma 2.12, given
x ∈ Σ, and p = q in P(Rm ),
α
δ(Mn p, Mn q) δ(Mn p, Mn q)
Ex = Ex exp α log ≤ Ex e4 α (Mn ) .
δ( p, q) δ( p, q)
5.3 The Proof of LDT Estimates 195
If 0 < α ≤ 1
4n
, setting c := max{log
A
∞ , log
A−1
∞ }
Ex e4 α (Mn ) ≤ e4nαc ≤ ec = max{
A
∞ ,
A−1
∞ }.
1
lim Ex (log
A(n) p
) = L 1 (A),
n→+∞ n
Ex log ≤ Ex log
n δ( p, q) n
Mn p
Mn q
p ∧ q
1
(Mn p) ∧ (Mn q)
p
q
≤ Ex log
n
p ∧ q
Mn p
Mn q
1 1 1
≤ Ex log
∧2 A(n)
− Ex log
A(n) p
− Ex log
A(n) q
,
n n n
and the right hand side converges to L 1 + L 2 − 2 L 1 = L 2 − L 1 < 0. By Lemma 5.8,
we have
1 δ(Mn p, Mn q)
lim sup sup Ex log ≤ L 2 − L 1 < 0.
n→+∞ x∈Σ, p=q n δ( p, q)
Hence taking n large enough such that n (L 2 − L 1 ) < −1 the Lemma follows.
vα (Q nB f ) ≤ C σ n vα ( f ),
B −1
∞ ≤ C. By Lemma 5.4
B (±n)
∞ ≤ C n for all B ∈ V and n ∈ N. Thus,
by Lemmas 2.13 and 5.4(b), there exists a polynomial expression C(g1 , g2 ), with
degree < 11 in the variables
g1
,
g2
,
g1−1
and
g2−1
, such that
! α "
n δ(A(n) p, A(n) q) α δ(B (n) p, B (n) q)
κ (A) − κ n (B) ≤
α α sup Ex −
x∈Σ, p=q δ( p, q) δ( p, q)
≤ α C(A(n) , B (n) )
A(n) − B (n)
∞ ≤ α C 11 n
A(n) − B (n)
∞ ≤ α n C 12n−1
A − B
∞ .
Let Mn = A(n) and note that (Mn ) ≤ n log C. We claim that καn00 (A) < 1 for some
n 0 ∈ N and 0 < α0 ≤ 1 small enough. We will make use the following inequality
x 2 |x|
ex ≤ 1 + x + e .
2
Choose n 0 ∈ N as given by Lemma 5.9. For all x ∈ Σ, p = q in P(Rm ),
δ(Mn 0 p, Mn 0 q) α δ(Mn 0 p, Mn 0 q)
Ex = Ex exp α log
δ( p, q) δ( p, q)
! "
δ(Mn 0 p, Mn 0 q) α 2 δ(Mn 0 p, Mn 0 q) δ(Mn 0 p, Mn 0 q) α
≤ Ex 1 + α log + log2
δ( p, q) 2 δ( p, q) δ( p, q)
α2 # $
≤1−α+ Ex 16 (Mn 0 )2 exp(4 α (Mn 0 )) ≤ 1 − α + O(α 2 ).
2
The last inequality follows because Ex 16 (Mn 0 )2 exp(4 α (Mn 0 )) is finite and
uniformly bounded in x and 0 < α ≤ 1 by the constant 16 n 20 (log C)2 C 4 n 0 α .
Taking α > 0 sufficiently small the right-hand-side above becomes less than 1,
which implies that καn 0 (A) < 1. Hence, we can choose 0 < α1 < α20 and 0 < ρ < 1
such that for all α1 ≤ α ≤ α0 , καn 0 (A) ≤ ρ.
Next, we extend this inequality to all cocycles B ∈ V.
Pick ρ ∈ (ρ, 1) and choose δ > 0 such that α0 n 0 C 12n 0 −1 δ < ρ − ρ. Make the
neighborhood V small enough so that
A − B
∞ < δ for all B ∈ V. Then, using the
modulus of continuity for καn (B), for all B ∈ V and α1 ≤ α ≤ α0 ,
n
κ 0 (A) − κ n 0 (B) < ρ − ρ,
α α
5.3 The Proof of LDT Estimates 197
which implies
καn 0 (B) ≤ καn 0 (A) + καn 0 (A) − καn 0 (B) < ρ .
j
By Lemma 5.7, κα (B) ≤ C for all B ∈ V, 0 < α ≤ 4 1n 0 and 0 ≤ j ≤ n 0 . Shrinking
if necessary the constants α1 and α0 above, we may assume that α0 ≤ 4 1n 0 . Thus,
because the sequence {καn (B)}n≥0 is sub-multiplicative, letting σ = (ρ )1/n 0 we have
καn (B) ≤ C σ n for all B ∈ V, n ∈ N and α1 ≤ α ≤ α0 where C = C (α1 , C) < ∞.
The proposition follows then from the inequality in Lemma 5.5.
Q nB f − f, μ B 1
α ≤ C σ n
f
α .
Proof The argument below is an adaptation of the proof of Theorem 3.7 in [1].
Take the neighborhood V, and the constants 0 < α1 < α20 , C > 0 and 0 < σ < 1
given by Proposition 5.21. Enlarging the constants C > 0 and 0 < σ < 1 we
can assume that the conditions of Definition 5.6 are also satisfied with ρ = σ . By
Lemma 5.5, given B ∈ V and any K B -stationary measure ν B ,
vα (Q nB f − f, ν B 1) = vα (Q nB f ) = vα ( f ) καn (B) ≤ C σ n
f
α .
Q nB f − f, ν B 1
∞ ≤ C σ n
f
α .
(2)
The sequence {TB, f }≥0 is relatively compact in C(P(Rm )). Hence, the set S f of
its limit points in (C(P(Rm )),
·
∞ ) is non-empty. Take any g ∈ S f and any K B -
stationary probability measure ν B . We claim that g = f, ν B 1.
From (5.21) we have for all m ∈ N,
B f − g
∞ ≤ 3 C σ
f
α .
Q 2m m
B f − f, ν B 1
∞ ≤ 3 C σ
f
α
Q 2m m
∀ m ∈ N.
Claim (2): Defining ϕm,q (x) := Ex f (em , B (m) q) , because (K , μ) is strongly mix-
ing on L ∞ (Σ) we have
(1)
(T (2)
f − TB,m f )(x, q) = Ex f (en , (B (m) ◦ T n−m ) q) − Eμ f (em , B (m) q)
B,n,m
% &
= Ex Een−m f (en , (B (m) ◦ T n−m ) q) − Eμ f (em , B (m) q)
= Ex ϕm,q (en−m ) − ϕm,q dμ = (Q n−m
K ϕm,q )(x) − ϕm,q , 1 ≤ C σ
n−m
.
Proof In the proof of Proposition 5.22 we have shown that given a function f ∈
(2)
Hα (Σ × P(Rm )), if we denote by S f the set of limit points of {TB, f }≥0 , then
S f = { f, ν B 1} for any K B -stationary measure ν B .
Hence, given any other K B -stationary measure μ B , and f ∈ Hα (Σ × P(Rm )),
we have f, ν B = f, μ B . Since Hα (Σ × P(Rm )) is dense in L ∞ (Σ × P(Rm )), it
follows that ν B = μ B .
Proof (of Proposition 5.18) Assumption (A1) follows from the definition of X.
Like the Markov operator Q A defined in (5.19), the Laplace-Markov operator
Q A,z leaves invariant the subspaces Hα (Σ × P(Rm )), for all small enough α > 0.
Choose 0 < α1 < α0 ≤ 1 according to Proposition 5.22. Assumption (A2) is a
consequence of this proposition.
Assumption (A3) is automatically satisfied because
A
∞ < ∞ and
A−1
∞ <
∞ which imply that ξ A ∈ Hα (Σ ×P(Rm )) for all α > 0. Note that Q A,z = Q A ◦Dez ξa ,
where Dez ξa denotes the multiplication operator by e z ξa . This is a bounded operator
because Hα (Σ × P(Rm )) is a Banach algebra containing the function e z ξa .
Finally the next lemma proves (A4).
Q A,z f − Q B,z f
∞ ≤ C2 d∞ (A, B)α
f
α .
Proof A simple computation shows that for all z ∈ C with Re z ≤ b, and all
A, B ∈ GL(d, R),
A p
z −
B p
z ≤ b max{
A
b−1 ,
B
b−1 }
A − B
.
Hence
(Q A,z f − Q B,z f )(x, p) ≤ Ex
A p
z f (e1 , A p) −
B p
z f (e1 , B p)
≤
f
∞ Ex
A p
z −
B p
z +
B
b∞ Ex f (e1 , A p) − f (e1 , B p)
≤ b max{
A
∞ b−1 ,
B
b−1 }
A − B
f
+
B
b v ( f ) E δ(A p, B p)α
∞ ∞ ∞ ∞ α x
b−1 ,
B
b−1 }
A − B
f
≤ b max{
A
∞ ∞ ∞ ∞
+
B
b∞ vα ( f )
A − B
α∞ ≤ C2
f
α d∞ (A, B)α ,
where C2 = max{
B
b∞ , b
B
b−1
∞ , b
A
∞ }.
b−1
Choose the canonical basis {e1 , . . . , em } of Rm and consider the following norm
·
on the space of matrices Mat(m, R),
M
:= max1≤ j≤m
M e j
. Since this norm is
equivalent to the operator norm, for all B ∈ V, p ∈ P(Rm ) and n ∈ N,
B (n) p
≤
B (n)
B (n)
= max
B (n) e j
.
1≤ j≤m
In this last section we use the LDT estimates (Theorems 5.2 and 5.3) to derive the con-
tinuity of the Lyapunov exponents and of the Oseledets filtration/decomposition. We
give some simple generalizations of the continuity results and explain the method’s
limitations regarding the continuity of the LE in the reducible case.
For the reader’s convenience, we briefly review the relevant definitions as we explain
how Theorems 3.1, 4.7 and 4.8 are applicable to the context of this chapter.
Proof (of Theorem 5.1) Let (K , μ) be a strongly mixing Markov system, and consider
the associated Markov shift (X, Pμ , T ).
The collection C = {(Im∞ (K ), d∞ )}m∈N of totally irreducible cocycles over
(X, Pμ , T ) is a space of measurable cocycles in the sense of Definition 3.1.
Consider the space of LDT parameters P = N × E × I, where E is the set of
constant deviation functions ε(t) ≡ ε, 0 < ε < 1, and we use the set of exponential
functions I = { ι(t) ≡ M e−c t : M < ∞, c > 0 } to measure the deviation sets.
Define Ξ to be the set of observables ξ : X → R which depend only on finitely
many coordinates. Finally, take p = ∞.
We now check the four assumptions of Theorems 3.1, 4.7 and 4.8.
Given a cocycle A ∈ Im∞ (K ) and N ∈ N, let F N (A) be the lattice generated by
the sets {x ∈ X :
A(n) (x)
≤ c} and {x ∈ X :
A(n) (x)
≥ c}.
Recall that Ξ is said to be compatible with a cocycle A ∈ B∞ m (K ) when for any
N ∈ N, F ∈ F N (A), ε > 0, there is ξ ∈ Ξ such that 1 F ≤ ξ and X ξ dμ ≤ μ(F)+ε.
In our setting, the set Ξ is compatible with all cocycles A ∈ B∞
m (K ) because for any
www.Ebook777.com
202 5 Large Deviations for Random Cocycles
set F ∈ F N (A) its indicator function 1 F depends only on finitely many coordinates,
i.e., 1 F ∈ Ξ . Hence the compatibility inequality holds with ε = 0.
Recall that ξ ∈ Ξ is said to satisfy a base-LDT estimate if for every ε > 0 there
is an LDT parameter (n 0 , ε, ι) ∈ P such that for all n ≥ n 0 we have ε(n) ≤ ε and
n−1
{x ∈ X : 1 ξ(T j x) − ξ d x > ε(n)} < ι(n).
n j=0 X
for all cocycles B that are close enough to A (in the given topology on the space of
cocycles) and for all scales n ≥ 1. In our setting the L p -boundedness is automatic
because p = ∞ and the cocycle functions B and B −1 are bounded.
Recall that a cocycle A is said to satisfy a uniform fiber-LDT if for every ε > 0
there are δ > 0 and an LDT parameter (n 0 , ε, ι) ∈ P such that for any cocycle B
with dist(B, A) < δ and for all n ≥ n 0 , we have ε(n) ≤ ε and
{x ∈ X : 1 log
B (n) (x)
− L (n) (B) > ε(n)} < ι(n).
1
n
Given A ∈ Im∞ (K ) such that L 1 (A) > L 2 (A), by Theorem 5.3 the cocycle A
satisfies uniform fiber-LDT estimates w.r.t. P.
A simple computation shows that the modulus of continuity associated to the
choice of deviation function sets E and I above corresponds to Hölder continuity.
Hence, this theorem follows from Theorems 3.1, 4.7 and 4.8.
A−1
∞ < ∞. Note that the iterates of A are
We identify B∞
m (K , p) as a space of functions A : Σ
p+1
→ GL(m, R). Each such
function determines a locally constant cocycle over the Markov shift (X, Pμ , T ).
Given A ∈ B∞ m (K , p), we define  : Σ × Σ → GL(m, R)
p p
 (x0 , . . . , x p−1 ), (y0 , . . . , y p−1 ) := A(x0 , . . . , x p−1 , y p−1 ).
In particular, all conclusions above on the continuity of the LE, the Oseledets
filtration, and the Oseledets decomposition, apply to irreducible and locally constant
cocycles over strongly mixing Markov and Bernoulli shifts.
The abstract setting developed in Sect. 5.2 is general enough to deal with a cocycle
having singularities, i.e., points x ∈ X where the matrix A(x) is singular. Consider
the family of spaces Bam (K ), with 0 < a < ∞, consisting of all bounded measurable
functions A : Σ × Σ → GL(m, R) such that for some C > 0 and all x ∈ Σ,
ηaA (x) :=
A(x, y)−1
a K (x, dy) ≤ C.
Σ
da (A, B) :=
A − B
∞ +
ηaA − ηaB
∞ .
We need the irreducibility assumption in order to prove uniform fiber LDT estimates
in Theorem 5.3. The proof exploits the fact that for irreducible cocycles there is
some Banach algebra of measurable functions, independent of the cocycle, where
the associated Laplace-Markov operators act as quasi-compact and simple operators
(see Proposition 5.22). For reducible cocycles this fact may still be true, and it could
eventually lead to fiber LDT estimates. However, the Banach algebra would have to
be tailored to the cocycle, and hence the scheme of proof presented here would not
provide the required uniformity.
References 205
References
1. P. Bougerol, Théorèmes limite pour les systèmes linéaires à coefficients markoviens. Probab.
Theory Related Fields 78(2), 193–221 (1988). MR 945109 (89i:60122)
2. P. Bougerol, J. Lacroix, Products of random matrices with applications to Schrödinger opera-
tors, Progress in Probability and Statistics, vol. 8 (Birkhäuser Boston Inc., Boston, MA, 1985).
MR 886674 (88f:60013)
3. J.L. Doob, Stochastic Processes (Wiley Classics Library, Wiley, New York, 1990), Reprint of
the 1953 original, A Wiley-Interscience Publication. MR 1038526 (91d:60002)
4. H. Furstenberg, H. Kesten, Products of random matrices. Ann. Math. Stat. 31, 457–469 (1960).
MR 0121828 (22 #12558)
5. H. Furstenberg, Noncommuting random products. Trans. Am. Math. Soc. 108, 377–428 (1963).
MR 0163345 (29 #648)
6. Y. Guivarc’h, A. Raugi, Frontière de Furstenberg, propriétés de contraction et théorèmes de
convergence. Z. Wahrsch. Verw. Gebiete 69(2), 187–242 (1985). MR 779457 (86h:60126)
7. H. Hennion, L. Hervé, Limit Theorems for Markov Chains and Stochastic Properties of Dynam-
ical Systems by Quasi-compactness, Lecture Notes in Mathematics, vol. 1766 (Springer, Berlin,
2001)
8. A.S. Kechries, Kechries, Classical Descriptive Set Theory, Graduate Texts in Mathematics,
vol. 156 (Springer, New York, 1995)
9. S.G. Krein, Yu.I. Petunin, Scales of Banach spaces. Russ. Math. Surv. 21(2), 85 (1966)
10. É. Le Page, Théorèmes limites pour les produits de matrices aléatoires, Probability measures
on groups (Oberwolfach, 1981), Lecture Notes in Mathematics, vol. 928 (Springer, Berlin,
1982), pp. 258–303. MR 669072 (84d:60012)
11. E. Lukacs, Characteristic Functions (Hafner Publishing Co., New York, 1970), 2nd edn. revised
and enlarged. MR 0346874 (49 #11595)
12. S.V. Nagaev, Some limit theorems for stationary Markov chains. Teor. Veroyatnost. i Primenen.
2, 389–416 (1957). MR 0094846 (20 #1355)
13. K. Petersen, Ergodic Theory, Cambridge Studies in Advanced Mathematics, vol. 2 (Cambridge
University Press, Cambridge, 1983). MR 833286 (87i:28002)
14. A. Raugi, J. Rosenberg, Fonctions harmoniques et théorèmes limites pour les marches aléatoires
sur les groupes, Bulletin de la Société mathématique de France. Mémoire, no. 54, Société
mathématique de France (1977)
15. F. Riesz, B. Szőkefalvi-Nagy, Functional Analysis, Ungar (1955)
16. V.N. Tutubalin, Limit theorems for a product of random matrices. Teor. Verojatnost. i Primenen.
10, 19–32 (1965). MR 0175169 (30 #5354)
Chapter 6
Large Deviations for Quasi-Periodic Cocycles
Abstract We derive large deviations type estimates for linear cocycles over an
ergodic multifrequency torus translation. These models are called quasi-periodic co-
cycles. We make the following assumptions on the model: the translation vector
satisfies a generic Diophantine condition, and the fiber action is given by a matrix
valued analytic function of several variables which is not identically singular. The
LDT estimates obtained here depend on some uniform measurements on the cocycle.
The general results derived in the previous chapters regarding the continuity prop-
erties of the Lyapunov exponents and of the Oseledets filtration and decompositions
are then applicable. In particular we obtain local weak-Hölder continuity of these
quantities in the presence of gaps in the Lyapunov spectrum. The main new feature of
this work is allowing a cocycle depending on several variables to have singularities,
i.e. points of non invertibility. This requires a careful analysis of the set of zeros of
certain analytic functions of several variables and of the singularities (i.e. negative
infinity values) of pluri-subharmonic functions related to the iterates of the cocycle.
A refinement of this method in the one variable case leads to a stronger LDT estimate
and in turn to a stronger, nearly-Hölder modulus of continuity of the LE, Oseledets
filtration and Oseledets decomposition.
We introduce the quasi-periodic cocycles model and describe our assumptions on it.
We then formulate the main statements and relate them to recent results for similar
models.
Let T = R/Z be the one variable torus, which we may regard as the unit circle in the
complex plane. We use the notation e(x) = e2πi x and in fact we write f (x) instead
of f (e(x)) whenever f is a function on T.
t
k · ω ≥ (6.1)
|k|d+δ0
for some t > 0, δ0 > 0 and for all k ∈ Zd \ {0}, where for any real number x we
denote x := mink∈Z x − k .
Note that if δ0 is fixed and if for every t > 0, DCt denotes the set of frequency
vectors satisfying the condition (6.1) above, then the set DC := ∪t>0 DCt has full
measure.
Since A(x) is analytic on Td , it has an extension A(z) to Ard = Ar × · · · × Ar ,
where Ar = {z ∈ C : 1 − r < z < 1 + r } is the annulus of width 2r . Note that
the iterates A(n) (x) := A(x + (n − 1) ω) . . . A(x + ω) A(x) of the cocycle are also
analytic functions on Ard .
We denote by Crω (Td , R) the Banach space of real valued analytic functions on
Ar with continuous extension up to the boundary and norm f r := supz∈Ard f (z).
d
For every integer m ≥ 1, let Crω (Td , Mat(m, R)) be the vector space of matrix
valued analytic functions on Ard , with continuous extension up to the boundary.
Endowed with the norm Ar := supz∈Ard A(z), it is a Banach space.
In a previous work (see [7]) we studied GL(m, R)-valued analytic cocycles. Here
we will allow our cocycles to have singularities (i.e. points of non-invertibility), as
long as they are not identically singular (which in particular ensures that all Lyapunov
exponents are finite).
Let us then define Cm to be the set of cocycles A ∈ Crω (Td , Mat(m, R)) with
det[A(x)] ≡ 0. This condition implies in particular that all LE are finite.
The set Cm is open in Crω (Td , Mat(m, R)) and we let dist(A, B) := A − Br
be the induced distance on it. The collection C := {(Cm , dist)}m≥1 is the space of
cocycles for our quasi-periodic model.
For the reader’s convenience we briefly recall some definitions and notations
regarding the Lyapunov exponents, Oseledets filtrations and decompositions of a
cocycle A in a space of cocycles Cm .
6.1 Introduction and Statements 209
1
Λ j (A) := lim log∧ j A(x) for -a.e. x ∈ Td .
n→∞ n
On the space D(Td , Rm ) we consider the coarsest topology that makes the sets
D⊃τ (Td , Rm ) open, and the pseudo-metrics dist τ continuous.
We are ready to state a general result on the continuity of the LE, the Oseledets
filtration and the Oseledets decomposition for quasi-periodic cocycles.
Theorem 6.1 Assume that the translation ω ∈ DCt for some t > 0, and let m ≥ 1.
Then all Lyapunov exponents L j : Cm → R, with 1 ≤ j ≤ m, the Oseledets
filtration F : Cm → F(Td , Rm ), and the Oseledets decomposition E · : Cm →
D(Td , Rm ), are continuous functions of the cocycle A ∈ Cm .
Moreover, if A ∈ Cm has a τ -gap pattern then the functions Λτ , F τ and E ·τ are
weak-Hölder continuous in a neighborhood of A.
In the one-variable case d = 1, and for translations that satisfy a slightly stronger
(but still generic) Diophantine condition, we obtain a stronger modulus of continuity
(see Sect. 6.5).
Remark 6.1 The gap pattern hypothesis in Theorem 6.1, which for SL(2, R)-valued
cocycles simply means positivity of the top LE, is a necessary assumption for ob-
taining a modulus of continuity (see Chap. 8 in J. Bourgain’s monograph [3] and the
preamble to Sect. 2 in Ávila’s paper [1]). Moreover, according to a private conversa-
tion of the second author with Qi Zhou, it seems that even in the presence of a gap
pattern, an arithmetic condition on the frequency ω is also a necessary assumption
for proving Hölder or weak-Hölder continuity of the LE.
A more detailed review of related results was given in Sect. 1.6 of the introductory
Chap. 1. The reader may also consult the recent surveys [6, 14]. Here we focus mainly
on recent work on quasi-periodic models and on their difference with our results.
In some sense, the strongest result on continuity of the Lyapunov exponents
for quasi-periodic cocycles in the one-frequency translation (d = 1) case is due
to Ávila et al. [2]. The space of cocycles considered in this paper is the whole
Crω (T1 , Mat(m, C)) (hence identically singular cocycles are not excluded) and the
authors prove joint continuity in cocycle and frequency at all points (A, ω) with
ω irrational. A previous work of Jitomirskaya and Marx [13] established a similar
result for Mat(2, C)-valued cocycles, using a different approach. We note here that
both approaches rely crucially on the convexity of the top Lyapunov exponent of the
complexified cocycle, as a function of the imaginary variable, by establishing first
continuity away from the torus. This method immediately breaks down in the several
variables (d > 1) case which we treat here.
The approaches of [2, 13] are independent of any arithmetic constraints on the
translation frequency ω and they do not use large deviations. However, the results are
not quantitative, in the sense that the they do not provide any modulus of continuity of
the Lyapunov exponents. All available quantitative results, from the classic result of
6.1 Introduction and Statements 211
Goldstein and Schlag [9] to more recent results such as [7, 20, 21, 23] or our current
work, use some type of large deviations, whose derivation depends upon imposing
appropriate arithmetic conditions on ω.
We note that in the (more particular) context of Schrödinger cocycles, joint con-
tinuity in the energy parameter and the frequency translation was proven for the one
variable d = 1 case by Bourgain and Jitomirskaya [5] and for the several variables
d > 1 case by Bourgain [4]. Both papers used weaker versions of large deviation
estimates, proven under weak arithmetic (i.e. restricted Diophantine) conditions on
ω, although eventually the results were made independent of any such restrictions.
Continuity properties of the Lyapunov exponents were also established for certain
non-analytic quasi-periodic models (see [15, 16, 22]).
In this chapter we are dealing with both a base dynamics given by a translation
on the several variables torus and a fiber action which has singularities.
Our approach is based in an essential way upon establishing certain uniform esti-
mates on analytic functions of several variables and on pluri subharmonic functions.
The issue of singularity is especially delicate for several variables functions.
One obstacle, for instance, is the fact that an analytic function of several variables
may vanish identically along hyperplanes, while not being globally identically zero.
Related to this, a pluri subharmonic function may be −∞ along hyperplanes, while
not being globally −∞.
A crucial tool in our analysis is Theorem 6.3, which shows that the obstacle
described above for an analytic function can be removed with an appropriate change
of coordinates. Another crucial tool in our analysis is the observation that for the
pluri subharmonic functions corresponding to iterates of a cocycle, while they may
have singularities as described above, these singularities can be captured by certain
analytic functions.
Most of the work in this chapter is devoted to proving a uniform LDT estimate
for iterates of the cocycle. Uniform LDT estimates for cocycles with singularities
were obtained before by Jitomirskaya and Marx [12] for Mat(2, C)-valued cocycles.
Again, the approach in [12] is one-variable specific.
Let us now phrase the large deviation type estimate obtained in this chapter.
Theorem 6.2 Given A ∈ Cm and ω ∈ DCt , there are constants δ = δ(A) > 0,
n 0 = n 0 (A, t) ∈ N, C = C(A) < ∞, a = a(d) > 0 and b = b(d) > 0 such that if
B − Ar ≤ δ and n ≥ n 0 then
{x ∈ Td : 1 logB (n) (x) − L (n) (B) > C n −a } < e−n b . (6.2)
1
n
Once the above LDT is established, we simply need to verify that we are in the
context of the abstract continuity Theorem 1.6 in Chap. 3.
We note that the above LDT estimate is of independent interest. Such estimates
have been widely used in the study of discrete quasi-periodic Schrödinger opera-
tors, to establish positivity of Lyapunov exponents, estimates on Green’s functions,
Free ebooks ==> www.Ebook777.com
212 6 Large Deviations for Quasi-Periodic Cocycles
continuity of the integrated density of states and spectral properties (such as Ander-
son localization) for such operators (see Bourgain’s monograph [3], see also [15, 16]
and references therein).
The LDTs proven here, along with the other technical analytic tools, may then
prove useful for future projects on topics in mathematical physics related to larger
classes of discrete, quasiperiodic operators.
The rest of this chapter is organized as follows. In Sect. 6.2 we prove general
uniform estimates on analytic and pluri subharmonic functions. These abstract results
are then applied in Sect. 6.3 to quantities related to cocycles iterates, leading to the
proof of the LDT. In Sect. 6.4 we explain how our system satisfies the assumptions
of the general criterion in Chap. 3. Finally, in Sect. 6.5 we show that in the one-
variable case, the LDT proven in Sect. 6.3 can be used in an inductive argument that
eventually leads to a sharper LDT, and in turn, to a stronger modulus of continuity
for the Lyapunov exponents.
Lemma 6.1 Let f (x) ∈ Crω (Td , R) such that f (x) ≡ 0. Then there are constants
δ = δ( f ) > 0, S = S( f ) < ∞ and b = b( f ) > 0 such that if g(x) ∈ Crω (Td , R)
with g − f r < δ then
{x ∈ Td : g(x) < t} < S t b for all t > 0. (6.3)
Proof We may assume that f (x) is not constant, otherwise f (x) ≡ C, C > 0 and
(6.3) is then obvious.
www.Ebook777.com
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 213
where, according to the last line of its proof (see also Remark 5.1) S = S( f ) ∼
A( f ) · m( f ) and b = b( f ) = 3m(1 f ) .
Therefore, in order to obtain the uniform estimate (6.3), all we need to show is
that the above constants m = m( f ), c = c( f ), A = A( f ) depend uniformly on the
function f .
Indeed, let g ∈ Crω (T2 , R) such that g − f r < δ. By analyticity, for some
constant B = B( f ) depending only on α and r ,
∂ α g − ∂ α f 0 ≤ B δ,
Lemma 6.2 Let f be a bounded function satisfying the Łojasiewicz inequality with
constants S, b:
{x ∈ Td : f (x) < t} < S t b for all t > 0. (6.6)
Then
log f L 2 (Td ) ≤ C, (6.7)
Proof The argument is straightforward. From (6.6), the set {x : f (x) = 0} has zero
measure. Split the rest of the phase space into
E :={x : f (x) ≥ 1},
1 1
E n :={x : n+1 ≤ f (x) ≤ n }
2 2
for all n ≥ 0.
If x ∈ E, then log
f (x) ≤ log f ∞ < ∞.
If x ∈ E n , then log f (x) n + 1.
b n
Moreover, from (6.6), E n < S 21n = S 21b .
Then
∞
2 2
log f 2L 2 (Td ) = log f + log f
E n=0 E n
∞
2 (n + 1)2
≤ log f ∞ + S
n=0
(2b )n
2 22b (2b + 1)
= log f ∞ + S .
(2b − 1)3
We then conclude:
log f L 2 (Td ) log f ∞ + S(2b − 1)−3/2 .
Remark 6.2 The previous two lemmas imply that if f ∈ Crω (Td , R), f ≡ 0, then
there are constants δ = δ( f ), C = C( f ) < ∞ such that
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 215
logg L 2 (Td ) ≤ C
Proof Just use (in each variable) the translation invariance of the Lebesgue measure
on T.
Proof The assumption implies that for all 1 ≤ j ≤ d and for all x1 , . . . , x j−1 ,
x j+1 , . . . , xd ∈ T, the analytic function
is not identically zero. Since clearly for all 1 ≤ j ≤ d, the set of functions
Theorem 6.3 For any analytic function f ∈ Crω (Td , R) with f ≡ 0, there are
δ = δ( f, r ) > 0, C = C( f, r ) < ∞ and there is a matrix M ∈ SL(d, Z) such that for
any g ∈ Crω (Td , R) with f − gr < δ, and for any x1 , . . . , x j−1 , x j+1 , . . . , xd ∈ T
with 1 ≤ j ≤ d,
logg ◦ M(x1 , . . . , x j−1 , · , x j+1 , . . . , xd ) L 2x (T) ≤ C.
j
In other
words, up to some linear change of coordinates in the torus Td , the functions
2
log g with g near f are uniformly separately L -bounded.
Given δ > 0, define Σδ to be the set of all k ∈ Zd such that any geodesic circle
parallel to the vector through k is δ-dense in Td .
Proposition 6.1 Given δ > 0, there is a matrix M ∈ SL(d, Z) such that every
column of M is in Σδ .
Proof Take any matrix M ∈ SL(d, Z) with non-negative entries which is primi-
tive, and has a characteristic polynomial p M (λ) = det(M − λI ) irreducible over Z
(see Lemma 6.5). Then M has a dominant eigenvector ω ∈ int(Rd+ ). Consider the
canonical projection π : Rd → Td and define H = π(ω) (topological closure
in Td ), where ω = { t ω : t ∈ R }. Define also h = π −1 (H ). The set H is a
compact connected subgroup of Td , while h is a linear subspace of Rd , the Lie alge-
bra of H . The group H is invariant under the torus automorphism φ M : Td → Td ,
φ M (x) = M x (mod Zd ), and hence (by restriction and quotient) the map φ M induces
the toral automorphisms φ H : H → H and φ H : Td /H → Td /H . Thus M h = h,
and the linear maps of these toral automorphisms at the level of Lie algebras are
Φh : h → h, Φh (x) = M x, and Φ h : Rd /h → Rd /h, Φ h (x + h) = M x + h. The
characteristic polynomials of these linear automorphisms have integer coefficients
because they are associated with toral automorphisms. Finally, because h is invariant
under M, the characteristic polynomial p M (λ) factors as the product of the charac-
teristic polynomials of Φh and Φ h . Since the polynomial p M (λ) is irreducible, we
conclude that H = T , which implies that the line spanned by ω is dense in Td .
d
Lemma 6.5 The matrix Md is primitive and its characteristic polynomial is irre-
ducible over Z.
Proof Computing the characteristic polynomial of the matrix Md with the Laplace
determinant rule, we obtain
and since the right-hand-side has all entries positive, it follows that Md is primitive.
Writing μ = λ − 1, we get (λ − 1)d − λ = μd − μ − 1. Hence the irreducibility
of pd (λ) is equivalent to that of μd − μ − 1, which was established to hold for every
d ≥ 2 by Selmer (see Theorem 1 in [19]).
Remark 6.3 There is an alternative argument that does not require a change of coor-
dinates, but which is technically much more complicated. It involves proving, using
the uniform Łojasiewicz inequality (6.3), that for any g ∈ Crω (Td , R) which is close
enough to a function f ∈ Crω (Td , R) with f ≡ 0, and for any n 1,
logg(x1 , . . . , x j−1 , · , x j+1 , . . . , xd ) L 2x (T) ≤n
j
holds for all (x1 , . . . , x j−1 , x j+1 , . . . , xd ) ∈ Td−1 outside of a set of measure <
e−cn .
1/2
Remark 6.4 Besides being a very helpful technical tool in this chapter, the fact
that after a change of coordinates, a two variables analytic function which is not
identically zero, will not vanish identically along any horizontal or vertical line,
could be of independent use elsewhere.
We note here, for instance, that it can be used to derive the Łojasiewicz inequality
in two variables from the one-variable statement. That is because after a change of
coordinates, the one-variable statement is applicable along any horizontal line, hence
via a compactness argument and Fubini, the two-variables statement follows.
Of course, once one has a uniform Łojasiewicz inequality in one variable (see
[15] or [12]), this argument also provides a uniform statement in two-variables.
The several (instead of two) variables situation is similar.
is subharmonic in Ω.
Note that the function u(z) defined above is bounded from above on any compact
subdomain. However, unless say A : Ω → GL(m, C), the subharmonic function
u(z) = logA(z) may be unbounded from below or it may assume the value −∞.
A fundamental result in the theory of subharmonic functions is the Riesz repre-
sentation theorem, which we formulate below.
Theorem 6.4 Let u(z) be a subharmonic function in a domain Ω, and assume that
u(z) ≡ −∞. Then there is a unique Borel measure μ on Ω called the Riesz measure
of u, such that for every compactly contained subdomain Ω1 ,
u(z) = log |z − ζ | dμ(ζ ) + h(z),
Ω1
Let us assume for now that the subharmonic function u(z) is bounded:
Using Jensen’s formula for subharmonic functions (see Sect. 7.2 in [17]), for every
z ∈ Ω and r > 0, if the disk D(z, r ) ⊂ Ω, then we have
r
μ(D(z, t)) 1
dt = u(z + r e(θ ))dθ − u(z).
0 t 0
Since clearly
r
μ(D(z, t)) r
μ(D(z, t))
dt ≥ dt μ(D(z, r/2)),
0 t r/2 t
we conclude that
μ(D(z, r/2)) C. (6.9)
220 6 Large Deviations for Quasi-Periodic Cocycles
μ(Ω1 ) C(Ω, Ω1 ) C,
where C is the bound on u(z) and C(Ω, Ω1 ) is a constant that depends on how the
subdomain Ω1 is covered by disks contained in Ω.
An argument that uses the Poisson-Jensen representation formula (see Sect. 3.7
in [11]) leads to a similar bound on the L ∞ -norm (on a slightly smaller compactly
contained subdomain Ω2 ) for the harmonic part h(z) in the Riesz representation of
u(z). We conclude that if |u(z)| ≤ C on Ω, then
Then
μ(Ar/2 ) + h L ∞ (Ar/4 ) ≤ Cr C, (6.12)
Remark 6.5 Lemma 6.6 above says that in order to obtain a uniform measurement
like (6.10) for a subharmonic function u(z), it is enough to have an upper bound
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 221
everywhere and a lower bound at some point. Clearly assumption (6.11) is implied
(up to doubling the constant) by
Lemma 6.7 Under the assumptions above, the following estimates on the Fourier
coefficients of u(x) as a function on T hold:
û(k) S 1 for all k ∈ Z, k = 0. (6.15)
k
Free ebooks ==> www.Ebook777.com
222 6 Large Deviations for Quasi-Periodic Cocycles
Lemma 6.8 Let u(z) be a subharmonic function satisfying (6.14). Assume moreover
that there is a splitting u = u 0 + u 1 with u 0 L ∞ (T) < ε0 and u 1 L 1 (T) < ε1 . Then
u(x) has the following BMO bound:
Using the Poisson integral representation for harmonic functions and scaling, it
is easy to see that since h L ∞ (Ar/4 ) ≤ Cr C, then ∂x h L ∞ (T) Crr C.
We conclude that
Cr
μ(Ar/2 ) + h L ∞ (Ar/4 ) + ∂x h L ∞ (T) C,
r
or in other words, the assumption (6.14) above holds with S = O Crr C .
We also note that since the annulus Ar will be fixed throughout the chapter, the
uniform measurement in (6.14) will depend only on the bound C in (6.17). That is,
we may write
μ(Ar/2 ) + h L ∞ (Ar/4 ) + ∂x h L ∞ (T) C.
The following result is an immediate consequence of Lemma 6.8 and it shows that
a a weak a-priori concentration inequality for a subharmonic function can always be
boosted
to a stronger estimate. We use the notation u to denote the space average
T u(x)d x of the function u(x) on T.
www.Ebook777.com
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 223
Lemma 6.9 Let u(z) be a subharmonic function such that the bound (6.14) holds.
If u(x) satisfies the weak a-priori estimate:
{x ∈ T : u(x) − u > ε0 } < ε1 (6.18)
The Birkhoff ergodic theorem implies that if the translation by ω is ergodic, then for
any observable ξ ∈ L 1 (Td ), as n → ∞ the Birkhoff average
1
n−1
ξ(x + jω) → ξ for a.e. x ∈ Td .
n j=0
We now state the analogue of the boosting Lemma 6.9 for several variables. For
simplicity we consider d = 2 variables, but a similar result, proven the same way,
holds for any number d of variables. The meaning of the constant Cr below was
given in Sect. 6.2.3 (see Remark 6.6). We remind the reader that Cr depends only on
the annulus Ar , hence only on r .
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 225
Lemma 6.10 Let u(z) be a pluri-subharmonic function on Ar2 such that the uniform
measurement (6.20) holds for some constant C < ∞ and let S := Crr C.
If u(x) satisfies the weak a-priori estimate:
{x ∈ T2 : u(x) − u > ε0 } < ε1 (6.21)
A similar result was proven in [3] (see Lemma 4.12) for bounded pluri-
subharmonic functions, using a slicing argument and the corresponding one vari-
able result.
The reader may verify that the argument in [3] can be employed as long as the one
variable result, i.e. Lemma 6.9, can be applied uniformly to the functions u(·, z 2 ),
u(z 1 , ·), for all z 1 , z 2 ∈ Ar , and provided these functions are also uniformly bounded
in say L 2 (T). These conditions are of course ensured by the assumption (6.20).
We are now ready to formulate the main result of this section, a quantitative
Birkhoff ergodic theorem.
Let ω ∈ DCt and put n 0 := t −2 . There is a = a(d) > 0 so that for all n ≥ n 0
n−1
{x ∈ Td : 1 u(x + jω) − u > S n −a } < e−c n ,
a
(6.24)
n j=0
where S = O Crr C and c = O(1).
Proof We prove this statement for d = 2 variables, but the same argument holds for
any number of variables. We follow the same strategy used in the proof of Proposition
4.1 in [16], as the main ingredients of the argument depend only on having a uniform
decay on the Fourier coefficients of u (separately in each variable) and on the boosting
Lemma 6.10, both of which are ensured by the assumption (6.23). For the reader’s
convenience, we present the complete argument here.
The assumption (6.23) implies that the bound (6.17) holds in each variable and
with the same constant C, i.e. for all subharmonic functions u(·, z 2 ),u(z 1 , ·) with
z 1 , z 2 ∈ Ar . Hence by Remark 6.7, the bound (6.14) with S = O Crr C holds for all
these functions as well.
226 6 Large Deviations for Quasi-Periodic Cocycles
This ensures that we can apply Lemma 6.7 on the decay of the Fourier coefficients
in each variable and obtain
1 1
sup û(l1 , x2 ) ≤ S · and sup û(x1 , l2 ) ≤ S · (6.25)
x2 ∈T |l1 | x1 ∈T |l2 |
1
n−1
u((x1 , x2 ) + j (ω1 , ω2 ))
n j=0
1
n−1
= u + û(l1 , l2 ) · e((l1 , l2 ) · (x1 , x2 )) · e( j (l1 , l2 ) · (ω1 , ω2 ))
(l1 ,l2 )∈Z2
n j=0
(l1 ,l2 ) =(0,0)
= u + û(l1 , l2 ) · e((l1 , l2 ) · (x1 , x2 )) · K n ((l1 , l2 ) · (ω1 , ω2 )),
(l1 ,l2 )∈Z2
(l1 ,l2 ) =(0,0)
1
n−1
1 1 − e(ny)
K n (y) = e( j y) = .
n j=0 n 1 − e(y)
1
K n (y) ≤ min 1, (6.26)
ny
We will estimate the second sum above using the bounds (6.25) on the Fourier
coefficients of u(x1 , x2 ) and the first sum using the Diophantine condition on the
frequency (ω1 , ω2 ). The splitting point K will be chosen to optimize the sum of
these two estimates.
Clearly (6.25) implies:
1 2 1
û(l1 , l2 )2 = û(l1 , x2 )2 2 ≤ S ≤ S2 ,
(T)
Lx 2 |l1 | |l1 |2
l2 ∈Z
and 1 2
1
û(l1 , l2 )2 = û(x1 , l2 )2 2 ≤ S ≤ S2 .
(T)
Lx 1 |l2 | |l2 |2
l1 ∈Z
Then we have:
û(l1 , l2 )2 · K n ((l1 , l2 ) · (ω1 , ω2 ))2 ≤ û(l1 , l2 )2
|l1 |+|l2 |≥K |l1 |+|l2 |≥K
≤ û(l1 , l2 )2 + û(l1 , l2 )2 S2 1 .
(l1 ,l2 ) : |l1 |≥K /2 (l1 ,l2 ) : |l2 |≥K /2
K
1
û(l1 , l2 ) S .
|l1 | + |l2 |
We conclude:
1 n−1 1 K 1+δ0
u((x1 , x2 ) + j (ω1 , ω2 )) − u 2 2 < S + ≤ S n −a
n j=0 L (T ) K 1/2 nt
for n ≥ 1/t 2 and for some positive constant a which is universal here (as d = 2),
but which in general depends on d.
Applying Chebyshev’s inequality, we have:
n−1
{x ∈ T2 : 1 u(x + jω) − u > S n −a/5 } < n −8a/5 . (6.27)
n j=0
This estimate is of course too weak, since the size of the exceptional set is only
polynomially small. We will boost it by using Lemma 6.10.
Consider the average
1 1
n−1
v(z) := u(z + jω).
S n j=0
Since v = 1
S
u, from (6.27) we have:
{x ∈ T2 : v(x) − v > n −a/5 } < n −8a/5 .
The assumption (6.23) in Theorem 6.5 will not necessarily be satisfied by the pluri-
subharmonic functions associated to our model. That is because they are of the form
u(z) = logA(z), for some matrix valued analytic function A(z), and A(z) may
very well vanish identically along some lines parallel to coordinate axes. However,
they will satisfy a weaker property that will allow us to handle their singularities
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 229
via a change of coordinates, and replace them with pluri-subharmonic functions for
which (6.23) does hold. We describe this idea in the following proposition.
Proposition 6.2 Let f ∈ Crω (Td , R), f ≡ 0, ω ∈ DCt and C > 0. There
are constants δ = δ( f, r ) > 0, a = a(d) > 0, k0 = k0 ( f, r, d) ∈ N and
S = S( f, r, C) < ∞ such that if u : Ard → [−∞, ∞) is a pluri-subharmonic
function satisfying the bounds
1
m−1
−C + log g(z + jω) ≤ u(z) ≤ C for all z ∈ Ard , (6.28)
m j=0
for some g ∈ Crω (Td , R) with g − f r < δ and for some m ≥ 1, then
n−1
{x ∈ Td : 1 u(x + jω) − u > S n −a } < e−n
a
(6.29)
n j=0
Proof By Theorem 6.3, since f ≡ 0, there is M ∈ SL(d, Z) such that in the new
coordinates x = M x, the analytic
function
f (x ) does not vanish identically along
any hyperplane, and in fact, log f (x ) has the property that its L 2 -norms
separately
in each variable are uniformly bounded, or in other words that log f (x ) is
bounded. Moreover, this holds uniformly in a neighborhood of f .
We note that the conclusion (6.29) of the theorem is coordinate agnostic. Indeed,
if M ∈ SL(d, Z) then M −1 preserves the Haar measure on Td , while if ω ∈ DCt
then ω = M −1 ω ∈ DCt for t = t/M1d+1 , hence the Diophantine condition is
preserved up to a constant.
Furthermore, since M is linear, u ◦ M(z) is also pluri-subharmonic, and its domain
contains Ar , where r ∼ r/M. That is because the linear map induced by M
expands the imaginary direction, hence the width of the domain of u ◦ M(z) is
proportionally smaller. However, M is a fixed constant depending upon M, hence
only upon f .
Hence it is enough to prove (6.29) for u(x) replaced by u ◦ M(x) and for ω
replaced by M −1 ω.
Moreover, the assumption (6.28) holds for u ◦ M(x) provided we replace g(x) by
g ◦ M(x), which is still δ-close to f ◦ M(x).
Therefore, with no loss of generality, we may assume that for δ = δ( f, r ) small
enough, and for some finite constant C0 = C0 ( f, r ) we have: if g ∈ Crω (Td , R) with
g − f r < δ, then
|||log |g|||| ≤ C0 . (6.30)
230 6 Large Deviations for Quasi-Periodic Cocycles
m−1
u(x) ≤ 2C + 1 logg(x + jω) for all x ∈ Td .
m j=0
hence
sup u(z) + |||u||| ≤ 3C + C0 .
z∈Ard
This shows that the assumption (6.23) in Theorem 6.5 is now (after a change of
coordinates) satisfied.
We apply Theorem
6.5 to conclude that (6.29) holds with S = S( f, r, C) =
O Crr (3C + C0 ) and for all n ≥ n 0 , where n 0 := t −2 k0 . We choose k0 = k0 ( f, r, d)
so that k0 ≥ M2(d+1)
1 and so that the constant c = O(1) in (6.24) is absorbed.
We introduce some notations. For a cocycle A ∈ Crω (Td , Mat(m, R)), let
f A (z) := det[A(z)].
1
u (n)
A (z) := logA(n) (z).
n
6.3 The Proof of the Fiber LDT Estimate 231
Note that due to the analyticity of the cocycle A(z), the functions u (n)
A (z) are pluri-
subharmonic on Ard . This property is crucial in establishing the fiber LDT estimate.
We denote the space averages of these functions by
1
L (n)
1 (A) = u (n)
A (x) d x = logA(n) (x) d x.
Td Td n
Proposition 6.3 Given a cocycle A ∈ Crω (Td , Mat(m, R)) with f A = det[A] ≡ 0,
there are constants δ = δ(A) > 0 and C = C(A) < ∞, such that for any cocycle
B ∈ Crω (Td , Mat(m, R)), if B − Ar < δ, then
f B = det[B] ≡ 0, (6.31)
log f B L 2 (Td ) ≤ C (6.32)
1
n−1
−C + log f B (T i z) ≤ u (n)
B (z) ≤ C, (6.33)
n i=0
u (n)
B L 2 (Td ) ≤ C. (6.34)
Proof Clearly the map Crω (Td , Mat(m, R)) B → f B = det[B] ∈ Crω (Td , R)
is locally Lipschitz, hence we can choose δ = δ(A) > 0 small enough such that if
B − Ar < δ, then the analytic function g := f B satisfies the Łojasiewicz inequality
(6.3) with the same constants S = S( f A ), b = b( f A ) as f := f A (see Lemma 6.1).
In particular, f B ≡ 0 and from Remark 6.2 we have the uniform L 2 -bound:
log f B L 2 (Td ) ≤ C( f A ∞ , S, b) ≤ C2 (A).
The upper bound in (6.33) is clear: if B − Ar < δ, then Br ∼ Ar , and
since for every z ∈ Ard we have
n−1
B (n) (z) ≤ B(T i z) ≤ Brn ,
i=0
we conclude that
1
u (n)
B (z) = logB (n) (z) ≤ logBr < C1 (A).
n
232 6 Large Deviations for Quasi-Periodic Cocycles
Hence
n−1
n−1
f B (T i z) · I = det[B(T i z)] · I
i=0 i=0
= B(T n−1 z) . . . B(z) · adj(B(z)) . . . adj(B(T n−1 z))
= B (n) (z) · adj(B(z)) . . . adj(B(T n−1 z)).
Clearly
adj(B(z)) B(z)m−1 ≤ Brm−1 for all z.
This implies, for some C1 = C1 (A) ∼ logAr ,
1
n−1
1
u (n)
B (z) = logB (n) (z) ≥ −C1 + log f B (T i z),
n n i=0
(n)
n−1
u (x) ≤ 2C1 + 1 log f B (T i x).
B
n i=0
For GL(m, R)-valued cocycles, the functions u (n)A (x) are almost invariant under the
base dynamics T , in the sense that for all x ∈ Td ,
(n)
u (x) − u (n) (T x) ≤ C 1 .
A A
n
For a cocycle that has singularities but it is not identically singular, we establish
this property off of an exponentially small set of phases. It is in fact crucial to obtain
a bound on the measure of this exceptional set of phases, uniform in the cocycle.
6.3 The Proof of the Fiber LDT Estimate 233
Proposition 6.4 Let A ∈ Crω (Td , Mat(m, R)) such that det[A(x)] ≡ 0. Then there
are constants δ = δ(A) > 0 and C = C(A) < ∞, such that for any a ∈ (0, 1), if
B ∈ Crω (Td , Mat(m, R)) with B − Ar < δ, then
(n)
u (x) − u (n) (T x) ≤ C 1 (6.35)
B B
na
/ Bn , where Bn < e−n .
1−a
holds for all n ≥ 1 and for all x ∈
The exceptional set Bn may depend on the coycle B, but its measure does not.
Proof For any cocycle B ∈ Crω (Td , Mat(m, R)), let f B (x) := det[B(x)] ∈
Crω (Td , R). Then if B− Ar < δ we have f B − f A r < C δ, where C = C(A) > 0.
Using Lemma 6.1, there are constants δ, C, b > 0, all depending only on A, such
that if B − Ar < δ then
{x ∈ Td : f B (x) < t} < Ct b for all t > 0. (6.36)
In particular (or by Fubini), the set Z0 (B) := {x ∈ Td : f B (x) = 0} has zero measure,
and so does ∪n≥0 T −n Z0 (B) =: Z(B).
Hence if x ∈ / Z(B), then B(T n x) is invertible for all n ≥ 0 and we can write:
1 1
logB (n) (x) − logB (n) (T x)
n n
1 B(T n x)−1 · [B(T n x) · B(T n−1 x) · · · · · B(T x)] · B(x)
= log
n B(T n x) · B(T n−1 x) · · · · · B(T x)
1 C 1
≤ log[B(T n x)−1 · B(x)] ≤ + logB(T n x)−1 .
n n n
The last inequality follows from
1 1
logB (n) (T x) − logB (n) (x)
n n
1 B(T n x) · [B(T n−1 x) · · · · · B(T x) · B(x)] · B(x)−1
= log
n B(T n−1 x) · · · · · B(x)
1 C 1
≤ log[B(T n x) · B(x)−1 ] ≤ + logB(x)−1 .
n n n
234 6 Large Deviations for Quasi-Periodic Cocycles
We conclude that if x ∈
/ Z(B), then
1
logB (n) (x) − 1 logB (n) (T x)
n n
C 1 1
< + logB(T n x)−1 + logB(x)−1 . (6.37)
n n n
Therefore, in order to prove (6.35), we need to obtain a (uniform in B) upper
bound for B(x)−1 , where x is outside an exponentially small set.
Upper bounds on the norm of the inverse of a matrix are obtained from lower
bounds on the determinant via Cramer’s rule. Indeed, if x ∈
/ Z(B) then
1 1
B(x)−1 = adj(B(x)) · ≤ C1 · ,
det[B(x)] f B (x)
1
B(x)−1 ≤ C1 = C1 C 1/b e(1/b) n = C2 (A) e(1/b) n ,
1−a 1−a
t
hence
1 log C2 n 1−a 1
logB(x)−1 ≤ + 1/b ≤ C3 (A) a
n n n n
which, combined with (6.37), proves the proposition.
Theorem 6.6 Given A ∈ Crω (Td , Mat(m, R)) with det[A(x)] ≡ 0 and ω ∈ DCt ,
there are constants δ = δ(A) > 0, k0 = k0 (A) ∈ N, C = C(A, r ) < ∞, a = a(d) >
0 and b = b(d) > 0 such that if B − Ar ≤ δ and n ≥ n 0 := t −2 k0 , then
{x ∈ Td : 1 logB (n) (x) − L (n) (B) > C n −a } < e−n b . (6.38)
1
n
6.3 The Proof of the Fiber LDT Estimate 235
Proof Using Proposition 6.3, there are constants δ = δ(A), C = C(A), such that if
B is a cocycle near A: B − Ar < δ, then for all scales n ≥ 1
1
n−1
−C + log f B (T i z) ≤ u (n)
B (z) ≤ C.
n i=0
R−1
{x ∈ Td : 1 u (n) (x + jω) − u (n)
> S R −a } < e−c R
a
(6.39)
B B
R j=0
From the nearly almost invariance property given by Proposition 6.4, after possibly
decreasing δ, and for a constant C = C (A) < ∞, we have that if B − Ar < δ,
then
(n)
u (x) − u (n) (T x) ≤ C 1
B B
na
/ Bn , where Bn < e−n .
1−a
for all n ≥ 1 and for all x ∈
T Bn . Hence B̄n ≤ R e−n and if x ∈
R−1 −i 1−a
Let B̄n := ∪i=0 / B̄n then
(n)
R−1
u (x) − 1 u (n) R
B B (x + jω) ≤ C a (6.40)
R j=0 n
Pick R n a , say R = n a/(a+1) and let C = 2(C +S). The conclusion (6.38) of
the theorem then follows from (6.39) and (6.40) for some easily computable choice
of the new parameter a and the parameter b.
Remark 6.8 What determines all constants in the LDT estimate above, are precisely
some measurements on the function f A (x) := det[A(x)] and the parameter t of the
Diophantine condition DCt on the frequency ω.
We also note that unlike in the case of random cocycles, the fiber-LDT estimate
above was proven without assuming the existence of a gap between the first two
Lyapunov exponents. In particular, using exterior powers, we can derive an LDT
estimate for every Lyapunov exponent.
236 6 Large Deviations for Quasi-Periodic Cocycles
To prove the continuity of the LE, we use the abstract criterion given by Theorem 3.1,
which was formulated and proven in Chap. 3. Section 3.1 contains all the relevant
definitions and the precise formulation of this criterion. Under the same assumptions,
in Chap. 4 we obtained abstract criteria for the continuity of the Oseledets filtration
(Theorem 4.7) and of the Oseledets decomposition (Theorem 4.8).
For the reader’s convenience, we briefly review the relevant definitions. We then
explain how Theorems 3.1, 4.7 and 4.8 are applicable to the context of this chapter.
Proof (of Theorem 6.1) The torus Td together with the σ -algebra of Borel sets,
the Haar measure and thetranslation by a rationally independent vector ω form an
ergodic MPDS. Let C = m≥1 Cm be the space of analytic, not identically singular
cocycles over this ergodic system defined in Sect. 6.1.1.
We say that a cocycle A is uniformly L 2 -bounded if there are constants C =
C(A) < ∞ and δ = δ(A) > 0 such that
1
logB (n) 2 d < C
n L (T )
for all cocycles B that are close enough to A (in the given topology on the space of
cocycles) and for all scales n ≥ 1.
Estimate (6.34) in Proposition 6.3 shows that all cocycles in C are uniformly
L 2 -bounded.
Given a cocycle A ∈ Cm and N ∈ N, note that the sets {x ∈ Td : A(n) (x) ≤ c}
or {x ∈ Td : A(n) (x) ≥ c} for some 1 ≤ n ≤ N and c > 0 are closed, so the lattice
F N (A) generated by them consists only of closed sets.
We say that a set Ξ of observables and a cocycle A ∈ C are compatible,
iffor any
N ∈ N, F ∈ F N (A), ε > 0, there is ξ ∈ Ξ such that 1 F ≤ ξ and Td ξ d x ≤ F +ε.
Let Ξ := C0 (Td ) be the set of all continuous observables ξ : Td → R. By the
regularity of the Borel measure, there is an open set U ⊇ F such that U ≤ F + ε.
By Urysohn’s lemma, there is a continuous function ξ ∈ Ξ such that 0 ≤ ξ ≤ 1,
ξ ≡ 1 on F and ξ ≡ 0 on U . Then
1 F ≤ ξ and ξ d x ≤ U ≤ F + ε,
Td
which shows that Ξ is compatible with every cocycle in C , a property we call the
compatibility condition.
We call deviation size function any non-increasing map ε : (0, ∞) → (0, ∞), and
deviation measure function any sufficiently fast decreasing function ι : (0, ∞) →
(0, ∞). A triplet (n 0 , ε, ι), where n 0 ∈ N and ε, ι are deviation size or measure
functions is called an LDT parameter, while any set P containing such triplets is
called a set of LDT parameters.
6.4 Deriving Continuity of the Lyapunov Exponents 237
n−1
{x ∈ Td : 1 ξ(T j x) − ξ d x > ε(n)} < ι(n).
n j=0 Td
{x ∈ Td : 1 logB (n) (x) − L (n) (B) > ε(n)} < ι(n).
1
n
Theorem 6.6 shows that a uniform fiber-LDT estimate holds for all cocycles in C ,
with the parameter space P being the set of all triplets (n 0 , ε, ι) with n 0 ∈ N, ε(t) ≡
C t −a and ι(t) ≡ et where the constants n 0 , a, b are explicitly described in terms of
b
Hölder continuity.
The statements on the Oseledets filtration and decomposition follow directly from
the corresponding general criteria.
t
kω ≥ (6.41)
k (log k )1+
1
gr(gi ) > for all 0 ≤ i ≤ n − 1
κ
gi gi−1
>ε for all 1 ≤ i ≤ n − 1
gi gi−1
then
n−2
n−1 κ
(n)
logg + loggi − loggi gi−1 n · 2 .
ε
i=1 i=1
The argument we are about to present works only for one-variable translations
because it requires the following sharp version of the quantitative Birkhoff ergodic
Theorem 6.5, and this sharp version is only available in the one-variable setting.
Proposition 6.6 Let ω ∈ T satisfying the strong Diophantine condition (6.41) for
some t > 0 and let u : Ar → [−∞, ∞) be a subharmonic function satisfying the
bound
sup u(z) + u L 2 (T) ≤ C.
z∈Ar
There are constants c1 , c2 > 0 that depend only on C and r and there is n 0 ∈ N
that depends on t such that for all ε > 0 and n ≥ n 0 we have:
n−1
{x ∈ T : 1 u(x + jω) − u > ε} < e−c1 εn+c2 (log n) .
4
n j=0
6.5 Refinements in the One-Variable Case 239
This result was proven in [9] (see Theorem 3.8) for bounded subharmonic func-
tions. It remains valid in our setting, by the same argument following Lemma 6.8.
We can now phrase the sharper fiber-LDT in the one-variable case.
Theorem 6.7 Let A ∈ Crω (T, Mat m (R)) with det[A(x)] ≡ 0, and let ω ∈ T be a
frequency satisfying the strong Diophantine condition (6.41).
Assume that L 1 (A) > L 2 (A) and let ε > 0. There are δ = δ(A) > 0, p =
p(A) < ∞, n̄ 1 = n̄ 1 (A, ω, ε) ∈ N such that for all B ∈ Crω (T, Mat m (R)) with
B − Ar < δ and for all n ≥ n̄ 1 we have:
{x ∈ T : 1 log B (n) (x) − L (n) (B) > ε} < e−ε n/(log n) p .
n 1
Before starting the proof of this sharper fiber-LDT, we note that it cannot be ob-
tained along the same lines as Theorem 6.6, by using the sharper estimate in Propo-
sition 6.6, precisely because the nearly almost invariance property (Proposition 6.4)
is too weak, as a consequence of the singularities of the cocycle.
Proof We first explain the mechanics of the proof, then we detail the argument,
which bears some similarities with an argument used in [10].
We already have a version of the fiber-LDT estimate in Theorem 6.6, where the
deviation functions ε(t) ≡ C t −a , ι(t) ≡ e−t are both relatively coarse. Using this
b
LDT at a scale n 0 and the avalanche principle in Proposition 6.5, we derive an LDT
b1
estimate at a larger scale n 1 en 0 (where 0 < b1 < b), which has a much sharper
deviation size function ε(t), but a very coarse deviation measure function ι(t). This
will lead, via Lemma 6.8, to BMO estimates, which with the help of John-Nirenberg’s
inequality will prove the desired stronger LDT estimate.
Let γ := L 1 (A) − L 2 (A) > 0. Let n̄ 0 ∈ N, δ > 0 be such that the fiber-
LDT estimate in Theorem 6.6 holds for all cocycles B ∈ Crω (T, Mat m (R)) with
B − Ar < δ and for all n ≥ n̄ 0 .
Moreover, the Lyapunov exponents are already known to be continuous, and in
fact, the more precise finite scale uniform estimates from Lemma 3.4. in Chap. 3 hold
as well. Therefore, we can choose n̄ 0 and δ so that the following conditions hold for
all B ∈ Crω (T, Mat m (R)) with B − Ar < δ. Firstly,
1
log gr(B (m) (x)) ≥ γ /2 (6.42)
m
holds for all x outside a set of measure <e−m and for all m ≥ n̄ 0 . Also
b
(m 1 )
L (B) − L (m 2 ) (B) < γ /30 (6.43)
1 1
for all m 1 , m 2 ≥ n̄ 0 .
Fix any integer n 0 ≥ n̄ 0 . We will use the notation m n 0 for scales m such that
n 0 ≤ m ≤ 3n 0 .
240 6 Large Deviations for Quasi-Periodic Cocycles
B (m 2 +m 1 ) (x)
≥ e−n 0 γ /5 =: ε (6.44)
B (m 2 ) (T m 1 x) B (m 1 ) (x)
1
gr(B (m) (x)) ≥ en 0 γ /2 =: , (6.45)
κ
n−1
m i n 0 and n 1 = mi .
i=0
1
gr(gi ) = gr(B (m i ) (T qi x)) > . (6.48)
κ
6.5 Refinements in the One-Variable Case 241
By excluding a set of measure < n e−n 0 < e−1/2 n 0 the estimates (6.47) and (6.48)
b b
will hold for all indices i, so the avalanche principle of Proposition 6.5 applies and
we have:
n−2
n−1 κ
(n)
logg + loggi − loggi gi−1 n · 2 < n e−n 0 γ /10 ,
ε
i=1 i=1
which becomes
n−2
logB (n 1 ) (x) + logB (m i ) (T qi x) (6.49)
i=1
n−1
− logB (m i +m i−1 ) (T qi−1 x) < n e−n 0 γ /10 .
i=1
1
log B (m) (x) > L 1(m) (B) − m −a > −C,
m
for some C = C(A) < ∞, where the lower bound on L 1(m) (B) follows from (6.34).
Moreover, since by (6.33) we have that for all x ∈ T,
1
log B (m) (x) ≤ C,
m
log B (m) (x) ≤ C m C n 0 .
This estimate clearly applies to the function x → log B (m 2 +m 1 ) (x) and to the
function x → log B (m n−1 +m n−2 ) (T qn−2 x) which represent the terms corresponding
to i = 1 and i = n − 1 in the second sum in (6.49).
Let v(x) := log B (m 2 +m 1 ) (x) + log B (m n−1 +m n−2 ) (T qn−2 x). Then for x outside
a set of measure e−n 0 , the function v(x) has the bound
b
v(x) Cn 0 .
Free ebooks ==> www.Ebook777.com
242 6 Large Deviations for Quasi-Periodic Cocycles
b1
Since n 1 en 0 , with b1 < 1, estimate (6.49) then implies that for all x outside a
set of measure e−1/2 n 0 we have:
b
1 1
n−2
logB (n 1 ) (x) + logB (m i ) (T qi x) (6.50)
n1 n 1 i=1
1
n−2
n −n 0 γ /10 n0 n0
− logB (m i +m i−1 ) (T qi−1 x) < e +C .
n 1 i=2 n1 n1 n1
We are now ready to average (6.50) over some specific choices of the integers
m 0 , m 1 , . . . , m n−1 .
The goal is to apply Proposition 6.6 to u (m)
B (z) = m log B
1 (m)
(z) for m = n 0
and m = 2n 0 , which will allow us to replace the two sums in (6.50) by L (n 0)
1 (B) and
L (2n
1
0)
(B) respectively.
The reason we will not simply choose all integers m i to be n 0 , but instead we
will consider n 0 -many configurations and then average, is that otherwise the trans-
lates T qi x = x + qi ω would only be by multiples of n 0 , i.e. qi = i n 0 . However,
Proposition 6.6 involves all translations T j x = x + jω, 0 ≤ j < n 1 and not just the
translations by i n 0 ω, 0 ≤ i ≤ n − 1.
Firstly, pick all integers m i to be n 0 except for the last one, m n−1 , which is
chosen such that 2n 0 ≤ m n−1 ≤ 3n 0 . Note that in this case q0 = 0, qi = i n 0 for
1 ≤ i ≤ n − 1, hence we are getting the translates by multiples of n 0 .
Secondly, increase by 1 the size of the first block, decrease by 1 the size of the
last block, and keep all the other blocks the same. In other words, let m 0 = n 0 + 1,
m 1 = m 2 = · · · = m n−2 = n 0 and m n−1 n 0 is chosen so that all integers m i add
up to n 1 . In this case q0 = 0, qi = in 0 + 1 for 1 ≤ i ≤ n − 1, so we are getting the
translates by multiples of n 0 plus 1.
Continue to increase the first block by 1, decrease the last by 1 and keep the rest
the same, for n 0 steps.
In other words, for each 0 ≤ j ≤ n 0 − 1, choose the following integers:
1 1
n 0 −1
n−2
1
logB (n 1 ) (x) + logB (n 0 ) (T i n 0 + j x) (6.51)
n1 n 1 j=0 i=1 n 0
2
n 0 −1
n−2
1 n0
− logB (2n 0 ) (T (i−1) n 0 + j x) < C
n 1 j=0 i=2 2n 0 n1
www.Ebook777.com
6.5 Refinements in the One-Variable Case 243
1 (n−1)
n 0 −1
1 1
logB (n 1 ) (x) + logB (n 0 ) (T k x) (6.52)
n1 n1 k=n 0
n0
(n−2)
n 0 −1 (log n 1 )1/b1
2 1 n0
− logB (2n 0 ) (T k x) < C
n1 k=n 0
2n 0 n1 n1
Due to the uniform estimates (6.33), (6.34) in Propositions 6.3, 6.6 can be applied
to the functions
1 1
u (n 0)
B (z) = log B (n 0 ) (z) and u (2n
B
0)
(z) = log B (n 0 ) (z).
n0 2n 0
Let ε (log nn11) and we may of course assume that 1/b1 > 1. Moreover, since
1/b1
1 (n−1)
n 0 −1 (log n 1 )1/b1
1
logB (n 0 ) (T k x) − L (n 0)
1 (B) < ε (6.53)
n1 k=n 0
n0 n1
for x outside a set of measure < e−c1 εn 1 +c2 (log n 1 ) < e−c3 (log n 1 )
4 1/b1
.
Similarly we have:
1 (n−2)
n 0 −1 (log n )1/b1
1
logB (2n 0 ) (T k x) − L (2n 0) 1
1 (B) < (6.54)
n1 k=n 0
2n 0 n1
Combine (6.52), (6.53), (6.54), (6.55) to conclude that for x outside a set of
measure e−c3 (log n 1 ) 1 , we have
1/b
1 (log n 1 )1/b1
(n 1 ) (n 1 )
n logB (x) − L 1 (B) n1
. (6.56)
1
244 6 Large Deviations for Quasi-Periodic Cocycles
Let
(log n 1 )1/b1
, ε1 e−c3 (log n 1 ) ε08
1/b1
ε0
n1
and let
1
u(z) = u (n 1)
B (z) = log B (n 1 ) (z).
n1
Then u(z) is subharmonic on Ar , and by Proposition 6.3, for some C = C(A) <
∞ we have the bound
sup u(z) + u L 2 (T) ≤ C,
z∈Ar
which via Remark 6.7 implies (6.14) with S = O Crr C .
Moreover, from (6.56)
{x ∈ T : u(x) − u > ε0 } < ε1 .
All the assumptions of Lemma 6.8 are then satisfied and we conclude that
We ran this argument starting at any scale n 0 ≥ n̄ 0 and we obtained (6.57) for
b1 b1
n 1 en 0 . Therefore, if n̄ 1 en̄ 0 , then the estimate (6.57) holds for all n 1 ≥ n̄ 1 ,
which proves our theorem.
Remark 6.9 Using the terminology in Sect. 6.4, we have shown that for a one-
variable torus translation by a (strongly) Diophantine frequency, a uniform fiber-LDT
holds with deviation measure function
We may then conclude, as in the proof of Theorem 6.1, that if A ∈ Cm has a τ -gap
pattern, then in a neighborhood of A the functions Λτ , F τ and E ·τ have the modulus
of continuity
ω(h) := e−c [log 1/ h]/[log log 1/ h] .
b
That is, in the presence of gaps in the Lyapunov spectrum, the LE, the Oseledets
filtration and the Oseledets decomposition are locally nearly-Hölder continuous.
References
1. A. Ávila, Global theory of one-frequency Schrödinger operators. Acta Math. 215(1), 1–54
(2015). MR 3413976
2. A. Ávila, S. Jitomirskaya, C. Sadel, Complex one-frequency cocycles. J. Eur. Math. Soc.
(JEMS) 16(9), 1915–1935 (2014). MR 3273312
3. J. Bourgain, Green’s function estimates for lattice Schrödinger operators and applications.
Annals of Mathematics Studies, vol. 158 (Princeton University Press, Princeton, NJ, 2005).
MR 2100420 (2005j:35184)
4. J. Bourgain, Positivity and continuity of the Lyapounov exponent for shifts on Td with arbitrary
frequency vector and real analytic potential. J. Anal. Math. 96, 313–355 (2005). MR 2177191
(2006i:47064)
5. J. Bourgain, S. Jitomirskaya, Continuity of the Lyapunov exponent for quasiperiodic operators
with analytic potential. J. Stat. Phys. 108(5–6), 1203–1218 (2002). Dedicated to David Ruelle
and Yasha Sinai on the occasion of their 65th birthdays. MR 1933451 (2004c:47073)
6. D Damanik, Schrödinger operators with dynamically defined potentials: a survey (2015), 1–80
(to appear in Ergodic Theory and Dynamical Systems) (preprint)
7. P. Duarte, S. Klein, Continuity of the Lyapunov exponents for quasiperiodic cocycles. Comm.
Math. Phys. 332(3), 1113–1166 (2014). MR 3262622
8. J. Duoandikoetxea, Fourier analysis, Graduate Studies in Mathematics, vol. 29 (American
Mathematical Society, Providence, RI, 2001). Translated and revised from the 1995 Spanish
original by David Cruz-Uribe. MR 1800316 (2001k:42001)
9. M. Goldstein, W. Schlag, Hölder continuity of the integrated density of states for quasi-periodic
Schrödinger equations and averages of shifts of subharmonic functions. Ann. Math. (2) 154(1),
155–203 (2001). MR 1847592 (2002h:82055)
10. M. Goldstein, W. Schlag, Fine properties of the integrated density of states and a quantitative
separation property of the Dirichlet eigenvalues. Geom. Funct. Anal. 18(3), 755–869 (2008).
MR 2438997 (2010h:47063)
11. W.K. Hayman, P.B. Kennedy, Subharmonic functions. Vol. I. (Academic Press [Harcourt Brace
Jovanovich, Publishers], London, 1976). London Mathematical Society Monographs, No. 9.
MR 0460672 (57 #665)
12. S. Jitomirskay, C.A. Marx, Continuity of the Lyapunov exponent for analytic quasi-periodic
cocycles with singularities. J. Fixed Point Theory Appl. 10(1), 129–146 (2011). MR 2825743
(2012h:37095)
13. S. Jitomirskaya, C.A. Marx, Analytic quasi-perodic cocycles with singularities and the Lya-
punov exponent of extended Harper’s model. Comm. Math. Phys. 316(1), 237–267 (2012).
MR 2989459
14. S. Jitomirskaya, C.A. Marx, Dynamics and spectral theory of quasi-periodic Schrödinger-type
operators (2015), 1–44 (to appear in Ergodic Theory and Dynamical Systems) (preprint)
15. S. Klein, Anderson localization for the discrete one-dimensional quasi-periodic Schrödinger
operator with potential defined by a Gevrey-class function. J. Funct. Anal. 218(2), 255–292
(2005). MR 2108112 (2005m:82070)
246 6 Large Deviations for Quasi-Periodic Cocycles
16. S. Klein, Localization for quasiperiodic Schrödinger operators with multivariable Gevrey po-
tential functions. J. Spectr. Theory 4, 1–53 (2014)
17. B.Ya. Levin, Lectures on entire functions. Translations of Mathematical Monographs, vol.
150 (American Mathematical Society, Providence, RI, 1996) In collaboration with and with a
preface by Yu. Lyubarskii, M. Sodin, V. Tkachenko, Translated from the Russian manuscript
by Tkachenko. MR 1400006 (97j:30001)
18. C. Muscalu, W. Schlag, Classical and multilinear harmonic analysis, vol. I. Cambridge Studies
in Advanced Mathematics, vol. 137 (Cambridge University Press, Cambridge, 2013). MR
3052498
19. E.S. Selmer, On the irreducibility of certain trinomials. Math. Scand. 4, 287–302 (1956). MR
0085223 (19,7f)
20. K. Tao, Continuity of Lyapunov exponent for analytic quasi-periodic cocycles on higher-
dimensional torus. Front. Math. China 7(3), 521–542 (2012). MR 2915794
21. K. Tao, Hölder continuity of Lyapunov exponent for quasi-periodic Jacobi operators. Bull. Soc.
Math. France 142(4), 635–671 (2014). MR 3306872
22. Y. Wang, Z. Zhang, Uniform positivity and continuity of Lyapunov exponents for a class of C 2
quasiperiodic Schrödinger cocycles. J. Funct. Anal. 268(9), 2525–2585 (2015). MR 3325529
23. J. You, S. Zhang, Hölder continuity of the Lyapunov exponent for analytic quasiperiodic
Schrödinger cocycle with weak Liouville frequency. Ergodic Theory Dynam. Syst. 34(4),
1395–1408 (2014). MR 3227161
Chapter 7
Further Related Problems
An intrinsic limitation of the inductive method used to prove the continuity of the
Lyapunov exponents is that it can never lead to a better than Hölder modulus of conti-
nuity. That is because the modulus of continuity depends on the speed of convergence
of the finite to the infinite scale Lyapunov exponents. This speed of convergence is in
turn dependent upon the leap in the inductive procedure from one scale to the next,
which is determined by the strength of the large deviation type estimates. But the
exceptional sets in these estimates cannot be better than exponentially small, leading,
at best, to Hölder modulus of continuity. However, as we explain below, in general
this is optimal.
Indeed, in the quasi-periodic, one variable setting, consider the almost Mathieu
cocycle
λ cos(x) − E −1
Aλ,E (x) := ,
1 0
where λ, E are some real parameters, and the translation is a Diophantine number.
Bourgain has shown in [9] that for large enough λ, as a function of E, the Lyapunov
exponent of Aλ,E is α-Hölder continuous for any α < 21 . Moreover, due to the
presence of gaps in the spectrum of the associated almost Mathieu operator, the
Hölder exponent 21 is optimal (see [9, 10]).
In the random setting, more precisely for random Schrödinger cocycles, B.
Halperin has provided the following family of cocycles with arbitrary small Hölder
exponent. Consider the probability measure μa,b (E) := 21 δ X E + 21 δY E , where X E
and Y E are the SL(2, R)-matrices
Each of these measures determines a random irreducible cocycle over a full shift in
two symbols. Halperin has proven (see [33, Appendix 3]) that the Hölder exponent in
the modulus of continuity for the Lyapunov exponent as a function of the parameter
E is less or equal than
2 log 2
α(a, b) = ,
arccosh(1 + 21 |a − b|)
This formula implies the affirmed relation between the modulus of continuity of v(∞)
and L 1 .
7.1 Limitations and Counterexamples 249
Therefore, the almost Mathieu example for quasi-periodic cocycles, and the
Halperin example for random cocycles, show that Hölder is the optimal modulus
of continuity one can expect for the Oseledets filtration in these contexts.
More serious limitations are provided by the examples where the Lyapunov expo-
nents are known to be discontinuous. By Kingman’s Ergodic theorem, the first
Lyapunov exponent of a cocycle is the infimum of the corresponding finite scale
Lyapunov exponents. Since in general the finite scale Lyapunov exponents are con-
tinuous, the first Lyapunov exponent is an upper semi-continuous function of the
cocycle.
The Bochi-Mañé dichotomy says that for a residual set of volume preserving
C 1 -diffeomorphisms, the systems are either Anosov (uniformely hyperbolic) or else
have zero Lyapunov exponents. This result was first announced by R. Mañé in the
1980s, and later proved by Bochi [6] in the context of area preserving diffeomor-
phisms. Bochi and Viana [7] have generalized this dichotomy to higher dimensional
diffeomorphisms, including both the cases of symplectic and volume preserving
diffeomorphisms. This type of dichotomy provides residual sets of continuous
S L(2, R)-cocycles which are either uniformly hyperbolic or else have zero Lyapunov
exponents [5]. Consider an ergodic automorphism T : X → X on a probability space
(X, μ), where X is a compact metric space. Let C be the class of continuous cocycles
A : X → SL(2, R) with L 1 (A, μ) > 0 but which are not uniformly hyperbolic. By
the previous dichotomy, the Lyapunov exponent L 1 : C → R, A
→ L 1 (A, μ) is
discontinuous everywhere.
In the context of quasi-periodic cocycles Wang and You [37] have recently pro-
vided examples, for any 0 ≤ l ≤ ∞, of class C l quasi-periodic SL(2, R)-cocycles
where the first Lyapunov exponent L 1 : C l (T1 , SL(2, R)) → R is discontinuous.
A previous work of Young [38] had signaled the ubiquity of elliptic behavior for
quasi-periodic cocycles A : T → SL(2, R) over a rotation on the torus T, which is
a key ingredient to produce the discontinuities in [37].
As a last example, in the context of random cocycles, consider the following
family of probability measures μθ := θ δ X + (1 − θ ) δY , where 0 ≤ θ ≤ 1, and X
and Y are the SL(2, R)-matrices
0 −1 2 0
X= Y = . (7.1)
1 0 0 2−1
The probability measure μθ determines a random cocycle over a full shift in two
symbols. Let us denote its first Lyapunov exponent by L 1 (μθ ). We leave as an exercise
for the reader to check that L 1 (μ0 ) = log 2 while L 1 (μθ ) = 0 for all 0 < θ ≤ 1.
This shows that θ = 0 is a discontinuity point of the function θ
→ L 1 (μθ ).
In spite of these limitations on the continuity of the Lyapunov exponents, there
are positive results which enhance the regularity of the Lyapunov exponents at the
cost of restricting the space of cocycles.
In the context of random cocycles over Bernoulli shifts of finite type, Peres [32] has
proven the analiticity of the first Lyapunov exponent as a function of the probability
250 7 Further Related Problems
where
i. λ
= 0 is a coupling constant encoding the disorder of the system, and
ii. for every lattice point n ∈ Z, the potential vn (x) is given by
vn (x) = f (T n x),
Hλ (x) ψ = E ψ, (7.3)
for some energy (i.e. eigenvalue) E ∈ R and state (i.e. eigenvector) ψ = {ψn }n∈Z .
7.2 Some Connections to Mathematical Physics 251
Define the associated Schrödinger cocycle as the cocycle (T, Aλ,E ), where
λ f (x) − E −1
Aλ,E (x) := ∈ SL(2, R).
1 0
Note that the Schrödinger equation (7.3) is a second order finite difference equa-
tion. An easy calculation shows that its formal solutions are given by
ψn+1 ψ0
= A(n+1)
λ,E (x) · ,
ψn ψ−1
where A(n)
λ,E (x) are the iterates of Aλ,E (x), for all n ∈ N .
The Lyapunov exponents of a Schrödinger cocycle are regarded as functions of
the energy parameter E (or, when the disorder λ varies, as functions of λ and E).
Note that since Aλ,E (x) ∈ SL(2, R), we have L 2 (Aλ,E ) = −L 1 (Aλ,E ). In spectral
theory it is common to use the notation L(E) (or L(λ, E)) for the first Lyapunov
exponent, and to refer to it as the Lyapunov exponent.
There are direct connections between the LE of a Schrödinger cocycle and the
spectral properties of the corresponding operator. We refer the reader to the recent
survey paper [20] by Damanik, and only mention here some examples of such con-
nections. By Johnson’s theorem, x-almost surely we have that if L(E) = 0 then E
is in the spectrum of Hλ (x) and if E is in the spectrum of Hλ (x) but L(E) > 0,
then Aλ,E is not uniformly hyperbolic. More directly relevant to the topic of this
monograph, the quantitative continuity properties (such as Hölder continuity) of the
LE can be transferred to those of the integrated density of states (IDS) of Hλ (x). The
IDS represents the limiting distribution of the eigenvalues of the family of operators
{Hλ (x) : x ∈ X }. Its Hölder continuity is a crucial ingredient in obtaining localization
properties for random Schrödinger operators.
Let us define the IDS and describe its relationship with the LE. Denote by
Pn the coordinate restriction operator to {1, 2, . . . , n} ⊂ Z, and let Hλ(n) (x) :=
Pn Hλ (x) Pn∗ . By ergodicity, the following limit exists for μ-a.e. x ∈ X :
1
Nλ (E) := lim # (−∞, E] ∩ Spectrum of Hλ(n) (x) .
n→∞ n
The function E
→ Nλ (E) is called the integrated density of states of the family of
ergodic operators {Hλ (x) : x ∈ X }.
Thouless formula relates the LE and the IDS essentially via the Hilbert transform:
L(E) = log E − E d N (E ).
R
that any singular integral operator (including the Hilbert transform) preserves the
modulus of continuity of a function, provided that modulus of continuity satisfies
certain conditions. These conditions hold for all of the examples that showed up in
our applications, e.g. for Hölder, weak-Hölder and nearly-Hölder continuity.
Schrödinger cocycles over a Bernoulli shift with nontrivial probability distribution
are automatically irreducible (see [20]), hence by Furstenberg’s theorem they have
positive (maximal) LE for all coupling constant λ
= 0. Quasi-periodic Schrödinger
cocycles (with analytic sampling function) have positive LE for large enough cou-
pling constant λ 1 by Sorets-Spencer theorem and its extensions (see [10]).
Therefore, under these conditions, Le Page’s theorem (in the i.i.d. random case) and
Goldstein-Schlag theorem (in the quasi-periodic case with Diophantine translations)
provide Hölder (or weak-Hölder for the multifrequency torus translation) continuity
of the LE. This in turn implies Hölder (or weak-Hölder, respectively) continuity for
the IDS of the corresponding Schrödinger operator.
The results outlined above concerning random and quasi-periodic Schrödinger
operators have been known for some time. We now describe a more general and
not as well studied model, that of ergodic block Jacobi operators (also called strip
or band lattice operators). These types of operators describe the Hamiltonian of a
quantum particle on a band lattice of the form Z × S, where S may be a finite subset
of any integer lattice or a finite graph. To simplify matters, let S = {1, . . . , l}. Then
the block Jacobi operator acts on l 2 (Z × {1, . . . , l}, R) l 2 (Z, Rl ), that is, on square
#» #»
summable sequences of vectors ψ = {ψ n }n∈Z , by
#» #» #» #»
[Hλ (x) ψ ]n := −(Wn+1 (x) ψ n+1 + Wnt (x) ψ n−1 ) + λ Vn (x) ψ n , (7.4)
where we consider
i. an underlying ergodic base dynamics (X, μ, T ) (e.g. random or quasi-periodic);
ii. Wn (x) = W (T n x) for a bounded measurable ‘weight’ function W : X →
Mat(l, R),
iii. Vn (x) = F(T n x), for a bounded measurable ‘potential’ function F : X →
Mat(l, R).
These types of operators have been studied in [14, 23, 26, 33, 35] and in other
papers, mostly for more particular models (e.g. with W (x) ≡ I or with l = 2). In
[22] we obtained a Sorets-Spencer type theorem for the operator (7.4) over a one
frequency torus translation.
A recent result of Chapman and Stolz [18] provides a Thouless-type formula
(relating the LE and the IDS) which is applicable
in the full generality of the oper-
ator (7.4). Assuming that C W := X logdet W (x)μ(d x) > −∞, their result states
that
(L 1 + · · · + L l )(E) = l log E − E d N (E ) − C W , (7.5)
R
where L 1 , . . . , L l are the first l (i.e. the non-negative) Lyapunov exponents of (7.4).
www.Ebook777.com
7.2 Some Connections to Mathematical Physics 253
These considerations suggest that a similar strategy used to establish Hölder con-
tinuity of the IDS for Schrödinger operators may be applicable to block Jacobi
operators like (7.4) as well.
Question 1 Give sufficient conditions under which the IDS of block Jacobi operators
over random and quasi-periodic dynamics are Hölder continuous.
Note that the cocycle associated with the eigenvalue equation corresponding to
the operator (7.4) is more complex than the Schrödinger cocycles: it is higher dimen-
sional, i.e. Mat(m, R)-valued (m = 2l) and it may have singularities. However, it
fits the general setting of this monograph. This leads to continuity of the LE (with
no restrictions) and to a modulus of continuity (under appropriate conditions) of the
LE as functions of various input data: the energy E (most importantly), the cou-
pling constant λ and even the potential function F(x) or the weight W (x). The issue
of course is to determine those appropriate conditions that ensure the existence of
relevant spectral gaps and thus a modulus of continuity of the LE (or rather of the
Lyapunov spectrum blocks).
We note that some of the crucial ingredients used in establishing localization for
Schrödinger operators are already available or likely to follow from the results in this
monograph. Indeed, the quasi-periodic model satisfies LDT estimates, while in the
random case, a positive answer to Question 1 would provide the Hölder continuity
of the IDS.
We discuss some problems regarding the continuity of the LE for larger spaces of
cocycles and for different types of base dynamics.
In Chap. 6 we established weak-Hölder continuity of the Lyapunov exponents and
of the Oseledets filtration/decomposition for quasi-periodic, analytic cocycles with
singularities. We assumed the translation vector to be Diophantine. We restricted the
problem to the space of non identically singular cocycles. This restriction will be
removed in a project which is currently underway. The approach used there is much
more technically involved than the one used in Chap. 6. However, this extension will
have significant consequences regarding sharp lower bounds for Lyapunov exponents
or simplicity of the Lyapunov spectrum.
The results just mentioned above, on cocycles with singularities, apply to trans-
lations on the torus Td of any number d of variables.
254 7 Further Related Problems
Let us focus now on the one variable case d = 1. In Sect. 6.5 we obtained a
stronger result, nearly-Hölder continuity. In [21], for cocycles without singularities,
we obtained Hölder continuity. The authors of [35] obtained Hölder continuity for
the case of 2 × 2 Jacobi cocycles. This suggests one could expect Hölder continuity
in the setting of Chap. 6 as well.
Furthermore, a very delicate, related question concerns finding the optimal, or at
least an explicit expression for the Hölder exponent α. This problem was studied for
almost Mathieu cocycles (see [9], where α = 21 −), for Schrödinger cocycles given by
a fixed trigonometric polynomial of degree k (see [25] where α = 2k1 −) and also for
more general, Jacobi cocycles with trigonometric polynomials (or analytic functions)
entries (see [35], where α is similarly explicit in terms of certain characteristics of
these entries).
These considerations suggest the following questions.
This last question seems to be very hard even in the case of the Schrödinger cocycles,
which are SL(2, R)-valued.
In Chap. 5 we established the Hölder continuity of the maximal Lyapunov expo-
nent in the space of irreducible random cocycles, in a neighborhood of a cocycle
with a gap between its first two Lyapunov exponents. The same result applies to the
Oseledets filtration and decomposition.
In [8] (see also [36]), the authors establish continuity of the Lyapunov exponents
and of the Oseledets decomposition for GL(2, C)-valued i.i.d. random cocycles,
without any generic assumptions (such as irreducibility) on the cocycle. A similar
result concerning Lyapunov exponents was obtained in [30] for cocycles over Markov
shifts that depend only on one coordinate. Other related results were recently obtained
in [3, 4] for fiber-bunched cocycles. A higher dimensional version of the result in [8]
for Lyapunov exponents was also announced by A. Ávila, A. Eskin and M. Viana.
None of these results provide a modulus of continuity, as the proofs proceed by
contradiction.
These considerations lead to the following natural questions.
Question 4 Consider even the simplest random model, that of cocycles depending
on one coordinate, over a finite alphabet Bernoulli shift.
7.3 Continuity for Other Spaces of Cocycles 255
(a) Let A be a reducible random cocycle whose Lyapunov exponents have a certain
gap pattern, say simple Lyapunov spectrum.
Are the Lyapunov exponents (and the Oseledets filtration/decomposition) still
Hölder, or at least weakly-Hölder continuous locally near A?
(b) In the absence of gaps between Lyapunov exponents, what can be said about the
modulus of continuity of the Lyapunov exponents, even assuming irreducibility?
We note in connection with item (b) above that for both random and quasi-periodic
cocycles, all available quantitative results (including the ones in this book) depend
on the existence of a gap between Lyapunov exponents.
One should then perhaps begin the study of this problem by considering a simple
SL(2, R)-valued cocycle A with zero Lyapunov exponents and seeing if she could
find a sequence Ak → A for which L 1 (Ak ) → 0 at an arbitrarily slow rate.
A testing example for such continuity properties at a cocycle which is both
reducible and has zero Lyapunov exponents (hence no Lyapunov spectrum gaps),
is the following. Let Aε be the cocycle obtained by choosing with equal probability
either X ε or Y , where
sin ε − cos ε 20
X ε := and Y := .
cos ε sin ε 0 21
ii. Let T be an expanding map of the torus, i.e. either the doubling map T x := 2x
mod 1 on T or a hyperbolic toral automorphism T x := M x mod 1 on T2 , for
some M ∈ SL(2, Z). Assume that f (x) is a C 1 -function on T with T f = 0.
J. Bourgain and W. Schlag proved (see [17]) that for an appropriate range I of energies
and for small enough λ, the map I E
→ L 1 (Aλ, f,E ) is Hölder continuous. The
argument depends on the positivity of the first Lyapunov exponents, which is ensured
by the small λ assumption.
for some constants M, K > 0 and s ≥ 1 called the order of the Gevrey-class.
Denote by G s (Td ) the set of all Gevrey-class functions of order s, and note that
s = 1 represents the analytic class, while as s increases, G s (Td ) becomes larger.
S. Klein proved (see [28, 29]) that for large λ, the map E
→ L 1 (Aλ, f,E ) is
weakly-Hölder, provided the potential (i.e. sampling) function f satisfies a generic
transversality condition. That transversality condition is automatically satisfied by
non-constant analytic functions. A similar result holds even without a transversality
assumption on f , but the order s of the Gevrey-class is restricted in that case.
In all of these papers, LDT estimates are derived (and then used to prove continuity
of the LE). Given the availability of these LDT estimates, our abstract continuity
Theorem 3.1 applies as well. Indeed, consider the space of cocycles
C ( f, λ) := {Aλ, f,E : E ∈ I}
In this last section we address a few problems on the continuity of Lyapunov expo-
nents with respect to the frequency vector (for quasi-periodic cocycles), and to the
transition probability measure (for random cocycles).
258 7 Further Related Problems
References
1. A. Ávila, Global theory of one-frequency Schrödinger operators. Acta Math. 215(1), 1–54
(2015). MR 3413976
2. A. Ávila, S. Jitomirskaya, C. Sadel, Complex one-frequency cocycles. J. Eur. Math. Soc.
(JEMS) 16(9), 1915–1935 (2014). MR 3273312
3. L. Backes, A note on the continuity of Oseledets subspaces for fiber-bunched cocycles, preprint
(2015), 1–6
4. L. Backes, A.W. Brown, C. Butler, Continuity of Lyapunov exponents for cocycles with invari-
ant holonomies (2015), 1–34 (preprint)
5. J. Bochi, Discontinuity of the Lyapunov exponent for non-hyperbolic cocycles (1999), 1–14
(preprint)
6. J. Bochi, Genericity of zero Lyapunov exponents. Ergodic Theor. Dynam. Syst. 22(6),break
1667–1696 (2002). MR 1944399 (2003m:37035)
7. J. Bochi, M. Viana, Uniform (projective) hyperbolicity or no hyperbolicity: a dichotomy for
generic conservative maps. Ann. Inst. H. Poincaré Anal. Non Linéaire 19(1), 113–123 (2002).
MR 1902547 (2003f:37040)
8. C. Bocker-Neto, M. Viana, Continuity of Lyapunov exponents for random 2d matrices (2010),
1–38 (to appear in Ergodic Theory and Dynamical Systems) (preprint)
9. J. Bourgain, Hölder regularity of integrated density of states for the almost Mathieu operator
in a perturbative regime. Lett. Math. Phys. 51(2), 83–118 (2000). MR 1774640 (2003a:47072)
10. J. Bourgain, Green’s function estimates for lattice Schrödinger operators and applications.
Annals of Mathematics Studies, vol. 158 (Princeton University Press, Princeton, NJ, 2005).
MR 2100420 (2005j:35184)
11. J. Bourgain, Positivity and continuity of the Lyapounov exponent for shifts on Td with arbitrary
frequency vector and real analytic potential. J. Anal. Math. 96, 313–355 (2005). MR 2177191
(2006i:47064)
12. J. Bourgain, On the Furstenberg measure and density of states for the Anderson-Bernoulli
model at small disorder. J. Anal. Math. 117, 273–295 (2012). MR 2944098
13. J. Bourgain, An application of group expansion to the Anderson-Bernoulli model. Geom. Funct.
Anal. 24(1), 49–62 (2014). MR 3177377
14. J. Bourgain, S. Jitomirskaya, Anderson localization for the band model. Geometric Aspects
of Functional Analysis, Lecture Notes in Mathematics, vol. 1745 (Springer, Berlin, 2000), pp.
67–79. MR 1796713 (2002d:81053)
15. J. Bourgain, Continuity of the Lyapunov exponent for quasiperiodic operators with analytic
potential. J. Statist. Phys. 108(5–6), 1203–1218 (2002). Dedicated to David Ruelle and Yasha
Sinai on the occasion of their 65th birthdays
16. J. Bourgain, M. Goldstein, W. Schlag, Anderson localization for Schrödinger operators on
Z with potentials given by the skew-shift. Comm. Math. Phys. 220(3), 583–621 (2001). MR
1843776 (2002g:81026)
260 7 Further Related Problems
17. J. Bourgain, W. Schlag, Anderson localization for Schrödinger operators on Z with strongly
mixing potentials. Comm. Math. Phys. 215(1), 143–175 (2000). MR 1800921 (2002d:81054)
18. J. Chapman, G. Stolz, Localization for random block operators related to the XY spin chain.
Ann. Henri Poincaré 16(2), 405–435 (2015). MR 3302603
19. W. Craig, B. Simon, Log Hölder continuity of the integrated density of states for stochastic
Jacobi matrices. Comm. Math. Phys. 90(2), 207–218 (1983). MR 714434 (85k:47012)
20. D. Damanik, Schrödinger operators with dynamically defined potentials: a survey, preprint
(2015), 1–80, to appear in Ergodic Theory and Dynamical Systems
21. P. Duarte, S. Klein, Continuity of the Lyapunov exponents for quasiperiodic cocycles. Comm.
Math. Phys. 332(3), 1113–1166 (2014). MR 3262622
22. P. Duarte, S. Klein, Positive Lyapunov exponents for higher dimensional quasiperiodic cocy-
cles. Comm. Math. Phys. 332(1), 189–219 (2014). MR 3253702
23. I.Ya. Gol dsheı̆d, E. Sorets, Lyapunov exponents of the Schrödinger equation with quasi-
periodic potential on a strip. Comm. Math. Phys. 145(3), 507–513 (1992). MR 1162358
(93f:39007)
24. M. Goldstein, W. Schlag, Hölder continuity of the integrated density of states for quasi-periodic
Schrödinger equations and averages of shifts of subharmonic functions. Ann. of Math. (2)
154(1), 155–203 (2001). MR 1847592 (2002h:82055)
25. M. Goldstein, W. Schlag, Fine properties of the integrated density of states and a quantitative
separation property of the Dirichlet eigenvalues. Geom. Funct. Anal. 18(3), 755–869 (2008).
MR 2438997 (2010h:47063)
26. A. Haro, J. Puig, A Thouless formula and Aubry duality for long-range Schrödinger skew-
products. Nonlinearity 26(5), 1163–1187 (2013). MR 3043377
27. S. Jitomirskaya, C.A. Marx, Analytic quasi-perodic cocycles with singularities and the Lya-
punov exponent of extended Harper’s model. Comm. Math. Phys. 316(1), 237–267 (2012).
MR 2989459
28. S. Klein, Anderson localization for the discrete one-dimensional quasi-periodic Schrödinger
operator with potential defined by a Gevrey-class function. J. Funct. Anal. 218(2), 255–292
(2005). MR 2108112 (2005m:82070)
29. S. Klein, Localization for quasiperiodic Schrödinger operators with multivariable Gevrey
potential functions. J. Spectr. Theor. 4, 1–53 (2014)
30. E.C. Malheiro, M. Viana, Lyapunov exponents of linear cocycles over Markov shifts. Stoch.
Dyn. 15(3), 1550020, 27 (2015). MR 3349975
31. É. Le Page, Régularité du plus grand exposant caractéristique des produits de matrices aléa-
toires indépendantes et applications. Annales de l’institut Henri Poincaré (B) Probabilités et
Statistiques 25(2), 109–142 (1989) (fre)
32. Y. Peres, Analytic dependence of Lyapunov exponents on transition probabilities. Lyapunov
Exponents (Oberwolfach, 1990), Lecture Notes in Mathematics, vol. 1486 (Springer, Berlin,
1991), pp. 64–80. MR 1178947 (94c:60116)
33. B. Simon, M. Taylor, Harmonic analysis on SL(2, R) and smoothness of the density of states in
the one-dimensional Anderson model. Comm. Math. Phys. 101(1), 1–19 (1985). MR 814540
(87i:82087)
34. K. Tao, Continuity of Lyapunov exponent for analytic quasi-periodic cocycles on higher-
dimensional torus. Front. Math. China 7(3), 521–542 (2012). MR 2915794
35. K. Tao, M. Voda, Hölder continuity of the integrated density of states for quasi-periodic jacobi
operators, preprint (2015), 1–19
36. M. Viana, Lectures on Lyapunov Exponents. Cambridge Studies in Advanced Mathematics
(Cambridge University Press, 2014)
37. Y. Wang, J. You, Examples of discontinuity of Lyapunov exponent in smooth quasiperiodic
cocycles. Duke Math. J. 162(13), 2363–2412 (2013). MR 3127804
38. L.-S. Young, Lyapunov exponents for some quasi-periodic cocycles. Ergodic Theor. Dynam.
Syst. 17(2), 483–504 (1997). MR 1444065 (98c:58123)
Index
A E
Abstract continuity theorem, 15, 86 Expansion rift, 43
Abstract setting Expected value, 163
assumptions (A1)–(A4), 179 Exterior algebra, 25
assumptions (B1)–(B7), 178 k-vector, 25
constants, 180 norm, 26
space of observed Markov systems, 178 simple, 25
theorem, 185 unit, 26
Almost everywhere Cauchy, 121 dual wedge product, 27
Avalanche exterior power, 27
principle, 1, 63, 68 Hodge star operator, 27
complex version, 78 inner product, 26
flag version, 76 volume element, 27
practical version, 73 wedge product, 25
times, 3, 123
F
B Fekete’s subadditive lemma, 118
Birkhoff’s ergodic theorem, 7, 118, 223 Filtration
quantitative, 225 finite scale, 114
sharp, 238 measurable, 113
Bundle Oseledets, 114
base, 7 Flag, 117, 167
fiber, 7 α-angle, 43
measurable, 7 length, 31
trivial, 7 manifold, 31, 117
unit measurable section, 129 metrics ρ, d, δ, 31
orthogonal complement, 31, 117
orthogonal hyperplane, 43
D pull back action, 32
Deviation functions push forward action, 32
set measure, 13, 85 signature, 31, 116
size, 13, 85
Diophantine condition, 17, 208
DCt , 208 G
DC, 208 Grassmannian, 28
strong, 237, 238 α-angle, 41
© Atlantis Press and the author(s) 2016 261
P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6
262 Index
www.Ebook777.com