You are on page 1of 271

Free ebooks ==> www.Ebook777.

com

Atlantis Studies in Dynamical Systems


Series Editors: H. Broer · B. Hasselblatt

Pedro Duarte
Silvius Klein

Lyapunov Exponents
of Linear Cocycles
Continuity via Large Deviations · Volume 3

www.Ebook777.com
Free ebooks ==> www.Ebook777.com

Atlantis Studies in Dynamical Systems

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.

More information about this series at http://www.atlantis-press.com


Pedro Duarte Silvius Klein

Lyapunov Exponents
of Linear Cocycles
Continuity via Large Deviations
Free ebooks ==> www.Ebook777.com

Pedro Duarte Silvius Klein


Faculdade de Ciências Department of Mathematical Sciences
Universidade de Lisboa Norwegian University of Science and
Lisbon Technology (NTNU)
Portugal Trondheim
Norway

Atlantis Studies in Dynamical Systems


ISBN 978-94-6239-123-9 ISBN 978-94-6239-124-6 (eBook)
DOI 10.2991/978-94-6239-124-6
Library of Congress Control Number: 2016933219

© Atlantis Press and the author(s) 2016


This book, or any parts thereof, may not be reproduced for commercial purposes in any form or by any
means, electronic or mechanical, including photocopying, recording or any information storage and
retrieval system known or to be invented, without prior permission from the Publisher.

Printed on acid-free paper

www.Ebook777.com
In memory of João Santos Guerreiro and
Ricardo Mañé, professors whose friendship
and intelligence I miss
Pedro Duarte

To Florin Popovici and Șerban Strătilă who


taught me to seek and to appreciate good
mathematical exposition
Silvius Klein
Preface

The aim of this monograph is to present a general method of proving continuity


of the Lyapunov exponents (LE) of linear cocycles.
The method consists of an inductive procedure that establishes continuity of
relevant quantities for finite, larger and larger number of iterates of the system. This
leads to continuity of the limit quantities, the LE. The inductive procedure is based
upon a deterministic result on the composition of a long chain of linear maps called
the Avalanche Principle (AP). A geometric approach is used to derive a general
version of this principle.
The main assumption required by this method is the availability of appropriate
large deviation type (LDT) estimates for quantities related to the iterates of the base
and fiber dynamics associated with the linear cocycle. Crucial for our approach is
the uniformity in the data of these estimates.
We derive such LDT estimates for various models of random cocycles (over
Bernoulli and Markov systems) and quasi-periodic cocycles (defined by one or
multivariable torus translations). The random model, treated under an irreducibility
assumption, uses an existing functional analytic approach which we adapt so that it
provides the required uniformity of the estimates. The quasi-periodic model uses
harmonic analysis and it involves the study of (pluri) subharmonic functions.
This method has its origins in a paper of M. Goldstein and W. Schlag which
proves continuity of the Lyapunov exponent for the one-parameter family of
quasi-periodic Schrödinger cocycles, assuming a uniform lower bound on the
exponent. This is where the first version of the Avalanche Principle appeared, along
with the use and proof of the relevant LDT estimate.
The present work expands upon their approach in both depth and breadth.
Moreover, it reduces the general problem of proving continuity of the LE to one of
a different nature—proving LDT estimates. This may be treated independently and
by means specific to the underlying base dynamic of the the cocycle.
Our geometric approach to the AP also gives rise to a mechanism for studying
the most expanding singular direction of the composition of a long chain of linear
maps. This allows us to obtain a new proof of the Multiplicative Ergodic

vii
viii Preface

Theorem of Oseledets. Moreover, assuming the availability of the same LDT


estimates, this extension of the AP leads to continuity properties of the Oseledets
filtration and decomposition.
Most of the results presented in this research monograph are new. We assume
the reader to have a certain degree of familiarity with basic dynamical systems and
ergodic theory notions. The relevant concepts and definitions needed for the for-
mulation of the main results are introduced in Chap. 1. While each subsequent
chapter is to some extent self-contained and it may be read independently of the
rest, all the arguments in this work are based upon the results in Chaps. 2 and 3.
Besides the formulation and the proof of the AP, Chap. 2 contains Lipschitz esti-
mates on certain Grassmann geometrical quantities that are crucial in Chap. 4,
where we study the Oseledets filtration and decomposition and their continuity
properties. In Chap. 3 we establish the abstract continuity theorem (ACT) of the LE
and some other related technical results. In Chaps. 5 and 6, under appropriate
assumptions, we derive the relevant LDT estimates for random and respectively
quasi-periodic cocycles. The general results in Chaps. 3 and 4 are then applicable to
these models, and they imply continuity properties of the LE and of the Oseledets
filtration and decomposition for the corresponding spaces of cocycles.
Our work concludes in Chap. 7 with a list of related open problems, some of
which may be treated using the methods described in this monograph.
The first author was supported by National Funding from FCT—Fundação para
a Ciência e a Tecnologia, under the project: UID/MAT/04561/2013.
The second author was supported by the Norwegian Research Council project
no. 213638, “Discrete Models in Mathematical Analysis”.
Both authors are grateful to the Faculty of Sciences of the University of Lisbon
(FCUL) and to the Norwegian University of Science and Technology (NTNU) for
the support received and for facilitating their collaboration on this monograph.
We would like to thank José Pedro Gaivão and Wilhelm Schlag for reading
through parts of the manuscript.
And last but not least, many thanks to Teresa, Zé, Jaime, Daniel and Jaqueline
for their understanding.

Lisbon Pedro Duarte


Trondheim Silvius Klein
January 2016
Contents

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

3 Abstract Continuity of Lyapunov Exponents . . . . . . . . . . . . . . . . . 81


3.1 Definitions, the Abstract Setup and Statement . . . . . . . . . . . . . . 81
3.1.1 Cocycles and Observables . . . . . . . . . . . . . . . . . . . . . . 82
3.1.2 Large Deviations Type Estimates. . . . . . . . . . . . . . . . . . 84
3.1.3 Abstract Continuity Theorem of the Lyapunov
Exponents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.2 Upper Semicontinuity of the Top Lyapunov Exponent . . . . . . . . 86
3.3 Finite Scale Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.4 The Inductive Step Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.5 General Continuity Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.6 Modulus of Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4 The Oseledets Filtration and Decomposition . . . . . . . . . . . . . . ... 113
4.1 Introduction and Statements . . . . . . . . . . . . . . . . . . . . . . . ... 113
4.2 The Ergodic Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 116
4.2.1 Review of Grassmann Geometry Concepts
and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.2.2 The Ergodic Theorems of Birkhoff and Kingman . . . . . . 118
4.2.3 The Multiplicative Ergodic Theorem . . . . . . . . . . . . . . . 120
4.3 Abstract Continuity Theorem of the Oseledets Filtration . . . . . . . 141
4.3.1 Continuity of the Most Expanding Direction . . . . . . . . . . 142
4.3.2 Spaces of Measurable Filtrations and Decompositions . . . 149
4.3.3 Continuity of the Oseledets Filtration . . . . . . . . . . . . . . . 153
4.3.4 Continuity of the Oseledets Decomposition. . . . . . . . . . . 154
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5 Large Deviations for Random Cocycles. . . . . . . . . . . . . . . . . . . . . 161
5.1 Introduction and Statements . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.1.1 Description of the Model . . . . . . . . . . . . . . . . . . . . . . . 161
5.1.2 The Spectral Method . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.1.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.2 An Abstract Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.2.1 The Assumptions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.2.2 An Abstract Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.3 The Proof of LDT Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.3.1 Base LDT Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.3.2 Fiber LDT Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 191
5.4 Deriving Continuity of the Lyapunov Exponents . . . . . . . . . . . . 201
5.4.1 Proof of the Continuity . . . . . . . . . . . . . . . . . . . . . . . . 201
5.4.2 Some Generalizations. . . . . . . . . . . . . . . . . . . . . . . . . . 202
5.4.3 Method Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

www.Ebook777.com
Contents xi

6 Large Deviations for Quasi-Periodic Cocycles . . . . . . . . . . . . . . . . 207


6.1 Introduction and Statements . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6.1.1 Description of the Model . . . . . . . . . . . . . . . . . . . . . . . 207
6.1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6.2 Estimates on Unbounded Pluri-Subharmonic Functions. . . . . . . . 212
6.2.1 The Uniform Łojasiewicz Inequality . . . . . . . . . . . . . . . 212
6.2.2 Uniform L2 -Bounds on Analytic Functions . . . . . . . . . . . 215
6.2.3 Estimates on Unbounded Subharmonic Functions . . . . . . 218
6.2.4 Base LDT Estimates for Pluri-Subharmonic
Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
6.3 The Proof of the Fiber LDT Estimate . . . . . . . . . . . . . . . . . . . . 230
6.3.1 Uniform Measurements on the Cocycle . . . . . . . . . . . . . 230
6.3.2 The Nearly Almost Invariance Property . . . . . . . . . . . . . 232
6.3.3 The Statement and Proof of the LDT . . . . . . . . . . . . . . . 234
6.4 Deriving Continuity of the Lyapunov Exponents . . . . . . . . . . . . 236
6.5 Refinements in the One-Variable Case . . . . . . . . . . . . . . . . . . . 237
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
7 Further Related Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
7.1 Limitations and Counterexamples . . . . . . . . . . . . . . . . . . . . . . 247
7.2 Some Connections to Mathematical Physics . . . . . . . . . . . . . . . 250
7.3 Continuity for Other Spaces of Cocycles . . . . . . . . . . . . . . . . . 253
7.4 Continuity with Respect to Other Parameters. . . . . . . . . . . . . . . 257
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Acronyms

ACT Abstract continuity theorem


AP Avalanche principle
IDS Integrated density of states
LDT Large deviation type
LE Lyapunov exponent
MET Multiplicative ergodic theorem
MPDS Measure preserving dynamical system
MPT Measure preserving transformation

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 ,

© Atlantis Press and the author(s) 2016 1


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_1
2 1 Introduction

then the expansion rift behaves almost multiplicatively:

ρ(g0 , . . . , gn−1 )  ρ(g0 , g1 ) . . . ρ(gn−2 , gn−1 ) ,

or equivalently

g1 g0  . . . gn−1 gn−2 


gn−1 . . . g1 g0   . (1.1)
g1  . . . gn−2 

Moreover, the most expanding direction of the matrix product gn−1 . . . g1 g0 is


close to the most expanding direction of the first matrix g0 , that is:
κ
d(v(gn−1 . . . g1 g0 ), v(g0 ))  . (1.2)
ε

Let (X, μ, T ) be an ergodic dynamical system, and let A : X → Mat(m, R) be a


linear cocycle over this base dynamics. We denote by A(n) (x) its nth iterate and refer
to it as a block of length n. By Kingman’s ergodic theorem, the Lyapunov exponents
of A, denoted by L k (A), can be characterized as the almost everywhere limits

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).

Fig. 1.1 Blocks of different time scales in an inductive procedure


1.1 Prologue 3

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

d(v(A(n) (x)), v(A(n 0 ) (x)))  1 . (1.3)

By Kingman’s ergodic theorem for a.e. x ∈ X 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

d(v(A(n) (x)), v(A(n 0 ) (x)))  1.

Fig. 1.2 Consecutive blocks between avalanche times


4 1 Introduction

If n 0  n 0 , then (1.3) is proven. Otherwise, n 0 n 0 , and we repeat the procedure


above by further dividing up the the long block A(n 0 ) (x) of length n 0 , and so on. The
errors accumulated at each step form a convergent series whose sum is arbitrarily
small, provided we choose the initial scale n 0 to be large enough.
We may now define the μ-almost everywhere limit

v(∞) (A)(x) = lim v(A(n) (x)),


n→∞

and refer to it as the most expanding direction of A.


For simplicity, let us assume that the Lyapunov exponents of A have the following
gap pattern:

L 1 (A) = L 2 (A) > L 3 (A) > L 4 (A) = · · · = L m (A).

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

where ι(n, ε) → 0 fast as n → ∞.


Using conclusion (1.1) of the AP in the inductive procedure described before, and
after taking logarithms, we get that the finite scale quantity n1 logA(n) (x) is close to
certain Birkhoff averages of similar quantities at scales of order n 0  n, off of a very
small set of phases B. Moreover, (1.4) gives us a precise quantitative estimate on the
exceptional set B. This in turn leads to a quantitative description of the convergence
of the finite scale quantities (Lyapunov exponents and Oseledets filtrations) to their
limits, which implies a modulus (speed) of convergence.
If the LDT estimates (1.4) hold uniformly in a neighborhood of the cocycle A,
then the speed of convergence to the LE and the Oseledets filtration are also uniform
in a neighborhood of A. Since for a fixed initial finite scale, these quantities behave
continuously, we can transfer the continuity property to the limits.
1.1 Prologue 5

Assuming good behavior on average of the relevant quantities at an initial scale,


the role of the LDT estimates in this scheme is to provide a large enough set of phases
for which a successful, long enough run of the AP can be guaranteed. This then leads
to a good behavior at the next scale for many phases, hence on average.

1.2 The Main Concepts

Given a probability space (X, F, μ), a measure preserving transformation is an F-


measurable map T : X → X such that

μ(T −1 (A)) = μ(A), for all A ∈ F .

A measure preserving dynamical system (MPDS) is any triple (X, μ, T ) where


(X, μ) is a probability space (the σ -field F is implicit to X ) and T : X → X is
a measure preserving transformation.
We refer to elements of X as phases. The sequence of iterates {T n x}n≥0 is called
the orbit of the phase x.
Definition 1.1 We say that (X, μ, T ) is ergodic if there is no T -invariant measurable
set A = T −1 (A) such that 0 < μ(A) < 1.
Definition 1.2 We say that (X, μ, T ) is mixing when for all A, B ∈ F,

lim μ(A ∩ T −n (B)) = μ(A) μ(B) .


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

If (K , μ) is strongly mixing then the Markov shift (X, P, T ) is mixing. When Σ is


finite, this strong mixing property of (K , μ) is actually equivalent to the the mixing
property of (X, P, T ). These relations are discussed in Chap. 5.
Given a probability space (X, μ), we denote by L 1 (X, μ) the space of measurable
functions ϕ : X → C with finite first moment

Eμ (|ϕ|) := |ϕ| dμ < +∞ .
X

These functions will be called observables.


A simplified version of Birkhoff’s ergodic theorem reads as follows:
1.2 The Main Concepts 7

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

Kingman’s Ergodic Theorem below generalizes Birkhoff’s ergodic theorem.


Theorem 1.2 Let (X, μ, T ) be an ergodic MPDS. Given a sequence of measurable
functions f n : X → [−∞, ∞] such that f 1+ ∈ L 1 (X, μ) and

f n+m ≤ f n + f m ◦ T n for all n, m ≥ 0 ,



the sequence { f n dμ}n≥0 is sub-additive, i.e.,
  
f n+m dμ ≤ f n dμ + f m dμ for all n, m ≥ 0 ,

and for μ-a.e. x ∈ X , we have


 
1 1 1
lim f n (x) = lim f n dμ = inf f n dμ ∈ [−∞, +∞) .
n→∞ n n→∞ n n≥1 n

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

Gr k (Rm ) := { E ∈ Gr(Rm ) : dim E = k } .

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) } .

When E : X → Gr(Rm ) takes values in Gr k (Rm ) we say that B has constant


dimension k. The set X is called the base of B, and the linear subspace B(x) := E(x)
is called the fiber at the base point x ∈ X . To each measurable bundle B we associate
the projection onto the base πB : B → X , πB (x, v) = x.
The set B = X × Rm is a measurable bundle over X , called a trivial bundle.
Given two measurable bundles B and B  over X , we say that B is a sub-bundle of
B  when B ⊂ B  . By definition, every measurable bundle is a sub-bundle of some
trivial bundle.
Consider an MPDS (X, μ, T ) and a measurable bundle B over X .
Free ebooks ==> www.Ebook777.com
8 1 Introduction

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

Given a cocycle A on X ×Rm over an MPDS (X, μ, T ), a sub-bundle B ⊂ X ×Rm


is called A-invariant when for μ almost every x ∈ X , A(x) B(x) ⊂ B(T x). When
B is A-invariant, the restriction of FA : X × Rm → X × Rm to B is a linear cocycle
on B over (X, μ, T ).
The forward iterates FAn of a linear cocycle FA : B → B are given by FAn (x, v) =
(T x, A(n) (x)v), where
n

A(n) (x) := A(T n−1 x) . . . A(T x) A(x) (n ∈ N) .

When the base map T : X → X is invertible, we define backward iterates, even if


the linear maps A(x) are non-invertible, setting

A(−n) (x) := A(n) (T −n x)+ (n ∈ N) ,

where M + denotes the pseudo-inverse of a linear map M (see Definition 4.10).

Definition 1.5 A quasi-periodic cocycle is any cocycle A : Td → Mat(m, R) over


an ergodic torus translation T : Td → Td . If T x := x + ω (mod 1) then ω ∈ Rd is
called the frequency vector of the cocycle.

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 .

Definition 1.7 Let Σ be a compact metric space, K a Markov kernel on Σ, μ a K -


stationary probability measure on Σ and let P be the associated probability measure
on the space of sequences X = Σ Z .
A random Markov cocycle is any cocycle A : X → Mat(m, R) over the Markov
shift (X, P, T ), where the function A depends only on the coordinates x0 and x1 .

www.Ebook777.com
1.2 The Main Concepts 9

Consider a measurable bundle B of constant dimension m. A measurable filtration


of B is any finite sequence of sub-bundles B = B1 ⊃ B2 ⊃ · · · ⊃ Bk , with
constant decreasing dimensions. Given a linear cocycle FA : B → B, we say that
a measurable filtration of B is A-invariant if all the sub-bundles of the filtration are
A-invariant.
The first Multiplicative Ergodic Theorem (MET) of Oseledets [23] describes the
fiber asymptotic behavior of a linear cocycle over a possibly non-invertible ergodic
MPDS.
Theorem 1.3 (MET I) Let (X, μ, T ) be an ergodic MPDS and FA : B → B be a
μ-integrable linear cocycle. There exist numbers λ1 > λ2 > · · · > λk ≥ −∞ and
a measurable filtration B = B1 ⊃ B2 ⊃ · · · ⊃ Bk ⊃ Bk+1 = X × {0} invariant
under the cocycle such that for μ-almost every x ∈ X and every v ∈ B j (x)\B j+1 (x),

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

Let (X, μ, T ) be an ergodic MPDS. The numbers λ j in both Oseledets theorems


are called the distinct Lyapunov exponents of the linear cocycle.
Because norms behave sub-multiplicatively with matrix products, the sequence
of functions f n (x) := logA(n) (x) is sub-additive, in the sense that

f n+m ≤ f n + f m ◦ T n for all n, m ∈ 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:

L 1 (A) ≥ L 2 (A) ≥ · · · ≥ L m (A) ≥ −∞ .

These numbers are called the


 repeated Lyapunov exponents of A.
We define L (n) (n)
1 (A) := X n logA  dμ and call it the (maximal) finite scale
1

Lyapunov exponent of A. By Kingman’s ergodic theorem L 1 (A) = limn→∞ L (n) 1 (A).


All previous concepts correspond to basic notions in theory of linear cocycles.
The following concepts are specifically targeted to the statement of the main results
of this monograph.

Definition 1.8 A space of measurable cocycles C is a collection {(Cm , dist)}m∈N of


metric spaces such that:
(1) Each Cm is a space of F-measurable functions A : X → Mat m (R) with uni-
formly bounded norm, i.e., A ∈ L ∞ (X, μ).
(2) For all A ∈ Cm , we have ∧k A ∈ C(mk) , and the map Cm → C(mk) , A → ∧k A is
locally Lipschitz.
(3) For all A, B ∈ Cm , dist(B, A) ≥ B − A L ∞ .

By A ∈ C we mean that A ∈ Cm for some m ∈ N.

Definition 1.9 Let 1 < p ≤ ∞. A cocycle A ∈ C is called L p -bounded if there is


C < ∞, which we call its L p -bound, such that for all n ≥ 1 we have:
1 
 
 logA(n)  p < C .
n L

A cocycle A ∈ Cm is called uniformly L p - bounded if there are δ = δ(A) > 0 and


C = C(A) < ∞ such that for all B ∈ Cm with dist(B, A) < δ and all n ≥ 1 we
have: 1 
 
 logB (n)  p < C .
n L

Given a cocycle A ∈ C and an integer N ∈ N, denote by F N (A) the lattice


(w.r.t. union and intersection) generated by the sets {x ∈ X : A(n) (x) ≤ c} or
{x ∈ X : A(n) (x) ≥ c} where c ≥ 0 and 0 ≤ n ≤ N .
1.2 The Main Concepts 11

Let Ξ be a set of observables ξ : X → R.

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).

1.3 The Continuity Problem

Let X be a phase space and let F be a σ -algebra on X . A linear cocycle is determined


by a probability measure μ on (X, F ), an ergodic transformation T : X → X that
preserves μ, and a fiber action given by a measurable function A : X → Mat(m, R)
for some m ≥ 1. Therefore, the Lyapunov exponents (as well as the Oseledets
filtration and decomposition) may be regarded as functions of one of these input
data, individually or jointly.
Understanding the behavior of the LE as the input data is perturbed is considered
a difficult problem for most systems. Lyapunov exponents are known to be upper
semicontinuous functions in very general settings. However, their continuity depends
on the space of cocycles and its topology, and in some sense is rare (see [5]), unless
strong regularity assumptions on the cocycles are made.
In this monograph, the probability measure μ and the transformation T are fixed,
and we identify a linear cocycle with the function A : X → Mat(m, R) defining its
fiber action. We study the continuity properties of the Lyapunov exponents and of
the Oseledets filtration and decomposition as the function A varies in an appropriate
space which is endowed with a topology at least as fine as the uniform topology.
The Oseledets filtration and decomposition are in general only measurable in the
phase x. Their continuity as A varies is then understood in an average in x sense,
or in probability. Moreover, we are interested in obtaining quantitative continuity
properties, i.e. an explicit modulus of continuity.
We establish such continuity properties for any systems satisfying appropriate
large deviation type estimates, which are defined in the next section. Our main appli-
cations are to various types of random and quasi-periodic cocycles, for which the
relevant LDT estimates will be derived.
A related theme is that of joint continuity in various input data, e.g. in translation
vector (which determines the base dynamics) and fiber action function for quasi-
periodic cocycles, or in probability measure and fiber action function for i.i.d random
12 1 Introduction

cocycles. We will not study these types of problems in this monograph, but we will
discuss them briefly in the last chapter.

1.4 Large Deviations Type Estimates

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)

A fundamental result in harmonic analysis, concerning functions of bounded mean


oscillation (BMO), is John-Nirenberg’s inequality. Given f ∈ L 1 (T) let
 
 f BMO := sup  f −  f  I  I ,
I


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

where I (ε) := supt>0 (t ε − log M(t) + t μ) is called the rate function.


1.4 Large Deviations Type Estimates 13

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

and for any open set G ⊂ R,



1 1
lim inf log P Sn ∈ G ≥ − inf I (x) .
n→+∞ rn n x∈G

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

(2) The set I is a convex cone consisting of continuous functions ι : (0, ∞) →


(0, ∞) such that for every ι ∈ I,
(a) ι(t) is strictly decreasing and limt→+∞ ι(t) = 0,
(b) e−c0 t < ι(t) < t −10 as t → ∞, for some constant c0 > 0,
φι (2t)
(c) lim < 2, where φι (t) denotes the inverse of t → ψι (t) :=
t→∞ φι (t)
p−1
t ι(t)− 2 p−1 .
(3) For all ε > 0 there exists p = (n 0 , ε, ι) ∈ P such that ε(n 0 ) ≤ ε.

We will use the notation εn := ε(n) and ιn := ι(n) for integers n.


Condition (2)(b) says that any ι(t) ∈ I decreases to 0, as t → ∞, at least like a
power and at most like an exponential. Condition (2)(c) imposes a not too fast growth
on the inverse function φι (t).
In our applications of the main results, the constant p = 2 or p = ∞ and the set
of LDT parameters P = N × E × I, where E and I are sets of functions of one of the
following types.

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

(c) nearly exponentials ι(t) ≡ M e−c t/(log t) ,


b

where M < ∞ and c, b > 0.

We now define the base and fiber LDT estimates.

Definition 1.13 An observable ξ : X → R satisfies a base-LDT estimate w.r.t. a


space of parameters P if for every ε > 0 there is (n 0 , ε, ι) ∈ P 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 .
⎩ n j=0 X ⎭

Definition 1.14 A measurable cocycle A : X → Mat(m, R) satisfies a fiber-LDT


estimate w.r.t. a space of parameters P if for every ε > 0 there is (n 0 , ε, ι) ∈ P such
that for all n ≥ n 0 we have εn ≤ ε and
 
1 
μ x ∈ X:  (n) (n) 
logA (x) − L 1 (A) > εn < ιn .
n
1.4 Large Deviations Type Estimates 15

We use LDT estimates to prove continuity of the LE as functions of the cocycle,


where the space of cocycles is endowed with a distance. For this we need a stronger
form of the fiber-LDT, one that is uniform in a neighborhood of the cocycle, in the
sense that the estimate above holds with the same LDT parameter for all nearby
cocycles.

Definition 1.15 A measurable cocycle A satisfies a uniform fiber-LDT if for all


ε > 0 there are δ > 0 and (n 0 , ε, ι) ∈ P such that if B is a measurable cocycle with
dist(B, A) < δ and if n ≥ n 0 then εn ≤ ε and
 
1 
μ x ∈ X :  logB (n) (x) − L (n)
1 (B)  > εn < ιn .
n

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.

1.5 Summary of Results

The main result of this monograph is an Abstract Continuity Theorem (ACT).

Theorem 1.6 Consider an ergodic MPDS (X, μ, T ), a space of measurable cocy-


cles C , a set of observables Ξ , a constant 1 < p ≤ ∞, a set of LDT parameters
P = P( p) with corresponding spaces of deviation functions E, I and assume the
following:
1. Ξ is compatible with every cocycle A ∈ C .
2. Every observable ξ ∈ Ξ satisfies a base-LDT w.r.t. P.
3. Every A ∈ C with L 1 (A) > −∞ is uniformly L p -bounded.
4. Every cocycle A ∈ C with L 1 (A) > L 2 (A) satisfies a uniform fiber-LDT w.r.t. P.
Then every Lyapunov exponent L k : Cm → [−∞, ∞), 1 ≤ k ≤ m, is continuous.
Moreover, if for A ∈ Cm and 1 ≤ k ≤ m we have L k (A) > L k+1 (A), then there
p−1
exists a neighborhood V of A, and a modulus of continuity ω(h) := [ι (c log h1 )] 2 p−1
with ι = ι(A) ∈ I and c = c(A) > 0 such that:
(a) the map V  B → Λk (B)= (L 1 +· · ·+ L k )(B) ∈ R has modulus of continuity
ω, i.e., Λk (B1 ) − Λk (B2 ) ≤ ω(dist(B1 , B2 )), for all B1 , B2 ∈ V;
16 1 Introduction

(b) If T : X → X is invertible then for some α = α(A) > 0 and for all B1 , B2 ∈ V,

μ {x ∈ X : d(E k± (B1 )(x), E k± (B2 )(x)) > dist(B1 , B2 )α } < ω(dist(B1 , B2 )) ,

where d refers to the distance on the Grassmann manifold.

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.

Definition 1.16 A measurable section V : Σ → Gr(Rm ) is called A-invariant when

A(xn−1 , xn ) V (xn−1 ) = V (xn ) for Pμ -a.e. x = (xn )n∈Z .

Assuming (K , μ) is strongly mixing, the ergodicity of this Markov kernel implies


that the subspaces V (x) have constant dimension μ-a.e., denoted by dim(V ). We
say that this family is proper if 0 < dim(V ) < d.
We now introduce the concepts of irreducible and totally irreducible cocycle.

Definition 1.17 A cocycle A ∈ B∞ m (K ) is called irreducible w.r.t. (K , μ) if it admits


no measurable proper A-invariant section V : Σ → Gr(Rm ). A cocycle A ∈ B∞ m (K )
is called totally irreducible w.r.t. (K , μ) if the exterior powers ∧k A are irreducible
for all 1 ≤ k ≤ m − 1.

We denote by Im∞ (K ) the subspace of totally irreducible cocycles in B∞ m (K ).


This set is open (see Proposition 5.3).
In this application we consider the family C = {(Im∞ (K ), d∞ )}m∈N as the space
of measurable cocycles, the set E of constant deviation functions ε(t) ≡ ε, with 0 <
ε < 1, and the set of exponential functions I = { ι(t) ≡ M e−c t : M < ∞, c > 0 }
to measure the deviation sets.
We formulate the continuity statements, which follow from Theorem 5.1.

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).

Theorem 1.8 Given a Markov kernel on the compact metric space Σ, a K -


stationary measure μ and A ∈ B∞
m (K ), assume

(1) (K , μ) is strongly mixing,


(2) A is irreducible,
(3) L 1 (A) > L 2 (A).
Then there exists a neighborhood V of A in B∞ m (K ) and there exist C < ∞, k > 0
and ε0 > 0 such that for all 0 < ε < ε0 , B ∈ V and n ∈ N,

1 
Pμ  logB (n)  − L 1 (B) > ε ≤ C e−k ε n .
2

In Chap. 6 we apply the ACT to the space of quasi-periodic cocycles.


Let (Td , μ, T ) be a torus translation with frequency vector ω ∈ Rd . We warn
the reader that the letter ω denotes different quantities, hence its meaning should be
inferred from the context. The letter μ denotes the Haar measure on Td .
Let Ard := Ar ×· · ·×Ar , where Ar is the strip of width 2r around T. We denote by
ω
Cr (Td , Mat(m, R)) the space of real analytic functions A : Td → Mat(m, R) with
a holomorphic extension to Ard , which are continuous up to the boundary. Endowed
with the norm Ar := supz∈Ard A(z), this becomes a Banach space.
Consider the space of measurable cocycles C = {(Cm , dist)}m∈N , where Cm is
the set of cocycles A ∈ Crω (Td , Mat(m, R)) with det[A(x)] ≡ 0, which we call
non identically singular cocycles. We consider the following distance on this space
dist(A, B) := A − Br . Let E be the set of powers ε(t) ≡ C t −a for some C, a > 0,
and I be the set of sub-exponential functions ι(t) ≡ M e−c t , with M < ∞ and
b

0 < b < 1.

Definition 1.18 We call weak-Hölder any modulus of continuity of the form

ω(h) := M e−c (log(1/ h))


b

for some constants M, c and b as above.

Note that if b = 1 then we have ω(h) = M h c , which corresponds to the usual


modulus of Hölder continuity.
Fix any positive number δ0 > 0. Given t > 0, we denote by DCt the set of
frequency vectors ω ∈ Rd satisfying the following Diophantine condition:

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

Theorem 1.9 Consider the above space of measurable cocycles C consisting of


non-identically singular analytic quasi-periodic cocycles A : Td → Mat(m, R)
over a torus translation (Td , μ, T ) with frequency vector ω ∈ Rd .
If ω ∈ DCt for some t > 0 then all conclusions of Theorem 1.6 hold on the space
of measurable cocycles C with a weak-Hölder modulus of continuity.
As in the random setting, the key ingredient here is the following fiber uniform
LDT estimate of sub-exponential type, proven in Theorem 6.2.
Theorem 1.10 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
 
1 
μ x ∈T : d (n) (n) 
logB (x) − L (B) > C n −a
< e−n .
b

1.6 Literature Review

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.

2.1 Grassmann Geometry

Grassmann geometry is the geometric study of manifolds of linear subspaces of


an Euclidean space and of the action of linear groups (and algebras) on them. Its
foundations were laid in the masterpiece ‘Die lineale Ausdehnungslehre’ of Hermann
Grassmann, whose genius is still not fully understood, as explained in the survey [2].

2.1.1 Projective Spaces

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.

© Atlantis Press and the author(s) 2016 23


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_2
Free ebooks ==> www.Ebook777.com
24 2 Estimates on Grassmann Manifolds

The unit sphere S(V ) := {v ∈ V : v = 1} is a compact Riemannian manifold


of constant curvature 1 and diameter π . The natural projection π̂ : S(V ) → P(V ),
π̂ (v) = v̂, is a (double) covering map. Hence the projective space P(V ) has a natural
smooth Riemannian structure for which the covering map π̂ is a local isometry. Thus
P(V ) is a compact Riemannian manifold with constant curvature 1 and diameter π2 .
Given a linear map g ∈ L (V ) define P(g) := {v̂ ∈ P(V ) : g v = 0}. We refer
to the linear map ϕg : P(g) ⊂ P(V ) → P(V ), ϕg (v̂) := π̂ ( gg vv ), as the projective
action of g on P(V ). If g is invertible then ϕg : P(V ) → P(V ) is a diffeomorphism
with inverse ϕg−1 : P(V ) → P(V ). Through these maps, the group GL(V ), of all
linear automorphisms on V , acts transitively on the projective space P(V ).
We will consider three different metrics on the projective space P(V ). The Rie-
mannian distance, ρ, measures the length of an arc connecting two points on the
sphere. More precisely, given u, v ∈ S(V ),

ρ(û, v̂) := min{∠(u, v), ∠(u, −v)}. (2.1)

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

2.1.2 Exterior Algebra

Exterior Algebra was introduced by H. Grassmann in the ‘Ausdehnungslehre’. We


present here an informal description of some of its properties. See the book of
Stenberg [8] for a rigorous treatment of the subject.
Let V be a finite n-dimensional Euclidean space. Given k vectors v1 , . . . , vk ∈ V ,
their kth exterior product is a formal skew-symmetric product v1 ∧ · · · ∧ vk , in the
sense that for any permutation σ = (σ1 , . . . , σk ) ∈ Sk ,

vσ1 ∧ · · · ∧ vσk = (−1)sgn(σ ) v1 ∧ · · · ∧ vk .

These formal products are elements of an anti-commutative and associative graded


algebra (∧∗ V, +, ∧), called the exterior algebra of V . Formal products v1 ∧ · · · ∧ vk
are called simple k-vectors of V . The kth exterior power of V , denoted by ∧k V , is
the linear span of all simple k vectors of V . Elements of ∧k V are called k-vectors.
An easy consequence of this formal definition is that v1 ∧ · · · ∧ vk = 0 if and
only if v1 , . . . , vk are linearly dependent. Another simple consequence is that given
two bases {v1 , . . . , vk } and {w1 , . . . , wk } of the same 
k-dimensional linear subspace
of V , if for some real matrix A = (aij ) we have wi = kj=1 aij vj for all i = 1, . . . , k,
then
w1 ∧ · · · ∧ wk = (det A) v1 ∧ · · · ∧ vk .

More generally, two families {v1 , . . . , vk } and {w1 , . . . , wk } of linearly independent


vectors span the same k-dimensional subspace if and only if for some real number
λ = 0, w1 ∧ · · · ∧ wk = λ v1 ∧ · · · ∧ vk . Hence we identify the line spanned by a
simple k-vector v = v1 ∧ · · · ∧ vk , i.e., the projective point v̂ ∈ P(∧k V ) determined
by v, with the k-dimensional subspace spanned by the vectors {v1 , . . . , vk }, denoted
hereafter by v1 ∧ · · · ∧ vk .
The subspaces ∧k V induce the grading structure of the exterior algebra ∧∗ V , i.e.,
we have the direct sum decomposition ∧∗ V = ⊕dim k=0 ∧k V with (∧k V ) ∧ (∧k V ) ⊂
V

∧k+k V for all 0 ≤ k, k ≤ dim V . Geometrically, the exterior product operation
∧ : ∧k V × ∧k V → ∧k+k V corresponds to the algebraic sum of linear subspaces, in
the sense that given families {v1 , . . . , vk } and {w1 , . . . , wk } of linearly independent
vectors such that v1 ∧ · · · ∧ vk  ∩ w1 ∧ · · · ∧ wk  = 0, then

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

{e1 , . . . , en } of V , the family {eI : I ∈ Λnk , 0 ≤ k ≤ n} is an orthonormal basis of


the exterior algebra ∧∗ V .
Given vectors v1 , . . . , vk ∈ V let us call parallelepiped generated by these vectors
the set ⎧ ⎫
⎨k ⎬
P(v1 , . . . , vk ) := tj vj : tj ∈ [0, 1], j = 1, . . . , k .
⎩ ⎭
j=1

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,

v1 ∧ · · · ∧ vk  = Volk (P(v1 , . . . , vk )), (2.6)

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

v1 ∧ · · · ∧ vk  = v1 ∧ · · · ∧ vk  = v1  . . . vk 


= Volk (P(v1 , . . . , vk )) = Volk (P(v1 , . . . , vk )).

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)

Moreover, equality holds if and only if ei , fj  = 0 for all i = 1, . . . , r and j =


1, . . . , s.
A simple k-vector v1 ∧ · · · ∧ vk of norm one is called a unit k-vector. From the
previous considerations the correspondence v1 ∧ · · · ∧ vk → v1 ∧ · · · ∧ vk  is one-
to-one, between the set of unit k-vectors in ∧k V and the set of oriented k-dimensional
linear subspaces of V . In particular, if V is an oriented Euclidean space then the 1-
dimensional space ∧n V has a canonical unit n-vector, denoted by ω, and called the
volume element of ∧n V . In this case there is a unique operator, called the Hodge star
2.1 Grassmann Geometry 27

operator, ∗ : ∧∗ V → ∧∗ V defined by

v ∧ (∗w) = v, w ω, for all v, w ∈ ∧∗ V.

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,

∗(v1 ∧ · · · ∧ vk ) = v1 ∧ · · · ∧ vk ⊥ .

A dual product operation ∨ : ∧∗ V × ∧∗ V → ∧∗ V can be defined by

v ∨ w := ∗((∗v) ∧ (∗w)), for all v, w ∈ ∧∗ V.

This operation maps ∧k V × ∧k V to ∧k+k −n V , and describes the intersection opera-


tion on linear subspaces, in the sense that given families {v1 , . . . , vk } and {w1 , . . . , wk }
of linearly independent vectors with v1 ∧ · · · ∧ vk  + w1 ∧ · · · ∧ wk  = V , then

(v1 ∧ · · · ∧ vk ) ∨ (w1 ∧ · · · ∧ wk ) = v1 ∧ · · · ∧ vk  ∩ w1 ∧ · · · ∧ wk .

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 ,

∧k g(v1 ∧ · · · ∧ vk ) = g(v1 ) ∧ · · · ∧ g(vk ).

This construction is functorial in the sense that for all linear maps g, g : V → V ,

∧k id V = id∧k V , ∧k (g ◦ g) = ∧k g ◦ ∧k g and ∧k g∗ = (∧k g)∗ ,

where g∗ : V → V denotes the adjoint operator.


A clear consequence of these properties is that if g : V → V is an orthogonal
automorphism, i.e., g∗ ◦ g = id V , then so is ∧k g : ∧k V → ∧k V .
Consider a matrix A ∈ Mat(n, R). Given I, J ∈ Λnk , we denote by AI×J the square
sub-matrix of A indexed in I×J. If a linear map g : V → V is represented by A relative
to a basis {e1 , . . . , en }, then the kth exterior power ∧k g : ∧k V → ∧k V is represented
by the matrix ∧k A := (det AI×J )I,J relative to the basis {eI : I ∈ Λnk }. The matrix
∧k A is called the kth exterior power of A. Obviously, matrix exterior powers satisfy
the same functorial properties as linear maps, i.e., for all A, A ∈ Mat(n, R),

∧k In = I(nk) , ∧k (A A) = (∧k A )(∧k A) and ∧k A∗ = (∧k A)∗ ,

where A∗ denotes the transpose matrix of A.


28 2 Estimates on Grassmann Manifolds

Let n = dim V and {ei : i = 1, . . . , n} be an eigen-basis of a linear endomorphism


g : V → V with eigenvalues {λi : i = 1, . . . , n}, i.e., gei = λi ei for all i = 1, . . . , n.
Then the family {eI : I ∈ Λnk } is an eigen-basis of ∧k g : ∧k V → ∧k V with eigen-
values
λI = λi1 λi2 . . . λik , I = {i1 , . . . , ik } ∈ Λnk .

In other words, (∧k g)eI = λI eI for all I ∈ Λnk .

2.1.3 Grassmann Manifolds

Grassmannians, like projective spaces, are compact Riemannian manifolds which


stage the action of linear maps. For each 0 ≤ k ≤ n, the Grassmannian Gr k (V ) is
the space of all k-dimensional linear subspaces of V . Notice that the projective space
P(V ) and the Grassmannian Gr 1 (V ) are the same object if we identify each point
v̂ ∈ P(V ) with the line v = {λ v : λ ∈ R}. The full Grassmannian Gr(V ) is the
union of all Grassmannians Gr k (V ) with 0 ≤ k ≤ n. Denote by L (V ) the algebra of
linear endomorphisms on V , and consider the map π : Gr(V ) → L (V ), E → πE ,
that assigns the orthogonal projection πE onto E, to each subspace E ∈ Gr(V ). This
map is one-to-one, and we endow Gr(V ) with the unique topology that makes the
map π : Gr(V ) → π(Gr(V )) a homeomorphsim. With it, Gr(V ) becomes a compact
space, and each Grassmannian Gr k (V ) is a closed connected subspace of Gr(V ).
The group GL(V ) acts transitively on each Grassmannian. The action of GL(V )
on Gr k (V ) is given by · : GL(V ) × Gr k (V ) → Gr k (V ), (g, E) → g E. The special
orthogonal group SO(V ), of orientation preserving orthogonal automorphisms, acts
transitively on Grassmannians too. All Grassmannians are compact homogeneous
spaces.
For each 0 ≤ k ≤ n, the Plücker embedding is the map ψ : Gr k (V ) → P(∧k V )
that to each subspace E in Gr k (V ) assigns the projective point v̂ ∈ P(∧k V ), where
v = v1 ∧ · · · ∧ vk is any simple k-vector formed as exterior product of a basis
{v1 , . . . , vk } of E. This map is one-to-one and equivariant, i.e., for all g ∈ GL(V )
and E ∈ Gr(V ),
ψ(g E) = ϕ∧k g ψ(E). (2.8)

We will consider the metrics ρ, d, δ : Gr k (V ) × Gr k (V ) → [0, +∞) defined for


any given E, F ∈ Gr k (V ) by

ρ(E, F) := ρ(ψ(E), ψ(F)), (2.9)


d(E, F) := d(ψ(E), ψ(F)), (2.10)
δ(E, F) := δ(ψ(E), ψ(F)). (2.11)

which assign diameter π2 , 2 and 1, respectively, to the manifold Gr k (V ). These
distances are preserved by the action of orthogonal linear maps in SO(V ).
Free ebooks ==> www.Ebook777.com
2.1 Grassmann Geometry 29

Given k, k ≥ 0 such that k + k ≥ n = dim V , the intersection of subspaces is


an operation ∩ : Gr k,k (∩) ⊂ Gr k (V ) × Gr k (V ) → Gr k+k −n (V ) where:
Definition 2.1 The domain is defined by

Gr k,k (∩) := {(E, E ) ∈ Gr k (V ) × Gr k (V ) : E + E = V }.

Similarly, given k, k ≥ 0 such that k + k ≤ n = dim V , the algebraic sum of


subspaces is operation + : Gr k,k (+) ⊂ Gr k (V ) × Gr k (V ) → Gr k+k −n (V ) where:
Definition 2.2 The domain is defined by

Gr k,k (+) := {(E, E ) ∈ Gr k (V ) × Gr k (V ) : E ∩ E = {0}}.

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.

Proposition 2.2 The orthogonal complement operation E → E ⊥ is a d-isometric


involution on Gr(V ) which maps Gr k,k (+) to Gr n−k,n−k (∩) and satisfies for all
(E, E ) ∈ Gr k,k (+),
(E + E )⊥ = (E ⊥ ) ∩ (E )⊥ .

The composition semigroup L (V ) has two partial actions on Grassmannians,


called the push-forward action and the pull-back action. Before introducing them, a
couple of facts are needed.

Definition 2.3 Given g ∈ L (V ), we denote by Kg := {v ∈ V : g v = 0} the kernel


of g, and by Rg := {g v : v ∈ V } the range of g.

Lemma 2.1 Given g ∈ L (V ) and E ∈ Gr(V ),


1. if E ∩ (Kg) = {0} then the linear map g|E : E → g(E) is an isomorphism, and
in particular dim g(E) = dim E.
2. if E + (Rg) = V then the linear map g∗ |E ⊥ : E ⊥ → g−1 (E)⊥ is an isomorphism,
and in particular dim g−1 (E) = dim E.

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

g−1 (E)⊥ is clear. Since g∗ |E ⊥ is injective, dim g∗ (E ⊥ ) = dim(E ⊥ ). On the other


hand, by the transversality condition, g−1 (E) has dimension
 
dim g−1 (E) = dim (g|(Kg)⊥ )−1 (E ∩ Rg) + dim(Kg)
= dim(E ∩ Rg) + dim(Kg)
= dim(E) + dim(Rg) − n + dim(Kg) = dim(E).

Hence both g∗ (E ⊥ ) and g−1 (E)⊥ have dimension equal to dim(E ⊥ ), and the equality
follows. 

Given g ∈ L (V ) and k ≥ 0 such that k + dim(Kg) ≤ n = dim V , the push-


forward by g is the map ϕg : Gr k (g) ⊂ Gr k (V ) → Gr k (V ), E → gE, where:

Definition 2.4 The domain is defined by

Gr k (g) := {E ∈ Gr k (V ) : E ∩ (Kg) = {0}}.

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.

2.1.4 Flag Manifolds

Let V be a finite n-dimensional Euclidean space. Any strictly increasing sequence of


linear subspaces F1 ⊂ F2 ⊂ · · · ⊂ Fk ⊂ V is called a flag in the Euclidean space V .
Formally, flags are denoted as lists F = (F1 , . . . , Fk ). The sequence τ = (τ1 , . . . , τk )
of dimensions τj = dim Fj is called the signature of the flag F. The integer k is called
2.1 Grassmann Geometry 31

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

ρτ (F, F ) = max ρ(Fj , Fj ) (2.12)


1≤j≤k

dτ (F, F ) = max d(Fj , Fj ) (2.13)


1≤j≤k

δτ (F, F ) = max δ(Fj , Fj ) (2.14)


1≤j≤k

are equivalent to the Riemannian distance


√ on Fτ (V ). With these metrics, the flag
manifold Fτ (V ) has diameter π2 , 2 and 1, respectively. The group SO(V ) acts
isometrically on Fτ (V ) with respect to these distances.
Given a signature τ = (τ1 , . . . , τk ), if n = dim V , we define

τ ⊥ := (n − τk , . . . , n − τ1 ).

When τ = (τ1 , . . . , τk ) we will write τ ⊥ = (τ1⊥ , . . . , τk⊥ ), where τj⊥ = n − τk+1−i .


Definition 2.6 Given a flag F = (F1 , . . . , Fk ) ∈ Fτ (V ), its orthogonal complement
is the τ ⊥ -flag F ⊥ := (Fk⊥ , . . . , F1⊥ ).
The map ·⊥ : F(V ) → F(V ) is an isometric involution on F(V ), mapping Fτ (V )
onto Fτ ⊥ (V ). The involution character, (F ⊥ )⊥ = F for all F ∈ F(V ), is clear. As
explained in Sect. 2.1.2, the Hodge star operator ∗ : ∧k V → ∧n−k V is an isom-
etry between these Euclidean spaces. By choice of metrics on the Grassmannians,
see (2.10), the Plücker embeddings are isometries. Finally, the Plücker embedding
conjugates the orthogonal complement map ·⊥ : Gr k (V ) → Gr n−k (V ) with the
Hodge star operator. Hence for each 0 ≤ k ≤ n, the map ·⊥ : Gr k (V ) → Gr n−k (V )
is an isometry. The analogous conclusion for flags follows from the definition of
distance dτ .
Given g ∈ L (V ) and a signature τ such that τi + dim(Kg) ≤ n for all i, the
push-forward by g on flags is the map ϕg : Fτ (g) ⊂ Fτ (V ) → Fτ (V ), ϕg F :=
(g F1 , . . . , g Fk ), where:
Definition 2.7 The domain of ϕg is defined by

Fτ (g) := {F ∈ Fτ (V ) : Fk ∩ (Kg) = {0}}.


32 2 Estimates on Grassmann Manifolds

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 }.

The duality between duality between push-forwards and pull-backs is expressed


as follows.

Proposition 2.4 Given g ∈ L (V ), Fτ (g−1 ) = Fτ ⊥ (g∗ )⊥ and for all F ∈ Fτ (g−1 ),

(ϕg−1 F)⊥ = ϕg∗ (F ⊥ ).

2.2 Singular Value Geometry

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.

2.2.1 Singular Value Decomposition

Let V be a Euclidean space of dimension n.


Definition 2.9 Given g ∈ L (V ), the singular values of g are the square roots of the
eigenvalues of the quadratic form Qg : V → R, Qg (v) = g v√ 2
= gv, gv, i.e., the
eigenvalues of the positive semi-definite self-adjoint operator g∗ g.
Given g ∈ L (V ), let

s1 (g) ≥ s2 (g) ≥ · · · ≥ sn (g) ≥ 0,

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

Definition 2.10 The eigenvectors


√ of the quadratic form Qg , i.e., of the positive semi-
definite self-adjoint operator g∗ g, are called the singular vectors of g.
By the spectral theory of self-adjoint operators, for any g ∈ L (V ) there exists an
orthonormal basis consisting of singular vectors of g.
Proposition 2.5 Given g ∈ L (V ), let v ∈ V be such that g∗ g v = λ2 v with λ ≥ 0
and v = 1, i.e., v is a unit singular vector of g with singular value λ. Then there
exists a unit vector w ∈ V such that
(a) g v = λ w,
(b) g g∗ w = λ2 w, i.e., w is a singular vector of g∗ .
Proof Let v ∈ V be a unit singular vector of g. Then g∗ g v = λ2 v with λ ≥ 0
and λ2 = λ2 v, v = g∗ g v, v = g v2 , which implies that λ = g v. Since
(g g∗ ) (g v) = g (g∗ g) v = λ2 g v, if λ = 0 then setting w = g v/g v = λ−1 g v, we
have (g g∗ ) w = λ2 w, which proves that w is a singular vector of g∗ . By definition
g v = λ w. When λ = 0, take w to be any unit vector in Kg∗ . Notice that dim(Kg) =
dim(Kg∗ ). In this case v and w are singular vectors of g and g∗ , respectively, such
that g v = 0 = λ w. 
By the previous proposition, given g ∈ L (V ) there exist two orthonormal sin-
gular vector basis of V , {v1 (g), . . . , vn (g)} and {v1 (g∗ ), . . . , vn (g∗ )} for g and g∗ ,
respectively, such that

g vj (g) = sj (g) vj (g∗ ) for all 1 ≤ j ≤ n.

Denote by Dg the diagonal matrix with diagonal entries sj (g), 1 ≤ j ≤ n, seen as


an operator Dg ∈ L (Rn ). Define the linear maps Ug , Ug∗ : Rn → V by Ug (ej ) =
vj (g) and Ug∗ (ej ) = vj (g∗ ), for all 1 ≤ j ≤ n, where the ej are the vectors of the
canonical basis in Rn . By construction Ug and Ug∗ are isometries and the following
decomposition holds
g = Ug∗ Dg (Ug )∗ ,

known as the singular value decomposition (SVD) of g.


We say that g has a simple singular spectrum if its n singular values are all distinct.
When g has simple singular spectrum, the singular vectors vj (g) and vj (g∗ ) above
are uniquely determined up to a sign, and in particular they determine well-defined
projective points vj (g), vj (g∗ ) ∈ P(V ).
Definition 2.11 Given g ∈ L (V ), we call singular basis of g any orthonormal
basis {v1 , . . . , vn } of V formed by singular vectors of g ordered in such a way that
g vi  = si (g) for all i = 1, . . . , n.
Given g ∈ L (V ), consider singular bases {v1 , . . . , vn } and {v1∗ , . . . , vn∗ } for g and

g , respectively, such that

g vj = sj vj∗ with sj = sj (g) for all 1 ≤ j ≤ n.


Free ebooks ==> www.Ebook777.com
34 2 Estimates on Grassmann Manifolds

For any I = {i1 , . . . , ik } ∈ Λnk we have

(∧k g)(vi1 ∧ · · · ∧ vik ) = (si1 . . . sik ) (vi∗1 ∧ · · · ∧ vi∗k ).

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).

This name is justified by the following fact.


Proposition 2.7 Given a linear map g : V → V between Euclidean spaces, with
n = dim V , for any Borel set B ⊂ V ,

Voln (g(B)) = det + (g) Voln (B),

where Voln denotes the n-dimensional Hausdorff measure.


Proof Let {v1 , . . . , vn } be any basis of V and consider the parallelipiped B =
P(v1 , . . . , vn ). By Proposition 2.9 below and formula (2.6),

Voln (g(B)) = (gv1 ) ∧ · · · ∧ (gvn ) = (∧n g)(v1 ∧ · · · ∧ vn )


= ∧n gv1 ∧ · · · ∧ vn  = det + (g)Voln (B).

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

det + (g ◦ g) = det + (g ) det + (g).

Proposition 2.9 Let V and V be Euclidean spaces with n = dim V ≤ dim V .


Then for any linear map g : V → V

det+ (g) = s1 (g) . . . sn (g) = ∧n g.

www.Ebook777.com
2.2 Singular Value Geometry 35

Proof The squares si2 = si (g)2 (1 ≤ i ≤ n) are the eigenvalues of g∗ g. 


Next proposition provides a method to compute the volume expansion factor.
Proposition 2.10 Let g : V → V be a linear map between Euclidean spaces. Given
orthonormal bases {vi : i = 1, . . . , n} of V and {vi : i = 1, . . . , n} of the range gV ,
 
det + (g) = det gvi , vj  i,j .

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

det+ (g)2 = det(g∗ g) = det(UA∗ AU ∗ )


= det(A∗ A) = det(A)2 .

This proves that det+ (g) = det A . 

2.2.2 Gaps and Most Expanding Directions

Consider a linear map g ∈ L (V ) and a number 1 ≤ k < dim V .


Definition 2.12 The kth gap ratio of g is defined to be

sk (g)
gr k (g) := ≥ 1.
sk+1 (g)

We will also write gr(g) instead of gr 1 (g).


Definition 2.13 We say that g has a first singular gap when gr(g) > 1. More
generally, we say that g has a k singular gap when gr k (g) > 1.
In some occasions it is convenient to work with the inverse quantity, denoted by

σk (g) := gr k (g)−1 ≤ 1. (2.15)

Proposition 2.11 For any 1 ≤ k < dim V,

∧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)

More generally, given 1 ≤ k < dim V , we have:

Definition 2.15 If gr k (g) > 1 we define the g-most expanding k-subspace to be


 
vk (g) := Ψ −1 v(∧k g) ,

where Ψ stands for the Plücker embedding defined in Sect. 2.1.3.

The subspace vk (g) is the direct sum of all singular directions associated with the
singular values s1 (g), . . . , sk (g). We have

ϕg vk (g) = vk (g∗ ). (2.17)

Analogously, let n = dim V and assume gr n−k (g) > 1.


Definition 2.16 We define the g-least expanding k-subspace as

vk (g) := vn−k (g)⊥ .

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 vk (g) = vk (g∗ ). (2.18)

Let τ = (τ1 , . . . , τk ) be a signature with 1 ≤ τ1 < · · · < τk < dim V .


Definition 2.17 We define the τ -gap ratio of g to be

gr τ (g) := min gr τj (g).


1≤j≤k

When gr τ (g) > 1 we say that g has a τ -gap pattern.


Note that gr τ (g) > 1 means that g has a τj singular gap for 1 ≤ j ≤ k. Recall
that Fτ (V ) denotes the space of all τ -flags, i.e., flags F = (F1 , . . . , Fk ) such that
dim(Fj ) = τj for j = 1, . . . , k.
2.2 Singular Value Geometry 37

Definition 2.18 If gr τ (g) > 1 then the most expanding τ -flag is

vτ (g) := (vτ1 (g), . . . , vτk (g)) ∈ Fτ (V ).

Given g ∈ L (V ) the domain of its push-forward action on Fτ (V ) is


Definition 2.19 Fτ (g) := {F ∈ Fτ (V ) : Fk ∩ Kg = {0}}.
The push-forward of a flag F ∈ Fτ (g) by g is

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∗ ).

Proof Given F ∈ Fτ (g), we have Fj ∩ Kg = {0} for all j = 1, . . . , k. Hence


dim gFj = dim Fj = τj for all j, which proves that ϕg F ∈ Fτ (V ). To check that
ϕg F ∈ Fτ (g∗ ) we need to show that gFk ∩ Kg∗ = {0}. Assume g v ∈ Kg∗ , with
v ∈ Fk , and let us see that g v = 0. By assumption g∗ g v = 0, which implies
(g g∗ ) g v = 0. Since the self-adjoint map g g∗ induces an automorphism on Rg , we
conclude that g v = 0.
The second statement follows from (2.17). 

Given g ∈ L (V ), the domain of its pull-back action on Fτ (V ) is


Definition 2.20 Fτ−1 (g) := {F ∈ Fτ (V ) : F1 + Rg = V }.
The pull-back of a flag F ∈ Fτ (g) by g is

ϕg−1 F = g−1 F := (g−1 F1 , . . . , g−1 Fk ).

Definition 2.21 If gr τ ⊥ (g) > 1 the least expanding τ -flag is

vτ (g) := (vτ1 (g), . . . , vτk (g)) ∈ Fτ (V ).

Proposition 2.13 If gr τ (g) > 1 then vτ ⊥ (g) = vτ (g)⊥ .

Proof Let {v1 , . . . , vn } be a singular basis of g. Since this basis is orthonormal,

vn−k (g) = vk+1 , . . . , vn  = v1 , . . . , vk ⊥ = vk (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

Proof Given F ∈ Fτ−1 (g), we have Fj + Rg = V for all j = 1, . . . , k. Hence


dim g−1 Fj = dim Fj = τj for all j, which proves that ϕg−1 F ∈ Fτ (V ). To check that
ϕg−1 F ∈ Fτ−1 (g∗ ) just notice that g−1 F1 + Rg∗ ⊇ Kg + Kg⊥ = V .
The second statement follows from (2.18) and Proposition 2.13. 

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. 

2.2.3 Angles and Expansion

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 ),

δmin (E, F) := min δ(û, v̂), (2.19)


u∈E\{0},v∈F\{0}
Free ebooks ==> www.Ebook777.com
2.2 Singular Value Geometry 39

and the Hausdorff distance between subspaces E, F ∈ Gr k (V ),


 
δH (E, F) := max max δmin (û, F), max δmin (v̂, E) .
u∈E\{0} v∈F\{0}

Given a unit vector v ∈ V , v = 1, denote by πv , πv⊥ : V → V the orthogonal


projections πv (x) := v, x v, respectively πv⊥ (x) := x − v, x v.

Lemma 2.2 Given u, v ∈ V non-collinear with u = v = 1, denote by P the


plane spanned by u and v. Then
(a) πv − πu is a self-adjoint endomorphism,
(b) K(πv − πu ) = P⊥ ,
(c) πv − πu : P → P is anti-conformal with similarity factor
the restriction
sin ∠(u, v) ,
(d) πv⊥ − πu⊥  = πv − πu  = δ(û, v̂).

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

w.r.t. the canonical basis. Hence πv − πu is given by


   
v12 − u12 v1 v2 − u1 u2 β α
V −U = =
v1 v2 − u1 u2 v22 − u22 α −β

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,

πv⊥ − πu⊥ 2 = πv − πu 2 = πv⊥ (u)2 = u − v, u v2


= u ∧ v2 = δ(û, v̂)2 . 

Lemma 2.3 Let V be a Euclidean space of even dimension 2k and let E, F ∈ Gr k (V )


be subspaces such that V = E ⊕ F. Then the linear map πE − πF admits an invariant
decomposition V = P1 ⊕ · · · ⊕ Pk into pairwise orthogonal planes Pj such that

www.Ebook777.com
40 2 Estimates on Grassmann Manifolds

(1) each Pj is invariant under πE and πF ,


(2) Pj = Ej ⊕ Fj , where Ej = E ∩ Pj , Fj = F ∩ Pj and dim Ej = dim Fj = 1,
(3) (πE − πF )|Pj : Pj → Pj is anti-conformal.

Proof Choose unit vectors u0 ∈ E and v0 ∈ F such that ∠(u0 , v0 ) = max{∠(u, v) : u


∈ E \ {0}, v ∈ F \ {0}}. Then the function f (x) = u − v0 2 defined over the unit
sphere in E attains its maximum value at u0 . By the method of Lagrange multipliers,
πE (u0 − v0 ) is collinear with u0 , which implies that πE (v0 ) is also collinear with
u0 . Therefore πE (v0 ) = u0 , v0 u0 . By a similar argument, πF (u0 ) = u0 , v0 v0 . The
plane P spanned by the vectors u0 and v0 is invariant under both projections πE and
πF . Hence, by Lemma 2.2 the restriction πE − πF : P → P is anti-conformal. Now
the orthogonal complement P⊥ is also invariant under πE , πF and πE − πF . Defining
E0 = E ∩ P⊥ and F0 = F ∩ P⊥ , we have P⊥ = E0 ⊕ F0 and πE − πF = πE0 − πF0
over P⊥ , where πE0 and πF0 denote orthogonal projections on P⊥ . The claim of this
lemma follows proceeding inductively with πE0 − πF0 . 

Definition 2.22 Given E, F ∈ Gr(V ), we denote by πF : V → V the orthogonal


projection onto F, and by πE,F : E → F the restriction of πF to E.

Proposition 2.16 Given E, F ∈ Gr k (V ),


 
(a) δ(E, F) = 1 − det + (πE,F )2 = 1 − det + (πF,E )2 ,
(b) δH (E, F) = πE,F ⊥  = πF,E ⊥  = πE − πF ,
(c) δH (E, F) ≤ δ(E, F).

Proof Consider the unit k-vectors e = Ψ (E) and f = Ψ (F).


For (a) notice first that δ(E, F) = δ(e, f ) = 1 − e, f 2 . Since the exterior
power ∧k πF,E : ∧k F → ∧k E is also an orthogonal projection we have e, f  =
e, ∧k πF,E (f ) = ∧k πF,E  = det + (πF,E ).
Take an orthogonal reflexion g ∈ O(V ) such that g(F) = E and g(E) = F. We
have g−1 (E ⊥ ) = F ⊥ and πE,F ⊥ = g−1 ◦ πF,E ⊥ ◦ g. Therefore πE,F ⊥  = πF,E ⊥ .
We have δH (E, F) = πE,F ⊥  because for any unit vector u ∈ û, with û ∈ P(E),

πE,F ⊥ (u) = min δ(û, v̂).


v∈F\{0}

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

πE,F ⊥ (u)2 = 1 − πE,F (u)2 ≤ 1 − det + (πE,F )2 .

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

Σ(p̂) := {x̂ ∈ P(V ) : x, p = 0 for x ∈ x̂}.

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

ρmin (p̂, F) := min ρ(p̂, q̂).


q∈F\{0}

Proposition 2.17 For any p̂, q̂ ∈ P(V ),

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)

These concepts extend naturally to Grassmannians and flag manifolds.


Definition 2.24 Given E, F ∈ Gr k (V ), we define the α-angle between them

α(E, F) = αk (E, F) := α(Ψ (E), Ψ (F)),

where Ψ : Gr k (V ) → P(∧k V ) denotes the Plücker embedding (see Sect. 2.1.3).


Definition 2.25 We say that two k-subspaces E, F ∈ Gr k (V ) are orthogonal, and
we write E ⊥ F, iff α(E, F) = 0.
The Grassmannian orthogonal hyperplane of F is defined as

Σ(F) := {E ∈ Gr k (V ) : α(E, F) = 0}.

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 ),

α(E, F) = sin ρmin (E, Σ(F)) = δmin (E, Σ(F)).


42 2 Estimates on Grassmann Manifolds

Next we characterize the angle α(E, F). Consider the notation of Definition 2.22.

Proposition 2.19 Given E, F ∈ Gr k (V ),


(a) α(E, F) = α(E ⊥ , F ⊥ ),
(b) α(E, F) = det + (πE,F ) = det + (πF,E ),
(c) E ⊥ F iff there exists a pair (e, f ) of unit vectors such that e ∈ E ∩ F ⊥ and
f ∈ F ∩ E⊥,
(d) α(E, F) ≤ πE,F  = 1 − δmin (E, F ⊥ )2 .

Proof Given E, F ∈ Gr k (V ), take orthonormal bases {u1 , . . . , uk } and {v1 , . . . , vk }


of E and F, respectively, and consider the associated unit k-vectors u = u1 ∧ · · · ∧ uk
and v = v1 ∧ · · · ∧ vk , so that u ∈ Ψ (E) and v ∈ Ψ (F).
Using the Hodge star operator we obtain unit vectors ∗u ∈ Ψ (E ⊥ ) and ∗v ∈
Ψ (F ⊥ ). Hence

α(E ⊥ , F ⊥ ) = ∗u, ∗v = u, v = α(E, F),

which proves (a). Also



α(E, F) := u1 ∧ · · · ∧ uk , v1 ∧ · · · ∧ vk 
⎛ ⎞
u1 , v1  u1 , v2  . . . u1 , vk 
⎜ ⎟
⎜ u2 , v1  u2 , v2  . . . u2 , vk  ⎟
= det ⎜ .. .. .. .. ⎟
⎝ . . . . ⎠
uk , v1  uk , v2  . . . uk , vk 
= det + (πE,F ).

For the second equality above write ui = wi + kj=1 ui , vj  vj with wi ∈ F ⊥ and
use the anti-symmetry of the exterior product. For the third equality remark that the
matrix with entries ui , vj  represents πE,F w.r.t. the given orthonormal bases. By
symmetry, α(E, F) = det + (πF,E ). This proves (b).
From these relations, α(E, F) = 0 ⇔ K(πE,F ) = {0} ⇔ K(πF,E ) = {0}, which
explains (c).
Finally, because all singular values of πE,F are in [0, 1],

α(E, F) = det + (πE,F ) ≤ πE,F 


= max πE,F (u)
u∈E,u=1

= max 1 − πE,F ⊥ (u)2
u∈E,u=1

= 1 − δmin (E, F ⊥ )2 ,

which proves (d). 


2.2 Singular Value Geometry 43

Next we extend α-angles to flags. Consider a signature τ of length k.

Definition 2.26 Given flags F, G ∈ Fτ (V ), define

α(F, G) = ατ (F, G) := min α(Fj , Gj ).


1≤j≤k

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.

Comparing the two definitions, for all F, G ∈ Fτ (V )

ατ (F, G) = 0 ⇔ G ⊥ F.

Hence, the orthogonal flag hyperplane of F is defined as

Σ(F) := {G ∈ Fτ (V ) : α(G, F) = 0}.

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 ),

α(E, F) = sin ρmin (F, Σ(G)) = δmin (F, Σ(G)).

Consider a sequence of linear maps g0 , g1 , . . . , gn−1 ∈ L (V ). The following


quantities, called expansion rifts, measure the break of expansion in the composition
gn−1 . . . g1 g0 of the maps gj .
Definition 2.28 The first expansion rift of the sequence above is the number

gn−1 . . . g1 g0 
ρ(g0 , g1 , . . . , gn−1 ) := ∈ [1, +∞) .
gn−1  . . . g1 g0 

Given 1 ≤ k ≤ dim V , the kth expansion rift is

ρk (g0 , g1 , . . . , gn−1 ) := ρ(∧k g0 , ∧k g1 , . . . , ∧k gn−1 ) .

Given a signature τ = (τ1 , . . . , τk ), the τ -expansion rift is defined as

ρτ (g0 , g1 , . . . , gn−1 ) := min ρτj (g0 , g1 , . . . , gn−1 ) .


1≤j≤k

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

Definition 2.29 Given g, g ∈ L (V ), we define

α(g, g ) := α(v(g∗ ), v(g )) if g and g have a first gap ratio


αk (g, g ) := α(vk (g∗ ), vk (g )) if g and g have a k gap ratio
ατ (g, g ) := α(vτ (g∗ ), vτ (g )) if g and g have a τ gap pattern.

The following exotic operation is introduced to obtain an upper bound on the


expansion rift ρ(g, g ). Consider the algebraic operation a ⊕ b := a + b − a b on the
set [0, 1]. Clearly ([0, 1], ⊕) is a commutative semigroup isomorphic to ([0, 1], ·). In
fact, the transformation Φ : ([0, 1], ⊕) → ([0, 1], ·), Φ(x) := 1 − x, is a semigroup
isomorphism. We summarize some properties of this operation.
Proposition 2.21 For any a, b, c ∈ [0, 1],
(1) 0 ⊕ a = a,
(2) 1 ⊕ a = 1,
(3) a ⊕ b = (1 − b) a + b = (1 − a) b + a,
(4) a ⊕ b < 1 ⇔ a < 1 and b < 1,
(5) a ≤ b ⇒ a ⊕ c ≤ b ⊕ c,
−1
(6) b>0 ⇒ √ (ab √ ⊕ c) b ≤ a ⊕
√ c,
(7) a c + b 1 − a 2 1 − c 2 ≤ a 2 ⊕ b2 .
Proof Items (1)–(6) are left as exercises. For √ item consider the function f :
√ the last
[0, 1] → [0, 1] defined by f (c) := a c + b 1 − a2 1 − c2 . A simple computation
shows that √
b c 1 − a2
f (c) = a − √
1 − c2

The derivative f has a zero at c =√a/ a ⊕ b,√and one can check that this zero is a
global maximum of f . Since f (a/ a ⊕ b) = a2 ⊕ b2 , item (7) follows. 
Definition 2.30 Given g, g ∈ L (V ) with τ -gap patterns, the upper τ -angle between
g and g is defined to be

βτ (g, g ) := gr τ (g)−2 ⊕ ατ (g, g )2 ⊕ gr τ (g )−2 .

We will write βk (g, g ) when τ = (k), and β(g, g ) when τ = (1).


The next proposition relates norm expansion by the linear map g, and distance
contraction by the projective map ϕg , with angles and gap ratios.
Proposition 2.22 Given g ∈ L (V ) with σ (g) < 1, a point ŵ ∈ P(V ) and a unit
vector w ∈ ŵ,

(a) α(ŵ, v(g)) g ≤ g w ≤ g α(ŵ, v(g))2 ⊕ σ (g)2 ,
σ (g)
(b) δ(ϕg (ŵ), v(g∗ )) = δ(ϕg (ŵ), ϕg (v(g))) ≤ δ(ŵ, v(g)) .
α(ŵ, v(g))
2.2 Singular Value Geometry 45

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

gu = 1 − α 2 κ g. Hence

α 2 g2 ≤ α 2 g2 + gu2 = gw2 ,

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 ,

which proves (a).


Item (b) follows from

  g v ∧ gw g v ∧ gu v∗ ∧ gu


δ ϕg (ŵ), v(g∗ ) = = =
gv gw g gw gw

gu σ 1 − α g
2 σ δ(ŵ, v(g))
= ≤ = . 
gw α g α

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

which proves the first inequality.


For the second inequality, consider any ŵ ∈ P(g) and
√ a unit vector w ∈ ŵ such
that a := w, v = α(ŵ, v(g)) ≥ 0. Then w = a v + √ 1 − a2 u, where u is a unit
vector orthogonal to v. It follows that g w = a g v + 1 − a2 g u with g u ⊥ v∗ ,

and g u = κ g for some 0 ≤ κ ≤ σ (g). Therefore

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. 

Corollary 2.1 Given g, g ∈ L (V ) with a (k)-gap pattern,

∧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

Lemma 2.4 Given g, g ∈ L (V ) with (1)-gap patterns,



β(g, g ) gr(g)−2 ⊕ gr(g )−2
1≤ ≤ 1+ .
α(g, g ) α(g, g )2

Proof Just notice that


 
κ 2 ⊕ α 2 ⊕ (κ )2 α 2 + (κ 2 ⊕ (κ )2 ) κ 2 ⊕ (κ )2
≤ = 1+ . 
α α2 α2

Proposition 2.24 Given g, g ∈ L (V ) with a (1)-gap pattern



gr(g)−2 + gr(g )−2
α(g, g ) ≥ ρ(g, g ) 1 − .
ρ(g, g )2

Proof By Proposition 2.23

ρ(g, g )2 ≤ β(g, g )2 ≤ α(g, g )2 + σ (g)2 + σ (g )2 , 

which implies the claimed inequality.


These inequalities then imply the following more general fact.

Proposition 2.25 Given g0 , g1 , . . . , gn−1 ∈ L (V ), if for all 1 ≤ i ≤ n − 1 the


linear maps gi and g(i) = gi−1 . . . g0 have (1)-gap patterns, then


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

Proof By definition g(n−1) = gn−1 . . . g1 g0 , and by convention g(0) = id V . Hence


n−1 g(i+1) 
gn−1 . . . g1 g0  = i=0 g(i) 
. This implies that
n−1 n−1
gn−1 . . . g1 g0   1  g(i+1) 
=
gn−1  . . . g1  i=0
gi  i=0
g(i) 

n−1
gi g(i) 
= .
i=0
gi  g(i) 

It is now enough to apply Proposition 2.23 to each factor. 


48 2 Estimates on Grassmann Manifolds

2.3 Lipschitz Estimates

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.

2.3.1 Projective Action

Proposition 2.26 Given p, q ∈ V \ {0},

p q 1 1
 −  ≤ max{ , } p − q.
p q p q

Proof Given to vectors u, v ∈ V with u ≥ v = 1 we have


u v
 −  ≤ u − v.
u v

Assume for instance that p ≥ q, so that

max{p−1 , q−1 } = q−1 .


p q
Applying the previous inequality with u = q
and v = q
, we get

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. 

Given a linear map g ∈ L (V ), the projective action of g is given by the map


ϕg : P(g) → P(g∗ ), ϕg (p̂) := g!p.
For any non collinear vectors p, q ∈ V with p = q = 1, define

q − p, q p
vp (q) := .
q − p, q p

This is the normalized unit vector of the orthogonal projection of q onto p⊥ .


Free ebooks ==> www.Ebook777.com
2.3 Lipschitz Estimates 49

Proposition 2.27 Given g ∈ L (V ), and points p̂ = q̂ in P(V ),

δ(ϕg (p̂), ϕg (q̂)) gp ∧ gvp (q)


= .
δ(p̂, q̂) g p g q

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

δ(p̂, q̂) = p ∧ q = (sin θ ) p ∧ vp (q) = sin θ,

and
g p ∧ g q gp ∧ gvp (q)
δ(ϕg (p̂), ϕg (q̂)) = = (sin θ ) . 
g p g q g p g q

Given a point p̂ ∈ P(V ), we identify the tangent to the projective space at p̂ as


Tp̂ P(V ) = p⊥ , for any representative p ∈ p̂.

Proposition 2.28 Given g ∈ L (V ), x̂ ∈ P(g), and a representative x ∈ x̂, the


derivative of the map ϕg : P(g) → P(g∗ ) at x̂ is given by

g v − gg xx , g v gg xx 1


(Dϕg )x̂ v = = π⊥ (g v)
g x g x gx/gx

Proof The sphere S(V ) := {v ∈ V : v = 1} is a double covering space of


P(V ), whose covering map is the canonical projection π̂ : S(V ) → P(V ). With
the identification Tp̂ P(V ) = p⊥ , the derivative of π̂ , Dπ̂x : Tx S(V ) → Tx̂ P(V ),
is the identity linear map. The map ϕg lifts to the map defined on the sphere by
ϕg (x) := gg xx . Hence we can identify the derivatives (Dϕg )x̂ and (D"
" ϕg )x . A simple
calculation leads to the explicit expression above for (D" ϕg )x v. 

We will use the following closed ball notation

B(d) (p̂, r) := {x̂ ∈ P(V ) : d(x̂, p̂) ≤ r},

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.

Proposition 2.29 Given 0 < κ < 1 and g ∈ L (V ) such that gr(g) ≥ κ −1 ,


  √
(1) ϕg B(δ) (v(g), r)  ⊂ B(δ) (v(g∗ ), κ r/ 1 − r 2 ), for any 0 < r < 1,
(2) ϕg B(ρ) (v(g), a) ⊂ B(ρ) (v(g∗ ), κ tan a), for any 0 < a < π2 ,

r+ 1−r 2
(3) Lipρ (ϕg |B(δ) (v(g),r) ) ≤ κ 1−r 2
, for any 0 < r < 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

Using item (b) of the same proposition we get

σ (g) κr
δ(ϕg (x̂), v(g∗ )) ≤ δ(x̂, v(g)) ≤ √
α(x̂, v(g)) 1 − r2

By Proposition 2.28 we have

1 1 # ⊥ $
(Dϕg )x v = πv⊥∗ (g v) + π" ⊥
ϕg (x) − πv∗ (g v).
gx gx

Thus, by Lemma 2.2(d),

κ g δ(ϕg (x̂), v(g∗ )) g


(Dϕg )x  ≤ +
gx gx

κ κr κ (r + 1 − r 2 )
≤√ + = .
1 − r2 1 − r2 1 − r2

Since B(δ) (v(g), r) is a convex Riemannian √


disk, by the mean value theorem
κ (r+ 1−r 2 )
ϕg |B(δ) (v(g),r) has Lipschitz constant ≤ 1−r 2
with respect to distance ρ. 

2.3.2 Operations on Flag Manifolds

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

d(E1 , E2 ) := min{e1 − e2 , e1 + e2 },

where ej is a unit k-vector of Ej , for j = 1, 2.


Free ebooks ==> www.Ebook777.com
2.3 Lipschitz Estimates 51

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}.

The following numbers quantify the transversality of two linear subspaces.

Definition 2.32 Given E ∈ Gr r (V ) and F ∈ Gr s (V ), consider a unit r-vector e of


E, a unit s-vector f of F, a unit (n − r)-vector e⊥ of E ⊥ and a unit (n − s)-vector f ⊥
of F ⊥ . We define

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.

Remark 2.1 If r + s > n then θ+ (E, F) = 0. Similarly, if r + s < n then


θ∩ (E, F) = 0.

Remark 2.2 Given E, F ∈ Gr(V ), θ∩ (E, F) = θ+ (E ⊥ , F ⊥ ).

Next proposition establishes a Lispchitz modulus of continuity for the sum and
intersection operations on Grassmannians in terms of the previous quantities.

Proposition 2.30 Given r, s ∈ N and E, E ∈ Gr r (V ), F, F ∈ Gr s (V ),


 
1 1
(1) d(E + F, E + F ) ≤ max , , F )
(d(E, E ) + d(F, F )),
 +θ (E, F) θ+ (E 
1 1
(2) d(E ∩ F, E ∩ F ) ≤ max , (d(E, E ) + d(F, F )).
θ∩ (E, F) θ∩ (E , F )
Proof (1) Consider unit r-vectors e and e representing the subspaces E and E
respectively. Consider also unit s-vectors f and f representing the subspaces F and
F respectively. By Proposition 2.26

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 )

where K = max{e ∧ f −1 , e ∧ f −1 } = max{θ+ (E, F)−1 , max{θ+ (E , F )−1 }.


(2) reduces to (1) by duality (see Proposition 2.2). 

Next proposition gives an alternative characterization of the transversality mea-


surements θ+ (E, F) and θ∩ (E, F). Let, as before, πE : V → E denote the orthogonal
projection onto a subspace E ⊂ V , and define the restriction πE,F := πF |E : E → F.

www.Ebook777.com
52 2 Estimates on Grassmann Manifolds

Proposition 2.31 Given E ∈ Gr r (V ) and F ∈ Gr s (V ),


(1) θ+ (E, F) = det + (πE,F ⊥ ) = det + (πF,E ⊥ ).
(2) θ∩ (E, F) = det + (πE ⊥ ,F ) = det + (πF ⊥ ,E ).

Proof Notice that E ∩ F = K(πE,F ⊥ ) = K(πF,E ⊥ ). If E ∩ F = {0} then the three


terms in (1) vanish. Otherwise πE,F ⊥ and πF,E ⊥ are isomorphisms onto their ranges
R(πE,F ⊥ ) = F ⊥ ∩ (E + F) and R(πF,E ⊥ ) = E ⊥ ∩ (E + F). Take an orthonormal
basis {f1 , . . . , fs , fs+1 , . . . , fs+r , . . . , fn } such that {f1 , . . . , fs } spans F and the family
of vectors {f1 , . . . , fr , fs+1 , . . . , fs+r } spans E + F. Consider the unit s-vector f =
f1 ∧ · · · ∧ fs of F, and a unit r-vector e = e1 ∧ · · · ∧ er of E. Hence {fs+1 , . . . , fs+r }
is a basis of R(πE,F ⊥ ) and

θ+ (E, F) = (e1 ∧ · · · ∧ er ) ∧ (f1 ∧ · · · ∧ fs )


= πE,F ⊥ (e1 ) ∧ · · · ∧ πE,F ⊥ (er ) ∧ f1 ∧ · · · ∧ fs 
= det + (πE,F ⊥ ) fs+1 ∧ · · · ∧ fs+r ∧ f1 ∧ · · · ∧ fs  = det + (πE,F ⊥ ).

Reversing the roles of E and F, and because e ∧ f  is symmetric in e and f , we


obtain θ+ (E, F) = det + (πF,E ⊥ ), which proves (1).
By duality and Remark 2.2, item (2) reduces to (1). 

The measurement on the (∩) transversality admits the following lower bound in
terms of the angle in Definition 2.24.

Proposition 2.32 Given E ∈ Gr r (V ) and F ∈ Gr s (V ), if E + F = V then

θ∩ (E, F) ≥ αr (E, E ∩ F + F ⊥ ).

Proof Combining Lemmas 2.5 and 2.6 below we have

θ∩ (E, F) ≥ θ∩ (E, F ∩ (E ∩ F)⊥ ) = αr (E, (F ∩ (E ∩ F)⊥ )⊥ )


= αr (E, (E ∩ F) + F ⊥ ). 

Lemma 2.5 Given E ∈ Gr r (V ), E ∈ Gr r (V ) and F ∈ Gr s (V ) such that r + s ≥ n


and E ⊆ E then θ∩ (E , F) ≥ θ∩ (E, F).

Proof Because E ⊂ E , we have πF ⊥ ,E = πE ,E ◦ πF ⊥ ,E . Hence by Proposition 2.8

θ∩ (E, F) = det + (πF ⊥ ,E ) = det + (ππE (F ⊥ ),E ) det + (πF ⊥ ,E )


≤ det+ (πF ⊥ ,E ) = θ∩ (E , F),

where det + (ππE (F ⊥ ),E ) ≤ 1 because πE  ≤ 1. 

Lemma 2.6 Given E, E ∈ Gr r (V ), θ∩ (E , E ⊥ ) = αr (E , E).


2.3 Lipschitz Estimates 53

Proof Given orthonormal bases {v1 , . . . , vr } of E, and {v1 , . . . , vr } of E ,

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 

Next proposition gives a modulus of lower semi-continuity for the transversality


measurement θ∩ .

Proposition 2.33 Given E, E0 ∈ Gr r (V ) and F, F0 ∈ Gr s (V ),

θ∩ (E, F) ≥ θ∩ (E0 , F0 ) − d(E, E0 ) − d(F, F0 ).

Proof Consider unit vectors e ∈ Ψ (E ⊥ ), f ∈ Ψ (F ⊥ ), e0 ∈ Ψ (E0⊥ ) and f0 ∈ Ψ (F0⊥ ),


chosen so that

d(E, E0 ) = d(E ⊥ , E0⊥ ) = e − e0 ,


d(F, F0 ) = d(F ⊥ , F0⊥ ) = f − f0 .

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 ). 

Next proposition refines inequality (2.7).

Proposition 2.34 Given E, F ∈ Gr k (V ), and families of vectors {u1 , . . . , uk } ⊂ E


and {uk+1 , . . . , uk+i } ⊂ F ⊥ with 1 ≤ i ≤ m − k,
(a) u1 ∧ · · · ∧ uk ∧ uk+1 ∧ · · · ∧ uk+i  ≤ u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i ,
(b) u1 ∧ · · · ∧ uk ∧ uk+1 ∧ · · · ∧ uk+i  ≥ α(E, F) u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i .

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

det + (πF ⊥ ,E ⊥ ) ≤ m(∧i πF ⊥ ,E ⊥ ) ≤ ∧i πF ⊥ ,E ⊥  = 1.


54 2 Estimates on Grassmann Manifolds

Hence

u1 ∧ · · · ∧ uk ∧ uk+1 ∧ · · · ∧ uk+i  = u1 ∧ · · · ∧ uk ∧ πF ⊥ ,E ⊥ (uk+1 ) ∧ · · · ∧ πF ⊥ ,E ⊥ (uk+i )


= u1 ∧ . . . uk  πF ⊥ ,E ⊥ (uk+1 ) ∧ · · · ∧ πF ⊥ ,E ⊥ (uk+i )
≤ ∧i πF ⊥ ,E ⊥  u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i 
= u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i ,

which proves (a). By Proposition 2.19 we have

α(E, F) = α(F ⊥ , E ⊥ ) = det + (πF ⊥ ,E ⊥ ) ≤ m(∧i (πF ⊥ ,E ⊥ )).

Thus

u1 ∧ · · · ∧ uk ∧ uk+1 ∧ · · · ∧ uk+i  = u1 ∧ · · · ∧ uk ∧ πF ⊥ ,E ⊥ (uk+1 ) ∧ · · · ∧ πF ⊥ ,E ⊥ (uk+i )


= u1 ∧ . . . uk  πF ⊥ ,E ⊥ (uk+1 ) ∧ · · · ∧ πF ⊥ ,E ⊥ (uk+i )
≥ m(∧i πF ⊥ ,E ⊥ ) u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i 
≥ α(E, F) u1 ∧ · · · ∧ uk  uk+1 ∧ · · · ∧ uk+i ,

which proves (b). 

The angle α is a Lipschitz continuous function.

Proposition 2.35 Given u, u , v, v ∈ P(V ),



α(u, v) − α(u , v ) ≤ d(u, u ) + d(v, v ).

Proof Exercise. 

The intersection of complementary flags satisfying the appropriate transversality


conditions determines a decomposition of the Euclidean space V . We end this section
defining by this operation and proving a modulus of continuity for it.
Consider a signature τ = (τ1 , . . . , τk ) of length k with τk < dim V . We make the
convention that τ0 = 0 and τk+1 = dim V .

Definition 2.33 A τ -decomposition is a family of linear subspaces E· = {Ei }1≤i≤k+1


in Gr(V ) such that V = ⊕k+1
i=1 Ei and dim Ei = τi − τi−1 for all 1 ≤ i ≤ k + 1.

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

where dτi −τi−1 stands for the distance (2.10) in Gr τi −τi−1 (V ).


Given two flags F ∈ Fτ (V ) and F ∈ Fτ ⊥ (V ), we will define a decomposition,
denoted by F  F , formed out by intersecting the components of these flags. For that
we introduce the following measurement.
2.3 Lipschitz Estimates 55

Definition 2.34 Given two flags F ∈ Fτ (V ) and F ∈ Fτ ⊥ (V ), let

θ (F, F ) := min θ∩ (Fi , Fk−i+1



).
1≤i≤k


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 .

We prove by finite induction in i = 1, . . . , k + 1 that


%

Fi = Fj ∩ Fk−j+2 . (2.23)
j≤i

Since Fk+1 = V the proposition follows from this relation at i = k + 1.


For i = 1, (2.23) reduces to F1 = F1 ∩ V . The induction step follows from


Fi+1 = Fi ⊕ Fi+1 ∩ Fk−i+1 .

Since the following dimensions add up

dim Fi+1 = τi+1 = τi + (τi+1 − τi )



= dim Fi + dim(Fi+1 ∩ Fk−i+1 ),
56 2 Estimates on Grassmann Manifolds

it is enough to see that




Fi ∩ Fi+1 ∩ Fk−i+1 = Fi ∩ Fk−i+1 = {0},

which holds because of (2.22). 

Hence, by the previous proposition we can define:

Definition 2.35 Given flags F ∈ Fτ (V ) and F ∈ Fτ ⊥ (V ) such that θ (F, F ) > 0



we define F  F := {Fi ∩ Fk−i+2 }1≤i≤k+1 and call it the intersection decomposition

of the flags F and F .

Next proposition provides a modulus of lower semi-continuity for the transver-


sality measurement θ .

Proposition 2.37 Given F, F0 ∈ Fτ (V ) and F , F0 ∈ Fτ ⊥ (V ),

θ (F, F ) ≥ θ (F0 , F0 ) − dτ (F, F0 ) − dτ ⊥ (F , F0 ).

Proof Apply Proposition 2.33 at each subspace of the τ -decompositions. 

The modulus of continuity for the intersection map  : Fτ (V ) × Fτ ⊥ (V ) →


Dτ (V ) is established below.

Proposition 2.38 Given flags F1 , F2 ∈ Fτ (V ) and F1 , F2 ∈ Fτ ⊥ (V ),


 
1 1
dτ (F1  F1 , F2  F2 ) ≤ max , (dτ (F1 , F2 ) + dτ ⊥ (F1 , F2 )).
θ (F1 , F1 ) θ (F2 , F2 )

Proof Apply Proposition 2.30 at each subspace of the τ -decompositions. 

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 ).

Proposition 2.39 Given g0 , g1 ∈ L (V ), if gr τ (g0 ) > 1 and gr τ (g1 ) > 1 then

θ (vτ ⊥ (g1 ), vτ (g0∗ )) ≥ 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 ).

Proof Let n = dim V . Consider the flags F = vτ (g0∗ ) and F = vτ ⊥ (g1 ). We


have Fi = vτi (g0∗ ) and Fk−i+1 = vτk−i+1
⊥ (g1 ) = vn−τi (g1 ) = vτi (g1 )⊥ . Hence by
Lemma 2.6,
2.3 Lipschitz Estimates 57


θ∩ (Fi , Fk−i+1 ) = θ∩ (vτi (g0∗ ), vτi (g1 )⊥ ) = ατi (vτi (g0∗ ), vτi (g1 )) = ατi (g0 , g1 ),

and taking the minimum, θ (F, F ) ≥ ατ (g0 , g1 ). 

2.3.3 Dependence on the Linear Map

We establish a modulus of Lipschitz continuity for the most expanding direction of


a linear endomorphism with a first singular gap. For any 0 < κ < 1, consider the
set Lκ := {g ∈ L (V ) : gr(g) ≥ κ1 }. We denote by v : Lκ → P(V ) the map that
assigns the g-most expanding direction to each g ∈ Lκ .
The relative distance between linear maps g, g ∈ L (V ) \ {0} is defined as

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.

Proposition 2.40 The map v : Lκ → P(V ) is locally Lipschitz.


More precisely, given 0 < κ < 1 there exists ε0 = ε0 (κ) > 0, which increases as
κ decreases, such that for any g1 , g2 ∈ Lκ satisfying drel (g1 , g2 ) ≤ ε0 ,

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),

g1∗ g1 − g2∗ g2  ≤ g1∗  g1 − g2  + g1∗ − g2∗  g2 


= (g1∗  + g2 ) g1 − g2  ≤ 2 g1 − g2 .
58 2 Estimates on Grassmann Manifolds

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

h1 (p) =  p, p0  p0 + h1 (w) ≥ p, p0 



= 1 − p ∧ p0 2 ≥ 1 − δ02 ≥ 1/2,

and again, assuming δ0 is small,



h2 (p) ≥ h1 (p) − h1 − h2  ≥ 1 − δ02 − ε0 ≥ 1/2.

Thus, by Lemma 2.9 below, for all p̂ ∈ B(p̂0 , δ0 ),

d(ϕh1 (p̂), ϕh2 (p̂)) ≤ 2 h1 − h2 .

Choosing ε0 small enough, √κ δ0


2
+ 2 ε0 < δ0 . This implies that
1−δ02

 
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

Lemma 2.7 Let (X, d) be a complete metric space, T1 : X → X a Lipschitz con-


traction with Lip(T1 ) < κ < 1, x1∗ = T1 (x1∗ ) a fixed point, and T2 : X → X any
other map with a fixed point x2∗ = T2 (x2∗ ). Then

1
d(x1∗ , x2∗ ) ≤ d(T1 , T2 ),
1−κ

where d(T1 , T2 ) := supx∈X d(T1 (x), T2 (x)).

Proof

d(x1∗ , x2∗ ) = d(T1 (x1∗ ), T2 (x2∗ ))


≤ d(T1 (x1∗ ), T1 (x2∗ )) + d(T1 (x2∗ ), T2 (x2∗ ))
≤ κ d(x1∗ , x2∗ ) + d(T1 , T2 ),

which implies that


1
d(x1∗ , x2∗ ) ≤ d(T1 , T2 ). 
1−κ

Lemma 2.8 Given g1 , g2 ∈ L (V ), for any 1 ≤ i ≤ dim V ,

∧i g1 − ∧i g2  ≤ i max{1, g1 , g2 }i−1 g1 − g2 .

Proof Given any unit i-vector v1 ∧ · · · ∧ vi ∈ ∧i V , determined by an orthonormal


family of vectors {v1 , . . . , vi },

(∧i g1 )(v1 ∧ · · · ∧ vi ) − (∧i g2 )(v1 ∧ · · · ∧ vi )


= (g1 v1 ) ∧ · · · ∧ (g1 vi ) − (g2 v1 ) ∧ · · · ∧ (g2 vi )

i
≤ (g1 v1 ) ∧ · · · ∧ (g1 vj−1 ) ∧ (g1 vj − g2 vj ) ∧ (g2 vj+1 ) ∧ · · · ∧ (g2 vi )
j=1


i
≤ g1 j−1 g2 i−j g1 vj − g2 vj 
j=1

≤ i max{1, g1 , g2 }i−1 g1 − g2 . 

Given a dimension 1 ≤ l ≤ dim V and 0 < κ < 1, consider the set

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 }

Corollary 2.2 The map v : Ll,κ → Gr l (V ) is locally Lipschitz.


More precisely, given 0 < κ < 1 there exists ε0 > 0 such that for any g1 , g2 ∈ Ll,κ
such that g1 − g2  ≤ ε0 Cl (g1 , g2 )−1 , we have

16
d(vl (g1 ), vl (g2 )) ≤ Cl (g1 , g2 ) g1 − g2 .
1 − κ2

Proof By Lemma 2.8, drel (∧l g1 , ∧l g2 ) ≤ Cl (g1 , g2 ) g1 − g2 . Apply Proposi-


tion 2.40 to the linear maps ∧l gj : ∧l V → ∧l V , j = 1, 2. 

Given g ∈ L (V ) having k and k + r gap ratios, if a subspace E ∈ Gr k (V ) is close


to the g most expanding subspace vk (g) then the restriction g|E ⊥ has an r-gap ratio
and the most expanding r-dimensional subspace of g|E ⊥ is close to the intersection
of vk+r (g) with E ⊥ . Next proposition expresses this fact in a quantitative way.

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

Kr = {h ∈ L (V ) : h ≤ 1 and σr (h) ≤ κ}.

By uniform continuity of σr on Kr there exists δ0 > 0 such that for all h ∈ L (V ) if


there exists h0 ∈ Kr with h − h0  < δ0 then σr (h) ≤ 2 κ.
Recall that πF denotes the orthogonal projection onto a linear subspace F ⊂ V .
g
Given g ∈ L (V ) such that (a) holds, consider the map h = g ◦ πvk (g)⊥ . We have
h ∈ Kr because σr (h) = σr (g ◦ πvk (g)⊥ ) = σk+r (g) < κ.
g
Given E ∈ Gr k (V ) such that (b) holds, we define hE = g ◦ πE ⊥ . Then by items
(b) and (c) of Proposition 2.16

h − hE  ≤ πvk (g)⊥ − πE ⊥  ≤ δ(vk (g)⊥ , E ⊥ ) = δ(E, vk (g)) < δ0 ,

which implies that σr (g|E ⊥ ) = σr (hE ) ≤ 2 κ, and hence proves (1).


2.3 Lipschitz Estimates 61

To prove item (2) we use the triangle inequality

δ(vr (g|E ⊥ ), vk+r (g) ∩ E ⊥ ) ≤ δ(vr (hE ), vr (h))


+ δ(vr (h), vk+r (g) ∩ vk (g)⊥ )
+ δ(vk+r (g) ∩ vk (g)⊥ , vk+r (g) ∩ E ⊥ )
 
16 r
≤ + 0 + 2 δ(E, vk (g))
1 − 4 κ2
20r
≤ δ(E, vk (g)).
1 − 4 κ2

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

Proof Applying Proposition 2.26 to the non-zero vectors g1 p and g2 p, we get


g1 p g2 p
d(ϕg1 (p̂), ϕg2 (p̂)) ≤  − 
g1 p g2 p
≤ max{g1 p−1 , g2 p−1 } g1 p − g2 p
≤ max{g1 p−1 , g2 p−1 } g1 − g2 . 

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.

Lemma 2.10 Given g1 , g2 ∈ GL(V ), and p̂ = q̂ in P(V ),

δ(ϕg1 (p̂), ϕg1 (q̂)) δ(ϕg2 (p̂), ϕg2 (q̂))


− ≤ C(g1 , g2 ) g1 − g2 ,
δ(p̂, q̂) δ(p̂, q̂)

where C(g1 , g2 ) := (g1−1 2 + g2 2 g1−1 2 g2−1 2 ) (g1  + g2 ).


Free ebooks ==> www.Ebook777.com
62 2 Estimates on Grassmann Manifolds

Proof Given p ∈ p̂ and q ∈ q̂, by Proposition 2.27

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 . 

Lemma 2.11 Given g ∈ GL(V ) and p̂ = q̂ in P(V ),

1 δ(ϕg (p̂), ϕg (q̂))


≤ ≤ g2 g−1 2 .
g2 g−1 2 δ(p̂, q̂)

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 .

To finish the proof combine these inequalities with Proposition 2.27. 

Given g ∈ GL(V ), we define

(g) := max{logg, logg−1 }. (2.24)

Lemma 2.12 For every g ∈ GL(V ) and p̂ = q̂ in P(V ),


& '
δ(ϕg (p̂), ϕg (q̂))
−4 (g) ≤ log ≤ 4 (g).
δ(p̂, q̂)

Proof It follows from Lemma 2.11.

www.Ebook777.com
2.3 Lipschitz Estimates 63

Lemma 2.13 Given g1 , g2 ∈ GL(V ), 0 < α ≤ 1 and p̂ = q̂ in P(V ),


   
δ(ϕg1 (p̂), ϕg1 (q̂)) α δ(ϕg2 (p̂), ϕg2 (q̂)) α
− ≤ C1 (g1 , g2 ) g1 − g2 ,
δ(p̂, q̂) δ(p̂, q̂)

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 . 

2.4 Avalanche Principle

Consider a long chain of n linear maps g0 : V0 → V1 , g1 : V1 → V2 , etc.,


between Euclidean spaces Vi of the same dimension m. The AP relates the expan-
sion gn−1 . . . g1 g0  of the composition gn−1 . . . g1 g0 with the product of the indi-
vidual expansions gn−1  . . . g1  g0 . Given two quantities Mn and Nn depend-
ing on a large number n ∈ N, we say in rough terms that they are ε-asymptotic,
ε
and write Mn Nn , when e−n ε ≤ Mn /Nn ≤ en ε . In general it is not true that
ε
gn−1 . . . g1 g0  gn−1  . . . g1  g0  for some small ε > 0, unless some atypi-
cally sharp alignment of the singular directions of the linear maps gj occurs. Given the
chain of linear maps g0 , g1 , . . . , gn−1 , its rift ρ(g0 , . . . , gn−1 ) := ggn−1
n−1 ... g0 
 ... g0 
∈ [0, 1]
measures the break of expansion in the composition gn−1 . . . g1 g0 . The AP says that
given any such chain g0 , g1 , . . . , gn−1 , where the gap ratio of each map gj is large, and
the rift of any pair of consecutive maps is never too small, the rift of the composition
behaves multiplicatively, in the sense that for some small number ε > 0,
ε
ρ(g0 , g1 , . . . , gn−1 ) ρ(g0 , g1 ) ρ(g1 , g2 ) . . . ρ(gn−2 , gn−1 ),

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

Schrödinger cocycles. In its original version, the AP applies to chains of unimodular


matrices in SL(2, C), and the length of the chain is assumed to be less than some
lower bound on the norms of the matrices. Note that for unimodular matrices, the
gap ratio and the norm are two equivalent measurements. Still in this unimodular
setting, for matrices in SL(2, R), Bourgain and Jitomirskaya [4, Lemma 5] relaxed
the constraint on the length of the chain of matrices, and later Bourgain [3, Lemma
2.6] removed it, at the cost of slightly weakening the conclusion of the AP.
Later, Schlag [7, Lemma 1] generalized the AP to invertible matrices in GL(m, C).
Recently, C. Sadel has shared with the authors an earlier draft of [1], containing his
version of the AP for GL(m, C) matrices. Both of these higher dimensional APs
assume some bound on the length of the chains of matrices. A higher dimensional
AP without this assumption was proven by the authors [5, Theorem 3.1] for invertible
real matrices.
We present here a more general AP, which holds for (possibly non-invertible)
matrices in Mat(m, R). As a by-product of the geometric approach used in the proof,
we also obtain a quantitative control on the most expanding directions of the matrix
product, something essential in the proof of the continuity of the Oseledets decom-
position.

2.4.1 Contractive Shadowing

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

B(x, ε) := {z ∈ X : d(z, x) ≤ ε}.

Given an open set X 0 ⊂ X, define

X 0 (ε) := {x ∈ X 0 : d(x, ∂X 0 ) ≥ ε},

where ∂X 0 denotes the topological boundary of X 0 in (X, d).

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

(a) xj ∈ Xj0 and d(xj , ∂Xj0 ) = 1,


(b) gj has Lipschitz constant ≤ κ on Xj0 (ε),
(c) gj (xj ) ∈ Xj+1
0
(2 ε),
(d) gj (Xj (ε)) ⊂ B(gj (xj ), δ).
0

Then, setting g(n) := gn−1 ◦ · · · ◦ g1 ◦ g0 , the following hold:


(1) the composition g(n) is defined on B(x0 , ε) and Lip(g(n) |B(x0 ,ε) ) ≤ κ n ,
δ
(2) d( gn−1 (xn−1 ), g(n) (x0 ) ) ≤ 1−κ ,
(3) if x0 = gn−1 (xn−1 ) then g (B(x0 , ε)) ⊂ B(x0 , ε) and there is a point x ∗ ∈
(n)
δ
B(x0 , ε) such that g(n) (x ∗ ) = x ∗ and d (x0 , x ∗ ) ≤ (1−κ)(1−κ n) .

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

Fig. 2.2 Shadowing property for a chain of contractive mappings

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,

d(zji−1 , zji ) ≤ κ j−i−1 δ. (2.26)

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,

, zj+1 ) = d(gj (zji−1 ), gj (zji )) ≤ κ d(zji−1 , zji ) ≤ κ j−i δ.


i−1 i
d(zj+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(x, ∂X00 ) ≥ d(x0 , ∂X00 ) − d(x, x0 ) ≥ 1 − ε > ε.

This implies that d(g0 (x), g0 (x0 )) ≤ κ d(x, x0 ). Thus


2.4 Avalanche Principle 67

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

which proves that g0 (B(x0 , ε)) ⊂ X10 (ε).


Assume now that for every l ≤ i − 1,

(gl ◦ · · · ◦ g0 )(B(x0 , ε)) ⊂ Xl+1


0
(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),

d(g(i+1) (x), ∂Xi+1


0
) ≥ d(gi (xi ), ∂Xi+1
0
) − d(gi (xi ), g(i+1) (x))
≥ 2 ε − d(zi+1
0
, zi+1
i
) − d(zi+1
0
, g(i+1) (x))

i−1
≥ 2ε − l
d(zi+1 , zi+1
l+1
) − d(g(i+1) (x0 ), g(i+1) (x))
l=0
≥ 2 ε − (δ + κ δ + · · · + κ i−1 δ) − κ i d(x0 , x)
≥ 2 ε − (δ + κ δ + · · · + κ i−1 δ) − κ i ε
≥ 2 ε − (1 − κ) ε (1 + κ + · · · + κ i−1 ) − κ i ε = ε

which proves that g(i+1) (B(x0 , ε)) ⊂ Xi+1


0
(ε), and establishes the claim above.
Thus g is well-defined on B(x0 , ε), and, because of assumption (b), g(n) is a κ n
(n)

Lipschitz contraction on this ball. This proves (1).


Item (2) follows by (2.26). In fact


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

Finally we prove (3). Assume x0 = gn−1 (xn−1 ).


It is enough to see that g(n) (B(x0 , ε)) ⊂ B(x0 , ε), because by (1) g(n) acts as a
κ n -contraction in the closed ball B(x0 , ε). The conclusion on the existence of a fixed
point, as well as the proximity bound, follow from the classical fixed point theorem
for Lipschitz contractions.
Given x ∈ B(x0 , ε), we know from the previous calculation that

d(x0 , g(n) (x0 )) < δ + κ δ + · · · + κ n−2 δ.


68 2 Estimates on Grassmann Manifolds

Hence

d(g(n) (x), x0 ) ≤ d(g(n) (x), g(n) (x0 )) + d(g(n) (x0 ), x0 )


≤ κ n−1 d(x, x0 ) + δ + κ δ + · · · + κ n−2 δ
≤ δ + κ δ + · · · + κ n−2 δ + κ n−1 ε
≤ (1 − κ) ε (1 + κ + · · · + κ n−2 ) + κ n−1 ε
1 − κ n−1
= (1 − κ) ε + κ n−1 ε = ε.
1−κ

Thus g(n) (x) ∈ B(x0 , ε), which proves that g(n) (B(x0 , ε)) ⊂ B(x0 , ε). 

2.4.2 Statement and Proof of the AP

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

−2 gn−1 . . . g1 g0  g1  . . . gn−2  −2


e−n C κ ε ≤ ≤ en C κ ε ,
g1 g0  . . . gn−1 gn−2 

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

Xj0 := {v̂ ∈ Xj : α(v̂, v(gj )) > 0},



Yj0 := {v̂ ∈ Xj : α(v̂, v(gj−1 )) > 0}.

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

(A) ϕg0 , . . . , ϕgn−1 , ϕgn−1


∗ , . . . , ϕ ∗,
g0 x0 , . . . , xn−1 , xn∗ , . . . , x1∗ ,
∗ ∗
(B) ϕgn−1
∗ , . . . , ϕ ∗, ϕ , . . . , ϕ
g0 g0 gn−1 , xn , . . . , x1 , x0 , . . . , xn−1 ,

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

Similarly, by definition of Yj0 (ε),


#π $
v̂ ∈ Yj0 (ε) ⇔ δ(v̂, xj∗ ) ≤ cos ε .
2
Thus, because

#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,

α(ϕgj (xj ), xj+1 ) = α(v(gj∗ ), v(gj+1 )) = α(gj , gj+1 ) ≥ ε.

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,

and in the same way we infer that

∗ 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

Hence, (d) of Lemma 2.14 holds.


Therefore, because ϕg0∗ (x1∗ ) = x0 and ϕgn−1 (xn−1 ) = xn∗ , conclusion (3) of
Lemma 2.14 holds for both chains (A) and (B). The projective points v(g(n) )
and v(g(n)∗ ) are the unique fixed points of the chains of mappings (A) and (B),
respectively. Hence, by the shadowing lemma both distances d(x0 , v(g(n) )) and
d(xn∗ , v(g(n)∗ )) are bounded above by

δsh κ
δsh .
(1 − κsh ) (1 − κsh
2n
) ε

This proves conclusions (1) and (2) of the AP.


From Proposition 2.28 we infer that for any g ∈ L (V ),

s2 (g)
(Dϕg )v(g)  = = σ (g).
g

Hence, by conclusion (1) of the shadowing lemma

σ (g(n) ) = (Dϕg(n) )v(g(n) )  ≤ Lip(ϕg(n) |B(v(g0 ),εsh ) )


 
κ (4 + 2 ε) n
≤ (κsh ) ≤
n
.
ε2

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

which proves conclusion (3) of the AP.

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)

By (2) of the AP, applied to the chain of linear maps g0 , . . . , gi−1 ,

δsh
d(v(g(i)∗ ), v(gi−1

)) ≤  κ ε−1 .
(1 − κsh )(1 − κsh
2i
)

Hence, by Proposition 2.35

(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 )

which by Remark 2.4 is equivalent to (4). 

Next proposition is a practical reformulation of the Avalanche Principle.


2.4 Avalanche Principle 73

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

Since 0 < κ ≤ c ε2 , and c ε2 ≤ c (1−2 c2 ε2 ) ε2 = c (ε )2 we have 0 < κ ≤ c (ε )2 .


Thus, because ε ε , this proposition follows from conclusions (1), (2) and (4) of
Theorem 2.1. 

2.4.3 Consequences of the AP

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

g(j,i) := gj−1 ◦ · · · ◦ gi+1 ◦ gi .

With this notation the following relation holds for 0 ≤ i < k < j ≤ n,

g(j,i) = g(j,k) ◦ g(k,i) .

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

Proposition 2.43 Under the assumptions of Theorem 2.1, where 0 < κ  ε2 ,

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,

d(ϕgi∗ v(g(n,i+1) ), v(g(n,i) ))


= d( ϕgi∗ ◦ ϕg(n,i+1)∗ v(g(n,i+1)∗ ), ϕg(n,i)∗ v(g(n,i)∗ ) )
= d( ϕg(n,i)∗ v(g(n,i+1)∗ ), ϕg(n,i)∗ v(g(n,i)∗ ) )
≤ (κsh )n−i d( v(g(n,i+1)∗ , v(g(n,i)∗ )
κ (4 + 2 ε) n−i  
≤( ) d( v(g(n,i+1)∗ , v(gn−1
∗ ∗
) ) + d( v(gn−1 ), v(g(n,i)∗ )
ε 2
2 κ κ (4 + 2 ε) n−i
 ( ) .
ε ε2
which proves the proposition. 

Most expanding directions and norms of products of chains matrices under an


application of the AP admit the following modulus of continuity.

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

Assuming gi  ≥ gi , we have

gi  gi − gi  gi  gi 


≤ 1 + ≤ 1 + drel (gi , gi ) ≤ 1 + δ
gi  gi  gi  gi 

which implies
gi  1
≤ .
gi  1−δ

Because the case gi  ≤ gi  is analogous, we conclude that


 

log gi  ≤ log 1

δ
δ.
gi  1−δ 1−δ

Since the two chains of matrices satisfy the assumptions of the AP we have

gi gi−1  gi gi−1




≥ α(gi−1 , gi ) ≥ ε and ≥ α(gi−1 , gi ) ≥ ε.
gi  gi−1  gi  gi−1 

A simple calculation gives


 
gi gi−1  g 
drel ( gi gi−1 , gi gi−1

)≤ max 1, i drel (gi , gi )
gi gi−1  gi 
 
gi gi−1

 gi−1 
+ max 1, drel (gi−1 , gi−1 )
gi gi−1


gi−1 
2 δ δ
≤ .
(1 − δ)2 ε ε

Therefore, arguing as above,


log gi gi−1   δ .

gi gi−1  ε

Hence, by conclusion (4) of the AP we have


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

which proves (b). 

The next proposition is a flag version of the AP.


Let τ = (τ1 , . . . , τk ) be a signature with 0 < τ1 < τ2 < · · · < τk < m.
We call τ -block product any of the functions πτ,j : Mat(m, R) → R,

πτ,j (g) := sτj−1 +1 (g) . . . sτj (g), 1 ≤ j ≤ k,

where by convention τ0 = 0. A τ -singular value product, abbreviated τ -s.v.p., is any


product of distinct τ -block products. By definition, τ -block products are τ -singular
value products. Other examples of τ -singular value products are the functions

pτj (g) = s1 (g) . . . sτj (g) = ∧τj g.

Note that for every 1 ≤ j ≤ k we have:

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)

(4) for any τ -s.v.p. function π ,


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

d(vτj (g(n)∗ ), vτj (gn−1



)) = d(Ψ (vτj (g(n)∗ )), Ψ (vτj (gn−1

)))
= d(v(∧τj g(n)∗ ), v(∧τj gn−1

))  κ ε−1 .


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

Let us now prove (4).


For the τ -s.v.p. π(g) = pτ,j (g) = ∧τj g conclusion (4) is a consequence of the
corresponding conclusion of Theorem 2.1.
For the τ -block product π = πτ,j , since

log π(g) = log∧τj g − log∧τj−1 g,

conclusion (4) follows again from Theorem 2.1 (4).


Finally, since any τ -s.v.p. is a finite product of τ -block products we can reduce
(4) to the previous case. 

We finish this section with a version of the AP for complex matrices.


The singular values of a complex matrix g ∈ Mat(m, C) are defined to be the
eigenvalues of the positive semi-definite hermitian matrix g∗ g, where g∗ stands for the
transjugate of g, i.e., the conjugate transpose of g. Similarly, the singular vectors of g
are defined as the eigenvectors of g∗ g. The sorted singular values of g ∈ Mat(m, C)
are denoted by s1 (g) ≥ s2 (g) ≥ · · · ≥ sm (g). The top singular value of g coincides
with its norm, s1 (g) = g.
The (first) gap ratio of g is the quotient σ (g) := s2 (g)/s1 (g) ≤ 1. We say that
g ∈ Mat(m, C) has a (first) gap ratio when σ (g) < 1. When this happens the complex
eigenspace

{v ∈ Cm : g∗ g v = g v} = {v ∈ Cm : g v = g v}

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

Proof Make the identification Cm ≡ R2m , and given g ∈ Mat(m, Cm ) denote by


gR ∈ Mat(2m, R) the matrix representing the linear operator g : R2m → R2m in the
canonical basis.
We make explicit the relationship between gap ratios and angles of the complex
matrices and g, g ∈ Mat(m, C), and the gap ratios and angles of their real analogues
gR and (g )R .
Given g ∈ Mat(m, C), for each eigenvalue λ of g, the matrix gR has a correspond-
ing pair of eigenvalues λ, λ. Since g → gR is a C ∗ -algebra homomorphism, we have
(g∗ g)R = (gR )∗ (gR ). Therefore, for all i = 1, . . . , m, si (g) = s2i−1 (gR ) = s2i (gR ).
In particular, considering the signature τ = (2),

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

Abstract We devise an abstract, modular scheme to prove continuity of the


Lyapunov exponents for a general class of linear cocycles. The main assumption
is the availability of appropriate large deviation type (LDT) estimates which are uni-
form in the data. We provide a modulus of continuity that depends explicitly on the
sharpness of the LDT estimate. Our method uses an inductive procedure based on
the deterministic, general Avalanche Principle from the previous chapter. The main
advantage of this approach, besides the fact that it provides quantitative estimates,
is its versatility, as it applies to quasi-periodic cocycles (one and multivariable torus
translations), to random cocycles (i.i.d. and Markov systems) and to any other types
of base dynamics as long as appropriate LDT estimates are satisfied. Moreover, com-
pared to other available quantitative results for quasi-periodic or random cocycles,
this method allows for weaker assumptions.

3.1 Definitions, the Abstract Setup and Statement

An ergodic measure preserving dynamical system (X, F, μ, T ) consists of a set X , a


σ —algebra F, a probability measure μ on (X, F) and a transformation T : X → X
which is ergodic and measure preserving.
Two important classes of ergodic dynamical systems are the shift over a stochastic
process (i.e. a Bernoulli shift or a Markov shift) and the torus translation by an
incommensurable frequency vector.
A linear cocycle over an ergodic system (X, F, μ, T ) is a skew-product map on
the vector bundle X × Rm given by

X × Rm  (x, v) → (T x, A(x)v) ∈ X × Rm ,

where A : X → Mat(m, R) is a measurable function.


Hence T is the base dynamics while A defines the fiber action. Since the base
dynamics will be fixed, we identify the cocycle with just its fiber action A.

© Atlantis Press and the author(s) 2016 81


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_3
Free ebooks ==> www.Ebook777.com
82 3 Abstract Continuity of Lyapunov Exponents

The iterates of the cocycle are (T n x, A(n) (x) v), where

A(n) (x) = A(T n−1 x) · · · · · A(T x) · A(x).

Given an ergodic system (X, F, μ, T ), we introduce the main actors—a space of


cocycles and a set of observables—and we describe the main assumptions on them—
certain uniform large deviation type (LDT) estimates and a uniform L p -boundedness.
Then we formulate an abstract criterion for the continuity of the corresponding Lya-
punov exponents.

3.1.1 Cocycles and Observables

Definition 3.1 A space of measurable cocycles C is any class of matrix valued


functions A : X → Mat(m, R), where m ∈ N is not fixed, such that every A : X →
Mat(m, R) in C has the following properties:
1. A is F-measurable,
2. A ∈ L ∞ (μ),  
3. The exterior powers ∧k A : X → Mat( mk , R) are in C , for k ≤ m.
Each subspace Cm := { A ∈ C | A : X → Mat(m, R) } is a-priori endowed with
a distance dist : Cm × Cm → [0, +∞) which is at least as fine as the L ∞ distance,
i.e. for all A, B ∈ Cm we have

dist(B, A) ≥ B − A L ∞ .

We assume a correlation between the distances on each of these subspaces, in the


sense that the map
Cm  A → ∧k A ∈ C(mk)

is locally Lipschitz.

Let A ∈ C be a measurable cocycle. Since A ∈ L ∞ , we have log+ A ∈ L 1 ,


hence Furstenberg-Kesten’s theorem (the non-invertible, one-sided case, see Chap. 3
in [1]) applies. In particular, if we denote

1
L (n)
1 (A) := logA(n) (x) μ(d x), (3.1)
X n

then as n → ∞, L (n)1 (A) → L 1 (A) (the maximal Lyapunov exponent).


We call L (n)
1 (A) finite scale (maximal) Lyapunov exponents.

www.Ebook777.com
3.1 Definitions, the Abstract Setup and Statement 83

We will need stronger integrability assumptions on the measurable functions


1
n
logA(n) (x).
Let 1 ≤ p ≤ ∞. For simplicity of notations, later on we may assume that p = 2.

Definition 3.2 A cocycle A ∈ C is called L p -bounded if there is C < ∞, which


we call its L p -bound, such that for all n ≥ 1 we have:
1 
 
 logA(n)  p < C. (3.2)
n L

Definition 3.3 A cocycle A ∈ Cm is called uniformly L p -bounded if there are


δ = δ(A) > 0 and C = C(A) < ∞ such that for all B ∈ Cm with dist(B, A) < δ
and for all n ≥ 1 we have:
1 
 
 logB (n)  p < C. (3.3)
n L

It is not difficult to show that if a cocycle A ∈ C satisfies the bounds

logA±  L p ≤ C < ∞,

then for all n ∈ Z we have


 
logA(n)  L p ≤ C n . (3.4)

Hence if we assume that


logA±  ∈ L p (3.5)

holds for all cocycles A ∈ C , and if we endow Cm with the distance given by

dist p (A, B) := A − B L ∞ +  logA−1  − logB −1  L p

when 1 ≤ p < ∞, and by

dist∞ (A, B) := A − B L ∞ ,

when p = ∞, then every cocycle A ∈ C is uniformly L p -bounded.


In the applications we have in this book, the uniform L p -boundedness is automatic.
For instance, in the case of random cocycles, we assume from the beginning the
integrability condition (3.5) which implies uniform L p -boundedness.
However, we want that our scheme be applicable also to cocycles that are very
singular (i.e. non-invertible everywhere), which is the case of a forthcoming paper
on quasi-periodic cocycles, and that is why we make the weaker, uniform L p -
boundedness assumption.
84 3 Abstract Continuity of Lyapunov Exponents

Given a cocycle A ∈ C and an integer N ∈ N, denote by F N (A) the lattice


(w.r.t. union and intersection) generated by the sets {x ∈ X : A(n) (x) ≤ c} or
{x ∈ X : A(n) (x) ≥ c} where c ≥ 0 and 0 ≤ n ≤ N .
Let Ξ be a set of measurable functions ξ : X → R, which we call observables.

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

3.1.2 Large Deviations Type Estimates

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)

We will use the notation εn := ε(n) and ιn := ι(n) for integers n.


In the applications we have thus far, the constant p = 2 or p = ∞, the deviation
size functions are either constant functions ε(t) ≡ ε for some 0 < ε 1 or
powers ε(t) ≡ t −a for some a > 0, while the deviation set measure functions
are exponentials ι(t) ≡ e−c t , sub-exponentials ι(t) ≡ e−c t or nearly-exponentials
b

ι(t) ≡ e−c t/(log t) for some c, b > 0.


b

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

Definition 3.6 A measurable cocycle A ∈ C satisfies a fiber-LDT estimate if for


every ε > 0 there is p = p(A, ε) ∈ P, p = (n 0 , ε, ι), such that for all n ≥ n 0 we
have εn ≤ ε and
1 
μ {x ∈ X :  logA(n) (x) − L (n) 
1 (A) > εn } < ιn . (3.8)
n
We will need a stronger form of the fiber-LDT, one that is uniform in a neigh-
borhood of the cocycle, in the sense that estimate (3.8) holds with the same LDT
parameter for all nearby cocycles.
Definition 3.7 A measurable cocycle A ∈ Cm satisfies a uniform fiber-LDT if for
all ε > 0 there are δ = δ(A, ε) > 0 and p = p(A, ε) ∈ P, p = (n 0 , ε, ι), such that
if B ∈ Cm with dist(B, A) < δ and if n ≥ n 0 then εn ≤ ε and

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

3.1.3 Abstract Continuity Theorem of the Lyapunov


Exponents

We are ready to formulate the main result of this chapter.

Theorem 3.1 Consider an ergodic MPDS (X, F, μ, T ), a space of measurable


cocycles C , a set of observables Ξ , a constant 1 < p ≤ ∞, a set of LDT parameters
P = P( p) with corresponding spaces of deviation functions E, I and assume the
following:
1. Ξ is compatible with every cocycle A ∈ C .
2. Every observable ξ ∈ Ξ satisfies a base-LDT.
3. Every A ∈ C with L 1 (A) > −∞ is uniformly L p -bounded.
4. Every cocycle A ∈ C for which L 1 (A) > L 2 (A) satisfies a uniform fiber-LDT.
Then all Lyapunov exponents L k : Cm → [−∞, ∞), 1 ≤ k ≤ m, m ∈ N are
continuous functions of the cocycle.
Moreover, given A ∈ Cm and 1 ≤ k ≤ m, if L k (A) > L k+1 (A), then locally near
p−1
A the map L 1 + L 2 + · · · + L k has a modulus of continuity ω(h) := [ι (c log h1 )] 2 p−1
for some ι = ι(A) ∈ I and c = c(A) > 0.

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.

3.2 Upper Semicontinuity of the Top Lyapunov Exponent

Given are an ergodic system (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.
It is well know that the top Lyapunov exponent is upper semicontinuous as a
function of the cocycle. Our argument requires a much more precise version of
the upper semicontinuity, one that is uniform in the number n of iterates and in
the phase x. Such results are available, see [3, 5], and they are based on a stopping
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 87

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.

Proposition 3.1 (nearly uniform upper semicontinuity) Let A ∈ Cm be a measur-


able cocycle such that Ξ and A are compatible and every observable ξ ∈ Ξ satisfies
a base-LDT with corresponding LDT parameter in P.

(i) Assume that L 1 (A) > −∞ and that A is L 1 -bounded.


For every ε > 0, there are δ = δ(A, ε) > 0, n 0 = n 0 (A, ε) ∈ N and ι = ι(A, ε) ∈ I,
such that if B ∈ Cm with dist(B, A) < δ, and if n ≥ n 0 , then the upper bound

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

(i) Fix ε > 0. By Kingman’s subadditive ergodic theorem,

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

Then UN ∈ F N (A), U N ⊂ U N +1 and ∪ N U N has full measure. Therefore, there is


N = N (ε, A) such that μ(UN ) < ε.
We fix this integer N for the rest of the proof and denote the set U = U(ε, A) :=
U N . Therefore, U ∈ F N (A), μ(U ) < ε and we have: if x ∈ U then 1 ≤ n(x) ≤ N
and
logA(n(x)) (x) ≤ n(x)L 1 (A) + n(x)ε. (3.13)

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)

for all x outside a zero measure set and for all 1 ≤ n ≤ N .


Let t := e−N C/ε . Consider the set V := n=1 N
{x : A(n) (x) > t}. Clearly
2

 
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

If for some 1 ≤ n ≤ N and x ∈ X we have A(n) (x) ≤ t (< 1), then

1 
 logA(n) (x) > log 1/t ,
n n
hence

N
1  log 1/t
V ⊂ {x :  logA(n) (x) > }.
n=1
n n

Since A is L 1 -bounded, there is C = C(A) < ∞ such that for all n ≥ 1


1 
 
 logA(n)  1 < C.
n L

Then by Chebyshev’s inequality,

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,

and if 1 ≤ n ≤ N then for μ a.e. x ∈ V we have

logB (n) (x) ≤ logA(n) (x) + δ N C N −1 e N < logA(n) (x) + ε,


2
C/ε

provided we choose δ < δ(ε, C, N ) = δ(ε, A) small enough.


Let O := U ∩ V. Then O ∈ F N (A) and μ(O ) < 2ε. We conclude that for μ
almost every x ∈ O, we have:

logB (n(x)) (x) ≤ n(x) L 1 (A) + n(x) 2ε. (3.15)

Let n 0 = n 0 (ε, A) := CεN .


Fix x ∈ X and define inductively for all k ≥ 1 the sequence of phases xk =
xk (x) ∈ X and the sequence of integers n k = n k (x) ∈ N as follows:

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

Note that for all k ≥ 1, xk+1 = T n k +···+n 1 x and 1 ≤ n k ≤ N .


For any n ≥ n 0 (> N ≥ n 1 ), there is p ≥ 1 such that

n 1 + · · · + n p ≤ n < n 1 + · · · + n p + n p+1 ,

so n = n 1 + · · · + n p + m, where 0 ≤ m < n p+1 ≤ N .


For any cocycle B such that dist(B, A) < δ, let bn (x) := logB (n) (x). Then
clearly for μ-a.e. x ∈ X we have bn (x) ≤ n C, where C is a constant that depends
on A and bn (x) is a sub-additive process, meaning:

bn+m (x) ≤ bn (x) + bm (T n x)

for all n, m ≥ 1 and for μ almost every x ∈ X .


Using this sub-additivity and the definition of xk (x), n k (x), we have:


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:

bm (x p+1 ) ≤ mC < N C. (3.17)

For every 1 ≤ k ≤ p we have:

• Either xk ∈ O, so n k = n(xk ), in which case, using (3.15) we get:

bn k (xk ) = logB (n(xk )) (xk ) ≤ n(xk ) L 1 (A) + 2ε n(xk )


= (L 1 (A) + 2ε) n k .

• Or xk ∈
/ O, so n k = 1, in which case bn k (xk ) = logB(xk ) ≤ C.
Therefore,

bn k (xk ) = bn k (xk ) 1O (xk ) + bn k (xk ) 1O (xk )


≤ (L 1 (A) + 2ε) n k 1O (xk ) + C 1O (xk )
= (L 1 (A) + 2ε) n k − (L 1 (A) + 2ε) n k 1O (xk ) + C 1O (xk )
= (L 1 (A) + 2ε) n k − (L 1 (A) + 2ε) 1O (xk ) + C 1O (xk ),

where in the last equality we used the fact that n k = 1 when xk ∈ O .

www.Ebook777.com
3.2 Upper Semicontinuity of the Top Lyapunov Exponent 91

We conclude:

bn k (xk ) ≤ (L 1 (A) + 2ε) n k + (C − L 1 (A) − 2ε) 1O (xk )


< (L 1 (A) + 2ε) n k + 2C 1O (xk ). (3.18)

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

provided we choose x outside a set of measure ιn .


This ends the proof in the case L 1 (A) > −∞.
(ii) The case L 1 (A) = −∞. Let t be large enough, say t > C + 1. We apply again
Kingman’s subadditive theorem and for μ a.e. x ∈ X , define the integers

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

logA(n(x)) (x) ≤ −2tn(x). (3.21)

We show that (3.21) holds also for cocycles B in a neighborhood of A. This is


where the argument differs from the case L 1 (A) > −∞.
92 3 Abstract Continuity of Lyapunov Exponents

−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.

A(m) (x) < e−2t m .

In this case, using (3.22) we get B (m) (x) < e−2t m (1 + 1/t), so

log B (m) (x) < −2t m + 1/t. (3.23)

Case 2. A(m) (x) has a lower bound:

A(m) (x) ≥ e−2t m .

Then using again (3.22), we have

B (m) (x) − A(m) (x)


log B (m) (x) − log A(m) (x) ≤
A(m) (x)
−2mt
e
≤ e2t m = 1/t,
t
hence
log B (m) (x) < log A(m) (x) + 1/t. (3.24)

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)

which is the analogue of (3.15) when L 1 (A) > −∞.


The rest of the proof then follows exactly the same pattern as when L 1 (A) > −∞,
the role of L 1 (A) + 2ε being now played by −2t + 1/t, while the small set O is
simply U, since there was no extra small set excluded when deriving (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

Remark 3.2 If (X, μ, T ) is uniquely ergodic (e.g. an ergodic torus translation), we


may choose Ξ to be the set of all continuous functions on X . If the cocycle is
continuous, then using Urysohn’s lemma, it is easy to verify that the compatibility
condition between Ξ and A holds. For uniquely ergodic systems, the convergence in
Birkhoff’s ergodic theorem is uniform in the phase for all continuous observables.
Hence the base-LDT estimate (3.5) holds automatically for all ξ ∈ Ξ , with deviation
measure function ι(t) ≡ 0.
It follows that if (X, μ, T ) is uniquely ergodic, then the statements in Proposi-
tion 3.1 above hold for a.e. x. Because the cocycle is continuous, these statements
hold for all phases x. Hence we recover the corresponding result in [5].

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

then for all phases x outside a set of measure < ιn we have:

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

From (3.29) we get

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.

3.3 Finite Scale Continuity

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

Pick C1 > 2C0 + ε0 and n 01 ≥ n 0 such that e−C1 n 01 < δ0 .


Let n ≥ n 01 and Bi ∈ Cm with dist(Bi , A) ≤ δ0 (i = 1, 2) be arbitrary but fixed.
Assume that dist(B1 , B2 ) < e−C1 n . Apply the fiber-LDT to each Bi and conclude
that for all x outside a set Bin , with μ(Bin ) < ιn we have:

1
logBi(n) (x) > L (n)
1 (Bi ) − εn ≥ −C 0 − ε0 . (3.35)
n
96 3 Abstract Continuity of Lyapunov Exponents

Let Bn = B1n ∪ B2n , so μ(Bn ) < 2 ιn and if x ∈ Bn then we have:

B1(n) (x), B2(n) (x) > e−(C0 +ε0 ) n . (3.36)

Moreover, for μ-a.e. x ∈ Bn and all 0 ≤ j ≤ n − 1 we also have:

B1 (T j x) − B2 (T j x) ≤ dist(B1 , B2 ) < e−C1 n .

Therefore, for μ-a.e. x ∈ Bn we get:

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

1  (C0 +ε0 ) n C0 (n− j−1) −C1 n C0 j


n−1
≤ e e e e ≤ e−n (C1 −2C0 −ε0 ) .
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

Since ι ∈ I decays at most exponentially, we may assume, by choosing C1 large, that


e−n (C1 −2C0 −ε0 ) < ιn , so (3.37) and (3.38) imply
1/2


 (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

which proves (3.34). 


3.4 The Inductive Step Procedure 97

3.4 The Inductive Step Procedure

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.

Lemma 3.2 Let A ∈ C be an L 1 -bounded cocycle, and let C be its L 1 -bound. If


n 0 , n 1 , n, r ∈ N are such that n 1 = n · n 0 + r and 0 ≤ r ≤ n 0 , then
n0 n0
− 2C + L ((n+1)
1
n0 )
(A) ≤ L (n 1) (n n 0 )
1 (A) ≤ L 1 (A) + 2C (3.39)
n1 n1
 
Proof From the L 1 -boundedness assumption on A, for all m ≥ 1,  L 1(m) (A) ≤ C.
Since n 1 = n · n 0 + r and r ≥ 0, we have A(n 1 ) (x) = A(r ) (T n n 0 x) · A(n n 0 ) (x),
hence
A(n 1 ) (x) ≤ A(r ) (T n n 0 x) A(n n 0 ) (x) .

Taking logarithms, dividing by n 1 then integrating in x we get:


n n 0 (n n 0 ) r (r )
L (n 1)
1 (A) ≤ L 1 (A) + L (A).
n1 n1 1

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

which proves the right hand side of (3.39).


Now write (n + 1) n 0 = n 1 + q, where q = n 0 − r , so 0 ≤ q ≤ n 0 . Then

A((n+1) n 0 ) (x) = A(q) (T n 1 x) · A(n 1 ) (x),


A((n+1) n 0 ) (x) ≤ A(q) (T n 1 x) A(n 1 ) (x).

Taking logarithms, dividing by (n + 1) n 0 then integrating in x we get:


n1 q (q)
L ((n+1) n0 )
(A) ≤ L (n 1)
1 (A) + L (A).
1
(n + 1) n 0 (n + 1) n 0 1

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)

which proves the left hand side of (3.39). 


98 3 Abstract Continuity of Lyapunov Exponents

Lemma 3.3 Let B ∈ C satisfying a fiber-LDT with parameter p = (n 0 , ε, ι) ∈ P.


Let m 1 , m 2 , n ∈ N and η > 0 be such that m i ≥ n ≥ n 0 for i = 1, 2 and
 (m 2 +m 1 ) 
L (B) − L (m i ) (B) < η
1 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)

for all x outside a set of measure < 3 ιn .

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)

Combining (3.41)–(3.43), for x outside a set of measure < 3 ιn we get:

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

> e−(m 1 +m 2 )(η+2εn ) ,

which proves the lemma. 


3.4 The Inductive Step Procedure 99

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)

then for any integer n 1 such that


−1/3
0 ≤ n 1 ≤ n 0 · ιn 0 ,
n 1+ (3.47)

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

From (3.44), applying Lemma 3.3, we have:

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)

for all x outside a set of measure < 3ιn .


This estimate will ensure that the angles condition in the avalanche princi-
ple (Proposition 2.42) holds. Moreover, due to the assumption (3.45), applying
Lemma 3.1, for x outside a set of measure < ιn , we have:

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).

Then clearly g (n)


n−1= B (n 1 ) (x) and gi gi−1 = B (2n 0 ) (T (i−1)n 0 x) for all 1 ≤ i ≤ n − 1.
Let B̄n 0 := i=0 T −i n 0 Bn 0 , so μ(B̄n 0 ) < n ιn 0 and if x ∈ / B̄n 0 then

1
gr(gi ) > for all 0 ≤ i ≤ n − 1,
κap
gi gi−1 
> εap for all 1 ≤ i ≤ n − 1.
gi  gi−1 

Note also that condition (3.46) implies κap εap


2
.
Therefore, we can apply the avalanche principle (Proposition 2.42) and obtain:

 
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

/ B̄n 0 , where μ(B̄n 0 ) < nιn 0 ,


The AP applied to our data implies that for all x ∈

 
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

Divide both sides of (3.55) by n 1 = n · n 0 to get that for all x ∈


/ B̄n 0 we have

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

Using Cauchy-Schwarz and the L 2 -boundedness assumption we have


  
     
 f (x) μ(d x) =  f (x) μ(d x) +  f (x) μ(d x)
X B̄
n0 B̄n 0

 ιn 0 +  f  L 2 1B̄n  L 2 ≤ ιn 0 + C(A) (μ(B̄n 0 ))1/2


0
n0
≤ ιn 0 + C(A) (n ιn 0 )1/2 ≤ C ,
n1

where the last inequality follows from (3.47).


Therefore,

 (n 1 )   
 L (B) + n − 2 L (n 0 ) (B) − 2(n − 1) L (2n 0 ) (B) ≤  f (x) μ(d x) < C n 0 .
1 1 1
n n X n1
102 3 Abstract Continuity of Lyapunov Exponents

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

Using (3.57) for B and A we get:


 (n 1 ) 
 L (B) − L (n 1 ) (A)
1 1
 
<  L (n 1) (n 0 ) (n 0 )
1 (B) − L 1 (B) + 2[L 1 (B) − L 1
(2n 0 )
(B)] 
 
+  L (n 1) (n 0 ) (n 0 )
1 (A) − L 1 (A) + 2[L 1 (A) − L 1
(2n 0 )
(A)] 
 0) 
+  L (n (n 0 )
1 (B) − L 1 (A)

 (n 0 )   0) 
+ 2 L 1 (B) − L (2n 1
0)
(B) + 2 L (n 1 (A) − L 1
(2n 0 )
(A)
n0
< θ0 + 4η0 + C ,
n1

which establishes (3.50), that is, the conclusion (b++) of the proposition. 

3.5 General Continuity Theorem

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.

Proof Let 0 < ε < κ(A)/100 be arbitrary but fixed.


Since L (n)
1 (A) → L 1 (A) as n → ∞, there is n 02 = n 02 (A, ε) ∈ N such that for
all n ≥ n 02 we have
3.5 General Continuity Theorem 103

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)

Then (3.58)–(3.60) imply

L 1(n 0 ) (B) − L (2n


1
0)
(B) < 2 ι1/2
n 0 + ε =: η0 < 3ε. (3.61)

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)

while from Proposition 3.3 we get


(n ) (2n )
L 1 k+1 (B) − L 1 k+1 (B) < ηk+1 , (3.73)
 (n k+1 ) 
L (B) − L 1 k+1 (A) < θk+1 ,
(n )
(3.74)
1

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

< (θ0 + 4η0 ) + 10C n 0−0− < (ε + 12ε) + 10ε = 23 ε.

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 continuity at A of the top Lyapunov exponent L 1 .


Moreover, taking the limit as k → ∞ in (3.72), and using again (3.76) we have

L 1 (B) − L 2 (B) ≥ κ(A) − 46ε − 3ε > L 1 (A) − L 2 (A) − 50ε,

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

holds for an increasing sequence of scales n = n k+1 , k ≥ 0.


A slight modification of the argument shows that (3.77) holds in fact for all large
enough scales n.
106 3 Abstract Continuity of Lyapunov Exponents

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

If A has a different gap pattern, by taking appropriate exterior powers, we can


always reduce the problem to one where there is a gap between the first two Lyapunov
exponents.
For instance, if L 1 (A) = L 2 (A) > L 3 (A) ≥ · · · L m (A), consider instead the
cocycle ∧2 A. Clearly L 1 (∧2 A) = L 1 (A) + L 2 (A) and L 2 (∧2 A) = L 1 (A) + L 3 (A),
hence L 1 (∧2 A) − L 2 (∧2 A) = L 2 (A) − L 3 (A) > 0. Then there is a gap between
the first two Lyapunov exponents of ∧2 A. This implies, using Theorem 3.2 for ∧2 A,
that the block L 1 + L 2 is continuous and the gap L 2 − L 3 is lower semi-continuous
at A.
This argument shows that given a cocycle A ∈ C with any gap pattern, the
corresponding Lyapunov blocks are all continuous at A, while the corresponding
gaps are lower semicontinuous at A.
Moreover, the general assumptions made on the space of cocycles ensure that the
map 
 
Cm  B → L 1 (B) + · · · + L m (B) = logdet[B(x)] μ(d x)
X

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. 

3.6 Modulus of Continuity

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 .

Proposition 3.4 (uniform speed of convergence) Let A ∈ Cm be a measurable


cocycle for which L 1 (A) > L 2 (A). There are δ = δ(A) > 0, C = C(A), n 00 =
n 00 (A) ∈ N, ι = ι(A) ∈ I such that, with the above notations, for all n ≥ n 00 and
for all B ∈ Cm with dist(B, A) < δ we have:
108 3 Abstract Continuity of Lyapunov Exponents

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

(Since n 1  ψ(n 0 ), we have φ(n 1 )  n 0 , as φ is the inverse of ψ.)


The procedure continues in the same way, with intervals of scales defined induc-
tively by Nk+1 = [n − +
k , n k ]  ψ(Nk ) for all k ≥ 0. Again, each two consecutive
intervals of scales Nk and Nk+1 overlap. Therefore, if n ≥ n − 1 , then n = n k+1 ∈ Nk+1
−1/2
for some k ≥ 0, so there is n k ∈ Nk such that n k+1 = n k ιn k   n k + +. We then
have:

(n ) (2n k+1 ) nk φ(n k+1 )


L 1 k+1 (B) − L 1 (B) < C C ,
n k+1 n k+1
 (n k+1 )  nk
L (B) + L (n k) (2n k )
(B)  < C
1 1 (B) − 2L 1  Cι1/3
nk .
n k+1

which completes the proof. 

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.

Theorem 3.3 (modulus of continuity) Let A ∈ Cm be a measurable cocycle for


which L 1 (A) > L 2 (A). There are δ = δ(A) > 0, ι = ι(A) ∈ I and c = c(A) > 0
such that if we define the modulus of continuity function ω(h) := [ι (c log h1 )]1/3 ,
then for any cocycles Bi ∈ Cm with dist(Bi , A) < δ, where i = 1, 2, we have:
 
 L 1 (B1 ) − L 1 (B2 ) ≤ ω(dist(B1 , B2 )). (3.84)

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)

Proof Choose parameters δ0 = δ0 (A) > 0, n 00 = n 00 (A) ∈ N and ι = ι(A) ∈ I


such that both the finite scale uniform continuity Proposition 3.2 and the uniform
110 3 Abstract Continuity of Lyapunov Exponents

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)

Combining (3.86)–(3.89) we conclude:


 
 L 1 (B1 ) − L 1 (B2 )  ι1/3 ≤ [ι (1/(2C1 ) log(1/ h))]1/3
n
=: ω(h) = ω(dist(B1 , B2 )).

The more general assertion of the theorem follows by simply taking exterior
powers. Indeed, the cocycle ∧k A has the property

L 1 (∧k A) = (L 1 +· · ·+ L k−1 + L k )(A) > (L 1 +· · ·+ L k−1 + L k+1 )(A) = L 2 (∧k A),

hence (3.85) follows from (3.84) applied to ∧k A. 

Remark 3.4 If instead of uniform L 2 -boundedness we have uniform L p -boundedness


−1/3
(for some p > 1), then the range n 1 ≤ n 0 ιn 0 in (3.47) should be replaced by
−( p−1)/(2 p−1)
n 1 ≤ n 0 ιn 0 . This is required at some point in the proof of the inductive
step Proposition 3.3, where instead of Cauchy-Schwarz we apply Hölder’s inequal-
ity. Later, this range of values for the scale n 1 leads to the modulus of continuity
stated in the abstract continuity Theorem 3.1.
References 111

References

1. L. Arnold, Random Dynamical Systems, Springer Monographs in Mathematics (Springer, Berlin,


1998). MR 1723992 (2000m:37087)
2. P. Duarte, S. Klein, Continuity of the Lyapunov exponents for quasiperiodic cocycles. Commun.
Math. Phys. 332(3), 1113–1166 (2014). MR 3262622
3. A. Furman, On the multiplicative ergodic theorem for uniquely ergodic systems. Ann. Inst. H.
Poincaré Probab. Stat. 33(6), 797–815 (1997). MR 1484541 (98i:28018)
4. 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(1–2),
155–203 (2001). MR 1847592 (2002h:82055)
5. S. Jitomirskaya, R. Mavi, Continuity of the measure of the spectrum for quasiperiodic
Schrödinger operators with rough potentials. Comm. Math. Phys. 325(2), 585–601 (2014). MR
3148097
6. Y. Katznelson, B. Weiss, A simple proof of some ergodic theorems. Israel J. Math. 42(4), 291–
296 (1982). MR 682312 (84i:28020)
7. W. Schlag, Regularity and convergence rates for the Lyapunov exponents of linear cocycles. J.
Mod. Dyn. 7(4), 619–637 (2013). MR 3177775
Chapter 4
The Oseledets Filtration and Decomposition

Abstract In this chapter we obtain a new proof of the classical Multiplicative


Ergodic Theorem of V. Oseledets, using the Avalanche Principle. Furthermore, we
establish the continuity of the Oseledets filtration and decomposition as functions
of the cocycle, assuming the availability in the space of cocycles of appropriate
uniform large deviation type estimates. The same assumptions lead in the previous
chapter to an abstract continuity theorem of the Lyapunov exponents. This result and
other technical estimates derived there, along with the inductive scheme based on
the Avalanche Principle are the main ingredients of the arguments in this chapter.

4.1 Introduction and Statements

Let (X, μ, T ) be an ergodic dynamical system and let A : X → Mat(m, R) be a


measurable function defining a linear cocycle on the bundle space X × Rm by

X × Rm  (x, v) → (Tx, A(x)v) ∈ X × Rm .

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

{0} = Fk+1 (x)  Fk (x)  · · ·  F2 (x)  F1 (x) = Rm ,

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

The numbers λ1 , λ2 , . . . , λk , measuring the rate of expansion of the cocycle along


the invariant Oseledets subspaces, are the distinct Lyapunov exponents. The repeated

© Atlantis Press and the author(s) 2016 113


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_4
114 4 The Oseledets Filtration and Decomposition

Lyapunov exponents L1 (A) ≥ L2 (A) ≥ · · · ≥ Lm (A) are defined by the Furstenberg-


Kesten (or Kingman’s sub-additive ergodic) theorem.
Making further assumptions (e.g. the base dynamics and the fiber action are
invertible), there is a measurable and (T , A)-invariant decomposition (also called
splitting) into subspaces Rm = ⊕k+1j=1 Ej (x), such that for μ-a.e. x ∈ X and for every
v ∈ Ej (x) \ {0}, limn→±∞ n logA(n) (x) v = λj .
1

There are several methods of proving the multiplicative ergodic theorem.


One approach is due to Raghunathan in [14] who in particular has shown the
existence of the following limits of symmetric matrices
 1/2n  1/2n
lim A(n) (x)∗ A(n) (x) and lim A(n) (T −n x)A(n) (T −n x)∗ . (4.1)
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

Cm  B → E1 (B) ∈ L 1 (X, P(Rm ))

is continuous, with a modulus of continuity depending explicitly on the parameters


of the LDT estimates. In fact, a more precise pointwise statement holds. There are
constants δ > 0, α > 0 and a modulus of continuity function ω(h), all dependent
only on A, such that for any cocycles Bi , i = 1, 2 with dist(Bi , A) < δ,

μ {x ∈ X : d(E1 (B1 )(x), E1 (B2 )(x)) > dist(B1 , B2 )α } < ω(dist(B1 , B2 )),

where as h → 0, ω(h) → 0 at a rate that depends explicitly on the LDT estimates.

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].

4.2 The Ergodic Theorems

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.

4.2.1 Review of Grassmann Geometry Concepts


and Notations

A sequence of integers τ = (τ1 , . . . , τk ) with 1 ≤ τ1 < τ2 < · · · < τk < m is called


a signature. We make the convention that τ0 = 0 and τk = m.
4.2 The Ergodic Theorems 117

Let s1 (g) ≥ s2 (g) ≥ · · · ≥ sm (g) ≥ 0 denote the ordered (repeated) singular


values of a matrix g ∈ Mat(m, R). We say that g has a singular spectrum with a
τ -gap pattern, or shortly that it has a τ -gap pattern, when sτj (g) > sτj +1 (g) for all
j = 1, . . . , k. We say that it has an exact τ -gap pattern when furthermore sτj +1 (g) =
sτj+1 (g) for all j = 0, 1, . . . , k.
Analogously, let L1 (A) ≥ L2 (A) ≥ · · · ≥ Lm (A) ≥ −∞ denote the ordered
(repeated) Lyapunov exponents of a linear cocycle A. We say that A has a Lya-
punov spectrum with a τ -gap pattern, or shortly that it has a τ -gap pattern, when
Lτj (A) > Lτj +1 (A) for all j = 1, . . . , k. We say that it has an exact τ -gap pattern
when furthermore Lτj +1 (A) = Lτj+1 (A) for all j = 0, 1, . . . , k.
Given a matrix g ∈ Mat(m, R) with singular value gap ratio gr(g) := ss21 (g) (g)
> 1,
its most expanding direction is the point v(g) ∈ P(Rm ) determined by any singular
vector of g associated to the first singular value s1 (g) = g.
sk (g)
More generally, if 1 ≤ k ≤ m is such that gr k (g) := sk+1 (g)
> 1, the most
expanding k-subspace is the k-dimensional vector subspace vk (g) spanned by the
singular vectors of g associated to the first k singular values of g.
Finally, when g has a τ -gap pattern, hence gr τ (g) := min1≤j≤k gr τj (g) > 1, we
define the most expanding τ -flag vτ (g) := (vτ1 (g), . . . , vτ1 (g)) ∈ Fτ (Rm ).
A τ -flag in Rm is any finite strictly increasing sequence F = (F1 , . . . , Fk ) of vector
subspaces F1 ⊂ F2 ⊂ . . . ⊂ Fk ⊂ Rm such that dim Fj = τj for all j = 1, . . . , k.
The space of all τ -flags in Rm is denoted here by Fτ (Rm ).
The orthogonal complement F ⊥ of a flag F = (F1 , . . . , Fk ) is the flag F ⊥ =
(Fk , . . . , F1⊥ ) of its orthogonal complements, which has the complementary signa-

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.

Table 4.1 Table of notations Concept Takes values in Defined in


v(g) P(Rm ) 2.14
vk (g) Gr k (Rm ) 2.15
vτ (g) Fτ (Rm ) 2.18
F⊥ Fτ ⊥ (Rm ) 2.6
F  F Dτ (Rm ) 2.35
θ (F, F  ) R 2.34
118 4 The Oseledets Filtration and Decomposition

4.2.2 The Ergodic Theorems of Birkhoff and Kingman

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

A sequence of numbers {an }n≥0 in [−∞, +∞) is called sub-additive if

an+m ≤ an + am for all n, m ≥ 0.

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

Theorem 4.3 (Kingman’s Ergodic Theorem) Let (X, μ, T ) be an ergodic dynamical


system. Given a sequence of measurable functions fn : X → [−∞, +∞) such that
f1+ ∈ L 1 (X, μ) and

fn+m ≤ fn + fm ◦ T n for all n, m ≥ 0, (4.2)



the sequence { fn dμ}n≥0 is sub-additive, and for μ-a.e. x ∈ X, n1 fn (x) converges to
 
1 1
lim fn dμ = inf fn dμ ∈ [−∞, +∞).
n→∞ n n≥1 n

Let B ⊆ X ×Rm be a measurable bundle determined by some measurable function


E : X → Gr(Rm ). This means that

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

Definition 4.1 A linear cocycle on B over a measure preserving dynamical system


(X, μ, T ) is a measurable map FA : B → B, defined by a measurable family of
linear maps A(x) : E(x) → E(Tx), FA (x, v) := (Tx, A(x)v). We identify FA with
the pair (T , A) or simply with A.
Definition 4.2 A cocycle A is said to be μ-integrable if

log+ A(x) dμ(x) < +∞.
X

Proposition 4.1 Given a μ-integrable cocycle A, for μ almost every x ∈ X,

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,

L1 (∧i A) = Li (A) + L1 (∧i−1 A). (4.4)

Proof Consider the exterior power cocycles ∧i A where 1 ≤ i ≤ m + 1. Since

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,

∧i A(n) (x)


si (A(n) (x)) = .
∧i−1 A(n) (x)

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

Lk (A) = L1 (∧k A) − L1 (∧k−1 A) = −∞.

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. 

4.2.3 The Multiplicative Ergodic Theorem

Throughout this section let T : X → X be an ergodic invertible measure preserving


transformation on a probability space (X, F , μ).
Consider a measurable bundle B ⊆ X × Rm determined by some measurable
function E : X → Gr(Rm ) and a μ-integrable linear cocycle FA : B → B, defined
by a measurable family of linear maps A(x) : E(x) → E(Tx).
Given a vector v ∈ E(x) we define the Lyapunov exponents along v

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:

Proposition 4.3 For every x ∈ X, given vectors v, v ∈ E(x),


(a) λA (x, v) ≤ L1 (A),
(b) λA (x, c v) = λA (x, v) if c = 0,
(c) λA (x, v + v ) ≤ max{λA (x, v), λA (x, v )},
(d) if λA (x, v ) < λ− −  
A (x, v) = λA (x, v) then λA (x, v + v ) = λA (x, v + v ) = λA (x, v),
(e) λA (x, v) = λA (Tx, A(x) v).

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

Item (d) follows from the inequality


   
(n) A(n) (x)v  (n)  (n)A(n) (x)v 
A (x)v 1 − ≤ A (x)(v+v ) ≤ A (x)v 1 +
A(n) (x)v A(n) (x)v

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

Fλ (x) := { v ∈ E(x) : λA (x, v) ≤ λ },

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

This proposition will be proved using the Avalanche Principle.


Lemma 4.2 Given ε > 0 there exists r ∈ N such that for any n, n0 ∈ N with n0 ≥ r
and n ≥ r n0 there is a sequence of integers {mi }i≥0 for which
122 4 The Oseledets Filtration and Decomposition

(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

Notice that by the mean value theorem


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. 

Definition 4.4 Given


ε > 0, we call

an ε-doubling sequence any sequence {mi }i≥0
of integers such that
mi − 2 mi−1
< ε mi for all i ≥ 1.

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.

Proof By Birkhoff’s ergodic theorem, for μ-a.e. x ∈ X,

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

Note that if x ∈ Un0 , then


εm
#{ 0 ≤ j ≤ m − 1 : T j x ∈
/ Ω}< for all m ≥ n0 . (4.6)
4

We set Ω0 := Un0 ∩ T −n0 (Ω). Then

μ(X \ Ω0 ) ≤ μ(X \ Ω) + μ(Ω \ Un0 ) + μ(X \ T −n0 (Ω)) < ε,

and if x ∈ Ω0 then (4.6) holds and T n0 x ∈ Ω.


To prove (a), take x ∈ Ω0 and consider the sequence ai := 2i n0 .
For each i ≥ 1, applying (4.6) with m = ai , there is an integer mi in the range
(1 − ε/4)ai ≤ mi ≤ ai such that T mi x ∈ Ω. A straightforward computation shows
ε/4
that {mi }i≥0 is an ε -doubling sequence with ε = 1−ε/4 < ε.
Finally, to prove (b), we use Lemma 4.2 to get an integer r = r(ε) such that if
n ≥ r n0 and n0 ≥ r then there is an 4ε -doubling sequence {mi }i≥0 with m0 = n0 and
mk = n for some index k ≥ 1. Given x ∈ Ω0 , by the frequency bound (4.6) applied
m =
mi + ε 6mi , for each 1 ≤ i ≤ k − 1 there is mi ∈ N such that T mi x ∈ Ω and

with

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.

Proof The following limits exist for μ-a.e. x ∈ X,

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)

The almost sure convergences above imply that

lim μ(Ωn (ε)) = 1.


n→+∞

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)

This completes the proof. 

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

κap = e−mi (κ−2ε)(1−ε)/(1+ε) < e−mi (κ−2ε)/2 and εap = e−5 ε mi ,


κ
for which we have εap2 < e−mi (κ/2−11 ε)  1.
ap
Conclusions (1) and (2) of Proposition 4.5 imply the gap and angle conditions of
the AP. Therefore, by Proposition 2.42,
κap
d(v(A(mi ) (x)), v(A(mi+1 ) (x))) < = e−mi θ ,
εap

where θ := (κ − 2ε) 1−ε


1+ε
− 5ε. Note that θ → κ as ε → 0.
4.2 The Ergodic Theorems 125

By the definition of ε-doubling sequence we have mi+1 ≥ 2


1+ε
mi > 3
2
mi . Hence
mi ≥ (3/2)i n0 for all 0 ≤ i < k. Therefore

d(v(A(n0 ) (x)), v(A(n) (x))) = d(v(A(m0 ) (x)), v(A(mk ) (x)))



k−1
≤ d(v(A(mi ) (x)), v(A(mi+1 ) (x)))
i=0


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

Given 0 ≤ k ≤ m, we define a sequence of partial functions vk(n) (A) on X taking


values in Gr k (Rm ),

vk (A(n) (x)) if gr k (A(n) (x)) > 1,
vk(n) (A)(x) :=
undefined otherwise.

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

Proof Apply Proposition 4.4 to the cocycle ∧k A. 


Definition 4.5 Given a μ-integrable cocycle A, we say that x ∈ X is a μ-regular
point if whenever Lj (A) > Lj+1 (A), we have

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)

Because of this, and since


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

Since the right-hand-side above converges to 0, limn→+∞ cj,n = 0 for all j =


1, . . . , k. Equivalently,

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

Possibly changing the limit points vj , and extracting a subsequence of ns , we may


assume that vj,ns → vj as s → +∞, for all 1 ≤ j ≤ k. Pick any j such that v, vj  = 0.
Since the vectors A(ns ) (x) vj,ns are pairwise orthogonal,


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

Hence, taking logarithms, dividing by ns and passing to the limit we get

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

which contradicts (4.8). This shows that

1
λA (x, v) = lim logA(n) (x) v = γ ,
n→+∞ n

and concludes the proof. 

Definition 4.6 The adjoint of a cocycle (T , A) is the map FA∗ : B → B, defined


by FA∗ (x, v) = (T −1 x, A(T −1 x)∗ v). This cocycle is denoted by (T −1 , A∗ ) or by A∗ .

Remark 4.1 The adjoint cocycle satisfies for any n ∈ N and x ∈ X,

(A∗ )(n) (x) = A(n) (T −n x)∗ .

Proposition 4.9 If A is μ-integrable then the adjoint cocycle A∗ is also μ-integrable.


Moreover, the cocycle A and its adjoint A∗ have the same Lyapunov exponents,
Li (A) = Li (A∗ ) for all i = 1, . . . , m.

Proof The integrability of A∗ follows from the relation A = A∗ .


The second statement is a consequence of a linear operator and its adjoint sharing
the same singular values. In fact, by Proposition 4.2 and the previous remark

1
Li (A) = lim logsi (A(n) (x)) dμ(x)
n→+∞ n X

1
= lim logsi (A(n) (T −n x)) dμ(x)
n→+∞ n X

1
= lim logsi ((A∗ )(n) (x)) dμ(x) = Li (A∗ ). 
n→+∞ n X

Lemma 4.4 If L1 (A) > L2 (A) then for μ-almost every x ∈ X,



α v(∞) (A∗ )(x), v(∞) (A)(x) > 0.

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)

Define Ωi := T mi Ω0 . We have μ(X \ Ωi ) < ε, and for all i ≥ 0 and x ∈ Ωi ,


 
α v(mi ) (A∗ )(x), v(mi+1 −mi ) (A)(x) = α v(A(mi ) (T −mi x)∗ ), v(A(mi+1 −mi ) (x))
 
= α A(mi ) (T −mi x), A(mi+1 −mi ) (x)  e−5 mi ε .
4.2 The Ergodic Theorems 129

Notice that by Proposition 4.4, for i large, the distances

d(v(mi ) (A∗ )(x), v(∞) (A∗ )(x)) and d(v(mi+1 −mi ) (A)(x), v(∞) (A)(x))

are much smaller than e−5 mi ε . Hence on the set Ωi



α v(∞) (A∗ )(x), v(∞) (A)(x)  e−5 mi ε > 0.

Since Ωi has measure μ(Ωi ) > 1 − ε with arbitrary ε, this proves the lemma. 

Definition 4.7 Given a measurable sub-bundle v̂ : X → P(Rm ), we call unit mea-


surable section of v̂ : X → P(Rm ) any measurable function v : X → Rm such that
v(x) = 1 and v(x) ∈ v̂(x) for μ-a.e. x ∈ X.

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,

α0 := α(v̂(Tx), v(A(n) (T −n+1 x)∗ )) = α(v̂(Tx), v((A∗ )(n) (Tx)))


= α(v̂(Tx), v(n) (A∗ )(Tx)) = α(v̂(Tx), v̂n∗ (Tx)) > 0.

Hence by item (a) of Proposition 2.22

A(n) (T −n+1 x)∗ v(Tx) ≥ α0 A(n) (T −n+1 x)∗  > 0.

Finally, since

A(n) (T −n+1 x)∗ v(Tx) = A(n−1) (T −n+1 x)∗ A(x)∗ v(Tx),

we infer that A(x)∗ v(Tx) = 0. 

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

Proposition 4.10 If Lk (A) > Lk+1 (A) then for μ-a.e. x ∈ X,


(a) A(x)∗ [vk(∞) (A)(Tx)] = vk(∞) (A)(x),
(b) A(x)−1 [vk(∞) (A)(Tx)⊥ ] = vk(∞) (A)(x)⊥ .

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

v̂(Tx) ≈ v̂n−1 (Tx) = A(n−1) (Tx)∗ v̂n−1



(T n x).

Hence

A(x)∗ v̂(Tx) ≈ A(x)∗ v̂n−1 (Tx) = A(n) (x)∗ v̂n−1



(T n x)
≈ A(n) (x)∗ v̂n∗ (T n x) = v̂n (x) ≈ v̂(x).

Item (a) follows from taking limits in these proximity relations.


On the first occurrence of ≈ we use the continuity of the projective action of A(x)∗
and Lemma 4.5. This lemma asserts that A(x)∗ v(Tx) = 0 for μ-a.e. x ∈ X and any
unit measurable section v of v̂. By Proposition 2.26 we have
   
d A(x)∗ v̂(Tx), A(x)∗ v̂n−1 (Tx) ≤ C d v̂(Tx), v̂n−1 (Tx) → 0

where C is any constant larger than A(x)∗ v(Tx)−1 .


On the second occurrence of ≈, take 0 < κ < L1 (A) − L2 (A), 0 < ε  κ
arbitrary small and, by Egorov’s theorem, a measurable subset E ⊂ X such that v̂n∗
converges uniformly to v̂∗ on E. Then choose a sequence of times n ∈ N such that
T n x ∈ E and gr(A(n) (x)∗ ) = gr(A(n) (x)) ≥ en κ . Because of this large gap ratio,
A(n) (x)∗ acts as a strong contraction in a neighborhood of v̂n∗ (T n x). But for T n x ∈ E,

v̂n∗ (T n x) and v̂n−1 (T n x) are both very close to v̂∗ (T n x), and hence close to each other.
Thus
δ( A(x)∗ v̂n−1 (Tx), A(x)∗ v̂n (Tx) )  δ(v̂n−1 (Tx), v̂n (Tx))

converges to 0 as n → +∞.
On the last occurrence of ≈ we apply Proposition 4.4. 

Lemma 4.6 Given a measurable function f : X → R such that f −f ◦T ∈ L 1 (X, μ),


for μ-a.e. x ∈ X,
1
lim f (T n x) = 0.
n→+∞ n
4.2 The Ergodic Theorems 131

Proof First prove that lim inf n→+∞ n1


f (T n x)
= 0 using Poincaré’s recurrence the-
orem. Then notice that

1 
n−1
1 1
f (T n x) = f (x) − (f − f ◦ T )(T j x),
n n n j=0

and apply Birkhoff’s ergodic theorem. 


Lemma 4.7 Let T : X → X be an ergodic MPT on a probability space (X, μ),
and let f : X → (0, +∞) be a measurable non-integrable function. Then for μ-a.e.
x ∈ X,
1 
n−1
lim f (T j x) = +∞.
n→+∞ n
j=0

Proof Defining fn = max{f , n}, by Lebesgue’s monotone convergence theorem


 
lim fn dμ = f dμ = +∞.
n→+∞ X X

For each n ∈ N, since fn is μ-integrable


m−1  a full measure set Bn ⊆ X such that
there is
for all x ∈ Bn , limm→+∞ m1 f
j=0 n (T j
x) = X fn dμ. Thus B = ∩n∈N Bn is also a
full measure set. 
Given x ∈ B and L > 0, consider p ∈ N such that X fp dμ > L. Since x ∈ Bp ,
there is an order n0 = n0 (x) > p such that for n ≥ n0

1  1 
n−1 n−1
f (T x) ≥
j
fp (T j x) ≥ L,
n j=0 n j=0

which proves the lemma. 


Proposition 4.11 Assume L1 (A) > L2 (A) and let v, v∗ : X → Rm be unit measur-
able sections of v(∞) (A) and v(∞) (A∗ ), respectively. Then the functions logA v∗ 
and log(A ◦ T −1 )∗ v are μ-integrable, and
 
logA(x) v∗ (x) dμ(x) = logA(T −1 x)∗ v(x) dμ(x) = L1 (A).
X X

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

Let vn : X → Rm be a unit measurable section of v(n) (A). For notational simplicity


we will also write v̂∗ (x) := v(∞) (A∗ )(x) and v̂n (x) := v(n) (A)(x). By item (a) of
Proposition 2.22,
 
A(n) (x)v∗ (x) ≥ α v̂∗ (x), v̂n (x) A(n) (x),

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

α(v̂∗ (x), v̂(x)) α(v̂∗ (x), A(x)∗ v̂(Tx))


f (x) − f (Tx) = log = log
α(v̂∗ (Tx), v̂(Tx)) α(A(x) v̂∗ (x), v̂(Tx))
∗ ∗
v (x), A(x) v(Tx) A(x)v∗ (x)
= log ∗
A(x) v(Tx) (A(x)v∗ (x), v(Tx)
= logA(x)v∗ (x) − logA(x)∗ v(Tx).

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.

Thus, taking logarithms and dividing by n, (b) follows. 

Proposition 4.13 Given x ∈ X and unit vectors vk ∈ ∧k E(x) and vr ∈ ∧r E(x),

λ∧k+r A (x, vk ∧ vr ) ≤ λ∧k A (x, vk ) + λ∧r A (x, vr ),


λ− − −
∧k+r A (x, vk ∧ vr ) ≤ λ∧k A (x, vk ) + λ∧r A (x, vr ).

Proof By item (a) of Proposition 2.34,

log∧k+r A(n) (x)vk ∧ vr  ≤ log∧k A(n) (x)vk  + log∧r A(n) (x)vr .

Hence, dividing by n an passing to the limit, the inequalities follow. 


 
Proposition 4.14 Assume Lk (A) > Lk+1 (A). Given unit vectors vk ∈ ∧k vk(∞) (A)(x)
 
and vr ∈ ∧r vk(∞) (A)(x)⊥ ,

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

In view of Proposition 4.13, it is enough to prove that

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) .

Therefore, by Proposition 4.12 (b),

λ∧k A (x, vk ) + λ− ∧r A (x, vr )


1
= lim inf log∧k A(n) (x)vk  ∧r A(n) (x) vr 
n→+∞ n
1 1
≤ lim inf log∧k+r A(n) (x) (vk ∧ vr ) + lim inf log αn (x)−1
n→+∞ n n→+∞ n
1
≤ λ− ∧k+r A (x, vk ∧ vr ) − lim sup log αn (x) = λ−
∧k+r A (x, vk ∧ vr ).
n→+∞ n

Assume now that λ−


∧r A (x, vr ) = λ∧r A (x, vr ). Combining Proposition 4.13 with the
previous inequality

λ∧k+r A (x, vk ∧ vr ) ≤ λ∧k A (x, vk ) + λ∧r A (x, vr )


≤ λ∧k A (x, vk ) + λ−
∧r A (x, vr )
≤ λ−
∧k+r A (x, vk ∧ vr ),

which implies that λ−


∧k+r A (x, vk ∧ vr ) = λ∧k+r A (x, vk ∧ vr ). 

Definition 4.8 Given a μ-regular point x ∈ X of a cocycle A, we call limit singular


basis of the fiber E(x) any orthonormal basis {u1 , . . . , um } of E(x) obtained as a limit
point of a sequence of singular basis {u1,n , . . . , um,n } of A(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

L1 (∧i A) = · · · = Lk (∧i A) > Lk+1 (∧i A).

Since u1 ∧ · · · ∧ ui is a limit point of u1,n ∧ · · · ∧ ui,n , which is a sequence of vectors


in vk(n) (∧i A)(x), we infer that u1 ∧ · · · ∧ ui ∈ vk(∞) (∧i A)(x). The conclusion follows
then applying Proposition 4.8 to the cocycle ∧i A. 

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)⊥ .

Proof It is enough to see that



L1 (∧k A) + L1 ∧i A|v⊥k = L1 (∧i+k A). (4.10)

In fact, from (4.10), using Corollary 4.1,

L1 (A) + · · · + Lk (A) + L1 (A|v⊥k ) + · · · + Li (A|v⊥k ) = L1 (A) + · · · + Li+k (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,

L1 (∧i+k A) = λ∧i+k A (x, u1 ∧ · · · ∧ uk+i )


≤ λ∧k A (x, u1 ∧ · · · ∧ uk ) + λ∧i A (x, uk+1 ∧ · · · ∧ uk+i )
≤ L1 (∧k A) + L1 (∧i A|v⊥k ).

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

Hence, by Propositions 4.14 and 4.8,

L1 (∧i+k A) ≥ λ∧k+i A (x, u1 ∧ · · · ∧ uk ∧ uk+1 ∧ . . . ∧ uk+i )


= λ∧k A (x, u1 ∧ · · · ∧ uk ) + λ∧i A (x, uk+1 ∧ · · · ∧ uk+i )
= L1 (∧k A) + L1 (∧i A|v⊥k ).

Putting together these two inequalities we prove (4.10). 


Proposition 4.16 Consider integers 1 ≤ k < k + r ≤ m such that

Lk (A) > Lk+1 (A) = · · · = Lk+r (A) > Lk+r+1 (A).

Then for μ-almost every x ∈ X,

vr(∞) (A|v⊥k )(x) = vk(∞) (A)(x)⊥ ∩ vk+r


(∞)
(A)(x).

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

Lτj (A) > Lτj +1 (A), for all 1 ≤ j < k.

The size of these gaps is measured by

gapτ (A) := min Lτj (A) − Lτj +1 (A).


1≤j<k
4.2 The Ergodic Theorems 137

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

Proof Apply Proposition 4.6 at the dimensions i = τj , with j = 1, . . . , k. 

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.

Theorem 4.4 (Oseledets I) Let T : X → X be an invertible ergodic MPT of a


probability space (X, F, μ), and let FA : B → B be a μ-integrable linear cocycle
on a measurable bundle B ⊆ X × Rm .
Then there exist λ1 > λ2 > · · · > λk ≥ −∞ and a family of measurable functions
Fj : X → Gr(Rm ), 1 ≤ j ≤ k, such that for μ-almost every x ∈ X,
(a) A(x) Fj (x) ⊆ Fj (Tx) for j = 1, . . . , k
(b) {0} = Fk+1 (x)  Fk (x)  . . .  F2 (x)  F1 (x) = B(x)
1
(c) for every v ∈ Fj (x) \ Fj+1 (x), lim logA(n) (x) v = λj .
n→+∞ n

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).

Analogously, and using Proposition 4.15,



λA (x, v ) ≤ L1 A|v⊥τ = Lτj +1 (A) = Lτj+1 (A) = λj+1 (A) < λj (A).
j

Finally, applying item (d) of Proposition 4.3 we infer that

λ− 
A (x, v) = λA (x, v) = λA (x, u + v ) = λA (x, u) = λj (A).

This proves (c). 

Recall that Rg and Kg denote respectively the range and the kernel of any given a
linear map g : V → V  .

Definition 4.9 Given a linear map g : V → V  between Euclidean spaces V and


V  its pseudo inverse g+ : V  → V is the composition g+ := (g|Kg⊥ )−1 ◦ πRg of the
orthogonal projection πRg : V  → Rg with the inverse of g|Kg⊥ : Kg⊥ → Rg .

Lemma 4.9 For any linear map g : V → V  , and integer 0 ≤ k ≤ dim V ,

∧k (g+ ) = (∧k g)+ .

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

∧k (g+ ) = ∧k (g|Kg⊥ )−1 ◦ ∧k πRg = (∧k g|∧k Kg⊥ )−1 ◦ π∧k Rg


= (∧k g|K∧⊥ g )−1 ◦ πR∧k g = (∧k g)+ .
k

These cumbersome algebraic calculations have a natural geometric meaning. 

Definition 4.10 Given a cocycle A : X → Mat(m, R), we define for n > 0

A(−n) (x) := A(n) (T −n x)+ .


4.2 The Ergodic Theorems 139

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 ).

Theorem 4.5 (Oseledets II) Let T : X → X be an invertible ergodic MPT of a


probability space (X, F, μ), and let FA : B → B be a μ-integrable linear cocycle
on a measurable bundle B ⊆ X × Rm .
Then there exist λ1 > λ2 > . . . > λk+1 ≥ −∞ and a family of measurable
functions Ej : X → Gr(Rm ), 1 ≤ j ≤ k + 1, such that for μ-almost every x ∈ X,
(a) B(x) = ⊕k+1 j=1 Ej (x),
(b) A(x) Ej (x) = Ej (Tx) for j = 1, . . . , k, and A(x) Ek+1 (x) ⊆ Ek+1 (Tx),
1
(c) for every v ∈ Ej (x) \ {0}, lim logA(n) (x) v = λj ,
n→±∞ n
1

(d) lim log


sin min (⊕j≤l Ej (T n x), ⊕j>l Ej (T n x))
= 0, for any l = 2, . . . , k.
n→±∞ n

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

v∗τ (x) = (vτ(∞)


1
(A∗ )(x), . . . , vτ(∞)
k
(A∗ )(x)) ∈ Fτ (E(x)),
v⊥ (∞) ⊥ (∞) ⊥
τ (x) = (vτk (A)(x) , . . . , vτ1 (A)(x) ) ∈ Fτ ⊥ (E(x)).

According to Definition 2.35 we have

{Ej (x)}1≤j≤k+1 = v∗τ (x)  v⊥


τ (x).

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,

θ (v∗τ , v⊥ (∞) ∗ (∞) ⊥


τ ) = min θ∩ (vτi (A ), vτi (A) )
1≤i≤k

= min ατi (vτ(∞)


i
(A∗ ), vτ(∞)
i
(A))
1≤i≤k

= min α(v(∞) (∧τi A∗ ), v(∞) (∧τi A)) > 0.


1≤i≤k

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

We prove (c) through several reductions.


Consider first the case j = 1 and τ1 = 1. In this case τ0 = 0 and the intersection
sub-bundle Ej is the 1-dimensional A-invariant bundle v(∞) (A∗ ).
Let v∗ : X → P(Rm ) be a unit measurable section of this bundle. By invariance
we have A(x) v∗ (x) = ±v∗ (Tx) for μ-a.e. x ∈ X, and hence


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

On the other hand, since T is invertible, the Birkhoff averages


n−1
(n) −n ∗ −n
logA (T x) v (T x) = logA(T −j x) v∗ (T −j x)
j=0

also converge μ-almost everywhere to L1 (A). Now, inverting the relation

A(n) (T −n x) v∗ (T −n x) = A(n) (T −n x) v∗ (T −n x) v∗ (x),

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

By Proposition 4.13 we have


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

Finally consider the general case, where 2 ≤ j ≤ k. By Proposition 4.16,

Ej = vτ(∞)
j
(A∗ ) ∩ vτ(∞)
j−1
(A)⊥ = vτ(∞)
j −τj−1
(A|v⊥τ ).
j−1

We denote this A-invariant sub-bundle by Bj . Given a non-zero vector v ∈ Bj (x),


applying the previous case to the restricted cocycle A|Bj by Proposition 4.15

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).

This proves (c).


For the last item, (d), notice that ⊕j≤l Ej = vτ(∞)
l
(A∗ ), while ⊕j>l Ej = vτ(∞)
l
(A)⊥ .
Hence by Proposition 2.19 (d),


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)).

Thus item (d) follows by Proposition 4.12 (a). 

4.3 Abstract Continuity Theorem of the Oseledets


Filtration

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

3. Every cocycle with finite maximal Lyapunov exponent is uniformly L p -bounded,


where 1 < p ≤ ∞. For simplicity of notations we let p = 2.
4. Every cocycle A ∈ C such that L1 (A) > L2 (A) satisfies a uniform fiber-LDT
estimate: for all ε > 0 there are δ = δ(A, ε) > 0 and p = p(A, ε) ∈ P,
p = (n0 , ε, ι), such that if B ∈ Cm with dist(B, A) < δ and if n ≥ n0 then εn ≤ ε
and

1

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

e−c0 t < ι(t) < t −10 as t → ∞,

where c0 > 0 is some fixed constant.

4.3.1 Continuity of the Most Expanding Direction

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.

By Proposition 4.4, as n → ∞ the functions v(n) (A)(x) converge μ a.e. to a


measurable function v(∞) (A) : X → P(Rm ).
Let L 1 (X, P(Rm )) be the space of all Borel measurable functions F : X → P(Rm ).
Consider the distance

dist(F1 , F2 ) := d(F1 (x), F2 (x)) μ(dx),
X

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 → ∞,

v(n) (A) → v(∞) (A) in L 1 (X, P(Rm )). (4.13)

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

holds for all m, n ≥ n0 .


144 4 The Oseledets Filtration and Decomposition

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

be the exceptional set in the uniform fiber-LDT estimate, so μ [Bldt


n (B)] < ιn .
Let
Bgn (B) := Bldt
n (B) ∪ Bn (∧2 A).
usc

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

B(m2 +m1 ) (x)


> e−(m1 +m2 ) κ(A)/20 > e−n κ(A)/5 =: εn (4.16)
B(m2 ) (T m1 x) B(m1 ) (x)

holds provided that


−m1 ldt
x∈ m2 +m1 (B) ∪ Bm1 (B) ∪ T
/ Bldt ldt
Bm2 (B).

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

Clearly μ [Ban (B)]  n ιn , and if x ∈


/ Ban (B), then the angle condition will be ensured
for block components of the kind indicated above.
4.3 Abstract Continuity Theorem of the Oseledets Filtration 145

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 .

The following results are now easy to phrase and to prove.


−1/3 ap
Lemma 4.10 Let n1 , n0 ∈ N such that n0 ≥ n0 and 2n0 ≤ n1 ≤ ιn0 . If x ∈
/ Bn0 (B)
then κn
d(v(B(n1 ) (x)), v(B(n0 ) (x))) < 0 . (4.20)
εn0
146 4 The Oseledets Filtration and Decomposition

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

gn−1 = gn−1 (x) := B(m) (T (n−1) n0 x).

Then
g(n) = gn−1 . . . g1 g0 = B(n1 ) (x),

gi gi−1 = B(2n0 ) (T (i−1)n0 x)

for 1 ≤ i ≤ n − 2, while

gn−1 gn−2 = B(m+n0 ) (T (n−2)n0 x).


ap
/ Bn0 (B), we are in the setting described in Remark 4.2, hence the AP in
Since x ∈
the form described in Proposition 2.42 in Chap. 2 applies and we have:
κn0
d(v(B(n1 ) (x)), v(B(n0 ) (x))) = d(v(g(n) ), v(g0 ))  . 
εn0

/ Bn0 (B) then


Lemma 4.11 For all n0 ≥ n0 , m ≥ n02+ , if x ∈

κn0
d(v(B(m) (x)), v(B(n0 ) (x)))  . (4.21)
εn0

Proof Fix 0 < c  1. Let ψ(t) := t 2 .


Define inductively the following intervals of scales N0 := [n01+c , n03+c ] ⊂
−1/3
[2n0 , ιn0 ], N1 := ψ(N0 ) = [n02+2c , n06+2c ] and for all k ≥ 0, Nk+1 := ψ(Nk ).
These intervals overlap, so they cover up all integers ≥ n01+c .
Let m ≥ n02+2c . Then there is k ≥ 0 such that m = mk+1 ∈ Nk+1 , and so
mk+1  mk2 for some mk ∈ Nk . In fact, there is a backward “orbit” of integers
m0 ∈ N0 , m1 ∈ N1 , . . . , mk ∈ Nk such that mj+1  mj2 .
For any 0 ≤ j ≤ k, since mj ≥ m0 ≥ n0 , and since x ∈ / Bn0 (B), by (4.19) we
ap −1/3
have x ∈ / Bmj (B). Moreover, since mj+1  mj2 , we have 2mj ≤ mj+1 ≤ ιmj , hence
Lemma 4.10 is applicable to the scales mj , mj+1 and we get

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

Using (4.23), (4.22) and the triangle inequality we get

d(v(B(m) (x)), v(B(n0 ) (x)))


≤ d(v(B(m0 ) (x)), v(B(n0 ) (x)))

k
+ d(v(B(mj+1 ) (x)), v(B(mj ) (x)))
j=0

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. 

From Proposition 4.4 and by the proximity of the cocycle B to A, we already


know that its most expanding direction of v(∞) (B)(x) is well defined as the μ-a.e.
limit as n → ∞ of the finite scale most expanding direction v(n) (B)(x). We prove a
quantitative version of this convergence.

/ Bn (B), hence for x


Proposition 4.18 (speed of convergence) For all n ≥ n0 , if x ∈
1/2
outside a set of measure < ιn , we have
κn
d(v(B(n) (x)), v(∞) (B)(x)) < = e−3 nκ(A)/10 < ιn . (4.24)
εn

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. 

Proposition 4.19 (finite scale continuity) Given C2 > 0, there is a constant C1 =


C1 (A, C2 ) < ∞, such that for any B1 , B2 ∈ Cm with dist(Bi , A) < δ, where i = 1, 2,
if n ≥ n0 and if dist(B1 , B2 ) < e−C1 n , then for x outside a set of measure < ιn we
have
dist(v(B1(n) (x)), v(B2(n) (x))) < e−C2 n < ιn . (4.26)

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 .

Moreover, assuming dist(B1 , B2 ) < e−C1 n , with C1 to be chosen later,

g1 − g2  = B1(n) (x) − B2(n) (x) ≤ n eC0 (n−1) dist(B1 , B2 ) < e−(C1 −2C0 ) n .

If we choose C1 > 3C0 + C2 , then

g1 − g2  e−(C1 −2C0 ) n


drel (g1 , g2 ) := ≤ < e−C2 n  1.
max{g1 , g2 } e−C0 n

Proposition 2.40 in Chap. 2 applies, and we conclude:

16
d(v(g1 ), v(g2 )) ≤ drel (g1 , g2 )  e−C2 n .
1 − κn2

This proves (4.26), while (4.27) follows by integration in x. 

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

where ωι (h) := [ι (c log(1/h))]1/2 is a modulus of continuity function, and clearly


ωι (h) → 0 as h → 0.
Moreover,
dist(v(∞) (B1 ), v(∞) (B2 )) < ωι (dist(B1 , B2 )). (4.28)

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 ,

dist(v(B1(n) (x)), v(B2(n) (x))) < e−C2 n . (4.29)

Now apply Proposition 4.18 to B = Bi , i = 1, 2, to get that for x outside a set of


1/2
measure < ιn ,
d(v(Bi(n) (x)), v(∞) (Bi )(x)) < e−3 nκ(A)/10 . (4.30)

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

d(v(∞) (B1 )(x), v(∞) (B2 )(x)) < e−c0 n = hα ,

where α = Cc01 . This proves the pointwise estimate.


To prove (4.28), simply integrate in x and take into account the fact that since the
large deviation measure function ι decays at most exponentially, the corresponding
modulus of continuity function ωι will decay at most like a power of h, so we may
assume hα < ωι (h) as h → 0. 

4.3.2 Spaces of Measurable Filtrations and Decompositions

We introduce a space of measurable filtrations, i.e. a space of functions from the


phase space to the set of all flags. Thus the Oseledets filtration of a linear cocycle is
an element of this space. We endow the space of measurable filtrations with a natural
topology. Similarly, we define a space of measurable decompositions.
We start with an example that will motivate the formalism below. Let A be a linear
cocycle with exact gap pattern say τ = (2, 3), that is,

L1 (A) = L2 (A) > L3 (A) > L4 (A) = · · · = Lm (A).

The Oseledets filtration of A is a τ ⊥ = (m − 3, m − 2)-flag

{0} = F4 (A)(x)  F3 (A)(x)  F2 (A)(x)  F1 (A)(x) = Rm ,


150 4 The Oseledets Filtration and Decomposition

for μ-a.e. x ∈ X, thus defining a measurable function F(A) : X → Fτ ⊥ (Rm ).


The growth rate of the iterates of A along vectors in F3 (A)(x) is L4 (A) or less,
the growth rate along vectors in F2 (A)(x) is L3 (A) or less and the growth rate along
vectors in F1 (A)(x) is L1 (A) or less.
By the continuity of the Lyapunov exponents of a linear cocycle (which holds
under the assumptions in this section), if B is a small perturbation of A, then

L1 (B) ≥ L2 (B) > L3 (B) > L4 (B) ≥ · · · ≥ Lm (B),

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

{0} = F5 (B)(x)  F4 (B)(x)  F3 (B)(x)  F2 (B)(x)  F1 (B)(x) = Rm ,

for μ-a.e. x ∈ X, thus defining a measurable function F(B) : X → Fτ ⊥ (Rm ).


The subspaces F4 (B)(x), F3 (B)(x), F2 (B)(x) and F1 (B)(x) correspond to the Lya-
punov exponents L4 (B), L3 (B), L2 (B) and L1 (B) respectively.
In order to compare the Oseledets filtration of B with that of A, we would need
to “forget” the extra subspace F2 (B)(x) corresponding to the Lyapunov exponent
L2 (B), which appears precisely because the gap pattern τ  of B is finer than that of
A. In other words, we consider the projection F τ (B) of the Oseledets filtration F(B)
to the space of coarser τ ⊥ = (m − 3, m − 2)-flags valued filtrations

{0} = F5 (B)(x)  F4 (B)(x)  F3 (B)(x)  F1 (B)(x) = Rm .

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

ρτ,τ  (F) = ρτ,τ  (F1 , . . . , Fk ) := (Fi1 , . . . , Fik ),

where τj = τij for j = 1, . . . , k  .


With respect to the following normalized distance on the flag manifold Fτ (Rm )
(see (2.13) in Chap. 2),
dτ (F, F  ) = max d(Fj , Fj )
1≤j≤k

these projections are Lipschitz, with Lipschitz constant 1.


Let (X, F , μ) be a probability space and let T : X → X be an ergodic measure
preserving transformation.
We call measurable filtration of Rm any mod 0 equivalence class of an F -
measurable function F : X → F(Rm ). Two functions F, F  : X → F(Rm ) are
said to be equivalent mod 0 when F(x) = F  (x) for μ-a.e. x ∈ X. From now on we
will identify each mod 0 equivalence class with any of its representative measurable
functions.
Given any measurable filtration F of Rm , let τ (F)(x) denote the signature of the
flag F(x). We say that F has a T -invariant signature if τ (F)(x) = τ (F)(Tx) for μ-a.e.
x ∈ X. If this is the case, then by the ergodicity of (T , μ) the function τ (F)(x) is
constant μ-a.e.
Define F(X, Rm ) to be the space of mod 0 equivalence classes of measurable
filtrations with a T -invariant signature, which is a constant that we denote by τ (F).
We say that F has a τ -pattern when τ (F) ≥ τ .
Given a signature τ , let us define F⊃τ (X, Rm ) to be the subspace of measurable
filtrations in F(X, Rm ) with a τ -pattern.
By definition F⊃τ (X, Rm ) ⊆ F⊃τ  (X, Rm ), whenever τ ≥ τ  .
Given F ∈ F⊃τ (X, Rm ), the function

F τ (x) := ρτ (F),τ (F(x))

determines a measurable filtration with constant signature τ , which will be referred


to as the τ -restriction of F.
We endow F⊃τ (X, Rm ) with the following distance

 
distτ (F, F  ) := dτ F τ (x), (F  )τ (x) dμ(x). (4.31)
X

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,

Vδ,τ (F) := { F  ∈ F⊃τ (X, Rm ) : distτ (F, F  ) < δ },

where δ > 0, F ∈ F(X, Rm ) and τ = τ (F).


We note that this topology is not metrizable.
Proposition 4.20 Let C be a topological space. A map F : C → F(X, Rm ) is
continuous w.r.t. this topology if and only if for all A ∈ C such that F(A) has a
τ -pattern, there exists a neighborhood U ⊂ C of A such that F(U) ⊆ F⊃τ (X, Rm )
and the τ -restricted function F τ |U : U → F⊃τ (X, Rm ), B → F τ (B), is continuous
w.r.t. the distance distτ defined above.
Proof Assume first that F : C → F(X, Rm ) is continuous and take A ∈ C such
that F(A) ∈ F(X, Rm ). Consider the neighborhood Vδ,τ (F(A)) of F(A) where
δ > 0. By continuity of F, there exists a neighborhood U ⊂ C of A such that
F(U) ⊂ Vδ,τ (F(A)) ⊂ F⊃τ (X, Rm ). By definition of the topology in F(X, Rm ),
the set F⊃τ (X, Rm ) is open in F(X, Rm ), and the projection ρτ : F⊃τ (X, Rm ) →
F⊃τ (X, Rm ), ρτ (F) = F τ , is continuous. The restriction F τ |U : U → F⊃τ (X, Rm ),
B → F τ (B), is continuous because it coincides with the composition ρτ ◦ F.
The converse statement is a direct consequence of the definition. 
Recall that a τ -decomposition is a family of linear subspaces E· = {Ei }1≤i≤k+1 in
Gr(Rm ) such that Rm = ⊕k+1 i=1 Ei and dim Ei = τi − τi−1 for all 1 ≤ i ≤ k + 1. In
Chap. 2, we have denoted by Dτ (V ) the space of all τ -decompositions of Rm .
Given τ ≥ τ  , there is a natural projection ρτ,τ  : Dτ (Rm ) → Dτ  (Rm ), defined
by
ρτ,τ  (E· ) = ρτ,τ  (E1 , . . . , Ek+1 ) := (E1 , . . . , Ek  +1 ),

where Ej = ⊕ij ≤l<ij+1 El and τij = τj for j = 1, . . . , k  .


On the space of decompositions Dτ (Rm ) we consider the distance (see Defini-
tion (2.33)),
dτ (E· , E· ) = max dτj −τj+1 (Ej , Ej ).
1≤j≤k+1

By Proposition 2.30, the projections ρτ,τ  are locally Lipschitz.


An equivalence class mod 0 of an F-measurable function E· : X → D(Rm ) :=
∪τ Dτ (Rm ) will be called a measurable decomposition of Rm . Two measurable
decompositions E· , E· : X → D(Rm ) are equivalent mod 0 when E· (x) = E· (x)
for μ-a.e. x ∈ X. As before, we will identify each mod 0 equivalence class with any
of its representative measurable functions.
Given any measurable decomposition E· of Rm , its signature at a point x ∈ X is
the sequence of dimensions τ = (τ1 , . . . , τk ), where τj = dim ⊕l≤j Ej (x) for all
j = 1, . . . , k. We denote it by τ (E· )(x). We say that E· has a T -invariant signature if
τ (E· )(x) = τ (E· )(Tx) for μ-a.e. x ∈ X. In this case, by the ergodicity of (T , μ) the
function τ (E· )(x) is constant μ-a.e.
4.3 Abstract Continuity Theorem of the Oseledets Filtration 153

Define D(X, Rm ) to be the space of mod 0 equivalence classes of measurable


decompositions with a T -invariant signature, that we denote by τ (E· ).
We say that E· has a τ -pattern when τ (E· ) ≥ τ .
Given a signature τ , define D⊃τ (X, Rm ) to be the subspace of measurable decom-
positions in D(X, Rm ) with a τ -pattern.
By definition D⊃τ (X, Rm ) ⊆ D⊃τ  (X, Rm ), whenever τ ≥ τ  .
Given E· ∈ D⊃τ (X, Rm ), the function

E·τ (x) := ρτ (E· ),τ (E· (x))

determines a measurable decomposition with constant signature τ , referred to as the


τ -restriction of E· .
We endow D⊃τ (X, Rm ) with the following distance

 
distτ (E· , E· ) := dτ E·τ (x), (E· )τ (x) dμ(x). (4.32)
X

Finally, we endow the space D(X, Rm ) of all measurable decompositions of Rm


with the topology determined by the following neighborhood bases,

Vδ,τ (E· ) := { E· ∈ D⊃τ (X, Rm ) : distτ (E· , E· ) < δ },

where δ > 0, E· ∈ D(X, Rm ) and τ = τ (E· ).


Again, this topology is not metrizable, and a similar characterization of the con-
tinuity of a map E· : C → D(X, Rm ) holds.

Proposition 4.21 Let C be a topological space. A map E· : C → D(X, Rm ) is


continuous w.r.t. this topology if and only if for all A ∈ C such that E· (A) has a
τ -pattern, there exists a neighborhood U ⊂ C of A such that E· (U) ⊆ D⊃τ (X, Rm )
and the τ -restricted function E·τ |U : U → D⊃τ (X, Rm ), B → E·τ (B), is continuous
w.r.t. the distance distτ defined above.

4.3.3 Continuity of the Oseledets Filtration

We denote by F(A) the Oseledets filtration of a cocycle A ∈ Cm . If A has a τ gap


pattern, by the continuity of the Lyapunov exponents, any nearby cocycle B has the
same or a finer gap pattern τ  ≥ τ . Let F τ (B) denote the projection of the Oseledets
filtration of B to the space F⊃τ (X, Rm ) of measurable filtrations with a τ -pattern. We
are now ready to phrase and to prove the continuity of the Oseledets filtration.

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 )),

for any B1 , B2 ∈ Cm in a neighborhood of A, and for some α = α(A) > 0.


Moreover, the map Cm  A → F(A) ∈ F(X, Rm ) is continuous everywhere.

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

Cm  B → ∧τj B → v(∞) (∧τj B) ∈ L 1 (X, P(∧τj Rm ))

is continuous at A, with a modulus of continuity of the form ω(h) = [ι (c log h1 )]1/2 .


A similar pointwise estimate holds as well.
The Oseledets filtration of A was obtained in the proof of the Oseledets Theo-
 ⊥
rem 4.4 as F(A)(x) = vτ(∞) (A)(x) , where

vτ(∞) (A)(x) = vτ(∞)
1
(A)(x), . . . , vτ(∞)
k
(A)(x) ,

and
vτ(∞)
j
(A)(x) = Ψ −1 (v(∞) (∧τj A)(x)).

Moreover, since for any nearby cocycle B we clearly have



F τ (B)(x) = Ψ −1 (v(∞) (∧τ1 B)(x)), . . . , Ψ −1 (v(∞) (∧τk B)(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. 

4.3.4 Continuity of the Oseledets Decomposition

We denote by E· (A) the Oseledets decomposition of a cocycle A ∈ Cm . Assume that


A has a τ = (τ1 , . . . , τk ) gap pattern. By the construction in the proof of Theorem 4.5,
we have
E· (A)(x) = vτ(∞) (A∗ )(x)  vτ(∞) (A)(x)⊥ .
4.3 Abstract Continuity Theorem of the Oseledets Filtration 155

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

E·τ (B)(x) = vτ(∞) (B∗ )(x)  vτ(∞) (B)(x)⊥ .

We may immediately conclude from Sect. 4.3.3, or directly from he continuity of


the most expanding direction derived in Sect. 4.3.1 that the maps

B → vτ(∞) (B∗ ) and B → vτ(∞) (B)⊥

are continuous in a neighborhood of A, with an appropriate modulus of continuity.


However, this does not automatically guarantee the continuity of the intersection.
Indeed, by Proposition 2.38 in Chap. 2, the intersection map  : Fτ (V ) × Fτ ⊥ (V ) →
Dτ (V ) is locally Lipschitz, but with a Lipschitz constant that depends on the transver-
sality measurement of the subspaces, which may blow up for some phases x.
That is why we need to control these transversality measurements at finite scale
first. We will employ a similar scheme as in the establishing of the continuity of the
most expanding direction in Sect. 4.3.1.
Recall from Sect. 4.2.3 the nth scale partial functions vτ(n) (B) on X taking values
in Fτ (Rm ),
vτ (B(n) (x)) if gr τ (B(n) (x)) > 1
vτ(n) (B)(x) :=
undefined otherwise,

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

1. vτ (B(n) (x)) is well defined for all phases x ∈


/ Bn (B). Moreover, for all such x
we have
1
gr τ (B(n) (x)) = min gr(∧τj B(n) (x)) > ,
1≤j≤k κn
∧τj B(2n) (x)
ατ (B(n) (T −n x), B(n) (x))  min > εn .
1≤j≤k ∧τj B(n) (T −n x) ∧τj B(n) (x)

2. The sequence of partial functions vτ(n) (B) converges μ-a.e, as n → ∞, to a


function vτ(∞) (B) : X → Fτ (Rm ).
3. For all phases x ∈/ Bn (B), we have the following rate of convergence:
κ
dτ vτ(n) (B)(x), vτ(∞) (B)(x) <
n
. (4.33)
εn

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).

We define a finite scale decomposition which will be shown to converge to the (τ


restricted) Oseledets decomposition.
Consider the partial function on X taking values in Dτ (Rm ) and defined by

E·(n) (B)(x) := vτ(n) (B∗ )(x)  vτ(n) (B)(x)⊥

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.

/ Bn (B) and n ≥ n0 we have


Lemma 4.12 For all x ∈

θ vτ(n) (B∗ )(x), vτ(n) (B)(x)⊥ ≥ εn . (4.35)

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.

/ Bn (B) and n ≥ n0 we have


Proposition 4.22 (speed of convergence) For all x ∈

  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

F := vτ(∞) (B∗ )(x) ∈ Fτ (Rm ), F  := vτ(∞) (B)(x)⊥ ∈ Fτ ⊥ (Rm ),


F0 := vτ(n) (B∗ )(x) ∈ Fτ (Rm ), F0 := vτ(n) (B)(x)⊥ ∈ Fτ ⊥ (Rm ).



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  )

Applying (4.33) to B∗ we get:


κ
dτ (F0 , F) = dτ vτ(n) (B∗ )(x), vτ(∞) (B∗ )(x) <
n
, (4.38)
εn
158 4 The Oseledets Filtration and Decomposition

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

By Lemma 4.12 we have



θ (F0 , F0 ) = θ vτ(n) (B∗ )(x), vτ(n) (B)(x)⊥ ≥ εn , (4.40)

and by Lemma 2.37 in Chap. 2 combined with (4.38) and (4.39) we have:

θ (F, F  ) ≥ θ (F0 , F0 ) − dτ (F, F0 ) − dτ ⊥ (F  , F0 )


κn κn
≥ εn − −  εn . (4.41)
εn εn

We conclude by combining (4.37)–(4.41). 

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

Fi := vτ(n) (Bi∗ )(x), Fi := vτ(n) (Bi )(x)⊥ ,

hence E·(n) (Bi )(x) = Fi  Fi .


By Proposition 2.38 in Chap. 2, we have
 
d E·(n) (B1 )(x), E·(n) (B2 )(x) = d(F1  F1 , F2  F2 )

1 1
≤ max , (dτ (F1 , F2 ) + dτ ⊥ (F1 , F2 )). (4.43)
θ (F1 , F1 ) θ (F2 , F2 )
4.3 Abstract Continuity Theorem of the Oseledets Filtration 159

Applying (4.34) to B1∗ , B2∗ we get



dτ (F1 , F2 ) = dτ v(n) (B1∗ )(x), v(n) (B2∗ )(x) < e−C2 n , (4.44)

and applying (4.34) to B1 , B2 we get



dτ ⊥ (F1 , F2 ) = dτ ⊥ v(n) (B1 )(x)⊥ , v(n) (B2 )(x)⊥

= dτ v(n) (B1 )(x), v(n) (B2 )(x) < e−C2 n . (4.45)

By Lemma 4.12 we have, for i = 1, 2:



θ (Fi , Fi ) = θ vτ(n) (Bi∗ )(x), vτ(n) (Bi )(x)⊥ ≥ εn = e−n κ(A)/5 . (4.46)

Combining (4.43)–(4.46) we conclude:


 
d E·(n) (B1 )(x), E·(n) (B2 )(x)  en κ(A)/5 e−C2 n < e−C3 n ,

for an appropriate constant C3 , which proves the proposition. 

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 )),

for any B1 , B2 ∈ Cm in a neighborhood of A, and for some α = α(A) > 0.


Moreover, the map Cm  A → E· (A) ∈ D(X, Rm ) is continuous everywhere.

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

1. L. Arnold, Random Dynamical Systems. Springer Monographs in Mathematics (Springer,


Berlin, 1998). MR 1723992 (2000m:37087)
2. L. Backes, A note on the continuity of Oseledets subspaces for fiber-bunched cocycles (2015),
pp. 1–6
3. J. Bochi, The multiplicative ergodic theorem of Oseledets. http://www.mat.uc.cl/jairo.bochi/
docs/oseledets.pdf (2008) (online lecture notes)
4. C. Bocker-Neto, M. Viana, Continuity of Lyapunov exponents for random 2D matrices (2010),
pp. 1–38 (to appear in Ergodic Theory and Dynamical Systems)
5. I. Ya. Gol’dsheı̆d, G.A. Margulis, Lyapunov exponents of a product of random matrices.
Uspekhi Mat. Nauk 44, 5(269), 13–60 (1989). MR 1040268 (91j:60014)
6. S. Gouëzel, A. Karlsson, Subadditive and multiplicative ergodic theorems (2015), pp. 1–20
7. V.A. Kaı̆manovich, Lyapunov exponents, symmetric spaces and a multiplicative ergodic the-
orem for semisimple Lie groups, Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov.
(LOMI) 164 (1987) no. Differentsialnaya Geom. Gruppy Li i Mekh. IX, 29–46, 196–197. MR
947327 (89m:22006)
8. Y. Katznelson, B. Weiss, A simple proof of some ergodic theorems. Israel J. Math. 42(4),
291–296 (1982). MR 682312 (84i:28020)
9. F. Ledrappier, L.-S. Young, Stability of Lyapunov exponents. Ergodic Theory Dyn. Syst. 11(3),
469–484 (1991). MR 1125884 (92i:58096)
10. R. Mañé, Ergodic theory and differentiable dynamics, in Ergebnisse der Mathematik und ihrer
Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 8 (Springer-Verlag,
Berlin, 1987) (Translated from the Portuguese by Silvio Levy. MR 889254 (88c:58040))
11. G. Ochs, Stability of Oseledets spaces is equivalent to stability of Lyapunov exponents. Dyn.
Stab. Syst. 14(2), 183–201 (1999). MR 1698361 (2000g:37067)
12. 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)
13. K. Petersen, Ergodic Theory, Cambridge Studies in Advanced Mathematics, vol. 2 (Cambridge
University Press, Cambridge, 1983). MR 833286 (87i:28002)
14. M.S. Raghunathan, A proof of Oseledec’s multiplicative ergodic theorem. Israel J. Math. 32(4),
356–362 (1979). MR 571089 (81f:60016)
15. D. Ruelle, Ergodic theory of differentiable dynamical systems. Inst. Hautes Études Sci. Publ.
Math. (50), 27–58 (1979). MR 556581 (81f:58031)
16. D. Ruelle, Characteristic exponents and invariant manifolds in Hilbert space. Ann. Math.
115(2), 243–290 (1982). MR 647807 (83j:58097)
17. M. Viana, Lectures on Lyapunov Exponents (Cambridge University Press, Studies in Advanced
Mathematics, Cambridge, 2014)
18. P. Walters, An introduction to ergodic theory, in Graduate Texts in Mathematics, vol. 79
(Springer, New York, 1982). MR 648108 (84e:28017)
19. P. Walters, A dynamical proof of the multiplicative ergodic theorem. Trans. Am. Math. Soc.
335(1), 245–257 (1993). MR 1073779 (93c:28016)
Chapter 5
Large Deviations for Random Cocycles

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.

5.1 Introduction and Statements

We introduce a class of locally constant cocycles over Markov shifts satisfying a


strong mixing property. We formulate the results on the continuity of their Lyapunov
exponents and Oseledets splitting. To prove them we need base and fiber-LDT esti-
mates which are obtained through a spectral method of S.V. Nagaev, which we briefly
explain here. We finish the section with a short revision of the bibliography on the
subject of large deviations and other limit theorems for random linear cocycles.

5.1.1 Description of the Model

Let Σ be a compact metric space and let F be its Borel σ -field.

Definition 5.1 A Markov kernel is a function K : Σ × F → [0, 1] such that


(1) for every x ∈ Σ, E → K (x, E) is a probability measure on Σ, also denoted by
K (x),
(2) for every E ∈ F, the function x → K (x, E) is F-measurable.

© Atlantis Press and the author(s) 2016 161


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_5
162 5 Large Deviations for Random Cocycles

The iterated Markov kernels are defined recursively, setting


(a) K 1 = K , 
(b) K n+1 (x, E) = Σ K n (y, E) K (x, dy), for all n ≥ 1.
Each power K n is itself a Markov kernel on (Σ, F).
A probability measure μ on (Σ, F) is called K -stationary if for all E ∈ F,

μ(E) = K (x, E) μ(d x).

A set E ∈ F is said to be K -invariant when K (x, E) = 1 for all x ∈ E and


K (x, E) = 0 for all x ∈ X \ E. A K -stationary measure μ is called ergodic when
there is no K -invariant set E ∈ F such that 0 < μ(E) < 1. As usual, ergodic
measures are the extremal points in the convex set of K -stationary measures.

Definition 5.2 A Markov system is a pair (K , μ), where K is a Markov kernel on


(Σ, F) and μ is a K -stationary probability measure.

Let (K , μ) be a Markov system. There is a canonical construction, due to


Kolmogorov, of a probability space (X, F , Pμ ) and a Markov stochastic process
{en : X → Σ}n≥0 with initial distribution μ and transition kernel K , i.e., for all
x ∈ Σ and E ∈ F,
1. Pμ [ e0 ∈ E ] = μ(E),
2. Pμ [ en ∈ E | en−1 = x ] = K (x, E).
We briefly outline this construction. Elements in Σ are called states. Consider the
space X + = Σ N of state sequences x = (xn )n∈N , with xn ∈ Σ for all n ∈ N, and let
F + be the product σ -field F + = FN generated by the F-cylinders, i.e., generated
by sets of the form

C(E 0 , . . . , E m ) := { x ∈ X + : x j ∈ E j , for 0 ≤ j ≤ m },

where E 0 , . . . , E m ∈ F are measurable sets. The (topological) product space X +


is compact and metrizable. The σ -field F + coincides with the Borel σ -field of the
compact space X + .
Definition 5.3 Given any probability measure θ on (Σ, F), the following expression
determines a pre-measure
  
m
P+
θ [C(E 0 , . . . , E m )] := ··· θ (d x0 ) K (x j−1 , d x j )
Em E0 j=1

on the semi-algebra of F-cylinders. By Carathéodory’s extension theorem this pre-


measure extends to a unique probability measure P+ + +
θ on (X , F ).
5.1 Introduction and Statements 163

It follows from this definition that the sequence of random variables en : X + →


Σ, defined by en (x) := xn for x = (xn )n∈N , is a Markov chain with initial distribution
θ and transition kernel K w.r.t. the probability space (X + , F + , P+
θ ). It also follows
that the process {en }n≥0 is stationary w.r.t. (X, F + , P+
θ ) if and only if θ is a K -
stationary measure.
Consider now the space X = Σ Z of bi-infinite state sequences x = (xn )n∈Z ,
with xn ∈ Σ for all n ∈ Z, and let F be the product σ -field F = FZ generated
by the F-cylinders in X . Again the topological product space X is both metrizable
and compact, and the σ -field F is the Borel σ -field on the compact metric space X .
There is a canonical projection π : X → X + , defined by π(xn )n∈Z = (xn )n∈N ,
relating these two spaces.
Markov systems are probabilistic evolutionary models, which can also be studied
in dynamical terms. For that we introduce the shift mappings.

Definition 5.4 The one-sided shift is the map T : X + → X + , T (xn )n≥0 =


(xn+1 )n≥0 , while the two-sided shift is the map T : X → X , T (xn )n∈Z = (xn+1 )n∈Z .

The map T : X + → X + is continuous, and hence F + -measurable. It also


preserves the measure P+ + +
μ , i.e., T∗ Pμ = Pμ . Moreover, the Markov process {en }n≥0
on (X + , F + , P+μ ) is dynamically generated by the observable e0 in the sense that
en = e0 ◦ T n , for all n ≥ 0.
The two-sided-shift T : X → X is a homeomorphism, and hence F -
bimeasurable. The projection π : X → X + semi-conjugates the two shifts. The
two-sided-shift is the natural extension of the one-sided-shift. According to this
construction (see [13]), there is a unique probability measure Pμ on (X, F ) such
that T∗ Pμ = Pμ and π∗ Pμ = P+ +
μ . We will refer to the measures Pμ and Pμ as the
Kolmogorov extensions of the Markov system (K , μ).
The expected value of a random variable ξ : X + → R w.r.t. a probability measure
θ on (Σ, F) is denoted by

Eθ (ξ ) := ξ dP+θ .
X+

When θ = δx is a Dirac measure we will write Ex instead of Eδx .

Definition 5.5 Given a Markov system (K , μ) let Pμ be the Kolmogorov extension


of (K , μ) on X = Σ Z . The dynamical system (X, Pμ , T ) is called a Markov shift.

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

It follows from this definition that,

Proposition 5.1 If the Markov system (K , μ) is strongly mixing then the Markov
shift (X, Pμ , T ) is a mixing dynamical system.

Proof Consider a bounded measurable observable f : X → R depending only


on the coordinates x0 , . . . , x p , and write f (x) = f (x0 , . . . , x p ). Let g(x) =
g(x−q , . . . , x−1 ) be another bounded measurable observable depending only on the
coordinates x−q , . . . , x−1 with q ∈ N. Denote by {en }n∈Z the Markov process on
(X, Pμ ) with common distribution μ and transition kernel K . By the strong mixing
property

  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

converges uniformely (in x0 ) to

  
n+ p−1
··· f (xn , . . . , xn+ p ) μ(d xn ) K (x j , d x j+1 ) = Eμ ( f ).
Σ Σ j=n

Hence

Eμ [( f ◦ T n ) g] = Eμ [g(e−q , . . . , e−1 ) f (en , . . . , en+ p )]


  −1

= ··· g(x−q , . . . , x−1 ) Ex0 [ f (en , . . . , en+ p )] μ(d x−q ) K (x j , d x j+1 )
Σ Σ j=−q

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

for all x ∈ Σ and E ∈ F,

K (x, E) ≥ 1 − ε ⇒ ρ(E) ≥ ε.

Given E ∈ F, define

L ∞ (E) := { f ∈ L ∞ (Σ) : f |Σ\E ≡ 0 },

which is a closed Banach sub-algebra of (L ∞ (Σ),


·
∞ ).

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 .

Proof See [3, Sect. V-5].

Let (K , μ) be a Markov system. We introduce a space of measurable functions


A : Σ × Σ → GL(m, R).

Definition 5.7 The space B∞ m (K ) consists of all functions A : Σ ×Σ → GL(m, R)


such that A and A−1 are both measurable and uniformly bounded. On this space we
consider the metric d∞ (A, B) =
A − B
∞ .

Definition 5.8 The function A ∈ B∞


m (K ) determines a linear cocycle F A : X ×
Rm → X × Rm over the Markov shift (X, Pμ , T ), defined by

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 iterates of FA are the maps FAn : X × Rm → X × Rm ,

FAn (x, v) = (T n x, A(n) (x) v),


166 5 Large Deviations for Random Cocycles

with A(n) : X → GL(m, R) defined for all x = (xn )n∈Z by

A(n) (x) := A(xn−1 , xn ) . . . A(x1 , x2 ) A(x0 , x1 ).

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

A(xn−1 , xn ) V (xn−1 ) = V (xn ) for Pμ -a.e. x = (xn )n∈Z .

Assuming (K , μ) is strongly mixing, the ergodicity of this Markov kernel implies


that the subspaces V (x) have constant dimension μ-a.e., denoted by dim(V ). We
say that this family is proper if 0 < dim(V ) < d.
Next we introduce the concepts of irreducible and totally irreducible cocycle (see
Definition 2.7 in [1]).
Definition 5.10 A cocycle A ∈ B∞ m (K ) is called irreducible w.r.t. (K , μ) if it admits
no measurable proper A-invariant section V : Σ → Gr(Rm ). A cocycle A ∈ B∞ m (K )
is called totally irreducible w.r.t. (K , μ) if the exterior powers ∧k A are irreducible
for all 1 ≤ k ≤ m − 1.
We denote by Im∞ (K ) the subspace of totally irreducible cocycles in B∞
m (K ).

Proposition 5.3 The subspace Im∞ (K ) is open in B∞


m (K ).

Proof A cocycle A ∈ B∞ m (K ) is reducible (i.e. not irreducible) if it admits a mea-


surable proper A-invariant section V : Σ → Gr(Rm ). It is enough to prove that the
set of reducible cocycles is closed.
Let Ak → A be a convergent sequence of reducible cocycles in B∞ m (K ), and let
Vk : Σ → Gr(Rm ) be a measurable proper Ak -invariant section. We will prove that
A is also reducible.
Assume that (Σ, μ) is a complete probability space. Let Ω ⊂ X be a Borel
measurable set with Pμ (Ω) = 1 such that for all k ≥ 1 all x = (xn )n∈Z ∈ Ω and
n ∈ Z, Ak (xn−1 , xn ) Vk (xn−1 ) = Vk (xn ).
Fix any point s0 ∈ Σ. Extracting a subsequence we may assume that Vk (s0 )
converges to V0 ∈ Gr(Rm ) as k tends to ∞. Consider then the set

E := { s ∈ Σ : ∃ x ∈ Ω, n ∈ N such that x0 = s0 and xn = s }.

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

Let τ = (1 ≤ τ1 < · · · < τk < m) be a signature. If A ∈ Cm has a τ -gap pattern,


i.e., L τ j (A) > L τ j+1 (A) for all j, we define the Lyapunov τ -block

Λτ (A) := (Λτ1 (A), . . . , Λτk (A)) ∈ Rk .

A flag of Rm is any increasing sequence of linear subspaces. The corresponding


sequence of dimensions is called its signature. A measurable filtration is a measurable
function on X , taking values in the space of flags of Rm with almost sure constant
signature. We denote by F(X, Rm ) the Note that the Oseledets filtration of A, which
we denote by F(A), is an element of this space.
We denote by F⊃τ (X, Rm ) the subset of measurable filtrations with a signature τ or
finer. If F ∈ F⊃τ (X, Rm ) there is a natural projection F τ with signature τ , obtained
from F by simply ‘forgetting’ some of its components. This space is endowed with
the following pseudo-metric


distτ (F, F ) := dτ (F τ (x), (F  )τ (x)) μ(d x),
X

where dτ refers to the metric on the τ -flag manifold.


On the space F(X, Rm ) we consider the coarsest topology that makes the sets
F⊃τ (X, Rm ) open, and the pseudo-metrics dist τ continuous.
A decomposition of Rm is a sequence of linear subspaces {E j }1≤ j≤k+1 whose
direct sum is Rm . This determines the flag E 1 ⊂ E 1 ⊕ E 2 ⊂ · · · ⊂ E 1 ⊕ · · · ⊕ E k ,
whose signature τ also designates the signature of the decomposition.
168 5 Large Deviations for Random Cocycles

A measurable decomposition is a measurable function on X , taking values in


the space of decompositions of Rm with almost sure constant signature. We denote
by D(X, Rm ) the space of measurable decompositions. Note that the Oseledets
decomposition of A, which we denote by E · (A), is an element of this space.
We denote by D⊃τ (X, Rm ) the subset of measurable decompositions with a sig-
nature τ or finer. If E · ∈ D⊃τ (X, Rm ) there is a natural restriction E ·τ with signature
τ , obtained from E · by simply ‘patching up’ the appropriate components. This space
is endowed with the following pseudo-metric

distτ (E · , E · ) := dτ (E ·τ (x), (E · )τ (x)) μ(d x),
X

where dτ refers to the metric on the manifold of τ -decompositions.


On the space D(X, Rm ) we consider the coarsest topology that makes the sets
D⊃τ (X, 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 irreducible Markov cocycles.

Theorem 5.1 Let (K , μ) be a strongly mixing Markov system and let m ≥ 1.


Then all Lyapunov exponents L j : Im∞ (K ) → R, with 1 ≤ j ≤ m, the
Oseledets filtration F : Im∞ (K ) → F(X, Rm ), and the Oseledets decomposition
E · : Im∞ (K ) → D(X, Rm ), are continuous functions of the cocycle A ∈ Im∞ (K ).
Moreover, if A ∈ Im∞ (K ) has a τ -gap pattern then the functions Λτ , F τ and E ·τ
are Hölder continuous in a neighborhood of A.

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]

 x  ) := 2− inf{ |k| : k∈Z, xk =xk } ,


d(x, (5.1)

for all x = (xk )k∈Z and x  = (xk )k∈Z in X .


 is not compact.
Remark 5.1 Notice that unless Σ is finite (X, d)

Given k ∈ N, α > 0 and f ∈ L ∞ (X ) define


 
vk ( f ) := sup{  f (x) − f (y) : d(x,
 y) ≤ 2−k },
vα ( f ) := sup{ 2αk vk ( f ) : k ∈ N},

f
α :=
f
∞ + vα ( f ),
Hα (X ) := { f ∈ L ∞ (X ) : vα ( f ) < +∞ }.
5.1 Introduction and Statements 169

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,

From now on we denote by 1 the constant function 1.

Proposition 5.4 For all 0 ≤ α ≤ 1, (Hα (X ),


·
α ) is a Banach algebra with unit
element 1, and also a lattice.

Proof To see that (Hα (X ),


·
α ) is a normed algebra with unity, it is enough to verify
the following inequalities:

vk ( f g) ≤
f
∞ vk (g) +
g
∞ vk ( f ),
vα ( f g) ≤
f
∞ vα (g) +
g
∞ vα ( f ).

They imply that


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. 

Definition 5.11 We say that f : X → C is future independent


if f (x) = f (y) for any x, y ∈ X such that xk = yk for all k ≤ 0. Define the space

Hα (X − ) := { f ∈ Hα (X ) : f is future independent } . (5.2)

The space Hα (X − ) is a closed sub-algebra of Hα (X ), and hence a unital Banach


algebra itself.
Denote by F − the sub σ -field of F generated by cylinders in non-positive coor-
dinates. With this terminology, the subspace Hα (X − ) consists of all F − -measurable
functions in Hα (X ).
The base LDT theorem below is proved in Sect. 5.3.1.

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

Moreover, the constants C, k and ε0 depend only on K and


ξ
α , and if K is fixed
then they are uniform in ξ ranging over any bounded set in Hα (X − ).
170 5 Large Deviations for Random Cocycles

The fiber LDT theorem, proved in Sect. 5.3.2, has the following statement.

Theorem 5.3 Given a Markov system (K , μ) and A ∈ B∞


m (K ), assume

(1) (K , μ) is strongly mixing,


(2) A is irreducible,
(3) L 1 (A) > L 2 (A).
Then there exists V neighborhood of A in B∞
m (K ) and there exist C = C(A) < ∞,
k = k(A) > 0 and ε0 = ε0 (A) > 0 such that for all 0 < ε < ε0 , B ∈ V and n ∈ N,

1 
Pμ  log
B (n)
− L 1 (B) > ε ≤ C e−k ε n .
2

5.1.2 The Spectral Method

Consider a Markov system (K , μ) on a compact metric space Σ. Given some


F-measurable observable ξ : Σ → R, let ξ̂ : X + → R be the F + -measurable
function ξ̂ (x) = ξ(x0 ).
Given x ∈ Σ, let P+ x denote the probability on the measurable space (X , F )
+ +
+
that makes {en : X → Σ}n≥0 a Markov process with transition kernel K and initial
distribution with point mass δx (see Definition 5.3). Then {ξ̂ ◦T n }n≥0 is also a Markov
process on (X + , F + , P+x ).

Definition 5.12 We call sum process of the observable ξ : X + → R the following


sequence of random variables {Sn (ξ )}n≥0 on (X + , F + ),


n−1
n−1
Sn (ξ )(x) := ξ̂ ◦ T j (x) = ξ(x j ).
j=0 j=0

Definition 5.13 An observed Markov system on (Σ, F) is a triple (K , μ, ξ ) where


(K , μ) is a Markov system on (Σ, F), and ξ : Σ → R is an F-measurable function.

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

Given a class X of observed Markov systems (K , μ, ξ ) on a given measurable


space (Σ, F), we say that X satisfies uniform LDT estimates of exponential type if
there exist positive constants C, k and ε0 such that for every observed Markov system
(K , μ, ξ ) ∈ X, the observable ξ satisfies LDT estimates of exponential type with the
same constants C, k and ε0 .
Free ebooks ==> www.Ebook777.com
5.1 Introduction and Statements 171

Definition 5.15 Let η : X + → R be a random variable on (X + , F + ). The function


c(η, x, ·) : R → R,
c(η, x, t) := log Ex [et η ]

is called the second characteristic function of η, also known as the cumulant gener-
ating function of η (see [11]).

Proposition 5.5 Let η : X + → R be a F + -measurable random variable. Assume


there exist a, M > 0 such that for all x ∈ Σ, max{Ex [ea η ], Ex [ |η| ea η ]} ≤ M.
Then the cumulant generating function c(η, x, ·) satisfies
(1) c(η, x, t) is well-defined and analytic for t ∈ (−a, a),
(2) c(η, x, 0) = 0,
(3) dc
dt
(η, x, 0) = Ex (η),
(4) c(η, x, t) ≥ t Ex (η), for all t ∈ (−a, a),
(5) the function c(η, x, ·) : (−a, a) → R, t → c(η, x, t) is convex.

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 ).

Finally, (2), (3) and (5) imply (4). 

Given an observable ξ : Σ → R, the function cn (ξ, x, ·) : R → R defined by

cn (ξ, x, t) := log Ex [et Sn (ξ ) ],

is the cumulant generating function of Sn (ξ ). Under general conditions, e.g., if ξ is


bounded, this function is analytic in C, or at least analytic in a neighbourhood of 0.
Let us write Da (0) = { z ∈ C : |z| < a }.

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.

Proposition 5.6 Given an F-measurable observable ξ : Σ → R, let c(ξ, z) be a


limit cumulant generating function of the process {Sn (ξ )}n≥0 on Da (0). Then
(1) z → c(ξ, z) is analytic on Da (0),
(2) c(ξ, 0) = 0,
(3) dc
dt
(ξ, 0) = Eμ (ξ̂ ),
(4) c(ξ, t) ≥ t Eμ (ξ̂ ), for all t ∈ R,
(5) the function c(ξ, ·) : (−a, a) → R, t → c(ξ, t), is convex.

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

Taking the limit identity (3) holds.


Since convexity is a closed property, (5) follows from Proposition 5.5(5).
Finally, (2), (3) and (5) imply (4). 

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 .

Proposition 5.7 Let ξ : Σ → R be an F-measurable observable and let c(ξ, z) be


a limit cumulant generating function of the process {Sn (ξ )}n≥0 on Da (0).
2
Given h > ddt 2c (ξ, 0), there exist C, ε0 > 0 such that for all n ∈ N, x ∈ Σ and
0 < ε < ε0 ,

+ 1
  ε2
Px 
Sn (ξ ) − Eμ (ξ̂ ) > ε ≤ C e−n 2h .
n

In other words, ξ satisfies LDT estimates of exponential type.

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

which is always strictly convex.

Consider now a topological space X of observed Markov systems (K , μ, ξ ), on


a given measurable space (Σ, F).
Denote by H(Da (0)) the Banach space of analytic functions f : Da (0) → C
with a continuous extension up to its boundary. Endow this space with the usual max
norm
f
∞ = max|z|≤a | f (z)|.
174 5 Large Deviations for Random Cocycles

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).

For t = 0, the operator Q 0 : B → B, is a Markov operator. In particular it is


a positive operator which fixes the constant functions, e.g., Q 0 1 = 1, and whose
spectrum is contained in the closed unit disk. The key ingredient to estimate the
moment generating function Ex [et Sn (ξ ) ] via this spectral approach is the assumption
that the operator Q 0 : B → B is quasi-compact and simple. This means that the
eigenvalue 1 of Q 0 is simple and there exists a spectral gap separating this eigenvalue
from the rest of spectrum inside the open unit disk. Under this hypothesis, Q t is a
positive operator, whenever defined, and there exists a unique eigenfunction v(t) ∈ B
5.1 Introduction and Statements 175

such that Q t v(t) = λ(t) v(t), normalized by Eμ [v(t)] = 1, and corresponding to a


positive eigenvalue λ(t) of Q t . Hence, because the functions t → λ(t) and t → v(t)
are continuous in t (in fact analytic), we have
  
Eμ [et Sn (ξ ) ] = (Q nt 1) dμ ≈ Q nt v(t) dμ = λ(t)n v(t) dμ = λ(t)n .

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 ).

5.1.3 Literature Review

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.

5.2 An Abstract Setting

In this section we specialize an abstract setting in [7], from which we derive an


abstract theorem on the existence of uniform LDT estimates for Markov processes.
176 5 Large Deviations for Random Cocycles

5.2.1 The Assumptions

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

Definition 5.17 The operator T is called quasi-compact if there is a T -invariant


decomposition B = F ⊕ H such that dim F < +∞ and the spectral radius of
T |H is (strictly) less than the absolute value |λ| of any eigenvalue λ of T | F . T is
called quasi-compact and simple when furthermore dim F = 1. In this case σ (T | F )
consists of a single simple eigenvalue referred to as the maximal eigenvalue of T .

Consider a Markov system (K , μ) on a compact metric space Σ.

Definition 5.18 The following linear operator is called a Markov operator



(Q f )(x) = (Q K f )(x) := f (y) K (x, dy).
Σ

This operator acts on F-measurable functions on Σ, mapping L p functions to L p


functions, for any 1 ≤ p ≤ ∞. It is easy to verify that for all n ≥ 1 and f ∈ L ∞ (Σ),

(Q nK f )(x) = f (y) K n (x, dy).
Σ

Also, a probability measure μ on (Σ, F) is K -stationary if and only if


 
(Q K f ) dμ = f dμ, for all f ∈ L ∞ (Σ).
Σ Σ

We shall write Q instead of Q K when the kernel K is fixed.


The operator Q : L ∞ (Σ) → L ∞ (Σ) satisfies the following.
Proposition 5.8 For any f ∈ L ∞ (Σ),
(a) Q1 = 1, 
(b) Σ Q f dμ = Σ f dμ,
(c)
Q f
∞ ≤
f
∞ .

In particular if H0 = { f ∈ L ∞ (Σ) : Σ f dμ = 0} then L ∞ (Σ) = R1 ⊕ H0 is a
Q-invariant decomposition and ρ(Q| H0 ) ≤ 1.

Proof Since Σ K (x, dy) = 1, items (a) and (c) follow. Item (b) is a consequence
of μ being a K -stationary probability measure. 
5.2 An Abstract Setting 177

Definition 5.19 The following linear operator is called a Laplace-Markov operator



(Q ξ f )(x) = (Q K ,ξ f )(x) := f (y) eξ(y) K (x, dy).
Σ

It also operates on F-measurable functions on Σ, but the domain of Q ξ depends also


on the observable ξ .
Proposition 5.9 Given a Markov kernel K on (Σ, F) the following are equivalent:
(a) there is a K -stationary measure μ such that (K , μ) is strongly mixing,
(b) Q K : L ∞ (Σ) → L ∞ (Σ) is quasi-compact and simple.
Proof If (K , μ) is strongly mixing, by Definition 5.6 there exist constants C > 0
and 0 < ρ < 1 such that for all f ∈ L ∞ (Σ),


(Q K )n f −  f, μ 1
∞ ≤ C ρ n
f
∞ .

Defining

H0 = { f ∈ L ∞ (Σ) :  f, μ = 0 },

since (Q K )∗ μ = μ, this subspace is Q K -invariant. Thus, we have a Q K -invariant


decomposition L ∞ (Σ) = 1 ⊕ H0 such that
(Q K )n | H0
≤ C ρ n . This implies that
ρ(Q K | H0 ) ≤ ρ < 1.
Conversely, if Q K : L ∞ (Σ) → L ∞ (Σ) is quasi-compact and simple, there exists
a Q K -invariant decomposition L ∞ (Σ) = 1 ⊕ H0 such that ρ(Q K | H0 ) < 1. By
the Hahn-Banach Theorem there is a bounded linear functional Λ : L ∞ (Σ) → R
such that Λ(1) = 1, and Λ( f ) = 0 for all f ∈ H0 . We claim that Λ is a positive
functional, i.e., Λ( f ) ≥ 0 whenever f ≥ 0. Take any function f ∈ L ∞ (Σ) such
that f ≥ 0, and write f = c 1 + h with h ∈ H0 . Since Q K is a positive operator we
have

c 1 = lim (c 1 + (Q K )n h) = lim (Q K )n f ≥ 0,
n→+∞ n→+∞

which implies that c = Λ( f ) ≥ 0. Hence Λ is positive. By the Riez-Markov-



Kakutani Theorem there is a probability measure μ on Σ such that Λ( f ) = Σ f dμ,
for all f ∈ L ∞ (Σ).
Let us prove that μ is K -stationary. Given f ∈ L ∞ (Σ), write f = c 1 + h, with
h ∈ H0 . Hence Q K f = c 1 + Q K h, with Q K h ∈ H0 , and
 
(Q K f ) dμ = Λ(Q K f ) = c = Λ( f ) = f dμ.
Σ Σ

This proves that μ is K -stationary.


Now, because H0 is the kernel of Λ : L ∞ (Σ) → R, we get that for all f ∈

L (Σ), f ∈ H0 ⇔  f, μ = 0. Thus f −  f, μ 1 ∈ H0 , and taking ρ(Q K | H0 ) ≤
178 5 Large Deviations for Random Cocycles

ρ < 1, there is a constant C > 0 such that


(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
∞ .

This proves that (K , μ) is strongly mixing. 

We now discuss a setting, consisting of the assumptions (B1)–(B7) and (A1)–(A4)


below, where an abstract LDT theorem is proved, and from which Theorems 5.2
and 5.3 will be deduced. The context here specializes a more general setting in [7].
Let (X, dist) be a metric space of observed Markov systems (K , μ, ξ ) over the
compact metric space (Σ, d). Besides X, this setting consists of a scale of complex
Banach algebras (Bα ,
·
α ) indexed in α ∈ [0, 1], where each Bα is a space of
bounded Borel measurable functions on Σ. We assume that there exist seminorms
vα : Bα → [0, +∞) such that for all 0 ≤ α ≤ 1,
(B1)
f
α = vα ( f ) +
f
∞ , for all f ∈ Bα ,
(B2) B0 = L ∞ (Σ), and
·
0 is equivalent to 
·

 ∞,
(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

Proposition 5.10 If (Σ, d) has diameter ≤ 1 then the family of spaces

Hα (Σ) := { f ∈ L ∞ (Σ) : vα ( f ) < +∞ }, α ∈ [0, 1]

satisfies (B1)–(B7).

Proof (B1) holds by definition of the Hölder norm


·
α . For (B2) notice that v0 ( f )
measures the oscillation of f , and hence v0 ( f ) ≤ 2
f
∞ . Property (B3) is obvious.
Item (B4) follows from the inequality
5.2 An Abstract Setting 179

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 ).

Item (B7) follows from this convexity. 

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 .

(A3) The operators Q K ,z ξ act continuously on the Banach algebras Bα , uniformly


in (K , μ, ξ ) ∈ X and z small. More precisely, we assume there are constants
b > 0 and M > 0 such that for i = 0, 1, 2, |z| < b and f ∈ Bα ,

Q K ,z ξ ( f ξ i ) ∈ Bα and
Q K ,z ξ ( f ξ i )
α ≤ M
f
α .

(A4) The family of functions X  (K , μ, ξ ) → Q K ,z ξ , indexed in |z| ≤ b, is


Hölder equi-continuous in the sense that there exists 0 < θ ≤ 1 such that for
all |z| ≤ b, f ∈ Bα and (K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 ) ∈ X,


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.

Proposition 5.11 The function Q K ,∗ξ : Db → L (Bα ) is analytic and if f ∈ Bα


then
d
Q K ,z ξ ( f ) = Q K ,z ξ ( f ξ )
dz

for all (K , μ, ξ ) ∈ X, and α1 ≤ α ≤ α0 .

Proof Given b ∈ R, for all z, z 0 ∈ C,



ez b − ez0 b z
z−ζ
− b ez0 b = b2 eζ b dζ.
z − z0 z0 z − z0

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

Hence, by (A3), for all z ∈ Db ,


  
Q z f − Q z0 f z z − ζ 

− Q z0 ( f ξ )
α ≤
Q ζ ( f ξ )
α  2  |dζ |
z − z0 z0
z − z 0 
 
≤ M
f
α z − z 0 ,
5.2 An Abstract Setting 181

which proves that the following limit exists in L (Bα ),

Q z − Q z0
lim = Q z0 (ξ ·).
z→z 0 z − z0

Notice that (A3) also implies the operator Q z0 (ξ ·)( f ) := Q z0 (ξ f ) is in L (Bα ).




The next proposition focuses on the quasi-compactness and simplicity of Q z =


Q K ,z ξ , and it is proved using arguments in [1, 10].

Proposition 5.12 Consider a metric space X of observed Markov systems satisfying


(A1)–(A4) in the range [α1 , α0 ] ⊂ (0, 1] with setting constants C, σ , M, b and θ .
Given ε > 0 there exist C  , M  > 0 and 0 < b0 < b such that the following
statement holds: for all (K , μ, ξ ) ∈ X, z ∈ Db0 and α1 ≤ α ≤ α0 there exist:
a one dimensional subspace E z = E K ,z ξ ⊂ Bα , a hyperplane Hz = HK ,z ξ ⊂ Bα ,
a number λ(z) = λ K ,ξ (z) ∈ C, and a linear map Pz = PK ,z ξ ∈ L (Bα ) such that
(1) Bα = E z ⊕ Hz is a Q z -invariant decomposition,
(2) Pz is a projection onto E z , parallel to Hz ,
(3) Q z ◦ Pz = Pz ◦ Q z = λ(z) Pz ,
(4) Q z f = λ(z) f for all f ∈ E z ,
(5) z → λ(z) is analytic in a neighborhood of Db0 ,
(6) λ(z) ≥ 1 − ε.

Furthermore, for all f ∈ Bα ,


(7)
Q nz f − λ(z)n Pz f
α ≤ C  (σ + ε)n
f
α ,
(8)
Pz f
α ≤ C 
f
α ,  
(9)
Pz f − P0 f
α ≤ C  z 
f
α ,
and for all z ∈ Db0 and (K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 ) ∈ X,
 
(10) λ K 1 ,ξ1 (z) − λ K 2 ,ξ2 (z) ≤ M  d((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 .
θ

Given (K , μ, ξ ) ∈ X, define the operators



1
Pz = PK ,zξ := Rz (w) dw (5.3)
2πi Γ1

1
L z = L K ,zξ := w Rz (w) dw (5.4)
2πi Γ1

1
Nz = N K ,zξ := w Rz (w) dw (5.5)
2πi Γ0
182 5 Large Deviations for Random Cocycles

where Γ0 and Γ1 are the positively oriented circles

  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 .

Lemma 5.1 Given a Banach space (B,


·
) and linear operators T, T0 ∈ L (B),
if T0 is invertible with
T0−1
≤ C and
T − T0
≤ ε < C −1 then
C
1. T is invertible, with
T −1
≤ ,
1−Cε
C2
2.
T −1 − T0−1

T − T0
.
1−Cε

Proof Since T −1 = ∞ −1 n −1
n=0 (−1) (T0 (T − T0 )) T0 , we have
n



T0−1
C

T −1

T0−1
n+1
T − T0
n = ≤ .
n=0
1−
T0−1

T − T0
1−Cε

For (2) use the formula T −1 − T0−1 = −T −1 (T − T0 ) T0−1 . 


Lemma 5.2 There exist constants C0 > 0 and 0 < b0 < b, depending only on
C, M, σ and b, such that for (K , μ, ξ ) ∈ X, z ∈ Db0 , and for any of the five
operators Tz = Q K ,zξ , L z , Nz , Pz , and Rz (w) with w ∈
/ int(Γ0 ) ∪ int(Γ1 ),
1.
Tz
≤ C0 ,  
2.
Tz − T0
≤ C0 z .
Proof By the spectral decomposition theorem (see [15, Chap. XI]) applied to the
operator Q 0 , L 0 = P0 is the projection P0 f =  f, μ 1 and N0 = Q 0 − P0 . Notice
also that Q n0 = P0 + N0n for all n ≥ 1. Hence
L 0
=
P0
= 1,
N0
≤ C σ , and

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

Therefore, applying Lemma 5.1 to w I − Q z and w I − Q 0 , item (1) holds with


C2 C
C2 := 1−CC1 1C0 b0 , while (2) holds with C3 := 1−C11 C00 b0 . Of course we have to pick
0 < b0 < b small enough to make sure the denominators in constants C2 and C3 are
both positive. For the remaining operators Pz , L z and Nz we use the integral formulas
(5.3)–(5.5) to reduce to the previous case, using the same constants C2 and C3 as
before. 

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 

Notice that by Lemma 5.2, for all (K , μ, ξ ) ∈ X,

PK ,z ξ 1, μ0  ≥ 1 −
PK ,z ξ 1 − PK ,0 1
α ≥ 1 − C0 b0 .

Hence, for all z ∈ Db0 ,

   
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

It follows that for all n ∈ N,


Nzn
≤ C0 (σ + ε)n . This proves (7) with C  = C0 .
p p

Items (8) and (9) follow from Lemma 5.2.


To prove item (10), we claim that for all (K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 ) ∈ X, z ∈ Db0 ,
2α1 ≤ α ≤ α0 , and f ∈ Bα ,
θ
v α2 (Q K 1 ,zξ1 f − Q K 2 ,zξ2 f ) 
f
α dist ((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 . (5.7)

In fact by (B7), (B2) and (A4), we have


1 1
v α2 (Q K 1 ,zξ1 f − Q K 2 ,zξ2 f ) ≤ v0 (Q K 1 ,zξ1 f − Q K 2 ,zξ2 f ) 2 vα (Q K 1 ,zξ1 f − Q K 2 ,zξ2 f ) 2
1 1

Q K 1 ,zξ1 f − Q K 2 ,zξ2 f

2
vα (Q K 1 ,zξ1 f − Q K 2 ,zξ2 f ) 2
θ

f
α dist ((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 .
5.2 An Abstract Setting 185

Equation (5.7) implies, for all (K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 ), z, α and f as above, and


/ int(Γ0 ) ∪ int(Γ1 ),
all w ∈
θ
v α2 (R K 1 ,zξ1 (w) f − R K 2 ,zξ2 (w) f ) 
f
α dist ((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 .
(5.8)

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).

Thus, integrating (5.3) and (5.4), we obtain


θ

PK 1 ,zξ1 f − PK 2 ,zξ2 f
α2 
f
α dist ((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 ,
θ

L K 1 ,zξ1 f − L K 2 ,zξ2 f
α2 
f
α dist ((K 1 , μ1 , ξ1 ), (K 2 , μ2 , ξ2 )) 2 .

Finally, (10) follows from the previous inequalities and (5.6). 

Remark 5.3 The condition α1 < α20 and the assumption (A4) are only needed to
prove item (10) of Proposition 5.12.

5.2.2 An Abstract Theorem

In this subsection we state and prove an abstract LDT theorem.


Let (Bα ,
·
)α∈[0,1] be a scale of Banach algebras satisfying (B1)–(B7). Assume
X is a metric space of observed Markov systems for which assumptions (A1)–(A4)
hold. Take 0 < b0 < b according to Proposition 5.12.
Given (K , μ, ξ ) ∈ X, let

c K ,ξ (z) := log λ K ,ξ (z), (5.9)

where λ K ,ξ (z) denotes the maximal eigenvalue of Q K ,tξ .

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

Lemma 5.3 For all (K , μ, ξ ) ∈ X, n ∈ N, z ∈ Db0 and x ∈ Σ,



 
((Q K ,zξ )n 1)(x) = Ex e z Sn (ξ ) = e z Sn (ξ ) dP+
x.
X+

In particular, for all z ∈ Db0 ,


 
Eμ ((Q K ,zξ )n 1) = Eμ e z Sn (ξ ) .

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

We obtain the second identity averaging this relation in x0 w.r.t. μ. 

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.

Proposition 5.13 There exist C1 > 0 and a sequence δn converging geometrically


to 0 such that for all (K , μ, ξ ) ∈ X, z ∈ Db0 (0), x ∈ Σ and n ∈ N
  
n log λ K ,ξ (z) − log Ex e z Sn (ξ )  ≤ C1 |z| + δn .

Proof We will use the notations in Proposition 5.12, choosing


 ε>  0 small enough
so that σ + ε < 1 − ε. By Lemma 5.3, (Q nz 1)(x) = Ex e z Sn (ξ ) . By Lemma 5.2
there exists B > 0 such that for all z ∈ Db0 (0),
Pz − I
α ≤ B |z|. Hence
  z S (ξ )    
Ex e n − λ K ,ξ (z)n  ≤ (Q nz 1)(x) − λ K ,ξ (z)n 

Q nz 1 − λ K ,ξ (z)n Pz 1
α + λ K ,ξ (z)n
1 − Pz 1
α
=
Nzn 1
α + λ K ,ξ (z)n
1 − Pz 1
α
≤ C (σ + ε)n + B |z| λ K ,ξ (z)n .

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

B |z| λ K ,ξ (z)n + C (σ + ε)n



(1 − B |z|) λ K ,ξ (z)n − C (σ + ε)n
B |z| + δn
≤ ≤ 2 (B |z| + δn ),
1 − B |z| − δn
(σ +ε)n
 σ +ε n
where δn := C λ K ,ξ (z)n
≤C 1−ε
converges geometrically to zero. 

Proof (of Theorem 5.4) Combine Proposition 5.13 with Corollary 5.1. 

5.3 The Proof of LDT Estimates

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.

5.3.1 Base LDT Estimates

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 − .

Proposition 5.14 The family of spaces Hα (X − ) forms a scale of Banach algebras


satisfying (B1)–(B7).

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.

Proposition 5.15 The space X of observed Markov systems satisfies (A1)–(A4).


 determines the Markov operator Q K : L ∞ (X − ) → L ∞ (X − ),
The kernel K

(Q K f )( . . . , x−1 , x0 ) := f ( . . . , x−1 , x0 , x1 ) K (x0 , d x1 ).
Σ

This operator acts continuously on Ha (X − ). Its iterates are given by


 
n−1
(Q nK f )( . . . , x−1 , x0 ) = f ( . . . , x0 , x1 , . . . , xn ) K (x j , d x j+1 ). (5.11)
Σn j=0

Proposition 5.16 For all f ∈ Hα (X − ) and n ∈ N,


(1)
(Q K)n f
∞ ≤
f
∞ ,
(2) vα ((Q K)n f ) ≤ max{2
(Q K)n f
∞ , 2−n α vα ( f )}.

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

 x  ) ≤ 2−k , the inequality vk (Q n f ) ≤


and taking the sup in x, x  ∈ X − such that d(x,
vk+n ( f ) follows. Hence, for k ≥ 1,

2αk vk (Q n f ) = 2−n α (2α(k+n) vk+n ( f )) ≤ 2−n α vα ( f ) .

For k = 0 notice that v0 (Q n f ) is the variation of Q n f . Thus v0 (Q n f ) ≤ 2


Q n f
∞ .
Taking the sup in k ∈ N, item (2) follows. 

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

Proof Given a function f ∈ Hα (X − ), denote by f k : X − → C the following


function

f k ( . . . , x0 ) := f ( . . . , x−k , . . . , x0 ) dP−
μ ( . . . , x −k ).
X−

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)

Because (K , μ) is strongly mixing, there


 are constants C > 0 and 0 < ρ < 1 such
that for any function h ∈ L ∞ (Σ) with Σ h dμ = 0,

 
 h(y)K n (x, dy) ≤ C ρ n
h
∞ .
Σ

Now, if h ∈ L ∞ (X − ) is a function which depends only on the first coordinate x0


then by (5.11) Q n h also depends only on the first coordinate, and it is given by

(Q h)( . . . , x0 ) =
n
h(y)K n (x0 , dy).
Σ

Hence, if h has zero average, i.e., E−


μ (h) = 0, then


Q n h
∞ ≤ C ρ n
h
∞ . (5.13)

We claim that h = Q k ( f k − E− μ ( f ) 1) is a function with zero average that depends


only on the first coordinate. The first part of claim follows because Q preserves
averages (see Proposition 5.8) and, as remarked above, E− −
μ ( f k ) = Eμ ( f ). For
the second part notice two things: first Q preserves functions that depend only
on the first coordinate x0 ; second, Q maps a function f that depends only on
the coordinates x−k , . . . , x−1 , x0 to a function that depends only on the coordi-
nates x−k+1 , . . . , x−1 , x0 , in other words Q f looses dependence in x−k . Therefore,
from (5.13)
190 5 Large Deviations for Random Cocycles


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
∞ .

On the other hand, by item (2) of Proposition 5.16,

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
∞ .

Thus, for all f ∈ Hα (X − ),


Q n f − E−
μ ( f ) 1
α ≤ 4Cσ
f
α ,
n

which proves the proposition. 

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 ,

which implies (A4). 


Proof (of Theorem 5.2) By Theorem 5.4 every ξ ∈ X admits a neighborhood satisfy-
ing uniform LDT estimates of exponential type. However, as stated in Theorem 5.2,
we want to see that these LDT estimates hold uniformly for all observed Markov
system in X.
For δ > 0 small, denote by Bδ (0) the δ-ball around the origin in Hα (X − ), and
consider the analytic function λ̂ K : Bδ (0) → C, ξ → λ̂ K (ξ ) = maximal eigenvalue
Free ebooks ==> www.Ebook777.com
5.3 The Proof of LDT Estimates 191

of Q K ,ξ . Decreasing δ we can assume that λ̂ is bounded. Choose b0 > 0 small, such


that b0 L < δ, and notice that λ K ,ξ (z) = λ̂(z ξ ) for all z ∈ Db0 (0). Hence the family
of analytic functions {λ K ,ξ (z)}ξ ∈X is uniformly bounded over Db0 (0). Shrinking b0
even more, the derivatives λK ,ξ (z) and λK ,ξ (z) are also bounded. Thus, there exists
h > 0 such that (c K ,ξ ) (t) < h for all t ∈ [−b0 , b0 ] and ξ ∈ X. By Proposition 5.13,
conclusion (2) of Corollary 5.1 and Remark 5.4, there are constants ε0 and C > 0
such that for all ξ ∈ X, 0 < ε < ε0 and all n ∈ N,

1  2
P−  Sn (ξ ) − E− (ξ ) ≥ ε ≤ C e− 2ε h n .
μ μ
n

Consider the natural (measure preserving) projection π : X → X − . Since



1 
π −1  Sn (ξ ) − E− (ξ ) ≥ ε
μ
n
⎧ ⎫
⎨  ⎬
1 n−1

= x∈X :  ξ(T j (x)) − ξ dPμ  ≥ ε ,
⎩ n j=0 X ⎭

all observables ξ ∈ X satisfy a uniform base-LDT estimate.


The constants δ, b0 , h, ε0 and C depend on L and λ̂ K , i.e., on L and K . 

5.3.2 Fiber LDT Estimates

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

ξ A (x, y, p̂) := log


A(x, y) p
, (5.16)

for any unit vector p ∈ p̂.

www.Ebook777.com
192 5 Large Deviations for Random Cocycles

We can now introduce the metric space of observed Markov systems

X := { (K A , μ A , ±ξ A ) : A ∈ B∞
m (K ), A irreducible, L 1 (A) > L 2 (A) }.

This space is identified with a subspace of B∞


m (K ), and endowed with the distance

dist ((K A , μ A , ξ A ), (K B , μ B , ξ B )) := d∞ (A, B).

Proposition 5.18 The space X of observed Markov systems satisfies (A1)–(A4).

Next we define the scale of Banach algebras. Recall the following projective
distance (see (2.3))

p ∧ q

δ( p̂, q̂) := ,

p

where p ∈ p̂ and q ∈ q̂. Given 0 ≤ α ≤ 1 and f ∈ L ∞ (Σ × Σ × P(Rm )), let


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).

Proposition 5.19 The family of spaces Hα (Σ × Σ × P(Rm )) is a scale of Banach


algebras satisfying (B1)–(B7).

Proof (B1) holds by definition of the Hölder norm


·
α . For (B2) notice that
v0 ( f ) measures the maximum oscillation of f on the projective fibers, and hence
v0 ( f ) ≤ 2
f
∞ . Property (B3) is obvious. Assumption (B4) is a consequence of
the inequality

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. 

Definition 5.20 We define Hα (Σ × P(Rm )) to be the subspace of functions


f (x, y, p) in Hα (Σ × Σ × P(Rm )) that do not depend on the first coordinate x.

This subspace is clearly a closed sub-algebra of Hα (Σ × Σ × P(Rm )).


5.3 The Proof of LDT Estimates 193

Proposition 5.20 The family Hα (Σ × P(Rm )) is a scale of Banach sub-algebras


satisfying (B1)–(B7).

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

This is the Markov operator associated with the kernel (5.15).


Assumption (A1) follows from the definition of X.
Since (Q A f )(x, y, p) does not depend on the coordinate x, the Markov operator
Q A leaves invariant the subspace of functions f (x, y, p) that are constant in x. Next,
we are going to see that Q A acts invariantly on the subspace Hα (Σ × P(Rm )).
Given A ∈ B∞ m (K ) and 0 < α ≤ 1, define for all n ∈ N,
! α "
δ(A(n) p, A(n) q)
καn (A) := sup Ex ∈ [0, +∞] (5.20)
x∈Σ, p=q δ( p, q)

Lemma 5.4 Let A ∈ B∞


m (K ) and n ∈ N.

(a)
A(±n)
∞ ≤ max{
A
∞ ,
A−1
∞ }n .
(b)
A(n) − B (n)
∞ ≤ n max{
A
∞ ,
B
∞ }n−1
A − B
∞ .

Proof Item (a) is straightforward. To prove (b), we use the formula


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

The following lemma highlights the importance of the quantity (5.20).

Lemma 5.5 Given A ∈ B∞


m (K ), f ∈ Hα (Σ × P(R )) and n ∈ N,
m

vα (Q nA f ) ≤ καn (A) vα ( f ).

Proof For any f ∈ Hα (Σ × P(Rm )), and (x0 , p) ∈ Σ × P(Rm ),


 
n−1
(Q nA f )(x0 , p) = f (xn , A(xn−1 , xn ) . . . A(x0 , x1 ) p) K (x j , d x j+1 )
Σn j=0
 
= Ex0 f (en , A(n) p) .
194 5 Large Deviations for Random Cocycles

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). 

Lemma 5.6 The sequence {καn (A)}n≥0 is sub-multiplicative, i.e.,

καn+ (A) ≤ καn (A) κα (A) for n,  ∈ N.

In particular,
lim κ n (A)1/n = inf{ καn (A)1/n : n ∈ N }.
n→+∞ α

Proof Let us write Mn = A(n) . Given x ∈ Σ and p = q in P(Rm ),


 
δ(Mn+m p, Mn+m q) α
Ex
δ( p, q)
   
δ((Mn ◦ T m )Mm p, (Mn ◦ T m )Mm q) α δ(Mm p, Mm q) α
≤ Ex
δ(Mm p, Mm q) δ( p, q)
 α  
δ(Mm p, Mm q) δ(Mn p , Mn q  ) α

≤ Ex sup E K (x,·)
m ≤ καm καn ,
δ( p, q) p =q  δ( p  , q  )

and taking the sup we get καn+m ≤ καn καm . 

These constants become finite provided α is small enough.

Lemma 5.7 Given A ∈ B∞


m (K ) and n ∈ N for all 0 < α ≤
1
4n
,

καn (A) ≤ max{


A
∞ ,
A−1
∞ }.

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
∞ }.

Hence, taking the sup in x and p = q we obtain καn ≤ max{


A
∞ ,
A−1
∞ }. 

By the previous lemmas the operator Q A leaves the subspace Hα (Σ × P(Rm ))


invariant, for all small enough α > 0. To prove that Q A is quasi-compact and simple
all hypotheses of Theorem 5.3 are essential. The irreducibility and gap assumptions
are used in the following lemmas.

Lemma 5.8 Given A ∈ B∞


m (K ) such that (K A , μ A , ξ A ) ∈ X

1
lim Ex (log
A(n) p
) = L 1 (A),
n→+∞ n

with uniform convergence in (x, p) ∈ Σ × P(Rm ).

Proof See Lemma 3.1 in [1]. 

Lemma 5.9 Given A ∈ B∞ m (K ) such that (K A , μ A , ξ A ) ∈ X, there exists n ∈ N


such that for all x ∈ Σ and p = q in P(Rm ),

δ(A(n) p, A(n) q)
Ex log ≤ −1.
δ( p, q)

Proof We write Mn = A(n) . Given x ∈ Σ and p = q in P(Rm ),



1 δ(Mn p, Mn q) 1
(Mn p) ∧ (Mn q)

p

q

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. 

Proposition 5.21 Given A ∈ B∞ m (K ) such that (K A , μ A , ξ A ) ∈ X, there exists a


α0
neighborhood V of A in B∞
m (K ), and there are constants 0 < α1 < 2 < α0 , C > 0
196 5 Large Deviations for Random Cocycles

and 0 < σ < 1 such that

vα (Q nB f ) ≤ C σ n vα ( f ),

for all B ∈ V, α ∈ [α1 , α0 ], n ∈ N and f ∈ Hα (Σ × P(Rm )).

Proof We begin deriving a modulus of continuity for B → καn (B).


Fix a neighborhood V of A in B∞ m (K ) such that for all B ∈ V,
B
∞ ≤ C and

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. 

Next proposition implies (A2).

Proposition 5.22 Given A ∈ B∞ m (K ) such that (K A , μ A , ξ A ) ∈ X, there exist a


neighborhood V of A in B∞
m (K ), a range 0 < α1 < α20 < α0 ≤ 1 and there are
constants C > 0 and 0 < σ < 1 such that for all B ∈ V, α ∈ [α1 , α0 ] and
f ∈ Hα (Σ × P(Rm )),


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 .

Hence it is now enough to prove that


Q nB f −  f, ν B  1
∞ ≤ C σ n
f
α .

We define four families of transformations


(i)
TB,n,m : L ∞ (Σ × P(Rm )) → L ∞ (Σ × P(Rm )) i = 0, 1, 2, 3,

depending on B ∈ V, and n ≥ m, n, m ∈ N, which act continuously on the scale of


Banach spaces Hα (Σ × P(Rm )) with 0 < α ≤ α0 .
(0)  
(TB,n f )(x, p) := (Q nB f )(x, p) = Ex f (en , B (n) p) .
(1)  
(TB,n,m f )(x, p) := Ex f (en , (B (m) ◦ T n−m ) p) .
(2)  
(TB,m f )(x, p) := Eμ f (em , B (m) p) .
(2)
TB,m maps Hα (Σ × P(Rm )) onto the space Hα (P(Rm )) of α-Hölder continuous
(2)
functions, constant in x. In particular TB,m : Hα (Σ × P(Rm )) → C(P(Rm )) is a
compact transformation.
198 5 Large Deviations for Random Cocycles

(TB(3) f )(x, p) := f dν B , where ν B is any K B -stationary measure.
TB(3) maps L ∞ (Σ × P(Rm )) onto the space of constant functions. In particular the
linear transformation TB(3) : Hα (Σ × P(Rm )) → C(P(Rm )) has rank 1.
We claim that for all B ∈ V and all f ∈ Hα (Σ × P(Rm )) with 0 < α ≤ α0 , for
all n, m ∈ N with n ≥ m, and all (x, q) ∈ Σ × P(Rm ),
 (0) 
(1) (TB,n (1)
f )(x, q) − (TB,n,m f )(x, q) ≤ C σ m
f
α .
 (1) 
(2) (TB,n,m f )(x, q) − (TB,m (2)
f )(q) ≤ C σ n−m
f
α .
 (2) 
(3) (TB,m (2)
f )(q) − (TB,n f )(q) ≤ C σ m
f
α .
Let us finish the proof before proving these three claims. Setting n = 2m in (1)
and (2), and n =  in (3), for all B ∈ V and f ∈ Hα (Σ × P(Rm )) with 0 < α ≤ α0 ,
(2)
B f − TB, f
∞ ≤ 3 C σ
f
α .

Q 2m m
(5.21)

(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

On the other hand, since vα (Q 2m B f) ≤ Cσ


2m

f
α , we get vα (g) = 0, which

implies that g is constant. But Q B f, ν B  =  f, (Q 2m
2m
B ) ν B  =  f, ν B  implies that
g, ν B  =  f, ν B . Therefore g =  f, ν B  1, and also

B f −  f, ν B  1
∞ ≤ 3 C σ
f
α

Q 2m m
∀ m ∈ N.

This establishes the claim of the proposition.


To finish, we still have to prove the three claims:
Claim (1): Let Fn denote the sub σ -field generated by the random variables {e j : j ≥
n}. Note that for any random variable f : X → C
% &
Ex ( f ) = Ex Een ( f |Fn ) .

Then, using this fact we have,


 (0)    
(T f − T (1) f )(x, q) = Ex f (en , B (n) q) − f (en , (B (m) ◦ T n−m ) q) 
B,n B,n,m
 
≤ Ex  f (en , (B (m) ◦ T n−m )B (n−m) q) − f (en , (B (m) ◦ T n−m ) q)
#  α $

f
α Ex δ (B (m) ◦ T n−m )B (n−m) q, (B (m) ◦ T n−m ) q
' #  α $(
=
f
α Ex Een−m δ (B (m) ◦ T n−m )B (n−m) q, (B (m) ◦ T n−m ) q |Fn−m
5.3 The Proof of LDT Estimates 199
#  α $

f
α sup Ex δ B (m) p, B (m) q
x, p,q
! α "
δ(B (m) p, B (m) q)

f
α sup Ex =
f
α καm (B) ≤
f
α C σ m .
x, p=q δ( p, q)

 
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
.

Claim (3): Because μ is K -stationary,


 (2)      
(T (2)   (m)
q) − Eμ f (en , B (n) q) 
B,m f )(q) − (TB,n f )(q) = Eμ f (em , B
    
= Eμ f (em , B (m) q) − Eμ f (em , B (m) B (n−m) q) 
 
≤ Eμ  f (em , B (m) q) − f (em , B (m) B (n−m) q)
 

f
α Eμ δ(B (m) q, B (m) B (n−m) q)α
%  &

f
α Eμ Een−m δ(B (m) q, B (m) B (n−m) q)α
 

f
α sup Ex δ(B (m) q, B (m) p)α
x, p,q
! α "
δ(B (m) q, B (m) p)

f
α sup Ex =
f
α καm (B) ≤
f
α C σ m .
x, p=q δ(q, p)


Corollary 5.2 Given A ∈ B∞m (K ) such that (K A , μ A , ξ A ) ∈ X, the kernel K A on


the product space Σ × Σ × P(Rm ) has a unique stationary measure.

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 . 

The Laplace-Markov operator Q A,z of the observed Markov system (K A , μ A , ξ A )


is given by

(Q A,z f )(x, y, p) = f (y, z, A(y, z) p)
A(y, z)
z K (y, dz).
Σ
200 5 Large Deviations for Random Cocycles

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). 

Lemma 5.10 Given A, B ∈ B∞ m (K ) and b > 0, there is a constant C 2 > 0 such


that for all f ∈ Hα (Σ × P(Rm )), and all z ∈ C such that Re z ≤ b,


Q A,z f − Q B,z f
∞ ≤ C2 d∞ (A, B)α
f
α .

Moreover, C2 is bounded on a neighborhood of A.

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


Proof (of Theorem 5.3) By Theorem 5.4, there exists V ⊂ X neighborhood of


(K A , μ A , ξ A ), which we identify with a neighborhood of A ∈ B∞
m (K ), and there are
constants ε0 , C, h > 0 such that for all B ∈ V, 0 < ε < ε0 , (x, p) ∈ Σ × P(Rm )
and n ∈ N,

1  ε2
Px  
log
B p
− L 1 (B, μ) ≥ ε ≤ C e− 2 h n .
(n)
n

Integrating w.r.t. μ we get for all p ∈ P(Rm ),



1  ε2
Pμ  log
B (n) p
− L 1 (B, μ) ≥ ε ≤ C e− 2 h n .
n
Free ebooks ==> www.Ebook777.com
5.3 The Proof of LDT Estimates 201

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

Thus a simple comparison of the deviation sets gives



1  ε2
Pμ  log
B (n)
− L 1 (B, μ) ≥ ε  e− 2 h n
n

for all B ∈ V, 0 < ε < ε0 and n ∈ N. 

5.4 Deriving Continuity of the Lyapunov Exponents

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.

5.4.1 Proof of the Continuity

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

Given an observable ξ ∈ Ξ there exists p ∈ N such that ξ ◦ T p depends only


on negative coordinates, i.e., coordinates x j with − p ≤ j ≤ 0. This implies that
ξ ◦ T p ∈ Hα (X − ). By Theorem 5.2, the observable ξ ◦ T p satisfies a base-LDT
estimate w.r.t. P. Since  Sn (ξ ) − Sn (ξ ◦ T p ) converges uniformly to zero as n → ∞,
it follows that ξ satisfies base-LDT estimates too.
Recall that a cocycle A is said to be uniformly L p -bounded if there is a constant
C = C(A) < ∞ such that
)1 )
) )
) log
B (n)
) p < C
n L

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. 

5.4.2 Some Generalizations

Consider a compact metric space Σ.


A Markov kernel of order p ∈ N on Σ is a map K : Σ p → Prob(Σ) that assigns
a probability measure K (x0 , . . . , x p−1 , dy) on Σ to each tuple (x0 , . . . , x p−1 ) ∈ Σ p .
The concept of Markov kernel in Definition 5.1 corresponds to a Markov kernel of
order p = 1.
5.4 Deriving Continuity of the Lyapunov Exponents 203

Any Markov kernel K of order p on Σ determines the following Markov kernel


K̂ of order 1 on the product space Σ p ,

K̂ (x0 , . . . , x p−1 ) := δ(x1 ,...,x p ) K (x0 , . . . , x p−1 , d x p ).
Σ

A probability measure μ on Σ p is said to be K -stationary when it is K̂ -stationary.


We call a Markov system of order p any pair (K , μ), where K is a Markov kernel
of order p on Σ, and μ is a K -stationary probability on Σ p . We say that (K , μ) is
strongly irreducible when ( K̂ , μ) is a strongly irreducible Markov system on Σ p .
Given a Markov system (K , μ) of order p, let P̂μ denote the Kolmogorov exten-
sion of (K , μ) on the space of sequences X̂ := (Σ p )Z. Then, letting T̂ : X̂ → X̂
denote the shift homeomorphism, the triple X̂ , P̂μ , T̂ is a Markov shift.
Let X := Σ Z and consider the maps
ψ : X → X̂ , ψ{xn }n∈Z = {(xn , . . . , xn+ p−1 )}n∈Z ,
π : X̂ → X , π {(x0,n , . . . , x p−1,n )}n∈Z = {x0,n }n∈Z ,
which satisfy π ◦ ψ = id X .
Defining P μ := π ∗ P̂μ , these maps are bimeasurable isomorphisms conjugating
the shifts on X̂ , P̂μ and (X, Pμ ), where the measure Pμ is invariant under the shift
T : X → X . The triple (X, Pμ , T ) is called a Markov shift of order p.
Consider now the space B∞ m (K , p) of measurable functions A : X → GL(m, R)
which depend only on the coordinates (x0 , . . . , x p ) ∈ Σ p+1 with
A
∞ < ∞ and

A−1
∞ < ∞. Note that the iterates of A are

A(n) (x) = A(xn−1 , . . . , xn−1+ p ) . . . A(x1 , . . . , x1+ p ) A(x0 , . . . , x p ).

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 ).

Identifying  with a function  : X̂ → GL(m, R) we have  ◦ ψ = A. Hence the


cocycles (T̂ , Â) and (T, A) are conjugated.
The cocycle (T, A) over the Markov shift (X, Pμ , T ) will be called a random
Markov cocycle of order p.
Define Im∞ (K , p) to be the subspace of totally irreducible cocycles A ∈

Bm (K , p), i.e., the subspace of cocycles A such that  is totally irreducible over
( X̂ , P̂μ , T̂ ).
From these considerations and Theorem 5.1 we obtain the following result.
204 5 Large Deviations for Random Cocycles

Theorem 5.5 Let (K , μ) be a strongly mixing Markov system of order p ∈ N.


Then all Lyapunov exponents L j : Im∞ (K , p) → R, with 1 ≤ j ≤ m, the Ose-
ledets filtration F : Im∞ (K , p) → F(X, Rm ), and the Oseledets decomposition E · :
Im∞ (K , p) → D(X, Rm ), are continuous functions of the cocycle A ∈ Im∞ (K , p).
Moreover, if A ∈ Im∞ (K , p) has a τ -gap pattern then the functions Λτ , F τ and
τ
E · are Hölder continuous in a neighborhood of A.

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

Equip this space with the distance

da (A, B) :=
A − B
∞ +
ηaA − ηaB
∞ .

The collection C = {(Bam (K ), da )}m∈N is not a space of measurable cocycles,


because item 3 of Definition 3.1 fails. However, both the uniform fiber-LDT esti-
mates and the continuity statements about the LE can be extended to the spaces
Ima (K ) of totally irreducible cocycles in Bam (K ). More precisely, it can be proved
that Theorem 5.3 holds for all a ≥ 4, and Theorem 5.1 holds for all a ≥ 4 m.

5.4.3 Method Limitations

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.

6.1 Introduction and Statements

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.

6.1.1 Description of the Model

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.

© Atlantis Press and the author(s) 2016 207


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_6
208 6 Large Deviations for Quasi-Periodic Cocycles

Throughout this chapter, a  b will stand for a ≤ Cb for some context-universal


constant C, which may be discarded from subsequent estimates.
Let Td = (R/Z)d be the torus with d ≥ 1 variables. We denote by |·| the Haar
measure on Td . Let T x = x + ω be the translation on the torus Td by the vector
ω = (ω1 , ω2 , . . . , ωd ). We assume that 1, ω1 , ω2 , . . . , ωd are rationally independent,
hence T is ergodic. The map T defines the base dynamics and it is assumed fixed.
Let A : Td → Mat(m, R) be a matrix valued real analytic function.
The pair (T, A), acting on the vector bundle Td × Rm by (x, v) → (T x, A(x)v)
is called an analytic, quasi-periodic linear cocycle (quasi-periodic because of the
base dynamics, linear due to the linear fiber action, and analytic due to the analytic
dependence on the base point). As before, the frequency vector ω will be fixed, hence
we identify the cocycle with its fiber action given by the function A(x).
In order to be able to treat occurrences of small denominators, the translation
vector will be assumed to satisfy a generic Diophantine condition:

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

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 -a.e. x ∈ Td .
n→∞ n

Let τ = (1 ≤ τ1 < · · · < τk < m) be a signature. If A ∈ Cm has a τ -gap pattern,


i.e. L τ j (A) > L τ j+1 (A) for all j, we define the Lyapunov τ -block

Λτ (A) := (Λτ1 (A), . . . , Λτk (A)) ∈ Rk .

A flag of Rm is any increasing sequence of linear subspaces. The corresponding


sequence of dimensions is called its signature. A measurable filtration is a measurable
function on Td , taking values in the space of flags of Rm with almost sure constant
signature. We denote by F(Td , Rm ) the space of measurable filtrations. Note that the
Oseledets filtration of A, which we denote by F(A), is an element of this space.
We denote by F⊃τ (Td , Rm ) the subset of measurable filtrations with a signature
τ or finer. If F ∈ F⊃τ (Td , Rm ) then there is a natural projection F τ with signature
τ , obtained from F by simply ‘forgetting’ some of its components. This space is
endowed with the following pseudo-metric


distτ (F, F ) := dτ (F τ (x), (F  )τ (x)) d x,
Td

where dτ refers to the metric on the τ -flag manifold.


On the space F(Td , Rm ) we consider the coarsest topology that makes the sets
F⊃τ (Td , Rm ) open, and the pseudo-metrics dist τ continuous.
A decomposition of Rm is a sequence of linear subspaces {E j }1≤ j≤k+1 whose
direct sum is Rm . This determines the flag E 1 ⊂ E 1 ⊕ E 2 ⊂ · · · ⊂ E 1 ⊕ · · · ⊕ E k ,
whose signature τ also designates the signature of the decomposition.
A measurable decomposition is a measurable function on Td , taking values in
the space of decompositions of Rm with almost sure constant signature. We denote
by D(Td , Rm ) the space of measurable decompositions. Note that the Oseledets
decomposition of A, which we denote by E · (A), is an element of this space.
We denote by D⊃τ (Td , Rm ) the subset of measurable decompositions with a
signature τ or finer. If E · ∈ D⊃τ (Td , Rm ) then there is a natural restriction E ·τ with
signature τ , obtained from E · by simply ‘patching up’ the appropriate components.
This space is endowed with the following pseudo-metric

distτ (E · , E · ) := dτ (E ·τ (x), (E · )τ (x)) d x,
Td

where dτ refers to the metric on the manifold of τ -decompositions.


210 6 Large Deviations for Quasi-Periodic Cocycles

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.

6.1.2 Literature Review

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.

6.2 Estimates on Unbounded Pluri-Subharmonic Functions

In this section we derive certain uniform estimates on analytic functions of sever-


able variables and on pluri-subharmonic functions. These estimates are of a general
nature, and they will be applied in the next section to quantities related to iterates of
analytic cocycles. Uniformity is understood relative to some measurements which
are stable under small perturbations of the functions being measured. The main tech-
nical difficulties in establishing these estimates are related to the non-trivial nature
of the zeros of an analytic function of several variables, and correspondingly, to the
unboundedness of the pluri-subharmonic functions we study.

6.2.1 The Uniform Łojasiewicz Inequality

Throughout this chapter, a quantitative description of quasi-analyticity, the Ło-


jasiewicz inequality, will play a crucial role. We make the observation that this
property is uniform in a small neighborhood of such a function.

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

Łojasiewicz inequality (6.3) for a fixed, non-constant analytic function f (x) of


several variables has been established for instance in [9] (see Lemma 11.4), and for
smooth, transversal functions in [15] for d = 1 (see Lemma 5.3) and in [16] for d > 1
(see Theorem 5.1). Moreover, the constants S and b in [15, 16] depend explicitly on
some measurements of f , and it is easy to see that these measurements are uniform.
Assume for simplicity that d = 2, although a similar argument holds for any
d ≥ 1. Then recall from [16] that a smooth function f (x) is called transversal if for
any point x ∈ T2 there is a multi-index α = (α1 , α2 ) ∈ N2 , α = (0, 0) such that
the corresponding partial derivative is non-zero: ∂ α f (x) = 0. Clearly, non-constant
analytic functions are smooth and transversal.
By Lemma 5.1 in [16], for such a function f , there are m = m( f ) = (m 1 , m 2 ) ∈
N2 , m = (0, 0) and c = c( f ) > 0 such that for any x ∈ T2 we have
 α 
∂ f (x) ≥ c (6.4)

for some multi-index α = (α1 , α2 ) with α1 ≤ m 1 , α2 ≤ m 2 .


Let
 
A( f ) := max{∂ α f (x) : x ∈ Td , α = (α1 , α2 ), α1 ≤ m 1 + 1, α2 ≤ m 2 + 1}. (6.5)

Theorem 5.1 in [16] says that


   
{x ∈ T2 :  f (x) < t} < St b for all t > 0

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,

hence if α = (α1 , α2 ) with α1 ≤ m 1 + 1, α2 ≤ m 2 + 1 and m = (m 1 , m 2 ) = m( f )


from above, then
c( f )
∂ α g − ∂ α f 0 ≤ B( f )δ = ,
2
c( f )
provided δ = δ( f ) := 2B( f)
.
From (6.4) we conclude that if g − f r < δ, then for every x ∈ T2
 α 
∂ g(x) ≥ c
2
holds for some multi-index α = (α1 , α2 ) with α1 ≤ m 1 , α2 ≤ m 2 .
214 6 Large Deviations for Quasi-Periodic Cocycles

Moreover, for such functions g, the upper bound A(g) satisfies


c
A(g) ≤ A( f ) + ∼ A( f ),
2
which concludes the proof of the lemma. 

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)

where C is a finite explicit constant depending only on  f ∞ , S and b, i.e. on some


measurements of f .

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

holds for any g ∈ Crω (Td , R) with g − f r < δ.

6.2.2 Uniform L 2 -Bounds on Analytic Functions

Consider the following norm on measurable functions f : Td → R


 1/2
 
||| f ||| := sup  f (x1 , . . . , x j−1 , t, x j+1 , . . . , xd )2 dt .
x1 ,...,x j−1 ,x j+1 ,...,xd ∈T T
1≤ j≤d

We say that a measurable function f : Td → R is uniformly separately L 2 -bounded


if ||| f ||| < +∞.

Lemma 6.3 Given a translation T on Td , for any measurable function f : Td → R,


||| f ◦ T ||| = ||| f |||.

Proof Just use (in each variable) the translation invariance of the Lebesgue measure
on T. 

Definition 6.1 We say that a function f : Td → R vanishes along an axis if there


are 1 ≤ j ≤ d and x1 , . . . , x j−1 , x j+1 , . . . , xd ∈ T so that f (x1 , . . . , x j−1 , t,
x j+1 , . . . , xd ) = 0 for every t ∈ T.

 an analytic function f : T2 → R, if f does not vanish along


d
Lemma 6.4 Given
 
any axis then log f is uniformly separately L -bounded.

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

ϕ j;x1 ,...,x j−1 ,x j+1 ,...,xd (t) := f (x1 , . . . , x j−1 , t, x j+1 , . . . , xd )

is not identically zero. Since clearly for all 1 ≤ j ≤ d, the set of functions

{ϕ j;x1 ,...,x j−1 ,x j+1 ,...,xd (t) : x1 , . . . , x j−1 , x j+1 , . . . , xd ∈ T}

is compact, applying Remark 6.2 (with d = 1) to the one-variable  functions above, we


conclude that there is C = C( f ) < ∞ such that logϕ j;x1 ,...,x j−1 ,x j+1 ,...,xd  L 2 < C,
which shows that log f  is uniformly separately L 2 -bounded.
We note that in fact more can be shown, namely that for all  1 ≤ j ≤ d, the function 
H j : Td−1 → R, H j (x1 , . . . , x j−1 , x j+1 , . . . , xd ) := logϕ j;x1 ,...,x j−1 ,x j+1 ,...,xd  L 2 is
continuous, hence it has a maximum value, which leads to the same conclusion. 
216 6 Large Deviations for Quasi-Periodic Cocycles

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

Because M is irreducible, the lines spanned by the columns of M n approach the


line spanned by ω as n → +∞. Hence for n large enough, every column of M n
lies in Σδ . 

Consider the following family of matrices in SL(d, Z)


⎛ ⎞
1 0 0 ··· 0
1
⎜1 1 0 1⎟
··· 0
⎜ ⎟  
⎜0 1 1 0⎟
··· 0
⎜ ⎟ 11
Md = ⎜ . .. .. .. ⎟ if d > 2,
. . .. M2 = (6.8)
⎜ .. . . .⎟. . 12
⎜ ⎟
⎝0 0 0 ··· 1 0⎠
0 0 0 ··· 1 1
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 217

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

pd (λ) = det(Md − λI ) = (−1)d ((λ − 1)d − λ).

In particular, det(Md ) = pd (0) = (−1)d (−1)d = 1. Considering the permutation


matrix ⎛ ⎞
0 0 0 ··· 0 1
⎜1 0 0 ··· 0 0⎟
⎜ ⎟
⎜0 1 0 ··· 0 0⎟
⎜ ⎟
P =⎜. . . . . .⎟
⎜ .. .. .. . . .. .. ⎟
⎜ ⎟
⎝0 0 0 ··· 0 0⎠
0 0 0 ··· 1 0

we have Md ≥ I + P, where the partial order ≥ refers to component-wise comparison


of the entries of the two matrices. Hence
d  
 d
(Md ) ≥ (I + P) =
d d
P j,
j=0
j

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]). 

Proof (Proof of Theorem 6.3) Let f : Td → R be an analytic function such that


f ≡ 0. Take constants c and C such that 0 < c <  f ∞ and D f ∞ < C < +∞.
Let δ > 0 be such that for every g ∈ Crω (Td , R) with g − f r < δ, one still has
0 < c < g∞ and Dg∞ < C. Choose δ > 0 such that C δ < c, and pick a
matrix M ∈ SL(d, Z) such that every column of M lies in Σδ . Then any axis γ of Td
along the coordinate system defined by M has homotopy type in Σδ . We cannot have
g|γ ≡ 0, because the δ-density of γ implies the contradiction c < g∞ ≤ C δ < c.
Therefore, g ◦ M does not vanish along any axis, which implies that logg ◦ M  is
uniformly separately L 2 -bounded.
As before, by a compactness argument, Remark 6.2 (with d = 1) implies that for
some constant C = C( f ) < ∞, the inequality
 
logg(x1 , . . . , x j−1 , · , x j+1 , . . . , xd ) L 2x (T) <C
j

holds for all (x1, . . . , xj−1


 , x j+1 , . . . , xd ) ∈ T
d−1
and g ∈ Crω (Td , R) near f .
Therefore, logg  is uniformly bounded in a neighborhood of f . 
218 6 Large Deviations for Quasi-Periodic Cocycles

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

In other words, without a change of variables, while the uniform L 2 -bound in


Theorem 6.3 may not hold, we still get a polynomial bound off of an exponentially
small set of inputs and in each variable. When applying these estimates to quantities
related to cocycles iterates, the number n above will be correlated with the number
of iterates, and this non-uniform but polynomial L 2 -bound will be manageable.

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.

6.2.3 Estimates on Unbounded Subharmonic Functions

In this subsection we review some crucial estimates on subharmonic functions. That


is, we list the one variable tools to be employed later in the derivation of the base
LDT estimates for pluri-subharmonic observables.
We begin by reminding the reader some elementary facts about subharmonic
functions. She may also consult the monographs [11] or [17].

Definition 6.2 Let Ω be a domain of C. A function u : Ω → [−∞, ∞) is called


subharmonic in Ω if for every z ∈ Ω, u is upper semicontinuous at z and it satisfies
the sub-mean value property:
 1
u(z) ≤ u(z + r e(θ ))dθ,
0

for some r0 (z) > 0 and for all r ≤ r0 (z).


6.2 Estimates on Unbounded Pluri-Subharmonic Functions 219

Basic examples of subharmonic functions are log |z − z 0 | or more generally,


log | f (z)| for some analytic function f (z) or log |z − ζ | dμ(ζ ) for some positive
measure with compact support in C.
The maximum of a finite collection of subharmonic functions is subharmonic,
while the supremum of a collection (not necessarily finite) of subharmonic functions
is subharmonic provided it is upper semicontinuous. In particular this implies that if
A : Ω → Mat(m, C) is a matrix valued analytic function, then
 
u(z) := logA(z) = sup logA(z) v, w
v,w≤1

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

where h(z) is a harmonic function on Ω1 .

Let us assume for now that the subharmonic function u(z) is bounded:

|u(z)| ≤ C for all z ∈ Ω.

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

Therefore, we obtain the following measurement on the total Riesz mass of u:

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

μ(Ω1 ) + h L ∞ (Ω2 ) ≤ C(Ω, Ω1 , Ω2 ) C. (6.10)

We call the estimate in (6.10) a uniform measurement on the bounded subharmonic


function u(z), since it only depends on its bound and on its domain. It is precisely
this measurement that determines the parameters in the base-LDT estimates for the
observable u(x).
This chapter requires similar estimates for subharmonic functions that are un-
bounded from below. A uniform measurement like (6.10) on the Riesz mass and on
the harmonic part of such a subharmonic function was obtained by M. Goldstein and
W. Schlag (see Lemma 2.2 in [10]). The derivation is based on the Poisson-Jensen
formula and on considerations that involve Green’s functions. The result in [10] is
formulated for functions u : Ω → R. It holds, however, also for u : Ω → [−∞, ∞),
as long as u ≡ −∞. That is because the only requirements for the applicability of the
Poisson-Jensen formula (see Sect. 3.7 in [11]) are u ≡ −∞ and some assumptions
on the boundary of Ω.
We formulate the aforementioned result in [10] for a subharmonic function on an
annulus, as this is the context of our model.

Lemma 6.6 Let u : Ar → [−∞, ∞) be a subharmonic function, and let



u(z) = log |z − ζ | dμ(ζ ) + h(z)
Ar/2

be its Riesz representation on the smaller annulus Ar/2 . Assume that

sup u(z) − sup u(z) ≤ C (6.11)


z∈Ar z∈Ar/2

Then
μ(Ar/2 ) + h L ∞ (Ar/4 ) ≤ Cr C, (6.12)

where Cr is a constant that depends on the width r of the annulus.

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

sup u(z) + u L 2 (T) ≤ C. (6.13)


z∈Ar

Remark 6.6 We comment on the order of magnitude of the constant Cr C in (6.12),


as r → 0.
In the bounded case |u(z| ≤ C for all z ∈ Ar , from (6.9) and the fact that Ar/2 can
be covered by O( r1 ) many disks of radius O(r ), it follows that the total Riesz mass
of u is of order r1 C or less. The L ∞ bound on h will be of the same order, showing
that Cr C  r1 C.
In the unbounded case, under the assumption (6.11), an inspection of the proof of
Lemma 2.2 in [10] shows that the constant Cr in (6.12) depends only on the annulus
Ar , via certain estimates on its Green’s function and an argument involving Harnack’s
inequality. These considerations lead to an estimate on Cr which is exponential in r1
as r → 0. One can show, via some calculations involving elliptic integrals, that this
estimate on the order of Cr cannot be significantly improved, unless of course (6.11)
is strengthened.
We note, however, that throughout this chapter, the width r of the annulus Ar is
fixed. We do perform a change of coordinates of the multivariable torus Td , which in
turn affects the size of the domain of the relevant subharmonic functions. However,
this change of coordinates is performed only once. Hence for all intents and purposes,
the constant Cr in this chapter may be treated as a universal constant, and so the
uniform measurement on u(z) given by (6.12) depends only on the bound in (6.11)
or in (6.13).
We formulate the crucial estimates on a subharmonic function u(z) which are
needed in the proof of the LDT: a rate of decay of the Fourier coefficients of u(x)
and an estimate on its BMO norm derived under an appropriate splitting assumption.
The reader may consult [18] or [8] for background on the relevant harmonic analysis
topics.
Let u : Ar → [−∞, ∞) be a subharmonic function, and let

u(z) = log |z − ζ | dμ(ζ ) + h(z)
Ar/2

be its Riesz representation on the smaller annulus Ar/2 .


Assume that
μ(Ar/2 ) + h L ∞ (Ar/4 ) + ∂x h L ∞ (T) ≤ S. (6.14)

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:

u B M O(T)  ε0 + (S ε1 )1/2 . (6.16)

We make some comments regarding the proofs of these lemmas.


These types of results are available for bounded subharmonic functions, see
Chap. 4 in Bourgain’s monograph [3] or Sect. 1 in [4].
A careful inspection of their proofs shows that the boundedness of the subhar-
monic function u(z) is not strictly necessary: it is only used to derive estimates on
the Riesz mass μ(Ar/2 ) and on the L ∞ -norm of the harmonic part h in the Riesz
representation of u, that is, to derive the uniform measurement (6.10) on the annulus.
Moreover, the resulting constants appearing in the estimates on the Fourier coeffi-
cients and on the BMO norm of u(x) depend precisely on this uniform measurement
on u and on the bound on the derivative of h, that is, on ∂x h L ∞ (T) .
Therefore, the bound in (6.14) can be substituted for the boundedness of u(z) and
Lemmas 6.7 and 6.8 are proven along the same lines as their counterparts in [3, 4].

Remark 6.7 Let u : Ar → [−∞, ∞) be a subharmonic function and assume that


(6.13) holds, that is
sup u(z) + u L 2 (T) ≤ C. (6.17)
z∈Ar

Then from Lemma 6.6 we have

μ(Ar/2 ) + h L ∞ (Ar/4 ) ≤ Cr C.

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)

with ε1 ≤ ε04 , then for some absolute constant c > 0,

     1/2 1/4 −1/2


−1
−1 −1/2
{x ∈ T : u(x) − u > ε1/2 } < e−c ε0 +S ε1 ε0 < e−c S ε0 . (6.19)
0

  We may of course assume that u = 0. Denote the set in (6.18) by B, so


Proof
B < ε1 . Then u = u 0 + u 1 with u 0 := u · 1  and u 1 := u · 1B .
B
Clearly u 0  L ∞ (T) ≤ ε0 . Since (6.14) holds, we may apply Lemma 6.7 and (6.15)
1/2
implies u L 2 (T)  S, so u 1  L 1 (T) ≤ u L 2 (T) · 1B  L 2 (T) ≤ S ε1 .
Lemma 6.8 applies, and we have the BMO bound:
1/4
u B M O(T) ≤ ε0 + S ε1 .

The conclusion follows directly from John-Nirenberg inequality (see [18]). 

6.2.4 Base LDT Estimates for Pluri-Subharmonic


Observables

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

Moreover, if ξ is continuous, then the convergence above is uniform in x.


In order to establish fiber-LDT estimates, we need a quantitative version of this
convergence, one which applies to observables that admit a pluri-subharmonic ex-
tension. Moreover, the parameters describing this quantitative convergence should
depend explicitly on a certain uniform measurement of such observables.
We formulate and prove this quantitative version of the Birkhoff ergodic theorem
for pluri-subharmonic functions that are unbounded from below but otherwise satisfy
some bounds on average.
A similar result for bounded pluri-subharmonic functions was formulated and
proven in [16] (see also [3, 4] for results in the same spirit).
We begin with some general considerations on pluri-subharmonic functions.
224 6 Large Deviations for Quasi-Periodic Cocycles

Definition 6.3 Let Ω be a domain in Cd . A function u : Ω → [−∞, ∞) is called


pluri-subharmonic if it is upper semicontinuous and its restriction to any complex
line is subharmonic.

It follows from the definition above that the composition of a pluri-subharmonic


function with a linear function is pluri-subharmonic as well. Moreover, a pluri-
subharmonic function is subharmonic in each variable.
If f (z) is holomorphic (i.e. analytic in each variable), then u(z) = log | f (z)| is
pluri-subharmonic. Moreover, if A(z) is a holomorphic matrix-valued function, then
u(z) = logA(z) is pluri-subharmonic.
We note two important differences between subharmonic and pluri-subharmonic
functions.
Firstly, the zero set of an analytic function of one variable is discrete, which is
of course not the case for several variables analytic functions. Correspondingly, if
u(z) is pluri-subharmonic, then the set {z : u(z) = −∞} can be quite complex, i.e.
a variety of co-dimension 1 in Cd . In particular, this set may contain hyperplanes or
lines parallel to the Euclidian coordinate axes.
Secondly, the Riesz representation Theorem 6.4, which is an important tool in the
study of subharmonic functions, is not available for pluri-subharmonic functions.
Crucial to proving a quantitative Birkhoff ergodic theorem for observables on T
with subharmonic extension to Ar , is having the decay (6.15) on its Fourier coeffi-
cients from Lemma 6.7 and the boosting of a concentration inequality in Lemma 6.9.
Similar results for a pluri-subharmonic function u(z) on Ard can be obtained
through a slicing argument, provided we may apply Lemmas 6.7 and 6.9 in each
variable, and with the same measurements.
More precisely, this type of argument works if the measurement (6.14) applies
uniformly for all subharmonic functions Ar  z i → u(z 1 , . . . , z i−1 , z i , z i+1 , . . . , z d )
where 1 ≤ i ≤ d and z 1 , . . . , z i−1 , z i+1 , . . . , z d ∈ Ar .
Of course this would be automatic if the pluri-subharmonic function u(z) were
bounded. A weaker assumption is for (6.13) to hold uniformly in each variable. In
other words, we require that u(z) be bounded from above on Ard and that its L 2 -norm
in each variable be bounded as well. The latter assumption is equivalent to having a
bound on |||u||| (see Sect. 6.2.2 for the meaning of this norm).
To summarize, the assumption we make in this section on a pluri-subharmonic
function u : Ard → [−∞, ∞) is that for some constant C < ∞ we have

sup u(z) + |||u||| ≤ C. (6.20)


z∈Ard

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)

with ε1 ≤ ε08 , then for some absolute constant c > 0,

     1/4 1/8 −1/2


−1
−1 −1/4
{x ∈ T2 : u(x) − u > ε1/4 } < e−c ε0 +S ε1 ε0 < e−c S ε0 . (6.22)
0

For d variables, replace the powers 1/4 by 1/2d etc.

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.

Theorem 6.5 Let u : Ard → [−∞, ∞) be a pluri-subharmonic function satisfying

sup u(z) + |||u||| ≤ C. (6.23)


z∈Ard

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 |

for all l = (l1 , l2 ) ∈ Z2 with l1 = 0, l2 = 0.


Expand u(x) = u(x1 , x2 ) into a Fourier series

u(x1 , x2 ) = u + û(l1 , l2 ) · e((l1 , l2 ) · (x1 , x2 )).
(l1 ,l2 )∈Z2
(l1 ,l2 ) =(0,0)

Then the Birkhoff averages have the form

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)

where we denoted by K n (y) the Fejér kernel on T:

1
n−1
1 1 − e(ny)
K n (y) = e( j y) = .
n j=0 n 1 − e(y)

Clearly K n (y) has the following bound:

   1 
 K n (y) ≤ min 1, (6.26)
ny

where y was defined in (6.1).


We then have:
1 n−1 2
 
 u((x1 , x2 ) + j (ω1 , ω2 )) − u 2 2
n j=0 L (T )
    
= û(l1 , l2 )2 ·  K n ((l1 , l2 ) · (ω1 , ω2 ))2
(l1 ,l2 )∈Z2
(l1 ,l2 ) =(0,0)
6.2 Estimates on Unbounded Pluri-Subharmonic Functions 227
    
= û(l1 , l2 )2 ·  K n ((l1 , l2 ) · (ω1 , ω2 ))2
1≤|l1 |+|l2 |<K
    
+ û(l1 , l2 )2 ·  K n ((l1 , l2 ) · (ω1 , ω2 ))2 .
|l1 |+|l2 |≥K

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

Estimate (6.25) clearly implies:

  1
û(l1 , l2 )  S .
|l1 | + |l2 |

Then using the Diophantine condition on (ω1 , ω2 ) and (6.26), we obtain:


    
û(l1 , l2 )2 ·  K n ((l1 , l2 ) · (ω1 , ω2 ))2
1≤|l1 |+|l2 |<K
 1 1
 S2 · 2
1≤|l1 |+|l2 |<K
(|l1 | + |l2 |) n (l1 , l2 ) · (ω1 , ω2 )2
2

 1 (|l1 | + |l2 |)2(2+δ0 )


≤ S2 ·
1≤|l1 |+|l2 |<K
(|l1 | + |l2 |)2 n2 t 2
K 2(1+δ0 )
 S2 .
n2 t 2
228 6 Large Deviations for Quasi-Periodic Cocycles

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

Then v is a pluri-subharmonic function on Ar2 and it clearly satisfies the bounds


1
sup v(z) ≤ C ≤ 1 and |||v||| ≤ S1 |||u||| ≤ 1, hence
2
z∈Ar S

sup v(z) + |||v|||  1.


z∈Ar2

Since v = 1
S
u, from (6.27) we have:
   
{x ∈ T2 : v(x) − v > n −a/5 } < n −8a/5 .

Let ε0 := n −a/5 , ε1 := n −8a/5 = ε08 .


Applying Lemma 6.10 to the function v(z), we obtain the stronger estimate:
   
{x ∈ T2 : v(x) − v > n −a/20 } < e−c na/20 ,

where c = O(1). This completes the proof. 

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

holds for all n ≥ n 0 := t −2 k0 .

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

From (6.28) we have:

   
m−1

u(x) ≤ 2C + 1 logg(x + jω) for all x ∈ Td .
m j=0

Then using Lemma 6.3,

|||u||| ≤ 2C + |||log |g|||| ≤ 2C + C0 ,

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. 

6.3 The Proof of the Fiber LDT Estimate

Given any cocycle A ∈ Cm , we derive some uniform measurements on A which will


allow us to apply the base LDT estimate in Proposition 6.2 to all the iterates of any
cocycle B near A, in a uniform way. This, combined with an almost invariance under
the base dynamics property for the iterates of the cocycle, will lead to the proof of
the uniform fiber LDT estimate.

6.3.1 Uniform Measurements on the Cocycle

We introduce some notations. For a cocycle A ∈ Crω (Td , Mat(m, R)), let

f A (z) := det[A(z)].

Then clearly f A ∈ Crω (Td , R).


Moreover, for every scale n ≥ 1, let

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

The following proposition introduces some locally uniform measurements on a


cocycle. It shows that the above functions are bounded in the L 2 -norm, uniformly
in the scale, and uniformly in a neighborhood of a given non identically singular
cocycle. It also shows that the failure of the above functions to be bounded in the
L ∞ -norm is captured by Birkhoff averages of a one dimensional cocycle.

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)

and for all n ≥ 1 we have

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

To establish the lower bound, we use Cramer’s rule:

det[B(z)] · I = B(z) · adj(B(z)).

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

which establishes (6.33).


Moreover, (6.33) also implies that for all x ∈ Td ,

 (n)  
n−1
  
u (x) ≤ 2C1 + 1 log f B (T i x).
B
n i=0

Hence by (6.32) and the measure invariance of the translation T ,


 
u (n)  
B  L 2 (Td ) ≤ 2C 1 + log f B  L 2 (Td )  C 1 + C 2 . 

6.3.2 The Nearly Almost Invariance Property

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

B(x) ≤ B − Ar + Ar < δ + Ar ∼ Ar ,

thus logB(x) ≤ C(A) for all x ∈ Td .


Similarly,

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)

which holds because

adj(B(x))  B(x)m−1 ≤ Brm−1 ≤ C1 (A).

Fix a ∈ (0, 1) and apply (6.36) with


 1/b
1
e−(1/b) n .
1−a
t :=
C
 
  −n 1−a
 there is a set Bn ⊂ T with Bn < Ct = e , such that if x ∈
/ Bn then
d b
 Then
 f B (x) ≥ t. Thus

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. 

6.3.3 The Statement and Proof of the LDT

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

We apply Proposition 6.2 with f = f A , C = C(A), so the dependence of the


constants on the data will be: δ = δ( f A , r ) = δ(A), a = a(d), k0 = k0 ( f A , r, d) =
k0 (A) and S = S( f A , r, C) = S(A).
Since the map Crω (Td , Mat(m, R))  B → f B ∈ Crω (Td , R) is locally Lipschitz,
by possibly decreasing δ, we may assume that whenever B − Ar < δ we have
that  f B − f A r is small enough that the pluri-subharmonic function u(z) = u (n) B (z)
satisfies the assumption (6.28) with g = f B . Hence Proposition 6.2 applied to our
context says that for all R ≥ n 0 we have:

   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

6.4 Deriving Continuity of the Lyapunov Exponents

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

We say than an observable ξ ∈ Ξ satisfies a base-LDT estimate with parameter


space P, 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 ∈ Td :  1 ξ(T j x) − ξ d x  > ε(n)} < ι(n).
n j=0 Td

A torus translation by a rationally independent vector ω is uniquely ergodic, hence


the convergence in Birkhoff’s ergodic theorem is uniform for continuous observables.
This shows that the base-LDT estimate holds trivially for ξ ∈ Ξ , with ε ≡ ε and
ι ≡ 0.
Finally, a cocycle A ∈ Cm is said to satisfy a uniform fiber-LDT with parameter
space P, if for every ε > 0 there are δ > 0 and an LDT parameter (n 0 , ε, ι) ∈ P
which may only depend upon A and ε, such that for any B ∈ Cm with dist(B, A) < δ
and for all n ≥ n 0 , we have ε(n) ≤ ε and

   
{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

some measurements of A and the Diophantine condition on ω.


Theorem 3.1 in Chap. 3 says that given an ergodic system, a space of cocycles,
a set of observables and a set of LDT parameters, if the compatibility condition,
the uniform L 2 -boundedness, the base-LDT and the uniform fiber-LDT estimates
hold, then all Lyapunov exponents are continuous. Moreover, if A ∈ Cm has a
Lyapunov spectrum gap, i.e. for some 1 ≤ k ≤ m, L k (A) > L k+1 (A), then locally
near A, the map B → Λk (B) = (L 1 + · · · + L k )(B) has a modulus of continuity
ω(h) := [ι (c log h1 )]1/3 , for some c > 0 and some deviation measure function ι
corresponding to an LDT parameter in P.
Given that the deviation measure functions obtained in this chapter have the form
ι(t) ≡ e−t , the modulus of continuity is ω(h) = e−c [log(1/ h)] , i.e. we obtain weak-
b b

Hölder continuity.
The statements on the Oseledets filtration and decomposition follow directly from
the corresponding general criteria. 

6.5 Refinements in the One-Variable Case

Let us now consider the case of a one-variable torus translation by a frequency ω


that satisfies a stronger Diophantine condition, namely:
238 6 Large Deviations for Quasi-Periodic Cocycles

t
kω ≥     (6.41)
k  (log k )1+

for some t > 0 and for all k ∈ Z \ {0}.


Using the already established fiber-LDT estimate in Theorem 6.6 and the cor-
responding continuity result for Lyapunov exponents, we derive a sharper fiber-
LDT estimate under the additional assumption that the cocycle A has the property
L 1 (A) > L 2 (A). This in turn leads to a stronger modulus of continuity of the LE,
the Oseledets filtration and the Oseledets decomposition.
We note that our argument for proving this sharper fiber-LDT does require the gap
condition L 1 (A) > L 2 (A), since it depends essentially on the Avalanche Principle
proven in Chap. 2. In fact, the argument requires the full strength of the continuity
results in Chap. 3, specifically the finite scale uniform estimates in Lemma 3.4.
We remind the reader the particular estimate in the general AP which we use
here (see Chap. 2, specifically Proposition 2.42 in Sect. 2.4). We recall the notation
gr(g) = ss21 (g)
(g)
∈ [1, ∞] for the ratio of the first two singular values of a matrix
g ∈ Mat(m, R).
Proposition 6.5 There exists c > 0 such that given 0 < ε < 1, 0 < κ ≤ c ε2 and
g0 , g1 , . . . , gn−1 ∈ Mat(m, R), if

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

When it comes to other quantities a and b, a  b simply means K −1 a ≤ b ≤ K a,


where the constant K is context-universal.
Using (6.43) and the LDT estimate (6.38) applied at scales m 1 , m 2 , m 1 + m 2 , it
follows from Lemma 3.3 in Chap. 3 that if m 1 , m 2  n 0 , then

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)

holds for all x outside a set of measure  e−n 0 .


b

Moreover, from (6.42) we have, for m  n 0

1
gr(B (m) (x)) ≥ en 0 γ /2 =: , (6.45)
κ

for all x outside a set of measure < e−n 0 .


b

Note also that κ


= e−n 0 γ /10  1 . (6.46)
ε2
b1
Let 0 < b1 < b and let n 1  en 0 , hence n 0  (log n 1 )1/b1 .
Consider the n 1 -st iterate B (n 1 ) (x) = B(T n 1 −1 ) · · · · · B(T x) · B(x).
We will break down this block (i.e. product of matrices) of length n 1 into (differ-
ent configurations of) blocks of length  n 0 , and apply the avalanche principle in
Proposition 6.5 to the resulting configurations. The estimates (6.44), (6.45), (6.46)
will ensure that the hypotheses of the avalanche principle are satisfied outside a
relatively small set of phases.
Pick any n  nn01 many integers m 0 , m 1 , . . . , m n−1 such that


n−1
m i  n 0 and n 1 = mi .
i=0

Let q0 := 0 and qi := m i−1 + · · · + m 0 = m i−1 + qi−1 for 1 ≤ i ≤ n − 1.


Define gi = gi (x) := B (m i ) (T qi x) for 0 ≤ i ≤ n − 1.
Then g (n) = gn−1 . . . g0 = B (n 1 ) (x) and gi gi−1 = B (m i +m i−1 ) (T qi−1 x).
Using (6.44), if x is outside a set of measure < e−n 0 we have
b

gi gi−1  B (m i +m i−1 ) (T qi−1 x)


= > ε, (6.47)
gi  gi−1  B (m i ) (T m i−1 T qi−1 x) B (m i−1 ) (T qi−1 x)

while from (6.45), for a similar set,

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

We will compute an average of (6.49) to establish a relation between the finite


scale Lyapunov exponents L (n 1) (n 0 )
1 (B), L 1 (B) and L 1
(2n 0 )
(B) (see formula (6.55)
below). In this average, the first and the last terms appearing in the second sum
of (6.49) will be discarded. Let us now explain why these two terms are negligible.
First note that if m  n 0 , then applying (6.38), we have that for x outside a set of
measure < e−n 0 ,
b

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

we conclude that for x outside a set of measure < e−n 0 , we have


b

 
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:

m 0 = n 0 + j, m 1 = m 2 = · · · = m n−2 = n 0 and m n−1  n 0 ,

so that they all add up to n 1 . Then q0 = 0 and qi = in 0 + j.


Apply (6.50) for each of these n 0 configurations of integers m i , 0 ≤ i ≤ n − 1,
add up all the estimates and divide by n 0 to get:

 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

for all x outside a set of measure < n 0 e−1/2 n 0 .


b

Estimate (6.51) can be written as:

 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

for all x outside a set of measure < e−1/3 n 0 .


b

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

the number (n − 2) n 0 and respectively (n − 3) n 0 of translates in (6.52) are  n 1 ,


then up to an additional error of order nn01 , using Proposition 6.6 we obtain:

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

for x outside a set of measure < e−c3 (log n 1 )


1/b1
.
Now integrate (6.52) to get:
 (n 1 ) 
 L (B) + L (n 0 ) (B) − 2L (2n 0 ) (B)
1 1 1
n0 (log n 1 )1/b1
< C + C e−1/3 n 0 
b
. (6.55)
n1 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

u B M O(T)  ε0 + (S ε1 )1/2  ε0 ,

provided n 1 is large enough, depending on A, so that ε1 absorbs the constant S.


Therefore,
(log n 1 )1/b1
u B M O(T)  ε0  ,
n1

and so John-Nirenberg’s inequality implies that for all ε > 0


   
{x ∈ T : u(x) − u > ε} < e−c0 ε/u B M O(T) .

Writing the last estimate in terms of iterates of our cocycle we have:


 
   
{x ∈ T :  1 logB (n 1 ) (x) − L (n 1 ) (B) > ε} < e−c0 ε n 1 /(log n 1 )1/b1 . (6.57)
 n1 1 

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

ι(t) ≡ e−c t/[log t] ,


b
for some c, b > 0.
6.5 Refinements in the One-Variable Case 245

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

Abstract We describe limitations, counterexamples and extensions of the topics


presented in this monograph. We outline some connections with the spectral theory
of discrete Schrödinger operators with ergodic potentials. We formulate a few related
open problems, some of which may be studied using similar methods.

7.1 Limitations and Counterexamples

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

© Atlantis Press and the author(s) 2016 247


P. Duarte and S. Klein, Lyapunov Exponents of Linear Cocycles,
Atlantis Studies in Dynamical Systems 3, DOI 10.2991/978-94-6239-124-6_7
248 7 Further Related Problems
   
a − E −1 b − E −1
XE = YE = .
1 0 1 0

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|)

which tends to 0 as |a − b| → +∞.


The previous limitations on the modulus of continuity of the Lyapunov exponents
also apply to the continuity of the Oseledets filtration.
Assuming the cocycle A has an exact gap pattern τ = (τ1 , . . . , τk ), if the Oseledets
filtration has a certain modulus of continuity around A then, because the Oseledets
filtration F = (F1 , . . . , Fk ) is characterized by F j = vτ(∞)
j
(A)⊥ , the most expand-
ing directions vτ(∞)
j
(A) = Ψ −1 (v(∞) (∧τ j A)) ∈ Gr τ j (Rm ) have the same modulus of
continuity. Recall that Ψ denotes the Plücker embedding of Gr k (Rm ) into P(∧k Rm ).
Therefore, in order to understand how the limitations on the continuity of the Lya-
punov exponents imply similar limitations on the continuity of the Oseledets fil-
tration, it is enough to see that Hölder continuity of the most expanding direction
v(∞) (A) implies Hölder continuity of the first Lyapunov exponent L 1 (A).
To be more precise, consider the space Cm of all bounded measurable cocycles
A : X → GL(m, R), with bounded inverse, endowed with a metric such that

dist(A, B) ≥ sup A(x) − B(x).


x∈X

Recall that L 1 (X, P(Rm )) denotes the space of measurable functions F : X →


P(Rm ), where two functions are identified when they differ over azero measure set.
This is a complete metric space with the distance d(F, G) := d(F(x), G(x))
X
μ(d x).
Proposition 7.1 Given A ∈ Cm such that L 1 (A) > L 2 (A), if v(∞) : Cm →
L 1 (X, P(Rm )) is locally α-Hölder around A then the first Lyapunov exponent
L 1 : Cm → R is also locally α-Hölder in the same neighborhood of A.
Proof Consider a unit measurable section v A : X → Rm of v(∞) (A) : X → P(Rm ).
By Proposition 4.11,

L 1 (A) = logA(T −1 x)∗ v A (x) μ(d x).
X

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

vector (Bernoulli measure). More precisely, given matrices X 0 , X 1 , . . . , X k ∈


Mat(m, R) and a probability vector p = ( p0 , . . . , pk ) in the k-dimensional sim-
plex Δk , Y. Peres considers the measure μ p := kj=0 p j δ X j , and proves that the
function p
→ L 1 (μ p ) is analytic in the interior of the simplex Δk . The negative
example (7.1) justifies the need for excluding the boundary of the simplex.
In the same context of random cocycles over Bernoulli shifts, Le Page [31] has
proven the smoothness of the first Lyapunov exponent for cocycles determined by
probability measures which are absolutely continuous with respect to the group’s
Haar measure. In the same spirit we mention the work of Bourgain [12, Theorem
3], [13, Sect. 4] on the smoothness of the Lyapunov exponents for certain random
Schrödinger SL(2, R) cocycles (e.g. Anderson-Bernoulli).
Finally, there is the work of Ávila [1] who proves the analyticity of the first
Lyapunov exponent on the strata of some stratification of the space of analytic
quasi-periodic Schrödinger cocycles over an irrational rotation of the circle T.
These results indicate that the regularity of the Lyapunov exponents is dependent
upon the choice of the space of cocycles.

7.2 Some Connections to Mathematical Physics

We outline some connections to the spectral theory of discrete Schrödinger operators.


These types of operators describe the Hamiltonian of a quantum particle on the integer
lattice. We specialize to ergodic operators, whose potentials are dynamically defined.
We then describe some related problems for the more general case of block Jacobi
operators.
Let (X, μ, T ) be an ergodic dynamical system. A discrete ergodic Schrödinger
operator is an operator Hλ (x) on l 2 (Z) ψ = {ψn }n∈Z , defined by

[Hλ (x) ψ]n := −(ψn+1 + ψn−1 ) + λ vn (x) ψn , (7.2)

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),

for some bounded and measurable potential (or sampling) function f : X → R.


Due to the ergodicity of the system, the spectral properties of the family of oper-
ators {Hλ (x) : x ∈ X } are independent of x almost surely.
Consider the Schrödinger (i.e. eigenvalue) equation

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

The IDS is known to be log-Hölder continuous in a very general setting (see


[19]). A stronger modulus of continuity may be derived from that of the LE using
the Thouless formula. We refer the reader to Lemma 10.3 in [24], which shows
Free ebooks ==> www.Ebook777.com
252 7 Further Related Problems

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).

Question 2 Study the spectral properties (e.g. localization in the appropriate


regime) of the operator (7.4) with random or quasi-periodic underlying dynamics.

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.

7.3 Continuity for Other Spaces of Cocycles

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.

Question 3 Consider the space of m × m analytic, quasi-periodic cocycles with


singularities. Assume a generic arithmetic condition on the translation. Assume say,
simplicity of the Lyapunov spectrum.
(a) In the one variable case d = 1, are the Lyapunov exponents Hölder, instead of
just nearly-Hölder continuous?
(b) If so, can the Hölder exponent be explicitly given in terms of some characteristics
of the cocycle entries?
(c) For the several variables d > 1 model, can the modulus of continuity of the
Lyapunov exponents be improved from weak-Hölder to Hölder?

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

We know from [8] that ε


→ L(Aε ) is continuous near 0. What is the modulus of
continuity of this map?
In Chaps. 5 and 6 we applied the abstract continuity theorem of the Lyapunov
exponents derived in Chap. 3 to various models of random and quasi-periodic cocy-
cles, thus extending previous quantitative continuity results to more general spaces
of cocycles. In fact, most available results regarding quantitative continuity in the
cocycle, including results referring to other types of base dynamics, do fit our scheme,
as we explain below.
Consider the Schrödinger equation

− (ψn+1 + ψn−1 ) + λ f (T n x) ψn = E ψn (7.6)

associated to a discrete, one dimensional Schrödinger operator with dynamically


defined potential, as described in Sect. 7.2.
Let (Aλ, f,E , T ) be the associated Schrödinger cocycle, where
 
λ f (x) − E −1
Aλ, f,E (x) := ∈ SL(2, R).
1 0

We consider different types of base dynamics T as follows.

i. Let T be the skew translation on T2 given by

T (x1 , x2 ) := (x1 + x2 , x2 + ω),


256 7 Further Related Problems

where ω ∈ T satisfies a Diophantine condition.


Assume also that f (x1 , x2 ) is a non-constant real-analytic function on T2 .
J. Bourgain, M. Goldstein and W. Schlag proved (see [16])   that for every ε > 0,
there are λ0 = λ0 (ε, f ) and a set of frequencies Ωε with Ωε  > 1 − ε, such that for
all ω ∈ Ωε and for all λ ≥ λ0 , the map E
→ L 1 (Aλ, f,E ) is weakly-Hölder. That is,
as a function of the energy parameter E, the Lyapunov exponent is weakly-Hölder
continuous. We note that the assumption λ  1 ensures the positivity of the first LE,
and hence the gap between the Lyapunov exponents.

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.

iii. Let T be either a Diophantine translation on Td , d ≥ 1 or the skew-translation


defined above. Assume that f is a Gevrey-class function, that is, for all multi-
indices m ∈ Nd we have
 
sup ∂ m f (x)  K |m| (m!)s (7.7)
x∈Td

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}

for a given function f , coupling constant λ and appropriate interval of energies I.


Then if f is as above (i.e. real-analytic or Gevrey-class), there is λ0 such that
C ( f, λ) satisfies the assumptions of the ACT, provided λ ≥ λ0 for the translation
and skew-translation and λ ≤ λ0 for the expanding map base dynamics.
7.3 Continuity for Other Spaces of Cocycles 257

These considerations lead to some natural questions.


Question 5 Let T be as above (i..e a skew translation or an expanding map of the
torus) and assume that the cocycle is analytic. Study the continuity problem of the LE
for Schrödinger cocycles Aλ, f,E without restrictions on λ. In other words, if for fixed
λ and for some E 0 we have L 1 (Aλ, f,E0 ) > 0, is the map E
→ L 1 (Aλ, f,E ) (weakly-)
Hölder continuous near E 0 ? What about the continuity in (λ, E)?
Question 6 Instead of Schrödinger, hence SL(2, R)-valued cocycles, consider the
more general case of say GL(m, R)-valued cocycles over the same type of base
dynamics. Are the LE still (weakly-) Hölder in a neighborhood of a cocycle with
simple Lyapunov spectrum?
Regarding the last two question above, we note that when T is the skew transla-
tion, we could derive the fiber LDT estimates needed for the ACT in an inductive
procedure, if only an initial scale LDT estimate were available. Getting the initial
scale LDT estimate is precisely where the large λ assumption is used in the case of
Schrödinger cocycles, and we do not have an argument for it otherwise.
Let T be a translation on T by a well chosen frequency ω. As the work [37] of Y.
Wang and J. You shows, unlike in the analytic case, continuity of the LE exponents
does not hold in the larger space of C l , 1 ≤ l ≤ ∞ cocycles.  
f (x) −1
More precisely, assume ω of bounded type and let A f (x) := . Then
1 0
the discontinuity of the LE from [37] occurs in the space C := {A f : f ∈ C l (T)},
at a point A f0 where f 0 is some non transversal function.
As mentioned above, continuity of the LE for the one-parameter family of
Schrödinger cocycles {Aλ, f,E : E ∈ R} (hence f is fixed) was obtained for f in any
Gevrey-class (which is intermediate between analytic and C ∞ ) under a transversality
assumption. If, moreover, the order of the Gevrey-class is restricted, the transversality
condition is not necessary. All of these suggest the following questions.
Question 7 What is responsible for the continuity of the LE for quasi-periodic cocy-
cles: the regularity of the function, the transversality condition, or the choice of the
topology on the space of cocycles? What is the regularity class threshold between
positive and negative continuity results?
Question 8 Study the continuity of the LE for Gevrey-class, GL(m, R)-valued cocy-
cles over a translation or a skew-translation.
We note, as with Question 6, that the main obstacle for proving fiber LDT estimates
inductively is again the initial step.

7.4 Continuity with Respect to Other Parameters

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

In the context of quasi-periodic cocycles, the bundle map is determined by a


translation vector ω ∈ Rd as well as by a function A : Td → Mat(m, R), and hence
it is natural to ask about joint continuity in the pair (ω, A).
In the case of analytic quasi-periodic Schrödinger cocycles, Bourgain and Jit-
omirskaya [15] obtained joint continuity in frequency and energy (ω, E) at all
points with ω ∈ T irrational. Later Bourgain [11] extended this result to quasi-
periodic cocycles on Td with d > 1. In a much broader context, Jitomirskaya and
Marx [27] obtained joint continuity in (ω, A) for irrational ω and analytic cocycles
A : T → Mat(2, C). A generalization to cocycles of arbitrary dimension follows
from the work of Ávila et al. [2]. All of these results are non quantitative.
Recently, Tao [34] proved weak-Hölder joint continuity of the LE in (ω, A), for
ω ∈ DC and A : Td → Mat(2, R) analytic, non-identically singular and with a fixed
determinant. Here DC denotes any full measure set of vectors ω ∈ Rd satisfying a
Diophantine condition with a fixed exponent but with varying constant.

Question 9 Consider the space Cm of analytic functions A : Td → Mat(m, R)


which are non-identically singular. Is there some quantitative joint continuity of the
LE in (ω, A) ∈ DC×Cm , assuming for instance the Lyapunov spectrum to be simple?

Question 10 Consider the space Cm of analytic functions A : Td → GL(m, R).


Are the LE jointly continuous in (ω, A) ∈ {rationally indep. vectors} × Cm ?

In the random setting, Markov cocycles are determined by a stochastic kernel


K on some compact space of symbols Σ and a measurable function X x

A(x) ∈ Mat(m, R) on the space of sequences X = Σ Z , which depends only on two
coordinates x0 and x1 . The base dynamics for these cocycles is a Markov shift on the
space of sequences X . Bernoulli cocycles correspond to the special case where the
transition probabilities K (x, ·) = μ are constant in x, and the matrix valued function
A(x) depends only on the coordinate x0 . In this setting it is natural to ask about joint
continuity of the LE in (K , A) (for Markov cocycles) or in (μ, A) (for Bernoulli
cocycles).
In the context of Bernoulli cocycles, Bocker-Neto and Viana [8] proved joint con-
tinuity of the LE in (μ, A) for μ ∈ Prob(Σ) and A : Σ Z → GL(2, R) depending
only on the first coordinate. The generalization of this result to cocycles of arbi-
trary dimension has been announced by A. Ávila, A. Eskin and M. Viana (see the
introduction of the book [36]). Finally we mention the work of E. Malheiro and M.
Viana [30], in the context of random Markov cocycles, who prove joint continuity
of the LE in (K , A), where K is a stochastic matrix in some finite space of symbols
Σ, and A : Σ Z → GL(2, R) depends only on the first coordinate x0 .
These results provide no modulus of continuity.

Question 11 Consider the space Cm of measurable functions A : Σ Z → GL(m, R)


which depend only on the first coordinate x0 . Is there some quantitative joint conti-
nuity of the LE in (μ, A) ∈ Prob(Σ) × Cm , assuming for instance that the Lyapunov
spectrum is simple and the cocycle is irreducible?
7.4 Continuity with Respect to Other Parameters 259

Question 12 Consider the space Cm of measurable functions A : Σ Z → GL(m, R)


which depend only on the first two coordinates x0 and x1 . For some appropriate
topology on the space M (Σ) of stochastic kernels on Σ, is there some quantitative
joint continuity of the LE in (K , A) ∈ M (Σ) × Cm , assuming for instance that the
Lyapunov spectrum is simple and the cocycle is irreducible?
Some of these problems will be addressed in future projects. We believe that
the work in this monograph can be easily adapted to obtain positive answers to the
Questions 1, 9, 11 and 12.

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

duality, 29 random Bernoulli, 8


Hausdorff distance, 39 Halperin example, 247
intersection operation, 29 random Markov, 9, 165
metrics ρ, d, δ, 28 regular point, 126
minimum distance, 39 Schrödinger, 251
orthogonal hyperplane, 41 singular basis, 126
Plücker embedding, 28 Linear map
pull back action, 30 α-angle, 44
push forward action, 30 β-angle, 44
sum operation, 29 τ -block product, 76
τ -gap pattern, 36, 117
exact, 117
J τ -gap ratio, 36
Jacobi operator, 250, 252 τ -singular value product, 76
k singular gap, 35
k-gap ratio, 35
K adjoint, 27
Kingman’s ergodic theorem, 7, 118 expansion rift, 43, 63
first singular gap, 35
gap ratio, 35, 63, 117
L
kernel, 29
Laplace-Markov operator, 177
least expanding τ -flag, 37
Large deviation principle, 13
least expanding k-subspace, 36
Large deviation type estimates, 4, 13, 84,
most expanding τ -flag, 37, 117
169, 185
most expanding k-subspace, 36, 117
LDT
most expanding direction, 36, 117
base, 14, 85, 169
pseudo inverse, 138
quasi-periodic cocycle, 225, 229
range, 29
estimates, 14, 84, 185
relative distance, 57
exponential type, 170, 185
Lojasiewicz inequality, 212
fiber, 15, 85, 170
Lyapunov
parameters, 14, 85
τ -block, 167, 209
uniform fiber, 15, 85, 170, 211
Least expansion, 32 exponents, 9, 117, 119, 120, 137, 167
Linear cocycle, 8, 81, 119 finite scale, 82
τ -gap pattern, 117, 137, 149, 209 spectrum, 10, 117, 137
exact, 117, 137, 149 gap, 109
adjoint, 128
base dynamics, 8, 81
compatible with set of observables, 11, M
84 Markov
fiber action, 8, 81 kernel, 6, 161
integrable, 8, 119 Doeblin condition, 165
irreducible, 16, 166 ergodic measure, 162
iterates, 8, 81 finite order, 202
limit singular basis, 134 invariant set, 162
most expanding τ -flag, 136 stationary measure, 162
most expanding k-plane, 125 strongly mixing, 6
most expanding direction, 121 system, 162
non identically singular, 17, 208 finite order, 203
quasi-periodic, 8, 208 Kolmogorov extension, 163
almost Mathieu, 247 observable, 170
Schrödinger, 211 strong mixing, 163
with singularities, 208, 211 Markov operator, 176
Free ebooks ==> www.Ebook777.com
Index 263

Measurable bundle, 118 orthogonal hyperplane, 41


fiber, 118 space, 23
Measurable decomposition, 168
τ -pattern, 152, 168
space of, 149, 152, 168 S
Measurable filtration, 167 Schrödinger
τ -pattern, 151, 167 cocycle, 251
space of, 149, 151, 167 quasi-periodic, 211
Measure preserving dynamical system, 5 operator, 250
ergodic, 5, 81 quasi-periodic, 211
mixing, 5 Separately L 2 -bounded function, 215
Measure preserving transformation, 5 uniformly, 216
Modulus of continuity Sequence
Hölder, 17 ε-doubling, 122
nearly-Hölder, 245 Shift, 81, 163
weak-Hölder, 17 Bernoulli, 6
Multiplicative ergodic theorem, 113 Markov, 6
Singular value, 32
decomposition, 33
O singular basis, 33
Observable, 6, 84 volume expansion factor,det+ , 34
compatible with cocycle, 11, 84 Singular vector, 32
cumulant generating function, 171 Space of cocycles
future independent, 169 L p -bounded, 10, 83
Hölder measurable, 10, 82
norm, 168, 178 uniformly L p -bounded, 10, 83
space of, 168, 178 Stochastic kernel, 6
sum process, 170 Subharmonic function, 218
cumulant generating function, 172 BMO norm estimate, 222
Observed Markov system, 170 Fourier coefficients decay, 221
Operation ⊕, 44 Riesz representation theorem, 219
Operator uniform measurement of, 220, 222
Laplace-Markov, 177
Markov, 176
quasi-compact, 176 T
simple, 176 Torus translation, 5, 81
spectral radius, 176 Transversal
Oseledets intersection, 51
decomposition, 168 measurements θ+ , θ∩ , 51
filtration, 168 sum, 51
multiplicative ergodic theorem, 9, 120,
137, 139
U
Upper semicontinuity of the top LE, 86, 92
P
Pluri subharmonic function, 211, 223
uniform measurement of, 224 V
Projective Vector space
α-angle, 41 τ -decomposition, 54, 117
action, 24 intersection operation, 56, 117
metric δ, 38 space of, 117
metrics ρ, d, δ, 24 transversality, 55, 117

www.Ebook777.com

You might also like