Atlantis Series in Dynamical Systems

Series Editors: Henk Broer · Boris Hasselblatt

Jaap Eldering

Normally Hyperbolic
Invariant Manifolds
The Noncompact Case
Atlantis Series in Dynamical Systems

Volume 2

For further volumes:

http://www.springer.com/series/11155
Jaap Eldering

Normally Hyperbolic
Invariant Manifolds
The Noncompact Case
Jaap Eldering
Department of Mathematics
Utrecht University
Utrecht
The Netherlands

ISSN 2213-3526
ISBN 978-94-6239-002-7 ISBN 978-94-6239-003-4 (eBook)
DOI 10.2991/978-94-6239-003-4

Library of Congress Control Number: 2013939037

Ó Atlantis Press and the author 2013

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

Series Information

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.

Series Editors
Henk Broer
University of Groningen
Johann Bernoulli Institiute for Mathematics and Computer Science
Groningen, The Netherlands

Boris Hasselblatt
Department of Mathematics, Tufts University
Medford, USA

Atlantis Press
29, avenue Laumière
75019 Paris, France
For more information on this series and our other book series, please visit our
website www.atlantis-press.com
Preface

In this work, we prove the persistence of normally hyperbolic invariant manifolds.

This result is well known when the invariant manifold is compact; we extend this
to a setting where the invariant manifold as well as the ambient space are allowed
to be noncompact manifolds. The ambient space is assumed to be a Riemannian
manifold of bounded geometry.
Normally hyperbolic invariant manifolds (NHIMs) are a generalization of
hyperbolic fixed points. Many of the concepts, results, and proofs for hyperbolic
fixed points carry over to NHIMs. Two important properties that generalize to
NHIMs are persistence of the invariant manifold and existence of stable and
unstable manifolds.
We shall focus on the first property. Persistence of a hyperbolic fixed point
follows as a straightforward application of the implicit function theorem. For a
NHIM the situation is significantly more subtle, although the basic idea is the
same. In the case of a hyperbolic fixed point, we only have stable and unstable
directions. When we consider a NHIM, there is a third direction, tangent to the
manifold itself. The dynamics in the tangential directions is assumed to be dom-
inated by the stable and unstable directions in terms of the respective Lyapunov
exponents. Thus the dynamics on the invariant manifold is approximately neutral
and the dynamics in the normal directions is hyperbolic; hence the name normally
hyperbolic. The system is called r-normally hyperbolic, if the spectral gap con-
dition holds that the tangential dynamics is dominated by a factor r C 1. An r-
NHIM persists under C1 small perturbations of the system. The persistent manifold
will be Cr if the system is, but it may not be more smooth, even if the system is C?
or analytic. This can also be formulated as follows: r-normal hyperbolicity is an
‘open property’ in the space of Cr systems under the C1 topology. The description
above shows that the spectral properties of NHIMs and center manifolds are
similar. The difference is that NHIMs are globally uniquely defined, while center
manifolds are not.
There are two basic methods of proof for hyperbolic fixed points and center
manifolds: Hadamard’s graph transform and Perron’s variation of constants inte-
gral method. Both can be extended to prove persistence of NHIMs, as well as
existence of its stable and unstable manifolds. We employ the Perron method.

vii
viii Preface

Both methods of proof construct a contraction scheme to find the persistent

NHIM (and a similar contraction scheme can be used to find its stable and unstable
manifolds). Heuristically, we can construct the implicit function F(M, v) = Ut
(M) - M = 0, where M is the NHIM and Ut is the flow of the vector field v after
some fixed time t. Normal hyperbolicity of M implies that D1F is invertible.
Hence, there is a function M ~ ¼ Gð~vÞ that maps perturbed vector fields ~v to per-
~
sistent manifolds M, at least in a neighborhood of v. This idea does not work
directly for higher derivatives. An inductive scheme can be set up that typically
uses some form of the fiber contraction theorem. This scheme will break down
after r iterations, hence the limited smoothness. Example 1.1 shows that this is an
intrinsic problem.
To tackle the noncompact case, we replace compactness by uniformity condi-
tions. These include uniform continuity and global boundedness of the vector field
and the invariant manifold and their derivatives up to order r. We require addi-
tional uniformity conditions on the ambient manifold, namely ‘bounded geome-
try’. This means that the Riemannian curvature is globally bounded, and as a result
we have a uniform atlas which allows us to retain uniform estimates throughout all
constructions in the proof.

Organization of the Book

This book is organized as follows. In the introduction, we give a broad overview of

the theory of NHIMs with references to more details in the later chapters. We start
by describing how NHIMs are related to hyperbolic fixed points and center
manifolds. Then we give some basic examples and motivation for studying the
noncompact case. We give a brief overview of the history and literature and
compare the two methods of proof in the basic setting of a hyperbolic fixed point.
Then we continue to introduce the concept of bounded geometry and a precise
statement of the main result of this work and discuss its relation to the literature.
We describe a few extensions and details of the results and conclude the chapter
with notation used throughout this book.
Chapter 2 treats Riemannian manifolds of bounded geometry. We first intro-
duce the definition of bounded geometry and some basic implications. We
explicitly work out the relation between curvature and holonomy in Sect. 2.2. This
we use in Sect. 3.7 to prove the smoothness of the persistent manifold. In the
subsequent sections, we develop the theory required to prove persistence of
noncompact NHIMs in general ambient manifolds of bounded geometry. We
extend results for submanifolds to uniform versions in bounded geometry, to
finally show how to reduce the main theorem to a setting in a trivial bundle. A
number of these results are new and may be of independent interest, namely the
uniform tubular neighborhood theorem, the uniform smooth approximation of a
submanifold, and a uniform embedding into a trivial bundle.
Preface ix

In Chap. 3, we finally prove the main result in the trivial bundle setting. We first
state both this and the general version of the main theorem and discuss these in full
detail. We include a precise comparison with results in the literature, followed by
an outline of the proof. Section 3.3 contains a discussion of the differences to the
compact case and presents detailed examples to illustrate these. Then we start the
actual proof. We first prepare the system: we put it in a suitable form and obtain
estimates for the perturbed system. Then we prove that there exists a unique
persistent invariant manifold and that it is Lipschitz. Second, we set up an elab-
orate scheme in Sect. 3.7 to prove that this manifold is Cr smooth by induction
over the smoothness degree.
In Chap. 4, we discuss how the main result can be extended in a number of
different ways that may specifically be useful for applications. We show how time
and parameter dependence can be added and we present a slightly more general
definition of overflow invariance that might be applicable to systems that are not
overflowing invariant under the standard definition.
Finally, the appendices contain technical and reference material. These are
referenced from the main text where appropriate. Appendix A shows an important
idea that permeates this work: the implicit function theorem allows for explicit
estimates in terms of the input, hence it ‘preserves uniformity estimates’. This can
then directly be applied to dependence of a flow on the vector field. In Appendix
B, the Nemytskii operator is introduced as a technique to prove continuity of post-
composition with a function. This is an essential basic part in the smoothness
proof, together with the results on the exponential growth behavior of higher
derivatives of flows in Appendix C. Here, we also develop a framework to work
with higher derivatives on Riemannian manifolds. The last appendices include the
fiber contraction theorem of Hirsch and Pugh that is used in the smoothness proof,
Alekseev’s nonlinear variation of constants integral defined on manifolds, and a
brief overview of those parts of Riemannian geometry that we use.

What is (Not) New

Normal hyperbolicity can nowadays be called a classical subject; it was first

formulated and studied in the late 1960s and 1970s, see the historical overview in
Sect. 1.3. Although initial results were formulated for compact NHIMs, more
recent work by Sakamoto and especially Bates, Lu, and Zeng have brought this to
the noncompact setting, and even to semi-flows in Banach spaces.
The specific aspect that is new in this work is the differential geometric context
in which our results on noncompact NHIMs are formulated. Uniformity of the
ambient space seems not to have been addressed before in the literature, and our
use of bounded geometry allows us to extend persistence of NHIMs to this setting,
see Sect. 1.5 and Chap. 2. Additionally, some of our results on bounded geometry
appear to be new, including the uniform tubular neighborhood theorem and the
theorem on uniform smooth approximation of a submanifold.
x Preface

The ‘core’ persistence proof itself is based on the Perron method; it can
probably be replaced by the proof of Bates, Lu, and Zeng, that is based on the
graph transform, when taking into account the necessary bounded geometry
technical details from Chap. 2. Our proof uses ideas of Henry and Vander-
bauwhede and Van Gils to extend the Perron method to NHIMs and higher
smoothness. A novel aspect is that we develop these ideas on a trivial bundle with
a bounded geometry manifold as base. This requires a whole framework to be set
up, including representations of higher jets of a flow in Appendix C and formal
tangent bundles of spaces of curves (see Sect. 3.7.4) to study derivatives of the
Perron contraction operator. These ideas might be of interest in other contexts of
dynamics in noncompact differential geometry.
Finally, our Definition 4.4 of a priori overflowing invariance might be new
(although probably not surprising to experts in the field) and could prove useful for
certain applications where the original definition of over- or inflowing invariance
does not hold.

Acknowledgments

This book is based on my Ph.D. thesis work completed at Utrecht University. I am

grateful to having had Hans Duistermaat as my supervisor while working on this
project. His inspiring enthusiasm and deep insights have been greatly beneficial,
and I feel honored to having been his Ph.D. student. After Hans’ untimely death in
2010, the guidance of Erik van den Ban and Heinz Hanßmann has been instru-
mental in completing this project. I am indebted to both, and it was a pleasure
working together.
The members of my thesis committee have provided valuable comments, and I
would especially like to thank Charles Pugh for stimulating discussions and
finding an error in a lemma.
Finally, I would like to thank the editors Henk Broer and Boris Hasselblatt for
their support in preparing this book.

March 2013 Jaap Eldering

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Normally Hyperbolic Invariant Manifolds. . . . . . . . . . . . . . . . . 1
1.1.1 Persistence and (Un)stable Manifolds . . . . . . . . . . . . . . 3
1.1.2 The Relation to Center Manifolds . . . . . . . . . . . . . . . . . 5
1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 The Spectral Gap Condition . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Motivation for Noncompact NHIMs . . . . . . . . . . . . . . . 10
1.3 Historical Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.1 Hadamard’s Graph Transform. . . . . . . . . . . . . . . . . . . . 16
1.4.2 Perron’s Variation of Constants Method . . . . . . . . . . . . 16
1.4.3 Smoothness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Bounded Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Problem Statement and Results . . . . . . . . . . . . . . . . . . . . . . . . 20
1.6.1 Non-Autonomous Systems . . . . . . . . . . . . . . . . . . . . . . 24
1.6.2 Immersed Submanifolds. . . . . . . . . . . . . . . . . . . . . . . . 25
1.6.3 Overflowing Invariant Manifolds . . . . . . . . . . . . . . . . . 29
1.7 Induced Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.8 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2 Manifolds of Bounded Geometry . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.1 Bounded Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2 Curvature and Holonomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3 Submanifolds and Tubular Neighborhoods . . . . . . . . . . . . . . . . 51
2.4 Smoothing of Submanifolds . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.5 Embedding into a Trivial Bundle. . . . . . . . . . . . . . . . . . . . . . . 68
2.6 Reduction of a NHIM to a Trivial Bundle . . . . . . . . . . . . . . . . 73

3 Persistence of Noncompact NHIMs . . . . . . . . . . . . . . . . . . . . . . . . 75

3.1 Statement of the Main Theorems. . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Outline of the Proof. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

xi
xii Contents

3.3 Compactness and Uniformity . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.3.1 Non-Equivalent Metrics . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3.2 Non-Persistence of Embedded NHIMs. . . . . . . . . . . . . . 85
3.3.3 Non-Uniform Geometry of the Ambient Space . . . . . . . . 87
3.4 Preparation of the System. . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.5 Growth Estimates for the Perturbed System . . . . . . . . . . . . . . . 95
3.6 Existence and Lipschitz Regularity . . . . . . . . . . . . . . . . . . . . . 103
3.7 Smoothness. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.7.1 A Scheme to Obtain the First Derivative . . . . . . . . . . . . 111
3.7.2 Candidate Formal Derivatives. . . . . . . . . . . . . . . . . . . . 112
3.7.3 Uniformly Contractive Fiber Maps . . . . . . . . . . . . . . . . 113
3.7.4 Formal Tangent Bundles . . . . . . . . . . . . . . . . . . . . . . . 114
3.7.5 Continuity of the Fiber Maps . . . . . . . . . . . . . . . . . . . . 119
3.7.6 Application of the Fiber Contraction Theorem . . . . . . . . 125
3.7.7 Derivatives on Banach Manifolds . . . . . . . . . . . . . . . . . 128
3.7.8 Conclusion for the First Derivative . . . . . . . . . . . . . . . . 136
3.7.9 Higher Order Derivatives . . . . . . . . . . . . . . . . . . . . . . . 136

4 Extension of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

4.1 Non-Autonomous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2 Smooth Parameter Dependence . . . . . . . . . . . . . . . . . . . . . . . . 142
4.3 Overflowing Invariant Manifolds. . . . . . . . . . . . . . . . . . . . . . . 143
4.4 Full Normal Hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.5 Recovery of the Invariant Fibration and Splitting . . . . . . . . . . . 148

Appendix A: Explicit Estimates in the Implicit Function Theorem . . . 151

Appendix B: The Nemytskii Operator . . . . . . . . . . . . . . . . . . . . . . . . 157

Appendix C: Exponential Growth Estimates. . . . . . . . . . . . . . . . . . . . 161

Appendix D: The Fiber Contraction Theorem . . . . . . . . . . . . . . . . . . 173

Appendix E: Nonlinear Variation of Flows . . . . . . . . . . . . . . . . . . . . . 177

Appendix F: Riemannian Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . 179

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Chapter 1
Introduction

The basics of the theory of hyperbolic dynamics date back to the beginning of the 20th
century, and the general formulation of the theory of normally hyperbolic systems
was stated around 1970. Since then, many people have extended the theory, and even
more people have applied it to problems in all kinds of areas.
Normally hyperbolic invariant manifolds are important fundamental objects in
dynamical systems theory. They are useful in understanding global structures and
can also be used to simplify the description of the dynamics in, for example, slow-fast
or singularly perturbed systems.
In this work, we are specifically interested in noncompact normally hyperbolic
invariant manifolds. We extend classical results that were previously only formulated
for compact manifolds. However, in many applications the manifold is not compact,
so an extension of the theory to the general noncompact case allows one to attack
these problems in their natural context. The main result of this work is an extension
of the theorem on persistence of normally hyperbolic invariant manifolds to a general
noncompact setting in Riemannian manifolds of bounded geometry type.

1.1 Normally Hyperbolic Invariant Manifolds

We should first point out that the theory of (normally) hyperbolic systems can be
applied to both discrete and continuous dynamical systems. That is, if we have a
dynamical system (T, X, ) with X a smooth manifold and  : T × X → X the
evolution function, then the system1 is called discrete if T = Z and continuous if
T = R. In the discrete case, one typically has a diffeomorphism ϕ : X → X and the
full evolution function is defined as (n, x) = ϕ n (x), i.e. iterated application of ϕ.
In the continuous case, the map  is called a flow. It is generated by a vector field v ∈
X(X ) and in that case the map t : X → X is again a diffeomorphism for any t ∈ R.

1 For simplicity of presentation we ignore the facts that  may have a smaller domain of definition,
or that it is a semi-flow or semi-cascade, only defined on T ≥ 0.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 1

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4_1,
© Atlantis Press and the author 2013
2 1 Introduction

The two cases can be related by fixing a t ∈ R in the continuous case and then
view ϕ = t as generating a discrete system. The statements of definitions and
results are (almost) identical if formulated in terms of the evolution function . The
methods of proof share this similarity and can be translated into each other. We shall
adopt the continuous formulation in this work, and refer to the evolution parameter
t ∈ T = R as time. Even though our system is defined in terms of a vector field v,
we call x ∈ X a fixed point of the system when t (x) = x for any t ∈ R. This is
equivalent to saying that v(x) = 0, i.e. that it is a critical point of v; we adhere to
the former terminology to better preserve the analogy with discrete systems.
Before we proceed to explaining normally hyperbolic invariant manifolds, it
should be pointed out that these are a generalization of hyperbolic fixed points.
Many of the characteristic properties generalize as well, so we first sketch the basic
picture for hyperbolic fixed points. Let x be a fixed point of a vector field, v(x) = 0;
it is called hyperbolic if the derivative Dv(x) has no eigenvalues with zero real
part. This means that the eigenvalue spectrum splits into parts left and right of the
imaginary axis, that is, the stable and unstable eigenvalues, but no neutral ones. The
corresponding stable and unstable eigenspaces E ± are both invariant under the linear
flow of Dv(x) and these spaces are characterized by the fact that solution curves on
them converge exponentially fast towards the fixed point under forward or backward
time evolution respectively. It is a well-known result that there are corresponding
stable and unstable (local) manifolds, denoted Wloc S and W U respectively, which
loc
are the nonlinear versions of these, see Fig. 1.1. This situation can be generalized
to a normally hyperbolic invariant manifold by replacing the single fixed point by a
‘fixed set of points’, that is, a manifold which is, as a whole, invariant.
Let us start with a somewhat informal explanation of the concept of a normally
hyperbolic invariant manifold, which we shall from now on often abbreviate as a

Fig. 1.1 A hyperbolic fixed E+

point with (un)stable mani-
S , WU
folds Wloc loc
U
W loc

S
W loc

E−
1.1 Normally Hyperbolic Invariant Manifolds 3

Fig. 1.2 A normally hyper-

bolic invariant manifold. The
single and double arrows
indicate slow and fast flow
respectively

NHIM, as is common in the literature. If we have a dynamical system (T, Q, )

with phase space Q (which we shall often refer to as the ‘ambient manifold’) and
evolution map , then a manifold M ⊂ Q is called invariant under the system if it is
mapped to itself under evolution. In the continuous case this means that t (M) = M
for all times t ∈ R, that is, any point x ∈ M stays in M, so its complete orbit is
contained in M.
An invariant manifold M is then called normally hyperbolic if in the normal
directions, transverse to M, the linearization of the flow t has a spectrum separate
from the imaginary axis again. Although the precise definition is a bit more technical
than in the case of a hyperbolic fixed point, the geometric idea is the same. The normal
directions must separate into directions along which the linearized flow exponentially
converges towards M and directions along which it exponentially expands; no neutral
directions are allowed. Finally, the flow on M itself may expand or contract, but only
at rates that are dominated by the expansion and contraction in the normal directions.
Figure 1.2 shows part of a normally hyperbolic invariant manifold M that has only
stable normal directions. Note that the dynamics on M itself can be very complex;
it can have fixed points or even be chaotic. The only restriction is that the vertical
contraction rate is stronger than horizontal ones (and similarly for expansion), as
is indicated by the double and single arrows and visible from the convergence of
solution curves to the rightmost fixed point on M.

1.1.1 Persistence and (Un)stable Manifolds

There are two important properties that generalize from hyperbolic fixed points to
normally hyperbolic invariant manifolds. These are persistence of the fixed point and
the existence of stable and unstable manifolds. The generalization of these properties
is not a trivial statement nor easily proven in the generalized case of NHIMs, however.
Let us first focus on persistence. In case of a hyperbolic fixed point, this is trivially
stated and proven. If the fixed point x is hyperbolic, then it will persist as a nearby fixed
4 1 Introduction

point under small perturbations of the vector field v and stay hyperbolic. The proof is
a direct application of the implicit function theorem. If Dv(x) has no eigenvalues on
the imaginary axis, then certainly it has no zero eigenvalue, and therefore is a bijective
linear map. So a slightly perturbed vector field ṽ will again have a fixed point x̃ nearby
x and the eigenvalues of Dṽ(x̃) will be close to those of Dv(x) if ṽ −v is small in C 1 -
norm. Hence the eigenvalues are still separated byAZ the imaginary axis. For a NHIM
the situation is similar but technically much more involved due to the fact that there
is no control on the behavior of solution curves in the invariant manifold. A normally
hyperbolic manifold M does persist under C 1 small perturbations and the perturbed
manifold M̃ is again normally hyperbolic and close to M in a precise way. The most
important difference, however, is that M̃ generally has only limited smoothness, even
if M and the system were smooth or analytic.2 This smoothness is dictated by the
spectral gap condition, which is roughly the ratio between the normal exponential
expansion/contraction and the exponential expansion/contraction tangential to M.
This fact already indicates that the proof of persistence of a NHIM cannot be a
straightforward application of the implicit function theorem.
The stable and unstable manifolds generalize as well. That is, a normally hyper-
bolic invariant manifold M has stable and unstable manifolds W S (M) and W U (M)
such that solution curves on these converge exponentially fast towards M in forward
or backward time, respectively. Their intersection is precisely M. But there is actually
more structure: these manifolds—we consider W S (M) but everything is equivalent
for W U (M)—are fibrations of families of stable and unstable fibers to each point
m ∈ M, 
W S (M) = W S (m).
m∈M

We should be a bit careful with this last statement, as points m ∈ M are generally not
fixed points. These fibers W S (m) are invariant in the sense that the flow commutes
with the fiber projection π S :

∀ t ∈ R, m ∈ M, x ∈ W S (m) : π S ◦ t (x) = t ◦ π S (x).

In other words, each fiber is mapped into another single fiber under the flow, namely
the fiber over the flow-out of the base point m. This important fact means that if
we use the fibration for local coordinates, then in these coordinates the horizontal,
base flow decouples from the vertical, fiber flow. This is sometimes also called
an isochronous fibration [Guc75] as all points in a fiber have the same long-term
behavior. Each single fiber is as smooth as the system, but the dependence on the
base point m, and thus the smoothness of the fibrations as a whole, is generally not
better than continuous, see Fenichel [Fen74, Sect. I.G]. We do not investigate these
invariant fibrations in the present work, although the mentioned results should hold
for noncompact NHIMs as well.

2 I do not know whether loss of smoothness is generic for NHIMs. See [Has94, HW99] for the
case of Anosov systems.
1.1 Normally Hyperbolic Invariant Manifolds 5

1.1.2 The Relation to Center Manifolds

Normally hyperbolic invariant manifolds bear a close resemblance to center man-

ifolds. Their spectral properties are roughly equivalent; they differ in the fact that
NHIMs have an intrinsically global definition, while center manifolds are defined in
local terms.
A center manifold Wloc C (x) of a fixed point x is a local invariant manifold such

that its tangent space at the fixed point is the (generalized) eigenspace E 0 of the
eigenvalues with real part zero, that is,
C
Tx Wloc (x) = E 0 . (1.1)

We can extend the definition of center manifold a bit by including all eigenvalues λ
with real part bounded by |Re(λ)| ≤ ρ0 . An associated generalized center manifold
consists of solutions that converge or diverge from x at an exponential rate bounded
by ρ0 . Curves in the strongly3 stable or unstable manifold converge or diverge at
exponential rates larger than ±ρ0 , respectively. These conditions can directly be
compared to the description of NHIMs above, or Definition 1.8 (with ρ0 = ρ M ).
If we take a look at Fig. 1.2 again, then we see that both fixed points (indicated
with a dot) on M have (generalized) center manifolds; M itself is a center manifold
for these, but for the rightmost fixed point we can actually construct the center man-
ifold from any two solution curves converging to that fixed point from the left and
right. For example, the union of the two curves drawn in the figure that converge to it
could be taken as alternative center manifold. This reflects the well-known fact that
center manifolds are generally not unique. This is the main difference with the case
of NHIMs: center manifolds are only defined in terms of growth rates of solution
curves locally with respect to one fixed point, while NHIMs are globally invariant
objects, where the spectral splitting must hold everywhere along the invariant mani-
fold. This difference is effectively the reason that center manifolds are not uniquely
defined, while the perturbations of NHIMs are, see below. If we perturb the system
in Fig. 1.2 a bit, then the persistent NHIM must everywhere be close to the original
invariant manifold M. This enforces uniqueness; in Fig. 1.2 this is clearly visible:
the alternative choice of center manifold to the rightmost fixed point diverges far
from M. See also the example in Section 1.2.1.
There is a subtle question of smoothness both for center manifolds and NHIMs,
related to the spectral gap condition (1.11). Center manifolds are arbitrarily smooth
in a sufficiently small neighborhood of the fixed point x, but they are generally
not C ∞ , even though they satisfy an infinite spectral gap. See Van Strien’s short
note [Str79]. The reason is that the size of the neighborhood may depend on the
degree of differentiability C k . Persistent NHIMs generally have bounded smoothness
due a finite spectral gap; but even if they have an infinite spectral gap, the smoothness

3 We remove the eigenvalues associated to E 0 from E ± so that E − , E 0 , E + together disjointly

span the total tangent space at x.
6 1 Introduction

of a persistent NHIM is (generally) not C ∞ for the same reasons. See Remark 1.12
and Example 1.3.

1.2 Examples

We present a few examples. The first detailed example serves to show explicitly that
smoothness of a persistent manifold depends crucially on the spectral gap condition.
The next examples motivate the usefulness of a noncompact version of the theory of
normal hyperbolicity.

1.2.1 The Spectral Gap Condition

An invariant manifold is called an r -NHIM if the flow contracts or expands at expo-

nential rates along the normal directions, and if these rates dominate any contrac-
tion or expansion along tangential directions at least by a factor r . This separation
between growth rates along directions tangential and normal to the NHIM is encoded
in Eqs. (1.10) and (1.11).
Here we introduce a simple example where the growth rates can be identified with
eigenvalues λ of the linearization of the vector field at stationary points. Furthermore,
we consider the simplified case where only a stable normal direction is present. That
is, we consider a flow that contracts in the normal direction at an exponential rate of
at least ρY < 0 and along the invariant manifold it contracts at most at the rate ρ X
with the simplified spectral gap condition

ρY < r ρ X with ρ X ≤ 0, r ≥ 1. (1.2)

The spectral gap is fundamental to persistence of invariant manifolds: the compact

invariant manifolds that are persistent under any small perturbation are precisely those
that are normally hyperbolic4 [Mañ78]. Mañé only proved this inverse implication
for 1-normal hyperbolicity, the question is still open for r -normal hyperbolicity
with r > 1. A further property of normally hyperbolic invariant manifolds is that the
differentiability of a slightly perturbed manifold depends not only on the smoothness
of the original manifold and the perturbed vector fields, but also on the spectral gap.
The spectral gap determines an upper bound 1 ≤ r < ∞ on the smoothness of the
perturbed system, as r has to satisfy5 (1.11). This condition stems from the fact that

4 The definition of normal hyperbolicity in [Mañ78] is a bit more general than the definition in
this paper. That definition only requires a growth ratio r ≥ 1 along solution curves in the invariant
manifold, and not as a ratio of global growth rates ρ X , ρY , see also Remark 1.10.
5 The case r = ∞ would require ρ > 0; when ρ = 0, any finite order r can be obtained, but
X X
only for perturbations sufficiently small depending on r .
1.2 Examples 7

when the flow has exponential growth behavior eρt , then higher order derivatives will
generally have growth behavior ekρt and the interval inclusion [k ρ, ρ] ⊂ (ρY , ρ X ) is
required to show existence and uniqueness of the k–th derivatives via a contraction.
The optimal differentiability degree r can be extended to a real number by viewing
α– Hölder continuity as a fractional differentiability degree. That means that the
perturbed manifold can be shown to be C k,α when r = k + α satisfies the spectral
gap condition and the system is C k,α to start with.
The following example shows that this result is sharp. We construct a very simple
compact, normally hyperbolic invariant manifold, and then show that an arbitrarily
small perturbation yields a unique perturbed C k,α invariant manifold, where r = k+α
satisfies ρY = rρ X . This in fact precisely violates the spectral gap condition, since
that requires a strict inequality. The example could be adapted to obtain a perturbed

manifold with smoothness no better than C r for some r

< r , cf. Example 1.3.
A more qualitative exposition of this example can also be found in [Fen72, pp.
198–200] and [Hal69, p. 239, 251].
Example 1.1 (Optimal C k,α smoothness of persistent manifolds) Let the horizontal
space X = S 1 be the circle and the vertical space Y = R. Take two points x− = 0
and x+ = π in X and set the vector field v to zero at (x− , 0), (x+ , 0). We turn
these stationary points into hyperbolic fixed points, with v linear in neighborhoods
around them and Dv(x− , 0), Dv(x+ , 0) having eigenvalues λ− < 0 < λ+ along X ,
respectively, and one global eigenvalue λY < λ− in the vertical direction along Y ,
i.e. ẏ = v y (x, y) = λY y, see Fig. 1.3. We extend the horizontal component vx of the
vector field v to the whole space X × Y in such a way that it is C ∞ , independent of
y, and has no critical points except for x− , x+ . Hence, M = X × {0} is an invariant
manifold for the flow t of v.
First, we check that M is normally hyperbolic. The long time behavior of any
point m ∈ M is governed by its approach of the stable fixed point (x− , 0), except for
m = (x+ , 0). For m = (x+ , 0) we have Dt (m) = eDv(m) t , hence

λY λY

X
λ− x− χ λ+ x+

Fig. 1.3 An example invariant manifold exhibiting C k,α smoothness under perturbation
8 1 Introduction

Dt (m)|Tm X = eλ+ t ,

Dt (m)|Tm Y = eλY t .

More generally, consider a point m ∈ M in the neighborhood of either (x± , 0)

where v is linear. Then Dt (m) is given by
 
eλ± t 0
D (m) =
t
0 e λY t

for as long as t (m) stays in that neighborhood of (x± , 0) where the vector field is
linear. The transition time between these two neighborhoods is finite as v does not
have zeros and the transition map preserves vertical lines {x} × Y . The latter fact
is because vx is independent of y, that is, we have also found the invariant, foliated
stable manifold of M. Gluing together these Dt maps on the different domains,
we see that the resulting tangent flow splits again into independent horizontal and
vertical parts, which can be estimated by
 
∀t ≤ 0 : Dt (m)|Tm X  ≤ C X eλ− t ,
 
∀t ≥ 0 : Dt (m)|T Y  ≤ CY eλY t ,
m

where the constants C X , CY are determined by the flow t in the domain where v is
nonlinear. For any point m close to (x− , 0) this estimate is sharp, hence we expect
maximal smoothness r = λY /λ− for a generic perturbation.
Next, we add a perturbation term ε χ to the vector field v, so we have a perturbed
vector field ṽ = v + ε χ , where χ ∈ C0∞ is chosen with support on a small ball
intersecting M away from the fixed points and pointing upward. This will ‘lift’ the
invariant manifold as indicated in Fig. 1.3 for any ε > 0. Let M̃ denote this lifted
manifold, that is, M̃ is the image of the two heteroclinic solution curves that run from
(x+ , 0) to (x− , 0) together with these fixed points. The solution curve that runs to
the left is lifted up from the x– axis after entering the region supp χ .
We first investigate two claims: that M̃ is invariant and that it is the unique invariant
manifold that is close to M. The invariance is obvious; to the right of x+ nothing has
changed, so there M̃ = M. To the left of x+ we follow the original unstable manifold,
get pushed up within the domain of support of χ and after leaving that domain and
entering the linear flow around (x− , 0) we follow a standard curve ending at (x− , 0).
This is a solution curve of ṽ ∈ C ∞ , hence invariant and even smooth. Now assume
there exists another invariant manifold M
nearby and let (x
, y
) ∈ M
\ M̃. The
backward orbit of the point (x
, y
) must diverge to |y|  1. If x
= x− , then y
= 0
and this is clear. If x
= x− , then the backward orbit will end up at a point (x, y) with
x close to x+ and y = 0; since we are in the linear domain of (x+ , 0), this orbit will
then diverge (in reverse time) along the stable manifold towards |y|  1. Hence, M

is not close to M.
Next, we show that (for any ε > 0) the perturbed manifold M̃ is not more than
C k,α with k + α = λY /λ− , even though the original and perturbed systems are
1.2 Examples 9

C ∞ -smooth. To the left of (x− , 0), M̃ is given by the graph of the zero function
from X to Y (as the continuation from (x+ , 0) to the right along X = S 1 ). To the
right of (x− , 0), the solution curve is given by (x, y)(t) = (x0 eλ− t , y0 eλY t ), hence
y = C x λY /λ− where C depends on x0 , y0 only. So we can write M̃ as the graph of
the function 
0 if x ≤ 0,
h̃ : X → Y : x → λ /λ
Cx Y − if x > 0.

This function is exactly C k,α for k +α = r in x = 0. Note that the loss of smoothness
appears at a different place than the perturbation of the vector field. The relevant fact
is that the different solution curves approaching the stable limit point have finite
differentiability with respect to each other, and this depends on the horizontal and
vertical rates of attraction at (x− , 0). 
If we had assumed that ρY = ρ X , that is, r = 1, but with a non-strict inequality
ρY ≤ r ρ X , then normal hyperbolicity precisely fails and the invariant manifold
indeed need not persist. By the arguments above it can already be seen that the
persistent manifold can lose differentiability: when r = 1, the graph of the manifold
will be given by 
0 if x ≤ 0,
h̃(x) =
C x if x > 0,

which is clearly non-differentiable at x = x− = 0. We can extend the example above

to show that even more serious problems can occur.
Example 1.2 (Non-persistence of non-NHIMs) We consider Example 1.1 with ρY =
ρ X . If we perturb the system with a small circular vector field around x− = (0, 0),
then Dv(0, 0) will have two eigenvalues λY ± i ω with λY < 0 and ω ∈ R small.
Thus, the solution curves that should make up the invariant manifold around (0, 0)
will spiral in, which leads to the picture in Fig. 1.4. Note that the curves wind around
the origin infinitely often. At the origin this is not a manifold anymore, and cannot
be described by a function h̃ : X → Y . 
The idea to perturb around the stable fixed point x− also leads to the following
example.
Example 1.3 (Non-C ∞ persistence for r = ∞ NHIMs) We consider again
Example 1.1, but now with λ− = 0. Then we have ρ X = 0 and spectral gap r = ∞.
If we let λ− = ε depend on the perturbation parameter ε > 0, then this decreases
the spectral gap condition6 to a finite number r = λY /λ− . Even though r → ∞ as
the perturbation size ε goes to zero, we still have a finite spectral gap for any fixed
perturbation. We conclude that the corresponding perturbed manifolds are not C ∞ ,

6 The ratio r in the spectral gap is defined by a strict inequality, which we ignore here for simplicity
of presentation.
10 1 Introduction

Fig. 1.4 Breakdown of a Y

non-NHIM under a circular
perturbation

but have smoothness C r where r can be made arbitrarily large by decreasing the
perturbation size. 

1.2.2 Motivation for Noncompact NHIMs

Most of the literature on normal hyperbolicity and its applications treat com-
pact NHIMs only. This excludes possibly interesting applications. Settings where
a noncompact, general geometric version of normal hyperbolicity may be useful
include chemical reaction dynamics [Uze+02] and problems in classical and celes-
tial mechanics [DLS06].
We describe three examples where noncompactness naturally comes into play.
The first example, a normally attracting cylinder, is set in Euclidean space. This
example could be complicated a bit more by adding normal expanding directions
to get a fully normally hyperbolic system. Such situations show up in Hamiltonian
or reversible systems with invariant tori [Bro+09]. The second and third examples
are set in ambient manifolds with nontrivial topology, thus motivating the need for a
theory of noncompact NHIMs in such a geometric setting. The third example shows
an application to classical mechanics and actually motivated this work.
Let us first treat a simple example.
Example 1.4 (A normally attractive cylinder) Let us consider the infinite cylinder
y 2 + z 2 = 1 in R3 . If we define a very simple dynamics by

(ẋ, ṙ , θ̇ ) = (0, r (1 − r ), 1)

in cylindrical coordinates, then the cylinder is normally attractive and the motion on
the cylinder consists of only periodic orbits, see Fig. 1.5.
1.2 Examples 11

Fig. 1.5 A normally attract- z

ing cylinder

The dynamics on the cylinder is completely neutral, while it attracts in the normal
direction with rate −1. Hence, there exists a unique persistent manifold diffeomorphic
and close to the original cylinder. For any k ≥ 1, the persistent manifold has C k
smoothness if the perturbation is chosen sufficiently small. The perturbed manifold
must be uniformly close to the original cylinder; this rules out Example 3.9 of a
cylinder with exponentially shrinking radius.
The dynamics on the persistent manifold can be perturbed in arbitrary ways. It
could slowly spiral towards x– infinity, or develop attracting and repelling periodic
orbits on the cylinder. If the cylinder were higher dimensional, it could even become
chaotic. 
The previous example is set in Euclidean space, even though the invariant manifold
has not completely trivial topology. Let us next consider a case where the ambient
space is a manifold of nontrivial topology.
Example 1.5 (A pendulum with time-dependent perturbation) Consider a classical
pendulum described by its angle θ ∈ S 1 and angular velocity ω. The unstable top
position θ = π is a hyperbolic fixed point, hence a special case of a NHIM.
The phase space TS 1 is noncompact and nontrivial. Without perturbations we
could restrict ourselves to a compact energy surface; if we add a general time-
dependent perturbation, however, then energy is not preserved anymore and the full
phase space must be considered. We should include the noncompact time interval R
as well, leading to the full phase space TS 1 × R. Then the unperturbed unstable top
position corresponds to the one-dimensional NHIM

M = {(θ, ω, t) ∈ TS 1 × R | θ = π, ω = 0}. (1.3)

If the perturbation is globally sufficiently small in C 1 norm, then M uniquely persists

into a manifold M̃ nearby. This means that under any small perturbations there exists
a unique orbit of the pendulum that balances closely around the unstable top position
for all time t ∈ R. Note that this orbit stays uniformly close to the top position and
that the perturbation need not depend (quasi-)periodically on time. 
12 1 Introduction

Remark 1.6 The nontrivial topology of S 1 can easily be undone by modeling the
system on its universal cover R. The example can be changed to a spherical pendu-
lum modeled on S 2 ; this indicates more clearly the usefulness of having the theory
available in spaces with nontrivial topology, although in this case we still only need
to study the problem in a small neighborhood of the top position. It was pointed out
to me by Robert MacKay that TS 2 is still a trivializable tangent bundle, and that
more complex, non-trivializable examples including higher dimensional spheres can
be constructed in the setting of chemical reaction dynamics. ♦
The last example actually motivated this work.
Example 1.7 (Nonholonomic systems as singular perturbation limit) Let a classical
mechanical system be given by a smooth Riemannian manifold (Q, g) as configura-
tion space and a Lagrangian L : TQ → R. The vector field v on TQ is determined
by the Lagrange equations of motion, given in local coordinates by

 i d ∂L ∂L
L = − = 0. (1.4)
dt ∂ ẋi ∂ xi

A nonholonomic constraint can be placed on such a system by specifying a dis-

tribution7 D ⊂ TQ and adding reaction forces to [L] according to the Lagrange–
d’Alembert principle, that is, we require that a solution curve γ satisfies
 
L (γ )(t) ∈ D0 and γ̇ (t) ∈ D for all t ∈ R (1.5)

where D0 ⊂ T∗ Q denotes the annihilator of D. This means that we restrict the

velocities—but not the positions—of the system and adapt the vector field such that
it preserves D. Such constraints are called ‘nonholonomic’ if the distribution D is
not integrable. This means that some small positional changes can only be obtained
through long orbits due to the constraints. The prototypical example is that parallel
parking a car a small distance sideways requires repeated turning and moving forward
and backward.
As a concrete example of a nonholonomic system, let us consider a ball rolling on
a flat surface. The possible positions of the ball are specified by Q = S O(3) × R2 ,
i.e. orientation and position in the plane. If we enforce the constraint that the ball can
only roll and not slip, then its linear velocity is determined by its angular velocity
ω ∈ so(3), thus we have

so(3) × S O(3) × R2 ∼
= D ⊂ T S O(3) × R2 .

The addition of the nonholonomic reaction forces specified by the Lagrange–

d’Alembert principle can be argued for on physical grounds, and some experimental
verification has been done by Lewis and Murray [LM95] to check its correctness

7 Here, a distribution is meant in the sense of differential geometry as a subbundle of the tangent

bundle, not a generalized function (nor a probability distribution).

1.2 Examples 13

against the alternative vakonomic principle. Still, it would be nice to rigorously

derive these forces from fundamental principles; this would complement [RU57,
Tak80, KN90] which showed this for holonomic constraints. The nonholonomically
constrained system can be obtained from the unconstrained system by adding friction
forces, see [Kar81, Bre81, Koz92]. Heuristically, one could say that if a rolling ball
feels a strong contact friction force, then if this force is taken to infinity, it suppresses
all slipping. This can be viewed as a singular perturbation limit, where D precisely
is the invariant manifold, and it is normally attracting due to the dissipative friction
force.
The cited works prove this result, but only asymptotically on finite time intervals.
The extension of the theory of NHIMs to noncompact manifolds as developed in this
work can be applied here. It allows one to improve upon this result and make it exact
on infinite time intervals and general noncompact configuration spaces Q, as long as
these satisfy the ‘bounded geometry’ condition. One could think, for example, of a
gently sloping surface and a ball that is not perfectly round, or even a time-dependent
perturbation, as long as it is uniformly bounded in time. 

1.3 Historical Overview

As already mentioned, the theory of normally hyperbolic invariant manifolds is a

generalization of the theory of hyperbolic fixed points. The study of these dates
back to the beginning of the 20th century, or even the end of the 19th century.
From 1892 onwards, Poincaré published his works “Les méthodes nouvelles de la
mécanique céleste” [Poi92], in which he founded the theory of dynamical systems
and famously studied the three-body problem. This triggered further research in
nonlinear dynamical systems and persistence questions. Another important work
published in the same year is “The general problem of the stability of motion”
by Lyapunov; the original is in Russian, but translations in French [Lya07] and
English [Lya92] are available. In this work, he introduced the concept of characteristic
numbers, nowadays called ‘Lyapunov exponents’, to study ‘conditional stability’ of
nonlinear differential equations at a fixed point. Conditional stability corresponds to
the existence of stable (and unstable) linearized directions and Lyapunov proves the
existence of a stable manifold by means of a series expansion under the assumption
that the system is analytic.
In the beginning of the 20th century, the problem of stable manifolds was stud-
ied, without assuming analyticity, by Hadamard [Had01] and Cotton [Cot11]. Both
Frenchmen applied different methods to obtain the stable and unstable manifolds of
a hyperbolic fixed point. Later, the German mathematician Perron extended the ideas
of Cotton to allow for generic complex eigenvalues, possibly of higher multiplicity,
as long as the real parts of the eigenvalues are separated by zero (or even a number
r = 0), see [Per29; Per30]. Hadamard’s method is now named after him, and also
known as the ‘graph transform’. The other method was first formulated by Cotton,
although the idea of exponential growth of solution curves can be traced to Lyapunov.
14 1 Introduction

This method is commonly referred to as the Perron or Lyapunov–Perron method

in the literature. This seems to pay too little credit to Cotton, even though Perron
himself [Per29] does attribute the method to Cotton.8
From around 1960, renewed activity in the area of hyperbolic dynamics led to the
generalization of the theory of (un)stable manifolds for hyperbolic fixed points to per-
sistence and (un)stable fibrations for normally hyperbolic invariant manifolds. Many
authors have contributed to this subject, culminating in the seventies in the works by
Fenichel [Fen72] and Hirsch, Pugh, and Shub [HPS77]. These two works formulate
the theory slightly differently, but in broad generality and can be viewed as the basic
references nowadays; references to earlier works can be found in both. Both Fenichel
and Hirsch, Pugh, and Shub use Hadamard’s graph transform as their fundamental
tool. In these works, compactness of the invariant manifold is a basic assumption.
Noncompact, immersed manifolds are considered in Ref. [HPS77, Sect. 6], albeit
under the assumption that the immersion image is compact again.
The theory of normal hyperbolicity has seen some interesting developments
since these foundational works, and the applications have slowly started to flour-
ish, see [Wig94] for a list of subjects. A major development was the generalization
to semi-flows in Banach spaces. This situation can arise when one wants to study
partial differential equations as ordinary differential equations on appropriate func-
tion spaces. This technique has been applied to PDEs such as the Navier–Stokes or
reaction-diffusion equations.
In his book on parabolic PDEs, Henry extended the Perron method to apply to
semi-flows with a NHIM given as the horizontal submanifold9 X × {0} in a product
X × Y of Banach spaces [Hen81, Chap. 9]. Henry’s idea is to linearize only the
normal directions, but keep the horizontal flow along M in its general, nonlinear
form, while at the same time splitting the Perron contraction map into a two-stage
contraction map on horizontal and vertical curves separately. Henry obtains C 1,α
smoothness only. In the series of papers [BLZ98; BLZ99; BLZ08], Bates, Lu, and
Zeng study more general NHIMs of semi-flows in Banach spaces. They employ
Hadamard’s graph transform and allow so-called ‘overflowing invariant manifolds’,
as in [Fen72]. They also allow the NHIM to be noncompact and an immersed instead
of an embedded submanifold. In [BLZ99] the unperturbed NHIM is assumed to be C 2
to obtain C 1 persistence results, for the technical reason of constructing C 1 normal
bundle coordinates. In their later paper [BLZ08], this technicality is overcome,10
and existence of a NHIM is even proven when sufficiently close, approximately
normally hyperbolic invariant manifolds exist; the persistence result is then obtained
for compact NHIMs only, to circumvent the C 2 assumption.
Vanderbauwhede and Van Gils [VG87; Van89] introduced the technique of con-
sidering a scale (family) of Banach spaces of curves with exponential growth, and

8 These facts were pointed out to me by Duistermaat.

9 Henry actually has reversed notation where the ‘vertical’ manifold Y × {0} is the NHIM.
10 Their Hypothesis (H2) that a certain approximate splitting like (1.9) “does not twist too much”,

can be obtained from uniform Lipschitz continuity of the tangent spaces of the invariant manifold.
I am not sure if this is a significantly weaker hypothesis. See also the discussion in Remark 3.13.
1.3 Historical Overview 15

using the fiber contraction theorem (see Appendix D), proved smoothness of center
manifolds with the Perron method. Although not the same, center manifolds have
many properties in common with NHIMs and Sakamoto [Sak90] has built upon the
works of Henry and Vanderbauwhede and Van Gils to prove persistence and C k−1
smoothness for singularly perturbed systems in a finite-dimensional Rm ×Rn product
space setting. The loss of one degree of smoothness is again due to the construction
of normal bundle coordinates, although this fact is obscured by the explicit Rm × Rn
setting.
Singularly perturbed, or, slow-fast systems are another important class of applica-
tions. These describe systems where the dynamics is governed by multiple, separate
time scales, or when a system can be viewed as an approximation of an idealized,
restricted system. Singularly perturbed systems can be studied using the theory of
normal hyperbolicity by turning them into a regular perturbation problem via a rescal-
ing of time, see foundational work by Fenichel [Fen79] or the more introductory
expositions [Jon95; Kap99; Ver05].

1.4 Comparison of Methods

There are two well-known methods for proving the existence and smoothness of
invariant manifolds in hyperbolic-type dynamical systems. The Hadamard graph
transform and the variation of constants method, also known as the (Lyapunov–)
Perron method. Variations of both have been applied in many situations with some
form of hyperbolic dynamics. This ranges from the relatively simple problem of
finding the stable and unstable manifolds of a hyperbolic fixed point, to center man-
ifolds, partially hyperbolic systems, and normally hyperbolic systems. The quote
of Anosov [Ano69, p. 23] that “every five years or so, if not more often, someone
‘discovers’ the theorem of Hadamard and Perron, proving it either by Hadamard’s
method of proof or by Perron’s” is nowadays probably familiar to many researchers
in these areas; it illustrates the pervasiveness of these methods.
In this section, I describe the ideas that are common to both methods, as well as
their differences. I hope to elucidate the merits and weak points of both methods,
especially when applied to normally hyperbolic systems. Basically they seem to be
able to produce the same conclusions, but each method takes a different viewpoint
to the problem.
Let us first identify some basic common ideas. As a sample problem, we consider
finding the invariant unstable manifold W U of a hyperbolic fixed point, positioned at
the origin of Rn . The system is defined by either a diffeomorphism  in the discrete
case, or a flow t in the continuous case. Both methods use the splitting of the
tangent space into stable and unstable directions:

T0 Rn ∼
= Rn = U ⊕ S.
16 1 Introduction

Let (x + , x − ) denote coordinates in U ⊕ S according to projections π + , π − from

Rn onto the unstable and stable directions U and S, respectively. We shall use the
notation ± = π ± ◦ .

1.4.1 Hadamard’s Graph Transform

The graph transform is due to Hadamard. His paper [Had01] (in French, 4 pages) can
be used as a concise and basic introduction to the graph transform, applied to the stable
and unstable manifolds of a hyperbolic fixed point. He does not prove smoothness
or even continuity of these invariant manifolds, although continuity could easily be
concluded by introducing the Banach space of bounded continuous functions with
supremum norm.
The basic idea of the graph transform is to view the unstable manifold W U as
the graph of a function g : U → S. The graph, as a set, is invariant under  (or e.g.
1 in the continuous case). The diffeomorphism  can also be interpreted as a map
acting on functions g through its action on their graphs. This induces a mapping

T : g → g̃ implicitly defined by g̃ + (x, g(x)) = − (x, g(x)). (1.6)

Thus, by definition, any point (x, g(x)) on the graph of g gets mapped to a point
(x
, g̃(x
)) on Graph(g̃). The map T turns out to be well-defined and a contraction
on functions U → S that are sufficiently small in Lipschitz norm. The graph of
the unique fixed point g  of T must correspond to the unstable manifold, that is,
W U = Graph(g  ).
By considering the invariant sets, this method focuses on the geometry of the
problem. The method uses a diffeomorphism map ; the continuous case can be
studied by considering the flow map t for a fixed time t. The diffeomorphism can
easily be studied locally in charts on a manifold. Therefore this method lends itself
well to the generalized setting of normally hyperbolic invariant manifolds, where
the invariant manifold is intrinsically a global object. Even if this global object is
nontrivial, it can still be studied in local charts.

1.4.2 Perron’s Variation of Constants Method

This method is commonly referred to as the Perron or Lyapunov–Perron method.

Although in the literature this is attributed to Perron [Per29], he in turn cites
Cotton [Cot11] for the main idea.
This method focuses on the behavior of solution curves. The solutions on the
unstable manifold are precisely characterized by the fact that they stay bounded
under backward evolution. In the following, we explain the Perron method for the
1.4 Comparison of Methods 17

continuous case.11 We adopt the notation from the graph transform setting. A con-
traction operator T is constructed via a variation of constants integral. The non-
linear part of the vector field is viewed as a perturbation of the linear part. The
integral equation is split into the components along the stable and unstable direc-
tions. Then the integration of the unstable component is switched from the interval
[0, t] to [−∞, t], and only bounded functions are considered. Writing the vector
field v(x) = Dv(0) · x + f (x) in linearized form with nonlinearity f , this leads to
the following contraction operator on curves x = (x + , x − ) ∈ C 0 ([−∞, 0]; Rn ):
 

0 −

+ −
T : x (t), x (t) → x0+ − Dt−τ +
+ (0) f + x (τ ), x (τ ) dτ ,
t
 t − +


Dt−τ
− (0) f − x (τ ), x (τ ) dτ . (1.7)
−∞

This mapping T is well-defined and a contraction on curves x ∈ C 0 ([−∞, 0]; Rn )

whose stable component x − is bounded and sufficiently small. Note that T does not
depend on the stable component x0− of the initial conditions anymore. The fixed point
of T is a solution curve on W U with x0+ given as a parameter. The unstable manifold
is described, finally, by evaluating the stable component at zero, leading to a graph

g : U → S : x0+ → x − (0).

First of all, it must be noted that this method requires f to be small in C 1 -norm. We
can make f small by restricting to a sufficiently small neighborhood of the origin
and cutting off f outside of it. This cut-off does not influence the results: due to
the boundedness condition, curves x stay in the neighborhood. The method can be
generalized to a separation of stable and unstable spectra (i.e. a dichotomy) away
from the imaginary axis,12 and for example be applied to show existence of center
manifolds. In that case, uniqueness is lost as solutions will generally run out of small
neighborhoods. This makes the Perron method not directly applicable to normally
hyperbolic invariant manifolds. The center direction corresponds to the invariant
manifold, but solution curves are global objects that cannot be treated locally.
The Perron method can be extended to overcome this problem. Henry [Hen81,
Chap. 9] linearizes the vector field only in the normal directions of the invariant
manifold. Henry uses a two-step contraction scheme, but this can be reduced to a
single contraction T = T− ◦ T+ that is a composition of two maps. The maps T± are

11 Contrary to the graph transform (which is only intrinsically defined for mappings), the Perron
method can be formulated both for flows and discrete mappings. For the discrete case, the integral
must be replaced by a sum, the mapping  must be split into a linear and nonlinear part, and the
linearized flow must be replaced by iterates of the linearized mapping. See for example [APS02;
PS04].
12 This is for the continuous case. The imaginary axis of the spectrum of a vector field corresponds

(via the exponential map) to the unit circle for the spectrum of a diffeomorphism in the discrete
case.
18 1 Introduction

essentially the components of (1.7). Still, the results obtained are not quite as general
as those obtained with the graph transform. For the graph transform, the condition
of normal hyperbolicity can be formulated in terms of the ratio of the normal and
tangential growth rates of the flow along orbits, while for the Perron method it must be
formulated in terms of the ratio of global growth rates. This less general assumption
is required because the contraction operator (1.7) is studied on spaces of solution
curves with a fixed exponential growth behavior, see Definition 1.17.
Explicit time dependence can be added to the Perron method with only trivial
modifications. This allows one to study hyperbolic fixed points in non-autonomous
systems.13 An application is the study of invariant fibrations of, for example, normally
hyperbolic invariant manifolds. These have fibered stable and unstable manifolds.
Points in a single fiber are characterized by the unique orbit on the normally hyper-
bolic invariant manifold they are exponentially attracted to under forward or back-
ward evolution, respectively. Finding these fibers is turned into a non-autonomous
hyperbolic fixed point problem by following a point on the invariant manifold.

1.4.3 Smoothness

In the truly hyperbolic case—when the stable and unstable spectra are separated
by a neighborhood of the imaginary axis—the Perron method allows for a direct
proof of smoothness of the manifolds W U and W S , see [Irw70; Irw72] where this
is formulated for discrete systems. One first verifies that the contraction operator T
is as smooth as the system, still acting on continuous curves x. Then, by an implicit
function theorem argument, the fixed point depends smoothly on the (partial) initial
value parameter x0+ . To the best of my knowledge, there is no similarly simple
approach for the graph transform. The contraction map acts directly on graphs g, so
to obtain smoothness, one must consider the maps g ∈ C k (U ; S). A direct estimate
of contractivity in C k -norm requires higher than k-th order Lipschitz estimates on
the system.
When the spectra are not separated by the imaginary axis—this occurs for example
in normally hyperbolic systems—things become more complicated. The spectral gap
condition defines an intrinsic upper bound for the smoothness that one can generically
expect for a system, as was seen in Example 1.1. Both methods apply induction over
the smoothness degree in their proof. Formal derivatives of the contraction map T
are constructed. These are again contractions, but now on higher derivatives of the
fixed point mapping, while fixing the derivatives below. Finally, the fiber contraction
theorem (see Appendix D) can be used to conclude that these higher order derivatives
converge to a fixed point, jointly with all lower orders.

13 The term ‘fixed point’ in the context of a non-autonomous system is not definable in a coordinate-

free way: any orbit of the system can be made into a fixed point under a suitable time-dependent
coordinate transformation. However, there may be a preferred “time-independent” coordinate sys-
tem. Moreover, the hyperbolicity of an orbit with respect an intrinsic metric is independent of a
choice of coordinates.
1.4 Comparison of Methods 19

Explicit calculation of higher derivatives of T is very tedious; one should focus

on their form as dictated by Proposition C.3. For the graph transform, the relevant
terms that one obtains from (1.6) are, ignoring arguments,

k
Dk g̃ · D1 + + D2 + Dg + · · · = D2 − · Dk g + · · ·
 k
 
This leads to a contraction when D2 − ·D1 −1 +  < 1. The limit on k precisely
corresponds to the spectral gap condition, at least when we replace  by a sufficiently
high iterate  N of itself, or in the continuous case, if we take the flow map t at a
sufficiently large time t.
For the Perron method, the essential form of the derivatives of T is


D T (x) δx1 , . . . , δxk (t) = Dt−τ (0) · Dk f (x(τ )) δx1 (τ ), . . . , δxk (τ ) dτ.
k

(1.8)
The solution curve x as well as its variations δxi are of growth order eρt , so the
variation of f in the integrand is of growth order ekρt , even if Dk f itself is bounded.
This means that k-th order variations must be considered in spaces of growth order
ekρt and Dk T is only contractive on such spaces if both ρ and k ρ are contained in
the spectral gap.

1.5 Bounded Geometry

The main results of this work are formulated in a geometric context on differentiable
manifolds. Already in [Fen72; HPS77] the results are formulated in such a context.
This allows for more general situations than choosing Rn as ambient space. In the
compact case, it does not require a change in the basic proofs (as can be seen from
the approach taken in [Fen72]), but it does bring in some additional formalism. On
the other hand, pre-existing work on noncompact NHIMs [Hen81; Sak90; BLZ08;
BLZ99] explicitly assumes that the ambient space is Euclidean or Banach.
It turns out that if one switches to a noncompact setting in manifolds, then some
fundamental new ingredients must be added. Already in Euclidean space one must
assume uniformity of the vector fields as a replacement for compactness. Similar
additional uniformity assumptions are required for the ambient space—these do not
manifest themselves in Euclidean space. First, a choice of Riemannian metric (or
possibly a weaker form: a Finsler structure) is required since not all metrics are
equivalent anymore on a noncompact manifold, see Example 3.6. As an extension,
Example 3.7 shows that one cannot reduce the noncompact to a compact case by
compactification. Secondly, the ambient manifold and functions on it should satisfy
uniformity criteria that can be captured in terms of ‘bounded geometry’.14 For full

14 We do not claim that bounded geometry is a necessary condition to generalize the theory of
normal hyperbolicity to noncompact ambient spaces, only that it is sufficient. Section 3.3 does
contain some examples, though, that indicate that some form of bounded geometry is necessary.
20 1 Introduction

details see Sect. 3.3 on compactness and uniformity and Chap. 2 on bounded geom-
etry. Let us just give a quick overview here.
A Riemannian manifold has bounded geometry, loosely speaking, if it is globally,
uniformly well-behaved. More precisely, its curvature must be bounded and the
injectivity radius must be bounded away from zero, see Definition 2.1. Then there
exists a preferred set of so-called normal coordinate charts for which coordinate
transition maps are uniformly continuous and bounded, smooth functions. That is,
in k-th order bounded geometry we have a C k uniform atlas. As a consequence,
uniformly continuous and bounded submanifolds, vector fields, and other objects
can be defined and manipulated in a natural way in terms of these coordinates. Note
that Rn and compact manifolds have bounded geometry, see Example 2.3. Together
with corollaries 3.4 and 3.5 of the main theorem, this shows that bounded geometry
provides a natural generalization to the known settings of compact and Euclidean
spaces.
We use bounded geometry to obtain boundedness estimates on holonomy, see
Sect. 2.2. This is a fundamental ingredient in our proof of smoothness of the perturbed
manifold. Finally, we present more technical results in bounded geometry: a uniform
tubular neighborhood, uniform smoothing of submanifolds, and a trivializing embed-
ding of the normal bundle. We use these to reduce the full problem of persistence of
a normally hyperbolic submanifold M in an ambient manifold Q to the trivialized
situation X × Y , where M is represented by the graph of a small function h : X → Y
and Y is a vector space. Uniformity permeates all these constructions in order to
obtain uniform estimates required for the persistence proof in the trivialized setting.

1.6 Problem Statement and Results

The main problem in this work is the persistence of normally hyperbolic invariant
manifolds under small perturbations of the dynamical system. That is, given a flow
t defined by some vector field v and a normally hyperbolic invariant submanifold
M, we want to show that for any vector field ṽ sufficiently close to v, there exists
a unique manifold M̃ close to M that is invariant under the flow of ṽ; moreover we
would like to show that M̃ is normally hyperbolic again. To make this statement
precise, we need to define a lot of things: first of all, we need to rigorously define
normal hyperbolicity. Secondly, the statements about vector fields and manifolds
being ‘close’ need to be formalized and finally, we need to specify the ambient space
Q on which the system is defined.
We start with a Riemannian manifold (Q, g) as ambient space and a submani-
fold M. For technical reasons this manifold is assumed to be complete and of bounded
geometry (or at least in a δ > 0 neighborhood of M, since the whole analysis can be
restricted to such a neighborhood). Basically, these conditions impose uniformity of
the space, and fit in the principle of replacing compactness by uniform estimates, see
Sect.1.3 and Chap. 2 for more details. Note that Q = Rn with the standard Euclidean
metric is an easy (and typical) special case.
1.6 Problem Statement and Results 21

k,α
Let v ∈ X(Q) be a vector field on Q with v ∈ Cb,u , that is, v up to its k-th
derivative is uniformly continuous and bounded, and α-Hölder continuous if α = 0.
On Rn these statements make immediate sense; on general manifolds Q, results
from Chap. 2 are required, in particular Definition 2.9, to make sense of uniform
boundedness and continuity by means of normal coordinates. Let ṽ be another such
vector field. The closeness of v and ṽ will be measured using supremum norms. The
C 1 -norm is required to be small for the persistence result. Thus, even though we
consider the space of C k,α bounded vector fields, we endow this space with a C 1
topology. See Sect. 1.7 for some more remarks on this topology and a comparison
with standard topologies on noncompact function spaces. If we assume that ṽ − v
is small in C k,α -norm as well, then M̃ will be C k,α -close15 to M. These C 1 and C k
norm requirements and results are direct analogues of those in the implicit function
theorem.
Finally, we define normal hyperbolicity of a submanifold M with respect to a
continuous dynamical system (R, Q, ). The flow t should have a domain of
definition containing at least a neighborhood of the invariant manifold M. This
definition is easily adapted to the discrete case of a diffeomorphism  : Q → Q;
simply replace t ∈ R by t ∈ Z as iterated powers of .
Definition 1.8 (Normally hyperbolic invariant manifold) Let (Q, g) be a smooth
Riemannian manifold, t ∈ C r ≥1 a flow on Q, and let M ∈ C r ≥1 be a submanifold
of Q. Then M is called a normally hyperbolic invariant manifold of the system
(Q, t ) if all of the following conditions hold true:
1. M is invariant, i.e. ∀ t ∈ R : t (M) = M;
2. there exists a continuous splitting

T M Q = TM ⊕ E + ⊕ E − (1.9)

of the tangent bundle TQ over M with globally bounded, continuous projections

π M , π+ , π− and this splitting is invariant under the tangent flow Dt = DtM ⊕
Dt+ ⊕ Dt− ;
3. there exist real numbers ρ− < −ρ M ≤ 0 ≤ ρ M < ρ+ and C M , C+ , C− > 0 such
that the following exponential growth conditions hold on the various subbundles:
 
∀ t ∈ R, (m, x) ∈ TM : Dt (m) x  ≤ C M eρ M |t| x,
M
 
∀ t ≤ 0, (m, x) ∈ E + : Dt (m) x  ≤ C+ eρ+ t x, (1.10)
+
 
∀ t ≥ 0, (m, x) ∈ E : − Dt (m) x  ≤ C− eρ− t x.
−

15 We actually only obtain C k closeness for integer k ≤ r − 1 where r is the ratio in the spectral
gap condition (1.11). This is probably an artifact of the techniques we used, while C k,α closeness
with k + α = r should be obtainable.
22 1 Introduction

These exponential estimates imply that the tangent flow Dt must contract at a rate
of at least ρ− along the stable complementary bundle E − , expand16 as eρ+ t along
the unstable bundle E + , and may not expand or contract at a rate faster than ± ρ M ,
respectively, tangent along TM.
Remark 1.9 We added the condition that the projections π M , π+ , π− are globally
bounded. This is a natural extension to the noncompact case, and is automatically
satisfied in case M is compact.
Remark 1.10 This definition of normal hyperbolicity is not as general as could be.
Fenichel [Fen72, pp. 200–204] defines normal hyperbolicity in terms of ‘general-
ized Lyapunov type numbers’. It follows from his uniformity lemma that these are
essentially exponentiated versions of our Lyapunov exponents ρ. For example, his
ν is equivalent to our e−ρ+ . But Fenichel defines σ in terms of the ratio ρ M /ρ+
along orbits in M. His definition allows the expansion rate along TM to be large,
for example, as long as the expansion rate along E + is large enough to keep the
ratio σ (m) bounded, along the orbit through m. The definitions in [HPS77, Mañ78,
BLZ08] are equivalent in the compact context to the one in [Fen72]. Mañé’s work
shows that this definition is as general as possible, see below. ♦
When M is compact, normal hyperbolicity is a sufficient condition for the exis-
tence of a persistent manifold M̃ for a system generated by ṽ if ṽ − v1 is sufficiently
small. Conversely, Mañé [Mañ78] has proved that normal hyperbolicity (in the sense
of e.g. Fenichel’s definition) is also necessary: if a compact invariant manifold M is
persistent under any C 1 small perturbation, then M is normally hyperbolic (see also
Example 1.2 and the clear exposition in the introduction of [Fen72]). Definition 1.8,
however, only guarantees C 1 smoothness for the perturbed manifold M̃. To obtain
higher order smoothness, a more stringent condition of r -normal hyperbolicity must
be satisfied.
Definition 1.11 (r -normally hyperbolic invariant manifold) A manifold M is
called r -normally hyperbolic with r ≥ 1 a real number, if it satisfies M ∈ C r
and the conditions in Definition 1.8, but with the stronger inequalities

ρ− < −r ρ M ≤ 0 ≤ r ρ M < ρ+ . (1.11)

This means that the normal expansion and contraction must not just dominate the
tangential ones, but do so by a factor r . For r = 1 we recover the original definition,
while the generalized inequality (1.11) is called the spectral gap condition. If M is
r -normally hyperbolic and v and the perturbation ṽ are C r as well, then the persistent
manifold M̃ is C r smooth again. The example in Sect. 1.2.1 shows that this spectral
gap condition is sharp: even when everything is C ∞ , the perturbed manifold M̃ in

 
16 Note that expansion along E + could also be formulated as Dt (m) x  ≥ C+ eρ+ t x for
+
t ≥ 0 and (m, x) ∈ E . This is equivalent to the condition as stated, which says that there is
contraction for t ≤ 0, that is, in backward time. This latter formulation is preferable because it is
the form required in estimates.
1.6 Problem Statement and Results 23

that example is only C r when no more than r -normal hyperbolicity holds. Note that
r can be interpreted as a ‘fractional differentiability degree’ when writing r = k + α
with integer k ≥ 1 the normal degree of differentiability and 0 ≤ α ≤ 1 an additional
Hölder continuity exponent.
Remark 1.12 We explicitly exclude the case r = ∞ from Definition 1.11, even
though the spectral gap condition (1.11) could hold for r = ∞, if ρ M = 0. The reason
is that one can generally not expect to obtain a persistent manifold M̃ ∈ C ∞ in this
case. Even though for any order r < ∞ there exist persistent manifolds M̃ ∈ C r for
sufficiently small perturbations, the maximum perturbation size generally depends
on r and may shrink to zero when r → ∞. See Example 1.3 and the example
in [Str79] for the closely related case of center manifolds.
On the other hand, it is shown in [HPS77] that there is forced smoothness. If
M ∈ C 1 is an r -NHIM and the system is C r , then M must be C r . This also holds in
our noncompact setting, see Remark 3.3, 10 for a sketch of its proof. ♦
With these preliminary definitions in place, we are now ready state our main
theorem; it is restated in Chap. 3. We should point out that M is not required to
be an embedded submanifold; immersions are allowed as well, see Sect. 1.6.2. For
k,α
the details of the smoothness notation Cb,u on manifolds we refer to definitions 2.9
and 2.21.
Theorem 1.1 (Persistence of noncompact NHIMs in bounded geometry) Let
k ≥ 2, α ∈ [0, 1] and r = k + α. Let (Q, g) be a smooth Riemannian manifold of
k,α k,α
bounded geometry and v ∈ Cb,u a vector field on Q. Let M ∈ Cb,u be a connected,
complete submanifold of Q that is r -normally hyperbolic for the flow defined by v,
with empty unstable bundle, i.e. rank(E + ) = 0.
Then for each sufficiently small η > 0 there exists a δ > 0 such that for any
k,α
vector field ṽ ∈ Cb,u with ṽ − v1 < δ, there is a unique submanifold M̃ in the
η-neighborhood of M, such that M̃ is diffeomorphic to M and invariant under the
k,α
flow defined by ṽ. Moreover, M̃ is Cb,u and the distance between M̃ and M can be
made arbitrarily small in C -norm by choosing ṽ − vk−1 sufficiently small.
k−1

This result generalizes the well-known results in [Fen72, HPS77] to the case of
noncompact submanifolds of Riemannian manifolds. Again, our definition of normal
hyperbolicity is slightly less general than the definitions used in these works. We also
assumed that only the stable bundle E − is present, see also Sect. 4.4; note that we
thus only have the spectral gap condition ρ− < −r ρ M with ρ M ≥ 0. See also the
restatement of this theorem on page 76 and the list of remarks 3.3 for more details.
We borrow the idea to generalize the Perron method to NHIMs from Henry
[Hen81], and use the techniques of Vanderbauwhede and Van Gils [VG87] (see
[Van89] for a clear presentation) for proving higher order smoothness. This is sim-
ilar, but developed independently from Sakamoto’s work [Sak90] in which he used
the same ideas to study singular perturbation problems. We improve these results in
a couple of ways. First of all, we simplify the basics of the proof by reducing the
24 1 Introduction

two-step contraction argument to a single contraction mapping, still written as a com-

position of two separate maps acting on horizontal curves in M and vertical curves in
the normal bundle fiber, respectively. More importantly, we remove the restriction of
a trivial product structure X × Y . Thus, we neither require M to have a global chart in
a Banach space X , so M need not be topologically trivial, nor do we require a global
product, so the normal bundle of M need not be trivial either. On the other hand, the
results by Bates, Lu, and Zeng also allow M to be a general submanifold, but still
assume the ambient space to be a Banach space. Our results are for finite dimensional,
but not necessarily linear, Riemannian ambient spaces. In their paper [BLZ08], they
only require an approximate NHIM for finding a persistent invariant manifold. We
use this idea as well (see the setup of h small in the formulation of Theorem 3.2),
but we do not expand this idea any further. Finally, this work was initiated from the
(unfortunately never published) preprint by Duistermaat on stable manifolds [Dui76].
It seems to be a well-known belief by many experts that the theory of normal
hyperbolicity can be extended to a general noncompact setting [DLS06, p. 165]. The
idea is to replace compactness by uniform estimates. An important conclusion to
be drawn from the present work is that indeed this principle holds, but probably in
a more strict way than one would naively realize. Uniform estimates are not only
required for the vector field defining the system, but for the underlying ambient space
as well, in terms of bounded geometry. This becomes clear only when one leaves
the context of Euclidean ambient spaces, which trivially have bounded geometry.
On a Riemannian manifold, already the very definition of uniform continuity of
a vector field v and its derivatives requires some aspects of bounded geometry. It
should be noted though, that we do not prove that bounded geometry is a strictly
necessary condition for persistence of NHIMs; nonetheless, the results do suggest
that persistence of NHIMs may break down in ‘unbounded geometry’, see Sect. 3.3.
In Sect. 3.2 we present an outline of the proof and how it is reduced to a more
basic setting M
× Y of a trivial normal bundle. Here M
is a smoothed version of
M to rectify an artificial loss of smoothness, as occurs e.g. in Ref. [Sak90]. Below
we present some extensions to the main Theorem 1.1 above.

1.6.1 Non-Autonomous Systems

Our main theorem can be trivially extended to the non-autonomous, time-dependent

case. First, extend the configuration space with time t as additional variable, i.e.
Q̂ = Q ×R, and add the equation t˙ = 1. If the original system was time-independent,
then M̂ = M × R is a NHIM for the extended system, and all uniform assumptions
still hold, since the flow along the time direction is neutral and trivial. Note that this
argument does not work in the classical theory as M̂ is not compact.17 Now we can

17 If the perturbation is time-dependent, but in an (almost) periodic way, then this can still be treated

in the compact setting. One can extend the configuration space with the circle S 1 (or an n-torus in
the almost periodic case).
1.6 Problem Statement and Results 25

make any C 1 small perturbation, and obtain a persistent manifold M̃ in the extended
configuration space. The perturbation is allowed to be generally time-dependent, as
long as it is uniform in time, including derivatives. The resulting manifold M̃ will
still be invariant and close to the original M, although it will depend on time. That
is, if we assume local coordinates (x, y) ∈ Rn × Rm for Q such that M = Rn × {0}
locally, then we can write M̃ = Graph(h) for a function

h : Rn × R → Rm , y = h(x, t).

In other words, M̃ can be viewed as a graph over M (i.e. a section of the normal
bundle), but this graph now additionally depends on time. The manifold M̃ itself
is again normally hyperbolic when viewed in the extended space Q × R, see also
Sect. 4.1.
Such time-dependent invariant manifolds are called ‘integral manifolds’. These
have been studied as non-autonomous generalizations of stable and unstable man-
ifolds of hyperbolic fixed points [Pal75], but also as generalizations of compact
NHIMs [Hal61, Yi93]. The theory of noncompact NHIMs allows one to treat all
such integral manifolds in the same way as the autonomous case. One can, for exam-
ple, also start with an integral manifold that is normally hyperbolic: it will persist
just as well.

1.6.2 Immersed Submanifolds

In the main Theorem 1.1, we intentionally do not precisely state in what sense M is
a submanifold of Q. The implicit assumption that M is an embedded submanifold
can be weakened to M being an immersion, see also [HPS77, Sect. 6] and [BLZ99].
That is, M can be viewed as an abstract manifold together with an immersion map
ι : M → Q that need not be injective. This does not affect the theory as long as ι is
still locally injective: ι(M) including a neighborhood modeled on its normal bundle
N can be pulled back via the immersion ι to the abstract M. All local properties
are preserved, so we can study the system via this ‘covering’. We may not always
make a clear distinction between the abstract manifold M and its immersed image
ι(M) ⊂ Q; the discussion below shows that this distinction is not really necessary,
as long as we do not consider perturbations.
For a generic immersion one could expect a picture as in Fig. 1.6, where the
immersed manifold intersects itself transversely. Such situations cannot occur if M
is a NHIM. This follows from the exponential growth rates along tangent and normal
bundles of M. Let m ∈ ι(M) be an intersection point of two preimages m 1 , m 2 ∈ M.
If the tangent spaces along M at m
1 and m 2 are embedded

differently into Tm Q,
then one could find x ∈ Im Dι(m 1 ) \ Im Dι(m 2 ) . This would imply that x has a
component in Nm 2 and give contradictory growth rates for Dt (m) · x depending
on whether we view m as image of m 1 or m 2 , as the orbit of m ∈ ι(M) is uniquely
defined. Hence, at each point m ∈ ι(M) the tangent spaces Dι(m i ) of all preimages
26 1 Introduction

Fig. 1.6 An immersion with Dι (m1)

a transverse intersection

Dι (m2 )

Fig. 1.7 An allowed immer-

sion with tangential intersec-
tion

m i ∈ ι−1 (m) must coincide, see Fig. 1.7. Stated more abstractly, M must have contact
of order one with itself. More generally it holds that an immersed k-NHIM has contact
of order k with itself,18 see [HPS77, p. 68].
Next, each maximal set of ι(M) with constant number of preimages19 p ∈ N ∪
{∞},
M p = {m ∈ ι(M)|# ι−1 (m) = p}, (1.12)

is an invariant subset of ι(M). This is again due to uniqueness of the flow. If an orbit
would cross into a set of different preimage number, then a least one of the ‘lifts’ of
this orbit from ι(M) to the ‘cover’ M would have to enter or leave M. This cannot
happen as M itself is invariant. Hence, the conclusion is that self-intersections of
ι(M) must be invariant.
Immersed NHIMs may occur on themselves, or appear as a persistent manifold
under perturbation from an embedded manifold. An example of an embedded non-
compact NHIM that collapses under a small perturbation into an immersed manifold
can be found in Sect. 3.3. The same can happen with an immersed manifold with
compact image. The following example is taken from [HPS77, p. 130] and shows
that the injection map is relevant for how the NHIM persists. Note that this example
exhibits a Shil’nikov bifurcation, see Remark 1.15 below.
Example 1.13 (Perturbation of a compact non-injectively immersed NHIM) We con-
sider on R3 the vector field

18 The order of contact is defined as the degree up to and including which the Taylor expansions
of the objects agree.
19 The number of preimages must be countable if M is assumed to be second-countable.
1.6 Problem Statement and Results 27

ẋ = arctan(x 2 ) + ε,
ẏ = y,
ż = −z

and smoothly modify it outside the cylinder y 2 + z 2 = 1 such that it flows in the
negative x-direction and connects the basin of repulsion of the origin intersected with
x > 0 to the basin of attraction intersected with x < 0. The perturbation parameter
ε is initially set to zero.
Note that the x-axis is a NHIM (the arctangent is there to keep the vector field and
tangential growth rate bounded). Due to the modification, the two loops in Fig. 1.8
are also NHIMs of this system, both separately and their union. They start from the
origin along the positive x-axis, then diverge from it in opposite directions in the
x y-plane; once outside the cylinder y 2 + z 2 = 1 they start moving into the negative
x direction and finally return to the origin approximately along the x z-plane.
We can parametrize their joint image with an injection ι1 mapping M = {0, 1}×S 1
separately onto the two loops, but we can also parametrize with ι2 that maps M = S 1
onto the full figure eight image. If we perturb to ε > 0, then ι1 will result in Fig. 1.9
where the two loops are separated, while ι2 will result in Fig. 1.10 which has one
loop, but the middle of the figure eight does not intersect anymore. Figure 1.11 shows
how the two orbits from the separate loops closely pass the x–axis along hyperbolic

Fig. 1.8 A non-injectively z

immersed manifold with
compact image

Fig. 1.9 The persistent man-

ifold of ι1 consisting of two
separate loops
28 1 Introduction

Fig. 1.10 The persistent

manifold of ι2 consisting of
one figure eight loop without
self-intersection

Fig. 1.11 Projection onto the z

yz-plane showing the orbits
of the persistent manifold ι1 x
while passing the origin

trajectories. The single orbit of ι2 follows hyperbolic trajectories through the other
two quadrants. 
Remark 1.14 Note that these different persistent NHIMs do not contradict the
uniqueness property of persistence, since the (abstract) manifolds M were differ-
ent to begin with. Formulated differently, if we consider the universal cover of the
tubular neighborhood of ι1 (M) (deduplicating the origin as image point), then Fig. 1.9
shows the unique invariant manifold that stays in this tubular neighborhood cover.
We obtain a different persistent NHIM for any prescribed (possibly infinite) sequence
of concatenating the two loops of the original figure eight into an immersion from
S 1 (or R if the sequence is infinite).
Remark 1.15 This example shows a Shil’nikov bifurcation of a saddle-saddle node
and provides an alternative proof of the result of Shil’nikov [Šil69, Thm. 3] that for
every bi-infinite encoding of the homoclinic loops there exists a unique orbit close
to the original homoclinics. We encode the bi-infinite sequence in the immersion ι of
a NHIM. For each ι there exists a unique persistent manifold, and these correspond
to separate orbits after perturbation. In particular, the periodic orbits correspond to
1.6 Problem Statement and Results 29

those bi-infinite encodings that are actually periodic and there are countably infinitely
many of these. ♦
Finally, we present an example of an injectively immersed (but not embedded)
NHIM, see [HPS77, p. 68]. The mapping below is known as Arnold’s cat map.
Example 1.16 (Injectively immersed dense line in the torus) The matrix
 
21
A=
11

acting on the√two-torus T2 is an Anosov diffeomorphism. The line through 0 with

slope 21 (1 − 5) is densely immersed in the torus and it is a NHIM for this discrete
system. If we take its suspension, then
we have a
flow with a NHIM that is densely
immersed into the mapping torus [0, 1] × T2 / ∼ with identification (1, x) ∼
(0, A x). 

1.6.3 Overflowing Invariant Manifolds

In many applications of normally hyperbolic systems, the manifold M has a boundary

∂ M. A typical reason is that the system ceases to be normally hyperbolic across the
boundary. This happens, for example, when studying a singularly perturbed, or slow-
fast system and in the fast limit there are points on M with zero eigenvalues in the
normal direction. At such points, M is not normally hyperbolic anymore, so one must
restrict M such that these points are outside of M. Another, somewhat artificial but
practical example would be if the invariant manifold is noncompact and one would try
to use the classical theorems that are only applicable to compact manifolds by cutting
off M to a compact manifold with boundary. One can try to attack this latter case
with our more general theory for noncompact manifolds. The additional uniformity
assumptions should be checked then.
If M is a manifold with boundary, some persistence results can still be retained.
This idea was introduced by Fenichel [Fen72] in studying so-called overflowing
invariant manifolds. These are normally hyperbolic manifolds that are invariant under
backward time flow, or in other words, only under the forward flow, orbits can
leave, i.e. ‘overflow’ the manifold. The condition of overflowing invariant is slightly
stronger: the vector field must strictly point outward at the boundary. This weakened
version that the manifold is negatively invariant does come at the additional cost
that only stable normal directions are allowed. The time-reversed situation of an
inflowing invariant manifold with only unstable normal directions is equivalent. In
Sect. 4.3 we discuss how this idea can be incorporated into the Perron method proof.
The attention of the reader is also drawn to the following remark made in [Fen72,
p. 214]. If an open submanifold N ⊂ M is overflowing invariant, and the spectral
gap condition is satisfied on N with a higher ratio r N than on the whole of M, then
30 1 Introduction

the persistent manifold Ñ over N retains C r N smoothness, even if smoothness of M̃

will generally be lower.

1.7 Induced Topology

In this work the topologies for spaces of vector fields, submanifold embeddings, et
cetera, are (implicitly) defined by norms and distance functions. The norms we use are
uniform C k -norms for bounded functions, and families with additional exponential
growth rates. Let us call the topologies induced by these norms Cbk -topologies and
consider how they compare to two common topologies: the weak and strong Whitney
topologies for maps between manifolds, alternatively known as the compact-open
and fine topology, see [Hir76].
The weak topology has a subbasis generated by the set of functions g that are
close to some function f in C k -norm on compact subsets in local coordinate charts.
This means that for example the function family

f δ : R → R : x → δ exp(x 2 )

converges to zero for δ → 0 in this topology. On any compact set f δ will become
arbitrarily small when δ → 0 while it does not converge in uniform norm (nor with
additional exponential growth rate). Hence the weak topology is weaker than our
induced Cbk -topologies.
The strong topology has as basis all sets of functions g that are close to some
function f on a locally finite cover by compact sets K i , where g must approximate
f in C k -norm on each K i in local coordinates up to a given chart-dependent size εi .
For any function without compact support, a collection εi > 0 can be found that
converges faster to zero on each larger K i than the function to zero when x → ∞.
Hence the only sequences of functions R → R that converge to the zero function in
the strong topology are those with (eventually) compact support. A family f δ of func-
tions with noncompact support cannot converge to the zero function, as can be seen
by using a diagonal argument. The family f δ (x) = δ exp(−x 2 ), for example, does
not converge to the zero function in the strong topology. Given a locally finite cover
of R by compact sets K i , we choose xi ∈ K i and corresponding εi = exp(−xi2 )/i.
Then for any given δ > 0, we will have | f δ (xi )| > εi for some large i. On the
other hand, this family f δ obviously converges under the uniform norm with any
exponential growth rate. Thus, the strong topology is stronger than our induced Cbk -
topologies, see also the remark in [GG73, p. 43] for noncompact manifolds.
We conclude that the Cbk -topologies induced by our uniform norms are not equiva-
lent to either the weak or strong Whitney topology, because the weak topology allows
arbitrary behavior of functions outside compact sets, while the strong topology com-
pletely restricts that behavior. Our norms allow moderate variations at infinity. In
general, ‘moderate behavior’ is not well-defined on a general noncompact manifold,
as it depends on the choice of charts. In the setting of bounded geometry, though,
1.7 Induced Topology 31

the uniform, metric structure makes this behavior unambiguous; we can restrict to
normal coordinate charts and consider ‘moderate behavior’ with respect to these.
Note that these topologies are equivalent on compact domains.

1.8 Notation

Here, we will establish some notation and conventions to be used throughout this
work. See the index for more specific symbols.
• The letters I and J will denote intervals in R; I will typically represent an interval
that is unbounded on one side, while J will be bounded.
• ε, δ > 0 will denote (small) bounds for continuity-like estimates; C > 0 will
denote arbitrary bounds. The specific meaning of these symbols will vary depend-
ing on context. ε f (δ) will denote a uniform continuity modulus of the function f ,
that is, ε f : R≥0 → R≥0 satisfies

d( f (x2 ), f (x1 )) < ε f (d(x2 , x1 )) and lim ε f (δ) = 0. (1.13)

δ→0

Without subscript f this will denote an arbitrary continuity modulus.

• The D denotes a total derivative, while Di with index i ∈ N denotes a partial deriva-
tive with respect to the i-th argument, or, when a subscript symbol is appended, say
Dx , then this denotes a partial derivative with respect to the argument commonly
referred to by that symbol.
• We use the following symbols to denote classes of function spaces:
Cb bounded, continuous functions;
Cb,u bounded, uniformly continuous functions;
Ck k times continuously differentiable functions;
C k,α C k functions with α-Hölder continuous k-th derivative. We will conventionally write
k,α
r = k + α ∈ R≥1 ; the Hölder estimates are assumed to be uniform in Cb,u spaces.
L continuous, i.e. bounded, (multi)linear operators;
X vector fields;
 sections of a fiber bundle.

Unless otherwise specified, Cbk and Cbk,α spaces will be endowed with the canonical
norms that turn these into Banach spaces, that is,
 k 
  n  D f (x2 ) − Dk f (x1 )
 f k,α = sup D f (x) + sup . (1.14)
x x2 =x1 d(x2 , x1 )α
0≤n≤k

We define the operator norm on a multilinear operator A ∈ Lk (V1 × · · · × Vk ; W )

as
 A = sup A(v1 , . . . , vk ). (1.15)
vi ∈Vi
vi =1
32 1 Introduction

This multilinear operator norm can be extended to sections s of real-valued tensor

bundles by taking the operator norm pointwise of s(x) as a multilinear operator
into R.

• On a Riemannian manifold,  will denote the Christoffel symbols, while  will be

used for parallel transport along a curve given as argument, for example, (γ |ab )
will denote parallel transport along the curve γ restricted to the interval [a, b].
We shall denote induced parallel transport on products of the tangent bundle by
(γ |ab )⊗k .
• We shall often work with maps that are defined on the tangent space over a point
x ∈ M and denote this dependence on x by a subscript, for example h x : Tx M →
Tx M. If we want to refer to the whole family of such maps for all x ∈ M, then we
denote this by
h • : T• M → T• M,

particularly if we want to stress that this family satisfies some properties uniformly
in x.
• We use the notation B(x; δ) not only to indicate open balls of radius δ around a
single point x, but also B(M; δ) to indicate a (tubular) neighborhood of some set
or submanifold M, that is,

B(M; δ) = {x|d(M, x) < δ}.

The following definition of a scale of Banach spaces (cf. [VG87]) is fundamental

to the rest of this work.
Definition 1.17 Let X be a normed linear space and F = C(I ; X ) the space of
continuous functions from an interval I ⊂ R to X . We define a family of exponential
growth norms with parameter ρ ∈ R by

 f ρ = sup  f (t) e−ρ t for f ∈ F. (1.16)

t∈I

We define B ρ (I ; X ) to be the normed space consisting of all functions f ∈ F with

 f ρ < ∞. If X is a Banach space, then B ρ (I ; X ) is a Banach space as well.
Remark 1.18 When the interval I is bounded from below, then the embedding
B ρ1 (I ; X ) → B ρ2 (I ; X ) is continuous for ρ1 ≤ ρ2 . The time reversed version
when I is bounded above and ρ2 ≤ ρ1 holds, will frequently recur throughout this
work. See also Remark B.4 and the note on integrals of exponentials (1.18) below.
In Chap. 3 we shall use I = R≤0 and negative rates ρ, while in the appendices B
and C we use (the somewhat more natural) I = R≥0 ; though ρ’s can take both signs
there. ♦
The definition of an exponential growth norm can be generalized to curves map-
ping into a metric space. Let (X, d) be a metric space, then analogously to (1.17),
we define a family of exponential growth distance functions on F by
1.8 Notation 33

dρ ( f 1 , f 2 ) = sup d f 1 (t), f 2 (t) e−ρ t . (1.17)

t∈I

Note that this distance function might be infinite for some x1 , x2 ∈ F.

We will be working with exponential growth estimates of the form C eρt through-
out this paper. The pair of numbers C > 0, ρ ∈ R that determine such a growth
estimate will be referred to as exponential growth numbers, and ρ as an exponential
growth rate.
We will frequently encounter integrals over a time interval, where the integrand
obeys an exponential estimate. As long as the interval [a, b] is bounded in the direc-
tion of exponential growth and ρ = 0, these can be estimated as
 b
1

eρt dt ≤ exp sup ρ t . (1.18)

a |ρ| t∈[a,b]

We also state here some basic facts about uniformly Hölder continuous functions.
Lemma 1.19 (Product rule for Hölder continuity) Let f, g ∈ Cb,u α be defined on
α .
spaces such that the product f · g is well-defined. Then also f · g ∈ Cb,u
Proof Let  f 0 , g0 ≤ M and let C f,α , C g,α be the respective Hölder coefficients
of f, g. Then we have for all x1 = x2

 f (x2 ) g(x2 ) − f (x1 ) g(x1 ) ≤  f (x2 ) g(x2 ) − g(x1 ) +  f (x2 ) − f (x2 ) g(x2 )
≤ M (C f,α + C g,α ) x − yα ,

which exhibits the Hölder coefficient M (C f,α + C g,α ) for the product, and f · g is
clearly bounded by M 2 . 

α . Then it also holds that f ∈ C β

Lemma 1.20 Let f ∈ Cb,u b,u for any 0 < β < α.

Proof Let M be the bound on f , and Cα its α-Hölder coefficient. For x2 − x1  ≤ 1
the estimate for β follows automatically from that of α. For x2 − x1  > 1 we use
boundedness to obtain

 f (x2 ) − f (x1 ) ≤ 2 M ≤ 2 M x2 − x1 β .

Hence, Cβ = max(Cα , 2 M) suffices as β-Hölder coefficient. 

Typographical Conventions

As usual we close proofs with the symbol , while we shall use ♦ and  to denote
the end of (a series of) remarks or examples, respectively.
Chapter 2
Manifolds of Bounded Geometry

For noncompact normally hyperbolic systems, uniformity assumptions that were

implicit in the compact case must be made explicit. Not only assumptions on the
vector field, but on the underlying space as well. For this we need the concept of
bounded geometry; Sect. 3.3 contains a discussion and examples for why we require
this concept.
The class of manifolds of bounded geometry allows us to uniformly apply con-
structions that are well-known for compact manifolds. We single out the atlas of
normal coordinate charts and derive from the very definition of bounded geometry
that all constructions and estimates are uniform over all such charts. For complete-
ness, we present here all results that we need later on. Some of these results are
already present in the literature: the construction of a uniformly locally finite cover
and a subordinate C k uniformly bounded partition of unity, and bounded coordinate
transformations can be found in [Shu92; Sch01], for example, while [Roe88] includes
the result on finite coloring of the connectedness graph of a uniformly locally finite
cover and the construction of a trivial bundle embedding in Proposition 7.5 with a
sketch of the proof. I have not been able to find in the literature the results about the
existence of a uniform tubular neighborhood and the approximation of a submanifold
by a smoothed manifold. Submanifolds are allowed to be non-injectively immersed.
This chapter is organized as follows. First, the material is presented that is already
required for the global coordinate setting of Theorem 3.2. These include the basic
definitions of bounded geometry, related results on bounded coordinate transition
maps, uniform covers and partitions of unity, and an explicit relation between
holonomy and curvature. Then we continue to work towards the final goal of this
chapter: to reduce a noncompact normally hyperbolic system from a setting in general
manifolds to a trivial bundle setting, in order to generalize the persistence theorem
to the former setting. To this end, we need some more technical results: a uniform
tubular neighborhood, smooth approximation of a submanifold, and embedding into
a trivial bundle.
This chapter relies heavily on some more advanced concepts from differential
and specifically Riemannian geometry. On the other hand, the results are used as

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 35

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4_2,
© Atlantis Press and the author 2013
36 2 Manifolds of Bounded Geometry

tools in solving a dynamical systems problem. Appendix F provides a quick review

for non-experts of the most relevant geometric concepts used here. It also provides
further references to the literature. We shall assume the contents of this appendix
known from here on.
I suggest the reader to at least take a glance at the first two sections of this chapter
to familiarize himself with the basic definitions and results of bounded geometry,
without the need to go through the details of the proofs. Then, depending on his
interest, he can choose to delve into the more technical geometric details or skip to
Chap. 3 for the more analytical side of the proof of Theorem 3.2, and possibly return
later to read how Theorem 3.1 is reduced to the former.

2.1 Bounded Geometry

We follow the definition in [Eic91] to introduce bounded geometry. Recall that the
injectivity radius rinj (x) at a point x ∈ M is the maximum radius for which the
exponential map at x is a diffeomorphism, see also Appendix F.
Definition 2.1 (Bounded geometry) We say that a complete, finite-dimensional
Riemannian manifold (M, g) has k-th order bounded geometry when the following
conditions are satisfied:
(I) the global injectivity radius rinj (M) = inf rinj (x) is positive, rinj (M) > 0;
x∈M
(Bk ) the Riemannian curvature R and its covariant derivatives up to k-th order are
uniformly bounded,

∀ 0 ≤ i ≤ k : sup ∇ i R(x) < ∞,

x∈M

with operator norm of ∇ i R(x) as an element of the tensor bundle over x ∈ M.

Remark 2.2 The conditions (I) and (Bk ) are independent. We present a simple
example which exhibits zero infimum for the injectivity radius while all deriva-
tives of the curvature are globally bounded. Indeed, let M = R × S 1 be a cylinder
with metric g = dx 2 + e−2x dθ 2 in coordinates (x, θ), see also Fig. 3.3 on page 88.1
The injectivity radius rinj (M) is zero since the cylinder circumference shrinks to zero
with x → ∞. Global boundedness of the curvature and all of its derivatives follows
from a symmetry argument. The family

ϕξ,α : (x, θ) → (x + ξ, eξ θ + α) with ξ ∈ R, α ∈ [0, 2π )

1 This is a noncompact surface with constant negative curvature, hence it cannot be isometrically
embedded into R3 , see [Hil01]. The embedding is nearly isometric for x
0 though, so the figure
is still a good representation there.
2.1 Bounded Geometry 37

is a set of local isomorphisms that acts transitively on M. That is, for any two
points (x1 , θ1 ), (x2 , θ2 ) ∈ M there exist ξ, α and a neighborhood U  (x1 , θ1 ) such
that ϕξ,α : U → ϕξ,α (U ) is an isomorphism and ϕξ,α (x1 , θ1 ) = (x2 , θ2 ). For any
(x, θ) ∈ U and ν, w ∈ T(x,θ) M we have
∗
 
(ϕξ,α g)(x,θ) (ν, w) = g(x+ξ,eξ θ+α) Dϕξ,α (x, θ ) ν, Dϕξ,α (x, θ ) w
= dx(ν)dx(w) + e−2(x+ξ ) eξ dθ (ν) eξ dθ (w)
= g(x,θ) (ν, w)

so ϕξ,α∗ g = g on U . Since the curvature and its derivatives are locally determined,

this implies that these are constant across M, hence uniformly bounded (actually all
derivatives of R vanish). Note that these local isometries do not imply a finite global
injectivity radius since the size of the neighborhood U does depend on the points
(x 1 , θ1 ), (x2 , θ2 ) ∈ M. ♦
Example 2.3 (Manifolds of bounded geometry) The following are examples of man-
ifolds with bounded geometry of any (i.e. infinite) order.
• Euclidean space with the standard metric trivially has bounded geometry.
• A smooth, compact Riemannian manifold M has bounded geometry as well; both
the injectivity radius and the curvature including derivatives are continuous func-
tions, so these attain their finite minimum and maxima, respectively, on M. If
M ∈ C k+2 , then it has bounded geometry of order k.
• Noncompact, smooth Riemannian manifolds that possess a transitive group of
isomorphisms (such as hyperbolic space) have bounded geometry since the finite
injectivity radius and curvature estimates at any single point translate to a uniform
estimate for all points under isomorphisms. Note that the example in Remark 2.2
above shows that it is not sufficient to have local isometries.
More manifolds of bounded geometry can be constructed with these basic building
blocks in the following ways.
• The product of a finite number of manifolds of bounded geometry again has
bounded geometry, since the direct sum structure of the metric is inherited by
the exponential map and curvature. We give an outline of the proof. In a product
coordinate chart
(ϕ1 , ϕ2 ) : U1 × U2 → Rn 1 × Rn 2

with coordinates (x1 , x2 ), the metric has diagonal form

 
g1 (x1 ) 0
g(x1 , x2 ) = (g1 ⊕ g2 )(x1 , x2 ) = .
0 g2 (x2 )

The coordinate dependence on x1 , x2 is non-mixed and this is preserved under

taking derivatives and index contractions, so R will split into a direct sum of R1
and R2 again. This can be extended to derivatives of R.
38 2 Manifolds of Bounded Geometry

A geodesic in M1 × M2 is precisely given by γ = (γ1 , γ2 ) where γ1 , γ2 are

geodesics parametrized with constant speed in M1 , M2 , respectively. This follows
easily since minimization of length is equivalent to minimization of the energy
functional
 b  b
2 E(γ ) = g(γ̇ , γ̇ )dt = g1 (γ̇1 , γ̇1 ) + g2 (γ̇2 , γ̇2 )dt
a a

and this splits nicely into independent minimization

 problems for γ1 and γ2 . With
a little effort one sees that rinj (M) ≥ min rinj (M1 ), rinj (M2 ) .
• If we take a finite connected sum of manifolds with bounded geometry such that the
gluing modifications are smooth and contained in a compact set, then the resulting
manifold has bounded geometry again.
• We can endow the tangent bundle TM of a Riemannian manifold (M, g) with the
natural Sasaki metric [Sas58]. Let x i denote coordinates on an open neighborhood
U ⊂ M. These coordinate functions can be pulled back to TU and the one-forms
dx i can be viewed as additional coordinates vi such that the x i , v j together form
a complete set of induced coordinates on TU . With respect to these coordinates
the Sasaki metric is given by
 
ĝ(x, v) = gi j (x) dx i dx j + Dvi Dv j where Dvi = dvi + ijk v j dx k (2.1)

and ijk denote the Christoffel symbols on M, while the dvi are one-forms on the
manifold TU .
Bounded geometry of (M, g) is not inherited by (TM, ĝ) since the extended Rie-
mannian curvature R̂ contains unbounded terms v when expressed in terms of
R, see [GK02,
 Prop. 7.5]. These expressions
 do readily show that the restriction
Tr M = (x, v) ∈ TM | gx (v, v) ≤ r 2 satisfies curvature bounds of order k − 1 if
(M, g) has k-bounded geometry. The geodesic flow equation is given in induced
coordinates by [Sas58, Eq. (7.7)]. By application of Theorem A.6, one can then
show that the injectivity radius is bounded.
Note that Tr M is a manifold with boundary, but this is not problematic in our setting
as long as the invariant submanifold stays away from the boundary. Alternatively
one could try to use results from [Sch01]. 
When we say that a manifold has bounded geometry without specifying the order
k, then it is assumed that the order is infinite, k = ∞, or sufficiently large. When
k ≥ 1 we have the following result, see [Eic91, Thm. 2.4 and Cor. 2.5]. In case
k = ∞ the converse also holds [Roe88, Prop. 2.4].
2.1 Bounded Geometry 39

Theorem 2.4 (Boundedness of the metric) Let (M, g) be a Riemannian manifold

of k-bounded geometry. Then there exists a δ > 0 such that the metric up to its k-th
order derivatives and the Christoffel symbols up to its (k −1)-th order derivatives
are bounded in normal coordinates of radius δ around each xin M, and the bounds
are uniform in x.
This basic fact can be used to make the properties of all kinds of constructions uniform
over a noncompact manifold. Note that here and in the following, all uniformity
estimates are assumed globally valid, that is, independent of the point x ∈ M.
To stress this, we shall use notation f • , for example as in Definition 2.9, to indicate
that the family of maps { f x }x∈M satisfies continuity estimates independent of x
With Theorem 2.4 at hand, we shall exclusively use normal coordinates for local
coordinate calculations. To establish notation, we say that

ϕ = expx −1 : B(x; δ) ⊂ M → B(0; δ) ⊂ Tx M (2.2)

is a normal coordinate chart at x ∈ M. The radius δ will always be chosen smaller

than the injectivity radius rinj (M), so ϕ is a diffeomorphism. Each tangent space
Tx M carries the inner product gx , hence is isometric to Euclidean space Rn (but
identification requires a choice of basis).
Proposition 2.5 Let (M, g) be a Riemannian manifold of k ≥ 1 bounded geometry.
For every C > 1 there exists a δ > 0 such that the normal coordinate charts ϕx
in (2.2) are defined on B(x; δ) for each x ∈ M and the Euclidean distance d E on
the normal coordinates is uniformly C-equivalent to the metric distance d induced
by M, that is,

∀x1 , x2 ∈ B(x; δ) : C −1 d(x1 , x2 ) ≤ d E (ϕx (x1 ), ϕx (x2 )) ≤ C d(x1 , x2 ).

Proof Let δ < 21 rinj (M) and x ∈ M. We consider a normal coordinate chart ϕx on
B(x; 2δ). According to Theorem 2.4, the metric g and its derivatives are bounded
in normal coordinates. We have (exp∗x g)(0) = gx , the Euclidean inner product on
Tx M, while the total derivative D(exp∗x g)(ξ ) is bounded on B(0; 2δ)  ξ , say by
D(exp∗x g)(ξ ) ≤ C1 , independent of x ∈ M. By the mean value theorem this
induces the uniform bounds

1 − 2δ C1 ≤ (exp∗x g)(ξ ) ≤ 1 + 2δ C1 .

Let x1 , x2 ∈ B(x; δ) and let γ E be the straight curve between ϕx (x1 ) and ϕx (x2 )
in Tx M parametrized by arc length. This curve γ E attains the Euclidean distance
l E (γ E ) = d E (ϕx (x1 ), ϕx (x2 )). On the other hand, it gives an upper bound on the
metric distance
40 2 Manifolds of Bounded Geometry
 l E (γ E )  
d(x1 , x2 ) = inf l(γ ) ≤ (exp∗x g)γ E (t) γ  E (t), γ  E (t) dt
γ

0  
≤ 1 + 2δ C1 d E ϕx (x1 ), ϕx (x2 ) .

Let γ be a geodesic minimizing the distance d(x1 , x2 ). Then γ is contained in

B(x; 2δ): the distance from each xi to the boundary of B(x; 2δ) is at least δ, so if γ
would leave and re-enter B(x; 2δ) then its length would be at least 2δ. On the other
hand, x1 and x2 can be connected via x with a curve of length less than 2δ. Let us
write η = ϕx ◦ γ and assume that η is parametrized by arc length with respect to the
Euclidean metric gx . Then we obtain an inverse estimate to the one above:
 l E (η)  l E (η)
1  
d E (ϕx (x1 ), ϕx (x2 )) ≤ 1dt ≤ (1 − 2δ C1 )− 2 (exp∗x g)η(t) η (t), η (t) dt
0 0
 l E (η)
1  
= (1 − 2δ C1 )− 2 gγ (t) γ  (t), γ  (t) dt
0
− 21
≤ (1 − 2δ C1 ) d(x 1 , x2 ).

Finally, we complete the proof by choosing δ > 0 small enough that

 1 1
max (1 + 2δ C1 ) 2 , (1 − 2δ C1 )− 2 ≤ C. 

From here on, we shall frequently represent objects living in B(x; δ) ⊂ M on nor-
mal coordinate neighborhoods B(0; δ) ⊂ Tx M via the normal coordinate chart ϕx .
We will mostly use B(x; δ) to clearly indicate the base point, or B(0x ; δ) ⊂ Tx M
to stress the tangent space domain of the coordinates as well. In spaces of bounded
geometry, normal coordinate charts are the natural charts to works in and coordinate
transition maps are not just smooth, but uniformly bounded, as stated in the following
lemma.
Lemma 2.6 (Boundedness of transition maps) Let (M, g) be a Riemannian man-
ifold of k-bounded geometry with k ≥ 2. There exists a δ with 0 < δ < rinj (M) and
constants C, L > 0 such that for all x1 , x2 ∈ M with d(x1 , x2 ) < δ the following
holds.
1. The coordinate transition map

ϕ2,1 = ϕ2 ◦ ϕ1−1 : U → Tx2 M with U = ϕ1 (B(x1 ; δ) ∩ B(x2 ; δ)) ⊂ Tx1 M

(2.3)
is C k−1 bounded with ϕ2,1 k−1 ≤ C.
2. Let γ2,1 : [0, 1] → B(x1 ; δ) be the unique shortest geodesic connecting x1 and
x2 and let (γ2,1 ) be the associated parallel transport. Then the map

ϕ2,1 − (γ2,1 ) : U → Tx2 M

has C k−2 -norm bounded by the Lipschitz estimate

2.1 Bounded Geometry 41

ϕ2,1 − (γ2,1 )k−2 ≤ L d(x1 , x2 ). (2.4)

Remark 2.7 One degree of smoothness is lost because the exponential map is defined
in terms of the geodesic flow. This flow in turn is defined in terms of the Christoffel
symbols, which depend on derivatives of the metric, so these are only C k−1 bounded.
We lose another degree of smoothness in estimating ϕ2,1 −(γ2,1 ) since the Lipschitz
estimate follows from a uniform bound on one higher derivative of these. ♦
We shall first compare both ϕ2,1 and (γ2,1 ) to the identity in normal coordinates
and finally conclude with the triangle inequality that their difference must be small.
We compare ϕ2,1 to the parallel transport (γ2,1 ) since this is the most natural way
to identify the tangent spaces Tx1 M and Tx2 M.
Proof Let B(x1 ; δ), B(x2 ; δ) be two normal coordinate neighborhoods with non-
empty intersection. The coordinate transition map ϕ2,1 = ϕ2 ◦ ϕ1−1 = exp−1 x2 ◦ exp x1
can be studied as the exponential map expx1 : Tx1 M → M in normal coordinates
on B(x2 ; δ), since ϕ2 = exp−1 x2 . From here on, we will implicitly be working in
normal coordinates around x2 , using some choice of basis to isometrically identify
Tx2 M ∼= Rn .
Let x ∈ B(x1 ; δ) ∩ B(x2 ; δ), hence x1 ∈ B(x2 ; 2 δ). We choose δ ≤ 1, and small
enough so that the results of Theorem 2.4 and Proposition 2.5 (with C = 2) hold for
2 δ. The exponential map is given by the time-one geodesic flow projected on the
base manifold. For the base point x2 , this is the identity map, while for the base point
x1 we will show that it is a small perturbation thereof. The geodesic flow on TM is
given in local coordinates by

ẋ i = vi , (2.5)
v̇ = −
i i
jk (x) v v ,
j k

where ijk denote the Christoffel symbols with respect to the coordinates x i on M
and the v j are induced additional coordinates on TM, see the explanation above (2.1).
The Christoffel symbols are C k−1 bounded due to Theorem 2.4. Let ϒ t denote the
geodesic flow of (2.5) on TM restricted to B(x2 ; 2 δ). We denote by (x(t), v(t)) a
solution curve of ϒ t . The geodesic flow preserves the length of tangent vectors with
respect to the metric g, so we have v(t) ≤ 2 v(0) ≤ 2 δ with respect to the
Euclidean distance in the normal coordinates. This implies that the vector field (2.5)
is bounded in these induced coordinates. Hence, by Theorem A.6, ϒ t ∈ Cbk−1 is
bounded as well on the interval [0, 1]. Moreover, Dϒ t ∈ Cbk−2 exhibits a Lipschitz
estimate for the base point dependence ϕ2 (x1 ) E . By Proposition 2.5 the local
Euclidean distance is equivalent to the distance on M, so ϕ2 (x1 ) E ≤ 2 d(x1 , x2 ).
These conclusions directly translate to expx ( · ) = π ◦ ϒ 1 (x, · ) and we conclude
that ϕ2,1 = exp−1 k−1
x2 ◦ exp x1 ∈ C b with bound C > 0 uniform in x1 , x2 ∈ M and
ϕ2,1 − 1k−2 ≤ L d(x1 , x2 ) for some L  > 0.

42 2 Manifolds of Bounded Geometry

The parallel transport (γ2,1 ) is given by integrating the pullback of the con-
nection along γ2,1 . This yields a differential equation similar to (2.5) and similarly
leads to C k−1 boundedness estimates in normal coordinates and Lipschitz estimates
for the C k−2 -norm. Thus, the difference ϕ2,1 − (γ2,1 ) is C k−1 bounded, and has
C k−2 -norm that satisfies the Lipschitz estimate (2.4) for some L > 0. 
Definition 2.8 (M-small coordinate radius) Let (M, g) be a Riemannian manifold
of bounded geometry. We define δ > 0 to be M-small if Theorem 2.4 and Lemma 2.6
hold on all normal coordinate charts of radius δ.
Note that such a δ > 0 always exists. From now on, we shall always assume to have
selected such a δ for any given manifold of bounded geometry and restrict its atlas
to include these normal coordinate charts only.
Lemma 2.6 shows that normal coordinate transformations respect C k boundedness
of functions in coordinate representations. Thus, it is natural to consider manifolds
of bounded geometry as the class of C k bounded manifolds with respect to this
restricted atlas. This also makes the following definition natural.
Definition 2.9 (Ck bounded maps) Let X, Y be Riemannian manifolds of k + 1-
bounded geometry and f ∈ C k (X ; Y ). We say that f is of class Cbk when there
exist X, Y -small δ X , δY > 0 such that for each x ∈ X we have f (B(x; δ X )) ⊂
B( f (x); δY ) and the representation

f˜x = exp−1
f (x) ◦ f ◦ expx : B(0; δ X ) ⊂ Tx X → T y Y (2.6)

in normal coordinates is of class Cbk and the associated C k -norms of f˜• are bounded
k (X ; Y ) and C k,α (X ; Y ) functions
uniformly in x ∈ X . We define the classes of Cb,u b,u
analogously when X, Y are of k +2-bounded geometry.
Remark 2.10 We shall say that a vector field v ∈ X(X ) is of class Cbk , also denoted
by v ∈ Xkb (X ), when v ∈ Cbk with respect to coordinates on TM induced by normal
coordinates on M. This is slightly different from normal coordinates on TM induced
by the metric (2.1). Note that since v ≤ r is assumed bounded, we could restrict
to the submanifold Tr M of bounded geometry and consider v ∈ Cbk (M; Tr M), but
this is less practical.
Remark 2.11 The manifolds X, Y need to have bounded geometry of one or two
degrees higher than the smoothness of the maps to preserve boundedness and uniform
continuity estimates under normal coordinate transformations. This shall from now
on always be an implicit assumption.
Remark 2.12 (Locally/globally defined continuity modulus) The continuity modulus
ε f of a function f ∈ Cb,u
k (X ; Y ) is only defined on the interval
[0, δ X ) ⊂ R. On the
other hand, D f (x) is globally well-defined in terms local charts and assumed to
k

be bounded. We shall want to compare Dk f at points x1 , x2 far apart. If we have

isometric isomorphisms
2.1 Bounded Geometry 43

∼ ∼
ϕ : Tx1 X → Tx2 X and ψ : T f (x1 ) Y → T f (x2 ) Y,

then this allows us to compare

Dk f (x2 ) ◦ ϕ ⊗k − ψ ◦ Dk f (x1 ) ≤ Dk f (x2 ) + Dk f (x1 ). (2.7)

Note that the right-hand expression does not depend on the choice2 of isomorphisms.
Thus, with such isomorphisms at hand, we can use (2.7) to heuristically extend
the local to a global continuity modulus. That is, for nearby points x1 , x2 we use
an estimate in terms of local charts; if this is not possible, then the points must be
separated by a distance larger than a δ as in Definition 2.8. Since the functions we
consider are globally bounded, we then use some (non-canonical) choice to identify
the vector bundle fibers over x1 , x2 that the function lives in and estimate by the
right-hand side of (2.7). This estimate is crude but independent of the choice of
k,α
identification and will always satisfy our needs. For example, if f ∈ Cb,u (X ; Y ),
with Hölder coefficient Cα locally for d(x1 , x2 ) ≤ δ then we have

Cα d(x1 , x2 )α if d(x1 , x2 ) < δ,
D f (x2 ) − D f (x1 ) ≤
k k
2  f k α else.
δα d(x ,
1 2x )
 2f 
This shows that we can heuristically consider max Cα , δ α k as a global Hölder
coefficient. ♦
The following proposition shows that we may measure continuity of the deriva-
tives of a function f using local parallel transport. With the remark above we see how
it can be extended to a global continuity modulus if a (non-unique) choice is made
for how to connect non-close points x1 , x2 by a path; this idea will be developed in
Sect. 3.7.4.
Proposition 2.13 (Equivalence of continuity moduli) Let X, Y be Riemannian
manifolds of bounded geometry and f ∈ Cbk (X ; Y ). Then the following statements
are equivalent:
k,α
1. f ∈ Cb,u (X ; Y ) according to Definition 2.9;
2. we have the continuity estimate

∃ ε f, ∈ C α (R+ ; R+ ), δ0 > 0 : ∀ x1 , x2 ∈ X, d(x1 , x2 ) ≤ δ0 :

Dk f˜x2 (0) · (γ2,1 )⊗k − (η2,1 ) · Dk f˜x1 (0) ≤ ε f, (d(x1 , x2 )),
(2.8)

2 In practice, we shall use isomorphisms defined by parallel transport onX = Y , cf. Proposition 2.13.
This is a non-canonical choice, since it depends on the path connecting x1 , x2 . A canonical choice
that depends continuously on x1 , x2 cannot be made in general, since it would imply that the tangent
bundle is trivializable.
44 2 Manifolds of Bounded Geometry

where (η2,1 ) and (γ2,1 )⊗k denote parallel transport along the unique short-
est geodesic between f (x1 ), f (x2 ) and x1 , x2 , respectively, and ε f, denotes a
uniform or α-Hölder continuity modulus.
Proof We first prove the statement in case Y is a normed linear space, hence no
parallel transport term (η2,1 ) appears.
Let δ0 ≤ δ X as in Definition 2.9 (thus, in particular δ0 is X -small), and let
d(x1 , x2 ) ≤ δ0 . Then we have the Lipschitz estimate ϕ2,1 −(γ2,1 ) ≤ L d(x1 , x2 )
while the normal coordinate representations (2.6) of f at x1 , x2 are related by f˜x1 =
f˜x2 ◦ ϕ2,1 . This leads to

Dk f˜x2 (0) · (γ2,1 )⊗k − Dk f˜x1 (0)

= Dk f˜x2 (0) · (γ2,1 )⊗k − Dk [ f˜x2 ◦ ϕ2,1 ](0)
 ⊗k
≤ Dk f˜x2 (0) · (γ2,1 )⊗k − Dk f˜x2 (ϕ2,1 (0)) · Dϕ2,1 

k−1
 
+ Dl f˜x2 (ϕ2,1 (0)) · Pl,k D• ϕ2,1 (0) 
l=1
≤ D f˜x2 (0) − Dk f˜x2 (ϕ2 (x1 )) + Dk f˜x2 (ϕ2 (x1 )) (γ2,1 ) − Dϕ2,1 k
k


k−1
 
+ Dl f˜x2 (ϕ2 (x1 )) Pl,k D• ϕ2,1 (0) 
l=1

 k 
k−1
 
≤ ε f (d(x2 , x1 )) +  f k L d(x1 , x2 ) +  f l Pl,k D• ϕ2,1 (0) ,
l=1

where ε f denotes the continuity modulus of f and its derivatives according to Def-
inition 2.9, and the Pl,k denote (l, k)-linear maps according to Proposition C.3. We
used the fact that both (γ2,1 ) and Dϕ2,1 (0) act on the k-tensor bundle as a k-tuple
of copies. By assumption  f k is bounded, and to estimate the Pl,k terms, we note
that l < k, so each of the Pl,k contains at least a factor Di ϕ2,1 (0) with i ≥ 2. Since
ϕ2,1 is close to (γ2,1 ) and Di (γ2,1 ) = 0 for i ≥ 2, it follows that
 
Pl,k D• ϕ2,1 (0)  ≤ C L d(x1 , x2 )

for some constant C independent of x1 , x2 . This shows that the continuity modulus
ε f, of (2.8) can be estimated by the continuity modulus ε f plus additional Lipschitz
terms. We can reverse the estimates above to arrive at the same conclusion when
expressing ε f in terms of ε f, . Hence, the continuity statements are equivalent for
any α ≤ 1.
If Y is a Riemannian manifold of bounded geometry, we just apply the same
estimates in the codomain. To this end, we must have d( f (x2 ), f (x1 )) ≤ δY , so we
choose δ0 small enough that
2.1 Bounded Geometry 45

d( f (x2 ), f (x1 )) ≤ D f  d(x2 , x1 ) ≤  f 1 δ0 ≤ δY

holds with δY as in Definition 2.9. 

The definition of bounded geometry can be extended to vector bundles, see
also [Shu92, p 65].
Definition 2.14 (Vector bundle of bounded geometry) Let (M, g) be a manifold
of bounded geometry and δ be M-small as in Definition 2.8. We say that a vector
bundle π : E → M with fiber F has k-th order bounded geometry when there exist
preferred trivializations
 
τ : π −1 B(m; δ) → B(m; δ) × F for each m ∈ M (2.9)

such that if we have a transition function ϕ2,1 = τ2 ◦τ1 −1 between two trivializations
on B(m 1 ; δ) and B(m 2 ; δ), then the function g : B(m 1 ; δ) ∩ B(m 2 ; δ) → L(F)
defined by ϕ2,1 (m, f ) = g(m) f satisfies g ∈ Cbk independent of the points
m 1 , m 2 ∈ M.
Remark 2.15 Note that we could have replaced B(m; δ) by arbitrary (preferred)
coordinate charts. The relevant property is that we express uniformity of the transition
functions in terms of uniformity of the function g with respect to the underlying
coordinate charts of M, which are normal coordinates in our case. ♦
It follows from Lemma 2.6 that the tangent bundle TM has bounded geometry of
order k −2 if (M, g) has bounded geometry of order k ≥ 2. One order of smoothness
is lost (beyond the one expected) as noted in Remark 2.7.
We introduce the concept of a uniformly locally finite cover of a manifold of
bounded geometry. This is a natural extension of a locally finite cover. Uniformity
means that we require a global bound K on the number of sets in the cover that
intersect any small open ball.
Lemma 2.16 (Uniformly locally finite cover) Let (M, g) be a Riemannian mani-
fold of bounded geometry.
 Then for  δ2 > 0 small enough and any 0 < δ1 ≤ δ2 , M has a countable cover
B(xi ; δ1 ) i≥1 such that
1. ∀ i  = j : d(xi , x j ) ≥ δ1 ;
2. there exists an explicit global bound K such that for each x ∈ M the ball B(x; δ2 )
intersects at most K of the B(xi ; δ2 ).
Note that the second result implies both that the cover is locally finite with fixed
neighborhood size, and that each set in the cover overlaps with at most K others, cf.
Lebesgue covering dimension.
Proof Using Proposition 2.5, choose δ > 0 such that Euclidean distance in normal
coordinates on each B(x; δ) is C = 2 equivalent to the metric distance and set
δ2 ≤ δ/3.
46 2 Manifolds of Bounded Geometry

Let {Mk }k∈N be a compact exhaustion of M. Cover Mk with a sequence of balls

B(xi ; δ1 ), where d(xi , x j ) ≥ δ1 . This sequence is finite, because an infinite sequence
{xi }i≥0 must have an accumulation point in Mk , which contradicts d(xi , x j ) ≥ δ1 .
Choosing the first xi ’s in Mk+1 to coincide with those of Mk , it follows that the union
of all balls B(xi , δ1 ) is a countable cover of M such that ∀ i  = j : d(xi , x j ) ≥ δ1 .
Let x ∈ M arbitrary. Any ball B(xi ; δ2 ) that intersects B(x; δ2 ) must be
completely contained in B(x; 3 δ2 ). Each of these balls has an exclusive subset
B(xi ; δ1 /2),
 so in normal coordinates around x, each has an exclusive  volume of 
at least Vol B(0; δ1 /(2C)) , while B(x; δ) has volume of at most Vol B(0; C 3 δ2 ) .
With n = dim(M), this leads to the explicit upper bound

(3 C δ2 )n δ n
2
K ≤ = 24 . (2.10)
(δ1 /(2 C)) n δ1

Thus, only finitely many can intersect B(x; δ2 ). These estimates are uniform and do
not depend on x ∈ M so the bound K is global. 
Lemma 2.17 (Uniform partition of unity) Let M be a manifold with a uniformly
locally finite cover with δ1 < δ2 and δ2 sufficiently small, as per Lemma 2.16.
Then there exists a partition of unity by functions χ• ∈ Cb,u k (B(x ; δ );
i 2 [0, 1])
subordinate to this cover.
We shall also apply this lemma to submanifolds which have a uniformly locally finite
cover due to Corollary 2.26 on page 54.
Proof Let δ2 be small enough that by Lemma 2.6 coordinate transition maps are
k . Define a standard radially symmetric smooth bump function ϕ ∈ C ∞ (Rn ;
Cb,u
[0, 1]) that is identically one on B(0; δ1 ) and has compact support in B(0; δ2 ), hence
k . We set ϕ = ϕ ◦ exp−1 by isometric identification T M ∼ Rn and zero
ϕ ∈ Cb,u i xi xi =
outside B(xi ; δ2 ). We have ϕ• ∈ Cb,u in any coordinate patch. Define in the usual
k

way
 
χi = ϕi ϕn . (2.11)
n≥1

The sum is finite as at most K of the B(xn ; δ2 ) overlap any B(xi ; δ2 ). The balls
B(xi ; δ1 ) already cover M, so the denominator is at least one, from which it follows
that χ• ∈ Cb,uk . 
Corollary 2.18 Similar to a uniform partition of unity, we can construct a partition
by functions χ• ∈ Cb,u
k (B(x ; δ );
i 2 [0, 1]) whose squares sum to one.

In the proof of Lemma 2.17 we simply replace (2.11) by


 
χi = ϕi ϕn2 . (2.12)
n≥1
2.2 Curvature and Holonomy 47

2.2 Curvature and Holonomy

To prove smoothness of the persistent manifold in Sect. 3.7, we shall want to estimate
the holonomy along closed loops to be close to the identity, that is, if c is a closed
loop, then we want (c) − 1 to be small. To this end, we relate the holonomy to the
curvature and finally obtain an estimate in terms of a global bound on the curvature
and the area of a surface enclosed by c.
The result that curvature is the generator of holonomy dates back at least to
Ambrose and Singer [AS53] who formulated this in differential form in the 1950’s;
they cite an even older statement (without proof) by Élie Cartan [Car26]. More recent
work by Reckziegel and Wilhelmus [RW06] shows explicit integral formulas for this
relation, formulated on fiber bundles, a context far more general than is required here.
We shall present a formulation for Riemannian manifolds (M, g).
Let  denote the parallel transport functional, which takes C 1 curves to orthogonal
maps between the tangent spaces at their endpoints, see (F.3). If c is a closed loop,
then (c) is a linear endomorphism on Tc(0) M and we can measure (c) − 1.
Our goal is to bound this quantity by the integral of the curvature form R over a
surface with boundary precisely c. This result can be viewed as a generalization
of Stokes’ theorem where the curvature is the exterior derivative of the connection
form ω, while the connection on the other hand generates parallel transport along
the boundary of the surface A that the curvature is integrated over. Note though,
that we actually have R = dω + ω ∧ ω, so there is an additional term due to the
noncommutativity of the connection form.
Let  
γ : D = 0, t¯ × [0, s̄] → M : (t, s) → γ (t, s) (2.13)

parametrize the surface A = γ (D) ⊂ M. The idea is that γ is the homotopy of

a (closed) curve c. We shall only consider parallel transport along horizontal or
vertical lines in D; let us denote by st 2 ,s1 parallel transport along s → γ (s, t) with
s ∈ [s1 , s2 ] and by st2 ,t1 parallel transport along t → γ (s, t) with t ∈ [t1 , t2 ].
We shall calculate the holonomy along ∂ A with respect to a chosen frame on the
pullback bundle γ ∗ (TM). The final result will turn out to be independent of this
choice, hence it is covariantly defined. Let f be an orthonormal frame on γ ∗ (TM),
that is, f t,s : Tγ (t,s) M → Rn is an isometry of inner product spaces. The Levi-Civita
connection ∇ on M can be pulled back to the connection γ ∗ (∇) on γ ∗(TM) and it can
be expressed in terms of the connection form ω ∈ 1 D; End(Rn ) with respect to
the frame f . The curvature of γ ∗ (∇) is equal to the curvature R of ∇ pulled back to D,
so we have dω + ω ∧ ω = γ ∗ (R) f , where the subscript f indicates that everything is
expressed with respect to the chosen frame. In the same notation, parallel transport
along a curve s → c(s) satisfies the linear, homogeneous differential equation3

3 If the frame f is induced by local coordinates, then ω will precisely be given by the Christoffel
symbols and we recover Eq. (F.4).
48 2 Manifolds of Bounded Geometry

d
f (c|s0 ) = −ω(ċ(s)) ◦ f (c|s0 ), f (c|00 ) = 1Rn , (2.14)
ds

which has a unique solution s → f s,0 = f (c|s0 ). This can be viewed as time-
dependent flow in End(Rn ).
Let us define the parallel transport term

P(s) = s̄,s
t¯
◦ s0,t¯ ◦ s,0
0 : Tγ (0,0) M → Tγ (t¯,s̄) M, (2.15)

see Fig. 2.1. The holonomy defect can be expressed as

 
1 − (∂ A) = 1 − (∂ D) = 1 − P(s̄)−1 ◦ P(0) = P(s̄)−1 ◦ P(s̄) − P(0) ,

where (∂ D) is defined using the pullback connection. We use the fundamental

theorem of calculus to write
 s̄ dP(s)
P(s̄) − P(0) = ds. (2.16)
0 ds

Expressing everything with respect to the frame f , we see that the first and last factor
of P(s) are easily differentiated using (2.14):

d s,0 ∂ d ∂
f 0 = −ω( ) ◦ f s,0 and s̄,s s̄,s
f t¯ = f t¯ ◦ ω( ). (2.17)
ds ∂s 0
ds ∂s

The middle term f st¯,0 can be differentiated by viewing s as parameter in the differ-
ential Eq. (2.14). Variation of constants yields (see e.g. [DK00, App. B] for a proof
of the differentiable dependence of a flow on parameters)
 t¯
d d ∂ 
f t¯,t = f t¯,t ◦ − ω( ) ◦ f st,0 dt
s s
ds 0 ds ∂t

t¯ ∂ ∂ d ∂  ∂ ∂
= − f st¯,t ◦ dω( , ) + ω( ) + ω( , ) ◦ f st,0 dt
0 ∂s ∂t dt ∂s ∂s ∂t

Fig. 2.1 The path of the (t , s)

parallel transport term P(s)
in D

(t , s)

(0, 0)
2.2 Curvature and Holonomy 49
∂ ∂

using standard rules for exterior derivatives. Next we note that
 ∂  ∂s , ∂t = 0, and
integrate by parts the term dtd ω( ∂s )

t¯ ∂ ∂ ∂ ∂ ∂ ∂ 
= − f st¯,t ◦ − ω( ) ◦ ω( ) + dω( , ) + ω( ) ◦ ω( ) ◦ f st,0 dt
0 ∂t ∂s ∂s ∂t ∂s ∂t
 ∂  ¯
t
− f st¯,t ◦ ω( ) ◦ f st,0
∂t t=0
 t¯
 ∂ ∂  ∂ ∂
= f t¯,t ◦ (dω + ω ∧ ω)
s
, ◦ f st,0 dt − ω( ) ◦ f st¯,0 + f st¯,0 ◦ ω( ).
0 ∂t ∂s ∂s ∂s
(2.18)

We see that this variation depends on the curvature form γ ∗ (R) f = dω + ω ∧ ω

along the path and two additional boundary terms. If we view γ as a homotopy of
paths with homotopy parameter s and we keep the path endpoints γ (0, s) and γ (t¯, s)
fixed for all s ∈ [0, s̄], then these boundary terms vanish and the result (2.18) agrees
with [RW06, Cor. 3].
Instead, we insert (2.17) and (2.18) into (2.16). Then these boundary terms cancel
against the terms from (2.17) and we finally obtain
  t¯
s̄ ∂ ∂
P(s̄) f − P(0) f = s̄,s
f t¯ ◦ f st¯,t ◦ γ ∗ (R) f (, ) ◦ f st,0 ◦ f s,0
0 dtds
0 0 ∂t ∂s
 
= s̄,s
t¯
◦ st¯,t ◦ γ ∗ (R) ◦ st,0 ◦ s,0
0 . (2.19)
D f

The integrand on the last line is a two-form on D with values in L(Tγ (0,0) M;
Tγ (t¯,s̄) M). This final expression is clearly independent of a choice of frame, so we
have recovered an explicit integral formula relating holonomy along a null-homotopic
loop to the curvature.
We conclude from (2.19) that if c is a closed, null-homotopic loop, and the cur-
vature globally bounded, then (c) − 1 can be estimated by Rsup times the
surface area of any null-homotopy γ of c. Note that we do not require γ to be
an embedding; the integral is intrinsically defined on D by pullback. Furthermore,
γ is required to be C 1 only. This follows from the fact that both sides of the equation
are continuous with respect to γ in C 1 -norm; alternatively, an explicit calculation
∂2 γ
requires that the mixed partial derivative ∂s ∂t is continuous to perform integration
by parts. Both lead to the to the following result.
Lemma 2.19 (Exponential growth bound on holonomy) Let (M, g) be a man-
ifold of bounded geometry with normal coordinate radius δ that is M-small as in
Definition 2.8. Fix T > 0 and ρ > 0 and let x1 , x2 be two C 1 curves on M with
derivatives bounded by N such that dρ (x1 , x2 ) eρ T ≤ δ < rinj (M). Denote by γt the
unique shortest geodesic connecting x1 (t) to x2 (t) for any t ∈ [0, T ].
50 2 Manifolds of Bounded Geometry

If δ is sufficiently small, then the closed loop η = x2 |0T ◦ γT ◦ x1 |0T ◦ γ0−1 satisfies
the holonomy bound

eρ T
(η) − 1 ≤ C̃ R0 N dρ (x1 , x2 ) (2.20)
ρ

where C̃ depends on the geometry of M only.

Proof The two-parameter family (s, t) → γt (s) defines a null-homotopy of the
closed loop η. The map s → γt (s) is defined through the exponential map as
 
γt : [0, 1] → M : s → expx1 (t) s exp−1
x1 (t) (x 2 (t)) .

Since expx is a local diffeomorphism at least for d(x1 (t), x2 (t)) < δ eρ t < rinj (M),
that depends smoothly on x, it follows that (s, t) → γt (s) defines a homotopy
between the curves x1 , x2 restricted to the interval [0, T ]. The map γt (s) has contin-
uous mixed derivatives with respect to s, t (even though the double derivative with
respect to t does not exist since x1 , x2 ∈ C 1 only), so integration by parts is allowed
in (2.18).
We estimate the surface area mapped by γt (s). We use shorthand notation
ξ = s exp−1x1 (t) (x 2 (t)) ∈ Tx1 (t) M and denote by Dx expx the derivative of the expo-
nential map with respect to the base point parameter x. Then

d
γt (s) =D expx1 (t) (ξ ) · exp−1 x1 (t) (x 2 (t)),
ds
d
γt (s) =Dx expx1 (t) (ξ ) · ẋ1 (t) + D expx1 (t) (ξ )·
dt
 
s Dx (exp−1 −1
x1 (t) )(x 2 (t)) · ẋ 1 (t) + s D expx1 (t) (x 2 (t)) · ẋ 2 (t) .

Since M has bounded geometry, D expx and its inverse are bounded by Theorem 2.4,
while Dx expx and its inverse are bounded by Lemma 2.6, say by C > 1. This leads
to estimates
 
d 
 γt (s) ≤ C d(x1 (t), x2 (t)),
 ds 
 
d 
 γt (s) ≤ C ẋ1 (t) + C s[C ẋ1 (t) + C ẋ2 (t)] ≤ 3 C 2 N ,
 dt 

so the holonomy bound satisfies

2.2 Curvature and Holonomy 51
 1 T d d
(η) − 1 ≤ R0  γt (s)  γt (s)dtds
0 0 ds dt
 T
≤ R0 3 C 3 N dρ (x1 , x2 ) eρ t dt
0
eρ T
≤ 3 C 3 R0 N dρ (x1 , x2 ) . 
ρ

Remark 2.20 It should be possible to obtain C̃ = 1 if the curves xi are generated by

a flow  and we choose as homotopy (s, t) → t (γ (s)), where γ is the geodesic
connecting x1 (0) and x2 (0). In our applications, though, the curves x1 , x2 need not
be solutions to exactly the same flow, while the current result is sufficient for our
purposes. ♦

2.3 Submanifolds and Tubular Neighborhoods

From this section on, we shall prove results that—although they may be of interest
independently within bounded geometry—are building up towards the final section
of this chapter, where we prove how to reduce Theorem 3.1 on persistence in general
manifolds of bounded geometry to the setting of a trivial bundle. These results form
the more technical part of this chapter and are not required elsewhere.
In the following, we assume that (Q, g) is an ambient manifold that has bounded
geometry of large or infinite order and M ∈ C k will denote a submanifold of Q. Only
a finite order l > k of bounded geometry is required of (Q, g), but for simplicity we
shall assume l = ∞. Recovering the explicit additional order l −k would amount to
tediously tracking the details throughout all the proofs; it should be sufficient if l is
larger than k by some number between 2 and 10.
Let ι : M → Q be a C 1 immersion. With abuse of notation we denote by
Tx M = Im(Dι(x)) and N x = Im(Dι(x))⊥ the tangent and normal spaces of M with
respect to the immersion. Note that even if ι is not injective, the original point x ∈ M
uniquely selects the tangent and normal spaces in Tι(x) Q.
Definition 2.21 (Uniformly immersed submanifold) Let ι : M → Q be a C k≥1
immersion of M into the Riemannian manifold (Q, g) of bounded  geometry. Denote

by Mx,δ the image under ι of the connected component of x in ι−1 B(ι(x); δ)∩ι(M) .
k immersed submanifold when there exists a δ > 0 such that
We define M to be a Cb,u
for all x ∈ M, the connected component Mx,δ is represented in normal coordinates
on B(ι(x); δ) ⊂ Q by the graph of a function h x : Tx M → N x and the family of
functions h • ∈ Cb,u
k (T M; N ) has uniform continuity and boundedness estimates
• •
independent of x. We define Cbk≥1 immersions in a similar way.
52 2 Manifolds of Bounded Geometry

Nx Nx
B (0; δ )
Nx
hx

x′ x′
Tx M Tx M
ξ′ 0 0

Fig. 2.2 An immersed submanifold represented by the graph of h x in normal coordinates. In the
left figure, another part of M intersects transversely on the right; the right figure contains an orbit
of the geodesic flow along a normal vector at x  ∼= ι(x  )

Remark 2.22 By taking the connected component Mx,δ in M, we allow for immersed
submanifolds that intersect, or nearly intersect themselves. See Fig. 2.2 on the left:
Mx,δ is described by the graph of h x , while on the right side, a different part of
M embeds into this same neighborhood B(ι(x); δ). See Fig. 3.2 on page 86 for
an example of a nearly self-intersecting submanifold. If we want to rule out such
cases, we can assume that Mx,δ is the unique component of M ∩ B(ι(x); δ). This
will turn M into an embedded submanifold, but more strongly, the nearly self-
intersecting case is also ruled out. We will refer to this as a uniformly embedded
submanifold. ♦
Remark 2.23 The sets Mx,δ play a similar role as ‘plaques’ in [HPS77, p. 72–73].
♦
Remark 2.24 In case k = 1, boundedness is automatically implied by uniform conti-
nuity. This follows from the representation in normal coordinates. We have Dh x (0) =
0, so by uniform continuity there exists a δ > 0 such that Dh x (ξ ) < ε = 1 when
ξ  < δ, hence Dh x is bounded. Put another way, there is no intrinsic measure for
the ‘size of the derivative or tangent’ of a submanifold. ♦
Note that the function h x is only defined on that part of the domain B(0; δ) ⊂ Tx M
where its graph is contained in B(0; δ) ⊂ Tι(x) Q, as can be seen in Fig. 2.2. In the
splitting Tι(x) Q = Tx M ⊕ N x , we denote with p1 , p2 orthogonal projections onto
the Tx M and N x subspaces, respectively.
k≥1
From now on we shall continually assume that M ∈ Cb,u is a uniformly immersed
submanifold of Q. We will often identify M with its image ι(M) ⊂ Q, as well as
identify points x ∈ M with ι(x), keeping in mind the definition of Mx,δ to track local
injectivity. Furthermore, denote by d M the distance on M induced by the pulled back
Riemannian metric ι∗ (g). This distance function measures whether points are close
when viewed along the domain of the immersion, disallowing ‘shortcuts’ through Q.
2.3 Submanifolds and Tubular Neighborhoods 53

It also distinguishes different points with the same immersion image. Note that it is
different from the distance d on Q pulled back to M. This we denote by d Q = ι∗ (d)
but it is not a distance on M when ι is not injective. Still, we have the following local
result, which will be useful for later estimates.
Lemma 2.25 (Local equivalence of distance) Let M ∈ Cb,u 1 be a uniformly
immersed submanifold of the bounded geometry manifold (Q, g). Then d Q and d M
are locally equivalent in the following sense:
1. ∀x1 , x2 ∈ M : d Q (x1 , x2 ) ≤ d M (x1 , x2 );
2. for any C  > 1 there exists a δ > 0 such that for all d M (x1 , x2 ) < δ, we have the
local converse d M (x1 , x2 ) ≤ C  d Q (x1 , x2 ).
Proof The first assertion follows directly from the fact that any path in M induces
a path of equal length in Q via the immersion ι.
For the second part, we first note that if δ is small enough and d M (x1 , x2 ) < δ, then
we must have x2 ∈ Mx1 ,δ . If this would not be the case, then any path γ connecting
x1 , x2 through M cannot be contained in Mx1 ,δ . But this implies that the path runs
out of B(x1 ; δ), so its length is greater than δ. This contradicts the assumption that
d M (x1 , x2 ) < δ. Hence, x2 can be represented as a point on the graph of h x1 in
B(x1 ; δ).
Let C > 1, ε > 0 be constants to be fixed later and let δ be small enough
such that the metric coefficients are bounded by C in normal coordinate charts, that
Proposition 2.5 holds with C, and we have h • 1 ≤ ε as in Remark 2.24. We consider
the normal coordinate chart on B(x1 ; δ) and construct a path in M to find an upper
bound for d M (x1 , x2 ). Let x2 = (ξ, h x1 (ξ )) and define γ (t) = (t ξ, h x1 (t ξ )) for
t ∈ [0, 1]. We estimate the length of γ as
 1
√

l(γ ) ≤ g 1 + h • 21 ξ dt ≤ C 1 + ε2 ξ ,

0

while the Euclidean norm can be estimated by the distance in Q as

ξ  ≤ (ξ, h x1 (ξ )) ≤ C d(x1 , x2 ).

We conclude that

d M (x1 , x2 ) ≤ l(γ ) ≤ C 3/2 1 + ε2 d Q (x1 , x2 )

√
and for any C  > 1 we can find C > 1, ε > 0 such that C 3/2 1 + ε2 < C  . 
A uniform submanifold of a bounded geometry manifold can be shown to possess
a uniformly locally finite cover as a corollary of Lemma 2.16, without the need to
show that the submanifold itself has bounded geometry. As a consequence, it also
has (square-sum) partitions of unity.
54 2 Manifolds of Bounded Geometry

Corollary 2.26 (Uniform cover of a submanifold) Let M ∈ Cb,u 1 be a uniformly

immersed submanifold of the bounded geometry manifold (Q, g).

Then for δ2 > 0 small enough and any δ1 ∈ (0, δ2 ], M has a uniformly locally
finite cover by balls of radius δ2 in terms of the distance d Q , such
 that the balls of
radius δ1 already cover M. That is, there exist {xi }i≥1 such that i≥1 Mxi ,δ1 covers
M with a uniform bound K on the maximum number of sets Mxi ,δ2 covering any set
Mx,δ2 with x ∈ M.
Proof The proof follows the ideas of Lemma 2.16. As an additional requirement, let
δ > 0 be sufficiently small such that each Mx,δ is represented in normal coordinates
by the graph of h x . Under this assumption, the open sets Mx,δ are induced by d Q
and correspond to the connected component of x of the pre-image of B(ι(x); δ).
Consequently, we can locally push the argument to ι(M) ⊂ Q to conclude that there
is an upper bound K on the number of sets Mxi ,δ2 that intersect any set Mx,δ2 . 
Even though we do not require submanifolds to have bounded geometry for the
results in this section, the lemma below will be needed in the final reduction to a
trivial bundle. The essential idea of the proof is to use Gauß’ second fundamental
form to relate curvature of the submanifold to second derivatives of its immersion
map.
Lemma 2.27 (Submanifold of bounded geometry) Let M ∈ Cbk≥2 be a uniformly
immersed submanifold of the bounded geometry manifold (Q, g). Then (M, ι∗ (g))
is a Riemannian manifold with bounded geometry of order k − 2.
Remark 2.28 We lose two orders of smoothness in the bounded geometry definition.
This is due to bounded geometry being defined in terms of the curvature, which
depends on second order derivatives of the metric, and in this case also on second
order derivatives of the embedding through Gauß’ second fundamental form. ♦
Proof Let δ be sufficiently small such that for each x ∈ M we have the representation
Mx,δ = Graph(h x ) with Dh x  ≤ 1.
The Riemann curvature tensor R M of M can be expressed as a sum of the curvature
R on Q and the second fundamental form of the (local) embedding, see e.g. [Jos08,
Thm. 3.6.2]:

g(R M (X, Y ) Z , W ) = g(R(X, Y ) Z , W ) + g(S(Y, Z ), S(X, W ))

− g(S(Y, W ), S(X, Z )), (2.21)

where
S : TM × TM → N : X, Y → (∇ X Y )⊥ (2.22)

is the second fundamental form, and it is indeed pointwise defined. In normal coor-
dinates we find
Sx (X, Y ) = D2 h x (0)(X, Y ). (2.23)
2.3 Submanifolds and Tubular Neighborhoods 55

Since h ∈ Cbk and g, g −1 ∈ Cbk as well, it follows that S ∈ Cbk−2 and by (2.21) then
that R M ∈ Cbk−2 , so condition (Bk−2 ) of Definition 2.1 is satisfied.
Condition (I) on the injectivity radius follows from an implicit function argument
applied to the geodesic flow using Theorem A.6. We consider local coordinates
around x ∈ M by projecting the representation M ∩ B(x; δ) onto Tx M in normal
coordinates in Q. That is, we have the coordinate chart map

κx : B(0; δ/2) ⊂ Tx M → M : ξ → expx (ξ, h x (ξ ))

and the corresponding embedding into normal coordinates Tx M → Tx Q : ξ →

(ξ, h x (ξ )) of Q. We calculate explicit estimates for the exponential map expxM using
Christoffel symbols of the connection ∇ M on M in the coordinates in chart κx .
Let X, Y be vector fields on M. Their representation in κx is mapped to normal
coordinates B(x; δ) on Q as
 T
X (ξ ) → X̃ (ξ ) = 1, Dh x (ξ ) · X (ξ ).

Hence, from the covariant derivative on M in normal coordinates B(x; δ) ⊂ Q we

can recover the Christoffel symbols in local coordinates κx as
 ∂  
∇ XM Y = p1 ◦ X i (ξ ) i Y (ξ ) + (ξ, h x (ξ )) X̃ (ξ ), Ỹ (ξ ) ,
∂ξ

where the first term has reduced to derivatives with respect to ξ ∈ Tx M ⊂ Tx Q

only, and : B(x; δ) → L2 (Tx Q; Tx Q) are the Christoffel symbols in normal
coordinates at x ∈ Q. Thus, the Christoffel symbols
 
M (ξ )(X, Y ) = p1 ◦ (ξ, h x (ξ )) X̃ , Ỹ (2.24)

of M in κx coordinates are uniformly bounded on sufficiently small balls B(0; δ  ) ⊂

Tx M. The Euclidean geodesic flow at time one defines the (trivial) Euclidean expo-
nential map, which is an isomorphic diffeomorphism (with infinite injectivity radius
actually). Since we study a small perturbation of this flow in local coordinates, given
by the additional term (2.24), and the perturbation is at least Cbk−1 and C 1 small,
the perturbed geodesic flow of M can be made close enough that expxM is still a dif-
feomorphism on B(0; δ  ) for some δ  > 0. Hence, rinj (x) ≥ δ  , but these estimates
depend only on the perturbation size, so they hold uniformly for all x ∈ M. 
To obtain the final result of this section, the tubular neighborhood theorem 2.33,
we first need to work out some details on local coordinates. If M is a submanifold
of Q, it is natural to consider a specific splitting on the normal coordinates at points
x ∈ M, namely Tx Q = Tx M ⊕ N x , where N is the normal bundle over M. We shall
require bounds, not just on coordinate transformations, but more specifically bounds
on how well this splitting is preserved. The lemmas are formulated in a more general
56 2 Manifolds of Bounded Geometry

context of splittings of tangent spaces at any two nearby points, while the results for
coordinates along M follow as an easy corollary.
Lemma 2.29 (Coordinate transformations of splittings) Let (Q, g) be a smooth
Riemannian manifold of bounded geometry, let C be sufficiently large and let δ, ζ > 0
be sufficiently small. Let x1 , x2 ∈ Q and let Txi Q = Hi ⊕ Vi , i = 1, 2 be split-
tings along ‘horizontal’ and ‘vertical’ perpendicular subspaces with dim(H1 ) =
dim(H2 ). Assume that d(x1 , x2 ) < δ and that, for i  = j, Hi is represented in
tangent normal coordinates at x j by the graph of L i ∈ L(H j ; V j ) with L i  ≤ ζ .
Then the coordinate transformation ϕ2,1 in Lemma 2.6 is of the form
 
ϕ2,1 = O H ⊕ OV + ϕ̃2,1 with ϕ̃2,1 k ≤ C ζ + d(x1 , x2 ) , (2.25)

where O H , OV are orthogonal transformations between the Hi and Vi with i = 1, 2,

respectively.
We first prove the following result and use it to prove Lemma 2.29.
Lemma 2.30 (Approximation of orthogonal maps) Let V be a finite-dimensional
inner product space and define the map

f : L(V ) → Sym(V ) : A → A T A − 1. (2.26)

There exists an ε > 0 and a tubular neighborhood B(O(V ); η) ⊂ G L(V ) with fiber
projection π, such that on {A ∈ G L(V ) |  f (A) < ε,  A ≤ 2}, the map ϕ : A →
(π(A), f (A)) is a smooth diffeomorphism. As a direct corollary, if  f (A) < ε and
A ≤ 2 then U = π(A) ∈ O(V ) is an orthogonal approximation of A in the sense
that U − A ≤  f (A).
Proof The map f is smooth and invariant under the left action of the orthogo-
nal maps O(V ), while O(V ) = ker( f ). Since G L(V ) is a Lie group, we have
the canonical trivialization TG L(V ) = G L(V ) × gl(V ) by left multiplication.
The similar trivialization TO(V ) = O(V ) × o(V ) can be viewed as a subbundle of

O(V ) × o(V ) ⊕ Sym(V ) = TG L(V )| O(V ) ,

where o(V ) is identified with the skew-symmetric linear maps. We restrict the
exponential map exp : TG L(V ) → G L(V ) to O(V ) × Sym(V ). At 1 ∈ O(V )
this restriction has bijective derivative, hence it is a local diffeomorphism. Since
exp is O(V )-invariant, it defines a diffeomorphism onto a tubular neighborhood
B(O(V ); η) ⊂ G L(V ) of O(V ) of size η > 0 and a corresponding smooth fiber
projection map π : B(O(V ); η) → O(V ).
Now D f (1) : a → a T + a has image precisely Sym(V ). Thus, if we restrict
f to the fiber over 1 ∈ O(V ) in the tubular neighborhood, then D f (1)|Sym(V ) =
2 and f is a diffeomorphism with D f −1  ≤ 1 in some neighborhood of 0 ∈
π −1 (1); if necessary, we reduce η > 0 for D f −1  ≤ 1 to hold on B(O(V ); η) ∩
2.3 Submanifolds and Tubular Neighborhoods 57

π −1 (1). By O(V ) invariance of f , this holds globally on all (fibers) of the tubular
neighborhood. Since, Dπ and D f have complementary image at O(V ), ϕ = (π, f )
is a diffeomorphism on B(O(V ); η).
The set B(0; 2) \ B(O(V ); η) ⊂ L(V ) is compact, so  f ( · ) attains its nonzero
minimum on it. Let ε be smaller than this minimum. Then, if  f (A) < ε, we
must have A ∈ B(O(V ); η) and hence A = exp(U, a) for a unique (U, a) ∈
O(V ) × Sym(V ). By O(V )-invariance, we can assume w.l.o.g. that U = 1 and use
the mean value theorem to estimate
 
 
A − 1 ≤ D f −1   f (A) − f (1) ≤  f (A) < ε.

In other words, when A is sufficiently close to being orthogonal, measured according

to f , then it is close to an orthogonal map U in operator norm. 
Proof of lemma 2.29 Extending the results of Lemma 2.6, let

O = (γ2,1 ) : Tx1 Q → Tx2 Q

denote the orthogonal linear map induced by parallel transport. We decompose ϕ2,1 =
O + ϕ̂2,1 , where ϕ̂2,1 can be made arbitrarily small. Moreover, we write
 
A B
O= ∈ L(H1 ⊕ V1 ; H2 ⊕ V2 ),
C D

with the idea that B, C should be small and A, D should approximate orthogonal
maps O H , OV , respectively. Orthogonality of O implies
     T 
AT C T A B A A + C T C AT B + C T D
1=O O= T
· = .
B T DT C D B T A + DT C B T B + DT D

For the operator norm we have A, B, C, D ≤ O = 1, so if we assume
for the moment that B, C can be made sufficiently small, then, by writing A T A−1 =
−C T C and D T D − 1 = −B T B, Lemma 2.30 implies that we can find O H , OV
such that
      
 A0   A0 OH 0 

O − O H ⊕ OV  ≤  O −  +  − 
0 D   0 D 0 OV 
≤ B + C + A − O H  + D − OV 
   
   
≤ B + C + C T C  +  B T B . (2.27)

In normal coordinates around x2 we have H1 = Graph(L), so C : H1 → V2 is

represented by L in these coordinates. The metric g is close to the identity in these
coordinates, so C ∼
= L can be assumed bounded by 4 ζ ≤ 1, as measured in the
−1
metric on Q. The same argument can be made for B T by considering ϕ1,2 = ϕ2,1 ,
58 2 Manifolds of Bounded Geometry

since  
AT C T
O −1 = O T = .
B T DT

We conclude that both B, C ≤ 4 ζ when δ is chosen small, hence O can be
approximated by O H ⊕ OV , and the error from (2.27) can be absorbed into ϕ̃2,1 :

ϕ̃2,1 = ϕ̂2,1 + (O − O H ⊕ OV ).

The errors introduced in ϕ̃2,1 from lemmas 2.6 and 2.30 are Lipschitz small in terms
of d(x1 , x2 ) and ζ , respectively, so these add up to the estimate in (2.25). 
k≥1
Corollary 2.27 Let M ∈ Cb,u be a uniformly immersed submanifold of a smooth
Riemannian manifold (Q, g) of bounded geometry. Let x 1 , x2 ∈ M and let Txi Q =
Txi M ⊕ N xi , i = 1, 2, be the respective splittings in horizontal and vertical direc-
 hold ford M (x1 , x2 ) < δ. If moreover M ∈ Cb ,
tions. Then the results of Lemma 2.29 2

then we have a Lipschitz estimate Dϕ̃2,1 (0) ≤ C d(x1 , x2 ).

Proof This follows immediately from the local representation Mx2 ,δ = Graph(h 2 )
as Tx1 M is represented in tangent normal coordinates at x2 by L = Dh 2 (ξ ), where
x1 = (ξ, h 2 (ξ )). And Dh 2 (ξ ) becomes small when δ is small. The same holds with
x1 , x2 interchanged.  
If M ∈ Cb2 , then we can estimate Dh • (ξ ) ≤ D2 h •  ξ  ≤ C d(x1 , x2 ).
Hence, the Lipschitz result in Lemma 2.29 transforms into a Lipschitz estimate in
d(x1 , x2 ) only. 
Below we define when a mapping is approximately isometric, see for example
also [Att94, p 505]. The Lyapunov exponents of a dynamical system are preserved
under these quasi-isometries since the exponential growth dominates any bounded
factors when measuring sizes. This property is required when we transfer a noncom-
pact normally hyperbolic system to a different space and want normal hyperbolicity
to be preserved.
Definition 2.30 (quasi-isometry) Let M, N be manifolds with distance metrics
d M , d N and let ϕ : M → N be a diffeomorphism. we call ϕ a C-quasi-isometry with
C > 1, if

∀x, y ∈ M : C −1 d M (x, y) ≤ d N (ϕ(x), ϕ(y)) ≤ C d M (x, y). (2.28)

We simply call ϕ a quasi-isometry if there exists an unspecified C > 1.

We conclude this section with a version of the tubular neighborhood theorem that
is appropriate in the bounded geometry setting.
Theorem 2.31 (Uniform tubular neighborhood) Let M ∈ Cbk≥2 be a uni-
formly immersed submanifold of the bounded geometry manifold (Q, g). Then for
η > 0 sufficiently small (but depending explicitly on M and Q), the η-sized tubular
2.3 Submanifolds and Tubular Neighborhoods 59

neighborhood B(M; η) = {y ∈ Q | d(y, M) ≤ η} can be represented on the η-sized

normal bundle N≤η of M by a diffeomorphism ϕ, locally on each N≤η | Mx,δ and we
have ϕ, ϕ −1 ∈ Cb,u
k−1
(hence ϕ is a quasi-isometry).
When moreover M is uniformly embedded, i.e. Mx,δ = M ∩ B(x; δ) for each
x ∈ M as in Remark 2.22, then ϕ is a global diffeomorphism.
In case M is compact, the standard proof uses the fact that the exponential map
has bijective differential at the zero section, and then by compactness it must be a
diffeomorphism on a uniform neighborhood N≤η of the zero section. Here, to get a
uniform neighborhood N≤η on which ϕ = exp | N≤η is a diffeomorphism, we require
bounds on second order derivatives (that is curvature, cf. Lemma 2.27) of M so that
it has curvature radius r bounded from below, hence cut locus points can only occur
at least at distance r away from M, making ϕ injective for η < r . See Fig. 2.2 for a
representation of a submanifold M in normal coordinates around x and a ray of the
normal bundle at a nearby point x  .
Note that for an immersed submanifold, we define the normal bundle as
  
N = (x, ν) ∈ M × TQ|ι(x) = π(ν), ν⊥Im Dι(x) . (2.29)

This can again be viewed as immersed into TQ.

Proof We set ϕ = exp | N and in the following we will implicitly apply Theorem 2.4
and Proposition 2.5 to choose 0 < δ ≤ 1 small enough such that the metric g up to
its second order derivatives is bounded, as well as that the Christoffel symbols are
bounded. Also, we choose δ such that M is uniformly locally representable by graphs
according to Definition 2.21. We will in sequence prove local and global injectivity,
surjectivity of ϕ and finally that ϕ, ϕ −1 ∈ Cb,u
k−1
.
We claim that for some η > 0, ϕ is locally injective on N≤η , the normal bundle
restricted to size η. Let ν, ν  ∈ N be such that ϕ(ν) = ϕ(ν  ) and denote by x =
π(ν), x  = π(ν  ) their base points in M. We consider normal coordinates at x, hence
we have ν = (0, σ2 ) for some σ2 ∈ N x , while ν  is given by (σ1 , σ2 ) ∈ Tx M ⊕ N x .
From Corollary 2.31 it follows for small δ that ν  is nearly mapped onto N x in normal
coordinates at x. Since M ∈ Cb2 , the deviation from mapping onto N x is Lipschitz
small in d(x  , x) ∼
= ξ , so we have

σ1  ≤ C ξ  σ2 . (2.30)

Now, ϕ(ν  ) = ϕ(ν) can only hold if the respective horizontal coordinates along Tx M
are equal. By definition of normal coordinates around x, we have exp(ν) = (0, σ2 ).
Therefore it is sufficient to prove that some η > 0 exists as a lower bound for
   
ν  ∈ R | ϕ(ν  )1 = 0, π(ν  )  = x .

We view the exponential map as the time-one geodesic flow, which is given in local
coordinates by (2.5). The geodesic flow along ν  = (σ1 , σ2 ) starting at (ξ  , h(ξ  )) is
60 2 Manifolds of Bounded Geometry

a small perturbation of the flow along (0, σ2 ) starting at (0, 0). The latter has solution
curve t → (0, t σ2 ) ∈ Tx M ⊕ N x .
By Theorem 2.4 we arrange for g − 1,   ≤ 2 and D  ≤ C in local
coordinates. We have estimates
     
 (x  (t)) − (x(t)) ≤ D  x  (t) − x(t) ≤ C x  (t) − x(t),
  

√  
σ (t) ≤ g(σ  (t), σ  (t)) = g(σ  (0), σ  (0)) ≤ 2 σ  (0).

With these, we obtain the Gronwall-like estimates

d   
x  (t) − (0, t σ  ) ≤ σ  (t) − (0, σ  ),
2 2
dt
d    
σ  (t) − (0, σ  ) ≤ D  x  (t) − (0, t σ  ) σ  (t)2
2 2
dt      
+   σ  (t) + (0, σ2 ) σ  (t) − (0, σ2 )
 √  
≤ 4 C η2 + 4 2 η σ  (t) − (0, σ  ),2

for which Gronwall’s inequality yields

    
σ (t) − (0, σ  ) ≤ (σ  (0), 0) eC̃ η t .
2 1

Now, if η is chosen sufficiently small, then using (2.30), we have for all 0 ≤ t ≤ 1
that  t
     1 
x (t) − x1 (0) ≤ C ξ   η eC̃ η t dτ ≤ ξ  .
1
0 2

This shows that there exists an explicit η > 0 such that ϕ is injective on N≤η
restricted to a neighborhood Mx,δ , and by construction this η is uniform over M.
Later modifications to choose η smaller will only depend on the global geometry of
Q, but not on any details of M.
If moreover Mx,δ = M ∩ B(x; δ) is the unique connected component of M in each
normal coordinate chart, then ϕ is injective globally on N≤η . This follows easily by
taking η < 2δ . Then, any ν  and ν that have the same image, must have base points
x  , x separated by a distance less than δ, as ϕ = exp will only map onto points at
most η away from the base point. Therefore, x  must lie in B(x; δ) and in M, hence
on Graph(h) = Mx,δ . This case was already treated.
Finally, we will show that ϕ is surjective onto B(M; η) when η < rinj (Q). Take
y ∈ B(M; η), then M ∩ B(y; rinj (Q)) contains a nonempty compact set, so there
exists an x ∈ M such that d(y, M) = d(y, x). This distance must be realized by a
(unique) geodesic γ . We will derive a contradiction if γ  (0)  ∈ N x , by showing that
then the minimum distance is not attained at x. Let (ξ1 , ξ2 ) ∈ Tx M ⊕ N x be the
normalized tangent vector of γ  (0). By assumption we have ξ1  = 0, so ξ2  < 1.
We parametrize
γ : [0, d(y, x)] → Tx Q : t → t (ξ1 , ξ2 )
2.3 Submanifolds and Tubular Neighborhoods 61

by arc length in normal coordinates, thus d(x, γ (t)) = t. Consider the Euclid-
ean distance in normal coordinates at x of γ (t) to its vertical projection onto
M = Graph(h). This shows that

d E (γ (t), M) ≤ t ξ2 − h(t ξ1 ) ≤ t ξ2  + o(t ξ1 )

as Dh(0) = 0 and h ∈ Cb2 . The Euclidean distance is C-equivalent to the g-induced

distance, so we have

d(γ (t), M)
lim ≤ lim C ξ2  + o(t)/t = C ξ2 .
t↓0 d(x, γ (t)) t↓0

By assumption ξ2  < 1, so we can restrict to a small enough neighborhood B(x; δ)

such that 1 < C < ξ2 −1 and conclude that d(γ (t), M) < d(x, γ (t)) for some
t > 0, which shows that a shorter (broken) geodesic from y to M exists. This
completes the contradiction and proves that ϕ is surjective.
k−1
Finally, ϕ ∈ Cb,u follows directly from the fact that it is the restriction of the
exponential map to N≤η ∈ C k−1 and in induced normal coordinate charts, we have
k−1
exp ∈ Cb,u . For ϕ −1 we use a formula and arguments similar to (A.2), showing that if
Dϕ −1 is uniformly bounded, then ϕ −1 ∈ Cb,u k−1
holds as well. Now Dϕ = 1 in induced
normal coordinates at M, so by uniform continuity, there exists some η > 0 such
that Dϕ stays away from non-invertibility on N≤η , hence Dϕ −1 stays bounded.
 This
automatically implies that ϕ is a quasi-isometry with C = max(Dϕ, Dϕ −1 ). 

2.4 Smoothing of Submanifolds

It is well-known, at least in the compact case, that r -normal hyperbolicity is a

persistent property under C 1 small perturbations of class C r , that is, the persist-
ing manifold is again r -normally hyperbolic and specifically C r , see [HPS77, Thm.
4.1] or [Fen72, Thm. 2]. In other words, r -normal hyperbolicity is an ‘open prop-
erty’ in the space of C r systems with C 1 topology. Therefore, it is natural to only
assume that the original manifold is C r , but not smoother. Even if we start out with an
r -NHIM M ∈ C ∞ , then after a perturbation we will generally only have a manifold
Mε ∈ C r . We could, however, also have tried to obtain this manifold Mε by first
perturbing M to an intermediate manifold Mε/2 and then perturb that manifold to
Mε . When applying a persistence theorem in the second step, we can only assume
the initial manifold to be C r .
This restricted C r smoothness assumption forces us to be careful about the precise
smoothness of each and every object. For example, a vector field on a C r manifold
can only be C r −1 . This could probably be overcome by considering discrete-time
mappings instead of flows, but we need other smoothness improvements as well. For
example, we want to model the persisting manifold as a section of the normal bundle
62 2 Manifolds of Bounded Geometry

N ∈ C r −1 of the original manifold, which is not smooth enough. So here we need a

smoothing argument as well, cf. [Fen72, p 205].
We solve these problems by constructing an approximate, smoothed manifold
Mσ ∈ C ∞ . This allows M to be modeled as a small section of the normal bundle
Nσ of Mσ , so the system in a neighborhood of M can be transferred to Nσ while
preserving smoothness and normal hyperbolicity properties. With this construction
we need not worry about smoothness in the proof, while the conclusions are preserved
up to C r smoothness. Uniform estimates must be preserved though, so standard
methods for constructing Mσ and Nσ do not readily apply or need a careful analysis.
First, we construct σ > 0 close approximations Mσ ∈ C ∞ to M by globalizing a
local chart construction of smoothing by convolution with a mollifier. Then we use
the fact that Mσ has uniformly bounded ‘second-order derivatives’ to show that for a
sufficiently small σ > 0, Mσ has a normal bundle diffeomorphic to a neighborhood
B(M; δ) ⊂ Q of uniform size.
We recall some standard techniques on Rn , see for example [Hör03, p 25]. Let
ϕν ∈ C0∞ (Rn ; R≥0 ) be a mollifier function, with support in B(0; ν) and integral
normalized to one for any ν > 0. We also define a generic cut-off function χα,β ∈
C ∞ (R; [0, 1]) such that 
1 if x ≤ α,
χα,β (x) = (2.31)
0 if x ≥ β.

Note that ϕν , χα,β ∈ Cb,u

k for any k ≥ 0, as they are constant outside compact sets.

k≥0 
Lemma 2.34 (Smoothing by convolution) Let r, δr > 0, f ∈ Cb,u B(0; r +

2 δr ) ⊂ R ; R , and fix l > k and ε > 0. If the mollifier support radius ν > 0 is
m n

chosen sufficiently small, then f can be approximated by a function f˜ such that

1. f˜ = f outside B(0; r + δr );
2. f˜ ∈ Cb,u
l ∩ C ∞ on B(0; r ) and wherever f ∈ C l ∩ C ∞ ;
b,u
3.  f˜ − f k ≤ ε;
4.  f˜l ≤ C(ν, l)  f 0 on B(0; r ), for some C(ν, l) > 0.
Note that C(ν, l) may grow unboundedly as ν → 0 or l → ∞.
Proof A function that is Cbl+1 is automatically Cb,u
l , that is, uniformly continuous

up to one degree less, so we only need to prove f˜ ∈ Cbl ∩ C ∞ for l shifted by one.
We construct f˜ by a combination of convolution and cut-off. Let
χ̂(x) = χr,r +δr (x) for x ∈ Rm and define

f˜(x) = (1 − χ̂ (x)) f (x) + χ̂(x) f (x − y) ϕν (y)dy. (2.32)
Rm

When ν < δr/2, this f˜ is smooth on B(0; r ) and equal to f outside B(0; r + δr ).
The convolution approximates f in C k -norm, as for any 0 ≤ j ≤ k and x ∈
B(0; r + δr )
2.4 Smoothing of Submanifolds 63

D j (ϕν ∗ f )(x) − D j f (x)


≤ D j f (x − y) − D j f (x)ϕν (y)dy ≤ εD j f (ν),
B(0;ν)

so by uniform continuity of f up to kth derivatives, ν can be chosen small enough

such that (ϕν ∗ f ) − f k ≤ ε on B(0; r + δr ). The map x → χ̂(x) is Cb,u
k , so we

can estimate for j ≤ k

j  
 j  
D f˜(x) − D j f (x) ≤
j
Di χ̂ (x) · D j−i ϕν ∗ f − f (x) ≤ C j ε(ν).
i
i=0

Hence, we can construct f˜ close enough to f in C k -norm by choosing ν small

enough.
Uniform continuity of f˜ follows from uniform continuity of ϕν ∗ f as χ̂ ∈ Cb,u
k

on its compact support. We find for 0 ≤ j ≤ k

D j (ϕν ∗ f )(x2 ) − D j (ϕν ∗ f )(x1 )


≤ D j f (x2 − y) − D j f (x1 − y) ϕ(y)dy ≤ εD j f (x2 − x1 ).
B(0;ν)

To estimate bounds for higher derivatives of f˜ within B(0; r ), we note that χ̂ = 1

and let the derivatives act on ϕν in the convolution: these are bounded on the compact
domain of support, but bounds will depend on the size ν and degree l, while  f 0
can be factored out. 
The smoothing technique in Lemma 2.34 is formulated for Euclidean space.
To adapt it to manifolds in a uniform setting, we need to have uniformly sized coordi-
nate charts, as well as uniform behavior of the function under these smoothing oper-
ations. We cannot simply use local coordinates and a partition of unity, because the
images on different charts cannot be glued together on the target manifold. Instead, we
will apply this smoothing operation sequentially on each coordinate chart in a cover.
We require a cover that is locally finite with a global upper bound K on the number
of charts covering a point, so that each point undergoes only a bounded number
of smoothing operations and hence the final smoothed manifold Mσ differs by a
controllable amount from the original M.
When the graph representation of M in one chart is modified, we need control
on how much the graph is modified in overlapping charts. To this end, we extend
Lemma 2.29 and Corollary 2.31.
Lemma 2.35 (Graph difference under coordinate transformations) Let (Q, g)
be a smooth Riemannian manifold of bounded geometry. Let x1 , x2 ∈ Q and let
Txi Q = Hi ⊕ Vi , i = 1, 2 be splittings along horizontal and vertical perpendicular
subspaces with dim(H1 ) = dim(H2 ). Assume that d(x1 , x2 ) < δ and that, for i  = j,
64 2 Manifolds of Bounded Geometry

Hi is represented in normal coordinates at x j by the graph of L i ∈ L(H j ; V j ) with

L i  ≤ ζ .
k≥1  
Let f 1 , g1 ∈ Cb,u B(0; δ) ⊂ H1 ; V1 with  f 1 1 , g1 1 ≤ ε and  f 1 k , g1 k ≤
C.
When δ, ζ, ε > 0 are sufficiently small, then there exists a constant C̃ such that
the graphs of f 1 , g1 are (partially) represented by functions f 2 , g2 ∈ Cb,u k (H ; V )
2 2
and  f 2 − g2 k ≤ C̃  f 1 − g1 k . This result is uniform for all x1 , x2 ∈ Q.
Remark 2.36 The functions f i , gi may only be defined on parts of B(xi ; δ); all
claims should thus be read as only for those points where the respective functions
are defined. ♦
Proof Let (ξi , ηi ) ∈ Hi ⊕ Vi , i = 1, 2 denote normal coordinates, decomposed
in the split directions at xi ∈ Q. By Lemma 2.29, transformations between these
coordinates are of the form (2.25), where ϕ̃2,1 k+1 can be made uniformly small as
δ, ζ → 0.
We aim to apply the implicit function theorem to find a function f 2 on B(x2 ; δ)
whose graph corresponds to that of a function f 1 on B(x1 ; δ). We define

X = H1 × V2 , Y = H2 × Cb,u
k
(H1 ; V1 ), Z = H2 × V2 , and
F : X × Y → Z : (ξ1 , η2 ), (ξ2 , f 1 ) → ϕ2,1 (ξ1 , f 1 (ξ1 )) − (ξ2 , η2 ). (2.33)

Note that X and Z are isomorphic vector spaces, so we can apply the implicit function
theorem with Y as parameter space. Moreover, if we have two functions f 1 , f 2 whose
graphs represent the same manifold on the intersection B(x1 ; δ) ∩ B(x2 ; δ), then we
have  −1   
F p1 ◦ ϕ2,1 ξ2 , f 2 (ξ2 ) , f 2 (ξ2 ), ξ2 , f 1 = 0

for all ξ2 ∈ H2 where this is defined, so the implicit function

 −1 
G(ξ2 , f 1 ) = p1 ◦ ϕ2,1 (ξ2 , f 2 (ξ2 )), f 2 (ξ2 )

encodes the representation f 2 . We verify the conditions of the implicit function

theorem:
 
  O H + p1 · Dϕ̃2,1 · (1 + D f 1 ) 0
D1 F (ξ1 , η2 ), (ξ2 , f 1 ) =
OV · D f 1 1

is unitary when ϕ̃2,1 = f 1 = 0. When δ, ε are sufficiently small, then these functions
are still small enough such that D1 F is invertible with uniformly bounded inverse,
using Lemma A.1. Furthermore, F ∈ Cb,u k , as the dependence on ξ , η , ξ is clearly
1 2 2
k
Cb,u , while the omega Lemma [AMR88, p 101] guarantees joint Cb,u k -dependence

on f 1 as well. Note that compactness of the domain of f 1 is not required, as we

assume these functions to be uniformly bounded and thus have compact image.
2.4 Smoothing of Submanifolds 65

The implicit function theorem has a corresponding formulation as a uniform

contraction principle. The latter formulation shows that the implicit function G must
be unique, while existence holds if ϕ̃2,1 , f 1 , ξ2 are sufficiently close to zero, due to
a priori estimates. We apply Corollary A.4 as an extension of the implicit function
theorem to conclude that G ∈ Cb,u k . This means that f = p ◦ G( · , f ) ∈ C k on
2 2 1 b,u
suitable neighborhoods. Using formula (A.2) for DG, we can moreover conclude that
f 2 depends Lipschitz on f 1 . This follows from explicit control on the boundedness
and continuity estimates, while variation with respect to f 1 only introduces additional
k +1-order derivatives of ϕ̃2,1 , which can be assumed uniformly bounded. Hence,
there exists some constant C̃ such that g2 − f 2 k ≤ C̃ g1 − f 1 k and all estimates
are uniform. 
Corollary 2.37 (Graph size under coordinate transformations) Under the assump-
tions of Lemma 2.35, there exist constants A, B such that we have the estimate

 f 2 k ≤ A  f 1 k + B (2.34)

on amplification of the size of a graph under coordinate transformations.

Proof We choose g1 = 0 in Lemma 2.35 and set A = C̃. There exists a uniform
bound B such that g2 k ≤ B, and when ζ, ε are sufficiently small, then for δ  = 10
9
δ

we have B(0; δ ) ⊂ Dom(g2 ). Hence, we easily deduce

 f 2 k ≤  f 2 − g2 k + g2 k ≤ C̃  f 1 k + B

for all x ∈ B(0; δ  ) where f 1 , f 2 are defined. 

Theorem 2.38 (Uniform smooth approximation of a submanifold) Let M ∈
k≥1
Cb,u be a uniformly immersed submanifold of a smooth Riemannian manifold (Q, g)
of bounded geometry.
Then for each σ > 0 and integer l ≥ k, there exists a uniformly immersed
l ∩ C ∞ and δ > 0 such that M − M ≤ σ with respect
submanifold Mσ ∈ Cb,u σ k
to normal coordinate charts of radius δ along both M and Mσ . If M is a uniformly
embedded submanifold, then so is Mσ .
The proof relies on finding a (uniformly locally finite) cover of M and then in
each chart make smooth the graph representation h • . All estimates are uniform,
independent of the point x ∈ M, hence so is the final result. Smoothing is done
sequentially in each chart, so we must be careful to check how smoothing in one
chart influences the graph representation in other charts. This makes the technical
estimates quite involved, but the basic idea is that we have uniform control on the
size of changes in each h • by the convolution kernel parameter ν in Lemma 2.34, as
well as the size of this change in other charts.
Proof This proof contains a lot of interdependent size estimation parameters. Giving
explicit choices and dependencies would clutter the proof needlessly, so we make a
few remarks on beforehand. Any δ’s denote sizes of normal coordinate balls and ε’s
66 2 Manifolds of Bounded Geometry

are used for sizes of (changes in) functions in these coordinates. The parameter ν
from Lemma 2.34 depends on most of the foregoing, while only the C l bound (but
not the C k bound) of Mσ depends on the choice of ν. The various ε’s will be fixed
later, and depend on σ and global properties of M and (Q, g), but not on δ’s. Also
note that everything is independent of points x ∈ M, Q.
We fix 2 δ1 = δ2 = 21 δ3 , ε∞ = 2 ε0 , and C∞ = C0 + 1 and choose
δ3 , ζ, ε∞ , εϕ > 0 sufficiently small such that all the following statements hold
true.
1. By Proposition 2.5 and Lemma 2.25, distances and metrics are C = 2 equivalent
on balls B(x; δ3 ), and d Q , d M are locally equivalent, all up to order l + 1.
2. By assumption of M ∈ Cb,u k , we have for each x ∈ M the representation M
x,δ3 =
graph(h x ) with h • 1 ≤ ε0 and h • k ≤ C0 . 
3. By Corollary 2.26, there exists a uniformly locally finite cover i≥1 Mxi ,δ2 of M
with all xi separated by at least δ1 , the balls Mxi ,δ1 already covering M, and bound
K on the maximum number of Mxi ,δ2 intersecting any Mx,δ2 . Formula (2.10)
shows that K depends on the ratio δ2 /δ1 , but does not increase when δ3 → 0.
4. By Lemma 2.29, all coordinate transition maps ϕ2,1 between any x1 , x2 ∈ Q,
d(x1 , x2 ) < δ3 are C l bounded. When the graph representations Hi = Graph(L i )
are bounded by ζ > 0, then these are of the form ϕ2,1 = O H ⊕ OV + ϕ̃2,1 , with
ϕ̃2,1 l+1 ≤ εϕ . And by Corollary 2.31, this holds for the coordinate transfor-
mations between the xi chosen for the cover of M.
5. By Lemma 2.35, there exists a constant C̃ that estimates the graph change under
coordinate transformations from the previous point, when  f 1 , g1 ≤ ε∞ and
 f k , gk ≤ C∞ , while Corollary 2.37 holds on balls of size δ  = 10 9
δ3 > δ2 .
6. If M was a uniformly embedded submanifold, then let Mx,δ3 be the unique
connected component of M in B(x; δ3 ) for each x ∈ M.
Let M j denote a modification of M after applying smoothing operations in the
j
first j charts, and let h i denote the graph representation of M j in chart i. So, initially
j
we have M 0 = M and h i0 = h i . Note that the sequence {h i } j≥1 is constant after some
finite index j (i), since it is changed at most K times by smoothing in overlapping
j (i)
charts. Thus, the final graphs are given by h i∞ = h i .
Initially, we have h • 1 ≤ ε0 and h • k ≤ C0 ,and we assume that  throughout

 j  j
the sequential smoothings it holds for all i, j that h i  ≤ ε∞ and h i  ≤ C∞ ,
1 k
and therefore the final h i∞ satisfy these estimates as well. Let ε0 be sufficiently small,
such that by a mean value theorem estimate

h • (ξ ) = h • (ξ ) − h • (0) ≤ Dh •  ξ  ≤ ε0 ξ ,

we have B(0; δ2 ) ⊂ Dom(h • ).

We apply the convolution smoothing Lemma 2.34 with some choice r > δ1 and
r + 2 δr < δ2 to sequentially make smooth h ii−1 in the coordinate chart B(xi ; δ3 )
l ∩ C ∞ on B(0; r ). The h j representations of the M j
to obtain h ii ∈ Cb,u • x• ,δ3 overlap,
2.4 Smoothing of Submanifolds 67

j
so we must be careful that (at most) K repeated smoothing operations keep the h •
within the bounds required to apply this lemma, while at the same time we must
l at some
ensure that each point on the sequence of manifolds M j is smoothed to Cb,u
stage, even though M j changes to Mσ = M ∞ throughout the sequential smoothings.
Let us first show that each point is smoothed. We can keep track of each original
point x ∈ M as a sequence of points x j ∈ M j throughout the smoothings, and once
M j is smoothed around x j , then the convolution lemma guarantees that smoothness
is preserved around the sequence x j under further smoothing in other charts. Let
∼
 : M → Mσ denote the diffeomorphism that assigns to x ∈ M the final point
∞
x ∈ Mσ . Each point x ∈ M is element of a graph h i in at least one ball B(xi ; δ1 ),
so if the corresponding sequence of points x j ∈ M j moves less than r − δ1 , then it
is smoothed in B(xi ; r ). Therefore, we choose ν in Lemma 2.34 small enough, such
that
r − δ1
 f˜ − f k ≤ ε(ν) < .
2K

The factor 2 K accounts for at most K charts in which x j is moved and C = 2 to

correct for equivalence of distance in charts. Hence, the manifold is smoothed to
l ∩ C ∞ at each point.
Cb,u
j
Next, we show that each h i is defined at least on B(0; r + 2 δr ) and satisfies the
bounds ε∞ and C∞ . Initially, we have h i 1 ≤ ε0 , h i k ≤ C0 , and Graph(h i ) =
Mxi ,δ3 is well-defined in B(xi ; δ3 ). So for ε0 ≤ 10
1
, say, we must have h i 0 ≤ 10
1
δ3
j
and Dom(h i ) ⊃ B(0; 10 δ2 ). The only reason that the domain of some h i decreases
11

is if either the graph moves outside of B(0; δ3 ) or the modified manifold cannot
j
be represented by a graph anymore. The latter cannot occur if Dh i stays bounded,
j
while the former can be controlled by bounding h i 0 . Both can be controlled by
j j−1
estimating the C 1 changes h i − h i . First, in coordinate chart i = j we can directly
j j−1
use the convolution smoothing Lemma 2.34 to conclude that h i − h i 1 ≤ ε(ν).
In any other chart i  = j, this change is amplified by a bounded factor C̃, as per
Lemma 2.35, so we have
j j−1
h i − h i 1 ≤ C̃ ε(ν).

When we choose ν small enough that

 1 
K C̃ ε(ν) < min ε0 , δ3 , 1
10
holds, then this leads to

j 2
h i 0 ≤ h i 0 + K C̃ ε(ν) ≤ δ3 ,
10
j
h i 1 ≤ ε0 + ε0 = ε∞ ,
j
h i k ≤ h i k + K C̃ ε(ν) ≤ C0 + 1 = C∞ .
68 2 Manifolds of Bounded Geometry

j
This shows that indeed the assumed bounds ε∞ and C∞ hold, and that h i is defined
at least on the ball B(0; r + 2 δr ).
The sequential smoothings create and preserve Cb,ul ∩ C ∞ smoothness, while
every ‘point’ x j ∈ M j is touched by these operations. Moreover, Lemma 2.34 and
Corollary 2.37 together guarantee that the smoothing in each chart keeps
j j−1 j−1
h i l ≤ A h j 0 + B ≤ A C(ν, l) h j 0 + B ≤ A C(ν, l) ε∞ + B

bounded with a uniform estimate, at least on charts B(xi ; δ2 ).

Finally, we want to estimate the sizes and distance between the graphs h • , h ∞ • in
split coordinate charts of radius δ = δ2 −r along either M or Mσ . If x is a point either
on M or Mσ , then it is contained in at least one ball B(xi ; r ) and B(x; δ) ⊂ B(xi ; δ2 ).
If we also set ε∞ ≤ ζ and consider the coordinate transformation ϕ from normal
coordinates at xi to x, then Lemma 2.35 and Corollary 2.37 hold and can be used to
estimate
 ∞    ∞   
h  ≤ A h ∞  + B and h − h x  ≤ C̃ h ∞ − h i  ≤ C̃ 2 K ε(ν)
x l i l x k i k

for all points in the domains of h x and h ∞ x − h x within B(0; δ). So if we set
C̃ K ε(ν) < σ , then Mσ is C close to M in normal coordinate charts of radius δ
2 k

along either M or Mσ , while at the same time Mσ ∈ Cb,u l ∩ C ∞.

If M is a uniformly embedded submanifold, then δ3 was chosen small enough

such that Mx,δ3 is the unique connected component of M in B(x; δ3 ) for any x ∈ M.
We now show that the same holds for Mσ with balls of radius δ. Let x̃ ∈ Mσ be
arbitrary and x = −1 (x̃) ∈ M. We take ỹ ∈ Mσ ∩ B(x̃; δ) and want to prove that
ỹ ∈ (Mσ )x̃,δ . By the uniform estimates made before, both Mx,δ2 and (Mσ )x̃,δ2 can be
represented by graphs hx , h ∞
x respectively
  δ2 ) ⊂ Tx M ⊕ N x . We
in coordinates B(0;
have y = −1 ( ỹ) ∈ B x̃; δ+(r −δ1 ) ⊂ B x; δ+2(r −δ1 ) , so y ∈ B(x; δ2 )∩ M =
Mx,δ2 = Graph(h x ).
By the construction of Mσ we have x ∈ B(xi ; r ) in some chart i, but also
(Mσ )x,δ2 = Graph(h x ). Let y ∈ B(x; δ)∩ Mσ ; we want to prove that y ∈ (Mσ )x,δ =
Graph(h ∞ x ). Since y ∈ B(x; δ), then for its original it must hold that
   
y 0 ∈ B x; δ + (r − δ1 ) ⊂ B xi ; r + δ + (r − δ1 ) ,

hence y 0 ∈ Mxi ,r +δ+(r −δ1 ) = Graph(h x ). Following the change of M to Mσ in

coordinates around x, we see that y ∈ Graph(h ∞ x ) must hold. 

2.5 Embedding into a Trivial Bundle

Let π : N → M be the normal bundle over M immersed in (Q, g), a Riemannian

manifold of bounded geometry. We are going to construct a trivial bundle N over
M that contains N and preserves uniform properties. As a second step, we extend a
2.5 Embedding into a Trivial Bundle 69

normally hyperbolic vector field to this trivial bundle setting. This procedure is also
alluded to in [Sak94, p. 333–334], but especially in the case of bounded geometry
requires a more careful inspection.
Theorem 2.39 (Uniform embedding of a normal bundle in a trivial bundle) Let
k≥1
M ∈ Cb,u be a uniformly immersed submanifold of the bounded geometry manifold
(Q, g). Then there exists an embedding λ : N → N of the (nontrivial) normal bundle
π : N → M into a larger, trivial vector bundle N = M × Rn̄ . The embedding map
k−1
λ ∈ Cb,u is a quasi-isometry when restricted to N≤η for any η > 0 and the splitting
N = λ(N ) ⊕ N ⊥ is Cb,u k , where N ⊥ is chosen perpendicular to λ(N ) according to

the standard Euclidean metric on Rn̄ .

Note that λ only has smoothness C k−1 since N ∈ C k−1 is the normal bun-
dle of M ∈ C k . The image bundle λ(N ) ⊂ N has smoothness C k , though, since
its construction only involves the immersion M → Q. This increase of smooth-
ness is possible because we do not view λ(N ) as a normal bundle with respect to
the differentiable structure of Q anymore. This can be compared to the remark in
[Fen72, p. 205] and the reference to [Whi36, Lem. 23] therein.
The idea of the proof is to use normal coordinate charts of Q covering M to
 maps of N . In such charts B(x; δ), we have, from Defi-
construct local trivialization
nition 2.21, Mx,δ = expx Graph(h x ) ⊂ M and trivialization maps τx for the vector
bundle trivialization diagram

τx
N ⊃ N | Mx,δ / Mx,δ × N x
xxx
xx
π
xxxp1 (2.35)
xx
{x
 x
M ⊃ Mx,δ

Then we take a uniformly locally finite cover of M by sets Mxi ,δ . The trivializations
on each Mxi ,δ induce a spanning set of sections, i.e. a frame. Using the uniformity
of the cover, we can globally glue these frames together to obtain rank(N ) = n̄ =
(K+1) rank(N ). Here, K is the maximum number of overlapping charts in the cover.
This identifies λ(N ) as the subbundle of N = M ×Rn̄ spanned by these glued frames.
Proof Let δ be Q-small as in Definition 2.8, as well as sufficiently small such that M
is given as the graph of h • : T• M → N• in normal coordinates as in Definition 2.21.
For any x ∈ M we have a trivialization map
∼
τx = (expx , p2 ) ◦ D exp−1
x : N | Mx,δ → M x,δ × N x , (2.36)

where we canonically identified T(Tx Q) ∼ = (Tx Q)2 and apply expx only on the base
Tx Q and p2 on the fibers of T(Tx Q). In a normal coordinate representation (see
page 52, Fig. 2.2 on the right) this just means that we project the normal fiber N x  at
any point x  = expx (ξ, h x (ξ )) ∈ Mx,δ onto N x . By Corollary 2.31 this projection
70 2 Manifolds of Bounded Geometry

is approximately orthogonal and bounded away from non-invertibility for small δ,

k−1
hence τ• ∈ Cb,u and it is a quasi-isometry, but only on a finitely sized neighborhood
N≤η | Mx,δ since it acts
 linearly on the fibers of N | Mx,δ . We then choose a uniformly
locally finite cover i≥1 Mxi ,δ of M such that the sets Mxi ,δ/2 already cover M.
k−1
Next, we prove the existence of a finite set of Cb,u sections that everywhere
span N . Let G = (V, E) be the (possibly infinite) graph whose vertices are sets
in thecover, V = {Mxi ,δ }i≥1 , and  edges are added between overlapping sets, i.e.
E = (A, B) ∈ V 2 | A ∩ B  = ∅ . Each set in the cover overlaps at most K other
sets, so the maximal degree of G is bounded by K . Therefore, we can ‘color’ the
vertices of G with numbers {0, . . . , K } such that no two connected vertices have the
same number. Sequentially for each i ≥ 1, set the number of vertex i to one of the
numbers {0, . . . , K } that is not already taken by its neighbors. We thus obtain a map
c : N → {0, . . . , K } such that each preimage c−1 (k) labels a collection of mutually
disjoint sets of the cover.
Let n denote the rank of N and let N = M × Rn̄ be a trivial bundle with rank n̄ =
n (K + 1). On each Mxi ,δ we have an orthogonal frame ei of n sections that span N ,
induced by the local trivialization, while on N we have the global orthogonal frame ē
of standard unit sections. The latter can also be viewed as a (K+1)-tuple of n-frames
{ēk }0≤k≤K on Rn̄ = (Rn ) K +1 . Since all spaces have (the standard Euclidean) inner
products, the dual frames can be canonically identified as the inverse ei∗ = ei−1 :
N xi → R N and a projection ēk∗ = pk : Rn̄ → Rn onto the k-th n-tuple of all
n̄ coordinates respectively. Let the functions χi be a square-sum partition of unity
subordinate to the cover according to Corollary 2.18 and define the embedding

λ : N≤η → N̄≤η : (m, ν) → χi (m) ēc(i) ei∗ τxi (m, ν). (2.37)
i≥1

k−1
This mapping is Cb,u as a composition of such maps (and can be extended, albeit
non-boundedly so, to a map N → N ). The τxi are quasi-isometries, while ēc(i) ei∗ is
isometric on each Mxi ,δ . Each frame ēc(i) is orthogonal to the frame of
any overlap-
ping set Mx j ,δ since c( j)  = c(i) and the χi squared sum to one. Thus i χi ēc(i) ei∗
is an isometry, and so λ is a quasi-isometry.
Let p and p ⊥ = 1− p denote the projections from N onto N and N ⊥ , respectively.
One can verify that 
∗
p = χi χ j ēc(i) ēc( j) (2.38)
i, j≥1


is the projection onto N by noting that λ(N ) equals the image of i≥1 χi ēc(i) , while
the identities 
ēk∗ ēl = δkl and χi2 = 1
i≥1
2.5 Embedding into a Trivial Bundle 71

can be used to show that p 2 = p. Formula (2.38) shows that both p, p ⊥ ∈ Cb,u
k ,

hence the splitting N = λ(N ) ⊕ N ⊥ is Cb,u

k . 

Next, we must extend a vector field v on N to the larger bundle N . There are
additional directions along the fibers of N ⊥ and the extended vector field v̄ must be
such that it is normally hyperbolic in these directions as well. On the other hand,
the uniform boundedness of v must be preserved. We do not assume here that M
is the exact invariant manifold, since these results shall be applied after application
of Theorem 2.38, which has smoothed and slightly altered M such that it is not the
original NHIM anymore. The extension v̄ will keep N invariant and is identical to v
on N , so in the end, we can conclude that the perturbed manifold is contained in N
and restrict to the original setting again.
Lemma 2.40 (Normally hyperbolic extension of a vector field) Let λ : N →
N = M × Rn̄ be a trivializing embedding of vector bundles as in Theorem 2.39 and
l,α
let v ∈ Cb,u be a vector field on N with l + α ≤ k − 2. Let N |≤η be the restriction
to some radius η > 0. Then v can be extended to a vector field v̄ on N , such that v̄ is
l,α
Cb,u on N ≤η , leaves N invariant, and contracts at a given exponential rate ρ < 0
along the fiber direction of N ⊥ towards N ⊂ N .
To extend v to a vector field v̄ on N with the required properties, we must do two
things. First of all, v must be extended from N through λ to the whole of λ(N ) ⊕ N ⊥
and secondly, a normal component along the fibers of N ⊥ must be added to make v̄
contracting, thus normally hyperbolic in that direction. The idea can be expressed in
local coordinates (m, y, z) ∈ λ(N ) ⊕ N ⊥ as
 
v̄(m, y, z) = v̂(m, y, z) + v⊥ (m, y, z) = v(m, y), ρ z ,

where v̂(m, y, z) = v(m, y) points ‘horizontally’ along λ(N ) and v⊥ (m, y, z) = ρ z

is the ‘vertical’ component along the N ⊥ fibers. By construction, the latter has the
required contraction property in the N ⊥ direction, while it preserves λ(N ) ⊕ {0}
as an invariant manifold. We shall make this intuitive idea rigorous by introducing
an appropriate bounded connection to lift v to v̂ for z  = 0; the second term v⊥ is
canonically defined.
k−1
Proof The embedding map λ is a quasi-isometry and of class Cb,u , hence the
pushforward λ∗ (v) = Dλ · v ◦ λ−1 is a Cb,u
l,α
vector field on λ(N ) ⊂ N . From now
on we identify N with λ(N ) as well as v with its pushforward.
Let g be the standard Euclidean metric on N and ∇ the compatible, trivial,
flat connection. The restricted metric g ⊥ = g| N ⊥ is preserved by the connection
∇ ⊥ = p ⊥ · ∇. We create the pullback bundle
 
E = π N∗ (N ⊥ ) = (y, z) ∈ N × N ⊥ |π N (y) = π N ⊥ (z) ∼= N ⊕ N⊥ = N.
72 2 Manifolds of Bounded Geometry

Note that we identify this pullback of N ⊥ along the projection π N : N → M with

the vector bundle p : N → N , that is, we view N as the base manifold and N
as bundle over N with fibers π N−1 (y) = Nπ⊥N (y) . We naturally endow E with the
pullback connection ∇ˆ = π N∗ (∇ ⊥ ). With this connection, v can be lifted to a unique
vector field v̂ ∈ X(N ) that is horizontal along N on E = ∼ N , and thus the flow of v̂
preserves the norm along the fibers of E, that is, v̂ ·  ·  E = 0. More heuristically,
we can say that the pullback π N : N → M introduces a trivial additional base
coordinate y ∈ Nm to the bundle π N ⊥ : N ⊥ → M.
l,α
To prove that v̂ ∈ Cb,u , we first recover an explicit representation of the Christoffel
symbols of ∇ˆ in terms of trivial coordinates on E ∼ = N = M × Rn̄ , and then a
representation for the lifted vector field v̂. Let s ∈ (N ), s ≡ s0 ∈ Rn̄ be a constant
section. Let m ∈ B(m 0 ; δ) denote normal coordinates in M. Define s ⊥ = p ⊥ · s ∈
(N ⊥ ) and ŝ = π N∗ (s ⊥ ) ∈ (E). Let X̂ ∈ TN a tangent vector in the base of E and
X = Dπ N ( X̂ ) ∈ TM. Then we find for the covariant derivative ∇ˆ on E
  
ˆ ∗
 ⊥ ⊥
 ∗ ⊥ ⊥ ∂s ∂ p⊥ 
∇ X̂ ŝ = π N p · ∇ X ( p · s) = π N p · X p i
+ s .
∂m i ∂m i

∂ p⊥
We read off that the Christoffel symbols are given by p ⊥ ∂m i
(Dπ N )i , so they are
k−1
Cb,u . The horizontal lift

∂ p⊥
v̂(m, y, z) = v(m, y) − p ⊥ · ( z)(Dπ N v(m, y))i , (2.39)
∂m i
l,α l,α
then, is also Cb,u since all functions involved are at least Cb,u in these coordinates,
and z is bounded on N ≤η .
We define v⊥ as the Euler vector field along the fibers of N ∼ = E, taking values in
Vert(E). Each fiber E (m,y) = Nm⊥ is a linear space, so the tangent space at any point
is canonically identified with the fiber itself, which allows us to canonically define

v⊥ : N → TN : (m, y, z) → ρ z ∈ Tz Nm⊥ = Vert(E)(m,y,z) . (2.40)

This vector field leaves N ⊂ N invariant, while generating a flow that attracts towards
N at the exponential rate ρ < 0. It is clear that v⊥ ∈ Cb,uk for any k when z is bounded

on N ≤η .
We conclude that the vector field v̄ = v̂ + v⊥ indeed leaves N invariant. Since v̂
is neutral in the fiber directions of E and v⊥ contracting at rate ρ, it follows that v̄ is
contracting with rate ρ as well. The combined vector field v̄ is defined in terms of v
l,α l,α
and other functions that are all at least Cb,u , hence v̄ ∈ Cb,u . 
2.6 Reduction of a NHIM to a Trivial Bundle 73

2.6 Reduction of a NHIM to a Trivial Bundle

Having set up the theory of bounded geometry spaces, we are finally in a position
to reduce a general normally hyperbolic system to the setting of a trivial bundle.
That setting is required to apply our basic Theorem 3.2 on persistence of NHIMs.
k,α
Let M ∈ Cb,u with k ≥ 2 and 0 ≤ α ≤ 1 be a uniformly immersed or embedded
submanifold in (Q, g) of bounded geometry and furthermore assume that M is an
k,α
r -NHIM with r = k + α for the vector field v ∈ Cb,u on Q.
Remark 2.41 The bounded smoothness requirement k ≥ 2 is dictated by
Theorem 2.33. It is not present in the compact case where the normal bundle can
be “jiggled slightly” [Fen72, Prop. 2] to make it sufficiently smooth to model a C k
flow for k ≥ 1. Hypotheses 2 and 3 in [BLZ99, p. 987] require similar conditions in
the noncompact setting in Banach spaces, see also the discussion after Corollary 3.5
and Remark 3.13. I have not investigated in detail whether the Cb2 requirement is
necessary, or if M ∈ Cb,u1 could be modeled as a sufficiently small section of the

normal bundle of a smooth approximate manifold. ♦

We reduce this system to a trivial bundle in the following steps:
1. approximate M by a smoothed manifold Mσ ;
2. construct a tubular neighborhood of M in the normal bundle N of Mσ ;
3. embed N into a trivial bundle N = Mσ ×Rn̄ , and construct an extended, normally
hyperbolic vector field v̄;
4. after application of the basic persistence theorem in the enlarged bundle, push the
results to the original setting and conclude that M persists.
Proof of Theorem 3.1. Assume the hypotheses of the theorem. First, Theorem 2.38
gives a smooth, C k -close approximation Mσ ∈ Cb,u l of M, where the choice l =
k + 10 suffices and σ > 0 will be fixed later. The C k,α bounds of Mσ are uniformly
close to those of M for all σ small. Then, Theorem 2.33 says that there exists a tubular
∼
neighborhood ϕ : N≤η → B(Mσ ; η) where the size η > 0 depends only on the C 2
bounds of Mσ . These bounds are of the same order as those of M, independent of σ .
Hence, we can choose σ so small that Mσ − Mk ≤ σ ≤ η2 and the neighborhood
B(M; η/2) is fully within the tubular neighborhood N≤η of Mσ . The map ϕ is a C l−1
bounded (local) diffeomorphism and a quasi-isometry, so by the reasoning4 before
Definition 2.32, a pullback by ϕ does not change the normal hyperbolicity growth
rates of the vector field v. The bounded continuous splitting (1.9) of T M Q is also
preserved.
Next, as a result of Theorem 2.39, N is embedded into the trivial bundle

λ : N → N = Mσ × Rn̄ = N ⊕ N ⊥ .

4 This is similar to Fenichel’s argument in [Fen72, p 203] that normal hyperbolicity is independent
of a choice of metric when M is compact.
74 2 Manifolds of Bounded Geometry

The embedding is a quasi-isometry and so preserves hyperbolicity properties of v.

k,α
The extended vector field v̄ ∈ Cb,u on N ≤η is constructed in Lemma 2.40 as a lift of
v, so the flow preserves N≤η and intertwines with the projection onto N , while in the
perpendicular direction along the fibers of N ⊥ it has the same normal hyperbolicity
properties as v. Boundedness of the invariant splitting T M Q = TM ⊕ N + ⊕ N − is
also preserved under these quasi-isometries. The additional directions along N ⊥ are
stable and invariant, and have bounded projections by construction. Thus, M is an
r -NHIM for v̄ as well.
The invariant manifold M is given by the graph of a section h ∈ (N ) ⊂ (N ),
k,α
and from Theorem 2.38 it follows that hk ≤ σ while h ∈ Cb,u . By Lemma 2.27,
X = Mσ has bounded geometry of order l − 2 = k + 8, which is sufficiently smooth
for the conditions of Theorem 3.2, while Y = Rn̄ is clearly a Banach space. Hence,
we are in the trivial bundle setting, and all conditions are satisfied.
A small perturbation of v in the original setting in Q corresponds to a small pertur-
bation of v̄ ∈ X(N ), while N≤η is preserved under the flow by construction. There-
fore, after application of Theorem 3.2 we recover a unique persistent invariant mani-
fold M̃ and by construction M̃ = Graph(h̃) ⊂ N , so we can restrict the system to N .
Then it can be transferred back to Q under the quasi-isometries of the embedding
λ : N → N and the tubular neighborhood map ϕ : N≤η → B(M; η) ⊂ Q.
All size estimates can be transferred between the settings (with bounded factors)
due to the near isometry and uniform C k,α boundedness of ϕ and λ. We conclude
k,α
that M̃ is a Cb,u submanifold of Q for appropriate estimates σ, ε, and δ, where σ
must be chosen sufficiently small as well to make  M̃ − Mk−1 small.
This completes the reduction from the general setting in bounded geometry to
that of a trivial bundle and proves Theorem 3.1. 
Chapter 3
Persistence of Noncompact NHIMs

This chapter contains the main proof of persistence of noncompact normally

hyperbolic invariant manifolds, formulated in Theorem 3.2. This theorem is for-
mulated in a specific setting: we assume that the invariant manifold M is (nearly) the
zero section of a trivial vector bundle. This is a slightly more general formulation
than in [Hen81; Sak90]. There, it is assumed that in a product X × Y of Euclidean
(or Banach) spaces, the invariant manifold M is given as the graph of a function
h : X → Y . We shall also assume that Y is a vector space, but we let X instead
be a Riemannian manifold that is finite-dimensional and has bounded geometry. In
Chap. 2 on bounded geometry, we extended the result obtained here to a setting where
M is assumed to be a general submanifold of a Riemannian manifold (Q, g) that is
again finite-dimensional and of bounded geometry. We assume the basic statements
from Sect. 2.1 to be known.
This chapter is organized as follows. First we state the two main theorems; both
the general version with M a submanifold of Q and the X × Y trivial bundle ver-
sion to be proved in this chapter. We provide detailed remarks on these theorems
and compare them to the literature. Then we present an outline of the proof of
Theorem 3.2. Section 3.3 presents some thoughts on replacing the classical com-
pactness by uniformity conditions, and presents examples that indicate the necessity
of various assumptions we impose.
In Sect. 3.4 we transform the (still somewhat geometrical) formulation of
Theorem 3.2 into a more explicit setup suitable for analysis. In the subsequent section,
we prove (with relatively little work) the existence and uniqueness of the persistent
manifold M̃ = Graph(h̃). It automatically follows that h̃ is bounded and uniformly
Lipschitz.
The last sections are devoted to the tougher job of proving C k,α smoothness,
exhausting the spectral gap. A formal scheme is set up, and we work out the details
for C 1 smoothness. Higher, C k smoothness follows along the same lines by induction.
The addition of Hölder continuity to obtain C k,α smoothness is included as a natural
extension to (uniform) continuity that slightly simplifies the spectral gap estimates.
See the proof outline and the introduction of Sect. 3.7 for more details.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 75

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4_3,
© Atlantis Press and the author 2013
76 3 Persistence of Noncompact NHIMs

3.1 Statement of the Main Theorems

The main theorem on persistence was already formulated in the introduction. We

state it again to directly compare it to the trivialized bundle version of Theorem 3.2.
The main theorem is reduced to this trivialized version in Sect. 2.6; in this chapter we
shall prove the latter version. Then we formulate corollaries of these theorems, both
to present simpler versions and to compare our result to well-known results from the
literature.
Theorem 3.1 (Persistence of noncompact NHIMs in bounded geometry) Let
k ≥ 2, α ∈ [0, 1] and k + α. Let (Q,g) be a smooth Riemannian manifold of bounded
k,α k,α
geometry and v ∈ Cb,u a vector field om Q. Let M ∈ Cb,u be a connected, complete
submanifold of Q that is r-normally hyperbolic for the flow defined by v, with empty
unstable bundle, i.e. rank (E+ ) = 0.
Then for each sufficiently small η > 0 there exists a δ > 0 such that for any
k,α
vector field ṽ ∈ Cb,u with ||ṽ − v||1 < δ there is a unique submanifold M̃ in the η
-neighborhood of M, such that M̃ is diffeomorphic to M and invariant under the flow
k,α
defined by ṽ Moreover M̃ is Cb,u and the distance between M̃ and M can be made
arbitrarily small in C k−1 -norm by choosing ||ṽ − v||k−1 sufficiently small.
Theorem 3.2 (Persistence of noncompact NHIMs in a trivial bundle) Let k ≥ 2,
α ∈ [0, 1] and r = k + α. Let (X, g) be a smooth, complete, connected Riemannian
k,α
manifold of bounded geometry and Y a Banach space. Let vσ ∈ Cb,u be a family
of vector fields defined on a uniformly sized neighborhood of the zero-section in
X × Y with family parameter σ ∈ (0, σ0 ]. Let the submanifold Mσ = Graph(h σ ) be
k,α
given as the graph of a function h σ ∈ Cb,u (X ; Y ) and let Mσ be an r -NHIM with
+
rank(E ) = 0 for the flow defined by vσ where all estimates are uniform in σ and
additionally h σ 2 ≤ σ holds.
Then for each sufficiently small η > 0 there exist σ1 , δ > 0 such that for any
k,α
σ ∈ (0, σ1 ] and any vector field ṽ ∈ Cb,u with ṽ − vσ 1 < δ, there is a unique
 
 
submanifold M̃ = Graph(h̃), h̃ : X → Y , h̃  ≤ η such that M̃ is invariant under
0 
k,α  
the flow defined by ṽ. Moreover, h̃ ∈ Cb,u and h̃  can be made arbitrary small
k−1
by choosing hk and ṽ − vσ k−1 sufficiently small.
Remark 3.3 Let us make some remarks on these theorems.
1. The spectral gap condition contained in Definition 1.11 of r -normal hyperbolicity
is essential to the proof. The C k,α -smoothness result is optimal, see Sect. 1.2.1.
2. In Theorem 3.1, both M and M̃ are assumed to be (non-injectively) immersed
according to Definition 2.21. If M is assumed uniformly embedded according to
Remark 2.22, then M̃ will be uniformly embedded again when δ is sufficiently
small.
3. The additional family parameter σ in Theorem 3.2 is required to reduce Theorem 3.1
to this case. If the unperturbed manifold is given as M = X × {0}, i.e. the zero
3.1 Statement of the Main Theorems 77

section, then the family vσ can simply be taken constant and all σ dependence
can be dropped from the formulation.
4. We only obtain a C k−1 -norm estimate for the perturbation distance of M̃ away
from M, even though M̃ ∈ C k,α is preserved. This is due to a linearization along Y
and the smoothing convolution used to restore C k,α smoothness after linearization.
See Sect. 3.4, in particular Remarks 3.13 and 3.15, for more details. I fully expect
it to hold that M̃ and M are C k,α close when ṽ − vk,α is small.
5. The minimum smoothness requirement k ≥ 2 is a stronger assumption than k ≥ 1
in the well-known compact case. This seems to be intrinsic to the noncompact
case. If the spectral gap condition only holds for some 1 ≤ r < 2, then we can
still obtain a perturbed manifold M̃. This manifold M̃ will generally not have
better than C r smoothness, though.
6. We allow both values α = 0 and α = 1, where α = 0 is considered an empty
condition (besides the boundedness and uniform continuity). Thus, if r = k + 1
satisfies the spectral gap condition (1.11), then we can choose both C k,1 or C k+1,0
as resulting smoothness for M̃, if M had the same smoothness. Thus, if M was
sufficiently smooth, then the choice M̃ ∈ C k+1,0 yields the best result. Note,
though, that by Rademacher’s theorem, Lipschitz functions are differentiable
almost everywhere, so the difference is not that big.
Finally, it should also be noted that the spectral gap condition is a strict inequality
on r , so if we can choose r = k integer, then we can also find an α > 0 such that
r  = k + α satisfies the spectral gap as well. This shows that in this context Cb,uk
k,α
‘integer’ smoothness really is a special case of Cb,u ‘fractional’ smoothness.
7. Both Riemannian manifolds Q in Theorem 3.1 and X in Theorem 3.2 are assumed
to be finite-dimensional; multiple results on bounded geometry crucially depend
on this fact. On the other hand, we allow Y to be an infinite-dimensional Banach
space simply because everything naturally generalizes to that setting. Note that
we do not allow semi-flows as in [Hen81; BLZ08], so the case that Y is infinite-
dimensional may not be that useful.
8. These results are weaker than those in the well-known compact case in a
few aspects. First of all, we use a stricter notion of normal hyperbolicity, see
Remark 1.10. This seems to be a fundamental restriction of the Perron method;
the more general definition of normal hyperbolicity is successfully applied to
noncompact manifolds in [BLZ08]. Secondly, we only include the stable nor-
mal bundle E − . Adding the unstable bundle E + as well should be possible, see
Sect. 4.4 for more details.
9. While we do prove that the NHIM persists into a new invariant manifold M̃,
we do not prove that M̃ is again normally hyperbolic. I fully expect this to be
true though: the perturbed flow satisfies slightly perturbed exponential growth
conditions and the spectral gap is an open condition, so should be preserved
under sufficiently small perturbations. It remains to be proven that M̃ again has a
continuous invariant splitting (1.19) with bounded projections. This is one possible
reason for breakdown of normal hyperbolicity [HL06]. In Sect. 4.5 we sketch how
to recover the invariant stable fibration and the invariant splitting.
 78 3 Persistence of Noncompact NHIMs

k,α
10. In both theorems, we assumed that M ∈ Cb,u . Just as in the compact case, there
is forced smoothness, see [HPS77]. That is, if we only have M ∈ Cb2 , while it is
k,α k,α
r -normally hyperbolic and the system is Cb,u , then we have in fact M ∈ Cb,u .
Note that we do require k ≥ 2 unlike the compact case, see also point 3.3.
This statement can be verified by reviewing the persistence proof under zero per-
turbation. We construct a smoothed approximation of M in Theorem 3.1 to model
everything on, and in Theorem 3.2 we smoothen the linearization of the vertical
part vY of the original vector field. These can be made as smooth as required, while
k,α
we only use the perturbed flow (3.16) or the Cb,u smooth decomposition (3.24)
(hence with remainder f˜ ∈ C ) in the rest of the proof. Therefore the resulting
k,α
b,u
k,α
manifold will satisfy M̃ = M ∈ Cb,u . ♦
These two theorems reduce to the corollaries formulated below, when M is com-
pact or when Q = Rm+n with standard Euclidean metric and M = Rm × {0}. The
statements then significantly reduce in complexity, and are comparable to well-known
results.
Firstly, the case that M is compact. Then we can take a (pre)compact neigh-
borhood B(M; ε) ⊂ Q of M and thus conclude that bounded geometry holds on
B(M; ε), ignoring irrelevant boundary problems. Any C k,α function on B(M; ε) is
k,α
automatically Cb,u , so Theorem 3.1 reduces to the following corollary. For simplicity
we leave out α-Hölder continuity and the C k distance estimate between M and M̃.
Corollary 3.4 (Persistence of compact NHIMs) Let k ∈ Z≥2 , let (Q, g) be a
smooth Riemannian manifold and v ∈ C k be a vector field on Q. Let M ∈ C k be
a connected, compact submanifold of Q that is k-normally hyperbolic for the flow
defined by v, with empty unstable bundle, i.e. rank(E + ) = 0.
Then for each sufficiently small η > 0 there exists a δ > 0 such that for any
vector field ṽ ∈ C k with ṽ − v1 < δ, there is a unique submanifold M̃ in the
η-neighborhood of M, such that M̃ is diffeomorphic to M and invariant under the
flow defined by ṽ. Moreover, M̃ is C k .
This corollary closely resembles [Fen72, Thm. 1] in the absence of a boundary ∂ M
(a boundary is allowed when M is overflowing invariant, see also Sects. 1.6.3 and 4.3).
Note that our definition of normal hyperbolicity is less general (see Remark 1.10),
and that we exclude unstable normal directions and the case1 k = 1, while we do
allow M to be an immersed submanifold. The persistence result of Hirsch, Pugh,
and Shub [HPS77, Thm. 4.1 (f )] is similar to that in Fenichel’s work; it additionally
includes Hölder smoothness and allows immersed submanifolds as well.
Secondly, the case of M = Rm × {0} ⊂ Q = Rm+n . Again, Rm+n has bounded
k if it can (locally) be
geometry with one trivial, global chart. Thus, any object is Cb,u
k functions, but with common global bound and continuity modulus.
described by Cb,u
Then Theorem 3.2 reduces to the following corollary. We again suppress Hölder

1 This was for technical reasons in the noncompact setting, see Remark 3.3, 5, and could be repaired
in the compact setting.
3.1 Statement of the Main Theorems 79

continuity and drop the σ parameter dependence, which was only relevant for the
reduction of Theorem 3.1.
Corollary 3.5 (Persistence of a trivial NHIM in Euclidean space) Let k ∈ Z≥2 .
k be a vector field on Rm+n and let M = Rm × {0} be a k-NHIM for the
Let v ∈ Cb,u
flow defined by v, with empty unstable normal bundle.
Then for each sufficiently small η > 0 there exists a δ > 0 such that for any vector
k with ṽ − v < δ, there is a unique submanifold M̃ = Graph(h̃),
field ṽ ∈ Cb,u
  1
 
h̃ : Rm → Rn , h̃  ≤ η such that M̃ is invariant under the flow defined by ṽ.
0  
k and h̃ 
Moreover, h̃ ∈ Cb,u can be made arbitrary small by choosing ṽ − vk
 
k−1
sufficiently small.
This theorem can be compared, for example, to [Sak90, Thm. 2.1]. Sakamoto’s
theorem is specifically targeted to singular perturbation problems. His conditions
are more specific and concrete: the invariant manifold M is assumed to consist of
stationary points and normal hyperbolicity is formulated in terms of the eigenvalues
of normal derivatives of the vector field at M. He starts with an invariant manifold
that is the graph of a nonzero function h ∈ Cbk ; this he reduces to the zero graph
case M = Rm × {0}, while he incurs a loss of one degree of smoothness, obtaining
a Cbk−1 persistent manifold and he requires k ≥ 3, see [Sak90, p. 50]. He does allow
both stable and unstable normal bundles.
In their series of papers [BLZ98; BLZ99; BLZ08], Bates, Lu, and Zeng obtained
multiple results on noncompact NHIMs, including a persistence result. Most impor-
tantly, they work in Banach spaces with semi-flows, which adds nontrivial obstacles
to be overcome, and allows application to PDE problems. On the other hand, my
setting allows the ambient space to be a manifold, albeit finite-dimensional. They
use the graph transform instead of the Perron method. This allows for the more gen-
eral definition of relative normal hyperbolicity as in Remark 1.10. They include both
stable and unstable normal directions, while they do not prove Hölder regularity.
Finally, in [BLZ08] the interesting idea is developed to start with an approximate
NHIM only.
If we ignore these differences, then their results fit in between the formulations of
Theorem 3.1 and Corollary 3.5. Their invariant manifold M is immersed in a Banach
space, but not necessarily described by the graph of a function h. Their hypothe-
sis [BLZ08, p. 363] that the splitting does not twist too much is a bounded Lipschitz
condition on the (approximate) splitting of the (un)stable and tangent bundles over M.
This condition is similar, but slightly weaker than our condition M ∈ Cb,u 2 , see also

Remarks 3.3, 5 and 3.13.

Although the results of Bates, Lu, and Zeng are more general and complete in
many aspects, I think that these cannot easily be generalized to prove a version of
Theorem 3.1, set in an ambient manifold of bounded geometry. One could hope
to use the Nash embedding theorem to obtain the ambient manifold (Q, g) as an
isometrically embedded subspace of some Rn . Then the dynamical system must be
extended from Q to Rn , such that M is still normally hyperbolic as a submanifold
80 3 Persistence of Noncompact NHIMs

of Rn ; this procedure can be compared to the reduction in Sect. 2.6. The problem
that arises is that the Nash embedding theorem provides no control on the extrinsic
curvature of the embedding2 , so M need not be Cb2 in Rn . It might be possible to
work around this by proving a ‘bounded geometry version’ of the Nash embedding
theorem. Alternatively, their proof could likely be adapted to work with a uniform
atlas of charts, e.g. using theory similar to that developed in Chap. 2.
We should also mention the paper [JS99] by Jones and Shkoller. They generalize
Fenichel’s results on persistence of overflowing invariant manifolds to semi-flows on
infinite-dimensional Riemannian manifolds. They do assume the invariant manifold
itself to be compact.

3.2 Outline of the Proof

The proof of Theorem 3.2 is lengthy and involves a lot of details. We therefore first
present an overview of the separate steps involved in the proof.
First, in Sect. 3.4 we bring the system into a form that is suitable for application
of further analytical techniques. That is, we decompose the vector fields along X and
Y directions and linearize the vertical direction, leading to equations

ẋ = v X (x, y),
ẏ = vY (x, y) = A(x) y + f (x, y)

with a C 1 -small term f . This allows us to apply a generalization of the classical

Perron method for hyperbolic fixed points (see Sect. 1.4.2 for a quick overview) to
NHIMs, first presented by Henry [Hen81, Chap. 9]. In the case of a hyperbolic fixed
point, we could fully linearize the system; here, we can only linearize the normal
directions, while we keep the full nonlinear form in the directions along X . These
cannot be linearized because we have no control to localize the dynamics in the
directions along X .
In the theorem, the invariant manifold M is given as a (small) graph h : X → Y .
A coordinate change to represent the invariant manifold as M = X × {0} would
(re)introduce a loss of smoothness that we carefully worked around in Chap. 2 by
means of a uniformly smoothed submanifold. The graph h can be chosen arbitrarily
close to the zero section, and together with the small perturbation ṽ−v, this influences
the exponential growth rates (1.10) only slightly. We recover Eqs. (3.16) for the
perturbed system that satisfy slightly perturbed exponential estimates (3.17), even
when we decouple the equations for x and y by inserting curves y(t) and x(t),

2 This can be seen from the result that the Nash embedding can be obtained into an arbitrarily

small ball. As an explicit example, take Q = R with standard metric and embed it into R2 via the
map r (θ) = arctan(θ)/π + 21 in polar coordinates. Since the integral of r (θ) diverges both when
θ → ±∞, we obtain (after arc length reparametrization) an isometric embedding of R ‘curled up’
into B(0; 1), while the extrinsic curvature grows unbounded for θ → −∞.
3.2 Outline of the Proof 81

respectively, that are ‘close’ to solution curves of the original system. We directly
include the perturbation ṽ−v into the horizontal component of the vector field; for the
vertical component we include the perturbation in the nonlinear term f˜. This gives
rise to a nonlinear, horizontal flow
y (t, t0 , x0 ) and a linear, vertical flow x (t, t0 ) y0
that depend on a curve in the other space and satisfy estimates
 
∀t ≤ t0 : D
y (t, t0 , x0 ) ≤ C X eρ X (t−t0 ) ,
∀t ≥ t0 : x (t, t0 ) ≤ CY eρY (t−t0 ) ,

where ρ X , ρY are close to the original exponential rates −ρ M , ρ− .

The next step in Sect. 3.6 is to define a pair of maps (3.32) between curves in X and
Y in terms of these flows and the decomposed vector fields (3.16). The composition
T = TY ◦ (TX , pr1 ) of these maps will be a contraction on bounded curves in Y
depending on a parameter x0 ∈ X , but we measure these curves with  · ρ norms
for some exponent ρ with ρY < ρ < ρ X . Lemma 3.27 shows that the fixed points
of T are precisely the vertical parts y(t) of solution curves of the perturbed vector
field ṽ that stay in the tubular neighborhood y ≤ η of M and have initial value
x0 for their horizontal part x(t). The maps TX and TY generalize the center-unstable
and stable components, respectively, of the Perron integral in the fixed point case, cf.
Sect. 1.4.2. The nonlinear flow along the invariant manifold is used in TX , but now
depends on the vertical component y(t) too, while in the vertical, normal directions
we use a variation of constants integral to separate the nonlinear terms from the
linearized flow, just as in the classical Perron method.
ρ
This setup leads to a fixed point map : X → Bη (I ; Y ) that maps an initial
value x0 ∈ X to the unique bounded curve in Y that corresponds to a full solution
curve (x, y) such that x(0) = x0 . If we now evaluate the vertical solution curve
y = (x0 ) at t = 0, then we obtain the vertical component y0 = y(0) ∈ Y of the
initial value corresponding to x0 ∈ X . All these solution curves stay close to M and
form an invariant manifold, so the graph of

h̃ : X → Y : x0 → (x0 )(0)

must describe the unique perturbed invariant manifold M̃. Application of the contrac-
tion principle immediately implies that h̃, and therefore M̃, is Lipschitz continuous.
In Sect. 3.7 we continue to prove that M̃ is C k,α . We start that section with a more
detailed overview of this smoothness part of the proof and in Sect. 3.7.1 we present
a scheme to obtain the first derivative in a number of steps. Higher smoothness then
follows along the same lines, just with more complex expressions, see Sect. 3.7.9.
Let us focus here on the basic ideas.
Smoothness of M follows directly from smoothness of . We study the derivatives
of by formal differentiation of the fixed point equation; this leads to Eq. (3.42).
But let us consider for a moment a simpler heuristic formulation, similar to Eq. (1.8)
for the hyperbolic fixed point in Sect. 1.4.3. Then the derivatives of the Perron fixed
point map are
82 3 Persistence of Noncompact NHIMs

  t  
Dk T (y) δy1 , . . . , δyk (t) = (t, τ ) · Dk f˜(y(τ )) δy1 (τ ), . . . , δyk (τ ) dτ.
−∞

Even if Dk f˜ is bounded, it acts as a multilinear map on a k-tuple of variations δyi ,

each having exponential growth of order ρ, so the result has exponential growth of
order k ρ. This is canceled by the exponential growth of (t, τ ) if ρY < kρ. Then
Dk T can be viewed as a contraction on Dk , but only when Dk is viewed as a
map into B k ρ (I ; Y ). To obtain continuity of the maps Dk T , we have to add another
arbitrarily small term μ < 0 to the exponent, i.e. k ρ + μ; in case of α-Hölder
continuity we need μ = α ρ. These key facts show how the spectral gap condition
limits smoothness; see Sect. 1.2.1 for a detailed discussion and an example that shows
that our smoothness result is in fact sharp.
The technique of using a scale of Banach spaces as developed by Vanderbauwhede
and Van Gils [VG87], and the fiber contraction theorem of Hirsch and Pugh [HP70]
can be applied, and we obtain each Dk as a fixed point in the appropriate space.
The final conclusion h̃ ∈ C k follows by evaluating Dk at t = 0.
We have to be very careful however: higher derivatives of maps between mani-
folds are difficult to define (at least in a practical way), so we develop some theory
to describe higher derivatives using normal coordinates in Appendix C and gener-
alize results from Rn to this setting. Secondly, the derivatives of TX , TY only exist
as ‘formal derivatives’ on ‘formal tangent bundles’. We endow these formal tan-
gent bundles with a topology induced by parallel transport. This allows us to study
continuity of the formal derivatives DTX , DTY at the cost of introducing additional
 
holonomy terms. Finally, we do obtain the Dk maps as derivatives of .

3.3 Compactness and Uniformity

The classical results on normally hyperbolic invariant manifolds [Fen72; HPS77]

assume the invariant manifold to be compact. This is used to obtain uniform bound-
edness and continuity of the vector field and other objects. Here, instead, we assume
these objects to have the required uniformity directly, replacing the compactness
requirement. In this section, we expose some of the issues that need to be dealt
with and we present accompanying examples. We focus here on those issues that
are not (clearly) present in the literature and only show up when considering general
manifolds. See Sect. 1.2.2 for motivation and examples of noncompact NHIMs.
The primary requirement in the noncompact case—well-known to experts in the
field—is that the vector field defining the system must be uniformly bounded, includ-
ing all its spatial derivatives up to the order of the smoothness result requested.
Secondly, the vector field should be uniformly continuous, and uniformly α-Hölder
continuous when α = 0. The case α = 0 really is a special case, whose proof
needs more care. Hölder continuity provides an explicit continuity estimate, which
3.3 Compactness and Uniformity 83

is tailored to the problem; ‘plain’ uniform continuity does not provide this, forcing
us to use an (arbitrarily) small amount of the spectral gap to compensate.
In the compact case, Fenichel [Fen72, p. 200] argues that persistence of the invari-
ant manifold should be independent of the choice of a Riemannian metric. Indeed,
he proves that the exponential growth rates are independent of such a choice, as all
metrics are equivalent. In a noncompact setting, however, non-equivalent metrics do
exist and we do expect persistence to depend on the choice of metric, since it deter-
mines which perturbations are globally C 1 small. Moreover, we make a technical
uniformity assumption of bounded geometry (see Chap. 2, also for example spaces of
bounded geometry) on both the underlying space and the invariant manifold. These
assumptions are automatically satisfied in the compact case. It is not clear though to
what extend they are essential in the noncompact case.
The remainder of this section is devoted to examples that show multiple aspects
that should be treated carefully in the noncompact setting, while being trivially
fulfilled in the compact case. Some interesting examples can also be found in the early
work [Hop66] by Hoppensteadt. He presents counterexamples to uniform stability of
solutions in a time-dependent singular perturbation setting when the stability criteria
do not have sufficient uniformity.

3.3.1 Non-Equivalent Metrics

As a simple example of two metrics leading to different results in the noncompact

case, let us consider the following.
Example 3.6 (Non-equivalent metrics) Let X = R2 , Y = R with on the one hand
the usual Euclidean metric ge and on the other hand a metric gs induced by a dif-
feomorphism similar to stereographic projection from the sphere (with North Pole
removed) onto X = R2 .
Let the vector field be given by
 x 
v(x, y) = arctan(x) , λy (3.1)
x

with λ < 0. This makes the vertical, y direction uniformly attracting with exponent λ,
while in the plane X × {0} the origin is an expanding fixed point and the exponential
growth rate is everywhere non-negative. Thus, M = X × {0} is an r -NHIM for
arbitrarily large r and v ∈ Cb,u
k for any k ≥ 1 on a tubular neighborhood of M of

size η = 1, say. See Fig. 3.1 on the left side.

In polar coordinates (s, θ ) on X we have
 
v(s, θ, y) = arctan(s), 0, λ y .
84 3 Persistence of Noncompact NHIMs

Fig. 3.1 A normally hyperbolic flow with respect to Euclidean and sphere induced metrics

Now, instead of the usual stereographic projection map (ϕ, θ ) → (s = tan(ϕ/2), θ )

with ϕ = 0 at the South Pole, we take

f : [0, π ) → R≥0 : ϕ → − log(1 − ϕ/π ) (3.2)

and the corresponding diffeomorphism

that acts trivially along the directions of θ
and y coordinates. This diffeomorphism induces a metric gs on X by pushforward of
the standard metric on the sphere. The system with metric gs is most easily studied
by pullback to the sphere with North Pole removed, see Fig. 3.1 on the right side.
This is an equivalent formulation since
is an isometry by construction. The vector
field is then represented on S 2 × R by
   
(
∗ v)(ϕ, θ, y) = π(1 − ϕ/π ) arctan − log(1 − ϕ/π ) , 0, λ y . (3.3)

This shows that the vector field is still Cb1 , although not Cb,u
1 anymore in this met-

ric. More importantly, the system can be extended to include the North Pole as an
attracting fixed point. The rate of attraction along the perpendicular y direction has
not changed from λ, but along the horizontal directions of the sphere, the attraction
rate now is π
lim D1 (
∗ v)(ϕ, 0, 0) = − . (3.4)
ϕ→π 2

In both metrics the normal exponential attraction rate is ρY = λ < 0 since the linear
flow on Y is decoupled from X . The exponential growth rate on X does depend on
the choice of metric. For the Euclidean metric ge , we have ρ X = 0. This follows
from an analysis of the radial component ṡ = arctan(s) of the system. At s = 0
this has unstable exponent 1, while away from s = 0 the tangent flow is uniformly
bounded away from both zero and infinity, so there the Lyapunov exponent is zero.
With respect to the metric gs , it follows from 3.4 that the Lyapunov exponent is
ρ X = − π2 for solutions approaching planar infinity. Thus, we see that if λ ≥ − π2 ,
then the system is not normally hyperbolic with respect to the metric gs , while
3.3 Compactness and Uniformity 85

X × {0} is an r -NHIM for any r ≥ 1 under the metric ge . On the other hand, if
λ < − π2 , then the system is still only r -normally hyperbolic with respect to gs
for r < π/(2λ). Again, we can construct an explicit perturbation similar to that in
Example 1.1. If we add a small vertical perturbation ε χ with support away from
the poles (compare with Fig. 1.3), then along meridians passing through supp χ , the
invariant manifold is lifted and approaches the North Pole ϕ = π approximately
along a graph y = C (π − ϕ)−2λ/π , while on meridians not passing through supp χ ,
the invariant manifold stays at y = 0. See the perturbed flow lines in Fig. 3.1. This
results in unbounded C r derivatives of the perturbed invariant manifold at the North
Pole for r > −2λ/π . 
We conclude that in the noncompact setting, normal hyperbolicity explicitly
depends on the choice of metric since metrics need not be equivalent. Moreover,
the allowed size of the perturbations depends on the metric.
Example 3.7 (Perturbation sizes depend on the choice of metric) We extend
Example 3.6 above. Let λ < − π2 so that the system is normally hyperbolic with
respect to both metrics, and set
 
w(s, θ, y) = 0, 0, arctan(s) sin(θ ) . (3.5)

Then the vector field v + ε w is a C 1 small perturbation of 3.1 with respect to ge

and perturbs the original manifold M smoothly to a manifold M̃ that has height
converging to y = −ε sin(2 θ ) π/(2λ) along radials when s → ∞. The pullback
of M̃ to the sphere, however, has a discontinuity at the North Pole, since s → ∞
corresponds to ϕ → π , and so the North Pole is approached at different constant
heights along these radials. This apparent contradiction that
∗ ( M̃) is not a C 1
small perturbation with respect to gs stems from the fact that w is not C 1 small
with respect to this metric. The vector field w has unbounded derivatives since gs
‘squeezes’ distances when approaching planar infinity, that is, the North Pole. 
Thus, non-equivalent metrics also lead to different classes of C 1 small perturba-
tions under which the invariant manifold persists. Moreover we see that one cannot
simply get rid of noncompactness by a compactification argument. The metric gs is
induced by a one-point compactification to the sphere, but leads to different normal
hyperbolicity properties than the noncompact case with metric ge . Any other choice
of the diffeomorphism
would lead to the same problems, since pullback of the
metric ge must introduce a singularity at the North Pole. This cannot be equivalent
to a metric that extends regularly there.

3.3.2 Non-Persistence of Embedded NHIMs

Let us give another example which shows that an embedded invariant manifold need
not persist. This example clarifies the remarks already made in the introduction in
86 3 Persistence of Noncompact NHIMs

Sect. 1.6.2: noncompact embedded NHIMs can perturb into immersed manifolds. On
the one hand, this example shows that it is natural to consider immersed NHIMs. It
also shows that a noncompact NHIM must have a uniformly sized tubular neighbor-
hood that does not self-intersect, in order to guarantee perturbation as an embedded
manifold. Further details can be found in Sect. 2.3 where the concept of a uniformly
embedded submanifold is defined.
In the example presented here, the unperturbed manifold is normally hyperbolic
but noncompact and ‘touches’ itself in the limit to infinity, see Fig. 3.2, the top image.
In this case, we can find arbitrarily small perturbations that will let the two persisting
branches collapse into one at a finite point.
Example 3.8 (A non-uniformly embedded NHIM) Let (x, y) ∈ R2 = Q. For x ≤ 0
we define the vector field v of the system in polar coordinates, and for x ≥ 21 in
Cartesian coordinates as

(ṙ , θ̇ ) = v(r, θ ) = (1 − r , − sin(θ/2)) if x = r cos(θ ) ≤ 0,

(3.6)
(ẋ, ẏ) = v(x, y) = (e−x , −y)if x ≥ 21 .

We glue these vector fields together in a smooth way somewhere between x = 0

and x = 21 . Then the manifold as shown at the top in Fig. 3.2 is a NHIM. The flow
attracts uniformly in the normal direction with rate −1 (except that the rate may
deviate slightly around the glued area), while along the manifold, the flow has an
expanding fixed point at (−1, 0) and the contraction in the direction of x → ∞ is
weaker than exponential. Explicitly solving the flow for x ≥ 21 yields

t (x, y) = (log(e x + t) , y e−t ),

Fig. 3.2 Collapse of a nearly y

self-intersecting invariant
manifold

x
3.3 Compactness and Uniformity 87

which exhibits the rates of contraction in the normal and tangential directions by
considering either projection in

1  
lim log πx,y ◦ D
t (x, y) .
t→∞ t

Let us now introduce the very simple perturbation vector field w(x, y) = (−ε , 0)
for x ≥ 1 and smoothly cut off to zero left of x = 1. When this perturbation is added
to the vector field v, the vertical line x = − log(ε) becomes a stable, invariant set,
see Fig. 3.2 the bottom image. The upper and lower branch of the original NHIM
will both converge to the newly created fixed point (− log(ε) , 0). On the right side
of this point the manifold is given by the single line y = 0.
Each branch separately persists as a C ∞ manifold, as could (naively) be expected.
The problem is that we have no control on the distance between the two branches,
so for any ε > 0, these branches will collapse at some point where the persisting
object ceases to be an embedded manifold. As already remarked, there are two ways
to address this issue. One can abandon the implicit assumption that the NHIM is an
embedded submanifold and replace this by immersed submanifolds; this idea was
introduced already in [HPS77]. If one insists on having embedded submanifolds,
even under perturbations, then one must eliminate the possibility of these ‘collapses’
occurring. A sufficient condition is the existence of a uniformly sized tubular neigh-
borhood of the invariant manifold that does not intersect itself. Global control on the
perturbation distance of the invariant manifold will imply that the perturbed manifold
stays inside this tubular neighborhood and thus will not self-intersect. 

3.3.3 Non-Uniform Geometry of the Ambient Space

The previous examples were set in Euclidean space. The next two examples show that
additional uniformity conditions must be imposed on a nontrivial ambient space. It is
not enough to assume uniform continuity and boundedness for the dynamical system.
The first example is an extension to the previous one and shows that the ambient space
must have a uniformly finite injectivity radius. The second example indicates that
even if the ambient space has finite injectivity radius and trivial topology, persistence
might be lost due to non-bounded curvature of the ambient space.
Example 3.9 (Zero injectivity radius) We construct as ambient space Q a cylinder
whose radius shrinks exponentially. That is, we take Q = R × S 1 with metric
g(x, θ ) = dx 2 + e−2x dθ 2 . See Fig. 3.3 for an impression, but note that the metric
induced by the embedding in R3 is not (and cannot be made) the same as g. The
vector field
v(x, θ ) = (1, 0) (3.7)
88 3 Persistence of Noncompact NHIMs

Fig. 3.3 A non-uniform

cylinder with a winding curve

generates a simple flow along the cylinder, and each solution curve is a NHIM purely
due to the fact that all curves flow into an exponentially shrinking tube, while there
is no contraction along the curve.
Let us consider the invariant manifold M = { θ = 0 }. We add a perturbation to
the vector field that is given by θ̇ = ε for x ≥ 0 and is smoothly cut off to zero left
of x = 0. This perturbation is smooth and C 1 small with respect to the metric3 g.
When the original curve M enters the region x ≥ 0, it is modified to a curve M̃ that
starts winding around the cylinder, as indicated in Fig. 3.3.
This clearly cannot be represented in a tubular neighborhood of M in Q since
the curve would leave the neighborhood ‘above’ and reenter ‘from below’. On the
other hand, the normal bundle of M can be viewed as a covering of Q, and on that
covering, M̃ is represented by the function θ (x) = ε x, which is still a bounded
graph with norm θ (x) = ε x e−x , but which winds around since θ ∈ [0, 2 π ).
Thus, a globally finite injectivity radius seems a necessary requirement if we want
the perturbed manifold to be represented in a diffeomorphic tubular neighborhood
of M. 
The second example indicates that a finite injectivity radius is not enough;
unbounded curvature of the ambient manifold might lead to loss of persistence of the
NHIM. It should be pointed out that this example satisfies all properties of normal
hyperbolicity with uniform estimates up to C 1 smoothness, except that the vector
field has no uniformly continuous derivative. I have not been able to add this final
property to create a complete counterexample where persistence fails in the absence
of the curvature property of bounded geometry only.
Example 3.10 (Unbounded curvature) Let Q = R3 with metric
   
g(x, y, z) = dx 2 + exp − 2 |x| arctan(x z) dy 2 + exp − 2 |x| dz 2 , (3.8)

but with component functions symmetrically smoothed around x = 0. This Rie-

mannian manifold is invariant under translations in y and has a mirror symmetry
involution in any plane of fixed y. Hence, each submanifold { y = y0 } is geodesically
invariant.

3 Measuring the C 1 size with respect to g requires taking covariant derivatives and may introduce

results not directly apparent in coordinates (x, θ). A perturbation term θ̇ = ε exp(x) would still be
globally small in this metric.
3.3 Compactness and Uniformity 89

Let the vector field be

 
x, 0, − arctan(z) for |x| ≤ 1,
v(x, y, z) =   (3.9)
sign(x), 0, 0 for |x| ≥ 1,

and smoothly glued together in a neighborhood of the boundary |x| = 1, see Fig. 3.4.
Thus, the whole system is invariant under translations in y, and within any plane
{ y = y0 }, the point (x, z) = (0, 0) is a hyperbolic fixed point with eigenvalues 1
and −1 in the x and z direction, respectively. The system also has a mirror symmetry
around x = 0; from now on we only consider x ≥ 0.
The plane M = { z = 0 } is a NHIM; it is clearly invariant under the flow, and
similar to the exponentially shrinking cylinder, the metric contracts in the z direction
along solution curves x → ∞, while no contraction occurs along the manifold. On
TM the metric reduces to g|TM = dx 2 + dy 2 while the flow is linear in time. On a
neighborhood of the y-axis, finally, normal hyperbolicity follows from the attraction
along the z directions due to the term − arctan(z) in (3.9).
The vector field v and its covariant derivative are uniformly bounded with respect
to the metric. For x ≤ 2, this follows from the fact that g including its inverse
and derivatives, as well as v and its derivatives are bounded. For x ≥ 2, explicit
calculations in local coordinates show that v = 1, while we have
⎛ ⎞
0 0 0
∇ v = ⎝ 0 − 1+(x z)2 − arctan(x z) 0 ⎠ ,
x z

0 0 −1

expressed in an orthonormal frame, which is bounded as well. The second covari-

ant derivative ∇ 2 v is unbounded, though. This indicates that ∇v is probably not
uniformly continuous for a reasonable definition of uniform continuity, cf. Defini-
tion 2.9, although I have not completely investigated this question.
The Ricci scalar curvature of Q is unbounded and on M = { z = 0 } it is given by
 
S = −2 1 + x 4 e2x .

Clearly, this implies that the Riemannian curvature is unbounded too.

Remark 3.11 In hindsight, it should probably not come as a complete surprise that
∇ 2 v is unbounded. The Riemannian curvature is unbounded, and since it is the gen-
erator of holonomy (see Sect. 2.2), it can thus generally be expected that holonomies
along infinitesimal loops act as an unbounded family of operators on v. These are
expressed in local coordinates by second covariant derivatives of v:

R(∂i , ∂ j ) v = ∇∂i ∇∂ j v − ∇∂ j ∇∂i v. ♦

90 3 Persistence of Noncompact NHIMs

We proceed with checking that exp : N → Q has finite injectivity radius on the
normal bundle4 N of M. Then all assumptions for persistence are fulfilled, except for
bounded curvature (and uniform continuity of ∇v). As each submanifold { y = y0 }
is invariant, we can restrict our investigation to y = 0, such that x denotes the
coordinate along the base manifold of the normal bundle; let t denote the normalized
coordinate in the vertical direction. The exponential map of (x, t) is generated by
the geodesic flow as follows: start at x with vertical unit vector and then follow a
geodesic for time t. For x = 2, say, this flow is well-defined and stays inside the
region x ≥ 1 for some bounded time |t| ≤ r . The diffeomorphism group

ϕ(x, z) = (x + ξ, z eξ ) with ξ ∈ R

translates along x while simultaneously scaling z, see also Remark 2.2. In the region
x ≥ 1 this is an isometry, so the exponential mapping defined for x = 2, |t| ≤ r can
be isometrically mapped onto the whole region x ≥ 2, |t| ≤ r . For x on the compact
interval [0, 2] the exponential map must have a finite injectivity radius too, so there
exists a global r > 0 such that exp : N≤r → Q is diffeomorphic onto its image.
Now we add a perturbation in a similar spirit to that in Sect. 1.2.1: we lift M by
a local, vertical perturbation of the vector field, varying along y. In a neighborhood
of the plane x = 2 we add a small vertical component
 −1
ż = ε (2 − cos(y)) exp if |x − 2| ≤ 1.
1 − (x − 2)2

In the region x ≤ 1 the flow is unmodified, so there the perturbed manifold M̃ must
coincide with the original M; otherwise it would not stay in a bounded neighborhood
of M under the backward flow. Around x = 2, the flow lifts M̃ to at least a height
 
3 −1
z≥ε exp dx ≥ ε/4
1 1 − (x − 2)2

and the height z depends on y, see Fig. 3.5. Then in the region x ≥ 3 the manifold
M̃ continues along ẋ = 1 at the same y, z coordinates. Now we have ε/4 ≤ z ≤ 6 ε,
so for all small ε > 0 the y component of the metric along the flow on the invariant
manifold eventually shrinks at an exponential rate that is stronger than in the z
direction, while M̃ has variable height z(y) independent of x. Hence the Lipschitz
norm of M̃ can be estimated by measuring z  (y) with respect to the metric g along the
manifold. But z  (y) is nonzero and constant along x in coordinates, while horizontal
distances along y shrink faster than vertical distances. This means that the Lipschitz
norm of M̃ grows unbounded for x → ∞. Moreover, the normal exponential growth

4 We should actually show that the injectivity radius of Q is finite, i.e. rinj (Q) > 0, at least in a
neighborhood of M. I have not been able to do this. Finite injectivity radius of the normal bundle
does allow us to construct a tubular neighborhood to model persistent manifolds close to M, though.
3.3 Compactness and Uniformity 91

x=1 x=3

Fig. 3.4 Perturbation of a normally hyperbolic system in unbounded geometry

z
y

z
y

Fig. 3.5 The graph of the perturbed manifold with respect to the Euclidean metric (top) and an
approximate image of the same graph with respect to the metric g (bottom)

rate does not dominate the tangential rate anymore, so the perturbed manifold is not
normally hyperbolic anymore. 

3.4 Preparation of the System

As a first step towards proving Theorem 3.2 we shall bring the system in a form
suitable to apply analytical tools to it. Let
92 3 Persistence of Noncompact NHIMs

ẋ = v X (x, y) ∈ Tx X,
ẏ = vY (x, y) ∈ Y, (3.10)

be the decomposition of the vector field v along X and Y . The invariant manifold is
given as the graph M = {y = h(x)}. We dropped the explicit dependence on σ from
the notation. The full flow of v will be denoted by ϒ t , while
,  are reserved for
flows defined in terms of the horizontal and vertical components of v, respectively. To
shorten notation we write g(x) = (x, h(x)). We shall always assume that y ≤ η.
Our goal is to establish a linearized form

vY (x, y) = A(x) y + f (x, y) (3.11)

k,α
for the vertical part of (3.10) such that f is small and A, f ∈ Cb,u , while the flows

and  generated by
t t

ẋ = v X (g(x)),
ẏ = A(x(t)) y with g(x(t)) a solution curve on M, (3.12)

should satisfy exponential growth estimates (1.10) as in Definition 1.8 of normal

hyperbolicity with exponents ρ X , ρY close to the original −ρ M , ρ− , respectively; the
corresponding constants C̃ M , C̃− may differ arbitrarily from the original C M , C− .
We first identify the invariant splitting and associated flows on T M (X × Y ) to
be able to relate these exponential growth rates, see also Fig. 3.6. By definition of
normal hyperbolicity (without an unstable bundle) we have

T M (X × Y ) = TM⊕E − , 1 = πTM + π E − , Dϒ t = Dϒ M
t
⊕ Dϒ−
t

with associated exponential growth rates (1.10). On the other hand we have the
splitting
T(X × Y ) = πY∗ (TX ) ⊕ π X∗ (TY ) ∼
= TX × (Y × Y )

that is naturally induced by the trivial bundle structure. The identification Dg =

1TX + Dh : TX → TM is bounded linear with bounded inverse, so the associated

Fig. 3.6 The splitting Tm Em Y

(X × Y ) = Tm M ⊕ E m−
with m = g(x)
πE ,N

Tm M
m
X
x
h
3.4 Preparation of the System 93

vector field g ∗ (v) = v X ◦ g on X generates a flow

t suchthat D
t
 has the same
exponential growth rate as Dϒ M , up to a bounded factor Dg  · Dg due to
t −1

the norms on the different tangent spaces. Recall that vσ depends on a parameter
σ ∈ (0, σ0 ]. We choose the bound σ1 small enough such that for all σ ≤ σ1 we have
 
∀ t ≤ 0 : D
t  ≤ C̃ M e−ρ M t with C̃ M = 2 C M .

Let N = π X∗ (TY )| M denote the vertical bundle over M, whose fibers can be
canonically identified with Y . Just as above, we want to project the flow Dϒ− t onto

N while preserving the exponential growth rate. The projection π E − along TM is uni-
formly bounded for all σ . This means that the angle between TM and E − is bounded
away from zero. Since TM can be chosen arbitrarily close to the horizontal TX by
choosing σ sufficiently small, it follows that the projection DπY | E − : E − → N and
its inverse π E −,N are bounded for all σ ≤ σ1 when σ1 is sufficiently small, see also
Fig. 3.6. To this end, let (0, ϕ) ∈ Tg(x) (X × Y ) and consider the identity

ϕ = DπY · (0, ϕ) = DπY · (πTM + π E − ) · (0, ϕ).

We have πTM · (0, ϕ) ∈ TM so πTM · (0, ϕ) = (ξ, Dh(x) ξ ) where ξ = Dπ X · πTM ·

(0, ϕ) ∈ Tx X Now we have estimates

ξ  = Dπ X πTM (0, ϕ) ≤ πTM ϕ,

πTM · (0, ϕ) = Dh(x) ξ  ≤ σ πTM ϕ,
DπY · π E − · (0, ϕ) = ϕ − DπY · πTM · (0, ϕ) ≥ (1 − σ πTM )ϕ

−
 thatDπY | E − has an inverse π E −,N : N → E for which
from which it follows
we have the bound π E −,N  ≤ (1 − σ1 πTM )−1 π E −  ≤ 2 π E −  if we choose
σ1 ≤ 2π1TM  .
Consider the flow

ˆ t = DπY ◦ Dϒ t ◦ π E −,N : N → N ,
 (3.13)

generated by DvY ◦ π E −,N along solution curves g(x(t)). Both DπY | E − and π E −,N
are uniformly bounded, so the exponential estimates of Dϒ− ˆ t up to
t carry over to 

a constant factor:
 
 ˆ t
∀ t ≥ 0 :   ≤ C̃− eρ− t with C̃− = 2 π E −  C− . (3.14)

ˆ t on X and N , respectively, that are

We have thus constructed flows
t and 
generated by
v X ◦ g and Â(x) = DvY (g(x)) · π E −,N (g(x))
94 3 Persistence of Noncompact NHIMs

with a solution curve x(t) of the vector field v X ◦ g inserted. These flows are of
the form (3.12) and satisfy exponential estimates (1.10) inherited from the invariant
bundle splitting.
The vector field v X ◦ g already has sufficient smoothness5 , but  is not smooth
k,α
enough since the projection π E −,N is only continuous. We construct A ∈ Cb,u as a
smoothed approximation of D y vY ◦ g. This term is C -close to Â, since
0

     
 
 Â − D y vY ◦ g  = (Dx vY ◦ g) · Dπ X · π E −,N 0 ≤ Dx vY ◦ g0 π E −,N 0
0
(3.15)
because DπY · π E −,N = 1 N and Dx vY ◦ g is small. Lemma 3.17 will imply that
the flow  t of this approximation has exponential growth estimates close to those
of ˆ t . The following lemma will be used to obtain A from D y vY ◦ g. We apply it
 
with l = k − 1 to obtain A ∈ Cbk (X ; L(Y )) such that  A − D y vY ◦ g k−1 ≤ ε(ν).
This lemma is a (strongly) simplified version of Theorem 2.38; the notation of l, k
is reversed to match the context here.
Lemma 3.12 (Uniform smoothing of a vector bundle section) Let (X, g) be a Rie-
mannian manifold of bounded geometry and V a Banach space. Let f ∈ Cb,ul (X ; V )

be a section of the trivial vector bundle π : X × V → X .

Then for any k > l and ε > 0 there exists a smoothed function f˜ ∈ Cb,u
k (X ; V )
 
such that  f˜ − f l ≤ ε. (The bounds on higher than derivatives will generally
depend on ε.)
Proof We apply convolution smoothing of Lemma 2.34 in each chart of a cover of
X and glue these together.
Let 0 < δ1 < δ2 < δ3 and let {B(xi ; δ2 )}i≥1 be a uniformly locally finite cover
of X obtained from Lemma 2.16, such that the δ1 -sized sets already cover X , and
the δ3 -sized sets still
 have normal coordinate charts. Lemma 2.17 yields a uniform
partition of unity i≥1 χi subordinate to this cover.
In each chart B(xi ; δ2 ) we apply Lemma 2.34 to f with r = δ2 and 2 δr ≤ δ3 −δ2 .
k on each chart with uniformly bounded C k -norms and  f˜ − f 
We obtain f˜i ∈ Cb,u i l
can be made as small as required by choosing the parameter ν small. We glue these
together to one function 
f˜ = χi f˜i
i≥1

defined globally on X with the functions χi ∈ Cb,u

k . Together with the uniform bound

on the number of charts in the cover that intersect any one point,  this guarantees that
f˜ satisfies estimates equivalent to those of the f˜i . Note that  f˜l does not depend
on the smoothing parameter ν, but the higher derivative norms do. 

5 It may seem impossible to define a C k vector field v on the tangent bundle TM of a C k manifold
M since TM ∈ C k−1 . See [PT77, App. 1] or [PT83, p. 398] for a method to endow an invariant
submanifold M ∈ C k with a compatible topology that makes v| M ∈ C k . We effectively used this
in our definition of v X ◦ g ∈ C k .
3.4 Preparation of the System 95

Remark 3.13 (On loss of smoothness). We must carefully construct the system 3.11
in order not to lose one degree of smoothness, while at the same time retaining
exponential growth rates and proximity estimates.
The invariant complementary bundle E − is only continuous, while the normal
bundle of M is only C k−1 , even if disguised in coordinate expressions. We use
k−1,α
the linearization at y = 0, but not directly, since D y v y ( · , 0) ∈ Cb,u artificially
decreases the smoothness as well. The loss of smoothness in [Sak90] occurs for
these reasons. Note that even though we retain C k,α smoothness by a convolution
smoothing, this does not preserve higher than C k−1 bounds. This seems to be an
artifact of the proof, inherent to the partial linearization along Y .
In the proof of Theorem 3.1 we construct a smoother, approximate manifold Mσ
exactly to circumvent these problems. In the trivial bundle setting of Theorem 3.2
then, we must be careful not to pick a representation that reintroduces this loss of
smoothness. On the other hand, we do not seem to obtain optimal results in the
sense that we require h2 small, while the classical results in the compact case only
require M = Graph(h) ∈ C 1 . Similarly, h ∈ Cbk with k ≥ 3 is assumed in [Sak90,
p. 50], while hypothesis H2 in [BLZ99, p. 987] is imposed to bound ‘twisting’ of
the invariant manifold. This requirement seems closely related to our condition on h,
and is necessary for the same reason as in our Theorem 3.1: to construct a tubular
neighborhood of uniform size. I do not know whether these stronger assumptions
can be weakened or removed. ♦

3.5 Growth Estimates for the Perturbed System

We shall finally put all the ingredients together to obtain exponential growth estimates
for perturbed flows contained in the tubular neighborhood y ≤ η of X × Y . We
write the perturbed vector field ṽ on X × Y as

ẋ = ṽ X (x, y),
ẏ = ṽY (x, y) = A(x) y + f˜(x, y), (3.16)
 
where f˜(x, y) = f (x, y) + ṽY (x, y) − vY (x, y) .

Let us assume that the conditions of Theorem 3.2 hold true. First, if ρ M = 0,
then for any fixed r we can always slightly increase6 to ρ M > 0, such that the
growth rates (1.10) and spectral gap condition ρ− < −r ρ M still hold true; this way,
we get rid of degenerate exponentials in integrals. We have some ‘spectral space’
ρ = r ρ M − ρ− > 0 that we use to define modified exponential growth numbers

6 We have ρ M ≥ 0 from (1.10). Note that we are interested in −ρ M for the stable side of the
spectrum. In the rest of this chapter, all exponential rates will be negative.
96 3 Persistence of Noncompact NHIMs

ρ
ρ X = −ρ M − , C X = 2 C̃ M ,
4
ρ
ρY = ρ− + , CY = C̃− . (3.17)
4
This allows us to get all perturbed flows within these slightly modified growth rates,
while we reserve another ρ/2 spectral space for later use, such as proving (higher
order) differentiability. Note that both ρY , ρ X are negative since we focus on the
stable normal bundle.
We first fix some notation to be used throughout the proof:
• Cv denotes the global C k,α bound on v and ṽ.
• ε(ν) denotes the perturbation size in Lemma 3.12 depending on the smooth-
ing convolution parameter ν from Lemma 2.34, while Cv (ν) denotes the C k,α
bound onA, f˜, which may grow due to smoothing when ν → 0. We also have
 Ak−1 ,  f˜k−1 ≤ Cv .
• ζ denotes a small bound both on the derivative of f˜ and on perturbations of the
horizontal vector field on X , that is, we impose bounds
 
 
sup D f˜(x, y) ≤ ζ and sup ṽ X (x, y) − v X (x, h(x)) ≤ ζ,
x∈X x∈X
y≤η y≤η

and the size of ζ will be controlled by δ, σ1 , ν, and η.

Let us point out here that multiple parameters must be chosen small, some dependent
on other small parameters. The following graph shows all dependencies; an arrow
indicates that the choice of a parameter influences the choice of the object pointed
to.
δO iS
bDDSSSS z< σO 1 dII kkkuku:5 νO
DD zSzSSS kIkIkk uu
Dz SSk I u
zzzDkDkDkkkk SuSuSuSuISIII
zzkkk D uo u S
η oDkbD ζO (β, T ) (3.18)
DD t t:
DD t
DD tt
D tttt
other (small) constants and bounds

The constants and bounds include (3.17) and Cv , and are all fixed. Note that there are
no circular dependencies, so we are free to choose any of these parameters smaller
if necessary without the risk of having unsatisfiable constraints.
By invariance of M = {y = h(x)} we have

dy
vY (x, h(x)) = = Dh(x) · v X (x, h(x)). (3.19)
dt

This can be used to estimate vY (x, h(x)) ≤ σ1 Cv and derived estimates, such as
(taking the derivative with respect to x)
3.5 Growth Estimates for the Perturbed System 97

Dx vY ◦ g ≤ D y vY Dh + D2 hvY ◦ g

 
+ Dh Dx v X ◦ g + D y v X ◦ gDh ≤ 4 Cv σ1 (3.20)

where σ1 ≤ 1 has been assumed. Together with previous estimates, this leads to

 f (x, 0) = vY (x, 0) − A(x) · 0

≤ Dh(x)v X ◦ g0 ≤ σ1 Cv ,
Dx f (x, y) = Dx vY (x, y) − DA(x) y
≤ Dx vY (x, y) − Dx vY (x, h(x)) + Dx vY ◦ g0

+ D(A − D y vY ◦ g)0 + Dx D y vY 0 y
 
≤ εDx vY (η + σ1 ) + 4 Cv σ1 + ε(ν) + Cv η,
D y f (x, y) = D y vY (x, y) − A(x)
≤ D y vY (x, y) − D y vY (x, h(x)) + D y vY ◦ g − A0
≤ εD y vY (η + σ1 ) + ε(ν),
ṽ X (x, y) − v X (x, h(x)) ≤ ṽ X (x, y) − v X (x, y) + v X (x, y) − v X (x, h(x))
≤ δ + εv X (η + σ1 ).

Hence,  f ( · , 0)0 can be made small independently of η, while  f 1 ,

ṽ X ( · , y) − v X ◦ g0 ≤ ζ can be obtained for any ζ > 0 depending on δ, η, σ1 ,
ε(ν), and the continuity moduli εDvY , εv X . So if we set
 
ζ = 5 Cv σ1 + 2 εDvY (η + σ1 ) + ε(ν) + Cv η + ε(ν) + εv X (η + σ1 ) + δ, (3.21)

then  f 1 , ṽ X ( · , y) − v X ◦ g0 ≤ ζ hold and ζ is small when δ, σ1 , ν, η are.

We need the following result to control the C k−1 distance of the perturbed manifold
M̃ to M.
Proposition 3.14 For any ε > 0, the nonlinearity f˜ and its partial derivatives with
respect to x ∈ X can be bounded as
 
 
∀ 0 ≤ i ≤ k − 1 : Dix f˜ ≤ ε (3.22)

by choosing η, ν, σ1 , hk , and ṽ − vk−1 small enough.

The idea of the proof is the following. If M is described exactly by h(x) ≡ 0,
then by invariance we have vY (x, 0) ≡ 0 (cf. (3.19)), hence Dix vY (x, 0) ≡ 0 as well.
We adapt the proof to incorporate small perturbations introduced by the nonzero
function h and the convolution smoothing of A.
Proof Note that f˜ is defined by (3.16) and (3.11) as
 
f˜(x, y) = vY (x, y) − A(x) · y + ṽY (x, y) − vY (x, y) , (3.23)
98 3 Persistence of Noncompact NHIMs

where A is defined as a convolution smoothing of D y vY ◦ g such that A − D y vY ◦ g

k−1 ≤ ε(ν). The term in brackets obviously becomes small when ṽ − vk−1
does. For the second term note that y ≤ η, while Ak−1 is bounded close to
D y vY ◦ gk−1 , which in turn can be estimated by D y vY k−1 ≤ Cv and hk−1
after application of Proposition C.3.
For the first term in (3.23) we use the continuity modulus of Dix vY to estimate
 
Dix vY (x, y) ≤ Dix vY (x, h(x)) + εDix vY y − h(x) ,

while y − h(x) ≤ η + σ1 . We insert (3.19) and apply Proposition C.3 another time
to obtain
    • 
Dix vY (x, h(x)) = Dix Dh(x) · v X (x, h(x)) − y vY (x, h(x)) · Pm,i−l D h(x) .
Dlx Dm
l≥0,m≥1
l+m≤i

This expression can be made small since Dlx Dmy vY  ≤ Cv and each term contains
at least one factor D j h(x) for some 0 ≤ j ≤ k. 
Remark 3.15 Note that we cannot improve the result to a C k size estimate, since
Ak ≤ C(ν) may grow with ν → 0, while compensating this by choosing η smaller
would introduce a circular dependency in (3.18). ♦
As the next step, we will derive exponential growth estimates for the perturbed
system (3.16). More generally, we consider the horizontal flow
y and vertical, linear
flow x generated by

ẋ = ṽ X (x, y), (3.24a)

ẏ = A(x) y, (3.24b)

with specific curves y : I → Y and x : I → X substituted, respectively. The follow-

ing series of lemmas and propositions show that these flows are small perturbations
of the flows of (3.12) and satisfy exponential growth rates (3.17). We prove the non-
linear case on X and the linear case on Y separately, since we use C 1 smoothness
for the nonlinear case, while only continuity can be assumed for the linear case.
Lemma 3.16 (Growth estimates for a perturbed system) Let X be a Riemannian
manifold and let the system ẋ = v(t, x) with v, Dx v ∈ Cb0 have flow
with expo-
nential growth estimate

∀ x0 ∈ X, t ≤ t0 : D
(t, t0 , x0 ) ≤ C eρ(t−t0 ) . (3.25)

Let ṽ = v + r be a perturbed system generating a flow

. ˜ For each ρ̃ < ρ and
˜ satisfies the
C̃ > C, there exists a δ > 0, such that if r 0 , Dx r 0 < δ, then

growth estimate (3.25) with ρ̃ and C̃ inserted.

3.5 Growth Estimates for the Perturbed System 99

Note that this lemma is formulated in backward time.

Proof Choose T > 0 sufficiently large such that C̃ eρ(−T ) ≤ eρ̃(−T ) . By continuous
dependence of the solutions of differential equations on parameters (see Theorem A.6
and Remark A.7), a C 1 small perturbation r results in a C 1 small perturbed flow
˜
on compact time intervals −T ≤ t − t0 ≤ 0. This result is uniform in t0 , t when
v, r ∈ Cb1 , where differentiation is understood with respect to x only. Hence we
obtain
sup ˜ t0 , x0 ) e−ρ(t−t0 ) ≤ C̃
D
(t,
x0 ∈X
−T ≤t−t0 ≤0

if δ is chosen sufficiently small. Writing t −t0 = −(n T +τ ) with n ∈ N, τ ∈ [0, T ),

we use the group property of the flow to obtain
 n
˜ t0 , x0 ) ≤ C̃ eρ(−T ) C̃ eρ(−τ ) ≤ eρ̃ n(−T ) C̃ eρ(−τ ) ≤ C̃ eρ̃(t−t0 ) .
D
(t, 

Lemma
 3.17 (Perturbation of linear flow) Let Y be a Banach space and let A ∈
Cb0 R; L(Y ) generate a flow (t, t0 ) with growth estimate

∀ t ≥ t0 : (t, t0 ) ≤ C eρ(t−t0 ) . (3.26)

 
Let ρ̃ > ρ be given and set δ = ρ̃−ρ
C > 0. If B ∈ C b R; L(Y ) is globally bounded
0

˜ t0 ) of Ã(t) = A(t) + B(t) satisfies (3.26) with ρ̃ inserted.

by δ, then the flow (t,
˜ is
Proof The variation of constants integral equation for 
 t
˜ t0 ) = (t, t0 ) +
(t, ˜
(t, τ ) B(τ ) (τ, t0 )dτ. (3.27)
t0

˜ with an approach inspired by Gronwall’s lemma.

We shall prove the estimate for 
Note that our variation of constants formula (3.26) is slightly different from the
standard context of Gronwall’s lemma, since we do not have a bound for A.
We denote by ψ(t, t0 ) = C eρ(t−t0 ) the bound on . Now ψ̃(t, t0 ) = C eρ̃(t−t0 )
satisfies the integral equation
 t
ψ̃(t, t0 ) = ψ(t, t0 ) + ψ(t, τ ) δ ψ̃(τ, t0 )dτ (3.28)
t0
100 3 Persistence of Noncompact NHIMs

when δ C = ρ̃ − ρ. We verify this by calculating the right-hand side:

 t
ρ(t−t0 )
Ce + C eρ(t−τ ) δ C eρ̃(τ −t0 ) dτ
t0
  t 
= C eρ(t−t0 ) 1 + δ C e(ρ̃−ρ)(τ −t0 ) dτ
t0
 δ C  (ρ̃−ρ)(t−t0 ) 
= C eρ(t−t0 ) 1 + e −1
ρ̃ − ρ
= C eρ(t−t0 ) e(ρ̃−ρ)(t−t0 )
= C eρ̃(t−t0 ) .

Next, we prove by contradiction that

˜ t0 ) ≤ ψ̃(t, t0 ).
(t,

Thus, let
˜ t0 ) > ψ̃(t, t0 ).
t1 = inf {t ∈ R|t ≥ t0 and (t,

Note that  ˜ is the solution of a differential equation, hence continuous. We write

˜
(t, t0 ) = ψ̃(t, t0 ) + f (t), so we may assume that f (t) ≤ 0 for t ∈ [t0 , t1 ], but
there exist t ∈ (t1 , t2 ] arbitrary close to t1 such that f (t) > 0. Let f |[t1 ,t2 ] attain its
supremum at t, thus we have

sup f = f (t) > 0.

[t1 ,t]

We insert these estimates into the integral equality (3.27) and obtain
 t  
˜ t0 ) = ψ̃(t, t0 ) + f (t) ≤ ψ(t, t0 ) +
(t, ψ(t, τ ) δ ψ̃(τ, t0 ) + f (τ ) dτ
t0
 t
≤ ψ̃(t, t0 ) + ψ(t, τ ) δ f (τ )dτ
t1
≤ ψ̃(t, t0 ) + (t1 − t) δ sup ψ(t, τ ) sup f,
τ ∈[t1 ,t] [t1 ,t]

where we used that ψ̃(t, t0 ) satisfies (3.28) and that f |[t0 ,t1 ] ≤ 0. Now we choose t2
and therefore t sufficiently small that (t − t1 ) δ supτ ∈[t1 ,t] ψ(t, τ ) ≤ q < 1, which
leads to the contradiction

f (t) ≤ q sup f < f (t). 

[t1 ,t]
3.5 Growth Estimates for the Perturbed System 101

Proposition 3.18 (Perturbation of X -flow estimate) If δ, η, σ1 are sufficiently

small, then the flow of (3.24a) satisfies the modified exponential growth esti-
mates (3.17) for any y ∈ Bη (R; Y ) inserted.

Here Bη (R; Y ) denotes the closed ball of radius η in the space of bounded continuous
functions R → Y .
Proof Define the non-autonomous system v(t, x) = ṽ X (t, x, y(t)). This system is
a C 1 small perturbation of (v X ◦ g)(x), uniformly in t:

v(t, x) − (v X ◦ g)(x) ≤ ṽ X (x, y(t)) − v X (x, y(t)) + v X (x, (t)) − v X (x, h(x))
≤ δ + εDv (η + σ1 ),
Dx v(t, x) − D(v X ◦ g)(x) ≤ Dx ṽ X (x, y(t)) − Dx v X (x, y(t))
+ Dx v X (x, y(t)) − Dx v X (x, h(x))
≤ δ + εDv (η + σ1 ),

where εDv denotes the uniform continuity modulus of v and its first derivative,
which can be made small by choice of η, σ1 . We apply Lemma 3.16 to obtain
exponential growth numbers C X , ρ X for (3.24a) by choosing δ + εDv (η + σ1 )
sufficiently small. 
The following definition and lemma for flows on X are again formulated in back-
ward time, similar to Lemma 3.16.
Definition 3.19 (Approximate solution) Let X be a Riemannian manifold and
v(t, x) a time-dependent vector field on X . We call a continuous curve x : R → X
a (β, T )-approximate solution of v if for each interval [t2 , t1 ] ⊂ R with t1 − t2 ≤ T
and associated exact solution curve ξ of v with initial condition ξ(t1 ) = x(t1 ), it
holds that
sup d(x(t), ξ(t)) < β. (3.29)
t2 ≤t≤t1

It would have been easier to define approximate solutions as C 1 curves x such that
ẋ(t) − v(t, x(t)) < β. We shall want to work with C 0 -norms, though, and C 1
curves do not form a complete space under such norms. We use this continuous
curve definition to avoid any complications associated with non-completeness. We
still have the following result, as a discretized variant on variation by constants
estimates.
Lemma 3.20 (Growth of approximate solutions) Let X be a Riemannian mani-
fold, v(t, x) a time-dependent vector field on X , and x a (β, T )-approximate solution
of v. Assume that v generates a flow
t,t0 that satisfies the exponential growth esti-
mate (3.25) with C ≥ 1, ρ < 0. Let ξ0 denote the exact solution of v with initial
condition ξ0 (0) = x(0).
Then the distance dρ (x, ξ0 ) is finite on the interval (−∞, 0], and explicitly
bounded by
102 3 Persistence of Noncompact NHIMs

 C
dρ (x, ξ0 ) ≤ β 1 + . (3.30)
1 − eρ T

Proof Let ξi with i ∈ N be the associated exact solutions of x that satisfy (3.29)
 on
the interval [−(i + 1)T, −i T ]. We have d ξi (−(i + 1)T ), ξi+1 (−(i + 1)T ) < β.
Hence, dρ (ξi , ξi+1 ) < β C eρ(i+1)T on the interval (−∞, (i + 1)T ] by the exponen-
tial growth estimate.
Thus on each interval [−(i + 1)T, −i T ] we can use the triangle inequality to
estimate


i−1
dρ (x, ξ0 ) ≤ dρ (x, ξi ) + dρ (ξ j , ξ j+1 )
j=0


i−1
< β eρ i T + β C eρ( j+1)T
j=0
 C
≤β 1+ .
1 − eρ T

The union of all such intervals is (−∞, 0] hence (3.30) follows. 

Proposition 3.21 (Perturbation of Y -flow estimate) Let x be a (β, T )-approximate
solution to v X ◦ g. If T is sufficiently large and σ1 , ν, β are sufficiently small, then
the flow x of A(x(t)) has exponentially bounded growth as specified in (3.17), that
is, x (t, t0 ) ≤ CY eρY (t−t0 ) for all t ≥ t0 .
Proof Let ρ̃ = 21 (ρ− +ρY ) < ρY and choose T > 0 sufficiently large that C̃− eρ̃ T ≤
eρY T . Let xi be an exact solution to v X ◦ g such that supt∈[ti ,ti +T ] d(x(t), xi (t))
 ≤  β
ˆ  ˆ t
per Definition 3.19, hence the flow  of Â(xi (t)) satisfies (3.14), that is,   ≤
C̃− eρ− t . We decompose A(x(t)) = Â(xi (t)) + B(t) and estimate

B(t) = A(x(t)) − Â(xi (t))

≤ A(x(t)) − A(xi (t)) + A(xi (t)) − (D y vY ◦ g)(xi (t))
+ (D y vY ◦ g)(xi (t)) − Â(xi (t))
≤ DA d(x(t), xi (t)) + ε(ν) + 4 Cv σ1 2 π E − .

Note that A1 is bounded close to D y vY 1 ≤ Cv , and (3.15) and (3.20) were used
to estimate the third term. We thus have B(t) ≤ δ for any δ > 0 when σ1 , ν, β
are sufficiently small. Hence by Lemma 3.17, we have x (τ, τ0 ) ≤ C̃− eρ̃(τ −τ0 )
for any τ, τ0 ∈ [ti , ti + T ].
Now we cover the interval [t0 , t] by intervals [t0 + (i − 1)T, t0 + i T ] with cor-
responding exact solutions xi that approximate x. As in the proof of Lemma 3.16,
we write t − t0 = n T + τ and use the group property of the flow to obtain
3.5 Growth Estimates for the Perturbed System 103
 n
x (t, t0 ) ≤ C̃− eρ̃ T C̃− eρ̃ τ ≤ eρY n T C̃− eρY τ = C̃− eρY (t−t0 ) .

We note that CY = C̃− to complete the proof. 

Using these results, we choose T sufficiently large and δ, η, σ1 , β, ν sufficiently
small that the modified flows D
y , x satisfy exponential growth rates (3.17) when
curves y ∈ Bη (R; Y ) and (β, T )-approximate solutions x ∈ C 0 (R; X ) are inserted.
Lemma 3.22 (Variation of linear flow) Let X be a metric space and Y a Banach
α (X ; L(Y )) be a family of linear operators on Y that depends
space and let A ∈ Cb,u
uniformly α-Hölder continuous on x ∈ X , with Hölder coefficient Cα and 0 < α ≤ 1.
Let x denote the flow of A under a curve x ∈ C(I ; X ), and assume that it satisfies
the exponential growth condition (3.26).
Then the variation of the flow satisfies the Hölder-like estimate

Cα C 2 ρ(t−τ )
(1 − 2 )(t, τ ) ≤ e dρ (x1 , x2 )α eα ρ τ (3.31)
−α ρ

when dρ (x1 , x2 ) is finite.

Proof Let dρ (x1 , x2 ) be finite, let 1 , 2 be the associated flows of A and denote
ϒ = 1 − 2 . We have for ϒ the differential equation

d  
ϒ(t, τ ) = A(x1 (t)) ϒ(t, τ ) + A(x1 (t)) − A(x2 (t)) 2 (t, τ ), ϒ(t, t) = 0.
dt
By variation of constants we obtain
 t
ϒ(t, τ ) ≤ 1 (t, σ ) Cα d(x1 (σ ), x2 (σ ))α 2 (σ, τ )dσ
τ
 t
≤ Cα C 2 eρ(t−τ ) dρ (x1 , x2 )α eα ρ σ dσ
τ
Cα C2
≤ eρ(t−τ ) dρ (x1 , x2 )α eα ρ τ . 
−α ρ

3.6 Existence and Lipschitz Regularity

We start with proving Lipschitz estimates for two mappings onto curves in X, Y ,
respectively. These mappings will be combined to a contraction mapping T . Its fixed
points parametrized by x0 ∈ X will correspond to the unique solution curves of the
modified system (3.16) that stay bounded.
104 3 Persistence of Noncompact NHIMs

Let B ρ (I ; Y ) denote the Banach space of exponentially bounded, continuous

curves in Y on the interval I = (−∞, 0] and recall that in this chapter we always
ρ
assume ρ < 0. Additionally, we denote by Bη (I ; Y ) = B ρ (I ; Y ) ∩ Bη (I ; Y ) the
subset of curves y(t) which are moreover globally smaller than η. The closure of
ρ ρ
Bη (I ; Y ) is given by B̄η (I ; Y ) = B ρ (I ; Y ) ∩ Bη (I ; Y ).
ρ
Proposition 3.23 The space B̄η (I ; Y ) is a closed subspace of the Banach space
B ρ (I ; Y ), hence a complete metric space.
Proof Consider the evaluation mapping evt : B ρ (I ; Y ) → Y : y → y(t). For each
fixed t ∈ I this is a continuous mapping as y(t) ≤ yρ eρ t with eρ t a finite
number.
Let R = B(0; η) be the closed ball in Y , then we have

B̄ηρ (I ; Y ) = ev−1
t (R)
t∈I

as an intersection of closed preimages under evt , hence closed. 

For curves x(t) in X , we cannot construct a similar space B ρ (I ; X ) as X is not a
normed linear space. Instead we construct a (not necessarily complete) metric space.
Let B ρ (I ; X ) = C 0 (I ; X ), dρ denote the space of continuous curves equipped
ρ
with the metric (1.17) (which is allowed to take the value ∞) and let Bβ (I ; X ) be
the subset of curves x ∈ C 0 (I ; X ) that are (β, T )-approximate solutions to v X ◦ g
according to Definition 3.19. We suppress the dependence on T from the notation
(note that T in the equation below is a completely different object); both β and T were
fixed once and for all to fulfill the requirements of Proposition 3.21, we keep just the
subscript β as a reminder and to distinguish from the space B ρ (I ; X ). Let ρ < ρ X .
Then exact solutions of v X ◦ g have finite dρ distance on I , and by Lemma 3.20 the
ρ
distance of any two curves x1 , x2 ∈ Bβ (I ; X ) is finite, too.
 
We write T = TY ◦ TX , pr1 with

ρ ρ
T : B̄η (I ; Y ) × X → Bη (I ; Y ),
ρ ρ
TX : B̄η (I ; Y ) × X → Bβ (I ; X ), (3.32)
ρ ρ ρ ρ
TY : Bβ (I ; X ) × Bη (I ; Y ) → Bη (I ; Y ) ⊂ B̄η (I ; Y ),

for any ρY < ρ < ρ X . The map TX is defined by the flow

y of ṽ X ( · , y(t)) with
initial value x0 ∈ X , that is,

TX (y, x0 )(t) =
y (t, 0, x0 ). (3.33)

In [Hen81], the TY part of the contraction operator is indirectly defined by another

contraction. Instead, here, we will set up TY as a direct mapping
3.6 Existence and Lipschitz Regularity 105
 t
TY (x, y)(t) = x (t, τ ) f˜(x(τ ), y(τ ))dτ, (3.34)
−∞

where x is the flow of A(x(t)). This should ease proving smoothness properties of
T , which will subsequently imply smoothness of the invariant manifold.
ρ
Remark 3.24 Note that Bβ (I ; X ) is not a Banach space or even a complete metric
space.
 It will only appear as an intermediate space in the composition T = TY ◦
TX , pr1 though, so this does not affect the Banach fixed point arguments, as long
ρ
as the mappings TX , TY compose to a contraction T on B̄η (I ; Y ) uniformly in the
parameter x0 ∈ X . ♦
The following two propositions show that the maps TX , TY do indeed map into
their specified codomains, when parameters are chosen sufficiently small.
Proposition 3.25 If ζ is chosen such that (3.36) holds and δ, σ1 are sufficiently
small, then TY maps into Bη (I ; Y ).
ρ
Proof The conditions of Proposition 3.21 are satisfied for any x ∈ Bβ (I ; X ), so
the flow x of system (3.24b) satisfies exponential growth estimates with numbers
ρY , C Y .
Now, TY maps into Bη (I ; Y ) since
 t  
TY (x, y)(t) ≤ (t, τ )  f˜(x(τ ), 0) + D y f˜ y(τ ) dτ
−∞
CY
≤ (Cv σ1 + δ + ζ η) (3.35)
−ρY

which can be made smaller than η by choosing ζ such that

CY ζ 1
≤ (3.36)
−ρY 2

holds, as well as δ, σ1 sufficiently small. 

This shows that TY is well-defined in (3.32) after choosing ζ, δ, σ1 possibly smaller.
Note that the choice of ζ does not depend on any of the other small bounds. Similarly,
ρ
we verify that TX maps into Bβ (I ; X ).
Proposition 3.26 If δ, σ1 , and η are chosen sufficiently small, then TX maps into
ρ
Bβ (I ; X ).
ρ
Proof Let y ∈ Bη (I ; Y ) and x0 ∈ X . The curve x = TX (y, x0 ) is generated by the
vector field ṽ X ( · , y(t)) which is a small perturbation of v X ◦ g, since

ṽ X ( · , y(t)) − v X ◦ g0 ≤ ζ.

106 3 Persistence of Noncompact NHIMs

Let t2 − t1 ≤ T and let

t denote the flow of v X ◦ g. We apply the nonlinear variation
of constants estimate (E.2) and obtain for t ∈ [t1 , t2 ]

  t2  
d x(t),
t−t2
(x(t2 )) ≤ D
t−τ (x(τ )) ζ dτ
t
 t2
≤ C X eρ X (t−τ ) ζ dτ
t
C X ζ ρ X (t−t2 )
≤ e .
−ρ X

Thus, if we choose ζ sufficiently small that

C X ζ ρX T
e < β, (3.37)
−ρ X

then x is (β, T )-approximated by the exact solution

t,t2 (x(t2 )) on the interval [t1 , t2 ]
ρ
so x ∈ Bβ (I ; X ). 
The basic argument for the Perron method is encoded in the following lemma:
a y-bounded solution curve of the system (3.16) is equivalent to y ∈ Bη (I ; Y ) being
a fixed point of T , while this map T will be shown to be a contraction.
Lemma 3.27 (Cotton–Perron) Let x ∈ C(I ; X ), y ∈ Bη (I ; Y ) bounded, and
x0 ∈ X . Then the following statements are equivalent:
1. the pair (x, y) is a solution curve for the modified system (3.16) with partial
initial condition x(0) = x0 ;
2. y ∈ Bη (I ; Y ) is a fixed point of T ( · , x0 ) and x = TX (y, x0 ).
Proof The proof goes along the same lines as the classical Perron method for hyper-
bolic fixed points. As an intermediate step, we introduce the operator
 t
T̂Y (x, y, t0 )(t) = x (t, t0 ) y(t0 ) + x (t, τ ) f˜(x(τ ), y(τ ))dτ (3.38)
t0

and the following statement that is equivalent to those in the lemma:

3. the pair (x, y) is a fixed point of (TX , T̂Y ) for each t0 ∈ (−∞, 0].
Equivalence of 1 and 3 (with t0 = 0) is a direct consequence of equivalence of
differential and integral equations; the equation for y has been rewritten as a variation
of constants integral with respect to the nonlinear term f˜. If (x, y) is a fixed point of
(TX , T̂Y ) for t0 = 0, then this holds for any t0 ∈ (−∞, 0]. Note that the initial value
for y is left unspecified in both statements.
We finish by proving the implications 3 ⇒ 2 ⇒ 1. For the first we take the limit
t0 → −∞ in T̂Y (x, y, t0 ). Since x decays exponentially and y is bounded, it follows
that this limit is well-defined:
3.6 Existence and Lipschitz Regularity 107

∀ t ∈ I : lim T̂Y (x, y, t0 )(t) = TY (x, y)(t).

t0 →−∞

Hence, a fixed point of (TX , T̂Y ) is a fixed point of (TX , TY ). The last implication
can readily be verified by calculating the time derivatives of x = TX (y, x0 ) and
y = TY (x, y) to show that (x, y) is a solution of (3.16) with x(0) = x0 . 
Next, we prove that both TX , TY are Lipschitz, while the Lipschitz constant of TY
can be made arbitrarily small.
Lemma 3.28 Let ρY < ρ ≤ ρ X and 0 < qY < 1. If ζ is sufficiently small, then
Lip(TY ) ≤ qY .
ρ ρ
Proof Let (xi , yi ), i = 1, 2 be curves from Bβ (I ; X ) × Bη (I ; Y ). Let i be the
corresponding flows of A along the curves xi . Then the application of Lemma 3.22
in the Lipschitz case α = 1 leads to a Lipschitz estimate on TY for any ρY < ρ ≤ ρ X :

TY (x1 , y1 )(t) − TY (x2 , y2 )(t)

 t
≤ 1 (t, τ ) f˜(x1 (τ ), y1 (τ )) − 2 (t, τ ) f˜(x2 (τ ), y2 (τ ))dτ
−∞
 t
≤ 1 (t, τ ) − 2 (t, τ )  f˜(x1 (τ ), y1 (τ ))
−∞
+ 2 (t, τ )  f˜(x1 (τ ), y1 (τ )) − f˜(x2 (τ ), y2 (τ ))dτ
 t
Cv CY2 ρY (t−τ )
≤ e dρ (x1 , x2 ) eρ τ  f˜0
−∞ −ρ
 
+ CY eρY (t−τ ) D f˜0 dρ (x1 , x2 ) + y1 − y2 ρ eρ τ dτ
 
≤ ζ C dρ (x1 , x2 ) + y1 − y2 ρ eρ t .

Here C < ∞ depends only on the constants and additional integration factors
(ρ − ρY )−1 in the last integral, hence ζ C ≤ qY when ζ is small enough. 
Lemma 3.29 Let ρY < ρ < ρ X . If ζ, η are sufficiently small, then Lip(TX ) ≤ q X
for some q X > 1 independent of all small parameters.
ρ
Proof Let ξ1 , ξ2 ∈ X and y1 , y2 ∈ Bη (I ; Y ). For i = 1, 2, define vi (t, · ) =
ṽ X ( · , yi (t)) and let xi ∈ C 1 (I ; X ) be a solution of the system vi with initial condition
ξi . We compare the systems v1 , v2 :
 
v1 (t, · ) − v2 (t, · ) ≤ D y ṽ X  y1 (t) − y2 (t) ≤ Cv y1 (t) − y2 (t).

The flow
y1 has exponential growth numbers ρ X , C X . We view v2 as a small per-
turbation of v1 and apply the nonlinear variation of constants estimate (E.2) to obtain
108 3 Persistence of Noncompact NHIMs
 0
ρX t
d(x1 (t), x2 (t)) ≤ C X e d(ξ1 , ξ2 ) + C X eρ X (t−τ ) Cv y1 (τ ) − y2 (τ )dτ
t
 0
ρX t
≤ CX e d(ξ1 , ξ2 ) + C X Cv y1 − y2 ρ eρ X (t−τ ) eρ τ dτ
t
C X Cv
≤ C X eρ X t d(x1 , x2 ) + y1 − y2 ρ eρ X t .
ρX − ρ

Now
  C X Cv
dρ TX (y1 , ξ1 ), TX (y2 , ξ2 ) ≤ sup C X d(ξ1 , ξ2 ) e(ρ X −ρ) t + y1 − y2 ρ e(ρ X −ρ) t
t≤0 ρX − ρ
C X Cv
≤ C X d(ξ1 , ξ2 ) + y1 − y2 ρ
ρX − ρ

exhibits a Lipschitz constant q X for TX that does not depend on any of the small
parameters. 
Proposition 3.30 (Extension of solution is bounded in Y ) Let (x, y)(t) be a solu-
tion of the perturbed system (3.16) satisfying y ∈ Bη (I ; Y ) for t ≤ 0. For ζ, δ, σ1
sufficiently small, the forward extension to t ≥ 0 has y ∈ Bη (R; Y ).
Proof First of all, choose ζ, δ, σ1 sufficiently small such that by (3.35), we have
y ∈ Bη/2 (I ; Y ). Proceeding by contradiction, let t0 be the first time after which y(t)
becomes larger than η, thus

t0 = sup {t ∈ R|∀ τ ≤ t : y(τ ) ≤ η}.

The curve x(t) is a (β, T )-approximate solution to v X ◦ g on the interval

 (−∞, t0 ],
so from Proposition 3.25 we conclude that y ∈ Bη/2 (−∞, t0 ]; Y . The continuity
of y contradicts the assumption that t0 is the supremum. 

Completing the Proof of Existence and Lipschitz Regularity

We finally put things together and prove that a unique persistent manifold M̃ exists
and that it is Lipschitz.
Since TX satisfies a fixed Lipschitz estimate, we can choose ζ small enough to
ρ
obtain q X ·qY < 1. Thus, T is a contraction on B̄η (I ; Y ) for each fixed ρY < ρ < ρ X ;
ρ
ζ will depend on ρ though. According to Proposition 3.23, B̄η (I ; Y ) is a complete
metric space, so the Banach fixed point theorem shows that there is a unique y ∈
ρ ρ
B̄η (I ; Y ) fixed point of T ; it holds moreover that y ∈ Bη (I ; Y ). This contraction
also depends (uniformly) on the parameter x0 ∈ X , hence we obtain a fixed point
map
∞ : X → Bηρ (I ; Y ), (3.39)
3.6 Existence and Lipschitz Regularity 109

satisfying the relation

∀ x0 ∈ X : ∞ (x0 ) = T ( ∞ (x0 ), x0 ). (3.40)

The superscript ∞ indicates that this map is obtained as a limit of applying the
uniform contraction T . The parameter dependence in T is Lipschitz, so the map ∞
will be Lipschitz as well.
By Proposition 3.30, the fixed point y = ∞ (x0 ) is bounded by η for all time
ρ
and Bη (I ; Y ) = Bη (I ; Y ) as sets, so y is the unique η-bounded solution with partial
initial data x(0) = x0 . In combination with the evaluation map y → y(0), we obtain
the mapping
h̃ : X → B(0; η) ⊂ Y : x0 → ∞ (x0 )(0). (3.41)

Its graph M̃ = Graph(h̃) is the unique invariant manifold of the modified sys-
tem (3.16) and is Lipschitz as well.
  by graphs of small functions X → Y , it follows
Since both M and M̃ are described
that they are homeomorphic and h̃ 0 ≤ η. We can choose another, arbitrarily small
η instead. This requires us to choose smaller δ  , σ1 parameters as well. But as can
 
be seen from (3.18), η does not depend on δ, σ1 , so the newly found h̃  0 ≤ η will
actually be unique in the original η-sized neighborhood as well.

3.7 Smoothness

To study smoothness of ∞ , we can formally differentiate the fixed point rela-

tion (3.40) with respect to x0 to obtain contractive mappings T (k) for the k-þorder
derivatives of maps and then apply the fiber contraction theorem, see Appen-
dix D. If we assume that satisfies (3.40), then Proposition C.3 shows that (at least
formally)
  • 
Dk (x0 ) = x0 T ( (x 0 ), x 0 ) · Pl,k−m D (x 0 ) ,
Dly Dm (3.42)
l,m≥0
l+m≤k
(l,m) =(0,0)

which can be rewritten as a fiber contraction map on Dk by isolating that term

(l = 1, m = 0) on the right-hand side as

Dk (x0 ) = D y T ( (x0 ), x0 ) · Dk (x0 ) + · · ·

All the remaining terms are expressions in the lower order derivatives Dn (x0 ) for
n < k only; these form the base space in the fiber contraction theorem.
ρ
The derivatives Dk and Dly Dm
x0 T in (3.42) do not exist on the space Bη (I ; Y ) as
codomain, however. Indeed, if they did, we could have applied the implicit function
110 3 Persistence of Noncompact NHIMs

theorem right away. Instead, the derivatives Dk are only well-defined on spaces7
B kρ+μ (I ; Y ), where μ < 0 is an arbitrarily small additional exponential growth
rate. Derivatives of the maps TX , TY do not exist at all. By using the fiber contraction
theorem and interpreting the Dk TX , Dk TY as ‘formal derivatives’ in some appropriate
way, we can still show, though, that the Dk are higher derivatives of that converge
to the derivatives of ∞ under iteration of the fiber contraction maps (3.42). The
gap condition ρY < r ρ X will show up in the requirement that (3.42) is contractive
for k ≤ r (and finally k + α ≤ r when considering Hölder continuity). In case of
uniform continuity (i.e. when α = 0) we make use of the strict inequality to seize
some of the spectral space left for the terms μ < 0.
The interpretation of DT and its constituents DTX and DTY as true derivatives is
ρ ρ
obstructed already by the fact that neither Bη (I ; Y ) nor Bβ (I ; X ) are smooth Banach
manifolds8 , hence these can never be the (co)domain of differentiable maps. Thus,
the chain rule
D n+1 = D y T · D n + Dx0 T

cannot be used to conclude the existence of D n+1 from D n by induction. On the

other hand, we can find ‘formal tangent bundles’ of these spaces on which DTX , DTY
are defined as ‘formal derivatives’, and we even have explicit formulas (3.43) for these
maps. From here on we shall use the notation D f to indicate a formal derivative and

D f to indicate that a function f is truly differentiable. We shall not make precise
the notion of ‘formal’, but heuristically these formal objects can be seen as limits of
well-defined real smooth manifolds and derivatives, see Sect. 3.7.7.
First, we outline the procedure of obtaining ∞ as a truly differentiable map by
careful manipulation of these formal derivatives. This is followed by the details of
working out the definitions and estimates. Finally, we show how everything gen-
eralizes to higher derivatives. This last step adds more complexity, but requires no
fundamentally new ideas.
Higher derivatives of functions involving variables or values in X need to be treated
with some care, as these are not naturally defined. In such expressions, the deriva-
tives are with respect to normal coordinates at the base point in domain and range,
according to Definition 8.6. I should point the reader to Appendix C: it establishes
the essential basic ingredient for this section on (higher) smoothness, namely how
exponential growth estimates carry over to continuity and higher derivatives of the
flow. Additionally, building on bounded geometry and Definition 2.9, a framework
is set up to work with these notions on the manifold X .

7 Note that the spaces B kρ+μ (I ; Y ) are to be understood as the codomains of the maps , hence
the Dk as multilinear operators into these. The spaces Yη play the same role in [Van89, Def. 3.10].
8 At least, they are not smooth Banach manifolds in a natural way, see the discussion in Sect. 3.7.4.
3.7 Smoothness 111

3.7.1 A Scheme to Obtain the First Derivative

The map T is not differentiable. Instead, we shall use the scheme below to obtain
ρ
differentiability of ∞ . The sequence { n }n≥0 of maps X → Bη (I ; Y ) is defined
by (x0 ) = T ( (x0 ), x0 ) and ≡ 0. We prove the differentiability of the n
n+1 n 0

by induction and finally conclude that ∞ is differentiable as well.

1. First, we propose candidate formal derivatives (3.43) of TX , TY . These are obtained
naturally by standard differentiation and variational techniques, postponing for
the moment the question of which spaces these maps are well-defined on. We
define DT in terms of the formal derivatives DTX and DTY .
  
2. The pair (T, DT ) acts as a uniform fiber contraction on pairs of maps


( n , D n ) : TX → Bηρ (I ; Y ) × B ρ+μ (I ; Y )


when ρY < ρ + μ ≤ ρ < ρ X holds, both in case of μ = 0 and μ < 0 small.

ρ ρ
3. There are appropriate formal tangent bundles of the spaces Bη (I ; Y ), Bβ (I ; X ) on
which these formal derivatives are well-defined. Moreover, these formal tangent
bundles can be endowed with a topology such that DTX , DTY , and DT are
  
uniformly continuous into bundles with slightly larger exponential growth rate
ρ+μ. Under appropriate assumptions (and with μ = α ρ) these formal derivatives
are α-Hölder continuous.
4. The fiber contraction theorem D.1 can be applied. It follows from 2 that DT has
ρ 
a unique fixed point D ∞ : TX → T Bη (I ; Y ), and from 3 that the map
 

    
→ DT ( ) · D ∞ : C 0 X ; Bηρ (I ; Y ) → b L TX ; T Bηρ+μ (I ; Y )
  

 ρ+μ 
into bounded sections of the bundle π : L TX ; T Bη (I ; Y ) → X is contin-
   
uous. Thus we can conclude that D n converges in b L TX ; B ρ+μ (I ; Y )

to the unique fixed point D ∞ , simultaneously with n → ∞ . See (3.45)

and (3.47) for precise definitions of these spaces. Moreover, D ∞ is uniformly

or Hölder continuous.
5. There is a family of maps, given by restricting the domain I of curves,

T b,a : Bηρ ([a, 0]; Y ) × X → Bηρ ([b, 0]; Y )

that approximate T , and moreover these T b,a are differentiable maps between
Banach manifolds whose derivatives DT b,a approximate the formal deriva-
tive DT .

6. With the continuous embedding B ρ (I ; Y ) → B ρ+μ (I ; Y ) and the previous point,
we show that if n : X → B ρ+μ (I ; Y ) is differentiable, then
112 3 Persistence of Noncompact NHIMs

D n+1 = D y T · D n + Dx0 T
  

is the derivative of n+1 : X → B ρ+μ (I ; Y ).

7. Finally, we use Theorem D.2 to conclude that since the sequence n converges
to ∞ and its derivatives satisfy D n → D ∞ , it must hold that D ∞ = D ∞
 
as a map into B ρ+μ (I ; Y ).
In the subsequent sections we shall work out the details of this scheme. With some
care, the same ideas generalize to higher derivatives.

3.7.2 Candidate Formal Derivatives

We first explicitly give the candidate mappings for the derivatives of TX , TY . From
now on, we will use shorthand notation x y (t) = TX (y, x0 )(t) =
y (t, 0, x0 ). The
spaces that these maps act on will be made more precise in the following sections;
ρ ρ
δx, δy denote variations of curves x ∈ Bβ (I ; X ) and y ∈ Bη (I ; Y ), respectively,
and δx0 ∈ Tx0 X .
 
Dx0 TX (y, x0 ) δx0 (t) = D
y (t, 0, x0 ) · δx0 (3.43a)

 0
 
D y TX (y, x0 ) δy (t) = D
y (t, τ, x y (τ )) D y ṽ X (x y (τ ), y(τ )) δy(τ )dτ,
 t
(3.43b)

  t
Dx TY (x, y) δx (t) = x (t, τ ) Dx f˜(x(τ ), y(τ )) δx(τ )
 −∞
 
+ Dx x · δx (t, τ ) f˜(x(τ ), y(τ ))dτ, (3.43c)


  t
D y TY (x, y) δy (t) = x (t, τ ) D y f˜(x(τ ), y(τ )) δy(τ )dτ, (3.43d)
 −∞
 t
 
Dx x · δx (t, τ ) = x (t, σ ) DA(x(σ )) δx(σ ) x (σ, τ )dσ. (3.43e)
 τ

The correctness of these expressions pointwise in t can be checked by variation of

constants and follows from Theorem E.2. Note also that the expressions above are
linear in the variations δx, δy, δx0 . The map (3.43e) is only included in the list for
its occurrence in (3.43c).
3.7 Smoothness 113

3.7.3 Uniformly Contractive Fiber Maps

We establish uniform boundedness of the formal derivative maps (??) as linear oper-
ators on δx, δy, and δx0 . The estimates are straightforward generalizations of those
in Sect. 3.6. The operator norms are induced by  · ρ norms. We have the following
list of estimates:

 
Dx0 TX (y, x0 ) = sup D
y (t, 0, x0 )δx0  e−ρ t ≤ sup C X eρ X t e−ρ t ≤ C X ,
 t∈I t∈I
δx0 =1
D y TX (y, x0 ) ≤ sup
 t∈I
δyρ =1
 0
 
D
y (t, τ, x y (τ )) D y ṽ X (x y (τ ), y(τ )) δy(τ )dτ · e−ρ t
t
 0
C X Cv
≤ sup C X eρ X (t−τ ) Cv eρ(τ −t) dτ ≤ ,
t∈I t ρX − ρ
 t  
 
D y TY (x, y) ≤ sup x (t, τ ) D y f˜(x(τ ), y(τ )) δy(τ )dτ · e−ρ t
 t∈I −∞
δyρ =1
 t
CY ζ
≤ sup CY eρY (t−τ ) ζ eρ(τ −t) dτ ≤ ,
t∈I −∞ ρ − ρY
 t
 
 Dx x · δx (t, τ ) ≤ x (t, σ ) DA(x(σ )) δx(σ ) x (σ, τ )dσ
 τ
 t
≤ CY eρY (t−σ ) Cv δxρ eρ σ CY eρY (σ −τ ) dσ
τ
C2 C
δxρ eρY (t−τ ) eρ t ,
v
≤ Y
−ρ
 t  
 
Dx TY (x, y) ≤ sup x (t, τ ) Dx f˜(x(τ ), y(τ )) δx(τ )
 t∈I −∞
δxρ =1
  
+  Dx x ·δx (t, τ ) f˜(x(τ ), y(τ)) dτ · e−ρ t

 t 2
C Cv ζ ρY (t−τ ) ρ t
≤ sup CY eρY (t−τ ) ζ eρ τ + Y e e dτ · e−ρ t
t∈I −∞ −ρ
CY ζ C 2 Cv ζ
≤ + Y .
ρ − ρY ρ · ρY
114 3 Persistence of Noncompact NHIMs

These estimates show that

D y T = Dx TY · D y TX + D y TY ,
   
Dx0 T = Dx TY · Dx0 TX (3.44)
  

are bounded linear maps when ρY < ρ < ρ X . Since we have some spectral elbow
room, we can first choose a value for ρ and then choose μ < 0 sufficiently close to
zero, such that this inequality holds both for ρ and ρ + μ. If ζ is sufficiently small,
then D y T  ≤ q < 1 can be satisfied. This shows that (T, DT ) is a uniform fiber
 ρ 
contraction on δy ∈ B ρ+μ (I ; Y ) over base curves y ∈ Bη (I ; Y ), and with additional
parameter x0 ∈ X . It can also be viewed as a fiber mapping of maps D over base

maps . Let us define
  μ   
S0 = C 0 X ; Bηρ (I ; Y ) and S1 = b L TX ; B ρ+μ (I ; Y ) , (3.45)

μ
where S0 is equipped with the supremum norm and S1 is interpreted as bounded
sections of the bounded geometry bundle over X of linear maps between TX and the
trivial bundle π : X × B ρ+μ (I ; Y ) → X , equipped with the (supremum/operator)
norm
D  = sup D (x0 )L(Tx0 X ;B ρ+μ (I ;Y )) .
 x0 ∈X 

Then (T, DT ) can also be viewed as a fiber mapping



μ μ
(T, DT ) : S0 × S1 → S0 × S1 ,

 
( , D ) → (x0 , δx0 ) → T ( (x0 ), x0 ), (3.46)

  
D y T ( (x0 ), x0 ) · D (x0 ) + Dx0 T ( (x0 ), x0 ) · δx0 .
  

As such, it is again a uniform fiber contraction since the contraction was uniform
in y and x0 to begin with, and the supremum norm does not affect the contraction
factor q < 1.

3.7.4 Formal Tangent Bundles

Derivatives of the maps TX , TY should be defined between tangent bundles of the

ρ ρ ρ
spaces Bη (I ; Y ) and Bβ (I ; X ). This is problematic for both spaces: Bη (I ; Y ) is a
ρ
subspace of the Banach space B (I ; Y ), but it has empty interior. The restriction
ρ
to (β, T )-approximate solutions of v X ◦ g creates a similar problem for Bβ (I ; X ),
but here, construction of the tangent bundle faces an additional obstruction. There
3.7 Smoothness 115

is no clear way to define local coordinates around a solution curve x. The obvious
method would be by constructing a tubular neighborhood of x and represent nearby
ρ
curves x̃ in the tubular neighborhood. But the metric on Bβ (I ; X ) allows x̃ to diverge
exponentially from x even if dρ (x, x̃) is small. Thus the tubular neighborhood would
need to be of infinite size to contain x̃(t) for all t ∈ I , which is generally not possible.
Since any finite-size tubular neighborhood does not contain a full neighborhood of
the curve x, we cannot use local coordinates to define tangent spaces.
Instead, we shall construct formal tangent bundles. These are just convenient
spaces to model variations of curves on; they are natural extensions of true Banach
tangent spaces, see Sect. 3.7.7. The primary role of these bundles is to introduce a
topology that allows us to show that DT is uniformly or Hölder continuous.
ρ 
A formal tangent bundle of Bη (I ; Y ) can be constructed rather easily: B ρ (I ; Y )
is a Banach space, so its tangent bundle is canonically identified as TB ρ (I ; Y ) =
ρ
B ρ (I ; Y ) × B ρ (I ; Y ). We then define the formal tangent bundle of Bη (I ; Y ) by
restricting the base:

T Bηρ (I ; Y ) = TB ρ (I ; Y )| Bηρ (I ;Y ) ∼
= Bηρ (I ; Y ) × B ρ (I ; Y ) (3.47)


with induced topology and norm.

ρ
To define a formal tangent bundle of Bβ (I ; X ), we consider variations δx of a
curve x as sections of a pullback bundle: δx ∈ (x ∗ (TX )). That is, δx ∈ C(I ; TX )
is such that δx(t) ∈ Tx(t) X for each t ∈ I . We equip this space with the norm that is
natural for our problem, namely

δxρ = sup δx(t) e−ρ t ,

t∈I
 
and denote it by B ρ (I ; x ∗ (TX )) = (x ∗ (TX )),  · ρ . The curves δx ∈ B ρ (I ; x ∗
ρ
(TX )) form the formal tangent space over one curve x ∈ Bβ (I ; X ). The complete
ρ
formal tangent bundle is then defined as the coproduct over all curves x ∈ Bβ (I ; X ),

ρ

TBβ (I ; X ) = B ρ (I ; x ∗ (TX )). (3.48)
 ρ
x∈Bβ (I ;X )

A curve δx lives above a specific base curve x, so there is no direct way of com-
paring two curves δx1 , δx2 with different base curves x1 , x2 ; (3.48) was constructed
as a coproduct without topological structure. We add a topology based on parallel
ρ
transport. This requires the base curves x ∈ Bβ (I ; X ) to be differentiable, so we
consider the bundle 
ρ
TBβ (I ; X )C 1 (3.49)

116 3 Persistence of Noncompact NHIMs

ρ
restricted to differentiable9 base curves x ∈ Bβ (I ; X ) ∩ C 1 . Variational curves
ρ ∗  ρ
δx 0∈ B (I ; x (TX )) are isometrically mapped onto curves δx ∈ B (I ; Tx(0) X ) =
C (I ; Tx(0) X ),  · ρ by

˜ x : δx → δx,
  
δx(t) = (x|t0 )−1 δx(t). (3.50)

Let the normal coordinate radius δ X be X -small as in Definition 2.8. If we now restrict
all base curves x under consideration to a small neighborhood
ρ
Ux̄ = {x ∈ Bβ (I ; X ) ∩ C 1 |x(0) ∈ B(x̄; δ X ), (3.51)

i.e., the curves x that start in the open ball B(x̄; δ X ) ⊂ X , then there exists a unique
shortest geodesic γx̄,x(0) from x(0) to x̄ for each x ∈ Ux̄ . Parallel transport along
these geodesics induces a local trivialization 10 of TB( x̄; δ ). This in turn induces a
ρ  X

local trivialization of TBβ (I ; X ) C 1 :


ρ τx̄
TBβ (I ; X )|Ux̄ / Ux̄ × B ρ (I ; Tx̄ X ) (3.52)

ttt
t
tt
π ttt
t
tt p1
 t tt
ρ yt
Bβ (I ; X )|C 1 ⊃ Ux̄

 
˜ x (δx) . The tran-
The trivialization map is given by τx̄ (x, δx) = x, (γx̄,x(0) ) ◦ 
sition maps between overlapping local trivializations Ux̄1 ∩ Ux̄2 = ∅ are induced by
transition functions
 
ϕ2,1 : B(x̄2 ; δ X ) ∩ B(x̄1 ; δ X ) × Tx̄2 X → Tx̄1 X : (ξ, ν) → (γx̄2 ,ξ ◦ γξ,x̄1 ) · ν

between local trivializations of TB(x̄i ; δ X ). The map ϕ2,1 is uniformly Lipschitz by

Lemma 2.6 and linear in the fiber. This induces a Lipschitz continuous transition
function τx̄2 ◦ τx̄−1
1
that depends on the base curve x ∈ Ux̄1 ∩ Ux̄2 only through
x(0) ∈ X ; this dependence is uniform since X has bounded geometry. Thus the
bundle satisfies Definition 2.14, and the order of bounded geometry is actually equal
to k − 2 when X has k-th order bounded geometry.

9 This does not cause problems since TX actually maps into curves x ∈ C 1 . The fiber contraction
theorem only requires that the base space has a globally attractive fixed point. Since the fixed point
is a C 1 curve, we can simply restrict to this subset of curves.
10 We make the specific choice to trivialize TB( x̄; δ ) by parallel transport along geodesics. Any
X
other trivialization with uniformly bounded transition maps would also suffice for our purposes
ρ
and induce a trivialization of TBβ (I ; X )|Ux̄ (see also the alternative viewpoint on this trivializa-

tion below). This explicit choice is somewhat natural in this context, though, and it shows that a
trivialization with these properties does exist.
3.7 Smoothness 117

ρ 
We endow the bundle TBβ (I ; X )C 1 with the topology induced by these local

trivializations. Note that this topology is induced by a locally defined distance func- 
ρ
tion, so we can express uniform and Hölder continuity ofmaps on TBβ (I ; X )C 1 .
ρ 
That is, if (x1 , δx1 ) and (x2 , δx2 ) are elements of TBβ (I ; X )C 1 such that dρ (x1 , x2 ) <

δ X , then the topology is induced by the locally defined distance function
 
˜ x2 (δx2 ) − 
d (x1 , δx1 ), (x2 , δx2 ) = dρ (x2 , x1 ) + (γx1 (0),x2 (0) )  ˜ x1 (δx1 )ρ .
(3.53)
The transition functions τx̄2 ◦ τx̄−1
1
are uniformly Lipschitz, so they preserve uniform
and Hölder continuity moduli up to a constant. Therefore, overlapping trivializations
define the same topology on their intersection, with compatible local distances. To
summarize, we have
ρ ρ 
Proposition 3.31 The spaces T Bη (I ; Y ) and TBβ (I ; X )C 1 are well-defined nor-
 
med vector bundles of bounded geometry, and they have a (local) distance structure.
The topologies introduced above allow us to express uniform and Hölder con-
ρ
tinuity of the maps (??). The topology on T Bη (I ; Y ) is clear and explicit from
ρ   ρ
the topology on B ρ (I ; Y ). For TBβ (I ; X )C 1 let x ∈ Bβ (I ; X ) ∩ C 1 be a curve

and δx ∈ B ρ (I ; x ∗ (TX )) a variational curve at x. The topology is induced by the
isometric representation

˜ x · δx ∈ B ρ (I ; Tx̄ X )
 = (γx̄,x(0) ) · 
δx
ρ 
of δx. Uniform continuity of maps (??) that have TBβ (I ; X )C 1 as (co)domain can

thus be checked by switching to a local trivialization, that is, substitute


δx(t) = (x|t0 ) · (γx(0),x̄ ) · δx(t)

and then use the known topology on Ux̄ × B ρ (I ; Tx̄ X ). In explicit calculations of
ρ
continuity with respect to the base Bβ (I ; X )∩C 1 , we shall thus add parallel transport
terms such as those above to the maps (??) and let these act on δx  ∈ B ρ (I ; Tx̄ X ).

Alternative Viewpoints
ρ 
Instead of the immediate trivialization (3.52) of the bundle TBβ (I ; X )C 1 , we can also

introduce an intermediate viewpoint that corresponds to only applying the parallel
transport term  ˜ x , but not (γx̄,x(0) ) in the local neighborhood B(x̄; δ X ). We view
ρ ρ 
ev0 : Bβ (I ; X ) → X as a bundle; this identifies TBβ (I ; X )C 1 as a bundle over X

as well, via ev0 ◦ π . Let B ρ (I ; TX ) X denote the space of (continuous, exponential
growth) functions δx : I → TX such that π ◦ δx  is constant into X , viewed as a
bundle over X .
118 3 Persistence of Noncompact NHIMs

ρ  ˜
TBβ (I ; X )C 1  / B ρ (I ; X )|C 1 × X B ρ (I ; TX ) X
β (3.54)

qqq q
qqq
qqqqq
π qqqp qq
qqq evt ◦ p2
 xqqq 1 xqqq
ρ
Bβ (I ; X )|C 1 TX
oo
oo
ev0
o oooo
o π
 woooo
X

ρ 
This commutative diagram shows that TBβ (I ; X )C 1 can be identified via 
˜ with

ρ
the fiber product bundle Bβ (I ; X )|C 1 × X B ρ (I ; TX ) X over X . This identification
is natural in the sense that no local trivialization of X or TX is used. The second
component B ρ (I ; TX ) X of this bundle contains the variational curves δx.  This is a
(nontrivial) bundle over X , but its projection onto the base π ◦evt : B ρ (I ; TX ) X → X
factors through TX . The fact that π ◦ evt is constant for t ∈ I simply expresses that
 ∈ B ρ (I ; TX ) X maps into a fixed tangent space Tξ X . This shows that a local
each δx
trivialization σ : TX | B(x̄;δ X ) → B(x̄; δ X ) × Rn naturally lifts to a local trivialization

σ̃ : B ρ (I ; TX )| B(x̄;δ X ) → B(x̄; δ X ) × B ρ (I ; Rn ).

We have chosen local trivializations of TX by parallel transport along geodesics, i.e.

(γx̄,ξ ), since this construction is compatible with the bounded geometry of X in
the sense that trivialization chart transitions are Cbk maps by Proposition
 2.13.
ρ
We also introduce a reformulation of the topology on TBβ (I ; X )C 1 using frames,

as an alternative to the explicit formulation in terms of parallel transport above. This
allows us to abstract away these ideas into a lighter notation in the next section and
only recall the full details when required.
Let ex̄ : Rn → Tx̄ X be a choice11 of orthonormal frame at x̄ ∈ X . We can extend
this to an orthonormal frame e on TB(x̄; δ X ) by parallel transport of the frame ex̄
along geodesics emanating from x̄. As a second step, we further extend the frame e
along any curve x ∈ Ux̄ , again by parallel transport.12
We adopt the notation v f = f −1 · v to express a vector v ∈ Tx X with respect
to a frame f at x, and use this notation more generally on the tensor bundle of X .
Now let v be a vector field and ω a one-form on X , then the construction of e above
leads to

11 The precise choice does not matter and will drop out in the final, relevant equations. The relative
choice of frame along curves is what matters.
12 Note that e does not define a (global) frame on TX . The choice of frame at x(t) depends not just
ρ
on the point x(t) ∈ X , but on the whole curve x ∈ Bβ (I ; X )|C 1 . Another curve x̃ with x(t) = x̃(t)
will generally induce a different frame in Tx(t) X .
3.7 Smoothness 119

v(x(t))e = e−1
x̄ · (γx̄,x(0) ) · (x|t ) · v(x(t)),
0

ω(x(t))e = ω(x(t)) · (x|t0 ) · (γx(0),x̄ ) · ex̄ , (3.55)

and naturally extends to the tensor bundle of X .

3.7.5 Continuity of the Fiber Maps

We prove the uniform and Hölder continuous dependence on x, y, and x0 of the

maps (??) using a combination of techniques. One is the variation of constants
formula to get expressions for the variation of flows when changing a parameter.
Such variations require us to compare the variational curves over different base
curves; for this, we use the topologies of the formal tangent bundles in Sect. 3.7.4,
while we measure the variation of vector fields with the formulation of continuity via
parallel transport in Proposition 2.13. Together these lead to holonomy terms along
the base paths (see Fig. 3.7), in addition to the variation of constants terms that would
simply occur in Rn . These holonomy terms can be estimated with Lemma 2.19 and
do not essentially alter the estimates.
We use Nemytskii operator techniques as laid out in Appendix B to conclude that
functions such as A and f can be interpreted as uniformly continuous maps onto
curves with some μ < 0 exponential growth norm. Instead of uniform continuity,
we can also obtain Hölder continuity if the original maps are Hölder continuous and
if we view the Nemytskii operator as a mapping into a space with norm  · αρ . In
other words, we replace the uniform continuity modulus by the explicit α-Hölder
continuity modulus. Hölder continuity precisely fits the problem, so in that case there
is no need anymore to add a small μ < 0 to the exponential growth norms.

One Example in Full Detail

As an example, let us consider continuity of the map (3.43c) with respect to x ∈

ρ
Bβ (I ; X ), that is, x → Dx TY (x, y). To be able to explicitly use the topology on
ρ  
TBβ (I ; X )C 1 , we switch to a local trivialization neighborhood Ux̄  x1 , x2 as


Fig. 3.7 Paths involved in the x 2 (t)

holonomy term

γt
γ0
x
x1 (t)
120 3 Persistence of Noncompact NHIMs

in (3.51). We choose x̄ = x1 (0) to simplify expressions; any other choice for x̄

 ∈ B ρ (I ; Tx̄ X ) be the
can be obtained by a transition of trivialization charts. Let δx
representation of an arbitrary variational curve in the fiber of this trivialization.
Note that (3.43c) is defined in terms of (3.43e). We estimate continuity of the
separate components and build towards the full expression. Let us first focus on the
ρ
continuity of x → x (t, τ ), which is a map Bβ (I ; X ) → L(Y ) for fixed t, τ ∈ I .
Proposition 3.32 For any μ < 0, the variation

ϒ t,τ = x2 (t, τ ) − x1 (t, τ ) (3.56)

of the linear flow x on Y satisfies continuity estimate (3.57).

Proof We extend the ideas from the proof of Lemma C.8. The variation ϒ t,τ satisfies
the differential equation

d t,τ
ϒ = A(x2 (t)) x2 (t, τ ) − A(x1 (t)) x1 (t, τ )
dt  
= A(x2 (t)) ϒ t,τ + A(x2 (t)) − A(x1 (t)) x1 (t, τ ),

which leads to a variation of constants integral that can be estimated as


 t,τ  t    
ϒ  ≤ x (t, σ )A(x2 (σ )) − A(x1 (σ ))x (σ, τ )dσ
2 1
τ
 t
≤ CY eρY (t−σ ) ε Ã (dρ (x2 , x1 )) eμ σ CY eρY (σ −τ ) dσ
τ
eμ τ
≤ CY2 eρY (t−τ ) ε Ã (dρ (x2 , x1 )) . (3.57)
−μ

Here we use ideas from Appendix B; we applied Corollary B.3 to obtain A as a uni-
ρ
formly continuous fiber mapping Bβ (I ; X ) → B μ (I ; L(Y )) with continuity modu-
lus ε Ã (that depends on μ). 
ρ
Thus, the flow xt,τ depends uniformly continuously on x ∈ Bβ (I ; X ) when viewed
as a flow with ρY -exponential growth and measured with an additional exponential
factor eμ τ .
Remark 3.33 In the previous proposition, if A is α-Hölder continuous, then we can
replace μ by α ρ to obtain a similar, α-Hölder continuous result using Lemma B.2.
♦
To show that x → Dx x (t, τ ) is continuous as well, we first write down the

corresponding variation in the bundle trivialization chart:
3.7 Smoothness 121
 
 − Dx x1 · δx
Dx x2 · δx  (t, τ )
 
 t
 
=  ) x2 (t, σ )
x2 (t, σ ) DA(x2 (σ )) (x2 |σ0 ) (γx2 (0),x̄ ) δx(σ
τ
− (2  1)dσ
 t
 
=  ) x2 (t, σ ) − (2  1)dσ,
x2 (t, σ ) DA(x2 (σ ))e δx(σ (3.58)
τ

where the notation (2  1) means that we take the first expression and replace all
2’s by 1’s (note that γx2 (0),x̄ = γ0 in the first term and (γx1 (0),x̄ ) = 1 in the second
term). The last line is just a rewrite in terms of the frame as in (3.55) and suppresses
all parallel transport terms. We separately estimate continuity of the three factors
in the integrand, and insert the estimate of Proposition 3.32 for the variation • in
the first and third factor. Note that δx is the same over both curves x1 and x2 in this
trivialization.
For the middle factor DA(x(t))e , we again apply Nemytskii operator techniques
from Appendix B. But in this case we have to combine these with holonomy terms,
due to the fact that comparison of DA at nearby points ξ2 , ξ1 ∈ X only makes sense
after identification of the tangent spaces Tξ2 X and Tξ1 X .
1 according to Definition 2.9. Then for any μ < 0,
Proposition 3.34 Let A ∈ Cb,u
the map
  ρ  
x → t → DA(x(t))e : Ux̄ ⊂ Bβ (I ; X ) → B μ I ; L(Tx̄ X ; L(Y )) (3.59)

1,α
is uniformly continuous. If moreover A ∈ Cb,u , then the map (3.59) is α-Hölder with
μ replaced by α ρ.
Proof Let x1 , x2 ∈ Ux̄ . We introduce another frame f to directly compare DA
at points x1 (t), x2 (t). Let f x1 (t) = ex1 (t) : Rn → Tx1 (t) X and define f x2 (t) =
(γt ) · f x1 (t) . Thus, the frames e and f at x2 (t) are both defined in terms of the
frame ex̄ ; ex2 (t) by parallel transport along x2 ◦ γ0 and f x2 (t) by parallel transport
along γt ◦x1 , see Fig. 3.7. Since f x1 (t) = ex1 (t) , we can rewrite the difference of (3.59)
at points on these curves as
 
DA(x2 (t))e − DA(x1 (t))e = DA(x2 (t))e − DA(x2 (t)) f
 
+ DA(x2 (t)) f − DA(x1 (t)) f .

The first term can be estimated by the holonomy defect along the loop

γ0−1 ◦ x2 |0t ◦ γt ◦ x1 |t0

using Lemma 2.19 and the second term using the continuity of DA and
Proposition 2.13. Together, this leads to
122 3 Persistence of Noncompact NHIMs

DA(x2 (t))e − DA(x1 (t))e 

   
≤ DA 1 −  γ0−1 ◦ x2 |0t ◦ γt ◦ x1 |t0  + εDA d(x2 (t), x1 (t))
 
≤ Cv C dρ (x2 , x1 ) eρ t + εDA d(x2 (t), x1 (t)) .

If dρ (x2 , x1 ) eρ t ≥ δ X , then we use the boundedness estimate 1 − (γ ) ≤ 2

for any closed loop γ and Remark 2.12 to effectively extend the local to a global
continuity modulus. We can recover any α-Hölder continuity from the Lipschitz
holonomy estimate, again by using the fact that the holonomy is bounded by 2 in
combination with Lemma 1.20.  
With the same arguments as in Lemma B.2, it follows that x → t → DA(x(t))e
is uniformly or α-Hölder continuous, and we denote its continuity modulus by ε DA .
Note that ε DA  does not depend on the trivialization chart since all estimates are
uniform with respect to these charts. 
Proposition 3.35 For any μ < 0 and uniformly in x̄ ∈ X , the map x → Dx x (t, τ )
 ρ
satisfies continuity estimate (3.60) in a trivialization neighborhood Ux̄ ⊂ Bβ (I ; X ).
Proof We combine the estimates from Propositions 3.32 and 3.34 and obtain
for (3.58)
 
 − Dx x1 · δx
 Dx x2 · δx  (t, τ )
 
 t
 
≤ x2 (t, σ ) DA(x2 (σ ))e δx(σ  ) x2 (σ, τ ) − (2  1)dσ
τ
 t
 
≤ x2 (t, σ ) − x1 (t, σ ) DA(x2 (σ ))e  x2 (σ, τ )
τ
   
+ x1 (t, σ ) DA(x2 (σ ))e − DA(x1 (σ ))e  x2 (σ, τ )
   
+ x1 (t, σ ) DA(x1 (σ ))e  x2 (σ, τ ) − x1 (σ, τ ) δx   eρ σ dσ
ρ
 t μ σ
e
≤ CY2 eρY (t−σ ) ε A (dρ (x2 , x1 )) Cv CY eρY (σ −τ )
τ −μ
+ CY eρY (t−σ ) εDA
 (dρ (x 2 , x 1 )) e
μσ
CY eρY (σ −τ )
eμ τ 
δx

+ CY eρY (t−σ ) Cv CY2 eρY (σ −τ ) ε A (dρ (x2 , x1 ))   eρ σ dσ
−μ ρ
ρY (t−τ )
 
(ρ+μ)τ   
≤Ce ε(dρ (x2 , x1 )) e δx .
ρ
(3.60)

1
We absorbed all constants and integration factors such as −μ into the general constant
C and combine the continuity moduli into one; Cv is a global bound on all vector
fields including A and its derivatives. 
We finally plug estimate (3.60) into Eq. (3.43c). We repeat the Nemytskii and
holonomy arguments for f˜ and Dx f˜ (just as for A and DA) to obtain a uniform
continuity estimate for
3.7 Smoothness 123
 
x → Dx TY (x, y) : Ux̄ → L B ρ (I ; Tx̄ X ); B ρ+μ (I ; Y ) .

1,α
both for any μ < 0, or with μ = α ρ when A, f ∈ Cb,u . That is, Dx TY ( · , y) is a

ρ
map that given a curve x ∈ Bβ (I ; X )|C 1 , linearly maps a variational curve δx over
ρ
x to a variational curve δy in the trivial bundle T Bη (I ; Y ). We can formulate this

more abstractly as
 ρ  ρ  
α
Dx TY ( · , y) ∈ b,u Bβ (I ; X )|C 1 ; L TBβ (I ; X )C 1 ; B ρ+μ (I ; Y ) ,
 

that is, Dx TY ( · , y) is a uniformly α-Hölder bounded section of the bounded geometry


bundle  ρ   ρ
π : L TBβ (I ; X )C 1 ; B ρ+μ (I ; Y ) → Bβ (I ; X )|C 1 .


Continuity in the Other Cases

We treated the continuity for one of the maps (??) with respect to a single variable.
The continuity in all other cases can be shown in a similar fashion. Many arguments
can be repeated, but each of these maps also has its own peculiar details which
makes that I have not been able to find one general, abstract way to prove continuity
of all of these maps at once. In this section we shall focus on these specific details
and not repeat the recurring elements. Let me reiterate that the uniform continuity
results hold for any μ < 0 sufficiently small, and these can be replaced by α-Hölder
continuity when μ is replaced by α ρ and the spectral gap condition (1.11) is satisfied
for r = 1 + α.
First of all, note that continuity with respect to the combined variables follows
directly from continuity with respect to each separate variable since we have explicit
uniform or Hölder continuity moduli. If f (x, y) has continuity moduli εx , ε y with
respect to x, y, respectively, then

 f (x2 , y2 ) − f (x1 , y1 ) ≤  f (x2 , y2 ) − f (x1 , y2 ) +  f (x1 , y2 ) − f (x1 , y1 )

≤ εx (d(x2 , x1 )) + ε y (d(y2 , y1 ))
 
≤ (εx + ε y ) d((x2 , y2 ), (x1 , y1 ))

shows that εx + ε y is a continuity modulus for f . We assumed w.l.o.g. that εx , ε y are

non-decreasing, while all choices of distance on the product space are equivalent, so
we leave it unspecified.
Let us start with the easy cases. Continuity of the map (3.43c) as a function of y,
that is,
 
y → Dx TY (x, y) : Bηρ (I ; Y ) → L B ρ (I ; x ∗ (TX )); B ρ+μ (I ; Y ) ,

124 3 Persistence of Noncompact NHIMs

requires no additional details: only f and Dx f depend on y ∈ Y , and we can reapply

ρ
the arguments above to show that these depend continuously on y ∈ Bη (I ; Y ). No
ρ
holonomy terms are present since T Bη (I ; Y ) is a trivial bundle. That is, we can
 ρ ρ
directly compare Dx TY at different y1 , y2 ∈ Bη (I ; Y ); keeping x ∈ Bβ (I ; X ) fixed
 ρ
means that everything is situated in the fixed fiber T x Bβ (I ; X )|C 1 = B ρ (I ; x ∗ (TX ))

and no holonomy terms are required.
Continuity of the map (3.43d), i.e. D y TY (x, y), both with respect to x and y follows

along the same lines. Neither case requires holonomy arguments; we just apply the
Nemytskii technique to D y f and reuse Proposition 3.32 to show continuity with
respect to x. 
ρ
The formal derivatives (3.43a) and (3.43b) of TX map into TBβ (I ; X )C 1 ; here

we have to apply holonomy arguments in the codomain. Let us first focus on
 
x0 → Dx0 TX (y, x0 ) : X → L Tx0 X ; B ρ+μ (I ; Tx̄ X )


ρ+μ 
with a local trivialization Ux̄ × B ρ+μ (I ; Tx̄ X ) within13 the bundle TBβ (I ; X )C 1

with additional μ in the exponential growth norm on the fibers. Note that Dx0 TX (y, · )

could actually be considered as a bundle map on the vector bundle TX that is linear
on each tangent space Tx0 X . We consider a local trivialization of TB(x̄; δ X ) ⊂
TX by parallel transport along geodesics: this is equivalent to trivialization by a
normal coordinate chart for the purpose of measuring continuity, while it matches
ρ
the trivialization of TBβ (I ; X )|Ux̄ . This will lead to a holonomy term.

For any x0 ∈ B(x̄; δ X ) we have TX (y, x0 ) ∈ Ux̄ by construction, so let e denote
ρ
the frame introduced by the trivialization of TBβ (I ; X )|Ux̄ , i.e. by parallel transport

along solution curves TX (y, x0 ). On the other hand, let f denote a frame introduced
by local parallel transport. We define

f x1 (t) = ex1 (t) and f x2 (t) = (γt ) · f x1 (t) .

 
It follows from Lemma C.10 that Dx0 TX (y, · ) = t → D
y (t, 0, · ) satisfies the

correct type of continuity estimates, but with respect to local charts (or equivalently,
with respect to f determined by local parallel transport) instead of the choice of
ρ
frame e, defined by the topology of TBβ (I ; X )C 1 . To examine the difference, let

x0,1 , x0,2 ∈ B(x̄; δ X ) denote two initial conditions and xi = TX (y, x0,i ), i = 1, 2,
ρ
their respective solution curves for a fixed y ∈ Bη (I ; Y ). We also fix x0,1 = x̄ for
convenience. Then we have

13 Embeddings B ρ → B ρ+μ are continuous, so we can view Ux̄ × B ρ+μ (I ; Tx̄ X ) as a local
ρ
trivialization of a subset of TBβ (I ; X ) with ρ  = ρ + μ.

3.7 Smoothness 125

D
y (t, 0, x0,2 )e − D
y (t, 0, x0,1 )e
 
= D
y (t, 0, x0,2 )e − D
y (t, 0, x0,2 ) f
 
+ D
y (t, 0, x0,2 ) f − D
y (t, 0, x0,1 ) f
 
= (γ0−1 ◦ x2 |0t ) − (x1 |0t ◦ γt−1 ) · D
y (t, 0, x0,2 ) · (γ0 )
 
+ D
y (t, 0, x0,2 ) f − D
y (t, 0, x0,1 ) f .

This shows that uniform and Hölder continuity with respect to the topology of
ρ
TBβ (I ; X )C 1 is equivalent to the same continuity with respect to normal coor-

dinate charts, since the additional holonomy term can be estimated in the same way
as in Proposition 3.34. Continuity of
 
y → Dx0 TX (y, x0 ) : Bηρ (I ; Y ) → L Tx0 X ; B ρ+μ (I ; Tx̄ X )


follows in the same way, if we first apply Corollary C.12 to obtain the continuity
estimates with respect to the frame f .
Finally, we consider continuity of the map (3.43b),
 
D y TX (y, x0 ) ∈ L B ρ (I ; Y ); B ρ+μ (I ; Tx̄ X )


ρ
with respect to y ∈ Bη (I ; Y ) and x0 ∈ X . We apply Corollary C.12 and Lemma 3.29
to conclude that D
y (t, τ, x y (τ )) depends α-Hölder or uniformly continuously
on y. Lemma 3.29 in combination with a Nemytskii operator argument shows that
D y ṽ X induces a uniformly continuous map
   
y → t → D y ṽ X (x y (t), y(t)) : Bηρ (I ; Y ) → B μ I ; L(Y ; TX )

with μ replaced by α ρ in the Hölder case. For continuity with respect to x0 ∈ X we

need to replace application of Corollary C.12 by that of Lemma C.10 for dependence
of x y = TX (y, x0 ) on x0 . Again the continuity estimates obtained are with respect to
the frame e and we use Lemma 2.19 to estimate the additional holonomy term when
switching to the frame f .

3.7.6 Application of the Fiber Contraction Theorem

In Sect. 3.7.3 we already established that the fiber mapping (T, DT ) in formula (3.46)

is uniformly contractive. With the results of the previous sections on formal tangent
bundles and continuous formal derivatives, we can now apply the fiber contraction
theorem, see Appendix D.
126 3 Persistence of Noncompact NHIMs

Proposition 3.36 For any μ < 0, the fiber mapping (3.46) has a unique, globally
μ
attractive fixed point ( ∞ , D ∞ ) ∈ S0 × S1 , while it also holds that D ∞ ∈ S10 .
 
μ
Proof In the notation of Theorem D.1 we take X = S0 and Y = S1 as in (3.45) with
ρ, μ such that ρY < ρ + μ < ρ < ρ X holds. The fiber mapping is F = (T, DT ), as

in (3.46). The first two conditions of Theorem D.1 are satisfied due to the arguments
in Sect. 3.7.3, while the third condition that DT is continuous can be obtained from

the results in Sect. 3.7.5 as follows.
First, note that (T, DT ) is a well-defined, uniformly contractive fiber mapping

both when acting on B ρ (I ; Y ) and on B ρ+μ (I ; Y ) variational curves. Thus, for each
μ
n ≥ 0 we have D n ∈ S10 → S1 , where the embedding is continuous. The same
 ∞
conclusion holds for D by a simple uniform contraction argument. Next, we

view DT as a map

μ
DT : S0 × S10 → S1 . (3.61)


Note that we set μ = 0 in the domain only. To obtain continuity of (3.61) with respect
to the base variable ∈ S0 , it is sufficient to check that the maps
 
y → D y T (y, x0 ) : Bηρ (I ; Y ) → L B ρ (I ; Y ); B ρ+μ (I ; Y ) ,

 
y → Dx0 T (y, x0 ) : Bηρ (I ; Y ) → L Tx0 X ; B ρ+μ (I ; Y ) (3.62)


are uniformly continuous, uniformly in x0 ∈ X . Continuity of (3.61) with respect

μ
to the base S0 (with fixed fiber part D ∈ S10 → S1 ) then follows from the

interpretation of (3.62) as acting on maps ( , D ) with the supremum norm on

S0 . The maps (3.62) are defined by the chain rule formula (3.44) in terms of the
ρ
derivative maps (??). A variation of y ∈ Bη (I ; Y ) can be distributed over the product
(we only estimate the variation of D y T with respect to y, but the variation of Dx0 T
 
is completely analogous),

D y T (y2 , x0 ) − D y T (y1 , x0 )ρ+μ,ρ

 
 
≤ Dx TY (x y2 , y2 ) · D y TX (y2 , x0 ) − D y TX (y1 , x0 ) ρ+μ,ρ
  
 
+  Dx TY (x y2 , y2 ) − Dx TY (x y1 , y1 ) · D y TX (y1 , x0 )ρ+μ,ρ (3.63)
  
≤ Dx TY (x y2 , y2 )ρ+μ,ρ+μ · D y TX (y2 , x0 ) − D y TX (y1 , x0 )ρ+μ,ρ
  
+ Dx TY (x y2 , y2 ) − Dx TY (x y1 , y1 )ρ+μ,ρ · D y TX (y1 , x0 )ρ,ρ .
  

The  · ρ2 ,ρ1 denote operator norms on linear (bundle) maps from B ρ1 to B ρ2 spaces.
In the factor that is not varied we can simply take the operator norm between functions
of either ρ or ρ + μ exponential growth: in Sect. 3.7.3 we have seen that the fiber
3.7 Smoothness 127

maps are uniformly bounded linear in both cases. The factor that is varied satisfies a
uniform continuity estimate in  · ρ+μ,ρ -norm, a result from Sect. 3.7.5. Note that
ρ
we use the topology defined in Sect. 3.7.4 on the intermediate space TBβ (I ; X )C 1 ,

as well as a local trivialization to express the difference D y TX (y2 , x0 )−D y TX (y1 , x0 ).
 
As a result of the fiber contraction theorem, we conclude that there is a unique,
globally attractive fixed point ( ∞ , D ∞ ) of the fiber mapping (3.46). Note that

D ∞ is already well-defined as an element of S10 , although it is only proven to be

μ
attractive in S1 . 
As a next step, we show that the fixed point map D ∞ that we found is actually

continuous. This follows from a standard uniform contraction argument.
μ
Proposition 3.37 For any μ < 0, the map D ∞ ∈ S1 is uniformly continuous. If

we set μ ≤ α ρ and the assumptions of Theorem 3.2 are satisfied with r ≥ 1 + α,
then it is α-Hölder continuous.
Proof First note that it is sufficient to prove the statement for μ < 0 sufficiently
small, or μ = α ρ in case of α-Hölder continuity; by continuous embedding of
exponential growth spaces, it then automatically follows for any μ that is more neg-
ative. We use local trivializations by parallel transport to express continuity moduli
of functions with domain TX .
The assumptions of Theorem 3.2 imply that the spectral gap condition ρY <
ρ + μ < ρ < ρ X is satisfied. Since D ∞ is (the fiber part of) the fixed point of the

uniform contraction (T, DT ), we have for any two x1 , x2 ∈ B(x̄; δ X ) ⊂ X that


D ∞ (x2 ) − D ∞ (x1 )ρ+μ

 
= D y T ( (x2 ), x2 ) · D ∞ (x2 ) + Dx0 T ( ∞ (x2 ), x2 ) − (2  1)ρ+μ
∞
  
≤ D y T ( (x2 ), x2 ) − D y T ( (x1 ), x1 )ρ+μ,ρ · D ∞ (x2 )ρ
∞ ∞
  
+ D y T ( ∞ (x1 ), x1 )ρ+μ,ρ+μ · D ∞ (x2 ) − D ∞ (x1 )ρ+μ
  
+ Dx0 T ( ∞ (x2 ), x2 ) − Dx0 T ( ∞ (x1 ), x1 )ρ+μ
 
 
≤ εD y T (L + 1)d(x2 , x1 ) + q D ∞ (x2 ) − D ∞ (x2 )ρ+μ
  
 
+ εDx T (L + 1)d(x2 , x1 ) .


Here L = Lip( ∞ ) denotes the Lipschitz constant of ∞ ∈ S0 , while q < 1 is

ρ
the uniform contraction factor of D y T on the fibers of Bη (I ; Y ) × B ρ+μ (I ; Y ). We

saw in Sect. 3.7.5 that the maps D y T, Dx T have appropriate continuity moduli into
 
B ρ+μ (I ; Y ). Finally, we move the contraction term to the left-hand side, divide by
1 − q, and obtain
128 3 Persistence of Noncompact NHIMs

1   
D ∞ (x2 ) − D ∞ (x1 )ρ+μ ≤ εD y T (L + 1)d(x2 , x1 )
  1−q 
 
+ εDx T (L + 1)d(x2 , x1 ) .


μ
This shows that D ∞ ∈ S1 has the same type of continuity modulus as DT . 
 

3.7.7 Derivatives on Banach Manifolds

We can recover the maps (??) as true derivatives on Banach manifolds if we restrict
to bounded time intervals J ⊂ I = R≤0 . The maps TX , TY naturally restrict to such
intervals, either exactly, or in a well-behaved approximate way. By restricting to
ρ ρ
intervals J = [a, 0] with a < 0, the spaces Bη (J ; Y ) and Bβ (J ; X ) become Banach
manifolds and the restrictions of TX , TY become continuously differentiable maps
on these.
ρ ρ
Lemma 3.38 For any −∞ < a < 0, the spaces Bη (J ; Y ) and Bβ (J ; X ) with
J = [a, 0] a bounded interval are well-defined Banach manifolds.
ρ
Proof We first treat the easy case Bη (J ; Y ). For any −∞ < a < 0, the norms  · ρ
ρ
and  · 0 are equivalent on B ρ (J ; Y ). The set Bη (J ; Y ) is an open ball of radius η
ρ
in the Banach space B (J ; Y ), so it follows that Bη (J ; Y ) is a Banach manifold as
0
ρ
an open subset of B (J ; Y ).
In the same way, the metrics dρ and d0 are equivalent on B ρ (J ; X ), but here we
need to do a little more work to show the following. 
ρ
Proposition 3.39 The set Bβ (J ; X ) is open in B ρ (J ; X ).
ρ
Proof Let x ∈ Bβ (J ; X ), hence by Definition 3.19, x is approximated on each
interval of length |[t1 , t2 ]| ≤ T by t →
(t, t2 , x(t2 )), where
denotes the flow of
v X ◦ g, the horizontal part of the unperturbed vector field (3.10). The map
 
(t, t2 ) → d x(t),
(t, t2 , x(t2 ))

is continuous, and since it is defined on a compact subset of J × J , it attains its

supremum  
η1 = sup sup d x(t),
(t, t2 , x(t2 )) ,
t2 ∈J t∈[t2 −T,t2 ]

so it must hold that η1 < β. Let x̃ ∈ B(x; η2 ) ⊂ B ρ (J ; X ) with

e−ρ a
η2 = (β − η1 ) .
1 + C X eρ X T
3.7 Smoothness 129

We apply the triangle inequality and obtain

     
d x̃(t),
(t, t2 , x̃(t2 )) ≤ d x̃(t), x(t) + d x(t),
(t, t2 , x(t2 ))
 
+ d
(t, t2 , x(t2 )),
(t, t2 , x̃(t2 ))
≤ e ρ a η2 + η1 + C X e ρ X T e ρ a η2
 
≤ 1 + C X eρ X T eρ a η2 + η1 < β.

This shows that all functions in the ball B(x; η2 ) ⊂ B ρ (J ; X ) are still (β, T )-
ρ
approximate solutions of v X ◦ g, and thus Bβ (J ; X ) is open. 
ρ
From here on we shall not always precisely distinguish between Bβ (J ; X ) and
B ρ (J ;
X ) anymore.
We introduce a local coordinate chart κx around a curve x ∈ B ρ (J ; X ) using the
exponential map (see also [Kli95, Sect. 2.3]):
 
κx : Ux ⊂ B ρ (J ; X ) → B ρ (J ; x ∗ (TX )) : ξ → t → exp−1
x(t) (ξ(t)) . (3.64)

The vector bundle x ∗ (TX ) is trivial, so the space of sections B ρ (J ; x ∗ (TX )) is

isomorphic to B ρ (J ; Rn ). An explicit trivialization of x ∗ (TX ) (and thus isomorphism
of sections) can be obtained, for example if x ∈ C 1 , using parallel transport as
in (3.50) and identification of Tx(0) X ∼ = Rn by a choice frame, but we refrain
from making such a choice here; one reason is that curves x ∈ B ρ (J ; X ) are only
assumed continuous. The chart κx bijectively covers a full rinj (X ) neighborhood of
x with respect to the metric d0 , hence a neighborhood of size rinj (X ) e−ρ a > 0 with
respect to dρ . Recall that δ X is X -small as in Definition 2.8; let us choose a radius
δa = δ X e−ρ a , such that all bounded geometry results also hold true in these induced
charts κx . Then the coordinate transition map

κx2 ◦ κx−1
1
: B ρ (J ; x1∗ (TX )) → B ρ (J ; x2∗ (TX )) (3.65)

is a bijection between isomorphic Banach spaces that is as smooth as the exponential

map of X . 
Remark 3.40 We could choose isomorphisms τx : B ρ (J ; x ∗ (TX )) → B ρ (J ; Rn )
to obtain one fixed Banach space B ρ (J ; Rn ) as model for the manifold B ρ (J ; X ).
The τx are linear isometries so they preserve norms and smoothness, hence there is
no need to explicitly make this identification. Specifically, note that the construction
of B ρ (J ; x ∗ (TX )) as the pullback along a curve x that is merely continuous, does
not influence the smoothness of coordinate transformations on B ρ (J ; X ). ♦
We shall again call charts in this atlas ‘normal coordinate charts’, since they are
induced by normal coordinates on X along the curve x ∈ B ρ (J ; X ). By construction
all bounded geometry results carry over to these induced charts. In particular, we
can measure maps in terms of their coordinate representations. We will use this fact
without always explicitly mentioning it.
130 3 Persistence of Noncompact NHIMs

The tangent space of B ρ (J ; X ) at a point x can be canonically identified as

Tx B ρ (J ; X ) ∼
= B ρ (J ; x ∗ (TX )) (3.66)

as follows. Let s → xs : (−ε, ε) ⊂ R → B ρ (J ; X ) be a C 1 family of curves

such that x0 = x and let ξs = κx (xs ) be their representation in the coordinate chart
B ρ (J ; x ∗ (TX )). The chart κx is induced by normal coordinates, so

dρ (x0 , xs ) = sup d(x0 (t), xs (t)) eρ t = sup  exp−1

x0 (t) (x s (t)) e
ρt
= ξs ρ
t∈J t∈J

shows that  · ρ is the canonical norm on the chart B ρ (J ; x ∗ (TX )). Then v =
d 
 ρ ρ ∗
ds ξs s=0 represents a tangent vector in Tx B (J ; X ), while v ∈ B (J ; x (TX )) by
construction.
This completes our exposition of the manifold structure of B ρ (J ; X ). We shall
again exclusively make use of induced normal coordinate charts (3.64), in order to
use results on bounded geometry.
The map TX can be restricted to curves on any subinterval J = [a, 0] ⊂ I = R≤0 .
Let us introduce the restriction operator on curves

ρa : C(I ; Z ) → C(J ; Z ) : z → z| J . (3.67)

ρ ρ
This operator acts naturally on Bη (I ; Y ) and Bβ (I ; X ) and there is a natural family
of restrictions TXa of TX such that

TXa ◦ ρa = ρa ◦ TX for any − ∞ < a < 0. (3.68)

Proposition 3.41 Let J = [a, 0] with −∞ < a < 0. Then

ρ
TXa : Bηρ (J ; Y ) × X → Bβ (J ; X ) (3.69)

is a differentiable map between Banach manifolds with partial derivatives given by

a natural restriction of the maps (3.43a) and (3.43b).
ρ
Proof Let s → y + s δy ∈ Bη (J ; Y ) be a one-parameter family of curves and let
ρ
κx : B(x; δa ) ⊂ Bβ (J ; X ) → B ρ (J ; x ∗ (TX ))

be an induced normal coordinate chart centered around the curve x = TXa (y, x0 ).
The map TX is Lipschitz, so for s sufficiently small, TXa (y + s δy, x0 ) maps into
B(x; δa ). The vector field ṽ X ( · , (y + s δy)(t)) depends smoothly on the parameter
s and generates xs = TXa (y + s δy, x0 ). We apply Theorem E.2 with
3.7 Smoothness 131

d  
ṽ X x(t), (y + s δy)(t) = D y ṽ X (x(t), y(t)) · δy(t)
ds s=0

to obtain (3.43b) as the pointwise derivative of evt ◦ TXa , for any t ∈ J .

Now we only need to show that (3.43b) viewed as derivative pointwise in t satisfies
linear approximation estimates, uniformly for all t ∈ J with respect to dρ . We work
in the local chart κx , so xs (t) is represented in the normal coordinate chart centered
at x(t), while the curve s → y + s δy is canonically represented in B ρ (I ; Y ) with
derivative δy.  
Since s → κx ◦ TXa (y + s δy, x0 ) (t) ∈ C 1 (R; Tx(t) X ), we can apply the mean
value theorem to estimate
 
 TXa (y + s δy, x0 ) − TXa (y, x0 ) − D y TX (y, x0 ) · s δy (t)

  
≤  D y TX (y + σt δy, x0 ) − D y TX (y, x0 ) · s δy (t) (3.70)
 
≤ D y TX (y + σt δy, x0 ) − D y TX (y, x0 ) δyρ eρ t |s|
 

for some σt ∈ (0, s). Note that σt will in general depend on t ∈ J , so there is (a priori)
not one curve y + σ δy such that (3.70) holds for all t ∈ J at once. In Sect. 3.7.5 we
ρ
showed that D y TX : T Bη (I ; Y ) → TB ρ+μ (I ; X )|C 1 is continuous; on the bounded
  
interval J the norms  · ρ and  · ρ+μ are equivalent, so D y TX is continuous into
ρ ∗ 
B (J ; x (TX )) as well. Using this fact, we plug the result above into the definition
of (directional) derivative and verify

1
lim T a (y + s δy, x0 ) − TXa (y, x0 ) − D y TX (y, x0 ) · s δyρ
s→0 s X 
|s|
≤ lim sup D y TX (y + σt δy, x0 ) − D y TX (y, x0 ) δyρ = 0.
s→0 t∈J s  

Therefore, the derivative of TXa at (y, x0 ) in the direction of δy is given by

D y TX (y, x0 ) · δy restricted to the interval J . The limit is uniform on δyρ = 1 and

this map is continuous and linear in δy, so TXa is continuously partially differentiable
with respect to y.
If we use a local chart around x0 ∈ X , then we find in the same way that TXa is
continuously partially differentiable with respect to x0 . Thus, TXa is (continuously)
differentiable. 
The map TY does not have a similarly natural restriction since it depends on the
complete ‘history’ of the curves x, y through the integral from −∞. The dependence
on earlier times is exponentially suppressed, though. Therefore, we construct a family
of restrictions that approach TY when the amount of additional history in the input
goes to infinity. Let −∞ < a ≤ b < 0 and define the family TYb,a of restrictions as
132 3 Persistence of Noncompact NHIMs

ρ
TYb,a : Bβ ([a, 0]; X ) × Bηρ ([a, 0]; Y ) → Bηρ ([b, 0]; Y ),
  t
(x, y) → t → x (t, τ ) f˜(x(τ ), y(τ ))dτ for each t ∈ [b, 0]. (3.71)
a

Proposition 3.42 The family TYb,a approximates TY in the sense that for any fixed
b ∈ (−∞, 0], we have
TYb,a ◦ ρa → ρb ◦ TY (3.72)
ρ ρ
when a → −∞, uniformly in x, y ∈ Bβ (I ; X ) × Bη (I ; Y ).
Proof This follows from straightforward estimates:
 a
TYb,a ◦ ρa (x, y) − ρb ◦ TY (x, y)ρ ≤ sup e−ρ t x (t, τ ) f˜(x(τ ), y(τ ))dτ
t∈[b,0] −∞
 a
≤ sup e−ρ t CY eρY (t−τ ) ζ dτ
t∈[b,0] −∞
CY ζ ρY (b−a)−ρ b
≤ e . 
−ρY
In a same way we define approximate families for (3.43c) and (3.43d), denoted
by Dx TYb,a and D y TYb,a , respectively.
 

Corollary 3.43 The families Dx TYb,a and D y TYb,a approximate (3.43c) and (3.43d)
 
in the same way as in Proposition 3.42.
Proposition 3.44 Let −∞ < a ≤ b < 0. Then TYb,a is a differentiable map between
Banach manifolds.
Proof We shall only show that TYb,a is continuously partially differentiable respect
to x. Continuous partial differentiability with respect to y follows along the same
lines and total differentiability then is a direct consequence of these (also in the
Banach manifold setting, see [Lan95, Prop. 3.5]).
ρ ρ
Let x ∈ Bβ (J ; X ) and y ∈ Bη (J ; Y ) with J = [a, 0]. Let κx be an induced
normal coordinate chart around x and let

xs = x + s δx ∈ B ρ (J ; x ∗ (TX ))
ρ
be a one-parameter family of curves in Bβ (J ; X ), represented in the chart κx (for s
sufficiently
 small). Then δx ∈ B ρ (J ; x ∗ (TX )) is naturally identified as the derivative
d
x  .
ds s s=0
We shall show that the partial derivative Dx TYb,a is given by the formal deriva-
tive (3.43c), but with J as domain of integration and interpreted as a mapping into
B ρ ([b, 0]; Y ). Again, we split the full expression into manageable pieces and apply
the mean value theorem.
3.7 Smoothness 133
 
 TYb,a (xs , y) − TYb,a (x, y) − Dx TYb,a (x, y) · s δx (t)
 t

≤ x (t, τ ) f˜(xs (τ ), y(τ )) − x (t, τ ) f˜(x(τ ), y(τ ))
s
a

− x (t, τ ) Dx f˜(x(τ ), y(τ )) sδx(τ ) − (Dx x · sδx)(t, τ ) f˜(x(τ ), y(τ ))dτ

 t
 
≤ xs (t, τ ) − x (t, τ ) − Dx x · s δx (t, τ )  f˜(x(τ ), y(τ ))
a 

+ x (t, τ )  f˜(xs (τ ), y(τ )) − f˜(x(τ ), y(τ )) − Dx f˜(x(τ ), y(τ )) s δx(τ )

+ xs (t, τ ) − x (t, τ )  f˜(xs (τ ), y(τ )) − f˜(x(τ ), y(τ ))dτ

and application of Theorem E.2 shows that formula (3.43e) for Dx xs · δx is the

derivative of xs . We use this for a mean value theorem estimate14 in the first and
third15 term to arrive at
 t
   
≤  Dx xσ · s δx (t, τ ) − Dx x · s δx (t, τ ) ζ
a  

+ CY e ρY (t−τ )
Dx f˜(xσ (τ ), y(τ )) − Dx f˜(x(τ ), y(τ )) |s| δxρ eρ τ
 
+  Dx xσ · s δx (t, τ ) Dx f˜(xσ (τ ), y(τ )) |s| δxρ eρ τ dτ

 t
≤ C eρY (t−τ ) ε(|σ | δxρ ) e(ρ+μ)τ |s| δxρ ζ
a
 
+ CY eρY (t−τ ) εDx f |σ | δxρ eρ τ |s| δxρ eρ τ
CY2 Cv
+ δxρ eρY (t−τ ) eρ t ζ |s| δxρ eρ τ dτ.
−ρ

We applied Proposition 3.35 to estimate the variation of Dx ; the induced normal


coordinate charts and Proposition 2.13 allow us to freely switch between parallel
transport and normal coordinates for estimating differences. All exponential norms
are equivalent on the compact interval J , so with the usual estimates we see that this
expression is o(|s|), uniformly for all δxρ = 1. 
We have thus converted the map T = TY ◦ (TX , pr1 ) to a Banach manifold setting
by defining it on curves restricted to compact time intervals. Although all estimates

14 The intermediate point σ in the mean value theorem implicitly depends on both t and τ and will

be different in each term. This does not affect the uniform estimates, so we suppress this dependence
in the notation.
15 We applied the intermediate value theorem to both factors in the third term. This is not strictly

necessary: we could also have applied it to only one of these, and apply a uniform continuity
estimate to the other term. That would still have yielded a size estimate ε(|s|) |s| = o(|s|). When
we generalize to higher derivatives, we shall make use of this fact: at least one of the factors will
be differentiable and yield a factor |s|, while the other term(s) can be estimated by a continuity
modulus ε(|s|).
134 3 Persistence of Noncompact NHIMs

were already in place, this technicality allows us to draw the conclusions of the final
points 6 and 7 in the scheme in Sect. 3.7.1.
Lemma 3.45 (The n have true derivatives) Fix μ < 0 and let n : X →
ρ
Bη (I ; Y ) be differentiable into B ρ+μ (I ; Y ). Recursively define n+1 (x0 ) = T ( n
(x0 ), x0 ). Then n+1 is again differentiable into B ρ+μ (I ; Y ).
Proof We define D n+1 ∈ S10 using (3.46) and proceed to show that it is the
μ
derivative of n+1 as a function D n+1 ∈ S1 by a direct estimate

 n+1 (x0 + h) − n+1 (x0 ) − D n+1 (x0 ) · hρ+μ ≤ ε h,

with x0 , x0 + h ∈ X represented in normal coordinate charts.

First, we use the Nemytskii operator technique to get rid of the infinite tail
t → −∞. For any given ε > 0, we have on (−∞, b] the crude estimate
 
sup  n+1 (x0 + h) − n+1 (x0 ) − D n+1 (x0 ) · h (t) e−(ρ+μ) t
t≤b

≤  n+1 (x0 + h) − n+1 (x0 )ρ + D n+1 (x0 ) · hρ e−μ b
 
≤ Lip( n+1 ) + D n+1  h e−μ b ≤ ε h

for some b(ε) that is sufficiently negative. We use the differentiability of T b,a =
TYb,a ◦ (TXa , pr1 ) on the finite interval [b, 0] that is left. We define n+1
b,a = T
b,a ◦

ρa ◦ and estimate
n

 
sup  n+1 (x0 + h) − n+1 (x0 ) − D n+1 (x0 ) · h (t) e−(ρ+μ) t
t∈[b,0]

≤ ρb ◦ n+1 (x0 + h) − n+1

b,a (x 0 + h)ρ+μ
+ ρb ◦ n+1 (x0 ) − n+1
b,a (x 0 )ρ+μ
 
+  ρb ◦ D n+1 (x0 ) − D n+1
b,a (x 0 ) · hρ+μ
+  n+1 n+1 n+1
b,a (x 0 + h) − b,a (x 0 ) − D b,a (x 0 ) · hρ+μ .

This holds for all a ≤ b and the first three terms can be made arbitrarily small when
a → −∞ due to Proposition 3.42 and Corollary 3.43, while the last term is o(h)
since n+1
b,a is differentiable by the chain rule. If the estimate o(h) is independent
of a, then we can finally, for any ε > 0 and h ≤ δ sufficiently small, estimate this
by ε h.
That the term o(h) is independent of a follows from another application of the
mean value theorem:
3.7 Smoothness 135

 n+1 n+1 n+1

b,a (x 0 + h) − b,a (x 0 ) − D b,a (x 0 ) · hρ+μ
≤ D n+1 n+1
b,a (ξ ) − D b,a (x 0 )ρ+μ h
where d(ξ, x0 ) ≤ h
= DT b,a
( (ξ ), ξ ) · ρa ◦ D n (ξ )
n

− DT b,a ( n (x0 ), x0 ) · ρa ◦ D n (x0 )ρ+μ h

  
≤ DT b,a ( n (ξ ), ξ ) − DT b,a ( n (x0 ), x0 )ρ+μ,ρ D n (x0 )ρ
 
+ DT b,a ( n (x0 ), x0 )ρ+μ,ρ+μ D n (ξ ) − D n (x0 )ρ+μ h
≤ ε(d(ξ, x0 )) h

since the continuity estimates for the formal derivatives DT directly translate into

the same estimates for the true derivative counterparts DT b,a on restricted intervals.

Thus, we can now conclude by induction, starting at 0 ≡ 0, that for each n ≥ 0
the map n ∈ S0 is differentiable when viewed as map into B ρ+μ (I ; Y ), while the
results in Sect. 3.7.5 show that we actually have
1,α  
n ∈ Cb,u X ; B ρ+μ (I ; Y ) . (3.73)

Finally, we have uniformly convergent sequences

μ
n → ∞ ∈ S0 and D n → D ∞ ∈ S1 (3.74)


by the fiber contraction theorem, so now we apply Theorem D.2 (taking into
account Remark D.3) to conclude that D ∞ is the derivative of ∞ as a map

X → B ρ+μ (I ; Y ). It was already shown in Proposition 3.37 that D ∞ is bounded

and continuous, just as the D n in (3.73).
Remark 3.46 (on topologies used) The convergence in (3.74) is with respect to uni-
form supremum norms as in Definition 2.9. These induce a topology that is stronger
than the weak Whitney (or compact-open) topology, cf. Sect. 1.7. The convergence
in Theorem D.2 is with respect to the weak Whitney topology, both the assumption
and result. This is sufficient, since we are primarily interested in the result that ∞ is
differentiable, not in what sense n and its derivatives converge to ∞ . On the other
hand, we did already have convergence of D n → D ∞ = D ∞ with respect to

these stronger uniform norms, so clearly n → ∞ in uniform C 1 -norm as well. ♦
136 3 Persistence of Noncompact NHIMs

3.7.8 Conclusion for the First Derivative

The evaluation map ev0 : B ρ+μ (I ; Y ) → Y is bounded linear so the graph (3.41) of
the persistent invariant manifold also satisfies

h̃ = ev0 ◦ ∞ ∈ Cb,u
1,α
(X ; Y ).

The size of Dh̃ can be estimated using the fixed point equation for D ∞ . This yields


q
D ∞ ρ ≤ Dx TX ,
 1−q  0

and the contraction factor q < 1 can be made arbitrarily small by choosing ζ small.
As indicated in (3.18), ζ is in turn controlled by δ, σ1 from Theorem 3.2, and ν from
Lemma 3.12, which can be chosen arbitrarily small. This completes the proof of all
statements in Theorem 3.2 for r = 1 + α with α ∈ [0, 1]. Note that this is the case
k = 1 as in Remark 3.3, 5.

3.7.9 Higher Order Derivatives

To obtain higher order smoothness of the perturbed invariant manifold, we con-

sider Eq. (3.42) for k > 1. The principal term governing the contraction is still
D y T ( (x0 ), x0 ), now acting on multilinear maps Dk (x0 ) ∈ Lk TX ; B kρ+μk
  
(I ; Y ) . The remaining terms only depend on lower order derivatives of (x0 ),
hence they do not influence the contractivity estimate in the fiber contraction theo-
rem. It must be verified, though, that these terms depend continuously on the lower
order derivatives as mappings into B kρ+μk (I ; Y ). Note again that we set μk = α ρ
in case of α-Hölder continuity; in case of uniform continuity (denoted by α = 0) we
choose a sequence {μ j }1≤ j≤k such that the following hold true:
1. μ j < 0 for each j;
2. the spectral gap condition ρY < k ρ + μk < ρ < ρ X still holds;
3. there exists a ρ̃ < ρ such that

k ρ + μk < k ρ̃ and j ρ̃ ≤ j ρ + μ j for any j < k. (3.75)

It follows that the sequence μ j ’s is strictly decreasing (i.e. increasing in absolute

value), and that we have continuous embeddings B k ρ̃ → B kρ+μk and B jρ+μ j →
B j ρ̃ ; in the first embedding we reserved some spectral space to apply Corollary B.3.
These choices—as well as more ideas in this section—are inspired by [Van89,
Sect. 3], which is an interesting read for comparison in a simpler setting.
3.7 Smoothness 137

We reuse the scheme already defined in Sect. 3.7.1 for the first order deriva-
tives (??). Let us walk through these items step by step and indicate the changes that
need to be made.
1. Candidate functions for the higher order derivatives can be found by formal dif-
ferentiation and application of Theorem E.2. This is a straightforward procedure,
although tedious and quite unenlightening to perform. Let us show just one exam-
ple16 :
 
D2y TX (y, x0 ) δy1 , δy2 (t)

 0   
= D
y (t, τ, x y (τ )) · D2y ṽ X (x y (τ ), y(τ )) δy1 (τ ), δy2 (τ )
t
   
+ Dx D y ṽ X (x y (τ ), y(τ )) δy1 (τ ), D y TX (y, x0 )δy2 (τ )

  
2
+ D
y (t, τ, x y (τ )) · D y TX (y, x0 )δy2 (τ )

 τ 
 
+ D
y (t, σ, x y (σ )) · D y Dx ṽ X (x y (σ ), y(σ )) · δy2 (σ ) · D
y (σ, τ, x y (τ ))dσ
t

· D y ṽ X (x y (τ ), y(τ )) · δy1 (τ ) dτ. (3.76)

2. For contractivity in the fibers we still only need to consider the map D y T as

in (3.44), since that is the principal term in (3.42). This map is contractive for any
ρ  ∈ (ρY , ρ X ), hence also for ρ  = k ρ + μ for any μ ≤ 0 sufficiently small, when
ρ ∈ (ρY , ρ X ) is chosen appropriately. The other terms in (3.42) are bounded maps
as well, and linear in the D j (x0 ). It follows from Proposition C.3 that each of

these terms has weighted degree


k−1
i · pi = k − m
i=1

with respect to the D j (x0 ), while they incur an additional exponential factor

em ρ t from taking m derivatives with respect to x0 ∈ X , due to Lemma C.1. Thus
the combined exponential growth rates sum to k ρ, and ρY < k ρ implies that the
variation of constants integrals still converge, so these terms are bounded maps
into B kρ spaces. This still holds if we add μ j ’s that satisfy the conditions set out
above.
In the notation of Appendix C we define spaces of higher order derivatives,
μ   
Sk = b Lk TX ; B kρ+μ (I ; Y ) (3.77)

16 Even though it is not obvious from (3.76), this expression is in fact symmetric in δy1 , δy2 . To
verify this for the termscontaining Dx D yṽ X , one should change the order of integration of τ, σ and
expand the expression D y TX (y, x0 )δy2 (τ ) using (3.43b).

138 3 Persistence of Noncompact NHIMs

with norms
Dk  = sup Dk (x0 )Lk (Tx
0 X ;B
kρ+μ (I ;Y ))
 x0 ∈X 

extending (3.45). Similarly, we define as extensions of (3.46), higher order fiber

mappings
μ μ
F (k) = (T, D1 T, . . . , Dk T ) on S0 × S1 1 × · · · × Sk k . (3.78)
 

μ
These are again uniform fiber contractions with respect to the final factor Sk k as
fiber, for any choice of μ j ≤ 0 sufficiently small.
3. Instead of trying to construct higher order formal tangent bundles, we represent
the higher derivatives on ‘formal tensor bundles’

ρ

TBβ (I ; X )k = B ρ (I ; x ∗ (TX ))⊗k . (3.79)
 ρ
x∈Bβ (I ;X )

Note that this choice of representation along base curves x matches our choice
to represent higher derivatives as in Definition C.6 when the former is evaluated
at a fixed t. The trivializations, then, are defined by tensor products of parallel
transport terms (x|t0 )⊗k , again when restricted to curves x ∈ C 1 . The resulting
holonomy terms can be estimated by either the kþpower of the single holonomy
term, or, k − 1 factors can be bounded by (γ ) ≤ 2 such that the remaining
factor fulfills the required α-Hölder estimate. Then, all details in Sect. 3.7.5 can
be repeated to obtain uniform or α-Hölder continuity of the higher derivatives
of TX , TY as maps on these formal tensor bundles. Note that we can break each
expression into parts such that only one factor is varied for the continuity estimate
and thus only once adds either α ρ or μ to the exponential growth rate. Thus, the
spectral gap condition is still satisfied.
μ μk−1
4. We apply the fiber contraction theorem to (3.78) with base S0 ×S1 1 ×· · ·×Sk−1
μk
and fiber Sk . In case of α = 0, we again seize some of the unused spectral space
for the carefully chosen μ j ’s, such that condition 3 of Theorem D.1 holds. Let
us assume by induction that F (k−1) already is a globally attractive fiber map. The
μ
conditions (3.75) imply that if we insert elements D j ∈ S j j into F (k) , then their
 μ
exponents sum at most to k ρ̃, so the mapping onto the fiber Sk k is continuous
by application of Corollary B.3. For α-Hölder continuity we can simply choose
μ j = α ρ for all 1 ≤ j ≤ k. Thus, we find a globally attractive fixed point

μ μ
( ∞ , D ∞ , . . . , Dk ∞ ) ∈ S0 × S1 1 × · × Sk k with Dk ∞ ∈ Cb,u
α
.
  

ρ
5. We constructed a manifold structure on B ρ (J ; X ) (and a trivial one on Bη (J ; Y ) as
well) with an atlas of charts induced by normal coordinate charts of the underlying
3.7 Smoothness 139

manifold X . We represent higher17 derivatives in these induced normal coordinate

charts B ρ (J ; x ∗ (TX )). Thus, we have for example
 
Dkx TXa (x, y) ∈ Lk B ρ (J ; x ∗ (TX )); B ρ (J ; Y ) .

This precisely matches the representation of the formal higher derivatives on the
tensor space B ρ (I ; x ∗ (TX ))⊗k in point 3. Higher differentiability of the restricted
maps TXa and TYb,a follows as in Sect. 3.7.7.
6. Lemma 3.45 can be generalized to prove by induction over n that higher derivatives
Dk n exist; we define Dk n+1 ∈ Sk0 by (3.78).
7. By induction we may assume that it was already proven that
k−1  
n → ∞ ∈ Cb,u X ; B (k−1)ρ+μk−1 (I ; Y ) as n → ∞.

We apply Corollary D.4

 to conclude that n → ∞ as sequence of C k functions
∞
and that ∈ Cb,u X ; B
k kρ+μk (I ; Y ) . The convergence is with respect to our
uniform supremum norms, see Remark 3.46.
Just as in Sect. 3.7.8, this finalizes the proof of all statements in Theorem 3.2, but
k,α
now for r = k + α with k > 1 and α ∈ [0, 1]. It follows that h̃ ∈ Cb,u , but the
last part that remains to be shown, though, is that h̃k−1 can be made as small as
desired.
From the fixed point Eq. (3.42) it follows that

1   • ∞ 
Dk−1 ∞ (x0 ) ≤ Dly Dm ∞
x0 T ( (x 0 ), x 0 ) · Pl,k−m D (x 0 ) .
1−q
l,m≥0
l+m≤k−1
(l,m) =(0,0),(1,0)

Each term with l ≥ 1 contains at least one factor D j ∞ (x0 ) with j < k −1; these can
be assumed to be small by induction. The one remaining term with (l, m) = (0, k −1)
can be expanded using Proposition C.3. This yields, suppressing arguments ∞ (x0 )
and x0 ,

k−1
Dx TY · P j,k−1 (D•x0 TX ).
j
x0 T =
Dk−1
j=1

Since all terms are uniformly bounded in appropriate norms, it suffices to show
j
that the Dx TY can be made small. Recall formula (3.43c) and the fiber contraction
estimate for Dx TY in Sect. 3.7.3, where we saw that Dx TY could be made small
by choosing  f˜, Dx f˜ ≤ ζ small. The higher derivatives Dx TY , too, contain
j

17 It would probably be more natural to consider the higher derivatives as maps into Banach
manifolds with exponents k ρ, but these norms are equivalent anyways.
140 3 Persistence of Noncompact NHIMs

a factor Dix f˜ with 0 ≤ i ≤ j in each term, so by Proposition 3.14 these can be made
small. Hence, Dk−1 ∞ and consequently h̃k−1 can be made uniformly small.
Note that h̃k cannot be made small though, see Remark 3.15.
Chapter 4
Extension of Results

In this chapter we discuss some ways to extend the main result of Theorem 3.1 to
slightly more general situations. These extensions are known from the compact and
Euclidean settings, but a bit scattered over the literature. We try to collect a number
of these results here, while extending them to our noncompact setting.

4.1 Non-Autonomous Systems

We proved the main theorem for an autonomous system and perturbation. The Perron
method admits without difficulty a time-dependent formulation; we refrained from
including this, since it would only have cluttered the already detailed proof, while
time-dependence is easily added as an afterthought, as already noted in Sect. 1.6.1.
Let us assume that M is an r -NHIM for the (time-independent) vector field v
on (Q, g) and that all assumptions of Theorem 3.1 are fulfilled. We can allow time-
dependent perturbations by the standard trick to extend the phase space of the system
by R  t. Define
Q̂ = R × Q with metric ĝ = dt 2 + g. (4.1)

Then ( Q̂, ĝ) is again of bounded geometry. We trivially extend the vector field v to
 
v̂(t, x) = 1, v(x) ∈ T(t,x) Q̂ (4.2)

and set M̂ = R× M ⊂ Q̂. Then the flow  ˆ of v̂ has the same hyperbolicity properties
as  since the additional flow along t˙ = 1 is completely neutral and decoupled
from the original system. It follows that M̂ is again an r -NHIM for the dynamical
ˆ R). Note that we need a theory for noncompact NHIMs to perform
system ( Q̂, ,
this extension by the time interval R. Now we can choose a perturbed vector ṽ that
k,α
depends explicitly on time, as long as ṽ ∈ Cb,u ( Q̂) is close to v̂. This means that
the perturbation must be small in C k,α -norm (including derivatives with respect to

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 141

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4_4,
© Atlantis Press and the author 2013
142 4 Extension of Results

time), uniformly for all time. As a result we find that the perturbed manifold M̃ will
depend on time, i.e. it is not exactly of the form M̃ = R × M for some M ⊂ Q. We
do find that M̃ is uniformly close to M̂ = R × M, however, so M̃ is approximately
of this product form.
Remark 4.1 A direct application of Theorem 3.1 requires the perturbed vector field
to be C k,α with respect to time, too, since t ∈ R is added to the phase space variables.
Note that the result thus depends C k,α smoothly on time as well. A closer inspection of
k,α
the proof shows that this can in fact be replaced by the condition that v̂(t, · ) ∈ Cb,u ,
uniformly in t ∈ R, just as in Remark A.7. In that case the resulting manifold M̃
cannot be expected to be differentiable with respect to time anymore, but it still
satisfies all uniform C k,α smoothness and boundedness properties with respect to
x ∈ Q. In particular, M̃ is still uniformly close to M̂, uniformly for all t ∈ R. ♦
Instead of starting with an autonomous system v, we can also take an initial non-
autonomous system v̂ and perturb that. As long as v̂ truly describes a non-autonomous
system, that is, it is defined on a space R × Q and has component 1 along R, then
normal hyperbolicity is easily tested. The R-component of the flow is trivially neutral,
while the other Q-component must be checked in a context where, for example, also
the invariant splitting (1.9) may depend on time, but this introduces no fundamental
changes.

4.2 Smooth Parameter Dependence

Another interesting question for applications is if the persistent manifold depends

smoothly on the perturbation parameter. This result can be obtained in a similar way
as time-dependence, now adding a parameter p ∈ P to the phase space with trivial
dynamics ṗ = 0. The noncompact theory is not essential here, but it does allow for
a simple proof.
Let again (Q, g) and v = v( p, x) describe the system, where p ∈ P denotes
the parameter. For simplicity we assume that P = Rn and that p = 0 corresponds
to the unperturbed system for which we have M as r -NHIM. We consider again an
extended system Q̂ = P × Q and M̂ = P × M. The extended vector field we choose
slightly differently: we use an external scaling parameter α ≥ 0 to slowly ‘turn on’
the parameter dependence. Let χ ∈ C ∞ (R≥0 ; [0, 1]) be a radial cut-off function
such that χ (r ) = 1 for r ≤ 1 and χ (r ) = 0 for r ≥ 2, and define
  
v̂α ( p, x) = 0, v χ ( p ) α p, x (4.3)

as a vector field
 on Q̂. Note
 that M̂ is an r -NHIM for v̂0 by trivial extension. One
can verify that v̂α − v̂0 r can be chosen small with α. Uniformity with respect to p
follows automatically from χ having compact support. As a result of Theorem 3.1
we conclude that there exists an α > 0 such that v̂α has a C r family of invariant
4.2 Smooth Parameter Dependence 143

manifolds 
M̃ = M̃ p
, (4.4)
p
∈P

where M̃ p
is the invariant manifold corresponding to the vector field v( p, · ) with
p = χ ( p ) α p
. This parametrizes a full neighborhood B(0; α) ⊂ P.

4.3 Overflowing Invariant Manifolds

Overflowing invariance is a useful tool to study invariant manifolds whose normal

hyperbolicity properties break down beyond a certain domain, see also Sect. 1.6.3.
We shall indicate here how our main result can be extended to overflowing invariant
manifolds. We provide conditions for persistence that are slightly weaker than those
in the literature. These might prove useful for some applications.
The following definition extends that in [Fen72] and is equivalent to Definition 2.1
in [BLZ99].
Definition 4.2 (Overflowing invariant manifold) Let (Q, g) be a Riemannian
manifold, M ⊂ Q a C 1 submanifold with boundary ∂ M ∈ C 1 , and v ∈ C 1 a
vector field on Q with flow . Let n denote the outward normal at ∂ M. Then M is
called overflowing invariant under v if the following hold:
1. backward orbits stay in M, i.e. ∀ m ∈ M, t < 0 : t (m) ∈ M;
2. the vector field v points uniformly strictly outward at ∂ M, i.e. there exists some
ε > 0 such that ∀ m ∈ ∂ M : gm (v, n) ≥ ε.
Definition 1.8 of normal hyperbolicity can be adapted to this setting (only condi-
tion (1) is necessary): we assume that only stable normal directions are present and
we only require M to be negatively invariant, while the exponential rate conditions
must hold along orbits as long as they stay inside M.
Remark 4.3 Note that the uniformity in condition (2) reduces to the standard ‘strictly
outward’ if ( M̄) = M ∪ ∂ M is compact. This is the natural generalization for
noncompact manifolds, since the condition is used to guarantee that under small
perturbations and in a small tubular neighborhood the vector field is still pointing
outward. ♦
The Perron method uses orbits as fundamental objects and constructs a contraction
operator on these. The essence of Definition 4.2 is to guarantee condition (1) that
backward orbits stay inside ( M̄), even under a small perturbation of the vector field.
This provides an idea to slightly weaken the overflow invariance definition into an a
priori argument. If any orbits considered in the Perron method proof stay inside M̄,
then all assumptions throughout the proof are still valid and we obtain a persistent
manifold M̃. To make this idea explicit, we choose the trivial bundle setting of
Theorem 3.2 and introduce the following weakened definition.
144 4 Extension of Results

Definition 4.4 (A priori overflowing invariance) Let (X, g) be a Riemannian

manifold and let M ⊂ X an open submanifold, i.e. of the same dimension, with
boundary ∂ M ∈ C 1 . Let Y be a Banach space, and v ∈ C 1 a vector field on X × Y
with flow . Let n denote the outward normal at ∂ M. Let ṽ be a perturbation of v.
Then M is called a priori overflowing invariant for the pair (v, ṽ) if the following
hold:
1. backward orbits of v stay in M, i.e. ∀ m ∈ M, t < 0 : t (m) ∈ M;
2. the vector field ṽ points (non-strictly) outward at a tubular neighborhood over
∂ M, i.e. there exists some η > 0 such that
 
∀ (m, y) ∈ ∂ M × Y≤η : g Dπ X · v(m, y), n(m) ≥ 0.

Remark 4.5 Note that Definition 4.2 implies 4.4 when ṽ − v 1 is small enough and
ṽ ∈ Cb,u
1 .

Remark 4.6 A useful generalization of Definition 4.4 to the setting of Theorem 3.1
is less trivial. There we do not have canonical vertical fibers over ∂ M in the tubular
neighborhood, nor the associated projection of v onto TX at ∂ M. We cannot simply
take a non-vertical fiber; the Perron method adapts the curves x and y separately, so
it may happen that while x(0) ∈ ∂ M is kept fixed, y(0) is updated to a new value
such that (x(0), y(0)) lies outside of the tubular neighborhood over M̄, and control
is lost. ♦
Let us demonstrate the application of this more general definition with the fol-
lowing simple example, see also Fig. 4.1.
Example 4.7 (Persistence under a priori overflowing invariance) Let X × Y =
R × R and let the unperturbed vector field be given by
 
v(x, y) = − (x − 1)2 , (x 2 − 4) y .

M
X
0
−1 1 2

Fig. 4.1 A priori overflowing invariance for the manifold M

4.3 Overflowing Invariant Manifolds 145

Note that M = (−1, 1) ⊂ X is strictly overflowing invariant at its left boundary

x = −1 (we could choose other values as well), but non-strictly so at the right
boundary x = 1, which is a degenerate stationary point. The vector field is normally
attracting over the interval (−2, 2) and uniformly so over any closed subinterval. Note
that there does not exist a subinterval of X that is overflowing invariant according to
Definition 4.2.
Let us choose a family vδ of perturbations of v such that vδ − v 1 ≤ δ and v = vδ
on a neighborhood of (1, 0) ∈ X × Y . Then M satisfies Definition 4.4 for this family
vδ and application of Theorem 4.8 below shows that for δ sufficiently small, there
exists a unique negatively invariant manifold M̃ = Graph(h̃) for the flow of vδ such
that h̃ : [−1, 1] ⊂ X → [−η, η] ⊂ Y . For any r ≥ 1 there exists a δ such that
h̃ ∈ C r holds. 
Theorem 4.8 (Persistence under overflowing invariance) Let k ≥ 2, α ∈ [0, 1]
and r = k + α. Let (X, g) be a smooth, complete, connected Riemannian manifold of
k,α
bounded geometry and Y a Banach space. Let vδ ∈ Cb,u be a family of vector fields
defined on a uniformly sized neighborhood of the zero-section in X × Y such that
vδ − v0 1 ≤ δ. Let M satisfy Definition 4.4 for the pair (v0 , vδ ) for any δ ∈ (0, δ0 ]
and let M be r -normally attracting for the flow defined by v0 , that is, M satisfies the
overflowing invariant version of Definition 1.11 with rank(E + ) = 0.
Then for each sufficiently small η > 0 there exist δ1 > 0 such that for any
δ ∈ (0, δ1 ], there is a unique manifold with boundary M̃ = Graph(h̃), h̃ : M → Y ,
h̃ 0 ≤ η such that M̃ is negatively invariant under the flow defined by vδ . More-
k,α
over, h̃ ∈ Cb,u and h̃ k−1 can be made arbitrary small by choosing vδ − v0 k−1
sufficiently small. The function h extends continuously to ∂ M.
Remark 4.9 In this overflowing invariance setting, the condition that rank(E + ) = 0
is really necessary and not an artifact of our proof. The same results hold for inflowing
invariance with no stable normal directions present. Definition 4.4 can be extended
to full normal hyperbolicity with both stable and unstable normal directions present.
This requires full invariance of a tubular neighborhood of M under both the forward
and backward orbits.

Remark 4.10 We can restrict to a smaller open subset U of X that contains M̄, so
k,α
we do not need vδ ∈ Cb,u to hold on all of X . If this subset U is not convex, though,
we may run into difficulties when applying the mean value theorem, see Fig. 4.2:
an intermediate point ξ ∈ M̄ on the line between x1 , x2 may be selected, so we need

Fig. 4.2 A nonconvex subset

M⊂X
x1
M
x2
146 4 Extension of Results

to make sure that the uniform estimates still hold there. Thus the need for U ⊃ M̄
to be convex, see also the remark in [Hen81, p. 289]. ♦
Proof The proof of Theorem 3.2 requires minimal changes. Note that regardless
of the modifications and smoothing preparations performed in Sect. 3.4, the vector
field ṽ X is precisely the horizontal component of the perturbed vector field vδ . In
Sect. 3.6 where we proved existence and uniqueness of M̃, we take η small enough
 4.4. This guaranteesρ that x = TX (y, x0 )
that it satisfies condition (2) of Definition
is a solution curve such that x (−∞, 0] ⊂ M̄ for any y ∈ Bη (I ; Y ) and x0 ∈ M̄.
Hence, the contraction mapping T = TY ◦(TX , pr1 ) is well-defined with intermediate
space Bβ (I ; M̄) and we find a unique Lipschitz continuous fixed point map ∞ :
ρ
M̄ → Bη (I ; Y ).
No essential changes are needed with respect to the smoothness proof in Sect. 3.7.
The formal derivatives (3.43) are well-defined along all curves x and y that are
considered, since the derivatives of ṽ X , A, f are defined on an open neighborhood
of M̄. In Sect. 3.7.7 we use the mean value theorem to prove that the restricted
maps T b,a have true derivatives. Remark 4.10 is not problematic here, since T b,a is
defined on the finite interval J = [a, 0] and thus we can restrict to arbitrarily small
open neighborhoods along the curves x, y when restricted to J . Hence we find that
∞ ∈ Cb,u k,α
on M. 

4.4 Full Normal Hyperbolicity

We made the assumption in our main theorems that the unstable bundle E + was
absent, that is, that M was a normally attracting invariant manifold. As already noted
in Remark 3.3, 8, it should be possible to generalize this to the case of full normal
hyperbolicity where both stable and unstable normal directions are present. Let us
indicate here how this more general result can be obtained.
Assume that in Theorem 3.1 we have an invariant splitting (1.9) with both stable
and unstable bundles E ± present. The reduction principle in Sect. 2.6 leads to a
formulation of Theorem 3.2 with a trivial bundle
 
π : X × Y × Z → X, (4.5)

where X is a smoothed approximation of M and approximate, boundedly smooth

representations of the bundles E ± are embedded into the trivial bundles X ×Y , X × Z
with Y, Z Banach spaces. This means that M is again represented as the graph of an
approximate zero section h σ : X → Y × Z ; now, the subbundles X × Y and X × Z
are approximately invariant under vσ . The deviation from invariance is controlled
by σ , the parameter of the smoothing approximation of M. We find linear operators
A± (x) on Y and Z respectively, that approximate the linearizations of vY and v Z , and
4.4 Full Normal Hyperbolicity 147

corresponding flows  ± with approximate growth rates. We add a map1

 ∞  
TZ (x, y, z)(t) = x+ (t, τ ) f˜+ x(τ ), y(τ ), z(τ ) dτ (4.6)
t

ρ
with z ∈ Bη (R≥0 ; Z ) and adapt the other maps to incorporate z as an argument. We
use Lemma 3.30 and extend all curves in X, Y, Z to the full real line. This should
yield a contraction
 
T = (TY , TZ ) ◦ TX , pr1 , pr2 on Bηρ (R; Y ) × Bηρ (R; Z ), (4.7)

again with x0 ∈ X as initial value parameter. We obtain a pair ( − , + ) of fixed

point maps, and after evaluation at t = 0 we find

(h̃ − , h̃ + ) : X → Y × Z , (4.8)

whose graph describes the persistent invariant manifold M̃. See e.g. [VG87] for an
application of this technique in the center manifold setting.
An alternative method to obtain M̃ is by constructing it as the intersection of
its (center-)stable and (center-)unstable manifolds. Here we make use of over- and
+
inflowing invariance. First we construct a tubular neighborhood P≤η of M along
+
(a smooth approximation of) the unstable bundle E , see Fig. 4.3. This manifold
is approximately invariant and overflowing according to Definition 4.2 if η is suffi-
ciently small since it lies approximately along the unstable direction of M. Careful
modifications to the setup of the proof in Sects. 3.4 and 3.5 will allow us recover the
local unstable manifold Wloc U of M̃ as the graph of

+
h̃ U : P≤η → ē− ,

Fig. 4.3 An overflowing

+
E−
tubular neighborhood P≤η
along the unstable manifold
E+

M
η

P ≤+ η

1 Note that since t ≤ τ , we have a reverse flow  + (t, τ ) for the unstable directions, which indeed
satisfies the growth estimates (1.10).
148 4 Extension of Results

+
where ē− is an extension of the bundle E − over P≤η . Vice versa we can find the
S −
local stable bundle Wloc as a graph over P≤η . Their intersection yields

k,α
M̃ = Wloc
U
∩ Wloc
S
∈ Cb,u

since the intersection is uniformly transversal. This is a standard trick which is also
applied in [Fen72; HPS77] using the graph transform.

4.5 Recovery of the Invariant Fibration and Splitting

In our main Theorem 3.1 we did not prove that the persistent manifold is again
normally hyperbolic. As stated in Remark 3.3, 9, we must recover the invariant
splitting of the perturbed manifold M̃ to be able to conclude that M̃ again satisfies
Definition 1.8 of normal hyperbolicity. We will sketch how to find the invariant fibers
of the stable manifold of M̃ in the setting of Theorem 3.2 with a trivial bundle Q =
X × Y . These fibers are easily recovered using the Perron method, and smoothness
of each single fiber essentially comes for free. We then readily obtain the vertical
part Ẽ − of the invariant splitting as the tangent planes to these fibers at M̃. Since
T M̃ is already invariant, we have

T M̃ Q = T M̃ ⊕ Ẽ − (4.9)

which is invariant under the tangent flow, cf. (1.9). Note that we should still prove
boundedness and uniform continuity of this splitting.
Each single invariant fiber can be found by application of the non-autonomous
Perron method. Let ϒ̃ : R × M̃ → M̃ denote the flow of the perturbed system,
restricted to the persistent manifold, and let γ (t) = ϒ̃ t (m) denote a solution curve
in M̃. Consider the pullback bundle
π
Em = γ ∗ (TQ) −→ R. (4.10)

Note that γ ∗ (T M̃) is a subbundle of Em in a natural way, and that it is invariant under
the tangent flow Dϒ̃ t . Next we consider γ ∗ (E − ) as a subbundle of Em : even though
E − was defined over the original manifold M, it can simply be translated along the
canonical fiber Y to be identified with a bundle over M̃ and then be pulled back
along γ . Thus we have a splitting

TM̃ Q = T M̃ (X × Y ) = T M̃ ⊕ E − .

The original splitting TM ⊕ E − was bounded, hence when M̃ is sufficiently close to

M in C 1 norm, then this splitting will be bounded again. It is uniformly continuous
since both subbundles are. In order to recover a smooth stable fibration later, we will
4.5 Recovery of the Invariant Fibration and Splitting 149

Y S
W loc (γ (t ))

Eγ−(t )
Tγ (t ) M̃
M̃
X

S ( M̃) modeled on the bundle E = γ ∗ (T M̃ ⊕ Ê − )

Fig. 4.4 The invariant fibers of Wloc m

moreover want to construct a smooth approximation Ê − , similar to the convolution

S ( M̃) of the perturbed system
smoothing of  in Sect. 3.4. The stable manifold Wloc
consists of single fibers Wloc (m) that can be modeled as graphs of maps
S

S
h̃ m : Ê m− → Tm M̃,

S after application of the Perron method to the

see Fig. 4.4. We recover the map h̃ m
non-autonomous system defined on Em by pulling back the vector field ( ∂t∂ , ṽ) along2

ϕ : γ ∗ (TQ) → R × Q : (t, x) → expγ (t) (x) with x ∈ Tγ (t) Q (4.11)

That is, we look at the system in normal coordinates that follow the solution curve
γ . This means that the vector field ϕ ∗ ( ∂t∂ , ṽ) on Em has constant speed 1 on the base
R, i.e. the time axis, and preserves the zero section, i.e. the origin of the coordinates
at γ (t). Since D expγ (t) (0) = 1, the exponential growth conditions are preserved
under pullback. Hence, the splitting

Em = γ ∗ (T M̃) ⊕ γ ∗ ( Ê − )

is approximately invariant with growth rates close to ρ X and ρY respectively. Now

we apply the Perron method on the interval R≥0 . That is, we have the operator

2 We only can (and need to) perform the pullback in an η-sized tubular neighborhood of γ . Outside
this neighborhood we smoothly cut off the vector field to a suitable linearization. Therefore we only
recover the local stable manifold.
150 4 Extension of Results
 ∞
T (x, y, y0 )(t) = Dϒ̃ t−τ (γ (τ )) f X (τ, x(τ ), y(τ ))dτ, (4.12)
t
 t 
γ (t, 0) y0 + γ (t, τ ) f Y (τ, x(τ ), y(τ ))dτ
0

which is a contraction on pairs of curves (x, y) ∈ B ρ (R≥0 ; Em ) with the partial initial
condition y0 ∈ Em,0 = Ê m− as parameter. Here we chose ρ X < ρ < ρ X and used the
fact that f X ,Y (t, 0, 0) = 0. The functions f X ,Y are C 1 -small nonlinear terms. These
contain the linear and higher nonlinear perturbations away from the linear flows
Dϒ and γ along γ ∗ (T M̃) and γ ∗ ( Ê − ) respectively, corrected for the bounded
coordinate changes introduced by ϕ in (4.11). Thus, we find h̃ m S by evaluating the

x-component of the fixed point map at t = 0 for a given y0 .

We can establish that h̃ m S ∈ C k,α by an implicit function argument3 on the Banach
b,u
ρ
space B (R≥0 ; Em ), see [Irw70]. The unpublished preprint [Dui76] extends these
ideas to the geometric context of Banach bundles over R.
Note that this result is uniform in the family parameter m ∈ M̃, but it does not
include smoothness with respect to m. To obtain smoothness, we extend the pullback
bundle (4.10) to  π
E= Em = ϒ ∗ (TQ) −→ R × M̃. (4.13)
m∈ M̃

This bundle can be endowed with a topology by parallel transport along the curves γ
as in Sect. 3.7.4. Then each derivative with respect to m adds an exponential growth
factor eρ M̃ t since the curves γ can diverge at that rate in forward time. Note that
ρ M̃ corresponds to ρ M in (1.11), that is, the upper bound on the spectrum of the
tangential flow. This leads to the spectral gap condition

ρY < ρ X − r ρ M̃ ≈ (r + 1)ρ X , (4.14)

from which it follows that the mapping m → h̃ m S has C k−1,α smoothness. This
b,u
implies that the invariant splitting is at least uniformly continuous and bounded. The
exponential growth rates with respect to the splitting follow from small perturbation
estimates similar to those in Sects. 3.4 and 3.5. This completes the requirements for
normal hyperbolicity of M̃ in Definition 1.8.

3 We immediately recover higher smoothness instead of having to go through an elaborate scheme

involving the fiber contraction theorem as in Sect. 3.7. The reason is that B kρ (R≥0 ; Em ) →
B ρ (R≥0 ; Em ) is a continuous embedding when ρ < 0 and we look at R≥0 , that is, higher powers
of negative exponential growth decay even stronger.
Appendix A
Explicit Estimates in the Implicit Function
Theorem

In this appendix, we carefully examine the implicit function theorem. We extend this
standard theorem to classes of functions with additional properties such as bound-
edness and uniform and Hölder continuity. The crucial ingredient is the explicit
formula (A.2) for the derivative of the implicit function, which allows us to transfer
regularity conditions onto the implicit function.
As an application of the implicit function theorem in Banach spaces, we will
establish existence, uniqueness and smooth dependence on parameters for the flow
of a system of ordinary differential equations. Essentially, these are standard results
from differential calculus, see e.g. Zeidler [Zei86, p. 150,165] or [Irw72, Rob68]. We
consider a general setting of ODEs in Banach spaces and show smooth dependence,
both on the initial data, as well as on the vector field itself. Moreover, our extension
of the implicit function theorem yields boundedness and uniform continuity results.
We start with some results on inversion of linear maps.
Lemma A.1 (Invertibility of linear maps) Let X be a Banach space and let A ∈
L(X ) be a continuous linear operator with continuous inverse. Let B ∈ L(X ) be
another linear operator such that B <  A1−1  . Then A + B is also a continuous
linear operator with continuous inverse, given by the absolutely convergent series
 −1  n   n
A+B = − A−1 B A−1 = A−1 − B A−1 . (A.1)
n≥0 n≥0

 
Proof First of all, note that there exists an M ≥ 1 such that A, A−1  ≤ M. The
base of the geometric series can be estimated in operator norm as −A−1 B  < 1, so
the series is absolutely convergent and the limit is a well-defined continuous linear
operator, whose operator norm can be estimated as
    n  −1 
      A 
(A + B)−1  ≤ A−1  −A−1 B  ≤   < ∞.
n≥0
1 −  A−1 B 

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 151

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
152 Appendix A: Explicit Estimates in the Implicit Function Theorem

That the limit is again a well-defined linear operator follow from the fact that L(X )
is a Banach space.
Applying A + B to the left-hand side of (A.1), we see that the candidate is a right
inverse:
  n   n   n+1
(A + B) A−1 − B A−1 = − B A−1 − − B A−1 = 1.
n≥0 n≥0 n≥0

Similarly the candidate can be shown to be a left inverse of A + B. Now we have

that the candidate is continuous and a full inverse and furthermore, A + B itself is
clearly a continuous operator as the sum of two continuous operators, so the proof
is completed. 
Corollary A.2 (Linear inversion is analytic) Let I : A → A−1 be the inversion
map defined on continuous, linear mappings A ∈ L(X ) with continuous inverse,
where X is a Banach
 space. The map I is analytic with radius of convergence
ρ(A) ≥ 1/ A−1 . When X is finite-dimensional, I is a fortiori a rational map.
Proof Extending well-known results on analytic functions to Banach spaces (see
e.g. [Muj86]), we read off from (A.1) that the inversion map
 I can
 be given around A
by an absolutely convergent power series with ρ(A) ≥ 1/ A−1  and is thus analytic.
When X is finite-dimensional, det(A) = 0 implies that A−1 is a rational expression
in the matrix coefficients of A according to Cramer’s rule. 
The inversion map I is locally Lipschitz, like every C 1 mapping:

   n    −1 2
      A 
(A + B)−1 − A−1  ≤ −A−1 B  A−1  ≤   B.
n≥1
1 −  A−1 B 

However, when we  restrict

 to a domain bounded away from non-invertible operators
A, that is, when  A−1  ≤ M, then the Lipschitz constant is bounded for small B.
This implies that when A = A(x) depends on a parameter via a certain continuity
modulus, then A(x)−1 will have the same continuity modulus up to the Lipschitz
constant, at least in small enough neighborhoods.
The standard implicit function theorem on Banach spaces can be stated as
Theorem A.3 (Implicit function theorem) Let X be a Banach space, Y a normed
linear space, and let f ∈ C k≥1 (X × Y ; X ). Let (x0 , y0 ) ∈ X × Y and assume
that f (x0 , y0 ) = 0 and that D1 f (x0 , y0 )−1 ∈ L(X ) exists as a continuous, linear
operator.
Then there exist neighborhoods U ⊂ X of x0 and V ⊂ Y of y0 , and a unique
function g : V → U such that f (g(y), y) = 0. Furthermore, the map g is C k and
the derivative of g is given by the formula

Dg(y) = −D1 f (g(y), y)−1 · D2 f (g(y), y). (A.2)

Appendix A: Explicit Estimates in the Implicit Function Theorem 153

See [Zei86, pp. 150–155] for a proof. Note that we do not need to assume that Y is
a complete space, as the contraction theorem is only applied on X . Recall that we
use notation where D denotes a total derivative, while Di with index i ∈ N denotes
a partial derivative with respect to the i-th argument.
Formula (A.2) for the derivative of the implicit function g will be crucial for the
extension of the implicit function theorem to many classes of regularity, extending
C k smoothness. We use the Lipschitz estimate for the inversion map and require that
the regularity conditions are preserved under composition, addition, multiplication
and localization of functions. By Proposition C.3, the derivatives of g are expressed
in terms of D1 f (g(y), y)−1 acting on a polynomial expression of same or lower
order derivatives of f and strictly lower order derivatives of g.
As an example, let us take Cbk,α functions. Using Lemma 1.19 and induction over
k, this function class is preserved under products. For composition, we check Hölder
continuity,
 α
 f (g(x2 )) − f (g(x1 )) ≤ C f C g x2 − x1 α ≤ (C f C gα ) x2 − x1 α

for 0 < α ≤ 1, when x2 − x1  ≤ 1. In case x2 − x1  > 1 however, we can

directly use the boundedness of f :

 f (g(x2 )) − f (g(x1 )) ≤  f (g(x2 )) +  f (g(x1 )) ≤ 2  f 0 x2 − x1 α .

Thus, Hölder continuity is preserved with some new Hölder constant, while bound-
edness is trivially preserved as well. We conclude that if f ∈ Cbk,α , and (D1 f )−1 is
globally bounded, then we can read off from formula (A.2) that g ∈ Cbk,α . The same
k functions, or any other class of functions whose
results hold for the class of Cb,u
properties are preserved when inserted into (A.2). Together, interpreting α = 0 as
an empty condition, these lead to
k,α
Corollary A.4 Let in the Implicit Function Theorem A.3, f ∈ Cb,u with k ≥ 1 and
 

0 ≤ α ≤ 1. Assume moreover that D1 f (x, y) −1  ≤ M is bounded on U × V
k,α
for some constant M < ∞. Then g ∈ Cb,u , and the boundedness and continuity
estimates depend in an explicit way on those of f .
Remark A.5 Formula (A.2) only provides control on the derivatives of the implicit
function, but the size of g itself can be controlled by choice of the neighborhood U .
In our applications, this will match up with choosing coordinate charts around the
origin in Rn . ♦
Let us now consider an ordinary differential equation

ẋ = f (t, x), x(t0 ) = x0 , (A.3)

k,α
where x takes values in a Banach space B and f ∈ Cb,u (R × B; B) with k ≥
1, 0 ≤ α ≤ 1. We consider solutions x ∈ X = C (I ; B) equipped with the
0
154 Appendix A: Explicit Estimates in the Implicit Function Theorem

supremum norm, which turns X into a Banach space.1 We choose I to be a closed

interval I = [a, b] ⊂ R. The Picard integral operator
 t
T : X → X : x(t) → F(x)(t) = x0 + f (τ, x(τ )) dτ (A.4)
t0

has exactly the solution curves of (A.3) as fixed points. It also implicitly depends on
k,α
f ∈ Cb,u (R × B; B) and (t0 , x0 ) ∈ I × B. From now on we denote by Dx a partial
derivative with respect to the argument that is typically described by the variable x.
This T is a contraction for |I | = b − a small enough:
 t 
 

T (x1 ) − T (x2 ) = sup  f (τ, x1 (τ )) − f (τ, x2 (τ )) dτ 

t∈I t0
 t
≤ sup Dx f (τ, ξ(τ ))x1 (τ ) − x2 (τ ) dτ
t∈I t0
≤ sup |t − t0 |Dx f x1 − x2 
t∈I
≤ |I | Dx f  x1 − x2 .

We restrict T to a bounded subset of argument functions f ,

k,α
F ⊂ Cb,u (R × B; B), sup  f k,α ≤ R.
f ∈F

Thus, choosing |I | ≤ 2R1

turns T into a q = 21 contraction, which shows that there
is a unique x ∈ X satisfying T (x) = x and therefore (A.3).
Next, we consider small perturbations of both (t0 , x0 ) and f . To apply the implicit
function theorem, we define F(x) = x − T (x). This function has a unique zero and
DF(x) is invertible, as

F(x + δx)(t) − F(x)(t)

 t
 
= δx(t) − Dx f τ, ξ(τ ) · δx(τ ) dτ
t0


t   
= δx(t) − Dx f τ, x(τ ) · δx(τ ) + O ξ(τ ) − x(τ ) δx(τ ) dτ
t0
   
= DF(x) · δx (t) + o δx (A.5)
 
The neglected terms are o δx since Dx f is uniformly continuous on I , so DF(x)
exists. From the expression above, we can also easily read off continuity of DF(x)
as a linear operator, by writing DF(x) = 1 + A(x) and noticing that

1 Note that any actual solution x will be C 1 at least, but only x ∈ C 0 is required. This makes X a
complete space without the need to introduce norms more complicated than the supremum norm.
Appendix A: Explicit Estimates in the Implicit Function Theorem 155

A(x) ≤ |I | Dx f  < 21 ,

 
thus DF(x) is a bounded, invertible linear operator such that DF(x)−1  ≤ 2.
By similar estimates, the derivatives of F with respect to the parameters t0 , x0 ,
and f can be calculated as

Dt0 F(x) = f (t0 , x(t0 )),

Dx0 F(x) = −1,
 t (A.6)
 
D f F(x) · δ f (t) = − δ f (τ, x(τ )) dτ.
t0

Note that these are all bounded linear operators; Dt0 F(x) is because  f  ≤ R.
Hence, F ∈ Cb1 as a function of x, t0 , x0 , f , so by the implicit function theorem, the
solution x(t; t0 , x0 , f ) depends Cb1 on t0 , x0 , f .
k,α
Next, we establish Cb,u dependence on the initial conditions t0 , x0 and Cbk depen-
dence on f and t0 , x0 together. Uniform and Hölder dependence on f are lost because
k,α
the variations δ f ∈ Cb,u are not uniformly equicontinuous. The first derivatives can
be differentiated another k−1 times with respect to each of the variables, using similar
estimates as in (A.5). These derivatives are continuous as f is uniformly continuous
on the interval I . Uniform and Hölder continuity with respect to t0 , x0 can be read
off directly from the expressions (A.5), (A.6) or their higher order derivatives, as
k,α
f ∈ Cb,u . The implicit function theorem only gives an explicit formula (A.2) for
the derivative. Here, this translates into the fact that no boundedness follows for the
C 0 -norm of the solution curve, only for the norms on the derivatives.
We have thus shown that the conditions of Corollary A.4 of the implicit function
theorem have been satisfied, so there exists a neighborhood of (t0 , x0 , f ) in I × B ×F
such that for each (t 0  , x 0  , f  ) in that neighborhood there is a unique solution to (A.3)
and the solutions x depend in a Cb,u k,α
way on t0 , x0 and Cbk on all of t 0  , x 0  , f  . Note
that this result is obtained only on the interval I . We can however extend these
results to any bounded interval, by using the composition property of a flow; the
estimates may grow with interval size though. Hence, we have the following result,
see also [DK00, App. B].
Theorem A.6 (Uniform dependence on parameters of ODE solutions) Let an
k,α
ordinary differential equation (A.3) be given, where f ∈ F ⊂ Cb,u (R × B; B)
with k ≥ 1, 0 ≤ α ≤ 1, B a Banach space, and F a bounded subset. Let I ⊂ R
be a bounded interval and X = C 0 (I ; B) the Banach space of (solution) curves,
endowed with the supremum norm.
Then the flow  is a Cbk mapping
 
 : I × B × F → X : (t0 , x0 , f ) → t → x(t) .

k,α
The boundedness is understood to hold only for the derivatives. Moreover,  ∈ Cb,u
holds as a mapping from I × B for fixed f ∈ F.
156 Appendix A: Explicit Estimates in the Implicit Function Theorem

Remark A.7 Differentiable dependence on time can be dropped from this theorem.
That is, let us instead assume that f (t, x) and its derivatives Dix f (t, x), i ≤ k with
respect to x are bounded continuous with respect to (t, x). Then the flow is a Cbk
mapping  
 : B × F → X : (x0 , f ) → t → x(t)

when I ⊂ R is a bounded interval. This result follows directly from the proof, since
we only used differentiability with respect to t for differentiable dependence of 
on t.
Remark A.8 Instead of a Banach space B, we can also choose the setting of a
Riemannian manifold (M, g). Solving for the flow of a differential equation is defined
in terms of local charts, so by standard arguments the C k smoothness result extends
to this setting.
If we assume moreover in the context of Chap. 2 that (M, g) has bounded geometry
k,α
and that f ∈ Cb,u , then we can obtain stronger results close to those of Theorem A.6.
In any single normal coordinate chart the results of Theorem A.6 hold. To extend the
flow beyond one chart, we use the fact that coordinate chart transitions are uniformly
k,α
C k -bounded maps. It follows that  ∈ Cb,u on any domain such that all image
curves are covered by a uniformly bounded number of charts. This includes the
domain M × I for any finite interval I ⊂ R, since f itself is assumed bounded. The
bounds and continuity moduli will depend on |I | though.
Alternatively, uniform (Hölder) continuity estimates independent of charts can be
obtained by using Proposition 2.13 to express continuity moduli in terms of parallel
transport. See Lemma C.10, which is proven via a variation of constants method. ♦
Appendix B
The Nemytskii Operator

The Nemytskii operator creates a mapping on curves from a simple function between
spaces. That is, in its simplest form, if we have a function f : Rn → Rm , then the
associated Nemytskii operator

F : C(R; Rn ) → C(R; Rm ), F(x)(t) = f (x(t)),

maps curves x in Rn to curves y = F(x) = f ◦ x in Rm . See also [Van89, pp. 103–

109] for a clear presentation.
We investigate continuity of the Nemytskii operator for certain classes of curves.
The following definition of the Nemytskii operator in a somewhat more abstract
context on bundles over R allows e.g. for the map f to be time-dependent.
Definition B.1 (Nemytskii operator) Let I ⊂ R and let X, Y be normed vector
bundles2 over I . Furthermore, let f : X → Y be a bundle map, i.e. a fiberwise
mapping that covers the identity on I , but which is not necessarily linear in the
fibers. We define the corresponding Nemytskii operator

F : (X ) → (Y ) : x → f ◦ x, (B.1)

mapping continuous sections of X to continuous sections of Y .

In the previous definition as well as in the following lemma, we need not restrict
to vector bundles; we shall also require the case that X is a trivial fiber bundle
ρ
with a metric space as fiber (e.g. the bundle Bβ (I ; X ) in the context of Chap. 3).
Recall that the space of sections (X ) can be endowed with an exponential growth
distance (1.17) or norm (1.16), respectively. This turns (X ) into a metric (or normed
linear) space denoted by  ρ (X ) with exponent ρ ∈ R. The distance dρ (x1 , x2 ) may

2 For our purposes, a sufficient definition of a normed vector bundle π : X → R is that there exist
local trivializations τ : π −1 (U ) → U × F that are isometric with respect to the norms on X and
the normed linear space F. Note that we canonically have such trivializations by parallel transport,
see (3.52) and Proposition 3.34.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 157

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
158 Appendix B: The Nemytskii Operator

be infinite for some x1 , x2 ∈  ρ (X ) if X is a trivial metric fiber bundle. This is not a

problem, since it is only used to obtain (local) continuity estimates for sections such
that dρ (x1 , x2 ) < ∞.
Lemma B.2 (Continuity of the Nemytskii operator) Let X, Y be normed vector
bundles over I = R≥0 , or alternatively let X be a trivial fiber bundle of a metric
space. Let f ∈ C 0 (X ; Y ) be a continuous fiberwise mapping and let F : (X ) →
(Y ) be defined as in (B.1). Let ρ1 , ρ2 ∈ R and assume that one of the following
holds:
1. ρ2 > 0 and f is bounded into the normed vector bundle Y ;
2. ρ2 ≥ α ρ1 and f is α-Hölder continuous with 0 < α ≤ 1, uniformly with respect
to the fibers.
Then F is continuous as a map  ρ1 (X ) →  ρ2 (Y ) and under (2), F is moreover
α-Hölder continuous again.
Proof We first prove the statement under assumption (i). Fix x1 ∈  ρ1 (X ), let ε > 0
be given, and let x2 ∈  ρ1 (X ) be arbitrary. As f is bounded and ρ2 > 0, we can
choose a T > 0 such that

∀ t > T :  f (x1 (t)) − f (x2 (t)) e−ρ2 t ≤ 2  f  e−ρ2 T ≤ ε.

This leaves only the compact interval [0, T ] for which we still have to show that
 f (x1 (t)) − f (x2 (t)) e−ρ2 t ≤ ε. Let us denote g : I → R : t  → e−ρ2 t , then the
continuity estimate of f · g : X → Y is uniform on the compact set x1 ([0, T ]).
Hence, there exists a δ  > 0 such that for all t ∈ [0, T ] and ξ2 ∈ π −1 (t) ⊂ X ,

d(x1 (t), ξ2 ) ≤ δ  =⇒ e−ρ2 t  f (x1 (t)) − f (ξ2 ) ≤ ε.

We have that d(x1 (t), x2 (t)) ≤ e|ρ1 T | dρ1 (x1 , x2 ), so choosing δ = e−|ρ1 T | δ  yields
the required estimate for dρ1 (x1 , x2 ) ≤ δ. This proves that F is continuous at x1 .
Secondly, assume (2) and let Cα be the Hölder coefficient of f . Then we can
estimate

F(x1 ) − F(x2 )ρ2 = sup e−ρ2 t  f (x1 (t)) − f (x2 (t))

t≥0
 α
≤ sup e−ρ2 t Cα dρ1 (x1 , x2 ) eρ1 t = Cα dρ1 (x1 , x2 )α ,
t≥0

which shows that F is α-Hölder continuous again with coefficient Cα . 

Corollary B.3 Let the assumptions of Lemma B.2 with condition (1) be satisfied.
If f is fiberwise uniformly continuous with continuity modulus independent of the
fiber, then also F is uniformly continuous.
Proof This follows easily: in the proof above, the uniform continuity on the compact
set x1 ([0, T ]) can be replaced by the uniform continuity modulus of f itself. This
Appendix B: The Nemytskii Operator 159

does not depend on x1 , x2 anymore, only on their distance, so it leads to a uniform

continuity modulus of F. 
Remark B.4 The previous results also hold under time inversion. That is, if we
consider the interval I = R≤0 and invert the inequalities for ρ1 , ρ2 in conditions (1)
and (2), then Lemma B.2 and Corollary B.3 still hold true. We use this time inverted
version in Chap. 3. ♦
Appendix C
Exponential Growth Estimates

In this appendix we investigate the growth rate of higher order derivatives of a

general flow on a Riemannian manifold. Basically, if the growth of the tangent flow
is proportional to exp(ρ t), then the growth of the r -thorder derivative is of order
exp(r ρ t). This even extends to ‘fractional’ derivatives, that is, the C k,α -norm (which
includes α-Hölder continuity bounds) has this growth behavior for r = k + α. These
results will be used to obtain continuity and higher order smoothness of the persisting
NHIM. The particular exponential growth behavior exp(r ρ t) will precisely prescribe
the spectral gap condition: to construct a contraction on the r -thderivative, the normal
contraction of order exp(ρY t) must dominate the higher order exp(r ρ X t) along the
invariant manifold, hence ρY < r ρ X is required.3
These results are based on estimating variation of constants integrals and similar
in spirit to Gronwall’s lemma. We work on Riemannian manifolds, however. This
complicates matters with a lot of technicalities, but the basic ideas are still the same.
We do require uniform bounds and bounded geometry of the manifold, see Chap. 2.
Let us first show the idea for a flow on Rn and then introduce some concepts and
notation to finally treat the general case.
Lemma C.1 (Exponential growth estimates for a flow) Let t,t0 ∈ C k≥1 be
the flow of a time-dependent vector field v on Rm . Let v(t, · ) ∈ Cbk (Rm ) with
all derivatives jointly continuous
 in (t, x) ρ(t−t
∈ R × Rm and uniformly bounded by

V < ∞. Suppose that D (x) ≤ C1 e
t,t 0  0 ) for all x ∈ Rm , t ≥ t and fixed
0
C1 > 0, ρ = 0. Then for each n, 1 ≤ n ≤ k there exists a bound Cn > 0 such that

 n t,t  Cn en ρ(t−t0 ) if ρ > 0,
m 
∀ x ∈ R , t ≥ t0 : D  (x) ≤
0  (C.1)
Cn eρ(t−t0 ) if ρ < 0.

3 We formulate all statements in this section with respect to exponentially bounded flows in the
(more natural) forward time direction. That is, we work with t ∈ R≥t0 and typical exponents ρ > 0.
In our applications in Chap. 3 we use the time-reversed statements. See also Remark 1.18.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 161

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
162 Appendix C: Exponential Growth Estimates

Proof Let D denote the partial derivative with respect to the spatial variable x ∈ Rm .
We suppress the time dependence in the notation of v since we have the bound
Dn v ≤ V for all 1 ≤ n ≤ k, uniformly in space and time.
Since t,t0 is a flow, we have

Dt0 ,t0 (x) = 1 and Dn t0 ,t0 (x) = 0, 2 ≤ n ≤ k. (C.2)

For 1 ≤ n ≤ k we can write, suppressing arguments t0 , x,

d n t  n
 
D  = Dn (v ◦ t ) = Dv ◦ t · Dn t + Dl v ◦ t · Pl,n D1 t , . . . , Dn−1 t ,
dt
l=2
(C.3)
where the Pl,n are homogeneous, weighted polynomials as in Definition C.2 below.
In the first equality, the switching of partial derivatives is well-defined, because the
spatial derivative in the middle expression is well-defined and the resulting function
continuous. In the right-hand expression we have already used Proposition C.3 and
separated the homogeneous term with Dn t (when l = 1). The result is a linear
differential equation for Dn t with the inhomogeneous terms in the sum consisting
of lower order derivatives Di t , i < n, only.
For n = 1, statement (C.1) is already true by assumption and in that case we also
see that (C.3) is a homogeneous linear differential equation. Denote by x (t, t0 ) the
solution operator for this system with initial point x ∈ Rm , then
 
Dt,t0 (x) = x (t, t0 ) Dt0 ,t0 (x) = x (t, t0 ) · 1 = x (t, t0 ). (C.4)

This solution operator acts by left-composition on linear maps, so we read off that
x (t, t0 ) = Dt,t0 (x) and find the estimate x (t, t0 ) ≤ C1 eρ(t−t0 ) . Now we turn
to the induction step. For n > 1, we still have essentially the same solution operator
x (t, t0 ) for the homogeneous part, only now acting by composition on multilinear
maps Dn t,t0 (x) ∈ Ln (Rm ): the solution operator is not influenced by considering
multilinear maps, as Dv and  act by linear composition from the left, essentially
on tangent vectors. Therefore, the same growth estimate for x (t, t0 ) still holds.
 on thei ρDt  , i < n and by
The inhomogeneous terms in (C.3) depend i t
 i only
the induction hypothesis we can estimate D   ≤ Ci e . Using variation of
t

constants, the solution can now be written as

 
n
t  
Dn t (x) = x (t, τ ) · Dl v ◦ τ · Pl,n D1 τ , . . . , Dn−1 τ dτ, (C.5)
t0 l=2

where the homogeneous part of the solution is zero because Dn t0 ,t0 (x) = 0 for
n > 1. Given that the weighted degree of Pl,n is n, we can directly estimate
Appendix C: Exponential Growth Estimates 163

 
 n t  t n
 
D  (x) ≤ x (t, τ ) Dl vP l,n D1 τ , . . . , Dn−1 τ  dτ
t0 l=2
 t  
≤ C1 eρ(t−τ ) V R {Ci }i<n en ρ(τ −t0 ) dτ
t0
  en ρ(t−t0 ) − eρ(t−t0 )
= C1 V R {Ci }i<n . (C.6)
(n − 1) ρ

The bound R depends on finite sums and products of finite terms, so is finite again.
When ρ > 0, the denominator is positive and the numerator can be estimated by
en ρ(t−t0 ) ; when ρ < 0, the numerator can be estimated by eρ(t−t0 ) , adding a minus
sign to both parts of the fraction. Thus, in both cases (C.1) holds. This completes the
induction step. 
Before generalizing this lemma to Riemannian manifolds, we first refine some
previous notation. Instead of Rm , we more generally consider linear spaces V, W and
spaces Lk (V ; W ) of (multi)linear maps for the (higher order) derivatives of maps
f : V → W.
Definition C.2 (Homogeneous weighted polynomial) Let Pa,b (y1 , . . . , yn ) be a
polynomial in the variables y1 to yn . We call P a homogeneous weighted polynomial
of degree (a, b) if it is a homogeneous polynomial of degree a and moreover, each
p p
term y1 1 . . . yn n has weighted degree


n
i · pi = b. (C.7)
i=1

As a consequence, such a polynomial cannot have factors yn for n > b and the factor
yb can only occur as a term on itself when a = 1.
This definition can now be used to denote the higher derivatives of a composition
of two functions f, g on vector spaces.
Proposition C.3 (Higher order derivatives of compositions of functions) Let the
mapping x → f (g(x), x) be given with f : V × U → W and g : U → V two suffi-
ciently differentiable functions between vector spaces U, V, W . Then the k-th order
derivative of this mapping with respect to x is of the form
 d k   1 
f (g(x), x) = 2 f (g(x), x) · Pl,k−m D g(x), . . . , D
Dl1 Dm k−m
g(x) ,
dx
l,m≥0
l+m≤k
(l,m) =(0,0)
(C.8)
where Pl,k−m is a homogeneous weighted polynomial of degree (l, k−m) with l higher
order derivatives Di g(x) in each term, and weighted degree k−m: the total number
of derivatives that either produced an additional Dg(x) term or differentiated an
existing one.
164 Appendix C: Exponential Growth Estimates
   
Remark C.4 We will shorten the notation Pl,k D1 g(x), . . . , Dk g(x) = Pl,k D• g(x) .

Remark C.5 Note that Dl1 Dm

 ∗ ⊗l  ∗ ⊗m 2 f (g(x), x) is actually an element of the tensor product
space W ⊗ V ⊗ U and Pl,k−m an element of the (l, k −m)-linear maps
⊗l
 ∗ ⊗k−m
V ⊗ U , or (l, k −m) tensors, so the composition is indeed a mapping in
 ∗ ⊗k
W⊗ U = Lk (U ; W ), as expected. ♦
Proof This is easily proven by induction. For k = 1 we have

d
f (g(x), x) = D1 f (g(x), x) · Dg(x) + D2 f (g(x), x),
dx
which satisfies (C.8). For the induction step we have
 d k+1 d   • 
f (g(x), x) = 2 f (g(x), x) · Pl,k−m D g(x)
Dl1 Dm
dx dx
l,m≥0
l+m≤k
(l,m) =(0,0)

 • 
= Dl+1
1 D2 f (g(x), x) · D g(x) · Pl,k−m D g(x)
m 1

l,m≥0
l+m≤k
(l,m) =(0,0)
 
+ Dl1 D2m+1 f (g(x), x) · Pl,k−m D• g(x)
d  • 
+ Dl1 Dm2 f (g(x), x) · Pl,k−m D g(x)

dx
  
= D1 D2 f (g(x), x) · Pl+1,k+1−m D• g(x)
l+1 m

l,m≥0
l+m≤k
(l,m) =(0,0)
 
+ Dl1 D2m+1 f (g(x), x) · Pl,k+m D• g(x)
 • 
+ Dl1 Dm2 f (g(x), x) · Pl,k+1−m D g(x)
  
= D1 D2 f (g(x), x) · Pl,k+1−m D• g(x) .
l m

l,m≥0
l+m≤k+1
(l,m) =(0,0)

This is again of the form (C.8): k − m = (k + 1) − (m + 1), so all terms can be

absorbed in the new sum for k + 1. 
Let us make a few remarks on the form of (C.8). The P0,m for m < k are zero,
because then we have too few derivatives with respect to x; we have P0,k = 1 though.
After Definition C.2 it was already noted that in a polynomial of weighted degree
k, the factor yk can only occur as a term on itself, up to a constant factor. More
specifically in this case, Dk g(x) occurs exactly once, in the term
Appendix C: Exponential Growth Estimates 165

D1 f (g(x), x) · Dk g(x).

This can easily be seen by direct calculation or induction. Finally, when the compo-
sition mapping is of the form x → f (g(x)), then we only have terms with m = 0
and all polynomials in (C.8) have weighted degree k in that case.
The next step is to generalize Lemma C.1 to a Riemannian manifold (M, g). Here
we first need to define what we mean by higher derivatives of the flow. The tangent
flow Dt is well-defined as a mapping on TM, but higher derivatives live on higher
order tangent bundles Tk M. These abstract bundles make doing explicit estimates
as in the proof of Lemma C.1 difficult. Instead, we reuse the idea of Definition 2.9
and introduce a different representation of higher derivatives in terms of normal
coordinate charts.
Definition C.6 (Higher derivative on Riemannian manifolds) Let M, N be
Riemannian manifolds and f : M → N a smooth map. With the notation f x =
exp−1
f (x) ◦ f ◦ expx of f represented in normal coordinate charts, we define for k ≥ 1
and x ∈ M the higher order derivative

Dk f (x) = Dk f x (0) = Dk exp−1
f (x) ◦ f ◦ expx (0) (C.9)

as an element of Lk (Tx M; T f (x) N ).

Remark C.7 Definition C.6 can be viewed as creating a more explicit representation
of the jet bundle of the trivial fiber bundle π : M × N → M. A map f : M → N
is a section of this trivial bundle and the k-jet of f at a point x is fixed in terms
of the derivatives in (C.9) up to order k, in the normal coordinate chart centered
at x. We shall see below that this representation is still a (global) bundle, while the
explicit choice of normal coordinate charts introduces a convenient norm to measure
the jets. ♦
Let us make a few remarks on this choice of representation of higher derivatives.
First of all, for k = 1 this definition coincides with the ordinary tangent map, as
D expx (0) = 1Tx M by the natural identification T0 (Tx M) ∼ = Tx M. Furthermore,
this representation of derivatives admits operator norms, and all this behaves nicely
under composition of maps by virtue of the property ( f ◦ g)x = f g(x) ◦ gx for local
coordinate charts:
    
 2   
D ( f ◦ g)(x) = D2 f g(x) ◦ gx (0)
    
 
= D2 f g(x) (0) Dgx (0), Dgx (0) + D f g(x) (0) D2 gx (0) 
 
 
= D2 f (g(x)) · Dg(x)⊗2 + D f (g(x)) · D2 g(x)
   
   
≤ D2 f (g(x)) · Dg(x)2 + D f (g(x)) · D2 g(x),
166 Appendix C: Exponential Growth Estimates

that is, these operator norms as defined via normal coordinate charts are truly norms
and satisfy the usual product rules for compositions of (multi)linear maps.
The operator norms are induced by the norms on the tangent spaces of TM, which
in turn are induced by the metric. These norms depend smoothly on the base point,
so they glue together to a smooth function  ·  : TM → R≥0 that we will call a
‘bundle norm’ on the tangent bundle4 or sometimes refer to as just a norm on TM.
Higher derivatives can be viewed as partial sections of the vector bundle

Lk (TM; TN ) = TN  (TM ∗ )⊗k . (C.10)

That is, we define Lk (TM; TN ) as a bundle over M × N with fiber Lk (Tx M; T y N )

over the point (x, y) ∈ M × N . This is indicated by the operator , which differs
from the usual tensor product ⊗ in the sense that the new bundle is constructed
on the product of the base spaces instead of one common base. Now the k-th order
derivative (as in Definition C.6) of a map f : M → N is a section of the bundle (C.10)
restricted to the base submanifold Graph( f ) ⊂ M × N and the derivative Dk f (x) is
the point in the section over (x, f (x)). More generally, we can define vector bundles
of (l, k)-linear maps

Ll,k (TM; TN ) = (TN )⊗l  (TM ∗ )⊗k (C.11)

and the disjoint union of all these bundles. The bundle norms on TM and TN together
naturally induce bundle operator norms on these. From here on, we set M = N and
assume that f = t is a flow.
To finally generalize Lemma C.1 to Riemannian manifolds, there is still one issue
to tackle. When taking the time-derivative as in (C.3), the target base point t (x)
changes. This suggests that a covariant derivative is required. The Ll,k (TM; TM) are
smooth manifolds in a natural way, however, so both the tangent vector dtd Dk t (x)
and the differential of  ·  are well defined in
 this interpretation
 and independent of
a connection, and certainly their product dtd Dk t (x) is. The tangent exponential
maps D exp at x and t (x) together induce a local coordinate chart on Ll,k (TM; TM)
in a neighborhood of Ll,k (Tx M; Tt (x) M). We will use these local coordinates for
explicit calculations.
The dependence on the base point of the norms and normal coordinate charts
in (C.9) introduces additional terms when formulating equations (C.3) and (C.5) on
a Riemannian manifold. Under the assumption that (M, g) is of bounded geometry,
however, all these additional terms will be globally bounded. Hence, these will only
contribute to the overall constants Cn in Lemma C.1, but not influence the basic
result.
Lemma C.8 (Exponential growth estimates on a Riemannian manifold) Let
t,t0 ∈ C k≥1 be the flow of a time-dependent vector field v on a Riemannian manifold

4 Note that this is stronger than a Finsler manifold as the Finsler structure F : TM → R≥0 is allowed
to be asymmetric, that is, on each tangent space, F need only scale linearly for positive scalars. I
did not investigate whether it is possible to generalize this theory to Finsler manifolds.
Appendix C: Exponential Growth Estimates 167

(M, g) of (k+3)-bounded geometry. Let v(t, · ) ∈ Xkb (M) with all derivatives jointly
continuous in (t, x) ∈ R × M  and uniformly
 bounded by V < ∞ with respect to
Definition C.6. Suppose that Dt,t0 (x) ≤ C1 eρ(t−t0 ) for all x ∈ M, t ≥ t0 and
fixed C1 > 0, ρ = 0. Then for each n, 1 ≤ n ≤ k there exists a bound Cn > 0 such
that 
 n t,t  Cn en ρ(t−t0 ) if ρ > 0,

∀ x ∈ M, t ≥ t0 : D  (x) ≤
0  (C.12)
Cn eρ(t−t0 ) if ρ < 0.

Proof The proof is basically the same as the proof of Lemma C.1, with additional
technicalities due to M being a manifold. We will focus on these.
Equation (C.3) can be formulated in terms of the tangent normal coordinate chart

D exp−1 ∼
y : TB(y; δ) ⊂ TM → T(T y M) = (T y M)
2

with y = 
t (x) fixed. Note that we are finally interested in the growth behavior

of t → D t (x); this is defined in a coordinate-free way, so it is not influenced
 n

by our choice of intermediate coordinates. In these normal coordinates, both the

metric and its derivatives are bounded due to Theorem 2.4, and the vector field is C k
bounded by assumption. We have

d
D exp−1
y ◦D  (x)
n t
dt
d  −1
= D exp−1
y ◦D expt (x) ◦ ◦ expx (0)
n t
dt
d n
  • 
= D exp−1
y ◦ Dl exp−1
 t (x) ◦ exp y (0) · Pl,n D exp−1
y ◦ ◦ expx (0) .
t
dt
l=1

This splits the dependence on t in the target base point t (x) from that in the
derivatives Dn t itself. Note that the sum must be interpreted as a sum of terms in
the single fiber Ln (Tx M; T y M) over the base point (x, y). By using the coordinate
map D exp−1 y , we transferred the problem to fixed linear spaces, which allows us to
make sense of the differentiation with respect to t. In other words, D exp−1 y induces
locally trivializing coordinates for Ln (Tx M; TM) in a neighborhood of y with x
fixed. As D exp−1y is linear on the fibers, we can distribute it over the sum to further
obtain

 d

n
 −1  •
= D exp−1
y ◦Dl
exp  (x)
t ◦ exp y (0) · Pl,n D exp −1
y ◦ t
◦ exp x (0)
dt
l=1

d
  n
−1
 −1  −1 d
= D exp y ◦D expt (x) ◦ exp y (0) · D exp y ◦ ◦ expx (0) +
n t
... .
dt dt
l=2
(C.13)
168 Appendix C: Exponential Growth Estimates

In the last line, the homogeneous part is separated from the non-homogeneous terms
as in (C.3).
Working out the details of the homogeneous part, we obtain5

d
 −1  
D exp−1 −1
y ◦D expt (x) ◦ exp y (0) · D exp y ◦  ◦ expx (0)
n t
dt
d
 −1  n t
= D exp−1
y ◦D expt (x) ◦ exp y (0) · D  (x)
dt
 nd
+ D D exp−1y ·D exp−1
y ◦ ◦ expx (0)
t
dt
d
  n t
= D exp y ◦ D exp−1
−1
 (x)
t ◦ exp y (0) · D  (x)
dt
  n

+ D D exp−1y · Dl D exp−1 • t
y ◦v ◦ exp y (0) · Pl,n (D  (x)).
l=1

Note that again all terms l ≥ 2 in the sum are inhomogeneous terms that we will
add to those already present in (C.13). The homogeneous term is some linear vector
field acting (from the left) on Dn t (x) and it is precisely the vector field generating
Dt (x), which is the original case n = 1. Hence, we can again define the operator
xt,t0 as post-composition with Dt (x) and write the flow of Dn t (x) using a vari-
ation of constants integral with all the non-homogeneous terms. These terms again
contain only lower order derivative flows Dl t (x), l < n.
We can now take the operator norm of this expression. In principle we should be
careful that this bundle norm depends on the changing target point t (x). The normal
coordinates were chosen around y = t (x), however, and in these coordinates the
derivative of the metric at the origin (corresponding to y) is zero, hence the norm
has zero derivative. We can thus simply apply the operator norm to the variation of
constants integral and obtain estimates as in (C.6). The additional factors introduced
by differentiation of normal coordinate transition maps are bounded by Lemma 2.6
under the assumption that (M, g) is of (k+3)-bounded geometry. The inhomogeneous
terms still contain at least one factor Dl t (x), so the result in case ρ < 0 holds as
well. 
These exponential growth results can be extended further to uniform and Hölder
continuity in the highest derivatives. The Hölder continuity then is with respect to
the growth rate (k + α)ρ, where k is the order of the derivative and 0 < α ≤ 1 the
Hölder constant. Thus, α-Hölder continuity can be viewed as a fractional derivative;
Lipschitz continuity (when α = 1) can indeed be viewed as almost differentiability to
one higher order. The case α = 0 we shall identify with uniform continuity. Here we

 −1
5 The time derivative of D exp−1
y ◦ D expt (x) ◦ exp y (0) actually turns out to be zero in local
coordinates. This follows from an analysis of the exponential map as the time-one geodesic flow
in normal coordinates around y. This result is not relevant for us, so we leave out this tedious
calculation.
Appendix C: Exponential Growth Estimates 169

have no explicit modulus of continuity, which requires an arbitrarily small additional

μ > 0 in the exponent kρ + μ to compensate.
Remark C.9 (On using a global continuity modulus) In the next lemma, as well
as in Corollary C.12 below, we shall make abuse of notation in writing expressions
such as s(x2 ) − s(x1 ), where s is a section of a vector bundle, cf. (C.15), that is,
we compare objects that live in different fibers of a vector bundle.6 This notation
should be interpreted according to Remark 2.12. That is, if x1 , x2 are M-close in
the spirit of Definition 2.8, then this is well-defined in terms of local charts, and for
continuity estimates this is equivalent to an estimate by identification of the vector
bundle over x1 , x2 via parallel transport, cf. Proposition 2.13. If x1 and x2 are not
close, then we can use any choice of isometric identification of the vector bundle over
these points, such as the construction of parallel transport along solutions curves in
Sect. 3.7.4. In this case the notation can effectively be interpreted as an estimation
by the sum of the norms of the separate terms with the triangle inequality. When
applying this lemma, we shall always have such an isometric identification at hand,
hence these arguments can be made rigorous, and the notation provides a sensible
heuristic then. ♦
Lemma C.10 (Exponential growth estimates with Hölder continuity) Let t,t0 ∈
C k≥1 be the flow of a time-dependent vector field v on a Riemannian manifold (M, g)
of (k +3)-bounded geometry. Let D denote the partial derivative with respect to the
spatial variable x ∈ M as in Definition C.6 and let v(t, · ) ∈ Xk,α b,u (M), 0<α≤1
 
with all derivatives jointly continuous in (t, x) ∈ R×M. Suppose that Dt,t0 (x) ≤
C1 eρ(t−t0 ) for all x ∈ M, t ≥ t0 and fixed C1 > 0, ρ > 0.
Then in addition to the results of Lemma C.8, there exists a bound Ck,α > 0 such
that  
 
∀ t ≥ t0 : Dk t,t0  ≤ Ck,α e(k+α) ρ(t−t0 ) . (C.14)
α

If instead v(t, · ) ∈ Xkb,u (M), i.e. the special case α = 0, then for each μ > 0 there
exists a continuity modulus εk,μ such that
 
 
∀ t ≥ t0 : Dk t,t0 (x2 ) − Dk t,t0 (x1 ) ≤ εk,μ (d(x1 , x2 )) e(k ρ+μ)(t−t0 ) , (C.15)

 
that is, x → t → Dk t,t0 (x) is uniformly continuous in x, in  · k ρ+μ -norm.
Remark C.11 We restricted this lemma to the case ρ > 0 only. A result similar to
that in C.8 for ρ < 0 could be obtained for completeness sake, but it clutters the
already detailed proof, while we do not need the result. ♦

6 Note that the higher derivatives Dk t (x) of a flow are actually interpreted as elements of a
bundle of type (C.10). These bundles are still naturally induced by the tangent bundles of underlying
manifolds, so all bounded geometry techniques, such as uniformity of normal coordinate charts,
unique local trivializations by parallel transport, are induced on these bundles as well.
170 Appendix C: Exponential Growth Estimates

Proof The idea of the proof is essentially the same as that of Lemma C.8. The addi-
tional difficulty is that (Hölder) continuity requires finite, non-differential estimates
when comparing any two flows starting from different initial points x1 , x2 ∈ M.
Let d(x1 , x2 ) < δ M where δ M is M-small as in Definition 2.8. We drop t0 from
the notation and define ξi (t) = t (xi ), i = 1, 2 as the solution curves with xi as
initial conditions. We want to study the growth behavior of

t → Dk t (x2 ) − Dk t (x1 ). (C.16)

Note that this difference is defined with respect to coordinate charts at source and
target that contain x1 , x2 and ξ1 (t), ξ2 (t), respectively, but not in general.
We denote by γt the unique shortest geodesic that connects ξ1 (t) to ξ2 (t) when
d(ξ1 (t), ξ2 (t)) < δ M . Next, we set
 
ϒ t = Dk t (x2 ) · (γ0 )⊗k − (γt ) · Dk t (x1 ) ∈ Lk Tx1 M; Tξ2 (t) M (C.17)

to be the difference of the respective k-th order derivative flows, parallel transported
to matching spaces at their source and target. It is easily verified that ϒ t satisfies
initial conditions ϒ t0 ,t0 = 0 for any k ≥ 1.
Due to Proposition 2.13, the formulation in (C.17) with parallel transport to mea-
sure variation of the flows isequivalent
 to measuring (C.16) in normal coordinate
charts. Hence, if we study ϒ t  in charts, we may drop7 the parallel transport
terms at the cost of an (unimportant) global factor in the estimates. We assume
that d(ξ1 (t), ξ2 (t)) < δ M and study ϒ t in a normal coordinate chart covering both
points. Taking the difference of (C.3) with x1 , x2 inserted, we see that ϒ t satisfies
the differential equation

d t 
ϒ = Dv ◦ t (x2 ) · ϒ t + Dv ◦ t (x2 ) − Dv ◦ t (x1 ) · Dk t (x1 )
dt
 k
 
+ Dl v ◦ t (x2 ) · Pl,k D• t (x2 ) − (x2  x1 ).
l=2

This equation provides a variation of constants integral for ϒ t based on the flow
x2 (t, t0 ):

7 We could include the parallel transport terms, repeat similar arguments as in the proof of
Lemma C.8 and express everything in (induced) normal coordinate charts, but this would clut-
ter the proof here even more. These terms would all be bounded and Lipschitz continuous by
bounded geometry, hence not essentially alter the result.
Appendix C: Exponential Growth Estimates 171
 t 
ϒt = x2 (t, τ ) · Dv ◦ τ (x2 ) − Dv ◦ τ (x1 ) · Dk τ (x1 )
t0

k 
τ
 • τ 
+ x2 (t, τ ) · D v ◦  (x2 ) · Pl,k D  (x2 ) − (x2  x1 ) dτ.
l

l=2
(C.18)
We proceed by induction over k. For k = 1 we only have the first term of the
integrand. Using that Dv is uniformly α-Hölder, we have

 t t    
ϒ  ≤ x (t, τ ) Dv ◦ τ (x2 ) − Dv ◦ τ (x1 ) Dτ (x1 ) dτ
2
t0
 t  α
≤ C1 eρ(t−τ ) Dvα τ (x2 ) − τ (x1 ) C1 eρ(τ −t0 ) dτ
t0
 t  α
≤ C12 Dvα eρ(t−τ ) C1 eρ(τ −t0 ) x2 − x1  dτ
t0
eα ρ(t−t0 )
≤ C12+α Dvα eρ(t−τ ) x2 − x1 α .
αρ

Next, in the induction step for k > 1, we get the additional terms from (C.18) in
the integrand. These are (up to constants) a product of the flow , Dl v ◦ τ (x) and
Di (x)’s with weighted degree k. The terms Dl v ◦ τ (x) are uniformly α-Hölder
continuous in x analogous to the case k = 1 above. Each of the Di (x)’s satisfies the
Hölder estimate of this lemma by the induction hypothesis and the growth estimates
of Lemma C.1. Hence, for each term in the integrand, we obtain Hölder continuity
with respect to x with growth behavior at most eρ(t−τ ) e(k+α)ρ(τ −t0 ) . Integration then
yields the stated result.
Finally, the uniformly continuous case is an extension along the same lines as
Corollary B.3. The map x → Dl v ◦ τ (x) is uniformly continuous when measured
in  · μ -norm and by induction the alternative result (C.15) follows. 
Finally, we extend the Hölder continuous growth estimates to a parameter depen-
dent version. This is formulated to exactly fit the context of derivatives of TX with
ρ
respect to y ∈ Bη (I ; Y ), such as in (3.43b). Note that Remark C.9 applies again.
Corollary C.12 (Exponential growth with Hölder continuity and a parameter)
Assume the setting of Lemma C.10. Let the vector field v furthermore depend on a
α for all 0 ≤ l ≤ k, uniformly in
third variable y ∈ Y such that Dlx v(t, x, · ) ∈ Cb,u
t, x and that all original bounds are uniform in y as well. Let η ∈ B ρ (R; Y ) denote

a curve in Y and t,t t,t0
η the flow of v(t, · , η(t)). Assume that η  → t  → η (x) is
0

uniformly Lipschitz with respect to the distance function dρ on curves C(R≥t0 ; X ).

Then the map η → Dk η is Hölder continuous in the sense that there exists a
bound Ck,α,Y > 0 such that
 
 0 (x) ≤ C (k+α) ρ(t−t0 )
∀ t ≥ t0 , x ∈ X : η → Dk t,t
η  Y ,α e . (C.19)
α
172 Appendix C: Exponential Growth Estimates

In case of uniform continuity (i.e. α = 0), then for each μ > 0 there exists a continuity
modulus εk,μ,Y such that
 
 0 (x) − Dk t,t0 (x) ≤ ε (k ρ+μ)(t−t0 )
∀ t ≥ t0 , x ∈ X : Dk t,t
η2 η1  k,μ,Y (d(η1 , η2 )) e .
(C.20)
In both cases we interpret the continuity moduli as globally defined using Remark 2.12.
Proof The proof closely follows that of Lemma C.10; let us indicate the differences.
We define the variation

ϒ t = Dk tη2 (x) · (γ0 )⊗k − (γt ) · Dk tη1 (x) (C.21)

and study it by a variation of constants integral in local charts, similar to (C.18). In

this case we obtain
 t 

   
ϒt = η2 (t, τ ) · Dv τ, τη2 (x), η2 (τ ) − Dv τ, τη1 (x), η1 (τ ) · Dk τη2 (x)
t0

k 
   
+ l
Dv τ, τη2 (x), η2 (τ ) · Pl,k D• τη2 (x) − (2  1) dτ. (C.22)
l=2

As in Lemma C10, the factors η2 (t, τ ) and Dl τη2 (x) satisfy appropriate exponential
growth conditions. By induction over l < k the maps η → Dl tη (x) are α-Hölder
continuous, while all Dl v are uniformly α-Hölder in x, y, and η → Dl tη (x) is
 
uniformly Lipschitz by assumption, so η → Dl v t, tη (x), η(t) is also α-Hölder
when measured in  · α ρ -norm (or in  · μ -norm in case of uniform continuity, see
Appendix B).
In each term of the integrand, we can estimate the variation with respect to η as a
sum of the variations with respect to each factor (a product rule). The factor that is
being varied adds eα ρ(τ −t0 ) (or eμ(τ −t0 ) in case α = 0) to the overall growth estimate.
The proof is completed by inserting all these estimates into (C.22) and again using
the fact that we have a finite number of globally bounded terms. 
Appendix D
The Fiber Contraction Theorem

In this appendix, we give a proof of the fiber contraction theorem. This result is
originally due to Hirsch and Pugh [HP70]; the proof presented here is taken from
Vanderbauwhede [Van89, p. 105]. The fiber contraction theorem is a convenient
general tool to obtain convergence of functions in C k -norm when a direct contraction
in C k -norm is not available. Instead, one inductively constructs contractions for the k-
thderivative with all lower order derivatives assumed fixed. If this contraction depends
continuously on the lower order derivatives, then the fiber contraction theorem can
be applied to conclude that the sequence of the function together with its derivatives
converges to a fixed point. With the additional theorem on the differentiability of
limit functions, it can then be concluded that the sequence converges in C k -norm.
Theorem D.1 (Fiber contraction theorem) Let X be a topological space, (Y, d)
a complete metric space and let F : X × Y → X × Y be a fiber mapping, that is,
F(x, y) = F1 (x), F2 (x, y) , with the following properties:
1. F1 has a unique, globally attracting fixed point x  ∈ X , that is,

∀ x ∈ X : lim F1n (x) = x  ;

n→∞

2. there is a neighborhood U ⊂ X of x  , such that F2 : U × Y → Y is a uniform

contraction on Y with contraction factor q < 1; let y  ∈ Y denote the unique
fixed point of F2 (x  , · ) : Y → Y , as given by the Banach fixed point theorem;
3. the mapping F2 ( · , y  ) : X → Y is continuous.
If only properties (1) and (2) are assumed, then (x  , y  ) is the unique fixed point
of F. If moreover (3) holds, then this fixed point is globally attractive.
Proof The point (x  , y  ) is clearly the unique fixed point of F, where property (1)
implies uniqueness of x  as fixed point of F1 and (2) uniqueness of y  under
F2 (x  , · ).
The point x  is by assumption attractive under F1 , thus for the final conclusion
of global attractivity, it remains to show that y → y  under F.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 173

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
174 Appendix D: The Fiber Contraction Theorem

Let (x, y) ∈ X × Y be arbitrary and consider the sequence (xn , yn ) = F n (x, y)

for n ≥ 0. Since xn → x ∗ ∈ U , there exists an N ∈ N such that xn ∈ U for all
n ≥ N . By shifting the sequence (xn , yn ), we can assume without loss of generality
that xn ∈ U for all n ≥ 0 and use property (2) to estimate

d(yn+1 , y  ) = d(F2 (xn , yn ), F2 (x  , y  ))

≤ d(F2 (xn , yn ), F2 (xn , y  )) + d(F2 (xn , y  ), F2 (x  , y  ))
≤ q d(yn , y  ) + αn . (D.1)

On the other hand, αn = d(F2 (xn , y  ), F2 (x  , y  )) → 0 as n → ∞ from proper-

ties (1) and (3). Let ᾱk = supn≥k αn , then we also have ᾱk → 0.
For each k ∈ N, let δk,k = d(yk , y  ) and recursively define δn+1,k = q δn,k + ᾱk .
From (D.1) we see that d(yn , y  ) ≤ δn,k when n ≥ k. Now the map f : δ → q δ + ᾱ
is a contraction for any ᾱ ∈ R, so it has a unique, attractive fixed point δ  (ᾱ) and
solving the equation f (δ  ) = δ  yields

ᾱ
δ = .
1−q

Let ε > 0 be given and choose k large enough that ᾱk < 21 (1−q)ε. As limn→∞ δn,k =
δk we see that there exists some N such that

2 ᾱk
∀n ≥ N: δn,k < 2δk = < ε.
1−q

From this we conclude that d(yn , y  ) < ε for all n ≥ N . 

The following theorem is quite standard. We shall extend it to smooth manifolds
and higher derivatives, though.
Theorem D.2 (Differentiability of limit functions) Let Y be a Banach space and
let C k (Rn ; Y ) denote the space of C k functions Rn → Y equipped with the weak
Whitney topology. Let { f n }n≥0 be a sequence in C 1 (Rn ; Y ) that converges to f ∈
C 0 (Rn ; Y ) with respect to the C 0 topology, and assume that there is a function
g ∈ C 0 (Rn ; L(Rn ; Y )) such that D f n → g.
Then D f = g, or in other words, f n → f in C 1 (Rn ; Y ) with respect to the weak
Whitney topology.
Proof By the fundamental theorem of calculus we have
 t  t
d
f n (x + t h) = f n (x) + f n (x + τ h) dτ = f n (x) + D f n (x + τ h) · h dτ.
0 dτ 0

Uniform convergence of D f n → g on the compact set {x + τ h|τ ∈ [0, t]} allows us

to take the limit n → ∞ inside the integral to obtain
Appendix D: The Fiber Contraction Theorem 175
 t
f (x + t h) = f (x) + g(x + τ h) · h dτ,
0

and by differentiation with respect to t we conclude that g(x) · h is the directional

derivative of f at x along h.
Note that g(x) : Rn → Y is a bounded linear operator by assumption, so let us
verify that it is the total derivative, D f (x) = g(x), that is,

 f (x + h) − f (x) − g(x) · h
lim = 0.
h→0 h

Using the mean value theorem, we have

 f (x + h) − f (x) − g(x) · h ≤ sup g(x + ξ h) − g(x)h

ξ ∈[0,1]

and g is continuous, so indeed differentiability holds and D f (x) = g(x). 

Remark D.3 The statement that f is differentiable at x is local, so this result imme-
diately translates to maps C 1 (X ; Y ) with X a smooth manifold by considering a local
coordinate chart around x ∈ X .
This theorem could probably be generalized even further such that X, Y are
allowed to be Banach manifolds. The fact that g is continuous linear by assump-
tion mitigates possible convergence problems when having to consider infinitely
many independent partial derivatives. We should be careful though, since the weak
Whitney (or compact-open) topology is not clearly defined anymore when X is
infinite-dimensional. ♦
Corollary D.4 Assume the setting of Theorem D.2. Let { f n }n≥0 be a sequence in
C k≥2 (Rn ; Y ) that converges to f in C k−1 (Rn ; Y ) and let Dk f n → g converge in
C 0 (Rn ; Lk (Rn ; Y )). Then f n → f converges in C k (Rn ; Y ).
This is a trivial extension of Theorem D.2 when using the natural identification
L(Rn ; Lk−1 (Rn ; Y )) ∼
= Lk (Rn ; Y ).
Appendix E
Nonlinear Variation of Flows

In this appendix we collect two results on variation of nonlinear flows. The first is
a generalization of Lagrange’s variation of constants formula and the second is an
application of it to calculate the derivative of a flow with respect to parameters. Both
results are formulated for fully nonlinear flows.
The classical variation of constants integral due to Lagrange is well known.
Although Lagrange applied this method to the nonlinear problem of orbital mechan-
ics, a less known result of Alekseev [Ale61] (see also [LL69, p. 78]) generalizes the
variation of constants integral to the full nonlinear case.
Theorem E.1 (Nonlinear variation of constants) Let X be a smooth manifold and
let t,t0 (x) be the flow generated by the time-dependent vector field v(t, x), locally
Lipschitz in x. Let r (t, x) be an arbitrary (not necessarily small) perturbation, locally
Lipschitz in x as well. Then rt,t0 (x) is the flow generated by v + r if and only if it
satisfies the nonlinear variation of constants formula
 t
rt,t0 (x) = t,t0 (x) + D(t, τ, rτ,t0 (x)) r (τ, rτ,t0 (x)) dτ. (E.1)
t0

Proof Using uniqueness of solutions, it is sufficient to show that (E.1) satisfies the
differential equation and initial conditions rt,t (x) = x. The latter follows automat-
ically from t,t (x) = x. For the first part, we differentiate

d
t,τ  ∂ t,τ  d τ,t0
 ◦ rτ,t0 (x) =  (y) τ,t0 + Dt,τ (rτ,t0 (x)) ·  (x)
dτ ∂τ y=r (x) dτ r
= −Dt,τ (rτ,t0 (x)) · v(τ, rτ,t0 (x))
+ Dt,τ (rτ,t0 (x)) · (v + r )(τ, rτ,t0 (x))
= Dt,τ (rτ,t0 (x)) · r (τ, rτ,t0 (x)).

This expression yields (E.1) when integrated from t0 to t. 

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 177

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
178 Appendix E: Nonlinear Variation of Flows

Notice that (E.1) looks ill-defined on a manifold, but should be read as integration
from the point x along the vector field defined by the integrand, which is indeed, for
each τ ∈ [t0 , t], exactly defined to be the tangent vector to the curve τ → t,τ ◦
rτ,t0 (x), making the equation self-consistent. If (X, g) is a Riemannian manifold,
then this formula yields the distance estimate

  t  
d rt,t0 (x), t,t0 (x) ≤ D(t, τ, τ,t0 (x)) r (τ, τ,t0 (x)) dτ. (E.2)
r r
t0

As a differential variant of the previous result, we state the following.

Theorem E.2 (Differentiation of a flow) Let t,t s (x 0 ) be a flow on a manifold X ,
0

defined by a vector field vs (t, x) that also depends on time and an external parameter
s ∈ R. Let (s, x) → vs (t, x) ∈ Cb1 with derivative jointly continuous in (s, t, x).
Then the derivative of the flow with respect to s is given by
 t
d t,t0 d
 (x0 ) = s (x(τ ))
Dt,τ vs (τ, x(τ )) dτ, (E.3)
ds s t0 ds

for any fixed t, t0 , and where x(τ ) = τ,t

s (x 0 ).
0

See [DK00, Thm. B.3] for a proof of the formula for differentiation of a flow with
respect to a parameter. This is a slightly modified case where the vector field is time-
dependent. Theorem A.6 and Remark A.7 show that the result can be generalized to
the non-autonomous case and differentiable time-dependence of v is not required.
Appendix F
Riemannian Geometry

In this appendix we recall standard facts from Riemannian geometry and establish
some notational conventions. This appendix is targeted at the reader who has basic
knowledge of Riemannian manifolds, but wants to have a quick refresh. For more
detailed expositions see for example [GHL04, Jos08], or [Lan95] for a more abstract
presentation in the context of Banach manifolds. We shall not try to be exhaustive or
as general as possible in this overview.
A Riemannian manifold (M, g) is a pair of a smooth (or at least C 1 , respectively
2
C for defining curvature) manifold together with a metric g: a family of positive-
definite bilinear forms gx on each tangent space Tx M. The metric is a generalization
of the Euclidean inner product on Rn and gx depends in a smooth way on the point
x ∈ M in the manifold. The metric can be used to measure angles and lengths of
tangent vectors, so we can define the length of a piecewise C 1 curve γ : [a, b] → M
as  b
l(γ ) = gγ (t) (γ  (t), γ  (t)) dt.
a

This length functional induces the distance function

d(x, y) = inf l(γ ) (F.1)

γ

on M, where the infimum is taken over all piecewise C 1 curves γ connecting the
points x and y. This turns M into a metric space.
Simple examples of Riemannian manifolds are Rn with the standard Euclidean
inner product and the sphere S n−1 ⊂ Rn with the induced metric on its tangent
bundle. Due to the Nash embedding theorem, any C k≥3 Riemannian manifold can
actually be realized as a submanifold of Rn equipped with the induced metric.
Each Riemannian manifold (M, g) has an associated linear connection, or, covari-
ant derivative ∇ on the tangent bundle TM. This so-called Levi-Civita connection is
uniquely defined by the requirements that it is torsion-free and compatible with the
metric, i.e.

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 179

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
180 Appendix F: Riemannian Geometry

∇ X Y − ∇Y X = [X, Y ] and X g(Y, Z ) = g(∇ X Y, Z ) + g(Y, ∇ X Z )

for all smooth vector fields X, Y, Z on M. The connection is given in local coordinates
x i by the Christoffel symbols  ijk ,

∇∂ j ∂k =  ijk ∂i ,

where we used the Einstein summation convention for the repeated index i. The
connection can be extended to the tensor bundle of M so that it satisfies the Leibniz
rule.
A connection, more generally on a vector bundle π : E → M, can also be viewed
as a choice of a horizontal subbundle in TE. There is a naturally defined vertical
subbundle Vert(E) ⊂ TE where Vert(E)ξ = Tξ E x for ξ ∈ E x = π −1 (x). A
horizontal bundle Hor(E) is any subbundle complementary to the vertical bundle,
so
TE = Hor(E) ⊕ Vert(E).

This definition of a connection is related to the definition via the covariant derivative.
The horizontal bundle precisely corresponds to the tangent plane to a section s of E
that is flat at a given point x ∈ M:
 
Hor(E)s(x) = Im Ds(x) ⇐⇒ (∇• s)(x) = 0.

The Levi-Civita connection induces two important concepts: the geodesic flow
and parallel transport. Intuitively, the geodesic flow says how to follow a straight line
from an initial point along a given direction, while parallel transport defines how to
keep a tangent vector fixed while carrying it along a path.8 Both maps are defined
in local coordinates as solutions of (subtly different) differential equations involving
the Christoffel symbols.
The geodesic flow ϒ t is a flow on the tangent bundle TM and defined in local
coordinates x i by
ẋ i = vi ,
(F.2)
v̇i = − ijk (x) v j vk .

Here, the vi denote the induced additional coordinates on the tangent bundle. This
geodesic flow need not be complete, that is, defined for all times. However, by the
Hopf–Rinow theorem, the geodesic flow is complete if and only if M is complete
as a metric space with respect to (F.1). In the following we shall assume that M is
complete to simplify the exposition.
If we restrict the geodesic flow map to the tangent space T p M at a fixed point p ∈
M and to time t = 1, and finally project onto M, then we obtain the exponential map

8 If the path is a geodesic, then parallel transport carries the initial velocity vector to the velocity
vector along the entire path.
Appendix F: Riemannian Geometry 181

exp p = π ◦ ϒ 1 |T p M : T p M → M.

We have D exp p (0 p ) = 1T p M , so by the inverse function theorem, exp p is a local

diffeomorphism at 0 p . The local inverse φx = exp−1x of the exponential map can be
viewed as a coordinate chart since T p M ∼ = Rn isometrically. An explicit identifi-
cation would require a choice of orthonormal basis in T p M, which we shall refrain
from.
Such coordinates are called normal coordinates, and locally around the point p
these coordinates make M resemble Rn as close as possible, in the sense that the
metric at p in these coordinates is equal to the Euclidean metric and the Christoffel
symbols are zero. The exponential map is only a local diffeomorphism, and generally
there is a maximum radius r > 0 such that exp p : B(0; r ) ⊂ T p M → M is a
diffeomorphism onto its image. This is called the injectivity radius rinj ( p) of M at
the point p. The global injectivity radius of M is then defined as

rinj (M) = inf rinj ( p).

p∈M

If M is noncompact then this global injectivity radius need not be positive. The
shortest path from p ∈ M to any point x within distance rinj ( p) is uniquely realized
by one geodesic curve. In normal coordinates these curves are rays emanating from
the origin. That is, let v = exp−1
p (x) and γ (t) = exp p (t v) with t ∈ [0, 1], then
d( p, x) = l(γ ) = v.
Let γ : [a, b] → M be a C 1 curve, then parallel transport is a linear isometry (i.e.
it preserves the metric g)

(γ ) : Tγ (a) M → Tγ (b) M (F.3)

between the tangent spaces at the endpoints. We use the notation (γ |at ) for parallel
transport along a part of the curve. Parallel transport is defined in local coordinates
x i by the differential equation

d
(γ |at )i = − ijk (γ (t)) γ  (t) j (γ |at )k with (γ |aa ) = 1. (F.4)
dt

In (F.4) the x i are local coordinates around the point γ (t) with additional induced
coordinates ∂i on the tangent bundle. The representation (γ |at )i is defined by
(γ |at ) = (γ |at )i ∂i . Put more abstractly, parallel transport defines a horizontal
extension of a vector v ∈ Tγ (a) M to a section of the pullback bundle γ ∗ (TM), that
is, a vector field v(t) defined along γ (t), which has covariant derivative zero.
On a Riemannian manifold there is the concept of curvature. A manifold is flat,
i.e. it has zero curvature, if it is (locally) isometric to Rn . The Riemann curvature R
measures non-flatness on an infinitesimal level. It is given by

R(X, Y ) Z = ∇ X ∇Y Z − ∇Y ∇ X Z − ∇[X,Y ] Z ,
182 Appendix F: Riemannian Geometry

which measures how much the direction of a vector Z changes when parallel trans-
porting it around an infinitesimal loop spanned by the directions X, Y . There is a
relation between the curvature and parallel transport that is important to us. If we
consider holonomy, that is, parallel transport along a closed loop γ , then the deficit
(γ ) − 1 is (heuristically put) equal to the curvature form R integrated over any
surface enclosed by γ . This relation can be seen as an application of Stokes’ the-
orem and the differential statement is that the curvature R is the generator of the
infinitesimal holonomy group [AS53, RW06].
References

[Ale61] V.M. Alekseev, An estimate for the perturbations of the solutions of ordinary differential
equations. Vestnik Moskov. Univ. Ser. I Mat. Mech. (2), 28–36 (1961).
[AMR88] R. Abraham, J.E. Marsden, T. Ratiu, Manifolds, tensor analysis, and applications, 2nd
edn, in Applied Mathematical Sciences, vol. 75 (Springer, New York, 1988)
[Ano69] D.V. Anosov, Geodesic flows on closed Riemann manifolds with negative curvature,
in Proceedings of the Steklov Institute of Mathematics, No. 90, (Translated from the
Russian by S. Feder, American Mathematical Society, Providence, R.I., 1967), 1969.
[APS02] B. Aulbach, C. Pötzsche, S. Siegmund, A smoothness theorem for invariant fiber bundles.
J. Dyn. Differ. Equ. 14(3), 519–547 (2002)
[AS53] W. Ambrose, I.M. Singer, A theorem on holonomy. Trans. Am. Math. Soc. 75, 428–443
(1953)
[Att94] O. Attie, Quasi-isometry classification of some manifolds of bounded geometry. Math.
Z. 216(4), 501–527 (1994)
[Bro+09] H.W. Broer, M.C. Ciocci, H. Hanßman, A. Vanderbauwhede, Quasi-periodic stability
of normally resonant tori. Physica D 238, 309–318 (2009)
[BLZ98] P.W. Bates, K. Lu, C. Zeng, Existence and persistence of invariant manifolds for semi-
flows in Banach space. Mem. Am. Math. Soc. 135(645), viii+129 (1998).
[BLZ99] P.W. Bates, K. Lu, C. Zeng, Persistence of overflowing manifolds for semiflow. Comm.
Pure Appl. Math. 52(8), 983–1046 (1999)
[BLZ08] P.W. Bates, K. Lu, C. Zeng, Approximately invariant manifolds and global dynamics of
spike states. Invent. Math. 174(2), 355–433 (2008)
[Bre81] V.N. Brendelev, On the realization of constraints in nonholonomic mechanics. J. Appl.
Math. Mech. 45(3), 481–487 (1981)
[Car26] É. Cartan, Les groupes d’holonomie des espaces généralisés. Acta Math. 48, 1–42 (1926)
[Cot11] É. Cotton, Sur les solutions asymptotiques des équations différentielles. Ann. Sci. École
Norm. Sup. 28(3), 473–521 (1911)
[DLS06] A. Delshams, R. de la Llave, T.M. Seara, Orbits of unbounded energy in quasi-periodic
perturbations of geodesic flows. Adv. Math. 202(1), 64–188 (2006)
[DK00] J.J. Duistermaat, J.A.C. Kolk, Lie groups. Universitext (Springer, Berlin, 2000)
[Dui76] J.J. Duistermaat, in “Stable Manifolds” Unpublished preprint, Utrecht University,
Department of Mathematics (1976)
[Eic91] J. Eichhorn, The Banach manifold structure of the space of metrics on noncompact
manifolds. Differ. Geom. Appl. 1(2), 89–108 (1991)
[Fen72] N. Fenichel, Persistence and smoothness of invariant manifolds for flows. Indiana Univ.
Math. J. 21, 193–226 (1971/1972).

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 183

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
© Atlantis Press and the author 2013
184 References

[Fen74] N. Fenichel, Asymptotic stability with rate conditions. Indiana Univ. Math. J. 23, 1109–
1137 (1973/74).
[Fen79] N. Fenichel, Geometric singular perturbation theory for ordinary differential equations.
J. Differ. Equ. 31(1), 53–98 (1979)
[GG73] M. Golubitsky, V. Guillemin, Stable mappings and their singularities, in Graduate Texts
in Mathematics, vol. 14 (Springer, New York, 1973)
[GHL04] S. Gallot, D. Hulin, J. Lafontaine, Riemannian Geometry. Universitext, 3rd edn.
(Springer, Berlin, 2004).
[GK02] S. Gudmundsson, E. Kappos, On the geometry of tangent bundles. Expo. Math. 20(1),
1–41 (2002)
[Guc75] J. Guckenheimer, Isochrons and phaseless sets. J. Math. Biol. 1(3), 259–273 (1974/75).
[Had01] J. Hadamard, Sur l’itération et les solutions asymptotiques des equations différentielles.
Bull. Soc. Math. Fr. 29, 224–228 (1901)
[Hal61] J.K. Hale, Integral manifolds of perturbed differential systems. Ann. Math. 73(2), 496–
531 (1961)
[Hal69] J.K. Hale, Ordinary differential equations, in Pure and Applied Mathematics, vol. XXI
(Wiley-Interscience, New York, 1969)
[Has94] B. Hasselblatt, Regularity of the Anosov splitting and of horospheric foliations. Ergodic
Theory Dyn. Syst. 14(4), 645–666 (1994)
[HL06] À. Haro, R. de la Llave, Manifolds on the verge of a hyperbolicity breakdown. Chaos
16(1), 013120–013128 (2006)
[Hen81] D. Henry, in Geometric Theory of Semilinear Parabolic Equations, Lecture Notes in
Mathematics, vol. 840 (Springer, Berlin, 1981)
[Hil01] D. Hilbert, Ueber Flächen von constanter Gaussscher Krümmung. Trans. Amer. Math.
Soc. 2(1), 87–99 (1901)
[Hir76] M.W. Hirsch, Differential topology, in Graduate Texts in Mathematics, No. 33 (Springer,
New York, 1976).
[Hop66] F.C. Hoppensteadt, Singular perturbations on the infinite interval. Trans. Am. Math.
Soc. 123, 521–535 (1966)
[Hör03] L. Hörmander, The analysis of linear partial differential operators, in I, Classics in
Mathematics, Distribution theory and Fourier analysis. Reprint of the second (1990)
edition (Springer, Berlin, 2003).
[HP70] M.W. Hirsch, C.C. Pugh, Stable Manifolds and Hyperbolic Sets, Global analysis (Proc.
Sympos. Pure Math., Vol. XIV, Berkeley, California, 1968). Amer. Math. Soc., Provi-
dence, R.I., 133–163 (1970).
[HPS77] M.W. Hirsch, C.C. Pugh, M. Shub, Invariant manifolds, Lecture Notes in Mathematics,
vol. 583 (Springer, Berlin, 1977)
[HW99] B. Hasselblatt, A. Wilkinson, Prevalence of non-Lipschitz Anosov foliations. Ergodic
Theory Dyn. Syst. 19(3), 643–656 (1999)
[Irw70] M.C. Irwin, On the stable manifold theorem. Bull. London Math. Soc. 2, 196–198 (1970)
[Irw72] M.C. Irwin, On the smoothness of the composition map. Quart. J. Math. Oxford Ser.
23(2), 113–133 (1972)
[Jon95] C.K.R.T. Jones, Geometric singular perturbation theory, in Dynamical Systems
(Springer, Berlin, 1995), pp. 44–118 (Montecatini Terme, Lecture Notes in Math., vol.
1609, 1994).
[Jos08] J. Jost, Riemannian geometry and geometric analysis. Universitext, 5th edn. (Springer,
Berlin, 2008).
[JS99] A.D. Jones, S. Shkoller, Persistence of invariant manifolds for nonlinear PDEs. Stud.
Appl. Math. 102(1), 27–67 (1999)
[Kap99] T.J. Kaper, An introduction to geometric methods and dynamical systems theory for
singular perturbation problems, in Proceedings of Symposia in Applied Mathematics,
Analyzing Multiscale Phenomena using Singular Perturbation Methods, ed. by M.D.
Baltimore (American Mathematical Society, Providence, 1999), pp. 85–131
References 185

[Kar81] A.V. Karapetian, On realizing nonholonomic constraints by viscous friction forces and
Celtic stones stability. J. Appl. Math. Mech. 45(1), 42–51 (1981)
[Kli95] W.P.A. Klingenberg, Riemannian geometry, 2nd ed., in de Gruyter Studies in Mathe-
matics, vol. 1 (Walter de Gruyter & Co., Berlin, 1995).
[KN90] V.V. Kozlov, A.I. Neı̆shtadt, Realization of holonomic constraints. Prikl. Mat. Mekh.
54(5), 858–861 (1990)
[Koz92] V.V. Kozlov, On the realization of constraints in dynamics. Prikl. Mat. Mekh. 56(4),
692–698 (1992)
[Lan95] S. Lang, Differential and Riemannian manifolds, in Graduate Texts in Mathematics, 3rd
edn, vol. 160 (Springer, New York, 1995)
[LL69] V. Lakshmikantham, S. Leela, Differential and integral inequalities: Theory and appli-
cations, vol. I, in Ordinary Differential Equations (Academic Press, New York, 1969),
Mathematics in Science and Engineering, vol. 55-I.
[LM95] A.D. Lewis, R.M. Murray, Variational principles for constrained systems: theory and
experiment. Int. J. Non-Linear Mech. 30(6), 793–815 (1995)
[Lya07] A.M. Lyapunov, Problème général de la stabilité du mouvement. Ann. Fac. Sci. Toulouse
Sci. Math. Sci. Phys. 9(2), 203–474 (1907).
[Lya92] A.M. Lyapunov, The general problem of the stability of motion. Int. J. Control 55(3),
521–790 (1992). Trans. by A.T. Fuller, Édouard Davaux’s French translation (1907) of
the 1892 Russian original.
[Mañ78] R. Mañé, Persistent manifolds are normally hyperbolic. Trans. Amer. Math. Soc. 246,
261–283 (1978)
[Muj86] J. Mujica, Complex analysis in Banach spaces, in North-Holland Mathematics Studies,
vol. 120 (North-Holland Publishing Co., Amsterdam, 1986). Holomorphic functions
and domains of holomorphy in finite and infinite dimensions, Mathematical Notes, 107
[Pal75] J. Kenneth, Palmer, linearization near an integral manifold. J. Math. Anal. Appl. 51,
243–255 (1975)
[Per29] O. Perron, Über Stabilität und asymptotisches Verhalten der Integrale von Differential-
gleichungssystemen. Math. Z. 29(1), 129–160 (1929)
[Per30] O. Perron, Die Stabilitätsfrage bei Differentialgleichungen. Math. Z. 32(1), 703–728
(1930)
[Poi92] H. Poincaré, Les méthodes nouvelles de la mécanique céleste (Gauthier-Villars, Paris,
1892). Tome I (Les Grands Classiques).
[PS04] C. Pötzsche, S. Siegmund, C m -smoothness of invariant fiber bundles. Topol. Methods
Nonlinear Anal. 24(1), 107–145 (2004)
[PT77] J. Palis, F. Takens, Topological equivalence of normally hyperbolic dynamical systems.
Topology 16(4), 335–345 (1977)
[PT83] J. Palis, F. Takens, Stability of parametrized families of gradient vector fields. Ann.
Math. 118(2, 3), 383–421 (1983).
[Rob68] W.J. Robbin, On the existence theorem for differential equations. Proc. Amer. Math.
Soc. 19, 1005–1006 (1968)
[Roe88] J. Roe, An index theorem on open manifolds. I, II, J. Diff. Geom. 27(1), 87–113, 115–136
(1988).
[RU57] H. Rubin, P. Ungar, Motion under a strong constraining force. Commun. Pure Appl.
Math. 10, 65–87 (1957)
[RW06] H. Reckziegel, E. Wilhelmus, How the curvature generates the holonomy of a connection
in an arbitrary fibre bundle. Results Math. 49(3–4), 339–359 (2006)
[Sak90] K. Sakamoto, Invariant manifolds in singular perturbation problems for ordinary differ-
ential equations. Proc. Roy. Soc. Edinburgh Sect. A 116(1–2), 45–78 (1990)
[Sak94] K. Sakamoto, Smooth linearization of vector fields near invariant manifolds. Hiroshima
Math. J. 24(2), 331–355 (1994)
[Sas58] S. Sasaki, On the differential geometry of tangent bundles of Riemannian manifolds.
Tôhoku Math. J. 10(2), 338–354 (1958)
186 References

[Sch01] T. Schick, Manifolds with boundary and of bounded geometry. Math. Nachr. 223, 103–
120 (2001)
[Shu92] M.A. Shubin, Spectral theory of elliptic operators on noncompact manifolds. Astérisque
5(207), 35–108 (1992), in Méthodes semi-classiques, vol. 1 (Nantes, 1991).
[Šil69] L.P. Šil nikov, A certain new type of bifurcation of multidimensional dynamic systems.
Dokl. Akad. Nauk SSSR 189, 59–62 (1969).
[Tak80] F. Takens, Motion under the influence of a strong constraining force, Global theory of
dynamical systems, in Proceedings of the International Conference, Northwestern Uni-
versity, Evanston, Ill., Lecture Notes in Mathematics, 1979, vol. 819 (Springer, Berlin,
1980), pp. 425–445.
[Uze+02] T. Uzer, C. Jaffé, J. Palacián, P. Yanguas, s Wiggin. The geometry of reaction dynamics.
Nonlinearity 15(4), 957–992 (2002)
[Van89] A. Vanderbauwhede, Centre manifolds, normal forms and elementary bifurcations,
Dynamics reported, in Dyn. Report. Ser. Dynam. Systems Appl., vol. 2 (Wiley, Chich-
ester, 1989), pp. 89–169.
[Ver05] F. Verhulst, Methods and applications of singular perturbations, in Texts in Applied Math-
ematics, vol. 50 (Springer, New York, 2005) (Boundary layers and multiple timescale
dynamics).
[Str79] S.J. van Strien, Center manifolds are not C ∞ . Math. Z. 166(2), 143–145 (1979)
[VG87] A. Vanderbauwhede, S.A. van Gils, Center manifolds and contractions on a scale of
Banach spaces. J. Funct. Anal. 72(2), 209–224 (1987)
[Whi36] H. Whitney, Differentiable manifolds. Ann. Math. 37(2, 3), 645–680 (1936).
[Wig94] S. Wiggins, Normally hyperbolic invariant manifolds in dynamical systems, in Applied
Mathematical Sciences ed. by G. Haller, I. Mezić, vol. 105 (Springer, New York, 1994)
[Yi93] Y. Yi, A generalized integral manifold theorem. J. Diff. Equ. 102(1), 153–187 (1993)
[Zei86] E. Zeidler, Nonlinear functional analysis and its applications, in Fixed-point theorems
(Springer, New York, 1986). (Translated from the German by P.R., Wadsack)
Index

A, 94 b (approximate solution distance), 101

Bq, 32 dx ; dy ; dx0 , 112
Bqb ðI : X Þ, 104 dM, 42
Bqg ðI; YÞ, 104 ef, 31, see also continuity modulus
Cbk g (tubular neighborhood size), 58, 76
definition in bounded geometry, 42 l, 109, 136
function space, 31 m (mollifier size), 62, 96
topology, 30 q, 22, 33, see also exponential growth rate
k qa (restriction operator), 130
Cb;u , 31
qX, qY,, 96
Cf,a, 33 r (approximation parameter), 62, 76
Hor, Vert, 180 f (vector field change), 96
I, 31, 104 exp, 39, 181
J, 31, 128 ~f , 95
K, 45
hx, 51
Ll;k , 166 r, 22, 76, see also r-normal hyperbolicity
Mx,d, 51 rinj, 36, 181, see also injectivity radius
N (normal bundle), 59 vr, hr, 76
Pa,b, 163 ~vX , 95
S, 114, 138 kkq ; dq , 32
T, 17, 104
kkq2 ;q1 , 126
Tb,a, 111, 134
TX , TY, 104 , 32, 39
T Bqg ðI; Y Þ; T Bqb ðI; X Þ, 115
e e
Ux , 116 A
X, 31 Ambient manifold, 3, 19, 87
D, 110, see also formal derivative Approximate solution, 101
e
C
Christoffel symbols, 32, 41, 180 B
fiber bundle section, 31 Bounded geometry, 19, 35
Uy, 98 definition for manifold, 36
P, 32, 181, see also parallel transport definition for vector bundle, 45
~ 116
P, submanifold of, 54
Wx, 98 Bundle norm, 166
H, 108

J. Eldering, Normally Hyperbolic Invariant Manifolds, Atlantis Series 187

in Dynamical Systems 2, DOI: 10.2991/978-94-6239-003-4,
Ó Atlantis Press and the author 2013
188 Index

C Injectivity radius, 36, 87, 181

Center manifold, 5 Invariant fibration, see stable (unstable)
Christoffel symbols, 41, 180 fibration
Comparison
of methods, 15
of results, 14, 78 L
of topologies, 30, 135 Lyapunov exponent, 13, 22, see also expo-
Continuity modulus, 31 nential growth rate
in bounded geometry, 42, 43, 169 Lyapunov-Perron method,
Contraction operator, 17, 104 see Perron method
Convolution smoothing, 62
Coordinate transition map, 40
graph change under, 63, 65 M
on a Banach manifold, 129 Mollifier function, 62
Curvature, 36, 47, 181

N
D Nemytskii operator, 157
Dynamical system NHIM, 2, see also normally hyperbolic
continuous, 1 invariant manifold
discrete, 1, 21 Non-autonomous system, 24, 141
Noncompactness, 10, 19, 24, 82
Nonholonomic system, 12
E Normal coordinates, 39, 129, 181
Exponential growth numbers, 33, 95 Normal hyperbolicity, 21
Exponential growth rate, 6, 21, 33, 161 history, 13
r-normal hyperbolicity, 22, 61
Normally hyperbolic invariant manifold, 1
F definition, 21
Fiber contraction theorem, 173 parameter dependence, 142
Fixed point, 2 persistence, 3, 76
Forced smoothness, 23, 78
Formal tangent bundle, 114
topology, 115 O
Overflow invariance, 29, 143
a priori, 144
G persistence, 145
Geodesic flow, 41, 180
Graph transform, 16
P
Parallel transport, 47, 181
H differential equation, 47, 181
Hölder continuity, 23, 33 of a frame, 118
Higher order derivative, 18, 136, 161 Partition of unity (uniform), 46
chain rule, 163 Perron method, 14, 16, 106
on a Riemannian manifold, 165 Persistence, see normally hyperbolic
Holonomy, 47, 119 invariant
Hyperbolic fixed point, 2, 15 manifold
non-, 9, 85

I
Immersed submanifold, 25 Q
Implicit function theorem, 152 Quasi-isometry, 58
Index 189

R T
Riemannian manifold, 179 Trivial bundle, 69, 76

S U
Shil’nikov bifurcation, 28 Uniform tubular neighborhood, 58
Smoothness, 3, 7, 18, 109 Uniformly embedded submanifold, 52, 86
forced, see forced smoothness Uniformly immersed submanifold, 51
loss of, 62, 77, 95 Uniformly locally finite cover, 45
non-C?, 6, 9, 23 of a submanifold, 54
of a flow, 155 Unstable manifold, see stable manifold
Spectral gap condition, 4, 6, 22
Stable (unstable) fibration, 4, 148
Stable (unstable) manifold, 2, 15, 147