Introduction To The H-Principle, 2ed, Cieliebak, Eliashberg, GSM 239, 2024

GRADUATE STUDIES
I N M AT H E M AT I C S 239
Introduction to
the h-Principle
Second Edition
K. Cieliebak
Y. Eliashberg
N. Mishachev
Introduction to
the h-Principle
Second Edition
GRADUATE STUDIES
I N M AT H E M AT I C S 239
Introduction to
the h-Principle
Second Edition
K. Cieliebak
Y. Eliashberg
N. Mishachev
EDITORIAL COMMITTEE
Matthew Baker
Marco Gualtieri
Gigliola Staffilani (Chair)
Jeff A. Viaclovsky
Rachel Ward
2020 Mathematics Subject Classification. Primary 58Axx.
For additional information and updates on this book, visit

www.ams.org/bookpages/gsm-239
Library of Congress Cataloging-in-Publication Data

Names: Cieliebak, Kai, 1966– author. | Eliashberg, Y., 1946– author. | Mishachev, N. (Nikolai
M.), 1952– author.
Title: Introduction to the h-principle / K. Cieliebak, Y. Eliashberg, N. Mishachev.
Description: Second edition. | Providence, Rhode Island : American Mathematical Society, [2023]
| Series: Graduate studies in mathematics, 1065-7339 ; volume 239 | Includes bibliographical
references and index.
Identifiers: LCCN 2023043257 | ISBN 9781470461058 (hardback) | ISBN 9781470476175 (paper-
back) | ISBN 9781470476182 (ebook)
Subjects: LCSH: Geometry, Differential. | Differentiable manifolds. | Differential equations–
Numerical solutions.
Classification: LCC QA641 .E62 2024 | DDC 516.3/6–dc23/eng/20231023
LC record available at https://lccn.loc.gov/2023043257
Copying and reprinting. Individual readers of this publication, and nonprofit libraries acting
for them, are permitted to make fair use of the material, such as to copy select pages for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for permission
to reuse portions of AMS publication content are handled by the Copyright Clearance Center. For
more information, please visit www.ams.org/publications/pubpermissions.
Send requests for translation rights and licensed reprints to reprint-permission@ams.org.

c 2024 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at https://www.ams.org/
10 9 8 7 6 5 4 3 2 1 29 28 27 26 25 24
To
Vladimir Igorevich Arnold
who introduced us to the world of singularities
and
Misha Gromov
who taught us how to get rid of them
Contents
Preface to the Second Edition xiii
Preface to the First Edition xv
Intrigue 1
Part 1. Holonomic Approximation
Chapter 1. Jets and Holonomy 9

§1.1. Maps and sections 9
§1.2. Coordinate definition of jets 10
§1.3. Invariant definition of jets 11
§1.4. The space X (1) 12
§1.5. Holonomic sections of the jet space X (r) 13
§1.6. Geometric representation of sections of X (r) 14
§1.7. Holonomic splitting 14
Chapter 2. Thom Transversality Theorem 17

§2.1. Generic properties and transversality 17
§2.2. Stratified sets and polyhedra 18
§2.3. Thom Transversality Theorem 19
Chapter 3. Holonomic Approximation 25

§3.1. Main theorem 25
§3.2. Holonomic approximation over a cube 28
§3.3. Holonomic extension 28
vii
viii Contents
§3.4. Gluing Lemma 30

§3.5. Proof of Theorem 3.2.1 33
§3.6. Proof of the Gluing Lemma 35
§3.7. Parametric holonomic approximation 40
§3.8. Foliated holonomic approximation 41
§3.9. Refinement of the Holonomic Approximation Theorem 43
Chapter 4. Applications 45
§4.1. Functions without critical points 45
§4.2. Smale’s sphere eversion 46
§4.3. Approximate integration of tangential homotopies 48
§4.4. Open manifolds 50
§4.5. Directed embeddings of open manifolds 53
§4.6. Directed embeddings of closed manifolds 55
§4.7. Approximation of differential forms by closed forms 56
Chapter 5. Multivalued Holonomic Approximation 59
§5.1. Multifolds 59
§5.2. Holonomic approximation by multivalued sections 63
Part 2. Differential Relations and Gromov’s h-principle

Chapter 6. Differential Relations 69
§6.1. What is a differential relation? 69
§6.2. Open and closed differential relations 71
§6.3. Formal and genuine solutions to a differential relation 72
§6.4. Extension problem 72
§6.5. Approximate solutions to systems of differential equations 73
Chapter 7. Homotopy Principle 75
§7.1. Philosophy of the h-principle 75
§7.2. Different flavors of the h-principle 78
Chapter 8. Open Diff V -invariant Differential Relations 81
§8.1. Natural fibrations 81
§8.2. Diff V -invariant differential relations 85
§8.3. Local and global h-principle for open Diff V -invariant relations 85
Chapter 9. Applications to Closed Manifolds 89
§9.1. Microextension trick 89
Contents ix
§9.2. Smale–Hirsch h-principle 89

§9.3. Sections transverse to a distribution 91
§9.4. Multivalued h-principle 93
Chapter 10. Foliations 95
§10.1. Definition and examples 95
§10.2. Haefliger structures and h-principle for foliations 98
§10.3. Foliated h-principle 102
Part 3. Singularities and Wrinkling

Chapter 11. Singularities of Smooth Maps 109
§11.1. Thom–Boardman singularities 109
§11.2. The Morse Lemma 112
§11.3. Tangency of a submanifold to a foliation 114
§11.4. Fibered character of Σ1 -singularities 115
§11.5. An -singularities of functions 118
§11.6. Morin’s normal forms 119
§11.7. More about the geometry of maps with Σ1,...,1 -singularities 123
Chapter 12. Wrinkles 127
§12.1. Standard wrinkles 127
§12.2. Properties of standard wrinkles 129
Chapter 13. Wrinkled Submersions 135
§13.1. Definitions 135
§13.2. Main results 138
§13.3. Wrinkled immersion associated with a multifold 142
§13.4. Chopping wrinkles 144
§13.5. Corrugation 148
Chapter 14. Folded Solutions to Differential Relations 155
§14.1. Surgery of singularities 155
§14.2. Folded solutions to differential relations 164
Chapter 15. The h-principle for Sharp Wrinkled Embeddings 169
§15.1. Sharp wrinkled embeddings 169
§15.2. Integration of tangential rotations 176
§15.3. Preliminary steps in the proof 177
§15.4. Reduction to the main lemma 179
x Contents
§15.5. Proof of the main lemma 181

§15.6. The h-principle for folded embeddings 187
Chapter 16. Igusa Functions 193

§16.1. Leafwise Igusa functions and their formal analogues 193
§16.2. Formal extension of framed ζ-FLIFs 198
§16.3. Constructing a locally holonomic ζ-FLIF 199
§16.4. Proof of the main theorems 208
§16.5. Applications to pseudo-isotopy theory 213
Part 4. The Homotopy Principle in Symplectic Geometry
Chapter 17. Symplectic and Contact Basics 219

§17.1. Linear symplectic and complex geometries 219
§17.2. Symplectic and complex manifolds 224
§17.3. Symplectic stability 229
§17.4. Contact manifolds 232
§17.5. Contact stability 237
§17.6. Lagrangian and Legendrian submanifolds 240
§17.7. Hamiltonian and contact vector fields 241
§17.8. Characteristic foliations 243
Chapter 18. Symplectic and Contact Structures on Open Manifolds 245

§18.1. Classification problem for symplectic and contact structures 245
§18.2. Symplectic structures on open manifolds 246
§18.3. Contact structures on open manifolds 248
§18.4. Two-forms of maximal rank on odd-dimensional manifolds 249
Chapter 19. Symplectic and Contact Structures on Closed Manifolds 251

§19.1. Symplectic structures on closed manifolds 251
§19.2. Contact structures on closed manifolds 253
§19.3. Folded symplectic and contact structures 255
Chapter 20. Embeddings into Symplectic and Contact Manifolds 259

§20.1. Isosymplectic embeddings 259
§20.2. Equidimensional isosymplectic immersions 268
§20.3. Isocontact embeddings 271
§20.4. Subcritical isotropic embeddings 276
Contents xi
Chapter 21. Microflexibility and Holonomic R-approximation 279

§21.1. Local integrability 279
§21.2. Homotopy extension property for formal solutions 281
§21.3. Microflexibility 281
§21.4. Theorem on holonomic R-approximation 283
§21.5. Local h-principle for microflexible Diff V -invariant relations 283
Chapter 22. First Applications of Microflexibility 285
§22.1. Subcritical isotropic immersions 285
§22.2. Maps transverse to a contact structure 286
Chapter 23. Microflexible A-invariant Differential Relations 289
§23.1. A-invariant differential relations 289
§23.2. Local h-principle for microflexible A-invariant relations 290
Chapter 24. Further Applications to Symplectic Geometry 293
§24.1. Legendrian and isocontact immersions 293
§24.2. Generalized isocontact immersions 294
§24.3. Lagrangian immersions 296
§24.4. Isosymplectic immersions 297
§24.5. Generalized isosymplectic immersions 299
Part 5. Convex Integration

Chapter 25. One-Dimensional Convex Integration 303
§25.1. Example 303
§25.2. Convex hulls and ampleness 304
§25.3. Main lemma 305
§25.4. Proof of the main lemma 306
§25.5. Parametric version of the main lemma 311
§25.6. Proof of the parametric version of the main lemma 312
Chapter 26. Homotopy Principle for Ample Differential Relations 317
§26.1. Ampleness in coordinate directions 317
§26.2. Iterated convex integration 318
§26.3. Principal subspaces and ample differential relations in X (1) 320
§26.4. Convex integration of ample differential relations 321
Chapter 27. Directed Immersions and Embeddings 323
§27.1. Criterion of ampleness for directed immersions 323
xii Contents
§27.2. Directed immersions into almost symplectic manifolds 324

§27.3. Directed immersions into almost complex manifolds 325
§27.4. Directed embeddings 326
Chapter 28. First Order Linear Differential Operators 331
§28.1. Formal inverse of a linear differential operator 331
§28.2. Homotopy principle for D-sections 332
§28.3. Nonvanishing D-sections 333
§28.4. Systems of linearly independent D-sections 334
§28.5. Two-forms of maximal rank on odd-dimensional manifolds 336
§28.6. One-forms of maximal rank on even-dimensional manifolds 338
Chapter 29. Nash–Kuiper Theorem 341
§29.1. Isometric immersions and short immersions 341
§29.2. Nash–Kuiper theorem 342
§29.3. Decomposition of a metric into a sum of primitive metrics 343
§29.4. Approximation theorem 344
§29.5. One-dimensional Approximation Theorem 345
§29.6. Adding a primitive metric 346
§29.7. End of the proof of the Approximation Theorem 348
§29.8. Proof of the Nash–Kuiper theorem 349
Bibliography 351
Index 359
Preface to the
Second Edition
The book is significantly expanded for its second edition, and several old
parts are rewritten. The main addition to the original book is Part 3 en-
titled “Singularities and Wrinkling”, which is devoted to the method of
wrinkling and its applications. In particular, we discuss there the construc-
tion of maps with prescribed fold singularities and prove a generalized form
of Igusa’s results about families of functions with generalized Morse critical
points. The main ingredient in the proof is the multivalued holonomic ap-
proximation theorem, which is introduced in the new Chapter 5. We also
added a new chapter (Chapter 10) on foliations, and the “foliated” language
is now used throughout the book. Several other chapters and sections have
been significantly rewritten, including Chapter 3 on holonomic approxima-
tion, Chapter 20 on symplectic and contact embeddings, Section 2.3 on the
Thom Transversality Theorem, and Section 8.1 on natural fibrations.
In the more than 20 years since the first edition of this book, many more
instances of the h-principle were discovered. Especially interesting new man-
ifestations of flexibility phenomenon were found in symplectic and contact
geometry. It became clear to us that it is impossible to address new sym-
plectic geometric applications of the h-principle in the format of the current
book. Hence, we decided to leave the symplectic part of this book mostly as
is and are currently working on a separate book [CEM25] devoted to new
developments in symplectic and contact flexibility.
Throughout the rest of the book we made numerous small changes and
corrections of typos and errors. We are grateful to many readers who over
the years notified us of various typos and errors in the book. We thank
Daniel Álvarez-Gavela, Álvaro del Pino Gómez, and Zehan Hu for their
comments on the manuscript of the current edition.
xiii
Preface to the
First Edition
A partial differential relation R is any condition imposed on the partial

derivatives of an unknown function. A solution of R is any function which
satisfies this relation.
The classical partial differential relations, mostly rooted in physics, are usu-
ally described by (systems of) equations. Moreover, the corresponding sys-
tems of equations are mostly determined : the number of unknown functions
is equal to the number of equations. Given appropriate boundary condi-
tions, such a differential relation usually has a unique solution. In some
cases this solution can be found using certain analytical methods (potential
theory, Fourier method, and so on).
In differential geometry and topology one often deals with systems of partial
differential equations, as well as partial differential inequalities, which have
infinitely many solutions whatever boundary conditions are imposed. More-
over, sometimes solutions of these differential relations are C 0 -dense in the
corresponding space of functions or mappings. The systems of differential
equations in question are usually (but not necessarily) underdetermined. We
discuss in this book homotopical methods for solving this kind of differen-
tial relations. Any differential relation has an underlying algebraic relation
which one gets by substituting derivatives by new independent variables. A
solution of the corresponding algebraic relation is called a formal solution of
the original differential relation R. Its existence is a necessary condition for
the solvability of R, and it is a natural starting point for exploring R. Then
one can try to deform the formal solution into a genuine solution. We say
that the h-principle holds for a differential relation R if any formal solution
of R can be deformed into a genuine solution.
xv
xvi Preface to the First Edition
The notion of h-principle (under the name w.h.e.-principle) first appeared

in [Gr71] and [GE71]. The term “h-principle” was introduced and pop-
ularized by M. Gromov in his book [Gr86]. The h-principle for solutions
of partial differential relations exposed the soft/hard (or flexible/rigid) di-
chotomy for the problems formulated in terms of derivatives: a particular
analytical problem is soft or abides by the h-principle if its solvability is
determined by some underlying algebraic or geometric data. The softness
phenomena was first discovered in the 1950s by J. Nash [Na54] for isometric
C 1 -immersions, and by S. Smale [Sm58, Sm59] for differential immersions.
However, instances of soft problems appeared earlier (e.g., H. Whitney’s pa-
per [Wh37]). In the 1960s several new geometrically interesting examples
of soft problems were discovered by M. Hirsch, V. Poénaru, A. Phillips, S.
Feit, and other authors (see [Hi59], [Po66], [Ph67], [Fe69]). In his disser-
tation [Gr69], in the paper [Gr73], and later in his book [Gr86], Gromov
transformed Smale’s and Nash’s ideas into two powerful general methods
for solving partial differential relations: continuous sheaves (or the covering
homotopy) method and the convex integration method. The third method,
called removal of singularities, was first introduced and explored in [GE71].
There is an opinion that “the h-principle is the hardest part of Gromov’s

work to popularize” (see [Be00]). We have written our book in order to im-
prove the situation. We consider here two geometrical methods: holonomic
approximation, which is a version of the method of continuous sheaves, and
convex integration. We do not pretend to cover here the content of Gro-
mov’s book [Gr86], but rather we want to prepare and motivate the reader
to look for hidden treasures there. On the other hand, the reader interested
in applications will find that with a few notable exceptions (e.g., Lohkamp’s
theory [Lo95] of negative Ricci curvature and Donaldson’s theory [Do96]
of approximately holomorphic sections) most instances of the h-principle
which are known today can be treated by the methods considered in the
present book.
The first three parts1 of the book are devoted to a quite general theorem
about holonomic approximation of sections of jet-bundles and its applica-
tions. Given an arbitrary submanifold V0 ⊂ V of positive codimension, the
Holonomic Approximation Theorem allows us to solve any open differen-
tial relations R near a slightly perturbed submanifold V0 = h(V0 ) where
h : V → V is a C 0 -small diffeomorphism. Gromov’s h-principle for open
Diff V -invariant differential relations on open manifolds, his directed embed-
ding theorem, as well as some other results in the spirit of the h-principle
are immediate corollaries of the Holonomic Approximation Theorem.
1 In the current edition of the book these are the first, second, and fourth parts.
Preface to the First Edition xvii
The method for proving the h-principle based on the Holonomic Approx-
imation Theorem works well for open manifolds. Applications to closed
manifolds require an additional trick, called microextension. It was first
used by M. Hirsch in [Hi59]. The holonomic approximation method also
works well for differential relations which are not open, but microflexible.
The most interesting applications of this type come from symplectic geom-
etry. These applications are discussed in the third part of the book. For
convenience of the reader the basic notions of symplectic geometry are also
reviewed in that part of the book.
The fourth2 part of the book is devoted to convex integration theory. Gro-
mov’s convex integration theory was treated in great detail by D. Spring
in [Sp98]. In our exposition of convex integration we pursue a different
goal. Rather than considering the sophisticated advanced version of convex
integration presented in [Gr86], we explore only its simple version for first
order differential relations, similar to the first exposition of the theory by
Gromov in [Gr73]. Nevertheless, we prove here practically all the most
interesting corollaries of the theory, including the Nash–Kuiper theorem on
C 1 -isometric embeddings.
Let us list here some available books and survey papers about the h-principle.
Besides Gromov’s book [Gr86], these are Spring’s book [Sp98], Adachi’s
book [Ad93], Haefliger’s paper [Ha71], Poénaru’s paper [Po71], and, most
recently, Geiges’ notes [Ge01].
Acknowledgments. The book was partially written while the second au-
thor visited the Department of Mathematics of Stanford University, and the
first author visited the Mathematical Institute of Leiden University and the
Institute for Advanced Study at Princeton. The authors thank the host
institutions for their hospitality. While writing this book the authors were
partially supported by the National Science Foundation. The first author
also acknowledges the support of The Veblen Fund during his stay at the
IAS.
We are indebted to Ana Cannas da Silva, Hansjorg Geiges, Simon Gober-
stein, Dusa McDuff, and David Spring who read the preliminary version
of this book and corrected numerous misprints and mistakes. We are very
thankful to all the mathematicians who communicated to us their critical
remarks and suggestions.
2 In the second edition it is the fifth.

Intrigue
Examples.
A. Immersions. A smooth map f : V → W of an n-dimensional manifold

V into a q-dimensional manifold W , n ≤ q, is called an immersion if its
differential has the maximal rank n at every point. Two immersions are
called regularly homotopic if one can be deformed to the other through a
smooth family of immersions.
A1. For an immersion f : S 1 → R2 we denote by G(f ) its tangential degree,
i.e., the degree of the corresponding Gauss map S 1 → S 1 . Then two immer-
sions f, g : S 1 → R2 are regularly homotopic if and only if G(f ) = G(g);
see [Wh37] and Section 7.1.
A2. On the other hand, any two immersions S 2 → R3 are regularly homo-
topic; see [Sm58] and Section 4.2. In particular, the standard 2-sphere in
R3 can be inverted outside in through a family of immersions.
A3. Consider now pairs of immersions (f0 , f1 ) : D 2 → R2 which coincide

near the boundary circle ∂D 2 . What is the classification of such pairs up
to regular homotopy in this class? The answer turns out to be quite unex-
pected:
There are precisely two regular homotopy classes of such pairs. One is rep-
resented by the pair (j, j) where j is the inclusion D 2 → R2 , the second one
is represented by the pair (f, g) where the immersions f and g are shown in
Figure 0.1; see [El72].
A4. Let M1 , M2 be any two orientable connected 3-manifolds with non-

empty boundaries and ϕ : ∂M1 → ∂M2 a diffeomorphism. Then there exist
1
2 Intrigue
Figure 0.1. The immersions f and g.
immersions fj : Mj → R3 , j = 1, 2, such that f1 |∂M1 = f2 ◦ ϕ; see [El70]

and Section 14.2.
B. Isometric C 1 -immersions. Is there a regular homotopy ft : S 2 → R3

which begins with the inclusion f0 of the unit sphere and ends with an
isometric immersion f1 into the ball of radius 12 ? Here the word isometric
means preserving the length of all curves. The answer, of course, is negative
if f1 is required to be C 2 -smooth. Indeed, in this case the Gaussian curvature
of the metric on S 2 should be ≥ 4 at least somewhere. However, surprisingly,
the answer is “yes” in the case of C 1 -immersions (when the curvature is
not defined but the curve length is); see [Na54, Ku55] and Chapter 29.
C. Mappings with a prescribed Jacobian. Let Ω be an n-form on

a closed oriented stably parallelizable n-dimensional manifold M such that
Ω = 0, and let
M
η = dx1 ∧ · · · ∧ dxn
be the standard volume form on Rn . Then there exists a map f : M → Rn
such that f ∗ η = Ω; see [GE73].
D. Families of functions with simple singularities. An individual

function f : M → R on a manifold M can always be C ∞ -perturbed to have
only nondegenerate (Morse) critical points. In 1-parametric families one un-
avoidably may also encounter the so-called birth-death type critical points,
as in the family ft (x) = x3 + tx, t ∈ [−1, 1]. In families depending on more
parameters, the possible singularities become more and more complicated;
in fact, as it is proven in singularity theory (see, e.g., [Tho55],[AGZV85]),
they become (smoothly) unclassifiable with a growing number of parameters
and dimensions. However, it turns out that if topology permits, then by a
C 0 (and sometimes even C 1 ) perturbation one can get rid of all singularities
except Morse and birth-death ones; see [Ig84, Ig87, EM97, Lu09, EM12]
Intrigue 3
and Chapter 16. For instance, one has the following statement (see Section
16.5):
Let ft : Rn → R be a family of functions parameterized by t in the sphere
S k−1 = ∂D k . Suppose that for each t ∈ S k−1 the function ft has no critical
points and coincides with the coordinate function x1 outside a compact set.
Then the family ft extends to t ∈ D k such that for each t ∈ D k the func-
tion ft has only Morse and birth-death critical points and coincides with the
coordinate function x1 outside a compact set.
All the above statements are examples of the homotopy principle, or the h-
principle. Despite the fact that all these problems are asking for the solution
of certain differential equations or inequalities, they can be reduced to prob-
lems of a purely homotopy theoretic nature which then can be dealt with
using the methods of algebraic topology. For instance, the regular homotopy
classification of immersions S 2 → R3 can be reduced to the computation of
the homotopy group π2 (RP 3 ), which is trivial.
We are teaching in this book how to deal with these problems. In particu-
lar, the three general methods which we describe here will be sufficient to
handle all the above examples, except A3 and C, though we hope that after
studying the book, the reader would be able to solve A3 and C as advanced
exercises.
Another, maybe even more important goal of this book is to teach the reader
how to recognize the problems which may satisfy the h-principle. Of course,
in the most interesting cases this is a very difficult question. As we will
see below there are plenty of open problems where one can neither establish
the h-principle nor find a single instance of rigidity. Nevertheless we are
confident that the reader should develop a pretty good intuition for the
problems which may satisfy the h-principle.
Here are some more examples where the h-principle holds, fails, or is un-
known.
Examples.
E. Totally real, Lagrangian and ε-Lagrangian embeddings. Let T 2 =

(R/Z) × (R/Z) be the 2-torus with the cyclic coordinates x1 , x2 ∈ R/Z.
Given an embedding f : T 2 → C2 , consider the vectors
∂f ∂f
v1 (x) = (x) and v2 (x) = (x), x ∈ T 2 .
∂x1 ∂x2
The embedding f is called real or totally real if these vectors are linearly
independent (over C) for each x ∈ T 2 . It is called Lagrangian if the real
planes generated by the vectors v1 (x), v2 (x) and iv1 (x), iv2 (x) are orthogonal
4 Intrigue
for each x ∈ T 2 . For 0 < ε ≤ π2 , an embedding f is called ε-Lagrangian if

the angle between these planes is greater than π2 − ε for each x ∈ T 2 . Thus
Lagrangian embeddings are real, and real embeddings coincide with those
that are (π/2)-Lagrangian. Identifying C2 with R4 , we can view a 2 × 2
complex matrix as a pair of vectors in R4 and thus consider GL(2, C) as
a subspace of the Stiefel manifold V4,2 which is formed by pairs of vectors
linearly independent over C. With any embedding f : T 2 → C2 we associate
the map vf : T 2 → V4,2 defined by the formula
vf (x) = (v1 (x), v2 (x)) ∈ V4,2 .
Then f is real if and only if the image vf (T 2 ) is contained in GL(2, C).
E1. Both real and ε-Lagrangian embeddings satisfy the h-principle:
Let f : T 2 → C2 be any embedding. Suppose that the map
vf : T 2 → V4,2
is homotopic to a map
w : T 2 → GL(2, C) ⊂ V4,2 .
Then for any ε > 0 the embedding f is isotopic to an ε-Lagrangian embed-
ding. Moreover, let f, g : T 2 → C2 be two ε-Lagrangian embeddings which
are smoothly isotopic and such that the maps
vf , vg : T 2 → GL(2, C)
are homotopic inside GL(2, C). Then f and g are isotopic via an ε-Lagran-
gian isotopy; see [Gr86] and Section 27.4.
E2. On the other hand, the h-principle is wrong for Lagrangian embed-
dings. Indeed, any two Lagrangian embeddings T 2 → C2 are Lagrangian
isotopic [DRGI16], whereas the h-principle would predict the existence of
knotted Lagrangian tori.
F. Free maps. A map T 2 → Rn is called free if the five vectors

∂f ∂f ∂2f ∂ 2f ∂2f
(x), (x), (x), (x), (x) ∈ Rn
∂x1 ∂x2 ∂x1 ∂x2 ∂x21 ∂x22
are linearly independent for all x ∈ T 2 . Of course, the minimal dimension
n for which free embeddings may exist is equal to 5.
It is an open problem whether there exists a free map T 2 → R5 . In par-
ticular, we do not know whether the h-principle holds for free maps to R5 .
On the other hand, free maps to R6 satisfy the h-principle. We invite the
reader to guess what this statement really means, or look at [GE71].
Intrigue 5
G. Contact and Engel structures. A contact structure on a (2n + 1)-

dimensional manifold M is a completely nonintegrable tangent 2n-plane field
ξ. A completely nonintegrable tangent 2-plane field on a 4-manifold N is
called an Engel structure. In the first case complete nonintegrability means
that the Lie brackets of vector fields tangent to ξ generate T M at each point
of M . In the second case it means that two successive Lie brackets of vector
fields tangent to ξ generate T N at each point of N .
In both cases, some forms of the h-principle hold even on closed mani-
folds. For instance, any tangent 2n-plane field equipped with an almost
complex structure on a (2n + 1)-manifold is homotopic through such plane
fields to a contact structure (see [Lu77, BEM15] and Section 19.2). A
similar existence h-principle holds for Engel structures on parallelizable 4-
manifolds [Vo09, CPPP17].
Part 1
Holonomic
Approximation
Chapter 1
Jets and Holonomy
1.1. Maps and sections

It is customary to visualize a map f : Rn → Rq as its graph Γf ⊂ Rn × Rq .
This graph may be considered as the image of a map Rn → Rn × Rq given
by the formula x → (x, f (x)). Mathematicians call this map a section, while
physicists prefer to call it a field (or an Rq -valued field ). Hence any map
f : Rn → Rq can be thought of as a section Rn → Rn × Rq of the trivial
fibration p : Rn × Rq → Rn . Similarly, any map V → W , where V and W
are smooth manifolds, can be considered as a W -valued field, or as a section
V → V × W of the trivial fibration p : V × W → V . We will also consider
arbitrary smooth fibrations (= fiber bundles) p : X → V with fiber W and
sections of these fibrations, i.e., maps f : V → X such that p ◦ f = idV .
In all cases the image of a section contains all the information about this
section, and we will use the term section both for the section as a map and
for its image. If p : X → V and p : X → V are two fibrations over the
same base V , then the map f : X → X is called fibered (over V ) if the
diagram
f
X A / X
AA | |
AA ||
A |
p AA
|
~|| p
V
commutes.
In what follows we usually denote the (always finite) dimensions of V , W ,
and X by n, q, and n + q. By a section or a map we mean, as a rule, a
C ∞ -smooth section or map. A family of sections or maps is a continuous
(with respect to the parameter) family of sections or maps. However, such
9
10 1. Jets and Holonomy
a family is supposed to be smooth in the cases when we need to differentiate

with respect to the parameter.
1.2. Coordinate definition of jets

Given a (smooth) map f : Rn → Rq and a point x ∈ Rn , the string of
derivatives of f up to order r

Jfr (x) = f (x), f (x), . . . , f (r) (x)
is called the r-jet of f at x. Here f (s) consists of all partial derivatives

D α f , α = (α1 , . . . , αn ), |α| = α1 + · · · + αn = s, written lexicographically.
Let dr = d(n, r) be the number of all partial derivatives D α of order r of a
function Rn → R. The r-jet Jfr (x) of a map f : Rn → Rq can be viewed as
a point in the space
Rq × Rqd1 × Rqd2 × · · · × Rqdr = RqNr ,
where Nr = N (n, r) = 1 + d1 + d2 + · · · + dr .
(n+r−1)! (n+r)!
Exercise. Prove that d(n, r) = (n−1)!r! and N (n, r) = n!r! .
The space x × RqNr can be viewed as the space of all a priori possible values
of r-jets of maps f : Rn → Rq at the point x ∈ Rn . In this context the space
Rn × RqNr = Rn × Rq × Rqd1 × Rqd2 × · · · × Rqdr
is called the space of r-jets of maps Rn → Rq , or the space of r-jets of
sections Rn → Rn × Rq , and is denoted by J r (Rn , Rq ). For example,
J 1 (Rn , Rq ) = Rn × Rq × Mq×n ,
where Mq×n = Rqn is the space of (q × n)-matrices.
Given a section f : Rn → Rn × Rq , the section
Jfr : Rn → J r (Rn , Rq ), x → Jfr (x),
of the trivial fibration
pr : J r (Rn , Rq ) = Rn × RqNr → Rn
is called the r-jet of f , or the r-jet extension of f .
Note that for any point z ∈ J r (Rn , Rq ) there exists a unique Rq -valued
polynomial f (x1 , . . . , xn ) of degree ≤ r such that Jfr (pr (z)) = z. Hence,
there exists a canonical trivialization
Jr
Rn × Pr (n, q) → J r (Rn , Rq )
of the fibration pr : J r (Rn , Rq ) → Rn , where Pr (n, q) is the space of all
polynomial maps Rn → Rq of degree ≤ r.
1.3. Invariant definition of jets 11
X ϕ
n q
R xR
p
V
Figure 1.1. The trivialization ϕ, the sections f , g, and their images
ϕ∗ f , ϕ∗ g.
Exercise. Draw the 1-jet of the function f : R → R, f (x) = ax + b.
1.3. Invariant definition of jets

In order to define the space J r (V, W ) of r-jets of sections V → V × W of
a trivial fibration p : V × W → V and, more generally, the r-jet space of
sections V → X of an arbitrary smooth fibration p : X → V , we need to
define jets invariantly.
Following Gromov’s book [Gr86], we will use the notation Op A as a re-

placement of the expression an open neighborhood of A ⊂ V . In other
words, Op A is an arbitrarily small but nonspecified open neighborhood of
a subset A ⊂ V .
Let v ∈ V . Two local sections f : Op v → X and g : Op v → X of the

fibration X → V are called r-tangent at the point v if f (v) = g(v) and
Jϕr ∗ f (ϕ(v)) = Jϕr ∗ g (ϕ(v))
for a local trivialization ϕ : U → Rn × Rq of X in a neighborhood U of the
point x = f (v) = g(v). Here ϕ∗ f and ϕ∗ g are the images of the sections f
and g (see Figure 1.1).
It follows from the chain rule that the r-tangency condition does not depend
on the specific choice of the local trivialization. The r-tangency class of a
section f : Op v → X at a point v ∈ V is called the r-jet of f at v and
denoted by Jfr (v). Thus we have correctly defined the set X (r) of all r-
jets of sections V → X and the set-theoretic fibrations pr0 : X (r) → X and
pr = p ◦ pr0 : X (r) → V . Moreover, the extensions
ϕr : (pr0 )−1 (U ) → J r (Rn , Rq )
of the local trivializations ϕ : U → Rn ×Rq which send the r-tangency classes

of local sections of X to the r-tangency classes of their images in J r (Rn , Rq )
define a natural smooth structure on X (r) such that pr : X (r) → V becomes a
smooth fibration. This fibration is called the r-jet extension of the fibration
p : X → V . The section
Jfr : V → X (r) , v → Jfr (v),
is called the r-jet of a section f : V → X, or the r-jet extension of f .
It is important to understand that the chain of inclusions

Rn × Rq = J 0 (Rn , Rq ) ⊂ J 1 (Rn , Rq ) ⊂ J 2 (Rn , Rq ) ⊂ · · · ⊂ J r (Rn , Rq ) ⊂ · · ·
is not invariant with respect to reparametrizations of Rn × Rq fibered over
Rn . Indeed, the fibered reparametrizations act transitively on the space of
sections Rn → Rn × Rq . Hence for a general fibration X → V the chain
X = X (0) ⊂ X (1) ⊂ X (2) ⊂ · · · ⊂ X (r) ⊂ · · ·
does not exist as an invariant object. On the other hand, the r-tangency of
two sections implies their s-tangency for all 0 ≤ s < r, and therefore the
chain of projections
X = X (0) ← X (1) ← X (2) ← · · · ← X (r) ← · · ·
is invariantly defined.
Exercise. Prove that the projection prr−1 : X (r) → X (r−1) carries a

natural structure of an affine bundle.
If X is a trivial fibration V × W → V , then the space of r-jets of sections

(or maps V → W ) will be denoted by J r (V, W ).
1.4. The space X (1)

According to the invariant definition of the jet space, the points of X (1) are 1-
tangency classes of sections, and therefore they can be viewed as nonvertical
tangent n-planes Px ⊂ Tx X. Here nonvertical means
Px ∩ Vertx = {x},
where Vertx is the q-dimensional tangent space to the fiber of the fibration
X → V over p(x). If we fix a point in X (1) , i.e., a nonvertical n-plane Px ,
then the fiber of the fibration X (1) → X over x can be identified with
Hom(Px , Vertx ) ≈ Hom (Rn , Rq ) = Rqn .
1.5. Holonomic sections of the jet space X (r) 13
If X = V × W and x = (v, w), then Vertx = Tw (v × W ), and moreover, we

can set Px = Tv (V × w). Therefore J 1 (V, W ) → V × W is a vector bundle
with the fiber Hom (Tv V, Tw W ) over x = (v, w). In particular,
J 1 (V, R) = T ∗ (V ) × R and J 1 (R, W ) = R × T (W ).
In the general case Px cannot be canonically chosen, and therefore the fibra-
tion X (1) → X does not have a canonical vector bundle structure, although
the affine structure does survive.
Note that the sections of the fibration p10 : X (1) → X may be identified with
connections on X. For example, there exists a natural inclusion X → X (1)
for X = V × W which corresponds to the standard flat connection on the
trivial bundle V × W → V .
1.5. Holonomic sections of the jet space X (r)

Given a section F : V → X (r) , we will denote by bs F the underlying section
pr0 ◦ F : V → X. A section F : V → X (r) is called holonomic if F = Jbs r .
F
In particular, holonomic sections Rn → J r (Rn , Rq ) have the form

x → x, f (x), f (x), . . . , f (r) (x) .
The correspondence f → Jfr defines the derivation map
J r : Sec X → Sec X (r) .
Its one-to-one image J r (Sec X) coincides with the space
Hol X (r) ⊂ Sec X (r)
of holonomic sections, i.e., we have

Jr
Sec X Hol X (r) → Sec X (r) .
Note that the C 0 -topology on Sec X (r) induces via J r the C r -topology on
Sec X.
A homotopy of holonomic sections of X (r) is called a holonomic homotopy.
Any fibered map g : X → Y between two fibrations X → V and Y → V

can be extended to a map g (r) : X (r) → Y (r) which sends the r-jet of a
(local) section ϕ : V → X to the r-jet of the section g ◦ ϕ : V → Y . It
is important to observe that the induced map Sec X (r) → Sec Y (r) sends
Hol X (r) to Hol Y (r) .
q
R
n
R
Figure 1.2. A section F : Rn → J 1 (Rn , Rq ) as a pair (graph, plane

field along graph).
1.6. Geometric representation of sections of X (r)

A section F : V → X (1) can be viewed geometrically as a section
f = bs F : V → X
together with a field τ of nonvertical n-planes along f ; see Figure 1.2. Such
a section is holonomic if and only if the field τ is tangent to f (V ).
Similarly, a section F : V → X (s) can be viewed as a pair (Fs−1 , τs ) where
Fs−1 = pss−1 ◦ F : V → X (s−1)
and τs is a field of nonvertical n-planes along Fs−1 in T X (s−1) . Continuing
inductively, we interpret a section V → X (r) as the sequence
{f, τ1 , τ2 , . . . , τr },
where τs is a nonvertical n-plane field in T X (s−1) along
Fs−1 = {f, τ1 , τ2 , . . . , τs−1 }.
The section F : V → X (r) is holonomic if and only if for each s = 1, . . . , r
the field τs is tangent to Fs−1 (V ).
This interpretation of sections V → X makes geometrically clear the analy-

tically evident fact that a random section F : V → X (r) is not holonomic and
that the holonomic sections are rather exotic objects in the space Sec X (r)
of all sections of the jet-bundle X (r) .
1.7. Holonomic splitting

The following observation shows that given a local holonomic section
F : U → X (r) , V ⊃ U Rn ,
there are plenty of holonomic sections U → X (r) which are parallel to F .
1.7. Holonomic splitting 15
As we have already mentioned in Section 1.2, the space J r (Rn , Rq ) has a

canonical parametrization
J r : Rn × Pr (n, q) → J r (Rn , Rq ), (u, f ) → (u, Jfr (u)),
where Pr (n, q) is the space of all polynomial maps Rn → Rq of degree
≤ r. Such a parametrization has the following nice property, which we call
holonomic trivialization: The images of the horizontal fibers
Rn × f ⊂ Rn × Pr (n, q)
are holonomic sections Jfr .
In particular,
1.7.1. (Holonomic splitting) Any holonomic section F : V → X (r) has
a holonomically trivialized tubular neighborhood over any open ball U ⊂ V .
In other words, there exists an embedding
PF : U × RK → X (r) |U
onto a neighborhood of F |U , where
q(n + r)!
K = dim Pr (n, q) = ,
n!r!
such that PF (u, 0) = F (u), u ∈ U , and for each z ∈ RK the map
u → PF (u, z), u ∈ U,
is a holonomic section U → X (r) |U .
Proof. Let us denote by T (X|U ) the vertical tangent bundle of the fibration
X|U , i.e., the bundle formed by tangent planes to the fibers of the fibration
X|U → U , and by Y the induced vector bundle F ∗ T (X|U ) over U . Let
U × Rq → Y be a trivialization of the bundle Y over the ball U . By choosing
a coordinate system in U we can identify Y (r) with J r (Rn , Rq ) and define a
map
J r : Rn × Pr (n, q) → J r (Rn , Rq ) = Y (r)
by the formula J r (u, f ) = (u, Jfr (u)). There exists a neighborhood Ω of the
section F |U in X|U and a fiberwise diffeomorphism g : Y → Ω. The required
embedding
PF : U × RK → X (r)
can now be defined as the composition
Jr g (r) i
Rn × Pr (n, q) → Y (r) → Ω(r) → X (r) ,
where g (r) : Y (r) → Ω(r) is the r-jet extension of g and i is the inclusion
Ω(r) → X (r) .
This observation is a key to the Thom Transversality Theorem which we

discuss in the next chapter.
Chapter 2
Thom Transversality
Theorem
2.1. Generic properties and transversality

It is convenient to express the idea of abundance of maps or sections which
satisfy a certain property P by saying that a generic map or section satisfies
this property. More precisely, we say that a generic section from a space S
has a property P if the space of maps from S which have this property is
open and everywhere dense in S, or more generally if it can be presented as
a countable intersection of open and everywhere dense sets. The space of
smooth sections of a fibration, and most other functional spaces considered
in this book, are so-called Baire spaces in which sets of generic maps are
dense (in particular, they are nonempty).
A map f : V → W is called transverse to a submanifold Σ ⊂ W if for each

point x ∈ V one of the following two conditions holds:
• f (x) ∈
/ Σ or
• f (x) ∈ Σ and the tangent space Tf (x) W is generated by Tf (x) Σ and
df (Tx V ).
If codim Σ > dim V , then the second condition can never be satisfied, and
thus transversality just means that f (V ) ∩ Σ = ∅.
The implicit function theorem guarantees that if a map f : V → W is
transverse to Σ, then f −1 (Σ) is a submanifold of V of the same codimension
in V as that of Σ in W .
17
18 2. Thom Transversality Theorem
2.2. Stratified sets and polyhedra

A closed subset S of a manifold V is called stratified if it is presented as

N
a union Sj of locally closed submanifolds Sj , called strata, such that for
0
each k = 0, . . . , N we have

N
Sk = Sj ,
j=k
where Sk is the closure of the stratum Sk . The dimension of a stratified set

is the maximal dimension of its strata and its codimension is the minimal
codimension of its strata.
Examples.
1. Each manifold V with boundary has a stratification with two strata S0 =
Int V and S1 = ∂V .
2. Given a smooth triangulation of a manifold V , any closed subset which is
a union of simplices of the triangulation is stratified by the strata which are
interiors of the simplices. We will call stratified sets of this kind polyhedra.
3. Any real analytic or semianalytic set (i.e., a set defined by a system of
analytic equations and inequalities) can be stratified; see [GM87, p. 43].
2.2.1. (The space of matrices of bounded rank) Let us denote by

Σi , i = 0, . . . , m,
the algebraic subset of the space Mq×n of q × n matrices which consists
of matrices of rank ≤ m − i where m = min(n, q), and by Si the space
of matrices of rank = m − i. Then Si , i = 0, . . . , m, are locally closed

m
submanifolds of Mq×n of codimension i(|q − n| + i), and the union Sj is
j=i
a natural stratification of Σi .
Proof. The condition that the rank of a q × n matrix is precisely equal to

m − i is expressed by equating to zero i(|q − n| + i) minors of order m − i + 1
enveloping a nonzero minor of order m − i. It is straightforward to check
that this system has maximal rank.
2.2.2. (Corollary) Let Σi ⊂ J 1 (V, W ) be the space of 1-jets of maps of

rank ≤ min(n, q) − i.
Then Σi is a stratified subset of codimension i(|q − n| + i).
2.3. Thom Transversality Theorem 19

N
A map f : V → W is called transverse to a stratified set Σ = Sj ⊂ W if
0
it is transverse to each stratum Sj , j = 0, . . . , N . For a transverse map f
the preimage f −1 (Σ) of a stratified subset Σ ⊂ W is a stratified subset of V
of the same codimension.
2.3. Thom Transversality Theorem

We begin with an almost obvious lemma which is the simplest case of Sard’s
theorem [Sa42].
2.3.1. Let f : V → W be a C 1 -smooth map. If q = dim W > n = dim V ,
then the image f (V ) has zero q-dimensional Lebesgue measure.
Proof. It is sufficient to consider the case when V is the closed n-disc D =

D n and W = Rq . There is a constant C > 0 such that for any integer N > 0
the ball D can be covered by CN n balls Bi of radius 1/N . Set
Δ = max da f .
a∈D
i ⊂ Rq of radius ≤
The image of each ball Bi is contained in a ball B Δ
N.
i is bounded by
Hence the total volume of balls B
CN n Δq σn C
= → 0,
Nq N q−n N →∞
where σn is the volume of the unit ball in Rn . Therefore the image f (V )
can be covered in W by a set of an arbitrarily small measure.
2.3.2. (Thom Transversality Theorem) Let X → V be a smooth fibra-

tion and Σ a stratified subset of the r-jet space X (r) . Then for a generic
section f : V → X its r-jet extension Jfr : V → X (r) is transverse to Σ.
Proof of Theorem 2.3.2 in the case codim Σ > n = dim V . We need to

prove that for a generic holonomic section F : V → X (r) , the image F (V )
does not intersect Σ. Take a closed n-disc D ⊂ V . Note that the space
HolD (X (r) \ Σ) of holonomic sections F : V → X (r) for which F (D) ∩ Σ = ∅
is open. Thus if we show that HolD (X (r) \Σ) is everywhere dense in Sec X (r)
this would imply the theorem. Indeed, V can be covered by countably
many closed balls Di , i = 1, . . . , hence the space Hol(X (r) \ Σ) can then be
presented as the intersection

HolDi (X (r) \ Σ)
i
of countably many open everywhere dense sets.

Choose any section F : V → X (r) . According to Lemma 1.7.1, a neighbor-

hood of F in X (r) |D can be holonomically trivialized, i.e., there exists an
embedding
PF : D × RK → X (r) |D
q(n+r)!
onto a neighborhood of F |D , where K = dim Pr (n, q) = n!r! , such that
PF (u, 0) = F (u), u ∈ D, and for each z ∈ RK the map
u → PF (u, z), u ∈ D,
is a holonomic section D → X (r) |D . Let π denote the projection
D × RK → RK
and set
=π ◦ P −1 (Σ)
Σ F
={z ∈ RK | the section u → PF (u, z) is not transverse to Σ}.
By assumption we have dim Σ < dim X (r) − n = K. Hence Lemma 2.3.1
implies that Σ is
has measure 0 in RK . In particular, the complement RK \ Σ
everywhere dense in R . Therefore, any open neighborhood of F |D contains
K
a holonomic section D → X (r) \ Σ, which then can be extended to a section

F ∈ Hol (X (r) \ Σ) approximating F over the whole V .
Proof of Theorem 2.3.2 in the case codim Σ ≤ n = dim V . Denote

by Σ(1) the subset of the jet space X (r+1) which consists of the (r + 1)-jets
at v ∈ V of sections f : Op v → X for which Jfr is not transverse to Σ
at v. The assertion that a generic section in Hol X (r) is transverse to Σ is
equivalent to the assertion that a generic section in Hol X (r+1) belongs to
Hol (X (r+1) \ Σ(1) ). Hence, this case of Theorem 2.3.2 can be deduced from
the one considered above and the following lemma.
2.3.3. (Singularity of tangency to a stratified subset) Let Σ ⊂ X (r)
be a stratified subset. Suppose that k = codim Σ ≤ n = dim V . Then Σ(1)
is contained in a stratified subset of X (r+1) of codimension n + 1.
We will illustrate the ideas involved in the proof of Lemma 2.3.3 in two
special cases and will leave the general case to the reader as an (advanced)
exercise; see [BM15].
Case r = 0. We may assume that X → V is the trivial fibration Rn × Rq →

Rn . Then X (1) = J 1 (Rn , Rq ) = Rn × Rq × Mq×n , where Mq×n is the space
of (q × n)-matrices. We denote by (x1 , . . . , xn ) the coordinates on Rn , by
(y1 , . . . , yq ) the coordinates on Rq , and by (x1 , . . . , xn , y1 , . . . , yn , Y = (yij )),
i = 1, . . . , q, j = 1, . . . , n the corresponding coordinates on X (1) = Rn ×
Rq × Mq×n . Denote k = codim Σ and = n + q − k = dim Σ. Let Σ be
given locally by a system F1 = · · · = Fk = 0 of maximal rank. Choose

local vector fields Z1 = (z11 , . . . , z(n+q)1 )T , . . . , Z = (z1 , . . . , z(n+q) )T
which span T Σ. Denote
⎛ ⎞
z(n+1)1 . . . z(n+1)
:= ⎝ . . .
Z ... ... ⎠.
z(n+q)1 . . . z(n+q)
Then rankZ ≥ − n = q − k, so we can assume without a loss of gener-
ality that the first m = min(0, q − k) columns Z j = (z(n+1)j , . . . , z(n+q)j )T ,
j = 1, . . . , m of the matrix Z are linearly independent. We define the ma-
trix Z by taking the m first columns of Z. For a map f : Rn → Rq the
nontransversality
condition off to Σ implies F = (F1 , . . . , Fk ) = 0 and
∂f ∂f m < q. Hence, the subset Σ(1) ⊂ J 1 (Rn , Rq )
rank ∂x1 , . . . , ∂xn , Z1 , . . . , Z
satisfies the conditions

F (x1 , . . . , xn , y1 , . . . , yq ) = 0 and rank Y Z < q.
Note that the last condition is equivalent to rank Y < q − m, where Y ∈

M(q−m)×n denotes the projection of Y to the complement of the columns
of Z. According to Lemma 2.2.1, the above equations therefore define a
stratified subset of X (1) of codimension

n + 1, q ≥ k,
k + (n + m − q + 1) =
n + 1 + k − q > n + 1, q < k.
Case codim Σ = 1. We may assume as above that p : X → V is the trivial

fibration Rn × Rq → Rn . Let us denote the coordinates in the space X (r) by
x1 , . . . , xn , y1α , . . . , yqα ,

n
where α = (α1 , . . . , αn ) are multi-indices with αi ≥ 0 and | α| = αi ≤ r.
i=1
The coordinate yiα corresponds to the partial derivative
∂ | α| fi ∂ | α| fi
= α1
∂x α ∂x1 · · · ∂xαnn
of the i-th coordinate function fi , i = 1, . . . q, of a section V → X. The
projections
pr : X (r) → V and prs : X (r) → X (s) , s = 0, . . . , r − 1
(see 1.3) are given by dropping the coordinates yiα with |α| ≥ s + 1 for
s = −1, 0, . . . , r − 1 and i = 1, . . . , q. Without loss of generality we may
assume that Σ is a submanifold. Locally near a point z ∈ Σ the submanifold
Σ can be defined by an equation F = 0. Let us denote by Xsr (z) and X r (z)
the fibers of the projections prs and pr through the point z.
Suppose that Σ is not transverse at z to the fiber X r (z). Then for x = pr (z)
any local section J r : Op x → X (r) , which is C 1 -close to a section Jfr :
f
Op x → X (r) with Jfr (x) = z, is transverse to Σ. Hence in this case Σ(1) ⊂
X (r+1) does not intersect a neighborhood of the fiber Xrr+1 (z). Suppose now
that Σ is transverse at z to the fiber X r (z) and set
S = max {s = 0, . . . , r | Σ is transverse at z to the fiber Xs−1
r
(z)},
r = X r . Then
where X−1
∂F
(z) = 0
∂yiα
for any i = 1, . . . , q when | α| > S, and there exist i ∈ {1, . . . , q} and a
with | α
multi-index α | = S such that
∂F
(z) = 0 where ȳ = yiα .
∂ ȳ
Since
∂ |α| f
i
Jfr (x) = x, (x) , 1 ≤ i ≤ q, 0 ≤ |α| ≤ r,
∂xα
the tangent space to a section Jfr at the point z is generated by the vectors
∂Jfr ∂ | α|+1 fi
vk = (x) = ek , (x) , k = 1, . . . , n,
∂xk ∂xk ∂xα
where e1 , . . . , en denotes the standard basis of Rn . Nontransversality of Jfr
to Σ = {F = 0} at the point z = Jfr (x) is thus expressed by the equations
∂F
q
∂F ∂ | α|+1 fi
0 = dF (vk ) = (z) + (z) (x), k = 1, . . . , n.
∂xk ∂yiα ∂xk ∂xα
i=1 |α|≤r
In coordinates x1 , . . . , xn and yiβ , 1 ≤ i ≤ q, 0 ≤ |β| ≤ r + 1 on X (r+1) , the

set Σ(1) ⊂ X (r+1) is therefore defined in a neighborhood of the fiber Xrr+1 (z)
by the system of n + 1 equations
⎧
⎪
⎨F (z) = 0,
(1) ∂F q ∂F α+ek
⎪
⎩ ∂xk (z) + ∂yiα (z)yi = 0, k = 1, . . . , n.
i=1 |α|≤S
Note the triangular character of this system: In the sum only α with |α| ≤ S
α (z) = 0 for | α| > S and all i. For the same reason,
∂F
appears, because ∂y
i
the first equation contains no variables yiα with |α| > S. The remaining
equations can be viewed as an inhomogeneous system of n linear equations
in the variables yiβ with |β| = S + 1 with coefficients depending the on xj
and yiα with α ≤ S.
We claim that the rank of this linear system is equal to n. To see this,
note first that the matrix of this system consists of blocks corresponding to
i = 1, . . . , q. So it suffices to show that the block corresponding to i = i has
rank n, where i is the index from above for which there exist α with |
α| = S
such that ∂yα (z) = 0. Abbreviating A := ∂yα (z), this block corresponds
∂F α ∂F

i
i
to the linear transformation

{yβ }|β|=S+1 → { Aα yα+ek }1≤k≤n .
i i
|α|=S
Let B be the matrix of this transformation with coefficients bβk , |β| = |S| +1,
k = 1, . . . , n. Consider the homogeneous polynomial

Q(u) = Q(u1 , . . . , un ) = Aα uα , uα = uα1 1 · · · uαnn .
|α|=S
Note that Q is not the zero polynomial (because Aα = 0), and the coefficients
of the polynomials

Q(u)uk = Aα uα+ek = bβk uβ , k = 1, . . . , n,
|α|=S |β|=|S|+1
encode the rows of B. Suppose now that the rows of B are linearly depen-
dent, i.e., there exists (c1 , . . . , cn ) = 0 such that
0 = c1 Q(u)u1 + · · · + cn Q(u)un = (c1 u1 + · · · + cn un )Q(u).
But this is impossible because the ring of polynomials has no zero divisors.
Hence, rank B = n, and the claim is proved.
The claim together with the condition dF = 0 and the triangular character
of system (1) implies that system (1) has rank n + 1, and therefore Σ(1) has
codimension at least n + 1.
An alternative proof of Theorem 2.3.2 can be found in [Gr86].
Chapter 3
Holonomic
Approximation
The Holonomic Approximation Theorem which we discuss in this chapter

shows that in some sense there are unexpectedly many holonomic sections
near any submanifold A ⊂ V of positive codimension.
3.1. Main theorem

Question. Let p : X → V be a smooth fibration. Is it possible to approx-
imate any section F : V → X (r) by a holonomic section? In other words,
given an r-jet section and an arbitrarily small neighborhood of the image
of this section in the jet space, can one find a holonomic section in this
neighborhood ?
The answer is evidently negative (excluding, of course, the situation when

the initial section is already holonomic). For instance, in the case r = 1 and
X = V × R the question has the following geometric reformulation. Given
a function and a nonvertical n-plane field along the graph of this function,
can one C 0 -perturb this graph to make it almost tangent to the given field?
The problem of finding a holonomic approximation of a section of the r-jet

space near a submanifold A ⊂ Rn (i.e., over Op A ⊂ Rn ) is also usually
unsolvable. The only exception is the zero-dimensional case: any section
can be approximated near any point by the r-jet of the respective Taylor
polynomial map.
25
26 3. Holonomic Approximation
In contrast, the following theorem says that we can always find a holonomic
approximation of a section F : V → X (r) near a slightly deformed subman-
⊂ V if the original submanifold A ⊂ V is of positive codimension.
ifold A
3.1.1. (Holonomic Approximation Theorem) Let A ⊂ V be a polyhe-
dron of positive codimension and let
F : Op A → X (r)
be a section. Then for arbitrarily small δ, ε > 0 there exists a δ-small (in
the C 0 -sense) diffeotopy
hτ : V → V, τ ∈ [0, 1],
and a holonomic section
F : Op h1 (A) → X (r)
such that
dist(F(v), F |Op h1 (A) (v)) < ε
for all v ∈ Op h1 (A) (see Figure 3.1).
Figure 3.1. The sets A, h1 (A), Op A (gray), and Op h1 (A) (deep gray).
Moreover, the relative version is also true: if F is already holonomic over

Op B ⊂ V , where B is a subpolyhedron of A, then the isotopy hτ can be
assumed to be fixed on B and F can be chosen equal to F on B.
Remarks.
1. If A (and thus V ) is noncompact, then instead of arbitrarily small numbers
ε, δ > 0 one can take arbitrarily small positive functions
δ, ε : V → R+ .
Later in the book similar situations will appear frequently, and we will always
silently assume that our arbitrarily small numbers become arbitrarily small
functions in the case of a noncompact polyhedron A.
2. Let us recall that we use the notation Op A as a replacement of the

expression an open neighborhood of A and the term polyhedron in the sense
that A is a subcomplex of a certain smooth triangulation of the manifold V .
3.1. Main theorem 27
3. We assume that the manifold V is endowed with a Riemannian metric and

the bundle X (r) is endowed with a Euclidean structure in a neighborhood
U of the section F (V ) ⊂ X (r) .
4. A diffeotopy (i.e., a family of diffeomorphisms) hτ : V → V , τ ∈ [0, 1], is

called δ-small if h0 = IdV and
dist(hτ (v), v) < δ
for all v ∈ V and τ ∈ [0, 1].
5. We assume that the image h1 (A) is contained in the domain of definition

of the section F .

Exercise. Construct the required h1 and F in the case V = R2 , A = I ×0
(I = [0, 1]), and
F : Op A → J 1 (R2 , R), (x1 , x2 ) → (x1 , x2 , f (x1 , x2 ), 0, 0)
where f (x1 , x2 ) = x1 . In other words, approximate the steep path
bs F (Op I) ⊂ R2 × R
by an almost flat path.
The parametric and relative parametric versions of Theorem 3.1.1 are also
true:
3.1.2. (Parametric Holonomic Approximation Theorem) Let A ⊂ V
be a polyhedron of positive codimension, B ⊂ A a subpolyhedron and
Fz : Op A → X (r) , z ∈ Im
a family of sections parametrized by a cube I m = [0, 1]m . Suppose that the
sections Fz are holonomic for all z ∈ ∂I m and are holonomic over Op B ⊂ V
for all z ∈ I m . Then for arbitrarily small δ, ε > 0 there exists a family of
δ-small diffeotopies
hτz : V → V, τ ∈ [0, 1], z ∈ I m ,
and a family of holonomic sections
Fz : Op h1z (A) → X (r) , z ∈ I m ,
such that
• hτz (v) = v and Fz (v) = Fz (v) for (z, v) ∈ (I m × Op B) ∪ (∂I m ) × A;
• dist(Fz (v), Fz |Op h1z (A) (v)) < ε for all (z, v) such that v ∈ Op h1z (A).
Remark. Note that what we call here and below a relative parametric
version is relative with respect to (I m × B) ∪ (∂I m × A).
In Section 3.8 we reformulate Theorem 3.1.2 in a foliated (= leafwise) form.
3.2. Holonomic approximation over a cube

Using induction over the skeleton of the polyhedron A and taking into ac-
count that the fibration X → V is trivial over simplices, we reduce the
relative version of Theorem 3.1.1 to its special case for the pair (A, B) =
(I k , ∂I k ) ⊂ Rn .
3.2.1. (Holonomic approximation over a cube) Let I k ⊂ Rn , k < n,
be the unit cube in the coordinate subspace Rk ⊂ Rn of the first k coordinates.
For any section
F : Op I k → J r (Rn , Rq )
which is holonomic over Op ∂I k and for arbitrarily small positive numbers
δ, ε > 0 there exists a δ-small (in the C 0 -sense) diffeomorphism
h : Rn → Rn , h(x1 , . . . , xn ) = (x1 , . . . , xn−1 , xn + ϕ(x1 , . . . , xn )),
and a holonomic section
F : Op h(I k ) → J r (Rn , Rq )
such that
• h = Id and F = F on Op ∂I k ;
• F − F |Op h(I k ) C 0 < ε.
Theorem 3.2.1 will be deduced from the Gluing Lemma 3.4.1 which we for-
mulate below. We begin by introducing the notion of a holonomic extension.
3.3. Holonomic extension

Given a closed subset A ⊂ V , we say that a section F : A → X (r) admits a
holonomic extension along A if there exists a holonomic section F : Op A →
X (r) such that F |A = F . Any two holonomic extensions Op A → X (r)
of a section F : A → X (r) can be joined by a homotopy in the space of
holonomic extensions, and moreover, the space of holonomic extensions of
F is contractible.
Let π : V → B be a fibration. A holonomic extension of a section F : V →
X (r) along the fibers of π is a continuous family of holonomic sections
Fb : Op π −1 (b) → X (r) , b ∈ B,
such that for each b ∈ B the sections Fb and F coincide over the fiber π −1 (b).
Here by continuity of a family of sections Fb : Op π −1 (b) → X (r) , b ∈ B, we
mean the continuity of the section
F : Op V ⊂ V × B → X (r) × B,
3.3. Holonomic extension 29
where V = {(v, π(v)), v ∈ V } is the graph of the projection π and the

restriction of F to Op V ∩ V × b coincides with Fb .
3.3.1. (Holonomic extension over points) For any section F : V →
X (r) each restriction F |v : v → X (r) admits a holonomic extension. More-
over, suppose that for a closed set A ⊂ V a section F : V → X (r) is
holonomic over Op A. Then there exists a family of holonomic extensions
Fv : Op v → X (r) , v ∈ V,
such that Fv = F |Op v for v ∈ A.
Indeed, locally one can take as the section Fv : Op v → X (r) the Taylor
polynomial map corresponding to F (v) with respect to some local coordinate
system centered at v. The global result follows, using a partition of unity
and the contractibility of the space of holonomic extensions.
The above statement also holds parametrically for families of sections.
Remarks.
1. Given a fibration p : X → V and a submanifold A ⊂ V , consider the
preimage XA = p−1 (A) → A and its r-jet space (XA )(r) . On the other hand,
we have the preimage XA = (pr )−1 (A) under the map pr : X (r) → V and
(r)
(r)
a canonical projection π : XA → (XA )(r) . We call a section F : V → X (r)
holonomic along A if the section
F |A (r) π
FA : A −→ XA −→(XA )(r)
is holonomic.
2. Any fibered section F : Z × V → Z × X (r) can be extended (in many
ways) to a section F : Z × V → (Z × X)(r) . If F is fiberwise holonomic (i.e.,
F |z×V : z × V → z × X (r) is holonomic for all z), then any such extension F
is holonomic along z × V for every z. However, there is no guarantee that F
admits holonomic extensions along the fibers z × V ; see the exercises below.
On the other hand, if a section F is obtained from a fiberwise holonomic
section F by differentiating the section f = bs F : Z × V → Z × X with
respect to V and Z,1 then F will be holonomic and thus automatically
holonomically extendable along z × V .

Exercises.
(1)
1. Prove that any section F : A → XA which is holonomic along A admits
a holonomic extension along A.
1 We remind the reader (see Section 1.1) that we assume all the families to be smooth in the
case when we need to differentiate with respect to the parameter.

(2)
2. Give an example of a section F : A → XA which is holonomic along A
but has no holonomic extension along A.
3.4. Gluing Lemma

In the Gluing Lemma below we will consider the cube
I k = I k × 0 ⊂ Rk × Rn−k = Rn , 1 ≤ k < n,
as the family
{y × I l }y∈I k−l , y = (x1 , . . . , xk−l )
of l-dimensional cubes, l = 0, 1, . . . , k − 1. We assume that there is a family

of holonomic extensions of a section
along the cubes {y×I l }y∈I k−l and try to glue these extensions in the direction
of the coordinate t = xk−l to obtain a family of holonomic extensions along
larger cubes
{z × I l+1 }z∈I k−l−1 , z = (x1 , . . . , xk−l−1 ).
We recommend that the reader keeps in mind the two simplest cases
n = 2, k = 1, l = 0 and n = 3, k = 2, l = 1
while reading the statements and proofs in this and the next section. We
will illustrate these cases with pictures.
Given a subset A ⊂ Rn , we will denote its cubical δ-neighborhood by Nδ (A).

Let us fix a positive θ < 1, and for y = (x1 , . . . , xk−l ) ∈ I k−l ⊂ I k ⊂ Rn set
Uδ (y) = Nδ (y × I l ), Vδ (y) = Nδ (y × ∂I l ),

Aδ (y) = Uδ1 (y) \ Vδ (y) ∩ (y × Rn−k+l ), where δ1 = θδ;
see Figures 3.2 and 3.3.
Remark. In all our considerations below in this chapter we can proceed

with any fixed positive θ < 1. However, for some further generalizations in
Section 23.2 it will be convenient to take θ ≤ 14 .
3.4. Gluing Lemma 31
x2
x1
Figure 3.2. The sets Uδ (y) (grey) and Aδ (y) (black) in the case
n = 2, k = 1, l = 0.
x3
x2
x1
Figure 3.3. The sets Uδ (y), Vδ (y) (light grey) and Aδ (y) (dark
grey) in the case n = 3, k = 2, l = 1.
3.4.1. (Gluing Lemma) Suppose that a section

is holonomic over Op ∂I k and for a nonnegative integer l < k is holonomi-
cally extended along the cubes
y × I l ⊂ Rn , y ∈ I k−l ,
i.e., for a positive δ we have a family of holonomic sections
Fy = Jfry : Uδ (y) → J r (Uδ (y), Rq ), y ∈ I k−l ,
such that
• Fy |(y×I l )∪Vδ (y) = F |(y×I l )∪Vδ (y) ;
• Fy = F |Uδ (y) for y ∈ Op ∂I k−l .
Then for an arbitrarily small ε > 0 there exist an integer N > 0 and a
family of holonomic sections
Fz : Ωz → J r (Rn , Rq ), z = (x1 , . . . , xk−l−1 ) ∈ I k−l−1 ,
where
N

N
Ωz = Op Aδ (z, ci ) ∪ z × I l+1
\ Aδ (z, ci ),
i=1 i=1
2i−1
ci = 2N , i = 1, . . . , N , (see Figures 3.4 and 3.5) such that
• Fz = F on Ωz ∩ Op ∂I k ;
• Fz − F |Ω C 0 < ε.
z
The proof of the Gluing Lemma will be given in Section 3.6.
Remark. Note that for l = k −1 we have z ∈ I 0 , so the family Fz consists

of one section F : Ω → J r (Rn , Rq ).

N
Figure 3.4. The sets Aδ (ci ) ∪ I l+1 and Ω (gray) in the case
i=1
n = 2, k = 1, l = 0.

N
Figure 3.5. The set Aδ (ci ) ∪ I l+1 in the case n = 3, k = 2, l = 1.
i=1
3.5. Proof of Theorem 3.2.1 33
3.5. Proof of Theorem 3.2.1

The following lemma is a corollary of the Gluing Lemma 3.4.1.
3.5.1. (Induction Step for Theorem 3.2.1) Let Ul be a neighborhood of

the cube I k ⊂ Rn . Suppose that a section
F l : Ul → J r (Rn , Rq )
is holonomic over Op ∂I k and holonomically extended along the cubes
y × I l ⊂ Rn , y ∈ I k−l .
Then for any given neighborhood U (F l ) of the image F l (Ul ) ⊂ J r (Rn , Rq )

there exists a diffeomorphism
hl+1 : Rn → Rn , hl+1 (x1 , . . . , xn ) = (x1 , . . . , xn−1 , xn + ϕl+1 (x1 , . . . , xn ))
supported in Ul and a section

l+1 → J r (Rn , Rq ),
Fl+1 : U l+1 ⊂ Ul
Op hl+1 (I k ) ⊂ U
such that
• hl+1 = Id and Fl+1 = F l on Op ∂I k ;
l+1 ) ⊂ U (F l );
• F l+1 (U
• F l+1 is holonomically extendable along the cubes hl+1 (z × I l+1 ),
z ∈ I k−l−1 .
Remark. In particular, for l = k − 1 the new section Fk is holonomic as

a whole section.
Proof. It follows from Lemma 3.4.1 that there exists a diffeomorphism
hl+1 : Rn → Rn , hl+1 (x1 , . . . , xn ) = (x1 , . . . , xn−1 , xn + ϕl+1 (x1 , . . . , xn ))
supported in Ul such that hl+1 = Id on Op ∂I k and for each z ∈ I k−l−1 the

image hl+1 (z × I l+1 ) is contained in Ωz (see Figure 3.6). Then the section
Fz constructed in Lemma 3.4.1 is defined on Op hl+1 (z × I l+1 ), hence the
section

Fz (z, t, x), (z, t, x) ∈ (Op hl+1 (I k )) ∩ (I k−l−1 × Rn−k+l+1 )
F (z, t, x) =
l+1
F (z, t, x), (z, t, x) ∈ Op (∂I k )
has the required properties.

Figure 3.6. The image h(I) in the case n = 2, k = 1, l = 0.
Proof of Theorem 3.2.1. We will prove Theorem 3.2.1 by successively

applying Lemma 3.5.1 for l = 0, 1, . . . , k − 1.
For l = 0 we set U0 = Nδ (I k ), F 0 = F , and U (F 0 ) = Nε (F 0 (U0 )). Lemma
3.3.1 provides a family of holonomic extensions along the points y ∈ I k
and therefore we can apply Lemma 3.5.1 with l = 0. For l = 1 we cannot
apply Lemma 3.5.1 directly because the new section F1 is defined near the
deformed cube h1 (I k ) rather than the original one. Note, however, that any
diffeomorphism h : Rn → Rn induces bundle isomorphisms
h∗ , h∗ = (h−1 )∗ : J r (Rn , Rq ) → J r (Rn , Rq )
1 ), F 1 =
and therefore we can straighten F1 assuming that U1 = (h1 )−1 (U
h∗1 (F1 ), and U (F 1 ) = h∗1 (U (F 0 )). The section
F 1 : U1 → J r (Rn , Rq )
coincides with F near ∂I k and is holonomically extendable along the inter-

vals y × I 1 ⊂ Rn , y ∈ I k−1 . Therefore, we can apply Lemma 3.5.1 with
l = 1. Next, we repeat the process with l = 2, . . . , k − 1. Each time we
apply 3.5.1 after a preliminary straightening of the section Fl . In the last
step we apply 3.5.1 to
k−1 ),
Uk−1 = h∗k−1 (U F k−1 = h∗k−1 (Fk−1 ),
U (F k−1 ) = h∗k−1 (U (F k−2 )) = (h1 ◦ · · · ◦ hk−1 )∗ (U (F 0 ))
and get the holonomic section
k → J r (Rn , Rq ),
Fk : U k ⊂ Uk−1 ,
Op hk (I k ) ⊂ U Fk (U
k ) ⊂ U (F k−1 ).
The diffeomorphism h and section F required in Theorem 3.2.1 are then

given by the formulas
h = h1 ◦ · · · ◦ hk , F = (h1 ◦ · · · ◦ hk−1 )∗ F k .
3.6. Proof of the Gluing Lemma 35
3.6. Proof of the Gluing Lemma

We will consider first the case l = k −1 in which F is holonomically extended
along the cubes t × I k−1 ⊂ I k , and thus we need to construct an entirely
holonomic section F . Then, in the general case, we will rewrite the proof
almost literally, just incorporating all the way the variable z ∈ I k−l−1 into
the notation.
A. The case l = k − 1. In this case y = t ∈ I and the notation which we

introduced at the beginning of this section takes the form
Uδ (t) = Nδ (t × I k−1 ), Vδ (t) = Nδ (t × ∂I k−1 ),

Aδ (t) = Uδ1 (t) \ Vδ (t) ∩ (t × Rn−1 ).

We also set Wδ (t) = Uδ (t) \ Uδ1 (t) ∪ Vδ (t).
In what follows δ is fixed and we will write U (t), A(t), V (t), and W (t)
instead of Uδ (t), Aδ (t), Vδ (t), and Wδ (t), respectively.
We are given a section F : Op I k → J r (Rn , Rq ) which is holonomic over

Op ∂I k and a family of holonomic sections Ft : Uδ (t) → J r (Uδ (t), Rq ), t ∈ I,
such that
• Ft |(t×I k−1 )∪Vδ (t) = F |(t×I k−1 )∪Vδ (t) ;
• Ft = F |Uδ (t) for t ∈ Op ∂I.
Set Ft = F1 for t > 1. Note that
max Ft+σ (x) − Ft (x) → 0,
t∈I, x∈U (t+σ)∩U (t) σ→0
hence we have the following

3.6.1. (Interpolation Property) For any ε > 0 there exist a number
σ = 1/N and a family of holonomic sections
Ftτ : U (t) → J r (Rn , Rq ), t ∈ I, τ ∈ [0, σ],
such that
(a) Ft0 = Ft for all t ∈ I;
(b) Ftτ |W (t) = Ft |W (t) for all t ∈ I and τ ∈ [0, σ];
(c) Ftτ − Ft C 0 < ε for all t ∈ I and τ ∈ [0, σ];
(d) Ftτ |Op (t×I k−1 ) = Ft+τ |Op (t×I k−1 ) for all t ∈ I and τ ∈ [0, σ] (see
Figures 3.7 and 3.8).
Remark. Note that 3.6.1(d) automatically implies σ < δ. In fact, in
most cases σ δ.
Figure 3.7. The graphs of the sections Ft (schematically) in the

case n = 2, k = 1, l = 0.
Figure 3.8. The graphs of the sections Ft (left picture) and the
sections Ft and Ftσ (right picture) over I k ∩ U (t) (schematically)
in the case n = 2, k = 1, l = 0.
For i = 0, 1, . . . , N set
Bi = iσ × I k−1 .
For i = 1, . . . , N set
Fiold = Fiσ Finew = Fiσ

σ
,
σ 2i − 1
ci = iσ − = , Ai = A(ci ), Δi = ((i − 1)σ, iσ),
2 2N
Δ−i = ((i − 1)σ, ci ], Δi = [ci , iσ),
+
i = U (iσ) ∩ (Δi × Rn−1 ),

U
U i ∩ (Δ− × Rn−1 ), U
− = U + = U
i ∩ (Δ+ × Rn−1 ),
i i i i
U i ∩ (ci × Rn−1 ) = U
= U − ∩ U
+
i i i
(see Figure 3.9).

Figure 3.9. The set W (iσ) ⊂ U (iσ) (left picture, gray color) and
i , U
the sets U −, U
+ (right picture, gray color), U
, Ai in the case
i i i
n = 2, k = 1, l = 0.
The set Ui \ Ai lies in W (iσ) and therefore, according to the Interpolation
Property 3.6.1, the section Finew coincides with Fiold on Ui \ Ai . Hence the
formula

−,
Fiold (x), x ∈ U
F(x) = i
+,
Finew (x), x ∈ U i

N
i \ Ai ) (see Figure 3.10).
i = 1, . . . , N , defines a holonomic section over (U
i=1
We also have
old
Fi+1 = Finew
over Op Bi for i = 1, . . . , N − 1, hence F extends continuously to

N
N
i \ Ai ) ∪
(U Op Bi = Ω
i=1 i=0
(see Figures 3.11 and 3.12). By Property 3.6.1(c) we have F − F |Ω C 0 <
ε.
Figure 3.10. The section F over U

i \ Ai in the case n = 2, k = 1,
l = 0.

N
i \ Ai ) ∪ Op Bi (gray color) in the
N
Figure 3.11. The set (U
i=1 i=0
case n = 2, k = 1, l = 0.
over
N
i \ Ai ) ∪ Op Bi in the
N
Figure 3.12. The section F (U
i=1 i=0
case n = 2, k = 1, l = 0.
Exercises.
1. Prove that for σ > δ1 one can construct an approximating section F on

Op I k .
2. The previous exercise may lead to the (dubious) conclusion that by choos-
ing a sufficiently small δ1 , one always can construct the approximating sec-
tion F over Op I k . Why does this idea fail?
B. The general case. We will proceed parametrically with z ∈ I k−l−1 to

produce the families of diffeomorphisms hz and holonomic sections Fz . We
repeat the previous proof almost literally, just systematically inserting the
parameter z in our notation.
Recall that for y = (z, t) ∈ I k−l−1 × I and δ > 0, we have the following
notation:
Uδ (z, t) = Nδ (z × t × I l ), Vδ (z, t) = Nδ (∂(z × t × I l )) ,

Aδ (z, t) = Uδ1 (z, t) \ Vδ (z, t) ∩ ((z, t) × Rn−k+l ),
where δ1 = θδ for a fixed positive θ < 1. We also set

Wδ (z, t) = Uδ (z, t) \ Uδ1 (z, t) ∪ Vδ (z, t).
As in the nonparametric case, we fix δ and write U (z, t), A(z, t), V (z, t),
and W (z, t) instead of Uδ (z, t), Aδ (z, t), Vδ (z, t), and Wδ (z, t), respectively.
Set Fz,t = Fz,1 for t > 1. Note that

max Fz,t+σ (x) − Fz,t (x) → 0,
(z,t)∈I k−l , x∈U (z,t+σ)∩U (z,t) σ→0
hence, similarly to Property 3.6.1, we have

3.6.2. (Parametric Interpolation Property) For any ε > 0 there exist
a number σ = 1/N and a family of holonomic sections
τ
Fz,t : U (z, t) → J r (Rn , Rq ), (z, t) ∈ I k−l−1 × I, τ ∈ [0, σ],
such that
z,t for all (z, t) ∈ I
0 =F
(a) Fz,t k−l−1 × I;
τ |
W (t) = Fz,t |W (t) for all (z, t) ∈ I
(b) Fz,t k−l−1 × I and τ ∈ [0, σ];
τ −F
(c) Fz,t z,t C 0 < ε for all (z, t) ∈ I
k−l−1 × I and τ ∈ [0, σ];
τ |
Op (z×t×I l−1 ) = Fz,t+τ |Op (z×t×I l−1 ) for all (z, t) ∈ I
(d) Fz,t k−l−1 × I
and τ ∈ [0, σ].
Similar to the nonparametric case, we set for i = 0, 1, . . . , N ,

Bz,i = z × iσ × I l ,
and for i = 1, . . . , N and z ∈ I k−l−1
old new σ
Fz,i = Fz,iσ , Fz,i = Fz,iσ ,
z,i = U (z, iσ) ∩ (z × Δi × Rn−k+l ),
U
U ( z, i) ∩ (z × Δ− × Rn−k+l ),
− = U + = U
U z,i ∩ (z × Δ+ × Rn−k+l ),
z,i i z,i i
U z,i ∩ (z × ci × Rn−1 ) = U
= U − ∩ U
+
z,i z,i z,i
where we keep using the notation
σ 2i − 1
ci = iσ − = , Δ−i = ((i − 1)σ, ci ], Δi = [ci , iσ).
+
2 2N
\A
The set Uz,i z,i lies in W (z, iσ) and therefore, according to the Interpo-
lation Property 3.6.2, the section Fz,inew coincides with F old on U \ A .

z,i z,i z,i
Hence the formula

old (x),
Fz,i x∈U − ,
Fz (x) = z,i
new (x), x ∈ U
Fz,i + ,
z,i

N
i = 1, . . . , N , defines a family of holonomic sections Fz over z,i \ Az,i ).
(U
i=1
We also have
old new
Fz,i+1 = Fz,i
over Op Bz,i for i = 0, . . . , N − 1, hence Fz extends continuously to

N
N
z,i \ Az,i ) ∪
(U Op Bz,i = Ωz .
i=1 i=0
3.7. Parametric holonomic approximation

It turns out that the Inductive Lemma 3.5.1 implies also the parametric
version of Theorem 3.2.1. Namely, we have
3.7.1. (Parametric holonomic approximation over a cube) Let Fu ,
u ∈ I m , be a family of sections
Op I k → J r (Rn , Rq )
parametrized by the cube I m . Suppose that k < n and the sections Fu are
holonomic over Op ∂I k for all u ∈ I m and are holonomic over Op I k for
u ∈ Op ∂I m . Then for arbitrarily small δ, ε > 0 there exists a family of
δ-small (in the C 0 -sense) diffeomorphisms
hu : Rn → Rn , hu (x1 , . . . , xn ) = (x1 , . . . , xn−1 , xn + ϕu (x1 , . . . , xn )),
and a family of holonomic sections
Fu : Op hu (I k ) → J r (Rn , Rq )
such that
• hu = Id and Fu = Fu on Op ∂I k for all u ∈ I m ;
• hu = Id and Fu = Fu for u ∈ Op ∂I m ;
• Fu − Fu |Op hu (I k ) C 0 < ε.
Proof. Consider the cube

I m+k = I m × I k ⊂ Rm × Rn = Rm+n .
Let J r (Rm+n |Rn , Rq ) be the bundle over Rm ×Rn whose restriction to u×Rn ,
u ∈ Rm , equals J r (Rn , Rq ). The family of sections
Fu : I k → J r (Rn , Rq )
3.8. Foliated holonomic approximation 41
can be viewed as a section

F : I m+k → J r (Rm+n |Rn , Rq ).
The section F lifts to a section
F : I m+k → J r (Rm+n , Rq ),
so that π ◦ F = F , where
π : J r (Rm+n , Rq ) → J r (Rm+n |Rn , Rq )
is the canonical projection. Moreover, the section F can be chosen holonomic

near ∂I m+k . Hence we can apply Theorem 3.2.1 to get an ε-approximation F
of F over a displaced cube h(I m+k ) for a δ-small diffeomorphism
h : Rm+n → Rm+n of the form
h(x1 , . . . , xm+n ) = (x1 , . . . , xm+n−1 , xm+n + ϕ(x1 , . . . , xm+n )).
Then the composition

F = π ◦ F : I m+k → J r (Rn+m |Rn , Rq )
can be viewed as the required family {Fu }u∈I m of holonomic approximations
of the family {Fu } near {hu (I k )}.
In the same way as Theorem 3.2.1 implies the Holonomic Approximation

Theorem 3.1.1, by induction over skeleta, Theorem 3.7.1 implies the Para-
metric Holonomic Approximation Theorem 3.1.2.
3.8. Foliated holonomic approximation

It will be useful to reformulate Theorem 3.1.2 in a stronger foliated (= leaf-
wise) form. A k-dimensional
foliation F on an n-dimensional manifold V
is a partition V = La into connected subsets La , called the leaves, which
locally look like the partition of Rn into the sets Rk × s, s ∈ Rn−k . Equiva-
lently (by Frobenius’s theorem), F can be defined as an integrable k-plane
distribution on V . See Section 10.1.A for further discussion of foliations.
Remark. For k+m ≥ n = dim V , an m-dimensional submanifold M ⊂ V
is called transverse to a k-dimensional foliation F of V if M is transverse to
all leaves of F in the sense of Section 2.1. For k + m < n this is impossible
unless M is empty; in this case we say that M is transverse to F if Tx M ∩
Tx L = {0} for each x ∈ M , where L denotes the leaf through x. This
definition corresponds, on the level of tangent spaces, to calling two linear
subspaces S, P ⊂ L of a vector space L transverse if S + P = L or S ∩ P =
{0}. Similarly, one defines transversality of a submanifold to a distribution,
and transversality of two foliations or distributions.
Let (V, F ) be a manifold equipped with a foliation and p : X → V a fibration.

(r)
The leafwise r-jet extension prF : XF → V of the fibration X → V is formed
by classes of r-tangency along the leaves of local sections of X. There is a
(r)
canonical projection pF : X (r) → XF and we have a commutative diagram
pF
X (r)B / X (r)
BB F
BB |||
BB |
BB || r
~ | pF
p r
|
V.
(r)
The leafwise r-jet extension Jfr|F : V → XF of a section f : V → X is
(r)
defined as the composition prF ◦ Jfr . Sections F : V → XF of the form
(r)
F = Jfr|F = pF ◦ Jfr : V → XF
for some f : V → X are called leafwise holonomic sections.
Theorem 3.7.1 generalizes to the foliated case as follows.

3.8.1. (Foliated holonomic approximation over a cube) Suppose that
the base Rn of the trivial fibration p : X = Rn × Rq → Rn is endowed with a
(r)
foliation F transverse to Rk ⊂ Rn , k < n. Let F : Op I k → XF be a section
which is leafwise holonomic on Op ∂I . Then for arbitrarily small δ, ε > 0
k
there exists a δ-small (in the C 0 -sense) leafwise diffeotopy hτ : Rn → Rn ,

τ ∈ [0, 1], and a leafwise holonomic section F : Op h1 (I k ) → XF such that
(r)
• hτ = Id and F = F on Op ∂I k ;
• F − F |Op h1 (I k ) C 0 < ε.
Proof. After a coordinate change, we may assume that the foliation F

contains everywhere the last coordinate direction ∂xn . First, we lift F to a
section G : Op I k → X (r) . Next, we apply Lemma 3.3.1 for k ≤ codim F
and Theorem 3.2.1 for k > codim F to get a δ-small diffeomorphism
h : Rn → Rn , h(x1 , . . . , xn ) = (x1 , . . . , xn−1 , xn + ϕ(x1 , . . . , xn ))
(in Lemma 3.3.1 we have h = Id) and a holonomic section G : Op h(I k ) →

X (r) with G − G|Op h(I k ) C 0 < ε. Then hτ = (1 − τ )Id + τ h is a leafwise
diffeotopy and the composition pF ◦ G is the required leafwise holonomic
section F .
In the same way as Theorem 3.2.1 implies the Holonomic Approximation

Theorem 3.1.1, by induction over skeleta, Theorem 3.8.1 implies
3.9. Refinement of the Holonomic Approximation Theorem 43
3.8.2. (Foliated Holonomic Approximation Theorem) Suppose that

that the base V of a fibration p : X → V is endowed with a foliation F . Let
A ⊂ V be a polyhedron of positive codimension such that all its strata are
transverse to the leaves of the foliation F . Let B ⊂ A be a closed subset and
(r)
F : Op A → XF a section which is leafwise holonomic on Op B. Then for
any positive functions δ, ε : V → R there exists a δ-small leafwise diffeotopy
hτ : V → V, τ ∈ [0, 1] and a leafwise holonomic section
F : Op h1 (A) → XF
(r)
such that
• hτ = Id and F = F on Op B;
• F − F |Op h1 (A) < ε.
3.9. Refinement of the Holonomic Approximation Theorem

In [ÁG18] Daniel Álvarez-Gavela proved the following improved version
of the Holonomic Approximation Theorem 3.1.1, which is important for
applications in symplectic geometry:
3.9.1. (Holonomic approximation for s-holonomic sections) Let A ⊂
V be a polyhedron of positive codimension, s < r, and F : V → X (r) a
section of the r-jet bundle pr : X (r) → V such that the projection
F (s) = prs ◦ F : V → X (r) → X (s)
is a holonomic section. Then for arbitrarily small δ, ε > 0 there exists a δ-
small (in the C 0 -sense) diffeotopy hτ : V → V , τ ∈ [0, 1], and a holonomic
section F : V → X (r) such that
(a) dist(F (v), F (v)) < ε for all v ∈ Op h1 (A);
(b) dist(F (s) (v), F (s) (v)) < ε for all v ∈ V ;
(c) hτ = id and F (s) = F (s) outside a slightly bigger neighborhood
Op A ⊃ Op h1 (A).
Remarks.
1. Theorem 3.9.1 also holds in relative and parametric forms.
2. The whole essence of Theorem 3.9.1 is condition (b). For s = 0 or without
condition (b), Theorem 3.9.1 is a essentially equivalent to Theorem 3.1.1.
3. D. Álvarez-Gavela in [ÁG18] proved also another extension of Theorem
3.1.1 for so-called ⊥-holonomic sections in the sense of [Gr86]. For instance,
if one begins with a section F : Op I n−1 → J r (Rn , Rq ) which coincides with
a holonomic section, except for one of its coordinate functions corresponding
∂r
to the pure r-th derivative ∂x r , then for the approximating holonomic section
n
F = J r (f) one can arrange in (b) that all derivatives of f are ε-close to the
∂ f r
corresponding components of F except ∂x r . Of course, closeness of the
n
remaining derivative in (a) can be achieved only on Op h1 (I n−1 ).

Chapter 4
Applications
The first two examples below illustrate Gromov’s homotopy principle for
open Diff V -invariant differential relations over open manifolds which we
formulate and prove later in Part 2 (see Section 8.3).
4.1. Functions without critical points

Let V be the annulus {1/4 ≤ x21 + x22 ≤ 4} ⊂ R2 .
4.1.1. There exists a family of functions ft : V → R, t ∈ [0, 1], such that
grad ft = 0, f0 (x1 , x2 ) = −x21 − x22 and f1 (x1 , x2 ) = x21 + x22 (see Figure 4.1).
Figure 4.1. The functions f0 and f1 .
Proof. The 1-jet space J 1 (V, R) equals V × R × R2 and we will identify

the last factor, which corresponds to the gradient of a function, with the
45
46 4. Applications
complex line C. Note that grad f0 = − grad f1 . The family of sections of

J 1 (V, R) defined by the formula
Ft = ((1 − t)f0 + tf1 , eiπt grad f0 )
joins F0 = Jf10 with F1 = Jf11 . For t = 0, 1 the section Ft is not holonomic.
We can reparametrize the family Ft to make it independent of t, and thus
holonomic, for t ∈ Op ∂I. Applying the Parametric Holonomic Approxima-
tion Theorem 3.1.2 with the unit circle A = S 1 ⊂ V one can construct a
family of holonomic ε-approximations Ft = J 1 : Ut → J 1 (V, R), where Ut is
ft
a neighborhood of a perturbed circle ht (S ). Moreover, one can choose Ft
1 1
and Ut such that Ut = V and Ft = Ft for t ∈ Op ∂I. For sufficiently small
ε the functions ft do not have critical points on Ut because
grad ft ≈ eiπt grad f0 = 0 near S 1 .
Let {ϕτt : V → V, τ ∈ [0, 1]}t∈[0,1] be a family of isotopies such that for each
t ∈ [0, 1] the isotopy ϕτt , τ ∈ [0, 1], shrinks V into the neighborhood Ut and
ϕτ0 = ϕτ1 = IdV . Then the family gt = ft ◦ ϕ1t consists of functions without
critical points on V and interpolates between f0 and f1 .
Exercise. Try to construct the required family explicitly.
4.2. Smale’s sphere eversion

Let dim V ≤ dim W . A map f : V → W is called an immersion if rank f =
dim V everywhere on V . If dim V = dim W , then an immersion V → W
is the same as a local diffeomorphism. Two immersions are called regularly
homotopic if they can be connected by a family of immersions.
For 0 < δ < 1 denote by V the thickened sphere

(1 − δ)2 ≤ x21 + x22 + x23 ≤ (1 + δ)2
in R3 . Let
inv : R3 \ 0 → R3 \ 0, inv(x) = x/x2 ,
be the inversion,
r : R3 → R3 , r(x1 , x2 , x3 ) = (x1 , x2 , −x3 ),
the reflection, and iV : V → R3 the inclusion.
4.2.1. (Smale’s sphere eversion, [Sm58]) The map
r ◦ inv ◦ iV : V → R3 ,
which inverts V outside in, is regularly homotopic to the inclusion iV : V →
R3 .
4.2. Smale’s sphere eversion 47
Remarks.
1. This counter-intuitive statement is a corollary of S. Smale’s celebrated
theorem [Sm58]. Equivalently it can be formulated by saying that the 2-
sphere in R3 can be turned inside out via a regular homotopy, i.e., via a
family of smooth but possibly self-intersecting surfaces. One can follow the
proof below to actually construct this eversion. However, there are much
more efficient ways to do that. The explicit process of the eversion became
the subject of numerous publications, videos, and computer programs.
2. The map
inv ◦ iV : V → R3 ,
which also everts V inside out, is not regularly homotopic to the inclusion
iV : V → R3 because these maps induce on V the opposite orientations.
Proof. Let f0 = iV and f1 = r ◦ inv ◦ iV . Both df0 and df1 have rank 3 and
induce the same orientation on V . Hence the sections
df0 , df1 : V → J 1 (V, R3 ) = V × R3 × M3×3
can be viewed as maps

V → R3 × SO(3).
These maps are homotopic because π2 (SO(3)) = 0.1 Let Ft be a homotopy
connecting F0 = df0 and F1 = df1 . The deformation Ft can be assumed
holonomic for t near ∂I. Applying Theorem 3.1.2 with A = S 2 one can
construct a family of holonomic ε-approximations
Ft = Jf1 : Ut → J 1 (V, R3 ),

t
where Ut is a neighborhood of a perturbed sphere h1t (S 2 ). Moreover, one

can choose Ft and Ut such that Ut = V and Ft = Ft for t ∈ Op ∂I. If ε is
sufficiently small, then ft is a regular homotopy. As in the previous example,
we can compose ft with a family of contractions of V into the neighborhoods
Ut and get the desired regular homotopy gt : A → R3 which connects f0 and
f1 .
Exercise (S. Smale, [Sm58]). Prove that every immersion S 2 → R3 is

regularly homotopic to the standard embedding S 2 → R3 .
1 In fact, the homotopy between df and df can be easily constructed explicitly without
0 1
referring to the computation of π2 (SO(3)).
48 4. Applications
4.3. Approximate integration of tangential homotopies

Let
π : Grn W → W
be the Grassmannian bundle of n-planes tangent to a q-dimensional manifold
W , q > n, and V an n-dimensional manifold. Given a monomorphism
(= fiberwise injective homomorphism) F : T V → T W , we will denote by
GF the corresponding map V → Grn W . Thus the tangential Gauss map
associated with an immersion f : V → W can be written as Gdf .
In what follows we assume that V ⊂ W is an embedded submanifold and
denote by f0 the inclusion iV : V → W . We also assume that the manifolds
W and Grn W are endowed with Riemannian metrics.
A homotopy Gt : V → Grn W such that G0 = Gdf0 and π ◦ Gt = f0 is called
a tangential homotopy of the inclusion f0 .
4.3.1. (Approximate integration of tangential homotopies) Let K ⊂
V be a polyhedron of positive codimension and Gt : V → Grn W a tangential
homotopy. Then one can approximate Gt near K by an isotopy of embed-
dings in the following sense: for arbitrarily small δ, ε > 0 there exists a
δ-small diffeotopy {hτ : V → V }τ ∈I and an isotopy of embeddings
ft : Op V K
→ W, t ∈ I, where K = h1 (K) and f0 = f0 | ,
Op V K
such that the homotopy
Gdft : Op V K
→ Grn W
is ε-close to the tangential homotopy Gt |Op V K .
Remark. The relative and parametric versions of Theorem 4.3.1 are also
true.
Proof. Let us first assume that the homotopy Gt is small in the following
sense: the angle between Gt1 (v) and Gt2 (v) is less than π4 for all v ∈ V and
t1 , t2 ∈ I.
Let X be a tubular neighborhood of V in W and π : X → V the normal
projection. Let us recall (see Section 1.4) that the space X (1) of 1-jets of
sections V → X can be interpreted as the space of n-planes tangent to X
which are nonvertical, i.e., transverse to the fibers of the fibration π : X →
V . The assumption on the angles implies that Gt (v) is nonvertical for all
t, v. Hence the inclusion f0 : V → X together with the tangential homotopy
Gt : V → Grn W can be viewed as a homotopy of sections Ft : V → X (1) ,
t ∈ I.
For arbitrarily small ε and δ one can construct, using Theorem 3.1.1, a
holonomic ε -approximation F of F1 over Op h1 (K), where {hτ : V → V }τ ∈I
4.3. Approximate integration of tangential homotopies 49
is a δ -small diffeotopy. The 0-jet part f = p10 ◦ F of the section F is

automatically an embedding because it is a section of the normal bundle.
Identifying the fibers of the fibration π : X → V with normal spaces to V ,
we consider the linear homotopy ft , t ∈ I, connecting f0 |Op h1 (K) with f.
If ε , δ are chosen sufficiently small, then the isotopy ft has the required
properties.
In the general case we subdivide the interval I,
−1
N
I= Ij , Ij = [j/N, (j + 1)/N ],
j=0
so that on each subinterval Ij the tangential homotopy Gt is small in the

above sense, and we consecutively repeat the above construction on each of
these intervals.
More precisely, let us extend the homotopy Gt to a homotopy
Gt : X0 → Grn W
defined on a tubular neighborhood X0 of V in W . We can assume that the
homotopy Gt is also small on each interval Ij , j = 0, . . . , N − 1. We begin
with the interval I0 , set V0 = V and K0 = K, denote by π0 : X0 → V0 the
normal projection, and use Theorem 3.1.1 to construct a δ1 -small diffeotopy
{hτ0 : V0 → V0 }τ ∈I and a family of sections
ft1 : Op h10 (K0 ) → X0 , t ∈ I0 ,
such that Gdft1 (v) is ε1 -close to Gt (ft1 (v)), v ∈ Op h10 (K0 ). Set
1
1
V1 = f1/N Op h10 (K0 ) and K1 = f1/N 1
h0 (K0 ) .
If ε1 is chosen sufficiently small, then for any v1 ∈ V1 and t ∈ I1 the angle
between the planes Gt (v1 ) and Tv1 V1 is still bounded by π4 . Hence, we can
repeat now the first step on the interval I1 by choosing a tubular neighbor-
hood X1 of V1 in X0 and for sufficiently small ε2 , δ2 construct a δ2 -small
diffeotopy {hτ1 : V1 → V1 }τ ∈I and a family of sections
ft2 = Op h11 (K1 ) → X1 , t ∈ I1 ,
of the fibration π1 : X1 → V1 such that Gdft2 (v1 ) is ε2 -close to Gt (ft2 (v1 ))
for all v1 ∈ V1 and t ∈ I1 . Set
1
2
V2 = f2/N Op h11 (K1 ) and K2 = f2/N2
h1 (K1 ) .
Continuing this process for i = 2, . . . , N −1, we construct a δ-small diffeotopy
hτ : V → V, τ ∈ I, which deforms K0 into
= π0 ◦ · · · ◦ πN −1 (KN );
K
see Figure 4.2.
50 4. Applications
K0 π0 (K1) π 0 π1(K2)
V0
π 0(V1 )
Figure 4.2. The sets K0 , V0 , π0 (K1 ), π0 (V1 ), and π0 ◦ π1 (K2 ).
The required isotopy ft , t ∈ I, can now be defined by the formula

⎧
⎪
⎪ft1 |Op K , t ∈ I0 ;
⎪
⎪
⎨f 2 ◦ f 1 | t ∈ I1 ;
,
ft = t 1/N Op K
⎪
⎪... ...
⎪
⎪
⎩f N ◦ · · · ◦ f 2 ◦ f 1 |
t 2/N , t ∈ IN −1 .
1/N Op K
Remark. By choosing a connection in the principal SO(n)-bundle asso-

ciated with the tautological n-dimensional vector bundle over Grn W , we can
canonically lift any tangential homotopy Gt : T V → Grn W to a homotopy
of fiberwise isometric monomorphisms Ft : T V → T W such that GFt = Gt .

Exercise. Prove that one can approximate the homotopy Ft near K ⊂ V
by an almost isometric isotopy ft : Op V K → W .
4.4. Open manifolds

For further applications we need some information about open manifolds.
A manifold V is called open if there are no closed manifolds among its

connected components. In particular, any path connected manifold V with
nonempty boundary is open in this sense.
We say that a path p : [0, ∞) → V connects v = p(0) with ∞ if either p is

proper or lim p(t) ∈ ∂V . A point v ∈ V is said to have an exit to infinity if
t→∞
there exists a path which connects v with infinity.
4.4. Open manifolds 51
Given an open manifold V , a stratified subset V0 ⊂ V is called a pseudo-core

of V if codim V0 ≥ 1 and there exists an isotopy ϕt : V → V which brings
V to Op V0 . If, moreover, the isotopy ϕt can be chosen to be fixed on V0 ,
then we will call V0 a (genuine) core of V .
Figure 4.3. Paths connecting the barycenters with infinity.
4.4.1. (Existence of a pseudo-core) If an n-dimensional manifold V is

open, then the (n − 1)-skeleton V n−1 ⊂ V of any triangulation is a pseudo-
core.
Figure 4.4. A deformation which brings V into an arbitrarily small

neighborhood of (n − 1)-skeleton of a triangulation.
Proof. The case n = 1 is obvious, so let us assume n ≥ 2. Fix a triangula-

tion of V and disjoint paths [0, ∞) → V which connect all the barycenters
of the n-simplices with ∞; see Figure 4.3. Using these paths we can de-
form V by an isotopy ϕt into the complement of the set of barycenters of
the n-simplices and after that into an arbitrarily small neighborhood of the
(n − 1)-skeleton V n−1 ; see Figure 4.4. Note that such an isotopy is not fixed
on V n−1 and the image ϕ1 (V ) is not necessarily homotopy equivalent to
V n−1 ,
52 4. Applications
Remarks.
1. Theorem 4.4.1 belongs to mathematical folklore (see also [Gr86, p. 37,
Exercise (b)]).
2. Using this construction, it is easy to prove that every open manifold V
admits a smooth function without critical points (see [Gr86, p.37, Exercise
(c)]). Indeed, it suffices to start with a function whose critical points do not
belong to the (n − 1)-skeleton of the fixed triangulation.
In what follows the existence of a pseudo-core will usually be sufficient for

us when working with open manifolds. But in some cases we will need a
genuine core.
4.4.2. (Existence of a core) Every open manifold V admits a core V0 . If
V is connected and noncompact with ∂V = ∅, then we can choose V0 such
that ∂V ⊂ V0 .
We outline two versions of the proof of this statement.
A. (M. Morse and J. Milnor). We may assume without loss of generality

that V is connected. Assume first that V is noncompact. Then V admits
a proper Morse function f : V → R+ = [0, ∞) without local maxima. This
follows from [Mi65], noting that cancellation of local maxima works in any
dimension. Choose a Riemannian metric on V , and let V0 be the stratified
set which is the union of stable manifolds of the gradient vector field ∇f for
all critical points of the function f . We can pick a function ϕ : V → R+
with ϕ|V0 ≡ 0 such that the flow of the vector field −ϕ∇f is fixed on V0 and
brings V to Op V0 in time 1. Since codim V0 ≥ 1, this shows that V0 ⊂ V is
a core of V . If in addition ∂V = ∅, then we can choose the function f to
be constant and equal to its minimum on ∂V , hence ∂V ⊂ V0 .
In the case that V is compact with ∂V = ∅, we apply the above construction
to a Morse function f : V → [0, 1] with regular level set ∂V = f −1 (1) and
no other local maxima.
B. (J. H. C. Whitehead). In [Whi61] Whitehead proved that every

noncompact connected n-dimensional triangulated manifold V has a core
V0 ⊂ V n−1 with ∂V ⊂ V0 . Whitehead’s proof is, in a sense, an improved
version of the construction of a pseudo-core in Theorem 4.4.1.
Let us cite the necessary definitions from [Whi61]. An infinite graph G is
called reducible if it contains an edge e such that G \ Int(e) has no com-
pact component. A ray is a graph which is isomorphic to the half line R+
decomposed into the 1-simplices [i, i + 1]. A special tree is a locally finite
graph which is the union of a ray and a (possibly empty) set of mutually
4.5. Directed embeddings of open manifolds 53
disjoint finite trees each of which has just one vertex in common with the
ray. Every connected infinite locally finite graph contains a ray. (Exercise:
Prove this.) Hence, it is a special tree if and only if it is irreducible.
The idea of the proof is to consider the graph G whose vertices are the
barycentres of n-simplices and whose edges connect the barycentres of neigh-
boring n-simplices (i.e., G is the 1-skeleton of the cell complex dual to the
fixed triangulation of V ), enumerate the edges of the graph G, and define
the sequence of graphs G = G0 , G1 , . . . as follows. If Gr is irreducible, then
Gr+1 = Gr ; otherwise Gr+1 = Gr \ Int(e) where e is the first edge of Gr
such that Gr+1 , thus defined, has no compact component. As a result, the
original graph G will turn, without losing a single vertex, into a finite or
countable set of disjoint special trees G ⊂ G. Then one can check that
V0 = V n−1 \ Int(L) is the required core, where V n−1 ⊂ V is the (n − 1)-
skeleton and L ⊂ V n−1 consists of all (n − 1)-simplices that intersect G .
4.5. Directed embeddings of open manifolds

Let A ⊂ Grn W be an arbitrary subset. An immersion f : V → W is called
A-directed if Gdf sends V into A. If V is an oriented manifold, then we
can also consider A-directed immersions where A is an arbitrary subset in
the Grassmannian Gr n W of oriented tangent n-planes to a q-dimensional
manifold W . Note that by an embedding of an open manifold, we always
mean an embedding onto a locally closed submanifold of the target manifold.
For A-directed embeddings Gromov proved in [Gr86] via his convex inte-
gration technique the following theorem.
4.5.1. (A-directed embeddings of open manifolds) Let V be an open
n-manifold and A ⊂ Grn W an open subset. If f0 : V → W is an embedding
whose tangential lift
G0 = Gdf0 : V → Grn W
is homotopic to a map
G1 : V → A ⊂ Grn W,
then f0 can be isotoped to an A-directed embedding f1 : V → W . More-
over, given a core K ⊂ V of the manifold V , the isotopy ft can be chosen
arbitrarily C 0 -close to f0 on Op K.
Remarks.
1. Gromov’s proof is discussed in detail in [Sp00]. C. Rourke and B. Sander-
son gave two independent proofs of this theorem in [RS97] and [RS00].
54 4. Applications
2. The parametric version of Theorem 4.5.1 is also true. The relative version
for (V, V0 ) is false in general, but it is true if each connected component of
V \V0 has an exit to ∞ other than V0 , i.e., when Int V \Op V0 has no compact
connected components.
Proof. Theorem 4.5.1 follows almost immediately from Theorem 4.3.1. In-
deed, let K ⊂ V be a core of V , i.e., a codimension ≥ 1 subcomplex in V
such that V can be compressed into an arbitrarily small neighborhood of
K by an isotopy fixed on K. Using Theorem 4.3.1 we can approximate Gt
near K = h1 (K) by an isotopy ft : Op V K → W . For a sufficiently close
approximation the image Gdf1 (Op V K) belongs to A. In order to construct
the required isotopy ft , we first compress V into Op V h1 (K) and then apply
ft .
Remark. The theorem is valid also in the case when the homotopy
Gt covers an arbitrary isotopy gt : V → W instead of the constant isotopy
gt ≡ f0 . Indeed, one can apply the previous proof to the pull-back homotopy
(dĝt )−1 ◦ Gt and the pull-back set A = ĝ ∗ (A), where ĝt : W → W is a
1
diffeotopy which extends the isotopy gt : V → W underlying Gt .
The following version of Theorem 4.5.1 will be useful for applications which
we consider in Section 20.1.
4.5.2. Let A ⊂ Grn W be an open subset, V an open manifold, and f0 : V →

W an embedding whose differential F0 = df0 is homotopic via a homotopy
of monomorphisms Ft : T V → T W , bs Ft = f0 , to a map F1 with
GF1 (V ) ⊂ A.
Then f0 can be deformed by an isotopy ft : V → W to a A-directed em-

bedding f1 : V → W such that F1 is homotopic to df1 through a homotopy
of monomorphisms Ft : T V → T W , bs Ft = ft , with GFt (V ) ⊂ A for all
t ∈ [0, 1].
Proof. Let ft : Op V K

→ W be the isotopy constructed in Theorem 4.3.1.
It is sufficient to construct a homotopy Ψt between F1 |Op V K and df1 such
that GΨt (Op V K) ⊂ A for all t. The existence of the underlying tangential
homotopy G t = GΨt follows from the C 0 -closeness of the maps GF1 |
Op V K
and Gdf1 , while the existence of the covering homotopy Ψt for G t follows
from the C -closeness of the homotopies GFt |Op V K and Gdft .
0
4.6. Directed embeddings of closed manifolds 55
4.6. Directed embeddings of closed manifolds

For some special sets A ⊂ Grn W Theorem 4.5.1 implies a theorem about
A-directed embeddings of closed manifolds.
Let n < m ≤ q. An open set A ⊂ Grn W is called m-complete if there exists

an open set A ⊂ Grm W such that A = Grn L.
A
L∈
Example. Suppose that n < m < q. Let the set A0 ⊂ Grn Rq0 consist of
all n-planes intersecting trivially the subspace L = 0 × Rq−m ⊂ Rq and let
A = Rq × A0 ⊂ Rq × Grn Rq0 = Grn Rq .
Then A is m-complete with A = Rq × A0 , where A0 consists of all m-planes
η ⊂ Rq such that L ∩ η = {0}.
4.6.1. (A-directed embeddings of closed manifolds) Let A ⊂ Grn W
be an open set which is m-complete for some m, n < m < q. Then the
statement of Theorem 4.5.1 holds for any closed n-dimensional manifold
V.
Proof. Let us fix some notation. Denote by Grm,n W the manifold of all

(m, n)-flags on W , where each flag is a pair of tangent planes (Lm , Ln ) in
Tw W such that Ln ⊂ Lm . Denote by π : Grm,n W → Grm W and π :
Grn,m W → Grn W the projections. Set
Ā = {(L, L) | L ∈ A ⊂ Grm W, L ∈ Grn L } ⊂ Grn,m W,
where A is the set implied by the definition of m-completeness. Note that
π(Ā) = A and π(Ā) = A.
Let Gt : V → Grn W be a homotopy between the tangential lift G0 = Gdf0
of the embedding f0 and the map G1 : V → A. Suppose first that the map
G1 : V → A ⊂ Grn W
lifts to a map
Ḡ1 : V → Ā ⊂ Grm,n W.
Then the homotopy Gt lifts to a homotopy Ḡt : V → Grm,n W , t ∈ [0, 1].
We have Gt = π ◦ Ḡt . Set Gt = π ◦ Ḡt , t ∈ [0, 1]. Let N be the total
space of the vector bundle over V whose fiber over a point v ∈ V is the
normal space to G1 (v) in G1 (v). The embedding f0 can be extended to an
embedding f0 : Op N V → W such that the tangential lift Gdf0 coincides
with G0 over V . Hence we can apply Theorem 4.5.1 to construct an isotopy
ft : Op N V → W such that f1 is an A-directed embedding. Then the
restriction ft = ft |V is an isotopy between f0 and an A-directed embedding
f1 : V → W .
56 4. Applications
In general, we cannot guarantee the existence of the global map Ḡ1 : V → Ā

which covers the map G1 : V → A. However, one can avoid this problem
using the following localization trick. Theorem 4.5.1 allows us to construct
the required isotopy ft in a neighborhood Op K of the (n − 1)-skeleton of
a triangulation of V . Moreover, the proof of Theorem 4.5.1 provides an
isotopy whose tangential lift Gdft is C 0 -close to Gt |Op K . Hence, one can
assume from the very beginning that our original homotopy Gt is constant
over Op K and we need to construct a required isotopy ft on each top-
dimensional simplex Δ of the triangulation keeping ft constant on Op ∂Δ.
If the simplices of the triangulation of V are sufficiently small, then the map
G1 : Δ → A ⊂ Grn W lifts to a map Ḡ1 : Δ → Ā ⊂ Grm,n W . It remains to
notice that the previous (global) construction of the isotopy ft also works
in the extension form.
Example. Suppose that n < m < q. Theorem 4.6.1 implies that any
closed n-dimensional submanifold V ⊂ Rq whose tangent planes can be
rotated into planes projecting injectively to Rm × 0 ⊂ Rq along 0 × Rq−m
can be perturbed via an isotopy so that its projection to Rm × 0 becomes
an immersion.
Two subbundles ξ, η ⊂ T W are called transverse if the composition map

ξ → T W → T W/η is surjective if dim ξ + dim η ≥ dim W , and injective if
dim ξ + dim η ≤ dim W .
4.6.2. (Generalization: closed submanifolds transverse to distribu-
tions) Let ξ ⊂ T W be a distribution on a q-dimensional manifold W with
codim ξ = k. Let n < k. Then for any closed n-dimensional submanifold
V ⊂ W whose tangent bundle T V is homotopic inside T W to a subbun-
dle τ ⊂ T W transverse to ξ, one can perturb V via an isotopy to make it
transverse to ξ.
Remark. The relative and the parametric versions of Theorem 4.6.2 are
also true.
4.7. Approximation of differential forms by closed forms

A. Formal primitive of a differential form. Let us recall that a dif-
ferential p-form ω on a manifold V can be viewed as a section of the fi-
bration Λp V → V . In particular, 1-forms are sections of the fibration
Λ1 V = T ∗ V → V .
The exact p-forms on V and the holonomic sections of the fibration

(Λp−1 V )(1) → V
4.7. Approximation of differential forms by closed forms 57
are closely related to each other. Indeed, the exterior differentiation

d
Sec Λp−1 V → Sec Λp V
can be written as the composition
J1
D
Sec Λp−1 V → Sec (Λp−1 V )(1) → Sec Λp V,
is induced by a homomorphism of bundles over V ,
where the map D
D
(Λp−1 V )(1) → Λp V,
which is called the symbol of the operator d. For example, for p = 2 choosing
local coordinates, we can identify the fiber of (Λ1 Rn )(1) → Λ1 Rn with the
space of n × n matrices, the fiber of Λ2 Rn → Rn with the space of skew-
symmetric n × n matrices, and D(A) = A − AT .
The map D : (Λp−1 V )(1) → Λp V is an affine fibration. In particular, any

section ω : V → Λp V can be lifted, uniquely up to homotopy, to a section
Fω : V → (Λp−1 V )(1) such that D ◦ Fω = ω. It is useful to think of Fω as a
formal primitive of ω. Therefore we can say that any p-form has a formal
primitive, or that any p-form is formally exact.
Note that there are no restrictions on the underlying section bs Fω . In other

words, given any p-form ω and an arbitrary (p − 1)-form α : V → Λp−1 V ,
one can construct a formal primitive Fω such that bs Fω = α.
B. Approximation of differential forms by closed forms. The propo-

sitions which we formulate below are rather technical. An important appli-
cation will be given in Section 18.2.
4.7.1. (Approximation by exact forms) Let K ⊂ V be a polyhedron of
codimension ≥ 1 and ω a p-form. Then there exists an arbitrarily C 0 -small
= h1 (K)
diffeotopy hτ : V → V such that ω can be C 0 -approximated near K
by an exact p-form ω = dα. Moreover, given a (p − 1)-form α on V , one
can choose α 0
to be C -close to α near K.
Proof. Take a formal primitive Fω for ω such that bs Fω = α, choose its

holonomic approximation Jα1 along K = h1 (K) ⊂ V , where hτ is an arbi-
to the whole manifold V . Then
trarily C 0 -small diffeotopy, and extend α
= d
ω α is the desired exact form.
Proposition 4.7.1 implies

4.7.2. (Approximation by closed forms) Let K ⊂ V be a polyhedron of
codimension ≥ 1. Let ω be a p-form on V and a ∈ H p (V ) a fixed cohomology
class. Then there exists an arbitrarily C 0 -small diffeotopy hτ : V → V ,
58 4. Applications
= h1 (K) by a closed
t ∈ [0, 1], such that ω can be C 0 -approximated near K
p-form ω ∈ a.
Indeed, one can take a closed form Ω ∈ a, apply the previous proposition to
the form θ = ω − Ω, and then take ω = θ + Ω.
The parametric versions of Propositions 4.7.1 and 4.7.2 are also valid. In
particular,
4.7.3. (Parametric approximation by exact forms) Let K ⊂ V be
a polyhedron of codimension ≥ 1 and {ωu }u∈Dk a family of p-forms such
that {ωu = dαu }u∈∂Dk . Then there exists a family of arbitrarily C 0 -small
diffeotopies
{hτu : V → V, τ ∈ [0, 1]}u∈Dk , where {hτu = IdV , τ ∈ [0, 1]}u∈∂Dk
such that {ωu }u∈Dk can be C 0 -approximated near K u = h1u (K) by a family
of exact p-forms {ωu = dαu }u∈Dk such that {αu = αu }u∈∂Dk . Moreover,
given a family of (p − 1)-forms {αu }u∈Dk on V , which extends the family
{αu }u∈∂Dk , one can choose the family {
αu }u∈Dk to be C 0 -close to {αu }u∈Dk
near Ku.
The proof is analogous to the proof of Proposition 4.7.1.

4.7.4. (Parametric approximation by closed forms) Let a ∈ H p (V )
be a fixed cohomology class. Let K ⊂ V be a polyhedron of codimension ≥ 1
and {ωu }u∈Dk a family of p-forms such that {dωu = 0, ωu ∈ a}u∈∂Dk . Then
there exists a family of arbitrarily C 0 -small diffeotopies
{hτu : V → V, τ ∈ [0, 1]}u∈Dk , where {hτu = IdV , τ ∈ [0, 1]}u∈∂Dk
u = h1 (K) by a family
such that {ωu }u∈Dk can be C 0 -approximated near K u
ωu ∈ a}u∈Dk such that {
of closed p-forms { ωu = ωu }u∈∂Dk .
Proof. The convexity of the space of closed p-forms on V which repre-

sent the class a ∈ H p (V ) allows us to construct a family of closed forms
{Ωu ∈ a}u∈Dk which extends the family {ωu }u∈∂Dk . Hence, it remains
to apply Proposition 4.7.3 to {θu = ωu − Ωu }u∈Dk and then take
ωu = θu + Ωu }u∈Dk .
{
Chapter 5
Multivalued Holonomic
Approximation
This chapter is a continuation of Chapter 3. We recall that that the Holo-

nomic Approximation Theorem 3.1.1 holds in a neighborhood of a codimen-
sion 1 stratified subset. It turns out that one can prove a global approxi-
mation result at the expense of making the approximating section multival-
ued. More precisely, we will show that any section F : V → X (r) can be
C 0 -approximated by a holonomic section F : V → X (r) , where V is a non-
Hausdorff covering space of V called a multifold. This result can be seen as
an implementation of the program outlined by René Thom in [Tho59] for
constructing multivalued solutions to differential inequalities, see also the
review of Thom’s works in [Lau19]. We discuss applications of this result
in Chapter 13.
5.1. Multifolds
A. Non-Hausdorff manifolds. So far in this book we assumed that all
manifolds satisfy the Hausdorff condition: any two distinct points can be
separated by disjoint open neighborhoods. As we argue in this chapter,
generalizing some of the results to the non-Hausdorff case allows us to sim-
plify proofs in the Hausdorff case. We note that most of our results about
maps between smooth manifolds (e.g., Smale–Hirsch immersion theory) have
straightforward generalizations to the case when the target manifold is non-
Hausdorff (the situation is much more subtle when the source manifold is
non-Hausdorff).
59
60 5. Multivalued Holonomic Approximation
The non-Hausdorff manifolds V that appear in this book have quite mild
loci NH(V ) of non-Hausdorff points, i.e., points x ∈ V for which there
exists a point x ∈ V , x = x, such that x and x do not have disjoint
neighborhoods. In particular, NH(V ) will always be a Hausdorff stratified
subset of V of positive codimension, and for each x ∈ NH(V ) there will be
only finitely many points x which cannot be separated from x by disjoint
neighborhoods. A prototypical example of such a non-Hausdorff manifold is
given by two open intervals glued along open subintervals.
B. The multifolded non-Hausdorff interval I. Fix an integer N ≥ 1.

Let 0 = a0 < a1 < · · · < aN < 1 and 0 < b1 < · · · < bN < bN +1 = 1
be two sequences of points on the interval [0, 1], such that ai < bi for all
i = 1, . . . , N . Take disjoint copies of the intervals
A0 := [a0 , b1 ), B1 := (a1 , b1 ), A1 := (a1 , b2 ), B2 := (a2 , b2 ), . . . ,
BN := (aN , bN ), AN := (aN , bN +1 ].
Choose any δ ∈ (0, 12 min (bi − ai )) and identify
1≤i≤N
• Ai and Bi+1 along (bi+1 − δ, bi+1 ), i = 0, . . . , N − 1, and

• Bi and Ai along (ai , ai + δ), i = 1, . . . , N .
b1 b 2 bN
a1 a2 aN
0 1
Figure 5.1. The multifolded interval I.
We will call the resulted non-Hausdorff manifold a standard multifolded in-

terval and denote it by I or, in more detail, by I(ai , bi , N, δ). The multifolded
interval I(ai , bi , N, δ) consists of N + 1 direct intervals A0 , . . . , AN and N
reverse intervals B1 , . . . , BN . The non-Hausdorff interval I comes with the
canonical projection π : I → I.
C. The multifolded
non-Hausdorff cube I n . For a small positive num-
ber ρ ∈ 0, 2 denote I1−ρ
1 k := [ρ, 1 − ρ]k . Take the product I n−1 × I and
for every p ∈ I n−1 \ I1−ρ
n−1
stick together the non-Hausdorff interval p × I
into the usual interval p × I, i.e., identify points (p, x), (p, x ) ∈ p × I if
p ∈ I n−1 \ I1−ρ
n−1
and π(x) = π(x ). We will call the resulting non-Hausdorff
5.1. Multifolds 61
manifold a standard multifolded cube and denote it by I n or, in more detail,

by I n (ai , bi , N, δ, ρ). Similarly to a multifolded interval, a multifolded cube
I n consists of N + 1 direct layers
I n−1 × A0 , . . . , I n−1 × AN
and N reverse layers
I n−1 × B1 , . . . , I n−1 × BN ,
glued together over their overlaps above I n−1 \ I1−ρ
n−1
.
The non-Hausdorff cube I n comes with the canonical projection π : I n → I n

fibered over I n−1 ,

π / I n = I n−1 × I
I n OO
OOO p
OOO
p pppp
O pp π
π OOOO
' x pp
p
I n−1 .
Note that for any x ∈ I1−ρ

n−1
the preimage π −1 (x × I) ⊂ I n is the standard
multifolded interval I, while for x ∈ I n−1 \ I1−ρ
n−1
the preimage π −1 (x × I) is
the standard interval I. Moreover, given a splitting I n−1 = I k × I n−k−1 , for
any x ∈ I1−ρ
k the preimage
π −1 (x × I n−k−1 × I) ⊂ I n
is the standard multifolded cube I n−k , while for x ∈ I k \ I1−ρ
k the preimage
−1
π (x × I n−k−1 × I) is the standard cube I n−k .
D. Multifolds and multisections. A multifold over an n-dimensional

manifold V is a non-Hausdorff manifold V together with the projection
π : V → V such that there are disjoint embeddings ϕj : I n → IntV and
ϕj : I n → V indexed by a finite or countable set J such that
• the map π|V \ j (In )
ϕ
is a diffeomorphism;
j
• there exists a multifolded cube I n such that the diagram

j
ϕ
In /V

π
π
ϕj
In /V
commutes for each j ∈ J;

• in the case of a countable J the map ϕ := j ϕj is proper.
Remark. Strictly speaking, the standard multifolded cube I n is not a

unique object but a family which depends on several parameters. It will be
convenient for us to keep this ambiguity, so that in the preceding definition
I n may depend on j.
Given a multifold π : V → V over V and a diffeomorphism h : V → V , we

define the push-forward multifold πh : Vh → V as its pullback under h−1 ,
i.e., by the commuting diagram

V
h / Vh

π h
π

V
h /V
with a diffeomorphism h : V → Vh . Note that NH(Vh ) = h(NH(V ).
A multifold π : V → V over a foliated manifold (V, F ) is called compatible

with the foliation F if the embeddings ϕj : I n → V in the definition of the
multifold V send fibers x × I n−k , x ∈ I k , of the projection I n = I k × I n−k →
I k to the leaves of F through ϕj (x).
A multisection of a fibration p : X → V is a map f : V → X, where

π : V → V is a multifold over V , such that the diagram
?X
f

p
π
V /V
commutes. Alternatively, a multisection can be viewed as an ordinary sec-
tion of the pull-back bundle π ∗ X → V . If the base V is equipped with
a foliation F , then we call a multisection V → X foliated if the multifold
π : V → V is compatible with the foliation F .
Given a multifold π : V → V , any section f : V → X yields a multisection
f ∧ := f ◦ π : V → V → X.
We will say that a multisection f : V → X coincides with a section f :
V → X over a subset A ⊂ V if f = f ∧ over π −1 (A) ⊂ V . Similarly, a
section f : V → X and a multisection f : V → X are called C 0 -close if the
multisections f ∧ and f are C 0 -close.
A multisection F : V → X (r) is called holonomic if there exists a multisec-
tion f : V → X such that F = J r. Given a foliation F on V , a foliated
f
(r)
multisection F : V → XFis called leafwise holonomic if there exists a
multisection f : V → X such that F = J r .
f |F
5.2. Holonomic approximation by multivalued sections 63
5.2. Holonomic approximation by multivalued sections

A. The nonparametric case.
5.2.1. (Multivalued Holonomic Approximation Theorem) Let X →
V be a fibration and A ⊂ V a closed subset such that ∂V ⊂ A. Then any sec-
tion F : V → X (r) , which is holonomic over Op A can be C 0 -approximated
by a holonomic multisection F : V → X (r) for some multifold π : V → V
over V such that F coincides with F over Op A.
Using the Holonomic Approximation Theorem 3.1.1, we can proceed induc-

tively over the skeleton of a triangulation of the manifold V and reduce
Theorem 5.2.1 to its special case for the pair (V, A) = (I n , ∂I n ):
5.2.2. (Multivalued holonomic approximation over a cube) Let
X → I n be a fibration. Then for any section F : I n → X (r) which is
holonomic over Op ∂I n there exist a multifolded cube I n and a C 0 -small
diffeomorphism
h : I n−1 × I → I n−1 × I 1 , h|Op ∂I n = id
fibered over I n−1 such that F can be C 0 -approximated by a holonomic multi-
section F : Ihn → X (r) with F = F over Op ∂I n . Here Ihn is the pushforward
of I n by h.
Proof. We apply the Parametric Holonomic Approximation Theorem 3.1.2

to the family of cubes Op (I n−1 × t), t ∈ I. Thus we get a family of dif-
feotopies
hτt : Op (I n−1 × t) → Op (I n−1 × t), τ ∈ [0, 1], t ∈ [0, 1],
fibered over I n−1 . We can choose this family such that the formula hτ (x, t) =
hτt (x, t) defines a diffeomorphism hτ : I n → I n fibered over I n−1 for all
τ ∈ [0, 1]; see Figure 5.2 for h1 .
We choose the isotopies hτt constant for t close to 0 and 1, as well as near
∂I n−1 ×t, so that the diffeomorphism hτ is the identity on Op (∂I n ). We also
get a family of holonomic sections Φt : Op (h1 (I n−1 ×t)) → X (r) approximat-
ing F |Op (h1 (I n−1 ×t)) , t ∈ [0, 1], and equal to F for t close to 0 and 1, as well as
near ∂I n−1 × t. For a large N we choose, as in Section 5.1.B, two sequences
of points 0 = a0 < a1 < · · · < aN < 1 and 0 < b1 < · · · < bN < bN +1 = 1,
ai < bi , so that
max(aj+1 − aj ), max(bj+1 − bj ) min(bj − aj ).
j j j
Define the intervals Aj , Bj as in Section 5.1.B. We construct the required

holonomic multisection F : Ihn → X, h = h1 , by sticking together direct
Figure 5.2. h1 (I n−1 × t), t ∈ I, n = 2.
layers
FAj := Φcj |h1 (I n−1 ×Aj ) , cj = (aj + bj+1 )/2, j = 0, . . . , N
and appropriately adjusted inverse layers
FBj := Φdj |h1 (I n−1 ×Bj ) , dj = (aj + bj )/2, j = 1, . . . , N
as depicted in Figure 5.3 in the one-dimensional case. Note that the number
of layers (i.e., partial holonomic sections) in the construction depends on
the desired degree of approximation.
Figure 5.3. From partial sections to a multisection.

5.2. Holonomic approximation by multivalued sections 65
B. The foliated case. Let us turn now to the foliated case.

5.2.3. (Foliated multivalued holonomic approximation over a cube)
Consider a fibration X → I n . For k < n − 1 consider the foliation F of
I n by the fibers of the projection s : I n = I k × I n−k → I k onto the first
(r) (r)
factor. Let F : I n → XF be a section of the leafwise r-jet bundle XF
which is leafwise holonomic over Op ∂I n . Then for any ε, δ > 0 there exist
a multifolded cube π : I n → I n , a δ-small (in the C 0 sense) diffeomorphism
h : I n−1 × I → I n−1 × I, h|Op ∂I n = id,
fibered over I n−1 , and a leafwise holonomic multisection
(r)
F : Ihn → XF
such that |F − F | < ε and F = F over Op (∂I n ). Here Ihn is the pushforward
of I n by h.
The proof of Theorem 5.2.3 repeats the proof of Theorem 5.2.2, taking into
account that the diffeotopies hτt preserve the foliation F . As a consequence,
we obtain
5.2.4. (Foliated Multivalued Holonomic Approximation Theorem)
Suppose that that the base V of a fibration p : X → V is endowed with a
(r)
foliation F . Let A ⊂ V be a closed subset and F : V → XF a section of
(r)
the leafwise r-jet bundle XF over V such that F |Op A is leafwise holonomic.
Then there exists a multifold π : V → V compatible with the foliation F and
(r)
a leafwise holonomic multisection F : V → XF such that F is C 0 -close to
F and F = F over Op A.
Proof. In [Th74] Thurston proved (see the jiggling lemma in [Th74]) that
for any foliation F on V there exists a triangulation of V such that:
(a) each l-simplex Δl is transverse to the foliation F ;
(b) for l > s = codim F the foliation F ∩ Int Δl is equivalent to the
restriction of the standard foliation on Rl (generated by the pro-
jection Rl → Rs ) to the open ball Int D l .
Using (a) and the Foliated Holonomic Approximation Theorem 3.8.2, we
can construct a leafwise diffeotopy hτ and a leafwise holonomic section F in
a neighborhood of the image of (n − 1)-skeleton of Thurston’s triangulation.
Note that properties (a) and (b) remain true after the perturbation of the
triangulation. Hence, we can apply Theorem 5.2.3 to the n-simplices of the
perturbed triangulation.
Remark. The existence of a triangulation with properties (a) and (b)
seems obvious at first glance, but the naive idea of constructing a sequence
of barycentic subdivisions of an arbitrary initial triangulation does not work.

The proof requires a more sophisticated partition and jiggling; see [Th74].
One can avoid referring to Thurston’s triangulation by observing that any
subdomain in Rn can be approximated by a domain with piecewise smooth
boundary whose boundary strata are transverse to the foliation of Rn by
the fibers of the projection Rn → Rk . We can then take a covering of V by
cubes such that the restriction of the foliation F to each cube is standard
and argue by induction, replacing each of the cubes by a slightly smaller
subdomain with boundary strata transverse to the foliation.
Remark. Analogous results to our Theorems 5.2.1 and 5.2.4 are proved,
in slightly different language, in the preprint [FPT23] by A. Fokma, Á. del
Pino, and L. Toussaint.
Part 2
Differential Relations
and Gromov’s
h-principle
Chapter 6
6.1. What is a differential relation?

The language of jets is a vehicle for extending the Cartesian geometrization
of algebraic equations to differential equations.
A differential relation of order r imposed on sections f : V → X of a

fibration X → V is a subset R of the jet space X (r) .
Example (Partial differential equations). Any system of ordinary
(n = 1) or partial (n > 1) differential equations
Ψ(x, f, D α f ) = 0
imposed on unknown functions
yj = fj (x1 , . . . , xn ), j = 1, . . . , q,
and their derivatives
∂ |α| fj
D α fj = , α = (α1 , . . . , αn ), |α| = α1 + · · · + αn ≤ r,
∂xα1 1. . . ∂xαnn
may be considered as a differential relation R in the r-jet space J r (Rn , Rq ),
defined by the system of algebraic (= nondifferential) equations
Ψ(x, y, zα ) = 0,
where the variables
x = (x1 , . . . , xn ), y = (y1 , . . . , yq ), and zα = (z1,α , . . . , zq,α )
are coordinates in J r (Rn , Rq ). This way any system of differential equations
can be thought of as a subset of the jet space J r (Rn , Rq ).
69
70 6. Differential Relations
Roughly speaking, differential equations and systems of differential

equations correspond to submanifolds of codimension ≥ 1 in the jet space
J r (Rn , Rq ), while strict differential inequalities correspond to open subsets.
Exercise. Draw the differential relations in J 1 (R, R) which correspond
to the differential equation y = y 2 and the differential inequality y ≥ y 2 .
Example (Immersions and submersions). Let V and W be smooth mani-

folds, n = dim V , and q = dim W .
Let n ≤ q. Recall that a map f : V → W is said to be an immersion if
rank f = n everywhere on V , i.e., the differential df : T V → T W is fiber-
wise injective. The implicit function theorem implies that any immersion is
locally equivalent to the inclusion Rn → Rq .
The immersion relation Rimm ⊂ J 1 (V, W ) over each point x = (v, w) ∈
V × W consists of monomorphisms (= fiberwise injective bundle homo-
morphisms) Tv V → Tw W , or equivalently of nonvertical n-planes Px ⊂
Tv V × Tw W such that
dim (Px ∩ (Tv V × 0)) = 0.
Locally, with respect to a chart
ϕ : V × W ⊃ U1 × U2 → Rn × Rq ,
the relation Rimm over each point x = (v, w) ∈ V × W consists of matrices
A ∈ Mq×n of rank n.
Let n ≥ q. A map f : V → W is called a submersion if rank f = q every-
where on V , i.e., the differential df : T V → T W is fiberwise surjective. The
implicit function theorem implies that any submersion is locally equivalent
to the projection Rn → Rq .
The submersion relation Rsubm ⊂ J 1 (V, W ) over each point x = (v, w) ∈ V ×
W consists of epimorphisms (fiberwise surjective bundle homomorphisms)
Tv V → Tw W , or equivalently of nonvertical n-planes Px ⊂ Tv V × Tw W such
that
dim (Px ∩ (Tv V × 0)) = n − q.
Locally, with respect to a chart
ϕ : V × W ⊃ U1 × U2 → Rn × Rq ,
the relation Rsubm over each point x = (v, w) ∈ V × W consists of matrices
A ∈ Mq×n of rank q.
Note that for n = q

immersion = submersion = local diffeomorphism.
6.2. Open and closed differential relations 71
6.2. Open and closed differential relations

A differential relation R ⊂ X (r) is called open or closed if it is open or closed
as a subset of the jet space X (r) .
As we already mentioned above, closed subsets R which are submanifolds

(or, more generally, stratified subsets) of positive codimension correspond
to systems of differential equations. Such a relation is called determined,
overdetermined, or underdetermined depending on whether codim R = q,
> q, or < q. This classification corresponds to the usual classification of
systems of (differential or algebraic) equations.
Examples. The Laplace equation Δf = 0 defines a closed determined dif-

q
ferential relation in J 2 (Rn , R1 ). The differential equation ẋ2i = 1 defines
1
a closed underdetermined differential relation in J 1 (R, Rq ). The relations
Rimm and Rsubm are open differential relations in J 1 (V, W ).
Exercises.
1. Let X = Λ1 Rn and the differential relation Rclo ⊂ (Λ1 Rn )(1) define

closed 1-forms on Rn . When is this relation underdetermined/determined/
overdetermined?
2. Let the differential relation Riso ⊂ J 1 (Rn , Rq ) define isometric

immersions f : Rn → Rq , i.e., f ∗ h = g where g and h are the standard
metrics on Rn and Rq . When is this relation underdetermined/determined/
overdetermined?
An open differential relation arises, for example, when one tries to find ε-
approximate solutions of a closed differential relation R. In this case our
open relation Rε is the ε-neighborhood of R ⊂ X (r) .
Another rich source of open differential relations is supplied by singularity

theory, where one tries to construct functions, maps or sections for which
certain expressions involving derivatives never vanish, and thus we are led to
a differential relation R which is the complement X (r) \ Σ of a submanifold
(or, more generally, of a stratified subset) Σ ⊂ X (r) . This Σ is usually called
a singularity . Solving R = X (r) \ Σ means finding Σ-nonsingular holo-
nomic sections V → X (r) . For example, the differential relation Rimm (and
similarly Rsubm ) is the complement of a stratified subset Σ1 ⊂ J 1 (V, W ).
6.3. Formal and genuine solutions to a differential relation

A section F : V → R ⊂ X (r) is called a formal solution of the differential
relation R.
Examples.
1. A formal solution to a system of differential equations is a solution of the
underlying system of algebraic equations obtained by substituting deriva-
tives with new independent functions.
2. A formal solution of the immersion relation Rimm is a monomorphism
(= fiberwise injective bundle homomorphism) T V → T W . A formal solution
of the submersion relation Rsubm is an epimorphism (= fiberwise surjective
bundle homomorphism) T V → T W .
A (genuine) solution of a differential relation R ⊂ X (r) is a section f :

V → X such that Jfr (V ) ⊂ R. Alternatively, we can define solutions of
R as holonomic sections F = Jfr : V → R. We will call the holonomic
sections V → R ⊂ X (r) r-extended solutions, or just r-solutions, when the
distinction between the solutions of R as sections of X or X (r) is not clear
from the context.
We will denote the space of solutions of R by Sol R, the space of r-solutions

of R by Hol R, and the space of formal solutions of R by Sec R. The r-jet
extension gives a one-to-one correspondence J r : Sol R → Hol R.
Exercises.
1. Write down the formal and the genuine solutions of the differential equa-
tions y = y and y = y 2 . Draw these formal and genuine solutions in the
jet space J 1 (R, R).
∂2f ∂2f
2. Write down all formal solutions of the Laplace equation ∂x21
+ ∂x22
= 0.
3. Write down all formal solutions of the isometry relation Riso ⊂ J 1 (R2 , R3 ).
6.4. Extension problem

Consider the following boundary value problem. Let R ⊂ X (r) be a differen-
tial relation and f its solution over Op B ⊂ V (i.e., over an open neighbor-
hood of B). Given a bigger subset A ⊃ B, we want to extend the solution
f as a solution of R over Op A. A formal solution of the extension problem
is a section F : Op A → R which coincides with Jfr over Op B. The special
6.5. Approximate solutions to systems of differential equations 73
case of the extension problem when A D k and B = ∂A S k−1 will be

particularly important to us.
In what follows we will use the term solution (or formal solution) of R also
in the sense of a solution (formal solution) of an extension problem for R.
Global solutions of R over V correspond to the case A = V, B = ∅.
Remark. The classical boundary value problem (or Cauchy problem) for
a relation R usually deals with a closed relation R of positive codimension
(PDE) and initial data at the boundary ∂V which can be interpreted as a
formal solution F : ∂V → R. In nonpathological cases the section F can
be trivially extended to a local formal solution F : Op ∂V → R. How-
ever, reduction to the extension problem described above requires a genuine
solution F : Op ∂V → R.
6.5. Approximate solutions to systems of differential

equations
Let X → V be a fibration and R ⊂ X (r) a closed differential relation
of positive codimension. In particular, R may correspond to a system of
PDEs. A section f : V → X is called an ε-approximate solution of R if
f is a solution of the open relation Rε = Uε (R), where Uε (R) is the ε-
neighborhood of R in X (r) . The Holonomic Approximation Theorem 3.1.1
obviously implies the following result:
6.5.1. (Local approximate solution to boundary value problem)
Let F : Op ∂V → R be a formal solution of R. Then for arbitrarily small
δ, ε > 0 there exists a δ-small diffeotopy hτ : V → V , τ ∈ [0, 1], and an
ε-approximate solution f of R over Op h1 (∂V ) such that
Jfr − F |Op h1 (∂V ) C 0 < ε.
In turn, the Multivalued Holonomic Approximation Theorem 5.2.1 implies

the following extension theorem:
6.5.2. (Approximate multivalued solution of extension problem)
Let F : V → R be a formal solution of R such that F |Op ∂V is a genuine
solution of R. Then for arbitrarily small ε > 0 there exists a multifold
π : V → V and a homotopy of multivalued sections
Ft : V → Uε (F (V )) ⊂ Uε (R), t ∈ [0, 1],
such that F0 = F ∧ , Ft |Op ∂V = F |Op ∂V , and F1 is a multivalued solution
(multivalued holonomic section) of Uε (R).
Example. Let R be given by a system of partial differential equations in
Rn . Let F : S → R be the initial data along a hypersurface S = ∂D ⊂ Rn
and we want to find an approximate solution of the Cauchy problem for R

on D. Suppose that there exists an extension F : Op D → R to a formal so-
lution of R over Op D. Then Theorem 6.5.1 implies that for any ε > 0 there
exists an ε-approximate solution of R near a C 0 -small perturbed hypersur-
face S = h(S). In other words, instead of an approximate solution near S
we can find an approximate solution for R near a C 0 -small perturbed (in the
direction normal to S) hypersurface S = ∂ D (see Figure 6.1). Furthermore,
applying Theorem 6.5.2, we can extend F to an approximate multivalued
solution
F :D →R
of the slightly perturbed near the boundary Cauchy problem.
and their neighborhoods Op S and Op S.

Figure 6.1. Hypersurfaces S, S
Chapter 7
Homotopy Principle
7.1. Philosophy of the h-principle

The existence of a formal solution is a necessary condition for the solvability
of a differential relation R, and thus before trying to solve R one should
check whether R admits a formal solution. The problem of finding formal
solutions is of a purely homotopy-theoretical nature. This problem may
be simple or highly nontrivial, but in any case it is important to treat the
homotopical problem first, and look for genuine solutions only after existence
of formal solutions has been established.
Exercises.
1. Let R ⊂ J 1 (R, R) be defined by the nonequality f = 0. Let A = [0, 2π],

B = ∂ A, and f |Op B = sin x. Find a formal solution of the extension
problem and prove that there are no genuine solutions.
2. Let R ⊂ J 1 (R2 , R) be defined by the nonequality grad f = 0. Let

f |Op ∂D2 = x21 − x22 . Prove that the extension problem does not even have
formal solutions.
3. Let R be the same relation as in the previous Exercise. Let A be the

annulus a ≤ r ≤ b where r = x21 + x22 and B = ∂A. Let f |Op B = (r − r0 )2
where r0 = (a + b)/2. Find a formal solution of the extension problem and
prove that there are no genuine solutions.
4. Consider the relation Rimm ⊂ J 1 (S 1 , R2 ). When are two immersions

S 1 → R2 homotopic in Sec Rimm ?
75
76 7. Homotopy Principle
It seems, at first thought, that the existence of a formal solution cannot

be sufficient for the genuine solvability of R. As we already said above,
finding a formal solution is an algebraic, or homotopy-theoretical problem,
which is a dramatic simplification of the original differential problem. Thus
it came as a big surprise when it was discovered in the second half of the
twentieth century that there exist large and geometrically interesting classes
of differential relations for which the solvability of the formal problem is
sufficient for genuine solvability. Moreover, for many of these relations the
spaces of formal and genuine solutions turned out to be much more closely
related than one could expect. This property was formalized in [GE71] and
[Gr71] as the following:
• Homotopy principle (h-principle). We say that a differential
relation R satisfies the h-principle, or that the h-principle holds
for solutions of R, if every formal solution of R is homotopic in
Sec R to a genuine solution of R.
A similar definition can be given for solutions of the extension problem for
R and for families of solutions. For example, we say that
• R satisfies the 1-parametric h-principle if every family of formal
solutions {Ft }t∈I of R, which joins two genuine solutions F0 and
F1 , can be deformed inside Sec R, keeping F0 and F1 fixed, into a
family {Ft }t∈I of genuine solutions of R.
In fact, it is useful to consider different degrees of the h-principle when one
wants to establish closer and closer connections between formal and genuine
solutions. For instance, different forms of the h-principle may include some
approximation and extension properties. We will discuss the different flavors
of the h-principle in the next section.
Exercise (H. Whitney, [Wh37]). Prove the 1-parametric h-principle for
Rimm ⊂ J 1 (S 1 , R2 ).
Hints.
(a) Let f0 , f1 : S 1 → R2 be two immersions and Fτ = (fτ , vτ ) a homotopy
of formal immersions which connects (f0 , f˙0 ) and (f1 , f˙1 ), i.e., vτ (t) = 0 for
all τ, t. Consider some extensions
f¯0 , f¯1 : S 1 × (−ε, ε) → R2
of f0 , f1 and an extension
F̄τ : S 1 × (−ε, ε) → J 1 (S 1 × (−ε, ε), R2 )
of Fτ and apply the Parametric Holonomic Approximation Theorem 3.1.2
(compare Section 4.2).
7.1. Philosophy of the h-principle 77
(b) (Whitney’s proof) We may assume from the very beginning that the
lengths of the curves f0 and f1 are equal to 1 and |vτ (t)| = 1 for all t ∈
[0, 1]/{0, 1} S 1 and τ ∈ [0, 1]. Consider the family of immersions
gτ : [0, 1] → R2 , τ ∈ [0, 1],
given by
t
gτ (t) = fτ (0) + vτ (σ)dσ, t ∈ [0, 1].
0
t
Adjust the family of vector fields vτ = 0 to guarantee that 0 vτ (σ)dσ = 0,
and thus get closed regular curves gτ for all τ .
Remark. S. Smale in [Sm59] generalized Whitney’s theorem to the case
of immersions S n → Rq , n < q. Smale’s covering homotopy method was
different from Whitney’s method. In fact, Whitney’s method is closer, in
some sense, to the idea of convex integration; see Part V of the book.
Exercise. Prove the h-principle for the differential equation y = y and
disprove it for the differential equation y = y 2 (in both cases we consider
global solutions). Disprove the h-principle in both cases for the extension
problem with A = D 1 and B = ∂D 1 .
The examples in the last exercise are trivial and, of course, not typical of
situations where the h-principle is useful. In fact, the h-principle is rather
useless in the classical theory of (ordinary or partial) differential equations
because there, as a rule, it fails or holds for some trivial (or at least well
known) reasons, as in the above examples.
By contrast, for many differential relations rooted in topology and geometry

the notion of h-principle appears to be fundamental, whether it holds or not.
For a given differential relation a priori we often have no obvious reasons
both for the validity or failure of the h-principle. Paradoxically, it appears
that sometimes we need extremely sophisticated tools for disproving the
h-principle. For example, modern symplectic geometry was born in a long
battle for establishing the borderline between the areas where the h-principle
holds and where it fails. Since the beginning of the 1980s the symplectic
rigidity army scored a lot of victories which brought to life the whole new
area of symplectic topology. However, in recent years there were several
unexpected breakthroughs on the flexibility side, bringing us closer to the
ultimate goal of having flexibility and rigidity results in close proximity.
While the majority of symplectic flexibility results discussed in Part 4 of this
book, go back the work of Gromov in the 1970s and 1980s (see, however,
the discussion in Chapter 19), we plan to address the new developments in
our forthcoming book [CEM25].
This book is devoted to two general methods of proving the h-principle:

holonomic approximation (and its variation, called wrinkling) and convex
integration.
7.2. Different flavors of the h-principle

Let R ⊂ X (r) be a differential relation.
A. Parametric h-principle. We say that:

• the (multi )parametric h-principle holds for R if for every relative
spheroid
ϕ0 : (D k , S k−1 ) → (Sec R, Hol R), k = 0, 1, . . . ,
there exists a homotopy
ϕt : (D k , S k−1 ) → (Sec R, Hol R), t ∈ [0, 1],
fixed on S k−1 such that ϕ1 (D k ) ⊂ Hol R. In other words, the
inclusion Hol R → Sec R is a weak homotopy equivalence.
Using the language of homotopy groups, we can say that the parametric
h-principle for a differential relation R means that
πk (Sec R, Hol R) = 0, k = 0, 1, . . . .
In particular, the map Hol R → Sec R induces an isomorphism
π0 (Hol R) → π0 (Sec R).
The epimorphism on π0 means that any formal solution is homotopic (in
Sec R) to a genuine solution, and the monomorphism on π0 means that two
genuine solutions which are homotopic in Sec R are also homotopic in Hol R.
B. Local h-principle. Let A be an arbitrary subset of V . We say that:

• the (local ) h-principle holds for R near A if for every formal solu-
tion F0 : Op A → R there exists a homotopy Ft : Op A → R, t ∈
[0, 1], such that F1 is a genuine solution;
• the parametric (local ) h-principle holds for R near A if for every
relative spheroid
ϕ0 : (D k , S k−1 ) → (Sec Op A R, Hol Op A R), k = 0, 1, . . . ,
ϕt : (D k , S k−1 ) → (Sec Op A R, Hol Op A R), t ∈ [0, 1],
fixed on S k−1 such that ϕ1 (D k ) ⊂ Hol Op A R.
7.2. Different flavors of the h-principle 79
C. Relative h-principle, or h-principle for extensions. For subsets

B ⊂ A ⊂ V we denote by Sec Op (A,B) R the space of formal solutions F :
Op A → R which are holonomic near B. We say that:
• the (relative) h-principle holds for R near the pair (A, B) if for
every formal solution F0 ∈ Sec Op (A,B) R there exists a homotopy
through formal solutions
Ft ∈ Sec Op (A,B) R, t ∈ [0, 1],
such that Ft |Op B = F |Op B for all t ∈ [0, 1] and F1 is a genuine
solution;
• the parametric (relative) h-principle holds for R near the pair
(A, B) if for every relative spheroid
ϕ0 : (D k , S k−1 ) → (Sec Op (A,B) R, Hol Op A R), k = 0, 1, . . . ,
ϕt : (D k , S k−1 ) → (Sec Op (A,B) R, Hol Op A R), t ∈ [0, 1],
fixed on S k−1 such that ϕ1 (D k ) ⊂ Hol Op A R and for every p ∈ D k
the homotopy ϕt (p) : Op A → R is fixed near B.
D. C0 -dense h-principle. We say that:

• the C 0 -dense h-principle holds for R if the (usual) h-principle holds
for R and if for every formal solution F0 : V → R and an arbitrarily
small neighborhood U ⊂ X of the underlying section f0 = bs F0 the
homotopy Ft , t ∈ [0, 1], in R which brings F0 to a genuine solution
F1 can be chosen in such a way that bs Ft (V ) ⊂ U, t ∈ [0, 1].
In a similar way one defines the C 0 -dense versions of all previously defined
h-principles.
E. Fibered h-principle. The parametric h-principle can also be reformu-

lated in a fibered version. Moreover, sometimes when we are working with
the parametric situation, the differential relation itself may also depend on
the parameter and in this case the fibered version is more convenient than
the parametric one. Let us formulate the corresponding definition.
Let P be the space of parameters. Any subset R ⊂ P × X (r) is called a

fibered differential relation imposed on the fibered (over P ) sections
f : P × V → P × X; (p, v) → (p, fp (v))
which depend continuously on p ∈ P . Formal solutions of R are sections
P ×V → R ⊂ P ×X (r) and (extended) solutions of R are fiberwise holonomic
sections P × V → R. We say that

• the fibered h-principle holds for a fibered relation R ⊂ P × X (r) if
every fibered formal solution F0 : P × V → P × X (r) is homotopic
via a homotopy of fibered formal solutions Ft to a fibered genuine
solution F1 : P × V → R.
If P is a manifold with nonempty boundary ∂P , then we usually assume that
the formal solution F is holonomic over ∂P ×V and require the homotopy Ft
to be fixed near ∂P ×V . The parametric h-principle can now be reformulated
as follows. The parametric h-principle holds for R if and only if for every
k = 0, 1, . . . the fibered h-principle holds for Rk = D k × R ⊂ D k × X (r) .
F. Foliated (leafwise) h-principle. Let F be a foliation on V and R ⊂

(r)
XF a subset in the leafwise jet space. We say that
(r)
• the foliated h-principle holds for R ⊂ XF relative to a closed
(r)
subset A ⊂ V if any leafwise formal solution F0 : V → R ⊂ XF
which is leafwise holonomic on Op A is homotopic via a homotopy
(r)
of leafwise formal solutions Ft : V → XF fixed on Op A to a
leafwise holonomic section (genuine leafwise solution) F1 : V → R.
We will discuss the relation between the parametric, fibered, and foliated
h-principles further in Section 10.3.
Chapter 8
Open Diff V -invariant

8.1. Natural fibrations

The class of natural fibrations (or natural fiber bundles) was introduced
(without the term) in Gromov’s paper [Gr69]. The term was introduced by
Nijenhuis in [Ni72] where he gave a categorial definition, see Section 8.1.C.
Since that time natural fibrations were intensively studied in the work of
several authors; see the monograph [KMS93] for details.
A. Definition and examples. Given a fibration p : X → V , we denote

by Diff V X the group of fiber-preserving diffeomorphisms hX : X → X, i.e.,
hX ∈ Diff V X if and only if there exists a diffeomorphism hV : V → V such
that the diagram
hX
X / X
p p
hV
V / V
commutes. We are interested in the situation when the action of Diff V on

V can be canonically lifted to an action on X via fiber-preserving diffeo-
morphisms, i.e., the fibration Diff V X → Diff V , hX → hV has a canon-
ical section. Moreover, we would like this lifting property to hold also
for the pseudo-group of local diffeomorphisms. Namely, we call a fibra-
tion p : X → V natural if for any diffeomorphism h : U → U between
two open subsets U, U ⊂ V there is given a fiber-preserving diffeomorphism
81
82 8. Open Diff V -invariant Differential Relations
h∗ : p−1 (U ) → p−1 (U ) such that for any two diffeomorphisms

h
U −→ U −→ U
h
we have (h ◦ h)∗ = h∗ ◦ h∗ , and for any point v ∈ V the germ of h∗ along
p−1 (v) depends only on the germ of h at v. The latter property allows us to
extend the lifting property to local diffeomorphisms (i.e., equidimensional
immersions) U → U ; we will refer to their lifts as fiberwise isomorphisms.
Examples.
1. The trivial bundle X = V × W → V is natural. Here h∗ = h × idW .
2. The tangent bundle T V → V is natural. The corresponding lift h∗ here
is provided by the differential h∗ = dh : T U → T U of a diffeomorphism
h : U → U for U, U ⊂ V .
3. The cotangent bundle T ∗ V → V is natural. The corresponding lift h∗
here is provided by the inverse h∗ = (dh∗ )−1 : T ∗ U → T ∗ U of the codif-
ferential dh∗ : T ∗ U → T ∗ U of a diffeomorphism h : U → U . A similar
construction defines a natural bundle structure on X = Λp T ∗ V , the p-th
exterior power of the cotangent bundle T ∗ V .
4. If X → V is natural, then X (r) → V is also a natural bundle. The
(r)
corresponding lift h∗ : (X|U )(r) → (X|U )(r) is given by
(r)
h∗ (s) = Jhr∗ ◦s̄◦h−1 (h(v)),
where s ∈ X (r) , v = pr (s) ∈ V , and s̄ : Op v → X is a local section near v
which represents the r-jet s, i.e., Js̄r (v) = s.
B. Universal natural principal bundle. It turns out that for any given
n there is a universal principal bundle P D (V ) → V over any n-dimensional
manifold V , such that any natural bundle over V is associated to it. Denote
by D0 (n) the group of germs of diffeomorphisms δ : (Rn , 0) → (Rn , 0) at 0.
8.1.1. (Universal natural principal bundle) For any dimension n and
any n-dimensional manifold V there is a canonical natural principal D0 (n)-
bundle P D (V ) such that all natural bundles over V are associated with
P D (V ) for some representation of D0 (n) in the group of automorphisms
of their standard fibers. Moreover, any local diffeomorphism h : V → V
lifts to a fiberwise isomorphism
P D (h) := h∗ : P D (V ) → P D (V ),
such that P D (h ◦ h) = P D (h ) ◦ P D (h) for any two local diffeomorphisms
h
V −→ V −→ V .
h
8.1. Natural fibrations 83
In other words, P D is a functor from the category of n-dimensional mani-

folds and local diffeomorphisms to the category of natural D0 (n)-principal
bundles over n-dimensional bases and fiberwise isomorphisms.
Proof. For an n-dimensional manifold V consider the space P D (V ) of pairs

(v, ϕ), where v ∈ V and ϕ is the germ of a diffeomorphism
(Op Rn 0, 0) → (Op V v, v).
The group D0 (n) acts freely on P D (V ) by the formula δ ·(v, ϕ) = (v, ϕ◦δ −1 ),
and the orbits of this action are the fibers of the projection (v, ϕ) → v.
Hence, P D (V ) → V is a principal D0 (n)-bundle. Any local diffeomorphism
h : V → V lifts to a fiberwise isomorphism P D (h) : P D (V ) → P D (V ) by
the formula
P D (h)(v, ϕ) = (h(v), h ◦ ϕ).
Clearly, P D (V ) → V is a natural principal fiber bundle with the infinite-
dimensional standard fiber D0 (n), and for local diffeomorphisms h : V → V
and h : V → V the functorial property P D (h ◦ h) = P D (h ) ◦ P D (h) is
satisfied.
The previous definition lacks the description of the topology in P D (V ); the
following alternative definition of P D (V ) fills this gap. For a domain U ⊂ Rn
we can identify P D (U ) with U × D0 (n) via parallel transport. We endow
D0 (n) with the C ∞ -topology and use this identification to define the product
topology on P D (U ). Choose a coordinate atlas {(Ui , ϕi )} for V , where
ϕi : Ui → Gi := ϕi (Ui ) ⊂ Rn ,
and let hij := ϕj ◦ ϕ−1
i : Gij → Gji , where Gij := ϕi (Ui ∩ Uj ), be the corre-
sponding transition maps. We can then construct P D (V ) by gluing P D (Gj )
and P D (Gj ) with gluing maps Hij = P D (hij ) : P D (Gij ) → P D (Gji ).
Let us now show that any natural fiber bundle p : X → V is associated with
the principal bundle P D (V ) → V . Fix a point v0 ∈ V and denote by W the
fiber p−1 (v0 ) ⊂ X. By fixing any diffeomorphism
f0 : (Op Rn 0, 0) → (Op V v0 , v0 )
we can define an action δ∗ of D0 (n) on W by the formula
δ∗ w := (f0 ◦ δ ◦ f0−1 )∗ w, δ ∈ D0 (n), w ∈ W.
Define a map Φ : P D (V ) × W → X as follows. Given (v, ϕ) ∈ P D (V ) and
w ∈ W , consider the germ of the diffeomorphism
ϕ := ϕ ◦ f0−1 : (Op V v0 , v0 ) → (Op Rn 0, 0) → (Op V v, v)
and define
Φ((v, ϕ), w) = (ϕ)∗ w ∈ p−1 (v) ⊂ X.
Note that Φ((v, ϕ), w) = Φ((v , ϕ ), w ) if and only if v = v and ϕ = ϕ ◦δ for

some δ ∈ D0 (n) such that δ∗ w = w . Hence, Φ descends to an isomorphism
from the associated bundle P D (V ) ×D0 (n) W to X.
C. Equivalent definitions of natural fibrations. Proposition 8.1.1 leads

to two other equivalent definitions of a natural fibration. First, we can define
a natural fibration over an n-dimensional base V with standard fiber W as
any fibration associated with the D0 (n)-principal fibration P D (V ) → V with
respect to an action of D0 (n) on W .
Yet equivalently, we can define natural fibrations in a functorial way, as was

done by Nijenhuis in [Ni72]. Let Mn be the category of n-dimensional
manifolds and local diffeomorphisms, and F Mn the category of smooth
fibrations over n-dimensional manifolds and fiberwise isomorphisms. Then a
natural fibration is a covariant functor F : Mn → F Mn which associates to
each n-dimensional manifold V a fiber bundle pV : F (V ) → V , and to a local
diffeomorphism h : V1 → V2 a fiberwise diffeomorphism F (h) : F (V1 ) →
F (V2 ) covering h, such that the following axiom is satisfied. If i : U → V is
an inclusion of an open subset, then pU : F (U ) → U is the restrction of the
fibration pV : F (V ) → V to U , i.e., pU = pV |p−1 (U ) : p−1
V (U ) → U, and F (i)
V
is the inclusion of total spaces. In particular, the identity map V → V lifts
to the identity map F (V ) → F (V ).
D. Finiteness theorem. Denote by Lr (n) the group of r-jets at 0 of dif-

feomorphisms in D0 (n). In particular, L1 (n) = GL(n) and L0 (n) = {1} (the
trivial group).
Similarly to the construction of the D0 (n)-principal bundle, we can construct
for any r ≥ 0 a principal Lr (n)-bundle P r (V ) → V . The total space P r (V )
consists of pairs (v, ϕ), where v ∈ V and ϕ is the r-jet of a diffeomorphism
(Op Rn 0, 0) → (Op V v, v)
at 0 ∈ Rn . In particular, P 1 (V ) is the principal GL(n)-bundle of tangent

n-frames associated to the tangent bundle T V . For any r ≥ 0 the principal
bundles P r (V ) → V are associated with the bundle P D (V ) with respect to
the canonical projection D0 (n) → Lr (n), and thus natural.
According to the the finiteness theorem (Palais and Terng, [PT77]) the
converse is also true: the structure group of a natural fiber bundle X → V
always coincides with Lr (n) for some r = r(X), i.e., for any v ∈ V the
actions of two germs
f1 , f2 : Op v → Op v , v = f1 (v) = f2 (v),
8.3. Local and global h-principle for open Diff V -invariant relations 85
coincide if Jfr1 (v) = Jfr2 (v). Therefore (see [PT77], [KMS93]), all natural
bundles over V with a fiber W are in a bijective correspondence with smooth
left actions of the jet groups Lr (n), r = 0, 1, . . . on W .

q
In fact, r max n−1 , nq + 1 for n > 1, and r ≤ 2q + 1 for n = 1, where q
is the dimension of the fiber; see [Za87], [ET79].
8.2. Diff V -invariant differential relations

Recall from Example 4 in Section 8.1 that for a natural fibration X → V
the r-jet fibration X (r) → V is also natural. A differential relation R ⊂ X (r)
is called Diff V -invariant if the action
(r)
s → h∗ (s)
for any local diffeomorphism h : U → U for U, U ⊂ V leaves R invariant.
In other words, a Diff V -invariant differential relation R ⊂ X (r) is a natu-
ral subbundle of the natural fiber bundle X (r) . Equivalently, a differential
relation R is Diff V -invariant if it can be defined in a V -coordinate free form.
(r)
The action s → h∗ (s) preserves the set of holonomic sections
(r)
h∗ (Jfr (v)) = Jhr∗ ◦f ◦h−1 (h(v))
for f ∈ SecX, h ∈ Diff V . In particular, the group Diff V acts on the space
Sol R of solutions of a Diff V -invariant differential relation R.
Recall that any natural bundle X → V over an n-dimensional manifold V

is determined by an action α of the group D0 (n) on the standard fiber W
of the fibration. This action lifts to an action αr of D0 (n) on the standard
fiber W r of the r-jet bundle X (r) , which is the space of r-germs at 0 of maps
Rn → W . A Diff V -invariant relation R ⊂ X (r) is determined by a subset
R ⊂ W r which is invariant with respect to the action αr .
Examples. The relations Rimm and Rsubm are Diff V -invariant. For any
A ⊂ Grn W the relation RA which defines A-directed maps V → W (see
4.5) is Diff V -invariant.
8.3. Local and global h-principle for open Diff V -invariant

relations
8.3.1. (Local h-principle for open Diff V -invariant relations) Let
X → V be a natural fiber bundle. Then any open Diff V -invariant differ-
ential relation R ⊂ X (r) satisfies all forms of the local h-principle near any
polyhedron A ⊂ V of positive codimension.
Proof. We first consider the nonparametric case, i.e., we will prove that
π0 (Sec Op A R, Hol Op A R) = 0.
We need to show that, given a section F ∈ Sec Op A R, there exists a section
G ∈ Hol Op A R which is homotopic in Sec Op A R to F . According to Theorem
3.1.1, there exists an arbitrarily C 0 -small diffeotopy hτ : V → V, τ ∈ [0, 1],
and a section F ∈ Hol Op h(A) R such that F is C 0 -close to F over Op h(A). In
particular, we may assume that the linear homotopy Ft between F |Op h1 (A) =
F 0 and F = F1 lies in R. The desired section G ∈ Hol Op A R can then be
defined by the formula G = H 1 (F ) where
H τ = ((hτ )r∗ )−1 = ((hτ )−1 )r∗ : X (r) → X (r)
is the induced action of the straightening diffeotopy (hτ )−1 on the natural
fibration X (r) → V . The required homotopy in R, which connects F and
G over a neighborhood of A, consists of two stages: first H τ (F ), τ ∈ [0, 1],
and then H 1 (F t ), t ∈ [0, 1].
Note that although the holonomic approximation gives us a section F over

Op h(A) which is C 0 -close in X (r) to the initial formal solution F over
Op h(A), it does not imply the same property for G over Op A, as the C 0 -
closeness to F fails for the straightened solution H 1 (F) on Op A. However,
the C 0 -approximation over Op A in X survives after straightening, and this
implies the local C 0 -dense h-principle for R near A.
For the general parametric, relative, and relative parametric cases we need
to use the corresponding parametric/relative versions of the Holonomic Ap-
proximation Theorem 3.1.1. However, the proof for these cases differs only
in notation. Let us consider, for example, the parametric case.
In order to prove the equality

πm (Sec Op A R, Hol Op A R) = 0, m ≥ 1,
we need to show that, given a family of sections
Fz ∈ Sec Op A R, z ∈ I m ,
such that for z ∈ Op ∂I m the section Fz is holonomic, there exists a family
Gz ∈ Hol Op A R which is homotopic in Sec Op A R to the family Fz , z ∈
I m , relative to ∂I m . According to Theorem 3.1.2, there exists a family of
arbitrarily C 0 -small diffeotopies hτz : V → V , τ ∈ [0, 1], z ∈ I m , and a
family of sections Fz ∈ Hol Op h(A) R such that hz = Id for z ∈ ∂I m and Fz is
C 0 -close to F over Op h1z (A). In particular, we may assume that the linear
homotopy Fzt between Fz |Op h1z (A) and Fz lies in R. The desired family of
8.3. Local and global h-principle for open Diff V -invariant relations 87
sections Gz ∈ Hol Op A R can be then defined by the formula G = Hz1 (F)

where
Hzτ = ((hτz )r∗ )−1 = ((hτz )−1 )r∗ : X (r) → X (r)
is the induced action of the straightening diffeotopy (hτz )−1 on the natural
fibration X (r) → V . The required homotopy in R which connects Fz and
Gz over a neighborhood of A consists of two stages: first Hzτ (Fz ), τ ∈ [0, 1],
and then Hz1 (Fzt ), t ∈ [0, 1].
8.3.2. (Local h-principle implies global for open manifolds) Let V
be an open manifold and X → V a natural fibration. Let R ⊂ X (r) be a
Diff V -invariant differential relation. Then the parametric local h-principle
for R implies the parametric global h-principle for R.
Proof. We consider only the nonparametric case, i.e., prove that

π0 (Sec R, Hol R) = 0.
The general parametric case differs only in notation. We need to show
that, given a section F ∈ Sec R, there exists a section G ∈ Hol R which is
homotopic in Sec R to F . Let K ⊂ V be a core of the manifold V . The
local h-principle near K implies the existence of a section GK ∈ Hol Op K R
which is homotopic in Sec Op K R to F |Op K . Let hτ : V → V, τ ∈ [0, 1],
be an isotopy compressing V into a neighborhood U of K such that GK is
defined over U . The desired section G ∈ Hol R can now be defined by the
formula G = H 1 (GK ), where
H τ = ((hτ )r∗ )−1 = ((hτ )−1 )r∗ : X (r) → X (r)
is the induced action of the decompressing diffeotopy (hτ )−1 on the natural
fibration X (r) → V . The required homotopy in R which connects the global
sections F and G consists of two stages: first H τ (F ), τ ∈ [0, 1], and then
H 1 (GtK ), t ∈ [0, 1], where GtK is the homotopy which connects the local
sections FOp K and GK in R.
Remark. Note that the local C 0 -dense h-principle near K does not sur-
vive after an expansion, so we cannot derive the global C 0 -dense h-principle
for R from the local C 0 -dense h-principle.
Theorems 8.3.1 and 8.3.2 imply

8.3.3. (Gromov h-principle for open Diff V -invariant relations over
open manifolds [Gr69]) Let V be an open manifold and X → V a natural
fiber bundle. Then any open DiffV -invariant differential relation R ⊂ X (r)
satisfies the parametric h-principle. In particular, immersions, submersions,
k-mersions (maps of rank ≥ k), and immersions V → W directed by an open
set A ⊂ Grn (W ) (see 4.5) satisfy the parametric h-principle as long as the
underlying manifold V is open.
Besides the applications listed in Theorem 8.3.3, the following special case
of 8.3.3 will be important for us.
8.3.4. (Maps tranverse to a distribution) Let τ ⊂ T W be a tangent
distribution and R ⊂ J 1 (V, W ) the differential relation consisting of homo-
morphisms T V → T W transverse to τ . The relation R is DiffV -invariant,
and thus the h-principle holds for R (i.e., for maps V → W transverse to
τ ) if V is open.
The relative version of 8.3.3 holds in the following form:

8.3.5. (Relative version of 8.3.3) Let R ⊂ X (r) be an open Diff V -
differential relation over an open manifold V . Let B ⊂ V be a closed subset
such that each connected component of the complement V \ B has an exit to
∞. Then the relative parametric h-principle holds for R and the pair (V, B).
Chapter 9
Applications to Closed
Manifolds
9.1. Microextension trick

The microextension trick, which goes back to M. Hirsch, sometimes allows us
to reformulate problems about closed manifolds in terms of open manifolds.
For example, if dim W > dim V , then the construction of an immersion
V → W homotopic to a map f : V → W is equivalent to the construction of
an immersion E → W , where E is the total space of the normal bundle to
T V in f ∗ T W . The manifold E is open, hence the h-principle 8.3.3 applies.
Above we already used the microextension trick for directed embeddings of
closed manifolds; see Theorem 4.6.1.
9.2. Smale–Hirsch h-principle

The h-principle for immersions V → W obviously fails if V is closed and
n = q: a closed n-dimensional manifold never admits an immersion into Rn
even if it is parallelizable, which is equivalent to the existence of a formal
solution of the immersion problem.
Exercise. Prove that an immersion of a neighborhood of ∂D 2 into R2 ,

which is shown in Figure 9.1, cannot be extended to an immersion of the
disk D 2 into R2 , while it extends to a formal immersion.
However, in the case n < q one gets the C 0 -dense parametric h-principle via
the microextension trick.
89
90 9. Applications to Closed Manifolds
9.2.1. (Smale–Hirsch h-principle for immersions [Hi59]) The relative

parametric C 0 -dense h-principle holds for immersions of an n-dimensional
manifold V into a manifold W of dimension q > n.
Figure 9.1. An immersion S 1 × (−ε, ε) → R2 which cannot be

extended to an immersion D2 → R2 although it has a formal ex-
tension to D2 .
Proof. In the nonparametric case let F : T V → T W be a formal solution of

the differential relation Rimm ⊂ J 1 (V, W ). Set f = bs F . Suppose that F is
holonomic over Op A for a closed subset A ⊂ V , i.e., F = df on Op A. The
homomorphism F identifies T V with a subbundle λ of rank n of the induced
bundle f ∗ T W . Let ν be the normal bundle to λ in T W with respect to a
choice of a Riemannian metric on W , N the total space of the bundle ν, and
π : N → V the projection. Then T N ∼ = π ∗ T V ⊕ π ∗ ν, hence F canonically
lifts to a fiberwise isomorphism F : T N → T W . Moreover, we can make F
holonomic on Op N A. The restriction of an immersion N → W to the zero
section V ⊂ N is an immersion V → W , and therefore, applying the relative
C 0 -dense local h-principle 8.3.1 near V , we can deform F to an immersion
V → W , keeping it fixed on Op A.
Given a family of formal solutions Fz : T V → T W parametrized by the
points z of a disc D k , we denote by N the total space of the bundle ν0
normal to the homomorphism F0 . Then the isomorphism F0 : T N → T W
which extends F0 can be canonically prolongated along radii of the disc D k
to a family of isomorphisms Fz : T N → T W for all z ∈ D k , hence the above
argument applies parametrically.
Remarks.
1. If the bundle (f ∗ T W )/T V contains a trivial one-dimensional subbundle
θ, then it is sufficient to extend V to V × R. In fact, we can proceed
inductively over skeleta of a triangulation of V and in this case there is no
problem with the existence of θ locally over a neighborhood of a simplex.
2. Note that the microextension trick does not work for submersions because
the restriction of a submersion is not, in general, a submersion. In fact, the
h-principle is false for submersions of closed manifolds.

9.3. Sections transverse to a distribution 91
Similarly to Theorem 9.2.1 one can prove the following generalization of

Theorem 9.2.1 (cf. Theorem 4.6.2).
9.2.2. (Immersions transverse to a distribution) Let ξ be a subbundle
of T W . If dim V < codim ξ, then all forms of the h-principle hold for
immersions V → W transverse to ξ.
9.3. Sections transverse to a distribution

We will need the following modification of the local h-principle 8.3.1.
9.3.1. (Special local h-principle) Let X → V × R be a natural fibration
and R ⊂ X (r) an open differential relation which is invariant with respect
to diffeomorphisms of the form
(x, t) → (x, h(x, t)), x ∈ V, t ∈ R.
Then R satisfies all forms of the local h-principle near V × 0 and the global
parametric h-principle over V × R.
The proof follows from the Holonomic Approximation Theorem 3.1.1 accord-
ing to the same scheme as the proof of Theorem 8.3.1, with the additional
observation that the perturbation h implied by Theorem 3.1.1 has the special
form required in Proposition 9.3.1.
Note that Proposition 9.3.1 is a version of the Main Flexibility Theorem

from [Gr86, p. 78].
Given two manifolds X, Y and tangent distributions τ ⊂ T X and η ⊂ T Y ,

we say that a homomorphism F : T X → T Y sends τ transversely to η if the
composition
F |τ
τ −→ T Y −→ T Y /η
is injective when dim τ + dim η ≤ dim Y , and surjective when dim τ +
dim η ≥ dim Y . We say that a map f : X → Y sends τ transversely to η if
the differential df : T X → T Y has the above property.
9.3.2. (Gromov [Gr86]) Let X → V be a fibration and τ a subbundle of
the tangent bundle T X. If
dim τ + dim V < dim X,
then sections V → X transverse to τ satisfy all forms of the h-principle.
Remarks.
1. Here the respective differential relation R is not Diff V -invariant.
92 9. Applications to Closed Manifolds
2. The condition dim τ + dim V < dim X is crucial and in general cannot
be weakened, even for open V (because it would imply the corresponding
statement for any closed submanifold of V ). See, however, Theorem 22.2.1.
3. For a trivial fibration V × W → V and τ equal to the vertical tangent
bundle of the fibration V × W → W , Theorem 9.3.2 is just Hirsch’s theorem
about immersions V → W , dim V < dim W (see Theorem 9.2.1 above).
Proof of Theorem 9.3.2. Using a sufficiently small triangulation of the

manifold V , we can reduce the problem to its following relative version:
V = D n , X = D n × Rq , and the section V → X is already transverse to τ
near ∂V = ∂D n .
The microextension trick that we are going to apply below differs from the
one we have used in the proof of Theorem 9.2.1: now we will extend both
the source and the target manifolds.
Let R ⊂ X (1) be the differential relation of transversality to τ and F : V →

R a formal solution which is already holonomic near ∂V . Set f = bs F .
Let ξ be the subbundle of T X|f (V ) defined by F . Consider the fibration
X × R → V × R and the subbundle τ × R ⊂ T (X × R). Let R ⊂ (X × R)(1)
be the differential relation which defines the sections transverse to τ × R.
The subbundle
ν = (T X|f (V ) )/(τ |f (V ) ⊕ ξ)
is trivial (this is the only place where we need the localization) and dim ν ≥
1. Therefore, we can extend F to a formal solution F : V × R → R which
is holonomic near ∂V × 0 ⊂ V × R.
Let A be the subgroup of Diff (V × R) consisting of diffeomorphisms fibered

over V . The relation R is open and A-invariant. Therefore, according to
Proposition 9.3.1, the local h-principle holds for R near V = V × 0 ⊂ V × R.
The condition
dim τ + dim V < dim X
implies
dim (τ × R) + dim (V × R) ≤ dim (X × R),
hence a section V × R → X × R which is transverse to τ × R defines a
section V → X which is transverse to τ . Therefore, the local h-principle for
R implies the h-principle for R.
9.4. Multivalued h-principle 93
Similarly we can prove

9.3.3. (h-principle for τ -ξ-immersions) Let τ be a subbundle of T V and
ξ a subbundle of T W . If dim τ < codim ξ, then all forms of the h-principle
hold for maps V → W which send τ transversely to ξ.
In particular, we have
9.3.4. (h-principle for τ -immersions) Let τ be a subbundle of T V . If
dim τ < dim W , then the h-principle holds for τ -immersions f : (V, τ ) →
W , i.e., maps f : V → W with rank(df |τ ) = dim τ .
9.4. Multivalued h-principle

We finish this chapter with the following multivalued h-principle which yields
multivalued approximate solutions of an open differential relation on any
manifold, open or closed, in the absolute and relative cases. While it is
just a reformulation of the Multivalued Holonomic Approximation Theorem
5.2.1, it is in some sense a realization of the program of René Thom in
[Tho59]. In the words of F. Laudenbach in [Lau19], in that paper Thom
“anticipated” the h-principle.
9.4.1. (Multivalued h-principle for open differential relations) Let
X → V be a fibration and R ⊂ X (r) an open differential relation. Then
for any formal solution F : V → R there exists a multifold V → V and a
homotopy of multivalued sections
Ft : V → R, t ∈ [0, 1],
such that F0 = F ∧ and F1 is a multivalued solution (multivalued holonomic
section) of R.
The parametric, relative, and foliated versions of the multivalued h-principle

9.4.1 are also true.
Chapter 10
Foliations
The theory of foliations is of interest to us for two reasons, one meaningful

and one technical. The meaningful aspect is related to the h-principle for the
existence and classification of foliations on manifolds. For open manifolds,
this is the Gromov–Phillips theorem [Gr69, Ph70]. The culmination of this
h-principle are Thurston’s famous theorems for foliations on closed manifolds
[Th74,Th76b]. The technical aspect is related to the fact that on manifolds
with a given foliation, one can define a foliated h-principle as a generalization
of the parametric and fibered h-principle. There are a large number of
surveys and books devoted to the theory of foliations; see for example [La74,
CC00]. We give here only the most basic definitions, allowing us to use the
language of foliations and formulate the h-principle for foliations and the
foliated h-principle.
10.1. Definition and examples

A. C ∞ -smooth foliations. Let 0 < p < n. A p-dimensional C ∞ -smooth
foliation F on an n-dimensional manifold V is a partition V = La into
connected subsets La , called leaves, such that around every
point of M there
is a chart V ⊃ U →Rn such that the partition U ∩ ( La ) coincides with
the partition Rn = s∈Rn−p Rp × s. The number q = n − p is called the
codimension of F . The tangent bundle τ (F ) to the foliation F is the p-
dimensional distribution formed by the tangent spaces to the leaves of F .
We will identify the normal bundle ν(F ) = T V /τ (F ) with the orthogonal
complement to τ (F ) ⊂ T V , assuming that V is endowed with a Riemann-
ian metric. For a manifold with boundary, we assume that the foliation is
95
96 10. Foliations
orthogonal to the boundary. The exceptions include foliations of codimen-

sion 1, where in some cases it is convenient to assume that (all or some)
boundary components are leaves.
Any one-dimensional distribution is integrable, that is, it is the tangent bun-

dle of a one-dimensional foliation. In contrast, distributions of dimension
greater than 1 are in general not integrable. For p ≥ 2 the tangent distribu-
tion τ (F ) of any foliation F is involutive, i.e., in a neighborhood of any point
the exterior algebra generated by differential 1-forms annihilating τ (F ) is
closed under the operation of exterior differentiation. Frobenius’s theorem
(see, e.g., [Wa83]) asserts that, conversely, any involutive p-dimensional
distribution integrates to a foliation.
Examples.
1. Any submersion f : V → W between manifolds of dimension n and q

induces an (n − q)-dimensional foliation of V by the preimages f −1 (w) of
points w ∈ W . In particular, any fibration p : X → V , where dim X = n + q
and dim V = n, induces a q-dimensional foliation of X by the fibers of the
fibration p. By definition, any foliation is locally induced by a submersion.
2. Any nowhere vanishing vector field X on a manifold V defines a one-
dimensional foliation by its trajectories. In general, this one-dimensional
foliation is not induced (globally) by any submersion V → W .
3. The theory of foliations of dimension greater than 1 essentially started
with the following example of Reeb [Re52]. Let β : [0, 1] → [0, 1] be a
decreasing C ∞ -function such that
(a) β(0) = 1, β (0) = 0, and
(b) β(1) = β (1) = β (1) = · · · = 0.
Consider the foliation Rβ on Int(D n × R) ⊂ Rn × R with leaves defined by
the equation
1
z= + C, C ∈ R,
β(x2 )
where x ∈ D n and z ∈ R. We extend the foliation Rβ to the manifold D n ×R

by adding the boundary leaf ∂D n × R. As the foliation Rβ is invariant with
respect to translations along the second factor, we can define the quotient
foliation Rβ on D n × S 1 , where S 1 = R/Z. The foliation Rβ is called the
Reeb component. Given any diffeomorphism f : ∂D n × S 1 → ∂D n × S 1 , we
get a foliation Rβ,f on (D n × S 1 ) ∪f (D n × S 1 ). In particular, we get Reeb
foliations on S 2 × S 1 , S 3 and all Lens spaces L(p, q).

10.1. Definition and examples 97
B. C r,s -smooth foliations. Although the class of smoothness is important

in some aspects of foliation theory, in what follows (as in the above defini-
tion) we usually assume all foliations to be C ∞ -smooth. However, some
discussion of foliations of finite smoothness is necessary, at least in connec-
tion with the Reeb foliation. The careful consideration of Reeb foliation
Rβ,f shows that for foliations the notion of smoothness requires some cau-
tion, namely, one should distinguish between tangent and normal smooth-
ness. In fact, the question of two smoothnesses appears already for one-
dimensional foliations on the plane. Let us define a C s,r -smooth foliation
for 0 ≤ r ≤ s ≤ ∞.
foliation F on a (C ∞ -smooth) n-dimensional

A p-dimensional C s,r -smooth
manifold V is a partition V = La into connected C s -smooth leaves La such
that around every point of M there is a C r chart V ⊃ U → Rn such that
the partition U ∩ ( La ) coincides with the partition Rn = s∈Rn−p Rp × s.
In the literature on foliation theory, as a rule, foliations of class C r are
understood as C ∞,r -smooth foliations, i.e., transversely C r -smooth foliations
with C ∞ -smooth leaves. In particular, C ∞ -smooth foliations are usually
understood as C ∞,∞ -smooth foliations. In some sources the description of
the Reeb foliation on S 3 gives only a C ∞,r -smooth foliation with r = 0 or
r = 1.
Exercises.
1. Let F be a one-dimensional foliation on R2 which is vertical for x ≤ 0
and consists of the graphs of the functions y = f (x) + C, C ∈ R, for x > 0.
Show that F is C ∞,0 -smooth for f (x) = ln x, C ∞,1 -smooth for f (x) = 1/x,
and C ∞,∞ = C ∞ -smooth for f (x) = exp(1/x).
2. Show that the Reeb foliation Rβ,f defined above is C ∞ -smooth. Verify
that if we modify condition (b) of Example 3 above to
β(1) = β (1) = β (1) = · · · = β (r−1) (1) = 0, β (r) (1) = c = 0
(r = 1, 2, . . . ), then the resulting foliation Rβ,f has class C ∞,r on S 3 and

C ∞,∞ on S 2 × S 1 .
If F is a C r,r -smooth foliation, then τ (F ) is a C r−1 -smooth distribution. On

the other hand, if τ is an integrable C r−1 -smooth distribution, i.e., τ = τ (F ),
then for F we can guarantee only C r,r−1 -smoothness because integration
does not increase smoothness in the normal direction.
98 10. Foliations
10.2. Haefliger structures and h-principle for foliations

In this section we formulate a theorem on the existence of foliations, and
show that in the case of open manifolds this theorem is a simple corollary
of the Holonomic Approximation Theorem 3.1.1. The proof of this corollary
is our main goal. We will not discuss the proof of Thurston’s h-principle
for foliations on closed manifolds. The reader can find all details related to
this topic in Thurston’s papers [Th74, Th76b], and alternative proofs for
foliations of codimension ≥ 2 in [EM98] and of codimension 1 in [Mei17].
A. Equivalence relations for foliations. Two foliations F0 and F1 of

the same codimension q on a manifold V (without boundary) are said to be
• homotopic if they are homotopic through integrable plane fields;
• concordant if there exists a foliation F of codimension q (a concor-
dance) on V × I transverse to ∂(V × I) = V × 0 ∪ V × 1 such that
F |V ×0 = F0 and F |V ×1 = F1 ;
• integrably homotopic 1 if there exists a concordance F transverse to
V × t for every t ∈ I.
Remarks.
1. Of course, integrable homotopy is the strongest of the three equivalences
which implies the other two. On a closed manifold V , it is an overly strong
equivalence: any two integrably homotopic foliations are isotopic ([Mi70]).
2. The two one-dimensional foliations on the torus T 2 = S 1 × S 1 formed
by the parallels and meridians are homotopic through the family of winding
foliations, but they are not integrably homotopic. At the same time, these
foliations are concordant. (As an exercise, construct the concordance!)
3. In his remarkable note [Thu72] Thurston constructed a family (= homo-
topy) of foliations on S 3 which are pairwise nonconcordant. Therefore,
homotopy does not imply concordance.
4. Conversely, strong concordance implies homotopy for foliations of codi-
mension > 2, see [EM98]. Here, two foliations F0 and F1 of the same
codimension q on a manifold V are called strongly concordant if there exists
a concordance F such that the inclusion i : ν(F ) → T (V × I) is homotopic
through a family of bundle monomorphisms it : ν(F ) → T (V × I) fixed over
V × 0 ∪ V × 1 to a monomorphism ν(F ) → T V × I → T (V × I).
1 Although the term integrable homotopy is universally accepted in the literature on foliations,
it should not mistaken with an integrable homotopy of distributions, which is just a homotopy of
foliations.
10.2. Haefliger structures and h-principle for foliations 99
Foliations behave contravariantly under mappings transverse to a foliation

on the target manifold. We denote by f ∗ F the foliation on V induced by
a map f : V → W transverse to a foliation F on W . If ft : V → W is a
family of maps transverse to a foliation F on W , then ft∗ F is an integrable
homotopy connecting f0∗ F and f1∗ F .
B. Haefliger structures. Given a fiber bundle ν, we will use the notation

Xν for the total space of ν.
A Haefliger structure of codimension q (or Γq -structure, see [Hae70]) on a
manifold V is a pair H = (ν, H) where ν is a q-dimensional vector bundle,
called the normal bundle of H, and H is the germ of a foliation of codi-
mension q on the total space Xν near the zero section V ⊂ Xν which is
transverse to the fibers of ν (see Figure 10.1). Note that the foliation H
need not to be transverse to V .
Figure 10.1. Haefliger structure.
A Haefliger structure H = (ν, H) on a manifold V is called regular if H is

transverse to V . For a regular Haefliger structure H = (ν, H) the intersection
F = H ∩ V is a foliation. Conversely, for any foliation F on V we can
canonically construct a regular Haefliger structure
HF = (ν(F ), HF = (exp ◦ i)∗ F ),
where i : ν(F ) → T V is the inclusion and exp : T V → V is the exponen-

tial map with respect to some Riemannian metric (see Figure 10.2). The
foliation HF is transverse to V and HF ∩ V = F . We will identify the
foliation F with the regular Haefliger structure HF . Similarly, a general
Haefliger structure H can be viewed as a singular foliation whose singular
locus Σ(H) consists of the points of V where H is not transverse to V .
100 10. Foliations
Figure 10.2. Regular Haefliger structure.
Another extreme class of Haefliger structures are the flat ones. A Haefliger
structure (ν, H) is called flat if the zero section is one of the leaves of H,
see Figure 10.3. In this case Σ(H) = V . Note that the bundle ν admits
a flat Haefliger structure if and only if its structure group GL(q) can be
reduced to a discrete group in the larger group D0 (q) of germs at the origin
of diffeomorphisms (Rq , 0) → (Rq , 0).
Figure 10.3. Flat Haefliger structure.
Two Haefliger structures H0 and H1 on V are said to be integrably homotopic

if there exists a Haefliger structure H on V × I (of the same codimension
as H0 and H1 ) such that H|V ×0 = H0 and H|V ×1 = H1 . Note that in the
literature (see e.g., [La74]) an integral homotopy of Haefliger structures is
usually called a homotopy. We prefer the term integral homotopy because
the foliation H is automatically an integral homotopy between H0 and H1 .
Unlike foliations, Haefliger structures behave contravariantly under all maps
of the underlying manifolds, without any transversality conditions. Indeed,
given a codimension q Haefliger structure (ν, H) on a manifold W and a map
ϕ : V → W , we obtain a codimension q Haefliger structure ϕ∗ H = ( on
ν , H)
∗
V by taking the pull-back bundle ν = ϕ ν and the (germ of the) foliation
H = ϕ ∗ H, where the bundle map ϕ : Xν → Xν covers ϕ. A homotopy
ϕt : V → W, t ∈ [0, 1], induces an integrable homotopy between ϕ∗0 H and
ϕ∗1 H. By standard homotopy theory (see [Bro65], [Ha70b]), there exists
for each q a universal space BΓq with a universal Haefliger structure HU such
that for any manifold V all Haefliger structures of codimension q on V can be
10.2. Haefliger structures and h-principle for foliations 101
induced from the structure HU on BΓq , and the integrable homotopy classes
of these structures are in one-to-one correspondence with the homotopy
classes of maps V → BΓq . Associating to each Haefliger structure (ν, H)
the bundle ν gives rise to a canonical fibration BΓq → BOq , where BOq is
the classifying space of real rank q vector bundles, see [Ha70b].
C. Augmented Haefliger structures. An augmented Haefliger structure

H of codimension q on a manifold V is given by a triple (ν, H, i) where (ν, H)
is a Haefliger structure and i : ν → T V is a bundle monomorphism (augmen-
tation) such that i(ν|∂V ) ⊂ T (∂V ). Two augmented Haefliger structures H0
and H1 are called integrably homotopic if the underlying Haefliger struc-
tures H0 and H1 are integrably homotopic and the monomorphisms i0 and
i1 are homotopic. An augmented Haefliger structure H = (ν, H, i) on a
manifold V is called regular if the underlying H = (ν, H) is regular, i.e.,
H = HF = (ν(F ), HF ), and i : ν = ν(F ) → T V is the canonical inclusion.
For a foliation F on V we can canonically equip HF with the augmentation
i : ν(F ) → T V , and therefore we may identify the foliation F with the
regular augmented Haefliger structure HF .
D. Foliations on open manifolds. The following h-principle for foliations

on open manifolds is, in fact, a direct consequence of 8.3.4; the only problem
is the translation into the language of Haefliger structures.
10.2.1. (Gromov–Phillips–Haefliger [Gr69, Ph70, Hae70]) Any aug-
mented Haefliger structure H on an open manifold V is integrably homotopic
to a regular Haefliger structure, i.e., a foliation. Two foliations F0 and F1
on an open manifold V are integrably homotopic if and only if the associated
augmented Haefliger structures HF0 and HF1 are integrably homotopic.
Proof. Let H = (ν, H, i) and H = (ν, H). Let Φ : T V → ν be a left inverse

for i, i.e., Φ ◦ i = idν . The epimorphism Φ gives us a formal solution
: V → Rτ ⊂ J 1 (V, Xν ),
Φ
where Φ covers the diagonal map V → V × Xν and the relation Rτ defines
sections transverse to τ (H), i.e., transverse to the foliation H. According
to 8.3.4, there exists a homotopy Φ t : V → Rτ such that Φ 0 = Φ and
∗
Φ1 = dϕ1 is a genuine solution, thus Φt H is an integrable homotopy between
H = H0 = Φ ∗ H and the regular Haefliger structure H1 = Φ ∗ H. The
0 1
homotopy of augmentations is determined by right inverses of Φ t.
Similarly, the relative version of 8.3.4 proves the 1-parametric part of the
theorem.
A theorem of Haefliger in [Hae70] asserts that the homotopy fiber of the

map ν : BΓk → BOk is k-connected. In particular, any two codimension
102 10. Foliations
k foliations on a manifold which has the homotopy type of a CW complex

of dimension ≤ k are homotopic as Haefliger structures. In view of Theo-
rem 10.2.1 this implies
10.2.2. (Haefliger [Hae70]) Let W be an (n + k)-dimensional manifold,
V ⊂ W a k-dimensional submanifold, and F0 , F1 two n-dimensional folia-
tions on Op V . Suppose that T F0 , T F1 ⊂ T W |Op V are homotopic as not
necessarily integrable distributions. Then they are (integrably) homotopic as
foliations.
E. Foliations on closed manifolds. In his famous papers [Th74, Th76b]
W.P. Thurston proved the following analogue of 10.2.1 for foliations on
closed manifolds:
10.2.3. (h-principle for foliations on closed manifolds) Any aug-
mented Haefliger structure H on a closed manifold V is integrably homo-
topic to a regular structure, i.e., to a foliation. Two foliations F0 and F1 on
a closed manifold V are concordant if and only if the associated Haefliger
structures HF0 and HF1 are integrably homotopic.
Theorems 10.2.1 and 10.2.3 reduce the problem of homotopy of a distribution
ξ on V to a foliation to the homotopy theoretical problem of whether the
classifying map ν : V → BOq of the normal bundle ξ ⊥ to ξ lifts to the
classifying space BΓq for Haefliger structures, see the end of Section 10.2.B.
For instance, this is the case if the structure group of ξ ⊥ is discrete.
10.3. Foliated h-principle

In this section we introduce the notion of a foliated h-principle and show
that, for differential relations on a natural bundle that are invariant under
leafwise diffeomorphisms, it is equivalent to the parametric and fibered h-
principles.
As in Section 3.8 let (V, F ) be a manifold equipped with a foliation, X → V
(r) (r)
a fibration, and XF → V its leafwise r-jet extension. A subset R ⊂ XF
is called a leafwise differential relation. Its formal solutions are sections F :
V → R. Genuine solutions can be equivalently defined either as holonomic
sections F : V → R, or as sections f : V → X with Jfr|F (V ) ⊂ R.
(r)
A leafwise differential relation R ⊂ XF is said to satisfy the foliated h-
principle if any formal solution of R is homotopic to a genuine one, and
the relative foliated h-principle if for any closed subset A ⊂ V any formal
solution F : V → R which is holonomic over Op A is homotopic rel. Op A
to a genuine solution through formal ones.
Recall that any natural bundle over an n-dimensional manifold V is associ-
ated with the canonical principal D0 (n)-bundle P D (V ) over V with respect
10.3. Foliated h-principle 103
to an action α of D0 (n) on the standard fiber of the fibration, hence it yields

a corresponding natural bundle over any other n-dimensional manifold. The
following argument extends this construction to foliated manifolds.
Given an n-dimensional foliation F on a manifold V , we define the principal
D0 (n)-bundle P D (V, F ) over V formed by germs of embeddings (Rn , 0) →
(V, v), v ∈ V , into leaves of the foliation F . By a natural bundle over a
foliated manifold (V, F ) we mean any bundle associated with P D (V, F ) with
respect to an action α of D0 (n) on its standard fiber W . Consider now two
manifolds (V1 , F1 ) and (V2 , F2 ) with n-dimensional foliations (the dimension
of the manifolds could be different) and two natural bundles X1 → V1 and
X2 → V2 associated with the same action α on their common standard fiber.
Let h : V1 → V2 be a map which sends the leaves of F1 to leaves of F2 and
which restricts to leaves as immersions. Then it lifts to a map h∗ : X1 → X2
covering h which restricts to fibers as diffeomorphisms.
Similarly to the nonfoliated case, if X → V is a natural bundle over a
(r)
foliated manifold (V, F ), then XF → V is also a natural bundle. If X is
(r)
associated with an action α of D0 (n) on W , then XF is associated with
the lifted action α(r) of D0 (n) on its fiber W (r) , the space of r-germs at 0
of maps Rn → W . If a subset R ⊂ W (r) is invariant with respect to the
(r)
action α(r) , then it defines a differential relation R = R(V, F ) ⊂ XF which
is invariant with respect to local leafwise diffeomorphisms.
10.3.1. (Parametric, fibered, and foliated h-principles) Let α be

an action of the group D0 (n) on a manifold W , and R ⊂ W (r) an α(r) -
invariant subset of W (r) . Then the relative parametric, fibered, and foliated
h-principles for the corresponding differential relation R are equivalent.
Proof. The relative fibered h-principle is a special case of the relative foli-
ated h-principle when the foliation is given by the fibers of a fibration. In
turn, the relative parametric h-principle is the specialization of the relative
fibered h-principle to the case of a trivial fibration over a disk D k . Hence,
we only need to deduce the foliated h-principle from the relative parametric
one.
Let V be an (n + k)-dimensional manifold endowed with a n-dimensional
foliation F , and A ⊂ V a closed subset. Consider the natural bundle X → V
(r)
associated to the action α and the differential relation R = R(V, F ) ⊂ XF
associated to R. Let F : V → R = R(V, F ) be a formal solution which is
leafwise holonomic on a neighborhood Ω ⊃ A. We cover V by coordinate
charts (Ui , Φi ), i = 1, 2, . . . , where the closures U i are compact and Φi :
Ui → Rk+n are embeddings which send the foliations F ∩ Ui to the foliations
by the fibers of the projection πk : Rk+n → Rk . Choose a covering V = Ui
104 10. Foliations

by smaller open sets Ui with U i Ui and denote

V1 := Φ1 (U1 ), V 1 := Φ1 (U 1 ), A1 := Φ1 (A ∩ U1 ), Ω1 := Φ1 (Ω ∩ U1 ).
Let H1 denote the foliation on V1 by the fibers of the projection πk . Choose
a partition of Rk+n by sufficiently small (k + n)-cubes
Q = ε(a + I k+n ), a ∈ Zk+n ,
such that there exist finitely many cubes Q1 , . . . , QN1 , QN1 +1 , . . . , QN1 +N2
for which

N1

N1+N2
A1 ⊂ Int(Q1 := Qj ) ⊂ Ω1 and V 1 ⊂ Int(Q2 := Qj ) ⊂ V1 .
1 1
On each cube Qj (and similarly on each of its subcubes) the leafwise differ-
ential relation R = R(V1 , H1 ) can be viewed as a differential relation in the
last n coordinates, regarding the first k coordinates as parameters. There-
fore, we can apply the relative parametric h-principle to successively deform
the push-forward formal solution F1 = (Φ1 )∗ F of the differential relation
R(V1 , H1 ) to make it holonomic: First, we make it holonomic near the ver-
tices of cubes which are in Q2 \ Q1 . This can be done via a homotopy fixed
on Op Q1 . Next, by a homotopy fixed on a neighborhood of Q1 ∪ Skel0 (Q2 ),
we extend the deformation to make F1 holonomic on neighborhoods of the 1-
cubes in Q2 \ Q1 . Here we denote by Skelj (Q2 ) the j-dimensional skeleton of
Q2 . Continuing the process we construct a homotopy F1t : V1 → R(V1 , H1 )
which is fixed on Op A1 , connects F10 = F1 with F11 , and is such that F11 |Op Q2
is holonomic. Consider the pull-back homotopy

F1t := Φ∗1 F1t : Op U 1 → R(V, F )|Op U ,
1
and extend it to a homotopy V → R(V, F ) which is fixed on Op A. We keep

the notation F1t for the extended homotopy. To complete the construction,
we repeat the process for all the other coordinate neighborhoods Ui begin-

ning with U2 , the homotopy F1t instead of F , and the closed subset A ∪ U 1
instead of A.
As a corollary, we obtain
10.3.2. (h-principle for leafwise immersions, Gromov [Gr69]) Let
V be a manifold with an n-dimensional foliation F , and W a manifold
of dimension q > n. Then the h-principle holds for leafwise immersions
(V, F ) → W (i.e., maps V → W which are immersions on leaves of F ).
In particular, given any nowhere vanishing vector field X on a manifold V ,
there exists a map f : V → R2 such that df (X) = 0.
Proof. According to 10.3.1, the relative parametric h-principle implies the

foliated one. Hence, the statement follows from Theorem 9.2.1.
10.3. Foliated h-principle 105
Remark. We have previously proved (see 9.3.4) a stronger version of

10.3.2 for not necessarily integrable tangent distributions instead of folia-
tions.
Part 3
Singularities and
Wrinkling
Chapter 11
Singularities
of Smooth Maps
The local theory of singularities of smooth maps is a large and rich subject.
However, rather than cultivating the beautiful germs, our main efforts will
go into the development of methods of how to get rid of them. We review
in this chapter the local theory of only the simplest singularities which will
be important for our purposes, and refer the reader to the book [AGZV85]
for a more comprehensive study of the local theory of singularities.
11.1. Thom–Boardman singularities

Consider manifolds V and W of dimension n and q, respectively, and a
smooth map f : V → W . For k ≥ 0 a point p ∈ V is said to be of type Σk ,
or simply a Σk -point, if
corank dp f := min(n, q) − rank dp f = k.
In other words, k measures how much the rank drops compared to the
maximal possible one. Points of type Σ0 are also called regular, and points of
type Σk with k ≥ 1 are called Σk -singular. We say that f is Σk -nonsingular
if it has no points of type Σj for j ≥ k.
Remark. While our notation follows R. Thom’s original convention, it
differs from some books in singularitiy theory where the index k denotes
dim Ker dp f rather than the corank. For q ≥ n the two numbers coincide,
but if q < n, then dim Ker dp f = corank dp f + (n − q).
In Section 2.2 we defined the subset Σk ⊂ J 1 (V, W ) consisting of 1-jets

of germs of type Σk . Equivalently, recalling that J 1 (V, W ) fibers over
109
110 11. Singularities of smooth maps
J 0 (V, W ) = V × W with fiber Hom(Tx V, Ty W ) over (x, y) ∈ V × W , Σk

is the associated bundle whose fiber over (x, y) is the set of homomorphisms
Tx V → Ty W of corank k. Thus the set Σk (f ) := (Jf1 )−1 (Σk ) is the set of
Σk -points of f . Note that we have

Σk (f ) = Σj (f ).
j≥k
By the Thom Transversality Theorem 2.3.2, for a generic map f : V → W

the set Σk (f ) is stratified by smooth strata Σj (f ), j ≥ k, of codimension
j(|n − q| + j). If, in addition, the map f is Σk+1 -nonsingular, then Σk (f ) is
a submanifold which is closed as a subset.
Exercise. Verify that a generic map V → W is Σ2 -nonsingular if

n 3n
q< + 2 or q > − 2,
2 2
and in particular if q ≤ 3 for any n.
If Jf1 is transverse to Σk , then one can further consider the locus
Σk,j (f ) := Σj (f |Σk (f ) ).
The condition that p ∈ Σk,j (f ) can be expressed in terms of the 2-jet of the
map f at the point p. The corresponding singularities Σk,j ⊂ J 2 (V, W ) were
explicitly defined by Boardman in [Boa67].
Exercise. For Jf1 transverse to Σk show that dim Σk (f ) < q if k ≥ 1,

and Σk,j (f ) = ∅ if j > k + max(0, n − q).
More generally, for any multi-index I = (i1 , . . . , i ) with

i1 + max(0, n − q) ≥ i2 ≥ · · · ≥ i ≥ 0,
J.M. Boardman defined in[Boa67] stratified subsets
ΣI = Σi1 ,...,i ⊂ J (V, W ),
called Thom–Boardman singularities, such that if J −1 (f ) is transverse to
Σi1 ,...,i−1 ⊂ J −1 (V, W ), then
Σi1 ,...,i (f ) := (Jf )−1 (Σi1 ,...,i ) = Σi (f |Σi1 ,...,i−1 ).
To formulate Boardman’s formula, let us introduce the following notation.
Given a multi-index (k1 , . . . , k ) with k1 ≥ k2 ≥ · · · ≥ k ≥ 0, we denote by
μ(k1 , . . . , k ) the number of sequences j1 ≥ · · · ≥ j ≥ 0 such that js ≤ ks
for all s and j1 > 0.
11.1. Thom–Boardman singularities 111
11.1.1. (Boardman’s formula) Let I = (i1 , . . . , i ). Denote i1 := i1 +

max(0, n − q) and suppose that i1 ≥ i2 ≥ · · · ≥ i ≥ 0. Then
codim ΣI
= (i1 + q − n)μ(i1 , . . . , i ) − (i1 − i2 )μ(i2 , . . . , i ) − · · · − (i−1 − i )μ(i ).
Examples.
1. For = 1 and i1 = k, Boardman’s formula recovers our earlier formula
codim Σk = k(|n − q| + k): If q ≥ n, then k = k and Boardman’s formula
gives (k + q − n)μ(k) = (k + |q − n|)k. If q ≤ n, then k = k + n − q, and we
get (k + q − n)μ(k) = k(k + n − q) = k(k + |n − q|).
2. If all coranks are equal to 1, then abbreviating 1 := (1, . . . , 1), we get

n−q+ , n ≥ q,
codim Σ1 =
(q − n + 1) , q > n.
Indeed, if q ≥ n, then 1 = 1 and μ(1 ) = yields codim Σ1 = (q − n + 1) .
If q ≤ n, then 1 = 1 + n − q and μ(j, 1−1 ) = j yields
codim Σ1 = μ(n − q + 1, 1−1 ) − (n − q)μ(1−1 )
= (n − q + 1) − (n − q)( − 1) = n − q + .

Exercise. Show that for q > 3n−12 a generic map f : V → W between
manifolds of dimension n and q has no ΣI -singular points except Σ1,0 , and
for n ≥ q = 2 a generic map has only Σ1,0 and Σ1,1 singularities.
We will discuss the geometry of the singularities Σ1,...,1 in more detail in

Section 11.7. In the case n ≥ q generic singularities of type Σ1,0 are called
folds, and of type Σ1,1,0 cusps. Here we use the word generic in the sense
of transversality of the corresponding jet section to the singularities Σ1 and
Σ1,1 in the jet spaces. These two types of singularities will play an important
role in the book. See Figures 11.1 and 11.2 for folds and cusps in the case
where n = q = 2.
x
x=z2
x=z2
z
z y
y y
Figure 11.1. The fold (y, z) → (y, z 2 ).

x
3 3
x=z +3yz x=z +3yz
z
y
y y
Figure 11.2. The cusp (y, z) → (y, z 3 + 3yz).
A generic singular point of type Σ1,1,1,0 (in the case n ≥ q) is sometimes

called a swallowtail, and the singularity Σ1,0 for n = 2, q = 3 is called the
Whitney umbrella. Figure 11.3 shows f (Σ1 (f )) for a swallowtail f : R3 →
R3 , and f (R2 ) for a Whitney umbrella f : R2 → R3 .
Figure 11.3. Swallowtail and Whitney umbrella.
11.2. The Morse Lemma

Two germs f : V → W at p ∈ V and f : V → W at p ∈ V are called

(smoothly) equivalent if f = g ◦ f ◦ h −1 for germs of diffeomorphisms
, f(
h : (V, p) → (V , p) and g : (W, f (p)) → (W p)).
Exercise.
1. Regular points. Use the implicit function theorem to show that if p is a
regular point for f : V → W , then the germ of f at p is equivalent to the
germ at 0 of the map

(x1 , . . . , xn , 0, . . . , 0) if n < q,
(x1 , . . . , xn ) →
(x1 , . . . , xq ) if n ≥ q.
Thus all germs at regular points between manifolds of fixed dimensions are
equivalent. Note that regularity is an open condition.
11.2. The Morse Lemma 113
2. Critical points of a function. For a function f : V → R we have Σ2 (f ) =

∅, and Σ1 (f ) is just the set of critical points of f . Verify that transversality
of Jf1 : V → J 1 (V, R) to Σ1 means that all critical points are nondegenerate
(also often called Morse).
11.2.1. (Morse Lemma [Mo39]) Suppose that a function ϕ : Op Rn 0 → R

has a nondegenerate critical point at the origin with ϕ(0) = 0. Then there
exists a diffeomorphism h : Op Rn 0 → Op Rn 0 such that h(0) = 0 and

s
n
ϕ ◦ h(x) = Qn,s (x) := − x2i + x2j .
i=1 j=s+1
The number s is called the (Morse) index of the critical point of ϕ. There
are many proofs of this classical result. We will use Moser’s homotopical
method (see [Mo69]), which will also be useful for several other results in
this book.
Proof. We have ϕ(x) = Q(x) + ψ(x), where ψ(x) = δ(x)x2 for a smooth
function δ : Op 0 → R which vanishes at the origin and Q is a nondegenerate
quadratic form. Consider the family of functions ϕt (x) = Q(x) + tψ(x),
t ∈ [0, 1]. Let us look for an isotopy ht : Op 0 → Op 0 with h0 = Id such
that ht (0) = 0 and ϕt ◦ ht = ϕ0 = Q for all t ∈ [0, 1]. Differentiating this
equation, we get
dϕt
(2) 0 = dϕt (y) · Xt (y) + (y) = ∇ϕt (y), Xt (y) + ψ(y),
dt
where y = ht (x) and Xt (y) = dh dt (x) is the time-dependent vector field
t
generating the isotopy ht . We look for the vector field Xt in the form
Xt = gt ∇ϕt for a function gt : Op 0 → R. Then (2) becomes
∇ϕt (y)2 gt (y) = −ψ(y).
By assumption we have ψ(y) = δ(y)y2 , and ∇ϕt (y)2 = γt (y)y2 for a
positive smooth function γt on Op 0. So the equation is solved by gt = − γδt ,
and thus Xt = − γδt ∇ϕt . Note that Xt (0) = 0 for all t ∈ [0, 1]. Therefore,
we can integrate the vector field Xt on a sufficiently small neighborhood of
the origin to obtain the desired isotopy ht : Op Rn 0 → Op Rn 0, t ∈ [0, 1],
with h0 = Id and ϕt ◦ ht = Q. It remains to diagonalize Q by a linear
transformation.
One of the advantages of the above proof is that it works without any changes
in parametric form. In particular, we get
11.2.2. (Parametric Morse Lemma) Let E → B be a vector bundle,

and ϕ : E → R a smooth function which vanishes with its differential along
the zero section B ⊂ E. Suppose that the restriction of ϕ to each fiber has a
nondegenerate critical point at the origin. Then there exist a split Euclidean
vector bundle E + ⊕ E − → B and a diffeomorphism h : Op E + ⊕E − B →
Op E B fibered over B such that h|B = Id : B → B and
ϕ ◦ h(b, e+ , e− ) = e+ 2 − e− |2 , e± ∈ Eb± .
11.3. Tangency of a submanifold to a foliation

Let W be a manifold of dimension k + n endowed with an n-dimensional
foliation F , and V ⊂ W a submanifold of dimension m ≥ k. By definition
of a foliation, for each point v ∈ V there is a neighborhood U ⊂ W and a
submersion sU : U → Rk such that the foliation F |U is given by the fibers
of the submersion sU . The singularities of the restriction sU |V ∩U (which
are independent of the choice of the submersion sU ) are called tangency
singularities of V to F . Similarly to the singularities of maps, there are
thus tangency singularities of Thom–Boardman type ΣI . The corresponding
singular loci will be denoted by ΣI (V, F ). If f : V → W is an embedding,
we will also write ΣI (f, F ) instead of ΣI (f (V ), F ).
A submanifold or embedding is called folded with respect to F if all its
tangency singularities are folds, i.e., of type Σ1,0 . The Parametric Morse
Lemma 11.2.2 implies the following stability result.
11.3.1. Let Ft , t ∈ [0, 1], be a family of n-dimensional foliations on a (k+n)-
dimensional manifold W , and V ⊂ W a submanifold of dimension m ≥ k.
Suppose that V is folded with respect to all foliations Ft . Then there exists a
diffeotopy gt : Op V → Op V such that g0 = Id, gt (V ) = V , and gt∗ Ft = F0
for all t ∈ [0, 1].
Proof. After a diffeotopy of V , we can assume that Σ1 (V, Ft ) = Σ1 (V, F0 ) =:

Σ and T Ft |Σ = T F0 |Σ for all t ∈ [0, 1]. Note that dim Σ = k − 1. Let
us first construct the required isotopy on Op Σ. There is an embedding
h : Σ := Σ × (−ε, ε) → W such that h(σ, 0) = σ and h is transverse to Ft for
all t ∈ [0, 1]. Note that given any codimension k foliation L on a manifold
X transverse to a submanifold Y ⊂ X of dimension k, the germ of L along
Y is diffeomorphic to the germ of the normal bundle to Y in X along its
0-section. Indeed, one can choose a Riemannian metric for which the germ
of L along Y coincides with the germ of the foliation by normal fibers of
a tubular neighborhood. As the space of metrics is contractible, the claim
holds for any family of foliations Lt .
Denote by E the vector bundle T F0 |Σ . Let Σ ⊂ E be its 0-section and E
the foliation by its fibers. Let E be the pullback of E under the projection
11.4. Fibered character of Σ1 -singularities 115
π : Σ × R → Σ and E the corresponding foliation by its n-dimensional fibers.

Applying the above remark to the family of foliations Ft on Op W Σ, we find
an isotopy ht : Op E Σ → Op W Σ which is the identity on Σ and such that
h∗t Ft = E, t ∈ [0, 1]. Denote Vt = h−1
t (V ) ⊂ E. Note that all Vt are folded
with respect to E with fold Σ. Hence, to prove the proposition we need
to find a E leafwise isotopy ft : Op E Σ → Op E Σ, t ∈ [0, 1], which fixes Σ
and such that ft (V0 ) = Vt . The required isotopy gt can then be defined as
gt := ht ◦ ft ◦ h−1
0 .
To construct ft , notice that all Vt are transverse to the foliation of E by the
(germs along Σ of the) (n + 1)-dimensional fibers Gσ := Eσ × R ⊂ E, where
Eσ is the fiber of E through the point σ ∈ Σ. Indeed, for each σ ∈ Σ, we
have
Span(Tσ V, Gσ ) = Span(Tσ Σ, Eσ ) = Tσ E
because Σ is transverse to the foliation E. Denote Vt,σ = Vt ∩ Gσ , so that
1 ≤ dim Vt,σ = m − k + 1 ≤ n (because Σ = ∅ for m = k + n). The
folding condition for Vt with respect to E is equivalent to the folding of
each Vt,σ , t ∈ [0, 1], σ ∈ Σ, with respect to the codimension 1 foliation of
Gσ = Eσ ×R by the fibers Eσ ×s, s ∈ R. Denoting by π : Gσ = Eσ ×R → Eσ
the projection, the map π|Vt,σ : Vt,σ → Eσ has the maximal rank = dim Vt,σ .
Hence, the implicit function theorem (see Exercise 1 at the beginning of
Section 11.2) implies that by a family of germs of fiberwise diffeomorphisms
of the fiber bundle E, which are fixed on Σ, we can transform the images
π(Vt,σ ) ⊂ Eσ into germs of linear subspaces Lσ ⊂ Eσ independent of t.
This transforms the family Vt,σ to the family of graphs of (germs at the
origin of) functions ψt,σ : Lσ → R which all have the origin as their Morse
critical point. Hence, we can apply the parametric Morse Lemma 11.2.2 to
transform these functions by a family diffeomorphisms Lσ → Lσ to a family
of nondegenerate quadratic forms qσ . This gives the desired isotopy ft .
Thus, we have reduced our original problems to the case when the foliations
Ft coincide on a neighborhood of the fold Σ ⊂ V . To normalize Ft elsewhere
on Op V , we just need to apply once more the remark about the germ of a
codimension k foliation transverse to a k-dimensional submanifold.
11.4. Fibered character of Σ1 -singularities

Let V and W be manifolds of dimensions n and q, respectively.
11.4.1. (Fibered form for Σ1 ) Consider a smooth map f : V → W and
denote d := min(n, q) − 1. Then the germ of f at any point p ∈ Σ1 (f ) is
equivalent to the germ at the origin of a fibered map
(3) h : Rd × Rn−d → Rd × Rq−d , (y, x) → (y, hy (x)),
where hy : Op Rn−d 0 → Rq−d is a family of maps parametrized by y ∈ Op Rd 0

and h0 = hy=0 has a critical point at the origin.
In other words, for n ≥ q a map near its Σ1 -singular point looks like a
(q − 1)-parametric family of functions Rn−q+1 → R, and for q ≥ n a map
near its Σ1 -singular point looks like an (n − 1)-parametric family of paths
R → Rq−n+1 . In the former case the function for y = 0 has a critical point
at the origin, and in the latter case the path for y = 0 has velocity 0 at the
origin. We will refer to (3) as a fibered form of a map near its Σ1 -singularity.
Proof. Take a point p ∈ Σ1 (f ) and denote J := dp f (Tp V ). Then J ⊂

Tf (p) W is a subspace of dimension d = min(n, q) − 1. Let us choose local
coordinates (u1 , . . . , ud , . . . uq ) on Op f (p) centered at f (p) such that the dif-
ferentials df (p) u1 , . . . , df (p) ud generate the dual space J ∗ . Then the implicit
function theorem guarantees that the functions
y1 := u1 ◦ f, . . . , yd := ud ◦ f
can be extended to a local coordinate system (y1 , . . . , yd , x1 , . . . , xn−d ) on
Op p. In these coordinates the map f takes the required form (y, x) →
(y, hy (x)) with y = (y1 , . . . , yd ) and x = (x1 , . . . , xn−d ).
Exercises.
1. Suppose that h is in the form (3) and n ≥ q. Show that the origin is of
type Σ1,0 if and only if 0 is a Morse critical point of the function h0 .
2. Suppose that n ≥ q and p ∈ Σ1,0 (f ) is a fold point of a map f : V → W .
Show that the germ of f at p is equivalent to the germ at the origin of the
product map
(4) Id × Qn−q+1,s : Rq−1 × Rn−q+1 → Rq−1 × R,
where Qn−q+1,s is the quadratic form of index s defined in Lemma 11.2.1.
Note, however, that by a change of coordinates on W one can switch the
sign of the quadratic form, so the index is only well defined up to replacing
s by n − q + 1 − s.
As another useful application of the fibered form (3), let us consider the
equidimensional case n = q.
11.4.2. (Equations for Σ1,...,1 (h) in the case n = q) Let n = q and h
have the form (3). Then the submanifold ΣI (h) for I = (1, . . . , 1, 0) is given

j
by the equations
∂hy (x) ∂ j hy (x) ∂ j+1 hy (x)
= 0, . . . , = 0, = 0.
∂x ∂xj ∂xj+1
11.4. Fibered character of Σ1 -singularities 117
Proof. The one-dimensional kernel of dh at the points of Σ1 (h) is generated

by the vector field ∂x∂
|Σ1 (h) . On the other hand, if a submanifold N is given
by equations g1 = · · · = gk = 0, then the condition of tangency of the vector
∂
field ∂x to N is expressed by adding to the above system the equations
∂g1 ∂gk ∂hy
∂x = · · · = ∂x = 0. The equation ∂x = 0 defines Σ (h). Hence, the claim
1
follows by induction on j.
Exercise. Consider a map f : R2 → R2 of the form f (y, x) = (y, h(y, x)).

Show that transversality of Jf2 to Σ1,1 at (0, 0) ∈ Σ1,1 (f ) means that
∂ 3h ∂ 2h
(0, 0)
= 0 and (0, 0) = 0.
∂3x ∂y∂x
Examples.
1. Fix integers r ≥ 1 and s ∈ [0, n − 1], and for any d ≥ r − 1 consider the d-
parametric family of functions hn,r,d,s
t : Rn−1 × R → R, t = (t1 , . . . , td ) ∈ Rd ,
given by the formula

r−1
(5) hn,r,d,s
t (x, z) =z r+1
+ tj z j + Qn−1,s (x)
j=1
(the family is independent of the coordinates tj with j ≥ r). Consider the

map H n,r,d,s : Rd × Rn−1 × R → Rd × R given by the formula
H n,r,d,s (t, x, z) = (t, hn,r,d,s

t (x, z)), t, x, z ∈ Rd × Rn−1 × R.
It is Σ2 -nonsingular and Σ1r+1 -nonsingular, and its Σ1j -singularities for j ≤
r are given by the equations
∂ i hn,r,d,s ∂hn,r,d,s
t
= 0, i = 1, . . . , j, t
= 0, m = 1, . . . , n − 1.
∂z i ∂xm
Explicitly, Σ1r (H n,r,d.s ) = {x = 0, t1 = · · · = tr−1 = 0} and for j < r the
Σ1j -singular locus is given by the equations
x = 0,

t1 = − (r + 1)z r + (r − 1)tr−1 z r−2 + · · · + 2t2 z ,
! "
r(r + 1) r−1 (r − 1)(r − 2)
(6) t2 = − z + tr−1 z r−3
+ · · · + 3t3 z ,
2 2
...
!! " ! " ! " "
r + 1 r−j+1 r−1 j+1
tj = − z + tr−1 z r−j−1
···+ tj+1 z .
j j j
2. Suppose that d ≥ mr − 1. Consider the d-parametric family of paths

γtm,r : R → Rm , t ∈ Rd , and the map Γd,m,r : Rd+1 → Rd × Rm given by the
formulas
(7)

r
r
r
r−1
γtm,r (z) = tj z j , tr+j z j , . . . , t(m−2)r+j z j , t(m−1)r+j z j +z r+1 ,
1 1 1 1
d,m,r m,r
Γ (t, z) = (t, γt (z)).
(γtm,r is independent of the parameters tj for j ≥ mr.) Then Γd,m,r is Σ2 -
nonsingular, and Σ1k -nonsingular for k > r. Its Σ1j -singularities for j ≤ r
are given by the equations
∂ i γtm,r
= 0, i = 1, . . . , j.
∂z i
In particular, Σ1r (Γd,m,r ) = {z = 0, t1 = 0 = · · · = trm−1 = 0}. We leave it
to the reader as an exercise to write explicit equations for the singularities
Σ1j (Γd,m,r ) for j < r.
For d = 1, m = 2, and r = 1 the map Γ1,2,1 : R2 → R3 has the form
(t, z) → (t, tz, z 2 ). This is the Whitney umbrella; see the end of Section 11.1.
11.5. An -singularities of functions

If p is a critical point of a function f : V → R (i.e., dp f = 0), then there
is an invariantly defined second differential d2p f which is a quadratic form
on Tp V . For a vector T ∈ Tp V we have d2p f (T ) = lim f (p+tTt2)−f (p) , and in
t→0
any local coordinate system (x1 , . . . , xn ) centered at p the matrix of d2p f is

1 ∂2f
the Hessian 2 ∂x2 (p) . If d2p f is nondegenerate, then the critical point p
i
is Morse, and this is a generic case for an individual function. However, for
families of functions we generically encounter more complicated singularities.
In this section we consider the simplest case of the so-called An -singularities,
corresponding to the case when corank d2p f = 1. Denote λp := Ker d2p f .
Then we have an invariantly defined cubic differential which is a cubic form
d3p f : λp → R, defined as d3p f (T ) = lim f (p+tTt3)−f (p) . Similarly, if d3p f = 0,
t→0
we can define a quartic form d4p f : λp → R by d4 fp (T ) = lim f (p+tTt4)−f (p) ,
t→0
etc. We say that a critical point p ∈ V has type A1 if d2p f is nondegenerate,
type A2 if dim Ker d2p f = 1 and d3p f = 0, and type Ar for r ≥ 3 if
dim Ker d2p f = 1, djp f = 0 for 3 ≤ j ≤ r, and dr+1
p f = 0.
11.6. Morin’s normal forms 119
Exercises.
1. Recall that the 1-jet bundle J 1 (V, R) → V is the Whitney sum R ⊕ T ∗ V
of the trivial line bundle over V and the cotangent bundle T ∗ V . Under
this identification, the singularity Σ1 ⊂ J 1 (V, R) corresponds to R ⊕ V ,
where V ⊂ T ∗ V denotes the zero section. The Σ1 -singular points are crit-
ical points of functions. Recall the projections prs : J r (V, R) → J s (V, R).
The preimage (p21 )−1 (Σ1 ) ⊂ J 2 (V, R) can be viewed as the space of second
differentials. Denote by A1 the open subset of (p21 )−1 (Σ1 ) formed by non-
degenerate second differentials, and by A1 the locally closed codimension 1
submanifold of (p21 )−1 (Σ1 ) formed by second differentials of corank 1. Set
B := (p21 )−1 (Σ1 ) \ (A1 ∪ A1 ). Define inductively for r = 2, . . . closed subsets
Ar ⊂ (pr+1 −1
2 ) (A1 ) ⊂ J
r+1 (V, R) such that J r+1 (p) ∈ A if and only if p is
f r
a type Ar critical point of the function f . Show that Ar are locally closed
submanifolds of J r+1 (V, R) of codimension n + r − 1.
2. Show that if p is an Ar -singular point of f , then its germ at p is equivalent
to the function Qn−1,s (x1 , . . . , xn−1 ) + xr+1
n for some s = 0, . . . , n − 1.
3. Show that a generic B-nonsingular d-parametric family of functions V →
R consists of functions with critical points of type Ar for r ≤ d + 1. Show
that B is a stratified subset of J 2 (V, R) of codimension n + 4, so for d ≤ 3
a generic d-parametric family of functions has only Ar -critical points with
r ≤ d + 1.
4. Verify that the family of functions hn,r,d,s
t : Rn−1 × R → R, t ∈ Rd , d ≥
r − 1, defined in (5) consists of B-nonsingular functions, and for j = 1, . . . , r
the locus of their Aj -singular points coincides with Σ1j (H n,r,d,s ) given by
equations (6).
11.6. Morin’s normal forms

A. Local normal forms. The following theorem of B. Morin shows that
any generic family of B-nonsingular (i.e., with corank of the second differ-
ential at critical points ≤ 1) functions is locally equivalent to one of the
families hn,r,d,s
t introduced in (5).
11.6.1. (B. Morin [Mo65]) Let ft : Op Rn 0, t ∈ Op Rd 0, be a family of
functions such that f0 has an Ar -singularity (r ≤ d + 1) at the origin, and
the family Jfr+1
t
is transverse to Ar ⊂ J r+1 (Rn , R). Then the family ft is
equivalent to the family hn,r,d,s
t from (5) for some integer s ∈ [0, n].
According to Lemma 11.4.1 any germ of a map f : Rn → Rq at a Σ1 -singular

point can be viewed either as a (q − 1)-parametric family of functions on
Rn−q+1 if q ≤ n, or as an (n − 1)-parametric family of paths in Rq−n+1 if
q ≥ n. Hence, one can expect that Theorem 11.6.1 yields normal forms of
Σ1r -singularities at least in the case n ≥ q. This is indeed the case with
the help of the following lemma, whose proof we leave to the reader as an
exercise.
11.6.2. Suppose that a fibered map h : Rn = Rq−1 × Rn−q+1 → Rq−1 × R,

(y, x) → (y, hy (x)), has a transverse Σ1r -singularity at the origin. Then the
function h0 : Rn−q+1 → R has an Ar -singularity at the origin.
11.6.3. (B. Morin [Mo65]) Let f : Op Rn 0 → Rq be the germ of a map

with transverse Σ1r -singularity at the origin, r ≥ 1.
(a) If n ≥ q, then f is equivalent to the germ at 0 of the map H n−q+1,r,q−1,s
defined in Section 11.4 for some integer s ∈ [0, n − q].
(b) If n ≤ q, then f is equivalent to the germ at 0 of the map Γn−1,q−n+1,r
defined in (7).
Specializing to the case n = q = 2 (and also summarizing our previous dis-

cussion for this case), we get the celebrated theorem of H. Whitney, some-
times (e.g., in [AGZV85]) taken to mark the birth of singularity theory.
11.6.4. (H. Whitney [Whi55]) A generic map f : V → W between

surfaces has only fold and cusp singularities. Its germ is equivalent to
(t, z) → (t, z 2 ) at a fold point, and to (t, z) → (t, z 3 + tz) at a cusp point.
Specializing instead to the case r = 1 and q = 2n − 1, we have t ∈ Rn−1 and
γtn,1 (z) = (t1 z, t2 z, . . . , tn−1 z, z 2 ),
so we get another theorem of Whitney (see [Whi43]): the only generic

singularity of a map Rn → R2n−1 is given by
(t1 , . . . , tn−1 , z) → (t1 , . . . , tn−1 , t1 z, t2 z, . . . , tn−1 z, z 2 ).
For maps R2 → R3 , this is the Whitney umbrella; see Figure 11.3.
Exercise. Let ϕ : [a, b] → [c, d] be a C ∞ -function with the following

properties:
• ϕ(a) = c, ϕ(b) = d, and ϕ > 0 at a and b;
• ϕ has exactly two critical points M and m, a < M < m < b, which
are both nondegenerate with values c < ϕ(m) < ϕ(M ) < d.
# $
Prove that ϕ is equivalent to the function [−1, 1] → − 23 , 23 , z → z 3 − z3 ,
and deduce from this fact Whitney’s Theorem 11.6.4.
11.6. Morin’s normal forms 121
B. Leafwise singularities of a function on a foliated manifold. Recall

that to a foliated manifold (W, F ) we have associated the bundle JFr (W, R)
of leafwise r-jets of functions and, similarly to the nonfoliated case, we can
define subsets Ar ⊂ JFr+1 (W, R) corresponding to functions with leafwise
Ar -singularities. As a consequence of Morin’s normal forms we get
11.6.5. (Leafwise singularities) Let (W, F ) be an (n + k)-dimensional
manifold with an n-dimensional foliation F . Suppose that a function f :
j+1
W → R has leafwise only Aj -critical points and its jet extensions Jf,F are
transverse to the singularity sets Aj , j = 1, . . . , k + 1. Denote by V the locus
of leafwise critical points of f . Then V is a k-dimensional submanifold of
W which has type Σ1j tangency to F at the leafwise Aj -critical points.
In the following, when considering functions with leafwise Aj -critical points

we will always assume transversality of their jet extensions to the corre-
sponding submanifolds in the leafwise jet space.
C. Semilocal normal forms. Let us formulate some semilocal results con-

cerning normal forms of singularities of submanifolds and functions on a
foliated manifold.
We will be especially interested in the case of a function f which has only
A1 and A2 leafwise critical points. According to Lemma 11.6.5, the locus
of leafwise critical points of f is in that case a k-dimensional submanifold
folded with respect to F , with the fold Σ equal to the locus of A2 -critical
points. We denote by Vert the bundle T F |V . It will be convenient for us to
fix a Riemannian metric on W and view the leafwise quadratic differential
d2F f as a self-adjoint operator on Vert. The bundle Vert is transverse to V
over V \ Σ, and λ = (Vert ∩ T V )|Σ is a one-dimensional bundle tangent to V
and transverse to Σ in V . The quadratic differential d2F f defines a splitting
Vert = Vert+ ⊕ Vert− over V \ Σ according to the sign of the eigenvalues
of d2F f . Over Σ, where one eigenvalue degenerates to 0, we have a splitting
Vert|Σ = λ ⊕ Ver+ ⊕ Ver− , where the bundle Ver = Ver+ ⊕ Ver− is the
orthogonal complement to λ in Vert|Σ .
If there are only A1 -singularities, then by the Parametric Morse Lemma
11.2.2 there is a diffeomorphism ϕ : Op Vert V → Op W V which restricts the
identity to V , sends fibers of Vert to leaves of F , and is such that
f ◦ ϕ(v, X) = f (v) + X+ 2 − X− 2 ,
X = (X+ , X− ) ∈ (Vert+ ⊕ Vert− )v , v ∈ V.
A similar normal form, generalizing Whitney’s normal form for a cusp sin-
gularity, exists for a function with Ar critical points; see [Arn76]. For our
purposes we will need only the case r = 2. As we already stated, in this case
V is folded with respect to F along the locus Σ of A2 -critical points. Note
Figure 11.4. Characteristic coorientation of the fold.
that the line bundle λ = T F ∩ T V |Σ coincides with the kernel bundle of

the leafwise quadratic differential d2F f on T F |Σ . Moreover, the bundle λ is
trivial and trivialized by the cubic leafwise differential of f at its A2 -critical
points. Let λ+ ⊂ λ be the unit vector field which defines the trivialization.
Denote by μ the one-dimensional orthogonal complement to Vert|Σ ⊕ T Σ
in T W |Σ . The line bundle μ can be canonically trivialized as follows. The
leaves of F through Σ in a sufficiently small tubular neighborhood U ⊃ Σ
in W form a hypersurface Σ ⊂ U which divides U into two parts U± , where
U− is the part which contains V ∩ U . We trivialize μ by the unit vector
field μ+ which is outward normal to U− ; see Figure 11.4. We call μ+ the
characteristic vector field of the fold Σ.
Note that for an appropriate choice of the metric, the bundle Vert|Σ ⊕ μ can
be viewed as the normal bundle to Σ in W .
11.6.6. (Foliated Whitney Theorem) Let f : W → R be a function with

only A1 and A2 leafwise critical points. Then there exists a diffeomorphism
ϕ : Op Vert⊕μ Σ → Op W Σ which restricts as the identity to Σ, sends fibers
of Ver ⊕ λ to leaves of F , and such that
f ◦ ϕ(σ, X) = f (σ) + X+ 2 − X− 2 + t3 + 3st,

X = (X+ , X− , sμ+ , tλ+ ) ∈ (Ver+ ⊕ Ver− ⊕ μ ⊕ λ)σ , σ ∈ Σ.
We finish this section with a semilocal normal form for equidimensional folds
and cusps. It is a corollary of Lemma 11.4.1 and the Exercise in part A of
this section. In order to formulate the theorem we need to slightly adjust
the normal form for a cusp singularity.
11.7. More about the geometry of maps with Σ1,...,1 -singularities 123
# 1 1$
Let us fix a# family$ of functions θs : [−1, 1] → [−1, 1], s ∈ − 4 , 4 , such that
for all s ∈ − 4 , 4 we have
1 1
% &
1 1
θs (z) = z 3 + sz for z ∈ − , ,
2 2
% &
3 3
(8) θs (z) = z for z ∈ / − , ,
4 4
% &
1 1
θs (z) > 0 for z ∈ / − , .
2 2
# $
Choose a nondecreasing function δ : [−1, 1] → − 14 , 14 such that δ(t) = t
near 0 and δ(t) = ± 14 near ±1. Denote
Θ : [−1, 1] × [−1, 1] → [−1, 1] × [−1, 1], Θ(t, z) = (t, θδ(t) (z)).
Note that Θ has a cusp singularity at the origin and Σ1 (Θ) = {(t, z) ∈
[−1, 1]2 | δ(t) = −3z 2 }.
11.6.7. (Semilocal normal form for folds and cusps) Consider a map
f : V → W where dim V = dim W = n.
(a) Let U ⊂ Σ1,0 (f ) be a compact codimension 0 submanifold with (possibly
empty) boundary. Then there exists a split tubular neighborhood N = U ×
[−1, 1] of U = U × 0 in V and an immersion F : N → W such that
f |N (y, z) = F (y, z 2 ).
(b) Let U ⊂ Σ1,1,0 (f ) be a compact codimension 0 submanifold with (possibly
empty) boundary. Assume that the normal bundle to U in Σ1 (f ) is trivial.
Then there exists a split tubular neighborhood N = U × [−1, 1] × [−1, 1] of
U = U × 0 × 0 in V and an immersion F : N → W such that
f |N (y, t, z) = F (y, Θ(t, z)).
11.7. More about the geometry of maps with

Σ1,...,1 -singularities
Let f : V → W , n = dim V ≥ q = dim W , be a map with only Σ1,...,1 -
singularities. Then under the assumption of transversality we have a chain
of submanifolds
V ⊃ Σ1 (f ) ⊃ Σ1,1 (f ) ⊃ · · · ⊃ Σ1q
of dimensions dim Σ1j (f ) = q − j.
The kernel bundle K := (Ker df )|Σ1 (f ) over Σ1 (f ) has rank n − q + 1. Note
that over Σ1,0 (f ) = Σ1 (f ) \ Σ1,1 (f ) the bundle K is transverse to T Σ1 (f )
and its rank is complementary to the dimension of Σ1 (f ). Over Σ1,1 (f )
we have the one-dimensional subbundle λ := K ∩ T Σ1 (f ) ⊂ T Σ1 (f ). The
Σ1(f)
K
Σ111(f)
Σ 11(f )
Figure 11.5. Chain of singularities Σ1 (f ) ⊃ Σ1,1 (f ) ⊃ Σ1,1,1 (f ).
subbundle λ is transverse to Σ1,1 (f ) over Σ1,1,0 (f ) = Σ1,1 (f ) \ Σ1,1,1 (f ) and

tangent to it over Σ1,1,1 (f ), etc.
Assume now that n = q, so the codimension of Σ1 (f ) in V is equal to 1. We
claim that there is a preferred coorientation of Σ1 (f ) in V along Σ1,1 (f ).
Indeed, let us choose adapted coordinates (3), i.e., view the map near a
point p ∈ Σ1,1 (f ) ⊂ Σ1 (f ) as a fibered map Rn−1 × R → Rn−1 × R. Let π
be the projection Rn → Rn−1 . Then the map
π ◦ f |Σ1 (f ) : Σ1 (f ) → Rn−1
is a map with a fold along Σ1,1 (f ). Hence, f (Σ1 (f )) locally lies on one side
of f (Σ1,1 (f )). Let ν be an outward normal vector at f (p) to the boundary
f (Σ1,1 (f )) of f (Σ1 (f )) in Rn−1 . Since d(π◦f ) : T V → T Rq−1 is a surjection,
there exists a vector ν at ∂ ∈ Σ1,1 (f ) which is transverse to Σ1 (f ) in V and
such that d(π ◦ f )( ν ) = ν. Any such vector field ν is called a characteristic
vector field of the cusp. If the map f in a neighborhood of a cusp point is
presented in its Whitney–Morin normal form
(y, t, z) → (y, t, z 3 + tz), y ∈ Rn−2 , t, z ∈ R,
then
Σ1 (f ) = {t = −3z 2 }, π ◦ f (Σ1 (f )) = {(y, t) | t ≤ 0},
and one can choose the vector field ∂
∂t as the characteristic field ν of the
cusp locus.
Exercise. Show that for n ≥ q there exists a canonical coorientation of
Σ1j (f ) in Σ1j−1 (f ) along Σ1j+1 (f ), j = 2, . . . , q − 1.
Consider again a cusp singularity of a map f : V → W in the case n ≥ q = 2.

Suppose that f is in its Whitney normal form
(t, z, x) → (t, z 3 + tz + Qn−2,s (x)), t, z ∈ R, x ∈ Rn−2 .
Considering t as a parameter, we can view f as a family of functions ht :
Rn−1 → R, ht (z, x) = z 3 + tz + Qn−2,s (x). The function ht has no critical
11.7. More about the geometry of maps with Σ1,...,1 -singularities 125
'
points when t > 0, and it has two critical points (z = ± − 3t , x = 0)
of index s and s + 1 when t < 0. Hence, a cusp point corresponds to a
cancellation of two critical points of neighboring indices in a 1-parametric
family of functions.
More generally, for n ≥ q any singular point p ∈ Σ1r admits a similar
interpretation. Indeed, consider the Morin normal form for a Σ1r -singularity
of a map Rn = Rr−1 × R × Rn−r → Rr :
(t1 , . . . , tr−1 , z, x) → (t1 , . . . , tr−1 , z r+1 + tr−1 z r−1 + · · · + t1 z + Qn−q,s (x)).
Then according to (6) the singularity Σ1r−1 (f ) is given by the system of
equations
r(r + 1) 2
tr−1 = − z ,
2
. . .,
t1 = −((r + 1)z r + (r − 1)tr−1 z r−2 + · · · + 2t2 z).
Viewing the coordinate tr−1 as a parameter, we consider f as a 1-parametric
family of maps gtr−1 : Rn−1 → Rr−1 . Then gtr−1 is Σ1r−1 -nonsingular if
'
2tr−1
tr−1 > 0, and it has two Σ 1 r−1 -singular points z = ± − r(r+1) when tr−1 <
0. Moreover, one can check that these two points have opposite characteristic
coorientations of Σ1r−2 in Σ1r−3 .
Chapter 12
Wrinkles
While critical points of a generic smooth function are Morse, a generic 1-

parametric family of functions has at isolated moments also embryos. When
passing through these moments, two new critical points of neighboring in-
dices are born or die. More generally, in a 1-parametric family of maps
V n → W q , n ≥ q ≥ 1, a new component of the singular locus is generically
born in the form of a wrinkle, which is a (q − 1)-dimensional sphere of fold
points with an equator of cusps. We explore wrinkles and their embryos in
this chapter.
12.1. Standard wrinkles

A. The map wn,q,s . Recall our notation for the standard quadratic form
of index s on Rk ,

s
k
Qk,s (x) = − x2i + x2j .
i=1 j=s+1
For n ≥ q ≥ 1 and 0 ≤ s ≤ n − q consider the 1-parametric family of maps
wt = wn,q,s
t
: Rn = Rq−1 × R1 × Rn−q → Rq = Rq−1 × R1 , t ∈ R1 ,

t
wn,q,s (y, z, x) := y, z 3 + 3(|y|2 − t))z + Qn−q,s (x) .
Note that
t
wn,q,s (y, z, x) = −wn,q,n−q−s
t
(−y, −z, x),
127
128 12. Wrinkles
where x = (x1 , . . . , xn−q ) = (xn−q , . . . , x1 ), and thus, if no additional orien-

tation information is provided, the maps wn,q,s t t
and wn,q,n−q−s are geomet-
rically indistinguishable. Hence, in the definition of wn,q,s t we can consider
the reduced index s which takes values 0, . . . , [(n − q)/2].
B. Standard wrinkles, core, saucer and rim. The standard (n, q, s)-
wrinkle (= standard wrinkle of index s + 1/2) is the germ of the map

1
w = wn,q,s = wn,q,s : (y, z, x) → y, z 3 + 3(|y|2 − 1)z + Qn−q,s (x)
along its core, the q-dimensional unit disk
D q = D q × 0 = {z 2 + |y|2 ≤ 1, x = 0} ⊂ Rq × Rn−q = Rn .
Equivalently, and more conveniently for us, the standard wrinkle can be
viewed as the restriction wn,q,s |Uε to the product domain
Uε := {z 2 + |y|2 < (1 + ε)2 , |x| < ε} ⊂ R × Rn−q
for a sufficiently small ε > 0. The image
D q := wn,q,s (D q ) = {(y, u) ∈ Rq−1 × R | u2 ≤ 4(1 − |y|2 )3 } ⊂ Rq
will be called the saucer of the wrinkle, see Figure 12.1. Note that the image
wn,q,s (Uε ) collapses onto the saucer as ε → 0.
Figure 12.1. Saucer wn,3,s (D 3 ).
The boundary ∂ D q of the saucer is singular along the rim

S q−2 := wn,q,s (S q−2 × 0).
Note, however, that there is a well-defined (q −1)-dimensional tangent plane
to the boundary of the saucer at every point of the rim.
√
t
Note that for each t > 0 the germ of wn,q,s along the disk t D q ⊂ Rn is
equivalent to the germ of wn,q,s along the disk D q ⊂ Rn . For n = q we have
s = 0 and we will refer to wn,n := wn,n,0 as the standard (n, n)-wrinkle.
12.2. Properties of standard wrinkles 129
C. Half-wrinkles and embryos. Let Rk+ = [0, ∞) × Rk−1 ⊂ Rk . Note

that wn,q,s (Rn+ ) = Rq+ . We will refer to the germ of the map w+ = wn,q,s |Rn+ :
Rn+ → Rq+ along the q-dimensional unit half-disk
q q
D+ = D+ × 0 = {y1 ≥ 0, z 2 + |y|2 ≤ 1, x = 0} ⊂ Rq+ × Rn−q = Rn+
as the standard half-wrinkle of index s+1/2. Equivalently, the standard half-
wrinkle can be viewed as the restriction wn,q,s |Uε,+ to the product domain
Uε,+ = {y1 ≥ 0, z 2 + |y|2 < (1 + ε)2 , |x| < ε} ⊂ Rq+ × Rn−q = Rn+
for a sufficiently small ε > 0. Note that half-wrinkles are defined only when
n ≥ q ≥ 2.
The standard (n, q, s)-embryo (= standard embryo of index s + 1/2) is the
germ at 0 ∈ Rn of the map

w0 = wn,q,s
0
: (y, z, x) → y, z 3 + 3z|y|2 + Qn−q,s (x) .
As in the case of wrinkles, it will be more convenient for us to view the stan-
dard embryo as a map defined on a sufficiently small open ε-ball centered
at 0 ∈ Rn . Note that in contrast to folds and cusps, which are stable singu-
larities for individual maps, embryos are unavoidable only in 1-parametric
families of maps. Under a generic perturbation, an individual embryo either
disappears or gives birth to a wrinkle.
D. Regularized differential of standard wrinkles and embryos. Al-

though the differential dw = dwn,q,s : Rn → Hom(Rn , Rq ) of the standard
wrinkle degenerates along Σ(w) = ∂D q , it has a regularization
R(dw) : Rn → Epi(Rn , Rq )
which agrees with dw outside Uε ⊃ D q and is canonical up to homotopy rel.
∂Uε . For this, we simply replace in the Jacobi matrix of w the entry aq,q =
3(|y|2 + z 2 − 1) by a function γ : Rn → R which agrees with 3(|y|2 + z 2 − 1)
outside Uε and does not vanish along Uε ∩(Rq ×0). See Figure 12.2 for w1,1,0
and w2,1,0 . The regularized differential R(dw0 ) of an embryo is defined in a
similar way.
12.2. Properties of standard wrinkles

A. Singularities. Let us compute the Thom–Boardman stratification of
the standard wrinkle w = wn,q,s . The rank of the differential dw(y, z, x)
drops to q − 1 if x = 0 and
∂ 3
0= z + 3(|y|2 − 1)z = 3(|y|2 + z 2 − 1),
∂z
so the singular set is the (q − 1)-sphere
Σ1 (w) = ∂D q = {|y|2 + z 2 = 1, x = 0}.
130 12. Wrinkles
z z
x
z
Figure 12.2. Regularization: replacing aq,q by γ for w1,1,0 and w2,1,0 .
Along this sphere we have

Ker dw(y, z, 0) = Span{∂z , ∂x1 , . . . , ∂xn−q } for (y, z, 0) ∈ Σ1 (w),
which intersects T(y,z,0) Σ1 (w) in the one-dimensional space Span{∂z } if z =
0, and trivially if z = 0. So we have Σ1 (w) = Σ1,0 (w) ∪ Σ1,1 (w) with
Σ1,0 (w) = {|y|2 + z 2 = 1, z = 0, x = 0},
Σ1,1 (w) = {|y|2 = 1, z = 0, x = 0}.
Thus w has folds of index s + 1 along the lower hemisphere ∂D q ∩ {z < 0},
folds of index s along the upper hemisphere ∂D q ∩ {z > 0}, and cusps of
index s + 1/2 along the the equatorial (q − 2)-sphere ∂D q ∩ {z = 0}. We will
sometimes use the term wrinkle also for the singularity Σ(wn,q,s ) = S q−1 ,
which is the boundary of the core of the wrinkle.
For t > 0 the wrinkle wt = wn,q,s
t has singularities
Σ1,0 (wt ) = {|y|2 + z 2 = t, z = 0, x = 0},
Σ1,1 (wt ) = {|y|2 = t, z = 0, x = 0}.
Note that
3
t
wn,q,s ({|y|2 + z 2 ≤ t, x = 0}) = {|y| ≤ t, |z| ≤ 2(t − |y|2 ) 2 } ⊂ Rq .
In particular,
t
wn,q,s ({|y|2 + z 2 ≤ t, x = 0}) ⊂ wn,q,s
1
({|y|2 + z 2 ≤ 1, x = 0}).
B. Visualization.
1. For q = 1, the standard wrinkle wn,1,s is the restriction of the function
wn,1,s : R1 × Rn−1 → R, (z, x) → z 3 − 3z + Qn−1,s (x)
to a neighborhood Op D 1 of the gradient trajectory D 1 = D 1 ×0 ⊂ R1 ×Rn−1
connecting a pair of critical points of indices s + 1 and s.
See Figures 12.3, 12.4, and 12.5, where the z-direction is drawn vertically.
w
2
z
1 1
2
Figure 12.3. The wrinkle w1,1,0 .
Figure 12.4. Level curves of the wrinkle w2,1,0 .
Figure 12.5. Level surfaces of the wrinkle w3,1,1 .
2. For n = q = 2 one can draw the graph of the second coordinate function
(y, z) → z 3 +3(y 2 −1)z of the wrinkle w2,2,0 , see Figure 12.6 (the black round
spot in the center is an artefact of using spherical coordinates for plotting).
3. One way to visualize the standard wrinkle w = wn,q,s is to look at the
images of the (q − 1)-spheres Srq−1 = {|y|

2 + z 2 = r 2 , x = 0}. For (y, z, 0) ∈

Sr let us write (y, u) = w(y, z, 0) = y, z(r2 + 2|y|2 − 3) , so that w(Srq−1 )
q−1
is described by the equation

u2 = (r2 − |y|2 )(3 − r2 − 2|y|2 )2 .
132 12. Wrinkles
Figure 12.6. The wrinkle w2,2,0 : graph of second coordinate function.
For r < 1 the second factor on the right hand side is positive and we see
that w(Srq−1 ) is an embedded sphere. For r > 1 the right
hand side vanishes
along the concentric spheres {|y| = r} and {|y| = (3 − r2 )/2 < r}, so
w(Srq−1 ) is an immersed sphere. For r = 1 the equation becomes
u2 = 4(1 − |y|2 )3 = 4(1 + |y|)3 (1 − |y|)3 ,
r<1 r>1
r=1
u u u
y y y
Figure 12.7. The images w2,2,0 (Sr ).

Figure 12.8. The image w3,3,0 (Sr ) for r < 1 (embedding), r = 1

(cubic cusp along equator) and r > 1 (immersion).
so w(S1q−1 ) is a sphere with a cubic cusp along {|y| = 1, u = 0}. See

Figure 12.7 for n = q = 2, and Figure 12.8 for n = q = 3.
4. Another way to visualize the map w = wn,q,s is to view w as the family
of functions
uy = R1 × Rn−q → R, uy (z, x) = z 3 + 3(|y|2 − 1)z + Qn−q,s (x),
parametrized by y ∈ Rq−1 . The function uy has a pair of nondegenerate

critical points of indices s + 1, s for |y| < 1, an embryonic critical point for
|y| = 1, and no critical points for |y| > 1. So uy possesses a pair of critical
points over the ball {|y| < 1} which cancels as y crosses the (q − 2)-sphere
{|y| = 1}. See Figure 12.9 for w3,2,0 .
Figure 12.9. The wrinkle w3,2,0 as the family of functions uy .
Exercise. Find ε0 > 0 with the following property: wn,q,s |O p ∂Uε does
not extend to a submersion Uε → Rq if ε ≤ ε0 , and it does if ε > ε0 .
134 12. Wrinkles
C. Fibered character of the standard wrinkle. Suppose we are given

a splitting Rq−1 = Rk × Rq−k−1 , k ≤ q − 1. The standard wrinkle wn,q,s
defines a map
wn,q,s
Rk × Rq−1−k × RR1 × Rn−q / Rk × Rq−1−k × R
RRR p
RRR
RR ppppp
pk RRRRR ppppk
RR( w pppp
Rk
fibered over Rk , where pk is the projection onto the first k coordinates. Let
us write y = (y , y ) ∈ Rk × Rq−k−1 . Then we have

wn,q,s (y , y , z, x) = y , y , z 3 + 3(|y |2 − (1 − |y |2 ))z + Qn−q,s (x)
t(y )
= y , wn−k,q−k,s (y , z, x) , t(y ) := 1 − |y |2 .
Hence, wn,q,s fibers over a neighborhood of the unit disc D k ⊂ Rk with re-
spect to the projection pk and its fibers are (n − k, q − k, s)-wrinkles over
points of Int D k , and embryos over points of ∂D k . The standard wrinkle
wn,q,s |Uε viewed as a map fibered over Rk will be referred as the standard
fibered wrinkle. Its regularized fiberwise differential defines a fibered epi-
morphism
R(dRk wn,q,s )
Rk × T (Rq−1−k × R1 × Rn−q ) / Rk × T (Rq−1−k × R)
SSS nn
SSS
SSS
SSS nnnnn
SSS nnn pk
wnnn
pk
SS)
Rk
Note that a half-winkle, which is the germ of the map wn,q,s |Rn+ : Rn+ → Rq+
along the half-disc
q q
D+ = D+ × 0 = {y1 ≥ 0, z 2 + |y|2 ≤ 1, x = 0},
fibers over a neighborhood of the half-disc D+ k = {y ≥ 0} ∩ D k ⊂ Rk . The
1
restriction of this fibered map to the (k − 1)-dimensional disc
D k−1 := D+
k
∩ {y1 = 0}
is the standard (n − 1, q − 1, s)-wrinkle fibered over D k−1 .
Chapter 13
Wrinkled Submersions
We define in this chapter the notion of a wrinkled submersion and prove

the corresponding h-principle, which is what can be salvaged from the
Smale–Hirsch-Phillips-Gromov theory for immersions and submersions of
open manifolds in the case of closed manifolds.
13.1. Definitions
A. Wrinkled submersions. Let n ≥ q ≥ 1. We continue using the nota-
tion
Uε = {|y|2 + z 2 < (1 + ε)2 , |x| < ε} ⊂ Rq−1 × R × Rn−q
from the previous chapter and write D q for the core {|y|2 + z 2 ≤ 1, x = 0} ⊂
Rn of the standard wrinkle wn,q,s : Uε → Rq .
Given an n-dimensional manifold V and q-dimensional manifold W , a map
f : V → W is called a wrinkled submersion if there exist disjoint embeddings
ϕi : Uε → V , i = 1, . . . , N , such that
• f |V \i ϕi (∂Dq ) is a submersion;
• each composition f ◦ ϕi : Uε → W can be factored as
wn,q,s gi
Uε / Rq / W,
where wn,q,s is the standard wrinkle of index s + 1/2 and gi is an

immersion.
As it will be clear from our constructions below (see Section 13.4), the im-
mersions gi in the above definition can always be adjusted to be embeddings.
The singular set Σ(f ) of a wrinkled submersion f is the union of the (q − 1)-
spheres Si = ϕi (∂D q ), with the (q − 2)-equators Ci ⊂ Si of cusp points
135
136 13. Wrinkled Submersions
dividing Si into two hemispheres Si,− and Si,+ of fold points of indices
s + 1 and s, respectively. The spheres Si bound the embedded q-discs Di =
ϕi (D q ), the cores of the wrinkles. The images f ◦ ϕi (Uε ) ⊂ W are contained
in neighborhoods of the saucers Di := f (Di ). The boundary ∂ Di = f (∂Di )
is singular along the rim f (Ci ) of the saucer f (Di ), and there is a well defined
tangent plane Tf (v) (∂ Di ) = df (Tv V ) for each point v ∈ Ci . Sometimes we
will use the term wrinkle for the components of the singular locus Σ(f ) of
a wrinkled submersion f : V → W .
Example. In the case W = R, a wrinkled submersion f : V → R is a
Morse function whose critical points are organized in pairs of neighboring
indices connected by a gradient-like trajectory Di , the core of the corre-
sponding wrinkle.
Note that the saucers of a wrinkled submersion f are allowed to intersect

each other. In general, they can also have self-intersections, as we allow gi
in the definition of a wrinkle to be an immersion.
The differential df : T V → T W is not surjective at points of the singular
locus Σ(f ) = Si , but it can be regularized on every neighborhood Ui =
ϕi (Uε ), canonically up to homotopy, to an epimorphism R(df ) : T V → T W
such that R(df ) = df outside of these neighborhoods.
B. Fibered wrinkled submersions. We will define fibered wrinkled sub-

mersions only for the case of a pair of trivial fibrations. The general case
of arbitrary fibrations over the same base can be treated similarly, or as a
special case of the foliated wrinkled submersions defined in part C below.
Let n ≥ q, k < q, dim V = n − k, dim W = q − k, and let P be a space of
parameters, dim P = k. Let
f : P × V → P × W, (p, v) → (p, fp (v))
be a map fibered over P . We say that f is a fibered wrinkled submersion if
there exist disjoint embeddings ϕi : Uε → P × V , i = 1, . . . , N , which send
fibers of the projection
Uε → (Rk × Rq−k−1 ) × R × Rn−q → Rk
to fibers of the projection P × V → P , and such that
• f |(P ×V )\ ϕi (∂Dq ) is a fiberwise submersion;
i
• each composition f ◦ ϕi : Uε → P × W can be factored as

wn,q,s gi
Uε / Rk × Rq−k = Rq / P × W,
where wn,q,s is the standard wrinkle of index s+1/2 fibered over Rk

(see Section 12.2.C) and gi is an immersion covering an immersion
13.1. Definitions 137
g i : Rk → P such that the following diagram commutes:

gi
Rq / P ×W
pk p
gi
Rk / P.
As it will be clear from our constructions below in Sections 13.4 and 13.5,
a fibered wrinkled submersion can always be modified to make gi and g i in
the above definition embeddings.
The fiberwise differential dP f : P × T V → P × T W of a fibered wrinkled
submersion f : P × V → P × W can be regularized on every neighborhood
Ui = ϕi (Uε ), canonically up to homotopy, to an epimorphism R(dP f ) :
P × T V → P × T W fibered over P .
In the fibered case we will allow the manifold P to have boundary and f to
have fibered half-wrinkles, i.e., to admit embeddings
ψi : (Uε,+ , ∂Uε,+ ) → (P × V, ∂P × V ),
where Uε,+ = Uε ∩ {y1 ≥ 0} and ∂Uε,+ = Uε ∩ {y1 = 0}, which send fibers of
the projection pk : Uε,+ → Rk+ to fibers of the projection P × V → P such
q
that f ◦ ψi factors as gi ◦ wn,q,s
+ , where w +
n,q,s : Uε → R+ is the standard
+
half-wrinkle fibered over Rk+ and gi : (Rq+ , Rq−1 ) → (P × W, ∂P × W ) is an

immersion fibered over P . Note that in this case the restriction
f |∂P ×V : ∂P × V → ∂P × W
is itself a wrinkled submersion fibered over ∂P , whose wrinkles are the inter-
sections of the half-wrinkles of f with the boundary ∂P ×V . The regularized
fiberwise differential R(dP f ) is also defined in this case.
C. Foliated (leafwise) wrinkled submersions. Consider an n-dimen-

sional manifold V with a foliation F of codimension k, and a manifold W
of dimension q ≤ n − k = dim F . A map f : (V, F ) → W is called a
foliated (or leafwise) submersion if the restriction of f to each leaf of the
foliation F is a submersion. We say that a map f : (V, F ) → W is a foliated
(or leafwise) wrinkled submersion if there exist disjoint leafwise embeddings
ϕi : (Uε , Fk ) → (V, F ), i = 1, . . . , N , where Fk is the foliation of Uε by the
fibers of the projection
Uε → (Rk × Rq−k−1 ) × R × Rn−k → Rk ,

such that
• f |V \(ϕi (∂Dq )) is a leafwise submersion;
i
• each composition f ◦ ϕi : Uε → W can be factored as

wn,k+q,s gi
Uε / Rk × Rq πk / Rq /W
where wn,k+q,s : Uε → Rk+q = Rk × Rq is the standard wrinkle of

index s + 1/2 fibered over Rk , πk : Rk × Rq → Rq is the projection
to the second factor, and gi : Rq → W is an immersion.
The leafwise differential dF f : τ (F ) → T W of a foliated wrinkled submer-
sion f can be regularized on every neighborhood Ui = ϕi (Uε ), canonically
up to homotopy, to a leafwise epimorphism R(dF f ) : τ (F ) → T W .
The special case of a foliated wrinkled submersion when W = R is called a
foliated (or leafwise) wrinkled function. Thus a foliated wrinkled function
f : V → R has only leafwise critical points of type A1 and A2 , and in
addition its Morse (= A1 ) critical points are organized in pairs of neighboring
indices connected by a gradient-like trajectory, where each pair is allowed to
degenerate into an embryo (= A2 ) point. The leafwise critical point locus of
f is a union of spheres Si contained in coordinate neighborhoods, and each
Si is folded with respect to F with fold locus consisting of embryo points.
13.2. Main results

A. Main technical theorem. The following theorem is the main technical
result of this chapter.
13.2.1. (Fibered wrinkled submersion of a cube) Let
f0 : I k × I n−k → I k × W,
dim W = q − k, be a map fibered over I k which restricts to a submersion on
the fibers near the boundary K = ∂(I k × I n−k ). Let F0 : I k × T (I n−k ) →
I k ×T W , bs F0 = f0 , be an epimorphism fibered over I k such that F0 = dI k f0
on Op K. Then there is a homotopy Ft : I k ×T (I n−k ) → I k ×T W , t ∈ [0, 1],
of fibered epimorphisms fixed over Op K such that ft = bs Ft is C 0 -close
to f0 , f1 : I k × I n−k → I k × W is a fibered wrinkled submersion, and
F1 = R(dI k f1 ).
In Section 13.3 we will prove Theorem 13.2.1 for n = q, and in Section 13.5
for n > q.
B. Main theorems. The following h-principles for wrinkled submersions

are the main results of this chapter.
13.2. Main results 139
13.2.2. (Nonparametric h-principle for wrinkled submersions) Let

dim V = n ≥ dim W = q. Let f0 : V → W be a smooth map which is covered
by an epimorphism F0 : T V → T W such that F0 = df0 near a (possibly
empty) closed subset A ⊂ V . Then there is a homotopy Ft : T V → T W ,
t ∈ [0, 1], of surjective homomorphisms fixed over Op A such that ft = bs Ft
is C 0 -close to f0 , f1 is a wrinkled submersion V → W , and F1 = R(df1 ).
Proof. Fix a triangulation of the manifold V . On the (n − 1)-skeleton

we apply the Holonomic Approximation Theorem 3.1.1, and then on the
n-simplices Theorem 13.2.1 with k = 0.
13.2.3. (h-principle for foliated wrinkled submersions) Let (V, F )

be an n-dimensional manifold endowed with a codimension k foliation F ,
A ⊂ V a closed subset, and W a manifold of dimension q ≤ n − k = dim F .
Let F0 : TF V → T W , bs F0 = f0 , be a leafwise epimorphism such that
F0 = dF f0 on Op A. Then there is a homotopy Ft : TF V → T W , t ∈ [0, 1],
of leafwise epimorphisms fixed over Op A such that ft = bs Ft is C 0 -close to
f0 , f1 is a foliated wrinkled submersion (V, F ) → W , and F1 = R(dF f1 ).
Proof. Fix a Thurston triangulation of the foliated manifold (V, F ) (see

5.2.4). On the (n − 1)-skeleton we apply the Foliated Holonomic Approx-
imation Theorem 3.8.2, and then on the n-simplices Theorem 13.2.1 with
k = codim F .
The following special case when W = R will play an important role in

Chapter 16.
13.2.4. (h-principle for foliated wrinkled functions) Let (V, F ) be
an n-dimensional manifold endowed with a codimension k foliation F and
A ⊂ V a closed subset. Let α0 be a nonvanishing leafwise 1-form and f0 :
V → R a function such that α0 = dF f0 on Op A. Then there is a homotopy
αt of nonvanishing leafwise 1-forms and a family of functions ft : V → R,
t ∈ [0, 1], such that ft is C 0 -close to f0 , f1 is a wrinkled function, and
α1 = R(dF f1 ).
13.2.5. (Relative fibered h-principle for wrinkled submersions) Let
n ≥ q > k, dim V = n − k, dim W = q − k, and let P be a space of
parameters, dim P = k. Let f0 : P × V → P × W be a map fibered over P
such that
f0∂ := f0 |∂P ×V : ∂P × V → ∂P × W
is a wrinkled submersion fibered over ∂P . Let F0 : P × T V → P × T W be
a fibered epimorphisms with bs F0 = f0 such that
F0 |∂P ×T V = R(d∂P f0∂ ) : ∂P × V → ∂P × W.
Then there exist a homotopy of fibered epimorphisms Ft : P ×T V → P ×T W ,

t ∈ [0, 1], fixed on ∂P × T V such that ft = bs Ft is C 0 -close to f0 , f1 is a
fibered wrinkled submersion P × V → P × W , and F1 = R(dP f1 ).
Proof. The idea is to extend the wrinkles of f0∂ from ∂P × V into P × V

as half-wrinkles and then apply Theorem 13.2.3 to the complement.
Choose a small interior collar C := [0, 1] × ∂P ⊂ P along 0 × ∂P = ∂P and
adjust f0 so that f0 (y1 , p, v) = (y1 , f0∂ (p, v) for all y1 ∈ [0, 1], p ∈ ∂P, v ∈ V .
By definition of a fibered wrinkled submersion there are disjoint embeddings
ϕi : Uε → ∂P × V , Uε ⊂ Rn−1 , j = 1, . . . , N , such that the compositions
f0∂ ◦ ϕj can be factored as f0∂ ◦ ϕj = gj ◦ wn−1,q−1,s , where w = wn−1,q−1,s :
Uε → Rq−1 is the standard fibered wrinkle and gj : Rq−1 → ∂P × W is a
fibered equidimensional immersion. Slightly renaming the variables we write

w(y , z, x) = y , z 3 + 3z(|y |2 − 1) + Qn−q,s (x)
for y = (y2 , . . . , yq−1 ), z ∈ R, x ∈ Rn−q . We view the wrinkle w as fibered
with respect to the projection (y2 , . . . , yq−1 , z, x) → (y2 , . . . , yk ). Denote
Uj := [0, 1] × ϕj (Uε ) ⊂ [0, 1] × ∂P × V = C × V . Consider the map
w = Id × w : [0, 1] × Uε → [0, 1] × Rq−1
and the standard half-wrinkle w+ : [0, 1] × Uε → [0, 1] × Rq−1 ,

w+ (y, z, x) y, z 3 + 3z(|y|2 − 1) + Qn−q,s (x)
for y = (y1 , y ) = (y1 , . . . , yk ), y1 ∈ [0, 1], (y , z, x) ∈ Uε . Note that w|0×Uε =
w+ |0×Uε .
Let wt = (1 − 2t)w + 2tw+ , t ∈ [0, 12 ], be the linear homotopy connecting w
and w+ , We define the required fibered homotopy ft on Uj , j = 1, . . . , N ,
by the formula

ft (y1 , u) = y1 , gj (wt (y1 , ϕ−1
j (u)) , y1 ∈ [0, 1], u ∈ ϕj (Uε ).
Let us extend ft arbitrarily to a fibered homotopy ft : P × V → P × W ,

t ∈ [0, 12 ], fixed on ∂P × V . The homotopy ft can be covered by a homotopy
of epimorphisms P × T V → P × T W fixed on ∂P × T V such that F1/2 |Uj
is the regularization of the half-wrinkle f1/2 |Uj . Denote

N
U := (P × V ) \ Uj .
1
Applying Theorem 13.2.3 to U (foliated by the fibers p × V ), F1/2 , and

f1/2 = bs F1/2 , we can extend the homotopy (Ft , ft ) for t ∈ [ 12 , 1] as fixed

N
on Op (∂P × V ∪ Uj ) so that f1 : P × V → P × W is the required
1
13.2. Main results 141
fibered wrinkled submersion, and F1 = R(dP f1 ) is its fibered regularized

differential.
C. Homotopy equivalence for wrinkled submersions. As we have

seen in Section 10.3.C, a fibered h-principle can usually be reformulated as
a parametric one, and the parametric h-principle we usually associate with
a (weak) homotopy equivalence of spaces of genuine and formal solutions
to the considered problem. Hence, one would expect that Theorem 13.2.5
should imply that the space of generalized wrinkled submersions Wr(V, W )
of V to W (where wrinkles are allowed to degenerate to embryos) is homo-
topy equivalent to its formal counterpart, the space of bundle epimorphisms
T V → T W . Indeed, this is true for an appropriate definition of the space
Wr(V, W ). Namely, one needs to define an element of Wr(V, W ) as a triple
(f, {Dj , πj }) where f : V → W is a wrinkled immersion, {Dj } a collection
of unparameterized cores for the wrinkles of f , and {πj : Dj → ∂− Dj } a
collection of submersions fixed on the lower hemisphere ∂− Dj .
Exercise (h-principle for wrinkled functions). Given a manifold V with a

fixed Riemannian metric let us denote by Wr(V ) the space of wrinkled func-
tions, i.e., functions with Morse or embryonic critical points whose Morse
critical points are organized in pairs of points of neighboring indices con-
nected by a chosen gradient trajectory. Define an appropriate topology
on Wr(V ) and prove that the regularized differential defines a homotopy
equivalence Rd : Wr(V ) → Ω1=0 (V ), where Ω1=0 (V ) denotes the space of
nonvanishing 1-forms on V .
D. Outline of the proof of Theorem 13.2.1. The proof of Theorem

13.2.1 uses three auxiliary constructions which we present in the next three
sections. The first two constructions, the wrinkled right inverse for multi-
folds (13.3) and chopping of wrinkles (13.4), are quite universal and useful
for other applications besides wrinkled submersions. The third construction,
corrugation of wrinkles (13.5), is rather special and is intended exclusively
for wrinkled submersions in the case of n > q. In fact, for the case of wrin-
kled immersions (n = q) we need only the first construction of the wrinkled
right inverse, so we will consider this case at the end of Section 13.3 be-
fore discussing chopping and corrugation. Thus, the reader interested only
in the case of immersions can skip these two sections which are somewhat
technical.
We note, however, that even in the equidimensional case the chopping con-
struction is useful if one would like to arrange the saucers of the wrinkles to
be embedded, which is useful for some applications.
13.3. Wrinkled immersion associated with a multifold

A. Wrinkled right inverse for multifolds. To relate wrinkles and mul-
tifolds we will construct for every multifold π : V → V a homotopically
:V →V
canonical wrinkled immersion, called its wrinkled right inverse, w
such that the regularization R(dw) : T V → T V of the wrinkled immer-
sion w = π ◦ w : V → V is homotopic to the identity through fiberwise
isomorphisms. We start with the following obvious one-dimensional lemma.
b1 δ /2
a1+ δ /2
ã 1 b̃
1
Figure 13.1. The function w1 .
13.3.1. (Wrinkled right inverse for a multifolded interval) For any

multifolded interval I = I(ai , bi , N, δ) there exists a wrinkled immersion w1 :
I → I such that (see Figure 13.1)
• w1 is the identity on Op ∂I;
• Σ(w1 ) consists of nondegenerate maxima a1 , . . . ,
aN and nondegen-

erate minima b1 , . . . , bN ;
ai ) = bi − δ/2 and w1 (bi ) = ai + δ/2;
• w1 (
ai −
• for sufficiently small ε the restrictions of w1 to the intervals (

ε, bi + ε) are wrinkles;
• there exists a wrinkled immersion w 1 : I → I (the wrinkled lift of
w1 ) such that the following diagram commutes:
I
@
1
w
π

w1
I / I.
Note that we can choose the function w1 in such a way that, in addition to
the properties listed in Lemma 13.3.1, the restriction of the derivative w1 to
ai −ε, bi +ε) is convex for all i, i.e., w1 |(ai −ε,bi +ε) > 0. Using an
the interval (
13.3. Wrinkled immersion associated with a multifold 143
appropriate homotopy w1t , t ∈ [0, ρ], between w10 = id : I → I and w1ρ = w1 ,

we obtain the following n-dimensional generalization of Lemma 13.3.1:
13.3.2. (Wrinkled right inverse for a multifolded cube) For any
multifolded cube I n = I n (ai , bi , N, δ, ρ) there exists a wrinkled immersion
w : I n = I n−1 × I → I n
fibered over I n−1 with N wrinkles such that
• w is the identity on Op ∂I n ;
• w|p×I = w1 for p ∈ I1−ρ
n−1
, and w|p×I = w1t for p ∈ ∂I1−t
n−1
with
t ∈ [0, ρ];
: I n → I n (the wrinkled lift of
• there exists a wrinkled immersion w
w) such that the following diagram commutes:
I n
| |>
|
w |
||
π
|| w
In / I n.
Note that the wrinkles (= singular sets) Σ(w) and Σ(w) of the maps w and
coincide in the domain I n , see Figure 13.2 for n = 2.
w
and w = π ◦ w
Figure 13.2. Wrinkles of the maps w in the domain
I n , n = 2.
Lemma 13.3.2 for multifolded cubes yields the corresponding statement for
multifolds:
13.3.3. (Wrinkled right inverse for a multifold) For any multifold

π : V → V there exists a homotopically canonical wrinkled immersion w :
V → V such that
• the regularization R(dw) : T V → T V is homotopic to the identity
through fiberwise isomorphisms;
• there exists a wrinkled immersion w
: V → V (the wrinkled lift of
w) such that the following diagram commutes:
?V

w

π
w
V / V.
The statement also holds in its fibered and relative versions. We note,
however, that in applications it will be more convenient for us to directly
use Lemma 13.3.2 rather than its more global version Lemma 13.3.3.
B. Proof of Theorem 13.2.1 in the case n = q. Suppose first that

k = 0. Using Theorem 5.2.2 with X = I n × W (and viewing the maps f0
and F0 , depending on the context, also as sections of X and X (1) ) we find
a multifolded cube π : I n → I n over I n such that the lift
F0 = F0 ◦ π : I n → I n → X (1)
with bs F0 = f0 ◦ π : I n → I n × W can be C 0 -approximated by J 1 for a
f1
multisection f1 : I n → I n × W with f1 = f0 ◦ π on Op ∂ I n . If the holonomic
multisection J 1 is sufficiently close to F0 , then the map f1 := p ◦ f1 : I n →
f1
In × W → W is an immersion. Let w : I n → I n be a wrinkled right inverse
to the projection π provided by Lemma 13.3.2. Then the composition

f1 : I n
w / I n f1 /W
is a wrinkled immersion and R(df1 ) is homotopic to F0 via a family of

fiberwise isomorphisms. Indeed, we have R(df1 ) = df1 ◦ R(dw)
∼ F0 ◦
= F0 ◦ dπ ◦ R(dw)
R(dw) ∼ F0 . The case k > 0 can be treated similarly,
using Theorem 5.2.3 instead of Theorem 5.2.2.
13.4. Chopping wrinkles

In this section we show that the wrinkles of an equidimensional wrinkled
immersion f : V → W can be made arbitrarily small (and their saucers
embedded) at the expense of increasing their number.
13.4. Chopping wrinkles 145
Assume that W is endowed with a Riemannian metric. We say that f has

σ-fine wrinkles if the saucer of each wrinkle has diameter < σ, and the
angle between the tangent spaces at any two points on the boundary of
the same saucer is < σ. Here we measure the angles between hyperplanes
at different nearby points via parallel transport along the uniquely defined
local geodesics.
13.4.1. (Wrinkled immersion with fine wrinkles) Suppose that dim V =
dim W . Then for any σ > 0 any wrinkled immersion f : V → W is homo-
topic through a fibered wrinkled immersions to a wrinkled immersion with
σ-fine wrinkles. Moreover, the homotopy can be chosen σ-small and sup-
ported in arbitrarily small neighborhoods of the wrinkle cores and covered by
a homotopy of the regularized differentials.
Proposition 13.4.1 is a corollary of the following more precise claim about
the standard wrinkle.
13.4.2. (Chopping a standard equidimensional wrinkle) Let
w = wn,n,0 : Uε → Rn−1 × R, w(y, z) = (y, z 3 + 3(|y|2 − 1)z),
be the standard equidimensional wrinkle. Then for any σ > 0 there exists a
wrinkled map w : Uε → Op w(Uε ) fibered over Rn−1 which coincides with w
on Op ∂Uε and such that its wrinkles are σ-fine, with saucers which project
onto round balls in {|y| ≤ 1} ⊂ Rn−1 .
Proposition 13.4.2 also implies the relative and fibered analogues of Propo-
sition 13.4.1.
While the proof of Proposition 13.4.2 is given below, we suggest that before
reading the proof the reader try to prove it as an exercise, using Figure 13.4
(in the case n = 2) as a hint. See also Figure 13.3 where the result of
chopping is illustrated for n = 1. Cutting off small (σ-fine) wrinkles, as in
Figure 13.4, we can successively push the original wrinkle into a position
where it can be removed.
Figure 13.3. Chopping of a one-dimensional wrinkle.
The proof uses two auxilliary lemmas. Denote

Δ := {|y| ≤ (1 + ε/2)2 } ⊂ Rn−1 , Δ− := {|y| ≤ (1 + ε/4)2 } ⊂ Δ.
Figure 13.4. Chopping: cutting off a small (σ-fine) wrinkle.
13.4.3. Let f, g : Δ → R be two C ∞ -functions such that g > f on

Δ \ Int Δ− . Then for any δ, ρ > 0 with δ < 4ε there exist an integer N ,
(possibly intersecting) closed discs D1 , . . . DN ⊂ Int Δ of radius δ, and C ∞ -
functions f0 := f, f1 , . . . , fN < g on Δ, such that for each j = 1, . . . , N
(a) fj > fj−1 on Δ \ IntDj ;
(b) fj − fj−1 C 2 (Δ) < ρ;
(c) fj C 2 (Δ) ≤ 2f C 2 (Δ) .
Proof. Choose a collection of balls D1 , . . . , DN of radius δ < 4ε centered at

N
points of IntD− such that IntDj ⊃ Δ− . Let θj : Δ → [0, 1], j = 1, . . . , N ,
1 !N "
(
N
∞
be C -functions supported in IntDj such that θj (Δ− = 1 and θj ≤
1 1
1 everywhere in Δ. Denote
m− := max(0, max(f (y) − g(y)), m+ := min (g(y) − f (y)).
y∈Δ y∈Δ\Δ−
Define inductively functions

m+
f0 := f fj := fj−1 + − (m+ + m− )θj , j = 1, . . . , N.
2N
Note that fj > fj−1 on Δ\IntDj for all j = 1, . . . , N , and fN < g everywhere
on Δ. Hence, condition (a) of the lemma is satisfied. Denote r := f −
gC 2 (Δ) . Then for all j = 1, . . . , N we have fj − fj−1 C 2 (Δ) ≤ Cr, and thus
fj C 2 (Δ) ≤ f C 2 (Δ) + N Cr, where the constant C depends only on the
partition functions θ1 , . . . , θN . But the constant r can be made arbitrarily
small by choosing a large K and applying the above argument successively
to the pairs f + j−1 j
K (g − f ), f + K (g − f ), j = 1, . . . , K, instead of f, g. This
allows us to fulfill conditions (b) and (c) at the expense of increasing the
total number of functions fj .
13.4.4. Given a disc D ⊂ IntΔ of radius δ > 0 and a concentric disc
D ⊂ D of radius δ/2, consider functions f, g : Δ → R such that g > f on
Δ \ IntD . Denote C := f C 2 (Δ) and ρ := f − gC 0 (Δ) . Then there exists
a wrinkled immersion H : Δ × [−1, 1] → Δ × R fibered over Δ with exactly
13.4. Chopping wrinkles 147
z=g(y)
[ 1,1] z=f(y)
D’
y
D D
Δ Δ
Figure 13.5. The wrinkled immersion H : Δ × [−1, 1] → Δ × R.
one σ = σ(δ, ρ, C)-fine wrinkle whose saucer projects onto the disc D, and
such that (see Figure 13.5)
(a) H(y, −1) = f (y) and H(y, 1) = g(y) for y ∈ Δ;
(b) H(Δ × [−1, 1]) ⊂ {(y, z) ∈ Δ × R | f (y) − 2ρ ≤ z ≤ f (y) + 2ρ};
(c) σ(δ, ρ, C) ≤ K(δ + Cδ + ρ/δ) for a constant K not depending on
(δ, ρ, C).
Proof. Let y0 be the center of D. We construct the function H in the

form H(y, z) = (y, hy (z)) for a family of functions hy : [−1, 1] → R, y ∈ Δ,
satisfying
• hy (−1) = f (y), hy (1) = g(y) and hy (±1) > 0 for all y ∈ Δ;
• hy > 0 for y ∈ Δ \ D;
• for y ∈ IntD the function hy has exactly one nondegenerate local
maximum and one nondegenerate local minimum (i.e., it is a one-
dimensional wrinkle extended as an immersion);
• for (y, z) ∈ Op (∂D × 0) the function hy has the form
f (y) + g(y)
hy (z) = + z 3 + 3(|y − y0 |2 − δ 2 )z.
2
Then H is a fibered wrinkled immersion with exactly one wrinkle whose
saucer projects onto the disc D. Property (a) in the lemma holds by con-
struction, and property (b) can be easily arranged. For property (c), note
first that the wrinkle has diameter 2δ. The variation of the tangent spaces
on the boundary of the saucer has two contributions: one from the variation
of the gradient of f over the disc D, which is of order Cδ; and another one
from the variation of f − g over D, which is of order ρ/δ.
Proof of Proposition 13.4.2. Denote Q := (Δ × R) ∩ {|y|2 + z 2 ≤ (1 +

4 ) }. We have Q ⊂ Uε = {|y| + z < (1 + ε) } ⊂ R
3ε 2 2 2 2 n−1 × R. Choose
a diffeomorphism ϕ : Δ × [0, 1] → Q fibered over Δ. The fibered map

w ◦ ϕ : Δ × [0, 1] → Δ × R has the form w ◦ ϕ(y, z) = (y, h(y, z)), y ∈ Δ, z ∈
[0, 1]. Denote f (y) = h(y, 0), g(y) = h(y, 1), y ∈ Δ. Apply Lemma 13.4.3
to the functions f, g and constants δ1 , δ2 which will be fixed later in the
proof to construct the functions f0 = f, f1 , . . . , fN < g. Denote tj := N j+1 .
For each j = 1, . . . , N we apply Lemma 13.4.4 to the functions fj−1 , fj to
construct a wrinkled immersion Hj : Δ×[tj−1 , tj ] → with exactly one wrinkle
such that Hj (y, tj−1 ) = fj−1 (y) and Hj (y, tj ) = fj (y). Using the fact that
fN < g, we can also the construct a genuine fibered immersion HN +1 :
Δ × [tN , 1] → Δ × R such that HN +1 (y, tN ) = fN (y) and Hj (y, 1) = g(y).
After adjusting the differentials dHj , j = 1, . . . , N , along Δ × tj the maps
H1 , . . . , HN +1 can be made to fit together into a fibered wrinkled immersion
H : Δ × [0, 1] → Δ × R with N wrinkles which project to round balls. By
taking appropriate parameters in Lemmas 13.4.3 and 13.4.4 we can make all
wrinkles σ-fine. Finally we replace back the map H to Q, i.e., we take the
map w := H ◦ϕ−1 : Q → Δ×R and extend it as a fibered immersion to Uε \Q
which coincides with w on Op ∂Uε , and such that w(U ε ) ⊂ Op w(Uε ).
13.5. Corrugation
The composition g ◦ wn,n of the equidimensional standard wrinkle wn,n :
Rn → Rn with a submersion g : Rn → Rq is again a wrinkled submersion
if n = q, but for n > q this need no longer be the case (e.g., if g is the
projection onto the last q coordinates). The corrugation process in this
section allows us to approximate g ◦ wn,n again by a wrinkled submersion.
Let n > q, and (x, y, z) be the coordinates corresponding to the splitting
Rn = Rn−q × Rq−1 × R = Rn−1 × R.
Let β : Rn−1 × R → Rn−1 × R,
y, z) = (x, y, β(x, y, z)),

β(x, β(x, y, z) = z 3 + 3(|x|2 + |y|2 − 1)z
be the standard wrinkle wn,n (up to renaming the coordinate y ∈ Rn−1 to

(x, y) ∈ Rn−q × Rq−1 ), D = D n = {|x|2 + |y|2 + z 2 ≤ 1} the core of the

wrinkle, Δ = β(D) and
the saucer of β,
Uε := {|x|2 + |y|2 + z 2 < (1 + ε)2 } ⊃ D.
Let g : Op Δ → Rq be a submersion which restricts as an immersion to fibers

of the projection π : Rn−q × Rq → Rn−q .
13.5. Corrugation 149
13.5.1. (Corrugating a wrinkle) In the setup above, for any τ > 0 there
exists a function γ : Rn → R such that
(a) γ = β outside Uε ;
(b) |γ − β| < τ ;
: Uε → Rq , where γ
(c) the map g ◦ γ (x, y, z) = (x, y, γ(x, y, z)), is a
wrinkled submersion;
(d) the regularized differential R(d(g◦
γ )) is homotopic as epimorphisms

to dg ◦ R(dβ) rel. Op (∂Uε ).
The corrugation process, further detailed in the proof, is shown schemati-

cally in Figure 13.6 and allegorically in Figure 13.7.
Figure 13.6. Corrugation: n = 2, q = 1.
Figure 13.7. Corrugation: n = 3, q = 2 (Kapunda museum, South Australia).

Proof. Choose a cutoff function ρ : Rn → [0, 1] which is equal to 0 outside

2
Uε , and equal to 1 on Uε/2 = {|x|2 + |y|2 + z 2 < 1 + 2ε }. We will construct
below a 2π-periodic function b : R → R such that the required function γ
will have the form

n−q
(9) γ(x, y, z) = β(x, y, z) + δρ(x, y, z) b(N xj )
j=1
for a sufficiently large integer N and a small positive δ > 0. Note that
property (a) of the proposition holds by definition of ρ.
For a large positive real number M (which will be fixed later) let us choose
a smooth even function aM : R → [0, ∞) such that aM (u) = M 2 u2 for
|u| ≤ 1/3M , aM (u) = |u| for |u| ≥ 1/2, and aM (u) > 1/2 for u ≥ 1/3M .
Define the function b = bM : R → R as equal to aM (2u) − π on [− π2 , π2 ]
and to π − aM (2u − 2π) on [ π2 , 3π
2 ], and extend it 2π-periodically to R, see
Figure 13.8.
−π/2 π/2
−3π/2 3π/2
1/9
1/2 1/2 −π
Figure 13.8. Function bM .

( (
( ∂ρ (
Note that σ := min ∂β |
∂z Uε \Uε/2 > 0. Denote R := max ( |
∂z Uε \Uε/2 ( and choose
1
σ
δ < (n−q)π min τ, R . This ensures property (b) of the proposition because
( (
(n−q
(
( (
|γ −σ| < δ ( b(N xj )( δ ≤ (n−q)π < τ . For later use, note that on Uε \Uε/2
( j=1 (
( (
( ∂ρ (
we have ∂γ
∂z > σ − π(n − q) ( ∂z ( > σ − δπ(n − q)R > 0.
Choose ) ∂β )
1 ) )
N > max) ) .
δ i ∂xi C 0 [Uε/2 ]
Recall that the core D of the wrinkle β projects to the (n − 1)-dimensional
unit disc B ⊂ Rn−1 , so its image B = π(B) ⊂ Rn−q under the projection
(x, y) → x is the (n − q)-dimensional unit disc. Consider the lattice L :=
= ∅ (because π 2 is irrational).
NZ ⊂ Rn−q , and observe that L ∩ ∂ B
π n−q
Denote P := L ∩ B = L ∩ IntB. For x ∈ Rn−q and r > 0 we denote by

1 ⊂ R
Ix (r) n−q the cube centered at x of side length r. Note that the cubes
Ip 3N for p ∈ P are pairwise disjoint, and there exists a constant M0 ≥ 1

To complete the
(which depends on N ) such that p∈P Ip 3M10 N ⊂ B.
proof we need two lemmas.
13.5.2. For δ, N, M0 as above and each M ≥ M0 , the composition g ◦ γ
:
Uε → Rq is a submersion outside the set
1
U = π −1 Ip .
3M N
p∈P
Proof of Lemma 13.5.2. Consider first u ∈ R with |u| ≤ π/2N . Then by

definition b(N u) = aM (2N u) − π and the lower bound on the derivative of
aM yields
1
|b (N u)| = 2|aM (2N u)| > 2 · =1
2
provided that |2N u| ≥ 3M 1
, or equivalently, |u| ≥ 6M1 N . By periodicity of b
this implies |b (N u)| > 1 for u outside the intervals
*π 1 π 1 +
I( , M, N ) = − , + , ∈ Z.
N 6M N N 6M N
Consider now a point (x, y, z) ∈ Uε \ U . Then at least one component xi
of x lies outside the intervals I( , M, N ), ∈ Z. We have noted above that
∂γ
∂z > 0 on Uε \ Uε/2 , so γ is an equidimensional immersion outside Uε/2 . For

(x, y, z) ∈ Uε/2 \ U , the condition ρ|Uε/2 = 1 and the preceding discussion
imply
( ∂γ ( ( ∂β ( ( ∂β (
( ( ( ( ( (
( (x, y, z)( = ( (x, y, z) + δN b (N xi )( > δN − ( (x, y, z)( > 0
∂xi ∂xi ∂xi
by the choice of N . This shows that γ is an equidimensional immersion on
Uε \ U , and therefore g ◦ γ is a submersion on Uε \ U .
13.5.3. For δ, N, M0 as above and M ≥ M 0 sufficiently
large, the restriction
g◦γ|U has a single wrinkle on each π −1 Ip 3M1 N , p ∈ P .
Proof of Lemma 13.5.3. Let us fix p ∈ P and abbreviate

1 1
IM = Ip ⊂ IM 0 = Ip .
3M N 3M0 N
After choosing M0 sufficiently large, we may assume that
π −1 (IM0 ) ∩ Uε/2 ⊂ IM0 × U ⊂ π −1 (IM0 ) ∩ Uε
for some open subset U ⊂ Rq . Consider a point
(p + x, y, z) ∈ π −1 (IM0 ) ∩ Uε/2 ,
πn−q
so that |xi | ≤ 6M10 N for i = 1, . . . , n − q. Recall that p = π N ,..., N
1
∈

P = L ∩ B with i ∈ Z. By definition of b we have
π
i
b N + xi = εi (4N 2 M 2 x2i − π),
N
where εi = 1 if i is even and εi = −1 if i is odd. It follows that
γ(p + x, y, z) = β(p + x, y, z) + 4δN 2 M 2 Qp (x) + Cp ,
n−q
where Qp (x) = n−qi=1 εi xi and Cp = −πδ
2
i=1 εi . For M ≥ M0 we consider
the dilation
∼ M0
ΔM : IM0 × U −→ IM × U ,
=
(p + x, y, z) → p + x, y, z .
M
It suffices to show that for M sufficiently large the map
◦ ΔM : IM0 × U → Rq
hM := g ◦ γ
has a single wrinkle. We factor the map hM as
hM = (g ◦ ΔM ) ◦ γ
M , M := Δ−1
γ M ◦γ
◦ ΔM .
M (p + x, y, z) = (p + x, y, γM (p + x, y, x)), where
Note that γ
M0
γM (p + x, y, z) = β p + x, y, z + 4δN 2 M02 Qp (x) + Cp
M
converges as M → ∞ in the C ∞ -topology to
γ∞ (p + x, y, z) = β(p, y, z) + 4δN 2 M02 Qp (x) + Cp .
On the other hand,
M0
g ◦ ΔM (p + x, y, z) = g p + x, y, z
M
∞
converges as M → ∞ in the C -topology to g(p, y, z). In other words,
g ◦ ΔM converges to the composition gp ◦ π ⊥ of the projection
π ⊥ : IM 0 × U → U , (p + x, y, z) → (y, z)
with the map
gp : U → Rq , (y, z) → g(p, y, z).
Altogether, the map hM converges as M → ∞ in the C ∞ -topology to the
composition
∞
γ π⊥ gp
h∞ : IM0 × U −→ IM0 × U −→ U −→ Rq .
The composition of the first two maps is
π⊥ ◦ γ
∞ : IM0 × U → Rq−1 × R, (p + x, y, z) → (y, γ∞ (p + x, y, z))
with
γ∞ (p + x, y, z) = z 3 + 3(|p|2 + |y|2 − 1)z + 4δN 2 M02 Qp (x) + Cp .
This shows that π ⊥ ◦ γ ∞ has a single (n, q, s)-wrinkle over the ball
{x = p, |p| + |y| + z ≤ 1} = D ∩ {x = p}, where s is the index of
2 2 2
the quadratic form Qp (which depends on p).

Now we use hypothesis on the submersion g, which implies that the map
gp : U → Rq is a local diffeomorphism onto an open subset. Therefore, the
composition h∞ = gp ◦ (π ⊥ ◦ γ ∞ ) has again a single wrinkle. Since wrinkles
are C ∞ -stable, it follows that hM has a single wrinkle for sufficiently large
has a single wrinkle on π −1 (IM ). This proves the lemma.
M , so g ◦ γ
The two lemmas together yield property (c) of the proposition. We leave
the verification of property (d) to the reader. This concludes the proof of
Proposition 13.5.1.
Proof of Theorem 13.2.1 for n > q. We prove in this section Theorem

13.2.1 for n > q and k = 0. For k > 0 the proof is similar and we leave it to
the reader to make the necessary adjustments.
For n > q, we can repeat the construction from Section 13.3.B, but now in
the composition

f1 : I n
w / I n f1 /W
the map f1 is a submersion (and not an immersion), and the map w is an

equidimensional wrinkled immersion. By definition of a wrinkled immer-
sion there are embeddings ϕi : Uε → I n , i = 1, . . . , N , such that w
is an
immersion in the complement of the singular locus

N
Sw := ϕi (∂D n )
1
and w ◦ ϕi factors as gi ◦ wn,n , where gi is an immersion Rn → I n . Hence,

the map f1 ◦ w is a submersion away from Sw and its singularities coincide
with the singularities of the restriction of the submersion f1 to Sw . Let us
analyze the map f1 ◦ ϕi = f1 ◦ (w ◦ ϕi ) = f1 ◦ (gi ◦ wn,n ) = hi ◦ wn,n , where
hi := f1 ◦ gi is a submersion R → W .
n
According to Lemma 13.4.2, we can chop the standard wrinkle wn,n : Rn−1 ×
R → Rn−1 × R into σ-fine wrinkles wj fibered over Rn−1 whose cores project
to σ-small balls in Rn−1 . Any fibered σ-fine wrinkle w := wj has the form
βw : (x, z) → (x, βw (x, z)).
If σ is sufficiently small, then the restriction hw := hi |Op Sw of the submer-
sion hi to a neighborhood Op Sw ⊂ Rn boundary Sw of the saucer can be
considered as a submersion hw : Op Sw → Rq which is C 1 -close to L|Op Sw
for a surjective linear map L : Rn → Rq . We only need to consider the
case when the restriction hw |Sw has singularities, i.e., there is at least one
point s ∈ Sw such that Ker dhw ⊂ Ts Sw . In this case hw is a leafwise im-
mersion on the one-dimensional vertical foliation of Rn (given by the last
coordinate). Moreover, hw is a leafwise equidimensional immersion on the
q-dimensional foliation of Rn parallel to T × R, where T ⊂ Rn−1 is the or-
thogonal complement to Ker(L|Rn−1 ×0 ) ⊂ Rn−1 . We can assume, after an
orthogonal rotation in Rn−1 , that T × R = 0 × Rq−1 × R = 0 × Rq . Then
there exist diffeomorphisms ϕ, ψ : Rn−1 × R → Rn−1 × R fibered over Rn−1
such that ψ ◦ βw ◦ ϕ = wn,n . Thus we can apply Proposition 13.5.1 with
β = ψ ◦ βw ◦ ϕ and g = hw ◦ ψ −1 to approximate g ◦ β = hw ◦ βw ◦ ϕ by a
wrinkled submersion.
Chapter 14
Folded Solutions to
This chapter is an introduction to the theory of mappings and other struc-

tures with prescribed singularities. Using the wrinkling technique developed
in Chapters 12 and 13 we prove an h-principle for existence of mappings with
prescribed (nonempty!) singularities of only fold type. We also consider
various generalizations of this result for folded solutions of Diff V -invariant
differential relations for section of a natural bunle X → V .
14.1. Surgery of singularities

We introduce in this section the surgery technique for Σ1 -singularities of
maps V → W . We will restrict our discussion to the equidimensional case
dim V = dim W and (with one exception) to surgeries involving only folds
and cusps, and refer the reader to the paper [El72] for the general case. We
always assume below transversality of the corresponding jet sections to all
considered singularities.
A. Surgery of a pair (Σ1 (f ), Σ1,1 (f )). . Assume dim V = dim W = n and

consider a family ft : V → W , t ∈ R, of maps which are Σ2 -nonsingular.
Then the map F : V ×R → W ×R given by the formula F (x, t) = (ft (x), t) is
also Σ2 -nonsingular, and for a generic family ft we get nested submanifolds
V ⊃ Σ1 (F ) ⊃ Σ1,1 (F ) ⊃ · · · ⊃ Σ1n (F ), each of codimension 1 in the previous
one. Let t : V × R → R denote the projection.
155
156 14. Folded Solutions to Differential Relations
14.1.1. (Singularities of a pair)

(a) The restriction t|Σ1,0 (F ) has no critical points.
(b) If for k ≥ 2 a point p ∈ Σ1k ,0 (F ) is a critical point of t|Σ1k (F ) , then p is
also critical for t|Σ1k−1 (F ) ;
(c) Generically, if a point p ∈ Σ1k ,0 (F ) is critical for t|Σ1k (F ) , then it is not
critical for t|Σ1k−2 (F ) , and it is Morse for t|Σ1k (F ) and t|Σ1k−1 (F ) .
Proof. Let us abbreviate Σ1j = Σ1j (F ) and Σ1j ,0 = Σ1j ,0 (F ).

(a) For p ∈ Σ1 let κp be the one-dimensional kernel of the differential dp F .
Note that dt|κp = 0. If p ∈ Σ1,0 then κp is transverse to Tp Σ1 , hence
dt|Tp Σ1 = 0 would imply dt|Tp (V ×R) = 0, which is not the case.
(b) Similarly, at a point p ∈ Σ1k ,0 we have Tp Σ1k−1 = Tp Σ1k ⊕ κp , hence
dt|Tp Σ1k = 0 implies dt|Tp Σ1k−1 = 0, i.e., p is also critical for t|Σ1k−1 .
(c) The claim that generically p is a regular point for t|Σ1k−2 follows from the
Thom Transversality Theorem 2.3.2. Indeed, let us write the map F near a
point p ∈ Σ1k (F ) in Morin normal form, using coordinates (τ1 , . . . , τn , z) in
the source and coordinates (y1 , . . . , yn , ζ) in the target:
F (τ1 , . . . , τn , z) = (τ1 , . . . , τn , z k+1 + τk−1 z k−1 + · · · + τ1 z) = (y1 , . . . , yn , ζ).
We have (see (6) in Section 11.4) Σ1k = {τ1 = · · · = τk−1 = z = 0} and
! "
1k ∂ ∂
Tp Σ = Span ,..., ,
∂τk ∂τn
! "
1k−1 ∂ ∂ ∂
Tp Σ = Span ,..., , ,
∂τk ∂τn ∂z
! "
1k−2 ∂ ∂ ∂
Tp Σ = Span ,..., , ,
∂τk−1 ∂τn ∂z
! "
1k 1k−1 ∂ ∂
dF (Tp Σ ) =dF (Tp Σ ) = Span ,..., ,
∂yk ∂yn
! "
1k−2 ∂ ∂
dF (Tp Σ ) = Span ,..., .
∂yk−1 ∂yn
The hypothesis p ∈ / Σ1k+1 implies that there is a neighborhood U p in Σ1k
whose image U := F (U ) is an (n+1−k)-dimensional submanifold of W ×R.
Let us slightly change our viewpoint. We will consider the parameter t as a
function on the target manifold and denote t := t ◦ F . Since F |U : U → U
is a diffeomorphism and t ◦ F = t, p is a critical point of t|Σ1k if and only
if is is a critical point of
t|Σ1k , which in turn
∂isequivalent 1to F (p) being a
critical point of t|U . First we note that dF ∂z = 0 on Σ , and therefore,
∂
∂z
t
≡ 0 on Σ1 . We also observe that p ∈ Σ1 is a critical point of t | Σ1k
14.1. Surgery of singularities 157
(and then automatically of t|Σ1k−1 ) if and only only if ∂y∂t

j
(F (p)) = 0 for
j = k, . . . , n, and it is a critical point of t|Σ1k−2 if and only if ∂y
∂t
j
(F (p)) = 0
for j = k − 1, . . . , n.
Consider the jet space J 1 (Op U , R)|U and its subspace of codimension n −
k + 2 given by the linear equations Yj = 0, j = k − 1, . . . , n, where Yj
are the coordinates on J 1 (Op U , R) corresponding to the derivatives ∂y∂ j .
As dim U = n − k + 1, the Thom Transversality Theorem 2.3.2 ensures
that for a generic perturbation of the function t these conditions are never
satisfied. As F (p) is a critical point of t|U , any perturbation of this function
keeping t(F (p)) fixed can be realized by taking a composition t ◦ ϕ with
a diffeomorphism ϕ : W × R → W → R fibered over W and such that
ϕ(F (p)) = F (p). As p is a regular point of the function t|V ×R , there is a
C ∞ -small diffeomorphism ψ : V × R → V × R fibered over V with ψ(p) = p
and a family ft : V → W which is C ∞ -close to ft such that
ϕ (F (ψ(x, t))) = (ft (x), t) for all (x, t) ∈ V × R.
Explicitly, writing ϕ(x, t) = (x, g(x, t)) and ψ(x, t) = (x, hx (t)), the last

displayed equation reads fs (x), g(fs (x), s) = (ft (x), t) with s = hx (t).
This is solved by taking for hx the inverse of the function s → g(fs (x), s),
which is invertible for ϕ sufficiently C 1 -close to the identity.
Hence, the map F(x, t) := (ft (x, t), t) satisfies F = ϕ ◦ F ◦ ψ. It follows that
critical points (and their degeneracy types) of t ◦ ϕ ◦ F |Σ1k (F ) correspond to
critical points of t ◦ F| 1 . Hence, we can realize any generic perturbation
Σ k (F )
of t|Op Σ1k (F ) , and in particular we can make p regular for t|Σ1k−2 (F ) . A

similar argument allows us to ensure that generically t|Σ1k ,0 has only Morse
critical points.
It remains to verify that if p is a Morse critical point for t|Σ1k and regular

for t|Σ1k−2 , then it is automatically a Morse critical point for t|Σ1k−1 . To see
k+1 2 ∂
this, recall that on Σ 1 k−1 we have τk−1 = − 2 z . We also have ∂z t
≡0
1 ∂
2 t
everywhere on Σ , so in particular ∂τj ∂z = 0 for all j = k, . . . , n.
By the above description of the tangent spaces, the condition that p

is not a critical point for t ◦ F |Σ1k−2 means that ∂τ∂k−1 t
(p) = 0. Hence,

2
∂2
∂
∂z 2
t| 1
Σ k−1 (p)
= 0. Together with the condition that t
∂τi ∂τj (p) is
k≤i,j≤n
2
nondegenerate (because p is Morse for t|Σ1k ) and the condition ∂τ∂j ∂z
t
(p) =
0, j = k, . . . , n, we conclude that p is a nondegenerate critical point of

t|Σ1k−1 .
V V
v
C C C
C C C
F F
Figure 14.1. Simultaneous surgery.
Lemma 14.1.1 can be summarized as follows: Generically, the singularities

Σ1j (ft ) always experience surgeries for pairs of consecutive values of j as t
increases.
Remark. The surgeries of Σ1k (ft ) and Σ1k+1 (ft ) could be of the same
index s, or the indices could be s and s + 1, respectively. In [El72] surgeries
of the first kind were called direct, and surgeries of the second kind inverse
(as they are inverse to each other).
Example. Consider the family of maps ft : Rn = Rn−1 × R → Rn ,

(10) ft (y, z) = y, z 3 + (Qn−1,s (y) − t)z , (y, z) ∈ Rn−1 × R, t ∈ R.
Then the pair (Σ1 (ft ), Σ1,1 (ft )) experiences simultaneous surgeries of index
s as t increases and changes its sign.
Let us discuss this double surgery in more detail. Let V be an n-dimensional

manifold and F ⊃ C two submanifolds of V of codimension 1 and 2, respec-
tively. Let (M, ∂M ) ⊂ (V, C) be a submanifold of dimension k ≤ n − 1 such
that Int M ⊂ V \ F and M is transverse to F along ∂M . Suppose that we
are given a normal vector field v to M in V such that v|∂M is tangent to F
and transverse to C, see Figure 14.1.
Then a simultaneous (or double) surgery of (F, C) along M with framing
v is defined as follows. Consider the split normal bundle ν = ν ⊕ Span(v)
to M in V . Let p : N → M and p : N → M be the normal disc bundles
associated with ν and ν. We view N and N as embedded in V in such a

way that ∂ N ⊂ ∂N , p (∂M ) ⊂ C, and p−1 (∂M ) ⊂ F . Then the surgery
−1
and p−1 (∂M ) with ∂N and smoothes
operation replaces p−1 (∂M ) with ∂ N
p−1 (∂M )) and ∂(p−1 (∂M )).
corners along ∂(
14.1.2. (Simultaneous surgery) Consider a generic map f : V → W ,
dim V = dim W = n. Let (M, ∂M ) ⊂ (V, Σ1,1,0 (f )) be a compact k-
dimensional submanifold with boundary. Suppose that the characteristic vec-
tor field for the cusp locus Σ1,1 (f ) is tangent to M and inwardly transverse
to ∂M along ∂M . Let v be a vector field transverse to M such that v|∂M

is tangent to Σ1 (f ) and transverse to Σ1,1 (f ). Then there exists a C 0 -small
homotopy fs , s ∈ [0, 1], of f = f0 which is supported in Op M such that
(Σ1 (f1 ), Σ1,1 (f1 )) is the result of simultaneous surgery of (Σ1 (f ), Σ1,1 (f ))
along M with framing v.
# 1 1$
Proof. Recall the family of functions
# 1 1$ θs : [−1, 1] → [−1, 1], s ∈ −4, 4 ,
and the function δ : [−1, 1] → − 4 , 4 introduced around (8) in Section
11.6.
Take a tubular neighborhood P of ∂M in Σ1,1 (f ) and apply Proposition
11.6.7 to find a split tubular neighborhood
Ω = P × [−1, 1] × [−1, 1] = N × [−1, 1], N = P × [−1, 1],
of P in V such that f |Ω can be written as F ◦ h, where F : Ω → W is an
immersion and h : P × [−1, 1] × [−1, 1] → P × [−1, 1] × [−1, 1] is defined by

h(p, t, z) = p, t, θδ(t) (z) , (p, t) ∈ P × [−1, 1], z ∈ [−1, 1].
∂
Recall that ∂t is the characteristic vector field of the cusp locus along P =
P × 0 × 0, and the vector field ∂z∂
is tangent to Σ1 (f ) along P . Adjusting, if
necessary, the manifold M by a C -small isotopy in Op ∂M fixed on ∂M , we
1
can assume that M ∩ Ω = {x ∈ ∂M, t ∈ [0, 1], z = 0}, so that M ∩ Ω serves

as an interior collar C+ := ∂M × [0, 1] of ∂M in M . Let us also attach to
M an exterior collar
C− := ∂M × [−1, 0] = {x ∈ ∂M, t ∈ [−1, 0], z = 0}
and denote
, := M ∪ C− ,
M C := C− ∪ C+ = ∂M × [−1, 1].
∂
We may assume that the vector field v in the proposition agrees with ∂z on Ω.
Hence, we can use v to extend the split tubular neighborhood Ω = N ×[−1, 1]
of ∂M to a split tubular neighborhood Ω = N × [−1, 1] of M , in V , where
N is the total space of an (n − k − 1)-disc bundle p : N → M , such that
p−1 (C) = N . We denote the coordinate corresponding to the second factor
in Ω = N × [−1, 1] still by z.
Recall that for t close to 1 we have δ(t) = 14 . Slightly abusing notation we
will view δ as a function on N = P × [−1, 1] via δ(p, t) = δ(t). Let us extend
δ to N by setting it equal to 14 on N \ N , and extend h from Ω to a map
h : Ω → Ω by setting h(u, z) = (u, θδ(u) (z)) for (u, z) ∈ N × [−1, 1]. As
θ 1 : [−1, 1] → [−1, 1] is a diffeomorphism, so is h|Ω\Ω
. Thus, by extending
4
the immersion F : Ω → W to Ω as F |Ω\Ω
= f ◦ h−1 we get a factorization
f |Ω = F ◦ h.
Choose a sufficiently small σ > 0 and define a C ∞ -function γ : N → R

by the following procedure. On C− the function γ is a smoothing of the
function min(t, −σ). We extend γ to M as the constant −σ. Finally, we
extend γ to N as γ ◦ p + dist2M ,
, where distM
is the distance function to M .
Note that the function γ is negative on M, and positive outside its tubular
neighborhood. Define a homotopy hs : Ω → Ω by the formula

hs (u, z) = u, sθγ(u) (z) + (1 − s)θδ(u) (z) , u ∈ N , z ∈ [−1, 1].
Then the homotopy fs : V → W which is equal to F ◦ hs on Ω and to f

elsewhere connects f0 = f with a map f1 such that f1 |V \Ω = f and the pair
(Σ1 (f1 ), Σ1,1 (f1 )) is the result of simultaneous surgery of (Σ1 (f ), Σ1,1 (f ))
along M with framing v.
Examples.
1. Index 1 surgery.
Σ 1(f) Σ 1(f)
Σ 1(f)
v
Σ 11(f) Σ 11(f)
w w
Σ 1(f)
Figure 14.2. Index 1 surgery.
Let f : V → W , dim V = dim W = n, be a smooth map and a, b ∈ Σ1,1,0 (f )

be two distinct points. Suppose the characteristic vectors va and vb at a and b
point into the same connected component C of V \Σ1 (f ). Let γ : [0, 1] → V ,
γ(0) = a, γ(1) = b, γ((0, 1)) ⊂ C be any embedded path transverse to Σ1 (f )
at the endpoints. Then one can perform a simultaneous index one surgery
of (Σ1 (f ), Σ1,1 (f )) in V , see Figure 14.2. In particular, any two wrinkles in
the same connected component can be connect-summed into one wrinkle, see
Figure 14.3.
2. Rounding a wrinkle. Let w : Rn−1 × R → Rn−1 × R be the standard
wrinkle. Recall that Σ1 (w) is a sphere which bounds a ball B, the core of
the wrinkle, and Σ1,1 (w) ⊂ Σ1 (w) is the equator of ∂B. The characteristic
vector field of the cusp locus points out of B. Note that for an arbitrary
small neighborhood U ⊃ B the cusp locus Σ1,1 (w) bounds an (n − 1)-disc
Figure 14.3. Connected sum of two wrinkles.
D such that IntD ⊂ U \ B and the characteristic vector field of the cusp is
inwardly tangent to D along ∂D.
By performing a simultaneous surgery of (Σ1 (w), Σ1,1 (w)) along D we elim-
inate the cusp locus and create two concentric spheres of fold points instead
of the wrinkle w. As the disc D can be chosen in an arbitrary small neigh-
borhood of B, this operation allows us to replace each wrinkle of a wrinkled
map by two concentric spherical folds, see Figure 14.4.
Figure 14.4. Rounding a wrinkle.
B. Creation and elimination of cusps of opposite sign. We will need

another surgery, associated with the pair (Σ1,1 (f ), Σ1,1,1 (f )). Recall from
Theorem 11.6.3 that Morin’s normal form for a map R3 → R3 with a generic
Σ1,1,1 -singularity at the origin is given by
(11) (t1 , t2 , z) → (t1 , t2 , z 4 + t1 z 2 + t2 z).
Viewing it as a family of maps hε (t2 , z) = (t2 , z 4 −εz 2 +t2 z) which depend on

the parameter ε = −t1 , we observe that Σ1 (hε ) = {t2 = −4z 3 + 2εz} ⊂ R2
and Σ1,1 (hε ) = {z = ± ε/6} ⊂ Σ1 (hε ). Thus, when ε is negative there
Figure 14.5. Birth of cusps.
are no cusps, and as ε increases and changes from negative to positive two
cusps are born, see Figure 14.5. For ε > 0 the critical values of the function
3/2
−4z 3 + 2εz are equal to ±δ, where δ := 4ε√ .
3 6
Hence, for a sufficiently small ε > 0 there exists a family of smooth functions
ξt : R → R, t ∈ [−1, 1], such that
• ξt (z) = z 2 for (t, z) ∈
/ {|t| ≤ 12 , |z| ≤ 12 };
• ξt (z) = z 4 − εz 2 + tz for |t| ≤ 16 , |z|
1
≤ 14 ;
• ξt has exactly one critical
# $point for t > δ, and no critical points in
the complement of − 14 , 14 for any t ∈ [−1, 1].
Let f : V → W , dim V = dim W = n, be a generic map with a nonempty
fold locus Σ1,0 (f ). Consider first the case n = 2 and a fold point p ∈ Σ1,0 (f ).
According to Proposition 11.6.7(a) there are local coordinates (t, z) ∈ U =
[−1, 1] × [−1, 1] near p and (t, z) near f (p) in which the map f takes the
form (t, z) → (t, z 2 ). Let us replace f |U by the map (t, z) → (t, ξt (z)) to get
a new map f : V → W which coincides with f outside U . The smooth curve
Σ1 (f|U ) is isotopic to Σ1 (f |U ) relative its boundary, and Σ1,1 (f) consists of
two cusp points with opposite characteristic coorientation of Σ1 (f).
In the general case n ≥ 2 we choose any closed coorientable hypersurface
C ⊂ Σ1,0 (f ). Let N = C × [−1, 1] be a split tubular neighborhood of
C in Σ1,0 (f ). According to Proposition 11.6.7 there exists a split tubular
neighborhood U = N × [−1, 1] of N in V such that f |U can be presented
as a composition f |U = g ◦ Q, where Q(x, t, z) = (x, t, z 2 ), (x, t, z) ∈ C ×
[−1, 1] × [−1, 1], and g : U → W is an immersion. Replacing Q by the
t, z) := (x, t, ξt (z)) and f |U by the composition g ◦ Q
map Q(x, yields a map
f : V → W which coincides with f outside U . The singular locus Σ1 (f|U ) is
isotopic to Σ1 (f |U ) relative its boundary, and Σ1,1 (f) consists of two parallel
copies of C with opposite characteristic coorientation of Σ1 (f). We will refer
to this operation as a cusp-creation surgery.
The inverse surgery, which eliminates a pair of cusps whose characteristic
coorientations are opposite can be performed via the following trick. Con-
sider again first the two-dimensional case. Given two cusp points a1 , a2 on
the same component of Σ1 (f ) we use cusp-creation surgery to create an ad-
ditional pair of cusp points b1 , b2 such that the four cusps are ordered as
a1 , a2 , b1 , b2 and the characteristic coorientation of ai coincides with that of
bi for i = 1, 2. Next we perform index one surgeries along a path γ1 con-
necting a1 and b1 , and along a path γ2 connecting a2 and b2 . As a result
all cusps a1 , b1 , a2 , b2 got eliminated while Σ1 (f ) remains unchanged up to
isotopy, see Figure 14.6 for the picture in the source. We leave it to the
reader to generalize this procedure to the cusp-elimination surgery for any
n ≥ 2, thus eliminating two parallel components of the cusp locus in Σ1 (f )
with opposite characteristic coorientations. See Figure 14.7 for a hint.
a2 a2 b2
a1 a1 b1
Figure 14.6. Elimination of cusps of opposite sign: the source picture.
Let S 2 ⊂ R3 be the unit sphere and p : S 2 → R2 be its orthogonal projection

to the equatorial plane. Then Σ1 (p) is the equator S 1 ⊂ S 2 . Use cusp-
creation surgery to create two cusp points a1 , b1 , and then cusp-elimination
surgery to kill them. Verify that the resulting folded map h : S 2 → R2 is
glued out of the two immersions f, g : D 2 → R2 discussed in Example A3 in
the Intrigue.
14.1.3. (Elimination of wrinkles) Let V, W be manifolds of dimension
n ≥ 2 and S ⊂ V a codimension 1 submanifold. Suppose that V is connected
and S is nonempty. Let f : V → W be a Σ2 -nonsingular map such that
Σ1 (f ) consists of the (nonempty!) fold S and a finite number of wrinkles.
Then f is homotopic to a map f : V → W whose only singularity is the fold
along S.
Proof. Using cusp creation surgery we first create finitely many pairs of
concentric spheres Ci+ , Ci− ⊂ S of cusps with opposite characteristic coori-
entation such that each wrinkle can be connected so some Ci± by an embed-
ded path in V \ S at whose ends the characteristic coorienations point into
Figure 14.7. Star wars: elimination of wrinkles.
the path. We connect-sum each wrinkle to S along this path and denote the
resulting map by f, see Figure 14.7. Note that up to an isotopy we have

Σ1 (f) = S and Σ1,1 (f) = i (Ci+ ∪ Ci− ). Using cusp-elimination surgery we
can now remove the cusp locus, leaving only a fold isotopic to S.
14.2. Folded solutions to differential relations

Let V be an n-dimensional manifold and S ⊂ V a closed codimension 1
submanifold. Choose a tubular neighborhood N ⊂ V of S and an orientation
reversing involution j : N → N which is fixed on S. Let V S be the manifold
with boundary S = S S obtained by cutting V open (or equivalently taking
the oriented real blow-up of V ) along S. If S divides V into two parts then
V S is the disjoint union of two manifolds V1 , V2 ⊂ V with boundary S, where
V1 ∪ V2 = V, V1 ∩ V2 = S. The involution j lifts to a tubular neighborhood
N S of the boundary ∂V S of V S as an involution without fixed points. We
will keep the notation j for this lift. Let us also define
VS := V /{v ∼ j(v), v ∈ Int N }, NS := Int N/{v ∼ j(v), v ∈ Int N }.
Then VS is a non-Hausdorff manifold with boundary S and NS is an open
collar of its boundary. Denote by πV,S the canonical projection V → VS . The
restriction πV,S |S : S → S is the identity map, but the map πV,S : V → VS
is only Lipschitz along S: in local coordinates (s, t) ∈ S × (−1, 1) = IntN it
is given by (s, t) → (s, |t|). The restriction πV,S |V \S : V \ S → VS \ S is a
non-Hausdorff branched cover with ramification locus ∂N . Choose an even
function ϕ : [−1, 1] → [0, 1] which is equal to t2 near 0, to |t| near 1, and
which has no other critical points besides the origin, and modify πV,S to a
14.2. Folded solutions to differential relations 165
smooth map
V,S : V → VS
π
V,S (s, t) = (s, ϕ(t)) on S × (−1, 1) ⊂ N ⊂ V .

with fold along S by setting π
According to Proposition 11.6.7(a) any map f : V → W , dim V = dim W =
n, which has the fold S = Σ1 (f ) as its only singularity can be factored, for
V,S
π gS
an appropriate choice of a tubular neighborhood N , as V → VS → W where
gS is an immersion. On the other hand, an immersion gS : VS → W can be
interpreted as an immersion g S : V S → V which satisfies g S ◦ j = g S on N S .
We will use, when it is convenient, all of the above interpretations of a map
which is folded along S.
Consider now a natural fibration p : X → V . As in Section 8.1.C, we

will think about the natural fibration as a functor X on the category of
n-dimensional manifolds such that X = X (V ). Recall that for any n-
dimensional manifold V the bundle X := X (V ) → V is associated with
D
the principal D0 (n)-bundle P (V ) via an action α of D0 (n) on its standard
fiber W , and a Diff V -invariant differential relation R ⊂ X
(r) is determined
by a subset R ⊂ W invariant with respect to the lifted action αr of D0 (n)
r
on the standard fiber W r of the fibration X (r) .

The involution j : N → N canonically lifts to an involution J : p−1 (N ) →
p−1 (N ). Note that X (VS ) = XS := X/{u ∼ J(u), u ∈ p−1 (IntN )} and we
have a commutative diagram
πX,p−1 (S)
X
⏐ −→ X⏐S
p⏐
.
⏐p
. S
πV,S
V −→ VS .
Let R ⊂ W r be an αr -invariant subset and R = R(V ) ⊂ X (r) the corre-

sponding Diff V -invariant differential relation. We will refer to a solution
fS : VS → XS of RS = R(VS ) as a solution of R = R(V ) folded along S. As
in the case of folded maps, a solution of RS can be equivalently defined as
a solution f S : V S → RS = R(V S ) which satisfies the relation j ∗ f S = f S
over N S ⊂ V S .
14.2.1. (Existence h-principle for folded solutions) Let V be a con-

nected closed manifold of dimension > 1, S ⊂ V a nonempty coorientable
codimension 1 closed submanifold, X = X (V ) → V a natural fibration, and
R ⊂ X (r) an open Diff V -invariant differential relation. Then the existence
h-principle holds for solutions of RS = R(VS ) folded along S.
Proof. Consider first the special case of the relation Rimm in the equidi-
mensional case. We need to prove that given manifolds V, W of dimension
n ≥ 2, a nonempty hypersurface S ⊂ V , and a map ϕ : VS → W covered
by a fiberwise isomorphism Φ : T VS → T W , there exists an immersion
f : VS → W such that Φ and df are homotopic as fiberwise isomorphisms.
Applying the Smale–Hirsch Immersion Theorem 9.2.1 we first deform (Φ, ϕ)
to arrange Φ = dϕ on NS ⊂ VS . Consider the map ϕ := ϕ ◦ π
V,S : V → W .
N has a fold singularity along S. Using Theorem 13.2.2
The restriction ϕ|
we C 0 -approximate ϕ|V \S , keeping it fixed on N , by a wrinkled map, still
whose regularized differential is homotopic to Φ. Finally, we
denoted by ϕ,
apply Lemma 14.1.3 to eliminate the wrinkles. This concludes the proof for
the equidimensional case of the immersion relation Rimm .
Let us now consider the case of a general Diff V -invariant differential re-
lation. Let Φ : VS → RS be a formal solution of RS = R(VS ) covering
a section ϕ : VS → XS . Appplying Gromov’s h-principle 8.3.3 for open
DiffNS -invariant differential relations to the open manifold NS , we can de-
form Φ to make it equal to the holonomic section Jϕr over NS ⊂ VS .
Next, we apply Theorem 5.2.1 to find a multisection ϕ : V ,S → XS which
coincides with ϕ on NS and such that its r-jet extension J r ϕ is C 0 -close
to Φ. If the approximation is sufficiently good, then openness of R(VS )
guarantees that ϕ is a genuine solution of the differential relation R(VS ) in
the formal class determined by Φ.
Let w : VS → V,S be the wrinkled immersion constructed in 13.3.3. Consider
w := w ◦π V,S : V → VS . The map w has a fold singularity along S and
a number of wrinkles elsewhere. Applying Lemma 14.1.3 we can eliminate
the wrinkles to get a map w : V → VS with a fold along S and no other
singularities. According to Proposition 11.6.7(a) we can factor the map w
as
V,S
π g
V → VS → VS ,
where g is an immersion. Now Diff V -invariance of the differential relation
R allows us to consider the induced section ψ := g ∗ ϕ : VS → X (r) , which is
the required folded solution of the differential relation R.
Here is a corollary of Theorem 14.2.1 (we leave the verification of the corre-
sponding homotopic conditions to the reader, see also [El70]).
14.2.2. (Maps with prescribed fold) Let V be a connected orientable
manifold of dimension n ≥ 1, and S ⊂ V a nonempty coorientable codimen-
sion 1 closed submanifold. Then a map V → Rn with fold along S (and
no other singularities) exists if and only if V is stably parallelizable (i.e.,
14.2. Folded solutions to differential relations 167
its tangent bundle is stably trivial), S divides V into two parts V+ and V− ,
V+ ∩ V− = S, and in addition one of the following conditions is satisfied:
(a) n = 1, 3, 7;
(b) n is even and χ(V+ ) = χ(V− );
(c) n is odd, the tangent bundle T V is trivial, and χ(M ) ≡ 0 (mod 4);
(d) n is odd, T V is nontrivial, and χ(M ) ≡ 2 (mod 4).
In particular, there is a map S 3 → R3 with only a fold singularity along any

given nonempty closed surface S ⊂ S 3 .
Some other corollaries of Theorem 14.2.1 are discussed in Section 19.3 below.
Exercise.
1. Prove the following generalization of Theorem 14.2.1, see [Gr86]: Let
V be a manifold with boundary and N ⊃ ∂V a tubular neighborhood of the
boundary ∂V . Suppose a finite group G acts on N without fixed points. Let
X → V be a natural fibration and R ⊂ X (r) a Diff V -invariant differential
relation. Then the existence h-principle holds for solutions ϕ : V → X of
R whose restriction to N is invariant with respect to G, i.e., g∗ (ϕ|N ) = ϕN
for g ∈ G. In particular, let V1 , . . . , Vk be orientable 3-manifolds with the
same nonempty boundary S, i.e., we are given diffeomorphisms ϕj : S →
∂Vj , j = 1, . . . , k. Then there are immersions fj : Vj → R3 such that all the
immersions fj ◦ ϕj : S → R3 , j = 1, . . . , k, coincide.
2. Generalize Part 1 to the statement of epimorphism between the k-th
homotopy groups of the spaces of formal and genuine solutions of R, invari-
ant with respect to G.
3. Show that injectivity on π0 fails for the problem of immersions S 2 → R2
with folds along the equator S 1 ⊂ S 2 . (Hint: See Figure 0.1).

Remark. It is proven in [EM11] that the full h-principle for maps V →
W between equidimensional manifolds with only a fold along S ⊂ V holds if
and only if there is no global involution j : V → V with S as its fixed point
locus.
Chapter 15
The h-principle for

Sharp Wrinkled
Embeddings
We adapt in this chapter the notion of a wrinkle to work in the context of

embeddings and prove the corresponding h-principle. This in turn implies
an h-principle for embeddings folded with respect to a foliation, which will
be a crucial ingredient in the proof of Igusa’s theorem in the next chapter.
15.1. Sharp wrinkled embeddings

A. Sharp Σ1r -singularities. Let n ≥ r ≥ 1. Consider the map
H n,r := H 1,r,n−1,0 : Rn = Rn−1 × R → Rn−1 × R
which defines a Σ1r -singularity Rn → Rn in its Morin normal form (see

Section 11.6),

r−1
H n,r
(t, z) = (t, h n,r
(t, z)) , h n,r
(t, z) := ht1,r,n−1,r (z) =z r+1
+ tj z j .
1
Its Σ1 -singular locus is given by the equation
∂hn,r (t, z) j−1

r−1
= jz tj + (r + 1)z r = 0.
∂z
1
169
170 15. The h-principle for Sharp Wrinkled Embeddings
Consider the function un,r : Rn−1 × R → R,

z ! n,r " z
r−1
2
n,r ∂h (t, ζ) 2 j−1 r
(12) u (t, z) = dζ = jζ tj + (r + 1)ζ dζ,
∂ζ
0 0 1
and define the map H n,r : Rn → Rn × R by

H n,r (t, z) = (H n,r (t, z), un,r (t, z)) = (t, hn,r (t, z), un,r (t, z)) .
As for each t the function z → un,r (t, z) is strictly increasing except at
finitely many values of z, the map H n,r is a topological embedding. At
n,r
points of Σ1 (H n,r ) we have ∂u∂z = 0, so H n,r and H n,r have the same
singular locus, although for H n,r its singularities are not generic.
We will call the function un,r the unfolding function for the singularity Σ1r ,
and refer to the singularity of H n,r as a sharp Σ1r -singularity. In particular,
we will be talking about sharp folds (r = 1) and sharp cusps (r = 2). See
Figures 15.1 and 15.2 in the case n = 2.
Figure 15.1. Sharp fold.
Figure 15.2. Sharp cusp.

15.1. Sharp wrinkled embeddings 171
Remarks.
1. Our terminology differs from some commonly used terms in singularity
theory. For instance, in the case n = r = 1 the map H 1,1 : R → R2 given
by the formula H 1,1 (z) = (z 2 , 43 z 3 ) is a semicubic parabola, and its singular
point is usually referred to as a cusp and not a sharp fold. For n = r = 2
the map H 2,2 : R2 → R3 given by the formula H(t, z) = (t, z 3 + tz, 95 z 5 +
2z 3 t + t2 z) (see Figure 15.2) is sometimes called an unfurled swallowtail (see
[Mu11]) rather than a sharp cusp.
2. Given a cusp f : R2 → R2 in its standard Whitney-Morin form f (t, z) =
(t, z 3 − 3tz)1 , the restriction of this map to the one-dimensional submanifold
Σ1 (f ) = {t = z 2 } is a sharp fold. On the other hand, for a map f : R3 → R3
with the swallowtail singularity Σ1,1,1 (f ) in its Morin form
f (t1 , t2 , z) = (t1 , t2 , z 4 + t2 z 2 + t1 z)
the restriction of f to Σ1 (f ) = {t1 = −4z 3 − 2t2 z} ⊂ R3 is not a topolog-
ical manifold (see Figure 11.3), hence it is not a sharp cusp, although its
projection onto the first two coordinates is a two-dimensional cusp.
3. It is sometimes useful to use other forms of unfolding functions for sharp
Σ1r singularities, see [AG18b]. Namely, given any C ∞ -function θ : R → R
with θ(0) = 0 and θ(t) > 0 for t = 0, we can define the unfolding function
z ! n,r " z r−1

n,r ∂h (t, ζ) j−1 r
uθ (t, z) = θ dζ = θ jζ tj + (r + 1)ζ dζ
∂ζ
0 0 1
instead of (12), where we use θ(t) = t2 . Using θ which is faster decaying at

0 makes the singularities sharper.
As we already mentioned above, the image V := H n,r (Rn ) is a topological

submanifold of Rn+1 with singular locus Σ := H n,r (Σ), Σ := Σ1 (H n,r ) ⊂ Rn .
15.1.1. At every point p ∈ Σ = Σ1 (H n,r ) there is a well defined n-dimen-
sional tangent space to V = H n,r (Rn ) at p = H n,r (p) satisfying
TpV = lim dq H n,r (Rn ).
Rn \Σq→p
Proof. Consider the coordinate functions v = hn,r (t, z) and u = un,r (t, z) =
z ∂hn,r (t,ζ) 2
∂ζ dζ of the map H n,r . It is sufficient to verify the existence of
0
1 Here and elsewhere we use slight modification of the parameters of normal forms at our
convenience.
∂
the limit lim d(t,z) (v, u) Span( ∂z . We have
(t,z)→Σ
! " ! "
∂ ∂hn,r (t, z) ∂hn,r (t, z)
d(t,z) (v, u) = 1, ,
∂z ∂z ∂z
and therefore,
! "
1 ∂
∂hn,r (t,z)
d(t,z) (v, u) −→ (1, 0).
∂z (t,z)→Σ
∂z
B. Embeddings with sharp corank 1 singularities. Let V, W be mani-

folds of dimensions n and q ≥ n+1, respectively. A smooth map f : V → W
is called an embedding with sharp corank 1 singularities if the following prop-
erties are satisfied:
• f is a topological embedding;
• the singular set Σ = Σ(f ) ⊂ V is a codimension 1 submanifold, and
for any p ∈ Σ the map f |Op V p splits as a composition f = g ◦ h,
where h(p) = 0, g is an embedding Rn+1 → W and h : Op V p →
Rn+1 is equivalent to the germ of H n,r for some r = 1, . . . , n.
We will use the notation Σ1r (f ) for the locus of sharp Σ1r -singularities of
an embedding f : V → W with sharp Σ1r -singularities.
Denote V := f (V ) and Σ := f (Σ). Lemma 15.1.1 implies that the tangent
bundle T (V \ Σ) continuously extends to points of Σ. It turns out that along
Σ there is also a canonical extended tangent bundle E ⊃ T V |Σ of dimension
n + 1.
15.1.2. (Extended tangent bundle) Consider a point p ∈ Σ1r (f ) ⊂

Σ ⊂ V . Then given any splitting f |Op V p = g ◦ h, h : Op V p → Rn+1 ,
g : Rn+1 → W as in the definition of a sharp Σ1r -singularity, the space
Ef (p) := d0 g(Rn+1 ) ⊂ Tf (p) W is independent of the choice of the splitting.
Proof. It is sufficient to consider the case when f = j ◦ H n,r , where

j is the inclusion Rn+1 → Rq . We denote coordinates in Rn+1 ⊂ Rq
by (t, v, u) where t ∈ Rn−1 , u, v ∈ R, so that the map H n,r is given by
(t, z) → (t, hn,r (t, z), un,r (t, z)). The tangent space T to H n,r (Rn ) at the ori-
gin coincides with Rn = {u = 0} ⊂ Rn+1 ⊂ Rq . We claim that for each split-
ting the extended tangent space E at the origin agrees with Rn+1 . For this,
it suffices to prove that for any function ψ : Rq → R satisfying ψ ◦ H n,r ≡ 0
we have d0 ψ|Rn+1 = 0. As we already know that d0 ψ|Rn = d0 ψ|T = 0,
it is sufficient to show that ψu := ∂ψ ∂u (0) = 0. Differentiating the identity
ψ(H n,r (t, z)) ≡ 0 along the z-axis we get

∂(ψ ◦ H n,r )
(0, z)
∂z
! n,r "2
∂hn,r ∂h
= ψv (H n,r (0, z)) (0, z) + ψu (H n,r (0, z)) (0, z)
∂z ∂z

= (r + 1)z r ψv (H n,r (0, z)) + (r + 1)z r ψu (H n,r (0, z)) = 0.
This implies
ψv (H n,r (0, z)) + (r + 1)z r ψu (H n,r (0, z)) = 0,
and differentiating this identity again we get
(r + 1)rz r−1 ψu (H n,r (0, z)) + o(z r ) = 0,
which implies ψu (0) = 0.
Assuming we have fixed a Riemannian metric on W , we call the line field

μ ⊂ E along Σ which is normal to V the principal normal line bundle to V .
If μ is trivial we denote by μ+ a unit vector field in μ.
Let us now consider the special case when the sharp singularity Σ1 (f ) is
only of fold type, i.e., Σ1 (f ) = Σ1,0 (f ). We will also assume the bundle line
μ to be trivial, although the discussion below can be modified to the case of
a general μ with minimal adjustments.
15.1.3. Let f : V → W be an embedding with only a sharp fold Σ = Σ1 (f ).
Denote V = f (V ), Σ = f (Σ) and U = Σ × (−ε, ε) × (−ε, ε), ε > 0. Then
there exists an embedding ϕ : U → Op Σ W such that V ∩ Op Σ W ⊂ ϕ(U ).
Proof. We denote coordinates on U = Σ × (−ε, ε) × (−ε, ε) by (p, t, s). By

identifying U with a neighborhood of Σ in the normal bundle to T Σ in the
extended tangent bundle E|Σ and taking the exponential map, we construct
an embedding ψ : U → W such that ψ(p, 0, 0) = f (p), ∂s ∂
= μ+ , and ∂t∂
is the inward unit normal vector field to T Σ in T V |Σ . Denote U = ψ(U ).

Let U1 be a tubular neighborhood of U in W and π : U1 → U the normal
projection. Denote V := π(V ∩ U1 ) ⊂ U . Note that if the neighborhood
U1 is chosen sufficiently small then V ∩ U1 is graphical over V , i.e., there
is a section s : V → U1 such that s(V ) = V ∩ U1 and π ◦ s is the identity.
We claim that s extends to a C ∞ -section U → U1 . Recall that according to
the definition of an embedding with sharp folds, for each point p ∈ Σ there
exists an embedding g : Rn+1 → Op W p with g(0) = p, dp g(Rn+1 ) = Ep , and
g(Rn+1 ) ⊃ V ∩Op W p. Hence, there exists a local section Sp : Op U p → U1 of
the normal fibration U1 → U which coincides with s on Op V p. This means
that the section s is smooth in the sense of Whitney, hence by Whitney’s
theorem in [Wh34] it extends to a smooth section S : U → U1 . Then the

composition ϕ := S ◦ ψ : Σ × (−ε, ε) × (−ε, ε) → U1 ⊂ W has the required
properties.
C. Sharp wrinkled embeddings. An embedding f : V → W with sharp

corank 1 singularities is called a sharp wrinkled embedding, or just a wrinkled
embedding, if
• f has only sharp Σ1,0 and Σ1,1,0 singularities;
• Σ1 (f ) is the disjoint union of (n − 1)-spheres Sj , j = 1, . . . , N , and
Σ1,1 (f ) is the union of their (n − 2)-dimensional equatorial spheres
Sj ⊂ Sj ;
• each sphere Si bounds an n-ball Di ⊂ V ;
• the characteristic vector field νi along Si is outwardly transverse to
Di ;
• the image Si = f (Si ) of each sphere Si admits a principal normal
vector field μ+
i which extends to Di = f (Di ) as a normal vector
field to V .
We consider the choice of balls Di and vector fields μ+ i to be part of the
structure of a wrinkled embedding f . The components Si of Σ1 (f ) are called
the sharp wrinkles, or simply the wrinkles, of f . The balls Di , i = 1, . . . , N ,
spanning the wrinkles are not required to be disjoint and could be nested.
We say that a wrinkle Si has depth d + 1 if it is surrounded by d nested
wrinkles. The depth of a wrinkled embedding is the maximal depth of its
wrinkles, so a wrinkled embedding of depth 0 is a smooth embedding. For a
point y in V \Σ1 (f ) we denote by d(y) the number of balls spanning wrinkles
to which the point y belongs, so that d(y) is the maximal multiplicity of a
wrinkle Si = ∂Di such that y ∈ Di .
Example (The standard sharp wrinkle). The standard sharp wrinkle is
the germ of the map
w = wn,q : Rn = Rn−1 × R → Rq = Rn−1 × R × R × Rq−n−1
given by

w(y, z) := y, z 3 + 3(|y|2 − 1)z, u(y, z), 0
near the n-dimensional unit disk
D n := {|y|2 + z 2 ≤ 1} ⊂ Rn ,
where the unfolding function u : Rn → R is defined by
z
u(y, z) := (|y|2 + ζ 2 − 1)2 dζ.
0
z=1 y=const
z=1
Figure 15.3. Standard sharp wrinkles for q = n + 1 = 2 and

q = n + 1 = 3.
Clearly, the standard sharp wrinkle is an example of a sharp wrinkled em-

bedding, see Figure 15.3.
The embryo of a standard sharp wrinkle is the germ at 0 of the map
z
Rn−1
×R→R , q
(y, z) → y, z + 3|y| z,
3 2
(|y|2 + ζ 2 )2 dζ, 0 .
0
Note that the standard sharp wrinkle is fibered over Rn−1 and its restric-
tion to the fiber over y is either a standard sharp wrinkle (if y ∈ Int D n ),
/ D n ), or an embryo (if y ∈ ∂D n ).
nonsingular (if y ∈
By Lemma 15.1.1 a wrinkled embedding f : V → W has a canonical Gauss

map
Gdf : V → Grn W, Gdf (x) = lim dy f (Ty V )
V \Σy→x
to the Grassmannian of n-planes in T W . If V is oriented this lifts to a map
to the oriented Grassmannian
: V → Gr
Gdf n W, (x) =
Gdf lim (−1)d(y) dy f (Ty V ).
V \Σy→x
Note that the Gauss map does not lift to a continuous map T V → T W .
A tubular neighborhood N of f (V ) in W carries a normal foliation N , i.e., a
foliation whose leaves are transverse to the tangent spaces to f (V ).
D. Homotopies and isotopies of wrinkled embeddings. We will dis-

tinguish between homotopies and isotopies of wrinkled embeddings. A
homotopy of wrinkled embeddings ft : V → W , t ∈ I = [0, 1], is a fam-
ily of topological embeddings such that the singularities of each ft are either
sharp wrinkles or sharp embryos. Thus, in a homotopy of fibered wrinkled
embeddings, fibered wrinkles can appear or disappear through embryos. A
homotopy in which this does not happen will be called an isotopy. This is
equivalent to having ft = ht ◦ f0 for a diffeotopy ht : W → W with h0 = id.
15.2. Integration of tangential rotations

A. Tangential rotations. Let f : V → W be a (possibly wrinkled) embed-
ding with Gauss map Gdf : V → Grn W , where π : Grn W → W is the Grass-
mannian bundle. A tangential rotation of f is a homotopy Gt : V → Grn W ,
t ∈ [0, 1], with G0 = Gdf and π ◦ Gt = f : V → W .
Now we can state the main theorem of this chapter.
15.2.1. (Wrinkled approximation of a tangential rotation [EM09])
Let V and W be smooth manifolds of dimension n and q > n, respectively.
Let Gt : V → Grn W be a tangential rotation of a (possibly wrinkled) embed-
ding f0 : V → W . Then f0 extends to a homotopy of wrinkled embeddings
ft : V → W such that Gdft : V → Grn W is arbitrarily C 0 -close to Gt for
all t ∈ [0, 1]. If Gt = df0 over Op C for a closed subset C ⊂ V , then we can
arrange ft = f0 on Op C.
See Figure 15.4 for an illustration in the case n = 1, q = 2. Note that

C 0 -closeness of Gdft to Gt includes C 0 -closeness of ft to f0 = π ◦ Gt .
Figure 15.4. Wrinkled approximation of a counterclockwise rotation.
B. Wrinkled directed embeddings. Recall that an embedding f : V →

W is called A-directed for a subset A ⊂ Grn W if Gdf (V ) ⊂ A. See Section
4.5 where we established an h-principle for A-directed embeddings when
A open and V is open (Theorem 4.5.1), and Section 4.6 where for certain
classes of open A we proved an h-principle for closed V (Theorem 4.6.1).
However, for a general open A and closed V that h-principle fails.
Example. For a closed oriented connected n-manifold V and an embed-
with
ding f : V → Rn+1 , the composition of the oriented Gauss map Gdf
/n R
the projection p2 : Gr n+1 =R n+1 × S → S is surjective, so f can be
n n
A-directed only if p2 (A) = S . Now choose f : V → Rn+1 such that the

n
: V → S n has mapping degree 0 and is thus homotopic to a
map p2 ◦ Gdf
constant map (e.g., the standard embedding T 2 → R3 has this property),
and take A = Rn+1 × U for an open disk U ⊂ S n . Then f is covered by a
and G1 (V ) ⊂ A, but f is not isotopic
tangential rotation Gt with G0 = Gdf
to an A-directed embedding.
15.3. Preliminary steps in the proof 177
On the other hand, Theorem 15.2.1 immediately yields the following h-

principle for A-directed wrinkled embeddings of possibly closed manifolds
(with A still open).
15.2.2. (h-principle for A-directed wrinkled embeddings [EM09])
Let V and W be smooth manifolds of dimension n and q > n, respectively,
and A ⊂ Grn W an open subset. Let Gt : V → Grn W be a tangential
rotation of an embedding f0 : V → W such that G1 (V ) ⊂ A. Then f0
extends to a homotopy of wrinkled embeddings ft : V → W , arbitrarily C 0 -
close to f0 , such that Gdf1 (V ) ⊂ A. If Gt = df0 over some closed subset
C ⊂ V , then we can arrange ft = f0 on C.
15.3. Preliminary steps in the proof

The remainder of this chapter is devoted to the proof of Theorem 15.2.1.
We begin with some preliminary reductions of the proof to simpler cases.
A. Reduction to simple tangential rotations. Let us equip W with

a Riemannian metric. Then we define the angle θ ∈ [0, π/2] between two
unoriented n-planes P, P contained in an (n + 1)-plane L ⊂ Tw W as the
smaller angle between their normal lines in L. We call a tangential rotation
Gt simple if for each v ∈ V the rotation Gt (v) takes place in an (n + 1)-
plane Lv ⊂ Tf0 (v) W and the angle between G0 (v) and Gt (v) is < π/4 for all
t ∈ [0, 1]. The proof of the following lemma is straightforward.
15.3.1. (Reduction to simple rotations) Any tangential rotation can
be C 0 -approximated by a finite sequence of simple rotations.
Hence, Theorem 15.2.1 follows by induction from the following result. Note
that the inductive procedure forces us to allow f0 to be wrinkled.
15.3.2. (Wrinkled approximation of simple tangential rotations)
Let V and W be manifolds with n = dim V < q = dim W . Let Gt : V →
Grn W be a simple tangential rotation of a wrinkled embedding f0 : V → W .
Then f0 extends to a homotopy of wrinkled embeddings ft : V → W such
that Gdft : V → Grn W is arbitrarily C 0 -close to Gt for all t ∈ [0, 1]. If
Gt = df0 over Op C for some closed subset C ⊂ V , then we can arrange
ft = f0 on Op C.
B. Approximation over a subset of positive codimension. Suppose
V ⊂ W is a smooth submanifold with inclusion f0 : V → W . Pick a tubular
neighbourhood X ⊂ W of V with projection p : X → V and normal foliation
N by the fibers of p. An isotopy ft : V → X extending f0 is called graphical
if each ft is transverse to N , i.e., ft = ft ◦ ht for sections ft : V → X and
diffeomorphisms ht : V → V . A tangential rotation Gt : V → Grn W over V
is called graphical if each Gt (v) is transverse to N .
15.3.3. (Smooth approximation of graphical tangential rotations

near a subset of positive codimension) In the preceding setup, let Gt :
V → Grn W be a graphical tangential rotation over V ⊂ W and K ⊂ V be
a stratified subset of positive codimension. Then the inclusion f0 : V → W
extends to an arbitrarily C 0 -small graphical isotopy ft : V → X ⊂ W such
that Gdft |Op K is arbitrarily C 0 -close to Gt |Op K for all t ∈ [0, 1].
Proof. The space X (1) of 1-jets of sections V → X can be described as the

space of pairs (x, P ) where x ∈ X and P ⊂ Tx X is an n-plane transverse
to N . Hence the graphical tangential rotation Gt defines a homotopy of
sections

Ft : V → X (1) , v → f0 (v), Gt (v)
covering the inclusion f0 : V → X. By the Holonomic Approximation
Theorem 3.1.1, there exists a C 0 -small diffeotopy ht : V → V with h0 = id
and a homotopy of sections ft : Op ht (K) → X with f0 = f0 |Op h0 (K) whose
1-jets J 1 : Op ht (K) → X (1) are arbitrarily C 0 -close to Ft |Op ht (K) . Extend
ft
the ft to sections ft : V → X that are C 0 -close to f0 with f0 = f0 . Then
ft := ft ◦ ht is a graphical isotopy extending f0 and C 0 -close to f0 . For
v ∈ Op K we have h (v) ∈ Op h (K), thus Gdf (v) = Gd ft ht (v) is C -
0
t t t
close to Gt ht (v) , hence to Gt (v) because ht is C 0 -small.
The notions of a graphical isotopy and tangential rotation carry over to the
case that f0 : V → W is a wrinkled embedding, using the normal foliation
N on a tubular neighbourhood X of f (V ). Then Lemma 15.3.3 implies
15.3.4. (Wrinkled approximation of graphical tangential rotations
near the singular set) Let Gt : V → Grn W be a graphical tangential
rotation of a wrinkled embedding f0 : V → W with singular set Σ = Σ(f0 ).
Then f0 extends to an arbitrarily C 0 -small graphical wrinkled isotopy ft :
V → W such that Gdft |Op Σ is arbitrarily C 0 -close to Gt |Op Σ for all t ∈ [0, 1].
Proof. Pick an n-dimensional submanifold V ⊂ W containing f0 (Σ) such

that V is tangent to f0 (V ) along f0 (Σ). Extend Gt |Σ to a graphical tan-
gential rotation Gt : V → Grn W and apply Lemma 15.3.3 to V ⊂ W , Gt
and K := f0 (Σ) ⊂ V . Let ft : V → W be the resulting graphical iso-
topy with f0 : V → W the inclusion. We can write ft = gt ◦ f0 for a
C 0 -small diffeotopy gt : W → W preserving the normal foliation N with
g0 = id. Then ft := gt ◦ f0 is a graphical wrinkled isotopy
extending
f0
and C 0 -close to f0 . For v ∈ Σ we have ft (v) = gt f0 (v) = ft f0 (v) , thus

Gdft (v) = Gdft f0 (v) is C 0 -close to Gt (v). This C 0 -closeness continues to
hold on a neighbourhood Op Σ ⊂ V .
15.4. Reduction to the main lemma 179
In the next section we will use Lemma 15.3.3 and Corollary 15.3.4 to reduce
the proof of Theorem 15.3.2 to a lemma about hypersurfaces in Rn+1 .
15.4. Reduction to the main lemma

In this section we reduce the proof of Theorem 15.2.1 to Lemma 15.4.1
below, which will be proved in the next section.
Consider a compact oriented hypersurface S ⊂ Rn+1 (possibly with bound-
ary) with inclusion map ιS : S → Rn+1 . Denote the composition of the ori-
ented Gauss map Gdι n Rn+1 = Rn+1 ×S n → S n
S with the projection p2 : Gr
(i.e., the classical Gauss map) by
:= p2 ◦ Gdι
G S : S → Sn.
For θ ∈ [0, π] denote by Uθ ⊂ S n the open θ-neighbourhood of the north
pole (0, . . . , 0, 1) with respect to the angular metric.
The hypersurface S ⊂ Rn+1 is called

• ε-horizontal if G(S) ⊂ Uε for a small given ε > 0;

• graphical if G(S) ⊂ Uπ/2 ;

• quasi-graphical if G(S) ⊂ Uπ .
These notions extend to a wrinkled embedding f : S → Rn+1 in the obvious
way. Now we can state the main lemma.
15.4.1. (Approximation of a family of hypersurfaces by a family
of ε-horizontal wrinkled embeddings) Let St ⊂ Rn+1 , t ∈ [0, 1], be a
smooth family of compact oriented quasi-graphical hypersurfaces such that
S0 is ε-horizontal, and each St is ε-horizontal near ∂St for all t. Then the
family of inclusions ιSt : St → Rn+1 can be C 0 -approximated by a family
of ε-horizontal wrinkled embeddings ft : St → Rn+1 of depth ≤ 1 such that
f0 = ιS0 , and ft = ιSt near ∂St for all t.
Before proving this lemma, let us show how it implies Theorem 15.3.2.
Proof of Theorem 15.3.2. Let Gt : V → Grn W be a simple tangential

rotation over a wrinkled embedding f0 : V → W . Cover f0 (V ) by coordinate
charts ψ : W ⊃ U → Rq such that

• ψ U ∩ f0 (V ) is C 1 -close to Rn × 0;
• ψ∗ N is the vertical foliation with leaves {x} × Rn−q ;
• the pushforward under ψ of the metric on W is C 0 -close to the
Euclidean metric on Rq ;
• ψ∗ Gt (v) is close to ψ∗ Gt (v ) in Grn Rq for all v, v ∈ f0−1 (U ) and
each t ∈ [0, 1].
Here to achieve the first three properties we choose the metric on W such
that the tangent spaces to f0 (V ) are almost
orthogonal to the normal fo-
liation N . Note that we cannot achieve ψ U ∩ f0 (V ) ⊂ Rn × 0 or strict
orthogonality if f0 is wrinkled. Pick a smooth triangulation of V such that
• the cusp locus Σ1,1 (f0 ) is contained in the (n − 2)-skeleton;
• the singular set Σ(f0 ) is contained in the (n − 1)-skeleton;
• the image f0 (Δi ) of each n-simplex Δi is contained in a coordinate
neighbourhood U from the covering above.
Let K be the (n − 1)-skeleton of this triangulation. By Lemma 15.3.3 and
Corollary 15.3.4, f0 extends to a C 0 -small graphical wrinkled isotopy ft :
V → W such that Gdft |Op K is C 0 -close to Gt |Op K for all t ∈ [0, 1]. For
each i pick an embedded closed n-ball Bi ⊂ Int Δi such that Gdft |Op ∂Bi is
C 0 -close to Gt |Op ∂Bi for all t ∈ [0, 1]. Pick a point vi ∈ Bi and a coordinate
chart ψ : W ⊃ U → Rq as above with f0 (Δi ) ⊂ U . By C 0 -closeness of ft to
f we may assume that ft (Δi ) ⊂ U for all t ∈ [0, 1].
Case q = n + 1. For each t ∈ [0, 1], let ψt : W ⊃ U → Rn+1 be the
composition of ψ with a rotation of Rn+1 mapping ψ∗ Gt (vi ) to Rn × 0.
Then the smooth submanifolds
Sti := ψt ◦ ft (Bi ) ⊂ Rn+1
have the following properties (in the terminology from above):
(i) S0i is ε-horizontal;
(ii) Sti is ε-horizontal near ∂Sti for all t ∈ [0, 1];
(iii) Sti is quasi-graphical for all t ∈ [0, 1].
Here the first two properties are clear from the construction. For property
(iii), recall that ft is graphical in the sense that it is transverse to the
vertical foliation N . Since ψ∗ N is the vertical foliation with leaves {x} × R,
it follows that ψ ◦ ft (Bi ) ⊂ Rn+1 is graphical. Since the tangential rotation
Gt is simple, the angle of the rotation in the definition of ψt can be chosen
to be ≤ (π/4) + ε and it follows that Sti = ψt ◦ ft (Bi ) is quasi-graphical.
So we are in the position to apply Lemma 15.4.1 to the family of hyper-
surfaces Sti , t ∈ [0, 1], for each i. Let gti : Sti → Rn+1 be the result-
ing families of ε-horizontal wrinkled embeddings. Then the compositions
ψt−1 ◦ gti ◦ ψt ◦ ft : Bi → W for all i fit together to the desired homotopy of
wrinkled embeddings ft : V → W .
Case q > n + 1. We construct the triangulation and vi ∈ Bi ⊂ Δi as
above. By definition of a simple tangential rotation, for each i the rotation
Gt (vi ) takes place in an (n + 1)-dimensional subspace Lvi ⊂ Tf0 (vi ) W . For
each i we compose the coordinate chart ψ : U → Rq with a rotation of
15.5. Proof of the main lemma 181
Rq preserving the vertical foliation {x} × Rn−q to get a chart ψ i satisfying

ψ i (Lvi ) = Rn+1 × 0. Then each p ◦ ψ i ◦ ft (Bi ) ⊂ Rn+1 is a graphical
hypersurface, where p : Rq → Rn+1 is the projection. For each t ∈ [0, 1], let
ρt be a rotation of Rn+1 mapping ψ∗i Gt (vi ) to Rn × 0. As above, simpleness
of Gt implies that the hypersurfaces Sti := ρt ◦ p ◦ ψ i ◦ ft (Bi ) ⊂ Rn+1 are
quasi-graphical. Applying the argument in the case q = n+1 to ρt ◦p◦ψ i ◦ft :
Bi → Rn+1 , keeping the components in Rq−n−1 unchanged, concludes the
proof of Theorem 15.3.2.
15.5. Proof of the main lemma

It remains to prove Lemma 15.4.1. We begin with its nonparametric version;
the parametric version will be an easy adaptation of this.
15.5.1. (Approximation of a hypersurface by an ε-horizontal wrin-
kled embedding) Let S ⊂ Rn+1 be a compact oriented quasi-graphical hy-
persurface which is ε-horizontal near ∂S. Then the inclusion ιS : S → Rn+1
can be C 0 -approximated by an ε-horizontal wrinkled embedding f : S → Rn+1
of depth ≤ 1 which agrees with ιS near ∂S.
The proof is achieved by the goffering construction in 3 steps.
Step 1: Preparation for goffering. Let G : S → S n be the coorienting

unit normal vector field along S and denote by N := (0, . . . , 0, 1) ∈ S n the
north pole. By hypothesis, G = −N on S and is ε-close to N near ∂S.
G
Define the height function
h : S → R, h(x) := xn+1 .
After removing from S a neighbourhood of the set of critical points of h
(on which S is already ε-horizontal), we may assume that h has no critical
= ±N on S.
points on S, i.e., G

For x ∈ S let v(x) be the orthogonal projection of G(x) onto Rn ×0, rescaled

to length 1 (see Figure 15.5). Since G(x) = ±N , this defines a unit vector
field v on S which is transverse to S, defines the coorientation of S, and is
horizontal in the sense that dh(v) ≡ 0. Extend v to a tubular neighbourhood
of S in Rn+1 . Let γ > 0 be such that the flow v t (x) of v is defined for all
|t| ≤ γ, x ∈ S, and defines an embeding [−γ, γ] × S → Rn+1 .
Now to any continuous function T : S → [−γ, γ] we associate the topological
embedding
IST : S → Rn+1 , IST (x) := v T (x) (x).
At a point x where T is differentiable we compute

dIST (x) = d(v T (x) )(x) + v v T (x) (x) dT (x).
Since the first term on the right hand side is bounded independently of T
and v is horizontal, the embedding IST will be ε-horizontal at x if |dT (x)|
is large. Now we cannot make |dT | large on all of S for a smooth function
T , but it becomes possible if we allow T to have vertical cusps as shown in
Figure 15.6. This is the basic idea of goffering. In the next step we describe
an explicit local model for such a function T and show that it gives rise to
a wrinkled embedding.
Step 2: The basic building block on a tube. The height function h

from Step 1 induces a pair of transverse foliations on S: the one-dimensional
foliation G by gradient lines of h, and the (n − 1)-dimensional foliation L by
level sets of h. A tube is an embedding ϕ : D n−1 × D 1 → S such that (see
Figure 15.5)
h ~
G
Figure 15.5. A tube domain.
• ϕ maps the horizontal slices D n−1 ×pt to leaves of L and the vertical
slices pt ×D 1 to leaves of G;
• S is ε-horizontal near the lids ϕ(D n−1 × ∂D 1 ).
For t ∈ R we consider the curve wt = (ut , st ) : R → R2 with components

15 z 2 1
(13) ut (z) = (ζ − t)2 dζ, st (z) = − (z 3 − 3tz).
8 0 2
Up to the scaling factors and interchanging u and s, w1 is the one-dimen-
sional standard embedded wrinkle. It shrinks as t decreases, becomes an
embryo at t = 0, and√ is a smooth embedding for t < 0. For t >√0 its
cusps occur at z = ± t, and a short computation shows that wt (± t) =
(±t5/2 , ±t3/2 ) (this is the reason for the scaling factors). Since ut is strictly
increasing, the image of wt is the graph of a continuous function at : R → R,
Figure 15.6. The function at for t = 1 and t = 0.3.
u → at (u). The preceding discussion and some short computations show

that the functions at have the following properties (see Figure 15.6):
(i) at ◦ ut (z) = st (z) is smooth in (t, z);
(ii) at (0) = 0 for all t;
(iii) for t < 0, at is smooth with negative derivative;
(iv) a0 is smooth with negative derivative everywhere except at z = 0
where the slope becomes −∞;
(v) for t > 0, at is smooth except for two vertical cusps at ±t5/2 with
values at (±t5/2 ) = ±t3/2 ;
(vi) for 0 < t ≤ 1, at (u) ≥ 4/5 for |u| < t5/2 and at (u) < 0 for |u| > t5/2 .
Fix a small δ > 0 and a smooth nonincreasing function β : [0, 1] → R such
that
1 r ∈ [0, 1 − 3δ],
β(r) =
1 − r − δ 1 ∈ [1 − 2δ, 1].
Consider now a tube ϕ : D n−1 × D 1 → S. With γ > 0 as in Step 1 we define
a continuous function
τ : D n−1 × D 1 → [−1, 1], τ (y, u) := γ aβ(|y|) (u)
and the corresponding goffering map

(14) I τ : D n−1 × D 1 → Rn+1 , I τ (y, u) := v τ (y,u) ϕ(y, u) .
To continue we need the following lemma.

15.5.2. The map g(y, z) := y, γ uβ(|y|) (z) can be extended from a neigh-
bourhood U of the set {z 2 ≤ β(|y|)} to a smooth homeomorphism g : D n−1 ×
D 1 → D n−1 × D 1 which equals the identity near ∂(D n−1 × D 1 ) such that
I τ ◦ g is smooth and equivalent to the standard sharp wrinkle.
Proof. We may assume after rescaling that γ = 1. Note first that g|U
is clearly smooth, and it can be extended to a homeomorphism because
z → ut (z) is a homeomorphism for each t. Now

I τ ◦ g(y, z) = v τ ◦g(y,z) ϕ ◦ g(y, z) ,
where τ ◦ g(y, z) = aβ(|y|) ◦ uβ(|y|) (z) = sβ(|y|) (z) is smooth by property (i) of
at , so I τ ◦ g is smooth. Next, we choose coordinates (y, s, u) ∈ Rn−1 × R × R
near the image ϕ(D n−1 × D 1 ) of the tube in which ϕ(y, u) = (y, 0, u) and
v = ∂s , so the flow of v is given by v s (y, u) = (y, s, u) and

I τ ◦ g(y, z) = y, sβ(|y|) (z), uβ(|y|) (z) .
Replacing t = β(|y|) in equation (13), we see that I τ ◦g agrees up to rescaling
with the map
z
2 2
(y, z) → y, z − 3β(|y|)z,
3
ζ − β(|y|) dζ .
0
By the choice of β this is a equivalent to the standard embedded wrinkle

(which has β(|y|) = 1 − |y|2 ), and the lemma follows.
Step 3: The goffering construction. Using the functions at from Step

2, we choose for all small σ > 0 a continuous family of functions
Wσ,t : R → [−1, 1], t ∈ [−δ, 1]
with the following properties (see Figure 15.7).
2 1 1 2
Figure 15.7. The function Wσ,1 for a small σ.
(i) Wσ,t is odd and 2-periodic with Wσ,t (k) = 0 for k ∈ Z.

(ii) Wσ,t is smooth except at the points 2k ± σt5/2 , k ∈ Z, for t ≥ 0.
(iii) For t < 0 the function Wσ,t satisfies Wσ,t (u) = at (u/σ) near u =
0 for t ∈ [−δ/4, 0), vanishes identically for t ≤ −δ/2, and has
≤ σ everywhere and W ≥ −1 on [2σ, 1].
derivative Wσ,t σ,t
(iv) For t ≥ 0 the function Wσ,t satisfies Wσ,t (u) = at (u/σ) on a

≤ σ on
neighbourhood of [−σt5/2 , σt5/2 ], and has derivative Wσ,t
≥ −1 on [2σ, 1]. It follows from property (vi) of
(σt5/2 , 1] and Wσ,t

at that Wσ,t ≥ 5σ on (−σt5/2 , σt5/2 ).
4
≤ −1 on (σ, 1].
(v) For t = 1 the function Wσ,1 has derivative Wσ,1
We cover the given hypersurface S ⊂ Rn+1 by the images of tube domains
∼
=
ϕα : D n−1 × D 1 −→ Uα ⊂ S, α = 1, . . . , K.
Consider δ > 0 so small and an odd positive integer N so large that S \ S is

ε-horizontal, where

S := n−1
ϕα (D1−3δ × D1−
1
1 ).
2N
α
After a small perturbation of the ϕα we may assume that the slices

ϕα (D n−1 × { 2k
N }) for α = 1, . . . , K and integers k are pairwise disjoint.
Given N we can therefore find σ > 0 such that the sets
* 2k − 2σ 2k + 2σ +
Uα,k = ϕα D n−1 × , , α = 1, . . . , K, k ∈ Z
N N
are pairwise disjoint.
We pick a nonincreasing cutoff function λ : [0, 1] → [0, 1] which equals 1 on
[0, 1 − 1/2N ] and 0 near 1. With β as in Step 2 and a small constant γ > 0
we define a continuous function τ = τγ,δ,N,σ : D n−1 × D 1 → R,
(15) τ (y, u) := γλ(|u|)Wσ,β(|y|) (N u).
For each α we denote
τα := τ ◦ ϕ−1
α : S → [−γ, γ]
(extended by 0 outside Uα ). Now for γ sufficiently small we consider

K
T := τα : S → [−Kγ, Kγ]
α=1
and the corresponding goffering map

IST : S → Rn+1 , IST (x) = v T (x) (x).
Lemma 15.5.1 now follows from the following lemma.
15.5.3. There exists a smooth homeomorphism g : S → S such that IST ◦ g
is a wrinkled embedding of depth ≤ 1. Moreover, IST is ε-horizontal for N
sufficiently large and σ sufficiently small.
Proof. The first assertion follows from Step 2 applied to each of the disjoint
sets Uα,k . To prove almost horizontality, we fix the constants γ, δ and vary
N and σ (which depends on N ). We will say that a term is bounded if it
is bounded independently of N and σ. At a point x ∈ S we have dIST =
d(v T ) + v(IST ) ◦ dT . Applying this to the gradient ∇h of the height function
h : S → R we get
dIST (∇h) = d(v T )(∇h) + dT (∇h)v(IST ).
Since the first summand on the right hand side is bounded and v is horizon-
tal, IST will be ε-horizontal at x if |dT (∇h)(x)| is large. By definition of a
tube domain, for each α we have dϕ−1 α (∇h)
= fα ∂u for a positive function
fα : D n−1 × D 1 → R. Hence, from T = α τ ◦ ϕ−1 α we get

dT (∇h)(x) = dτ ◦ dϕ−1
α (∇h)(x) = fα ∂u τ (yα , uα ),
α α
where the sum extends over all α such that x = ϕα (yα , uα ) is in the image
of ϕα . Let C > 0 be a constant such that
1
≤ |fα | ≤ C for all α = 1, . . . , K.
C
After rescaling we may assume that γ = 1. Abbreviating t = β(|y|) we then
have
u
(16) ∂u τ (y, u) = λ (|u|) Wσ,t (N u) + λ(|u|)N Wσ,t

(N u).
|u|
To prove ε-horizontality at a point x ∈ S we distinguish 3 cases.
Case 1: x has depth 1. Then there is a unique α0 such that x = ϕα0 (yα0 , uα0 )
with t = β(|yα0 |) > 0 and N uα0 ∈ (2k − σt5/2 , 2k + σt5/2 ) for some k ∈ Z.
Thus λ ≡ 1 near |uα0 | and property (iv) of Wσ,t implies ∂u τ (yα0 , uα0 ) ≥
all α = α0 with x = ϕα (yα , uα ) in the image of ϕα we have
4
5σ N . For
N uα ∈/ k∈Z [2k − 2σ, 2k + 2σ]. Since λ ≡ 1 near |uα |, properties (iii) and
(iv) of Wσ,t imply ∂u τ (yα , uα ) ≥ −N . As this occurs at most K − 1 times,
we get
4
dT (∇h)(x) ≥ N − (K − 1)C ≥ CN
5Cσ
provided that σ ≤ 4
Thus IST is ε-horizontal at x for N large.
5KC 2
.
Case 2: x has depth 0 and x ∈ S. Then by definition of S there exists
α0 such that x = ϕα0 (yα0 , uα0 ) with (yα0 , uα0 ) ∈ D1−3δ
n−1
× D1−
1
1 . Thus
2N
t = β(|yα0 |) = 1 and λ ≡ 1 near |uα0 |. Since x has depth 0, we must have
N uα 0 ∈
/ ∪k∈Z [2k −σ, 2k +σ] and property (v) of Wσ,t implies ∂u τ (yα0 , uα0 ) ≤
−N . For all α = α0 with x = ϕα (yα , uα ) in the image of ϕα properties (iii)
15.6. The h-principle for folded embeddings 187
and (iv) of Wσ,t imply ∂u τ (yα , uα ) ≤ N σ. As this occurs at most K − 1

times, we get
1 N
dT (∇h)(x) ≤ N − + (K − 1)Cσ ≤ −
C 2C
provided that σ ≤ 2(K−1)C 2 . Thus IS is ε-horizontal at x for N large.
1 T
Case 3: x has depth 0 and x ∈ S \ S. Note that the first term on the right
hand side of (16) is ≤ 0. Therefore, as in Case 2 we have ∂u τ (yα , uα ) ≤ N σ
for all α with x = ϕα (yα , uα ) in the image of ϕα , hence
1
dT (∇h)(x) ≤ KCN σ ≤
N

provided that σ ≤ KCN 2 . Now recall that S \ S is already ε-horizontal. By
1
definition of the vector fields v and ∇h this implies v(x), ∇h(x) < 0 (see
Figure 15.5), so a deformation in direction v(x) with negative dT (∇h)(x)
makes the hypersurface only more horizontal at x. Thus IST remains ε-
horizontal at x for N large.
This concludes the proof of Lemma 15.5.3, and thus of Lemma 15.5.1.
Proof of Lemma 15.4.1. In order to prove the 1-parametric version, we

need to make the following modifications in the preceding proof of Lemma
15.5.1. The family St , t ∈ I, can be viewed as a quasi-graphical hypersurface
S in I × Rn+1 fibered over I. Now all the above constructions need to be
done in the fibered category. In particular, instead of the cylinder D n × D 1
we need to use the fibered cylinder I × D n−1 × D 1 for the parameterization
of the fibered tubes. Notethat the hypersurface S is ε-horizontal near
the part ∂ S = {0} × S0 ∪ t∈[0,1] {t} × ∂St of its boundary, but not near
∂ S = {1} × S1 . So we modify the proof of Lemma 15.5.1 to keep the
deformation fixed near ∂ S, allowing it to vary freely near ∂ S.
15.6. The h-principle for folded embeddings

A. Sharp pleating and rounding of sharp wrinkles. Recall that a
sharp folded embedding is an embedding f : V n → W q , n < q, with sharp
corank 1 singularities of only fold (i.e., Σ1,0 ) type.
Let f : V → W be a sharp folded embedding (with a possibly empty singular
locus). Consider a closed cooriented hypersurface S ⊂ V \ Σ1 (f ) together
with a vector field μ normal to f (V ) in W . Choose a tubular neighborhood
U = S × [−1, 1] ⊂ V \ Σ1 (f ) of S = S × 0 such that the vector field ∂t ∂
|S
defines the coorientation of S in V , and extend the embedding f |U to an

embedding f : U × [−1, 1] → W such that f|U ×0 = f and ∂∂sf |U ×0 = μ. Here
we denote by t and s the coordinates corresponding to the second factors in
S × [−1, 1] and U × [−1, 1], respectively.
µ+
S
S
µ+
µ+
S
S
µ+
Figure 15.8. Sharp pleating.
Consider an embedding γ : [−1, 1] → [−1, 1] × [−1, 1] with sharp fold sin-

gularities as shown in Figure 15.8. We assume γ(t) = (t, 0) near t = ±1.
Consider the embedding f : V → W with sharp fold singularities which
coincides with f on V \ U and is given by f (p, t) = f(p, γ(t)) on U . We say
that the sharp folded embedding f : V → W is obtained by sharp pleating
of V with basis (S, μ). Note that the operation of pleating can be performed
in an arbitrarily small neighborhood of S. Moreover, it can be done to make
the corresponding Gaussian map Gdf : V → Grn W arbitrarily C 0 -close to
Gdf . The sharp fold of f consists of two parallel copies of S. Note that
switching the coorientation of S results in replacing the upper picture in
Figure 15.8 by the lower one, and replacing μ by −μ has the same effect.
The rounding of a wrinkle operation, which was discussed in the context of

equidimensional maps in Section 14.1, adapts in a straightforward way to
the context of embeddings with sharp wrinkles. The only additional thing
one needs to do is to choose an extension of the principal normal vector field
μ+ from the boundary ∂B of the core B of the wrinkle to its neighborhood
in V as a vector field normal to f (V ) in W , and use it for the construction
of the unfolding function.
Remark. The rounding of a sharp wrinkle close to an embryo results in

a sharp double fold which is equivalent to the result of sharp pleating with
basis (∂B, μ+ ).
The rounding of sharp wrinkles can be done continuously depending on

parameters, so a sharp wrinkled isotopy (see Section 15.1.D) can be modified
into an isotopy of sharp folded embeddings. Here each wrinkle with core Bi
is replaced by a sharp double fold which bounds an annulus neighborhood
of Bi ∪ Δi , where Δi is the (n − 1)-disc used in the rounding operation.
B. From sharp wrinkles to sharp double folds. Theorem 15.2.1 implies

the following result about sharp folded embeddings.
15.6.1. (Approximation of tangential rotations by sharp folded
embeddings) Let f : V → W be a sharp folded embedding and Gu : V →
Grn W , u ∈ [0, 1], its tangential rotation. Then there exist
• a sharp folded embedding f : V → W obtained from f by a sequence
of pleatings, and
• a diffeotopy Hu : W → W supported in Op f (V ) with H0 = Id
such that the homotopy Gd(Hu ◦ f ) : V → Grn W is arbitrarily C 0 -close to
Gu . If the rotation Gu is fixed on a neighborhood of a closed subset A ⊂ V ,
then we can arrange f = f and Hu = Id on Op f (A). In addition, we can
arrange that the basis (S, μ) of each sharp pleating consists of the boundary
sphere S of an embedded n-ball B ⊂ V \ Σ1 (f ) and the vector field μ extends
to a vector field normal to f (B) in W .
Proof. We will assume that the principal normal line bundle μ to V = f (V )

over Σ = f (Σ(f )) is trivial and leave the necessary adjustments for the case
of a nontrivial μ to the reader. For the applications in the next chapter only
the case of a trivial μ will be necessary. Let ϕ : U = Σ × (−ε, ε) × (−ε, ε) →
W be the embedding constructed in Lemma 15.1.3 with U := ϕ(U ) ⊃
V ∩ Op W Σ. Denote U0 = Σ × (−ε, ε) × 0 and U0 = ϕ(U0 ) ⊂ U . By
construction we have T U0 |Σ = T V |Σ . There exists a homotopy of sections
Gu : Op V → Grn W such that Gu (x) = Gu (f (x)), x ∈ V . So we can
apply Theorem 4.3.1 to construct an isotopy hu : U0 → Op W U0 such that
h0 = ϕ|U0 and Gd(x,t) hu is C 0 -close to Gu (x, t) for (x, t) ∈ U0 , u ∈ [0, 1].
There exists a diffeotopy αu : Op W V → Op W V which begins with α0 = Id
such that αu ◦ ϕ|U0 = hu , u ∈ [0, 1].
Recall that by construction we have f (Op V Σ) ⊂ U = ϕ(U ). Define a sharp
folded isotopy H u = αu ◦ f , u ∈ [0, 1].
u : Op V Σ → Op W U by setting H
Note that over a sufficiently small neighborhood Ω = Op V Σ the homotopy
GdH u approximates Gu . Define a tangential rotation G u : V → Grn W by
−1
the formula Gu (x) = (dαu ) Gu (αu (x)), where denotes the action of a
linear map on the Grassmanian. On Ω the tangential rotation G u is C 0 -close
to a constant one, hence by an additional adjustment we can arrange that
G u |Ω = G 0 |Ω .
Theorem 15.2.1 applied to f |V \Ω allows us to construct a sharp wrinkled
homotopy fu : V → W , starting at f0 = f , which is fixed on Op ∂Ω such
u . Recall that the definition of a sharp wrinkled
that Gdfu is C 0 -close to G
homotopy in the case when f0 is a genuine embedding implies that all wrin-
kles are born as embryos for certain values of the parameter u, live as sharp
V V
I I
Figure 15.9. Making wrinkles immortal.
wrinkles for certain periods of time, and either continue their existence until
the time u = 1 or disappear passing again through embryo states. Let us
modify fu by extending the existence of all sharp wrinkles to all u ∈ [0, 1]
in a close-to-embryo state. We call this procedure making the wrinkles im-
mortal, see Figure 15.9. As a result, we get a sharp wrinkled isotopy. Note
that all newly created wrinkles at u = 0 can be made arbitrarily close to
embryos, and in particular removable by a modification in their disjoint
neighborhoods.
By inductively applying the rounding of sharp wrinkles procedure to wrin-
kles of depth 1, 2, etc., we modify the sharp wrinkled isotopy fu to an
isotopy fu of sharp folded embeddings. Taking into account the Remark
in part A of this section, we conclude that f0 is obtained from f0 by a
sequence of sharp pleatings. Finally, we define the required sharp folded
isotopy Hu := αu ◦ fu : V → W .
C. Submanifolds folded with respect to a foliation. Let W be a

manifold of dimension n + k endowed with an n-dimensional foliation F ,
and V ⊂ W a submanifold of dimension k which is folded with respect to
F , i.e., it has only Σ1,0 type tangency singularities with respect to F , see
Section 11.3. We denote its fold locus by Σ.
The procedures of sharpening and smoothing of folds allow us to convert an
embedding which is folded with respect to a foliation F into an embedding
with sharp folds which is transverse to F , and vice versa. The procedures
are illustrated in Figure 15.10, and we leave the formal definition to the
reader. Note that under the conversion the principal normal line bundle
along the sharp fold Σ corresponds to the line bundle T F ∩ T V which is
tangent to V along Σ.
Example (Pleating). The above conversion construction transforms a
sharp pleating to a pleating of a submanifold with respect to a foliation.
As above, let V k ⊂ W k+n be a submanifold folded with respect to an
Figure 15.10. Smoothing of a sharp fold.
Q
Γ Γ̃
Figure 15.11. Pleating vs sharp pleating.
n-dimensional foliation F with fold Σ. Consider a closed cooriented hy-

persurface S ⊂ V \ Σ and a unit vector field μ+ ∈ T F |S . Denote Q =
[−1, 1] × [−1, 1] with coordinates (t, s), and let be the foliation of Q by the
fibers of the projection (t, s) → t. Pick a tubular( neighborhood S × Q ⊂ W
∂ (
of S × 0 × 0 = S such that S × [−1, 1] × 0 ⊂ V , ∂t S
defines the coorientation
of S in V , ∂s |S = μ, and F intersect the fibers x × Q for x ∈ S along the
∂
foliation .
Consider a curve Γ ⊂ Q with two fold tangencies to as shown in Fig-
ure 15.11. This is just a smoothing of the curve Γ used for the sharp pleat-
ing. Note that the two fold points have opposite characteristic vectors. By
replacing S ×[0, 1]×0 with S ×Γ, we get a manifold V folded with respect to
F with two parallel copies of S as additional components of the fold locus.
We say that V is obtained by pleating of V with basis (S, μ). Similarly to the
sharp pleating, the pleating can be performed in an arbitrarily small neigh-
borhood of S in W . Note that switching the coorientation of Σ switches
the direction of the characteristic vector field along the newly created fold
components.
15.6.2. (Pleated isotopy) Suppose V k ⊂ W k+n is folded with respect to
an n-dimensional foliation F along Σ ⊂ V . Let Ft , t ∈ [0, 1], be a homotopy
of foliations on W beginning with F0 = F , and Ω ⊂ W a neighborhood of
V . Then there exists a manifold V ⊂ Ω obtained from V by a sequence of
successive pleatings, and a diffeotopy ht : W → W supported in Ω such that
h∗t Ft = F on Op V . If the homotopy Ft is fixed on OpA ⊂ W for a closed
subset A ⊂ V , then one can arrange that V ∩ Op A = V ∩ Op A and the

isotopy ht is fixed on Op A. Moreover, any manifold obtained from V by
additional pleating admits a diffeotopy ht with the above properties.
Proof. By the sharpening of folds operation we convert the inclusion f :

V → W into a sharp folded embedding transverse to F . There exists a
tangential rotation Gt : V → Grk W starting from G0 = Gdf which is
transverse to T Ft |V . By applying Theorem 15.6.1 we construct a sharp
folded embedding f : V → W and a diffeotopy Ht : W → W such that
f is C 0 -close to f , Gdf is C 0 -close to Gdf , and Gd(Ht ◦ f ) is C 0 -close to
Gt . If the approximation is good enough, then Gd(Ht ◦ f ) is transverse to
Ft . By the smoothing of sharp folds construction we can deform the sharp
folded embedding f to a smooth embedding f which is folded with respect
to F and obtained from f by successive pleatings. Then the isotopy Ht ◦ f is
folded with respect to Ft . Equivalently, the submanifold V := f(V ) is folded
with respect to Ft := Ht∗ Ft . According to Proposition 11.3.1 there exists a
diffeotopy gt : Op V → Op V such that g0 = Id, gt |V = IdV , and gt∗ Ft = F .
We extend the diffeotopy gt to the whole manifold W as the identity outside
Ω. Then the diffeotopy ht := Ft ◦ gt has the required properties.
Chapter 16
Igusa Functions
As we already discussed in Chapter 11, a generic function has only nondegen-

erate (also called Morse, or A1 ) critical points, while a generic 1-parametric
family can also have at isolated values of the parameter embryonic (also
called birth-death, or A2 ) critical points. A function with only Morse and
embryonic critical points is called a generalized Morse function, or, as we
will call it in this book, an Igusa function, in honor of Kiyoshi Igusa.
Note that a wrinkled function W → R, see Section 13.1.C, is an example of
an Igusa function endowed with an additional structure: all its Morse critical
points are organized into pairs of points of neighboring indices connected by
a gradient-like trajectory. Theorem 13.2.4 which we proved in Chapter 13
is a foliated h-principle for such functions.
In this chapter we prove two other versions of a full relative parametric h-
principle for Igusa functions. They generalize Igusa’s results from [Ig84]
and [Ig87], see also [EM00, Lu09, EM12]. The foliated form of the cor-
responding h-principles allows us to avoid discussing the topology of the
corresponding spaces of Igusa functions and their formal analogues.
16.1. Leafwise Igusa functions and their formal analogues

A. Leafwise Igusa functions and associated data. Let W be an (n+k)-
dimensional manifold with an n-dimensional foliation F . A function ϕ :
W → R is called a leafwise Igusa function if restricted to leaves it has
only A1 (i.e., Morse) or A2 (i.e., birth-death) critical points. We denote by
T F ⊂ T W the subbundle tangent to the leaves of the foliation F and use
a Riemannian metric on W to identify T F with its dual T ∗ F . We review
193
194 16. Igusa Functions
below some of the notions which were already associated to a leafwise Igusa
function in Section 11.6.C.
Given a leafwise Igusa function ϕ : W → R consider the following data:
• the leafwise differential dF ϕ : W → T ∗ F of the function ϕ, which
we identify via the Riemannian metric with the leafwise gradient
∇F ϕ ∈ T F ;
• the set V = V (ϕ) = dF ϕ−1 (0) ⊂ W of leafwise critical points;
• the restricted bundle Vert := T F |V ;
• the leafwise quadratic differential d2F ϕ of ϕ, which is a quadratic
form on Vert; we use the Riemannian metric to view d2F ϕ as a field
of self-adjoint operators Vert → Vert;
• the set Σ ⊂ V of leafwise A2 -points of ϕ, i.e., points where d2F ϕ has
a one-dimensional kernel; by assumption V \Σ consists of A1 -points,
i.e., points where d2F ϕ is nondegenerate;
• the kernel line bundle λ = ker(d2F ϕ|Σ ) ⊂ Vert|Σ ;
• the third leafwise differential d3F ϕ : λ → R, which is a cubic form
on λ; by assumption this cubic form is nonvanishing, and therefore,
the bundle λ is trivial and canonically oriented by d3F ϕ; we denote
by λ+ the unit vector field in λ defining its orientation.
According to Lemma 11.6.5, for a generic leafwise Igusa function ϕ the

subset V is a k-dimensional submanifold of W which has fold type tangency
to F along its (k − 1)-dimensional submanifold Σ ⊂ V . The line bundle λ
along Σ is tangent to V and coincides with Vert ∩ T V |Σ .
The index of the leafwise quadratic differential d2F ϕ (v) at v ∈ V may take
the values 0, 1, . . . , n for v ∈ V \ Σ, and 0, 1, . . . , n − 1 for v ∈ Σ. Let
V \ Σ = V 0 ∪ ···∪ V n and Σ = Σ0 ∪ · · · ∪ Σn−1
be the decompositions of V \ Σ and Σ according to the index. Note that Σi

is the intersection of the closures of V i and V i+1 .
For v ∈ V \ Σ we have the canonical splitting
Vert(v) = Tv F = Vert+ (v) ⊕ Vert− (v),
where Vert+ (v) and Vert− (v) are the positive and negative eigenspaces of
d2F ϕ(v), and for σ ∈ Σ we have the canonical splitting
Vert(σ) = Ver(σ) ⊕ λ(σ) = Ver+ (σ) ⊕ Ver− (σ) ⊕ λ(σ)

16.1. Leafwise Igusa functions and their formal analogues 195
0
V
Σ0
V1
Figure 16.1. Framed leafwise Igusa function.
(Ver = Vert!), where Ver+ (σ) and Ver− (σ) are the positive and negative
eigenspaces of d2F ϕ (σ). For σ ∈ Σi we have
lim Vert+ (v) = Ver+ (σ) ⊕ λ(σ) and lim Vert− (v) = Ver− (σ),
V i v→σ V i v→σ
lim Vert− (v) = Ver− (σ) ⊕ λ(σ) and lim Vert+ (v) = Ver+ (σ).
V i+1 v→σ V i+1 v→σ
A framing of a leafwise Igusa function ϕ is an ordered set ξ = (ξ 1 , . . . , ξ n )

of unit vector fields in Vert such that:
(F1) ξ i is defined (only) on the union Σi−1 ∪ V i ∪ · · · ∪ Σn−1 ∪ V n ;
(F2) ξ i |Σi−1 = λ+ |Σi−1 ;
(F3) (ξ 1 , . . . , ξ i )|V i is an orthonormal basis for Vert− |V i .
In particular, ξ n is defined only on Σn−1 ∪V n and ξ 1 is defined only on V \V 0 .
The pair (ϕ, ξ) is called a framed leafwise Igusa function; see Figure 16.1.
Example (The standard wrinkle as a leafwise Igusa function). Recall the
standard wrinkle

wn+k,k+1,s : Rk+n = Rk × R × Rn−1 → Rk+1 , (y, z, x) → (y, ϕ(y, z, x)
with the function

s
n−1
ϕ(y, z, x) = z 3 + 3(|y|2 − 1)z − x2i + x2j .
i=1 j=s+1
Thus ϕ : Rk+n → R is a leafwise wrinkled function with respect to the

foliation by the fibers of the projection (y, z, x) → y. This is a special case
of a leafwise Igusa function with
V (ϕ) = {|y|2 + z 2 = 1, x = 0} = V s ∪ V s+1 ,
V s = V (ϕ) ∩ {z > 0}, V s+1 = V (ϕ) ∩ {z < 0},
Σ(ϕ) = {|y| = 1, x = 0, z = 0}.
Denote x− = (x1 , . . . , xs ) and x+ = (xs+1 , . . . , xn−1 ). Then the bundle

Vert− |V s is the restriction to V s of the fibration (y, z, x− , x+ ) → (y, z, x+ ),
and Vert− |V s+1 is the restriction to V s+1 of the fibration (y, z, x− , x+ ) →
∂
(y, x+ ). The vector field λ+ over Σ coincides with ∂z . The vector fields ξ i =
∂ s i ∂
∂xi , i = 1, . . . , s, form a framing of ϕ over V , while ξ = ∂xi , i = 1, . . . , s
∂
and ξ s+1 = ∂z form a framing over V s+1 .
Given any leafwise wrinkled function f : W → R on a foliated manifold
(W, F ), its singularities are contained in disjoint balls Bj such that the
restrictions f |Bj are equivalent to standard wrinkles. Hence, any leafwise
wrinkled function can be framed.
B. Formal leafwise Igusa functions (FLIFs). A formal leafwise Igusa

function (FLIF ) is a triple Φ = (Φ1 , Φ2 , λ+ ) where:
• Φ1 : W → T F is a vector field tangent to F which vanishes on a
subset V = V (Φ) ⊂ W ;
• Φ2 is a self-adjoint operator Vert = T F |V → Vert which has rank
n − 1 over a subset Σ = Σ(Φ) ⊂ V and rank n over V \ Σ;
• λ+ is a unit vector field in the line bundle λ := Ker(Φ2 ) ⊂ Vert|Σ
over Σ.
The leafwise 3-jet of a genuine leafwise Igusa function ϕ can be viewed as a
formal leafwise Igusa function Φ where Φ1 = ∇F ϕ is the leafwise gradient
vector field, Φ2 = d2F ϕ, and λ+ is the unit vector field in Ker(d2F ϕ|Σ ) oriented
by the third differential d3F ϕ. We denote this FLIF by J(ϕ). A FLIF Φ of
the form J(ϕ) is called holonomic. Hence, genuine leafwise Igusa functions
can be viewed as holonomic FLIFs. Usually we will not distinguish between
leafwise Igusa functions and the corresponding holonomic FLIFs.
For a FLIF Φ we define the splitting of Vert exactly as in the holonomic
case, with d2F ϕ replaced by Φ2 . A framing ξ for Φ is also defined as in the
holonomic case, and such a pair (Φ, ξ) is called a framed formal leafwise
Igusa function (framed FLIF ).
As in the holonomic case, for a generic FLIF Φ the set V ⊂ W is a k-
dimensional manifold and Σ ⊂ V is its (k − 1)-dimensional submanifold.
However, Σ has nothing to do with tangency of V to F ; see Figure 16.2.
Moreover, there is no control of the type of the tangency singularities be-
tween V and F .
The notion of a FLIF can be generalized without any changes if we replace
the foliation F by an arbitrary (not necessarily integrable) n-dimensional
distribution ζ ⊂ T W . We will refer to the corresponding object as a ζ-
FLIF. We call a ζ-FLIF locally holonomic if ζ is integrable on Op V and the
restriction of Φ to Op V is holonomic.
16.1. Leafwise Igusa functions and their formal analogues 197
0
Σ V
1
0
Σ0
V
Σ0 0
V
1
V Σ0
Figure 16.2. Framed FLIF.
C. Two h-principles for Igusa functions. According to the general phi-

losophy of the h-principle, the parametric h-principle for (framed) Igusa
functions should state that a (framed) FLIF is homotopic to a framed holo-
nomic one, and moreover, the homotopy can be chosen fixed on a neighbor-
hood of a closed subset where the FLIF is already holonomic. It turns out,
however (see Theorem 16.2.1), that framed FLIFs can always be extended
(i.e., in the framed case the existence of a formal solution is automatic), and
we get an unconditional statement rather than a classical h-principle. This
explains the role of the framing.
Here are the main results of this chapter.
Let W be an (n + k)-dimensional manifold endowed with an n-dimensional
foliation F , and A ⊂ W a (possibly empty) closed subset.
16.1.1. (Extension of framed leafwise Igusa functions) Let (ϕA , ξA )
be a framed leafwise Igusa function on Op A. Then there exists a framed
leafwise Igusa function (ϕ, ξ) on the whole W which coincides with (ϕA , ξA )
on Op A.
16.1.2. (h-principle for leafwise Igusa functions) Let Φ be a FLIF on
W which is holonomic over Op A. Then there exists a homotopy of FLIFs
fixed on Op A connecting the FLIF Φ with a holonomic one.
Theorems 16.1.2 and 16.1.1 were proved in [EM00] and [EM12], respec-
tively. They are generalizations of the results of K. Igusa in [Ig87] and
[Ig84], respectively, where equivalent versions of Theorems 16.1.2 and 16.1.1
were proved under the assumption that the codimension of the foliation is
less than its dimension.
Theorem 16.1.1 is equivalent to the fact that the space of framed Igusa
functions is contractible, which is the content of J. Lurie’s extension [Lu09,
Theorem 3.4.7] of K. Igusa’s result from [Ig87]. We note that the homo-
logical forms of both theorems follow from V. Vassiliev’s results in [Va92].
S. Kupers proved in [Ku19] a stronger form of Vassiliev’s theorem which
imply both theorems. As was already stated, the above foliated form of
the h-principle allows us to avoid a discussion of the topology on the corre-
sponding spaces of Igusa functions, cf. [Ig87].
We will concentrate below on proving Theorem 16.1.1, and then explain the
necessary adjustments for Theorem 16.1.1. Our proof of Theorem 16.1.1
consists of four main steps. First, in Section 16.2 we show that there is
always a formal extension of a framed Igusa function, i.e., an extension
as a framed FLIF Φ. Next, in Section 16.3 we deform Φ rel. Op A to a
locally holonomic framed ζ-FLIF Φ for a distribution ζ which is homotopic
to F . Next, we construct a homotopy of distributions ζt between ζ0 = ζ
and ζ1 = F which are integrable on Op V (Φ), and show that the homotopy
can be covered by a homotopy Φ t of locally holonomic ζt -FLIFs. Finally,
we use the foliated h-principle for wrinkled functions in Theorem 13.2.4 to
make Φ 1 holonomic away from V (Φ 1 ).
16.2. Formal extension of framed ζ-FLIFs

16.2.1. (Formal extension) Any framed ζ-FLIF (Φ, ξ) on Op A ⊂ W
on the whole manifold W .
ξ)
extends to a framed ζ-FLIF (Φ,
The proof is essentially Igusa’s argument in [Ig87] (see pp. 438–442). Its
main step is the following lemma which will be used inductively.
16.2.2. (Decreasing the index) Let 1 ≤ j ≤ n. Suppose W is a cobor-
dism between ∂− W and ∂+ W , and for a framed ζ-FLIF (Φ, ξ) on W one
on W
ξ)
has V i = ∅ for all i > j. Then there exists a framed ζ-FLIF (Φ,
with ˜)
such that (decorating objects belonging to Φ
• Φ = Φ on Op (∂− W );
• Φ 1 = Φ1 , in particular V = V ;
• V i ∩ ∂+ W = ∅ for all i ≥ j.
Proof. Recall that the j-dimensional bundle Vert− |V j is trivialized by the

framing ξ = (ξ 1 , . . . , ξ j ), and ξ j |Σj−1 = λ+ . The vector field ξ j extends
canonically to a neighborhood G of V j ∪ Σj−1 in V j ∪ Σj−1 ∪ V j−1 as a
unit eigenfield of Φ2 , which we still denote by ξ j . Let X j : Vert|G → Vert|G
be the self-adjoint linear operator which orthogonally projects onto the line
bundle spanned by ξ j . Choose neighborhoods H± of ∂± W ∩ V in V with
disjoint closures and pick a cut-off function θ : V → [0, 1] which is equal to 0
on H− ∪ (V \ G) and equal to 1 on V j ∩ H+ . See Figure 16.3. For a constant
16.3. Constructing a locally holonomic ζ-FLIF 199
∼
Φ j1 Φ ∼ j1
V V
j1 ∼ j1
Σ Σ
j ∼j
V V
∼ j1
Σj1 Σ
j1 ∼ j1
H_ V G V
H+
Figure 16.3. Decreasing the negative index.
C > 0 we define the self-adjoint operator Φ 2 := Φ2 + CθX j : Vert → Vert.

Then for C sufficiently large Φ coincides with Φ2 on V ∩ H− , has index ≤ j
2
everywhere, and has index < j on H+ . Since the kernel of Φ 2 on Σ j−1 is

generated by ξ j , we can define the new vector field λ + as ξ j on Σ j−1 and
+ j−1
λ on Σ \ Σ . Note that Φ = Φ on V \ (V ∪ Σ j j−1 ∪V j−1 ), and we have
V ∪Σ
j j−1 ∪V j−1
= V ∪Σ
j j−1 ∪V j−1
. On V j−1 the new negative eigenspace
/ − is spanned by the vectors ξ 1 , . . . , ξ j−1 , and on V j it is spanned by
Vert
ξ 1 , . . . , ξ j . Hence, the framing ξ of Φ determines a framing ξ of Φ.
Proof of Theorem 16.2.1. Let Φ = (Φ1 , Φ2 , λ+ ). Let B ⊂ C ⊂ W be

closed domains with smooth boundary such that A ⊂ Int B, B ⊂ Int C, and
Φ is defined over C. We apply Lemma 16.2.2 with j = n to the cobordism
C \ Int B to find a framed ζ-FLIF (Φ , ξ ) on C with V n (Φ ) ∩ ∂C = ∅. Let
B ⊂ Int C be a closed domain with smooth boundary such that B ⊂ Int B
and V n (Φ ) ∩ (C \ Int B ) = ∅. We again apply Lemma 16.2.2 with j = n − 1
to the cobordism C \ Int B to find (Φ , ξ ) on C with V i (Φ ) ∩ ∂C = ∅ for
i ≥ n − 1. Continuing inductively, we construct a framed ζ-FLIF (Φ, on
ξ)

C which agrees with (Φ, ξ) on B and satisfies V (Φ) ∩ ∂C = ∅ for all i > 0,
i
2|
i.e., Φ is positive definite. Now we extend Φ 1 in any generic way to
V ∩∂C
W , and then extend Φ 2 as a positive definite operator on Vert (which is
possible because the space of positive definite operators is contractible).
In what follows we will always assume that all considered FLIFs are holo-
nomic on Op A and all constructions and homotopies keep Φ|Op A fixed.
16.3. Constructing a locally holonomic ζ-FLIF

The goal of this section is the proof of the following proposition.
Φ
Vert
1
TΦ
Φ1
V W
Figure 16.4. Linearization of the section Φ1 .
16.3.1. Let W be an (n + k)-dimensional manifold with an n-dimensional

foliation F and A ⊂ W a closed subset. Let (Φ, ξ) be a framed FLIF which
is holonomic on Op A. Then there exists an n-dimensional distribution ζ ⊂
T W and a locally holonomic framed ζ-FLIF (Φ, ξ) such that
• ζ and (Φ, ξ) coincide on Op A with T F and (Φ, ξ), respectively;
• ζ is homotopic to T F relative Op A.
A. More structures associated to a FLIF. Let us introduce some addi-

tional objects and notions associated with a FLIF. Denote by Norm the nor-
mal bundle to V in W . For each point v ∈ V the tangent space to the graph
of the section Φ1 : W → T F is the graph of a linear map Tv W → Vertv ; see
Figure 16.4. Transversality of the section Φ1 : W → T F to the zero section
means that this map is an epimorphism with kernel Tv V , so it restricts to
an isomorphism Normv → Vertv . Thus we get a bundle isomorphism
ΓΦ : Norm → Vert.
The splitting Vert|Σ = Ver ⊕ λ induces via ΓΦ a splitting
Norm|Σ = Nor ⊕ Rn+ , n+ := (ΓΦ )−1 (λ+ ), Nor := (ΓΦ )−1 (Ver).
We denote by τ + the unit normal vector field to Σ in V which defines its

index coorientation, which means that it points in the direction of decreas-
ing index. Thus, on Σ i it points into V i ; see Figure 16.5. We denote by τ
the line bundle over Σ generated by τ + .
The unit vector field λ+ ∈ Vert|Σ , which generates the kernel of Φ2 |Σ , admits
+ as a unit eigenvector field for the self-adjoint operator
a unique extension λ
Φ to a sufficiently (small tubular neighborhood U = Σ × [−ε, ε] of Σ in V .
2
∂(
We assume that ∂t Σ
= τ + , where t is the coordinate corresponding to the
n+
Vi+1
Σi
τ+
i
V
Figure 16.5. The vector fields n+ and τ + .
+
/ ⊂ Vert|U be the orthogonal complement to λ
second factor in U . Let Ver
/
in Vert|U , so that Ver|Σ = Ver. Thus we get a splitting
+ ,
/ ⊕ Rλ
Vert|U = Ver
which induces via ΓΦ a splitting
/ ⊕ R
Norm|U = Nor n+ , + ),
+ := (ΓΦ )−1 (λ
n / := (ΓΦ )−1 (Ver).
Nor /
Denote Q := [−ε, ε] × [−ε, ε] with coordinates (t, s). We extend U = Σ ×

[−ε, ε] to a split tubular neighborhood
U = N × Q = N × [−ε, ε] × [−ε, ε]
of Σ in W , where N is a tubular neighborhood of Σ in the (total space of the
∂ (
bundle Nor. We assume that U = Σ × [−ε, ε] × 0 ⊂ V , ∂s Σ
= n+ , and the
/ coincides with the tangent bundle to the normal fibers of the
bundle Nor
tubular neighborhood N along U .
2
Consider the foliation of R2 by the parabolas s = 4tε +C, C ∈ R, and denote
by the induced foliation on Q = {|t|, |s| ≤ ε} ⊂ R2 ; see Figure 16.6. Denote
by N the foliation of N by normal fibers, and by LΦ the n-dimensional
product foliation N × on U = N × Q.
Set ∂± U := Σ × {±ε}. By adjusting the Riemannian metric, we can arrange
that for a sufficiently small tubular neighborhood Ω ⊃ V in W the leaves
of LΦ through points of Op (∂± U ) ⊂ V intersect Ω along the normal fibers
of the tubular neighborhood Ω. We extend LΦ to Ω \ U as the foliation by
normal fibers of the tubular neighborhood Ω ⊃ V .
Note that V is folded with respect to LΦ with fold Σ and the characteristic
vector field of the fold equal to n+ . In particular, LΦ is transverse to V
away from Σ.
The restriction of the tangent bundle T LΦ to V will be called the virtual
vertical bundle and denoted by VertΦ ; see Figure 16.7. It coincides with
l ε
−ε
−ε
Figure 16.6. The foliation .
Φ
Norm + Vert
n
Σ τ+
Figure 16.7. The virtual vertical bundle.
Norm over V \ U , and over U ⊂ V we have

/ ⊕ R ∂ 8t ∂
VertΦ |U = Nor τ +, τ+ := + .
∂t ε ∂s
According to our assumption, the vector field τ+ is orthogonal to V near
∂U , hence τ+ = ±
n+ near ∂± U . We also notice that the eigenvalue function
c : U → R for the eigenvector field λ + ) = cλ
+ , Φ2 (λ + , is positive near ∂+ U
and negative near ∂− U . By slightly adjusting the FLIF Φ, we can arrange
that the eigenvalue function c is equal to ±1 near ∂± U .
With these adjustments we define an isomorphism
ΔΦ : Vert → VertΦ
by the formula
⎧ −1
⎪
⎨Γ (Φ (Z)) over V \ U, Z ∈ Vert,
Φ 2
−1 /
(17) ΔΦ (Z) = ΓΦ (Φ2 (Z)) over U, Z ∈ Ver,
⎪
⎩ +
τ + .
over U, Z = λ
The following lemma is a straighforward consequence of the semilocal normal
forms Lemma 11.2.2 and Theorem 11.6.6 near the loci of leafwise A1 and
A2 singular points. We leave the details to the reader.
16.3.2. (Holonomic case) If Φ is holonomic, then for appropriate auxil-

liary choices we have LΦ = F |Op V , VertΦ = Vert, and ΔΦ = Id.
B. Making a framed FLIF balanced. A FLIF Φ is called balanced if

the composition
ΔΦ
Vert −→ VertΦ → T W |V
is homotopic rel Op A ∩ V to the inclusion Vert → T W |V through injective

homomorphisms.
The following stabilization procedure allows us to make any FLIF Φ bal-

anced. Given a FLIF Φ, consider a compact connected domain Ω ⊂ V \ Σ
with smooth boundary such that the bundles Vert± |Ω are trivial. Let
C ⊂ V \ Σ be an exterior collar of ∂Ω and set Ω := Ω ∪ C.
Let us assume first that Ω is contained in V i with i < n. We choose a
nowhere vanishing section θ of the bundle Vert+ over Ω and define the
= St− (Φ) with the following
negative stabilization of Φ over Ω as a FLIF Φ Ω,θ

properties (we decorate all objects belonging to Φ with):
1 = Φ1 and thus V = V ;
• Φ
2 = Φ2 over V \ Ω ;
• Φ
= Σ ∪ ∂Ω and Int Ω ⊂ V i+1 ;
• Σ
/ − |Int Ω = Span(Vert− |Int Ω , θ);
• Vert
+ |∂Ω = θ.
• λ
We will omit the reference to θ in the notation and write simply St− Ω (Φ)
when this choice is irrelevant.
In order to construct Φ 2 on Ω with these properties, we adjust the back-
ground metric on W to make θ an eigenvector field for Φ2 with eigenvalue
+1. The vector field θ remains an eigenvector field for Φ 2 but the eigenvalue

function changes to c : Ω → [−1, 1], where c is negative on Int Ω, equal to 1
near ∂Ω , and has ∂Ω as its regular 0-level set.
can be canonically
If the ζ-FLIF Φ is framed by ξ = (ξ 1 , . . . , ξ n ), then Φ
framed by ξ such that ξ = ξ over V \ Ω and Vert− |Int Ω is framed by ξ :=
/
(ξ 1 , . . . , ξ i , θ). We denote this framed FLIF by
−
St−
Ω (Φ, ξ) := StΩ, θ (Φ), ξ .
In the cases when Ω ⊂ V i and i > 0, we can similarly choose a section θ of

Vert− over Ω , and define the positive stabilization of Φ over Ω as a FLIF
= St+ (Φ) with the following properties:

Φ Ω, θ
1 = Φ1 and thus V = V ;
• Φ
2 = Φ2 over V \ Ω ;
• Φ
= Σ ∪ ∂Ω and Int Ω ⊂ V i−1 ;
• Σ
/ + |Int Ω = Span(Vert+ |Int Ω , θ);
• Vert
+ |∂Ω = θ.
• λ
If Φ is framed by a framing ξ = (ξ 1 , . . . , ξ n ), then we will always choose
θ = ξ i |Ω and define the framed positive stabilization by
St+ +
Ω (Φ, ξ) := (StΩ, ξ i (Φ), ξ),
Int Ω := (ξ 1 , . . . , ξ i−1 ).
where ξ|
16.3.3. (Balancing via stabilization) Any FLIF Φ can be stabilized to a
balanced one. If Φ is already balanced and χ(Ω) = 0, then St±
Ω (Φ) is balanced
as well. The statement holds also in the relative form.
Proof. Fix positive integers k, n. Let Vn (E) denote the Stiefel manifold of
n-frames in a vector space E. We need the following well-known facts about
the homotopy groups of Stiefel manifolds (see, e.g., [St51]):
(i) πi Vn (Rk+n ) = 0 for each i ≤ k − 1;
(ii) πk Vn (Rn+k ) = Z if k is even or n = 1, and Z/2 otherwise;
(iii) the canonical map πk Vn−1 (Rn+k−1 ) → πk Vn (Rn+k ) is an isomor-
phism if n > 2, and surjective if n = 2.
Note that, since V1 (Rk+1 ) S k , the surjective map πk V1 (Rk+1 )→πk V2 (Rk+2 )
in the case n = 2 of (iii) is the identity map Z → Z if k is even, and the
projection Z → Z/2 if k is odd.
Recall that dim V = k, dim W = k + n, and the bundle Vert → V has
rank n. Thus it follows from (i) and (ii) above that the obstruction to
the existence of a homotopy fixed over A ⊂ V between two monomor-
phisms Ψ1 , Ψ2 : Vert → T W |V is a cohomology class δ(Ψ1 , Ψ2 ; V, A) ∈
H k (V, A; πk Vn (Rn+k )), or more precisely a cohomology class with coeffi-
cients in the local system πk Vn (Tv W ), v ∈ V . Recall the homomorphism
ΔΦ : Vert → VertΦ → T W |V defined in equation (17). We claim that

Φ St± (Φ) χ(Ω)Θ if k is even,
(18) δ(Δ , Δ Ω ; Ω , ∂Ω ) =
±χ(Ω)Θ if k is odd
for an appropriate choice of a generator Θ of H k (Ω , ∂Ω ; πk Vn (Rn+k )) ∼

= Z,
where the equality has to be understood mod 2 if k is odd and n > 1.
1 1 1
f+ ~ f
0 0 0
U’ U’ U’
Figure 16.8. Positive and negative stabilization.
To prove (18), we first consider a positive stabilization. So we are given a

framing ξ = (ξ 1 , . . . , ξ n ) of Vert|Ω such that θ = ξ i is used for the stabiliza-
tion Φ = St+ (Φ). Taking the images of ξ, we view ΔΦ and ΔΦ as sections
Ω,θ
Ω → Vn (T W |Ω ) agreeing on ∂Ω . By (iii) above, the obstruction to a homo-

topy rel. ∂Ω from ΔΦ to ΔΦ arises from the case n = 1, given by considering
only the images of θ in V1 (T Ω ⊕ R). The map f0 : Ω → V1 (T Ω ⊕ R) corre-
sponding to ΔΦ is constant equal to the unit vector n + in the R-direction.
According to formula (17) and Figure 16.7, the map f1 Ω → V1 (T Ω ⊕ R)

corresponding to ΔΦ rotates n + to − n+ over the collar neighbourhood

C = Ω \ Ω via v → c(v) (sin θ(v) +
n + cos θ(v) τ + ), where τ+ is the extension
over C of the inward unit normal vector to Ω along ∂Ω (defining the index
coorientation). The obstruction to a homotopy rel. ∂Ω between these two
maps is the obstruction to the extension of f + : ∂(Ω × [0, 1]) → V1 (T Ω ⊕ R)
given by f + (v, 0) = f0 (v), f + (v, 1) = f1 (v), and f + (v, t) = n + for v ∈ ∂Ω

to a continuous section Ω × [0, 1] → V1 (T Ω ⊕ R). Figure 16.8 shows that
f + is homotopic to an inward pointing vector field along ∂(Ω × [0, 1]), so by
the Poincaré–Hopf theorem the obstruction to extending it equals its index
ind(f + ) = χ(Ω × [0, 1]) = χ(Ω).
In the case of a negative stabilization, an analogous discussion leads to a

map f − : ∂(Ω × [0, 1]) → V1 (T Ω ⊕ R) defined in the same way as f + , only
with τ replaced by −τ on the collar neighbourhood C; see Figure 16.8. Thus
f = R◦f with the bundle map R : V1 (T Ω ⊕R) → V1 (T Ω ⊕R) defined by
− +
R(x, τ ) = (−x, τ ). Since the map V1 (Rk ⊕ R) = S k → S k , (x, τ ) → (−x, τ )

has degree (−1)k , the obstruction is ind(f − ) = (−1)k ind(f + ) = (−1)k χ(Ω).
This concludes the proof of (18).
Equation (18) first shows that stabilization over a domain with vanishing
Euler characteristic does not change the obstruction class δ(J, ΔΦ ), where
J : Vert → T W |V denotes the inclusion. On the other hand, with the
exception of the case k = n = 1, this obstruction class can be changed in an
arbitrary way by an appropriate choice of Ω. Indeed, if k > 1, then one can
take for Ω either a disjoint union of balls or a regular neighborhood of an
embedded bouquet of circles to increase or decrease the obstruction by

(cf. similar arguments in [EGM11, CE12]). If k = 1 and n > 1, the sign
issue is irrelevant because the obstruction is Z/2-valued.
In the case k = n = 1 one may need two successive stabilizations in order
to balance a given ζ-FLIF. Indeed, the domain Ω in this case is a union of
some number of intervals, hence χ(Ω) = . Thus the positive stabilization
increases the obstruction class by , while the negative one decreases it by
. Suppose, for determinacy, that we want to stabilize over a domain in V 0 .
To change the obstruction class by − , we just negatively stabilize over the
union of intervals. To change it by + , we first negatively stabilize over
one interval I, and then positively stabilize over the union of + 1 disjoint
intervals in I.
C. Proof of Proposition 16.3.1. After applying Lemma 16.3.3 (which

does not change the foliation F ), we may assume that the framed FLIF
(Φ, ξ) is balanced. This implies that the isomorphism ΔΦ : Vert = T F |V →
VertΦ extends to a monomorphism T F → T W which is homotopic to the
inclusion T F → T W . Keeping the notation ΔΦ for the extension, we denote
ζ := ΔΦ (T F ). Note that ζ is homotopic to T F as distributions. By the
construction of VertΦ , we can additionally arrange that ζ|Op V = T LΦ .
We define a locally holonomic framed ζ-FLIF (Φ, ξ) as follows. Consider the
field of quadratic forms on the bundle Vert defined by
qv (X) := Φ2v (X), X, v ∈ V, X ∈ Vertv .
Recall the splitting U = N × Q of a neighborhood U ⊃ Σ in W which

was introduced in Section 16.3.A. Let (σ, X, t, s) where σ ∈ Σ, X ∈ Norσ ,
2
t, s ∈ [−ε, ε] be the corresponding coordinates. Denote u := s − 4tε , so
that the function u is constant on the leaves of the foliation LΦ . We will
use (σ, X, t, u) as new coordinates on U . In these coordinates we have U =
2
V ∩ U = {(σ, X, t, u) ∈ U | X = 0, u = − 4tε } and τ+ = ∂t
∂
.
Choose a positive function function b : R → R such that b(u)
= 1 for u> −ε
√ 3
√
εut
and b(u) = 1/ −u for u < −2ε. Denote ψ(u, t) := b(u) 3√ε + 2
2t
and
define a function ϕ : U → R by the formula
(19) (σ,t) (X)) + ψ(u, t), X ∈ Norσ , (σ, t) ∈ U,

ϕ(σ, X, t, u) = q(σ,t) (ΓΦ
where ΓΦ : Nor → Ver is the isomorphism defined above. The locus of LΦ -

leafwise critical points of ϕ coincides with U , while its leafwise quadratic
differential on VertΦ |U is equal to

1 ∂ 2ψ
d2(σ,t) ϕ(X + z
τ + ) = q(σ,t) (ΓΦ
(σ,t) (X)) + (u, t)z 2
2 ∂t2
2tz 2 4t2
(σ,t) (X)) + b(u) √ ,
= q(σ,t) (ΓΦ u=− .
ε ε
2
Recall that on ∂± U we have t = ±ε and thus u = − 4tε = −4ε < −2ε.
√
ε
Hence, near ∂± U we have b(u) = 2|t| and
d2(σ,t) ϕ(X + z (σ,t) (X)) ± z ,

τ ) = q(σ,t) (ΓΦ (σ, t) ∈ Op V ∂± U.
2
Let us extend the function ϕ|Op V U to any smooth function θ : V → R. Iden-

tifying a sufficiently small tubular neighborhood of V in W with a neigh-
borhood of the 0-section in the total space of the bundle Norm over V , we
define a function ϕ1 : Op V → R by the formula
(20) ϕ1 (v, X) = θ(v) + qv (ΓΦ
v (X)), v ∈ V, X ∈ Normv .
Let us compare the functions ϕ and ϕ1 on Op W (∂U ). Recall that near ∂U
the foliation LΦ coincides with the normal fibers of the tubular neighborhood
of V , and the bundle VertΦ coincides with Norm = Span(Nor, / τ+ ). The
functions ϕ and ϕ1 coincide on Op V ∂U , while on Op V ∂± U the leafwise
quadratic differential of ϕ1 ,
Φ
d2(σ,t) ϕ1 (X + z
τ + ) = q(σ,t) ΓΦ τ +)
(σ,t) (X) + q(σ,t) Γ(σ,t) (z

(σ,t) (X) ± z ,
= q(σ,t) ΓΦ 2
coincides with the quadratic differential of ϕ. Hence, the functions ϕ and ϕ1

can be glued together without affecting the LΦ -leafwise second differential
along V and without creating any additional LΦ -leafwise critical points. We
will keep the notation ϕ for the resulting function on Op V .
Let us now extend the leafwise gradient ∇LΦ ϕ as a section of the bundle ζ
to the whole W without zeroes outside Op V . Recall that we have extended
ΔΦ to an isomorphism T F → ζ ⊂ T W . Consider the section Φ1 := ΔΦ ◦Φ1 :
W → ζ. It vanishes transversely along V , and its differential along V defines
an isomorphism
ΓΦ := dΦ1 |Norm = ΔΦ ◦ ΓΦ : Norm → ζ|V = VertΦ .
−1 2
Over V \U we have VertΦ = Norm and ΔΦ = ΓΦ ◦Φ . Hence, ΓΦ |V \U =
Φ −1
Γ / ⊕ Span (
◦ Φ2 ◦ ΓΦ . Over U we have VertΦ = Nor τ ). The operator
ΓΦ leaves this splitting invariant and satisfies
Φ −1
ΓΦ |Nor
= (Γ ) ◦ Φ2 ◦ ΓΦ and ΓΦ (n+ ) = τ+ .
The LΦ -leafwise gradient ∇LΦ ϕ defines another section of ζ|Op W V = T LΦ .

Here the gradient is taken with respect to the metric on VertΦ which is
/ U via ΓΦ from the metric on Vert, and for
induced on Norm|V \U and on Nor|
/ Using (20), we compute
which τ+ is a unit vector field orthogonal to Nor.
for v ∈ V \ U and X, Y ∈ Normv :
∇LΦ ϕ(v, X), Y VertΦ = ΓΦ (∇LΦ ϕ(v, X)), ΓΦ (Y )Vert
= d(v,X) ϕ(Y ) = dqv (ΓΦ (X))(ΓΦ (Y )) = Φ2 ◦ ΓΦ (X), ΓΦ (Y )Vert .
Thus,
−1
(21) ∇LΦ ϕ(v, X) = ΓΦ ◦ Φ2 ◦ ΓΦ (X), v ∈ V \ U, X, Y ∈ Normv .
The differential d(∇LΦ ϕ) along V yields a section Norm → VertΦ , and from
(21) we conclude
−1
d(∇F ϕ)(X) = ΓΦ ◦ Φ2 ◦ ΓΦ (X) = ΓΦ (X), X ∈ Norm|V \U .
Similarly, using (19), we compute
d(∇LΦ ϕ)(X) = ΓΦ (X), / U,
X ∈ Nor|
and d(∇LΦ ϕ)(n+ ) = τ+ = ΓΦ (n+ ). Hence, ∇LΦ ϕ and Φ1 have the common
zero set V and their differentials along V agree because
d(∇LΦ ϕ)|Norm = ΓΦ = dΦ1 |Norm : Norm → VertΦ .
Therefore, we can modify the global section Φ1 : W → ζ to make it agree
with ∇LΦ ϕ on Op V , without creating any additional zeroes. Keeping the
notation Φ1 for this modified section, we thus get a locally holonomic ζ-FLIF
Φ = (Φ1 , d2LΦ ϕ|V , d3LΦ ϕ|Σ ).
It remains to observe that the framing ξ of the FLIF Φ induces a framing
ξ := (ΔΦ )∗ ξ of the ζ-FLIF Φ. Indeed, the bundle VertΦ serves as the vertical
bundle of the ζ-FLIF Φ, and by construction its negative subbundle VertΦ −
is the image ΔΦ (Vert− ) ⊂ VertΦ .
16.4. Proof of the main theorems

A. Pleating of framed leafwise Igusa functions. Consider a (k + n)-
dimensional manifold W endowed with an n-dimensional foliation F , and
a k-dimensional submanifold V ⊂ W which is folded with respect to F
with fold Σ ⊂ V . Recall from Section 15.6.B the pleating construction of
V with basis (S, μ), where S ⊂ V \ Σ is a cooriented codimension 1 closed
submanifold and μ ∈ T F |S 1 a unit vector field along S tangent to F . The
1 Here and in various other places in this book, “μ ∈ E” is a shorthand notation for “the
vector field μ takes values in the bundle E”.

16.4. Proof of the main theorems 209
next result shows that if V = V (Φ) for a locally holonomic ζ-FLIF Φ, then
the pleating can be realized by a homotopy of locally holonomic ζ-FLIFs.
16.4.1. (Pleating of framed leafwise Igusa functions)
(a) Let Φ be a locally holonomic ζ-FLIF and V the result of pleating V =
V (Φ) with a basis (S, μ), where μ ∈ Vert± |S . Then Φ can be deformed,
keeping it fixed outside Op W S, to a locally holonomic ζ-FLIF Φ such that
=V.
V (ϕ)
(b) Suppose that Φ is framed with a framing ξ, and if μ ∈ Vert− assume in
addition that μ = ±ξ i for one of the framing fields ξ i . Then the pleating of
V with basis (S, μ), followed by an additional compensating pleating, can be
realized by a deformation in Op W S of (Φ, ξ) to a locally holonomic framed
ζ-FLIF (Φ,
ξ).
Proof. By assumption, ζ integrates to a foliation F on a neighborhood

Ω ⊃ V and Φ|Ω is holonomic, Φ|Ω = J(ϕ). Let us first observe that if ϕu ,
u ∈ [0, 1], ϕ0 = ϕ, is a family of leafwise Igusa functions on Ω which have no
leafwise critical points on ∂Ω, then there is family Φu , Φ0 = Φ of ζ-FLIFs
extending J(ϕu ) from Ω to W without singularities outside Ω. In the proof
below we will call a deformation of ϕ with the above property admissible.
Consider first the case n = k = 1. Then S is a point and we have the
following model situation: W = R2 with coordinates (t, s), V = {s = 0},
F is the foliation by the fibers of the projection (t, s) → t, S is the origin,
μ = ∂s∂
, and ϕ(t, s) = ±s2 .
For 0 < ε < 12 take a nonincreasing cutoff function θ : [0, ∞) → [0, 1] which
is equal to 1 on [0, 2ε ] and equal to 0 on [ε, ∞). If ϕ = s2 , we define
! 4 "
s 3ε2 s2
s)
ϕ(t, := θ(|t|) − + ts + (1 − θ(|t|))s2 .
4 2
Then ϕ = ϕ for |t| ≥ ε, and we can interpolate ϕ
to ϕ for |s| ≥ ε1/4 without
creating additional critical points. We will denote the resulting function
which agrees with ϕ outside {|t| ≤ ε, |s| ≤ ε1/4 } still by ϕ. The leafwise
critical point locus V of ϕ
is given by
0 1
∂ϕ 2 3
V = = 0 = θ(|t|)(s3 − 3ε2 s + t) + 2(1 − θ(|t|))s = 0 ;
∂s
see Figure 16.9. Thus V is the result of pleating V = {s = 0} along S in
the direction μ. For ϕ = −s2 , we use instead the function −ϕ.

Unfortunately, this construction yields an Igusa function which cannot be
framed. Indeed, the framing over the arc V 1 ⊂ V connecting the points
(2ε3 , ε) and (2ε3 , ε) (where dim Vert− = 1) has to be the vector field ±μ,
but this vector field is inward pointing at one of the ends of this arc, which
3ε
ε
−ε ε−
2ε3 2
ε
t
−ε2 2ε 3
− 3ε
of the function ϕ.
Figure 16.9. The singular locus V
Γ1 −
+
−
µ
+
− +
Γ2
+ +
µ ξi
− −
_
Γ2
µ −
−
+ + ξi
Figure 16.10. Compensating pleatings.
contradicts condition (F2) in Section 16.1.A in the definition of a framing. A

similar problem arises if we apply the construction to the function −s2 : here
condition (F2) implies that the framing vectors at the points (±ε, 0) have
opposite signs, so they cannot match the original framing of V . However, the
problem can be fixed by an additional pleating as illustrated in Figure 16.10.
Note that for ϕ = s2 the compensating pleating takes place along a point
16.4. Proof of the main theorems 211
on the arc V 1 , while for ϕ = −s2 it takes place along a point outside of this
arc.
The general case can be reduced to the case n = k = 1. By the Parametric
Morse Lemma 11.2.2, the function ϕ can be written on Op W S in the form
ϕ(σ, t, X+ , X− ) = X+ 2 − X− 2 + g(σ, t).
Here (σ, t) ∈ S × [−1, 1] are coordinates on a tubular neighbourhood U ⊂ V
of S = S × 0 and X± ∈ Vert± , identifying a tubular neighborhood of U in
W with its neighborhood in the total space of the bundle Vert. Suppose
that μ takes values in Vert+ (the case μ ∈ Vert− is analogous). Then we
can write X+ = sμ(σ) + X+ with s ∈ R and X orthogonal to μ(σ), After
+
adjusting ϕ by an admissible homotopy, if necessary, we can assume that
2
ϕ(σ, t, sμ(σ) + X+ , X− ) = s2 + X+ − X− 2 + g(σ, t).
t, sμ(σ) + X+
Consider the function ϕ(σ, X ) = ϕ(t,
s) + X+ 2 − X 2 +
− −
t) as above. Its leafwise critical point locus V (ϕ)
g(σ, t) with ϕ(s, is the
result of pleating of V along S in the direction μ. It remains to observe that
functions of the family uϕ + (1 − u)ϕ, u ∈ [0, 1], have no leafwise critical
points outside a neighborhood of V which can be made arbitrarily small by
decreasing the parameter ε. Hence, ϕ is an admissible deformation of ϕ.
Suppose now that ϕ is framed, and if μ ∈ Vert− , then μ = ±ξ i for one of the
framing vector fields ξ i . Then, as in the case n = k = 1, the construction
can be done in the category of framed Igusa functions with an additional
pleating.
B. Proof of Theorem 16.1.1. Recall that W is an (n + k)-dimensional

manifold endowed with an n-dimensional foliation F , and (ϕA , ξA ) a framed
leafwise Igusa function on Op A for a (possibly empty) closed subset A ⊂ W .
Our goal is to extend (ϕA , ξA ) to a framed leafwise Igusa function on the
whole W .
First, we apply Theorem 16.2.1 to extend (ϕA , ξA ) to a framed FLIF (Φ, ξ)
on W . Without further mentioning, all the following constructions are done
relative to Op A.
Next, we apply Proposition 16.3.1 to obtain a locally holonomic framed ζ-
FLIF (Φ, ξ), where the n-dimensional distribution ζ is homotopic to T F . Let
ζt , t ∈ [0, 1], be a homotopy connecting ζ0 = ζ and ζ1 = T F . According to
Corollary 10.2.2 this homotopy can be made integrable on a neighborhood
Ω ⊂ W of the singular set V = V (Φ). We denote by Ft the family of
foliations on Ω which we get by integrating ζt , so that F1 = F |Ω . Using
Theorem 15.6.2, we find a manifold V ⊂ Ω obtained from V by a sequence of
successive pleatings, and a diffeotopy ht : W → W supported in Ω such that
h∗t Ft = F0 on Op V . According to Lemma 16.4.1, after possible additional
pleatings on V , (Φ, ξ) can be deformed in Ω to a locally holonomic framed ζ-

FLIF (Φ, ξ) such that V (Φ)
= V . Theorem 15.6.2 ensures that a diffeotopy
ht as above still exists after the additional pleatings.
By definition of locally holonomic, Φ = J(ϕ)
on a neighborhood Ω ⊂ Ω of V
for a leafwise Igusa function ϕ :Ω → R. So the function ψ = ϕ ◦ h−1
1 is an

F -leafwise Igusa function on Ω1 := h1 (Ω) which is framed by the framing
(h1 )∗ ξ, i.e., the image of the framing ξ under the diffeomorphism h1 . We
extend the holonomic F -FLIF J(ψ) on Ω1 to an F -FLIF Ψ = (Ψ1 , Ψ2 , λ+ )
on W by setting Ψ1 := (h1 )∗ Φ 1 and taking for Ψ2 , λ+ the corresponding
components of J(ψ) along the zero set V1 := h1 (V ) ⊂ Ω1 of Ψ1 . Hence, we
can apply the foliated h-principle 13.2.4 to the foliated manifold (W, F ), the
function ψ (arbitrarily extended to W ), and the leafwise 1-form Ψ1 on W
satisfying Ψ1 = dF ψ on Ω1 . This yields an extension of ψ from Ω1 to an
F -leafwise wrinkled function on ψ1 : W → R. Thus the extended function
ψ1 is leafwise Igusa with singular set

N
V (ψ1 ) = V1 ∪ Si ,
1
where Si are the singular sets of the wrinkles. Each Si is a k-sphere folded
with respect to F along its equator. According to the Example at the end
of Section 16.1.A each of these wrinkles can be framed. This completes the
proof of Theorem 16.1.1.
C. Proof of Theorem 16.1.2. Note that in the proof of Theorem 16.1.1

the framing ξ was essential only for extending the framed leafwise Igusa
function as a FLIF from Op A to W using Lemma 16.2.1. In all subsequent
steps of the proof the framing did not play any role, and its presence was
just an additional nuissance (e.g., in Lemma 16.4.1(b) in comparison with
16.4.1(a)). In Theorem 16.1.2 the formal extension is a hypothesis, hence its
proof repeats all the subsequent steps of the proof of Theorem 16.1.1, with
the simplification because we need not care about the framing. It remains
to verify that all the modifications in the proof preserve the homotopy class
of the FLIF. This is straightforward for all steps except the stabilization and
pleating constructions. But both of these constructions are always localized
in a neighborhood of a point p ∈ V \ Σ, where any FLIF can be framed. In
both cases the modifications are done in the category of framed FLIFs, where
the existence of a homotopy of framed FLIFs (related to the boundary of a
neighborhood of the point p) is guaranteed by Lemma 16.2.1. This yields, of
course, a homotopy of the unframed FLIFs, and Theorem 16.1.2 follows.
16.5. Applications to pseudo-isotopy theory 213
16.5. Applications to pseudo-isotopy theory

Theorems 16.1.1, 16.1.2 and 13.2.4 allow us, under certain conditions, to
reduce the singularities in multiparametric families of functions (which a
priori could be very complicated) to simple singularities of Morse and birth-
death type. This is useful for multiparametric Morse–Smale theory. In this
section we consider applications of this scheme to pseudo-isotopy theory, see
also [Ce70, HW73, Ig84, Ig87, EM97, EM00],
A. Pseudo-isotopies. Consider a compact manifold N , possibly with

boundary, and set M = N × I with I = [0, 1]. We denote by E(N ) be
the space of functions on M without critical points which coincide with the
projection t : M → I near ∂M = N × ∂I ∪ ∂N × I. We denote by E(N )
the formal analogue of E(N ), which is the space of nonvanishing 1-forms on
M which coincide with dt near ∂M , and by d : E(N ) → E(N ) the natural
inclusion given by ϕ → dϕ.
We also consider the pseudo-isotopy space P(N ), i.e., the space of diffeo-
morphisms M → M which are equal to the identity near N × 0 ∪ ∂N × I.
Consider the map h : P(N ) → E(N ) given by the formula h(f ) = t ◦ f . It
was observed by J. Cerf (see [Ce70]) that h is a homotopy equivalence.
Exercise. Find a homotopy inverse g : E(N ) → P(N ) of h.
B. The pseudo-isotopy space as a subspace of wrinkled, or Igusa

functions.
16.5.1. Consider a spheroid ψ : S k → E(N ) and the corresponding function
ψ : S k × M → I given by the formula Ψ(t, x) = ψ(t)(x), t ∈ S k , x ∈ M .
Foliate D k+1 × M by the fibers of the projection D k+1 × M → D k+1 . Then
(a) ψ extends to a framed leafwise Igusa function Ψ : D k+1 × M → I;
(b) ψ extends to a leafwise wrinkled function Ψ : D k+1 × M → I.
Proof. Claim (a) is a special case of Theorem 16.1.1. Claim (b) is a conse-
quence of Theorem 13.2.4 and Proposition 16.5.2 below, which goes back to
F. Laudenbach and A. Douady (see [La76]) and which was communicated
to us by K. Igusa.
16.5.2. (Laudenbach–Douady, Igusa) The map d : E(N ) → E(N ) is

contractible.
Proof. We will consider here only the case N = D n and leave to the reader
the general case, which uses the same idea but is more technical due to the
necessary cutoff constructions; see also [La76] and [EM00] for details.
We abbreviate E := E(D n ), E := E(D n ). The space E can be identified

with the space of nonvanishing vector fields on M which coincide with the
∂
vector field ∂t near ∂M , and the map d can be viewed as the inclusion
j : E → E of the subspace of gradient vector fields. Consider the subspace
F ⊂ E consisting of vector fields whose ∂t ∂
-component is everywhere > 0.
As this space deformation retracts to the point ∂t ∂
∈ F , it is sufficient to
prove that j : E → E is homotopic to a map E → F .
Let ∂+ M := D n × 1. Given v ∈ E and z ∈ M \ ∂+ M , we denote by v τ (z) the
forward trajectory of the vector field v originating at z. In other words, v τ (z)
τ
is the solution of the ODE dvdτ(z) = v(v τ (z)) with initial condition v 0 (z) = z.
The solution exists for τ ∈ [0, T (z)], where the exit time T (z) > 0 is defined
by v T (z) (z) ∈ ∂+ M .
For v ∈ E and s ∈ [0, 1], we define a vector field vs on M by
v sT (z) (z) − z
vs (z) := for s ∈ (0, 1], z ∈ M \ ∂+ M,
sT (z)
v0 (z) = v(z) for all z ∈ M , and vs (z) = ∂
∂t for z ∈ ∂+ M , s ∈ [0, 1]. Then
sT (z)
vs is continuous at s = 0 because lims→+0 vs (z) = lims→+0 v sT (z) (z)−z
=
(
dv s (z) ( ∂
ds s=0
= v(z), and it is continuous on ∂+ M because v = ∂t near ∂M .
Moreover, vs (z) = 0 for all s ∈ [0, 1] and z ∈ M because v (z) = z for all
τ
τ > 0 and z ∈ M \ ∂+ M . Finally, since z ∈ M \ ∂+ M has t-component

< 1 and v T (z) (z) ∈ ∂+ M has t-component 1, the ∂t∂
-component of the vector
field v1 is everywhere > 0.
∂
The vector fields vs do not yet belong to E because vs = ∂t holds near
∂M \ (0 × D ) but not near 0 × D . However, this is easily remedied by
n n
replacing vs by vs (z) := (1 − ϕv (z))vs (z) + ϕv (z) ∂t ∂

for a cutoff function
ϕv : M → [0, 1] supported in a neighborhood Uv of D n × 0 where v = ∂t ∂
,
equal to 1 near D × 0, and depending continuously on v. Note that vs is
n
nowhere zero because vs has positive ∂t ∂

component on Uv . Now vs ∈ E for
all s ∈ [0, 1], v0 = v and v1 ∈ F , so (s, v) → vs defines the desired homotopy
of j : E → E to a map E → F .
Remark. As we already mentioned above, Theorem 16.1.1 can be refor-
mulated as saying that the appropriately defined space IF(V, A) of framed
Igusa functions on a manifold V fixed on a neighborhood of a closed subset
A ⊂ V is contractible, and the same holds in relative form. Hence, the long
exact sequence of the pair E(N ) ⊂ IF(M, ∂M ) implies
πk+1 (if(M, ∂M ), E(N )) ∼
= πk (E(N )).
Similarly, Theorem 13.2.4 can be reformulated as a homotopy equivalence
between E(N ) and the appropriately defined space Wr(M, ∂M ) of wrin-
kled functions on M which coincide with the function t on Op ∂M . Hence,
16.5. Applications to pseudo-isotopy theory 215
Proposition 16.5.2 together with the long exact sequence of the pair E(N ) ⊂
E(N ) Wr(M, ∂M ) implies the isomorphism
πk+1 (Wr(M, ∂M ), E(N )) ∼= πk+1 (E(N )) ⊕ πk (E(N )).

Part 4
The Homotopy
Principle in Symplectic
Geometry
Symplectic and contact geometry lie on the borderline between the flex-
ible world, governed by the laws of the h-principle, and the rigid world,
which deals with the differential relations for which the homotopy restric-
tions sufficient for the existence of formal solutions are far from being suf-
ficient for genuine solvability. In the 1960s the conjectures of V.I. Arnold
(see [Ar65, Ar78]) were directing the development of symplectic geometry
towards rigidity, while the success of symplectic applications of Gromov’s
h-principle (see [Gr69]) put under a big question mark whether any rigid
phenomena may exist in the symplectic world, see the discussion in [Ar86]
and historical remarks in [Gr85]. For two decades after mid-1980s the rigid-
ity side dominated the development of symplectic topology. However, more
recently several new discoveries were made also on the flexible side. We will
consider in the subsequent chapters the flexible side of the symplectic story,
corresponding to its state at the time of the first edition of this book. More
recent developments will be discussed in our forthcoming book [CEM25].
We also briefly mention some rigid phenomena in Chapter 19, and refer the
readers who wish to see the rigid part of the symplectic world to Gromov’s
seminal paper [Gr85], Arnold’s paper [Ar86], as well as the books [MS98]
and [HZ94].
Chapter 17
Symplectic and
Contact Basics
This chapter contains a short introduction to symplectic and contact geom-

etry. We do not pretend to be systematic. However, the chapter contains all
the symplectic and contact information which is needed for our applications.
We stress the similarity and the relationship between symplectic and com-
plex geometries. The connection between the two geometries continues to
serve as one of the most important sources of all the recent developments in
symplectic geometry after Gromov’s paper [Gr85]. The proofs are mostly
only indicated, or even omitted. The reader who is not familiar with the
subject may consider the text as a long sequence of exercises, or may turn
to the books [AG01], [MS98], [HZ94], [CdS01], and [CE12].
17.1. Linear symplectic and complex geometries

A. Symplectic structures. A symplectic structure, or a symplectic form,
on a real vector space L is a nondegenerate 2-form ω ∈ Λ2 L∗ , i.e., a skew-
symmetric bilinear form
L × L → R.
The nondegeneracy condition means that the formula
Iω (X) = ω(X, ·) = X ω, X ∈ L,
defines an isomorphism Iω : L → L∗ between L and its dual space L∗ .
This condition automatically implies that L is of even dimension 2n. The
nondegeneracy of ω is equivalent to the condition that ω n = 0, i.e., that ω n
is a volume form. The pair (L, ω) is called a symplectic vector space. The
219
220 17. Symplectic and Contact Basics
group of all linear symplectomorphisms

Φ : L → L, Φ∗ ω = ω,
is denoted by Sp (L, ω).
Any symplectic vector space (L, ω) has a symplectic basis

u1 , . . . , un , v1 , . . . , vn
such that ω(ui , vi ) = 1 and w is equal to 0 on all other pairs of basic vectors;
with respect to this basis the form ω can be written as

n
ω = ω0 = pi ∧ qi , pi = ω(ui , ·), qi = ω(·, vi )
i=1
or, equally, ω(X, Y ) = X T Ω0 Y , where

% &
0 1
Ω0 = .
−1 0
In particular, as in the case of Euclidean structures, there exists, up to

isomorphism, only one symplectic structure on L and only one linear sym-
plectic group
Sp(2n) = Sp (R2n , ω0 ) Sp (L, ω).
Let us point out, however, that the space
S(L) = Λ2 L \ (Σ = {ω | ω n = 0}) GL (2n, R)/Sp (2n)
of all symplectic structures on L, in contrast to the space of Euclidean
structures G(L) GL (L)/O(L), is not contractible. Note that S(L) consists
of two identical (noncontractible) components which correspond to the two
orientations on L.
Exercise. Prove that S(R4 ) is homotopy equivalent to S 2 S 2 .
B. Symplectic orthogonal complement and classification of linear

subspaces. There exists a remarkable diversity of linear subspaces of a
symplectic vector space. Each linear subspace S in (L, ω) is characterized
up to isomorphism by two numbers (s, p) where dim S = s and rank ω|S = p.
Note that p is always even. An alternative description of this classification
can be given in terms of the symplectic orthogonal complement.
Let (L, ω) be a symplectic vector space, dim L = 2n. Given a linear subspace
S ⊂ L, the linear subspace
S ⊥ω = {X ∈ L | ω(X, Y ) = 0 for all Y ∈ S}
17.1. Linear symplectic and complex geometries 221
is called the symplectic orthogonal complement, or ω-orthogonal complement

of S. As in the case of the Euclidean orthogonal complement, for any S ⊂ L
we have
⊥ω
dim S + dim S ⊥ω = 2n and S ⊥ω = S.
However, in general S + S ⊥ω = L. In particular, the ω-orthogonal comple-

ment of a line always contains this line, while the ω-orthogonal complement
of a hyperplane is contained in the hyperplane.
A subspace S ⊂ (L, ω) has type (s, p) if and only if

dim (S ∩ S ⊥ω ) = s − p.
A subspace S ⊂ (L, ω) is called
• symplectic if it has type (s, s), or equivalently if S + S ⊥ω = L;
• isotropic if it has type (s, 0), i.e., ω|S = 0, or equivalently if S ⊂
S ⊥ω ;
• coisotropic if it has type (s, 2s − 2n), or equivalently if S ⊥ω ⊂ S;
• Lagrangian if it has type (n, 0), or equivalently if S = S ⊥ω .
Note that the dimension of an isotropic subspace is always ≤ n and the
dimension of a coisotropic subspace is always ≥ n. Lagrangian subspaces can
be also characterized as isotropic subspaces of maximal possible dimension,
or coisotropic subspaces of the least possible dimension. The symplectic
complement of an isotropic subspace is a coisotropic subspace and vice versa.
The intersection S∩S ⊥ω is always isotropic. Also note that for any fixed s we
have a stratification of the Grassmannian Grs L by strata which correspond
to (s, p)-subspaces; the stratum of maximal dimension corresponds to the
maximal possible p. In particular, for even s a generic s-dimensional linear
subspace of (L, ω) is symplectic.
C. Complex structures. A complex structure on a real vector space L is

an automorphism J : L → L such that J 2 = −IdL . This condition auto-
matically implies that L is of even dimension 2n. The pair (L, J) is called a
complex vector space. This definition, of course, coincides with the definition
of a complex vector space as a vector space over C; the correspondence is
(a + ib)v = av + bJv.
The group of all linear transformations
Φ : L → L, Φ−1 JΦ = J,
i.e., transformations which preserve J, is denoted by GL(L, J).
For any complex vector space (L, J) there exists a J-basis

u1 , . . . , un , v1 , . . . , vn
such that Jui = vi ; with respect to this basis we have
% &
0 −1
J = J0 = .
1 0
In particular there exists, up to isomorphism, only one complex structure
on L and only one linear complex group
GL(n, C) = GL (R2n , J0 ) GL (L, J).
As we will see below, the space
J (L) GL (2n, R)/GL (n, C)
of all complex structures on L is homotopy equivalent to the space S(L) of
symplectic structures on L. In particular, J(L) consists of two homeomor-
phic (noncontractible) components which correspond to the two orientations
of L.
Exercise. Prove that J (R4 ) is homotopy equivalent to S 2 S 2 .
D. Classification of real linear subspaces in a complex vector space.

Each linear subspace S in (L, J) is characterized, up to isomorphism, by two
numbers (s, p) where dim S = s and p = dim (S ∩ JS). Note that S ∩ JS
is J-invariant and p is always even.
A subspace S ⊂ (L, J) is called

• complex if it has type (s, s), or equivalently if S = JS; in other
words, a subspace is complex if it is invariant with respect to J;
• real or totally real if it has type (s, 0), or equivalently if S ∩JS = 0;
in other words, a subspace is totally real if it contains no complex
subspaces of positive dimension;
• co-real if it has type (s, 2s − 2n), or equivalently if S + JS = L.
Note that the dimension of a real subspace is always ≤ n, while the di-
mension of a co-real subspace is always ≥ n. The intersection S ∩ JS is
J-invariant and hence always complex. Also note that for any fixed s we
have a stratification of the Grassmannian Grs L by the strata which corre-
spond to (s, p)-subspaces; the stratum of maximal dimension corresponds to
the minimal possible p. In particular, a generic s-dimensional linear sub-
space of (L, ω) is real if s ≤ n and co-real if s ≥ n. The odd-dimensional
subspaces of type (s, s − 1) are called CR-subspaces, where the notation CR
stands for Cauchy–Riemann. Any real hyperplane in a complex space is
automatically a CR-subspace.
17.1. Linear symplectic and complex geometries 223
E. Hermitian structures and the homotopy equivalence J (L) ∼

S(L). A Hermitian structure on a (real) vector space L is a pair H = (J, ω)
where J and ω are compatible complex and symplectic structures on L, i.e.,
• ω is J-invariant, i.e., ω(JX, JY ) = ω(X, Y ) for all X, Y ∈ L;
• ω is J-positive, i.e., ω(X, JX) > 0 for all X ∈ L.
This definition is equivalent to the standard definition of a Hermitian struc-
ture on L as a positive definite Hermitian form H(X, Y ) on the complex
vector space (L, J); the correspondence is given by the formula
H(X, Y ) = ω(X, JY ) − iω(X, Y ).
A linear transformation L → L which preserves H is called unitary, and the
group of all unitary transformations is denoted by U(L, H).
For any Hermitian vector space (L, H) there exists a basis

u1 , . . . , un , v1 , . . . , vn
which is simultaneously a J-basis and a symplectic basis. With respect
to this basis we have H = H0 = (J0 , ω0 ). In particular, as in the case
of Euclidean spaces there exists, up to isomorphism, only one Hermitian
structure on L and only one unitary group
U(n) = U(R2n , H0 ) U(L, H).
The space H(L) of all Hermitian structures on L is a subspace of the product
J (L) × S(L) and thus we have natural projections
pJ : H(L) → J (L) and pS : H(L) → S(L).
17.1.1. (Homotopy equivalence J (L) ∼ S(L)) The projections pJ :
H(L) → J (L) and pS : H(L) → S(L) are surjective maps. Both maps pJ
and pS are fibrations with contractible fibers. Moreover, both pJ and pS are
homotopy equivalences, hence there exists a canonical homotopy equivalence
J (L) ∼ S(L).
Proof. The surjectivity of pJ and pS is evident. Each fiber

p−1 −1
J (J) pJ (J0 ) = GL(n, C)/ U(n)
is a convex subset of a vector space and thus contractible. The convexity of
p−1
J (J) also implies that pJ is a homotopy equivalence. Each fiber
p−1 −1
S (ω) pS (ω0 ) = Sp(2n)/ U(n)
is contractible, as can be seen from the polar decomposition (see, for exam-
ple, [MS98]). Therefore, there exists a continuous section f : S(L) → H(L)
of the fibration pS : H(L) → S(L); using fiberwise polar decompositions
with respect to the metrics gω (X, Y ) = ω(X, JY ) where (J, ω) = f (ω) we
can realize the contraction simultaneously in all fibers. Thus, the projection
pS is a homotopy equivalence.
Remarks.
1. The space Hω = p−1 S (ω) admits another interpretation which manifests
its contractibility. Namely, Hω can be identified with the so-called Siegel
upper half-space, i.e., the space of matrices of the form A + iB where A, B
are real symmetric n × n-matrices and B is positive definite; see [Si64] for
the details.

2. Instead of H(L) one can consider the bigger space H(L) of all pairs (J, ω)
satisfying only the positivity condition
ω(X, JX) > 0 for all X ∈ L.

A proposition similar to 17.1.1 is valid also for H(L) and the projections

pJ : H(L)
→ J (L) and pS : H(L) → S(L).
17.2. Symplectic and complex manifolds

A. Symplectic and complex vector bundles. Using the respective lin-
ear notions, one can define symplectic, complex, and Hermitian vector bun-
dles. For example, a symplectic vector bundle (X, ω) over a manifold V is
a real vector bundle p : X → V equipped with a symplectic form ωv on
each fiber Xv = p−1 (v) which smoothly depends on v ∈ V . Equivalently,
the symplectic structure on a real vector bundle X → V can be defined as
a section
V → Λ2 X \ Σ,
where Σ ∩ Λ2 Xv = {ω ∈ Λ2 Xv | ω n = 0}.
Two symplectic structures ω0 , ω1 on a real vector bundle X are homotopic

if they can be connected by a family {ωt }t∈I of symplectic structures. As
follows from Proposition 17.1.1, there is no difference from the homotopy
point of view between symplectic, complex, and Hermitian vector bundles,
i.e., there exists a canonical one-to-one correspondence between the homo-
topy classes of symplectic and complex structures on a given real vector
bundle X and, moreover, there are canonical homotopy equivalences
S(X) ∼ H(X) ∼ J (X),
where S(X), J (X), and H(X) are the spaces of symplectic, complex, and
Hermitian structures on X.
17.2. Symplectic and complex manifolds 225
Remark. For a given symplectic structure ω on a vector bundle X,

the connected component Jω ⊂ J (X) which corresponds to the connected
component Sω ⊂ S(X) of ω consists of all complex structures J ∈ J (X)
such that J is fiberwise compatible with a symplectic structure ω ∈ Sω (and
vice versa).
B. Almost symplectic and almost complex manifolds. An almost

symplectic (resp., almost complex, Hermitian) structure on an even-dimen-
sional manifold V is a symplectic (resp., complex, Hermitian) structure on
the tangent bundle T V . Equivalently, an almost symplectic structure on
V is a nondegenerate differential 2-form ω on V . All these structures have
local invariants and, in the almost complex case, were intensively studied,
similarly to the case of Riemannian metrics. As far as we know the differ-
ential geometry of almost symplectic structures is practically nonexistent.
C. Submanifolds of almost symplectic and almost complex mani-

folds. Using the respective linear notions, one can define (s, p)-submanifolds
of almost symplectic (resp., almost complex) manifolds, and in particular
isotropic, co-isotropic, Lagrangian, almost symplectic (resp., totally real, co-
real, almost complex ) submanifolds. Note that isotropic submanifolds S ⊂ V
are characterized by the condition ω|S ≡ 0. The isotropic submanifolds of
dimension < 12 dim V , i.e., non-Lagrangian isotropic submanifolds, are called
subcritical.
D. Symplectic and complex manifolds: infinitesimal description.

An almost symplectic structure ω is called integrable if dω = 0, i.e., ω is a
closed two-form. Such a differential form is called symplectic. An almost
complex structure J on V is called integrable if the Nijenhuis tensor
NJ (X, Y ) = [JX, JY ] − J[JX, Y ] − J[X, JY ] − [X, Y ]
vanishes. A Hermitian structure H = (J, ω) on V is called integrable if both

J and ω are integrable; this is equivalent to the equality ∇g J = 0 where ∇g
is the covariant derivative with respect to the metric g(X, Y ) = ω(X, JY ).
A manifold V provided with an integrable almost symplectic (resp., al-
most complex, Hermitian) structure is called a symplectic (resp., complex,
Kähler ) manifold. A Kähler manifold can equivalently be defined as a com-
plex manifold provided with a positive definite Hermitian form H such that
the imaginary part of H is closed; such a form H is called a Kähler metric.
An almost complex manifold V provided with a Kähler metric is called an
almost Kähler manifold.
E. Symplectic and complex manifolds: local description. Accord-

ing to a theorem of Newlander and Nirenberg (see [NN57]), any sufficiently
smooth integrable almost complex structure is locally equivalent to the stan-
dard complex structure on Cn . Thus, equivalently, a complex manifold can
be characterized by the existence of local charts {Ui → (R2n , J0 ) = Cn }
glued together by holomorphic maps (this is the standard definition of a
complex manifold).
For symplectic manifolds we have a similar situation. According to Dar-

boux’s theorem (see Theorem 17.3.2), any symplectic form is locally equiv-

n
alent to the standard symplectic form ω0 = dxi ∧ dyi on R2n . Thus,
1
equivalently, a symplectic manifold can be characterized by the existence of
local Darboux charts glued together by symplectomorphisms, i.e., diffeomor-
phisms which preserve the canonical form.
Note that for two-dimensional manifolds we have

almost symplectic structure = symplectic structure = area form;
almost complex structure = complex structure = conformal structure.
F. Submanifolds of symplectic and complex manifolds. Note that a

symplectic (complex) submanifold S ⊂ V of a symplectic (complex) mani-
fold V inherits an integrable symplectic (complex) structure. This contrasts
with the Riemannian case: the symplectic integrability condition is analo-
gous to the condition in the Riemannian case of being locally Euclidean, but
submanifolds of locally Euclidean manifolds need not, of course, be locally
Euclidean.
Let us mention the following property of (s, p)- submanifolds of (integrable)

symplectic manifolds (see, for example, [MS98]).
17.2.1. For any (s, p)-submanifold S ⊂ V of a symplectic manifold V the
(s − p)-dimensional distribution T S ∩ (T S)⊥ω ⊂ T S is integrable. The
corresponding foliation on S, called characteristic, consists of isotropic
leaves. In particular, any s-dimensional coisotropic submanifold S ⊂ V
carries a canonical (2n − s)-dimensional characteristic isotropic foliation.
Note that a hypersurface S ⊂ V is always coisotropic and thus carries a

canonical one-dimensional characteristic foliation.
The complex tangent subspaces T S ∩ J(T S) to a real hypersurface S of

an almost complex manifold V form a complex tangent distribution ξ of
real codimension 1, which is called a CR-structure on S. A coorientation
ν of S in V such that Jν is tangent to S defines a coorientation Jν of ξ
17.2. Symplectic and complex manifolds 227
in S. Choose a 1-form α on S such that ξ = Ker α and α(Jν) > 0. It

is straightforward to check that if the structure J is integrable, then the
formula
L(X, Y ) = dα(X, JY ) − idα(X, Y ), X, Y ∈ ξ,
defines an Hermitian form on ξ, called the Levi form. The Levi form is
independent of the choice of α up to multiplication by a positive function.
The cooriented hypersurface S is called strictly pseudo-convex if the form L
is positive definite.
G. Examples of symplectic manifolds.

1. Cotangent bundle An important example of a symplectic manifold is
provided by the cotangent bundle T ∗ M of a smooth manifold M . The
symplectic form ω on T ∗ M is the differential of the famous canonical 1-
form p dq. In coordinate notation the symplectic structure on T ∗ M can be
described as follows. If M = Rn , then R2n = T ∗ Rn is endowed with the
canonical symplectic structure

k
ω0 = d(p dq) = dpi ∧ dqi ,
1
where the coordinates q = (q1 , . . . , qn ) and p = (p1 , . . . , pn ) are chosen in
such a way that the projection T ∗ Rn → Rn is given by (p, q) → q. Let us
observe that any diffeomorphism f : Rn → Rn lifts to a symplectomorphism
f∗ : T ∗ Rn → T ∗ Rn by the formula
f∗ (p, q) = (f (q), (df ∗)−1 (p)).

Thus a coordinate atlas M = Uj on M lifts to a symplectic atlas
j

T ∗M = T ∗ Uj
j
with gluing symplectomorphisms defined by the above formula.

2. Symplectic stucture on Kähler manifolds Kähler geometry serves as a
rich source of examples of symplectic manifolds. The imaginary part of a
Kähler metric is a symplectic form, hence any Kähler manifold is automat-
ically symplectic. Moreover, complex submanifolds of Kähler manifolds are
Kähler, and thus symplectic. The complex affine space Cn and the complex
projective space CP n have canonical Kähler metrics ( ni=1 zi z̄i on Cn , and
the Fubini–Study metric on CP n ). Hence any affine complex manifold, i.e.,
a complex submanifold of Cn , and any projective complex manifold, i.e., a
complex submanifold of the complex projective space CP n , are symplectic.
The above examples of complex analytic origin can in a certain sense be
reversed. Note that according to Proposition 17.1.1 the space Jω (W ) of all
almost complex structures J : T W → T W compatible with ω|Tx W on every

tangent space Tx W, x ∈ W , of a symplectic manifold (W, ω) is nonempty
and contractible. In particular, any symplectic manifold (M, ω) admits a
compatible almost complex structure J, and thus (W, ω, J) is an almost
Kähler manifold.
3. Symplectic bundle over a symplectic manifold The following observation
belongs to W.P. Thurston (see [Th76]):
17.2.2. Let (V, ω) be a symplectic manifold and π : X → V a symplectic
on a neighborhood
vector bundle. Then there exists a symplectic structure ω
Op V of the 0-section V ⊂ X such that ω |V = ωV and the restriction of ω

to the fibers of the fibration X defines their linear symplectic structure.
Proof. Note that any vector space E has a canonical radial vector field
Z ∈ T E such that Zx (E) is the vector x ∈ T0 E parallel transported to the
point x. Similarly the total space E of any vector bundle has a tangent to its
fibers vector field which restricts to fibers as the canonical radial field. Let
Z ∈ T X be such a field for the vector bundle π : X → V . By definition the
total space X of a symplectic vector bundle admits a 2-form η such that its
restriction to the fibers of the fibration defines there the linear symplectic
structure, and the restriction of η to the 0-section V vanishes. Take the
1-form λ = Z η. Then the exact form η := dλ is fiberwise, coincides with
η, and also vanishes on V . Hence, the form η + π ∗ ωV is symplectic on a
neighborhood Op V of the 0-section and has the required properties.
H. Embeddings and immersions into (almost) symplectic and (al-

most) complex manifold. In an obvious way we define Lagrangian, iso-
tropic and coisotropic immersions into (almost) symplectic manifolds and
totally real and co-real immersions into (almost) complex manifolds.
When considering (almost) symplectic immersions into an (almost) sym-
plectic manifold (W, ωW ), we should distinguish between immersions which
induce on the source an (almost) symplectic structure, and immersions of
another symplectic manifold which induce on it the (almost) symplectic
structure which was a priori given. In the same way we should treat (al-
most) complex immersions into an (almost) complex manifold.
Let (W, ωW ) be a symplectic manifold. A map f : V → (W, ωW ) is called
symplectic if the form f ∗ ωW is nondegenerate (and thus symplectic). Such
a map is automatically an immersion. If V is already endowed with an-
other symplectic structure ωV , then we call a map f : (V, ωV ) → (W, ωW )
isometric symplectic, or isosymplectic, if f ∗ ωW = ωV .
In the same way one can define complex and isocomplex immersions into a
complex manifold.
17.3. Symplectic stability 229
Exercise. Define all kinds of the respective formal immersions (Lagran-

gian, symplectic, etc.) and the corresponding differential relations. Deter-
mine which of these relations are open, and which are invariant with respect
to Diff V .
17.3. Symplectic stability

By symplectic stability we mean the absence of nontrivial local invariants for
objects related to integrable symplectic structures. Here the locality refers
both to the manifold itself and to the space of symplectic structures.
For a smooth family of differential p-forms ωt on a manifold W the time-

derivative
d
ω̇t = ωt
dt
is again a family of p-forms. Note that the time-derivative commutes with
the exterior differentiation:
˙ t ).
d(ω̇t ) = (dω
For a (time-dependent) vector field vt on W denote by ϕt the flow on W
generated by vt , i.e., ϕt is determined by the differential equation
ϕ̇t (x) = vt (ϕt (x)), ϕ0 (x) = x.
The isotopy (flow) ϕt is always defined at least locally. For a time-indepen-
dent vector field v we will also use the notation etv for its flow ϕt .
Given a differential p-form ω and a vector field X on W , the derivative

LX ω = ω̇t | t=0 , ωt = (etX )∗ ω
is called the Lie derivative of the form ω along X. Let us recall E. Cartan’s
formula:
LX ω = X dω + d(X ω).
Suppose we have a homotopy ωt , t ∈ [0, 1], of differential p-forms on W , can
we find an isotopy ϕt : W → W such that ϕ∗t ωt = ω0 ? It is sufficient to find
a corresponding time-dependent vector field vt which generates the flow ϕt
(provided that vt is integrable). Differentiation of the equation ω0 = ϕ∗t ωt
with respect to t gives us the equation
d
0= (ϕ∗ ωt ) = ϕ∗t (Lvt ωt + ω̇t ) , or
dt t
Lvt ωt = −ω̇t for all t ∈ [0, 1].
The following proposition is the starting point for the symplectic stability
results.
17.3.1. (Solution to the equation Lvt ωt = −ω̇t for an exact homo-

topy of symplectic forms) Let ωt = ω0 + dαt be a smooth family of
symplectic forms on a manifold W and vt = −Iω−1
t
(α̇t ), the vector field that
is ωt -dual to the 1-form −α̇t . Then
Lvt ωt = −ω̇t for all t ∈ [0, 1].
Indeed, we have
˙ t ) = −ω̇t .
Lvt ωt = d(vt ωt ) = −d(α̇t ) = −(dα
Here are a few remarkable corollaries of this proposition:
17.3.2. (Symplectic Stability Theorems)

(Stability near a compact set) Let A ⊂ W be a compact subset. Let
ωt = ω0 + dαt be a family of symplectic forms on Op A ⊂ W such that
αt |T W |A = 0. Then there exists an isotopy ϕt : Op A → W , fixed on A, such
that ϕ∗t ωt = ω0 for all t ∈ [0, 1].
(Darboux’s Theorem) Any symplectic form is locally equivalent to the

n
form ω0 = dxi ∧ dyi on R2n = T ∗ Rn .
1
(Moser’s Theorem) Let ωt = ω0 + dαt be a family of symplectic forms on

a closed manifold W . Then there exists a canonical isotopy ϕt : W → W
such that ϕ∗t ωt = ω0 .
(Relative Moser’s Theorem) Let ωt be a family of symplectic forms on
a compact manifold W with boundary such that ωt = ω0 on Op ∂W and the
relative cohomology class [ωt − ω0 ] ∈ H 2 (W, ∂W ) vanishes for all t ∈ [0, 1].
Then there exists an isotopy ϕt : W → W which is fixed on Op ∂W and
such that ϕ∗t ωt = ω0 .
(Weinstein’s Theorem) Any isotropic immersion L → W (in particular,
any Lagrangian one) extends to an isosymplectic immersion Op L → W ,
where Op L is a neighborhood of the zero-section in the cotangent bundle
T ∗ L endowed with its canonical symplectic structure.
(Symplectic Neighborhood Theorem) Let f : (V, ωV ) → (W, ωW ) be an
isosymplectic immersion and E → V be the symplectic vector bundle whose
fiber over a point v ∈ V is the space (df (Tv V ))⊥ωW which is ωW -orthogonal
to df (Tv V ) ⊂ Tf (v) W . Then f extends to an isosymplectic immersion
f : (Op V, ωE ) → (W, ωW ),
where Op V is a neighborhood of the 0-section V in the total space of the
symplectic vector bundle E → V , and ωE is the symplectic form on Op V ⊂
E constructed in Lemma 17.2.2.
17.3. Symplectic stability 231
Remark. All the above results hold in parametric form. It is important,

however, to keep in mind that in the parametric version of the Symplectic
Neighborhood Theorem the symplectic bundle E also varies with the parame-
ter, and even though all of them are equivalent there could be a homotopical
obstruction if one tries to find a uniformization by a fixed symplectic model.
Proof.
(Stability near a compact set) The isotopy
ϕt : Op A → W
generated by the vector field vt = −Iω−1

t
(α̇t ) is fixed on A and thus is defined
on Op A for all t ∈ [0, 1].
(Darboux) All linear symplectic forms are equivalent. Hence, we can pull
back the symplectic form in local coordinates to ω on R2n which coincides
with ω0 at the origin. Then the linear homotopy
ωt = (1 − t)ω0 + tω
consists of symplectic forms on Op 0. Moreover, ωt = ω0 + dαt , where

αt (0) = 0 for all t ∈ [0, 1]. Hence, we can apply stability near A = {0}.
(Moser) The vector field vt = −Iω−1t
(α̇t ) here integrates on the whole mani-
fold W to an isotopy ϕt such that ϕ∗t ωt = ω0 for all t ∈ [0, 1].
(Relative Moser) The proof is the same as in the absolute case with the
additional remark that the forms αt can be chosen equal to 0 on Op ∂W ,
and then vt |Op ∂W = 0 as well.
(Weinstein) Let f : L → W be an isotropic immersion with dim L = k.
For any ω-compatible almost complex structure J on W , θ = J(f∗ T L) is
a transverse isotropic k-plane field along f such that the bundle T L ⊕ θ
is symplectic with respect to the symplectic form induced by ω. Thus f
extends to a symplectic immersion f : T ∗ L → W . The pullback f∗ ω is a
symplectic form on T ∗ L which vanishes on the zero-section L ⊂ T ∗ L, in
particular it is exact. Now apply stability near L to f∗ ω and the standard
form ω0 on T ∗ L to modify f to an isosymplectic immersion of Op L.
(Symplectic Neighborhood) By definition of the bundle E → V the
immersion f extends to an immersion f : Op V → W such that f∗ ωW
coincides with ωE on T E|V . Hence f∗ ωW − ωE = dα where the 1-form α
vanishes on T E|V , and therefore we can construct the required isotopy using
stability near V .
17.4. Contact manifolds

A. Contact forms and contact structures on manifolds. A 1-form α
on a (2n+1)-dimensional manifold V is called contact if the restriction of dα
to the (2n)-dimensional tangent distribution ξα = Ker α is nondegenerate
(and hence symplectic). Equivalently, we can say that a 1-form α is contact
if α ∧ (dα)n does not vanish on V . A codimension 1 tangent distribution ξ
on V is called a contact structure or a contact distribution if it can be locally
(and in the coorientable case globally) defined by the Pfaffian equation α = 0
for some choice of a contact form α. The pair (V, ξ) in this case is called a
contact manifold.
Example. The standard contact structure ξ0 on R2n+1 is defined by the
Darboux contact 1-form
n
α0 = dz − yi dxi
1
in the coordinates (x1 , . . . , xn , y1 , . . . , yn , z) (see Figure 17.1).
Figure 17.1. The standard contact structure dz − ydx = 0 on R3 .
B. Nonexistence of local contact geometry. Similarly to the symplec-

tic case, contact manifolds have no local geometry: according to a contact
version of Darboux’s theorem (see Theorem 17.5.2 below), any (2n + 1)-
dimensional contact manifold is locally isomorphic to the standard contact
R2n+1 . Thus, equivalently, a contact manifold can be defined by a con-
tact atlas which consists of Darboux charts glued by contactomorphisms,
i.e., diffeomorphisms preserving the standard contact structure ξ0 (but not
17.4. Contact manifolds 233
necessarily the contact form α0 !). The contact forms have no local invari-
ant either: any contact form on a (2n + 1)-dimensional manifold is locally
isomorphic to the standard contact form α0 on R2n+1 (see Theorem 17.5.2
below).
C. Orientations and conformal class CS(ξ) associated with a contact

structure ξ. Any contact form α on V = V 2n+1 defines an orientation of
V by the volume form α ∧ (dα)n , an orientation of the distribution ξ =
Ker α by the form (dα)n and (of course) a coorientation of ξ. Some of
these orientations survive when we pass to contact structures: any contact
structure ξ defines an orientation of the (2n + 1)-dimensional manifold V if
n is odd, and an orientation of ξ if n is even.
The symplectic structure dα|ξ on ξ = Ker α almost survives when we pass

from a contact form α to the contact structure ξ = Ker α which this form
defines: the conformal class of dα|ξ depends only on ξ (because d(f α)|ξ =
f dα|ξ for a function f : V → R). We denote this class by CS(ξ).
In a more invariant fashion we can canonically associate with any codimen-

sion 1 tangent distribution ξ ⊂ T V a defining 1-form α valued in the line
bundle λ = T V /ξ. The differential of this form restricted to ξ is a 2-form ω
on ξ valued in λ. Its value on vectors X, Y ∈ ξ is defined by first extending
Y tangent to ξ, then taking their Lie bracket
X, Y locally to vector fields X,
with the minus sign: −[X, Y ] = Y X − XY , and finally projecting it to
λ = T V /ξ. If ξ is a contact structure, then the form ω is symplectic. Note
that in this definition ω depends only on ξ. A choice of an R-valued contact
form trivializes (and, in particular, normalizes) λ and thus allows us to treat
ω as a usual R-valued symplectic form on ξ which is, as a trade-off, only a
conformal invariant of ξ.
Given a cooriented contact structure ξ+ = Ker α, the positive-conformal

class of the symplectic structure dα|ξ depends only on ξ+ ; we denote this
class by CS(ξ+ ).
D. Integral submanifolds of a contact distribution. According to

Frobenius’s theorem, the contact condition is a condition of maximal non-
integrability of the tangent hyperplane field ξ. In particular, all integral
submanifolds of ξ have dimension ≤ n. On the other hand, n-dimensional
integral submanifolds, called Legendrian, always exist in abundance; see
Section 17.6 below. Integral submanifolds of dimension < n are called sub-
critical.
E. Reeb vector field of a contact form. Any contact form α on V

defines a one-dimensional distribution
Ker dα ⊂ T V
and hence a one-dimensional foliation Rα on V , which is called the Reeb
foliation. Note that Rα is transverse to the contact distribution ξα = Ker α.
The condition α(X) = 1 defines a unique vector field X = Rα tangent to
Rα ; this vector field is called the Reeb vector field. The flow of Rα preserves
the contact form α.
The Reeb foliation Rα does not survive when we pass from a contact form
α to the underlying contact structure ξ = Ker α; however, some of its in-
variants do, see [EGH00].
F. Examples of contact manifolds.

1. Contactization of the cotangent bundle
The canonical contact structure on the 1-jet space J 1 (M, R) = T ∗ M × R is
defined by the contact form dz − p dq on T ∗ M × R, where the coordinate
z corresponds to the second factor, and where we identify the form p dq on
T ∗ M with its pullback to J 1 (M, R). For M = Rn this structure coincides
with the standard contact structure on R2n+1 .
2. Contactization of an exact symplectic manifold
Similarly to the above contactization of the cotangent bundle (T ∗ M, d(p dq))
any exact symplectic manifold (W, dα) can be contactized to a contact mani-
fold
(V = W × R, ξ = Ker (dz − α)),
or to a contact manifold
(V = W × (R/Z), ξ = Ker (dz − α)).
3. Prequantization of an integral symplectic manifold
The latter cyclic form of the contactization (or, as it is otherwise called,
prequantization) construction can be generalized to a symplectic manifold
(W, ω) whose form is not exact but integral, i.e., which belongs to a cohomol-
ogy class from H 2 (W ; Z). In this case the circle bundle V → W with first
Chern class equal to [ω] admits an S 1 -connection whose curvature is equal
to ω. This connection, viewed as a family of horizontal planes, represents
an S 1 -invariant contact structure ξ on V .
4. Space of contact elements of a smooth manifold
A point of the projectivized cotangent bundle P T ∗ M is a tangent hyperplane
to M , which can be identified with a line in T ∗ M . The canonical 1-form p dq
does not descend to P T ∗ M , but its kernel does and thus it defines a canonical
17.4. Contact manifolds 235
contact structure on P T ∗ M . This contact structure is not coorientable. The

double cover of P T ∗ M , which carries a coorientable contact structure, is
the associated spherical bundle ST ∗ M which can be viewed as the space of
cooriented tangent hyperplanes (cooriented contact elements).
Exercise. Find a contact embedding of the standard contact structures

on J 1 (M.R) into ST ∗ (M × R).
5. Strictly pseudo-convex hypersurfaces

As we already explained above in Section 17.2, complex geometry serves
as a rich source of examples of symplectic manifolds. It is a rich source
of examples of contact manifolds as well, because the CR-structure ξ =
T S ∩ JT S on a strictly pseudo-convex hypersurface S in a complex manifold
V is a contact structure on S. In particular, the standard sphere S 2n−1 ⊂ Cn
carries a canonical contact structure.
As in the symplectic case, the above example of complex analytic origin
can in a certain sense be reversed. If (V, ξ) is a contact manifold, then
we can endow the bundle ξ in a homotopically unique way with a complex
structure J compatible with ξ in the following sense: the Hermitian (Levi)
form ω(X, JY ) − iω(X, Y ), where ω ∈ CS(ξ), is positive definite. In other
words, any contact manifold can be viewed as a strictly pseudo-convex but
not necessarily integrable CR-manifold, or as a strictly pseudo-convex hy-
persurface in an almost complex manifold. Note that if dim V = 3, then
the CR-structure can be chosen integrable; see [El85].
Exercise. Find an explicit contactomorphism between the contact struc-

ture on S 2n−1 \ {p} ⊂ Cn defined by its CR-structure and the standard
contact structure on R2n−1 .
6. Symplectic bundle over a contact manifold

The following observation is the contact analogue of Lemma 17.2.2.
17.4.1. Let (V, ξ) be a contact manifold and π : E → V a symplectic vector
bundle. Then there exists a contact structure ξ on a neighborhood U of the
V = ξ, ξ is tangent to the fibers of the fibration E
0-section V such that ξ|
viewed as a subbundle of T E|V , and the given symplectic structure on these

fibers belongs to the conformal class CS(ξ).
Proof. We will consider here only the case of a cooriented ξ. Let α be the
contact form which defines ξ, and η a closed form on E such that
• its restriction to the fibers of the fibration defines there the given linear
symplectic structure;
• the restriction of η to the 0-section V vanishes.
Then the form η is exact, and by the Poincaré lemma there exists a primitive
β of η such that β|T E|V = 0. The form β + π ∗ α defines on a neighborhood
Op V a contact structure ξ with the required properties.
G. Symplectization and projectivization.

1. Symplectization
Let (V, ξ) be a (2n − 1)-dimensional contact manifold. The 2n-dimensional
manifold V = (T V /ξ)∗ \ V , called the symplectization of (V, ξ), carries a
natural symplectic structure ω induced by a tautological embedding V →
T ∗ V which assigns to each linear form
T V /ξ → R
the corresponding form
T V → T V /ξ → R.
Moreover, the symplectization V carries a canonical 1-form αξ induced from
the canonical 1-form p dq on T ∗ V by this embedding. A choice of a contact
form α (if ξ is coorientable) defines a section V → V and, in particular, a
splitting
V = V × (R \ 0).
In this case we can choose the positive half V × R+ of V and call it symplec-
tization as well. The symplectic structure ω can be written in terms of this
∂
splitting as d(τ α), τ > 0. Note that the vector field T = τ ∂τ is conformally
symplectic or, as it is also called, Liouville: we have
LT ω = ω as well as LT (τ α) = τ α.
All the notions of contact geometry can be formulated as the corresponding
symplectic notions, invariant or equivariant with respect to this conformal
action. For instance, any contact diffeomorphism of V lifts to an equivari-
ant symplectomorphism of V ; Legendrian submanifolds in V correspond to
cylindrical (i.e., invariant with respect to the R+ -action) Lagrangian sub-
manifolds of V .
2. Projectivization
From the point of view desribed above, contact geometry can be viewed as
projectivized symplectic geometry. Suppose the multiplicative group R+ or
R∗ = R \ 0, denoted G, acts on a 2n-dimensional symplectic manifold (W, ω)
by conformally symplectic transformation, i.e., λ∗ ω = λω, λ ∈ G. If the
quotient space V = W/G is Hausdorff, then it is automatically a (2n − 1)-
dimensional contact manifold. The contact plane ξv ∈ Tv V , v ∈ V , can be
defined as follows. Take any point w from the orbit v ⊂ W and consider
the space Nw = L⊥ w where Lw ⊂ Tw W is the tangent line to the orbit v at
ω
the point w ∈ v. Then ξv is the image dπ(Nw ), where π : W → V = W/G

17.5. Contact stability 237
is the canonical projection. Note that the standard contact structure on

the space V = P T ∗ M of contact elements or on the space V = ST ∗ M
of cooriented contact elements is obtained by the above construction from
W = (T ∗ M \ M, d(p dq)) using, respectively, the canonical action of R∗ or
R+ .
H. Embeddings and immersions into contact manifolds. An immer-

sion f : V → (W, ξ) is called isotropic if its differential df maps T V to
ξ ⊂ T W . An isotropic immersion is called Legendrian (resp., subcritical ) if
dim V = n (resp., dim V < n). It is important to observe that the differ-
ential df of an isotropic immersion f maps tangent spaces of V to subspaces
of ξ which are isotropic with respect to the conformal symplectic structure
CS(ξ). Indeed, we have f ∗ (dα) = df ∗ (α) = 0, where ξ = Ker α. In partic-
ular, the differential relation RLeg ⊂ J 1 (V, W ) responsible for Legendrian
immersions consists of monomorphisms (fiberwise injective homomorphisms)
Tv V → ξ onto Lagrangian subspaces of ξ.
For a contact manifold (W, ξ) a map f : V → (W, ξ) of an odd-dimensional
manifold V is called contact if it induces a contact structure on V . Such a
map is automatically an immersion which is transverse to ξ.
It is important to observe that the intersection df (T V ) ∩ ξ consists of sym-
plectic subspaces with respect to the conformal symplectic structure CS(ξ).
A monomorphism F : T V → T W is called contact if it is transverse to
ξ (i.e., F −1 (ξ) is a codimension 1 distribution on V ) and the intersection
F (T V )∩ξ consists of symplectic subspaces of ξ with respect to the conformal
symplectic structure CS(ξ).
If the manifold V itself has a contact structure, then we may consider iso-
metric contact or isocontact maps f : (V, ξV ) → (W, ξW ) which induce on
V the given structure ξV . If ξW and ξV are given by contact forms αW and
αV , then equivalently one can say that f is isocontact if f ∗ αW =ϕαV where
ϕ : V → R is a nonvanishing function. A monomorphism F : T V → T W is
called isocontact if ξV = F −1 (ξW ) and F induces a conformally symplectic
map ξV → ξW with respect to the conformal symplectic structures CS(ξV )
and CS(ξW ).
Exercise. Which of the differential relations RLeg , Rcont , Risocont are
open? Which are invariant with respect to Diff V ?
17.5. Contact stability

By contact stability we mean the absence of nontrivial local invariants for
objects related to contact structures. Here the locality refers both to the
manifold itself and to the space of contact structures.
One can prove for contact structures stability results similar to Theorem
17.3.2 in the symplectic case.
Suppose we have a homotopy αt , t ∈ [0, 1], of differential 1-forms on W . Can

we conformally realize this homotopy by an isotopy ϕt : W → W , i.e., to
satisfy the equation gt ϕ∗t αt = α0 for positive functions gt ? It is sufficient to
find a corresponding time-dependent vector field vt generating the isotopy
ϕt (provided vt is integrable). Differentiation of the equation α0 = gt ϕ∗t αt
with respect to t gives us
d ∗
0 = ġt ϕ∗t αt + gt (ϕ αt ) = gt ϕ∗t (Lvt αt + α̇t + ht αt ) ,
dt t
ġt
where ht ◦ ϕt = gt . Equivalently, we get
Lvt αt = −α̇t − ht αt for all t ∈ [0, 1].
Note that if vt and ht satisfying this equation are found, then gt can be
t
hs ◦ϕs ds
reconstructed by the formula gt = e0 .
17.5.1. (Solution to the equation Lvt αt = −α̇t − ht αt for a homotopy

of contact forms) Let αt be a family of contact forms on a manifold W
and ξt the family of contact structures defined by these forms. Let vt be the
vector field on V which is characterized by the conditions
• αt (vt ) = 0 and
−1
• vt = −I(dα t )|ξ
(α̇t |ξt ).
t
Then Lvt αt = −α̇t − ht αt for a function ht : W → R.
Proof. By Cartan’s formula we have
Lvt αt = vt dαt + d(αt (vt )) = vt dαt .
But
(vt dαt )|ξt = vt (dαt |ξt ) = −α̇t |ξt ,
hence
vt dαt = −α̇t − ht αt .
The following theorem can be deduced from Lemma 17.5.1 in the same way
as Theorem 17.3.2 was deduced from Proposition 17.3.1.
17.5. Contact stability 239
17.5.2. (Contact stability theorems)

(Stability near a compact subset) Let αt , t ∈ [0, 1], be a family of contact
forms given in a neighborhood Op A ⊂ W of a compact subset A ⊂ W such
that αt |T W |A = ξ0 . Then there exists an isotopy ϕt : Op A → W which is
fixed on A such that ϕ∗t αt = α0 for all t ∈ [0, 1].
(Darboux’s theorem) Any contact structure on a (2n + 1)-dimensional
manifold is locally equivalent to the standard contact structure ξ0 on R2n+1 .
Moreover, locally any contact form on a (2n + 1)-dimensional manifold is
n
equivalent to the standard contact form α0 = {dz − yi dxi } on R2n+1 .
1
(Gray’s theorem) Let ξt , t ∈ [0, 1], be a family of contact structures on

a closed manifold V . Then there exists an isotopy ϕt : V → V such that
ϕ∗t ξt = ξ0 .
(Relative Gray’s theorem) Let ξt , t ∈ [0, 1], be a family of contact struc-
tures on a compact manifold V with boundary such that ξt ≡ ξ0 on Op ∂V .
Then there exists an isotopy ϕt : V → V fixed on Op ∂V such that ϕ∗t ξt = ξ0 .
(Contact Weinstein’s theorem) Any isotropic immersion L → W , in
particular a Legendrian one, extends to an isocontact immersion Op L → W ,
where Op L is a neighborhood of the zero-section L in the 1-jet space J 1 (L)
endowed with its canonical contact structure.
(Contact neighborhood theorem) Let f : (V, ξ) → (W, ξ ) be an iso-
contact immersion. Let E → V be the symplectic vector bundle whose fiber
over a point v ∈ V is the space CS(ξf (v) )-dual to df (ξv ) ⊂ ξf (v) . Then f
extends to an isocontact immersion f : (Op V, ξ) → (W, ξ ) where Op V is
a neighborhood of the 0-section V in the total space of the symplectic vector
bundle E → V , and ξ is the contact structure on Op V ⊂ E constructed in
Lemma 17.4.1.
Remarks.
1 All the above statements hold parametrically. See, however, the
remark after Theorem 17.3.2.
2 The equation Lvt αt = −α̇t , which defines an isotopy preserving the
form α rather the contact structure Ker α, can be solved near any
submanifold L ⊂ V provided the Reeb vector field Rα is transverse
to L. In particular, the stability claim in Darboux’s theorem and
the contact Weinstein theorem holds for contact forms and not only
contact structures.

17.6. Lagrangian and Legendrian submanifolds

Lagrangian submanifolds of symplectic manifolds and Legendrian submani-
folds of contact manifolds play a central role in symplectic topology.
A section s : M → T ∗ M is a Lagrangian embedding if and only if s is a

closed 1-form. In fact, any Lagrangian submanifold L of T ∗ M can be viewed
as a multi-valued closed 1-form. Any Lagrangian submanifold L ⊂ T ∗ M can
equivalently be characterized by the condition that the restriction p dq|L is
a closed 1-form on L. A Lagrangian submanifold L ⊂ T ∗ M is called exact
if the closed 1-form p dq|L is exact. A Lagrangian immersion f : L → T ∗ M
is called exact if the closed 1-form f ∗ (p dq) is exact.
A section s : M → J 1 (M, R) is a Legendrian embedding if and only if s is

the 1-jet extension Jf1 of a function f : M → R. In other words, Legendrian
sections coincide with the holonomic sections of J 1 (M, R). Similarly to
the case of Lagrangian submanifolds of the cotangent bundle, a general
Legendrian submanifold of J 1 (M ) corresponds to a graph (wave front) of
a multi-valued function. The projection J 1 (M ) = T ∗ M × R → T ∗ M sends
Legendrian submanifolds of J 1 (M, R) onto (immersed) exact Lagrangian
submanifolds of T ∗ M . Conversely, any exact Lagrangian submanifold of
T ∗ M lifts, uniquely up to a translation along the R-factor, to a Legendrian
submanifold of J 1 (M, R).
More generally, let (W, ω) be a symplectic manifold with an exact symplectic

form ω = dα. A choice of a primitive α is sometimes referred to as a Liouville
structure on W . Let us associate with a Liouville manifold (W, dα) the
contact manifold
, = W × R, ξ = Ker (dz − α)),
(W
, → W be
where z denotes the coordinate along the second factor. Let π : W
the projection to the first factor. A Lagrangian immersion
f : L → (W, dα)
is called exact if the form f ∗ α is exact. If H1 (L; R) = 0, then any Lagrangian
immersion L → (W, dα) is exact. The projection f = π ◦ f of a Legendrian
immersion f : L → (W , , ξ) to W is an exact Lagrangian immersion. Con-
versely, any exact Lagrangian immersion f : L → W lifts uniquely, up to
a translation along the R-factor, to a Legendrian immersion f : L → W ,
by the formula f = (f, H), where dH = f ∗ α. The relation between exact
Lagrangian immersions into W and Legendrian immersions into W will be
exploited later in Chapter 24.
17.7. Hamiltonian and contact vector fields 241
We will show below (see Sections 22.1 and 24.1) that isotropic and, in par-
ticular, Legendrian immersions into a contact manifold satisfy all forms of
the h-principle. On the other hand, the literally understood h-principle fails
for isotropic immersions into a symplectic manifold. Indeed, in order that
a map f : L → (W, ω) be homotopic to an isotropic immersion, we have a
necessary cohomological condition f ∗ [ω] = 0 ∈ H 2 (L). Similarly, given a
map f : (L, A) → W which is an isotropic immersion on a neighborhood
Op A of a polyhedron A ⊂ L, vanishing of the relative cohomology class
[f ∗ ω] ∈ H 2 (L, A) is necessary in order that the map f be homotopic rel A
to an isotropic immersion. As we will see in Sections 22.1 and 24.3, the
isotropic and, in particular, Lagrangian immersions into symplectic mani-
folds satisfy a modified form of the h-principle augmented by these additional
necessary conditions.
Subcritical isotropic embeddings into contact and symplectic manifolds also

satisfy some forms of the h-principle (see Section 20.4). However, the prob-
lems of Lagrangian embeddings and Legendrian isotopy belong to the world
of symplectic and contact rigidity.
17.7. Hamiltonian and contact vector fields

A. Hamiltonian vector fields. A vector field X on a symplectic mani-
fold (M, ω) is called symplectic if the Lie derivative LX ω vanishes, which is
equivalent to the equation
d(X ω) = 0.
If the form X ω is exact, i.e., X ω = −dH, then the vector field X = XH
is called Hamiltonian and the function H is called the Hamiltonian function
for the vector field X. If M is noncompact, then we will always assume that
X and H have compact supports. This condition defines H uniquely. On
the other hand, if M is closed, then H is defined up to an additive constant.
A time-dependent Hamiltonian function

Ht : M → R, t ∈ [0, 1],
defines via ϕ̇t = XHt (ϕt ), ϕ0 = Id a symplectic isotopy
ϕt : M → M, t ∈ [0, 1].
The isotopy ϕt is called a Hamiltonian isotopy with the time-dependent
Hamiltonian function Ht , t ∈ [0, 1]. A symplectomorphism ϕ : M → M is
called Hamiltonian if it is the time 1 map of a Hamiltonian isotopy. The
group
Ham = Ham(M, ω)
of Hamiltonian diffeomorphisms of (M, ω) is a normal subgroup of the iden-

tity component of the group of symplectomorphisms of (M, ω). According
to a theorem of A. Banyaga (see [Ba78]), Ham = [Diff ω , Diff ω ].
It will be important for our applications to observe the following simple fact.
17.7.1. (Symplectic cutting-off ) Let X be a Hamiltonian vector field
on a symplectic manifold V . Then for any compact subset A ⊂ V and its
neighborhood U ⊃ A there exists a Hamiltonian vector field X̃A,U which is
supported in U and which coincides with X on A.
Indeed, any Hamiltonian function can be cut off to 0 away from any given
neighborhood U ⊃ A.
B. Contact vector fields. A vector field X on a contact manifold (V, ξ) is

called contact if its flow (which is always defined at least locally) preserves
the contact structure ξ. If ξ = Ker α, then this is equivalent to the equation
LX α = hα for a function h : V → R. Any contact vector field X on V
lifts to a Hamiltonian vector field X on the symplectization V of V . The
Hamiltonian KX which defines the field X equals αξ (X) where αξ is the
canonical 1-form on the symplectization V . Note that αξ (X) is a function
on V homogeneous of degree 1 (with respect to the canonical R+ -action on
V ). Conversely, any Hamiltonian H : V → R that is homogeneous of degree
1 defines a vector field which projects to a contact vector field on V . A
choice of a contact form α for ξ defines an embedding fα : V → V and a
decomposition
X = α(X)Rα + Yα ,
where Rα is the Reeb vector field and Yα ∈ ξ. The function KX,α : V → R
induced by the embedding fα from the Hamiltonian function KX is equal
to α(X). It is called the contact Hamiltonian for the contact vector field X
(with respect to α). Thus the contact Hamiltonian measures the component
transverse to ξ of the field X. It follows that the transverse component of X
completely determines X. In particular, there are no nonzero contact vector
fields which are everywhere tangent to ξ. Note that the Reeb vector field Rα
is itself a contact vector field which is defined by the contact Hamiltonian
K ≡ 1. Any contact vector field X which is transverse to ξ is the Reeb
vector field for the contact form α, Ker α = ξ, characterized by the equation
α(X) = 1.
Similarly to the symplectic case we have the following.

17.7.2. (Contact cutting-off ) Let X be a contact vector field on a contact
manifold V . Then for any compact subset A ⊂ V and its neighborhood
17.8. Characteristic foliations 243
U ⊃ A there exists a contact vector field X̃A,U which is supported in U and

which coincides with X on A.
Indeed, any contact Hamiltonian function can be cut off to 0 away from any
given neighborhood U ⊃ A.
17.8. Characteristic foliations

Recall from Section 17.2.F that a hypersurface Σ in a symplectic mani-
fold (V, ω) carries a one-dimensional characteristic foliation L defined by
ker(ω|Σ ). The following two properties follow directly from the definitions.
• If Σ = {H = 0} is the regular zero set of a function H, then the
Hamiltonian vector field of H along Σ generates L.
• A codimension 1 submanifold V ⊂ Σ is transverse to L if and only
if V is a symplectic submanifold of (V, ω).
Next, consider a hypersurface Σ in a contact manifold (V, ξ = ker α). As-
sume that Σ is transverse to the contact structure, so that ξΣ = ξ ∩ T Σ
defines a codimension 1 distribution on Σ. The 2-form dα|ξΣ has a one-
dimensional kernel ker(dα|ξΣ ) ⊂ ξΣ whose integral curves define the contact
characteristic foliation L of Σ. Again, we have the following two properties.
• If Σ = {H = 0} is the regular zero set of a function H, then the
contact vector field of H along Σ generates L.
• A codimension 1 submanifold V ⊂ Σ is transverse to L if and only
if V is a contact submanifold of (V, ξ).
Let V be the symplectization of the contact manifold (V, ξ) and p : V → V
be the canonical projection. For a hypersurface Σ ⊂ V consider its pre-
image Σ := p−1 (Σ) ⊂ V . Let L be the symplectic characteristic foliation on
Σ. Note that for any point p ∈ Σ where ξp = Tp Σ the line π −1 (p) ⊂ Σ is a
leaf of the characteristic foliation L. Any other leaf of L injectively projects
onto a leaf of the contact charactristic foliation L on Σ.
Chapter 18
Symplectic and
Contact Structures on
Open Manifolds
18.1. Classification problem for symplectic and contact

structures
A. Symplectic structures. A symplectic structure ω on V defines a vol-
ume form ω n and hence an orientation of V . Thus we should consider only
even-dimensional orientable manifolds. Given such a manifold V , we will
use the following notation:
• J = J (T V )—the space of almost complex structures on V = the

space of complex structures on the tangent bundle T V ;
• Ssymp = S(T V )—the space of almost symplectic structures on V
= the space of symplectic structures on the tangent bundle T V ;
• Ssymp —the space of symplectic structures on V ;
• Sasymp —the space of symplectic structures on V in a given cohomol-
ogy class a ∈ H 2 (V, R).
The existence of an almost symplectic (or, equally, an almost complex) struc-

ture on V is a necessary condition for the existence of a symplectic structure.
The existence of a homotopy in Ssymp which connects two given symplectic
forms ω0 and ω1 is a necessary condition for the existence of a symplectic
homotopy between ω0 and ω1 , and so on. Thus according to the philosophy
of the h-principle, the problem of classification of symplectic structures on
245
246 18. Symplectic and Contact Structures on Open Manifolds
a manifold V up to homotopy can be treated as the study of the homotopy

properties of the natural inclusions
Ssymp → Ssymp or Sasymp → Ssymp .
Remark. Let us recall that J is homotopy equivalent to Ssymp , and

hence, considering the classification of symplectic structures up to homotopy,
one can equally use Ssymp or J as the corresponding formal space. See also
the Remark in Section 17.2.A.
B. Contact structures. An almost contact structure on a (2n + 1)-dimen-

sional manifold V is a codimension 1 tangent distribution ξ on V together
with a symplectic form ω on ξ valued in the one-dimensional bundle λ =
T V /ξ. Note that if n is odd, then an almost contact structure defines an
orientation of V ; if n is even, it defines an orientation of ξ. Note that if
ξ is cooriented, then from the homotopical point of view defining a form ω
valued in λ is the same as defining a positive-conformal class of a symplectic
structure on ξ. Hence we can define a cooriented almost contact structure
on a (2n+1)-dimensional manifold V as a pair (ξ+ , ω) where ξ+ a cooriented
hyperplane distribution on V and ω a positive-conformal class of symplectic
structures on ξ+ .
Given an odd-dimensional manifold V , we will use the following notation:
• Scont —the space of almost contact structures on V ;
• Scont —the space of contact structures on V ;
• Scont
+
—the spaces of cooriented almost contact structures on V ;
• S+
cont —the spaces of cooriented contact structures on V .
According to the philosophy of the h-principle, the problem of homotopical

classification of contact structures on a manifold V can be treated as the
study of the homotopy properties of the natural inclusions
Scont → Scont or S+
cont → Scont .
+
18.2. Symplectic structures on open manifolds

Let us recall that Λp V is a natural vector bundle: any diffeomorphism h :
V → V lifts to Λp V as the exterior power dp h of the differential dh : T V →
T V . Hence we can consider Diff V -invariant subspaces of Λp V .
For a subspace R ⊂ Λp V and a cohomology class a ∈ H p (V ), we denote

by Cloa R a subspace of the space Sec R which consists of closed p-forms
ω : V → R in the cohomology class a.
18.2. Symplectic structures on open manifolds 247
18.2.1. Let V be an open manifold, a ∈ H p (V ) a fixed cohomology class,

and R ⊂ Λp V an open Diff V -invariant subset. Then the inclusion
Cloa R → Sec R
is a homotopy equivalence. In particular,
• any p-form ω : V → R is homotopic in R to a closed p-form ω ∈ a;
• any homotopy of p-forms ωt : V → R which connects two closed
forms ω0 , ω1 ∈ a can be deformed in R into a homotopy of closed
forms ω t ∈ a connecting ω0 and ω1 ∈ a.
Proof. The statement follows almost immediately from Theorem 4.7.4. We

explain the reduction in the nonparametric case; the general case differs
only in notation. Let K ⊂ V be a polyhedron of positive codimension, as
in Theorem 4.4.1. According to Proposition 4.7.2, there exists a diffeotopy
hτ : V → V and a closed form ω ∈ a which is arbitrarily C 0 -close to ω
over a neighborhood U of K = h (K) ⊂ V . Hence over the neighborhood
1
U the linear homotopy ωt between ω and ω lies in R. Let gt : V → V be a

diffeotopy which compresses V into U . Then ω = (g1−1 )∗ ω
is a section of R
and ω ∈ a. Applying consecutively the homotopies (gt−1 )∗ ω and (g1−1 )∗ ωt ,
we get the required homotopy which connects ω and ω in R.
For a 2n-dimensional manifold V let Rsymp ⊂ Λ2 V be defined in every fiber

by the condition β n = 0. Then Cloa Rsymp = Sasymp and Sec Rsymp = Ssymp .
The set Rsymp is open and Diff V -invariant. Therefore, applying 18.2.1, we
get the following homotopy principle for symplectic forms on open manifolds:
18.2.2. (Gromov [Gr69]) For any open manifold V the inclusion

Sasymp → Ssymp
is a homotopy equivalence. In particular,
• any 2-form β ∈ Ssymp is homotopic in Ssymp to a symplectic form
ω, ω ∈ a ∈ H 2 (V );
• if two symplectic forms ω0 , ω1 ∈ Sasymp are homotopic in Ssymp , then
ω0 and ω1 are homotopic in Sasymp .
Using almost complex structures (instead of almost symplectic), we can for-

mulate the existence theorem in the following way: any open almost complex
manifold (M, J) admits a symplectic structure ω which belongs to any pre-
scribed cohomology class a ∈ H 2 (M ) and such that J ∈ [Jω ] where [Jω ]
is the homotopy class of almost complex structures compatible with ω.
It is important to understand that ω may be noncompatible with the
original J.
Exercise. Prove that the inclusion Ssymp → Ssymp is a homotopy equiv-

alence.
18.3. Contact structures on open manifolds
For a subset R ⊂ Λp−1 ⊕ Λp V , let Exa R ⊂ Sec R be the subspace of pairs

(α, β) : V → Λp−1 V ⊕ Λp V
such that β = dα.
18.3.1. Let V be an open manifold and R an open Diff V -invariant subset
of Λp−1 ⊕ Λp V . Then the inclusion
Exa R → Sec R
is a homotopy equivalence. In particular, any section (α, β) : V → R is
homotopic in R to a section (ᾱ, dᾱ) : V → Exa R.
Proof. To simplify the notation, we will discuss only the nonparametric

case. Let K ⊂ V be a polyhedron of positive codimension, as in Theorem
4.4.1. According to Proposition 4.7.1, there exists a diffeotopy hτ : V → V
and a (p − 1)-form α such that the pair ( α) is arbitrarily C 0 -close to
α, d

(α, β) over a neighborhood U of K = h (K) ⊂ V . Hence over U the linear
1
homotopy (αt , βt ) between (α, β) and ( α) lies in R. Let gt : V → V be

α, d
a diffeotopy which compresses V into U . Then
(ᾱ, dᾱ) = (g1−1 )∗ (
α, d
α)
is a section of R. Consecutively applying the homotopies (gt−1 )∗ (α, β) and
(g1−1 )∗ (αt , βt ), we construct the required homotopy which connects (α, β)
and (ᾱ, dᾱ) in R.
Now suppose dim V = 2n + 1 and let the set

Rcont ⊂ Λ1 V ⊕ Λ2 V
be defined in every fiber by the condition α ∧ (β n ) = 0. Then we have the
commutative diagram
Rcont → Sec R
Exa ⏐ ⏐ cont
⏐ ⏐
. .
Scont
+
→ Scont
+
where the vertical arrows are natural homotopy equivalences. The set Rcont
is open and Diff V -invariant. Hence, Theorem 18.3.1 implies the following
homotopy principle for cooriented contact structures on open manifolds.
18.4. Two-forms of maximal rank on odd-dimensional manifolds 249
18.3.2. (Gromov [Gr69]) For any open manifold V the embedding
S+
cont → Scont
+
is a homotopy equivalence.
In particular, given an open manifold V , a nonvanishing 1-form α0 on V ,

and an almost complex structure on the bundle ξ0 = Ker α0 , there exists
a family of nonvanishing 1-forms αt on V and a family of almost complex
structures Jt on ξt = Ker αt , t ∈ [0, 1], such that α1 is a contact form and
J1 is compatible with the symplectic form dα1 |ξ1 . Note that even in the case
of an orientable three-dimensional manifold V , the latter statement implies
more than just the existence of a contact structure in every homotopy class
of a coorientable tangent plane field. It asserts, in addition, that such a
structure can be chosen to define an a priori given orientation of the manifold
V.
Theorem 18.3.2 can be generalized to cover the case of not necessarily coori-
entable contact structures. As we explained above in Section 17.4.C, any
tangent hyperplane field ξ on V can be defined by a Pfaffian equation
α = 0, where the 1-form α is valued in the not necessarily trivial line bundle
L = T L/ξ. The contact condition for ξ means, as usual, that dα|ξ is non-
degenerate, where the form dα is also valued in L. It is straightforward to
extend Theorem 18.3.1 to cover the case of relations
R ⊂ (Λp−1 ⊗ L) ⊕ (Λp V ⊗ L),
which then implies the corresponding generalization of Theorem 18.3.2 to

the general case of not necessarily coorientable contact structures.
Both h-principles 18.2.2 and 18.3.2 also hold in the relative version, when
one wants to extend a symplectic or contact structure from a neighborhood
of a subcomplex of codimension > 1.
18.4. Two-forms of maximal rank on odd-dimensional

manifolds
The rank of a differential 2-form is always even, and hence the maximal rank
of a 2-form on a manifold of dimension 2n + 1 equals 2n. Theorem 18.2.2
implies the h-principle for such forms on any (open or closed) manifold. It
was first proved by D. McDuff using the convex integration technique; see
Section 28.5.
Given an odd-dimensional manifold V , we will use the following notation

• Snon-deg —the space of 2-forms on V of maximal rank;
• Sanon-deg —the subspace in Snon-deg which consists of closed forms in
a given cohomology class a ∈ H 2 (V ).
18.4.1. (McDuff [MD87a]) The inclusion
Sanon-deg (V ) → Snon-deg (V )
is a homotopy equivalence. In particular, if V admits a 2-form of maximal
rank, then every two-dimensional cohomology class of V can be represented
by a closed nondegenerate form.
Proof. If V is orientable, then any nondegenerate 2-form on V extends in a

homotopically unique way to a nondegenerate 2-form on V × R. Conversely,
the restriction of a symplectic form on V ×R to V = V ×0 ⊂ V ×R is a non-
degenerate closed 2-form. Hence, the required homotopy equivalence follows
in this case from the h-principle for symplectic forms on open manifolds; see
18.2.2.
If V is nonorientable, then with any nondegenerate 2-form ω on V we as-
sociate its kernel Ker ω, a nonorientable line subbundle of T V . The form
ω homotopically canonically extends as a nondegenerate form to the total
space K of this line bundle. Hence we can apply Theorem 18.2.2 to produce
a symplectic form ω on K in a given cohomology class and then restrict it
back to the 0-section.
The same argument proves the relative version of the above h-principle. A
contact analogue of Theorem 18.4.1 will be discussed in Section 22.2.
Chapter 19
Symplectic and
Contact Structures on
Closed Manifolds
In general the problem of constructing symplectic or contact structures on

closed manifolds, or the problem of extending the structures from a codi-
mension 1 subpolyhedron do not abide by any h-principle. In this section
we review the current knowledge about this subject.
19.1. Symplectic structures on closed manifolds

A. Homotopy and isotopy. Two symplectic forms ω0 and ω1 on a mani-
fold V are called
• homotopic if they are homotopic in Ssymp ;
• formally homotopic if they are homotopic in Ssymp ;
• isotopic if there exists an isotopy ϕt : V → V such that ϕ∗1 ω0 = ω1 .
Theorem 17.3.2 implies that classification of symplectic structures on a
closed manifold V up to homotopy in a fixed cohomology class coincides
with the classification up to isotopy:
ω0 and ω1 are isotopic ⇔ they are homotopic in Sasymp .
B. Existence and uniqueness. For a closed symplectic manifold (V, ω)

the cohomology class [ω] ∈ H 2 (V ; R) represented by the closed form ω sat-
isfies the inequality [ω]n = 0. Hence, if understood literally, the h-principle
251
252 19. Symplectic and Contact Structures on Closed Manifolds
for the inclusions

Ssymp (V ) → Ssymp (V ) and Sasymp (V ) → Ssymp (V )
may fail for an almost symplectic closed manifold V of dimension > 2 be-
cause of the cohomological obstruction: a symplectic candidate V = V 2n
should at least have a cohomology class a ∈ H 2 (V ; R) with an = 0. How-
ever, with this modification the situation becomes far less clear.
(a) Does any closed manifold V 2n which admits an almost symplectic
structure ω0 and a cohomology class a ∈ H 2 (V ) with an = 0 have
a symplectic structure ω1 such that [ω1 ] = a ∈ H 2 (V ) and ω1 is
formally homotopic to ω0 ? Does V admit any symplectic structure
at all ?
(b) Let ω0 , ω1 be two symplectic structures on V such that ω0 is formally
homotopic to ω1 and the cohomology classes [ω0 ] and [ω1 ] coincide.
Are ω0 and ω1 isotopic?
(c) Let ω0 , ω1 be two symplectic structures on V such that ω0 is formally
homotopic to ω1 . Are ω0 and ω1 homotopic?
Exercise. Formulate questions (a)–(c) in terms of homotopy properties
of the inclusions
Ssymp (V ) → Ssymp (V ) and Sasymp (V ) → Ssymp (V ).
The answer to (a) is completely unknown for manifolds of dimension 2n > 4.

Thus it is still possible, though unlikely, that problem (a) abides by the h-
principle. For n = 2 the answer is known to be negative. For instance,
C.H. Taubes (see [Ta94]) proved that the connected sum of odd numbers
of copies of CP 2 has no symplectic structure, while it admits an almost
symplectic structure and also satisfies the cohomological condition.
D. McDuff constructed in [MD87b] an example of a 6-manifold which has
two nonisotopic but homotopic symplectic forms in the same cohomology
class. In particular, this implies the negative answer to (b) for manifolds
of dimension > 4. In dimension 4 problem (b) is still open. Note that C.
McMullen and C. Taubes (see [MT99]) constructed an example of a sim-
ply connected 4-manifold which admits symplectic forms whose first Chern
classes are not equivalent under the action of the diffeomorphism group.
The answer to (c) is negative for manifolds of dimension > 4. A counter-

example was constructed by Y. Ruan (see [Ru94]) using Gromov’s theory
of holomorphic curves in symplectic manifolds.
19.2. Contact structures on closed manifolds 253
C. Extension problems. Let ω be a symplectic form on Op ∂D 2n . If

n = 1 it obviously extends to D 2 . If n > 1, then besides the homotopical
obstruction, i.e., existence of a (not necessarily closed) nondegenerate form
extending ω, there is another obstruction which is similar to the cohomo-
logical obstruction for closed manifolds which we discussed above. Namely,
any extension ω of ω to D 2n is exact, ω
= dθ, and by Stokes’s theorem we
have

=
ω n
θ ∧ ω n−1 .
D 2n ∂D 2n
The first integral is positive. Hence, the second integral is positive as well.
But this integral is independent of the choice of a primitive of the form
ω|∂D2n , and hence it does not depend on the choice of ω . We will refer to
the positivity of this integral as the positivity condition.
(a) Let ω be a symplectic form on Op ∂D 2n , n > 1, which extends

to D 2n as a nondegenerate form and which satisfies the positivity
condition. Does it extend to D 2n as a symplectic form?
(b) Let ω be a symplectic form on D 2n which coincides with the stan-
n
dard symplectic form ω0 = dpi ∧ dqi on Op ∂D 2n . Is there a dif-
1
feomorphism f : D 2n → D 2n fixed on Op ∂D 2n such that f ∗ ω = ω0 ?
Can such an f be chosen isotopic to the identity relative to the
boundary?
The answer to (a) is negative in all dimensions. This can be deduced from
Gromov’s nonsqueezing theorem from [Gr86]; see [El15]. Problem (b) is
open in dimension 2n > 4. In the four-dimensional case a theorem of Gro-
mov (see [Gr85]) asserts that the answer is positive to the question about
existence of a diffeomorphism f with f ∗ ω = ω0 . However, it is unknown
whether f can be chosen isotopic to the identity.
19.2. Contact structures on closed manifolds

A. Homotopy and isotopy. Two contact structures ξ0 and ξ1 on a mani-
fold V are called
• homotopic if they are homotopic in Scont ;
• formally homotopic if they are homotopic in Scont ;
• isotopic if there exists an isotopy ϕt : V → V such that ϕ∗1 ξ0 = ξ1 .
Theorem 17.5.2 implies that the classification of contact structures on a

closed manifold V up to homotopy coincides with the classification up to
isotopy:
ξ0 and ξ1 are isotopic ⇔ they are homotopic.
B. Homotopy principle.
Question. Does the h-principle 18.3.2 hold for closed manifolds V ?
It turns out that the existence h-principle for contact structures on closed
manifolds does hold in all dimensions. In dimension 3 it goes back to
J. Martinet [Ma71] and R. Lutz [Lu77], and for the higher-dimensional
case it was proven in [BEM15].
However, as was discovered by D. Bennequin, the 1-parametric h-principle
fails, at least in the three-dimensional case.
19.2.1. (D. Bennequin [Be83]) There exists a contact structure ζ on S 3 ,
which is homotopic to the standard contact structure ξ on S 3 as a plane field,
which defines the same orientation of S 3 as ξ, but which is not equivalent
to ξ.
In other words, the h-principle 18.3.2 fails on the level of a monomorphism

on π0 . The notion of an overtwisted contact structure, introduced in [El89]
in the three-dimensional case, and in [BEM15] in higher dimensions, sheds
some light onto the phenomenon discovered by Bennequin. It turns out
that in the space of contact structures on a closed manifold there exists
an open-closed subspace of overtwisted contact structures for which the full
parametric h-principle holds.
A three-dimensional contact manifold is called overtwisted if there exists an
embedded disc D ⊂ V which is tangent to ξ along the boundary ∂D. A disc
with this property is also called overtwisted. D. Bennequin proved in [Be83]
that in the standard contact structure ξ on S 3 such an overtwisted disc
cannot be found. An equivalent characterization of an overtwisted contact
structure on a manifold V of dimension 2n + 1 > 3 is the existence of
a contact embedding (U × T ∗ Rn−1 , Ker(α + λ)) into V , where (U, Kerα)

n−1
is an overtwisted contact 3-manifold, and λ = pj dqj is the standard
1
Liouville form. Overtwisted contact structures could be also characterized
by existence of 2n-dimensional embedded overtwisted disc D with a certain
particular germ of contact structure.
19.2.2. ([El89] in dimension 3, and [BEM15] in higher dimensions))
Let V be a connected manifold and let
Contot (V, D) ⊂ Scont
19.3. Folded symplectic and contact structures 255
be the space of contact structures on V which all contain a fixed overtwisted

disc D ⊂ V . Let
Scont (V, D) ⊂ Scont (V )
be the space of almost contact structures on V which coincide with a genuine
contact structure from Contot (V, D) on Op D. Then the inclusion
Contot (V, D) → Scont (V, D)
The complementary class of tight contact structures belongs to the domain

of rigid contact topology, and the h-principle for tight contact structures
fails, at least for some contact manifolds (e.g., spheres), already on the level
an epimorphism on π0 .
C. Extension problem. Does the h-principle hold for the extension prob-
lem of contact structures from Op ∂D 2n+1 to D 2n+1 ? In particular,
(a) Suppose a contact form α on Op ∂D 2n+1 formally extends to D 2n+1 ,
i.e., there exists a pair α̃, ω on D 2n+1 such that the nonvanishing
form α̃ extends α, the 2-form ω extends dα, and ω|ξ̃=Ker ã is non-
degenerate. Does α extend to D 2n+1 as a contact form?
(b) Let α be a contact form on D 2n+1 which coincides over Op ∂D 2n+1
n
with the standard contact 1-form α0 = dz − pi dqi . Suppose
1
that α0 and α are formally homotopic relative to the boundary,
i.e., (α, dα) and (α0 , dα0 ) are homotopic through sections of Rcont
which coincide with (α0 , dα0 ) over Op ∂D 2n+1 . Is there a diffeotopy
of D 2n+1 fixed over Op ∂D 2n+1 which moves α into α0 ?
The situation with these problems is very much the same as in the closed
case. The answer to both questions is positive in the class of overtwisted
contact structures, and known to be negative at least in some cases when
the extension is required to be tight.
Problem. Let ξ is a germ of tight contact structure on the boundary
∂D 2n+1 of the (2n + 1)-disc D 2n+1 which formally extends to D 2n+1 . Does
it extend as a genuine tight contact structure?
The answer is positive in dimension 3 (see [El92]) but the problem is open
in higher dimensions.
19.3. Folded symplectic and contact structures

The h-principle 14.2.1 for folded solutions of Diff V -invariant differential
relations yields the following result for folded symplectic structures.
19.3.1. Let V1 and V2 be connected 2n-dimensional manifolds with nonempty

boundary and ϕ : Op ∂V1 → Op ∂V2 a diffeomorphism. Let η1 and η2 be
almost symplectic structures on V1 and V2 such that ϕ∗ η2 = η1 . Then there
exist homotopies η1t and η2t , t ∈ [0, 1], of almost symplectic structures on V1
and V2 such that
• η10 = η1 and η20 = η2 ;
• ϕ∗ (η2t |Op ∂V2 ) = η1t |Op ∂V1 for all t ∈ [0, 1];
• η11 and η21 are genuine symplectic structures.
Remark. An analogue of Theorem 19.3.1 in the contact case holds in
a much stronger form: any two germs of contact structures ξ1 and ξ2 on
Op ∂V1 and Op ∂V2 related by ϕ∗ ξ2 = ξ1 which formally extend to V1 and
V2 (the existence of such germs follows from Gromov’s h-principle 18.3.2)
admit, according to Theorem 19.2.2, extensions to V1 and V2 as genuine
contact structures in the prescribed formal homotopy classes.
A folded symplectic structure on a closed 2n-dimensional manifold V (see

[CdS10]) is a closed 2-form ω which near each point is equivalent either to

n
the standard Darboux form ωst = dxj ∧ dyj , or to the form
j=1

n
ωfold = x1 dx1 ∧ dy1 + dxj ∧ dyj .
j=2
The set of points where ω is equivalent to the degenerate form ωfold and
x1 = 0 is called the folding hypersurface. A folded symplectic manifold V
and the folding hypersurface Σ may be nonorientable, in which case the
homology class [Σ] ∈ H2n−1 (V ; Z/2) is nontrivial. If V is orientable, then Σ
is coorientable and divides V into two manifolds V+ and V− with common
boundary Σ (defined by the sign of ω n with respect to an orientation of V ).
Exercises.
1. Verify that for an orientable manifold V a folded symplectic structure can
be modified in a neighborhood of Σ to get two compact symplectic manifolds
(V+ , ω+ ) and (V− , ω− ) with boundary Σ such that j ∗ ω+ = ω− , where V =
V+ ∪Σ V− and j is an orientation reversing involution j : N (Σ) → N (Σ) of
a tubular neighborhood of Σ with fixed point set Σ. Conversely, two such
symplectic structures ω± on V± can be modified into a folded symplectic
structure on V with folding hypersurface Σ.
2. Let V be an orientable connected 2n-dimensional manifold. Show that V
admits a folded symplectic structure if and only if V admits a stable almost
complex structure, i.e., a complex structure on the stabilized tangent bundle
T V ⊕ ε2 , where ε2 is the trivial rank 2 real bundle (see [CdS10]).
19.3. Folded symplectic and contact structures 257
3. Let V be an orientable connected closed stably almost complex manifold.

Find a necessary and sufficient condition on a nonempty dividing hypersur-
face Σ ⊂ V to guarantee the existence of a folded symplectic structure on
V with folding hypersurface Σ.

Chapter 20
Embeddings into
Symplectic and
Contact Manifolds
20.1. Isosymplectic embeddings

Let us recall that when considering embeddings and immersions into a sym-
plectic manifold (W, ωW ) we distinguish between immersions and embed-
dings which induce on the source a symplectic structure, and immersions and
embeddings of another symplectic manifold which induce on it the symplec-
tic structure which was a priori given. Mappings of the first kind are called
symplectic, while those of the second kind are called isometric symplectic,
or isosymplectic.
Remark. The term isometric symplectic was used by G. D’Ambra and A.

Loi (see [DL01]) in a different sense. Namely, for symplectic manifolds en-
dowed with compatible almost complex structures, and therefore Riemann-
ian metrics, they considered the problem of finding immersions which are
simultaneously isosymplectic and isometric with respect to the given Rie-
mannian metrics. They proved in [DL01] an h-principle type result which
is a mixture of Gromov’s Theorem 20.1.1 and Nash–Kuiper’s C 1 -isometric
immersion theorem; see Theorem 29.2.1.
The symplectic condition is open (while the isosymplectic is not!). Hence,

for open manifolds Theorem 8.3.3 implies the parametric h-principle for
symplectic immersions, while Theorem 4.5.1 yields a theorem about directed
symplectic embeddings, i.e., embeddings of a 2m-dimensional open manifold
259
260 20. Embeddings into Symplectic and Contact Manifolds
V into a 2 -dimensional symplectic manifold (W, ωW ), > m, such that

the induced form f ∗ ωW is symplectic. We turn now to the problem of
isosymplectic embeddings and immersions.
Remark. In this section we assume without further mention that the
closed subsets A ⊂ V are neighborhood deformation retracts. This ensures
the existence of a neighborhood basis consisting of open sets U ⊃ A with
H ∗ (V, U ) ∼
= H ∗ (V, A). Alternatively, we could work with arbitrary closed
subsets and use Alexander–Spanier cohomology.
Let (V, ωV ) and (W, ωW ) be symplectic manifolds. A monomorphism (fiber-

wise injective homomorphism) F : T V → T W which covers a map f =
bs F : V → W is called symplectic if F ∗ ωW is nondegenerate and the equal-
ity f ∗ [ωW ] = [ωV ] holds for the cohomology classes. A monomorphism
F : T V → T W which covers a map f = bs F : V → W is called isosymplec-
tic if F ∗ ωW = ωV and the equality f ∗ [ωW ] = [ωV ] holds for the cohomology
classes.
Let us recall (see Section 4.4) that for an open manifold V a polyhedron
V0 ⊂ V is called a core of V if for an arbitrarily small neighborhood U of
V0 there exists an isotopy ht : V → V fixed on V0 which brings V to U .
Usually, but not always, we additionally require the polyhedron V0 to be of
positive codimension.
20.1.1. (Isosymplectic embeddings, Gromov [Gr86]) Let (V, ωV ) and
(W, ωW ) be symplectic manifolds of dimensions n = 2m and q = 2l, respec-
tively. Suppose that an embedding f0 : V → W satisfies the cohomological
condition f0∗ [ωW ] = [ωV ], and that the differential F0 = df0 is homotopic via
a homotopy of monomorphisms
Ft : T V → T W, bs Ft = f0
to an isosymplectic homomorphism F1 : T V → T W .
• Open case. If n ≤ q − 2 and the manifold V is open, then there
exists an isotopy ft : V → W such that the embedding f1 : V → W
is isosymplectic and the differential df1 is homotopic to F1 through
isosymplectic homomorphisms. Moreover, given a core V0 ⊂ V of
positive codimension, one can choose the isotopy ft to be arbitrarily
C 0 -close to f0 near V0 .
• Relative open case. Suppose n ≤ q − 2. Let A ⊂ V be a closed
subset such that each connected component of V \ A has an exit
to infinity, the embedding f0 : V → W is isosymplectic on Op A,
satisfies the relative cohomological condition [f0∗ ωW − ωV ] = 0 in
H 2 (V, A; R), and the homotopy Ft is fixed on T V |A . Then there
exists an isotopy ft : V → W fixed on Op A such that the embedding
20.1. Isosymplectic embeddings 261
f1 : V → W is isosymplectic and the differential df1 is homotopic

to F1 through isosymplectic homomorphisms. Moreover, given a
polyhedron V0 ⊂ V of positive codimension such that A ∪ V0 serves
as a core of V , one can choose the isotopy ft to be arbitrarily C 0 -
close to f0 near V0 .
• Closed case. If n ≤ q−4, then the above isotopy ft exists even if V
is closed. Moreover, one can choose the isotopy ft to be arbitrarily
C 0 -close to f0 .
Reduction of the closed case to the relative open case. Applying

the open case we can find an isotopy ft : V → W , t ∈ [0, 1], which makes
f1 isosymplectic on V \ IntD, where D ⊂ V is an embedded closed n-ball.
Note that the cohomology class [ωV − f1∗ ωW ] ∈ H 2 (V, V \ D; R) is trivial.
Indeed, if n > 2, the group itself is trivial, and if dim M = 2, we have
(f1∗ ωW −ωV ) = (ωV −f1∗ ωW ) = 0. The differential df1 is homotopic to a
D V \D
symplectic monomorphism F1 : (T V, ωV ) → (T W, ωW ) through a homotopy
of injective homomorphisms fixed on T (V \ D). We will keep the notation
Ft for this homotopy, and to simplify notation, we will identify V with
its image under f1 . Let Ft be a covering homotopy T W |V → T W |V such
that Ft |T V = Ft . Let B ⊃ D be a slightly larger ball in V . Consider a
(trivial) rank 2 symplectic bundle ν ⊂ (F1 (T B))⊥ωW ⊂ T W |B and denote
ν := F1−1 (ν) ⊂ T W |B . Then ν intersects T B trivially and is contained in
T B ⊥ωW over B \D. Hence, there exists an extension f1 : B ×D 2 → W of the
embedding f1 |B=B×0 which is isosymplectic on (B \ D) × D 2 with respect to
the split symplectic structure ωV ⊕ ωD2 and whose differential is homotopic
to an isosymplectic homomorphism. Applying the relative open case of
Theorem 20.1.1 we can deform the embedding f1 to an isosymplectic one,
while keeping it fixed on (B\D)×D 2 . Restricting the constructed embedding
to B yields the required isosymplectic embedding (V, ωV ) → (W, ωW ).
The relative open case of Theorem 20.1.1 will be deduced from the following
proposition.
20.1.2. Let (V, ωV ) and (W, ωW ) be symplectic manifolds of dimensions n =

2m and q = 2l > n, respectively, and let A ⊂ V be a closed subset. Let
f0 : V → W be an embedding which is isosymplectic on Op A and satisfies
the cohomological condition [ωV − f0∗ ωW ] = 0 in H 2 (V, A; R). Let M ⊂ V
be a codimension 0 compact submanifold with boundary. Suppose that the
differential F0 = df0 |M is homotopic via a homotopy of monomorphisms
Ft : T M → T W, bs Ft = f0 |M
fixed on Op A to an isosymplectic homomorphism F1 : T M → T W . Then

there exists an isotopy ft : V → W fixed on Op A such that
• f1 |M : M → W is isosymplectic;
• df1 |M is homotopic to F1 through isosymplectic homomorphisms
fixed on T M |Op A∩M ;
• [ωV − f1∗ ωW ] = 0 in H 2 (V, A ∪ M ; R).
Moreover, given a polyhedron V0 ⊂ M of positive codimension such that
(A ∪ V0 ) ∩ M serves as a core of M , one can choose the isotopy ft arbitrarily
C 0 -close to f0 near V0 .
Proof of Proposition 20.1.2. The proof consists of three steps.
Step 1. We can assume that A ∩ M is a subpolyhedron of M for its trian-

gulation. Let L1 be the 1-skeleton of V0 \ A.
20.1.3. There exists an isotopy ft : V → W fixed on Op A such that f1 |Op L1

is isosymplectic, [ωV − f1∗ ωW ] = 0 in H 2 (V, A ∪ L1 ; R), and df1 |Op (A∪L1 ) is
homotopic to F1 |T M |Op (A∪L1 ) through isosymplectic homomorphisms fixed on
T M |Op A .
Proof. We start with the 0-skeleton L0 . Following the homotopy Ft at

points of L0 and using Darboux’s theorem, Theorem 17.3.2, we make the
embedding isosymplectic on Op L0 . To simplify the notation, let us assume
that this is already the case for the embedding f0 .
By assumption we have ωV = f0∗ ωW + dα for a 1-form α supported away
from Op A. Take any 1-simplex σ ⊂ L1 . The embedding f0 |σ is isotropic,
hence its image has the standard symplectic neighborhood (see Theorem
17.3.2). Therefore, there are Darboux coordinates (x1 , y1 , . . . , xl , yl ) for the
symplectic form ωW |Op f0 (σ) such that the image f0 (σ) coincides with the
interval
I = {0 ≤ x1 ≤ 1, y1 = x2 = · · · = xl = yl = 0}.

Set Aσ = α. Take an integer N of the same sign as Aσ , choose a smooth
σ
function θ : [0, 1] → [0, 2πN ] such that θ = 0 near 0 and θ = 2πN near 1,
and for a sufficiently small ε > 0 define an isotopy
hσt : I → Op I, t ∈ [0, 1],
by the formula
hσt (x1 ) = (x1 , 0, tε sin θ(x1 ), tε(1 − cos θ(x1 )), 0, . . . , 0).
Note that the isotopy hσt is fixed near ∂I and

β σ = πN ε2 t2 ,
hσ
t (I)

q
where β σ = xi dyi is the primitive of the symplectic form ωW |Op I . In
1 '
Aσ
particular, for ε = N π we get

β σ = t2 Aσ ,
hσ
t (I)
and choosing the absolute value of N sufficiently large, we can make the
isotopy hσt be arbitrarily C 0 -close to the inclusion I → Op I. Now we define
an isotopy ft : L1 → W by setting ft |σ = hσt ◦ f0 |σ for each 1-simplex
σ ⊂ L1 . Again using 17.3.2 we can extend ft to an isotopy on Op L1 making
f1 isosymplectic on Op L1 . We extend this isotopy arbitrarily to the whole
V , still denoting it ft . We claim that the resulting embedding f1 satisfies

∗
f1 ωW = ωV
Δ Δ
for all 2-simplices Δ of the triangulation of M . To see this, consider the

map F : [0, 1] × Δ → W defined by F (t, x) := ft (x). Stokes’s theorem yields

0= F ∗ dωW = f1∗ ωW − f0∗ ωW − F ∗ ωW ,
[0,1]×Δ Δ Δ σ [0,1]×σ
where the last sum runs over the 1-simplices σ in the boundary of Δ. For
each such σ the map F : [0, 1] × σ → W is given by F (t, x) = hσt ◦ f0 (x) and
lands in the set Op I on which ωW = dβ σ . So again by Stokes’s theorem we
obtain

∗ ∗ σ ∗ σ σ
F ωW = F β = f1 β = β = Aσ = α,
[0,1]×σ ∂([0,1]×σ) σ σ
hσ
1 (I)
where the other boundary terms vanish because β σ |I = 0, and ft = f0 on

∂σ. Now the claim follows via

∗ ∗ ∗
f1 ωW = f0 ωW + α= (f0 ωW + dα) = ωV .
Δ Δ σ σ Δ Δ
It implies [ωV − f1∗ ωW ] = 0 in H 2 (V, A ∪ L1 ; R).
Again, to simplify the notation we assume that the conditions stated in

20.1.3 for f1 hold already for the embedding f0 .
Step 2.
20.1.4. There exists an isotopy ft : V → W fixed on Op (A ∪ L1 ) and a
family of 1-forms αt , t ∈ [0, 1], α0 = 0, supported in V \ Op (A ∪ L1 ) such
that
• f1 |M is symplectic;
• the form ωt := f1∗ ωW + dαt is symplectic on M for all t ∈ [0, 1];
• ω1 = ωV .
Proof. The homotopy Ft gives us a homotopy Gt = GFt between the tan-

gent lift G0 = Gdf0 of f0 and a map
G1 = GF1 : M → Asymp ⊂ Grn W.
The set Asymp ⊂ Grn W is open. Hence, according to Theorem 4.5.1, there
exists an isotopy ft : V → W , fixed on Op (A ∪ L1 ) and C 0 -small near V0 ,
such that f1 |M : M → W is a symplectic embedding whose tangential lift
Gdf1 |M is homotopic rel. Op (A ∪ L1 ) to G1 . According to Theorem 4.5.2,
one can additionally arrange that the differential df1 |M and the isosymplectic
homomorphism F1 are homotopic via a homotopy Φt for which GΦt (M ) ⊂
Asymp . This yields a homotopy of nondegenerate 2-forms which connects
(f1∗ ωW )|M and F1∗ ωW = ωV |M . This homotopy combined with the condition
[ωV − f0∗ ωW ] = 0 in H 2 (V, A ∪ L1 ; R) allows us to apply Theorem 18.2.2 to
get a homotopy ωt of closed 2-forms on V which are symplectic on M such
that ω0 = f1∗ ωW , ω1 = ωV , and [ωt − ω0 ] = 0 ∈ H 2 (V, A ∪ L1 ; R).
Step 3.
20.1.5. (Realization of an exact homotopy of symplectic forms by
an isotopy of symplectic embeddings) Let M be a compact manifold
with (possibly empty) boundary of dimension n = 2m and f0 : M → (W, ωW )
a symplectic embedding into a symplectic manifold (W, ωW ) of dimension
q = 2l > n. Let A ⊂ M be a compact subset. Set ω0 = f0∗ ωW . Let
ωt = ω0 + dαt , t ∈ [0, 1], be a homotopy of symplectic forms on M , where
α0 = 0 and αt |Op A = 0 for all t ∈ [0, 1]. Then there exists an arbitrarily
C 0 -small symplectic isotopy ft : M → (W, ωW ) fixed on Op A such that
ft∗ ωW = ωt .
Remark. If the manifold M is closed, then Moser’s theorem, Theorem
17.3.2, ensures the existence of an isotopy ht : M → M such that h∗t ωt = ω0 ,
t ∈ [0, 1]. However, this isotopy cannot be chosen C 0 -small. For example,
if most of the volumes of ω0 , ω1 are concentrated in disjoint balls B0 , B1 ,
then the length of the isotopy must be at least the distance between the two
balls. If M has a boundary, then Moser’s theorem is no longer true unless
αt vanishes on the boundary. Note that Proposition 20.2.2 below shows
what can be salvaged from Moser’s theorem in the case of manifolds with
boundary.
A differential 1-form r ds on M , where the functions r, s : M → R are

compactly supported in a coordinate neighborhood of M , is called primitive.
20.1.6. Let M be a compact manifold (possibly with boundary), A ⊂ M a
closed subset, and U ⊃ A an open neighborhood in M . Denote by Ω1 (M, U )
and Ω1 (M, A) the spaces of 1-forms on M with support in M \ U (resp.,
M \ A), endowed with the C 1 -norm. Then there exists an integer N and a
bounded linear operator
SM,U,A : Ω1 (M, U ) → Ω1 (M, A)⊕N , SM,U,A (α) = (β1 , . . . , βN )

such that the forms β1 , . . . , βN are primitive and N 1 βi = α.
Proof. Let us choose a smaller neighborhood U U of A and a finite

covering U1 , . . . , UK of M \ U by open subsets of M \ U diffeomorphic either
to the ball { xj < 1} ⊂ R or to the half-ball { x2j < 1, xn ≥ 0} ⊂ Rn+ .
2 n
Inscribe in Ui a slightly smaller covering U 1 , . . . , U

K of M \ U with U i Ui .
i and K
Let θ1 , . . . , θK be a partition of unity with Supp θi ⊂ U 1 θi ≡ 1 on
M \ U . Let κi : Ui → [0, 1] be cut-off functions supported in Ui and equal

to 1 on U i , and set αj := θj α for j = 1, . . . , K. Then K αj = α, and in
j=1
local coordinates x1 , . . . , xn on Uj we have

n
n
αj = rj, dx = rj, dsj,
=1 =1
j and sj, := κj x supported in Uj . Thus

with functions rj, supported in U
K n
α = j=1 =1 rj, dsj, is the required decomposition into a sum of N = Kn
primitive 1-forms. Continuity of the linear operator SM,U,A is clear from the
construction.
20.1.7. (Symplectic twisting) Let (M, ω0 ) be a symplectic manifold and

let Dε2 denote the disc of radius ε in the standard symplectic plane
(R2 , dx ∧ dy).
Then for any primitive 1-form r ds and any ε > 0 there exists a section
Φ : M → M × Dε2
of the trivial bundle M × Dε2 → M such that Φ(x) = (x, 0) outside Supp r ∪
Supp s and
Φ∗ (ω0 ⊕ dx ∧ dy) = ω0 + dr ∧ ds.
Proof. The image of the map ϕ = (r, s) : M → R2 is contained in a disc

2 of radius R > 0. Take an area-preserving immersion
DR
2
τR,ε : DR → Dε2 .
Then the map
ϕ = τR,ε ◦ ϕ : M → Dε2
pulls back the area form dx ∧ dy on Dε2 to the form dr ∧ ds on M . Hence
the corresponding section
Φ : M → M × Dε2 , Φ(x) = (x, ϕ(x))
satisfies the equation Φ∗ (ω0 ⊕ dx ∧ dy) = ω0 + dr ∧ ds.
The same argument proves the following parametric version of Lemma 20.1.7.
20.1.8. Let ωt , t ∈ P , be a family of symplectic forms on a manifold M ,
parametrized by a compact space of parameters. Then for any family rt dst ,
t ∈ P , of primitive 1-forms and any ε > 0 there exists a family of sections
Φt : M → M × Dε2 , t∈P
of the trivial bundle M ×
→ M such that Φt (x) = (x, 0) outside the set
Dε2
t∈P (Supp rt ∪ Supp st ) and
Φ∗t (ωt ⊕ dx ∧ dy) = ωt + drt ∧ dst .
Remark. The class of smoothness of the dependence on the parameter t
in the above lemma is the the same for given families ωt and rt dst and the
resulting family Φt . In particular, the statement holds also for families ωt
and rt dst continuously depending on the parameter t.
Proof of Lemma 20.1.5. Choose a neighborhood U of A such that

Supp αt ⊂ M \ U for all t ∈ [0, 1]. According to Lemma 20.1.6 there exists
an integer N and a linear operator SM,U,A : Ω1 (M, U ) → Ω1 (M, A)⊕N ,
SM,U,A (α) = (β1 , . . . , βN ) such that the forms β1 , . . . , βN are primitive,
N
1 βi = α, and maxj=1,...,N βj C 1 ≤ αC 1 for a constant C = C(M, U, A).
For any ε > 0 there is a sufficiently large integer K such that the 1-forms
γt,i := α t(i+1) − α ti satisfy γt,i C 1 ≤ ε for all t ∈ [0, 1], i = 0, . . . , K − 1.
K K
Applying the decomposition operator SM,U,A to the forms γt,i we get
SM,U,A (γt,i ) = (βt,i,1 , . . . , βt,i,N ) , t ∈ [0, 1], i = 0, . . . , K − 1,
N
where j=1 βt,i,j = γt,i and βt,i,j is a primitive 1-form with βt,i,j C 1 ≤ Cε.
Recall that by assumption the form ω0 + dα ti is symplectic for all t ∈ [0, 1]
K
and i = 0, . . . , K − 1. Hence, if ε is small enough, then the 2-forms ωt,i, :=

ω0 + dα ti + j=1 dβt,i,j are symplectic for any = 0, . . . , N and t ∈ [0, 1].
K
Let us denote m := m(i, j) = N i + j and write βt,m instead of βt,i,j and
ωt,m instead of ωt,i,j , m = 0, . . . , KN . So we have constructed families of

symplectic forms ωt,m , m = 0, . . . , KN , t ∈ [0, 1] satisfying
• ωt,0 = ω0 and ωt,KN = ωt for all t ∈ [0, 1];
• ωt,m = ωt,m−1 + dβt,m for families of primitive 1-forms βt,m sup-
ported in coordinate charts Um for all t ∈ [0, 1], m = 1, . . . , KN .
We will inductively construct isotopies ft,m : V → W, m = 0, . . . , KN ,
t ∈ [0, 1], such that (ft,m )∗ ωW = ωt,m . Then ft := ft,KN will be the required
isotopy. We begin with ft,0 ≡ f0 . Suppose that ft,m−1 for some m =
1, . . . , KN is already constructed. Thus (ft,m−1 )∗ ω = ωt,m−1 , and ωt,m =
ωt,m−1 + dβt,m for a family of primitive forms βt,m = rt dst supported in a
fixed coordinate chart Um ⊂ M . In view of the symplectic neighborhood
theorem (Theorem 17.3.2), there exists an ε > 0 such that the isotopy
ft,m−1 |Um : Um → W extends to an isosymplectic isotopy
ft,m−1 : (Um × Dε2 × Dεq−n−2 , ωt,m−1 ⊕ η2 ⊕ ηq−n−2 ) → (W, ωW )
onto a small neighborhood of ft,m−1 (Um ) in W . Here Dεp denotes the ball
of radius ε in Rp and the form ηp is the restriction to Dεp of the stan-
dard symplectic form on Rp . According to Lemma 20.1.8, there exists a
family of compactly supported sections Φt : Um → Um × Dε2 such that
Φ∗t (ωt,m−1 ⊕ η2 ) = ωt,m−1 + drt ∧ dst = ωt,m . We set

Φ := Φt × 0 : Um → Um × Dε2 × Dεq−n−2
and define the required isotopy of embeddings M → W by

ft,m−1 (Φ(x)), x ∈ Um ;
ft,m (x) :=
ft,m−1 (x), otherwise.
This finishes the proof of Lemma 20.1.5.
Now we can conclude the proof of Proposition 20.1.2. We successively apply

the three isotopies from Step 1, Step 2, and Step 3. Here we apply Lemma
20.1.5 with A ∪ L1 in place of A and arbitrarily extend the resulting isotopy
to a C 0 -small isotopy V → W fixed on Op (A ∪ L1 ). As a result, we get an
embedding f1 which is isosymplectic on M and whose differential restricted
to M is in the prescribed homotopy class of isosymplectic homomorphisms.
By the Darboux–Moser argument we can furthermore make f1 isosymplectic
on Op M by a C ∞ -small diffeotopy of V which is fixed on Op A ∪ M . It
only remains to verify the triviality of the cohomology class [ωV − f1∗ ωW ] in
H 2 (V, A ∪ M ; R). Note that the image of this class in the group H 2 (V, A ∪
L1 ; R) is trivial. Indeed, this property was guaranteed by the isotopy of
Step 1, while the isotopies of Steps 2 and 3 were fixed on Op (A ∪ L1 ).
Since the group H 1 (A ∪ M, A ∪ L1 ; R) is trivial, the exact cohomological
sequence of the triple (V, A ∪ M, A ∪ L1 ) implies that the homomorphism
H 2 (V, A∪M ; R) → H 2 (V, A∪L1 ; R) is injective. Therefore, [ωV −f1∗ ωW ] = 0

in H 2 (V, A ∪ M ; R), and Proposition 20.1.2 is proved.
Proof of Theorem 20.1.1 in the relative open case. Consider an ex-

haustion V = j≥1 Mj of V by compact codimension 0 manifolds Mj with
boundary such that IntMj+1 ⊃ Mj , each component of Mj \ (Mj−1 ∪ A)
intersects the boundary ∂Mj , and (V0 ∪ A ∪ Mj−1 ) ∩ Mj serves as a skeleton
of Mj , j = 1, . . . (we assume here M0 = ∅). Using Proposition 20.1.2,
we can find an isotopy ft1 which is fixed on Op A, C 0 -small on V0 ∩ M1 ,
and such that f11 is isosymplectic on Op (A ∪ M1 ), the cohomology class
[ωV −(f11 )∗ ωW ] ∈ H 2 (V, A∪M1 ) is trivial, and the differential d(f11 |M1 ) is ho-
motopic to F1 |T M1 through isosymplectic homomorphisms fixed over Op A.
The latter condition allows us to construct a homotopy Ft1 : T V → T W
connecting F01 := df11 with an isosymplectic homomorphism F11 through in-
jective homomorphisms fixed over Op (A ∪ M1 ). Hence, we can again apply
Proposition 20.1.2 to construct an isotopy ft2 which is fixed on Op (A ∪ M1 ),
C 0 -small on V0 ∩M2 , and connects f02 := f11 with an embedding f12 : V → W
which is isosymplectic on Op (A ∪ M2 ), satisfies the cohomology condition
[ωV −(f12 )∗ ωW ] = 0 in H 2 (V, A∪M2 ), and such that the differential d(f21 |M2 )
is homotopic to F11 |T M2 through isosymplectic homomorphisms fixed over
Op A. The required isotopy is constructed by inductively continuing this
process, and parametrizing the succesive isotopies over increasingly small
time intervals. This concludes the proof of Theorem 20.1.1
Remark. Using parametric versions of Theorem 4.5.1 and Lemma 20.1.5,

we can similarly prove a parametric version of Theorem 20.1.1.
20.2. Equidimensional isosymplectic immersions

Lemma 20.1.5 implies, in particular, that for a compact manifold V one can
realize any exact homotopy ωt = ω0 + dαt of symplectic forms on V by an
arbitrarily C 0 -small isotopy of embeddings
V → (V × R2 , ω0 ⊕ dx ∧ dy).
As we already mentioned, if V is closed, then Moser’s theorem, Theo-

rem 17.3.2, allows us to substitute V × R2 by V at the expense of C 0 -
approximation. The following proposition shows that for compact mani-
folds with nonempty boundary one can substitute V × R2 by (V × R, π ∗ ω )
where (V , ω
) is any open symplectic manifold which contains (V, ω0 ) as an
equidimensional symplectic submanifold and π is the projection V ×R → V .
As in the case of Moser’s theorem, this can be done only at the expense of
C 0 -approximation.
20.2. Equidimensional isosymplectic immersions 269
20.2.1. (V. Ginzburg, [Gi97]) Let (V , ω ) be a symplectic manifold with-

out boundary and V ⊂ V an equidimensional compact submanifold with
boundary. Let ωt = ω0 + dαt , t ∈ [0, 1], be a family of symplectic forms on
|V . Then there exists an isotopy of embeddings
V such that ω0 = ω
ft : V → V × R, t ∈ [0, 1],
such that f0 is the inclusion V → V = V × 0 → V × R and ft∗ (ω ⊕ 0) = ωt
⊕ 0 = π∗ω
for all t ∈ [0, 1], where ω is the pullback of ω
under the projection
V × R → V .
Remark. Any diffeomorphism of the form
(x, t) → (x, h(x, t)), x ∈ V , t ∈ R,
⊕0, hence the image of the isotopy ft can be shrunk into
preserves the form ω
an arbitrarily small neighborhood of the 0-section V × 0 ⊂ V × R. However,
the isotopy ft cannot, in general, be chosen C 0 -small (see the Remark after
Lemma 20.1.5).
Proof. By the proof of Lemma 20.1.5 and setting M := KN , we find fam-

ilies of symplectic forms ωt,m , m = 0, . . . , M , t ∈ [0, 1] satisfying
• ωt,0 = ω0 and ωt,M = ωt for all t ∈ [0, 1];
• ωt,m = ωt,m−1 + dβt,m for families of primitive 1-forms βt,m sup-
ported in coordinate charts Um for all t ∈ [0, 1], m = 1, . . . , M .
Starting with ft,0 := f0 , we will inductively construct isotopies ft,m : V →
V × R such that (ft,m )∗ (ω ⊕ 0) = ωt,m for all t ∈ [0, 1], m = 1, . . . , M .
Then ft := ft,M will be the required isotopy. Suppose that ft,m−1 is already
constructed for some m ≥ 1. Thus (ft,m−1 )∗ ( ω ⊕ 0) = ωt,m−1 , and ωt,m =
ωt,m−1 + dβt,m for a family of primitive forms βt,m = rt dst supported in a
fixed coordinate chart Um ⊂ M .
The condition (ft,m−1 )∗ (ω ⊕ 0) = ωt,m−1 implies that ft,m−1 is transverse to
the (one-dimensional) characteristic foliation Fω ⊕0 of the 2-form ω ⊕ 0 by
the fibers of the projection V × R → V . So we can extend the embeddings

ft,m−1 to embeddings ft,m−1 : V → V × R of a neighborhood V ⊂ V of V
that are still transverse to the characteristic foliation Fω ⊕0 on V × R. Set

ωt,m−1 = (ft,m−1 )∗ (
ω ⊕ 0).

The embeddings ft,m−1 can be extended to fiber preserving embeddings
Ft : V × R → V × R

such that Ft |V ×0 = ft,m−1 , so that we automatically have
Ft∗ (
ω ⊕ 0) = ωt,m−1 ⊕ 0,

where ωt,m−1
⊕0 is the pullback of ωt,m−1 under the projection V ×R → V .
The embedding
Φt = Id × (rt , st ) : V → V × R2 → V × R2
satisfies
Φ∗t (ωt,m−1

⊕ dx ∧ dy) = ωt,m−1 + drt ∧ dst = ωt,m .
Pick a smooth family of functions Ht : V × R → R with compact support
such that
Ht (v, rt (v)) = st (v) for v ∈ V.
Then the graph
ΓHt = {(v, u, Ht (v, u)) | v ∈ V , u ∈ R} ⊂ V × R2
contains the image Φt (V ). For each t ∈ [0, 1], denote by ϕH u the flow of
t

the time-dependent Hamiltonian Ht : V × R → R (here the flow time is u,
whereas t is a fixed parameter). Then
Ht
Ψt (v, u) = ϕH
u
t
(v), u, Ht ϕ u (v), u
defines diffeomorphisms Ψt : V × R → ΓHt and a short computation shows
Ψ∗t Ωt = ωt,m−1

⊕ 0,

where Ωt = (ωt,m−1 ⊕ dx ∧ dy)|ΓHt . Moreover, according to the Remark
preceding this proof, we can assume that the images Ψ−1 t (Φt (V )) are con-
tained in an arbitrarily small neighborhood of the 0-section V × 0. Hence
the embeddings
−1
Φt Ψt Ft
ft,m : (V, ωt,m ) −→(Γ Ht , Ωt ) −→(V × R, ωt,m−1 ⊕ 0) −→(V × R, ω
⊕ 0)
∗ (
satisfy ft,m ω ⊕ 0) = ωt,m and are thus the desired isotopy.
Remarks.
1. If the primitive forms rt dst in the preceding proof were supported in
V \ ∂V , then one could apply the construction without extending V to V .
Alternatively, one could just apply Moser’s theorem, Theorem 17.3.2.
2. The proof works also for a closed manifold V and an exact homotopy of
symplectic forms on V . In this case the result is of course not new: it is just
a version of Moser’s theorem, Theorem 17.3.2.
3. The diffeomorphisms Ψt are the key ingredient in the proof. They cannot
be made C 0 -small. Neither the application of symplectic twisting (Lemma
20.1.7) nor the smallness of the supports of the primitive forms can salvage
the C 0 -approximation property for ft . Indeed, suppose that f1 is ε-close
to f0 . Let f 1 : V → V be the immersion which is the composition of the
embedding f1 with the projection V × R → V . Choose a ball B ⊂ V such
20.3. Isocontact embeddings 271
−1
that dist(∂B, ∂V ) > 2ε. Then f 1 (B) ⊂ IntV . Let B1 be a connected
−1
component of f 1 (B). Since B1 ⊂ IntV , the restriction f 1 |B1 : B1 → B is a
covering, and thus a diffeomorphism. Hence,f 1 |B1 defines
a symplectomor-

phism (B1 , ω1 ) → (B, ω0 ), and we conclude B ω0n = B1 ω1n < V ω1n . This

gives a contradiction if we choose the family ωt such that B ω0n > V ω1n .
4. The condition ft∗ (

ω ⊕ 0) = ωt in Theorem 20.2.1 implies that the com-
position π ◦ ft : V → V of ft with the projection π : V × R → V is an
(equidimensional) immersion. In the case when the function Ht is supported
in V × R, any fiber of the characteristic foliation FΩt on ΓHt intersects the
image Φt (V ) only once (or does not intersect Φt (V ) at all) and π ◦ ft is an
embedding. However, in the general case, the characteristics on ΓHt may
intersect the image Φt (V ) many times because V is strictly bigger than V .
This is the reason why π ◦ ft may become an immersion rather than an
embedding.
In view of the last remark, Theorem 20.2.1 has as a corollary the following
version of Moser’s theorem, Theorem 17.3.2, for manifolds with boundary,
which is formulated as an Exercise in Gromov’s book [Gr86, p. 335].
20.2.2. (Realization of an exact homotopy of symplectic forms
by a homotopy of equidimensional immersions) Let ωt , t ∈ [0, 1],
be a family of symplectic forms on a compact manifold V with boundary.
Suppose that all these forms belong to the same cohomology class in H 2 (V ).
Let (V , ω
) be a symplectic manifold without boundary which contains V as its
equidimensional submanifold so that ω |V = ω0 . Then there exists a regular
homotopy ft : V → V such that f0 is the inclusion V → V and
ft∗ ω
= ωt , t ∈ [0, 1].
20.3. Isocontact embeddings

Let us recall that for a contact manifold (W, ξW ) a map f : V → W is
called contact if it induces a contact structure on V . A monomorphism
F : T V → T W is called contact if it is transverse to ξW and the inter-
section F (T V ) ∩ ξW consists of symplectic subspaces of ξW with respect to
the conformal symplectic structure CS(ξW ). If the manifold V itself has
a contact structure, then we also consider isometric contact or isocontact
maps f : (V, ξV ) → (W, ξW ) which induce on V the given structure ξV .
A monomorphism F : T V → T W is called isocontact if ξV = F −1 (ξW )
and F induces a conformally symplectic map ξV → ξW with respect to the
conformal symplectic structures CS(ξV ) and CS(ξW ).
Remark. As in the symplectic case, the term isometric contact was

used by G. D’Ambra (see [DA00]) in a different sense. Namely, for two
contact manifolds endowed with compatible CR-structures, and therefore
Riemannian metrics on the contact distributions, she considered the problem
of finding immersions which are simultaneously isocontact and isometric
with respect to the given Riemannian metrics. She proved in [DA00] an
h-principle type result which is a mixture of Theorem 20.3.1 and the Nash–
Kuiper C 1 -isometric immersion theorem; see Theorem 29.2.1.
As in the symplectic case, the contact condition is open, while the isocontact
one is not. Hence, for an open V , Theorem 18.3.2 implies the parametric
h-principle for contact immersions, while Theorem 4.5.1 (applied to the con-
tact Grassmannian Acont consisting of subspaces transverse to the contact
structure whose intersection with the contact hyperplanes is symplectic)
yields a theorem about contact embeddings. We turn now to the problem
of isocontact embeddings. Here is an analogue of Gromov’s isosymplectic
embedding theorem 20.1.1 for the contact case.
20.3.1. (Isocontact embeddings) Let (V, ξV ) and (W, ξW ) be contact

manifolds of dimension n = 2m + 1 and q = 2l + 1, respectively. Suppose
that the differential F0 = df0 of an embedding f0 : (V, ξV ) → (W, ξW ) is
homotopic (via a homotopy of monomorphisms Ft : T V → T W , bs Ft = f0 )
to an isocontact monomorphism F1 : T V → T W .
• Open case. If n ≤ q − 2 and the manifold V is open, then there
exists an isotopy of embeddings ft : V → W such that f1 : V → W
is isocontact and the differential df1 is homotopic to F1 through
isocontact monomorphisms. Moreover, given a core V0 ⊂ V , one
can choose the isotopy ft to be arbitrarily C 0 -close to f0 near V0 .
• Relative open case. Suppose n ≤ q − 2. Let A ⊂ V be a closed
subset such that each connected component of V \ A has an exit to
infinity (see Section 4.4) the embedding f0 : V → W is isocontact
on Op A, and the homotopy Ft is fixed on T V |A . Then there exists
an isotopy of embeddings ft : V → W fixed on Op A such that f1 :
V → W is isocontact and the differential df1 is homotopic rel. Op A
to F1 through isocontact monomorphisms. Moreover, given a poly-
hedron V0 ⊂ V of positive codimension such that each component of
V \ (A ∪ V0 ) has an exit to infinity, one can choose the isotopy ft
to be arbitrarily C 0 -close to f0 near V0 .
• Closed case. If n ≤ q−4, then the above isotopy ft exists even if V
is closed. Moreover, one can choose the isotopy ft to be arbitrarily
C 0 -close to f0 .
Proof. We will prove only the relative open case. The reduction of the
closed case to the open one via the microextension trick is similar to the
symplectic case, and even simpler because one does not need to worry about
the cohomological condition. The proof will follow the same strategy as in
the symplectic case, but with a few significant modifications. Since one does
not need the preliminary step to fix the cohomological condition, the proof
is based on the following two lemmas.
20.3.2. Under the assumptions of the relative open case of Theorem 20.3.1
there exists an isotopy ft : V → W connecting f0 with a contact embedding
f1 which is fixed on Op A, C 0 -close to f0 near V0 , and such that the contact
structures f1∗ ξW and ξV are homotopic via a homotopy of contact structures
fixed on Op A.
Proof. The homotopy Ft gives us a homotopy Gt = GFt between the tan-

gent lift G0 = Gdf0 of f0 and a map G1 : V → Acont ⊂ Grn W . We
recall that Acont consists of subspaces transverse to the contact structure
whose intersection with the contact hyperplanes is symplectic. The set
Acont ⊂ Grn W is open. Hence, according to Theorem 4.5.1, there exists
an isotopy ft : V → W fixed on Op A which is C 0 -close to f0 on Op V0 , and
such that f1 : V → W is a contact embedding whose tangential lift Gdf1
is homotopic rel. Op A to G1 . According to Theorem 4.5.2, one can addi-
tionally arrange that the differential df1 and the isocontact homomorphism
F1 are homotopic via a homotopy Ψt rel. Op A for which GΨt (V ) ⊂ Acont .
This fact and Theorem 18.3.2 imply the existence of a homotopy of contact
structures fixed on Op A which connects ξV and f1∗ ξW .
20.3.3. Let (V, ξ = {α = 0}) be a contact manifold without boundary, M ⊂
V a compact codimension 0 submanifold-with-boundary, and A ⊂ M a closed
subset. Let αt be a family of contact forms on M such that α0 = α|M and
αt = α0 on Op A for all t ∈ [0, 1]. Then for each ε > 0 there exists a family
of isocontact embeddings

ft : (M, {αt = 0} → V × Dε2 , {α ⊕ x dy = 0}
such that f0 is the inclusion M → V = V × 0 → V × Dε2 , and ft agrees on
Op A with the inclusion Op A → V = V × 0 → V × Dε2 .
Remark. Note that if V is endowed with a Riemannian metric such that
diam(M ) < ε, then the embeddings ft are 2ε-close to the inclusion.
Proof of 20.3.3. Applying the Partition Lemma 20.1.6 and arguing as in

the proof of Lemma 20.1.5, we can reduce the assertion to the case that
αt = α0 + rt dst for some smooth functions st , rt : M → R with compact
support in a coordinate neighborhood U ⊂ M which vanish on Op A. Pick a
cutoff function θ : M → [0, 1] supported in U with θ ≡ 1 on t∈[0,1] supp rt .
After replacing st by st + Cθ for a sufficiently large constant C, which does

not change the 1-form rt dst , we may therefore assume
{st = 0} ⊂ {rt = 0} for all t ∈ [0, 1].
The sections
Ft : M → M × R2 → V × R2 , Ft (v) = (v, rt (v), st (v))
are isocontact with respect to the contact structures defined by the contact
forms αt = α0 + rt dst on M and α := α ⊕ x dy on V × R2 . Pick a family of
compactly supported functions Ht : V × R → R which vanish on Op A × R
and satisfy
Ht v, st (v) = rt (v) for v ∈ M, t ∈ [0, 1].
In view of the condition {st = 0} ⊂ {rt = 0}, we can further arrange
Ht (v, 0) = 0 for all v ∈ V, t ∈ [0, 1].
For t ∈ [0, 1] consider the graph of Ht ,
Γt = {(v, x, y) | x = Ht (v, y)} ⊂ (V × R2 , α).
Its contact characteristic foliation Lt (see Section 17.8) is transverse to the
contact slices {y = const}, as well as to the contact submanifold Ft (M ) ⊂
Γt . Moreover, Lt agrees with the lines {(v, x) = const} outside a compact
set. Explicitly, Lt is generated by the contact vector field of Ht (v, x, y) =
Ht (v, y) − x which along Γt is given by
∂Ht ∂ ∂
Xt (v, x, y) = Xt (v, y) + dHt (Rα )x − − ,
∂y ∂x ∂y
where Xt is the contact vector field of Ht . The rescaled vector field y X has
last component −y ∂y ∂
, so its time s flow Ψst : Γt → Γt preserves {α = 0} and
contracts the y-component toward 0. In view of the condition Ht (v, 0) =
0 the flow Ψst then also contracts the x-component toward 0. Thus the
embeddings fts := Ψst ◦ Ft : M → M × R2 → V × R2 satisfy ft∗ {α = 0} =
{αt = 0} for all s ∈ R, t ∈ [0, 1], and for a sufficiently large s = s0 we have
fts0 (M ) ⊂ V × Dε2 . Hence, the isotopy ft := fts0 , t ∈ [0, 1], has the required
properties.
Proof of Theorem 20.3.1 in the relative open case. If V has boundary,

consider a slightly larger manifold V ⊃ V without boundary, and extend
the form α and the embedding f0 to V . We can assume that after the
extension each connected component of V \ A still has an exit to infinity.
After renaming V back to V , we may thus assume that V has no boundary.
We begin the proof by applying Lemma 20.3.2 to find an isotopy ft0 : V →
W fixed on Op A and C 0 -close to f0 on V0 connecting f00 = f0 with a
contact (but not necessarily isocontact) embedding f10 : V → (W, ξW ), and
a homotopy of contact structures ξt0 fixed on Op A connecting ξ00 = (f10 )∗ ξW

with ξ10 = ξV .
Our next task is to make the embedding isocontact on Op V0 while keeping it
C 0 -close to f0 on V0 . We cover V0 by finitely or countably many closed balls
M1 , M2 , . . . in V of of diameter < ε (or with diameters going to 0 at infinity
if an approximation in the fine Whitney topology is required). Let Vj ⊂ V
be open balls containing Mj , still of diameter < ε. Let ξV = {αV = 0} and
ξW = {αW = 0}. By the contact neighborhood theorem, Theorem 17.5.2,
the contact embedding f10 : V1 → (W, ξW ) extends for small ε > 0 to an
isocontact embedding
4
l−m 5
f10 : V1 × (Dε2 )l−m , (f10 )∗ αW ⊕ (−yj dxj ) = 0 → (W, ξW ).
j=1
We apply Lemma 20.3.3 with (V, M, A) replaced by (V1 , M1 , A ∩ M1 ) and

αt defining 1-forms for ξt0 to find isocontact embeddings

gt1 : (M1 , ξt ) → V1 × Dε2 × {0}l−m−1 , (f10 )∗ ξW
such that g01 and gt1 |Op (A∩M1 ) are the obvious inclusions. Then gt1 := f10 ◦
gt1 :
M1 → W is an isotopy of embeddings which is fixed on Op (A ∩ M1 ) and
connects g01 = f10 with an embedding g11 such that (g11 )∗ ξW = ξV |M1 . Since
gt1 lands in V1 ×Dε2 ×{0}l−m−1 and V1 has diameter < ε, the isotopy gt1 is C 0 -
small. We extend the isotopy to V (still denoting it gt1 ), keeping it C 0 -small
on V0 , and extend dg11 |T M1 to an isocontact monomorphism G11 : T V → T W
covering the extended g11 .
Next, we again apply Lemma 20.3.2 to construct an isotopy ft1 fixed on
Op (A∪M1 ) and C 0 -close to g11 on V0 connecting f01 = g11 with a contact (but
not necessarily isocontact) embedding f11 : V → (W, ξW ), and a homotopy ξt1
fixed on Op (A∪M1 ) connecting ξ01 = (f11 )∗ ξW with ξ11 = ξV . Continuing this
process, we inductively make the embedding isocontact on Mj , j = 2, . . . ,

while keeping it fixed on Op A ∪ j−1 1 Mi . After reparametrizing them over
shorter and shorter time intervals, the sequence of constructed isotopies
converges to an isotopy of embeddings V → W , agreeing with f0 on Op A
and C 0 -close to f0 on V0 , connecting f0 to an embedding f0 : V → W
which is isocontact on Op V0 and contact elsewhere. Moreover, there exists a
homotopy of contact structures ξt which is fixed on Op (A∪V0 ) and connects
f0∗ ξW with ξV . Denote A := A ∪ V0 .
To complete the proof we consider an exhaustion M ,1 ⊂ M ,2 ⊂ · · · ⊂ V ,
, ,j with boundary such that each compo-
Mj = V by compact manifolds M
nent of V \(A∪ M ,j ), j = 1, . . . , has an exit to infinity, and run the above pro-
cess again without worrying any more about the C 0 -approximation property.
Namely, we apply Lemma 20.3.3 with (V, M, A) replaced by (V, M ,1 , A ∩ M

,1 )
1 ,
to find an isotopy gt : M1 → W which is fixed on Op A and connects g0 = f0
1
to an embedding g11 such that (g11 )∗ ξW = ξV |M

1 . We continue by extend-
ing the isotopy in any way to V and extending dg11 |T M1 to an isocontact
monomorphism G11 : T V → T W covering g11 . Next, we again apply Lemma
20.3.2 to construct an isotopy ft1 which is fixed on Op (A1 ∪ M ,1 ) and con-
1 1
nects f0 = g1 with a contact (but not necessarily isocontact) embedding
f11 : V → (W, ξW ). Continuing this process, we consequently make the em-
,j , j = 1, . . . , while keeping it fixed on Op (A∪M
bedding isocontact on M ,j−1 ).
The sequence of constructed isotopies converges to the required isotopy con-
necting f0 to an embedding which is isocontact on the whole manifold V .
Remark. The parametric versions of Theorems 20.1.1 and 20.3.1 are

also valid with the same proof. One can also deduce the corresponding
results about isosymplectic and isocontact immersions. Later in Chapter
24 we will prove the results about isosymplectic and isocontact immersions
by a different method which will allow us to get rid of any dimensional
restrictions.
Exercise. Formulate and prove contact analogues of Theorems 20.2.1

and 20.2.2.
20.4. Subcritical isotropic embeddings
Let us point out a useful application of Theorems 20.1.1 and 20.3.1.
Let W be either a symplectic or a contact manifold of dimension # q, and $

V a smooth manifold of subcritical dimension, i.e., dim V < dim 2W −1 .
Let Mono emb be the space of monomorphisms T V → T W which cover
embeddings V → W , and Mono emb isot its subspace which consists of isotropic
monomorphisms F : T V → T W , i.e., of monomorphisms which send tangent
spaces to V to isotropic subspaces of T W in the symplectic case, and to
isotropic subspaces of the contact bundle ξ ⊂ T W in the contact case. Let
Mono emb
isot be the space of homotopies
emb
Mono isot = {Ft , t ∈ [0, 1] | Ft ∈ Mono
emb
, F0 = df0 , F1 ∈ Mono emb
isot }.
The space Emb isot of isotropic embeddings V → W can be viewed as a sub-

emb . Indeed, we can associate with f ∈ Emb isot the homotopy
space of Mono isot
emb .
Ft ≡ df, t ∈ [0, 1], from Mono isot
20.4. Subcritical isotropic embeddings 277
20.4.1. (Homotopy principle for subcritical isotropic embeddings)

The inclusion
emb
Emb isot → Mono isot
The above h-principle also holds in the relative and C 0 -dense forms.
Proof. In the symplectic case, any isotropic monomorphism extends in a

homotopically canonical way to an isosymplectic monomorphism
T (T ∗ V ) → T W,
and in the contact case it extends to an isocontact monomorphism
T (J 1 (V )) → T W,
where T ∗ V and J 1 (V ) are endowed with the canonical symplectic and
contact structures. The dimensional condition ensures in both the sym-
plectic and the contact cases that the dimensions of T ∗ V and J 1 (V ) are
≤ dim W − 2, hence Theorems 20.1.1 and 20.3.1 apply. Then we get the
required isotropic embeddings as the restrictions to the 0-section of the con-
structed isosymplectic and isocontact embeddings.
Warning. The analogues of Theorem 20.4.1 for Lagrangian and Legen-
drian embeddings are false.
Chapter 21
Microflexibility and
Holonomic
R-approximation
For further applications to symplectic geometry we will need a generalization

of the Holonomic Approximation Theorem which we discuss in Section 21.4.
21.1. Local integrability

A differential relation R ⊂ X (r) is called locally integrable if for any v ∈ V
and any section F : v → R there exists a local holonomic extension F of
F , i.e., a holonomic section F : Op v → R such that F (v) = F (v). In
other words, the Cauchy problem with the initial data (v, F (v)) has a local
solution.
A differential relation R ⊂ X (r) is called parametrically locally integrable if

given a map h : I k → V and a family of sections
Fp : h(p) → R, p ∈ I k ,
there exists a family of local holonomic extension
Fp : Op h(p) → R, Fp (h(p)) = Fp (h(p)), p ∈ I k .
Exercise. Prove that the parametric local integrability of R implies that

for any v ∈ V and any local section F : Op v → R there exists a homotopy
F τ : Op v → R, τ ∈ [0, 1],
279
280 21. Microflexibility and Holonomic R-approximation
such that F τ (v) = F (v) for all τ ∈ [0, 1], F 0 coincides with F and F 1
is holonomic. In other words, the Cauchy problem with the initial data
(v, F (v)) has a local solution in any homotopy class of local sections Op v →
R.
In fact, we need the following stronger relative version of parametric local

integrability: a differential relation R ⊂ X (r) is called parametrically locally
integrable if given a map h : I k → V , a family of sections
Fp : h(p) → R, p ∈ I k ,
and a family of local holonomic extensions near ∂I k
Fp : Op h(p) → R, Fp (h(p)) = Fp (h(p)), p ∈ Op (∂I k ),
there exists a family of local holonomic extensions
Fp : Op h(p) → R, Fp (h(p)) = Fp (h(p)), for all p ∈ I k ,
such that for p ∈ Op (∂I k ) these new extensions coincide with the original
extensions over Op h(p).
In what follows the term locally integrable always means this last stronger
version of local integrability.
Exercise. Prove that the local integrability of R implies the local h-
principle for R near any point v ∈ V .
Examples.
1. Any open differential relation is locally integrable; see (the parametric
version of) Lemma 3.3.1.
2. The differential relation Riso which defines isometric immersions of Rie-

mannian manifolds (V, gV ) → (W, gW ) is not locally integrable in general.
3. Symplectic and contact stability (Theorems 17.3.2 and 17.5.2) imply that
the following closed differential relations of symplectic-geometric origin are
locally integrable:
• the differential relation Risosymp which defines isosymplectic immersions
(V, ωV ) → (W, ωW );
• the differential relations RLag and Rsub-isotr which define Lagrangian
and subcritical isotropic immersions V → (W, ωW );
• the differential relation Risocont which defines isocontact immersions
(V, ξV ) → (W, ξW );
• the differential relations RLeg and Rsub-isotr which define Legendrian
and subcritical isotropic immersions V → (W, ξW ).

21.3. Microflexibility 281
21.2. Homotopy extension property for formal solutions

Let R ⊂ X (r) be a differential relation and A a compact subset of V .
21.2.1. (Homotopy extension property for formal solutions) Let
F : V → R be a section and FAτ : Op A → R, τ ∈ [0, 1], be a homotopy of
the section FA0 = F |Op A . Then FAτ extends to a homotopy F τ : V → R of
F.
Proof. The homotopy FAτ is defined over a neighborhood U ⊂ V . Take a

continuous function δ : V → R+ with a compact support in U such that
ρ ≡ 1 in a neighborhood U ⊂ U of A. Then set F τ (v) = F (v) for v ∈ V \ U
τ δ(v)
and F τ (v) = FA for v ∈ U .
Similarly, we can prove the following

21.2.2. (Homotopy extension property for formal solutions, rel-
ative case) Let (A, B), where B ⊂ A be a pair of compact subsets of V ,
FA : Op A → R a section and FBτ : Op B → R, τ ∈ [0, 1], a homotopy of
the section FB0 = FA |Op B . Then FBτ extends to a homotopy FAτ : Op A → R
of FA .
21.3. Microflexibility
The notion of a microflexible differential relation, which we introduce in this
section, roughly corresponds to Gromov’s notion of a microflexible sheaf; see
[Gr69] and [Gr86].
Let us recall that the term holonomic homotopy (or holonomic deformation)
is used in a sense of homotopy consisting of holonomic sections.
Let K m = [−1, 1]m . For a fixed n and any k < n, we denote by θk the pair
(K n , K k ). Let, as usual, V be an n-dimensional manifold. A pair (A, B) ⊂ V
is called θ-pair or, more precisely, θk -pair, if (A, B) is diffeomorphic to the
standard pair θk .1
A differential relation R ⊂ X (r) is called k-microflexible if for any sufficiently

small open ball U ⊂ V and any
• θk -pair (A, B) ⊂ U ,
• holonomic section F 0 : Op A → R, and
• holonomic homotopy F τ : Op B → R, τ ∈ [0, 1], of the section F 0
over Op B which is constant over Op (∂B),
1 We use the term θ-pair, because the shape of the set ∂K n ∪ K k resembles the letter θ.
there exists a number σ > 0 and a holonomic homotopy, constant over

Op (∂A),
F τ : Op A → R, τ ∈ [0, σ],
which extends the homotopy
F τ : Op B → R, τ ∈ [0, σ].
In other words, an initial stage of any holonomic deformation of the section
F 0 over Op B which is constant over Op (∂B) can be extended to a holonomic
deformation of F 0 over Op A which is constant over Op (∂A). If such an
extension exists for all τ ∈ [0, 1] then R is called k-flexible.
More generally, a differential relation R ⊂ X (r) is called parametrically k-

microflexible if for any sufficiently small open ball U ⊂ V and any families
parametrized by p ∈ I m of
• θk -pairs (Ap , Bp ) ⊂ U ,
• holonomic sections Fp0 : Op Ap → R, and
• holonomic homotopies Fpτ : Op Bp → R, τ ∈ [0, 1], of the sections
Fp0 over Op Bp which are constant over Op (∂Bp ) for all p ∈ I m and
constant over Op Bp for p ∈ Op (∂I m ),
there exists a number σ > 0 and a family of holonomic homotopies
Fpτ : Op Ap → R, τ ∈ [0, σ],
which extend the family of homotopies
Fpτ : Op Bp → R, τ ∈ [0, σ],
and are constant over Op (∂Ap ) for all p ∈ I m and constant over Op Ap for
p ∈ Op (∂I m ).
In what follows the term k-(micro)flexibility always means the parametric
k-(micro)flexibility.
A differential relation R is called (micro)flexible, if it is k-(micro)flexible for

all k = 0, . . . , n − 1 where n = dim V .
Examples.
1. Any open differential relation is microflexible.
2. The differential relation Rhol ⊂ (X (r) )(1) , which defines holonomic sec-
tions of X (r) , is flexible.
3. The differential relation Rclo ⊂ (Λp V )(1) , which defines closed p-forms
on V , is k-flexible if k = p. For k = p the relation Rclo is neither k-flexible
nor k-microflexible.
21.5. Local h-principle for microflexible Diff V -invariant relations 283
4. Symplectic and contact stability (Theorems 17.3.2 and 17.5.2) implies

the following.
• The differential relation Risosymp , which defines isosymplectic immer-
sions (V, ωV ) → (W, ωW ), is k-microflexible if k = 1.
• The differential relations RLag and Rsub-isotr , which define Lagrangian
and subcritical isotropic immersions, are k-microflexible if k = 1.
• The differential relation Risocont , which defines isocontact immersions
(V, ξV ) → (W, ξW ), is microflexible.
• The differential relations RLeg and Rsub-isotr , which define Legendrian
and subcritical isotropic immersions V → (W, ξW ), are microflexible.
21.4. Theorem on holonomic R-approximation

21.4.1. (Holonomic R-approximation Theorem) Let R ⊂ X (r) be a
locally integrable microflexible differential relation. Let A ⊂ V be a poly-
hedron of positive codimension and F : Op A → R a section. Then for
arbitrarily small δ, ε > 0 there exists a δ-small (in the C 0 -sense) diffeotopy
hτ : V → V , τ ∈ [0, 1], and a holonomic section F : Op h1 (A) → R such
that
dist(F (v), F |Op h1 (A) (v)) < ε
for all v ∈ Op h1 (A).
The proof of this theorem literally repeats the proof of the original Holo-
nomic Approximation Theorem 3.1.1. When working over a cube, the local
integrability provides the first step of the induction. Then the microflexibil-
ity implies the version of Interpolation Property 3.6.1 which is required for
the proof of Inductive Lemma 3.4.1. Finally we proceed inductively over the
skeleton of the polyhedron A. Note that the homotopy extension property
21.2.2 allows us to extend the holonomic solutions obtained at each step of
the induction as formal solutions to Op A.
The relative and the parametric versions of Theorem 3.1.1 also hold. In the
relative version the section F is assumed to be already holonomic over Op B,
where B is a subpolyhedron of A. The diffeotopy hτ and the section F can
be constructed in this case so that hτ is fixed on Op B and F coincides with
F on Op B.
21.5. Local h-principle for microflexible Diff V -invariant

relations
Using Theorem 21.4.1 instead of Lemma 3.1.1, we obtain, as in Section 8.3,
the following generalization of Theorem 8.3.1.
21.5.1. (Local h-principle) All forms of the local h-principle hold for lo-
cally integrable and microflexible Diff V -invariant differential relations near
any polyhedron A ⊂ V of positive codimension.
The proof repeats almost literally the proof of Theorem 8.3.1. The only
problem is the construction of a homotopy between formal and genuine
solutions which lies in R. The linear homotopy does not necessarily lie in R.
In general, the construction of a homotopy in R requires some additional
work. However, if R is a local neighborhood retract, then one can just
compress the linear homotopy into R by the retraction. This case is sufficient
for all our further applications. We leave the general case to the reader as
an exercise; the key to the construction of homotopy is the first exercise in
Section 21.1.
According to Theorem 8.3.2, Theorem 21.5.1 implies
21.5.2. (Gromov [Gr69]) Let V be an open manifold and X → V a
natural fiber bundle. Then any locally integrable and microflexible DiffV -
invariant differential relation R ⊂ X (r) satisfies the parametric h-principle.
When A is a manifold and V = A × R, then to prove the local h-principle

near A = A × 0 ⊂ V one does not need R to be invariant with respect to the
whole group Diff V , but only with respect to diffeomorphisms of the form
(x, t) → (x, h(x, t)), x ∈ V, t ∈ R.
Indeed, in this case we need to use only this kind of diffeomorphism in the
Holonomic R-approximation Theorem 21.4.1 (cf. the similar Theorem 9.3.1
for open differential relations over A × R). The following proposition is a
version of the Main Flexibility Theorem from [Gr86, p. 78].
21.5.3. (Gromov [Gr86]) Let X → V = A × R be a natural fibration and
R ⊂ X (r) a locally integrable and microflexible differential relation which is
invariant with respect to diffeomorphisms of the form
(x, t) → (x, h(x, t)), x ∈ V, t ∈ R.
Then R satisfies all forms of the local h-principle near A = A × 0 and
satisfies the parametric h-principle globally on V .
In Chapter 23 we consider more examples when the h-principle holds for

differential relations invariant only with respect to a certain subgroup of the
group Diff V .
Chapter 22
First Applications of
Microflexibility
22.1. Subcritical isotropic immersions

A. Immersions into contact manifolds. As an immediate application
of Theorems 21.5.1 and 21.5.2 we get all the forms of the local h-principle
for Legendrian/isotropic immersions Op A → (W, ξ) where, as usual, A ⊂ V
is a polyhedron of positive codimension and the parametric h-principle for
Legendrian/isotropic immersions V → (W, ξ) of open manifolds V . Any
formal/genuine subcritical isotropic immersion V → (W, ξ) can be extended,
at least locally, to a formal/genuine isotropic immersion V × R → (W, ξ),
and hence the microextension trick yields as usual all the forms of the h-
principle for subcritical isotropic immersions of closed manifolds. This is
not, however, too exciting. First, we already proved in Chapter 20 the h-
principle for subcritical isotropic embeddings, which is much stronger than
for immersions. Second, we will see in Section 24.1 that the h-principle holds
even for Legendrian immersions (see Theorem 24.1.3) of closed manifolds.
B. Immersions into symplectic manifolds. The local h-principle for

Lagrangian or isotropic immersions Op A → (W, ω) and the parametric h-
principle for Lagrangian/isotropic immersions V → (W, ω) of open manifolds
V become true if we incorporate the global algebraic condition [f ∗ ω] = 0
in the definition of a formal isotropic immersion. The proof follows from
Theorems 21.5.1 and 21.5.2, though not immediately because we need to
overcome the lack of microflexibility for k = 1 for isotropic immersions into
symplectic manifolds. The microextension trick yields all the forms of the
h-principle for subcritical isotropic immersions of closed manifolds. As in
285
286 22. First Applications of Microflexibility
the case of isotropic immersions into contact manifolds, all these results are
not too exciting. The only interesting thing is how one can fight the lack of
microflexibility. We will explain it later in Section 24.3 where we prove the
h-principle for Lagrangian immersions of closed manifolds.
22.2. Maps transverse to a contact structure

Theorem 9.3.3, which we proved earlier, asserts that the parametric h-
principle holds for mappings transverse to a tangent distribution, provided
that the sum of the dimension of the manifold and the dimension of the
distribution is less than the dimension of the target manifold. In his book
[Gr86] Gromov formulated a series of exercises which culminate in a theorem
which claims that if the distribution in question is completely nonintegrable,
then the h-principle holds even when the sum of dimensions of the mani-
fold and the distribution is greater than or equal to the dimension of the
target manifold. Here complete nonintegrability means that the successive
Lie brackets of vector fields tangent to the distribution span the tangent
space T W . Following Gromov’s scheme we consider below the special case
of maps transverse to a contact structure.
22.2.1. (Gromov [Gr86]) Let (W, ξ) be a contact manifold. Then the
df
maps f : V → W transverse to ξ (i.e., the maps for which T V → T W →
T W/ξ is a fiberwise surjective homomorphism) satisfy all forms of the h-
principle.
Proof. Denote by Rtrans the differential relation in J 1 (V, W ) which defines

the maps V → (W, ξ) transverse to ξ. We will prove the h-principle for
Rtrans using a certain form of the microextension trick.
Let us recall that the differential relation Rtang ⊂ J 1 (R, W ) which defines
the isotropic immersions R → (W, ξ), i.e., maps which are tangent to ξ, is
locally integrable and microflexible (see Sections 21.1 and 21.3). Consider
a mixed differential relation Rtrans-tang ⊂ J 1 (V × R, W ) which corresponds
to maps V × R → (W, ξ) transverse to ξ but tangent to ξ along each fiber
v × R, v ∈ V . The openness of the relation Rtrans and the local integra-
bility and microflexibility of the relation Rtang imply the local integrability
and microflexibility of the relation Rtrans-tang . The relation Rtrans-tang is
invariant with respect to diffeomorphisms of the form
(x, t) → (x, h(x, t)), x ∈ V, t ∈ R,
and hence according to Proposition 21.5.3 all forms of the local h-principle
hold for Rtrans-tang . The local h-principle for Rtrans-tang implies the h-
principle for Rtrans . Indeed, the restriction to V × 0 of any solution of
Rtrans-tang on Op (V × 0) ⊂ V × R is a solution of Rtrans on V . On the other
22.2. Maps transverse to a contact structure 287
hand, any formal/genuine solution of Rtrans over any simplex Δ in V can

be extended to a formal/genuine local solution of Rtrans-tang on Δ × R.
Similarly we can prove

22.2.2. (Immersions transverse to a contact structure) Let (W, ξ) be
a contact manifold and dim V < dim W . Then the immersions f : V → W
transverse to ξ satisfy all forms of the h-principle.
Proof. Set
Rimm-trans = Rimm ∩ Rtrans .
Consider a mixed differential relation
Rimm-trans-tang ⊂ J 1 (V × R, W )
which corresponds to immersions V ×R → (W, ξ) transverse to ξ but tangent
to ξ along each fiber v × R, v ∈ V . The openness of the relation Rimm-trans
and the local integrability and microflexibility of the relation Rtang imply
the local integrability and microflexibility of the relation Rimm-trans-tang . The
relation Rimm-trans-tang is invariant with respect to diffeomorphisms of the
form
(x, t) → (x, h(x, t)), x ∈ V, t ∈ R,
and hence according to Proposition 21.5.3 all forms of the local h-principle
hold for Rimm-trans-tang . The local h-principle for Rimm-trans-tang implies the
h-principle for Rimm-trans . Indeed, the restriction to V × 0 of any solution of
Rimm-trans-tang on Op (V × 0) ⊂ V × R is a solution of Rimm-trans on V . On
the other hand, any formal/genuine solution of Rimm-trans over any simplex
Δ in V can be extended to a formal/genuine local solution of Rimm-trans-tang
on Δ × R.
Remark. Theorems 22.2.1 and 22.2.2 remain true (with the same proofs)
for any distribution ξ on W for which the differential relation Rtang ⊂
J 1 (R, W ) which defines isotropic immersions R → (W, ξ), i.e., the maps
which are tangent to ξ, is microflexible. Thus to solve Gromov’s exercise
for a general completely nonintegrable distribution one needs to establish
a suitable microflexibility property for the relation Rtang . In fact, the re-
lation Rtang is not always microflexible. Indeed, R. Bryant and L. Hsu
(see [BH93]) have shown that most nonintegrable distributions (e.g., En-
gel structures, which are maximally nonintegrable 2-plane fields on four-
dimensional manifolds) possess rigid integral curves. In other words, the
corresponding space of integral curves contains isolated points. This, of
course, contradicts the microflexibility of the corresponding relation Rtang .
Despite this difficulty, Gromov’s exercise has been solved in the general case
by A. Bhowmick [Bh22].
288 22. First Applications of Microflexibility
Theorem 22.2.2 together with the h-principle for contact structures on open
manifolds (see Section 18.3.2) implies the following theorem of McDuff about
maximally nonintegrable tangent hyperplane distributions on even-dimen-
sional manifolds. This theorem is a contact analogue of Theorem 18.4.1.
Let ξ be a tangent hyperplane field on a 2n-dimensional manifold W . We

say that ξ is maximally nonintegrable if the form dα|ξ has the maximal rank
2n − 2, where α is a defining 1-form for ξ (which is valued in a nontrivial
line bundle T W/ξ if ξ is not coorientable). For simplicity we formulate the
h-principle below only for the case of coorientable distributions, leaving the
general case as an exercise to the reader.
22.2.3. (McDuff [MD87a]) Let ξ = Ker α be a hyperplane field on a
2n-dimensional manifold V and ω a 2-form whose restriction to ξ is of
maximal rank 2n − 2. Then V admits a maximally nonintegrable hyperplane
distribution ξ = Ker α d
such that (ξ, ω) and (ξ, α) are homotopic in the
space of pairs (tangent hyperplane distribution η, 2-form of maximal rank
on η).
Proof. Let ξ ⊂ T V = T (V ×R) be the Whitney sum of ξ, pulled back to V ,

and the trivial line bundle tangent to the second factor. The form ω on V
extends in a homotopically canonical way to a 2-form ω on V such that its
restriction to ξ is nondegenerate. Hence, Gromov’s h-principle for contact
structures on open manifolds provides a contact structure ξ on V in the
formal homotopy class prescribed by the pair (ξ, ω). Unfortunately, ξ is not
necessarily transverse to V = V × 0 ⊂ V . However, by applying Theorem
22.2.2 we can deform V via a regular homotopy to achieve transversality to
ξ. Then the distribution ξ induced on V by this transverse map from ξ has
the required properties.
A similar argument proves also the parametric and relative versions of the
h-principle 22.2.3.
Chapter 23
Microflexible
A-invariant Differential
Relations
We generalize in this section the local h-principle 21.5.1 to a class of differ-

ential relations which are invariant only with respect to a certain subgroup
of Diff V . Let the abstract definition of the allowable class of subgroups
not mislead the reader: the only interesting applications which we consider
below are concerned with the groups of symplectic and contact diffeomor-
phisms.
23.1. A-invariant differential relations

Let X → V be a natural fibration. Given a subgroup A ⊂ Diff V , the rela-
tion R is called A-invariant if h∗ (R) = R for all h ∈ A. For instance, the
relations Risosymp and Risocont which define isosymplectic and isocontact im-
mersions are not Diff V -invariant. However, they are invariant with respect
to the subgroups of symplectic and contact diffeomorphisms, respectively.
Let A be a Lie subgroup of the group of compactly supported diffeomor-

phisms of V and a its Lie algebra of vector fields. We call a (and A) capacious
if it satisfies the following two conditions:
(CAP1 ) for any v ∈ a, any compact subset A ⊂ V and its neighborhood

U ⊃ A there exists a vector field ṽA,U ∈ a which is supported in U
and which coincides with v on A;
289
290 23. Microflexible A-invariant Differential Relations
(CAP2 ) given any tangent hyperplane τ ⊂ Tx (V ), x ∈ V , there exists a

vector field v ∈ a which is transverse to τ .
Moreover, we require both properties (CAP1 ) and (CAP2 ) to hold paramet-
rically for any compact space of parameters.
The two examples of capacious subgroups most important for us are the
identity component of the group of compactly supported contact diffeo-
morphisms of a contact manifold, and the group of compactly supported
Hamiltonian diffeomorphisms of a symplectic manifold.
Remark. The notion of a capacious subgroup of diffeomorphisms is a
version of the notion of a set of sharply moving diffeotopies in Gromov’s
book [Gr86].
23.2. Local h-principle for microflexible A-invariant relations

23.2.1. (Local h-principle) Let A ⊂ Diff V be a capacious subgroup,
X → V a natural fibration, and R an A-invariant locally integrable and
microflexible differential relation. Then all forms of the local h-principle hold
for R near any subpolyhedron A ⊂ V of positive codimension. In particular,
if V is a symplectic (contact) manifold, then the local h-principle holds for
Ham V -invariant (Diff cont V -invariant) locally integrable and microflexible
differential relations.
Remark. As we will see below, only the invariance of the differential
relation R with respect to arbitrarily C 0 -small diffeomorphisms from the
capacious group A will be used in the proof. Hence, the h-principle 23.2.1
remains true if R is invariant only with respect to diffeomorphisms from an
arbitrarily small C 0 -neighborhood of Id in the group A.
Proof. The only problem here, compared with the proof of the h-principle
21.5.1 for microflexible DiffV -invariant relations, is that the fibered shifting
diffeotopy hτ , provided by the Holonomic R-approximation Theorem 21.4.1,
does not necessarily belong to the subgroup A. Let us recall here the scheme
of the proof of 21.4.1 and show how this problem could be corrected.
First we observe that the property (CAP2 ) guarantees that the polyhedron A
can be subdivided to ensure that each of its simplices Δ admits a transverse
vector field vΔ ∈ a. Next we choose near each simplex Δ a coordinate
system as in Theorem 3.2.1, which identifies a slightly smaller domain in
the simplex with the coordinate cube I k , and the vector field vΔ with the
coordinate vector field ∂x∂n . The key ingredient in the proof of Theorem
3.2.1 is Inductive Lemma 3.4.1 and its second version 3.5.1. No changes are
necessary in the proof of Inductive Lemma 3.4.1 (except that we substitute
23.2. Local h-principle for microflexible A-invariant relations 291
the openness condition by local integrability and microflexibility of R). As

a result we obtain a family of holonomic sections Fz : Ωz → R defined on
domains
N
N
Ωz = (Ui,z \ Ai,z ) ∪ Op Bi,z ,
i=1 1
where we use the notation introduced in Lemmas 3.6.1 and 3.6.2, i.e.,
Bi,z = z × iσ × I l , i = 0, . . . , N,
and for i = 1, . . . , N
i,z = Nδ (z × ci × I l ) ∩ {ci − σ/2 < t < ci + σ/2},
U

Ai,z = Uδ/4 (z × ci × I l ) \ Vδ (z, ci ) ∩ {(x1 , . . . , xk−l−1 ) = z, xk−l = ci },
2i−1 1
where ci = 2N and σ = N.
We also set

i =
U i,z , Ai =
U Ai,z , Bi = Bi,z , i = 1, . . . , N.
z∈I k−l−1 z∈I k−l−1 z∈I k−l−1
To finish the proof we need the following analogue of Lemma 3.5.1.
23.2.2. For i = 1, . . . , N there exist compactly supported diffeotopies hτi :
i → U
U i , τ ∈ [0, 1], such that
• hτi ∈ A;
i ) ∩ Ai = ∅.
• h1i (I k ∩ U
Proof of 23.2.2. By construction the vector field v = ∂x∂n belongs to a.

Using the property (CAP1 ) for each i = 1, . . . , N , we can find a field ṽi ∈ A
which coincides with vi on
Nδ/2 (Ai ) ∩ {ci − σ/4 ≤ t ≤ ci + σ/4}
and is supported in a slightly bigger subset of U i . Let eτ ṽi be the flow
generated by the vector field ṽi . Let l be any line parallel to the xn -axis
which intersects the set Ai . Set
λ = Ai ∩ l = {−δ/4 ≤ xn ≤ δ/4} ∩ l
and let λ̄ be the interval
{−3δ/4 ≤ xn ≤ −δ/4} ∩ l.
The flow eτ ṽi for the time T = δ/2 slides λ̄ along l with speed 1 and hence
i which is defined by rescaling
eT ṽi (λ̄) = λ. Therefore the isotopy hτi on U
the time parameter of this flow:
hτi = e(δ/2)τ ṽi , τ ∈ [0, 1],
disjoins I k from Ai .
292 23. Microflexible A-invariant Differential Relations
Now we can complete the proof of Theorem 23.2.1. Notice that the isotopies
hτi , i = 1, . . . , N , fit together into a smooth isotopy hτ ∈ A which is defined
on Op I k , and has the property that the image h1 (z × I × I l ) is contained
in the domain Ωz where the holonomic section Fz provided by Inductive
Lemma 3.4.1 is defined. Hence, one can use h1 to pull back the section Fz
to a neighborhood of z × I × I l and thus continue inductively in constructing
the solution of R precisely as in the proof of Theorem 3.2.1.
Chapter 24
Further Applications to
Symplectic Geometry
24.1. Legendrian and isocontact immersions

As an application of the h-principle 23.2.1 we get
24.1.1. (Local h-principle for isocontact immersions) Let (V, ξV ) and
(W, ξW ) be contact manifolds and A ⊂ V a polyhedron of positive codimen-
sion. Then all forms of the local h-principle hold for isocontact immersions
(Op A, ξV |Op A ) → (W, ξW ).
Indeed, the corresponding differential relation Risocont is locally integrable,

microflexible, and invariant with respect to the capacious group Diff cont (V ).
Theorem 24.1.1 and the microextension trick imply the following.
24.1.2. (Homotopy principle for isocontact immersions, Gromov
[Gr86]) If dim V < dim W , then all forms of the h-principle hold for iso-
contact immersions (V, ξV ) → (W, ξW ).
Proof. Let N be the normal bundle to F (ξV ) ⊂ ξW with respect to the

conformal symplectic structure CS(ξW ). Then N has the structure of a sym-
plectic vector bundle. According to Lemma 17.4.1, there exists a contact
structure ξN on a neighborhood Op V of the 0-section in the total space of
the bundle N such that (V, ξV ) is a contact submanifold of (Op A, ξN ), the
fibers of the bundle N are tangent to ξN |V and serve as orthogonal com-
plements of ξV in ξN with respect to the conformal symplectic structure
CS(ξN ). Then any isocontact homomorphism
F : (T V, ξV ) → (T W, ξW )
293
294 24. Further Applications to Symplectic Geometry
canonically extends to an equidimensional isocontact homomorphism

F : (Op V, ξN ) → (T W, ξW ).
On the other hand, any isocontact immersion (Op V, ξN ) → (W, ξW ) restricts
to V as an isocontact immersion (V, ξV ) → (W, ξW ).
Similarly, any formal Legendrian immersion F : T V → T W can be canoni-

cally extended to a formal equidimensional isocontact immersion
F : T (J 1 (V, R)) → T W.
Hence, we get
24.1.3. (Homotopy principle for Legendrian immersions; Gromov
[Gr71], Duchamp [Du84]) All forms of the h-principle hold for Legen-
drian immersions V → (W, ξ).
Of course, the h-principle for subcritical isotropic immersions into contact

manifolds also follows from Theorem 24.1.2. However, as we already noted
in Section 21.5, this h-principle also follows from Theorem 21.5.2 which we
proved earlier.
Remark. Parametric forms of all the h-principles considered in this sec-
tion remain true (with the same proof) in the fibered form (see Section
7.2.E), i.e., when the contact structures on the source, target, or both are
also allowed to vary with the parameter.
24.2. Generalized isocontact immersions

We will now generalize Theorem 24.1.2 for isometric mappings of arbitrary,
not necessarily contact distributions. Let η be an arbitrary tangent distri-
bution of codimension s on a manifold V and ξ a contact structure on a
manifold W . An immersion f : (V, η) → (W, ξ) is called isocontact if it
is transverse to ξ and df (η) ⊂ ξ. Let us now formulate the corresponding
formal notion. Let η ∗ ⊂ T ∗ V be the bundle conormal to η. Sections of η ∗
are 1-forms annihilating the distribution η. Let Dη be the vector bundle
which is pointwise generated by the sections dθ|η , where θ ∈ Sec η ∗ . Note
that any section of Dη has the form dθ|η , where θ ∈ Sec η ∗ . Indeed, for any
forms θi ∈ Sec η ∗ and functions fi : V → R, i = 1, . . . , k, we have

k
k
(
fi (dθi |η ) = d fi dθi (η = dθ|η ,
1 1

k
where θ = fi dθi ∈ Sec θ∗ . If g : (V, η) → (W, ξ) is an isocontact immersion
1
and ξ = Ker α, then α ◦ dg ∈ Sec η ∗ , and hence (g ∗ dα)|η is a section of the
bundle Dη. This motivates the following definition:
24.2. Generalized isocontact immersions 295
A monomorphism
F : (T V, η) → (T W, ξ = Ker α)
(
is called isocontact if it is transverse to ξ and F ∗ dα (η ∈ Sec Dη.
24.2.1. (Gromov [Gr86] and Datta [Da97]) Let (W, ξ) be a contact
manifold and η an arbitrary tangent distribution on a manifold V . Let
A ⊂ V be a polyhedron of positive codimension.
(a) All forms of the local h-principle hold for isocontact immersions
(Op A, η) → (W, ξ).
(b) If dim W > dim V , then all forms of the global h-principle hold for
isocontact immersions (V, η) → (W, ξ).
To prove Lemma 24.2.1 we will need the following lemma.

24.2.2. Let η = Ker α be a tangent hyperplane distribution on a manifold V
and
π:E→V
a vector bundle over V . Let us denote by η the subbundle of T (E)|V which
is the direct sum of η and the vector bundle E → V viewed as subbundle of
the bundle T (E)|V . Suppose that there exists a not necessarily closed 2-form
Ω on E such that Ω|η is nondegenerate and Ω|η = dα|η . Then η extends to
a contact structure η on E.
Proof of Lemma 24.2.2. Using an appropriate partition of unity, we can

reduce the proof to the case when the bundle E → V is trivial, E = V × Rs .
Let t1 , . . . , ts be the coordinates corresponding to the second factor. The
form Ω can be presented as the sum Ω = Ω + d α where α = π ∗ α, such that

Ω |τ = 0. In particular, the form Ω can be written as

k

∧β+
Ω =α dti ∧ βi .
1
Consider the 1-form

s
+
α=α ti αi .
1
The distribution η = Ker α coincides with η along V , and dαη = Ω|η . In
particular, η is a contact structure on Op V ⊂ E.
Proof of 24.2.1. The global h-principle in the case dim W > dim V fol-
lows from the local one via the standard microextension argument. So we
will prove here only the local h-principle and only in the nonparametric
case, and we will leave the general case as an exercise to the (already very
experienced) reader.
Let F : (T (Op A), η) → (W, ξ) be an isocontact homomorphism. Let ξ =

Ker α. Set β = f ∗ α and Ω = F ∗ dα. Set τ = Ker β and let E → Op A
be the vector bundle whose fiber over a point v ∈ Op A is the orthogonal
complement (with respect to some Riemannian metric) to F (τv ) in ξf (v) ,
where f : Op A → W is the map underlying the homomorphism F . Consider
the subbundle of τ ⊂ T (E)|Op A which is spanned by τ and the fibers of the
vector bundle E. The subbundle τ is symplectic. The linear symplectic
structure on its fibers is given by the pull-back Ω of the symplectic form
dα on ξ under the isomorphism F : τ → ξ which canonically extends the
injection F : τ → ξ. Hence, we can first use Lemma 24.2.2 to extend τ
to a contact structure τ on a neighborhood OpE A of the 0-section Op A
in E, and then apply the local h-principle 24.1.1 for isocontact immersions
g : (OpE A, τ ) → (W, ξ). Then the restriction of this map to a neighborhood
Op A of A in V is automatically an isocontact immersion (Op A, τ ) → (W, ξ).
Combined with the inclusion η → τ it gives us the required isocontact
immersion (Op A, η) → (W, ξ).
24.3. Lagrangian immersions

Let us recall that a symplectic manifold (W, ω) is called exact if ω = dα.
A Lagrangian immersion f : V → (W, ω = dα) is called exact if the closed
form f ∗ α is exact.
24.3.1. Let (W, ω = dα) be an exact 2n-dimensional symplectic manifold.
Then for any n-dimensional manifold V all forms of the h-principle hold
for exact Lagrangian immersions V → W . In particular, any family of
isotropic monomorphisms T V → (T W, ω) is homotopic in this class to a
family of differentials of exact Lagrangian immersions. Moreover, the same
is true when the symplectic form ω on W depends itself on the parameter.
Proof. Any isotropic monomorphism F : T V → T W homotopically canon-

ically lifts to an isotropic monomorphism into the contact bundle
ξ = Ker (dz − α)
on W × R. Conversely, any Legendrian immersion V → (W × R, ξ) projects
to an exact Lagrangian immersion into (W, dα). Thus it remains to apply
Theorem 24.1.3 (and the remark which follows it).
24.3.2. (Homotopy principle for Lagrangian immersions; Gromov

[Gr71], Lees [Le76]) For any n-manifold V and symplectic 2n-manifold
(W, ω), all forms of the h-principle hold for Lagrangian immersions V →
(W, ω) as long as the cohomological condition [f ∗ ω] = 0 is incorporated in
the definition of formal solutions of RLag .
24.4. Isosymplectic immersions 297
Proof. Let F : T V → T W be an isotropic monomorphism such that

[f ∗ ω] = 0. Using the Smale–Hirsch h-principle for immersions (see Theorem
9.2.1), we may assume from the very beginning that f = bs F is a differen-
tiable immersion. Let N be the total space of the normal bundle to f (V ) in
W . The immersions f : V → W can be extended to immersions f : N → W ,
so that the F = df ◦ G. The monomorphism G : T V → T N is isotropic
with respect to the induced symplectic form ω = f∗ ω. By assumption this
form is exact, ω = dα, and hence one can use 24.3.1 to finish the proof in
the nonparametric case.
Now let
Ft : T V → T W, t ∈ D k ,
be any family of isotropic monomorphisms such that [ft∗ ω] = 0, and Ft = dft
for t ∈ ∂D k , where ft is the map V → W underlying Ft . Using the Smale–
Hirsch h-principle for immersions, we may assume from the very beginning
that ft = bs Ft is a family of immersions. Let N be the total space of the
normal bundle to f0 (V ) in W . The immersions ft : V → W can be extended
to immersions ft : N → W so that the family Ft can be decomposed as
Ft = dft ◦ Gt , t ∈ D k . The monomorphism Gt : T V → T N is isotropic with
respect to the induced symplectic form ω t = ft∗ ω, t ∈ D k . By assumption
these forms are exact, ω t = dαt . Unfortunately, now we cannot use 24.3.1
because our original Lagrangian immersions ft : V → (W, ω), t ∈ ∂D k may
well be nonexact. This is, however, a minor problem because instead of αt
one can take a new family of primitives {αt − α t }t∈Dk for ω t , where α
t is a
family of closed 1-forms on N such that
t = p∗ (αt |V ) for t ∈ ∂D k ,
α
where p : N → V is the projection. With respect to this new family of
primitives the Lagrangian immersions ft : V → (W, ω), t ∈ ∂D k are exact
and we can apply 24.3.1 to finish the proof.
24.4. Isosymplectic immersions

Let now (V, ωV ) and (W, ωW ) be two symplectic manifolds, dim W ≥ dim V ,
and A a subpolyhedron of positive codimension in V .
24.4.1. Let U be an open subset of the product V × W such that the form
Ω = ωV ⊕ ωW is exact on U , Ω|U = dα. Then all forms of the local h-
principle hold for Lagrangian (isotropic) sections s : Op A → U ⊂ V × W .
Proof. As in the proofs of 24.3.1 and 24.3.2 we can reduce the problem for
Lagrangian (isotropic) sections Op A → U to the problem about Legendrian
(isotropic) sections Op A → (U × R, ξ = Ker (dz − α)). The differential
relation which corresponds to the Legendrian problem is locally integrable,
microflexible and invariant with respect to a sufficiently small neighborhood

of the identity in the capacious group Ham(V, ω). Hence, according to Theo-
rem 23.2.1 we deduce the required local h-principle.
Let Risosymp be the differential relation corresponding to the problem

of isosymplectic immersions (V, ωV ) → (W, ωW ). For a polyhedron A ⊂ V
of positive codimension we denote by Sec0Op A Risosymp the subspace in
SecOp A Risosymp which consists of sections F : Op A → J 1 (V, W ) which
satisfy the cohomological condition
[f ∗ ωW ] = [ωV |Op A ],
where f : Op A → W is the map underlying the section F .
24.4.2. (Local h-principle for isosymplectic immersions) All forms
of the local h-principle hold for the inclusion
HolOp A Risosymp → Sec0Op A Risosymp
near any polyhedron A of positive codimension.
Proof. To spare the reader from extra indices, we consider here only the
nonparametric case. Let F be a section from Sec0Op A Risosymp and f :
Op A → W its underlying map. Let us choose an open neighborhood U of
the 0-jet part of the section F , i.e., of the graph f (x) = (x, f (x)), x ∈ Op A.
Note that isosymplectic immersions Op A → W can be characterized by the
property that their graphs are isotropic with respect to the symplectic form
Ω = ωV ⊕ (−ωW ) on J 0 (V, W ) = V × W . By assumption the form Ω|U
is exact, and hence we can use Lemma 24.4.1 to C 0 -approximate f by an
isotropic section x → (x, g(x)), x ∈ Op A. Then g : Op A → W is the
required isosymplectic immersion.
Theorem 24.4.2 and the microextension trick imply the following theorem.
24.4.3. (Homotopy principle for isosymplectic immersions, Gro-
mov [Gr86]) All forms of the h-principle hold for isosymplectic immersions
(V, ωV ) → (W, ωW ) as long as dim V < dim W and the algebraic condi-
tion [f ∗ ωW ] = [ωV ] is incorporated in the definition of formal solutions of
Rsymp-iso .
Proof. Let U be the ωW -normal bundle to F (T V ) in T W . According to

Lemma 17.2.2, a neighborhood Op V of the 0-section V in the total space
of this bundle admits a symplectic form ωN such that ωN |V = ωV and the
fibers of the bundle N are ωN -orthogonal to T V in T N |V . Any formal
isosymplectic immersion
F : (T V, ωV ) → (T W, ωW )
24.5. Generalized isosymplectic immersions 299
then canonically extends to a formal equidimensional isosymplectic immer-

sion
F : (T (Op V ), ωN ) → (T W, ωW ).
Hence the result follows from Theorem 24.4.2.
24.5. Generalized isosymplectic immersions

The h-principle for isosymplectic immersions can be generalized to the case
when the form on the source manifold is not necessarily nondegenerate.
Let σ be a closed 2-form on a manifold V , and (W, ω) be a symplectic

manifold. An immersion f : (V, σ) → (W, ω) is called isosymplectic if f ∗ ω =
σ. A monomorphism F : (T V, σ) → (T W, ω) is called isosymplectic if F ∗ ω =
σ and the equality f ∗ [ω] = [σ] holds for the cohomology classes of the forms σ
and ω, where f : V → W is the map which underlies the monomorphism F .
Let us denote by Iso(V, σ; W, ω) and iso(V, σ; W, ω) the space of isosymplectic
immersions (V, σ) → (W, ω) and the space of isosymplectic homomorphisms
(T V, σ) → (T W, ω), respectively. The derivation map defines a natural
inclusion D : Iso(V, σ; W, ω) → iso(V, σ; W, ω).
24.5.1. Let (V, σ) and (W, ω) be as above and dim W > dim V . Then the
map
D : Iso(V, σ; W, ω) → iso(V, σ; W, ω)
To prove Theorem 24.5.1 we will need the following lemma, which is an

analogue of Lemma 24.2.2 which we proved above in the contact case.
24.5.2. Let η be a closed 2-form on a manifold W and π : E → W a vector
bundle over W . Suppose that there exists a nondegenerate, not necessarily
closed 2-form Ω on E with Ω|W = η. Then there exists a symplectic form ω
on Op W ⊂ E such that Ω|T E|W = ω|T E|W .
Proof of Lemma 24.5.2. The form Ω can be presented as the sum Ω =

Ω + π ∗ η, where Ω |W = 0. Suppose first that the fibration E → W is trivial,
so that E = W × Rs . Let t1 , . . . , ts be the coordinates corresponding to the
k
second factor. The form Ω can be presented in the form Ω = dti ∧ αi .
1
Then the closed 2-form

k
ω = π∗η + d(ti αi )
1
coincides on T E|W with Ω, and hence it is nondegenerate in a neighborhood
Op W ⊂ E.
The case of a nontrivial fibration E → W can be reduced to the just con-

sidered case by choosing an appropriate partition of unity on W .
Proof of Theorem 24.5.1. Let F : (T V, σ) → (T W, ω) be an isosym-

plectic immersion. Let E → V be the normal bundle to the formal im-
mersion F , i.e., its fiber over a point v ∈ V is the orthogonal comple-
ment to F (Tv V ) in Tf (v) W , where the monomorphism F covers the map
f : V → W . The injective homomorphism F canonically extends to an
isomorphism F : T E → T W . The 2-form σ = F ∗ ω is nondegenerate and
coincides with σ on v. Applying Lemma 24.5.2, we get a symplectic form η
on Op V such that σ |T N |V = η|T N |V . Then the local h-principle 24.4.2 allows
us to construct an isosymplectic immersion g : (Op V, η) → (W, ω). The re-
striction of g to V is the required isosymplectic immersion (V, σ) → (W, ω).
This finishes off the proof in the nonparametric case. The generalization to
the parametric case is straightforward.
Part 5
Convex Integration
Chapter 25
One-Dimensional
Convex Integration
25.1. Example
Let us call a path
r : I = [0, 1] → R2 , r(t) = (x(t), y(t)),
short if ẋ2 + ẏ 2 < 1. The graph of a short path in the space-time R3 = R×R2
(where the space is two dimensional) is a time-like world curve: its tangent
line at a point (t0 , x0 , y0 ) lies inside the light cone
(t − t0 )2 ≥ (x − x0 )2 + (y − y0 )2 ,
assuming that the speed of light c = 1 (see Figure 25.1).
x,y
Figure 25.1. The graph of a short path.
303
304 25. One-Dimensional Convex Integration
Exercise. Prove that any short path can be C 0 -approximated by a solu-

tion of the equation ẋ2 + ẏ 2 = 1. In other words, the world line of a particle
can be C 0 -approximated by the world line of a photon. (Hint: Any path of
length l can be C 0 -approximated by a path of length L for any L > l.)
The above exercise implies that the space of solutions I → R2 of the dif-
ferential equation ẋ2 + ẏ 2 = 1 is C 0 -dense in the space of solutions of the
differential inequality ẋ2 + ẏ 2 < 1.
The exercise illuminates the following idea: given a first order differential
relation for maps I → Rq , it is useful to consider a relaxed differential
relation which is the fiberwise convex hull of the original relation. A direct
implementation of this idea in the context of ordinary differential equations
(ODEs) is known to specialists in control theory as Filippov’s relaxation
theorem ([Fi67]). A far more subtle implementation in the context of partial
differential equations (PDEs) is known to specialists in differential topology
as Gromov’s convex integration theory ([Gr73], [Gr86]). It is interesting
to mention that for a long time specialists from both sides did not know
about the existence of a parallel theory. It was D. Spring who pointed out
the connection; see the discussion and historical remarks in [Sp98].
25.2. Convex hulls and ampleness

A differential relation R ⊂ J 1 (R, Rq ) can be thought of, in the spirit of
control theory, as a differential inclusion
ẏ ∈ Ω(t, y), t ∈ R, y ∈ Rq ,
where Ω(t, y) = R ∩ Pt,y and Pt,y Rq is the fiber of the projection
J 1 (R, Rq ) → J 0 (R, Rq ) = R × Rq
over a point (t, y) ∈ R × Rq . This fiber can be identified with the tangent
space Ty (t × Rq ) = Ty Rq , which, by definition, consists of all vectors in Rq
originating at the point y ∈ Rq .
Remark. We use here the letter t for a one-dimensional independent
variable, reserving the letter x for an n-dimensional independent variable
in our further generalizations in Section 26.2.
Given an affine space P , a set Ω ⊂ P , and a point y ∈ Ω, we will denote

by Conny Ω the path-connected component of Ω which contains y and by
Convy Ω the convex hull of Conny Ω. A set Ω ⊂ P is called ample if
Convy Ω = P for all y ∈ Ω. Note that according to this definition the empty
set is ample. A differential relation R ⊂ J 1 (R, Rq ) is called ample if R is
fiberwise ample, i.e., Ω(t, y) ⊂ Rq is ample for every (t, y) ∈ R × Rq .
25.3. Main lemma 305
Given a differential relation R ⊂ J 1 (R, Rq ) and a section

F = (f, ϕ) : R → R,
we will use the following notation:
• ConnF (t) R—the path-connected component of Ω(f (t)) which con-
tains F (t) in the fiber P(t,f (t)) of the fibration
p10 : J 1 (R, Rq ) → J 0 (R, Rq )
over the point (t, f (t));
• ConvF (t) R—the convex hull of ConnF (t) R;

• ConvF R—the differential relation ConvF (t) R in J 1 (R, Rq ).
t∈R
We will call a formal solution F = (f, ϕ) of R ⊂ J 1 (R, Rq ) short, if its

0-component f is a genuine solution of ConvF R.
Given a fiberwise path-connected differential relation R ⊂ J 1 (R, Rq ), we will

denote by Conv R the fiberwise convex hull of R.
25.3. Main lemma

25.3.1. (One-dimensional convex integration) Let R ⊂ J 1 (R, Rq ) be
an open differential relation and F = (f, ϕ) : I → R be a short formal
solution of R. Then there exists a family of short formal solutions
Fτ = (fτ , ϕτ ) : I → R, τ ∈ [0, 1],
which joins F0 = F to a genuine solution F1 such that
(a) fτ is (arbitrarily) C 0 -close to f for all τ ∈ [0, 1];
(b) Fτ (0) = F (0) and Fτ (1) = F (1) for all τ ∈ [0, 1].
Remark. If the formal solution F is already genuine near ∂I, then the
homotopy Fτ can be chosen fixed near ∂I. Indeed, we can apply Lemma
25.3.1 to a smaller interval [δ, 1 − δ] ⊂ I. Strictly speaking, the constructed
solution f1 is only C 1 -smooth at the points δ and 1−δ. However, the relation
R is open and hence f1 can be approximated by a C ∞ -smooth solution.
25.3.2. (Corollary) Let R ⊂ J 1 (R, Rq ) be an open ample differential
relation. Then for any formal solution F = (f, ϕ) : I → R which is genuine
near ∂I there exists a homotopy of formal solutions, fixed near ∂I,
Fτ = (fτ , ϕτ ) : I → R, τ ∈ [0, 1], F0 = F,
such that F1 is a genuine solution of R over I and f1 is (arbitrarily) C 0 -
close to f .
Indeed, the ampleness of the relation R implies that any formal solution of
R is short, and hence we can apply Lemma 25.3.1.
25.3.3. (Corollary) Let R ⊂ J 1 (R, Rq ) be an open and fiberwise path-
connected differential relation. Then the space of solutions I → Rq of R is
C 0 -dense in the space of solutions of Conv R. If, in addition, R is ample
and fiberwise nonempty then the space of solutions I → Rq of R is C 0 -dense
in the space of all maps I → Rq .
Indeed, the assumption that R is open and fiberwise path-connected guar-
antees the existence of a formal solution F = (f, ϕ) of R for any solution f
of Conv R, and hence we can apply Lemma 25.3.1.
25.4. Proof of the main lemma

A. Flowers. An abstract flower S is a union of a finite number of copies
I0 , I1 , I2 , . . . of the interval I = [0, 1] with their left ends identified to one
point denoted 0S . The interval I0 ⊂ S is called the stem of the flower, all
the other intervals I1 , . . . are called the petals; see Figure 25.2.
Figure 25.2. An abstract flower.
The ordering of petals is not essential for our purpose. We will denote by
∂S the union of free ends of the petals Ii , i = 1, . . . of the flower S.
A map ψ : S → Rq , and sometimes also its image Ψ = ψ(S), will be called

a flower. The parametrizing map ψ : S → Rq is a union of paths
ψi : I → Rq , ψi (0) = ψ(0S ).
Let us point out that we do not assume the parametrizing map to be one-to-
one. In particular, the map ψ may contract some of the petals or the stem
into the point ψ(0S ). Given a flower Ψ = ψ(S), we set ai = ψi (1), i = 1, . . .
and ∂Ψ = ψ(∂S).
B. Reduction of Lemma 25.3.1 to a special case. Denote by Dε the

standard ε-ball in Rq .
25.4.1. It is sufficient to prove Lemma 25.3.1 for the case when:
• R = R × Dεq × Ψ ⊂ J 1 (R, Rq ),
where Ψ = ψ(S) ⊂ Rq is a flower such that 0 ∈ Int Conv(∂Ψ);
• F = (0, ϕ) : I → R where ϕ ≡ ψ0 .
Remarks.
1. Here and in what follows we identify the section
ϕ : I → I × 0 × Rq ⊂ J 1 (R, Rq ), t → (t, 0, ϕ̄(t)),
with the map ϕ̄ : I → Rq .
2. The relation R = R × Dεq × Ψ is closed.
3. The formal solution F = (0, ϕ) is automatically short.
Proof. In the general case of Lemma 25.3.1 we can put z = y − f (t) and
consider, instead of R ∼ {ẏ ∈ Ω(t, y)}, the variation relation along f (t)
R ∼ {ż ∈ Ω(t,
z) = Ω(t, z + f (t)) − f˙(t)}
and its short formal solution (0, ϕ − f˙). To reduce further, we need the
following
25.4.2. (Sublemma) Let R ⊂ J 1 (R, Rq ) be an open differential relation
and
F = (0, ϕ) : I → R
be a short formal solution. Then there exists a number δ > 0 such that for
any t0 ∈ [0, 1 − δ] one can choose a flower
Ψ = Ψ(t0 ) ⊂ Pt0 ,0 Rq
with the properties
(a) 0 ∈ Int Conv(∂Ψ);
(b) ψ0 (t) = ϕ(t0 + δt), t ∈ I;
(c) [t0 , t0 + δ] × Dεq × Ψ ⊂ R for sufficiently small ε > 0.
Proof. Let t0 ∈ I. We can choose a finite set of points in Conn F (t0 ) R

such that 0 belongs to the interior of the convex hull of these points and
then connect the basepoint ϕ(t0 ) with the chosen points by some paths
I → Conn F (t0 ) R. These paths are petals of our flower Ψ = Ψ(t0 ), while
the path ψ0 (t) = ϕ(t0 + δt) is its stem. Then the flower Ψ has properties
(a) and (b). Using the openness of the relation R and the compactness of
the interval I, one can choose δ to satisfy (c).
Using Sublemma 25.4.2 and an appropriate subdivision of the interval I, we

reduce Lemma 25.3.1 to the required special case.
C. Uniform and weighted products of paths. Let us define the (uni-

form) product
p = p1 • · · · • pk : I → Rq
of k paths pi : I → Rq in the following natural way: p is the path of a
particle which successively goes along each pi (I) in 1/k seconds, i.e.,
p(t) = pi (k(t − (i − 1)/k)) , t ∈ ((i − 1)/k, i/k], p(0) = p1 (0).
If pi (1) = pi+1 (0), then p is continuous, otherwise p is only piecewise con-
tinuous.
We will need also a weighted product of paths. Let α1 + · · · + αk = 1, αi > 0.

The product p = p1 • · · · • pk weighted by (α1 , . . . , αk ) is the trajectory of a
particle which successively goes along each pi (I) in αi seconds, i.e.,
! "
t − ti−1
p(t) = pi , t ∈ (ti−1 , ti ], p(0) = p1 (0),
αi

i
where ti = αi , i = 1, . . . , k − 1, and t0 = 0. In particular, the uniform
1
product of k paths is the product weighted by (1/k, . . . , 1/k). We allow αi
to be equal to zero if pi is a constant path.
Given a path p : I → Rq , we will denote by pN the uniform product

p • ··· • p

of N factors, and by p(σ) dσ the path
t
t→ p(σ) dσ.
0
1
The path p : I → Rq is called balanced if 0 p(σ) dσ = 0.
In what follows we will need the two following evident properties of the
uniform product.
25.4.3. (“Multiplicativity” of the integral) Let pi : I → Rq , i =

1, . . . , N , be balanced paths. Then

(p1 • · · · • pN )(σ) dσ = (1/N ) p1 (σ) dσ • · · · • pN (σ) dσ.
25.4.4. (C 0 -norm of the uniform product and convergence to zero)

Let pi : I → Rq be as in 25.4.3. Then
) )
) )
) (p1 • · · · • pN )(σ) dσ ) 0
C
4) ) ) ) 5
) ) ) )
= (1/N ) max ) p1 (σ) dσ ) 0 , . . . , ) pN (σ) dσ ) 0 .
C C
In particular, if p is a balanced path, then

C0
pN (σ) dσ −→ 0.
N →∞
D. Piecewise linear solution. Let us recall that we consider a special

case of Lemma 25.3.1 described in 25.4.1.
By assumption we have 0 ∈ Int Conv(∂Ψ), where ∂Ψ = {a1 , . . . , ak }, so

we can write 0 as a convex combination 0 = α1 a1 + · · · + αk ak , αi > 0.
Let δ : I → Ψ be the product of constant paths ai : I → ai , weighted by
(α1 , . . . , αk ), so that δ is a piecewise constant discontinuous path. Then
1
δ(t) dt = 0.
0
In particular, δ is a balanced path.
Let ϕδ1 = δ • δ • · · · • δ be the uniform product of N factors. We define a

continuous piecewise linear path f1δ : I → Rq by the formula
t
δ
f1 (t) = ϕδ1 (σ) dσ.
0
As follows from 25.4.4,

1
f1δ C 0 = gC 0 , where g = δ(σ) dσ.
N
Therefore, for N > 1ε gC 0 the map f1δ is a piecewise linear solution of R.
Now we want to realize the same idea in constructing a smooth solution.

We will approximate the section ϕδ1 by a smooth section ϕ1 . In addition, we
need to satisfy the boundary conditions for F1 = (f1 , ϕ1 ).
E. Smooth solution. Let ψ = {ϕ, ψ1 , . . . , ψk } be the parametrizing map

for the flower Ψ. To ensure the smoothness of our further construction, we
will assume that ψi (t) = ϕ(0) near t = 0 and ψi (t) = ai near t = 1.
Consider the product

ψ = ψ1 • a1 • ψ1−1 • · · · • ψk • ak • ψk−1 ,
where the weights of constant paths ai are equal to (1−ρ)αi and the weights
of all other paths are equal to ρ/2k.
For what follows we need to balance the loop ψ, i.e., we need the equality
1
ψ(t)dt = 0.
0
This can be achieved by adjusting the weights of constant paths. Indeed,
let 1 1 1
d= ψ(t) dt = ψ(t) dt − δ(t) dt ∈ Rq .
0 0 0
Then d < Cρ, where
C = max ψ(t), t ∈ I.
If ρ is sufficiently small, then d ∈ Int{(1−ρ) Δ} where Δ = Conv{a1 , . . . , ak }.
Here the multiplication by (1 − ρ) means the homothety centered at the ori-
gin. Note that
(1 − ρ)Δ = Conv{(1 − ρ)a1 , . . . , (1 − ρ)ak }.
Hence we can present −d as the convex combination
1 (1 − ρ) a1 + · · · + α
−d = α k (1 − ρ) ak .
Therefore, if we assign the new weights
1 (1 − ρ), . . . , α
α q+1 (1 − ρ)
to the constant paths ai , then the integral of ψ(t) will be equal to 0.
Now apply the same balancing construction to the path

ψ = ψ1 • a1 • ψ1−1 • · · · • ψk • ak • ψk−1 • ϕ,
choosing the weights of constant paths ai equal to (1 − ρ)αi and the weights
of all other paths equal to ρ/(2k + 1). In other words, we choose the new
weights for constant paths ai while keeping the weights of all others paths
equal to ρ/(2k + 1), so that we get
1

ψ(t)dt = 0.
0
25.5. Parametric version of the main lemma 311
Let
ϕ1 = ψ • ψ • · · · • ψ •ψ

N −1
be the uniform product of N factors, and f1 be defined by the formula
t
f1 (t) = ϕ1 (σ) dσ.
0
Then 25.4.4 implies that
1
f1 C 0 =max{gC 0 , hC 0 },
N
where
g = ψ(σ) dσ and h = ψ(σ) dσ.
If N is sufficiently large, then F1 = (f1 , ϕ1 ) is a genuine solution of R and,
moreover, F1 satisfies the boundary conditions. The construction of the
homotopy (fτ , ϕτ ) is straightforward: the linear homotopy τ f1 (t) consists
of solutions of Conv R, and together with the canonical homotopies
ψi ◦ ψi−1 ∼ v0 = ϕ(0)
it gives us the required homotopy Fτ = (τ f1 , ϕτ ) in R. This finishes off the
proof of the Main Lemma 25.3.1.
25.5. Parametric version of the main lemma

25.5.1. (Parametric one-dimensional convex integration) Let R ⊂
I l × J 1 (R, Rq ) be an open fibered differential relation (see Section 7.2.E) and
F = F (p, t) = (f (p, t), ϕ(p, t)) : I l × I → R
be a fiberwise short formal solution of R, i.e., for each p ∈ I l the section
F (p, t) : p × I → Rp = R ∩ p × J 1 (R, Rq )
is a short formal solution of Rp . Suppose that f (p, t) smoothly depends on
p and consists of genuine solutions of Rp when p ∈ Op I l . Then there exists
a homotopy of fiberwise short formal solutions
Fτ = Fτ (p, t) = (fτ (p, t), ϕτ (p, t)) : I l × I → R, τ ∈ [0, 1],
which joins F0 = F with a genuine solution F1 of R such that for all τ
(a) fτ is (arbitrarily) C 0 -close to f ;
(b) Fτ (p, 0) = F (p, 0) and Fτ (p, 1) = F (p, 1) for all p ∈ I l ;
(c) Fτ is constant for p ∈ Op (∂I l ), and
(d) the first derivatives of f1 (p, t) with respect to the parameter p are
(arbitrarily) C 0 -close to the respective derivatives of f (p, t).
Remarks.
1. If the formal solution F is already genuine near ∂I, then the homotopy
Fτ can be chosen fixed near ∂I. Indeed, we can apply Lemma 25.5.1 to a
smaller set I l × [δ, 1 − δ] ⊂ I l × I.
2. We will use Lemma 25.5.1 in order to extend the convex integration of

ordinary differential relations (n = 1) to the convex integration of partial
differential relations (n > 1). The property (d) will be crucial for this goal.
25.6. Proof of the parametric version of the main lemma
The proof will follow the same scheme as in the nonparametric version.
A. Fibered flowers. A fibered flower is a fibered (over I l ) map

ψ : I l × S → I l × Rq
(where S is an abstract flower), as well as the set
Ψ = ψ(I l × S) ⊂ I l × Rq
parametrized by this map. Given a fibered flower Ψ, we will denote by Ψp
the flower ψ(p × S) ⊂ p × Rq , p ∈ I l .
B. Reduction to a special case.

25.6.1. It is sufficient to prove Lemma 25.5.1 for the case when:
• the fibered relation R consists of fibers
Rp = p × R × Dεq × Ψp ⊂ p × J 1 (R, Rq ), p ∈ I l ,
where Ψ = ψ(I l × S) ⊂ I l × Rq is a fibered flower such that 0 ∈
Int Conv(∂Ψp ) for each p ∈ I l ;
• F = (0, ϕ) : I l × I → R with ϕ ≡ ψ0 .
Remark. We identify here and further the fibered over I l section
ϕ : I l × I → I l × I × 0 × Rq ⊂ I l × J 1 (R, Rq ), (p, t) → (p, t, 0, ϕ̄(p, t)),
with the fibered map
I l × I → I l × Rq , (p, t) → (p, ϕ̄(p, t)).
Proof. As in the nonparametric case we may assume that f ≡ 0. To make

the further reduction we need the following.
25.6. Proof of the parametric version of the main lemma 313
25.6.2. (Sublemma) Let R ⊂ I l ×J 1 (R, Rq ) be an open fibered differential

relation and
F = (0, ϕ) : I l × I → R
be a fiberwise short formal solution of R. There exists a number δ > 0 such
that for any t0 ∈ [0, 1 − δ] one can choose a fibered flower
Ψ = Ψ(t0 ) ⊂ I l × Pt0 ,0 I l × Rq
such that for each p ∈ I l :
(a) 0 ∈ Int Conv(∂Ψp );
(b) ψ0 (p, t) = ϕ(p, t0 + δt), t ∈ I;
(c) p × [t0 , t0 + δ] × Dεq × Ψp ⊂ Rp for sufficiently small ε > 0.
Proof. Let t0 ∈ I. First take a fibered flower which consists of just its stem
parametrized by the map ψ0 : (p, t) → ϕ(p, t0 + δt) (the number δ will be
chosen later).
For every fixed p0 ∈ I l we can choose a flower Ψp0 as in Sublemma 25.4.2,
and using the openness of R extend Ψp0 over a neighborhood U of p0 ∈ I l
such that for all p ∈ U
• ψi (p, t) are paths in Conn F (p,t0 ) R, and
• 0 ∈ Int Conv(∂Ψp ).
Hence, we can choose a finite covering of I l by open sets Uj , j = 1, . . . , L,
such that over every Uj we have, as above, a flower fibered over Uj , which
we denote by ΨUj . Suppose that its petals are parametrized by the maps
U
ψi j : Uj × I → R, i = 1, . . . , Nj , j = 1, . . . , L.
Let Uj ⊂ Uj , j = 1, . . . , L, be slightly smaller open sets such that

L
Uj ⊂ Uj and Uj ⊃ I l .
1
For every j = 1, . . . , L choose a cut-off function β j : I l → [0, 1] which is

equal to 1 on Uj and equal to 0 on I l \ Uj , and for i = 1, . . . Nj set
U
ψij (p, t) = ψi j (p, β j (p) t) for p ∈ Uj and
ψij (p, t) = ϕ(p, t0 ) for p ∈ I l \ Uj .
The fibered (over I l ) flower with the stem ψ0 (p, t) = ϕ(p, t0 + tδ, p) and
petals parametrized by all the maps ψij for all i = 1, . . . , Nj , j = 1, . . . , L,
satisfies the properties (a) and (b). Therefore using the openness of the
relation R and the compactness of the interval I, one can choose δ to satisfy
(c).
Using Sublemma 25.6.2 and an appropriate subdivision of the interval I,

one can reduce the Parametric Main Lemma 25.5.1 to the required special
case.
C. Convex decomposition of a section. Let Ψ be a flower fibered over

I l . Let us set
ai (p) = ψi (1), i = 1, . . . , N, Δp = Conv ∂Ψp .
Let d : p → Int Δp be a section over I l . Then there exist functions
αi : I l → [0, 1], i = 1, . . . , N,
such that
α1 (p) + · · · + αN (p) = 1 and α1 (p)a1 (p) + · · · + αN (p)aN (p) = d(p).
Indeed, we can construct such a set of functions locally over a neighborhood
of each point p ∈ I l and then globalize the construction using a partition of
unity.
D. Construction of the homotopy Fτ . We can apply the proof of the

Main Lemma parametrically, working with convex decompositions of the
sections 0 and d(p) instead of convex decompositions of the vectors 0 and
d, and with weights αi (p), α i (p) which depend on the parameter p. This
way we construct a family of (balanced) paths ψ(p, t) and then a family
of functions ϕτ (p, t) such that the respective family of sections Fτ satisfies
properties (a) and (b). In order to satisfy property (c) it is sufficient to set
Fτ : = Fβ(p)τ , where the function β : I l → [0, 1] is equal to 0 near ∂I l and
equal to 1 on a slightly smaller cube I1l ⊂ I l , so that F0 (p, t) = F (p, t) is a
genuine solution of Rp for all p ∈ I l \ I1l .
E. Derivatives with respect to the parameter. Let us now turn to

property (d), which is specific for the parametric case and which is crucial
for further generalizations. For F = (0, ϕ) this property means that the
derivatives ∂p f1 (t, p) are arbitrarily close to 0. Take the uniform product

ϕ1 (p, ∗) = ψ(p) • · · · • ψ(p) • ψ(p)
of N factors and set t
f1 (p, t) = ϕ1 (p, σ) dσ.
0
Then
t t
∂p f1 (p, t) = ∂p ϕ1 (p, σ) dσ = ∂p ϕ1 (p, σ) dσ,
0 0
where

∂p ϕ1 (p, ∗) = ∂p ψ(p) • · · · • ∂p ψ(p) • ∂p ψ(p).
25.6. Proof of the parametric version of the main lemma 315
The path ∂p ψ(p) is balanced because

1 1
∂p ψ(p, σ) = ∂p ψ(p, σ)dσ = ∂p 0 ≡ 0.
0 0
Hence, according to 25.4.4 we have
1
∂p f1 C 0 = max{∂p gC 0 , ∂p hC 0 },
N
where t t
g(p, t) = ψ(p, σ) dσ, and h(p, t) = σ) dσ.
ψ(p,
0 0
Therefore, ∂p f1 → 0 when N → ∞.
Remark. The same proof is also valid in the case when F consists of
genuine solutions over a neighborhood of a closed subset A ⊂ ∂D l (instead
of the whole ∂D l ). In this case the homotopy Fτ is constant for p ∈ A
(instead of p ∈ ∂D l ).
Remark. M. Theillière has given a description of one-dimensional con-
vex integration by explicit formulas [Th22]. Building on this description,
F. van Doorn, P. Massot, and O. Nash have formalized one-dimensional
convex integration and Smale’s sphere eversion using the Lean 3 theorem
prover [MDN22]. In [MT23], Massot and Thellière show that convex in-
tegration implies holonomic approximation of order 1.
Chapter 26
Homotopy Principle
for Ample Differential
Relations
26.1. Ampleness in coordinate directions

A coordinate principal subspace in a fiber Mq×n = (Rq )n of the fibration
J 1 (Rn , Rq ) = J 0 (Rn , Rq ) × Mq×n → J 0 (Rn , Rq ) = Rn × Rq
is any q-dimensional affine subspace parallel to one of the factors Rq in the

product (Rq )n or, which is the same, the set of all q × n matrices with fixed
(n − 1) columns. Thus for every point z ∈ J 1 (Rn , Rq ) there are n coordi-
nate principal subspaces P 1 (z), . . . , P n (z) which go through z. A particular
coordinate principal subspace P i (z) over a point (x, f (x)) ∈ J 0 (Rn , Rq ) can
be interpreted as the space of all possible vector-derivatives ∂xi f under the
condition that all the other vector-derivatives ∂xj f, j = i, are fixed.
Let us recall that a set Ω ⊂ P where P is an affine space is called ample

if the convex hull of each path-connected component of Ω is P or if Ω is
empty.
A differential relation R ⊂ J 1 (Rn , Rq ) is called ample in the coordinate

directions if R intersects all coordinate principal subspaces along ample
sets.
317
318 26. Homotopy Principle for Ample Differential Relations
Examples.
1. If n < q, then the immersion relation Rimm ⊂ J 1 (Rn , Rq ), which consists
of all matrices of rank n, is ample in the coordinate directions. Indeed, for
any z = (x, y, a) ∈ Rimm and any coordinate principal subspace P = P i (z)
we have P ∩ R = P \ L where L is an (n − 1)-dimensional linear subspace in
P Rq spanned by all the columns of the matrix a except the ith column.
The codimension of L in P is less than 1 and hence Conv(P ∩ R) = P .
2. If n ≥ q, then the submersion relation Rsubm ⊂ J 1 (Rn , Rq ), which con-
sists of all matrices of rank q, is not ample in the coordinate directions.
Indeed, for any z = (x, y, a) ∈ Rsubm and any coordinate principal subspace
P = P i (z) we have P ∩ R = P \ L where L is a (q − 1)-dimensional linear
subspace in P Rq spanned by all the columns of the matrix a except the
ith column. Therefore P ∩ R consists of two open half-spaces, and thus is
not ample.
Exercise. Prove that the differential relation Rk-mers ⊂ J 1 (Rn , Rq ) which

consists of all matrices of the rank ≥ k is ample if k < q.
A singularity
Σ ⊂ J 1 (Rn , Rq ) = J 0 (Rn , Rq ) × Mq×n
is called thin in the coordinate directions if it intersects all the coordinate
principal subspaces along stratified subsets of codimension ≥ 2. In this case
the complement R = J 1 (Rn , Rq ) \ Σ is a differential relation ample in the
coordinate directions.
26.2. Iterated convex integration

26.2.1. (Convex integration over a cube) Let R ⊂ J 1 (Rn , Rq ) be an
open differential relation ample in the coordinate directions and
F = (f, ϕ) : I n → R ⊂ J 0 (Rn , Rq ) × Mq×n
a formal solution of R which is a genuine solution near ∂I n . Then there
exists a homotopy of formal solutions
Fτ = (fτ , ϕτ ) : I n → R, τ ∈ [0, 1],
which joins F0 = F with a genuine solution F1 of R such that for all τ
• fτ is (arbitrarily) C 0 -close to f ;
• Fτ coincides with F near ∂I n .
26.2. Iterated convex integration 319
Proof. Let (ϕ1 , . . . , ϕn ) be the columns of the matrix ϕ. We will integrate

the formal solution F = (f, ϕ) coordinate-wise, using Lemma 25.5.1.
At the first step we consider the cube I n as a family of intervals I × p, p ∈

I n−1 , parallel to the x1 -axis. Let us form a relation R1 ⊂ I n−1 × J 1 (R, Rq ),
fibered over I n−1 , which is defined over a small neighborhood of the graph
of the section f in I n × Rq in the following way. For t = x1 , p = (x2 , . . . , xn )
we define the set
Ωp (f (t, p)) = R1p ∩ f (t, p) × Rq ⊂ p × J 1 (R, Rq )
as the path-connected component of R∩P 1 (F (t, p)) which contains the point
F (t, p). Here we use the canonical identification P 1 (F (t, p)) Rq . In order
to expand Rp to a small neighborhood of the graph of f , one can slightly
decrease the (open!) sets Ωp (f (t, p)) in such a way that the new sets are still
ample and for a sufficiently small ε the product
Dεq (f (x)) × Ωp (f (x)) ⊂ J 1 (R, Rq )
is contained in R. Here Dεq (f (x)) denotes the ε-ball in Rq centered at f (x).
Now we can apply Lemma 25.5.1 to the fibered relation R1 and its fibered
formal solution (f (t, p), ϕ1 ), which is automatically short because the rela-
tion R is ample. As a result we get a genuine solution (f 1 (t, p), ∂t f 1 (t, p))
of the relation R1 and hence the new formal solution
F 1 = (f 1 ; ∂x1 f 1 , ϕ2 , . . . , ϕn )
of the relation R. This formal solution is homotopic to F in R, coincides

with f near ∂I n , while the section f 1 is C 0 -close to f . But what is most
important is that the section F 1 is holonomic with respect to the coordinate
x1 .
At the second step we consider the cube I n as a family of intervals parallel

to the axis x2 , form a relation R2 , fibered over I n−1 , and construct a new
formal solution
F 2 = (f 2 ; ∂x1 f 1 , ∂x2 f 2 , ϕ3 , . . . , ϕn )
of R. This formal solution is holonomic with respect to the coordinate
x2 , i.e., ϕ2 = ∂x2 f 2 . According to property 25.5.1´(d), the section ∂x1 f 2 is
(arbitrarily) C 0 -close to the section ∂x1 f 1 , and hence we can deform the
formal solution F 2 by a linear homotopy in R into a formal solution
F2 = (f 2 ; ∂x1 f 2 , ∂x2 f 2 , ϕ3 , . . . , ϕn )
which is holonomic with respect to both coordinates, x1 and x2 .

Thus using the parametric version of one-dimensional convex integration, we

can realize the following chain of homotopies (each arrow denotes a homo-
topy; f0 = f ):
(f0 ; ϕ1 , ϕ2 , . . . , ϕn ) → (f 1 ; ∂x1 f 1 , ϕ2 , . . . , ϕn ) → · · ·
→ (f i ; ∂x1 f i , . . . , ∂xi f i , ϕi+1 , ϕi+2 , . . . , ϕn )
→ (f i+1 ; ∂x1 f i+1 , . . . , ∂xi+1 f i+1 , ϕi+2 , . . . , ϕn ) → · · ·
→ (f n ; ∂x1 f n , . . . , ∂xn f n ).
Property 25.5.1(d) is crucial here: it allows us to realize each homotopy in
the chain as a homotopy in R.
Remark. The same proof is valid in the parametric case (for families of
sections over the cube I n ) and also in the case when F is a genuine solution
near a neighborhood of a closed subset A ⊂ ∂I n (instead of the whole ∂I n ).
In this case the homotopy Fτ can be chosen constant near A (instead of
∂I n ).
26.2.2. (Corollary: h-principle for ample differential relations over

a cube) Let R ⊂ J 1 (Rn , Rq ) be an open differential relation over the cube
I n ample in the coordinate directions. Then all forms of the relative h-
principle hold for R over the pair (I n , ∂I n ) and also over the pair (I n , A),
where A ⊂ ∂I n is any closed subset.
26.3. Principal subspaces and ample differential relations in

X (1)
Let p : X → V be a fibration. Let us recall that the fiber
Ex = (p10 )−1 (x), x ∈ X,
of the projection
p10 : X (1) → X (0) = X
can be identified with Hom(Tv V, Vertx ), where v = p(x) and Vertx is the
tangent space to the fiber of the fibration p : X → V at the point x ∈ X.
Given a hyperplane τ ⊂ Tv V and a linear map l : τ → Vertx , let us denote
by Pτl an affine subspace of Ex defined as
Pτl = {L ∈ Hom(Tv V, Vertx ) | L|τ = l} .
Affine subspaces of Ex of this type are called principal. Note that the
direction of the principal subspace Pτl is determined by the hyperplane τ ∈
Tv V , and thus the principal directions at a given fiber Ex are parametrized
by the projective space P (Tv∗ V ) ∼
= RP n−1 , where n = dim V .
26.4. Convex integration of ample differential relations 321
Alternatively the 1-jet space X (1) can be considered as the space of all
nonvertical n-planes ξ in T X. In this interpretation principal subspaces are
nonvertical n-planes which contain a fixed nonvertical (n − 1)-plane in Tx X.
If X = Rn × Rq → Rn = V is a trivial fibration, so that we have X (1) =

J 1 (Rn , Rq ), then our previously defined coordinate principal subspaces are
the principal subspaces directed by hyperplanes {xi = const} in Rn .
Any principal subspace in X (1) has a natural affine structure, but no natural
linear structure, even in the case of a trivial fibration X = V × W → V .
A differential relation R ⊂ J 1 (Rn , Rq ) is called ample if R intersects all

principal subspaces along ample sets.
Remark. Ampleness in coordinate directions is a less restrictive prop-

erty than ampleness, although for Diff V -invariant relations the two notions
of ampleness coincide. Differential relations that are ample in coordinate
directions but not ample come up in, e.g., the proof of the h-principle for
hyperbolic rank 4 distributions in dimension 6 by J. Martı́nez-Aguinaga and
Á. del Pino [MP21].
Example. The immersion relation Rimm ⊂ X (1) is ample if n < q. The

submersion relation Rsubm ⊂ X (1) is not ample. The k-mersion relation
Rk-mers ⊂ X (1) is ample if k < q.
A singularity Σ ⊂ X (1) is called thin if for any a ∈ Σ and any principal

subspace P through the point a the intersection P ∩ Σ is a manifold or,
more generally, a stratified subset of codimension ≥ 2 in P . If Σ is thin,
then the complementary differential relation R = X (1) \ Σ is ample.
26.4. Convex integration of ample differential relations

26.4.1. (Homotopy principle for ample differential relations) Let
R ⊂ X (1) be an open ample differential relation. Then all forms of the
h-principle hold for R.
Proof of Theorem 26.4.1. The induction over skeleta of a triangulation

of the base V reduces the h-principle 26.4.1 to the relative h-principle 26.2.2.

26.4.2. (Corollary: removal of a thin singularity) Let Σ ⊂ X (1) be

a thin singularity. Then all forms of the h-principle hold for Σ-nonsingular
sections of X (1) .
In particular, the k-mersion relation Rk-mers is ample if k < q (the respective

singularity Σ = X (1) \ Rk-mers is thin) and hence Theorem 26.4.1 implies the
h-principle for k-mersions V → W , k < dim W . The case k = n gives us
the h-principle for immersions V → W , dim V < dim W . Note that the h-
principle for submersions (of open manifolds) does not follow from Theorem
26.4.1 because the relation Rsubm is not ample.
Remarks.
1. We will discuss further applications of the convex integration method in
Chapters 27 and 28. In all these examples we will verify the ampleness of
the respective differential relations. The ampleness will then imply all forms
of the h-principle (relative, parametric, and C 0 -dense).
2. Suppose that the manifold V is covered by coordinate charts Ui , i =

1, . . . , N . We can relax the ampleness condition for the differential relations
R ⊂ X (1) in Theorem 26.4.1 by requiring instead that over each neighbor-
hood Ui , i = 1, . . . , N , the relation R is ample only in respective coordinate
directions.

Chapter 27
Directed Immersions
and Embeddings
27.1. Criterion of ampleness for directed immersions

Let us recall the definition of directed immersions from Section 4.5.
Let Grn W be the Grassmannian bundle of tangent n-planes to a manifold

W of dimension q > n. Let V be an n-dimensional manifold. Given a
monomorphism F : T V → T W , we denote by GF the corresponding map
V → Grn W . Let A ⊂ Grn W be an arbitrary subset. An immersion f :
V → W is called A-directed if Gdf sends V into A. If V is an oriented
manifold, then we can also consider A-directed immersions where A is an
n W of oriented tangent n-planes
arbitrary subset in the Grassmannian Gr
to a q-dimensional manifold W .
Given a subset A ⊂ Grn W , we will denote by RA the differential relation in

Rimm ⊂ J 1 (V, W ) which corresponds to A-directed immersions V → W , by
Aw the fiber A ∩ Grn (Tw W ), w ∈ W , and by Grn−1 Aw the set

Grn−1 (L) ⊂ Grn−1 Tw W.
L∈Aw
27.1.1. (Ampleness criterion) The relation RA is ample if and only if

for every w ∈ W and every S ∈ Grn−1 Aw the set
ΩS = {v ∈ Tw W | Span{S, v} ∈ Aw } ⊂ Tw W
is ample.
323
324 27. Directed Immersions and Embeddings
Proof. Let us check that the above condition implies the ampleness of RA .
Note that any principal subspace is a coordinate principal subspace for a
certain local coordinate system, and by choosing a local coordinate system,
we can work in J 1 (Rn , Rq ). For s = (x, y, a) ∈ RA let P = P i (a) be a co-
ordinate principal subspace over (x, y). The subspace P can be canonically
identified with Ty Rq . The intersection P ∩RA consists of all vectors v ∈ Ty Rq
such that Span{S, v} ∈ A, where S ⊂ Ty Rq is the (n − 1)-dimensional linear
subspace spanned by all the columns of the matrix a except the ith col-
umn. Thus this intersection is equal to ΩS , and hence ample. The opposite
implication follows from the Diff V -invariance of the relation RA .
Condition 27.1.1 can be reformulated in the following way:

27.1.2. The relation RA is ample if and only if for every w ∈ W and every
S ∈ Aw the set
ΩS = {v ∈ S ⊥ | Span{S, v} ∈ A} ⊂ S ⊥ ,
where S ⊥ is the orthogonal complement to S ⊂ Tw W , is ample.
For an oriented manifold V we can consider

n−1 Aω =
Gr n−1 (L) ⊂ Gr
Gr n−1 Tw W
L∈Aw
and the oriented version of the above criterions.
Exercise. Suppose that V is oriented and W = Rn+1 . Prove that the
oriented version of Condition 27.1.2 means that for every
n W = Rn+1 × S n
a ∈ A ⊂ Gr
and every great circle S 1 ⊂ y ×S n through a the intersection S 1 ∩A contains
an arc of length > π.
27.2. Directed immersions into almost symplectic manifolds

Let us recall that an almost symplectic structure on a manifold W of di-
mension q = 2k is a nondegenerate but not necessarily closed 2-form ω.
One can define symplectic, Lagrangian, isotropic, and coisotropic immer-
sions V → (W, ω) as A-directed immersions where A ⊂ Grn W is the re-
spective (symplectic, Lagrangian, etc.) Grassmannian of n-planes tangent
to W . We denote the corresponding differential relations in J 1 (V, W ) by
Rsymp , RLag , Risot and Rcoisot .
The relation Rsymp is open, while RLag , Risot and Rcoisot are closed. We
can take open neighborhoods of these relations considering for any positive
ε, ε < π/2, ε-Lagrangian, ε-isotropic, and ε-coisotropic immersions V →
(W, ω) as Aε -directed immersions, where Aε is the ε-neighborhood of the
27.3. Directed immersions into almost complex manifolds 325
respective set A in Grn W . We assume here that W is endowed with a

Riemannian metric. The respective differential relations in J 1 (V, W ) will be
denoted by RεLag , Rεisot , and Rεcoisot .
Exercise. Prove that the relations RεLag and Rεisot are ample and hence
all forms of the h-principle hold for ε-Lagrangian and ε-isotropic immersions
V → (W, ω).
Remark. The proof of the h-principle for isotropic and in particular
Lagrangian immersions into symplectic manifolds which we gave in Part III
of the book fails when we try to generalize it for immersions into almost
symplectic manifolds. On the other hand, the ampleness which is needed
for the application of convex integration does not depend a priori on the
closeness of the form ω and hence convex integration equally works for a
symplectic or almost symplectic target manifold W .
Exercise. Show that the relations Rsymp , Rεcoisot , and Rεisosymp are not
ample.
Problems. Is there any form of the h-principle for:

(a) Lagrangian and isotropic immersions into an almost symplectic
manifold?
(b) isosymplectic immersions between almost symplectic manifolds?
(c) coisotropic and isometric coisotropic immersions into almost sym-
plectic manifolds?
Comments to the Problems. We do not know the answer to most of

these questions. However, it seems to us that the positive answer to (a)
should not be difficult to prove. The answer to (b) is obviously negative
in the most general set-up, even locally in a neighborhood of a point. On
the other hand, when the source manifold is two dimensional, then Theorem
24.4.3 remains true even when the target manifold is only almost symplectic.
The problem in (b) is to find the conditions under which some kind of the
h-principle may hold. Theorem 24.5.1 implies the h-principle for isometric
coisotropic immersions into symplectic manifolds. We do not know whether
it remains true when the structure on the target manifold is not integrable.
27.3. Directed immersions into almost complex manifolds

Let us recall that a subspace S ⊂ Cn is called:
• complex, if iS = S;
• real or totally real, if S∩iS = 0; in other words, a subspace is totally
real if it contains no complex subspaces of positive dimension;
• co-real, if S + iS = L.
Let (W, J) be an almost complex manifold of (real) dimension q = 2k. One

can define complex and real immersions V → (W, J) as A-directed im-
mersions where A ⊂ Grn W are the respective Grassmannians Acomp , Areal ,
and Aco-real of complex, real, and co-real n-planes in T W . Denote the corre-
sponding differential relations in J 1 (V, W ) by Rcomp , Rreal , and Rco-real . For
any positive ε, ε < π/2, we also define ε-complex immersions V → (W, ω)
as Aεcomp -directed immersions where Aεcomp is the ε-neighborhood of Acomp
in Grn W . The corresponding differential relation in J 1 (V, W ) is denoted by
Rεcomp .
Exercise. Prove that the differential relation Rεcomp is not ample.
27.3.1. (Gromov [Gr86]) The relations Rreal and Rco-real are ample and
hence all forms of the h-principle hold for real and co-real immersions V →
(W, J).
Proof. Let n ≤ k. For a particular (n − 1)-dimensional subspace S ⊂ L ∈

Areal over a point w ∈ W , we have
ΩS = Tw W \ (S + iS),
where dim(S + iS) = 2n − 2 ≤ 2k − 2, and hence Rreal is the complement
of a thin singularity.
For n > k the set ΩS is the complement of S if S + iS = Tw W or is a
complement of S + iS if S + iS = Tw W . In both cases the codimension
of the singularity is ≥ 2 and hence Rco-real is the complement of a thin
singularity as well.
27.4. Directed embeddings

A differential relation RA ⊂ J 1 (V, W ) is called affine ample, if for any
S ∈ Grn−1 Aw , any hyperplane H ⊃ S and any affine hyperplane H ⊂ Tw W ,
parallel to H, the set ΩS ∩ H is ample in H .
Exercise.
1. Find an example of the ample differential relation RA which is not affine
ample.
2. Prove that if RA ⊂ J 1 (V, W ) is a complement of a thin singularity, then

RA is affine ample. In particular, the relation Rreal and Rco-real are affine
ample.

27.4. Directed embeddings 327
27.4.1. (Directed embeddings) Suppose that A ⊂ Grn W is an open

subset and the corresponding (open) differential relation RA ⊂ J 1 (V, W ) is
affine ample. Then every embedding f0 : V → W whose tangential lift
is homotopic over V to a map G1 : V → A can be isotoped to an A-directed
embedding f1 : V → W . Moreover, such an isotopy ft : V → W can be
chosen arbitrarily C 0 -close to the constant isotopy.
Here homotopy over V means that the underlying homotopy gt : V → W

for Gt is constant (i.e., Gt is a tangential homotopy, as in Section 4.3). In
fact, the theorem is also true, with an obvious modification of the proof, in
the case when Gt is a homotopy over embeddings gt : V → W . In this case
ft can be chosen arbitrarily C 0 -close to gt . We restrict ourselves to the case
gt = f0 only to clarify the main idea of the proof.
Theorem 27.4.1 also holds (with the same proof) in the relative and para-
metric versions. Here is, for example, the parametric version of Theorem
27.4.1.
27.4.2. (Families of directed embeddings) Suppose that f p : V → W ,
p ∈ D l , is a family of embeddings which are A-directed for p ∈ ∂D l , and
Gpt : V → Grn W , t ∈ [0, 1], is a homotopy of tangent lifts, constant over
∂D l , such that Gp0 = Gdf p and Gp1 sends V to A for each p ∈ D l . Then
there exists a family of isotopies ftp : V → W , constant over ∂D l , such that
f1p is an A-directed embedding for all p ∈ D l .
Proof of Theorem 27.4.1. Assuming that the manifold W is endowed

with a Riemannian metric, we can cover the homotopy Ḡt by a homotopy
of fiberwise isomorphisms Φt : T (W ) → T (W ), bs Φt = IdW . Then the
existence of the required isotopy ft follows from the following.
27.4.3. Let A ⊂ Grn W be an open subset such that the corresponding dif-
ferential relation RA ⊂ J 1 (V, W ) is affine ample. Let Φt : T W → T W ,
t ∈ [0, 1], be a homotopy of fiberwise isomorphisms such that bs Φt = IdW
for all t. Then for every A-directed embedding f0 : V → W there exists an
isotopy ft : V → W , t ∈ [0, 1], such that f1 is an A1 -directed embedding,
where A1 = Φ1∗ A. Moreover, such an isotopy ft can be chosen arbitrarily
C 0 -close to the constant isotopy.
Proof of Theorem 27.4.3. We begin with the following lemma which is

an immediate corollary of the Ampleness Criterion 27.1.1.
27.4.4. Let A ⊂ Grn W be an open subset such that the corresponding differ-
ential relation RA ⊂ J 1 (V, W ) is affine ample. Let V ⊂ W be an embedded
manifold and X a tubular neighborhood of V ⊂ W , fibered over V . Then
the differential relation RX

A ⊂ X
(1) , which defines the A-directed sections
of the fibration p : X → V , is open and ample.
Now we can proceed in the following way. Choose a sequence of maps

Φ(i) = Φti , 0 = t0 < t1 < · · · < tN = 1, such that the angle between
Φ(i) (L) and Φ(i+1) (L) is less than, say, π/4 for all n-planes L ∈ Grn W .
(i)
Set Ai = Φ∗ A. Consider a tubular neighborhood X of the submanifold
f0 (V ) ⊂ W and the differential relation RX A1 in X
(1) . This relation is open
and ample and hence we can apply convex integration to its formal solution
F = Φ(1) (df0 (V )). Let ft1 : V → X ⊂ W be the resulting embedding.
Now we can apply the same construction to a tubular neighborhood X of
ft1 (V ) ⊂ W , the differential relation RXA2 ⊂ X
(1) , and its formal solution
(2)
F = Φ (dft1 (V )). We can continue this way. The embedding ftN = f1 will
have the required properties. The approximation property follows from the
possibility to approximate at each step.
27.4.5. (Corollary: real embeddings, Gromov [Gr86]) Let (W, J) be

an almost complex manifold. Then the following hold.
(a) Every embedding f0 : V → (W, J) whose tangential lift
is homotopic over embeddings to a map
G1 : V → Areal ⊂ Grn W
(resp., G1 : V → Aco-real ⊂ Grn W )
can be isotoped to a real (resp., co-real ) embedding f1 : V → W .
(b) Let ft : V → (W, J), t ∈ [0, 1], be an isotopy which connects two
real (resp., co-real ) embeddings f0 and f1 . Suppose that there exists
a family of real (resp., co-real ) homomorphisms Ft : T V → T W
which covers the isotopy ft , t ∈ [0, 1] and such that the families
dft , Ft : T V → T W, t ∈ [0, 1],
are homotopic via families of monomorphisms fixed at t = 0, 1.
Then there exists an isotopy of real (resp., co-real ) embeddings ft :
V → W , t ∈ [0, 1] which connects f0 = f0 with f1 = f1 , is C 0 -close
to the isotopy ft , and such that the families
dft , Ft : T V → T W, t ∈ [0, 1],
are homotopic via families of monomorphisms fixed at t = 0, 1.
Remark. It is important to realize that just the existence of the family of
real monomorphisms Ft : T V → T W which covers the isotopy ft , t ∈ [0, 1],
is not sufficient for the existence of a real isotopy connecting f0 and f1 ; see
27.4. Directed embeddings 329
[Fd87] and [Pt88]. Thus the existence of homotopy of homotopies is crucial

and cannot be omitted. Algebro-topological consequences of this condition
were computed in several examples by Borrelli (see [Bo01]).
Chapter 28
First Order Linear

Differential Operators
All the examples below fit into the philosophy of Σ-nonsingular solutions;

see Section 6.2. The singularity Σ ⊂ X (1) in these examples will have the
form Σ = D −1 (S) where D is the symbol of a first order linear differential
operator D : Sec X → Sec Z and S ⊂ Z is a subset of the total space of the
vector bundle Z.
28.1. Formal inverse of a linear differential operator

Let X and Z be vector bundles over V . Note that in this case the fibration
X (1) → V has a natural linear structure. A first order linear differential
operator
D : Sec X → Sec Z
can be written as a composition
J1
D
Sec X → Sec X (1) → Sec Z,
is induced by a fiberwise homomorphism
where the map D
D
X (1) → Z
of vector bundles over V . The vector bundle homomorphism D = Symb D
is called the symbol of the operator D.
Suppose D = Symb D is a fiberwise epimorphism. Then D can be viewed as

an affine fibration D : X (1) → Z, and thus we have a homotopy equivalence
Sec X (1) Sec Z. In particular, any section s : V → Z can be lifted in a
homotopically canonical way to a section Fs : V → X (1) such that D◦Fs = s.
331
332 28. First Order Linear Differential Operators
It is useful to think of Fs as a formal inverse of s. Thus we can say that

the differential operator D with a surjective symbol is formally invertible. If
in addition the operator D is pure differential in the sense that D depends
only on derivatives of a section of X and does not depend on the values of
the section, i.e., the symbol D can be written as the composition
pr
X (1) → X (1) /X → Z,
then D is formally invertible over every fixed section γ : V → X, i.e., the
formal inverse Fs for s : V → Z can be chosen in a such way that bs Fs = γ.
Example (Formal primitive of a differential form (cf. Section 4.7). The
symbol D of the exterior differentiation
d : Sec Λp−1 V → Sec Λp V
is a fiberwise epimorphism
(Λp−1 V )(1) → Λp V.
Therefore, d is formally invertible. Moreover, d is pure differential and hence
it is formally invertible over any differential (p − 1)-form.
28.2. Homotopy principle for D-sections

Let D : Sec X → Sec Z be a differential operator. A section s : V → Z is
called D-section if s = Df for a section f : V → X. For example if D is the
exterior differentiation
d : Sec Λp−1 V → Sec Λp V,
then the D-sections are exact differential p-forms.
Given a subset S ⊂ Z, we will denote by SecD (Z \ S) the space of all
D-sections V → Z \ S.
Let us point out the following important but trivial h-principle.
28.2.1. (Homotopy principle for D-sections) Let D : Sec X → Sec Z
be a linear differential operator such that D = Symb D is a fiberwise epimor-
phism. Let S be a subset of Z. If the h-principle holds for Σ-nonsingular
sections V → X, where Σ = D (−1) (S) ⊂ X (1) , then it also holds for the
inclusion
SecD (Z \ S) → Sec(Z \ S),
i.e., for any section s0 ∈ Sec(Z \ S) there exists a homotopy
st : I → Sec(Z \ S)
such that s1 ∈ SecD (Z \ S). The same is also true for all forms of the
h-principle, except the C 0 -dense one (which is not defined for the inclusion
SecD (Z \ S) → Sec(Z \ S)).
28.3. Nonvanishing D-sections 333
If the operator D is pure differential, then the C 0 -dense h-principle for

Σ-nonsingular sections V → X implies that for any section f0 : V → X
one can choose the homotopy st in such a way that s1 = Df1 , where f1 is
arbitrarily C 0 -close to f0 (and similarly for all other forms of the C 0 -dense
h-principle).
The proof follows immediately from the homotopy equivalence

Sec(X (1) \ Σ) Sec(Z \ S).
The C 0 -dense version follows from existence of a formal inversion over any
fixed section f : V → X.
28.3. Nonvanishing D-sections

A section s : V → Z is called nonvanishing if s(v) = 0 for all v ∈ V .
We say that a linear differential operator D has principal rank ≥ 2 if

dim D(P ) ≥ 2 for each principal subspace P in X (1) . The following theorem
was first proved in [GE71] using the method of removal of singularities.
28.3.1. (Homotopy principle for nonvanishing D-sections) If the
linear differential operator D has principal rank ≥ 2 and D = Symb D is a
fiberwise epimorphism, then all the forms of the h-principle (excluding the
C 0 -dense one) hold for the inclusion
SecD (Z \ 0) → Sec(Z \ 0),
where 0 ⊂ Z is the zero-section. In particular, any nonvanishing section
s : V → Z can be deformed via a homotopy of nonvanishing sections to
a section Df . Moreover, if D is pure differential, then we can choose f
arbitrarily C 0 -close to any fixed section f0 : V → X.
Proof. According to Proposition 28.2.1, it is enough to prove the h-principle

for Σ-nonsingular sections V → X, where Σ = D −1 (0) = Ker D. The
inequality dim D(P ) ≥ 2 is equivalent to the inequality
codimP (P ∩ Ker D) ≥ 2,
and hence the singularity Σ is thin. Therefore we can apply Theorem 26.4.2.

28.3.2. (Corollary) Let n ≥ 3 and 2 ≤ p ≤ n − 1. Any nonvanishing

differential p-form on an n-dimensional manifold V can be deformed via a
homotopy of nonvanishing forms to a nonvanishing exact form.
Proof. Let us check that the inequalities n ≥ 3 and 2 ≤ p ≤ n − 1 im-

ply that the principal rank of the exterior differentiation d is ≥ 2. Let
X = Λp−1 V , Z = Λp V and D be the symbol of d. Let P be a coordinate
principal subspace which corresponds, say to the first coordinate x1 of a
local coordinate system x1 , . . . , xn on V . The dimension of P is equal to
Cnp−1 = (p−1)!(n−p+1)!
n!
. If the intersection (Ker D) ∩ P is not empty, then in
p−1
local coordinates it is defined by the system of Cn−1 equations
a1i1 ... ip−1 = const, 1 ∈
/ {i1 . . . ip−1 },
where the coordinates {a1i1 ... ip−1 } correspond to the derivatives ∂/∂x1 of the
coefficients of (p − 1)-forms. Hence for 2 ≤ p ≤ n − 1 and n ≥ 3, we have
p−1
rank D = Cn−1 ≥ 2,
and therefore Theorem 28.3.1 applies.
28.3.3. (Corollary) Let n ≥ 3 and V be endowed with a volume form Ω.

Then any nonvanishing vector field L on V is homotopic through nonvan-
ishing vector fields to a divergence free vector field.
Proof. By Cartan’s formula we have

LL Ω = d(L Ω) + L dΩ = d(L Ω).
Therefore the flow of the field L preserves Ω if and only if the (n − 1)-form
L Ω is closed. The correspondence v → ωv = v Ω is a fiberwise isomor-
phism T V → Λn−1 V . Thus Corollary 28.3.2 implies that ωL is homotopic
through nonvanishing forms to an exact form ω = ωL1 , and then L1 will be
a divergence free vector field.
Exercise. Let n ≥ 3, 2 ≤ p ≤ n − 1 and a ∈ H p (V ). Prove that
any nonvanishing differential p-form on an n-dimensional manifold V can
be deformed via a homotopy of nonvanishing forms to a closed form which
represents the class a. Hint: Consider the singularity Σ = D −1 (−ωa ), where
ωa is a closed p-form which represents a.
28.4. Systems of linearly independent D-sections

In all applications of convex integration which we considered so far, the
relation R was a complement of a thin singularity. In this chapter we will
consider an example when the singularity
Σ = D −1 (S) ⊂ X (1)
is not thin.
The sections of a vector bundle are called linearly independent if they are
pointwise linearly independent. In [Gr86] Gromov proved the following.
28.4. Systems of linearly independent D-sections 335
28.4.1. (Homotopy principle for systems of linearly independent

D-sections) Let D : Sec X → Sec Z be a differential operator of principal
rank ≥ 2 such that its symbol D = Symb D is a fiberwise epimorphism. Then
any system {si } = {s1 , . . . , sk } of linearly independent sections of the vector
bundle Z can be deformed via a homotopy of systems of linearly independent
sections to a system of sections {Dfi }.
Proof. It is sufficient to consider the case when Z is a trivial bundle and

{si } is a trivialization of Z. Write
X = X ⊕ ·
· · ⊕ X , Z = Z ⊕ ·
· · ⊕ Z , and D = D ⊕ ·
··⊕ D.
q q q
−1 (1)
Let Σ = D (S) ⊂ X , where D is the symbol of D, and S ⊂ Z is given
in the fibers of the fibration Z → V by the equation z1 ∧ · · · ∧ zq = 0. Then
the singularity Σ ⊂ X (1) is defined in the fibers Lv = Lv ⊕ · · · ⊕ Lv of the
(1)
fibration X → V by the equation
D(y 1 ) ∧ · · · ∧ D(y q ) = 0,
where y 1 , . . . , y q ∈ Lv .
According to Proposition 28.2.1, we just need to check that the differential

(1) (1)
relation R = X \ Σ is ample. A principal subspace P in X over a
point v ∈ V is the Cartesian product of q principal subspaces P1 , . . . , Pq in
X (1) over v. These spaces are parallel affine subspaces in the fiber Lv of the
bundle X (1) → V . Therefore the images
P1 = D(P1 ), . . . , Pq = D(Pq )
are parallel affine equidimensional subspaces in the fiber Zv of the bundle
Z. The operator D has principal rank ≥ 2 and hence
r = dim P1 = · · · = dim Pq ≥ 2.
To simplify the notation, we will further assume that r = 2; the case r > 2
can be considered in a similar way.
Choose a basis w1 , . . . wq in Zv such that w1 and w2 are parallel to Pi . For

each i = 1, . . . , q set
ai = Pi ∩ Span{w3 , . . . , wq }.
The coordinates of ai in w1 , . . . , wq are (0, 0, ai3 , . . . , aiq ).
For each i = 1, . . . , q fix an origin 0 in the affine subspace Pi such that

0 ∈ D −1 (ai ), and choose a basis v1i , . . . , vqi in (Pi , 0) such that
D(v1i ) = ai + w1 , D(v2i ) = ai + w2 ,
and the vectors v3i , . . . , vqi belong to the kernel of the map
D|Pi : (Pi , 0) → (Pi , ai ).
Let {αji } and {βj } be the coordinates in P and Zv which correspond to
the basis {vji } in P and to the basis {wj } in Zv . In these coordinates the
nonempty intersection R ∩ P is given by the nonequality
( 1 (
( α1 α12 . . . . . . α1q (
( 1 (
( α2 α22 . . . . . . α2q (
( 1 (
det A = (( a3 a23 . . . . . . aq3 (( = 0,
(. . . . . . . . . . . . . . .(
( (
( a1 a2 . . . . . . aqq (
q q
where aij are constants, i.e., they do not depend on the coordinates {αlk }.
Therefore,
R ∩ P = (R2q \ {det A = 0}) × Rq(q−2) .
The complement R2q \ {det A = 0} consists of two (nonempty) path con-
nected components: a positive one where det A > 0, and a negative one
where det A < 0. We need to prove that the convex hull of each of these
components coincides with R2q . Let {Aij }i=1,...,q
j=1,2 be the matrices which cor-
respond to the standard basis in R . For every matrix A0 there exists a
2q
constant a such that A0 belongs to the interior of the convex hull of 4q

matrices {±a · Aij }. Each of the matrices ±a · Aij belongs to the quadric
{det A = 0}. These matrices can be slightly moved into, say the positive
component so they would still contain A0 in the interior of their convex
hull.
28.4.2. (Corollary: systems of exact forms) Let {ωi }i=1,...,q be a sys-

tem of linearly independent differential p-forms on V . If 2 ≤ p ≤ n − 1,
where n = dim V , then {ωi } can be deformed via a homotopy of systems of
linearly independent forms to a system of exact linearly independent forms.
28.4.3. (Corollary: systems of divergence free vector fields) Let

n = dim V ≥ 3 and V be endowed with a volume form Ω. Any system of
linearly independent vector fields on V can be deformed via a homotopy of
systems of linearly independent vector fields to a system of divergence free
vector fields. In particular, every parallelizable manifold supports n = dim V
linearly independent divergence free vector fields.
28.5. Two-forms of maximal rank on odd-dimensional

manifolds
As we have already seen in Section 18.4, Gromov’s h-principle for symplectic
forms on open manifolds implies McDuff’s h-principle for closed 2-forms
28.5. Two-forms of maximal rank on odd-dimensional manifolds 337
of maximal rank on closed odd-dimensional manifolds. McDuff proved in

[MD87a] this h-principle using the convex integration technique by showing
that the corresponding differential relation is ample. We reproduce her
argument in this chapter.
Let V be a manifold of dimension n = 2m + 1. For a fixed 2-form ω0 on V

we define Sω0 ⊂ Λ2 V by the equation
(z + ω0 (v))m = 0
for each v ∈ V . Let D be the symbol of the exterior differentiation
d : Sec Λ1 V → Sec Λ2 V.
28.5.1. (McDuff [MD87a]) For any differential 2-form ω0 on V the dif-

ferential relation R = (Λ1 V )(1) \ Σ, where Σ = D −1 (Sω0 ), is ample.
According to Proposition 28.2.1, the ampleness of R implies that all forms

of the h-principle (excluding the C 0 -dense one) hold for the inclusion
Secd (Λ2 V \ Sω0 ) → Sec(Λ2 V \ Sω0 ),
where Secd (Λ2 V \ Sω0 ) is the space of exact sections V → Λ2 V \ Sω0 .
This is equivalent to the h-principle 18.4.1. In particular, if V supports a

2-form of maximal rank, then every two-dimensional cohomology class of V
can be represented by a (closed) nondegenerate form.
Proof of Theorem 28.5.1. To simplify notation, we assume that ω0 =

0. The proof can be easily rewritten for any ω0 . In local coordinates the
singularity Σ is defined by the equation
Ω = Σαi li = [Σi<j (yij − yji )dxi ∧ dxj ]m = 0,
where the coordinates {yij } correspond to the derivatives ∂aj /∂xi of the
coefficients of the 1-forms Σaj dxj and li is the exterior product of all basic
1-forms dxj , j = 1, . . . , n, excluding dxi . In a coordinate principal subspace
P which corresponds, say, to the first coordinate in V , only the y1j are
variables ; all the other yij = yij 0 are constants. In particular, the coefficient
α1 is constant. If α1 = 0, then Σ ∩ P is empty. Otherwise Σ ∩ P is defined

by the system of (n − 1) linear equations
{αi = 0}i=2,...,n ,
which can be written as AZ = 0, where A is a constant matrix, Z =
(z2 , . . . , zn ) and zj = y1j − yj1
0 . The condition Σ = P implies A = 0.
It is easy to check that aii = 0 and aij = ±aji . Hence, the rank of A is at
least 2 and thus the singularity Σ is thin. Therefore R is ample.
28.6. One-forms of maximal rank on even-dimensional

manifolds
Let V = V 2m be an even-dimensional manifold. As we already have shown

in Chapter 22, McDuff’s h-principle 22.2.3 for maximally nonintegrable hy-
perplane distributions can be deduced from two of Gromov’s h-principles:
Theorem 18.3.2 for contact structures on open manifolds, and Theorem
22.2.1 for mappings of closed manifolds transverse to a contact structure. In
this section we reconstruct McDuff’s original argument based on the convex
integration technique.
Let us recall that a 1-form α on V is called maximally nondegenerate if the

differential (2m − 1)-form α ∧ (dα)m−1 never vanishes. A pair of differential
forms (α, β) on V , where α is a 1-form and β is a 2-form, is called maximally
nondegenerate if the differential (2m − 1)-form α ∧ β m−1 never vanishes. A
pair (α, β) is called exact if β = dα.
Let the linear differential operator

D = (id, d) : Sec Λ1 V → Sec(Λ1 V ⊕ Λ2 V )
be defined by the formula α → (α, dα). Then the D-sections
V → Λ1 V ⊕ Λ2 V \ Sω0
are exact pairs (α, dα). Note that the symbol
D : (Λ1 V )(1) → Λ1 V ⊕ Λ2 V
of the operator D = (id, d) is fiberwise epimorphic.
For an even-dimensional manifold V = V 2m , let a subset S ⊂ Λ1 V ⊕ Λ2 V

be defined in the fibers of the fibration Λ1 V ⊕ Λ2 V → V by the equation
z1 ∧ (z2 )m−1 = 0.
28.6.1. (McDuff [MD87a]) The differential relation R = (Λ1 V )(1) \ Σ,

where Σ = D −1 (S), is ample.
According to Proposition 28.2.1, the ampleness of R implies that all forms

of the h-principle (excluding the C 0 -dense one) hold for the inclusion
SecD (Λ1 V ⊕ Λ2 V \ S) → Sec(Λ1 V ⊕ Λ2 V \ S),
where SecD (Λ1 V ⊕ Λ2 V \ S) is the space of D-sections
V → Λ1 V ⊕ Λ2 V \ Sω0 .
28.6. One-forms of maximal rank on even-dimensional manifolds 339
This is equivalent to the h-principle 22.2.3, and in particular, every non-

degenerate pair of forms (α, β) on V can be deformed via a homotopy of
nondegenerate pairs to an exact nondegenerate pair ( α). The C 0 -dense
α, d
h-principle holds in the following version: one can choose α to be arbitrarily
0 0
C -close to α (and similarly for all other forms of the C -dense h-principle).
Proof of Theorem 28.6.1. As in the proof of Theorem 28.5.1 the key ob-
servation is that the matrix of a system of linear equations which defines
the intersection of a principal subspace with Σ is almost skew-symmetric:
aij = ±aji and aii = 0. Hence its rank cannot be equal to 1, and thus the
corresponding singularity Σ is thin. To clarify the computation, we consider
only the case m = 2. The general case can be treated in a similar way (see
[MD87a]).
Let the coordinates {yij } correspond to the derivatives ∂aj /∂xi of the co-
efficients of the 1-forms Σaj dxj and zij = yij − yji , i < j. For m = 2 the
singularity Σ is defined in local coordinates by the equation ω1 ∧ ω2 = 0,
where
ω1 = a1 dx1 + a2 dx2 + a3 dx3 + a4 dx4 ,
ω2 = z12 dx1 ∧ dx2 + z13 dx1 ∧ dx3 + z14 dx1 ∧ dx4
+ z23 dx2 ∧ dx3 + z24 dx2 ∧ dx4 + z34 dx3 ∧ dx4 .
This equation is equivalent to the system of equations
a4 z23 − a3 z24 + a2 z34 = 0
0 + a4 z13 − a3 z14 = −a1 z34
a4 z12 + 0 − a2 z14 = −a1 z24
a3 z12 − a2 z13 + 0 = −a1 z23 .
In a principal subspace P which corresponds, say, to the first coordinate in
V , only the z1j are variables ; all the other zij , and also ai , are constants. In
particular, the first equation does not contain unknowns and hence Σ ∩ P =
∅ if this equation is not an identity. If Σ = P , then the system of the last
three equations, which contain three unknowns z12 , z13 , z14 , has rank ≥ 2
and hence the singularity Σ is thin.
Chapter 29
Nash–Kuiper Theorem
29.1. Isometric immersions and short immersions

Recall that a C r -smooth family g = {gx , x ∈ V } of positive quadratic forms
on Tx V, x ∈ V , is called a Riemannian C r -metric on V , r = 0, 1, . . . . The
pair (V, g) is then called a Riemannian C r -manifold. In what follows the
class of the Riemannian metric is not essential, and we will write Riemannian
manifold instead of Riemannian C r -manifold and so on.
We will consider also, as a technical tool, families of nonnegative quadratic

forms on V . Such a family will be called a semi-Riemannian metric on V .
Let (V n , g) and (W q , h) be Riemannian manifolds. A C 1 -smooth map f :

V → W is called isometric if f ∗ h = g, i.e., dx f : Tx V → fx (Tx V ) ⊂ Tf (x) W
is a linear isometry for every x ∈ V . Any isometric map is automatically
an immersion. Locally with respect to a frame of independent vector fields
{∂i }, i = 1, . . . , n, the isometry condition can be described by the system of
equations
f∗ ∂i , f∗ ∂j h = ∂i , ∂j g , 1 ≤ i ≤ j ≤ n,
where f∗ ∂i = df (∂i ) = ∂i f . Note that this system is overdetermined when
q < n(n+1)
2 .
A C 1 -map f : V → W is called strictly short if
f ∗ h < g, i.e., f∗ vh < vg ,
for all tangent vectors v ∈ T V . A C 1 -map f : V → W is called short if

f ∗ h ≤ g. A (strictly) short map is not necessarily an immersion.
341
342 29. Nash–Kuiper Theorem
Example. Given an arbitrary C 1 -map f : (V, g) → Rq , the composition

Ha ◦ f : (V, g) → Rq , where Ha (x) = ax is a homothety centered at the
origin, is strictly short for all sufficiently small a > 0.
29.2. Nash–Kuiper theorem

It is well known from classical differential geometry that for r > 1 the C r -
smooth isometric immersions of two-dimensional Riemannian C ∞ -manifolds
into R3 are very specific and rigid maps. For example, any isometric C 2 -
immersion of the standard sphere S 2 ⊂ R3 into R3 is congruent to the
standard embedding S 2 → R3 . Until the mid-1950s mathematicians mostly
believed that C 1 -smooth isometric immersions V n → W q are also rigid and
hard to construct, and, in particular, that the aforementioned uniqueness
survives also for isometric immersions S 2 → R3 which are only C 1 -smooth.
It was discovered by J. Nash in 1954 that the situation is, in fact, drastically
different when one passes to C 1 -smooth immersions. In contrast to C 2 -
immersions they appeared to be extremely flexible:
29.2.1. (Nash–Kuiper) If n < q, then any strictly short immersion
f : (V n , g) → (Rq , h),
where h is the standard metric on Rq , can be C 0 -approximated by isometric
C 1 -smooth immersions. Moreover, if the initial immersion f is an embed-
ding, then f can be approximated by isometric C 1 -embeddings.
For example, there exists a C 1 -isometric embedding of the standard sphere

S 2 into an arbitrarily small ball in R3 .
In [Na54] Nash proved this theorem for n ≤ q − 2, and later Kuiper in

[Ku55] extended the theorem to the case n = q − 1. The parametric version
of the theorem is also true and implies (together with the Example in Section
29.1) the following.
29.2.2. Isometric C 1 -immersions V n → Rq , n < q, satisfy the parametric
h-principle for all Riemannian manifolds V = (V, g).
Remark. The C 0 -dense h-principle will also hold if the shortness condi-
tion is incorporated in the definition of a formal isometric immersion.
Remark. There exist striking similarities between the Nash–Kuiper theo-
rem and the abundance of weak solutions to the Euler equations of fluid
dynamics discovered by Shnirelman, De Lellis, Székelyhidi, and others. See
[DS12] for a survey of this beautiful subject, which unfortunately lies be-
yond the scope of this book.
29.3. Decomposition of a metric into a sum of primitive metrics 343
The rest of this chapter is devoted to the proof of Theorem 29.2.1. We will
consider only the case when V is compact and f is an embedding. The case
of immersions follows formally from the case of embeddings. The proof can
be easily adjusted for noncompact manifolds and also for the parametric
case. Moreover, it can be generalized to a general target manifold (W, h)
without employing any additional ideas.
29.3. Decomposition of a metric into a sum of primitive

metrics
A quadratic form Q is called primitive if Q = l2 where l is a linear form. A
semi-Riemannian metric g on Rn is called primitive if g = α(x)(dl)2 , where
l = l(x) is a linear function on Rn and α is a nonnegative function with
compact support.
A semi-Riemannian metric g on a manifold V is called primitive if there

exists a local parametrization u : Rn → U ⊂ V such that supp g ⊂ U and
u∗ g is a primitive metric on Rn .
29.3.1. (Lemma) Any Riemannian metric g on a compact manifold V can
be decomposed into a finite sum of primitive metrics.
Proof. Choose a set of local parametrizations {ui : Rn → Ui ⊂ V } and a

partition of unity {αi }, αi ≡ 1, on V such that supp αi ⊂ Ui .
Let Ai = supp(αi ◦ ui ) and gi = (u∗i g)|Ai . For every i we can find positive
quadratic forms
Qij , j = 1, . . . , N (i),
on Rn such that for every x ∈ Ai the positive quadratic form
gxi = gi |Tx Rn =Rn
belongs to the interior of the convex hull of the forms Qij , j = 1, . . . , N (i).
Every positive quadratic form Qij is a sum of some primitive forms (lijk )2 ,
k = 1, . . . , n. Therefore,

g= αi gijk where gijk = (u−1 ∗
i ) (dlijk )
2
ijk
is the desired decomposition.
Remark. For any positive function β on V we can decompose a metric g

which is sufficiently C 0 -close to β(x)g into a sum of primitive metrics using
the same set of forms gijk .
29.4. Approximation theorem
A. The functions r( g , g) and dg (f, f ). Given a pair of metrics g and g

on V , we will denote by r( g , g) the function
vg
T V \ V → R, v → r(
g , g)(v) = .
vg
g , g) is defined also in the case when g is a semi-Riemannian
The function r(
metric on V . Note that
g , g1 ) ≤ r(
(r1) r( g , g2 ) if g1 ≥ g2 , and
g , g) ≤ r(
(r2) r( g , g) ≤ 1.
g + g1 , g + g1 ) if r(
Given a pair of maps f, f : (V, g) → Rq , we will denote by dg (f, f) the

function
f∗ v − f∗ vh
T V \ V → R, v → dg (f, f)(v) = ,
vg
where h is the standard metric on Rq . Note that
(d1) dg1 (f, f) ≤ dg2 (f, f) if g1 ≥ g2 .
g , g) and dg (f, f ) do not depend on the lengths of v and in

The functions r(
what follows we will consider the restriction of these functions to the g-unit
tangent bundle T1 V .
We will need the following lemma.

29.4.1. (Convergence Lemma) Let fi : V → Rq be a sequence of
C0
(smooth) maps. If fi → f and
dg (fi , fi+1 ) < ci ,
C1
with ci < ∞, then f is a C 1 -smooth map and fi → f .
i

∞
Indeed, the convergence of the series dg (fi , fi+1 ) is just the Cauchy con-
i=1
dition for first derivatives of the sequence fi .
B. Approximation Theorem. An embedding f : (V, g) → (Rq , h) is

called ε-isometric if
(1 − ε)g < f ∗ h < (1 + ε)g.
Instead of the theorem about C 1 -isometric embeddings we will first prove
the following.
29.5. One-dimensional Approximation Theorem 345
29.4.2. (Approximation Theorem) Let n < q. For any ε > 0, any short
embedding f : (V n , g) → (Rq , h) can be C 0 -approximated by ε-isometric
embeddings. Moreover, we can also control the C 1 -closeness in the following
sense: given a fixed decomposition of the semi-Riemannian metric
Δ = g − f ∗h
into a sum of N primitive metrics, then for any constant ρ > 0 we can
choose an approximating embedding f which satisfies the inequality
dg (f, f) < N r(Δ, g) + ρ.
Our proof of the Nash–Kuiper theorem will consist of two parts. First we will
use the first part of Approximation Theorem 29.4.2 to construct a sequence
of embeddings fi such that
C0 C0
fi → f and fi∗ h → g.
Then we will refine our choice of the sequence fi in order to have

dg (fi , fi+1 ) < ∞.
i
Then according to Convergence Lemma 29.4.1, f will be a C 1 -limit and

∗
f h = g.
The next three sections are devoted to the proof of Approximation Theorem
29.4.2.
29.5. One-dimensional Approximation Theorem

29.5.1. For any ε > 0 any short embedding f : (I, g) → (Rq , h) can be
C 0 -approximated by ε-isometric embeddings. Moreover, for any ρ > 0 the
approximating map f can be chosen in such a way that
dg (f, f) < r(Δ, g) + ρ
where Δ = g − f ∗ h.
Proof. Let f : (I, g) → (Rq , h) be a strictly short embedding. Let τ be the

orienting g-unit vector field on I, i.e., ∂τ t > 0 and τ g = 1.
The isometry condition

Riso ⊂ J 1 (I, Rq ) = I × Rq × Rq
over a point (t, y) ∈ I × Rq is the unit sphere
Ω(t, y) = {w ∈ Rq , wh = 1}.
f .
Figure 29.1. The relation Rf and R
Choose a normal vector field n to f (I) ⊂ Rq . Instead of the relation Riso

over I × Rq , we consider a smaller relation Rf ⊂ Riso over f (I) ⊂ I × Rq ,
which consists of vectors w ∈ Ω(t, y) such that
w ∈ Span{f∗ τ, n} and w, f∗ τ h ≥ f∗ τ 2h
(see Figure 29.1). The pair (f, f∗ τ / f∗ τ h ) is a short formal solution of
Rf . Let R f ⊂ J 1 (I, Rq ) be a small open neighborhood of Rf ⊂ J 1 (I, Rq ).
Applying one-dimensional convex integration (Lemma 25.3.1), one can con-
struct a solution f of R f which is arbitrarily C 0 -close to f .
If the map f is sufficiently C 0 -close to f , then f will also be an embedding,

because the angle between f∗ τ and f∗ τ is less than π2 − const.
f and f such that

For any ρ > 0, we can choose R
dg (f, f) < r(Δ, g) + ρ.
Indeed, using the Pythagorean theorem (see Figure 29.1) we have
'
dg (f, f)(τ ) = f∗ τ − f∗ τ h < 1 − f∗ τ 2h + ρ,
f → Rf . On the other hand,
where ρ → 0 if R
'
1 − f∗ τ 2h = (g − f ∗ h)(τ ) = τ g−f ∗ h = r(Δ, g)(τ ).
29.6. Adding a primitive metric

29.6.1. Suppose that n < q. Let f : (V, g) → (Rq , h) be a short embedding
such that Δ = g − f ∗ h is a primitive metric on V . Then for any ε the em-
bedding f can be C 0 -approximated by ε-isometric embeddings. Moreover, for
any ρ > 0, the approximating map f can be chosen to satisfy the inequality
dg (f, f) < r(Δ, g) + ρ.
29.6. Adding a primitive metric 347
n=2 n=3
Figure 29.2. The foliations P and L.
Proof. It is sufficient to consider the case (V, g) = (Rn , g).
We are going to reduce this version of the Approximation Theorem to the

parametric one-dimensional convex integration lemma (see 25.5.1). Let
g − f ∗ h = α(x)(dl)2 .
The map f is isometric on each leaf of the (n−1)-dimensional affine foliation

P = {l(x) = const}. Let v be the vector field on Rn normal with respect
to the metric g to the leaves of P. Integral trajectories of v form a one-
dimensional foliation L normal (with respect to the metric g) to the foliation
P (see Figure 29.2).
We can choose a global frame ∂i , i = 1, . . . , n, on V = Rn such that ∂1 is

tangent to L and ∂i , i = 2, . . . , n, are tangent to P. Therefore
⎧
⎪
⎨f∗ ∂i , f∗ ∂j h = ∂i , ∂j g , 2 ≤ i ≤ j ≤ n,
f∗ ∂1 , f∗ ∂j h = ∂1 , ∂j g = 0, 2 ≤ j ≤ n,
⎪
⎩
f∗ ∂1 , f∗ ∂1 h = ∂1 , ∂1 g − ∂1 , ∂1 α(x)(dl)2 .
Choose a normal vector field n to f (Rn ) ⊂ Rq . The map f can be considered

as a family of maps fp : Lp → Rq , where Lp are the leaves of L, and
hence we can apply the parametric version of the previous proof using the
parametric one-dimensional Lemma 25.5.1. According to property 25.5.1(d),
the derivatives ∂i f = f∗ ∂i , i = 2, . . . , n of the new map f can be made
arbitrarily close to the respective derivatives ∂i f = f∗ ∂i , i = 2, . . . , n, of the
initial embedding f . In particular, f will be an embedding if f is sufficiently
C 0 -close to f . On the other hand, the scalar products
f∗ ∂1 , f∗ ∂j h , 2 ≤ j ≤ n,
can be made arbitrarily small by choosing the relations R f sufficiently close

p
to Rfp . Therefore one can construct f such that f h will be arbitrarily close
∗
to g.
f sufficiently
As in the one-dimensional case, for any ρ > 0, by choosing R p
close to Rfp , we can construct f such that
dg (f, f) ≤ r(g − f ∗ h, g) + ρ.
29.7. End of the proof of the Approximation Theorem

N
Let g − f ∗ h = pi be a primitive decomposition of the metric g − f ∗ h. Let
i=1

k
gk = f ∗ h + pi , k = 1, . . . , N,
i=1
so that gN = g. Using 29.6.1 we can construct embeddings f1 , f2 , . . . , fN =

f such that fi∗ h is arbitrarily close to gi , i = 1, . . . , N , and, in particular,
f∗ h = fN
∗ h is arbitrarily close to g = g . Moreover, given a constant ρ > 0
N
we can construct embeddings f1 , f2 , . . . , fN such that for i = 1, . . . , N we
have
dgi (fi−1 , fi ) < r(pi , gi ) + ρ ,
where ρ = ρ/N and we set f0 = f . Using the inequalities
g ≥ gi , i = 1, . . . , N − 1, and r(pi , gi ) < 1, i = 1, . . . , N,
and the properties (r2) and (d1) from Section 29.4, we get
dg (f0 , f1 ) ≤ dg1 (f0 , f1 ) < r(p1 , g1 ) + ρ
≤ r(p1 + p2 + · · · + pN , g1 + p2 + · · · + pN ) + ρ = r(Δ, g) + ρ ,
dg (f1 , f2 ) ≤ dg2 (f1 , f2 ) < r(p2 , g2 ) + ρ
≤ r(p2 + p3 + · · · + pN , g2 + p3 + · · · + pN ) + ρ ≤ r(Δ, g) + ρ ,
...
dg (fN −1 , fN ) ≤ dgN (fN −1 , fN ) < r(pN , gN ) + ρ ≤ r(Δ, g) + ρ .
On the other hand,
dg (f, f) = dg (f0 , fN ) ≤ dg (f0 , f1 ) + dg (f1 , f2 ) + · · · + dg (fN −1 , fN ).
Therefore,
dg (f, f) < N r(Δ, g) + ρ.
29.8. Proof of the Nash–Kuiper theorem 349
29.8. Proof of the Nash–Kuiper theorem

∞
Choose constants ρi > 0 such that ρi < ∞ and choose a constant k > 0
1
such that k 2 f ∗ h > g.
Fix a decomposition of Δ = g − f ∗ h into a sum of N primitive metrics.
According to the shortness condition, Δ is a family of positive quadratic
forms. Fix a positive increasing sequence δi ↑ 1 such that
√
δ 1 + δ2 − δ1 + δ3 − δ2 + · · · < ∞.
Note that gi = f ∗ h + δi Δ → g. On the other hand, f ∗ h < g1 and thus the
embedding f0 = f : (V, g1 ) → (Rq , h) is strictly short. Using Theorem 29.4.2
we can C 0 -approximate the embedding f0 by an ε1 -isometric embedding
f1 : (V, g1 ) → (Rq , h) such that

d(f0 , f1 ) < N r(δ1 Δ, g1 ) + ρ1 = N δ1 r(Δ, g1 ) + ρ1

≤ N δ1 r(Δ, f ∗ g) + ρ1 = N k δ1 r(Δ, k 2 f ∗ g) + ρ1

< N k δ1 r(Δ, g) + ρ1 .
If ε1 is sufficiently small, then f1∗ h ≈ f0∗ h + δ1 Δ, and hence
f1∗ h < f0∗ h + δ2 Δ = g2 ,
which means that the embedding f1 : (V, g2 ) → (Rq , h) is strictly short.
Hence, for any ε2 > 0, we can C 0 -approximate the embedding f1 by an
ε2 -isometric embedding f2 : (V, g2 ) → (Rq , h). Moreover, choosing ε1 suf-
ficiently small at the previous step of the construction, we can make the
difference g2 − f1∗ h arbitrarily close to (δ2 − δ1 )Δ. Hence,

d(f1 , f2 ) < N r((δ2 − δ1 )Δ, g2 ) + ρ2 = N δ2 − δ1 r(Δ, g2 ) + ρ2

≤ N δ2 − δ1 r(Δ, f ∗ g) + ρ2 = N k δ2 − δ1 r(Δ, k 2 f ∗ g) + ρ2

< N k δ2 − δ1 r(Δ, g) + ρ2 ,
and so on. Note that according to the Remark to Lemma 29.3.1, the constant
N does not depend on i.
The sequence {fi } can be chosen C 0 -converging to a continuous map f . On
the other hand,
√
d(fi , fi+1 ) < N kr(Δ, g)( δ 1 + δ2 − δ1 + δ3 − δ2 + · · · ) + ρi < ∞.
i i
Therefore, Convergence Lemma 29.4.1 guarantees that the limit map f is a

C1
C 1 -smooth isometric embedding and fi → f .
Bibliography
[Ad93] M. Adachi, Embeddings and immersions, Translations of Mathematical Monographs,

vol. 124, American Mathematical Society, Providence, RI, 1993. Translated from the
1984 Japanese original by Kiki Hudson, DOI 10.1090/mmono/124. MR1225100
[ÁG18] D. Álvarez-Gavela, Refinements of the holonomic approximation lemma, Algebr.
Geom. Topol. 18 (2018), no. 4, 2265–2303, DOI 10.2140/agt.2018.18.2265. MR3797067
[AG18b] D. Álvarez-Gavela, The simplification of singularities of Lagrangian and Legendrian
fronts, Invent. Math. 214 (2018), no. 2, 641–737, DOI 10.1007/s00222-018-0811-3.
MR3867630
[Ar65] V. Arnold, Sur une propriété topologique des applications globalement canoniques
de la mécanique classique (French), C. R. Acad. Sci. Paris 261 (1965), 3719–3722.
MR193645
[Arn76] V. I. Arnold, Wave front evolution and equivariant Morse lemma, Comm. Pure Appl.
Math. 29 (1976), no. 6, 557–582, DOI 10.1002/cpa.3160290603. MR436200
[Ar78] V. I. Arnold, Mathematical methods of classical mechanics, Graduate Texts in Math-
ematics, vol. 60, Springer-Verlag, New York-Heidelberg, 1978. Translated from the
Russian by K. Vogtmann and A. Weinstein. MR690288
[Ar86] V.I. Arnold, First steps of symplectic topology, Russian Math. Surveys, 41 (1986) no.
6(251), 1–21.
[AGZV85] V. I. Arnold, S. M. Guseı̆n-Zade, and A. N. Varchenko, Singularities of differentiable
maps. Vol. I: The classification of critical points, caustics and wave fronts, Monographs
in Mathematics, vol. 82, Birkhäuser Boston, Inc., Boston, MA, 1985. Translated from
the Russian by Ian Porteous and Mark Reynolds, DOI 10.1007/978-1-4612-5154-5.
MR777682
[AG01] V. I. Arnold and A. B. Givental, Symplectic geometry [MR0842908 (88b:58044)],
Dynamical systems, IV, Encyclopaedia Math. Sci., vol. 4, Springer, Berlin, 2001, pp. 1–
138. MR1866631
[Ba78] A. Banyaga, Sur la structure du groupe des difféomorphismes qui préservent une
forme symplectique (French), Comment. Math. Helv. 53 (1978), no. 2, 174–227, DOI
10.1007/BF02566074. MR490874
[Be83] D. Bennequin, Entrelacements et équations de Pfaff (French), Third Schnepfenried
geometry conference, Vol. 1 (Schnepfenried, 1982), Astérisque, vol. 107, Soc. Math.
France, Paris, 1983, pp. 87–161. MR753131
351
352 Bibliography
[Be00] M. Berger, Encounter with a geometer. I, Notices Amer. Math. Soc. 47 (2000), no. 2,
183–194. MR1740350
[BM15] P. Bernard and V. Mandorino, Some remarks on Thom’s transversality theorem, Ann.
Sc. Norm. Super. Pisa Cl. Sci. (5) 14 (2015), no. 2, 361–386. MR3410613
[Bh22] A. Bhowmick, The h-principle for maps transverse to bracket-generating distributions,
arXiv:2205.04323, 2022.
[Bo01] V. Borrelli, On totally real isotopy classes, Int. Math. Res. Not. 2 (2002), 89–109, DOI
10.1155/S1073792802105125. MR1874320
[Boa67] J. M. Boardman, Singularities of differentiable maps, Inst. Hautes Études Sci. Publ.
Math. 33 (1967), 21–57. MR231390
[BEM15] M. S. Borman, Y. Eliashberg, and E. Murphy, Existence and classification of over-
twisted contact structures in all dimensions, Acta Math. 215 (2015), no. 2, 281–361,
DOI 10.1007/s11511-016-0134-4. MR3455235
[Bro65] E. H. Brown Jr., Abstract homotopy theory, Trans. Amer. Math. Soc. 119 (1965),
79–85, DOI 10.2307/1994231. MR182970
[BH93] R. L. Bryant and L. Hsu, Rigidity of integral curves of rank 2 distributions, Invent.
Math. 114 (1993), no. 2, 435–461, DOI 10.1007/BF01232676. MR1240644
[Ca70] E. Calabi, On the group of automorphisms of a symplectic manifold, Problems in
analysis (Sympos. in honor of Salomon Bochner, Princeton Univ., Princeton, N.J.,
1969), Princeton Univ. Press, Princeton, NJ, 1970, pp. 1–26. MR350776
[CdS01] A. Cannas da Silva, Lectures on symplectic geometry, Lecture Notes in Mathematics,
vol. 1764, Springer-Verlag, Berlin, 2001, DOI 10.1007/978-3-540-45330-7. MR1853077
[CdS10] A. Cannas da Silva, Fold-forms for four-folds, J. Symplectic Geom. 8 (2010), no. 2,
189–203. MR2670164
[CPPP17] R. Casals, J. L. Pérez, Á. del Pino, and F. Presas, Existence h-principle for Engel
structures, Invent. Math. 210 (2017), no. 2, 417–451, DOI 10.1007/s00222-017-0732-6.
MR3714508
[CC00] A. Candel and L. Conlon, Foliations. I, Graduate Studies in Mathematics, vol. 23,
American Mathematical Society, Providence, RI, 2000, DOI 10.1090/gsm/023.
MR1732868
[Ce70] J. Cerf, La stratification naturelle des espaces de fonctions différentiables réelles et
le théorème de la pseudo-isotopie (French), Inst. Hautes Études Sci. Publ. Math. 39
(1970), 5–173. MR292089
[CEM25] K. Cieliebak, Y. Eliashberg and N. Mishachev, Symplectic topology and the h-principle,
in preparation.
[CE12] K. Cieliebak and Y. Eliashberg, From Stein to Weinstein and back: Symplectic ge-
ometry of affine complex manifolds, American Mathematical Society Colloquium
Publications, vol. 59, American Mathematical Society, Providence, RI, 2012, DOI
10.1090/coll/059. MR3012475
[DA00] G. D’Ambra, An application of the h-principle to C 1 -isometric immersions
in contact manifolds, Comm. Anal. Geom. 8 (2000), no. 2, 347–373, DOI
10.4310/CAG.2000.v8.n2.a4. MR1753321
[DL01] G. D’Ambra and A. Loi, A symplectic version of the Nash C 1 -isometric embedding
theorem, Differential Geom. Appl. 16 (2002), no. 2, 167–179, DOI 10.1016/S0926-
2245(02)00067-0. MR1893907
[Da97] M. Datta, Homotopy classification of strict contact immersions, Ann. Global Anal.
Geom. 15 (1997), no. 3, 211–219, DOI 10.1023/A:1006589225203. MR1456508
[DS12] C. De Lellis and L. Székelyhidi, The h-principle and the equations of fluid dynamics,
Bull. Amer. Math. Soc. (N.S.) 49 (2012), no. 3, 347–375.
Bibliography 353
[DRGI16] G. Dimitroglou Rizell, E. Goodman, and A. Ivrii, Lagrangian isotopy of tori in S 2 ×S 2

and CP 2 , Geom. Funct. Anal. 26 (2016), no. 5, 1297–1358, DOI 10.1007/s00039-016-
0388-1. MR3568033
[Do96] S. K. Donaldson, Symplectic submanifolds and almost-complex geometry, J. Differential
Geom. 44 (1996), no. 4, 666–705. MR1438190
[Du84] T. Duchamp, The classification of Legendre immersions, preprint, 1984.
[Ee66] J. Eells Jr., A setting for global analysis, Bull. Amer. Math. Soc. 72 (1966), 751–807,
DOI 10.1090/S0002-9904-1966-11558-6. MR203742
[El70] Ja. M. Èliašberg, Singularities of folding type (Russian), Izv. Akad. Nauk SSSR Ser.
Mat. 34 (1970), 1110–1126. MR278321
[El72] Ja. M. Èliašberg, Surgery of singularities of smooth mappings (Russian), Izv. Akad.
Nauk SSSR Ser. Mat. 36 (1972), 1321–1347. MR339261
[El85] Ya. M. Eliashberg, Complexification of contact structures on 3-dimensional manifolds
(Russian), Uspekhi Mat. Nauk 40 (1985), no. 6(246), 161–162. MR815508
[El89] Y. Eliashberg, Classification of overtwisted contact structures on 3-manifolds, Invent.
Math. 98 (1989), no. 3, 623–637, DOI 10.1007/BF01393840. MR1022310
[El91] Y. Eliashberg, Filling by holomorphic discs and its applications, Geometry of low-
dimensional manifolds, 2 (Durham, 1989), London Math. Soc. Lecture Note Ser.,
vol. 151, Cambridge Univ. Press, Cambridge, 1990, pp. 45–67. MR1171908
[El92] Y. Eliashberg, Contact 3-manifolds twenty years since J. Martinet’s work (English,
with French summary), Ann. Inst. Fourier (Grenoble) 42 (1992), no. 1-2, 165–192.
MR1162559
[El15] Y. Eliashberg, Recent advances in symplectic flexibility, Bull. Amer. Math. Soc. (N.S.)
52 (2015), no. 1, 1–26, DOI 10.1090/S0273-0979-2014-01470-3. MR3286479
[EGM11] Y. Eliashberg, S. Galatius, and N. Mishachev, Madsen-Weiss for geometrically minded
topologists, Geom. Topol. 15 (2011), no. 1, 411–472, DOI 10.2140/gt.2011.15.411.
MR2776850
[EGH00] Y. Eliashberg, A. Givental, and H. Hofer, Introduction to symplectic field theory, Geom.
Funct. Anal. Special Volume (2000), 560–673, DOI 10.1007/978-3-0346-0425-3 4.
GAFA 2000 (Tel Aviv, 1999). MR1826267
[EM97] Y. Eliashberg and N. M. Mishachev, Wrinkling of smooth mappings and its appli-
cations. I, Invent. Math. 130 (1997), no. 2, 345–369, DOI 10.1007/s002220050188.
MR1474161
[EM98] Y. Eliashberg and N. M. Mishachev, Wrinkling of smooth mappings. III. Foliations
of codimension greater than one, Topol. Methods Nonlinear Anal. 11 (1998), no. 2,
321–350, DOI 10.12775/TMNA.1998.021. MR1659446
[EM00] Y. M. Eliashberg and N. M. Mishachev, Wrinkling of smooth mappings. II. Wrin-
kling of embeddings and K. Igusa’s theorem, Topology 39 (2000), no. 4, 711–732, DOI
10.1016/S0040-9383(99)00029-4. MR1760426
[EM01] Y. M. Eliashberg and N. M. Mishachev, Holonomic approximation and Gromov’s h-
principle, Essays on geometry and related topics, Vol. 1, 2, Monogr. Enseign. Math.,
vol. 38, Enseignement Math., Geneva, 2001, pp. 271–285. MR1929330
[EM11] Y. Eliashberg and N. Mishachev, Topology of spaces of S-immersions, Perspectives in
analysis, geometry, and topology, Progr. Math., vol. 296, Birkhäuser/Springer, New
York, 2012, pp. 147–167, DOI 10.1007/978-0-8176-8277-4 7. MR2884035
[EM12] Y. M. Eliashberg and N. M. Mishachev, The space of framed functions is contractible,
Essays in mathematics and its applications, Springer, Heidelberg, 2012, pp. 81–109,
DOI 10.1007/978-3-642-28821-0 5. MR2975585
[EM09] Y. M. Eliashberg and N. M. Mishachev, Wrinkled embeddings, Foliations, geometry,
and topology, Contemp. Math., vol. 498, Amer. Math. Soc., Providence, RI, 2009,
pp. 207–232, DOI 10.1090/conm/498/09753. MR2664601
354 Bibliography
[ET79] D. B. A. Epstein and W. P. Thurston, Transformation groups and natural bundles,

Proc. London Math. Soc. (3) 38 (1979), no. 2, 219–236, DOI 10.1112/plms/s3-38.2.219.
MR531161
[Et99] J. B. Etnyre, Tight contact structures on lens spaces, Commun. Contemp. Math. 2
(2000), no. 4, 559–577, DOI 10.1142/S0219199700000207. MR1806947
[Fe69] S. D. Feit, k-mersions of manifolds, Acta Math. 122 (1969), 173–195, DOI
10.1007/BF02392010. MR243541
[Fd87] T. Fiedler, Totally real embeddings of the torus into C2 , Ann. Global Anal. Geom. 5
(1987), no. 2, 117–121, DOI 10.1007/BF00127854. MR944776
[Fi67] A. F. Filippov, Classical solutions of differential equations with multi-valued right-hand
side, SIAM J. Control 5 (1967), 609–621. MR220995
[FPT23] A. Fokma, Á. del Pino, and L. Toussaint, Wrinkling and Haefliger structures,
arXiv:2309.15715, 2023.
[Ge01] H. Geiges, h-principles and flexibility in geometry, Mem. Amer. Math. Soc. 164 (2003),
no. 779, viii+58, DOI 10.1090/memo/0779. MR1982875
[GT99] H. Geiges and C. B. Thomas, Contact structures on 7-manifolds, Geometry, topology,
and dynamics (Montreal, PQ, 1995), CRM Proc. Lecture Notes, vol. 15, Amer. Math.
Soc., Providence, RI, 1998, pp. 53–67, DOI 10.1090/crmp/015/03. MR1619123
[Gi97] V. L. Ginzburg, A smooth counterexample to the Hamiltonian Seifert conjecture in R6 ,
Internat. Math. Res. Notices 13 (1997), 641–650, DOI 10.1155/S1073792897000421.
MR1459629
[Gi99] E. Giroux, Une infinité de structures de contact tendues sur une infinité de variétés
(French), Invent. Math. 135 (1999), no. 3, 789–802, DOI 10.1007/s002220050301.
MR1669264
[Gi01] E. Giroux, Structures de contact sur les variétés fibrées en cercles audessus d’une
surface (French, with French summary), Comment. Math. Helv. 76 (2001), no. 2,
218–262, DOI 10.1007/PL00000378. MR1839346
[GM87] M. Goresky and R. MacPherson, Stratified Morse theory, Ergebnisse der Mathematik
und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 14,
Springer-Verlag, Berlin, 1988, DOI 10.1007/978-3-642-71714-7. MR932724
[Gr59] J. W. Gray, Some global properties of contact structures, Ann. of Math. (2) 69 (1959),
421–450, DOI 10.2307/1970192. MR112161
[Gr69] M. L. Gromov, Stable mappings of foliations into manifolds (Russian), Izv. Akad.
[Gr71] M. L. Gromov, A topological technique for the construction of solutions of differential
equations and inequalities, Actes du Congrès International des Mathématiciens (Nice,
1970), Gauthier-Villars Éditeur, Paris, 1971, pp. 221–225. MR420697
[Gr73] M. L. Gromov, Convex integration of differential relations. I (Russian), Izv. Akad.
[Gr85] M. Gromov, Pseudo holomorphic curves in symplectic manifolds, Invent. Math. 82
(1985), no. 2, 307–347, DOI 10.1007/BF01388806. MR809718
[Gr86] M. Gromov, Partial differential relations, Ergebnisse der Mathematik und ihrer Gren-
zgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 9, Springer-Verlag,
Berlin, 1986, DOI 10.1007/978-3-662-02267-2. MR864505
[GE71] M. L. Gromov and Ja. M. Èliašberg, Elimination of singularities of smooth mappings
(Russian), Izv. Akad. Nauk SSSR Ser. Mat. 35 (1971), 600–626. MR301748
[GE73] M. Gromov and Y. Eliashberg, Construction of a smooth map with prescribed Jacobian,
Funk. Anal. and Applications, 5:2 (1973), 27–32
[Hae70] A. Haefliger, Feuilletages sur les variétés ouvertes (French), Topology 9 (1970), 183–
194, DOI 10.1016/0040-9383(70)90040-6. MR263104
Bibliography 355
[Ha70b] A. Haefliger, Homotopy and integrability, Manifolds–Amsterdam 1970 (Proc. Nuffic

Summer School), Lecture Notes in Math., Vol. 197, Springer, Berlin-New York, 1971,
pp. 133–163. MR285027
[Ha71] A. Haefliger, Lectures on the theorem of Gromov, Proceedings of Liverpool Singularities
Symposium, II (1969/1970), Lecture Notes in Math., Vol. 209, Springer, Berlin-New
York, 1971, pp. 128–141. MR334241
[HW73] A. Hatcher and J. Wagoner, Pseudo-isotopies of compact manifolds, Astérisque, No.
6, Société Mathématique de France, Paris, 1973. With English and French prefaces.
MR353337
[Hi59] M. W. Hirsch, Immersions of manifolds, Trans. Amer. Math. Soc. 93 (1959), 242–276,
DOI 10.2307/1993453. MR119214
[HZ94] H. Hofer and E. Zehnder, Symplectic invariants and Hamiltonian dynamics,
Birkhäuser Advanced Texts: Basler Lehrbücher. [Birkhäuser Advanced Texts:
Basel Textbooks], Birkhäuser Verlag, Basel, 1994, DOI 10.1007/978-3-0348-8540-9.
MR1306732
[Ho00] K. Honda, On the classification of tight contact structures. I, Geom. Topol. 4 (2000),
309–368, DOI 10.2140/gt.2000.4.309. MR1786111
[Ig84] K. Igusa, Higher singularities of smooth functions are unnecessary, Ann. of Math. (2)
119 (1984), no. 1, 1–58, DOI 10.2307/2006962. MR736559
[Ig87] K. Igusa, The space of framed functions, Trans. Amer. Math. Soc. 301 (1987), no. 2,
431–477, DOI 10.2307/2000654. MR882699
[Ku55] N. H. Kuiper, On C 1 -isometric imbeddings. I, II, Indag. Math. 17 (1955), 545–556,
683–689. Nederl. Akad. Wetensch. Proc. Ser. A 58. MR75640
[Lau19] F. Laudenbach, René Thom and an anticipated h-principle, Geometry in history,
Springer, Cham, 2019, pp. 469–491. MR3965771
[La76] F. Laudenbach, Formes différentielles de degré 1 fermées non singulières: classes
d’homotopie de leurs noyaux, Comment. Math. Helv. 51 (1976), no. 3, 447–464, DOI
10.1007/BF02568169. MR428335
[La74] H. B. Lawson Jr., Foliations, Bull. Amer. Math. Soc. 80 (1974), 369–418, DOI
10.1090/S0002-9904-1974-13432-4. MR343289
[Le76] J. A. Lees, On the classification of Lagrange immersions, Duke Math. J. 43 (1976),
no. 2, 217–224. MR410764
[Lo95] J. Lohkamp, Curvature h-principles, Ann. of Math. (2) 142 (1995), no. 3, 457–498,
DOI 10.2307/2118552. MR1356779
[Lu09] J. Lurie, On the classification of topological field theories, Current developments in
mathematics, 2008, Int. Press, Somerville, MA, 2009, pp. 129–280. MR2555928
[Lu77] R. Lutz, Structures de contact sur les fibrés principaux en cercles de dimension trois
(French, with English summary), Ann. Inst. Fourier (Grenoble) 27 (1977), no. 3, ix,
1–15. MR478180
[Ma71] J. Martinet, Formes de contact sur les variétés de dimension 3 (French), Proceedings
of Liverpool Singularities Symposium, II (1969/1970), Lecture Notes in Math., Vol.
209, Springer, Berlin-New York, 1971, pp. 142–163. MR350771
[MP21] J. Martı́nez-Aguinaga and Á. del Pino, Convex integration with avoidance and hyper-
bolic (4, 6) distributions, arXiv:2112.14632, 2021.
[MDN22] P. Massot, F. van Doorn, and O. Nash, Formalising the h-principle and sphere eversion
arXiv:2210.07746, 2022.
[MT23] P. Massot and M. Theillière, Holonomic approximation through convex integration,
Int. Math. Res. Not. IMRN (2023), no. 9, 7486–7501.
356 Bibliography
[MD87a] D. McDuff, Applications of convex integration to symplectic and contact geometry

(English, with French summary), Ann. Inst. Fourier (Grenoble) 37 (1987), no. 1, 107–
133. MR894563
[MD87b] D. McDuff, Examples of symplectic structures, Invent. Math. 89 (1987), no. 1, 13–36,
DOI 10.1007/BF01404672. MR892186
[MS98] D. McDuff and D. Salamon, Introduction to symplectic topology, 2nd ed., Oxford Math-
ematical Monographs, The Clarendon Press, Oxford University Press, New York, 1998.
MR1698616
[MT99] C. T. McMullen and C. H. Taubes, 4-manifolds with inequivalent symplectic forms and
3-manifolds with inequivalent fibrations, Math. Res. Lett. 6 (1999), no. 5-6, 681–696,
DOI 10.4310/MRL.1999.v6.n6.a8. MR1739225
[Mei17] G. Meigniez, Regularization and minimization of codimension-one Haefliger struc-
tures, J. Differential Geom. 107 (2017), no. 1, 157–202, DOI 10.4310/jdg/1505268031.
MR3698236
[Mi65] J. Milnor, Lectures on the h-cobordism theorem, Princeton University Press, Princeton,
NJ, 1965. Notes by L. Siebenmann and J. Sondow. MR190942
[Mi70] J. Milnor, Foliations and foliated vector bundles, mimeographed notes, 1970.
[Mo65] B. Morin, Formes canoniques des singularités d’une application différentiable
(French), C. R. Acad. Sci. Paris 260 (1965), 5662–5665. MR180982
[Mo39] A. P. Morse, The behavior of a function on its critical set, Ann. of Math. (2) 40 (1939),
no. 1, 62–70, DOI 10.2307/1968544. MR1503449
[Mo99] R. Montgomery, Engel deformations and contact structures, Northern California Sym-
plectic Geometry Seminar, Amer. Math. Soc. Transl. Ser. 2, vol. 196, Amer. Math.
Soc., Providence, RI, 1999, pp. 103–117, DOI 10.1090/trans2/196/07. MR1736216
[Mo69] J. Moser, On the volume elements on a manifold, Trans. Amer. Math. Soc. 120 (1965),
286–294, DOI 10.2307/1994022. MR182927
[Mu11] E. Murphy, Loose Legendrian Embeddings in High Dimensional Contact Manifolds,
ProQuest LLC, Ann Arbor, MI, 2012. Thesis (Ph.D.)–Stanford University. MR4172336
[Ni72] A. Nijenhuis, Natural bundles and their general properties. Geometric objects revisited,
Differential geometry (in honor of Kentaro Yano), Kinokuniya Book Store, Tokyo, 1972,
pp. 317–334. MR380862
[KMS93] I. Kolář, P. W. Michor, and J. Slovák, Natural operations in differential geometry,
Springer-Verlag, Berlin, 1993, DOI 10.1007/978-3-662-02950-3. MR1202431
[Ku19] A. Kupers, Three applications of delooping to h-principles, Geom. Dedicata 202
(2019), 103–151, DOI 10.1007/s10711-018-0405-7. MR4001810
[Na54] J. Nash, C 1 isometric imbeddings, Ann. of Math. (2) 60 (1954), 383–396, DOI
10.2307/1969840. MR65993
[NN57] A. Newlander and L. Nirenberg, Complex analytic coordinates in almost complex mani-
folds, Ann. of Math. (2) 65 (1957), 391–404, DOI 10.2307/1970051. MR88770
[Pa66] R. S. Palais, Homotopy theory of infinite dimensional manifolds, Topology 5 (1966),
1–16, DOI 10.1016/0040-9383(66)90002-4. MR189028
[PT77] R. S. Palais and C. L. Terng, Natural bundles have finite order, Topology 19 (1977),
no. 3, 271–277, DOI 10.1016/0040-9383(77)90008-8. MR467787
[Ph67] A. Phillips, Submersions of open manifolds, Topology 6 (1967), 171–206, DOI
10.1016/0040-9383(67)90034-1. MR208611
[Ph70] A. Phillips, Smooth maps transverse to a foliation, Bull. Amer. Math. Soc. 76 (1970),
792–797, DOI 10.1090/S0002-9904-1970-12555-1. MR263106
[Po66] V. Poenaru, On regular homotopy in codimension 1, Ann. of Math. (2) 83 (1966),
257–265, DOI 10.2307/1970430. MR192507
Bibliography 357
[Po71] V. Poénaru, Homotopy theory and differentiable singularities, Manifolds–Amsterdam

1970 (Proc. Nuffic Summer School), Lecture Notes in Math., Vol. 197, Springer, Berlin-
New York, 1971, pp. 106–132. MR285026
[Pt88] L. Polterovich, New invariants of embedded totally real tori and one problem of Hamil-
tonian mechanics, in the book “Methods of qualitative theory and the Theory of bi-
furcations” (Russian), Gorki, 1988, 84–90.
[RS97] C. Rourke and B. Sanderson, The compression theorem. I, II, Geom. Topol. 5 (2001),
399–429, 431–440, DOI 10.2140/gt.2001.5.399. MR1833749
[RS00] C. Rourke and B. Sanderson, Directed embeddings: A short proof of Gromov’s theorem,
preprint, 2000.
[Re52] G. Reeb, Sur certaines propriétés topologiques des variétés feuilletées (French), Pub-
lications de l’Institut de Mathématiques de l’Université de Strasbourg [Publications of
the Mathematical Institute of the University of Strasbourg], vol. 11, Hermann & Cie,
Paris, 1952. Actualités Scientifiques et Industrielles [Current Scientific and Industrial
Topics], No. 1183. MR55692
[Ru94] Y. Ruan, Symplectic topology on algebraic 3-folds, J. Differential Geom. 39 (1994),
no. 1, 215–227. MR1258920
[Sa42] A. Sard, The measure of the critical values of differentiable maps, Bull. Amer. Math.
Soc. 48 (1942), 883–890, DOI 10.1090/S0002-9904-1942-07811-6. MR7523
[Si64] C. L. Siegel, Symplectic geometry, Academic Press, New York–London, 1964.
MR164063
[Sm58] S. Smale, A classification of immersions of the two-sphere, Trans. Amer. Math. Soc.
90 (1958), 281–290, DOI 10.2307/1993205. MR104227
[Sm59] S. Smale, The classification of immersions of spheres in Euclidean spaces, Ann. of
Math. (2) 69 (1959), 327–344, DOI 10.2307/1970186. MR105117
[Sp98] D. Spring, Convex integration theory: Solutions to the h-principle in geometry and
topology, Monographs in Mathematics, vol. 92, Birkhäuser Verlag, Basel, 1998, DOI
10.1007/978-3-0348-0060-0. MR1488424
[Sp00] D. Spring, Directed embeddings and the simplification of singularities, Commun. Con-
temp. Math. 4 (2002), no. 1, 107–144, DOI 10.1142/S0219199702000609. MR1890079
[St51] N. Steenrod, The Topology of Fibre Bundles, Princeton Mathematical Series, vol. 14,
Princeton University Press, Princeton, NJ, 1951. MR39258
[Ta94] C. H. Taubes, The Seiberg-Witten invariants and symplectic forms, Math. Res. Lett.
1 (1994), no. 6, 809–822, DOI 10.4310/MRL.1994.v1.n6.a15. MR1306023
[Th22] M. Theillière, Convex integration theory without integration, Math. Z. 300 (2022),
no. 3, 2737–2770.
[Tho59] R. Thom, Remarques sur les problèmes comportant des inéquations différentielles glob-
ales (French), Bull. Soc. Math. France 87 (1959), 455–461. MR121807
[Tho55] R. Thom, Les singularités des applications différentiables (French), Ann. Inst. Fourier
(Grenoble) 6 (1955/56), 43–87. MR87149
[Thu72] W. Thurston, Noncobordant foliations of S 3 , Bull. Amer. Math. Soc. 78 (1972), 511–
514, DOI 10.1090/S0002-9904-1972-12975-6. MR298692
[Th74] W. Thurston, The theory of foliations of codimension greater than one, Comment.
Math. Helv. 49 (1974), 214–231, DOI 10.1007/BF02566730. MR370619
[Th76] W. P. Thurston, Some simple examples of symplectic manifolds, Proc. Amer. Math.
Soc. 55 (1976), no. 2, 467–468, DOI 10.2307/2041749. MR402764
[Th76b] W. P. Thurston, Existence of codimension-one foliations, Ann. of Math. (2) 104
(1976), no. 2, 249–268, DOI 10.2307/1971047. MR425985
[Us99] I. Ustilovsky, Infinitely many contact structures on S 4m+1 , Internat. Math. Res. No-
tices 14 (1999), 781–791, DOI 10.1155/S1073792899000392. MR1704176
358 Bibliography
[Va92] V. A. Vassiliev, Complements of discriminants of smooth maps: topology and ap-

plications, Translations of Mathematical Monographs, vol. 98, American Mathemati-
cal Society, Providence, RI, 1992. Translated from the Russian by B. Goldfarb, DOI
10.1090/mmono/098. MR1168473
[Vo09] T. Vogel, Existence of Engel structures, Ann. of Math. (2) 169 (2009), no. 1, 79–137,
DOI 10.4007/annals.2009.169.79. MR2480602
[Wa83] F. W. Warner, Foundations of differentiable manifolds and Lie groups, Graduate Texts
in Mathematics, vol. 94, Springer-Verlag, New York-Berlin, 1983. Corrected reprint of
the 1971 edition. MR722297
[Whi61] J. H. C. Whitehead, The immersion of an open 3-manifold in euclidean 3-space, Proc.
London Math. Soc. (3) 11 (1961), 81–90, DOI 10.1112/plms/s3-11.1.81. MR124916
[Wh34] H. Whitney, Analytic extensions of differentiable functions defined in closed sets,
Trans. Amer. Math. Soc. 36 (1934), no. 1, 63–89, DOI 10.2307/1989708. MR1501735
[Wh37] H. Whitney, On regular closed curves in the plane, Compositio Math. 4 (1937), 276–
284. MR1556973
[Whi43] H. Whitney, The general type of singularity of a set of 2n − 1 smooth functions of n
variables, Duke Math. J. 10 (1943), 161–172. MR7784
[Whi55] H. Whitney, On singularities of mappings of euclidean spaces. I. Mappings of the
plane into the plane, Ann. of Math. (2) 62 (1955), 374–410, DOI 10.2307/1970070.
MR73980
[Za87] A. Zajtz, The sharp upper bound on the order of natural bundles of given dimensions,
Bull. Soc. Math. Belg. Sér. B 39 (1987), no. 3, 347–357. MR925277
Index
Almost complex structure, 225 overtwisted, 254

integrable, 225 vector field, 242
Almost symplectic structure, 225 Contact structures
integrable, 225 formally homotopic, 253
Ampleness criterion, 323 homotopic, 253
isotopic, 253
Balanced path, 308 Contactization, 234
Boardman formula, 111 Contactomorphisms, 232
bs F , 13 Convex integration, xvi
iterated, 318
Canonical symplectic structure (form) one-dimensional, 305
on R2n , 226 parametric, 311
on a cotangent bundle, 227 ConvF R, 305
Capacious Lie subgroup, 289 Conv R, 305
Characteristic foliation, 226 Coordinate principal subspace, 317, 321
Cloa R, 246 Core, 51
Compatible complex and symplectic CR-structure, 226
structures, 223 CS(ξ), 233
Complex CS(ξ+ ), 233
manifold, 225, 226 Cusp, 111
structure, 221 creation surgery, 162
subspace, 222 elimination surgery, 162
vector space, 221 sharp, 171
Conny Ω, 304
Contact Darboux contact form, 232
cutting-off, 243 Darboux’s chart, 226
distribution, 232 dg (f, f), 344
form, 232 Diffeotopy, 27
Hamiltonian, 242 δ-small, 27
manifold, 232 Differential condition, 69
monomorphism, 237 Differential inclusion, 304
structure, 5, 232 Differential relation, xv, 69
cooriented, 233 Diff V -invariant, 85
359
360 Index
A-invariant, 289 Flower, 307

k-flexible, 282 abstract, 306
k-microflexible, 281, 282 fibered, 312
affine ample, 326 Fold, 111
ample, 304, 321 sharp, 171
ample in the coordinate directions, sharpening, 190
317 smoothing, 190
closed, 71 with respect to a foliation, 114
determined, xv, 71 Foliation, 41, 95
fibered, 79 concordance, 98
fiberwise path-connected, 305 integrable homotopy, 98
locally integrable, 279, 280 Formal inverse, 332
microflexible, 282 Formal primitive, 57
open, 71 Framed
overdetermined, 71 FLIF, 196
underdetermined, xv, 71 Igusa function, 195
Distribution, 56
integrable, 96 GF ,Gdf , 48
involutive, 96 Grn W , 48
Embedding h-principle for folded solutions, 165

ε-Lagrangian, 4 Haefliger structure, 99
co-real, 328 augmented, 101
contact, 272 Hamiltonian function, 241
directed, 326 time dependent, 241
isocontact, 272 Hamiltonian isotopy, 241
isosymplectic, 259 Hermitian structure, 223, 225
Lagrangian, 4 integrable, 225
real, 3, 328 H(L), 223
symplectic, 259 Holonomic R-approximation, 283
Embryo, 129 Holonomic approximation, 25
Engel structure, 5 multivalued, 63
Epimorphism, 70 Hol X (r) , 13
Exact Homotopy
Lagrangian immersion, 240 holonomic, 13
Lagrangian submanifold, 240 regular, 1, 46
symplectic manifold, 296 tangential, 48
Exa R, 248 Homotopy principle (h-principle), xv, 3,
Exit to infinity, 50 76
C 0 -dense, 79
Family of sections, 10 (multi) parametric, 78
continuous, 10 fibered, 80
smooth, 10 foliated, 102
Fiber bundle, 9 for
Fibration, 9 C 1 -isometric immersions, 342
trivial, 9 D-sections, 332
FLIF, 196 ample differential relations, 321
balanced, 203 ample differential relations over a
framed, 196 cube, 320
holonomic, 196 contact structures on open
locally, 196 manifolds, 248
stabilization of, 204 directed embeddings, 326
Index 361
divergence free vector fields, 334 ε-Lagrangian, 324

immersions transverse to contact ε-coisotropic, 324
structure, 287 ε-isotropic, 324
immersions transverse to co-real, 228
distribution, 91 coisotropic, 228
isocontact embeddings, 272 complex, 228, 326
isocontact immersions, 293 contact, 237
isosymplectic embeddings, 261 isocomplex, 228
isosymplectic immersions, 298 isocontact, 237, 294
Lagrangian immersions, 296 isometric, 2, 341
Legendrian immersions, 294 isosymplectic, 228, 299
linearly independent D-sections, isotropic, 228, 237
334 Lagrangian, 228
maps transverse to contact Legendrian, 237
structure, 286 real, 228, 326
maximally nondegenerate subcritical, 237
two-forms on odd-dimensional symplectic, 228
manifolds, 250, 337 Immersion relation, 70
microflexible Diff V -invariant Isotopy
relations, 283 Hamiltonian, 241
microflexible A-invariant relations, Legendrian, 241
290
nonintegrable hyperplane J , 245
distributions on J (L), 222
even-dimensional manifolds, 288, Jfr , 10
338 J r (Rn , Rq ), 10
nonvanishing D-sections, 333 J r (V, W ), 12
open Diff V -invariant relations, 85 J (X), 224
real and co-real embeddings, 328 k-mersion, 87
real and co-real immersions, 326 Kähler manifold, 225
sections transverse to distribution, Kähler metric, 225
91
subcritical isotropic embeddings, Λp V , 56
277 Liouville structure, 240
symplectic forms on open manifold, LX , 229
247
systems of divergence free vector Manifold
fields, 336 almost complex, 225
systems of exact forms, 336 almost Kähler, 225
local, 78 almost symplectic, 225
one-parametric, 76 complex, 225, 226
relative, 79 contact, 232
Smale–Hirsch, 89 Hermitian, 225
H(X), 224 Kähler, 225
open, 50
Igusa symplectic, 225, 226
function, 193 Map
leafwise, 193 free, 4
Igusa function short, 341
framed, 195 strictly short, 341
Immersion, 1, 46, 70 transverse to a distribution, 91
A-directed, 53 transverse to a stratified set, 19
362 Index
transverse to a submanifold, 17 Rcont , 237

Microextension trick, 89 Rco-real , 326
Monomorphism, 48 Rεcoisot , 325
contact, 237 Rεcomp , 326
isocontact, 237, 295 Reeb
isosymplectic, 260, 299 component, 96
symplectic, 260 foliation, 96, 234
Morin normal form, 119 vector field, 234
Morse Rεisot , 325
function, 113 RεLag , 325
index, 113 Removal of singularities, xvi
lemma, 113 r(
g , g), 344
Multifold, 61 Rhol , 282
Multisection, 62 Riemannian C r -manifolds, 341
Riemannian C r -metric, 341
Nash–Kuiper theorem, 342 Rimm , 70
Nijenhuis tensor, 225 Rimm-trans , 287
Ω(t, y), 304 Riso , 71
Op A, 11 Risocont , 237
Operator Risosymp , 280
formally invertible, 332 Risot , 324
pure differential, 332 Rk-mers , 318
RLag , 280, 324
Petals, 306 RLeg , 237
P i (z), 317 Rreal , 326
pJ , 223 Rsubm , 70
Pleating, 191 Rsub-isotr , 280
sharp, 187 Rsymp , 324
Polyhedra, 18 Rtang , 286
Positivity condition, 253 Rtrans , 286
pr , 11
pr0 , 11 Sanon-deg , 250
Primitive quadratic form, 343 Sasymp , 245
Primitive semi-Riemannian metric, 343 Scont , 246
Principal direction, 320 Section, 9
Principal subspace, 320 Σ-nonsingular, 71
Product of paths holonomic, 13
uniform, 308 transverse to a distribution, 91
weighted, 308 Sec X (r) , 13
Projectivization, 236 Semi-Riemannian metric, 341
prs , 12 Set
Pseudo-core, 51 m-complete, 55
Pseudo-isotopy, 213 ample, 304
pS , 223 stratified, 18
Pt,y , 304 Sharp
cusp, 171
r-jet, 10, 11 fold, 171
r-jet extension, 10, 12 pleating, 187
RA , 323 wrinkle, 174
Rclo , 71 Short formal solution, 305
Rcoisot , 324 Short path, 303
Rcomp , 326 Σ1r , 111
Index 363
Σi1 ,...,i , 110 totally real, 222

Σk , 18, 109 Surgery
Σk,j , 109 of a pair, 156
Singularity, 71 of singularities, 159
thin, 321 Symplectic
thin in the coordinate directions, 318 basis, 220
Smale’s sphere eversion, 46 cutting-off, 242
Solution, 72, 73 form, 219
r-extended, 72 manifold, 225, 226
formal, xv, 72, 73 orthogonal complement, 220
genuine, 72 structure, 219
Space subspace, 221
of r-jets, 10 twisting, 265
of complex structures, 222 vector field, 241
of symplectic structures, 220 vector space, 219
S+ Symplectic forms
cont , 246
S ⊥ω , 220 formally homotopic, 251
Scont , 246 homotopic, 251
S(L), 220 isotopic, 251
Snon-deg , 250 Symplectization, 236
Scont
+
, 246 Symplectomorphism, 226
Ssymp , 245 linear, 220
S(X), 224 θ-pair, 281
Ssymp , 245 Thom Transversality Theorem, 19
Stability theorems, 230, 239 Thom–Boardman singularities, 110
Standard contact structure on R2n+1 , Transverse subbundles, 56
232
Stem, 306 Vector bundle
Stratification, 18 complex, 224
Submanifold Hermitian, 224
(s, p), 225 symplectic, 224
almost complex, 225 Vector field
almost symplectic, 225 contact, 242
co-isotropic, 225 Hamiltonian, 241
co-real, 225 Liouville, 236
complex, 226 Reeb, 234
isotropic, 225 symplectic, 241
Lagrangian, 225
Wrinkle
Legendrian, 233
fibered, 128
subcritical, 233 immortal, 190
symplectic, 226 sharp, 174
totally real, 225 standard, 128
Submersion, 70 Wrinkled
Submersion relation, 70 embedding, 174
Subspace submersion, 135
(s, p), 221 fibered, 137
co-real, 222
coisotropic, 221 X (r) , 11
isotropic, 221
Lagrangian, 221
symplectic, 221
SELECTED PUBLISHED TITLES IN THIS SERIES
239 K. Cieliebak, Y. Eliashberg, and N. Mishachev, Introduction to the h-Principle,

Second Edition, 2024
238 Julio González-Dı́az, Ignacio Garcı́a-Jurado, and M. Gloria Fiestras-Janeiro,
An Introductory Course on Mathematical Game Theory and Applications, Second Edition,
2023
237 Michael Levitin, Dan Mangoubi, and Iosif Polterovich, Topics in Spectral
Geometry, 2023
235 Bennett Chow, Ricci Solitons in Low Dimensions, 2023
234 Andrea Ferretti, Homological Methods in Commutative Algebra, 2023
233 Andrea Ferretti, Commutative Algebra, 2023
232 Harry Dym, Linear Algebra in Action, Third Edition, 2023
231 Luı́s Barreira and Yakov Pesin, Introduction to Smooth Ergodic Theory, Second
Edition, 2023
230 Barbara Kaltenbacher and William Rundell, Inverse Problems for Fractional Partial
Differential Equations, 2023
229 Giovanni Leoni, A First Course in Fractional Sobolev Spaces, 2023
228 Henk Bruin, Topological and Ergodic Theory of Symbolic Dynamics, 2022
227 William M. Goldman, Geometric Structures on Manifolds, 2022
226 Milivoje Lukić, A First Course in Spectral Theory, 2022
225 Jacob Bedrossian and Vlad Vicol, The Mathematical Analysis of the Incompressible
Euler and Navier-Stokes Equations, 2022
224 Ben Krause, Discrete Analogues in Harmonic Analysis, 2022
223 Volodymyr Nekrashevych, Groups and Topological Dynamics, 2022
222 Michael Artin, Algebraic Geometry, 2022
221 David Damanik and Jake Fillman, One-Dimensional Ergodic Schrödinger Operators,
2022
220 Isaac Goldbring, Ultrafilters Throughout Mathematics, 2022
219 Michael Joswig, Essentials of Tropical Combinatorics, 2021
218 Riccardo Benedetti, Lectures on Differential Topology, 2021
217 Marius Crainic, Rui Loja Fernandes, and Ioan Mărcuţ, Lectures on Poisson
Geometry, 2021
216 Brian Osserman, A Concise Introduction to Algebraic Varieties, 2021
215 Tai-Ping Liu, Shock Waves, 2021
214 Ioannis Karatzas and Constantinos Kardaras, Portfolio Theory and Arbitrage, 2021
213 Hung Vinh Tran, Hamilton–Jacobi Equations, 2021
212 Marcelo Viana and José M. Espinar, Differential Equations, 2021
211 Mateusz Michalek and Bernd Sturmfels, Invitation to Nonlinear Algebra, 2021
210 Bruce E. Sagan, Combinatorics: The Art of Counting, 2020
209 Jessica S. Purcell, Hyperbolic Knot Theory, 2020
208 Vicente Muñoz, Ángel González-Prieto, and Juan Ángel Rojo, Geometry and
Topology of Manifolds, 2020
207 Dmitry N. Kozlov, Organized Collapse: An Introduction to Discrete Morse Theory, 2020
206 Ben Andrews, Bennett Chow, Christine Guenther, and Mat Langford, Extrinsic
Geometric Flows, 2020
205 Mikhail Shubin, Invitation to Partial Differential Equations, 2020
204 Sarah J. Witherspoon, Hochschild Cohomology for Algebras, 2019
For a complete list of titles in this series, visit the

AMS Bookstore at www.ams.org/bookstore/gsmseries/.
In differential geometry and topology one often deals with systems of partial
differential equations as well as partial differential inequalities that have infinitely
many solutions whatever boundary conditions are imposed. It was discovered in
the 1950s that the solvability of differential relations (i.e., equations and inequali-
ties) of this kind can often be reduced to a problem of a purely homotopy-theoretic
nature. One says in this case that the corresponding differential relation satis-
fies the h-principle. Two famous examples of the h-principle, the Nash–Kuiper
C1 -isometric embedding theory in Riemannian geometry and the Smale–Hirsch
immersion theory in differential topology, were later transformed by Gromov
into powerful general methods for establishing the h-principle.
The authors cover two main methods for proving the h-principle: holonomic
approximation and convex integration.The reader will find that, with a few notable
exceptions, most instances of the h-principle can be treated by the methods
considered here. A special emphasis is made on applications to symplectic and
contact geometry.
The present book is the first broadly accessible exposition of the theory and
its applications, making it an excellent text for a graduate course on geometric
methods for solving partial differential equations and inequalities. Geometers,
topologists, and analysts will also find much value in this very readable exposi-
tion of an important and remarkable topic.
This second edition of the book is significantly revised and expanded to almost
twice the original size. The most significant addition to the original book is the
new part devoted to the method of wrinkling and its applications. Several other
chapters (e.g., on multivalued holonomic approximation and foliations) are either
added or completely rewritten.
For additional information

and updates on this book, visit
www.ams.org/bookpages/gsm-239
GSM/239
www.ams.org

Introduction To The H-Principle, 2ed, Cieliebak, Eliashberg, GSM 239, 2024

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To The H-Principle, 2ed, Cieliebak, Eliashberg, GSM 239, 2024

Uploaded by

Copyright:

Available Formats

GRADUATE STUDIES

2020 Mathematics Subject Classiﬁcation. Primary 58Axx.

For additional information and updates on this book, visit

Library of Congress Cataloging-in-Publication Data

Preface to the Second Edition xiii

Preface to the First Edition xv

Part 1. Holonomic Approximation

Chapter 1. Jets and Holonomy 9

Chapter 2. Thom Transversality Theorem 17

Chapter 3. Holonomic Approximation 25

§3.4. Gluing Lemma 30

Part 2. Diﬀerential Relations and Gromov’s h-principle

§9.2. Smale–Hirsch h-principle 89

Part 3. Singularities and Wrinkling

§15.5. Proof of the main lemma 181

Chapter 16. Igusa Functions 193

Part 4. The Homotopy Principle in Symplectic Geometry

Chapter 17. Symplectic and Contact Basics 219

Chapter 18. Symplectic and Contact Structures on Open Manifolds 245

Chapter 19. Symplectic and Contact Structures on Closed Manifolds 251

Chapter 20. Embeddings into Symplectic and Contact Manifolds 259

Chapter 21. Microﬂexibility and Holonomic R-approximation 279

Part 5. Convex Integration

§27.2. Directed immersions into almost symplectic manifolds 324

A partial diﬀerential relation R is any condition imposed on the partial

The notion of h-principle (under the name w.h.e.-principle) ﬁrst appeared

There is an opinion that “the h-principle is the hardest part of Gromov’s

2 In the second edition it is the ﬁfth.

A. Immersions. A smooth map f : V → W of an n-dimensional manifold

A3. Consider now pairs of immersions (f0 , f1 ) : D 2 → R2 which coincide

A4. Let M1 , M2 be any two orientable connected 3-manifolds with non-

Figure 0.1. The immersions f and g.

immersions fj : Mj → R3 , j = 1, 2, such that f1 |∂M1 = f2 ◦ ϕ; see [El70]

B. Isometric C 1 -immersions. Is there a regular homotopy ft : S 2 → R3

C. Mappings with a prescribed Jacobian. Let Ω be an n-form on

D. Families of functions with simple singularities. An individual

E. Totally real, Lagrangian and ε-Lagrangian embeddings. Let T 2 =

for each x ∈ T 2 . For 0 < ε ≤ π2 , an embedding f is called ε-Lagrangian if

F. Free maps. A map T 2 → Rn is called free if the ﬁve vectors

G. Contact and Engel structures. A contact structure on a (2n + 1)-

Jets and Holonomy

1.1. Maps and sections

a family is supposed to be smooth in the cases when we need to diﬀerentiate

1.2. Coordinate deﬁnition of jets

is called the r-jet of f at x. Here f (s) consists of all partial derivatives

Exercise. Draw the 1-jet of the function f : R → R, f (x) = ax + b.

1.3. Invariant deﬁnition of jets

Following Gromov’s book [Gr86], we will use the notation Op A as a re-

Let v ∈ V . Two local sections f : Op v → X and g : Op v → X of the

of the local trivializations ϕ : U → Rn ×Rq which send the r-tangency classes

It is important to understand that the chain of inclusions

Exercise. Prove that the projection prr−1 : X (r) → X (r−1) carries a

If X is a trivial ﬁbration V × W → V , then the space of r-jets of sections

1.4. The space X (1)

If X = V × W and x = (v, w), then Vertx = Tw (v × W ), and moreover, we

J 1 (V, R) = T ∗ (V ) × R and J 1 (R, W ) = R × T (W ).

1.5. Holonomic sections of the jet space X (r)

The correspondence f → Jfr deﬁnes the derivation map

J r : Sec X → Sec X (r) .

Its one-to-one image J r (Sec X) coincides with the space

Hol X (r) ⊂ Sec X (r)

of holonomic sections, i.e., we have

A homotopy of holonomic sections of X (r) is called a holonomic homotopy.