GRADUATE STUDIES

I N M AT H E M AT I C S 175

Cartan for Beginners

Differential Geometry via
Moving Frames and Exterior
Differential Systems,
Second Edition

Thomas A. Ivey
Joseph M. Landsberg

American Mathematical Society

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175

Cartan for Beginners

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 GRADUATE STUDIES
I N M AT H E M AT I C S 175

Cartan for Beginners

Differential Geometry via Moving
Frames and Exterior Differential
Systems,
Second Edition

Thomas A. Ivey
Joseph M. Landsberg

American Mathematical Society

Providence, Rhode Island

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 EDITORIAL COMMITTEE
Dan Abramovich
Daniel S. Freed (Chair)
Gigliola Staffilani
Jeff A. Viaclovsky

2010 Mathematics Subject Classification. Primary 35N10, 37K35, 53A05, 53A20, 53A30,
53A55, 53C10, 53C15, 58A15, 58A17.

For additional information and updates on this book, visit

www.ams.org/bookpages/gsm-175

Library of Congress Cataloging-in-Publication Data

Names: Ivey, Thomas A. (Thomas Andrew), 1963– | Landsberg, J. M.
Title: Cartan for beginners : differential geometry via moving frames and exterior differential
systems / Thomas A. Ivey, Joseph M. Landsberg.
Description: Second edition. | Providence, Rhode Island : American Mathematical Society, [2016]
| Series: Graduate studies in mathematics ; volume 175 | Includes bibliographical references and
index.
Identifiers: LCCN 2016025601 | ISBN 9781470409869 (alk. paper)
Subjects: LCSH: Geometry, Differential. | Exterior differential systems. | AMS: Partial differential
equations – Overdetermined systems – Overdetermined systems with variable coefficients. msc |
Dynamical systems and ergodic theory – Infinite-dimensional Hamiltonian systems – Lie-Bäcklund
and other transformations. msc | Differential geometry – Classical differential geometry – Surfaces
in Euclidean space. msc | Differential geometry – Classical differential geometry – Projective
differential geometry. msc | Differential geometry – Classical differential geometry – Conformal
differential geometry. msc | Differential geometry – Classical differential geometry – Differential
invariants (local theory), geometric objects. msc | Differential geometry – Global differential
geometry – G-structures. msc | Differential geometry – Global differential geometry – General
geometric structures on manifolds (almost complex, almost product structures, etc.). msc
Classification: LCC QA641 .I89 2016 | DDC 516.3/6–dc23 LC record available at https://lccn.
loc.gov/2016025601

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. Permissions to reuse
portions of AMS publication content are handled by Copyright Clearance Center’s RightsLink
service. For more information, please visit: http://www.ams.org/rightslink.
Send requests for translation rights and licensed reprints to [email protected].
Excluded from these provisions is material for which the author holds copyright. In such cases,
requests for permission to reuse or reprint material should be addressed directly to the author(s).
Copyright ownership is indicated on the copyright page, or on the lower right-hand corner of the
first page of each article within proceedings volumes.

c 2016 by the American Mathematical Society. All rights reserved.
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 http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 21 20 19 18 17 16
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Contents

Introduction xi
Preface to the Second Edition xi
Preface to the First Edition xiv
Chapter 1. Moving Frames and Exterior Differential Systems 1
§1.1. Geometry of surfaces in E3 in coordinates 2
§1.2. Differential equations in coordinates 5
§1.3. Introduction to differential equations without coordinates 8
§1.4. Introduction to geometry without coordinates:
curves in E2 13
§1.5. Submanifolds of homogeneous spaces 16
§1.6. The Maurer-Cartan form 18
§1.7. Plane curves in other geometries 22
§1.8. Curves in E3 25
§1.9. Grassmannians 28
§1.10. Exterior differential systems and jet spaces 31
Chapter 2. Euclidean Geometry 39
§2.1. Gauss and mean curvature via frames 40
§2.2. Calculation of H and K for some examples 43
§2.3. Darboux frames and applications 46
§2.4. What do H and K tell us? 47
§2.5. Invariants for n-dimensional submanifolds of En+s 48
§2.6. Intrinsic and extrinsic geometry 52

v
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 vi Contents

§2.7. Curves on surfaces 55

§2.8. The Gauss-Bonnet and Poincaré-Hopf theorems 57
Chapter 3. Riemannian Geometry 65
§3.1. Covariant derivatives and the fundamental lemma of
Riemannian geometry 65
§3.2. Nonorthonormal frames and a geometric interpretation of
mean curvature 73
§3.3. The Riemann curvature tensor 77
§3.4. Space forms: the sphere and hyperbolic space 80
§3.5. Representation theory for Riemannian geometry 83
§3.6. Infinitesimal symmetries: Killing vector fields 85
§3.7. Homogeneous Riemannian manifolds 89
§3.8. The Laplacian 92
Chapter 4. Projective Geometry I: Basic Definitions and Examples 95
§4.1. Frames and the projective second fundamental form 96
§4.2. Algebraic varieties 101
§4.3. Varieties with degenerate Gauss mappings 109
Chapter 5. Cartan-Kähler I: Linear Algebra and Constant-Coefficient
Homogeneous Systems 115
§5.1. Tableaux 116
§5.2. First example 120
§5.3. Second example 122
§5.4. Third example 125
§5.5. The general case 126
§5.6. The characteristic variety of a tableau 129
Chapter 6. Cartan-Kähler II:
The Cartan Algorithm for Linear Pfaffian Systems 135
§6.1. Linear Pfaffian systems 135
§6.2. First example 137
§6.3. Second example: constant coefficient homogeneous systems 138
§6.4. The local isometric embedding problem 141
§6.5. The Cartan algorithm formalized:
tableau, torsion and prolongation 146
§6.6. Summary of Cartan’s algorithm for linear Pfaffian systems 149
§6.7. Additional remarks on the theory 151
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Contents vii

§6.8. Examples 154

§6.9. Functions whose Hessians commute,
with remarks on singular solutions 161
§6.10. The Cartan-Janet Isometric Embedding Theorem 164
§6.11. Isometric embeddings of space forms (mostly flat ones) 166
§6.12. Calibrated submanifolds 169
Chapter 7. Applications to PDE 175
§7.1. Symmetries and Cauchy characteristics 176
§7.2. Second-order PDE and Monge characteristics 184
§7.3. Derived systems and the method of Darboux 188
§7.4. Monge-Ampère systems and Weingarten surfaces 195
§7.5. Integrable extensions and Bäcklund transformations 204
Chapter 8. Cartan-Kähler III: The General Case 215
§8.1. Integral elements and polar spaces 216
§8.2. Example: triply orthogonal systems 223
§8.3. Statement and proof of Cartan-Kähler 226
§8.4. Cartan’s Test 229
§8.5. More examples of Cartan’s Test 232
Chapter 9. Geometric Structures and Connections 241
§9.1. G-structures 241
§9.2. Connections on FG and differential invariants
of G-structures 248
§9.3. Overview of the Cartan algorithm 253
§9.4. How to differentiate sections of vector bundles 254
§9.5. Induced vector bundles and connections 257
§9.6. Killing vector fields for G-structures 260
§9.7. Holonomy 262
§9.8. Extended example: path geometry 271
Chapter 10. Superposition for Darboux-Integrable Systems 285
§10.1. Decomposability 286
§10.2. Integrability 289
§10.3. Coframe adaptations 294
§10.4. Some results on group actions 299
§10.5. The superposition formula 300
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 viii Contents

Chapter 11. Conformal Differential Geometry 307

§11.1. Conformal geometry via Riemannian geometry 308
§11.2. Conformal differential geometry as a G-structure 311
§11.3. Conformal Killing vector fields 319
§11.4. Conformal densities and the Laplacian 322
§11.5. Einstein manifolds in a conformal class and the tractor
bundle 326
Chapter 12. Projective Geometry II: Moving Frames and Subvarieties
of Projective Space 331
§12.1. The Fubini cubic and higher order differential invariants 332
§12.2. Fundamental forms of Veronese, Grassmann, and Segre
varieties 336
§12.3. Ruled and uniruled varieties 339
§12.4. Dual varieties 341
§12.5. Secant and tangential varieties 346
§12.6. Cominuscule varieties and their differential invariants 350
§12.7. Higher-order Fubini forms 356
§12.8. Varieties with vanishing Fubini cubic 363
§12.9. Associated varieties 365
§12.10. More on varieties with degenerate Gauss maps 367
§12.11. Rank restriction theorems 369
§12.12. Local study of smooth varieties with degenerate
tangential varieties 372
§12.13. Generalized Monge systems 376
§12.14. Complete intersections 378
Appendix A. Linear Algebra and Representation Theory 381
§A.1. Dual spaces and tensor products 381
§A.2. Matrix Lie groups 386
§A.3. Complex vector spaces and complex structures 388
§A.4. Lie algebras 390
§A.5. Division algebras and the simple group G2 394
§A.6. A smidgen of representation theory 397
§A.7. Clifford algebras and spin groups 400
Appendix B. Differential Forms 405
§B.1. Differential forms and vector fields 405
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Contents ix

§B.2. Three definitions of the exterior derivative 407

§B.3. Basic and semi-basic forms 409
Appendix C. Complex Structures and Complex Manifolds 411
§C.1. Complex manifolds 411
§C.2. The Cauchy-Riemann equations 415
Appendix D. Initial Value Problems and the Cauchy-Kowalevski
Theorem 419
§D.1. Initial value problems 419
§D.2. The Cauchy-Kowalevski Theorem 420
§D.3. Generalizations 423
Hints and Answers to Selected Exercises 425
Bibliography 437
Index 445

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Introduction

Preface to the Second Edition

We were very happy with the reception the first edition received after its
appearance in 2003. In the years since, numerous readers have contacted
us with questions, corrections, and comments on the exposition in the first
edition. The book has been used in graduate classes at Texas A& M seven
times by the second author, who would like to thank the students for their
substantial input, especially the students from the 2012 and 2015 classes
where preliminary versions of Chapters 3 and 11 were used. Thanks to our
readers and students we have implemented many changes to improve and
correct the exposition. While many people have helped, we owe a special
debt to Colleen Robles, Matt Stackpole, Pieter Eendebak and Peter Vassil-
iou, who gave us numerous detailed comments, and to Robert Bryant and
Michael Eastwood for their help developing the new material. We are also
grateful to the AMS editorial staff, in particular Ed Dunne, Sergei Gelfand,
and Christine Thivierge, for their help and patience.
One feature of this edition is that we have attempted to make more con-
nections with the larger subject of differential geometry, mentioning major
theorems and open questions with references to the literature. There are
also three chapters of essentially new material:
Chapter 3 vastly expands the few pages on Riemannian geometry in
the second chapter of the first edition. Notable here is the emphasis on a
representation-theoretic perspective for the Riemann curvature tensor and
its covariant derivative. There is also a proof of Killing’s theorem describing

xi
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 xii Introduction

the space of Killing vector fields on a Riemannian manifold and a discus-

sion of homogeneous Riemannian manifolds, the latter following unpublished
notes of R. Bryant.
Chapter 10 is devoted to the latest development in the study of Darboux-
integrable exterior differential systems, namely the work of Anderson, Fels
and Vassiliou [6] on superposition formulas for such systems. These struc-
tures, which enable one to write down explicit solutions, are based on the
action of the system’s Vessiot group, which is generated by a set of vec-
tor fields which define its Vessiot algebra. After discussing the generalized
definition of Darboux integrability formulated by these authors, we explain
the construction of the Vessiot algebra by a sequence of delicate coframe
adaptations. These adaptations are illustrated using a running example,
and other recent applications of Darboux integrability in diverse settings,
including Toda lattice systems and wave maps, are detailed in the exercises.
In Chapter 11 we discuss conformal differential geometry. A central goal
of this chapter is to take steps to bring the EDS and parabolic geometry
perspectives together by discussing a particular geometry. It should enable
readers familiar with the parabolic geometry perspective to place this book
in better context and enable readers from an EDS perspective to start read-
ing the parabolic geometry literature (see [33] for an excellent introduction).
Conformal geometry is a beautiful subject and we only touch on several top-
ics: conformal Killing fields, the conformal Laplacian, and Gover’s work on
Einstein metrics in a given conformal class [78]. This chapter was heavily
influenced by the expositions [29] and [56], as well as conversations with M.
Eastwood, who we thank for his help.
In addition, we have re-arranged some of the material from the first
edition. Grassmannians are introduced earlier in section 1.9. The first
edition’s Chapter 3, on projective geometry, has been split into two chapters:
Chapter 4, which contains material that every differential geometer should
learn and material needed later in the book, and Chapter 12, which contains
more specialized and advanced topics. The chapter on G-structures (now
Chapter 9) has also been substantially re-written for clarity, and a new
section on G-Killing vector fields, based on conversations with R. Bryant,
has been added.
With the re-arrangement of material, the interdependence of chapters in
the second edition is described by the following diagram:

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Preface to the Second Edition xiii

Dependence of Chapters
1

2 5 7.1 4

3 6
7.2–7.5 12

8
9
10

11

Suggested uses of this book:

• a year-long graduate course covering moving frames and exterior
differential systems (chapters 1–9);
• a one-semester course on exterior differential systems and applica-
tions to partial differential equations (chapters 1 and 7–8);
• a one-semester course on the use of moving frames in algebraic
geometry (chapters 4 and 12, preceded by part of chapter 1);
• a one-semester beginning graduate course on differential geometry
(chapters 1, 2, 3 and 9)
• a year-long differential geometry course based on chapters 1, 2, 3
and 11, interspersed with an introduction to differentiable mani-
folds from e.g., [174, Vol. I].

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 xiv Introduction

Preface to the First Edition

In this book, we use moving frames and exterior differential systems to study
geometry and partial differential equations. These ideas originated about
a century ago in the works of several mathematicians, including Gaston
Darboux, Edouard Goursat and, most importantly, Elie Cartan. Over the
years these techniques have been refined and extended; major contributors
to the subject are mentioned below, under “Further Reading”.
The book has the following features: It concisely covers the classical
geometry of surfaces and basic Riemannian geometry in the language of
moving frames. It includes results from projective differential geometry that
update and expand the classic paper [85] of Griffiths and Harris. It provides
an elementary introduction to the machinery of exterior differential systems
(EDS), and an introduction to the basics of G-structures and the general
theory of connections. Classical and recent geometric applications of these
techniques are discussed throughout the text.
This book is intended to be used as a textbook for a graduate-level
course; there are numerous exercises throughout. It is suitable for a one-
year course, although it has more material than can be covered in a year, and
parts of it are suitable for a one-semester course (see the end of this preface
for some suggestions). The intended audience is both graduate students who
have some familiarity with classical differential geometry and differentiable
manifolds, and experts in areas such as PDE and algebraic geometry who
want to learn how moving frame and EDS techniques apply to their fields.
In addition to the geometric applications presented here, EDS techniques
are also applied in CR geometry (see, e.g., [120]), robotics, and control
theory (see [68, 69, 159]). This book prepares the reader for such areas,
as well as for more advanced texts on exterior differential systems, such as
[27], and papers on recent advances in the theory, such as [71, 144].

Overview. Each section begins with geometric examples and problems.

Techniques and definitions are introduced when they become useful to help
solve the geometric questions under discussion. We generally keep the pre-
sentation elementary, although advanced topics are interspersed throughout
the text.

In Chapter 1 we introduce moving frames via the geometry of curves in

the Euclidean plane E2 . We define the Maurer-Cartan form of a Lie group G
and explain its use in the study of submanifolds of G-homogeneous spaces.
We give additional examples, including the equivalence of holomorphic map-
pings up to fractional linear transformation, where the machinery leads one
naturally to the Schwarzian derivative.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Preface to the First Edition xv

We define exterior differential systems and jet spaces, and explain how
to rephrase any system of partial differential equations as an EDS using jets.
We state and prove the Frobenius system, leading up to it via an elementary
example of an overdetermined system of PDE.

In Chapter 2 we cover traditional material—the geometry of surfaces in

three-dimensional Euclidean space, submanifolds of higher-dimensional Eu-
clidean space, and the rudiments of Riemannian geometry—all using moving
frames. Our emphasis is on local geometry, although we include standard
global theorems such as the rigidity of the sphere and the Gauss-Bonnet
Theorem. Our presentation emphasizes finding and interpreting differential
invariants to enable the reader to use the same techniques in other settings.

We begin Chapter 31 with a discussion of Grassmannians and the Plücker

embedding.2 We present some well-known material (e.g., Fubini’s theorem
on the rigidity of the quadric) which is not readily available in other text-
books. We present several recent results, including the Zak and Landman
theorems on the dual defect, and results of the second author on complete
intersections, osculating hypersurfaces, uniruled varieties and varieties cov-
ered by lines. We keep the use of terminology and results from algebraic
geometry to a minimum, but we believe we have included enough so that
algebraic geometers will find this chapter useful.

Chapter 43 begins our multi-chapter discussion of the Cartan algorithm

and Cartan-Kähler Theorem. In this chapter we study constant coefficient
homogeneous systems of PDE and the linear algebra associated to the corre-
sponding exterior differential systems. We define tableaux and involutivity
of tableaux. One way to understand the Cartan-Kähler Theorem is as fol-
lows: given a system of PDE, if the linear algebra at the infinitesimal level
“works out right” (in a way explained precisely in the chapter), then exis-
tence of solutions follows.

In Chapter 54 we present the Cartan algorithm for linear Pfaffian sys-

tems, a very large class of exterior differential systems that includes systems
of PDE rephrased as exterior differential systems. We give numerous ex-
amples, including many from Cartan’s classic treatise [40], as well as the
isometric immersion problem, problems related to calibrated submanifolds,
and an example motivated by the variation of the Hodge structure.

1 Now Chapters 4 and 12

2 Now in Chapter 1
3 Now Chapter 5
4 Now Chapter 6
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 xvi Introduction

In Chapter 75 we take a detour to discuss the classical theory of charac-

teristics, Darboux’s method for solving PDE, and Monge-Ampère equations
in modern language. By studying the exterior differential systems associ-
ated to such equations, we recover the sine-Gordon representation of pseu-
dospherical surfaces, the Weierstrass representation of minimal surfaces, and
the one-parameter family of noncongruent isometric deformations of a sur-
face of constant mean curvature. We also discuss integrable extensions and
Bäcklund transformations of exterior differential systems, and the relation-
ship between such transformations and Darboux integrability.

In Chapter 66 we present the general version of the Cartan-Kähler The-

orem. Doing so involves a detailed study of the integral elements of an EDS.
In particular, we arrive at the notion of a Kähler-regular flag of integral ele-
ments, which may be understood as the analogue of a sequence of well-posed
Cauchy problems. After proving both the Cartan-Kähler Theorem and Car-
tan’s test for regularity, we apply them to several examples of non-Pfaffian
systems arising in submanifold geometry.

Finally, in Chapter 87 we give an introduction to geometric structures

(G-structures) and connections. We arrive at these notions at a leisurely
pace, in order to develop the intuition as to why one needs them. Rather
than attempt to describe the theory in complete generality, we present one
extended example, path geometry in the plane, to give the reader an idea
of the general theory. We conclude with a discussion of some recent gener-
alizations of G-structures and their applications.

There are four appendices, covering background material for the main
part of the book: linear algebra and rudiments of representation theory,
differential forms and vector fields, complex and almost complex manifolds,
and a brief discussion of initial value problems and the Cauchy-Kowalevski
Theorem, of which the Cartan-Kähler Theorem is a generalization.

Layout. All theorems, propositions, remarks, examples, etc., are numbered

together within each section; for example, Theorem 1.3.1 is the second num-
bered item in section 1.3. Equations are numbered sequentially within each
chapter. We have included hints for selected exercises, those marked with
the symbol  at the end, which is meant to be suggestive of a life preserver.

5 Now Chapter 8
6 Now Chapter 7
7 Now Chapter 9
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Preface to the First Edition xvii

Further reading on EDS. To our knowledge, there are only a small num-
ber of textbooks on exterior differential systems. The first is Cartan’s clas-
sic text [40], which has an extraordinarily beautiful collection of examples,
some of which are reproduced here. We learned the subject from our teacher
Bryant and the book by Bryant, Chern, Griffiths, Gardner and Goldschmidt
[27], which is an elaboration of an earlier monograph [26], and is at a more
advanced level than this book. One text at a comparable level to this book,
but more formal in approach, is [190]. The monograph [86], which is cen-
tered around the isometric embedding problem, is similar in spirit but covers
less material. The memoir [189] is dedicated to extending the Cartan-Kähler
Theorem to the C ∞ setting for hyperbolic systems, but contains an exposi-
tion of the general theory. There is also a monograph by Kähler [109] and
lectures by Kuranishi [119], as well the survey articles [82, 110]. Some dis-
cussion of the theory may be found in the differential geometry texts [174]
and [177].
We give references for other topics discussed in the book in the text.

History and Acknowledgements. This book started out about a decade

ago. We thought we would write up notes from Robert Bryant’s Tuesday
night seminar, held in 1988–89 while we were graduate students, as well
as some notes on exterior differential systems which would be more intro-
ductory than [27]. The seminar material is contained in §9.8 and parts of
Chapter 7. Chapter 2 is influenced by the many standard texts on the sub-
ject, especially [55] and [174], while Chapter 4 is influenced by the paper
[85]. Several examples in Chapter 6 and Chapter 8 are from [40], and the
examples of Darboux’s method in Chapter 7 are from [76]. In each case,
specific attributions are given in the text. Chapter 8 follows Chapter III of
[27] with some variations. In particular, to our knowledge, Lemmas 8.1.10
and 8.1.13 are original. The presentation in §9.7 is influenced by [15], [114]
and unpublished lectures of Bryant.
The first author has given graduate courses based on the material in
Chapters 6 and 7 at the University of California, San Diego and at Case
Western Reserve University. The second author has given year-long gradu-
ate courses using Chapters 1, 2, 4, 5, and 8 at the University of Pennsylvania
and Université de Toulouse III, and a one-semester course based on Chap-
ters 1, 2, 4 and 5 at Columbia University. He has also taught one-semester
undergraduate courses using Chapters 1 and 2 and the discussion of con-
nections in Chapter 8 (supplemented by [173] and [174] for background
material) at Toulouse and at Georgia Institute of Technology, as well as
one-semester graduate courses on projective geometry from Chapters 1 and
3 (supplemented by some material from algebraic geometry), at Toulouse,
Georgia Tech. and the University of Trieste. He also gave more advanced
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 xviii Introduction

lectures based on Chapter 3 at Seoul National University, which were pub-

lished as [129] and became a precursor to Chapter 3. Preliminary versions
of Chapters 5 and 8 respectively appeared in [126, 125].
We would like to thank the students in the above classes for their feed-
back. We also thank Megan Dillon, Phillipe Eyssidieux, Daniel Fox, Sung-
Eun Koh, Emilia Mezzetti, Joseph Montgomery, Giorgio Ottaviani, Jens
Piontkowski, Margaret Symington, Magdalena Toda, Sung-Ho Wang and
Peter Vassiliou for comments on the earlier drafts of this book, and An-
nette Rohrs for help with the figures. The staff of the publications division
of the AMS—in particular, Ralph Sizer, Tom Kacvinsky, and our editor,
Ed Dunne—were of tremendous help in pulling the book together. We are
grateful to our teacher Robert Bryant for introducing us to the subject.
Lastly, this project would not have been possible without the support and
patience of our families.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/01

Chapter 1

Moving Frames and

Exterior Differential
Systems

In this chapter we motivate the use of differential forms to study problems in

geometry and partial differential equations. We begin with familiar material:
the Gauss and mean curvature of surfaces in Euclidean space E3 in §1.1,
and Picard’s Theorem for local existence of solutions of ordinary differential
equations in §1.2. We continue in §1.2 with a discussion of a simple system
of partial differential equations, and then in §1.3 rephrase it in terms of
differential forms, which facilitates interpreting its solutions geometrically.
We also state the Frobenius Theorem.
In §1.4, we review curves in E2 in the language of moving frames. We
generalize this example in §§1.5–1.6, describing how one studies subman-
ifolds of homogeneous spaces using moving frames, and introducing the
Maurer-Cartan form. We give two examples of the geometry of curves in ho-
mogeneous spaces: classifying holomorphic mappings of the complex plane
under fractional linear transformations in §1.7, and classifying curves in E3
under Euclidean motions (i.e., rotations and translations) in §1.8. We also
include exercises on plane curves in other geometries. In §1.9, we discuss
Grassmannians, one of the most important classes of manifolds in all of ge-
ometry. We conclude, in §1.10, with the definitions of exterior differential
systems and integral manifolds. We prove the Frobenius Theorem, give a few
basic examples of exterior differential systems, and explain how to express
a system of partial differential equations as an exterior differential system
using jet bundles.

1
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2 1. Moving Frames and Exterior Differential Systems

Throughout this book we use the summation convention: unless other-

wise indicated, summation is implied whenever repeated indices occur up
and down in an expression.

1.1. Geometry of surfaces in E3 in coordinates

Let E3 denote Euclidean three-space, i.e., the affine space R3 equipped with
its standard inner product.
Given two smooth oriented surfaces S, S  ⊂ E3 , when are they “equiva-
lent”? For the moment, we will say that two surfaces are (locally) equivalent
if there exist a rotation and translation taking (an open subset of) S onto
(an open subset of) S  , so that the unit normals coincide.

Figure 1. Are these two surfaces equivalent?

It would be impractical and not illuminating to try to test all possible

motions to see if one of them maps S onto S  . Instead, we will work as
follows:
Fix one surface S and a point p ∈ S. We will use the Euclidean motions
to put S (and its unit normal) into a normalized position in space with
respect to p. Then any other surface S  will be locally equivalent to S at p
if there is a point p ∈ S  such that the pair (S  , p ) can be put into the same
normalized position as (S, p).
The implicit function theorem implies that there always exist coordinates
such that S is given locally by a graph z = f (x, y). To obtain a normalized
position for our surface S, first translate so that p = (0, 0, 0), and then use a
rotation to make Tp S the xy-plane, i.e., so that zx (0, 0) = zy (0, 0) = 0, and
the unit normal at p points in the negative z direction. We will call such
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.1. Geometry of surfaces in E3 in coordinates 3

coordinates adapted to p. At this point we have used up all our freedom of

motion except for a rotation in the xy-plane.

If coordinates are adapted to p and we expand f (x, y) in a Taylor series

centered at the origin, then functions of the coefficients of the series that
are invariant under this rotation are differential invariants.
In this context, a (Euclidean) differential invariant of S at p is a function
I of the coefficients of the Taylor series for f at p, with the property that,
if we perform a Euclidean change of coordinates
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
x̃ x a
⎝ỹ ⎠ = A ⎝y ⎠ + ⎝ b ⎠ ,
z̃ z c
where A is a rotation matrix and a, b, c are arbitrary constants, after which
S is expressed as a graph z̃ = f˜(x̃, ỹ) near p, then I has the same value
when computed using the Taylor coefficients of f˜ at p. Clearly a necessary
condition for (S, p) to be locally equivalent to (S  , p ) is that the values
of differential invariants of S at p match the values of the corresponding
invariants of S  at p .
For example, consider the Hessian of z = z(x, y) at p:
 
zxx zyx 
(1.1) Hessp = .
zxy zyy p
Assume we have adapted coordinates to p. If we rotate in the xy-plane, the
Hessian gets conjugated by the rotation matrix. The quantities
K0 = det(Hessp ) = (zxx zyy − zxy
2
) |p ,
(1.2)
H0 = 12 trace(Hessp ) = 12 (zxx + zyy ) |p
are differential invariants because the determinant and trace of a matrix are
unchanged by conjugation by a rotation matrix. Thus, if we are given two
surfaces S, S  and we normalize them both at respective points p and p as
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4 1. Moving Frames and Exterior Differential Systems

above, a necessary condition for there to be an orientation-preserving rigid

motion taking p to p such that the Taylor expansions for the two surfaces
at the point p coincide is that K0 (S) = K0 (S  ) and H0 (S) = H0 (S  ).
The formulas (1.2) are only valid at one point, and only after the surface
has been put in normalized position relative to that point. To calculate K
and H as functions on S it would be too much work to move each point
to the origin and arrange its tangent plane to be horizontal. But it is
possible to adjust the formulas to account for tilted tangent planes (see §3.2).
One then obtains the following functions, which are differential invariants
under Euclidean motions of surfaces that are locally described as graphs
z = z(x, y):
zxx zyy − zxy
2
K(x, y) = ,
(1 + zx2 + zy2 )2
(1.3)
1 (1 + zy2 )zxx − 2zx zy zxy + (1 + zx2 )zyy
H(x, y) = 3 ,
2 (1 + zx2 + zy2 ) 2
respectively giving the Gauss and mean curvature of S at p = (x, y, z(x, y)).
Exercise 1.1.1: By locally describing each surface as a graph, calculate the
Gauss and mean curvature functions for a sphere of radius R, a cylinder of
radius r (e.g., x2 + y 2 = r2 ) and the smooth points of the cone x2 + y 2 = z 2 .

Once one has found invariants for a given submanifold geometry, one
may ask questions about submanifolds with special invariants. For surfaces
in E3 , one might ask which surfaces have K constant or H constant. These
can be treated as questions about solutions to certain partial differential
equations (PDE). For example, from (1.3) we see that surfaces with K = 1
are locally given by solutions to the PDE
(1.4) zxx zyy − zxy
2
= (1 + zx2 + zy2 )2 .
We will soon free ourselves of coordinates and use moving frames and dif-
ferential forms. As a provisional definition, a moving frame is a smoothly
varying basis of the tangent space to E3 defined at each point of our sur-
face. In general, using moving frames one can obtain formulas valid at every
point, analogous to coordinate formulas valid at just one preferred point. In
the present context, the Gauss and mean curvatures will be described at all
points by expressions like (1.2) rather than (1.3); see §2.1.
Another reason to use moving frames is that the method gives a uni-
form procedure for dealing with diverse geometric settings. Even if one is
originally only interested in Euclidean geometry, other geometries arise nat-
urally. For example, consider the warp of a surface, which is defined to be
(k1 − k2 )2 , where the kj are the eigenvalues of (1.1). It turns out that this
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.2. Differential equations in coordinates 5

quantity is invariant under a larger change of coordinates than the Eucli-

dean group, namely conformal changes of coordinates, and thus it is more
relevant to study the warp in the context of conformal geometry.
Regardless of how unfamiliar a geometry initially appears, the method of
moving frames provides an algorithm to find differential invariants. Thus we
will have a single method for dealing with conformal, Hermitian, projective,
and other geometries. Because it is familiar, we will often use the geometry
of surfaces in E3 as an example, but the reader should keep in mind that
the beauty of the method is its wide range of applicability. As for the use
of differential forms, we shall see that when we express a system of PDE as
an exterior differential system, the geometric features of the system—i.e.,
those which are independent of coordinates—will become transparent.

1.2. Differential equations in coordinates

The first questions one might ask when confronted with a system of differ-
ential equations are: Are there any solutions? If so, how many?
In the case of a single ordinary differential equation (ODE), here is the
answer:
Theorem 1.2.1 (Picard1 ). Let f (x, u) : R2 → R be a function with f
and fu continuous. Then for all (x0 , u0 ) ∈ R2 , there exist an open interval
I  x0 and a unique function u(x) defined on I, satisfying u(x0 ) = u0 and
the differential equation
du
(1.5) = f (x, u).
dx
Moreover, any other solution of this initial value problem must coincide with
this solution on I.
In other words, for a given ODE there exists a solution defined near x0 ,
and this solution is unique given the choice of a constant u0 . Thus for an
ODE for one function of one variable, we say that solutions depend on one
constant. More generally, Picard’s Theorem applies to systems of n first-
order ODE’s involving n unknowns, where solutions depend on n constants.
The graph in R2 of any solution to (1.5) is tangent at each point to
∂ ∂
the vector field X = ∂x + f (x, u) ∂u . This indicates how determined ODE
systems generalize to the setting of differentiable manifolds (see Appendix
B). If M is a manifold and X is a vector field on M , then a solution to the
system defined by X is an immersed curve c : I → M such that c (t) = Xc(t)
for all t ∈ I. (This is also referred to as an integral curve of X.) Away
from singular points, one is guaranteed existence of local solutions to such
systems and can even take the solution curves as coordinate curves:
1 See, e.g., [172], p. 423
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6 1. Moving Frames and Exterior Differential Systems

Theorem 1.2.2 (Flowbox coordinates2 ). Let M be an m-dimensional C ∞

manifold, let p ∈ M , and let X ∈ Γ(T M ) be a smooth vector field which
is nonzero at p. Then there exists a local coordinate system (x1 , . . . , xm ),
defined in a neighborhood U of p, such that ∂x∂ 1 = X.
Consequently, there exists an open set V ⊂ U × R on which we may
define the flow of X, φ : V → M , by requiring that for any point (q, t) ∈ V ,
∂
∂t φ(q, t) = X|φ(q,t) . The flow is given in flowbox coordinates by
(x1 , . . . , xm , t) → (x1 + t, x2 , . . . , xm ).
With systems of PDE, it becomes difficult to determine the appropriate
initial data for a given system (see Appendix D for examples). We now
examine a simple PDE system, first in coordinates, and then later (in §6.2)
using differential forms.
Example 1.2.3. Consider the system for u(x, y) given by
ux = A(x, y, u),
(1.6)
uy = B(x, y, u),
where A, B are given smooth functions. Since (1.6) specifies both partial
derivatives of u, at any given point p = (x, y, u) ∈ R3 the tangent plane to
the graph of a solution passing through p is uniquely determined.
In this way, (1.6) defines a smoothly varying field of two-planes on R3 ,
just as the ODE (1.5) defines a field of one-planes (i.e., a line field) on R2 .
For (1.5), Picard’s Theorem guarantees that the one-planes “fit together”
to form a solution curve through any given point. For (1.6), existence of
solutions amounts to whether or not the two-planes “fit together”.
We can attempt to solve (1.6) in a neighborhood of (0, 0) by solving a
succession of ODE’s. Namely, if we set y = 0 and u(0, 0) = u0 , Picard’s
Theorem implies that there exists a unique function ũ(x) satisfying
dũ
(1.7) = A(x, 0, ũ), ũ(0) = u0 .
dx
After solving (1.7), hold x fixed and use Picard’s Theorem again on the
initial value problem
du
(1.8) = B(x, y, u), u(x, 0) = ũ(x).
dy
This determines a function u(x, y) on some neighborhood of (0, 0). The
problem is that this function may not satisfy our original equation.
Whether or not (1.8) actually gives a solution to (1.6) depends on
whether or not the equations (1.6) are “compatible” as differential equa-
tions. For smooth solutions to a system of PDE, compatibility conditions
2 See, e.g., [174] vol. I, p. 205
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.2. Differential equations in coordinates 7

arise because mixed partials must commute, i.e., (ux )y = (uy )x . In our
example,
∂ ∂u
(ux )y = A(x, y, u) = Ay (x, y, u) + Au (x, y, u) = Ay + BAu ,
∂y ∂y
(uy )x = Bx + ABu ,
so setting (ux )y = (uy )x reveals a “hidden equation”, the compatibility
condition
(1.9) Ay + BAu = Bx + ABu .
We will prove in §1.10 that the commuting of second-order partials in this
case implies that all higher-order mixed partials commute as well, so that
there are no further hidden equations. In other words, if (1.9) is an identity
in x, y, u, then solving the ODE’s (1.7) and (1.8) in succession gives a solution
to (1.6), and solutions depend on one constant.
Exercise 1.2.4: Show that, if (1.9) is an identity, then one gets the same
solution by first solving for ũ(y) = u(0, y).

If (1.9) is not an identity, there are several possibilities. If u appears in

(1.9), then it gives an equation which every solution to (1.6) must satisfy.
Fixing a point p = (x0 , y0 , u0 ) where (1.9) holds and the u-partial of (1.9) is
nonzero, locally solve (1.9) for a function u(x,y). Although this is the only
possible solution satisfying the initial condition, it still might not satisfy
(1.6), in which case there is no solution through p.
If u does not appear in (1.9), then it gives a relation between x and y,
and there is no solution defined on an open set around (x0 , y0 ).
Remark 1.2.5. For more complicated systems of PDE, it is not as easy to
determine if all mixed partials commute. The Cartan-Kähler Theorem (see
Chapters 5 and 7) will provide an algorithm which tells us when to stop
checking compatibilities.
Exercises 1.2.6:
1. Consider this special case of Example 1.2.3:
ux = A(x, y),
uy = B(x, y),
where A and B satisfy A(0, 0) = B(0, 0) = 0. Verify that solving the initial
value problems (1.7)–(1.8) with x0 = y0 = 0 gives
x y
(1.10) u(x, y) = u0 + A(s, y)ds + B(x, t)dt.
0 0
Under what condition does this function u satisfy (1.6)? Verify that the
resulting condition is equivalent to (1.9) in this special case.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8 1. Moving Frames and Exterior Differential Systems

2. Rewrite (1.10) as a line integral involving the 1-form

ω := A(x, y)dx + B(x, y)dy,
and determine the condition on ω which ensures that the integral is inde-
pendent of path.
3. Determine the space of solutions to (1.6) in the following special cases:
(a) A = − ux , B = − uy .
(b) A = B = ux .
(c) A = − ux , B = y.

1.3. Introduction to differential equations without

coordinates
Example 1.2.3 revisited. Instead of working on R2 × R with coordi-
nates (x, y, u), we work on the larger space R2 × R × R2 with coordinates
(x, y, u, p, q), which we denote J 1 (R2 , R), or J 1 for short. This space, called
the space of 1-jets of mappings from R2 to R, is given additional structure
and generalized in §1.10.
Let u : U → R be a smooth function defined on an open set U ⊂ R2 .
We associate to u the surface in J 1 given by
(1.11) u = u(x, y), p = ux (x, y), q = uy (x, y),
which we will refer to as the lift or 1-graph of u. The graph of u is the
projection of the lift (1.11) in J 1 to R2 × R.
We will eventually work on J 1 without reference to coordinates. As a
step in that direction, consider the differential forms
θ := du − pdx − qdy, Ω := dx ∧ dy
defined on J 1 . Suppose i : S → J 1 is a surface such that i∗ Ω = 0 at each
point of S. Since dx, dy are linearly independent 1-forms on S, we may use
x, y as coordinates on S, and the surface may be expressed as
u = u(x, y), p = p(x, y), q = q(x, y).
Suppose i∗ θ = 0. Then
i∗ du = pdx + qdy.
On the other hand, since u restricted to S is a function of x and y, we have
du = ux dx + uy dy.
Because dx, dy are independent on S, these two equations imply that p = ux
and q = uy on S. Thus, surfaces i : S → J 1 such that i∗ θ = 0 and i∗ Ω is
nonvanishing are 1-graphs of functions u : U → R.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.3. Introduction to differential equations without coordinates 9

Now consider the 3-fold j : Σ → J 1 defined by the equations

p = A(x, y, u), q = B(x, y, u).
Let i : S → Σ be a surface such that = 0 and i∗ Ω is nonvanishing. Then i∗ θ
the projection of S to R2 × R is the graph of a solution to (1.6). Moreover,
all solutions to (1.6) are the projections of such surfaces, by taking S as the
lift of the solution.
Thus we have a correspondence
solutions to (1.6) ⇔ surfaces i : S → Σ such that i∗ θ ≡ 0 and i∗ Ω = 0.
On such surfaces, we also have i∗ dθ ≡ 0, but
dθ = −dp ∧ dx − dq ∧ dy,
j ∗ dθ = −(Ax dx + Ay dy + Au du) ∧ dx − (Bx dx + By dy + Bu du) ∧ dy,
i∗ dθ = (Ay − Bx + Au B − Bu A)i∗ Ω.
(To obtain the second line we use the defining equations of Σ and to obtain
the third line we use i∗ (du) = Adx + Bdy.) Because i∗ Ω = 0, the equation
(1.12) Ay − Bx + Au B − Bu A = 0
must hold on S. This is precisely the same as the condition (1.9) obtained
by checking that mixed partials commute.
If (1.12) does not hold identically on Σ, then it gives another equation
which must hold for any solution. But since dim Σ = 3, in that case (1.12)
already describes a surface in Σ. If there is any solution surface S, it must
be an open subset of the surface in Σ given by (1.12). This surface will only
be a solution if θ pulls back to be zero on it. If (1.12) is an identity on
Σ, then we may use the Frobenius Theorem (see below) to conclude that
through any point of Σ there is a unique solution S (constructed, as in §1.2,
by solving a succession of ODE’s). In this sense, (1.12) implies that all
higher partial derivatives commute.
We have now recovered our observations from §1.2.

The general game plan for treating a system of PDE as an exterior

differential system (EDS) will be as follows:
One begins with a “universal space” (J 1 in the above example) where the
various partial derivatives are represented by independent variables. Then
one restricts to the subset Σ of the universal space defined by the system of
PDE by considering it as a set of equations among independent variables.
Solutions to the PDE correspond to submanifolds of Σ on which the vari-
ables representing what we want to be partial derivatives actually are partial
derivatives. These submanifolds are characterized by the vanishing of certain
differential forms.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10 1. Moving Frames and Exterior Differential Systems

These remarks will be explained in detail in §1.10.

Picard’s Theorem in the language of EDS. On R2 with coordinates (x, u),
consider θ = du − f (x, u)dx. Then there is a one-to-one correspondence
between solutions of the ODE (1.5) and curves c : R → R2 such that c∗ (θ) =
0 and c∗ (dx) is nonvanishing.
More generally, the flowbox coordinate theorem 1.2.2 implies:
Theorem 1.3.1. Let M be a C ∞ manifold of dimension m, and let
θ1 , . . . , θm−1 ∈ Ω1 (M ) be pointwise linearly independent in some open neigh-
borhood U ⊂ M . Then through p ∈ U there exists a curve c : R → U , unique
up to reparametrization, such that c∗ (θa ) = 0 for 1 ≤ a ≤ m − 1.
(For a proof, see [174].)
Later in this section we will discuss the Frobenius Theorem, which gener-
alizes this result, with curves replaced by higher-dimensional submanifolds.
On these submanifolds, not only do the 1-forms θa vanish, but so do their
exterior derivatives, and the wedge products of the θa with any other differ-
ential forms on M . This naturally leads one to consider the ideal generated
by these forms.

Differential ideals. Recall that Ω∗ (M ) denotes the space of smooth dif-

ferential forms on a manifold M . This is a graded algebra under the wedge
product.

Definition 1.3.2. An element in Ω∗ (M ) is homogeneous if all its terms have

the same degree.

Definition 1.3.3. A subspace I ⊂ Ω∗ (M ) is an algebraic ideal if it is a

direct sum of subspaces I k ⊂ Ωk (M ) and it is closed under wedge product
with arbitrary differential forms.

Notation 1.3.4. Each homogeneous component I k is a ring over C ∞ (M ).

It follows that the values of forms in I k at a point x ∈ M span a vector
subspace Ixk ⊂ Λk Tx∗ M . We let I k = x∈M Ixk ; similarly, if J is an ideal,
then J k ⊂ Λk Tx M denotes the union of subspaces spanned by forms in J k .
We will often write I to stand for I 1 .

In many instances each I k will be a vector subbundle of Λk T ∗ M , and

when it is not, we will usually restrict the ideal to subsets of M where
each I k has constant rank. Then I k is the space of smooth sections of I k .
We will rarely deal with ideals where I 0 = {0}, i.e., ideals which include
nontrivial functions; again, we will usually restrict to submanifolds on which
such functions vanish.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.3. Introduction to differential equations without coordinates 11

Definition 1.3.5. An algebraic ideal I is a differential ideal if it is closed

under exterior differentiation, i.e., the exterior derivative of any form in I
is also in I.
Notation 1.3.6. If ω 1 , . . . , ω k are 1-forms on M , then {ω 1 , . . . , ω k } ⊆
Ω1 (M ) denotes the linear span of the forms, i.e. for all x ∈ M ,
{ω 1 , . . . , ω k }x = span{ωx1 , . . . , ωxk } ⊂ Tx∗ M.

Let ω 1 , . . . , ω k be forms of arbitrary degree. Then

{ω 1 , . . . , ω k }alg ⊆ Ω∗ (M )
will denote the ideal generated algebraically by ω 1 , . . . , ω k , i.e., the smallest
algebraic ideal containing the generators. Similarly,
{ω 1 , . . . , ω k }diff ⊆ Ω∗ (M )
will denote the ideal generated algebraically by ω 1 , . . . , ω k and their exterior
derivatives, i.e., the smallest differential ideal containing the generators. For
example,
{ω 1 , ω 2 }alg = {α ∧ ω 1 + β ∧ ω 2 | α, β ∈ Ω∗ (M )},
{ω 1 , ω 2 }diff = {α ∧ ω 1 + β ∧ ω 2 + γ ∧ dω 1 + δ ∧ dω 2 | α, β, γ, δ ∈ Ω∗ (M )}.

Given a subbundle I ⊂ T ∗ M , we often define a differential ideal I gen-

erated by sections of I. In other words, given a local basis ω 1 , . . . , ω k of
sections of I defined on U ⊂ M , then
I|U = {ω 1 , . . . , ω k }diff .
Such a differential ideal is called a Pfaffian system of rank k (i.e., the rank
is the same as that of the vector bundle I).

The Frobenius Theorem. In §1.10 we will prove the following result,

which is a generalization, both of Theorem 1.3.1 and of the asserted existence
of solutions to Example 1.2.3 when (1.9) holds, to an existence theorem for
certain systems of PDE:
Theorem 1.3.7 (Frobenius Theorem, first version). Let Σ be a C ∞ man-
ifold of dimension m, and let θ1 , . . . , θm−n ∈ Ω1 (Σ) be pointwise linearly
independent. If there exist 1-forms αji ∈ Ω1 (Σ) such that dθj = αij ∧θi for all
j, then through each p ∈ Σ there exists a maximal connected n-dimensional
manifold i : N → Σ such that i∗ (θj ) = 0 for 1 ≤ j ≤ m − n. This manifold
is unique, in the sense that any other such connected submanifold through
p is a subset of i(N ).
When the hypotheses of Theorem 1.3.7 hold for the θ1 , . . . , θm−n ∈
Ω1 (Σ), we will call the ideal they generate Frobenius.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12 1. Moving Frames and Exterior Differential Systems

In order to motivate our study of exterior differential systems, we re-

word the Frobenius Theorem more geometrically as follows: Let Σ be an
m-dimensional manifold such that through each point x ∈ Σ there is an
n-dimensional subspace Ex ⊂ Tx Σ which varies smoothly with x. (Such
a structure is called a distribution.) We consider the problem of finding
submanifolds X ⊂ Σ such that Tx X = Ex for all x ∈ X.
Let Ex ⊥ be the annihilator of Ex inside the dual space Tx∗ Σ, and let
θxa , 1 ≤ a ≤ m − n, be a basis of Ex ⊥ . We may choose the θxa to vary
smoothly, to obtain m − n linearly independent forms θa ∈ Ω1 (Σ). Let
I = {θ1 , . . . , θm−n }diff . The submanifolds X tangent to the distribution E
are exactly the n-dimensional submanifolds i : N → Σ such that i∗ α = 0
for all α ∈ I. Call such a submanifold an integral manifold of I.
To find integral manifolds, we already know that if one exists, its tangent
space at any point x ∈ Σ that it contains is already uniquely determined,
namely it is Ex . The question is whether these n-planes can be “fitted
together” to obtain an n-dimensional submanifold. This information is con-
tained in the derivatives of the θa ’s, which indicate how the n-planes “move”
infinitesimally.
If we are to have i∗ θa = 0, we must also have d(i∗ θa ) = i∗ (dθa ) = 0. If
there is to be an integral manifold through x, or even an n-plane Ex ⊂ Tx Σ
on which α|Ex = 0, ∀α ∈ I, the equations i∗ (dθa ) = 0 cannot impose any
additional conditions, i.e., we must have dθa |Ex = 0 because we already have
a unique n-plane at each point x ∈ Σ. To recap, for all a we must have
(1.13) dθa = α1a ∧ θ1 + . . . + αm−n
a
∧ θm−n
for some αba ∈ Ω1 (Σ), because the forms θxa span Ex ⊥ .
Notation 1.3.8. Suppose I is an ideal and φ and ψ are k-forms. Then we
write φ ≡ ψ mod I if φ = ψ + β for some β ∈ I.

Now (1.13) may be restated as

(1.14) dθa ≡ 0 mod {θ1 , . . . , θm−n }alg .
The Frobenius Theorem states that this necessary condition is also sufficient:
Theorem 1.3.9 (Frobenius Theorem, second version). Let I be a differ-
ential ideal generated by the linearly independent 1-forms θ1 , . . . , θm−n on
an m-fold Σ, i.e., I = {θ1 , . . . , θm−n }diff . Suppose I is also generated al-
gebraically by θ1 , . . . , θm−n , i.e., I = {θ1 , . . . , θm−n }alg . Then through any
p ∈ Σ there exists a unique n-dimensional integral manifold of I. In fact,
in a sufficiently small neighborhood of p there exists a coordinate system
y 1 , . . . , y m such that I is generated by dy 1 , . . . , dy m−n .
We postpone the proof until §1.10.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.4. Introduction to geometry without coordinates: curves in E2 13

Definition 1.3.10. We will say a subbundle I ⊂ T ∗ Σ is Frobenius if the

ideal generated algebraically by sections of I is also a differential ideal. We
will say a distribution Δ ⊂ T Σ is Frobenius if Δ⊥ ⊂ T ∗ Σ is Frobenius.
Equivalently (see Exercise 1.3.11.2 below), Δ is Frobenius if ∀X, Y ∈ Γ(Δ),
[X, Y ] ∈ Γ(Δ), where [X, Y ] is the Lie bracket.

If {θa } fails to be Frobenius, not all hope is lost for an n-dimensional

integral manifold, but we must restrict ourselves to the subset j : Σ → Σ
on which (1.14) holds, and see if there are n-dimensional integral manifolds
of the ideal generated by j ∗ θa on Σ . (This was what we did in Exercise
1.2.6.1.)
Exercises 1.3.11:
1. Which of the following algebraic ideals are Frobenius?

I1 = {dx1 , x2 dx3 + dx4 }alg ,

I2 = {dx1 , x1 dx3 + dx4 }alg .

2. Show that the conditions for the span of differential forms being Fro-
benius is equivalent to the Frobenius condition for the corresponding dis-
tribution, i.e., Γ(Δ) ⊂ Γ(T Σ) satisfies [Γ(Δ), Γ(Δ)] ⊆ Γ(Δ) if and only if
Δ⊥ ⊂ T ∗ Σ satisfies dθ ≡ 0 mod Δ⊥ for all θ ∈ Γ(Δ⊥ ).
3. On R3 let θ = Adx + Bdy + Cdz, where A = A(x, y, z), etc. Assume
the differential ideal generated by θ is Frobenius, and explain how to find a
function f (x, y, z) such that the differential systems {θ}diff and {df }diff are
equivalent.

1.4. Introduction to geometry without coordinates:

curves in E2
We will return to our study of surfaces in E3 in Chapter 2. To see how to
use moving frames to obtain invariants, we begin with a simpler problem.
Let E2 denote the oriented Euclidean plane. Given two parametrized
curves c1 , c2 : R → E2 , we ask two questions: When does there exist a
Euclidean motion A : E2 → E2 (i.e., a rotation and translation) such that
A(c1 (R)) = c2 (R) as sets? And, when do there exist a Euclidean motion
A : E2 → E2 and a constant c such that A(c1 (t)) = c2 (t + c) for all t?
Instead of using coordinates at a point, we will use an adapted frame, i.e.,
for each t we take a basis of Tc(t) E2 that is “adapted” to Euclidean geometry.
This geometry is induced by the group of Euclidean motions—the changes
of coordinates of E2 preserving the inner product and orientation—which
we will denote by ASO(2).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 14 1. Moving Frames and Exterior Differential Systems

Figure 2. Are these two curves equivalent?

In more detail, the group ASO(2) consists of transformations of the form

 1  1  1
x t x
(1.15) → 2 + R 2 ,
x2 t x
where R ∈ SO(2) is a rotation matrix. The group ASO(2) can be repre-
sented as a matrix Lie group by writing
  
 1 0
(1.16) ASO(2) = M ∈ GL(3, R) M =  , t ∈ R2 , R ∈ SO(2) .
t R
Then its action on E2 is given by x → M x, where we represent points in E2
by x = t 1 x1 x2 .
We may define a mapping from ASO(2) to E2 by
   
1 0 x1
(1.17) → x = ,
x R x2
which takes each group element to the image of the origin under the trans-
formation (1.15). The fiber of this map over every point is a left coset of
SO(2) ⊂ ASO(2), so E2 , as a manifold, is the quotient ASO(2)/SO(2). Fur-
thermore, ASO(2) may be identified with the bundle of oriented orthonormal
bases of E2 by identifying the columns of the rotation matrix R = (e1 , e2 )
with an oriented orthonormal basis of Tx E2 , where x is the basepoint given
by (1.17). (Here we use the fact that for a vector space V , we may identify
V with Tx V for any x ∈ V .)
Returning to the curve c(t), we choose an oriented orthonormal basis
of Tc(t) E2 as follows: A natural element of Tc(t) E2 is c (t), but this may
not be of unit length. So, we take e1 (t) = c (t)/|c (t)|, and this choice also
determines e2 (t). To do this we must assume that the curve is regular:
Definition 1.4.1. A curve c(t) is said to be regular if c (t) never vanishes.
More generally, a map f : M → N between differentiable manifolds is
regular if df is everywhere defined and of rank equal to dim M .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.4. Introduction to geometry without coordinates: curves in E2 15

What have we done? We have constructed a map to the Lie group

ASO(2) as follows:

C : R → ASO(2),
 
1 0
t → .
c(t) (e1 (t), e2 (t))

We will obtain differential invariants of our curve by differentiating this

mapping, and taking combinations of the derivatives that are invariant under
Euclidean changes of coordinates.
Consider v(t) = |c (t)|, called the speed of the curve. It is invariant under
Euclidean motions and thus is a differential invariant. However, it is only
an invariant of the mapping, not of the image curve (see Exercise 1.4.2.2).
The speed measures how much (intrinsic) distance is being distorted under
the mapping c.
Consider de de1
dt . We must have dt = λ(t)e2 (t) for some function λ(t)
1

because |e1 (t)| ≡ 1 (see Exercise 1.4.2.1 below). Thus λ(t) is a differential
invariant, but it again depends on the parametrization of the curve. To
determine an invariant of the image alone, we let c̃(t) = c(φ(t)) be another
parametrization of the same curve for some regular φ : R → R. We calculate
ṽ(t)
that λ̃(t) = v(φ(t)) λ(φ(t)), so the quantity κ(t) = λ(t)
v(t) is unchanged under
reparametrization. This κ(t), called the curvature of the curve, measures
how much c is infinitesimally moving away from its tangent line at c(t).
The above discussion implies that a necessary condition for two curves
c, c̃ to have congruent images (i.e., differing by a Euclidean motion) is that
there exists a diffeomorphism ψ : R → R such that κ(t) = κ̃(ψ(t)). It will
follow from Corollary 1.6.15 that the images of curves are locally classified
up to congruence by their curvature functions, and that parametrized curves
are locally classified by κ, v.
Exercises 1.4.2:
1. Let V be a vector space with a nondegenerate inner product , . Let
v(t) be a curve in V such that F (t) := v(t), v(t) is constant. Show that
v  (t), v(t) = 0 for all t. Show the converse is also true.
t
2. Suppose that c is regular. Let s(t) = 0 |c (τ )|dτ and consider c paramet-
rized by s instead of t. Since s gives the length of the image of c : [0, s] → E2 ,
s is called an arclength parameter. Show that in this preferred parametriza-
de1
tion κ(s) = e2 , ds .
3. Show that κ(t) is constant if and only if the curve is an open subset of
a line (if κ = 0) or circle of radius κ1 .
4. Let c(t) = (x(t), y(t)) be given in coordinates. Calculate κ(t) in terms
of x(t), y(t) and their derivatives.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 16 1. Moving Frames and Exterior Differential Systems

5. Calculate the function κ(t) for an ellipse. Characterize the points on the
ellipse where the maximum and minimum values of κ(t) occur.
6. Can κ(t) be unbounded if c(t) is the graph of a polynomial?

Exercise 1.4.3 (Osculating circles):

(a) Calculate the equation of a circle passing through three points in the
plane.
(b) Calculate the equation of a circle passing through two points in the
plane and having a given tangent line at one of the points.
Parts (a) and (b) may be skipped; the exercise proper starts here:
(c) Show that for any curve c ⊂ E2 , at each point x ∈ c one can define an
osculating circle by taking the limit of the circle through the three points
c(t), c(t1 ), c(t2 ) as t1 , t2 → t. (A line is defined to be a circle of infinite
radius.) 
(d) Show that one gets the same circle if one takes the limit as t → t1 of
the circle through c(t), c(t1 ) that has tangent line at c(t) parallel to c (t).
(e) Show that the radius of the osculating circle is 1/κ(t).
(f) Show that if κ(t) is monotone, then the osculating circles are nested. 

1.5. Submanifolds of homogeneous spaces

Using the machinery we develop in this section and §1.6, we will answer the
questions about curves in E2 posed at the beginning of §1.4. The quotient
E2 = ASO(2)/SO(2) is an example of a homogeneous space, and our answers
will follow from a general study of classifying the maps into homogeneous
spaces.
Definition 1.5.1. Let G be a Lie group, H a closed Lie subgroup, and G/H
the set of left cosets of H. Then G/H is naturally a differentiable manifold
with the induced differentiable structure coming from the quotient map π
sending each g ∈ G to its coset (see [94], Theorem II.3.2). The space G/H
is called a homogeneous space.
Definition 1.5.2 (Left and right actions). Let X be a manifold, G a group,
and let σ : g → σ(g) associate a diffeomorphism σ(g) of X to each g ∈ G.
Then σ is called a left action if σ(a) ◦ σ(b) = σ(ab), or a right action if
σ(a) ◦ σ(b) = σ(ba). When G is a Lie group we require additionally that the
map G×X → X, given by (g, x) → σ(g)x, be smooth and that σ(g) : X → X
is a diffeomorphism. For a given x ∈ X, the set of g such that σ(g) fixes x
is the isotropy of x, and is a subgroup of G.

For example, the action of G on itself by left-multiplication is a left

action, while left-multiplication by g −1 is a right action.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.5. Submanifolds of homogeneous spaces 17

A homogeneous space G/H has a natural left G-action on it; the sub-
group stabilizing the coset [e] of the identity element is H, and the stabilizer
of any point is conjugate to H. Conversely, a manifold X is a homogeneous
space if it admits a smooth transitive action by a Lie group G. If H is the
isotropy group of a point x0 ∈ X, then X  G/H, and x0 corresponds to
[e] ∈ G/H. (See [94, 174] for additional facts about homogeneous spaces.)

In the spirit of Klein’s Erlanger Programm (see [93, 112] for historical
accounts), we will consider G as the group of motions of G/H. We will
study the geometry of submanifolds M ⊂ G/H, where two submanifolds
M, M  ⊂ G/H will be considered equivalent if there exists a g ∈ G such that
g(M ) = M  .
To determine necessary conditions for equivalence we will find differential
invariants as we did in §1.1 and §1.4. (Note that we need to specify whether
we are interested in invariants of a mapping or just of the image.) After
finding invariants, we will then interpret them as we did in the exercises in
§1.4.
We begin to derive invariants for maps f : M → G/H by constructing
lifts F : M → G as we did for curves in E2 .
Definition 1.5.3. A lift of a mapping f : M → G/H is defined to be a map
F : M → G such that the following diagram commutes:
G
F 
?
M -
f
G/H

Given a lift F of f , any other lift F̃ : M → G must be of the form

(1.18) F̃ (x) = F (x)a(x)
for some map a : M → H.

By associating the value of the lift F (x) with the action of F (x)−1 on
G/H, we may think of choosing a lift to G as analogous to putting a point
p on a surface in a normalized position, as we did in §1.1.
Given f : M → G/H, we will choose lifts adapted to the infinitesimal
geometry. To explain what this statement means, we first remark that
in the situations we will be dealing with, the fiber at x ∈ G/H of the
fibration π : G → G/H admits the interpretation of being a subset of the
space of frames, i.e., bases of Tx G/H. Since H fixes the point [e] ∈ G/H,
its infinitesimal action on tangent vectors gives a representation ρ : H →
GL(T[e] G/H), called the isotropy representation [187]. Now fix a reference
basis (v1 , . . . , vn ) of T[e] G/H. We identify the H-orbit of this basis (induced
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 18 1. Moving Frames and Exterior Differential Systems

from the action of H on g = Te G) with π −1 ([e]). Similarly, at other points

x ∈ G/H we have a group conjugate to H acting on Tx G/H.
Thus, a choice of lift may be considered as a choice of framing of G/H
along M , and we will make choices that reflect the geometry of M . For
example, if dim M = n, we may require the first n basis vectors of Tx G/H
to be tangent to M . In the above example of curves in E2 , we also normalized
the length of the first basis vector e1 to be one.
Once a unique lift is determined, differentiating that lift will provide
differential invariants. This is because we can classify maps into G, up to
equivalence under left multiplication, using the Maurer-Cartan form.

1.6. The Maurer-Cartan form

If you need to brush up on matrix Lie groups and Lie algebras, this would
be a good time to consult §A.2 and §A.4.
Given a Lie group G, we let g denote its Lie algebra, which may be
identified with Te G or with the space of left-invariant vector fields.
Exercise 1.6.1: Show that we have an identification T[e] G/H  g/h as
vector spaces and H-modules.

Definition 1.6.2. A differential form α ∈ Ωk (G) is left-invariant if for all

a ∈ G, we have L∗a (αg ) = αa−1 g , where La : G → G is the diffeomorphism
g → ag. (We similarly define left-invariant vector-valued differential forms
as well as left-invariant vector fields and k-vector fields.)

Proposition/Definition 1.6.3. On a Lie group G, there exists a unique

left-invariant g-valued 1-form ω ∈ Ω1 (G, g), called the Maurer-Cartan form
of G, such that ωe : Te G  g → g is the identity map.

Proof. Let X1 , . . . , Xn be a basis of Te G and let α1 , . . . , αn be the dual

basis of Te∗ G, so ωe = αi ⊗Xi . Then by left-invariance ωa = L∗a−1 (αi )⊗Xi ,
where we now think of the Xi as elements of the fixed vector space g. It
remains to verify that ωba = L∗b−1 ωa :

ωba =L∗(ba)−1 (αi )⊗Xi = (La−1 Lb−1 )∗ (αi )⊗Xi

= L∗b−1 L∗a−1 (αi )⊗Xi = (Lb−1 )∗ ωa .


Remark 1.6.4. The above proof shows more generally that a left-invariant
vector-valued form α ∈ Ωk (G, V ) is uniquely determined by αg for any
g ∈ G. In this way, the set of left-invariant k-forms may be identified with
Λk g∗ .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.6. The Maurer-Cartan form 19

Example 1.6.5. Let G ⊆ GL(N, R) be a matrix Lie group with Lie algebra
g ⊂ gl(N, R), and let g = (gji ) be the matrix-valued function which embeds
G into the vector space MN ×N of N × N matrices with real entries, with
differential dga : Ta G → Tg(a) MN ×N  MN ×N . The Maurer-Cartan form of
G is ωa = L∗a−1 (ωe ) = g(a)−1 dga , which is often written as
ω = g −1 dg.
In more detail, writing g = (gji ), we consider each gji : G → R as a func-
tion on G, and thus have the function (gji ) : G → MN ×N , so (ωa )ij =
(g(a)−1 )ik (dga )kj .

Exercise 1.6.6: Show that g −1 dg is indeed the Maurer-Cartan form of G.

In particular, g −1 dg is g-valued.
Example 1.6.7. Consider G = SO(2) ⊂ GL(2, R). Picking a constant a,
we may locally parametrize SO(2) by
 
cos θ − sin θ
g(θ) = , θ ∈ (a, a + 2π).
sin θ cos θ
Then  
−1 0 −dθ
ω=g dg = .
dθ 0

Definition 1.6.8. Let ω = (ωki ) and η = (ηki ) be matrices whose entries are
elements of a vector space V , so that ω, η ∈ V ⊗Mn×n . Define their matrix
wedge product ω ∧ η ∈ Λ2 V ⊗Mn×n by
(ω ∧ η)ij := ωki ∧ ηjk .

More generally, for ω ∈ Λk V ⊗ Mn×n , η ∈ Λj V ⊗ Mn×n the same formula

yields ω ∧ η ∈ Λk+j V ⊗Mn×n .

One thing that makes the Maurer-Cartan form ω especially useful to

work with is that its exterior derivative may be computed algebraically as
follows: If G is a matrix Lie group, then
dω = d(g −1 ) ∧ dg.
To compute d(g −1 ), consider the identity matrix e = (δji ) as a constant map
G → Mn×n and note that it is the product of two nonconstant Mn×n -valued
functions:
0 = d(e) = d(g −1 g) = d(g −1 )g + g −1 dg.
So, d(g −1 ) = −g −1 (dg)g −1 and
dω = −g −1 (dg)g −1 ∧ dg = −(g −1 dg) ∧ (g −1 dg) = −ω ∧ ω.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 20 1. Moving Frames and Exterior Differential Systems

Summary 1.6.9. On a matrix Lie group G, the Maurer-Cartan form ω

defined by ω = g −1 dg is a left-invariant g-valued 1-form and satisfies the
Maurer-Cartan equation
(1.19) dω = −ω ∧ ω.
Definition 1.6.10. If ω, θ are two g-valued 1-forms, define the g-valued
2-form [ω, θ] by
[ω, θ](X, Y ) = [ω(X), θ(Y )] − [ω(Y ), θ(X)].
The Maurer-Cartan equation holds on an abstract Lie group G in the fol-
lowing form:
1
(1.20) dω = − [ω, ω];
2
see [187, Prop. 3.12]
Remark 1.6.11. When F : M → G is a lift of a map f : M → G/H, and
G is a matrix Lie group, the change in the pullback of the Maurer-Cartan
form resulting from a change of lift (1.18) is
(1.21) F̃ ∗ (ω) = a−1 F ∗ (ω)a + a−1 da.
For an abstract Lie group, the analogous formula is
F̃ ∗ (ω) = Ad(a−1 ) ◦ F ∗ (ω) + a∗ ω.

As mentioned above, the Maurer-Cartan form will be our key to classi-

fying maps into homogeneous spaces of G. We first show how it classifies
maps into G:
Theorem 1.6.12 (Cartan). Let G be a matrix Lie group with Maurer-
Cartan form ω and let M be a manifold. Let f : M → G be an immersion,
so αf := f ∗ ω is a g-valued 1-form on M satisfying dα = −α ∧ α. Then a
map f˜ : M → G is congruent to f , in the sense that f˜ = La ◦ f for some
fixed a ∈ G, if and only if αf˜ = αf .
We prove more generally that:
Theorem 1.6.13 (Cartan). Let G be a matrix Lie group with Lie algebra
g and Maurer-Cartan form ω. Let M be a manifold on which there exists a
g-valued 1-form φ satisfying dφ = −φ ∧ φ. Then for any point x ∈ M there
exist a neighborhood U of x and a map f : U → G such that f ∗ ω = φ.
Moreover, any two such maps f1 , f2 must satisfy f1 = La ◦ f2 for some fixed
a ∈ G.

Proof. This is a good opportunity to use the Frobenius Theorem.

On Σ = M n × G, let π, ρ denote the projections to each factor and let
θ = π ∗ φ − ρ∗ ω. Write θ = (θji ), and let I ⊂ T ∗ Σ be the subbundle spanned
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.6. The Maurer-Cartan form 21

by the forms θji . Submanifolds of dimension n to which these forms pull back
to be zero are graphs of maps f : M → G such that φ = f ∗ ω. We check the
conditions given in the Frobenius Theorem. Calculating derivatives (and
omitting the pullback notation), we have
dθ = −φ ∧ φ + ω ∧ ω
= −φ ∧ φ + (θ − φ) ∧ (θ − φ)
≡ 0 mod I.
Thus, the system is Frobenius and by Theorem 1.3.7 there is a unique con-
nected maximal n-dimensional integral manifold through any (x, g) ∈ Σ.
Suppose f1 , f2 are two different solutions. Say f1 (x) = g. Let a =
gf2 (x)−1 . Then the graph of f = La ◦ f2 passes through (x, g) and f ∗ ω = φ.
By uniqueness, it follows that f1 = La ◦ f2 . 
Remark 1.6.14. If we assume in Theorem 1.6.13 that M is connected and
simply connected, then the desired map f may be extended to all of M
[187].

We may apply Theorem 1.6.13 to classify curves in E2 . In this case, the

pullback of the Maurer-Cartan form of ASO(2) ⊂ GL(3, R) under the lift
constructed in §1.4 takes the simple form
⎛ ⎞
0 0 0
(1.22) F ∗ ω = ⎝dt 0 −κ dt⎠ ,
0 κ dt 0
where t is an arclength parameter.
Corollary 1.6.15. For curves c, c̃ ⊂ E2 , if κ(t) = κ̃(t + a) for some constant
a, then c, c̃ are congruent.
Exercises 1.6.16:
1. Let CO(2) be the matrix Lie group parametrized by
 
t cos θ −t sin θ
g(θ, t) = , t ∈ (0, ∞), θ ∈ R.
t sin θ t cos θ
Explicitly compute the Maurer-Cartan form and verify the Maurer-Cartan
equation for CO(2).
2. Verify (1.21).
3. Verify (1.22) and prove Corollary 1.6.15.
4. Show that (1.20) coincides with (1.19) when G is a matrix Lie group.
5. Let g be a vector space with basis XB , 1 ≤ B ≤ dim g, and a multi-
plication given by [XA , XB ] = cC
AB XC on the basis and extended linearly.
Determine necessary and sufficient conditions on the constants cA BC for g,
with this Lie bracket, to be a Lie algebra.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 22 1. Moving Frames and Exterior Differential Systems

6. On a Lie group G with Maurer-Cartan form ω, show that

dωe (X, Y ) = −[X, Y ].
Conclude that, assuming G is connected, dω = 0 if and only if G is abelian.
7. On a matrix Lie group G with Lie algebra g, let {eA } be a basis of g and
write the Maurer-Cartan form as ω = ω A eA . (Note that each 1-form ω A is
left-invariant.) Write dω A = − 12 C̃BC
A ω B ∧ ω C , where C̃ A = −C̃ A . Show
BC CB
A are constants, and determine the set of equations
that the coefficients C̃BC
that these constants must satisfy because d2 = 0. Relate these equations to
your answer to problem 5.

1.7. Plane curves in other geometries

Equivalence of holomorphic mappings under fractional linear
transformations. Here is an example of a study of curves in a less famil-
iar homogeneous space, the complex projective line CP1 . To find differential
invariants in such situations, we generally seek a uniquely defined lift that
renders the pullback of the Maurer-Cartan form as simple as possible. Then,
after finding differential invariants, we interpret them.
Definition 1.7.1. A fractional linear transformation (FLT) is a map CP1 →
CP1 given in terms of homogeneous coordinates t[z1 , z2 ] by
     
z1 a b z1
→ , with ad − bc = 1.
z2 c d z2

The group of FLT’s is P SL(2, C) := SL(2, C)/{± Id}, which acts tran-
sitively on CP1 . Thus, CP1 is a homogeneous space P SL(2, C)/P , where
 
a b
P = ∈ P SL(2, C)
0 a−1
is the stabilizer of t[1, 0]. Although P SL(2, C), as presented, is not a matrix
Lie group, we may avoid problems by working locally as follows:
If Δ ⊂ C ⊂ CP1 is a domain, then P SL(2, C) acts on maps f : Δ → C
by
af + b
. f →
cf + d
(Since we will be working locally, there is no harm in considering f as a map
to C; then to think of f as a map to CP1 , write it as t[f, 1].)
Suppose f, g : Δ → C are two holomorphic maps with nonzero first
derivatives. When are these locally equivalent via a fractional linear trans-
formation, i.e., when does g = A ◦ f for some FLT A? (One can ask the
same question in the real category for analytic maps f, g : (0, 1) → RP1 ,
and the answer will be the same.) Note that in this example the target is of
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.7. Plane curves in other geometries 23

the same dimension as the source of the mapping, so we cannot expect an

analogue of curvature, but there will be an analogue of speed.
The coordinate approach to getting invariants would be to use an FLT to
normalize the map at some point z0 , say by requiring f (z0 ) = 0, f  (z0 ) = 1
and f  (z0 ) = 0. Since this is exactly the extent of normalization that
P SL(2, C) can achieve, then f  (z0 ) must be an invariant. Of course, this
is valid only at the point z0 .
Instead we construct a lift to P SL(2, C), which we will treat as SL(2, C)
in order to work with a matrix Lie group. As a first try, let
 
f −1
F = .
1 0
Since the projection to CP1 is the equivalence class of the first column, any
other lift F̃ of f must be of the form F̃ (z) = F (z)A(z), where
 
a(z) b(z)
A(z) = , a(z) = 0.
0 a(z)−1
We want to pick functions a, b to obtain a new lift whose Maurer-Cartan
form is as simple as possible. We have
F̃ −1 dF̃ = A−1 (F −1 dF )A + A−1 dA
 −1     −1   
a −b 0 0 a b a −b a b
= + dz
0 a −f  0 0 a−1 0 a 0 −a a−2
 
abf  + a−1 a a−2 ba + a−1 b + b2 f 
= dz.
−a2 f  −(abf  + a−1 a )

Choose a(z) = 1/ f  (z). Then F̃ −1 dF̃ takes the form
 
∗ ∗
dz.
1 ∗
(This is the analogue of setting f  (z0 ) = 1 in the coordinate approach.) The
function b is still free. We use it to set the diagonal term in the pullback of
the Maurer-Cartan form to zero, i.e., to set abf  + a−1 a = 0. This implies
a f 

= . b=−
2
a f 2(f  )3/2
Now our lift is unique and its Maurer-Cartan form is
 
0 12 Sf (z)
dz,
−1 0
where  
f  3 f  2
Sf =  −
f 2 f
is a differential invariant, called the Schwarzian derivative [1].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 24 1. Moving Frames and Exterior Differential Systems

Exercises 1.7.2:
1. Show that f is an FLT if and only if Sf ≡ 0. So, just as the curvature
of a curve in R2 measures the failure of a curve to be a line, Sf (z) is an
infinitesimal measure of the failure of a holomorphic map to be an FLT.
Since FLT’s map circles to circles, Sf may be thought of as measuring how
much circles are being distorted under f .
2. Calculate Sf for f (z) = aebz , and f (z) = z n . How do these compare
asymptotically? What does this say about how circles are distorted as one
goes out to infinity?

Exercises on curves in other plane geometries.

Exercises 1.7.3:
1. (Curves in the special affine plane) We consider the geometry of curves
that are equivalent up to translations and area-preserving linear transfor-
mations of R2 . These transformations are given by the matrix group
 
1 0 
ASL(2, R) = x ∈ R2 , A ∈ SL(2, R) ,
x A 

acting on R2 in the same way as ASO(2) acts in §1.4. Since the origin is
fixed by the subgroup SL(2, R), in this context we relabel R2 as the special
affine plane SA2 = ASL(2, R)/SL(2, R).
(a) Find differential invariants for curves in SA2 . (As with the Euclidean
case, one can consider invariants of a parametrized curve or invariants of
just the image curve.)
(b) What are the image curves with invariants zero? The image curves with
constant invariants?
(c) Let κA denote the differential invariant that distinguishes image curves.
Interpret κA (t) in terms of an osculating curve, as we did with the osculating
circles to a curve in §1.4.
(d) The preferred frame will lead to a unique choice of e2 . Give a geometric
interpretation of e2 . 
2. (Curves in the projective plane) Carry out the analogous exercise for
curves in the projective plane P2 = GL(3)/P , where P is the subgroup pre-
serving a line. Show that the curves with zero invariants are the projective

lines and plane conics. Derive the Monge equation (y  )−2/3 = 0, charac-
terizing plane conics among graphs y = f (x), by working in a local adapted
coordinate system. Note that one may do this exercise over R or C.
3. Carry out the analogous exercise for curves in the conformal plane
ACO(2)/CO(2), where equivalence is up to translations, rotations and di-
lations.
4. Carry out the analogous exercise for curves in L2 = ASO(1, 1)/SO(1, 1),
where SO(1, 1) is the subgroup of GL(2, R) preserving the quadratic form
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.8. Curves in E3 25


Q(v, w) = t v −1 0
0 1 w. Note that there will be three distinct types of
curves: spacelike curves, where Q(c (t), c (t)) > 0; timelike curves, where
Q(c (t), c (t)) < 0; and lightlike curves, where Q(c (t), c (t)) = 0. What are
the curves with constant invariants?

1.8. Curves in E3
The group ASO(3) and its Maurer-Cartan form. The group ASO(3)
is the set of transformations of E3 of the form x → t + Rx, i.e.,
⎛ 1⎞ ⎛ 1⎞ ⎛ 1⎞
x t x
⎝x2 ⎠ → ⎝t2 ⎠ + R ⎝x2 ⎠ ,
x3 t3 x3
where R ∈ SO(3) is a rotation matrix. Like ASO(2), it may be represented
as a matrix Lie group by writing
  
 1 0
(1.23) ASO(3) = M ∈ GL(4, R) M = , t ∈ R3 , R ∈ SO(3) .
t R
The action on E3 is given by x̂ → M x̂, where we represent points in E3 by
x = t (1, x1 , x2 , x3 ).
Having expressed ASO(3) as in (1.23), we may express an arbitrary
element of its Lie algebra aso(3) as
⎛ ⎞
0 0 0 0
⎜x1 0 −x2 −x3 ⎟
⎜ 1 1⎟ , ai , aij ∈ R.
⎝x2 x21 0 −x32 ⎠
x3 x31 x32 0
In this presentation, the Maurer-Cartan form of ASO(3) is
⎛ ⎞
0 0 0 0
⎜ω 1 0 −ω12 −ω13 ⎟
(1.24) ω=⎜⎝ω 2 ω12
⎟,
0 −ω23 ⎠
ω 3 ω13 ω23 0
where ω i , ωji ∈ Ω1 (ASO(3)). Recall from §1.6 that the forms ω i , ωji are left-
invariant and are a basis for the space of left-invariant 1-forms on ASO(3).
We identify ASO(3) with the space of oriented orthonormal frames of
E3 , as follows. Given a point x ∈ E3 and an oriented orthonormal basis
(e1 , e2 , e3 ) of Tx E3 , let
 
1 0
(1.25) g= ∈ ASO(3),
x R
where R = (e1 , e2 , e3 ) is a rotation matrix, and x = (x1 , x2 , x3 ) corresponds
to translation.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 26 1. Moving Frames and Exterior Differential Systems

With this identification, we obtain geometric interpretations of the left-

invariant forms. Substituting (1.25) and (1.24) into dg = g ω, and consider-
ing the first column, gives
(1.26) dx = ei ω i .
Thus, ω i has the geometric interpretation of measuring the infinitesimal
motion of the point x in the direction of ei . More precisely, let C(t) be any
lift of x(t) ⊂ E3 , so C ∗ (ω) = C(t)−1 C  (t)dt, and write
 
1 0 0 0
C(t) = ∈ ASO(3);
x(t) e1 (t) e2 (t) e3 (t)
then ω i (C  (t)) = x (t), ei . Similarly, ωji measures the infinitesimal motion
of ej toward ei , because the other columns of dg = g ω show that
(1.27) dej = ei ωji .
That these motions are infinitesimal rotations is reflected in the relation
ωji = −ωij , as illustrated by the following picture:

−ω 12 e2
ω 21

−e1 e1

ω 12
−e2 −ω 12

Recall from Appendix B that a form α ∈ Ω1 (P ) on the total space of a

fibration π : P → M is semi-basic for π if α(v) = 0 for all v ∈ ker π∗ .
Proposition 1.8.1. The forms ω i , 1 ≤ i ≤ 3, are semi-basic for the projec-
tion ASO(3) → E3 .

Proof. Let C(t), as above, be a curve in a fiber of ASO(3). We need to

show that ω i (C  (t)) = 0. If C(t) stays in one fiber, then dx dt = 0, but equation
(1.26) shows that dx = C  (t) dx = ω j (C  (t))e (t). The result follows because
dt j
the ej are linearly independent. 

Differential invariants of curves in E3 . We find differential invariants of

a regular curve c : R → E3 . For simplicity, we consider only the image curve,
so we can and will assume |c (t)| ≡ 1. Consequently, we have c ⊥ c (see
Exercise 1.4.2.1). To obtain a lift C : R → ASO(3) we may take e1 (t) = c (t),
e2 (t) = c (t)/|c (t)|, and this determines e3 (t). Our adaptations have the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.8. Curves in E3 27

effect that C ∗ (ω 1 ) is nonvanishing and C ∗ (ω 2 ) = C ∗ (ω 3 ) = 0. In terms of

the Maurer-Cartan form, we have:
d(x(t), e1 (t), e2 (t), e3 (t))
⎛ ⎞
0 0 0 0
(1.28) ⎜ω1 0 −ω12 −ω13 ⎟
= (x(t), e1 (t), e2 (t), e3 (t)) C ∗ ⎜
⎝0
⎟.
ω12 0 −ω23 ⎠
0 ω13 ω23 0
Exercise 1.8.2: Show that C ∗ (ω13 ) = 0.

All forms pulled back to R will be multiples of ω 1 , as ω 1 = dt furnishes

a basis of T ∗ R1 at each point. (We continue our standard abuse of notation,
writing ω 1 instead of C ∗ (ω 1 ).) So, we may write ω12 = κ(t)ω 1 and ω23 =
τ (t)ω 1 , where κ(t), τ (t) are functions called the curvature and torsion of the
curve. Traditionally one writes e1 = T, e2 = N, e3 = B; then (1.28) yields
the Frenet equations
⎛ ⎞
0 −κ 0
d(T, N, B) = (T, N, B) ⎝κ 0 τ ⎠ dt.
0 −τ 0
Curves with κ ≡ 0 are lines, and we may think of κ as a measurement of
the failure of the curve to be a line. Curves with τ ≡ 0 lie in a plane, and
we may think of τ as measuring the failure of a curve to lie in a plane. In
contrast to the example of plane curves, we needed a third-order invariant
(the torsion) to determine a unique lift in this case.
Theorem 1.3.1 implies that one can specify any functions (κ(t), τ (t)),
and there will be a curve having these as curvature and torsion (because on
ASO(3) × R the forms ω 1 − dt, ω 3 , ω 2 , ω12 − κ(t)ω 1 , ω13 , ω23 − τ (t)ω 1 satisfy
the hypotheses of the theorem). If the functions are nowhere vanishing the
curve will be unique up to congruence (see the exercises below).
Remark 1.8.3. Defining N as the unit vector in the direction of c (t)
means that κ cannot be negative, and N (along with the binormal B and
the torsion) is technically undefined at inflection points along the curve (i.e.,
points where c (t) = 0). However, it is still possible to smoothly extend the
frame (T, N, B) across inflection points, while satisfying the Frenet equations
for smooth functions (κ(t), τ (t)) where κ is allowed to change sign (see the
discussion in [16]). Such frames are sometimes called generalized Frenet
frames, and it is in this sense that ODE existence theorems provide a framed
curve with given curvature and torsion functions.
Exercises 1.8.4:
1. Using Theorem 1.6.12, show that if c, c̃ are curves with κ(t) = κ̃(t),
τ (t) = τ̃ (t), then c, c̃ differ by a rigid motion.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 28 1. Moving Frames and Exterior Differential Systems

2. Show that a curve c ⊂ R3 has constant κ and τ if and only if there

exists a line l ⊂ E3 with the property that every normal line of c intersects
l orthogonally. (A normal line is the line through c(t) in the direction of
N (t).)
3. (Bertrand curves) In the previous exercise we characterized curves with
constant invariants. Here we study the next simplest case, when there is a
linear relation among the curvature and torsion, i.e., nonzero constants a, b
such that aκ(t) + bτ (t) is constant for all t.
(a) Show that if such a linear relation holds, then there exists a second
curve c(t) with the same normal line as c(t) for all t.
(b) Show moreover that the distance between the points c(t) and c(t) is
constant. 
(c) Characterize the curves c where there exists more than one curve c with
this property.
4. Derive invariants for curves in En . How many derivatives does one need
to take to obtain a complete set of invariants?
5. (Curves on spheres) Show that a curve c with κ, τ = 0 is contained in a
sphere if and only if ρ2 + σ 2 is constant, where ρ = 1/κ and σ = ρ /τ . 
6. Let L3 = ASO(2, 1)/SO(2, 1), where SO(2, 1) is the subgroup of
GL(3, R) preserving the quadratic form Q whose associated matrix is
⎛ ⎞
−1
⎝ 1 ⎠.
1
Find differential invariants of curves in L3 . (As before, curves may be space-
like, timelike, or lightlike.) What are the curves with constant invariants?

1.9. Grassmannians
Perhaps the most important manifold is the Grassmannian G(k, V ) which
parametrizes k-dimensional subspaces of a vector space V . It plays a cen-
tral role in differential geometry, algebraic geometry, algebraic topology and
representation theory and helps bring these subjects together.
Fix index ranges 1 ≤ i, j ≤ k, and k + 1 ≤ s, t, u ≤ n.
Let V be a vector space over R or C and let G(k, V ) denote the Grass-
mannian of k-planes that pass through the origin in V . To specify a k-plane
E, it is sufficient to specify a basis v1 , . . . , vk of E. We continue our nota-
tional convention that {v1 , . . . , vk } denotes the span of the vectors v1 , . . . , vk .
After fixing a reference basis, we identify GL(V ) with the set of bases for V
by associating to g ∈ GL(V ) the columns of the matrix representing it, and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.9. Grassmannians 29

define a map
π : GL(V ) → G(k, V ),
(e1 , . . . , en ) → {e1 , . . . , ek }.
If we let ẽ1 , . . . , ẽn denote a reference basis of V , the fiber of this mapping
over π(Id) = {ẽ1 , . . . , ẽk } is the subgroup
 i 
g gi
Pk = g = j st | det(g) = 0 ⊂ GL(V ).
0 gs
More generally, for g ∈ GL(V ), π −1 (π(g)) = gPk , so the Grassmannian has
the structure of a homogeneous space.
Of particular importance is projective space PV = G(1, V ), the space of
all lines through the origin in V . When V = CN we write PN −1 = P(CN ).
We define a line in PV to be the projectivization of a two-dimensional linear
subspace of V and we define a (k − 1)-plane in PV as the projectivization
of a k-dimensional linear subspace in V . We may thus consider G(k, V ) as
the space of all Pk−1 ’s in PV .
Notation. If Y ⊂ PV , we let Ŷ ⊂ V denote its pre-image under the
projection V \{0} → PV , called the cone over Y . If Z ⊂ V , we let [Z] =
π(Z\{0}).
Exercises 1.9.1:
1. Show there is a canonical isomorphism G(k, V )  G(n − k, V ∗ ). In
particular, PV ∗ is the space of hyperplanes (i.e., Pn−1 ’s) in PV .
2. Show that the following map, called the Plücker embedding of the Grass-
mannian, is well-defined:
G(k, V ) → P(Λk V ),
{e1 , . . . , ek } → [e1 ∧ . . . ∧ ek ]. 
Proposition 1.9.2. Let E ∈ G(k, V ). There is a canonical identification
TE G(k, V )  E ∗ ⊗(V /E).

In particular, dim G(k, V ) = k(dim V − k).

First proof. Let C = {E(t)} ⊂ G(k, V ) denote a curve in G(k, V ), where

E(t) = [e1 (t) ∧ . . . ∧ ek (t)] with ei (t) curves in V . Write E(0) = E =
[e1 (0) ∧ . . . ∧ ek (0)] and el = el (0). It will be easier to differentiate in the

vector space Λk V , so let E(t) = e1 (t) ∧ . . . ∧ ek (t) ∈ Ĝ(k, V ) ⊂ Λk V . Then

k
  (0) =
E e1 ∧ . . . ∧ el−1 ∧ el (0) ∧ el+1 ∧ . . . ∧ ek .
l=1

We consider the term corresponding to a fixed l: If el (0) ∈ {e1 , . . . , el−1 , el+1 ,
. . . , ek }, the term is zero. If el (0) is a multiple of el , then the term in E
  (0)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 30 1. Moving Frames and Exterior Differential Systems

containing el (0) is a scalar multiple of E(0).

 Otherwise this term is nonzero

and linearly independent of E(0) and of the other terms in the summation
that are independent of E(0). Thus E  (0) will be zero if and only if E
  (0)

is a scalar multiple of E(0).
  (0) we associate a linear map f : E → V /E such that
To E

k

E (0) = e1 ∧ . . . ∧ el−1 ∧ f (el ) ∧ el+1 ∧ . . . ∧ ek ,
l=1

which descends to an element of TE G(k, V ). Now one just checks that this
correspondence is independent of our choices of bases and representative
curves el (t) with el (0) = f (el ). 

We do some preliminary work before our second proof. Write the

Maurer-Cartan form of GL(V ) as
 i 
ωj ωti
ω= .
ωjs ωts
Proposition 1.9.3. The forms ωis in the Maurer-Cartan form of GL(V ) are
semi-basic for the projection to G(k, V ).

Proof. A curve c(t) = (e1 (t), . . . , en (t)) ⊂ GL(V ) is vertical for the projec-
tion π : GL(V ) → G(k, V ) if and only if
ej (t) ⊂ {e1 (t), . . . , ek (t)}
for all 1 ≤ j ≤ k. But dej = ei ωji + es ωjs , and thus on a curve c(t),
dej
= ei ωji (c (t)) + es ωjs (c (t)).
dt
So, c(t) is vertical if and only if ωjs (c (t)) = 0 for all s, j. 

Second proof of 1.9.2. Let (ej , es ) be the dual basis to (ei , et ). Write
E = {ej }. For f = (e1 , . . . , en ) ∈ GL(V ), consider the tensor
Lf := ωis ⊗ei ⊗(es mod E) ∈ Tf∗ GL(V )⊗(E ∗ ⊗V /E).
In Exercise 1.9.4 you will show that L is basic for the projection π, so L
descends to a well-defined linear map TE G(k, V ) → E ∗ ⊗V /E, which is an
isomorphism because the ωis are semi-basic and independent, and the two
vector spaces have the same dimension. 

Exercise 1.9.4: Show that L is indeed basic, by showing that the actions
of Pk on the three terms cancel each other.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.10. Exterior differential systems and jet spaces 31

Third proof of 1.9.2. Recall from Exercise 1.6.1 that Tπ(Id) G/P  g/p.
In the case of G = GL(V ), we have g = V ∗ ⊗ V . Let E0 = π(Id) ∈
G(k, V ); then p  (E0∗ ⊗ E0 ) ⊕ ((V /E0 )∗ ⊗ E0 ) ⊕ ((V /E0 )∗ ⊗ (V /E0 )). (If
this confuses you, look at the block form of Pk above.) Writing V ∗ ⊗V =
(E0∗ + (V /E0 )∗ )⊗(E0 + V /E0 ), we obtain g/p  E0∗⊗(V /E0 ). In fact we may
drop the subscript 0, because the tangent space at g ∈ GL(V ) of the fiber
over E = π(g) is Lg∗ p, and the conjugate of g = gl(V ) is still isomorphic to
V ∗ ⊗V . 

Note that in the special case of PV = G(1, V ), we have Tx PV = x̂∗ ⊗

(V /x̂).
Definition 1.9.5. There are tautological vector bundles S → G(k, V ) and
Q → G(k, V ) where the fiber of S at E is the vector space E (so S is a
subbundle of the trivial bundle with fiber V ) and the fiber of Q at E is V /E.
The pointwise isomorphism of Proposition 1.9.2 extends to an isomorphism
of vector bundles S ∗ ⊗Q  T G(k, V ).

1.10. Exterior differential systems and jet spaces

In §1.3, we saw how a system of ODE or PDE could be replaced by an ideal
of differential forms, and solutions became submanifolds on which the forms
pull back to be zero. We now formalize this perspective, defining exterior
differential systems with and without independence condition.

Exterior differential systems with independence condition.

Definition 1.10.1. An exterior differential system with independence con-
dition on a manifold Σ consists of a differential ideal I ⊂ Ω∗ (Σ) and a
differential n-form Ω ∈ Ωn (Σ) defined up to scale. This Ω, or its equivalence
class under rescaling [Ω], is called the independence condition.
Definition 1.10.2. An integral manifold (or solution) of the system (I, Ω) is
an immersed n-fold f : M n → Σ such that f ∗ (α) = 0 ∀α ∈ I and f ∗ (Ω) = 0
at each point of M .

We also define the notion of an infinitesimal solution:

Definition 1.10.3. We say E ∈ G(n, Tx Σ) is an integral element of (I, Ω)
if Ω|E = 0 and α|E = 0 ∀α ∈ I. We let Vn (I, Ω)x denote the set of integral
elements of (I, Ω) at x ∈ Σ.

Integral elements are the potential tangent spaces to integral manifolds,

in the sense that the integral manifolds of an exterior differential system are
the immersed submanifolds M ⊂ Σ such that Tx M is an integral element
for all x ∈ M .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 32 1. Moving Frames and Exterior Differential Systems

Exercise 1.10.4: Let I n = I ∩Ωn (Σ). Show that Vn (I)x = {E ∈ Gn (Tx Σ) |

α|E = 0 ∀α ∈ I n }.

We now explain how to rephrase any system of PDE as an exterior

differential system with independence condition using the language of jet
bundles.

Jets.
Definition 1.10.5. Let t be a coordinate on R and let k ≥ 0. Two differ-
entiable maps f, g : R → R with f (0) = g(0) = 0 are said to have the same
k-jet at 0 if
df dg d2 f d2 g dk f dk g
(0) = (0), 2
(0) = 2 (0), . . . , k (0) = k (0).
dt dt dt dt dt dt
Let M, N be differentiable manifolds and f, g : M → N be two maps.
Then f and g are said to have the same k-jet at p ∈ M if
i. f (p) = g(p) = q, and
ii. for all maps u : R → M and v : N → R with u(0) = p and v(q) = 0,
the differentiable maps v ◦ f ◦ u and v ◦ g ◦ u have the same k-jet at 0.
Exercise 1.10.6: Show that to determine if f, g : M → N have the same
k-jet at p, it is sufficient to check derivatives up to order k with respect to
coordinate directions in any pair of local coordinate systems around p and
q.

Having the same k-jet at p is an equivalence relation on smooth maps.

We denote the equivalence class of f by jpk (f ), the space of k-jets where p
k (M, N ) and the space of all k-jets of all maps from M to N
maps to q, by Jpq
k
by J (M, N ). This is a smooth manifold, with local coordinates as follows:
Suppose M has local coordinates xi and N local coordinates ua . Then
J k (M, N ) has coordinates xi , ua , pai , paij , . . . , pai1 ,...,ik . We will abbreviate this
as (xi , ua , paI ), where I is a multi-index of length up to k whose entries range
between 1 and dim M . Then the point jxk0 (f ) ∈ J k (M, N ) has coordinates
∂ |I| f a
xi0 , ua = f a (x0 ) and paI = ∂xI
(x0 ) for 1 ≤ |I| ≤ k.
Furthermore, because we assume f is smooth, we may take the entries
in I to be nondecreasing. For example, coordinates on J 2 (R2 , R) would be
x1 , x2 ,u1 , p11 , p12 , p111 , p112 and p122 .
Exercise 1.10.7: Calculate the dimensions of (a) J 2 (R3 , R), (b) J 3 (R2 , R2 ),
(c) J k (Rn , Rm ). 

Note that Tx∗ M = Jx,0

1 (M, R), T M = J 1 (R, M ), and, in general,
x 0,x
J (M, N ) is a fiber bundle over M (as well as over M × N ). Any map
k
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.10. Exterior differential systems and jet spaces 33

f : M → N induces a section p → jpk (f ) of this bundle, called the lift of the

graph of f .
Canonical contact systems. On J k (M, N ) there is a canonical EDS with
independence condition, called the contact system, whose integral manifolds
are the k-graphs of maps f : M → N to J k (M, N ). We now describe this
system in the local coordinates used above.
Let Ω := dx1 ∧ . . . ∧ dxn and let I be the ideal generated differentially
by the 1-forms
θa := dua − pai dxi ,
θia := dpai − paij dxj ,
(1.29) ..
.
θia1 ,...,ik−1 := dpai1 ,...,ik−1 − pai1 ,...,ik dxik ,

which we will call contact forms. (Note the summation on ik .) We will use
multi-index notation to abbreviate the forms in (1.29) as

θIa := dpaI − paIj dxj .

The system (I, Ω) on J k (M, N ) is defined globally and is indepen-

dent of the coordinates chosen. It is called the canonical contact system
on J k (M, N ). Its integral manifolds are exactly the k-graphs of mappings
f : M → N to J k (M, N ). To see this, let i : X → J k be an n-dimensional in-
tegral manifold with local coordinates x1 , . . . , xn . On X, u = u(x1 , . . . , xn ),
pai = pai (x1 , . . . , xn ), etc., and i∗ θa = 0 implies that pai = ∂u
a
∂xi
for all
1 ≤ i ≤ n. Similarly, the vanishing of the other forms in the ideal force
the other jet coordinates to be the higher derivatives of u.
How to express any PDE system as an EDS with independence condition.
Given a kth-order system of PDE for maps f : Rn → Rs ,
 
∂ |I| ua
(1.30) F r xi , ua , = 0, 1 ≤ r ≤ R, 1 ≤ |I| ≤ k,
∂xI

we define a submanifold Σ ⊂ J k by the equations F r (xi , ua , paI ) = 0. The

lifts of solutions of (1.30) are precisely the integral manifolds of the pullback
of the contact system to Σ. Note that Ω tells us what the independent
variables should be.

Standard abuse of notation. Given an inclusion i : M → Σ, instead of

writing i∗ θ = 0 or i∗ Ω = 0 we will often simply say respectively “θ = 0 on
M ”or “Ω = 0 on M ”.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 34 1. Moving Frames and Exterior Differential Systems

Exterior differential systems. We generalize our notion of exterior dif-

ferential systems with independence condition as follows:

Definition 1.10.8. An exterior differential system on a manifold Σ is a

differential ideal I ⊂ Ω∗ (Σ). An integral manifold of the system I is an
immersed submanifold f : M → Σ such that f ∗ (α) = 0 ∀α ∈ I.

Note that for an exterior differential system, not only do we do not

specify the analog of independent variables, but we do not even specify a
required dimension for integral manifolds.
We define a k-dimensional integral element of I at x ∈ Σ to be an
E ∈ G(k, Tx Σ) such that α|E = 0 ∀α ∈ I. Let Vk (I)x denote the space of
k-dimensional integral elements to I at x.
Exercise 1.10.9: Let I = {x1 dx2 , dx3 }diff be an exterior differential system
on R3 . Calculate V1 (I)(1,1,1) , V1 (I)(0,0,0) , V2 (I)(1,1,1) and V2 (I)(0,0,0) .
Let Ω = adx1 + bdx2 + cdx3 , where a, b, c are constants, not all zero.
Calculate V1 (I, Ω)(1,1,1) and V1 (I, Ω)(0,0,0) .

Proof of the Frobenius Theorem. We now prove Theorem 1.3.9:

Proof. We follow the proof in [49], as that proof will get us accustomed to
calculations with differential forms.
We proceed by induction on n, the rank of the distribution annihilated
by I. If n = 1, then the distribution defines a line field and we are done by
Theorem 1.3.1. Assume that the theorem is true up to n − 1, and we will
show that it is true for n.
Let x : M → R be a smooth function such that θ1 ∧ . . . ∧ θm−n ∧ dx = 0
on a neighborhood U of p, and consider the ideal I  = {θ1 , . . . , θm−n , dx}diff .
Since I = {θ1 , . . . , θm−n }diff is Frobenius, I  is also Frobenius. By our
induction hypothesis, there exist local coordinates (y 1 , . . . , y m ) such that
I  = {dy 1 , . . . , dy m−n+1 }alg .
At this point we have an (n − 1)-dimensional integral manifold of I 
(hence, also of I) passing through p, obtained by setting y 1 , . . . , y m−n+1
equal to the appropriate constants. We want to enlarge it to an n-dimension-
al integral manifold.
Let 1 ≤ i, j ≤ m − n. We may write
dx = ai dy i + am−n+1 dy m−n+1 ,
θi = cij dy j + cim−n+1 dy m−n+1 ,

where ai , am−n+1 , cij , cim−n+1 are smooth functions. Without loss of gener-
ality, we may assume an−m+1 = ∂x/∂y m−n+1 = 0, so we may rewrite the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.10. Exterior differential systems and jet spaces 35

Figure 3. Curve with plane field to be enlarged to surface

second line as
θi = c̃ij dy j + f i dx
for some smooth functions c̃ij , f i . The matrix of functions c̃ij is invertible at
each point in a (possibly smaller) neighborhood Ũ of p, so locally we may
take a new set of generators for I, of the form
θ̃i = dy i + ei dx
for some smooth functions ei , and with θi = c̃ij θ̃j . Then dθ̃i = dei ∧ dx and,
since I is Frobenius,
dei ∧ dx ≡ 0 mod {θ̃j }.
Hence
dei = adx + bij dy j
for some functions a, bij . In particular, the ei are functions of the y j and
x only, and it follows that the θ̃i are defined in terms of the variables
y 1 , . . . , y m−n+1 only.
Let V ⊂ Ũ be the submanifold through p obtained by setting y m−n+2
through y m constant. Then I|V is a codimension-one Frobenius system.
Hence there are coordinates (ỹ 1 , . . . , ỹ m−n+1 ) on V that are functions of
the y 1 , . . . , y m−n+1 , such that I|V is generated by dỹ 1 , . . . , dỹ m−n . These
relationships extend to Ũ , so that
(ỹ 1 , . . . , ỹ m−n+1 , y m−n+2 , . . . , y m )
is the desired coordinate system. 

For a proof of the maximality assertion in the first version of the Frobe-
nius Theorem (1.3.7), see [187, Thm. 1.6.4].

Symplectic manifolds, contact manifolds and their EDS. What fol-

lows are two examples of classical exterior differential systems and a com-
plete local description of their integral manifolds.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 36 1. Moving Frames and Exterior Differential Systems

Symplectic manifolds. Let Σ = R2n with coordinates (x1 , . . . , xn , y 1 , . . . , y n )

and let

n
(1.31) φ= dxi ∧ dy i .
i=1

Consider the exterior differential system I = {φ}diff .

Exercises 1.10.10:
1. At what points is ∂x∂ 1 ∧ . . . ∧ ∂x∂n an integral element? What about
∂
∂y 1
∧ . . . ∧ ∂y∂n ?
2. Show that any graph y j = f j (x1 , . . . , xn ) of the form y j = f j (xj ) for
each j is an integral manifold.

We claim there are no (n + 1)-dimensional integral elements for I. First,

it is easy to check that φ is nondegenerate, i.e., φ(v, w) = 0 for all vectors
w ∈ Tp Σ only if v = 0. Now suppose E = {v1 , . . . , vn+1 } is an integral
element of dimension n + 1 at some point p ∈ Σ. Then the nondegeneracy
of φ implies that the forms αj = vj φ ∈ Tp∗ Σ are linearly independent.
However, this is impossible since αi |E = 0 for every αi .
Exercise 1.10.11: Alternatively, show that there are no (n+1)-dimensional
integral elements to I by relating φ to the standard inner product ,  on R2n .
Namely, let φ(v, w) = v, Jw, where J is the standard complex structure
defined in Exercise A.3.1. Then, if E is a integral element, show that , 
must be degenerate on E ∩ J(E).

An even-dimensional manifold with closed nondegenerate 2-form is called

a symplectic manifold, and the 2-form is called a symplectic form. The
following theorem shows that the above example on R2n is quite general.

Notation 1.10.12. For ω ∈ Ω2 (M ), we will write ω r for the r-fold wedge

product ω ∧ ω ∧ · · · ∧ ω of ω with itself.

Theorem 1.10.13 (Darboux). Suppose a closed 2-form ω ∈ Ω2 (M n ) is such

that ω r = 0 but ω r+1 = 0 in some neighborhood U ⊂ M . Then there exists
a coordinate system w1 , . . . , wn , possibly in a smaller neighborhood, such
that
ω = dw1 ∧ dw2 + . . . + dw2r−1 ∧ dw2r .
In particular, ω takes the form (1.31) when n = 2r.
Darboux’s Theorem implies that all symplectic manifolds are locally
equivalent, in contrast to Riemannian manifolds (see 3.1.13). Globally this
is not at all the case, and the study of the global geometry of symplectic
manifolds is an active area of research (see [60], for example).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 1.10. Exterior differential systems and jet spaces 37

Example 1.10.14. Given any differentiable manifold M , the cotangent

bundle T ∗ M is canonically a symplectic manifold.
Let π : T ∗ M → M be the projection and define a tautological 1-form
α ∈ Ω1 (T ∗ M ) as follows: α(v)(u) := u (π∗ (v)), where x ∈ M and u ∈ Tx∗ M .
then T ∗ M has local coordinates (xi , yj ) such
If M has local coordinates xi , 
∗
that if u ∈ Tx M , then u = j
 j yj (u)dx . So, in these coordinates α =
j ∗
j yj dx . Hence, ω = dα is a symplectic form on T M .

Contact manifolds.
Exercise 1.10.15: Consider the contact system on J 1 (Rn , R) = R2n+1 with
coordinates (z, x1 , . . . , xn , y 1 , . . . , y n ). Here θ = dz − Σi y i dxi generates the
contact system I = {θ}diff .
(a) Show that at any point ∂
∂y 1
∧ ...∧ ∂
∂y n is an integral element.
(b) At which points is ∂
∂x1
∧ . . . ∧ ∂x∂n an integral element?
(c) Show that any graph z = h(x1 , . . . , xn ), y j = f j (x1 , . . . , xn )
such that
f j = ∂h/∂xj is an integral manifold. (In fact, all n-dimensional integral
manifolds are locally of this form.)
(d) Show that there are no (n + 1)-dimensional integral elements for I.

As with the example above, this example is general:

Theorem 1.10.16 (Pfaff). Let M be a manifold of dimension n + 1, let θ ∈
Ω1 (M ) and I = {θ}diff . Let r ∈ N be such that (dθ)r ∧ θ = 0 but (dθ)r+1 ∧
θ = 0 in some neighborhood U ⊂ M . Then there exists a coordinate
system w0 , . . . , wn , possibly in a smaller neighborhood, such that I is locally
generated by

θ̃ = dw0 + wr+1 dw1 + . . . + w2r dwr

(i.e., θ is a nonzero multiple of θ̃ on U ). In fact, there exist coordinates

y 0 , . . . , y n such that

y 0 dy 1 + y 2 dy 3 + . . . + y 2r dy 2r+1 if (dθ)r+1 = 0,
θ=
dy 1 + y 2 dy 3 + . . . + y 2r dy 2r−1 if (dθ)r+1 = 0,

on U .
The integer r is sometimes referred to as the rank of θ [27, II.3]; to avoid
confusion with the rank of a Pfaffian system, we will call r the Pfaff rank.
If n = 2r, then Pfaff’s Theorem implies that M is locally diffeomorphic to
the jet bundle J 1 (Rr , R) and θ̃ is the pullback of the standard contact form.
Thus, r-dimensional integral manifolds of I are given in the coordinates
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 38 1. Moving Frames and Exterior Differential Systems

w0 , . . . , wn by
w0 = f (w1 , . . . , wr ),
∂f
wr+1 = ,
∂w1
..
.
∂f
. w2r =
∂wr
A 1-form θ on a (2n + 1)-dimensional manifold Σ is called a contact form if
it is as nondegenerate as possible, i.e., if θ ∧ (dθ)n is nonvanishing.
Since we use the 1-form θ to define an EDS, we really only care about
it up to multiplication by a nonvanishing function. A contact manifold is
defined to be a manifold with a contact form, defined up to scale—more
precisely, a rank-one subbundle I ⊂ T ∗ M any nonvanishing section of which
is a contact form. This generalizes the contact structure on J 1 (M, R), our
first example of a contact manifold.
Example 1.10.17. The projectivized tangent bundle PT M may be given
the structure of a contact manifold by taking the distribution α⊥ ⊂ T (T M )
annihilated by the tautological form and projecting to PT M .
Exercise 1.10.18 (Normal form for degenerate contact forms): On R3 , con-
sider a 1-form θ such that θ ∧ dθ = f Ω, where Ω is a volume form and f
is a function such that df |p = 0 whenever f (p) = 0. Show that there are
coordinates (x, y, z), in a neighborhood of any such point where θp = 0, such
that θ = dz − y 2 dx.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/02

Chapter 2

Euclidean Geometry

In this chapter we return to the study of surfaces in Euclidean space E3 =

ASO(3)/SO(3). Our goal is not just to understand Euclidean geometry
of surfaces, but to develop techniques for solving equivalence problems for
submanifolds of arbitrary homogeneous spaces. We begin with the problem
of determining if two surfaces in E3 are locally equivalent up to a Euclidean
motion. More precisely, given two immersions f, f˜ : U → E3 , where U
is a domain in R2 , when do there exist a local diffeomorphism φ : U →
U and A ∈ ASO(3) such that f˜ ◦ φ = A ◦ f ? Motivated by our results
on curves in Chapter 1, we first try to find a complete set of Euclidean
differential invariants for surfaces in E3 , i.e., functions I1 , . . . , Ir that are
defined in terms of the derivatives of the parametrization of a surface, with
the property that f (U ) differs from f˜(U ) by a Euclidean motion if and only
if (f˜ ◦ φ)∗ Ij = f ∗ Ij for 1 ≤ j ≤ r.
In §2.1 we derive the Euclidean differential invariants Gauss curvature K
and mean curvature H using moving frames. Unlike with curves in E3 , for
surfaces in E3 there is not always a unique lift to ASO(3), and we are led to
define the space of adapted frames. We calculate the functions H, K for two
classical types of surfaces (developable surfaces and surfaces of revolution)
in §2.2, and discuss basic properties of these surfaces.
Scalar-valued differential invariants turn out to be insufficient (or at least
not convenient) for studying equivalence of surfaces and higher-dimensional
submanifolds, and we are led to introduce vector bundle valued invariants.
This study is motivated in §2.4 and carried out in §2.5, resulting in the
definitions of the first and second fundamental forms, I and II. In §2.5 we
also generalize our discussion of adapted frames for surfaces in E3 to higher

39
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 40 2. Euclidean Geometry

dimensions and codimensions, interpret II and Gauss curvature, define the

Gauss map, and derive the Gauss equation for surfaces.
Relations between intrinsic and extrinsic geometry of submanifolds of
Euclidean space are taken up in §2.6, where we prove Gauss’s Theorem
Egregium, derive the Codazzi equation, and write down the Levi-Civita
connection for vector fields on surfaces.
In §2.7 we discuss curves on surfaces, which play a role in the Gauss-
Bonnet Theorem. In §2.8 we state and prove the Gauss-Bonnet and
Poincaré-Hopf theorems. We conclude this chapter with a discussion of
nonorthonormal frames in §3.2, which enables us to finally prove the formula
(1.3) and show that surfaces with H identically zero are minimal surfaces.
The geometry of surfaces in E3 is studied further throughout Chapters
6 and 7. Riemannian geometry is discussed further in Chapters 3 and 9.

2.1. Gauss and mean curvature via frames

Guided by Theorem 1.6.12, we begin our search for differential invariants of
immersed surfaces f : U 2 → E3 by trying to find a lift F : U → ASO(3)
which is adapted to the geometry of M = f (U ). The most naı̈ve lift would
be to take  
1 0
F (p) = ,
f (p) Id3
where Id3 denotes the 3 × 3 identity matrix. Any other lift F̃ is of the form
 
1 0
F̃ = F
0 r
for some map r : U → SO(3).
Let x = f (p); then Tx E3 has distinguished subspaces, namely f∗ (Tp U )
and its orthogonal complement. We use our rotational freedom to adapt
to this situation by requiring that e3 always be normal to the surface, or
equivalently that {e1 , e2 } span Tx M = f∗ (Tp U ). This is analogous to our
choice of coordinates at our preferred point in §1.1, but is more powerful
since it works on an open set of points in U .
We will call a lift such that e3 is normal to Tx M a first-order adapted
lift, and continue to denote such lifts by F . Our adaptation implies that
(2.1) F ∗ (ω 3 ) = 0,
(2.2) F ∗ (ω 1 ∧ ω 2 ) = 0 at each point.

The equation dx = ω 1 e1 + ω 2 e2 + ω 3 e3 (see (1.26)) shows that (2.1) can

be interpreted as saying that x does not move in the direction of e3 to first
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.1. Gauss and mean curvature via frames 41

order, and (2.2) implies that to first order x may move independently along
the span of e1 and e2 .
Let π : F 1 → U denote the bundle whose fiber over x ∈ U is the set of
oriented orthonormal bases (e1 , e2 , e3 ) of Tx E3 such that e3 ⊥ Tf (x) M . The
first-order adapted lifts are exactly the sections of F 1 . By fixing a reference
frame at the origin in E3 , ASO(3) may be identified as the bundle of all
oriented orthonormal frames of E3 , and F 1 is a subbundle of f ∗ (ASO(3)).
Throughout this chapter we will not distinguish between U and M when
the distinction is unimportant. In particular, we will usually consider F 1 as
a bundle over M .
Consequences of our adaptation. Thanks to the Maurer-Cartan equation
(1.19), we may calculate the derivatives of the left-invariant forms on
ASO(3) algebraically:

⎛ ⎞
0 0 0 0
⎜ω 1 0 −ω12 −ω13 ⎟
d⎜
⎝ω 2 ω12
⎟
0 −ω23 ⎠
ω 3 ω13 ω23 0
⎛ ⎞ ⎛ ⎞
0 0 0 0 0 0 0 0
⎜ω 1 0 −ω 2 −ω 3 ⎟ ⎜ω 1 0 −ω 2 −ω 3 ⎟
= −⎜⎝ω 2 ω12
1 1⎟ ∧ ⎜ 1 1⎟ .
0 −ω23 ⎠ ⎝ω 2 ω12 0 −ω23 ⎠
ω 3 ω13 ω23 0 ω 3 ω13 ω23 0
Write i : F 1 → ASO(3) for the inclusion map. By our definition of F 1 ,
i∗ ω 3 = 0, and hence
(2.3) 0 = i∗ (dω 3 ) = −i∗ (ω13 ∧ ω 1 + ω23 ∧ ω 2 ).
By (2.2), i∗ ω 1 and i∗ ω 2 are independent, and we can apply the Cartan
Lemma A.1.9 to the right hand side of (2.3). We obtain
 3    
∗ ω1 h11 h12 ∗ ω 1
(2.4) i = i ,
ω23 h21 h22 ω2
where hij = hji are some functions defined on F 1 . This h = (hij ) is analo-
gous to the Hessian at the origin in (1.1), but it has the advantage of being
defined on all of F 1 .
Given an adapted lift F : U → F 1 , we have
 3  1
∗ ω1 ∗ ω
F = hF F ,
ω23 ω2
where hF = F ∗ (h). We now determine the invariance of hF . Since F
is uniquely defined up to a rotation in the tangent plane to M , all other
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 42 2. Euclidean Geometry

possible adapted lifts are of the form

⎛ ⎞
1
(2.5) F = F ⎝ r ⎠ =: F R,
1
where r : U → SO(2) is an arbitrary smooth function.
We compare F∗ (ω) = F−1 dF with F ∗ (ω) = F −1 dF :
F−1 dF = (F R)−1 d(F R) = R−1 (F −1 dF )R + R−1 F −1 F dR
⎛ ⎞
⎛ ⎞ 0 0 0 0 ⎛ ⎞
1 ⎜ω 1
1 0 −ω1 −ω1 ⎟
2 3
= ⎝ r−1 ⎠ F ∗ ⎜ ⎝ω 2 ω12
⎟⎝ r ⎠
0 −ω23 ⎠
1 1
ω 3 ω13 ω23 0
⎛ ⎞
0
+ ⎝ r−1 dr ⎠ .
0
In particular,
 1  1
 ∗ ω −1 ∗ ω
F =r F , F∗ (ω13 , ω23 ) = F ∗ (ω13 , ω23 )r.
ω2 ω2
Since r−1 = t r, we conclude that
(2.6) hF̃ = r−1 hF r.
Thus, the properties of hF that are invariant under conjugation by a
rotation matrix are invariants of the mapping f . The functions 12 trace(hF )
and det(hF ) generate the ideal of functions on hF that are invariant under
(2.6). They are respectively called the mean curvature (first defined by
Sophie Germain) denoted by H, and the Gauss curvature (first defined by
a mathematician with better p.r.) denoted by K. We see immediately that
for two surfaces to be congruent it is necessary that they must have the
same Gauss and mean curvature functions at corresponding points, thus
recovering our observations of §1.1.
Remark 2.1.1. Instead of working with lifts to F 1 , one could work with
h : F 1 → S 2 R2 directly, calculating how h varies as one moves in the fiber.

Let k1 , k2 denote the eigenvalues of h; for the sake of definiteness, say

k1 ≥ k2 . These are called the principal curvatures of M ⊂ E3 , and are also
differential invariants. However, H, K are more natural invariants, because,
e.g., the Gauss curvature plays a special role in the intrinsic geometry of
the surface; see for example Theorem 2.6.2 below. Note also that if M is
smooth, then H, K are smooth functions on M while k1 and k2 may fail to
be differentiable at points where k1 = k2 , which are called umbilic points.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.2. Calculation of H and K for some examples 43

Exercises 2.1.2:  
cos θ sin θ
1. Let F̃ be as in (2.5) and let R = . Calculate F̃ ∗ ω12 in
− sin θ cos θ
terms of θ and compare with F ∗ ω12 .
2. Show that H, K are invariants of the image of f . 
3. Express k1 , k2 in terms of H and K. 

Example 2.1.3 (Surfaces with H = K = 0). Suppose f : U → E3 gives

a surface M with H, K identically zero. If H = K = 0, then the matrix h
is zero and ω13 , ω23 vanish. So, on Σ = ASO(3) define I = {ω 3 , ω13 , ω23 }diff
with independence condition Ω = ω 1 ∧ ω 2 . Such a surface is (as you may
already have guessed) a subset of a plane. For, if F : U → ASO(3) is a
first-order adapted frame for a surface with H = K = 0, then F (U ) will be
an integral surface of (I, Ω). Note that de3 = −ω13 e1 − ω23 e2 = 0, so e3 is
constant for such lifts. Therefore, for all x ∈ M there is a fixed vector e3
such that e3 ⊥ Tx M , and thus M is contained in a plane perpendicular to
e3 .

2.2. Calculation of H and K for some examples

The Helicoid. Let R2 have coordinates (s, t), fix a constant a > 0 and con-
sider the mapping f : R2 → E3 defined by

f (s, t) = (s cos t, s sin t, at).

The image surface is called the helicoid.

Exercise 2.2.1: Draw the helicoid. Note that the z-axis is contained in the
surface, as is a horizontal line emanating out from each point on the z-axis,
and this line rotates as we move up the z-axis.

We compute a first-order adapted frame F : R2 → ASO(3) for the

helicoid. Note that
fs = (cos(t), sin(t), 0),
ft = (−s sin(t), s cos(t), a),
so fs , ft  = 0 and we may take
fs
e1 = = (cos(t), sin(t), 0),
|fs |
ft 1
(2.7) e2 = =√ (−s sin(t), s cos(t), a),
|ft | s + a2
2
1
e3 = e1 × e2 = √ (a sin(t), −a cos(t), s).
Licensed to AMS.
s2 + a2
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 44 2. Euclidean Geometry

Since df = fs ds + ft dt = F ∗ (ω 1 )e1 + F ∗ (ω 2 )e2 , we obtain (omitting the F ∗

from the notation here and in what follows)
ω 1 = ds,
(2.8) 1
ω 2 = (s2 + a2 ) 2 dt.
Next, we calculate
 
de3 = −s(s2 + a2 )−3/2 (a sin(t), −a cos(t), s) + ((s2 + a2 )−1/2 (0, 0, 1) ds

+ (s2 + a2 )−1/2 (a cos(t), a sin(t), 0)dt.

So, using (2.7), (2.8), we obtain
ω13 = −a(s2 + a2 )−1 ω 2 ,
ω23 = −a(s2 + a2 )−1 ω 1 ,
2
and conclude that H(s, t) ≡ 0 and K(s, t) = − (s2 +a
a
2 )2 .

Surfaces with H identically zero are called minimal surfaces and are
discussed in more detail in §3.2 and §7.4.
Developable surfaces. A surface M 2 ⊂ E3 is said to be a tangential devel-
opable if it is describable as (a subset of) the union of tangent rays to a
curve. (These surfaces are also called tangential surfaces.)
Let c : R → E3 be a regular parametrized curve, and consider the surface
f : R2 → E3 defined by (u, v) → c(u) + vc (u). Since
df = (c (u) + vc (u))du + c (u)dv,
we see that f is regular, i.e., df is of maximal rank, when c (u) is linearly
independent from c (u) and v = 0. (We will assume v > 0.) Note that the
tangent space, spanned by c (u) and c (u), is independent of v.
Assume that c is parametrized by arclength. Then c (u), c (u) = 0,
and we can take
e1 (u, v) = c (u),
e2 (u, v) = c (u)/||c (u)||.
Using df = ω 1 e1 + ω 2 e2 , we calculate
ω 1 = du + dv,
ω 2 = vκ(u)du,
where κ(u) is the curvature of c.
Note that our frame is the same as if we were to take an adapted framing
of c (as in §1.8), so we have
de3 = (−τ (u)e2 )du.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.2. Calculation of H and K for some examples 45

Thus,
τ (u) 2
ω13 = 0, ω23 = ω ,
κ(u)v
τ (u)
showing that H(u, v) = 2κ(u)v and K ≡ 0.
Surfaces with K identically zero are called flat, and we study their ge-
ometry more in §2.4.
Developable surfaces are also examples of ruled surfaces (as is the heli-
coid). A surface is ruled if through any point of the surface there passes a
straight line (or line segment) contained in the surface.
Surfaces of revolution. Let U ⊂ R2 be an open set with coordinates u, v and
let f : U → E3 be a map of the form
x(u, v) = r(v) cos(u),
y(u, v) = r(v) sin(u),
z(u, v) = t(v),
where r, t are smooth functions. The resulting surface is called a surface of
revolution because it is constructed by rotating a generating curve (e.g., in
the xz-plane) about the z-axis. Call the image M .
Assuming that the generating curve is regular, we can choose v to be an
arclength parameter, so that (r (v))2 + (t (v))2 = 1. Let
e1 = (− sin u, cos u, 0),
(2.9)
e2 = (r (v) cos u, r (v) sin u, t (v)).
Note that ej ∈ Γ(U, f ∗ (T E3 )).
Exercises 2.2.2:
1. (a) Show that e1 , e2 in (2.9) is an oriented orthonormal basis of Tf (u,v) M .
(b) Calculate e3 such that e1 , e2 , e3 is an orthonormal basis of Tf (u,v) E3 .
2. Considering this frame as an adapted lift F : U → F 1 , calculate the
pullback of the Maurer-Cartan forms in terms of du, dv.
3. Calculate the Gauss and mean curvature functions of M .
4. Consider the surfaces of revolution generated by the following data. In
each case, describe the surface geometrically. (Take time out to draw some
pictures and have fun!) Calculate H, K and describe their asymptotic be-
havior.
(a) r(v) = constant, t(v) = v.
(b) r(v) = av, t(v) = bv, where a, b are constants such that a2 + b2 = 1.
(c) r(v) = cos v, t(v) = sin v.
(d) The generating curve in the xz-plane is a parabola, e.g., x − bz 2 = c.
(e) The generating curve is a hyperbola, e.g., x2 − bz 2 = c.
2 2
(f) The generating curve is an ellipse, e.g., xa2 + zb2 = 1.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 46 2. Euclidean Geometry

5. Find all surfaces of revolution with K ≡ 0. Give a geometric construction

of these surfaces. 
6. Find all surfaces of revolution with K ≡ 1 that intersect the xy-plane
perpendicularly. (Your answer should involve an integral and the choice of
one arbitrary constant.) Which of these are complete? 

2.3. Darboux frames and applications

Recall that k1 ≥ k2 are the eigenvalues of the second fundamental form
matrix h. Away from umbilic points (points where k1 = k2 ), k1 and k2 are
smooth functions, and we may further adapt frames by putting h in the
form  
k1 0
h= ,
0 k2
because a real symmetric matrix is always diagonalizable by a rotation ma-
trix. In this case F is uniquely determined. We will call such a framing a
Darboux or principal framing.
Notation. We will express the derivative of a function u on a framed
surface as du = u1 ω 1 + u2 ω 2 , where uj = ej (u). In particular, write dkj =
kj,1 ω 1 +kj,2 ω 2 to represent the derivatives of the kj in the following exercises.
Exercises 2.3.1:
1. Find all surfaces in E3 with k1 ≡ k2 , i.e., surfaces where each point is an
umbilic point.
2. Let c ⊂ E2 be the curve defined by intersecting a surface M 2 ⊂ E3 with
the plane through x parallel to e1 , e3 . Show that the curvature of c at x is
|k1 |.
3. Calculate F ∗ (ω12 ) in a Darboux framing as a function of the principal
curvatures and their derivatives. 
4. Suppose that X1 , X2 are vector fields on U such that f∗ Xi = ei . Show
that
k1,2 k2,1
[X1 , X2 ] = − X1 − X2 . 
k1 − k2 k1 − k2
5. Derive the Codazzi equation for Darboux frames, i.e., show that k1 , k2
satisfy the differential equation
(2.10)
2((k1,2 )2 + (k2,1 )2 ) − k1,1 k2,1 − k2,2 k1,2
k2,1,1 −k1,2,2 + +(k1 −k2 )k1 k2 = 0. 
k1 − k2
This to some extent addresses the existence question for principal curva-
ture functions. Namely, two functions k1 (u, v), k2 (u, v) that are never equal
cannot be the principal curvature functions of some embedding of U → E3
unless they satisfy the Codazzi equation. In particular, show that surfaces
with both H and K constant must be either flat or totally umbilic.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.4. What do H and K tell us? 47

6. Using the Codazzi equation, show that if k1 > k2 everywhere, and if

there exists a point p at which k1 has a local maximum and k2 a local
minimum, then K(p) ≤ 0.

Even among surfaces of revolution, there are an infinite number of non-

congruent surfaces with K ≡ 1 (see Exercise 2.2.2.6). We will see in Example
6.8.2 and again in §7.4 that surfaces with constant K > 0 are even more
flexible in general. Thus, the following theorem might come as a surprise:
Theorem 2.3.2. If M 2 ⊂ E3 is compact, without boundary, and has con-
stant Gauss curvature K > 0, then M is the round sphere.
Exercise 2.3.3: Prove Theorem 2.3.2. 

2.4. What do H and K tell us?

Since Darboux frames provide a unique lift for M and well-defined differen-
tial invariants, it is natural to pose the question:
Are surfaces M 2 ⊂ E3 with no umbilic points locally deter-
mined, up to a Euclidean motion, by the functions H and
K?
The answer is NO! Consider the following example:
The Catenoid. Let R2 have coordinates (u, v), let a > 0 be a constant, and
consider the following mapping f : R2 → R3 :
f (u, v) = (a cosh v cos u, a cosh v sin u, av).
The image is called the catenoid.
Exercise 2.4.1: Draw the catenoid. Calculate the mean and Gauss curva-
ture functions by choosing an adapted orthonormal frame and differentiating
as we did in §2.2. Conclude that the catenoid is a minimal surface. 

Now consider the map g : R2 → R2 given by s = a sinh v, t = u.

Exercise 2.4.2: Show that g ∗ (Khelicoid ) = Kcat , and of course the mean
curvature functions match up as well.

Although we have the same Gauss and mean curvature functions, the
helicoid is ruled, and it is not hard to check that the catenoid contains no
line segments. Since a Euclidean transformation takes lines to lines, we see it
is impossible for the helicoid to be equivalent to the catenoid via a Euclidean
motion.
We will see in Example 7.4.24 that, given a nonumbilic surface with
constant mean curvature, there are a circle’s worth of noncongruent surfaces
with the same Gauss curvature function and mean curvature as the given
surface. On the other hand, the functions H and K are usually sufficient
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 48 2. Euclidean Geometry

to determine M up to congruence. Those surfaces for which this is not the

case either have constant mean curvature or belong to a finite-dimensional
family called Bonnet surfaces, after Ossian Bonnet, who first investigated
them; see §7.4 for more discussion.
We will also see, in §2.6, that using slightly more information, namely
vector bundle valued differential invariants, one can always determine local
equivalence of surfaces from second-order information.

Flat surfaces. Recall that a surface is flat if K ≡ 0. The name is justified

by Theorem 2.6.2, which implies that the intrinsic geometry of such surfaces
is the same as that of a plane, and also by the following exercise:
Exercise 2.4.3: Show that if M is flat, there exist local coordinates x1 , x2
on M and an orthonormal frame (e1 , e2 , e3 ) such that ω 1 = dx1 , ω 2 = dx2 .


Here are some examples of flat surfaces:

Cylinders. Let C ⊂ E2 ⊂ E3 be a plane curve parametrized by c(u), and
assume X is a unit normal to E2 . Let f (u, v) = c(u) + vX.
Exercise 2.4.4: Find a Darboux framing for the cylinder and calculate its
Gauss and mean curvature functions.

Cones. Let C ⊂ E3 be a curve parametrized by c(u), and let p ∈ E3 \C. Let

f (u, v) = c(u) + v(p − c(u)). The resulting surface is called the cone over c
with vertex p.
Exercise 2.4.5: Show that cones are indeed flat. 

Remark 2.4.6. It turns out the property of being flat is invariant under a
larger group than ASO(3), and flat surfaces are best studied by exploiting
this larger group. This topic will be taken up in Chapter 4, where we classify
all flat surfaces: in the analytic category they are either open subsets of
cones, cylinders, or tangential surfaces to a curve. Even in the C ∞ category,
the only complete flat surfaces are cylinders; see [174], where there are also
extensive comments about the local characterization of flat C ∞ -surfaces.

2.5. Invariants for n-dimensional submanifolds of En+s

We already saw that for surfaces, the functions H, K alone were not sufficient
to determine equivalence. We now begin the study of vector bundle valued
functions as differential invariants for submanifolds of (oriented) Euclidean
space En+s = ASO(n + s)/SO(n + s).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.5. Invariants for n-dimensional submanifolds of En+s 49

Let (x, e1 , . . . , en+s ) denote an element of ASO(n + s). Define the pro-
jection
π : ASO(n + s) → En+s ,
(x, e1 , . . . , en+s ) → x.

Given an n-dimensional submanifold M ⊂ En+s , let π : F 1 → M denote the

subbundle of ASO(n + s)|M of oriented first-order adapted frames for M ,
whose fiber over a point x ∈ M is the set of oriented orthonormal bases such
that e1 , . . . , en are tangent to M (equivalently, en+1 , . . . , en+s are normal to
M ).
Using index ranges 1 ≤ i, j, k ≤ n and n + 1 ≤ a, b ≤ n + s, write the
Maurer-Cartan form on ASO(n + s) as
⎛ ⎞
0 0 0
ω = ⎝ ω i ωji ωbi ⎠ .
ω a ωja ωba

On F 1 , continuing our standard abuse in omitting pullbacks from the nota-

tion, ω a = 0 and thus dω a = −ωja ∧ ω j = 0, which implies

ωia = haij ω j ,

for some functions haij = haji : F 1 → R.

We seek quantities that are invariant under motions in the fiber Fx1 . The
motions in the fiber of F 1 are given by right-multiplication by
⎛ ⎞
1 0 0
(2.11) R = ⎝0 gji 0 ⎠ ,
0 0 uab

where (gki ) ∈ SO(n) and (uab ) ∈ SO(s). Let s : M → F 1 be an adapted lift

(a section of the bundle F 1 ); then any other adapted lift s̃ is of the form
s̃ = sR, and we compute

(2.12) s̃∗ (haij ) = (u−1 )ab gik gjl s∗ (hbkl ).

In this situation, if we were to look for scalar functions that are constant
on the fibers, we would get a mess, but there are simple vector bundle valued
functions that are constant on the fibers. Recall our general formula (1.21)
for how the Maurer-Cartan form changes under a change of lift. Under a
motion (2.11) we have

(2.13) ω i → (g −1 )ij ω j , ea → uba eb , ωia → gij (u−1 )ab ωjb .

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 50 2. Euclidean Geometry

To view our calculations more invariantly, let N M denote the normal bundle
of M , the bundle with fiber Nx M = (Tx M )⊥ ⊂ Tx En+s . Define
 : = ω a ω j ⊗ ea
II j

= haij ω i ω j ⊗ ea ∈ Γ(F 1 , π ∗ (S 2 T ∗ M ⊗N M )),


where we write ω i ω j for the symmetric product. Then (2.13) shows that II
is constant on fibers and thus is basic, i.e., if s1 , s2 : M → F are any two
1
 = s∗ (II).
sections, then s∗1 (II) 
2
Proposition/Definition 2.5.1. II  descends to a well-defined differential
invariant
II ∈ Γ(M, S 2 T ∗ M ⊗ N M ),
called the (Euclidean) second fundamental form of M .
When studying surfaces, we did not mention the vector bundle valued
first-order invariants of submanifolds described in the following exercises.
Exercises 2.5.2:
1. Consider

I˜ := ω i ω i ∈ Γ(F 1 , π ∗ (S 2 T ∗ M )).
i

Verify that I˜ descends to a well-defined differential invariant

I ∈ Γ(M, S 2 T ∗ M ),
which is called the first fundamental form or Riemannian metric of M .
2. Show that dvol := ω 1 ∧. . .∧ω n is invariant under motions in the fiber and
descends to a well-defined invariant, called the volume form of M . Show that
it is indeed the volume form induced by the Riemannian metric I. (Recall
that an inner product on a vector space V induces an inner product on Λn V ,
and thus a volume form up to a sign.)

Interpretations of II and K. We now give a more geometrical definition

of the second fundamental form for surfaces in E3 . This definition will be
extended to all dimensions and codimensions in §3.2.
The Gauss map. Let M 2 ⊂ E3 be oriented and let S 2 denote the unit sphere.
Since e3 is invariant under changes of first-order adapted frame, we obtain
a well-defined mapping
γ : M 2 → S2,
x → e3 (x),
called the Gauss map.
Proposition 2.5.3. II(v, w) = −de3 (v), we3 .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.5. Invariants for n-dimensional submanifolds of En+s 51

Thus, II admits the interpretation as the derivative of the Gauss map.

In what follows M 2 ⊂ E3 is a surface.
Exercises 2.5.4:
1. Prove Proposition 2.5.3.
2. Show that if the fibers of the Gauss map are all positive dimensional,
then they are (open subsets of) linear spaces, and thus flat surfaces are ruled
by lines. 
3. Show that if dim γ(M ) = 1, then the mean curvature H is not identically
zero.
4. Show that if there exists a line l ⊂ M , then Kx ≤ 0 for x ∈ l. If there
exist at least two distinct lines l1 , l2 ⊂ M and a point x ∈ (l1 ∩ l2 ) ⊂ M such
that Kx < 0, then either the lines are orthogonal or Hx = 0. If there exist at
least three lines passing through x and contained in M , then Kx = Hx = 0.
5. Calculate the images of the Gauss maps of the surfaces of revolution
studied in §2.2.
6. Let M be the hyperboloid x2 + y 2 = 1 + z 2 . Calculate the equation of
a line l on M passing through the point (1, 0, 0). Calculate γ(l), where γ is
the Gauss map of M .
7. Let M 2 ⊂ E3 be a surface with H = 0. Let gγ denote the Riemannian
metric induced on M via the pullback of the Gauss map from the unit sphere.
Let gγ be the pullback to M , via the Gauss map, of the standard metric on
the unit sphere S 2 . Thus in this situation, γ is a conformal mapping, i.e.,
angles are preserved: ∠(v, w) = ∠(dγx (v), dγx (w)) for all v, w ∈ Tx M . For
another proof, see §7.4.
Exercise 2.5.5: Let M n ⊂ En+1 be an oriented hypersurface. Define the
Gauss map of M and the analogous notions of principal curvatures, mean
curvature and Gauss curvature.
Remark 2.5.6 (A geometric interpretation of Gauss curvature). The round
sphere S 2 may be considered as the homogeneous space ASO(3)/ASO(2)
via the projection (x, e1 , e2 , e3 ) → e3 . As such, the form ω13 ∧ ω23 is the
pullback of the area from on S 2 because de3 = −(ω13 e1 + ω23 e2 ). Since
ω13 ∧ ω23 = Kω 1 ∧ ω 2 , we may interpret K as a measure of how much the area
of M is (infinitesimally) distorted under the Gauss map. This is because, for
a linear map A : V → V , the determinant of A gives, up to sign, the factor by
which volume is distorted under A. More precisely, if P is a parallelepiped
with one vertex at the origin, vol(A(P )) = | det A| vol(P ).

The Gauss equation. There is another way to calculate the Gauss curva-
ture of a surface, namely by differentiating ω12 . Using the Maurer-Cartan
equation, we obtain
(2.14) dω12 = −Kω 1 ∧ ω 2 ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 52 2. Euclidean Geometry

which is called the Gauss equation.

Exercise 2.5.7: Let c(t) ⊂ M 2 ⊂ E3 be a curve on a surface such that
|c (t)| = 1 and c (t) ⊥ Tc(t) M for all t. Show that |II(c (t), c (t)), e3 | =
κc(t) , where κ denotes the curvature of the curve.

The Euclidean Gauss Map.

Definition 2.5.8. Define the Gauss map of a submanifold M n ⊂ En+s as
follows: Let V = T0 En+s and define
γ : M → Gr(n, V )
by mapping x ∈ M to the translate of Tx M to the origin.

Figure 1. The shaded area of the surface maps to the shaded area
of the sphere

Exercise 2.5.9: Let M n ⊂ En+s . Show that dγ = II, the Euclidean second
fundamental form.

Thus if c(t) is a curve on M , then we may interpret II(c , c ) as measuring

how c is moving away from Tc(t) M infinitesimally.
Exercise 2.5.10: Suppose M contains a line l. Show that if x ∈ l, then
II(v, v)x = 0 if v is a tangent vector to l.

2.6. Intrinsic and extrinsic geometry

Definition 2.6.1. A Riemannian manifold is a differentiable manifold M
endowed with a smooth section g ∈ Γ(M, S 2 T ∗ M ) that is positive definite
at every point, called a Riemannian metric.

If M n ⊂ En+s , then the first fundamental form I is a Riemannian metric

on M . Which of our differential invariants for M depend only on the in-
duced Riemannian metric I? Such invariants are often called intrinsic, i.e.,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.6. Intrinsic and extrinsic geometry 53

depending only on the Riemannian structure of M , as opposed to extrinsic

invariants, which depend on how M sits in Euclidean space.

Intrinsic geometry of surfaces. By definition, I is intrinsic, while II is

necessarily extrinsic, since it takes values in a bundle that is defined only
by virtue of the embedding of M in to Euclidean space. However, one can
obtain intrinsic invariants from II. Given a surface M 2 ⊂ E3 , we have the
“great theorem” of Gauss:
Theorem 2.6.2 (Gauss’ Theorema Egregium). The Gauss curvature of a
surface M 2 ⊂ E3 depends only on the induced Riemannian metric.

Proof. Let f : U → M be a local parametrization, with U ⊂ R2 , and let

F : U → F 1 be a first-order adapted lift. Let η j = F ∗ (ω j ). Since η 1 ∧ η 2
is nonvanishing, it spans Λ2 Tx∗ M at each point. Thus there exist functions
a, b such that

dη 1 = aη 1 ∧ η 2 ,
dη 2 = bη 1 ∧ η 2 .

The proof is completed by the following exercises:

Exercises 2.6.3:
1. Show that there exists a 1-form α such that

dη 1 = −α ∧ η 2 ,
dη 2 = α ∧ η 1 .

2. If F̃ is a different adapted lift with η̃ j = F̃ ∗ (ω j ), show that dα̃ = dα.

Show that the function κ defined by dα = κη 1 ∧ η 2 is also unchanged, and
thus depends only on the Riemannian metric (η 1 )2 + (η 2 )2 .
3. Show that (η 1 , η 2 , α) = F ∗ (ω 1 , ω 2 , ω21 ), and thus κ = f ∗ K by (2.14). 

Let (M 2 , g) be a surface endowed with a Riemannian metric. Fix a local

orthonormal coframing η 1 , η 2 with dual framing e1 , e2 . Write η21 for the
form α in the proof of Theorem 2.6.2 and for notational convenience write
η12 = −η21 .
For X ∈ Tp M and Y ∈ Γ(T M ), write Y = y i ei for some functions y 1 , y 2 .
Define the vector

∇X Y := (dy 1 − y 2 η21 )(X)e1 + (dy 2 − y 1 η12 )(X)e2 .

Exercise 2.6.4: Show that ∇X Y is well-defined, i.e., independent of our

choice of orthonormal coframing.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 54 2. Euclidean Geometry

The differential operator ∇ : Γ(T M ) × Γ(T M ) → Γ(T M ) given by

(X, Y ) → ∇X Y is generalized to higher dimensions and studied in detail in
§3.1. Our main use of it in this chapter will be as follows:
First if c(t) is a curve on M , ∇c (t) c (t) is well defined:
Exercise 2.6.5: Show that if c : R → M is a smooth curve, then ∇c c is
well defined, despite c only being defined along c. 

Curves such that ∇c c = 0 are called geodesics. Define the geodesic
curvature κg (c(t)) := |∇c (t) c (t)|, a measure of the failure of c(t) to be a
geodesic.
The Codazzi equation. Given two functions H and K on an open subset U ⊂
R2 , does there exist (locally) a map f : U → E3 such that H and K are the
mean and Gauss curvature functions of M = f (U )? There are inequalities
on admissible pairs of functions because H, K are symmetric functions of
the principal curvatures (so, e.g., H = 0 implies K ≤ 0). However, as we
have already seen in (2.10), stronger restrictions are uncovered when one
differentiates and checks that mixed partials commute.
To derive these restrictions without assuming a Darboux framing, set
up an EDS on ASO(3) × R3 for lifts of surfaces equipped with second fun-
damental forms. Give R3 coordinates hij = hji and define

I = {ω 3 , ω13 − h11 ω 1 − h12 ω 2 , ω23 − h21 ω 1 − h22 ω 2 }diff

with independence condition Ω = ω 1 ∧ ω 2 . We calculate dω 3 ≡ 0 mod I, and

 3  1
ω1 ω
d −h
ω23 ω2
   3  1    1
0 −ω12 ω1 ω 0 −ω12 ω
=− 2 ∧ 3 − dh ∧ 2 +h 2 ∧
ω1 0 ω2 ω ω1 0 ω2
         1
0 −ω12 ω1 ω1 0 −ω12 ω
≡− 2 ∧h 2 − dh ∧ 2 +h 2 ∧ mod I.
ω1 0 ω ω ω1 0 ω2
Thus, h must satisfy the matrix differential equation
    1
0 −ω12 ω
(2.15) dh − h, 2 ∧ =0
ω1 0 ω2

(where [, ] denotes the commutator of matrices), which is called the Codazzi

equation.
Given a Riemannian metric g on M and an orthonormal framing, we saw
in the proof of Theorem 2.6.2 that F ∗ (ω12 ) is uniquely determined. Thus
we may interpret (2.15) as a system of equations for the possible second
fundamental forms II = hij ω i ω j⊗e3 for embeddings of M into E3 that induce
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.7. Curves on surfaces 55

the metric g. These restrictions are well-defined, since (2.15) is invariant

under changes of orthonormal framing.

2.7. Curves on surfaces

The interaction between the geometry of surfaces in E3 and the geometry
of curves lying on them was studied by early differential geometers such
as Dupin, Gauss, Minding and Monge (see [178] for more information).
We will use curves on surfaces to prove the Gauss-Bonnet Theorem and to
study Cauchy-type problems associated to the exterior differential systems
for surfaces in Chapter 6.
Let c(s) be a regular curve in E3 parametrized by arclength. Recall from
§1.8 that we can adapt frames so that
⎛ ⎞
0 0 0 0
⎜1 0 −κ 0 ⎟
d(c, T, N, B) = (c, T, N, B) ⎜
⎝0 κ 0 −τ ⎠ ds.
⎟

0 0 τ 0

Now say that c lies on a surface M ⊂ E3 . Let (e1 , e2 , e3 ) be a first-order

adapted lift of M (so e3 ⊥ T M ). Let θ denote the angle from e1 to T , and
let  be T rotated counterclockwise by π2 in Tp M , so that
    
T cos θ sin θ e1
= .
 − sin θ cos θ e2
Then the {e3 , }-plane is orthogonal to T . The angle between N and e3 is
traditionally denoted by ,1 so that
    
N cos  sin  e3
= .
B − sin  cos  
Since (T, , e3 ) gives an orthonormal frame of E3 , when we restrict this frame
to c we have
⎛ ⎞
0 0 0 0
⎜1 0 −κ̃g −κn ⎟
(2.16) d(x, T, , e3 ) = (x, T, , e3 ) ⎜
⎝0 κ̃g
⎟ ds
0 −τg ⎠
0 κn τg 0
for some functions κg (s), κn (s), τg (s). (This notation will become less mys-
terious in a moment.)
To interpret these functions, notice that
κ̃g = κ sin  = length of the orthogonal projection of κN onto ;
κn = κ cos  = length of the orthogonal projection of κN onto e3 .
1 This letter, pronounced “var-pi” by M. Spivak in our favorite introduction to differential

geometry [174], is not a sickly omega, but an alternate way of writing π.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 56 2. Euclidean Geometry

Exercise 2.7.1: Show that κ̃g = κg , the geodesic curvature.

In Exercise 2.8.11, you will show that geodesics are locally the curves
that are the shortest distance between two points on a Riemannian manifold.
Thus the notion of a small geodesic disk about a point, which we will use in
the proof of the Gauss-Bonnet Theorem, makes sense. Notice that if κg ≡ 0,
then all of the curvature of the curve lies in the normal direction, and  is
parallel to the binormal B of the curve.
Next, (2.16) shows that κn measures the curving of the surface in the
direction of T (by means of measuring how the surface normal e3 is bending);
it is called the normal curvature of the surface along c. Since it only depends
on the pointwise value of T , it is really an invariant of the surface. We say
c ⊂ M is an asymptotic line on M if κn ≡ 0.
Finally, Equation (2.17) shows τg measures what the torsion (as a curve
in R3 ) of a geodesic having tangent vector T would be; it is called the
geodesic torsion of c.
Exercises 2.7.2:
1. Show that
κn = −II(T, T ), e3 ,
(2.17) τg = II(, T ), e3 ,
(2.18) κg = −(dθ + ω12 )(T ). 
2. Show that κg ≡ 0 if and only if the osculating plane to c is perpendicular
to the surface at each point.
3. Show that ∇c c = κg (c).
4. Find formulas for κn , τg in terms of the principal curvatures k1 , k2 when
our surface is given a principal (Darboux) framing.
5. Find formulas for κn , τg when our surface is given a framing such that
e1 = T .
6. Calculate τg of a curve c such that c is a principal direction (i.e., a
direction where κn is a principal curvature) at each point along c. Such
curves are called lines of curvature.
Exercise 2.7.3: Prove the local Gauss-Bonnet Theorem: Let (M 2 , g) be
a Riemannian manifold and let U ⊂ M be a connected, simply connected
oriented open subset with smooth boundary ∂U . Show that

KdA = 2π − κg ds,
U ∂U

where dA is the area form on U , ds is the arclength measure on ∂U , and

∂U is oriented so that if T points along ∂U and N points into the interior
of U , then T ∧ N agrees with the orientation on U . 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.8. The Gauss-Bonnet and Poincaré-Hopf theorems 57

2.8. The Gauss-Bonnet and Poincaré-Hopf theorems

Given a compact oriented surface, the space of Riemannian metrics we can
put on it is infinite dimensional. With different metrics, the Gauss curvature
can have wildly different behavior. However, as we will see, the integral of
the Gauss curvature is independent of the metric and only depends on the
underlying topology of M .
First, we review some concepts from simplicial and differential topology.
A triangulation of the plane is the plane together with a set of triangular
tiles that fill up the plane. A triangulation of an open subset U in a surface
M is the pullback of some triangulation of a plane under a diffeomorphism
f : U → R2 such that ∂U is a union of edges, and a triangulation of a surface
is a covering by triangulated open sets such that the triangulations agree on
the overlaps.
Let T be a triangulation of M with V vertices, E edges and F faces.
Define
χ(M, T ) = V − E + F.
Exercise 2.8.1: Show that if T, T  are two triangulations of M , then
χ(M, T ) = χ(M, T  ). 

Since χ(M, T ) is independent of T , we will denote it by χ(M ).

Let X ∈ Γ(T M ) be a vector field with isolated zeros. Around such a zero
p ∈ M define the index of X at p as follows: Pick a closed embedded curve
retractible within M to p such that no other zero of X lies in the region U
enclosed by the curve. Choose a diffeomorphism f : U → D, where D is
the unit disc in E2 . Let θ be the counterclockwise angle between f∗ (X) and
some fixed vector v ∈ E2 . Define the integer
1
indX (p) = dθ,
2π ∂D
where the circle ∂D ⊂ E2 is oriented counterclockwise. The index is well-
defined by Stokes’ Theorem.
Intuitively, to obtain the index we draw a small circle around an isolated
zero. Travel around the circle once and count how many times the vector
field spins counterclockwise (counting clockwise spin negatively) when going
around the circle once. See [148] or [174] for more on indices of vector fields.
Suppose p1 , . . . , pr ∈ M are the isolated zeros of X. Define

χvf (M, X) = indX (pi ).
i
Below, we will indirectly prove that χvf (M, X) is independent of X. For a
direct proof, again see [148] or [174].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 58 2. Euclidean Geometry

Let M 2 be a compact oriented manifold without boundary. Let g be a

Riemannian metric on M and define
1
χmetric (M, g) = KdA.
2π M
Theorem 2.8.2 (Gauss-Bonnet and Poincaré-Hopf). Let (M 2 , g) be a com-
pact orientable Riemannian manifold, let T be a triangulation of M and let
X be a vector field on M with isolated zeros. Then
χmetric (M, g) = χ(M ) = χvf (M, X).

The equality χmetric (M, g) = χ(M ) is the Gauss-Bonnet Theorem, and

χvf (M, X) = χ(M ) is the Poincaré-Hopf Theorem. The common value of
these invariants is called the Euler characteristic of M , and is denoted by
χ(M ).
We first prove, for certain vector fields X, that χvf (M, X) = χ(M ). Next
we show that χmetric (M, g) = χvf (M, X), which, since the left hand side is
independent of X and the right hand side independent of g, shows that both
are well-defined. Combined with Poincaré-Hopf for certain vector fields, this
proves both theorems.
Just for fun, we afterwards give a direct proof that χmetric (M, g) = χ(M ).
(Actually, “we” here is a bit of a euphemism, as you, the reader, will do much
of the work in Exercise 2.8.3.)

Proof. Given a triangulation T , one can associate a vector field X to it

such that χvf (M, X) = χ(M ). Consider the following picture:

Notice that each vertex of the triangulation becomes a zero of index +1,
and each edge and face contains a zero of index −1 or +1, respectively.
To prove that χmetric (M, g) = χ(M ), we follow ([174], vol. III): Divide
M into two pieces, a subset U ⊂ M where X is complicated but the topology
of U is trivial, and M \U where X is simple but we know nothing about the
topology.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.8. The Gauss-Bonnet and Poincaré-Hopf theorems 59

Let p1 , . . . , pr be the zeros of X. Let Di () be an open geodesic disc (as

defined above) of radius  about pi , where  is small enough so that the discs
are contractible and don’t intersect each other. Let N () = M \( i Di ()),
a manifold with boundary. On N (), X is nonvanishing, so we may define
X
a global oriented orthonormal framing on N () by taking e1 = |X| with e2
1 2
determined by the orientation. We let η , η denote the dual coframing.
Now calculate


KdA = − dη12 = − η12 = η12 .
N( ) N( ) ∂N ( ) i ∂Di ( )

Let ẽj1 , ẽj2 be an oriented orthonormal framing in Dj (), and from now
on we suppress the j index. Let θ denote the angle between e1 and ẽ1 . Then
η̃12 = η12 − dθ wherever both framings are defined. Thus

KdA = lim KdA

M →0 N ( )

= lim η̃12 + dθ
→0 ∂Di ( )
i
 
(2.19) = lim KdA + dθ.
→0 Di ( ) ∂Di ( )
i i

Since K is bounded, as  → 0 the first expression in (2.19) tends to zero. As

 tends to zero the vector ẽ1 tends to a constant vector, so the second term
tends towards 2π times the index of X at pi . 

Exercise 2.8.3 (The Gauss-Bonnet formula): Let (M 2 , g) be an oriented

Riemannian manifold and let R ⊂ M be an open subset that is contractible
with ∂R the union of a finite number of smooth curves C1 , . . . , Cp , oriented
so that if T points along Cj and N points into R, then T ∧ N agrees with
the orientation on R. Let δi denote the angle between the terminal position
of the tangent vector to Ci and the initial position of the tangent vector to
Ci+1 (with the convention that Cp+1 = C1 ):
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 60 2. Euclidean Geometry

C2 δ1
δ2

C1
R
C3
δ4

δ3 C4

Prove the Gauss-Bonnet formula:


KdA = 2π − κg ds − δi .
R ∂R i

Then, use this formula to obtain a second proof of the Gauss-Bonnet The-
orem using the triangulation definition of the Euler characteristic.

Why are these theorems so wonderful? Take a plane in E3 , and draw a

circle in the plane. Now perturb the disk inside the circle—by stretching or
squashing, whatever you like—so that the boundary of the disk stays flat
(see Figure 2). What is the average curvature of the wildly curving surface

Figure 2. A distorted disk

you’ve made inside the circle? Zero! Next, take a round sphere, sit on it,
twist it, and fold it so it acquires regions of negative curvature. What’s the
average curvature of your distorted sphere? 4π, no matter how strong you
are!
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.8. The Gauss-Bonnet and Poincaré-Hopf theorems 61

Exercises 2.8.4:
1. Let M 2 ⊂ E3 . Show that the degree of the Gauss map is χ(M )/2.
2. What is ∂R κg ds, where R is the region enclosed by the lower dashed
curve (only half of which is pictured) in Figure 3?

Figure 3. What’s the average Gauss curvature?

Submanifolds of En+s . Consider the general frame bundle FGL (En+s ),

which we may identify with AGL(n + s) in the usual way. Let M n ⊂ En+s ,
and let F 1 ⊂ FGL (En+s ) be the bundle of first-order adapted frames, i.e.,
frames such that Tx M is spanned by e1 , . . . , en . Since the vectors e1 . . . en are
not necessarily orthonormal, the induced Riemannian metric on M takes the
form g = gij ω i ω j , where gij = g(ei , ej ). As a prelude to Chapter 4, define
the quotient normal bundle Ñ M as the bundle whose fiber at x ∈ M is
Tx En+s /Tx M . Then en+1 , . . . , en+s mod Tx M span Ñx M , but these vectors
are not necessarily perpendicular to Tx M .
Using index ranges as in §2.5, on F 1 we have
⎛ ⎞
0 0 0
(2.20) d(x, ej , ea ) = (x, ek , eb ) ⎝ω j ωkj ωbj ⎠ .
0 ωka ωba
(Note that the forms in the (n + s) × (n + s) submatrix at lower right are not
skew-symmetric.) Then the Maurer-Cartan equation 0 = dω a = −ωia ∧ ω i
again implies that ωia = haij ω j for some functions haij = haji on F 1 , so we
have
I = gij ω i ω j ∈ Γ(M, S 2 T ∗ M ),
II = ωja ω j ⊗ea ∈ Γ(M, S 2 T ∗ M ⊗ Ñ M ).
Now, for simplicity, assume M is a hypersurface. Fix an orientation
on M . Let N be a unit vector field perpendicular to the surface, and let
Q = II, N  ∈ Γ(S 2 T ∗ M ). Then the eigenvalues of g −1 ◦ Q are well-defined.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 62 2. Euclidean Geometry

These eigenvalues are the principal curvatures. The expression g −1 ◦ Q

should be understood as follows: Given a vector space V with a quadratic
form Q ∈ S 2 V ∗ , we may think of Q as a map V → V ∗ . Given a linear
map between two different vector spaces, it does not make sense to talk of
eigenvalues (in particular traces and determinants). But now say we have a
second, nondegenerate, quadratic form g ∈ S 2 V ∗ . We may think of g −1 as
a map g −1 : V ∗ → V and consider the composition g −1 ◦ Q : V → V . Then
eigenvalues are meaningful, and we can calculate the trace and determinant
of g −1 ◦ Q.
Now specialize further to the case where M is a surface.
Exercise 2.8.5: Show that
K = det(g −1 ◦ Q)
1
H = trace(g −1 ◦ Q). 
2
Coordinate formulas for H, K. Now we will finally prove the formulas (1.3).
Say M ⊂ E3 is given locally by a graph z = f (x, y), with f (0, 0) = 0 and
fx (0, 0) = fy (0, 0) = 0.
A simple coframing of T R3 along M is
ω 1 = dx,
ω 2 = dy,
ω 3 = dz − fx dx − fy dy.
Note that this coframing is first-order adapted in the sense that T M =
{ω 3 }⊥ . The dual framing is
e 1 = ∂ x + fx ∂ z ,
e 2 = ∂ y + fy ∂ z ,
e3 = ∂ z .
Then  
1 + fx2 fx fy
(gij ) = ,
fy fx 1 + fy2
and therefore
1
(2.21) dvolM = (det g) 2 ω 1 ∧ ω 2
1
= [(1 + fx2 )(1 + fy2 ) − (fx fy )2 ] 2 dx ∧ dy
1
= (1 + fx2 + fy2 ) 2 dx ∧ dy.

Computing de1 and de2 gives ω13 = d(fx ) and ω23 = d(fy ), so that h = (hij )
is just the Hessian of f .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 2.8. The Gauss-Bonnet and Poincaré-Hopf theorems 63

Exercise 2.8.6: Show that, relative to our framing for the graph z =
f (x, y),
   1
2 2 − 21 1 2 fxx fxy ω
Q = (1 + fx + fy ) (ω , ω ) ,
fyx fyy ω2
 
2 −1 1 + fx −fx fy
2
−1 2
(gij ) = (1 + fx + fy ) .
−fx fy 1 + fy2
Then confirm that H, K are given by (1.3). 

Geometric interpretation of H ≡ 0. Nonorthonormal frames are partic-

ularly useful if one wants to deform a submanifold.
Definition 2.8.7. M 2 ⊂ E3 is said to be minimal if for all x ∈ M there
exists an open neighborhood U , with x ∈ U ⊂ M , such that for any V ⊂ E3
that is a small deformation of U with ∂V = ∂U , we have area(V ) ≥ area(U ).
Theorem 2.8.8. M 2 ⊂ E3 is minimal if and only if H ≡ 0.

Proof. We show that minimal implies H ≡ 0. We use our area formula

(2.21) and deform the metric. We work locally, so we have U ⊂ R2 and
x : U → E3 giving the surface. Let u, v be coordinates on U .
Fix an orthonormal framing (e1 , e2 , e3 ) along x(U ) and let xt (u, v) be a
nontrivial deformation of x. For t sufficiently small, we may write
xt (u, v) = x(u, v) + t s(u, v)e3 (u, v) + O(t2 ),
where s : U → R is some function. For fixed t, calculate
dxt = dx + t e3 ds + t s de3 + O(t2 )
= e1 ω 1 + e2 ω 2 + t e3 (s1 ω 1 + s2 ω 2 ) − t s(e1 ω13 + e2 ω23 ) + O(t2 ).
Using the same ω 1 , ω 2 , we may write dxt = et1 ω 1 + et2 ω 2 where
et1 = (1 − t s h11 )e1 − t(s h12 e2 − s1 e3 ) + O(t2 ),
et2 = (1 − t s h22 )e2 − t(s h12 e1 − s2 e3 ) + O(t2 ).
This framing is no longer orthonormal. The metric now looks like
 
t 1 − 2t s h11 −2t s h12
g = + O(t2 ).
−2t s h12 1 − 2t s h22
1 
Exercise 2.8.9: Calculate dt d
|t=0 (det g t ) 2 and show dt
d
|t=0 U dvol(g t ) = 0
for all deformations if and only if H ≡ 0.

Now we wave our hands a little and ask you to trust that calculus in
infinite dimensions behaves the same way as in finite dimensions. That is,
a function has a critical point at a point where the derivative vanishes, and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 64 2. Euclidean Geometry

in our case its easy to see we are at a minimum. (If you don’t believe us,
consult a rigorous book on the calculus of variations, such as [72].) 
Exercise 2.8.10: More generally, for M n ⊂ En+s , define H  ∈ Γ(N M ), the
mean curvature vector, to be traceg (II). Show that M is minimal if and
 ≡ 0. In particular, show that straight lines are locally the shortest
only if H
curves between two points in the plane.
Exercise 2.8.11: More generally, let X n+s be a Riemannian manifold and
M n ⊂ X a submanifold, and the mean curvature vector define H  ∈ Γ(N M ),

to be traceg (II). Show that M is minimal if and only if H ≡ 0. In particular,
show that geodesics are locally the shortest curves between two points of X.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/03

Chapter 3

Riemannian Geometry

In this chapter we study local questions about Riemannian manifolds, em-

phasizing the use of representation theory to understand curvature. We
begin in §3.1 by establishing the fundamental lemma of Riemannian geom-
etry and the existence of the Levi-Civita connection from a moving frames
perspective. We state (but do not prove) the Gauss-Bonnet-Chern Theorem,
and state and prove the fundamental theorem for hypersurfaces in Euclidean
space.
We continue with standard exercises regarding Bianchi identities in §3.3,
followed by a brief discussion of the examples of the sphere and hyperbolic
space in §3.4. We then study representation theory relevant for Riemannian
geometry in §3.5. We discuss several topics in Riemannian geometry that
generalize to other G-structures: Killing vector fields in §3.6, homogeneous
Riemannian metrics in §3.7, and brief remarks on the Laplacian in §3.8.
Holonomy of Riemannian manifolds and symmetric spaces are covered in
Chapter 9. Our approach will facilitate the study of G-structures in Chapter
9.

3.1. Covariant derivatives and the fundamental lemma of

Riemannian geometry
We generalize the results of §2.6 to arbitrary Riemannian manifolds.

Intrinsic geometry in higher dimensions.

Frames for any manifold. Given an n-dimensional differentiable manifold
M n , let π : F (M ) → M denote the bundle whose fiber over x ∈ M is the set
of all bijective linear maps fx : Tx M → V , where V = Rn . We may write a

65
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 66 3. Riemannian Geometry

local section of F (M ) as s(x) = (x, fx ) = (x, η 1x , . . . , η nx ) where η i ∈ Ω1 (U )

for an open set U ⊂ M and the η ix form a basis of Tx∗ M for x ∈ U . In this
way, F (M ) may be considered as the bundle of all coframings of M . Since
the η i determine a dual framing (e1 , . . . , en ) of T M |U , we may also consider
F (M ) as the bundle of all frames of M .
The coframe bundle F (M ) is a right GL(V )-bundle: GL(V ) acts simply
transitively on the fibers by g · fx = g −1 ◦ fx . We now discuss to what extent
an analogue of the Maurer-Cartan form exists for F (M ).
Define the tautological V -valued 1-form η on F (M ): for f = (x, fx ) ∈
F (M ),
ηf (w) := fx (π∗ w), w ∈ Tf F (M ).
If we view η as a column vector of 1-forms, the components of η pulled back
to M via any smooth section are the η i introduced above.
The η i generalize the semi-basic forms ω i on the coframe bundle of Eu-
clidean space. We would also like to find analogues of the forms ωji , i.e.,
additional forms αji that fill out a coframing of F (M ) and satisfy

(3.1) dη i = −αji ∧ η j .

Forms αji satisfying (3.1) are called connection forms and we also write
α = (ei ⊗ η j ) ⊗ αji ∈ Ω1 (F , End(T M )). Such forms always exist; however,
unlike in the case of the Maurer-Cartan form, the αji will not be uniquely
defined unless we specify additional conditions beyond (3.1).
Exercise 3.1.1: Suppose η i is a local coframing defined on U ⊂ M . Define
a local trivialization
t : GL(V ) × U  F (M )|U
by (g, x) → g −1 η x . Show the existence of a desired coframing (η, α) on
F (M )|U as follows:
(a) Show that t∗ η = g −1 η.
(b) Show that there exist αji ∈ Ω1 (F (M )|U ) satisfying (3.1) such that
t∗ αji ≡ (g −1 dg)ij modulo {η i }. 
(c) Show that (η i , αji ) is a coframing of F (M )|U .
(d) Show that any other coframing satisfying (3.1) must be of the form
α̃ji = αji + Cjki η k for some functions C i = C i . 
jk kj

Frames for Riemannian manifolds. Over a Riemannian manifold M , con-

sider the bundle π : Fon (M ) → M of orthonormal coframes which we de-
fine as follows: endow V with an inner product ,  and define the fiber of
Fon (M ) ⊂ F (M ) over x ∈ M to be the linear maps (Tx M, gx ) → (V, , )
that are isometries. (We will sometimes denote the general frame bundle by
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.1. The fundamental lemma 67

FGL (M ) to distinguish it from Fon (M ).) The orthogonal group O(V ) acts
simply transitively on the fibers of Fon (M ).
Definition 3.1.2. For a bundle π : B → M , define the vertical tangent
space at u ∈ B to be ker π∗u ⊂ Tu B.
Exercise 3.1.3: Show that the vertical tangent space of Fon (M ) at each
point is isomorphic to so(V ). 

On Fon (M ) we have the tautological form η pulled back from F (M ). If

s : U → Fon (M ) is a smooth (local) section, then the forms η j = s∗ (η j )
provide a local coframing of U ⊂ M such that
g = (η 1 )2 + · · · + (η n )2 .

Thanks to the additional structure of the Riemannian metric, we can

add restrictions on the forms αji of (3.1) to make them uniquely defined as
follows:
Lemma 3.1.4 (The fundamental lemma of Riemannian geometry). Let M
be an n-dimensional Riemannian manifold and let η ∈ Ω1 (Fon (M ), V ) be
the tautological form. Then there exists a unique so(V )-valued 1-form θ ∈
Ω1 (Fon (M ), so(V )) such that
dη = −θ ∧ η.

In indices, writing η = (η i ), there exist unique forms ηji ∈ Ω1 (Fon (M ))

such that
(i) dη i = −ηji ∧ η j
and
(ii) ηji + ηij = 0.
To relate these perspectives, θ = (ei ⊗η j )⊗ηji where the ej are dual to
the η j and the factor in the parentheses is viewed as an endomorphism of
V.

Proof. On M , let η̄ be an orthonormal coframe, giving a local trivialization

of Fso (M ) such that η = g −1 π ∗ η̄ for some O(V )-valued function g on the
frame bundle. Then dη = −g −1 dg ∧ η + g −1 π ∗ dη̄. So, there exist 1-forms
αji on the frame bundle such that dη = −α ∧ η. Furthermore, if we write
αji = βji +γji where βji = −βij and γji = γij , then the γji are semi-basic because
g −1 dg is skew-symmetric. Let γji = Tjk i η k where T i = T j . As in Exercise
jk ik
i η k , where C i = −(T i + T i − T j ).
3.1.1.d, let α̃ji = αji + Cjk jk jk kj ki
i , the resulting α̃i are
Exercise 3.1.5: Show that with this choice of Cjk j
skew-symmetric, proving existence.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 68 3. Riemannian Geometry

To prove uniqueness, assume that ηji , η̃ji both satisfy (i) and (ii). The
Cartan Lemma applied to (η̃ji − ηji ) ∧ η j = 0 shows η̃ji − ηji = Cjk
i η k for some

i = C i . Moreover C i = −C j .
functions Cjk kj jk ik
i = 0.
Exercise 3.1.6: Show that Cjk 

In case M is a submanifold of En+r , our uniqueness argument implies

that s∗ (ηji ) = F ∗ (ωji ), where F is any extension of a local section s to a
first-order adapted framing F : M → FEn+r and the ωji are as in (2.20).
Note that if η i = s∗ (η i ), then one gets forms η ij defined by s∗ (θ) =
(ei ⊗η j )⊗η ij . Sometimes it is useful to work with the following “downstairs”
version of the fundamental lemma:
Lemma 3.1.7 (Fundamental lemma - downstairs version). Let (M n , g) be
a Riemannian manifold and let η i denote the tautological forms on Fon (M ).
Let U ⊂ M be an open neighborhood such that there exists a smooth section
s : U → Fon (U ), and for a differential form α on Fon (M ), denote s∗ (α) by
α. Then there exist unique forms η ij ∈ Ω1 (U ) such that
i. dη i = −η ij ∧ η j
and
ii. η ij + η ji = 0.
The ηji or η ij are referred to as connection forms. Note that η ij = Γijk η k
for some functions Γijk : M → R.
The upstairs version provides a coframing of Fon (M ), which we now
differentiate to obtain differential invariants. While dη i is given by (i) in
Lemma 3.1.4, we calculate dηji by using
(3.2) 0 = d2 η i = −(dηji + ηki ∧ ηjk ) ∧ η j .
Let Θ̃ij := dηji + ηki ∧ ηjk ∈ Ω2 (Fon (M )). The forms Θ̃ij are semi-basic by the
Cartan Lemma applied to Equation (3.2). Define
Θ̃ = (ei ⊗η j )⊗ Θ̃ij ∈ Ω2 (Fon (M ), π ∗ (End(T M ))) ,
where End(T M ) = T ∗ M ⊗ T M .
Exercise 3.1.8: Show that Θ̃ is basic, i.e., show that there exists Θ ∈
Ω2 (M, End(T M )) such that Θ̃ = π ∗ (Θ). 

The differential invariant Θ is called the Riemann curvature tensor.

Remark 3.1.9. The procedure we just followed, namely (i) define a sub-
bundle of F (M ) based on additional geometric structure, then (ii) attempt
to find a canonical coframing of this subbundle, and (iii) measure the failure
of the coframing to behave like the components of a Maurer-Cartan form,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.1. The fundamental lemma 69

will be generalized in Chapter 9. Riemannian geometry is special in that

the desired coframing exists and is unique.
Let so(Tx M ) be the space of endomorphisms A of Tx M that are com-
patible with gx , in the sense that
gx (Av, w) = −gx (v, Aw), ∀v, w ∈ Tx M,
and let so(T M ) ⊂ End(T M ) be the bundle with fiber so(Tx M ) at x. Then
Θ ∈ Γ(so(T M )⊗Λ2 T ∗ M ) = Ω2 (M, so(T M )).
Write
Θ = Θij ei ⊗η j
where 
1 i k
Θij = Rjkl η ∧ ηl = i
Rjkl ηk ∧ ηl
2
k>l
for some functions i
Rjkl : M → R.
The Riemannian metric g induces bundle isomorphisms  : T ∗ M → T M
and  : T M → T ∗ M defined by g(α, X) = α(X) and  is the pointwise
inverse. Tensoring with the identity map IdT ∗ M , we obtain a map ⊗IdT ∗ M :
so(T M ) → T ∗ M ⊗T ∗ M .
Exercise 3.1.10: Show that  ⊗ IdT ∗ M (so(T M )) = Λ2 (T ∗ M ).

Define R :=  ⊗ IdT ∗ M ⊗ IdΛ2 T ∗ M (Θ) and write

R = Rijkl (η i ∧ η j )⊗(η k ∧ η l ) ∈ Γ(M, Λ2 T ∗ M ⊗Λ2 T ∗ M ),
for some functions Rijkl .
The tensor field R ∈ Γ(M, Λ2 T ∗ M ⊗Λ2 T ∗ M ) is also called the Riemann
curvature tensor.
Exercises 3.1.11:
i . (If we
1. Show that, since we are using orthonormal frames, Rijkl = Rjkl
were using frames that were not orthonormal, then we would have Rijkl =
q
giq Rjkl ; see the end of this section.)
2. Show that when M is a surface, R1212 = K, the Gauss curvature.
3. Given Riemannian manifolds (M1 , g1 ) and (M2 , g2 ), form the product
Riemannian manifold (M1 × M2 , g1 + g2 ). Express the Riemann curvature
tensor of M1 × M2 in terms of R1 , R2 , the curvature tensors on M1 , M2 .
Definition 3.1.12. A Riemannian manifold (M n , g) is flat if in a neigh-
borhood of any point there exist local coordinates x1 , . . . , xn such that
g = (dx1 )2 + . . . + (dxn )2 . For example, the Riemannian metric on En
is flat.
Theorem 3.1.13. A Riemannian manifold (M n , g) is flat if and only if
Θ ≡ 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 70 3. Riemannian Geometry

Proof. First assume that dηji + ηki ∧ ηjk = 0. By Cartan’s Theorem 1.6.13,
around any point x ∈ M there exist an open set U and a map a : U → SO(n)
such that ηji = (a−1 da)ij on U . We have
dη j = −(a−1 da)jk ∧ η k .
Taking a new frame ηi = aij η j , where aij are the entries of a, we obtain
η i = daij ∧ η j − aij dη j
d
= daij ∧ η j − aij (a−1 da)jk ∧ η k = 0.
Thus ηi = dxi for some functions xi defined on a possibly smaller open set
U  . The other direction is straightforward 

Covariant differentiation. On a differentiable manifold, we have a well

defined derivative of any function f : M → R, namely df ∈ Ω1 (M ). However
we do not have a good notion of a “second derivative”. On a Riemannian
manifold we can obtain a well defined “second derivative”, as we now de-
scribe, however the price we will pay is that mixed partial derivatives will
no longer commute.
Say we have a fixed orthonormal coframe η i and we write df = fi η i for
some functions fi . Then we could compute dfi = fij η j for some functions
fij , and the tensor fij η i ⊗η j would be our naı̈ve candidate for the second
derivatives.
More generally, if β ∈ Ω1 (M ), and in our fixed coframe β = bj η j , we
could define N D(β) = η j ⊗ dbj (N D for “naı̈ve derivative”). Then in a
new frame η̃ j = (a−1 )ji η i for some a : M → O(n), we have b̃j = aij bi , and

N D(β) = η̃ i⊗db̃i = η j ⊗(dbj + bi (a−1 da)ij ). I.e., our naı̈ve derivative depends
on our choice of coframe. The difference of our two naı̈ve derivatives is
bi (a−1 da)ij ⊗η j . With the goal of finding a derivative independent of choice
of coframing, we search for a second quantity that varies in the same way
under a change of frame which we can use to cancel this term.
Exercise 3.1.14: Show that under a change of frame as above, η̃ji = (a−1 da)ij
+(a−1 )ik η il alj , so
η k ⊗bj η jk − η̃ k ⊗ b̃j η̃kj = η k ⊗bj (a−1 da)jk .
Definition 3.1.15. Given β ∈ Ω1 (M ) and a local orthonormal coframing
η j , write β = bj η j and write η ij for the pullbacks of the forms ηji from
Fon (M ). Define the covariant derivative
(3.3) ∇β := η j ⊗(dbj − bi η ij ) ∈ Γ(T ∗ M ⊗T ∗ M ).

By our discussion above, ∇β is independent of choice of orthonormal

coframing.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.1. The fundamental lemma 71

Exercise 3.1.16: Show that one can recover dβ from ∇β. 

Exercise 3.1.17: Show that if f is a smooth function on a Riemannian
manifold, then ∇df ∈ Γ(S 2 T ∗ M ).
Remark 3.1.18. Note that nothing we have done to define ∇ used the fact
that our connection form came from a Riemannian metric. In fact, given
any connection form θ ∈ Ω1 (F (M ), gl(V )) satisfying dη i = −θji ∧ η j on the
full frame bundle, then the covariant derivative defined as in (3.3), i.e.,
∇θ β = η j ⊗ (dbj − bi θji ),
is well-defined.
Define ∇ : C ∞ (M ) → Ω1 (M ) by ∇f = df . We may extend the definition
of ∇ to (any subbundle of) any tensor power of the cotangent bundle by the
Leibniz rule:
∇ : Γ(T ∗ M⊗k ) → Γ(T ∗ M⊗k ⊗T ∗ M )

k
bi1 ,...,ik η ⊗ · · · ⊗ η
i1 ik
→ η ⊗ · · · ⊗ η ⊗(dbi1 ,...,ik −
i1 ik
bi1 ,...,ij−1 , ,ij+1 ,...,ik ηij ).
j=1

We emphasize that the new 1-forms appearing in the tensor prod-

ucts generated by ∇ are put farthest to the right, e.g.,
∇(ai η i ⊗bj η j ) = η i ⊗bj η j ⊗(dai − ak ηik ) + ai η i ⊗η j ⊗(dbj − bk ηjk ).

We extend ∇ further to act on vector fields, i.e., ∇ : Γ(T M ) →

Γ(T M ⊗ T ∗ M ), by requiring it satisfy the compatibility condition that,
for any 1-form α,
∇(α(X)) = (∇α)(X, ·) + α(∇X).
In other words, for any vector field Y ,
Y d(α(X)) = (∇α)(X, Y ) + α(Y ∇X),
where Y pairs with the second factor of the tensor product in ∇X. Then
we again use the Leibniz rule to extend ∇ to act on sections of any bundle
constructed from tensor powers of T M and T ∗ M .
Exercise 3.1.19: Let X ∈ Γ(T M ) be a vector field and η i a local orthonor-
mal coframing with dual framing ei , so X = xi ei for some functions xi .
Show that
(3.4) ∇X = ei ⊗(dxi + xj ηji ) ∈ Ω1 (M, T M ) = Γ(T M ⊗T ∗ M ). 

Notation. For Y ∈ Γ(T M ), define ∇Y X := Y ∇X. In the above

notation, ∇Y X = [dxi (Y ) + xj ηji (Y )]ei . Let ∇Y : Γ(T M ) → Γ(T M ) denote
the corresponding differential operator.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 72 3. Riemannian Geometry

A wish list for our differential operator ∇. Here are properties we might like
∇ to satisfy for all X, Y ∈ Γ(T M ) and f ∈ C ∞ (M ):
(1) ∇(X + Y ) = ∇X + ∇Y (additivity).
(2) ∇X+Y = ∇X + ∇Y (linearity).
(3) ∇(f X) = X ⊗df + f ∇X (Leibniz rule).
(4) ∇f X Y = f ∇X Y , i.e., (∇X Y )p only depends on Xp and not on its
extension to a vector field.
(5) Mixed partials commute, in the sense that in coordinates ∇ ∂ ∇ ∂
∂xi ∂xj
−∇ ∂ ∇ ∂ .
∂xj ∂xi
Exercise 3.1.20: Show that the last property is equivalent to having ∇X ∇Y
−∇Y ∇X − ∇[X,Y ] = 0.

The first two properties are immediate from the definition. The next
two follow from the first two exercises below. The last property fails by
Exercise 3.1.21.4 below; however the weaker commutation of mixed partials
   
∂ ∂
∇ ∂ −∇ ∂ =0
∂xi ∂xj ∂xj ∂xi
holds. More generally,

(3.5) ∇X Y − ∇Y X = [X, Y ]

holds by Exercise 3.1.21.3 below. Note that the right-hand side of (3.5),
which is the Lie bracket of X and Y , is independent of the Riemannian
metric. A connection α whose associated differential operator satisfies (3.5)
is said to be torsion free.
Exercises 3.1.21:
1. Show that ∇Y (f X) = f ∇Y X + Y (f )X for f ∈ C ∞ (M ).
2. Show that ∇f Y (X) = f ∇Y X for f ∈ C ∞ (M ).
3. Show that (3.5) indeed holds. 
4. For X, Y ∈ Γ(T M ), define

Θ̃(X, Y ) = ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] .

Show that Θ̃ ∈ Γ(Λ2 T ∗ M ⊗ End(T M )) is a tensor, i.e. Θ̃(X, Y )Z is C ∞ -

linear in each of its three arguments. Show that in fact Θ̃ = Θ. 
5. Show that ∇ is compatible with the metric g in the sense that ∇g = 0.
(This property is equivalent to the requirement that the connection form θ
is so(V )-valued.)
6. Show that we can recover the connection forms ηji from the action of ∇
on vector fields. Because of this, ∇ is sometimes called a connection.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.2. Nonorthonormal frames and mean curvature 73

Thanks to Exercise 3.1.21.6, the covariant derivative on T M provides

another restatement of the fundamental lemma of Riemannian geometry as
follows:
Lemma 3.1.22 (Fundamental lemma, covariant derivative version). On a
Riemannian manifold (M, g) there exists a unique operator ∇ : Γ(T M ) →
Γ(T M ⊗ T ∗ M ) that extends to a differential operator on all bundles con-
structed tensorially from the tangent and cotangent bundles, satisfying, for
X, Y ∈ Γ(T M ) and f ∈ C ∞ (M ),
(1) ∇(X + Y ) = ∇X + ∇Y (additivity),
(2) ∇X+Y = ∇X + ∇Y (linearity),
(3) ∇(f X) = X ⊗df + f ∇X (Leibniz rule),
(4) ∇f X Y = f ∇X Y , i.e., (∇X Y )p only depends on Xp and not on its
extension to a vector field,
(5) ∇X Y − ∇Y X = [X, Y ] (torsion free),
(6) ∇g = 0 (compatible with the Riemannian metric).
The Fundamental Lemma of Riemannian Geometry 3.1.7 is sometimes
phrased that on a Riemannian manifold there exists a unique connection that
is torsion-free and compatible with the Riemannian metric. This connection
is called the Levi-Civita connection. Connections in general are discussed in
§9.4.
OK to here

3.2. Nonorthonormal frames and a geometric interpretation

of mean curvature
Nonorthonormal frames for Riemannian manifolds. Let M n be a
differentiable manifold and let F = FGL (M ) be the bundle of all framings of
M , as in §2.6. Suppose M happens to have a Riemannian metric g, but we
continue to use F (M ). (This will be desirable if, for example, we wish to
vary the metric on M .) We define the functions gij = gij (f ) := g(ei , ej ) on
F , where f = (x, e1 , . . . , en ) ∈ F . The fundamental lemma of Riemannian
geometry now takes the form:
Lemma 3.2.1 (Fundamental lemma, general frame version). There exist
unique forms ηji ∈ Ω1 (FGL (M )) such that
i. dη i = −ηji ∧ η j
and
ii. dgij = gik ηjk + gkj ηik ,
where η i are the tautological forms on FGL (M ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 74 3. Riemannian Geometry

Note that when gij = δij we recover the previous versions.

Exercise 3.2.2: Let (M, g) be oriented and let dvolM denote the induced
volume form. Show that
dvolM = det(gij )η 1 ∧ . . . ∧ η n ,
i.e., that the right hand side, which is defined on the frame bundle, descends
to be defined over M and to be the induced volume form. 

The Gauss-Bonnet-Chern Theorem.

Exercise 3.2.3: The Gauss-Bonnet Theorem for compact even-
dimensional hypersurfaces M n ⊂ En+1 is: let Kn denote the product of
the principal curvatures k1 , . . . , kn (i.e., the eigenvalues of h) and dV the
volume element. Then
1
Kn dV = vol(S n )χ(M ),
M 2
where vol(S n ) = 2n+1 π n/2 ( n2 )!/n! is the volume of the n-dimensional unit
sphere for n even. Prove the theorem, assuming the n-dimensional Poincaré-
Hopf Theorem. 

Here is a generalization by Chern [43] of the Gauss-Bonnet Theorem to

higher dimensions:
Theorem 3.2.4 (Gauss-Bonnet-Chern). Let M n be a compact, oriented
Riemannian manifold of even dimension. Then
Pfaff(Θ) = (2π)n χ(M ),
M
where Θij = 1
2 Rijkl η
k ∧ l
η.
The Pfaffian is defined in Exercise A.4.9.4. Note that, since the entries
of Θ are 2-forms and wedge products are commutative, the Pfaffian makes
sense with the multiplications being wedge products.
The proof of the Gauss-Bonnet-Chern Theorem is not too difficult; see,
e.g., [174]. The essential point for the proof is that if one puts two met-
rics g, g̃ on M , the forms Pfaff(Θ) and Pfaff(Θ) differ by an exact form and

therefore [Pfaff(Θ)] is well-defined as a cohomology class, i.e., M Pfaff(Θ) is
independent of the Riemannian metric. The same proof works for Riemann-
ian metrics on arbitrary vector bundles over M , giving rise to curvature rep-
resentations of the Euler class of a vector bundle (see [18] for the definition
of the Euler class). More generally, any elementary symmetric combination
of the eigenvalues of a skew-symmetric matrix with, e.g., positive imaginary
part leads to a characteristic class; that is, the corresponding cohomology
class obtained from Θ is independent of the Riemannian metric used. For
an excellent introduction to representing characteristic classes via curvature,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.2. Nonorthonormal frames and mean curvature 75

see [150, Appendix C]. There are further generalizations of Gauss-Bonnet-

Chern (e.g., the Atiyah-Singer index theorem), but discussion of them would
take us too far afield at this point; for further reading, see, e.g., [138].

The fundamental theorem for hypersurfaces. Let M n ⊂ En+s be a

submanifold. On the one hand, in §2.5 we derived differential invariants I, II
for the submanifold M . On the other, taking the Riemannian metric g = I
induced from the embedding, we obtain the Riemann curvature tensor. We
first observe that the Riemann curvature tensor (which one arrives at after
taking two derivatives at a point) is easily recovered from II (which one also
arrives at after taking two derivatives at a point):
Using the notation of §2.5 for M n → En+s , we take the orthonormal
coframe ω i so the ωji become connection forms. Then

Θij = −(dωji + ωki ∧ ωjk )

= −[−ωki ∧ ωjk − ωai ∧ ωja + ωki ∧ ωjk ]
= ωai ∧ ωja

= (haik ω k ) ∧ (hajl ω l )
a

= (haik hajl − hajk hail )ω k ∧ ω l .
a

Thus we easily recover the coefficients of Θ from those of II.

We formalize the above discussion by introducing the algebraic Gauss
map G, defined as follows:
Let V, W be vector spaces, where W has an inner product. Given a basis
ej of V ∗ and an orthonormal basis fa of W , let
G : S 2 V ∗ ⊗W → S 2 (Λ2 V ∗ ),

haij ei ej ⊗fa → (haij hakl − hail hajk )(ei ∧ ek )(ej ∧ el ),
a

where 1 ≤ i, j, k ≤ n = dim V and 1 ≤ a, b ≤ s = dim W and haij = haji . (If

we think of each ha as a matrix, the quantities in parentheses are the 2 × 2
minors.)
Exercise 3.2.5: Show that G is independent of the choices of bases.

Consider the local isometric embedding problem, which we will study

in detail in Chapter 6: Let (M n , g) be a Riemannian manifold. Can it be
locally realized as a submanifold of En+s so that g = I? (Note that g = I
implies G(h) = R.) Just by counting dimensions, the algebraic Gauss map
cannot possibly be surjective when n > 3 unless s is fairly large, namely at
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 76 3. Riemannian Geometry


least n2 . In this section we will deal with the especially restrictive case of
realizing M as a hypersurface.
We are now finally in a position to answer our original question regarding
the equivalence of surfaces in E3 (the case n = 2), and at the same time see
the generalization to hypersurfaces. On an oriented hypersurface in En+1 ,
the second fundamental form can be made scalar valued. Denote this by
h ∈ Γ(S 2 T ∗ M ).
Exercise 3.2.6: Show that when n = 2 the Codazzi equation (2.15) is
equivalent to ∇h ∈ Γ(S 3 T ∗ M ).
Exercise 3.2.7: More generally, for a hypersurface M n ⊂ En+1 , show that
h ∈ Γ(S 2 T ∗ M ) satisfies ∇h ∈ Γ(S 3 T ∗ M ). 
Theorem 3.2.8 (The fundamental theorem for hypersurfaces in En+1 ). Let
(M n , g) be a Riemannian manifold with curvature tensor R ∈
Γ(Λ2 T ∗ M ⊗Λ2 T ∗ M ), and let h ∈ Γ(S 2 T ∗ M ). Assume that
i. R = G(h) (Gauss)
and
ii. ∇h ∈ Γ(S 3 T ∗ M ) (Codazzi).
Then for every x ∈ M there exist an open neighborhood U containing x,
and an embedding f : U → En+1 as oriented hypersurface, such that f ∗ (I) =
g and f ∗ (II, en+1 ) = h, where en+1 is the outward unit normal vector
determined by the orientation. Moreover, f is unique up to a Euclidean
motion.
n
Corollary 3.2.9. Let M n , M ⊂ En+1 be two orientable hypersurfaces with
fundamental forms I, II and I, II. Suppose there exist a diffeomorphism
φ : M → M and unit vector fields en+1 , en+1 such that
i. φ∗ (I) = I,
and

ii. φ∗ II, ēn+1  = II, en+1 .
Then there exists a ∈ AO(n + 1) (the group ASO(n + 1) plus reflections)
such that φ = a |M .

Proof of 3.2.8. Let Σ = Fon (M )×AO(n+1)×S 2 Rn , where the last factor

has coordinates hij = hji . We use η’s to denote forms on the first factor and
ω’s for forms on the second factor and omit pullback notation in the proof.
Let Ω denote a volume form on Fon (M ) (e.g., wedge together all the entries
of the coframing of Fon (M )), and let

I = {η i − ω i , ηji − ωji , ω n+1 , ωjn+1 − hjk ω k }diff .

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.3. The Riemann curvature tensor 77

Integral manifolds of the EDS (I, Ω) are graphs of immersions i : Fon (M )|U →
AO(n + 1) that are adapted frame bundles of immersions j : U → En+1 , for
U ⊂ M , satisfying j ∗ (I) = g and j ∗ (II, en+1 ) = h. We obtain existence
by:
Exercise 3.2.10: Show that (I, Ω) is Frobenius. 

To prove uniqueness, fix an orthonormal frame (e1 , . . . , en ) at x. If two

such immersions j, j̃ exist, we can arrange (by composing j̃ with an element
of AO(n + 1)) that j(x) = j̃(x), j∗ ei = j̃∗ ei at x and the orientations at j(x)
match up. Thus, j and j̃ have lifts to Σ which are both integral manifolds
of (I, Ω) and pass through the same point (p, q) ∈ Σ. Thus, j = j̃ by the
uniqueness part of the Frobenius Theorem. 

3.3. The Riemann curvature tensor

The Cartan Lemma revisited. The Cartan Lemma (§A.1.9) may be
interpreted as follows: given X ∈ V ⊗V , if X is in the kernel of the skew-
symmetrization map V ⊗V → Λ2 V , then X ∈ S 2 V .
Let T = Tx M . Recall from §3.1 the Riemann curvature tensor Θx =
(dθ + θ ∧ θ)x ∈ so(T )⊗Λ2 T ∗ ⊂ T ⊗T ∗ ⊗Λ2 T ∗ satisfies Θx ∧ ηx = 0. In other
i (e ⊗η j )⊗η k ∧ η l is in the kernel of the skew-symmetrization map
words Rjkl i
T ⊗T ⊗Λ2 T ∗ → T ⊗Λ3 T ∗ .
∗

Exercise 3.3.1: Show that the above assertion implies:

i i i
(3.6) Rjkl + Rklj + Rljk = 0.

Equation (3.6) is called the first Bianchi identity.

We will be doing considerable symmetrization and skew-symmetrization
of tensors derived from the curvature tensor, and the positions we sym-
metrize and skew-symmetrize will be important, so we introduce notation
to keep track of position. The notation will be fully explained in §3.5. For
now, for a vector space V , we define a Young tableau (without repetitions)
of size d to be a collection of d left aligned boxes, each filled with dis-
tinct integers from the set {1, . . . , d}. The number of boxes in row j + 1
is at most the number of boxes in row j, and the tableau describes a sub-
space of V ⊗d by using the integers {1, . . . , d} to indicate which copy of V
in V ⊗ · · · ⊗ V = V1 ⊗V2 ⊗ · · · ⊗ Vd we are referring to. Any copies of V in
the same row of the tableau are symmetrized and then any positions in the
same column are skew-symmetrized. For example:
S1V = V,
S 1 2 V = S 2 1 V = S2V ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 78 3. Riemannian Geometry

S 1 V = S 2 V = Λ2 V .
2 1
Let S 2 1 V denote the kernel of the skew-symmetrization map S 1 V ⊗
3
S 2 V → S 2 V . Elements of S 2 1 V are tensors of the form u⊗v⊗w − u⊗w⊗
3 3 3
1
v + v⊗u⊗w − v⊗w⊗u. The first Bianchi identity implies Θx ∈ T ⊗S 2 1 T ∗ .
3

Note that when we lower indices via gx , Rijkl ∧ ∧ ηl ) ∈ (η i η j ) ⊗ (η k

∗ ∗
Γ(Λ T M ⊗ Λ T M ) (reflecting that so(T ) in bases consists of skew-
2 2

symmetric matrices). So we have R ∈ Γ(Λ2 T ∗ M ⊗ Λ2 T ∗ M ) ∩ (T ∗ M ⊗

S 2 1 T ∗ M ).
3

Exercise 3.3.2: Show that R ∈ Γ(S 2 (Λ2 T ∗ M )), i.e., Rijkl = Rklij .

Exercise 3.3.2 may be interpreted as saying that Rx ∈ S 1 3 T ∗ .

2 4
We explain Young tableaux and their uses in more detail in §3.5.

Extracting information from the Riemann curvature tensor.

Ricci, scalar, traceless Ricci, and Weyl curvature. In Riemannian geometry,
it is often convenient to work with smaller tensors, constructed from the
Riemann curvature tensor, that capture different ways a Riemannian metric
can fail to be flat.
Let Q ∈ S 2 V ∗ denote a nondegenerate quadratic form. (For the fol-
lowing decomposition, the signature of the inner-product does not matter.)
Recall that we may think of Q as a map V ⊗2 → R, via v ⊗ w → Q(v, w).
There are d2 distinct contractions from V ⊗d → V ⊗d−2 by feeding the vectors
in the i and j-th slots to Q, for 1 ≤ i < j ≤ d. However for tensors with
symmetries or skew-symmetries, some of the contractions may be redundant
or zero.
For example, let X ∈ V ⊗3 be X = u⊗v⊗w − u⊗w⊗v. Contracting the
first two factors gives Q(u, v)w − Q(u, w)v, contracting in the first and third
gives Q(u, w)v − Q(u, v)w, and contracting in the second and third gives
zero. Thus there is (up to a constant) only one useful contraction of X.
Fact: The decomposition of an irreducible GL(V ) module into irreducible
O(V )-modules is obtained by taking the images and kernels of such con-
tractions. When we restrict to GL(V )-submodules, many of these contrac-
tions will be redundant or zero. For example, restricted to Λd V , all such
contractions are zero, and thus Λd V is an irreducible O(V )-module. Re-
stricted to S d V , they are all the same, so S d V decomposes into S d−2 V and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.3. The Riemann curvature tensor 79

a component we will denote (S d V )0 (the “totally traceless” part). Contin-

uing, we see S d V decomposes into  d2  irreducible O(V )-modules, S d V =
S d V0 ⊕ S d−2 V0 ⊕ S d−4 V0 ⊕ · · · .
Ricci curvature.
 j
Definition 3.3.3. Let Rik = j Rijk . Define the Ricci curvature of M to
i k 2 ∗
be Ric = Rik η η ∈ Γ(S T M ).
Exercises 3.3.4:
1. Show that Ric is well-defined, i.e., independent of choice of orthonormal
coframe. 
2. Show that s := i Rii ∈ C ∞ (M ) is well-defined. It is called the scalar
curvature of M .
3. Show that when n = 2 (resp. 3) one can recover the full curvature tensor
R from s (resp. Ric), but this is not the case for n > 3.
4. Let Ric0 := Ric − ns g, which is called the traceless Ricci curvature. Show
that traceg (Ric0 ) = 0 so it lives up to its name.

Just as we can “subtract” the scalar curvature from Ricci to get a tensor
in a smaller space, we can subtract the Ricci from the Riemann curvature
tensor to get a tensor, called the Weyl curvature tensor and denoted W ,
which is the “totally traceless” part of the Riemann curvature tensor. The
formula for W is given in Equation (3.11) below. We will return to these
decompositions in §3.5.
Remark 3.3.5. Myers’ Theorem states that if for all x ∈ M and for all
v ∈ Tx M with gx (v, v) = 1, Ric(v, v) > C > 0, then M is compact; see, e.g.,
[149, Thm. 19.5]. The geometry and topology of manifolds with positive
scalar curvature is an active area of research; see e.g. [138].

Sectional curvature. The entire Riemann curvature tensor can be recovered

from just looking at curvatures of planes, in the following sense:
Definition 3.3.6. Fix x ∈ M . Let G(2, Tx M ) denote the Grassmann-
ian of two-planes in Tx M , and let v1 , v2 be an orthonormal basis of E ∈
G(2, Tx M ). Define the sectional curvature K : G(2, T M ) → R by K(E) :=
R(v1 , v2 , v1 , v2 ).
Exercises 3.3.7:
1. Show that K(E) is well-defined.
2. Show that one can recover R from K. 
3. What is the range of the sectional curvature function on
G(2, Tx (S 2 × S 2 )), where S 2 × S 2 has the product metric?
Remark 3.3.8. Sectional curvature has been studied considerably. The
Cartan-Hadamard theorem states that a Riemannian manifold with K ≤ 0
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 80 3. Riemannian Geometry

everywhere is diffeomorphic to Rn , and on it any two points are joined by

a unique geodesic (see, e.g., [149, Thm. 19.2]). A famous open conjecture
of Hopf asserts that there is no Riemannian metric on the differentiable
manifold S 2 × S 2 that has everywhere positive sectional curvature.
Remark 3.3.9. The scalar curvatures(x) is the average of the sectional
curvature in the sense that s(x) = Gr(2,Tx M ) K(E) dvol, where dvol is
the natural volume form on the orthogonal Grassmannian Gr(2, Tx M ) =
SO(Tx M )/(S(O(2) × O(n − 2)).

Covariant derivative of R. Write the covariant derivative of the curvature

tensor as
∇R = (∇R)ijklm (η i ∧ η j )⊗(η k ∧ η l )⊗η m .

Exercise 3.3.10: Show that ∇R satisfies the second Bianchi identity

(3.7) (∇R)ijklm + (∇R)ijmkl + (∇R)ijlmk = 0. 

The second Bianchi identity may be interpreted as saying that (∇R)x ∈

S1 ∗
3 5T .
2 4

3.4. Space forms: the sphere and hyperbolic space

The simplest examples of Riemannian manifolds beyond Euclidean space
are the sphere and hyperbolic space which we now describe.
Let S n ⊂ En+1 be the sphere of radius one, with its inherited met-
ric g. We have seen that En ∼ = ASO(n)/SO(n) as a homogeneous space.
Expressing E in this way facilitated a study of the geometry of its subman-
n

ifolds. We may similarly express S n as the quotient SO(n + 1)/SO(n). In

this manner, Fon (S n ) = SO(n + 1) with the base point projection given by
(e0 , e1 , . . . , en ) → e0 ∈ S n .
Let Ln+1 be Rn+1 equipped with a quadratic form
Q(x, y) = −x0 y 0 + x1 y 1 + . . . + xn y n
of signature (1, n). The space Ln+1 is called (n + 1)-dimensional Minkowski
space. Let O(V, Q) = O(1, n) denote the group of linear transformations
preserving Q (see Appendix A for details). Then
Ln+1 ∼
= ASO(1, n)/SO(1, n).
Define hyperbolic space to be
H n = {x ∈ Ln+1 | Q(x, x) = −1, x0 > 0}.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.4. Space forms: the sphere and hyperbolic space 81

Thus, H n may be considered as (one half of) the “sphere of radius −1” in
Ln+1 .
Exercise 3.4.1: Show that Q restricts to be positive definite on vectors
tangent to H n .

Thus, H n inherits a Riemannian metric from Ln+1 . Moreover, H n can

be expressed as the quotient SO(1, n)/SO(n). In this manner, Fon (H n ) =
SO(1, n), with the base point projection given by (e0 , e1 , . . . , en ) → e0 ∈ H n .

Let  = 0, 1, −1 respectively for X() = En , S n , H n . This handy notation

will enable us to study all three spaces and their submanifold geometry at
the same time. Then X() = G /SO(n), where G is ASO(n), SO(n + 1),
SO(1, n), respectively, with Lie algebra g = aso(n), so(n + 1), so(1, n). The
Maurer-Cartan form of G is
 
0 −ω j
ω=
ωi ωji

where 1 ≤ i, j ≤ n and ωji + ωij = 0.

As explained above, we identify G with Fon (X()), so that the ω i are
the tautological forms for the projection to X() and g := Σ(ω α )2 descends
to X() to give the Riemannian metric. The Maurer-Cartan equation for
dω i implies that the ωji are the (upstairs) Levi-Civita connection forms for
g. We also use the Maurer-Cartan equation to compute the curvature of
X():
dωji = −ωki ∧ ωjk − ω i ∧ (−ω j ).

Therefore, Θij = dωji + ωki ∧ ωjk = ω i ∧ ω j and

(3.8) Rijkl = (δik δjl − δil δjk ).

Exercises 3.4.2:
1. Show that the sectional curvature of X() is constant for all 2-planes.
(In particular, it is zero for En , positive for S n , and negative for H n .)
2. Compute Ric, s and Ric0 for the X()’s.
3. Compute ∇R for S n and H n .
4. Show that the above construction generalizes to space forms X(r) with
r ∈ R, i.e., replace  by an arbitrary real constant, and that the sectional
curvature function K : G(2, T M ) → R of a Riemannian manifold is constant
if and only if M = X(r) for some r, or an open subset thereof.
5. What are the geodesics in each of the X()’s? 
6. Let M n−1 ⊂ X n () be a hypersurface. Define its second fundamental
form and describe the hypersurfaces with II ≡ 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 82 3. Riemannian Geometry

7. Consider the surface S 1 × S 1 ⊂ S 3 defined by (x0 )2 + (x1 )2 = cos2 θ and

(x2 )2 + (x3 )2 = sin2 θ for some constant θ ∈ (0, π/2). Show that this surface,
which is known as a Clifford torus, is flat, and calculate its mean curvature.

Remark 3.4.3. In 1970 H. B. Lawson conjectured that any embedded min-

imal torus in S 3 is congruent to the Clifford torus. This conjecture was
proved in 2013 by S. Brendle [20].

Questions. In Exercise 3.4.2.2 we saw that space forms had zero traceless
Ricci curvature and constant scalar curvature. We ask more generally:
What are the Riemannian manifolds with s = 0?
What are the Riemannian manifolds with Ric0 = 0?
What are the Riemannian manifolds with W = 0?
What are the Riemannian manifolds with ∇R = 0?
These questions will lead us to study other geometries. For example, W is
related to conformal geometry, in fact metrics with W = 0 are conformally
equivalent to a flat metric; see §11.2. Metrics with Ric0 = 0 are solutions
of the following variational problem, which in the (1, n − 1) signature case
correspond to Einstein’s field equations for gravity, and so are called Einstein
(in any signature). For simplicity, only
 consider metrics on M with volume
equal to one, i.e., metrics g such that M dvolg = 1. Let S(g) := M sg dvolg ,
the integral of the scalar curvature. Then Einstein metrics are the critical
points for this functional.
Intuitively, metrics with ∇R = 0 are metrics where the curvature is in
some sense “constant”, and one would expect these to be homogeneous. In
Theorem 9.7.19 we will see something even stronger is true. We could also
ask for part of the curvature to be “constant”. For example, one can study
metrics with constant scalar curvature: this is related to the famous Yamabe
problem, to find a metric with constant scalar curvature that is conformally
equivalent to a given one (see [21] for a discussion). It turns out that there
is always such a metric, and moreover the metric is unique if the scalar
curvature is negative.
Exercises 3.4.4:
1. Assume n > 2. Show that if (M, g) is Einstein, then the scalar curvature
is constant.  Conclude that the only Riemannian manifolds with W = 0
and Ric0 = 0 are (open subsets of) spheres, planes, and hyperbolic space.
2. Calculate the Riemann curvature tensor, how it decomposes, and com-
pute ∇R for the product of two spheres S n (r1 ) × S N (r2 ) endowed with
the product metric. What are the geodesics? For which n, r1 , N, r2 is the
product metric Einstein?
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.5. Representation theory for Riemannian geometry 83

3. Show more generally that a product metric (M1 × M2 , g1 + g2 ) satisfies

Ric0 = 0 if and only if both original metrics are Einstein with dims1M1 =
s2
dim M2 .

3.5. Representation theory for Riemannian geometry

To better understand the symmetries of the curvature tensor and its co-
variant derivative, we take a short detour into representation theory, first
understanding the symmetries with respect to the general linear group and
then with respect to the orthogonal group.

Representation theory of the general linear group. We formalize the

discussion of §3.3 regarding Young tableau:
Definition 3.5.1. A Young tableau without repetitions Tπ associated to a
partition π = (p1 , . . . , pk ) of d is a collection of d left aligned boxes, with
pj boxes in the j-th row, and each box is filled with a distinct integer from
{1, . . . , d}. Since we will only discuss Young tableau without repetitions
here, we will just call Tπ a Young tableau for short.

For example, some Young tableau associated to π = (3, 2) are

5 3 1 1 2 3 1 2 5
2 4 , 4 5 , 3 4 .

To a Young tableau of size d, associate a GL(V )-submodule of V ⊗d as

follows: Let ρj : V ⊗d → V ⊗d denote the symmetrization of the slots corre-
sponding to the entries of the j-th row of a Young tableau Tπ and let σk
denote the skew-symmetrization of the slots corresponding to the entries of
the k-th column. (Here we write V ⊗d as V1 ⊗ · · · ⊗ Vd and the j-th slot means
Vj .) Let cTπ = σ1 ◦ · · · ◦ σp1 ◦ ρ1 ◦ · · · ◦ ρk : V ⊗d → V ⊗d . The map cTπ is called
a Young symmetrizer.
Facts: The space cTπ (V ⊗d ) is an irreducible GL(V )-submodule; see, e.g.,
[65, §6]. Write STπ V for cTπ (V ⊗d ). Different Young tableau from the same
partition give rise to isomorphic submodules, and the isomorphism class is
denoted Sπ V . We adopt the notation S∅ V = R for the trivial representation.
1 3
Example 3.5.2. If Tπ = 2 , then
1
cTπ (u⊗v⊗w) = σ12 [u⊗v⊗w + w⊗v⊗u]
2
1
= (u⊗v⊗w + w⊗v⊗u − v⊗u⊗w − v⊗w⊗u).
4
1 2
If Tπ = 3 , then cTπ (u⊗v⊗w) = 14 (u⊗v⊗w + v⊗u⊗w − w⊗v⊗u − w⊗u⊗v).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 84 3. Riemannian Geometry

We have the following interpretations of the Bianchi identities in terms

of representation theory:
Proposition 3.5.3. Let (M, g) be Riemannian and write R = Rijkl η i ∧ η j ⊗
η k ∧ η l and ∇R = (∇R)ijklm η i ∧ η j ⊗η k ∧ η l ⊗η m for the Riemann curvature
tensor and its covariant derivative. Let T ∗ = Tx∗ M for some x ∈ M . Then
(3.9) Rx ∈ S 1 3 T ∗ ,
2 4

(3.10) (∇R)x ∈ S 1 3 5 T ∗.
2 4

The Pieri rule (see, e.g., [65, §6]) implies that STπ V ⊗V ⊂ V ⊗d+1 decom-
poses into a direct sum all the ways of adding a box filled with the integer
d + 1 to Tπ and to still have a Young tableau.
Example 3.5.4.
S 1 3 V ⊗V = S 1 3 4 V ⊕ S 1 3 V ⊕ S 1 3 V.
2 2 2 4 2
4

While a priori ∇R could have nonzero components in any of the three

irreducible submodules of S 1 3 T ∗ ⊗T ∗ , the Bianchi identity says it lives in
2 4
the irreducible GL(T )-module S 1 3 5 T ∗.
2 4
It is remarkable that the homes of the Riemann curvature tensor and
its covariant derivative are both irreducible modules for GL(T ∗ ). Below we
compute their decompositions under O(T ).
Sometimes it will either be not important to keep track of positions, or
the positions will be understood, in which case we will simply write Sπ V for
the GL(V )-module corresponding to π, e.g., S32 V = S 1 2 3 V .
2 5
Dually, the decomposition of Sπ V ⊗V ∗ consists of all the modules whose
Young diagrams consist of the diagram of π with a box removed, plus a
component (Sπ V ⊗ V ∗ )0 , corresponding to the traceless part of Sπ V ⊗ V ∗ .
For example V ⊗V ∗ = R ⊕ (V ⊗V ∗ )0 corresponding to the splitting of the
space of endomorphisms of V into the multiples of the identity map and the
traceless endomorphisms. More generally Sπ V ⊗ Sμ V ∗ will have a similar
decomposition, subject to restrictions. For example S 2 V ⊗ Λ2 V ∗ = V ⊗
V ∗ ⊕ (S 2 V ⊗Λ2 V ∗ )0 = R ⊕ (V ⊗V ∗ )0 ⊕ (S 2 V ⊗Λ2 V ∗ )0 as a GL(V )-module.

Further decomposition under the orthogonal group. The contraction

of a module STπ V with the quadratic form is a module corresponding to a
Young tableau with two boxes in the same row erased. We denote the kernel
of the contraction by (STπ V )0 . This is consistent with our earlier notation
of taking trace, e.g., V ⊗V ∗ = (V ⊗V ∗ )0 ⊕ R as a GL(V )-module, because
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.6. Infinitesimal symmetries: Killing vector fields 85

we could view the contraction with the quadratic form as first identifying
one copy of V with V ∗ via the quadratic form and then taking the trace.
The contraction on S 2 V is just R.
As an O(T )-module T is isomorphic to T ∗ and we have the decomposition
S 1 3 T = (S 1 3 T )0 ⊕ (S 1 2 T )0 ⊕ R.
2 4 2 4

These are respectively the homes of Weyl, traceless Ricci and scalar curva-
ture.

To compute the Weyl tensor and determine the contribution of the Ricci
to the Riemann curvature tensor, we need to explicitly compute the inclusion
S 2 V ∗ ⊗S 2 V ∗ → S 2 (Λ2 V ∗ ). This is given by the Kulkarni-Nomizu product:
for h, k ∈ S 2 V ∗ and a basis v 1 , . . . , v n of V ∗ , define h  k ∈ S 2 (Λ2 V ∗ ) by
(h  k)ijkl = hil kjk − hik kjl + hjk kil − hjl kik. If h = k this is the map from
symmetric n × n matrices to symmetric n2 × n2 matrices given in bases
by taking the entries of the new matrix to be twice the 2 × 2 minors of the
original matrix. Then (when n > 2)
s 1
(3.11) R= gg+ Ric0 g + W.
2n(n − 1) n−2
When n = 2 the formula is R = 4s g  g.
Exercise 3.5.5: Verify that R − s
2n(n−1) g g− 1
n−2 Ric0 g ∈ (S 1 3 T ∗ )0 ,
2 4
which implies (3.11).
Exercise 3.5.6: What are the Weyl curvature tensors of the space forms
X()?

Sometimes it is useful to regroup the terms in (3.11). Define the Schouten

tensor by
s 1
(3.12) P:= g+ Ric0
2n(n − 1) n−2
1 s
= (Ric − g).
n−2 2(n − 1)
The advantage of the Schouten tensor is that R = W + P  g.

3.6. Infinitesimal symmetries: Killing vector fields

3.6.1. Flows that preserve a Riemannian metric. We now study Rie-
mannian manifolds (M, g) that admit continuous symmetries, that is, such
that there exists a one-parameter group of diffeomorphisms φt : M → M
with φ0 = Id that preserve the metric. It is easier to study this question
infinitesimally, so we look for a vector field X ∈ Γ(T M ) such that LX g = 0,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 86 3. Riemannian Geometry

where LX is the Lie derivative along X. The flow of X will provide the de-
sired one-parameter group and X = dφ dt |t=0 . We first derive the differential
t

equation for X satisfying LX g = 0 (Proposition 3.6.3), and then we set up

and analyze the EDS for such X, culminating in Theorem 3.6.5, which dates
back to Killing.
Definition 3.6.1. On a Riemannian manifold (M, g), a vector field X such
that LX g = 0 is called a Killing vector field.

Say we have a local orthonormal coframing (η 1 , . . . , η n ), so g = i (η i )2 .
Then

(3.13) LX g = 2 (X dη i )  η i + d(η i (X))  η i .
i
We first interpret the condition that LX g = 0. We work on the frame
bundle. To do so we need the following definition:
Definition 3.6.2. Let X ∈ Γ(T M ) be a vector field. Define the horizontal
lift X̃ ∈ Γ(Fon (M )) to be the vector field satisfying ηji (X̃) = 0, and π∗ (X̃) =
X. Thus we may write X̃ = xi ei , for some functions xi on Fon (M ), where
η i (ek ) = δki and ηji (ek ) = 0 for all i, j, k.

The (lifted) right hand side of (3.13) becomes


(3.14) (xj ηji + dxi )  η i .
i

This looks a lot like ∇X = ei ⊗(xj η ij + dxi ) from (3.3). In fact, recalling the
isomorphism  : Tx M → Tx∗ M induced by gx and Exercise 3.1.10, we have
(3.15) LX g = c 1 2 ◦ (⊗IdT ∗ M )(∇X).
Thus LX g = 0 if and only if (⊗IdT ∗ M )∇X is skew-symmetric. In summary:
Proposition 3.6.3. LX g = 0 if and only if ∇X ∈ Γ(so(T M, g)).
Exercises 3.6.4:
1. Show that the vector space of Killing vector fields is a Lie subalgebra of
Γ(T M ).
2. Show that LX = ∇X − ∇X, e.g., LX Y = ∇X Y − ∇Y X for Y ∈ Γ(T M )
and LX α = ∇X α − α ◦ ∇X for α ∈ Ω1 (M ).  In this formula ∇X is a
differential operator on sections of a tensor bundle and ∇X is the endomor-
phism of the tensor bundle induced from an element of gl(Tx M ). (Note:
this explains why, although in general to take a Lie derivative one must look
at a tensor along a curve, for LX g we do not, as ∇X g = 0 for all X, i.e.,
LX g = (∇X).g.) This observation is due to Kostant [116].
3. Use the previous exercise to obtain a second proof that LX g = 0 if and
only if ∇X ∈ Γ(so(T M )).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.6. Infinitesimal symmetries: Killing vector fields 87

3.6.2. An EDS for Killing vector fields. Given (M, g), we now set up an
EDS for Killing vector fields X ∈ Γ(T M ). We will solve for the horizontal lift
X̃ of a Killing vector field X ∈ Γ(T M ) to the orthonormal coframe bundle
F . Since the condition involves X̃ and its derivatives, we introduce new
2
variables to account for them: we work on a submanifold of F × Rn × Rn ×
n
Rn( 2 ) where the last three factors have coordinates (xi , xij , xi j ) respectively,
(k)
with all indices running from 1 to n and xi j = −xi k . On any integral
(k ) (j )
manifold, the xi will be functions such that X̃ = xi ei , and the remaining
variables will be such that dxi = xij η j + 12 xijk ηkj . Thus, our EDS will include
differential forms dxi − xij η j − 12 xijk ηkj whose vanishing ensures this.
Note that our initial value problem will be specifying the value of X at
a point and the value of its first derivatives at that point. The questions
will then be: is there an extension of X to a Killing vector field, and if there
is, is the extension unique?
Now we introduce the Killing condition, i.e., we restrict to Σ ⊂ F ×Rn ×
n
R × Rn( 2 ) defined by (3.14) assuming dxi = xij η j + 12 xijk ηkj , i.e.,
n2

 1
(3.16) 0= ((δki xj + xi k )ηjk + xik η k )  η i .
2 (j )
i

More precisely, since the forms ηjk , η i are all linearly independent, for (3.16)
to hold we must have
(3.17) xi k = −δki xj + δji xk ,
(j )
(3.18) xik + xki = 0,
n
and these are the defining equations for Σ. Thus Σ = F × Rn × R( 2 ) , where
n
R( 2 ) has coordinates xij = −xji . Set θi := dxi − xij η j + xj ηji . Our Pfaffian
EDS on Σ is generated by
(3.19) I = {θi }.
Since π ∗ (∇X) = ei ⊗(dxi + xj ηji ) ≡ ei ⊗xij η j mod I, where π : F → M is the
projection, we see that we are indeed requiring that ∇X be so(V )-valued.
Recalling dη j = −ηkj ∧ η k and dηji = −ηki ∧ ηjk + 12 Rjkl
i η k ∧ η l , we compute

1 i k
dθi = −dxij ∧ η j + xij ηlj ∧ η l + dxj ∧ ηji + xj (−ηki ∧ ηjk + Rjkl η ∧ ηl )
2
1
(3.20) ≡ [−dxil + xij ηlj − xjl ηji + xj Rjkli
η k ] ∧ η l mod I.
2
Now if we add
xj (− 12 Rjki − 12 Rjk i )η k
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 88 3. Riemannian Geometry

to the term in the brackets, it does not change the expression because
[xj (− 12 Rjki − 12 Rjk i )η k ] ∧ η l = 0. But now
1
−dxil + xij ηlj − xjl ηji + xj Rjkl
i
η k + xj (− 12 Rjki − 12 Rjk i )η k
2
is skew in i and l. Applying the Bianchi identity, this becomes
1
−dxil + xij ηlj − xjl ηji − xj Rljk
i
ηk ,
2
and these forms must vanish on any integral manifold, so we add the forms
1
θli := −dxil + xij ηlj − xjl ηji − xj Rljk i
ηk
2
to our ideal and start over. Let I˜ = {θi , θli }. We have arranged such that
dθi ≡ 0 mod I,˜ and it remains to compute the derivative of θi :
l
1 i
dθli = 0 + dxij ∧ ηlj + xij dηlj − dxjl ∧ ηji − xjl dηji + Rljk dxj ∧ η k
2
1 1
+ xj dRljk i
∧ η k + xj Rljk i
dη k
2 2
1
(3.21) ≡ − xj [(∇R)iljku η u ] ∧ η k
2
+ [xit Rlsj
t
− xtl Rtsj
i
− xts Rltj
i
− xtj Rlst
i
]η s ∧ η j mod I.

Let z = xij ei ⊗η j and rewrite Equation (3.21) as

(3.22) [X 3 (∇R)il•km + (z.R)ilkm ]η k ∧ η m ,
where 3 means contract X with the third factor in ∇R (indicated by the
dot).
The EDS becomes Frobenius if and only if the term in the brackets of
(3.22) is zero for all i, l, k, m. That is, there is at most one integral manifold
through a point where we specify X and the component of its first derivative
we are naı̈vely allowed to specify, and such an integral manifold exists if this
term is zero. Thus if we are at a point of Σ where xi , xij are such that X
contracted in the third slot of ∇R is zero and z.R = 0, there exists a unique
integral manifold. Note that although we are working at one point, if X is
defined everywhere, then it is determined by this data at just one point. In
summary:
Theorem 3.6.5. Let (M, g) be a Riemannian manifold of dimension n.
Then the space of Killing vector fields on M has dimension at most n + n2
with equality holding if and only if M is a space form.

Proof. If M has constant sectional curvature, then R = sg g and ∇R = 0.

Then z.R = s[(z.g)  g + g  (z.g)] = 0 as so(V ) annihilates the metric. To
see the other direction, so(V ) only annihilates trivial representations, in
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.7. Homogeneous Riemannian manifolds 89

particular, so(V ) will not annihilate any nonzero element of (S 1 3 V )0 or

2 4
(S 1 2 V )0 where Weyl and traceless Ricci live. So there will be some Z such
that Z.R = 0, and for that Z (which is to say most Z), we are no longer
free to choose X arbitrarily. 
Example 3.6.6. Consider the special case of flat space. Fix constants
xi , xij and coordinates (y 1 , . . . , y n ) on En , such that (dy 1 , . . . , dy n ) is an
orthonormal coframing. Then the general Killing vector field is of the form
∂
X = (xj + y i xji )
∂y j
which integrates to a translation plus a rotation.

3.6.3. Exercises.
(1) Give geometric interpretations in S n and H n of the Killing vector
fields at a point x that fix x.
(2) Let X ∈ K and show LX ◦ ∇ = 0.

3.7. Homogeneous Riemannian manifolds

Orthogonal Grassmannians and the Euclidean Gauss map.
Definition 3.7.1. Let V be equipped with an inner product and orientation.
Define the (orthogonal) Grassmannian
Gr(k, V ) = SO(n)/S(O(k) × O(n − k)),
where
S(O(k) × O(n − k))
 
A 0 
= A ∈ O(k), B ∈ O(n − k), det(A) det(B) = 1 .
0 B 
Exercises 3.7.2:
1. Show that, as differentiable
 manifolds, Gr(k, V ) = G(k, V ) := GL(V )/Pk ,
∗ ∗
where Pk = , and the blocking is (k, n − k). 
0 ∗
2. Show that Gr(k, V ) inherits additional structure from the metric on
V , including a Riemannian metric such that under any local section σ :
Gr(k, V ) → SO(V ), the forms σ ∗ (ωjt ), with 1 ≤ j ≤ k, k+1 ≤ t ≤ n, provide
an orthonormal coframing. Calculate the Riemann curvature tensor, as well
as the scalar, traceless Ricci and Weyl curvatures and ∇R for this metric.
3. Calculate the induced Riemann curvature tensor on Gr(k, V ) via its
Plücker embedding and compare it to the one above. (Note that Λk V
has an inner product on it induced from the one on V which induces a
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 90 3. Riemannian Geometry

Riemannian metric on PΛk V .) Calculate the sectional curvature function

K : G(2, Tp G(k, V )) → R.
4. Show that there is a canonical identification of the tangent space to
Gr(k, V ) at E with the space of linear maps E → E ⊥ .

Homogeneous Riemannian manifolds in general. This section follows

unpublished notes of R. Bryant.
Definition 3.7.3. Let G be a Lie group. A Riemannian manifold M is
G-homogeneous if it admits a transitive left G-action such that L∗a (g) = g
for all a ∈ G.

A G-homogeneous Riemannian manifold M may be written M = G/H

where H is the stabilizer of a point x ∈ M . The stabilizers are all isomorphic
groups, so the choice of point simply distinguishes x as the image of the
identity e ∈ G under the quotient G → G/H.
Exercise 3.7.4: With notation as above, show that the stabilizer of y ∈ M
is aHa−1 where La (x) = y.

If H is a compact Lie group and ρ : H → GL(V ) is a representation,

then one may obtain an H-invariant inner product on V by taking any inner
product and averaging it over H.
Write μ : H → GL(g/h) and μ : h → gl(g/h) for the adjoint actions of
H and h on T[e] G/H  g/h.
Proposition 3.7.5. Assume the map μ : h → gl(g/h) is injective (to avoid
trivialities). Then there exists a G-invariant metric on M = G/H if and
only if H is compact.

Proof. If H is compact, consider the H action on g/h  T[e] G/H. Take an

H-invariant inner product and translate it by the G action to get a metric on
G/H, i.e., given g[e] ∈ ∗ G/H,
S 2 T[e] define g[a] (v[a] , w[a] ) :=
g[e] (La−1 ∗ v[a] , La−1 ∗ w[a] ) and observe that this is independent of the choice
of a ∈ [a].
Now say there is a G-invariant metric on M ; then the action of H on
T[e] G/H preserves the metric at this point, so μ(H) ⊆ O(T[e] G/H). Since
O(T[e] G/H) is compact and μ(H) is a closed subgroup, we conclude μ(H)
is compact, but the kernel of μ is discrete so H is also compact. 

From now on, if H is assumed to be compact, motivated by mnemonics,

we will denote it instead by K.
On a Lie group G, the G-left-invariant metrics coincide with elements of
S+2 g∗ = S 2 T ∗ G, as if we take any inner product on g, it extends uniquely to
+ e
a metric on G by left translation. To obtain a left invariant metric on G/K
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.7. Homogeneous Riemannian manifolds 91

we may do the same, however we further need the inner product on g/k to
be K-invariant in order to left translate unambiguously, as in Proposition
3.7.5.
Notation. For a K-module W , W K denotes the K-invariant vectors in
W , the isotypic component of the trivial representation in W .

Proposition 3.7.6. Let MG denote the space of G-left invariant Riemann-

ian metrics on M = G/K. Then MG = S+ 2 (g/k)K .

Exercise 3.7.7: Show that if W is an irreducible K-module, with K com-

pact, then (S 2 W ∗ )K is one-dimensional. 
Exercise 3.7.8: If W is reducible but multiplicity free, show that
dim(S 2 W ∗ )K equals the number of irreducible components
 of W . If W =
U ⊕ m with U irreducible, show dim(S 2 W ∗ )K = m+1 , and in general, show
!r mj
2
that if W = j=1 Uj is the isotypic decomposition, then dim(S 2 W ∗ )K =
r mj +1
j=1 2 .

Example 3.7.9. The unitary group U (n) is defined in §A.2. Let

T
G = SU (n + 1) = {A ∈ M atn+1 (C) | AA = Id, detC A = 1},
 
detC (B)−1
K= | B ∈ U (n)}  U (n).
B
Write the map G → G/K as (e0 , . . . , en ) → [e0 ], where [v] = [w] if and only
if w = ζv for some ζ ∈ C with |ζ| = 1, so the image is S 2n+1 /S 1  CPn .
The group F = {ζ Idn+1 | ζ n+1 = 1} acts trivially on G/K = CPn . Here
g/k  Cn is an irreducible k-module so there will be a one-parameter family
of metrics.

Exercise 3.7.10: Compute the Riemann curvature tensor for each metric
in this one-parameter family as well as its decomposition, and ∇R. Which
metrics are Einstein?
 
1
Example 3.7.11. Let G = SU (n + 1) and K = | B ∈ SU (n)} 
B
SU (n). Take the map G → G/K, (e0 , . . . , en ) → e0 so G/K = S 2n+1 . Write
 
ix −z T
g= | x ∈ R, z ∈ Cn , X ∈ su(n) .
z − ix
n Id n +X
Then (g/k)∗  R ⊕ Cn , so there is a two-parameter family of SU (n + 1)-
invariant metrics on S 2n+1 .

Remark 3.7.12. The maps of Examples 3.7.9 and 3.7.11 are compatible.
Factoring we obtain the Hopf fibration S 1 → S 2n+1 → CPn .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 92 3. Riemannian Geometry

Exercise 3.7.13: Compute the Riemann curvature tensor for each metric
in this two-parameter family as well as its decomposition and ∇R. Which
metrics are Einstein?

3.8. The Laplacian

Invariant differential operators have several uses in differential geometry.
One can use them to obtain information about topology and curvature.
Already on a C ∞ -manifold, there is an invariant differential operator d =
dk : Ωk (M ) → Ωk+1 (M ). De Rham proved that one can use this to study
topology: Let HdR k (M ) := ker d / Image d k
k k−1 . Then HdR (M ) is a finite-
dimensional vector space naturally dual to Hk (M ), the singular homology
group, where the pairing is [α], [X] := X α. By Stoke’s Theorem, this
pairing is independent of the choice of representative.
On a Riemannian manifold (M n , g) with a volume form one has more
differential operators defined via the metric. Recall the star operator ∗ :
Ωk (M ) → Ωn−k (M ) defined in Appendix A. Define a first-order operator
δ : Ωk (M ) → Ωk−1 (M ) by
δα = (−1)n(k+1)+1 ∗ d ∗ α, α ∈ Ωk (M ), k ≥ 0.
(This is the adjoint of d with respect to the L2 inner product on k-forms
when M is compact [187].)
Now define a second-order differential operator, the Laplacian, by Δg α =
(dδ + δd)α.
For example, let f ∈ C ∞ (Rn ) = Ω0 (Rn ), where Rn has coordinates
x1 , . . . , xn .
Then
δf = 0,
df = fxi dxi ,
∗df = (−1)i fxi dx1 ∧ · · · ∧ dxi−1 ∧ dxi+1 ∧ · · · ∧ dxn ,
d ∗ df = fxi xi dx1 ∧ · · · ∧ dxi−1 ∧ dxi+1 ∧ · · · ∧ dxn ,
so

Δf = − fxi xi ,
i

and the operator agrees with the classical Laplacian (up to a sign which is
inserted to make the operator have nonnegative eigenvalues).
Since Δg : Ωk (M ) → Ωk (M ) is a map from a vector space to itself, we
can talk about eigenvalues and eigenvectors (eigenfunctions).
What is the geometric meaning of these eigenfunctions?
 In general, they
are critical for the “energy” functional E(α) = M (α, α)g dvolM .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 3.8. The Laplacian 93

Already the growth of the eigenvalues of the Laplacian on scalars con-

tains information about the curvature and shape of M .
The famous Hodge Theorem (see, e.g., [187, 102]) states that a basis of
ker Δ ⊂ Ωk (M ) (called the space of harmonic forms) descends to span the
de Rham cohomology HdR k (M ), so Harmonic forms give preferred represen-

tatives of cohomology classes.

The Laplacian and classical differential geometry. When (η 1 , . . . , η n )

is a coframing of M , recall that we write df = fj η j .
Exercise 3.8.1: (a) If M 2 ⊂ E3 is a surface and (e1 , e2 ) is a Darboux
framing, with principal curvature functions k1 , k2 , show that
f1 k2,1 − f2 k1,2
−Δg f = f11 + f22 + .
k1 − k2
(b) If g is a flat metric on M and (x1 , . . . , xn ) are coordinates such that dxj
gives an orthonormal coframing, show that
−Δg f = f11 + . . . + fnn .
(c) If x : M 2 → En is an isometric immersion (i.e., the metric g on M
agrees with the pullback from En ), then calculating the Laplacian of each
component of x as a vector-valued function gives

Δg x = −2H,
 = traceg II is the mean curvature vector.
where H

Isothermal coordinates. Let (M 2 , g) be a Riemannian manifold with coordi-

nates (x, y). Write g = a(x, y)dx2 + b(x, y)dxdy + c(x, y)dy 2 ; then one can
calculate K(x, y) by differentiating the functions a, b, c. In general one gets
a mess (although this was the classical way of calculating K).
Let (M n , g) be a Riemannian manifold. Coordinates (x1 , . . . , xn ) such
that
g = e2u ((dx1 )2 + . . . + (dxn )2 ),
where u = u(x1 , . . . , xn ) is a given function, are called isothermal coordi-
nates. That a Riemannian manifold (M, g) admits isothermal coordinates is
sometimes expressed by saying g is conformally flat. In Chapter 6 we will
show that every surface with an analytic Riemannian metric admits isother-
mal coordinates. In fact, this is true for C ∞ metrics as well; see ([174], vol.
IV). This will also follow from Exercise 11.2.1.
Specializing to surfaces with isothermal coordinates (x, y), the framing
e1 = e−u ∂x
∂
, e2 = e−u ∂y
∂
is orthonormal.
Exercise 3.8.2: Let (M 2 , g) have isothermal coordinates.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 94 3. Riemannian Geometry

(a) Show that the Gauss curvature is given by

K = −e−2u Δu,
where Δ is the Laplacian. In particular, if K = ±1, then Δu = ∓e2u .
(b) In surface theory it is often useful to introduce complex notation. Write
z = x + iy. Show that solutions to the equation K ≡ ±1 are given by
2|f  (z)|
u(z) = log ,
1 ± |f (z)|2
where f is a holomorphic function on some D ⊂ C with f  = 0 on D and
1 ± |f |2 > 0.
(c) Show that, in isothermal coordinates, Δf = 0 if and only if fxx +fyy = 0.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/04

Chapter 4

Projective Geometry I:
Basic Definitions and
Examples

Until the middle of the twentieth century, there was not such a distinc-
tion between algebraic and differential geometry. In this chapter we discuss
work of Darboux, Cartan, Segre, Terracini and others who approached the
geometry of subvarieties (and submanifolds) of projective space via local
differential geometry. This approach was revisited in the paper of Griffiths
and Harris [85], which began a synthesis of modern algebraic geometry and
moving frames techniques. An earlier version of this chapter, together with
Chapter 12, containing more algebraic results than presented here, consti-
tuted the monograph [129].
We study the local geometry of submanifolds of projective space and
applications to algebraic geometry. We describe moving frames for subman-
ifolds of projective space and define the projective second fundamental form
in §4.1. In §4.2 we give basic definitions from algebraic geometry. We also
give examples of homogeneous algebraic varieties and explain several con-
structions of auxiliary varieties from a given variety X ⊂ PV : the secant
variety σ(X), the tangential variety τ (X) and the dual variety X ∨ . In §4.3
we describe the basic properties of varieties with degenerate Gauss maps
and classify surfaces with this property.
In this chapter, unless we say otherwise, when we work over the complex
numbers, all tangent, cotangent, etc., spaces are the holomorphic tangent,
cotangent, etc., spaces (see Appendix C). We will generally use X to denote

95
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 96 4. Projective Geometry I: Basic Definitions and Examples

an algebraic variety and M to denote a complex manifold. We also use the

notations M̂ and [v] from §1.9.
Throughout this chapter we often abuse of notation by omitting the ⊗
when the tensor product is clear from the context. For example, we often
write ω0α eα for ω0α ⊗eα .

4.1. Frames and the projective second fundamental form

In this section we set up moving frames for submanifolds of projective space
and define the projective second fundamental form. Throughout this section
V = Cn+a+1 or Rn+a+1 , and M ⊂ PV is an n-dimensional complex or real
submanifold.

Tangent spaces and the projective Gauss map. Let v ∈ M̂ ⊂ V and

let x = [v] ∈ M . Define
T̂x M := Tv M̂ ⊂ V, the affine tangent space,
and
T̃x M := P(T̂v M ) ⊂ PV, the embedded tangent (projective) space.

Note that Tv M̂ = Tw M̂ for all nonzero v, w ∈ x̂, so T̂x M, T̃x M are well-
defined. For any submanifold of V , the affine tangent space is the naı̈ve
tangent space obtained by translating tangent vectors to the origin using the
identification Tv V  V . The affine and embedded tangent spaces should not
be confused with the (intrinsic) tangent space Tx M defined in Appendix B.
Recall from §1.9 that the intrinsic tangent space to PV has the description
Tx PV  x̂∗ ⊗(V /x̂).
Exercise 4.1.1: Let M ⊂ PV be a submanifold. Show that Tx M  x̂∗ ⊗
(T̂x M/x̂). Find the corresponding description of Tx∗ M .

Remark 4.1.2. Readers familiar with notation from algebraic geometry

should note that the line bundle with fiber x̂ at x ∈ M is OM (−1).

Define the normal space Nx M := Tx PV /Tx M . Note that Nx M = x̂∗ ⊗

(V /T̂x M ). Define the conormal space Nx∗ M to be the dual vector space, and
note that Nx∗ M = x̂⊗(T̂x M )⊥ ⊂ Tx∗ PV (where Here (T̂x M )⊥ ⊂ V ∗ denotes
the annihilator of T̂x M ⊂ V ) is such that Nx∗ M ⊗OM (1)x ⊂ V ∗ . While the
normal space is merely a quotient space, the projectivized conormal space
determines a linear subspace of PV ∗ .
Exercise 4.1.3: Show that the linear subspace of PV ∗ determined by Nx∗ M
(namely, (T̂x M )⊥ ⊂ PV ∗ ) admits the geometric interpretation as the space
of hyperplanes containing T̃x M .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.1. Frames and the projective second fundamental form 97

Define the (projective) Gauss map

γ : M → G(n + 1, V ),
x → T̂x M.
We call γ(M ) the Gauss image of M . Consider the differential
dγx : Tx M → TT̂x M G(n + 1, V ),
which describes how M is moving away from its embedded tangent space to
first order.
Proposition 4.1.4. For all v ∈ Tx M , dγx (v) ∈ Hom(T̂x M, V /T̂x M ) is such
that for all w ∈ x̂, w ∈ ker(dγx (v)). Thus we have an induced map
dγx (v) : T̂x M/x̂ → V /T̂x M,
i.e., dγx ∈ Tx∗ M ⊗Tx∗ M ⊗Nx M . Moreover, dγx ∈ S 2 Tx∗ M ⊗Nx M .
We (or more precisely, you) will prove Proposition 4.1.4 in Exercises
4.1.7 and 4.1.8.
Definition 4.1.5. Define IIM,x = dγx ∈ S 2 Tx∗ M ⊗ Nx M , the projective
second fundamental form of M at x.
Remark 4.1.6. Recall from Appendix A that, given a tensor A ∈ V ∗ ⊗W ,
we may consider A as a linear map, with either A : V → W or A : W ∗ →
V ∗ . In standard linear algebra texts, if one map is given first, the other
is denoted tA. In order to avoid prejudicing ourselves, we will use A to
denote both maps. In particular, we will slightly abuse notation, writing
both II : Nx∗ M → S 2 Tx∗ M and II : S 2 Tx M → Nx M when using II.

Frames. We continue with the projection

π : GL(V ) → PV,
(e0 , . . . , en+a ) → [e0 ].
Fix a reference basis of V as in §1.9 so that we may identify GL(V ) with
FPV , the set of bases of V . We may consider FPV as a frame bundle over PV ,
because (e1 mod e0 )⊗e0 , . . . , (en+a mod e0 )⊗e0 is a basis of T[e0 ] PV . (Here
e0 , . . . , en+a ∈ V ∗ is the dual basis to e0 , . . . , en+a .) Given M n ⊂ PV , let
FM0 := F
PV |π −1 (M ) , the bundle of 0th-order adapted frames to M .
The affine tangent space of M at each p ∈ M determines a flag
p̂ ⊂ T̂p M ⊂ V.
Define = F1 FM
→ M , the bundle of first-order adapted frames, to be the
1

frames respecting this flag, namely,

FM
1
= {(e0 , . . . , en+a ) ∈ FM
0
| [e0 ] ∈ M, T̂[e0 ] M = e0 , . . . , en }.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 98 4. Projective Geometry I: Basic Definitions and Examples

Consequently, on F 1 = FM 1 , de ≡ 0 mod{e , . . . , e }. In other words,

0 0 n
de0 (t)
dt | t=0 ∈ {e 0 , . . . , e n } for all curves [e 0 (t)] ∈ M with e0 (0) = e0 .
The fiber of π : F 1 → M over a point is isomorphic to the group
⎧ ⎛ 0 ⎞⎫
⎨ g0 gβ0 gν0 ⎬
G1 = g ∈ GL(V )| g = ⎝ 0 gβα gνα ⎠ ,
⎩ ⎭
0 0 gνμ
where we use the index ranges 1 ≤ α, β ≤ n, and n + 1 ≤ μ, ν ≤ n + a. Write
the pullback of the Maurer-Cartan form of GL(V ) to F 1 as
⎛ 0 ⎞
ω0 ωβ0 ων0
ω = ⎝ω0α ωβα ωνα ⎠ .
ω0μ ωβμ ωνμ

Since de0 = e0 ω00 +e1 ω01 +. . .+en+a ω0n+a , and de0 ≡ 0 mod{e0 , . . . , en } when
pulled back to F 1 , we must have ω0μ = 0 ∀μ.
Notation. Given f = (e0 , . . . , en+a ) ∈ F 1 , we let (e0 , . . . , en+a ) denote
the dual basis of V ∗ and eα = e0 ⊗(eα mod e0 ) ∈ T[e0 ] M denote the tangent
vector corresponding to eα (see Exercise 4.1.1). Note that for any section
s : M → F 1 , we have s∗ (ω0α )(eβ ) = δβα . Also note that eα depends on f and
not just eα . Similarly, for normal vectors, we let eμ = e0⊗(eμ mod T̂[e0 ] M ) ∈
N[e0 ] M .

We may recover the Gauss map γ from the map

F 1 → G(n + 1, V ),
f → [e0 ∧ · · · ∧ en ].

Exercise 4.1.7: Prove the first assertion in Proposition 4.1.4 by calculating

d(e0 ∧ · · · ∧ en ).

Thanks to our Pavlovian response to seeing something equal to zero (and

our usual abuse of notation omitting pullbacks in the notation here and in
what follows), we expand out

0 = dω0μ = −ωαμ ∧ ω0α .

The Cartan Lemma A.1.9 implies

ωβμ = qαβ
μ
ω0α
μ μ
for some functions qαβ = qβα defined on F 1 .
Exercise 4.1.8: Prove the “moreover” assertion in Proposition 4.1.4.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.1. Frames and the projective second fundamental form 99

Frame definition of the projective second fundamental form. While

our next definition of the projective second fundamental form is less elegant,
the technique we use to arrive at it is applicable to more general situations,
e.g., the definition of the cubic form in §12.1.
Say we did not know about the Gauss map, but we calculated ω0μ = 0
and dω0μ = 0 on F 1 to obtain functions qαβ μ
: F 1 → C. We would then like
to form an invariant tensor from these functions. To do so, we need to see
how they vary as we move in the fiber of π1 : F 1 → M . Let f˜ = (ẽ0 , ẽα , ẽν )
be another point in π1 −1 ([e0 ]). Then there exists g = (gB
A ) ∈ G such that
1

ẽ0 = g00 e0 ,
ẽα = gα0 e0 + gαβ eβ ,
ẽμ = gμ0 e0 + gμβ eβ + gμν eν .
We first determine how ω0α , ωαμ vary. Let “g·” denote the action of G both
on V and on the Maurer-Cartan form, and observe that
g · de0 = g · (e0 ω00 + eα ω0α )
= (g00 e0 )(g · ω00 ) + (gα0 e0 + gαβ eβ )(g · ω0α ).
On the other hand, since d commutes with the g-action,
g · de0 = d(e0 g00 )
= g00 (e0 ω00 + eα ω0α ).
Let h denote the inverse matrix to g. The analogous calculation for deα
combined with our calculation above yields
ω̃0α = hαβ g00 ω0β ,
ω̃αμ = gαβ hμν ωβν .
Substituting into the equation ω̃βμ = q̃βα
μ
ω̃0α , we obtain
μ
q̃αβ = h00 hγβ hδα gνμ qγδ
ν
.
Thus the quantity
 = q μ ω α ω β ⊗eμ ∈ π ∗ (S 2 T ∗ M ⊗N[e ] M )
II αβ 0 0 [e0 ] 0

is constant on the fiber and therefore descends to be a well-defined section

of S 2 T ∗ M ⊗N M , which is the projective second fundamental form.
Remark 4.1.9. When we compare the projective second fundamental form
to its Euclidean counterpart, unlike the Euclidean case we no longer have a
notion of ‘how fast’ M is moving away from its embedded tangent space to
first order in each direction, but only whether or not it is moving away. For
example, if M is a complex hypersurface, the only numerical information in
the projective second fundamental form is its rank because the only invariant
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 100 4. Projective Geometry I: Basic Definitions and Examples

of a quadratic form over C is its rank. For real projective hypersurfaces the
only numerical information is rank and signature.
Remark 4.1.10. If we are in the holomorphic category, we may compare
the projective second fundamental form with its Hermitian counterpart. In
this case the projective second fundamental form is simply the holomorphic
component of the Hermitian second fundamental form.
Remark 4.1.11. The projective second fundamental form admits the fol-
lowing interpretation. As remarked in Exercise 4.1.3, the conormal space
Nx∗ M may be identified (after tensoring with a line bundle) with the space
of linear forms on V annihilating T̂x M , so we may think of H ∈ N ∗ M as
x
determining a hyperplane H containing the tangent projective space T̃x M .
As such, as long as M  H, M ∩ H must be singular at x because it is
defined by the equations of M and the linear form defining H, but this
linear form has differential in the span of the differentials of the defining
equations of M . With this perspective II(H)  ∈ S 2 T ∗ M may be identified
x
as the quadratic part of the singularity.

Adopt the notation

|IIM,x | = II(Nx∗ M ) ⊂ S 2 Tx∗ M
to study the space of quadratic polynomials on Tx M determined by II.

The second fundamental form determines preferred tangent directions.

Let
Baseloc |IIM,x | = P{v ∈ Tx M | II(v, v) = 0},
(4.1)
Singloc |IIM,x | = {v ∈ Tx M | II(v, w) = 0 ∀w ∈ Tx M }.
Baseloc(IIM,x ) is often called the set of asymptotic directions, while the
dimension of Singloc(IIM,x ) is called the index of relative nullity. Note that
Baseloc(IIM,x ) is an intersection of quadric (i.e., degree two) hypersurfaces
in PTx M and that Singloc(IIM,x ) is a linear space.
Note that P Singloc(IIM,x ) ⊆ Baseloc(IIM,x ). For the projectivization
of a cylinder in affine 3-space we have equality, while for the projectivization
of a hyperbola in affine space, Baseloc(IIM,x ) consists of two points while
P Singloc(IIM,x ) is empty.
If we think of PTx M ⊂ PTx PV as the set of directions where there exists
a line having contact to order one with M at x (as defined in §4.2 below),
then Baseloc(IIM,x ) ⊂ PTx M is the set of directions where there exists a line
having contact to order two with M at x. We may interpret Singloc(IIM,x )
as ker dγx , so manifolds where Singloc(IIM,x ) = 0 at general points (defined
in §4.2 below) have degenerate Gauss maps. We discuss this further in §4.3.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.2. Algebraic varieties 101

M
M

Tx M Tx M

IP T x M IP T x M
Baseloc|II| Baseloc|II|

Exercises 4.1.12:
1. Show that Singloc |IIM,x | = {v ∈ Tx M | (s∗ ωαμ )(v) = 0, ∀α, μ}, where
s : M → F 1 is any section.
2. Show that if IIM,x ≡ 0 for all x ∈ M , then M is (an open subset of) a
linear Pn ⊂ PV . In particular, if M is analytic, then M is an open subset of
a linear space if and only if IIM,x = 0 at a general point x ∈ M as defined
below.

Remark 4.1.13. If M is given in local linear coordinates as a graph xμ =

f μ (x1 , . . . , xn ), then
∂ 2f μ ∂
IIM,x = (x)dxα dxβ ⊗ μ .
∂xα ∂xβ ∂x
See §12.1 for the proof.

4.2. Algebraic varieties

In this section we briefly cover basic definitions from algebraic geometry to
enable us to study global properties of submanifolds of projective space and
make initial connections with algebraic geometry. We also give examples of
homogeneous varieties and describe, given a variety X ⊂ PV , new varieties
constructed from X: tangential varieties, secant varieties, and dual varieties.
We will use coarse properties of the new varieties to determine properties of
the original variety. For example, in §4.3, we will see that if 0 < dim γ(X) <
dim X, then X cannot be a smooth variety.
Even if our original interest is only in complex manifolds, we will be
forced to deal with algebraic varieties. For example, the Gauss image γ(M )
of a compact complex submanifold of PV (other than a linear space or
quadric hypersurface) is never a complex manifold but it is an algebraic
variety.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 102 4. Projective Geometry I: Basic Definitions and Examples

Definitions. As a working definition, an algebraic variety is the projectiviza-

tion (i.e., the image under π : V \0 → PV ) of the common zero locus of a
collection of homogeneous polynomials on V . The ideal I(X) ⊂ S • V ∗ of a
variety X ⊂ PV is the set of all homogeneous polynomials vanishing on X.
We let Ik (X) = I(X) ∩ S k V ∗ .
By Chow’s Theorem (see, e.g., [151]), any compact complex submanifold
of a projective space is an algebraic variety. In the other direction, any
algebraic variety, minus at most a proper subvariety, is an analytic manifold.
The subvariety consists of points where the Jacobian of the local defining
equations drops rank; these are called the singular points of X, and we
denote the subvariety by Xsing .
We define the dimension of X to be the dimension of the corresponding
complex manifold Xsmooth = X\Xsing . (If X has several components, we
must define the dimension of each component.) A basic property of projec-
tive space is that if X, Y ⊂ PV are respectively subvarieties of codimension
a, b, then codim(X ∩ Y ) ≤ a + b, with equality holding if they intersect
transversely.
The degree of X is the maximum number of points of intersection with a
linear space of complementary dimension. If X is a hypersurface defined by
an equation {f = 0}, then the degree of X is the degree of the polynomial f .
Bezout’s Theorem (see, e.g., [151]) states that if X, Y ⊂ PV are two varieties
of complementary dimension that intersect transversely, then X ∩Y consists
of deg(X) deg(Y ) points.
A discussion of the Zariski topology and rational maps would take us
too far afield here, but readers familiar with such terminology should note
that the proper language for discussing Gauss maps and their higher-order
analogs is that of rational maps. For example, we could either define γ(X)
as the closure of γ(Xsmooth ) or consider γ as a rational map and define
γ(X) as its strict transform. Sometimes in statements of theorems we refer
to a Zariski open subset. The reader may omit the word Zariski without
any harm, because Zariski open subsets of algebraic varieties always contain
open subsets in the manifold topology. Throughout this chapter, closure
is denoted by an overline, and in all cases treated, one obtains the same
closure in either the manifold or the Zariski topology; see [133, p. 118] or
[151, Thm. 3.11].
Given a vector space V with additional structure S, e.g., a collection of
tensors or an ideal of polynomials on V , we define a point v ∈ V to be S-
generic if it is generic with respect to S. For example, if V is equipped with
a quadratic form Q, v ∈ V is Q-generic if Q(v, v) = 0, and more generally, if
V is equipped with a space S of quadratic forms, v is S-generic if all small
perturbations of v have similar behavior with respect to S. In particular,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.2. Algebraic varieties 103

all integer-valued invariants obtainable from v and S are locally constant if

one varies v. If the structure S is understood, we omit the quantifier S. If
X ⊂ PV is a variety, x ∈ X is a general point of X if all Z-valued differential
invariants of X (e.g., the rank of IIx : S 2 Tx∗ X → Nx X) are locally constant
in some neighborhood of X.

Example 4.2.1. If one considers the plane curve y = x3 , then all points
on the curve are general except the origin which is a flex point (also called
a point of inflection).

The general points of X form a nonempty Zariski open subset of X, as

do the S-generic points of V (for any S).
We will say a curve c has contact to order d with X at x if there exists
a curve c̃ ⊂ X such that the Taylor series of the curves c, c̃ agree to order
d at x. For those familiar with the definition of multiplicity in algebraic
geometry, a curve has contact to order d at x if the multiplicity of c ∩ X at
x is (d + 1).
Exercise 4.2.2: Let X be a hypersurface of degree d. Show that any line
intersecting X in d+1 points is contained in X. Similarly, show that if a line
has contact to order d at any point x ∈ X, then it is contained in X. Show
that the same conclusion holds if X may be described as the intersection of
hypersurfaces of degree at most d. 
Proposition 4.2.3. If X ⊂ PV is a variety defined by polynomials of degree
at most two, then Baseloc(IIX,x ) is the set of tangent directions to lines
contained in X that pass through x.

Proof. By Exercise 4.2.2 any line having contact to order two with X at x
is contained in X. 

Homogeneous algebraic varieties.

Definition 4.2.4. A subvariety X ⊂ PV is said to be projective rational

homogeneous (or homogeneous for short) if X is the orbit in PV of a point
under the action of a subgroup G ⊂ GL(V ). In this case we write X = G/P ,
where P is the isotropy subgroup.

Remark 4.2.5. A projective rational homogeneous variety is always the

projectivization of the orbit of a highest weight vector in V ; see, e.g., [65].
The group P is called a parabolic subgroup. A theorem of Kostant (see [133,
§16.2]) states that if X = G/P ⊂ PVλ is a rational homogeneous variety of
the G-module Vλ of highest weight λ (see Appendix A), then the ideal of X
is generated in degree two by V2λ ⊥ ⊂ S 2 (Vλ )∗ .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 104 4. Projective Geometry I: Basic Definitions and Examples

Remark 4.2.6. Borel and Remmert proved that any compact homogeneous
Kähler manifold is of the form G/P × T , where T is the quotient of Cg by
a lattice; see [2].

The Grassmannians G(k, W ) ⊂ P(Λk W ) are homogeneous algebraic va-

rieties; see e.g. [133, §6.10.3] for an explicit description of their ideals. For
example, if W = Cm and we identify Λ2 W with the space of skew-symmetric
m × m matrices, then G(2, W ) ⊂ P(Λ2 W ) may be thought of as the pro-
jectivization of the space of rank two skew matrices. Thus G(2, W ) may be
defined by the 3 × 3 minors. Kostant’s result above implies it is generated
in degree two, not three. In fact, the ideal of G(2, W ) is generated by the
Pfaffians (square roots of the determinants; see Exercise A.4.9.4) of the 4 ×4
minors centered along the diagonal.
Segre varieties. Let W = Cm and U = Cn . The Segre variety Seg(PW ×PU )
⊂ P(W⊗U ) is defined to be the projectivization of the decomposable tensors,
i.e., the image of the map
Seg : PU × PW → P(U ⊗W ),
([x], [y]) → [x⊗y].
It is homogeneous for the G = GL(U ) × GL(W ) action on U ⊗ W . If we
identify U ⊗W with the space of n × m matrices, Seg(PW × PU ) is the set
of rank one matrices up to scale. (Recall that every rank one matrix is the
product of a column vector with a row vector.) Thus, its ideal is generated
by the 2 × 2 minors. To have a 2 × 2 minor we need to choose two rows and
two columns, so I2 (Seg(PU × PW ))  Λ2 U ∗ ⊗Λ2 W ∗ .
If X ⊂ PU and Y ⊂ PW are varieties, we define the Segre product of X
and Y , Seg(X × Y ) ⊂ P(U ⊗W ), to be the image of X × Y under the map
Seg.
Exercises 4.2.7:
1. Show that I2 (G(2, W ))  Λ4 W ∗ .
2. Let Mr ⊂ P(U ⊗W ) denote the projectivization of the set of matrices
of rank at most r. Show that Mr is a variety, and Ir+1 (Mr )  Λr+1 U ∗ ⊗
Λr+1 W ∗ . In fact the ideal is generated in degree r + 1; see e.g., [133, §17.2].

Isotropic Grassmannians. Let Q ∈ S 2 V ∗ , ω ∈ Λ2 V ∗ be nondegenerate.

Define the corresponding isotropic Grassmannians
GQ−null (k, V ) := {E ∈ G(k, V ) | Q(v, w) = 0 ∀v, w ∈ E},
Gω−null (k, V ) := {E ∈ G(k, V ) | ω(v, w) = 0 ∀v, w ∈ E}.
In the case dim V = 2m and k = m, GQ−null (m, V ) has two isomorphic com-
ponents. The components are called the spinor varieties Sm . The isotropic
Grassmannians are naturally varieties, as they sit inside Grassmannians and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.2. Algebraic varieties 105

the isotropy conditions are defined by polynomials. They are homogeneous

for O(V, Q) and Sp(V, ω) respectively.
Veronese varieties. Let S 2 V denote the symmetric matrices and let v2 (PV )
⊂ P(S 2 V ) denote the projectivization of the rank one symmetric matrices.
Then v2 (PV ) is the image of PV under the injective mapping

v2 : PV → PS 2 V,
[v] → [v 2 ].

The d-th Veronese embedding of PV , vd (PV ) ⊂ PS d V , is defined by vd ([v]) =

[v d ]. Intrinsically, all these varieties are PV , but the embeddings are projec-
tively inequivalent. We may interpret vd (PV ) ⊂ PS d V as the set of degree d
polynomials on PV ∗ whose zero set is a hyperplane counted with multiplicity
d.
Given X ⊂ PV , we can consider the Veronese re-embeddings of X,
vd (X) ⊂ P(S d V ), which are the restrictions to X of the Veronese embed-
dings of PV . These will turn out to be useful in our study of osculating
hypersurfaces and complete intersections.
Exercise 4.2.8: If Z is a hypersurface of degree d, show that vd (Z) =
HZ ∩ vd (PV ), where HZ is the hyperplane in P(S d V ) annihilated by the
equation of Z (which is an element of S d V ∗ ).

Flag varieties. Let V = Cn and let 1 ≤ a1 < a2 < . . . < ap ≤ n − 1. A

(partial) flag of V is a sequence of linear subspaces 0 ⊂ E1 ⊂ E2 ⊂ · · · ⊂ Ep
with dim Ej = aj . Let Fa1 ,...,ap (V ) denote the space of all such flags. Note
that G(k, V ) = Fk (V ). At the other extreme is the variety of complete flags
F1,2,...,n−1 (V ). The flag varieties are homogeneous spaces of GL(V ).
Exercises 4.2.9:
1. Write F1,2,...,n−1 = GL(n)/B and explicitly determine B as a matrix Lie
group. If Fa1 ,...,ap = GL(n)/P , what is P as a matrix Lie group?
2. Determine T[Id] Fa1 ,...,ap . That is, write Ej = {e1 , . . . , ej } and generalize
our TE G(k, V ) = E ∗ ⊗V /E description.
3. Show that the flag variety F1,n (Cn+1 ) can be realized as the space of
rank one traceless matrices, up to scale.

The flag varieties embed as homogeneous subvarieties of projective space.

One way to see this is to consider the Segre product of Grassmannians:

S = Seg(G(a1 , V ) × G(a2 , V ) × · · · × G(ap , V )) ⊂ P(Λa1 V ⊗· · ·⊗Λap V )

and now to consider the subvariety defined by incidence relations

F = {(E1 , . . . , Ep ) ∈ S | E1 ⊂ E2 , E2 ⊂ E3 , . . . , Ep−1 ⊂ Ep },
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 106 4. Projective Geometry I: Basic Definitions and Examples

and observe that F = Fa1 ,...,ap (V ). The incidence relations are described by
polynomials, namely various contractions being zero.

Constructions of new varieties from old.

Tangential varieties. Let M n ⊂ PV = Pn+a be a smooth variety. Define
τ (M ) ⊂ PV by
(
τ (M ) := T̃x M,
x∈M
the tangential variety of M .
Notice that dim τ (M ) ≤ min{2n, n + a}. We will see that for most
smooth varieties equality holds. The tangential variety is indeed an algebraic
variety; see [89].
It is possible to define the tangential variety of an algebraic variety that
is not necessarily smooth. While there are several possible definitions, the
best one appears to be the union of tangent stars Tx X. This is because
the Fulton-Hansen Theorem 12.5.7 applies with this definition. Intuitively,
Tx X is the limit of secant lines through pairs of points that limit to x. More
precisely, let x ∈ X. Then P1∗ is a line in Tx X if there exist smooth curves
p(t), q(t) on X such that p(0) = q(0) = x and P1∗ = limt→0 P1p(t)q(t) , where
P1pq denotes the projective line through p and q. Tx X is the union of all P1∗ ’s
at x. In general τ (Xsmooth ) ⊆ τ (X), and strict containment is possible.
Exercises 4.2.10:
1. Let M ⊂ PV be a smooth curve that is not a line. Show that τ (M ) has
dimension two.
2. Let X be a variety and let x ∈ Xsmooth . Show that Tx X = T̃x X.

Joins, cones and secant varieties. Let Y, Z ⊂ PV be varieties (we allow the
possibility that Y = Z). For x, y ∈ PV , let P1xy denote the projective line
containing x and y. Define the join of Y and Z to be
J(Y, Z) = x∈Y,y∈Z,x=y P1xy .
If Z = Pk is a k-plane, J(Y, Pk ) is called the cone over Y with vertex Pk . If
Y = Z, σ(Y ) = J(Y, Y ) is called the secant variety of Y . We may similarly
define the join of k varieties to be the union of the corresponding Pk−1 ’s,
or by induction as J(Y1 , . . . , Yk ) = J(Y1 , J(Y2 , . . . , Yk )). In particular, write
σk (Y ) = J(Y, . . . , Y ) for the join of k copies of Y , called the k-th secant
variety of Y .
Aside 4.2.11. Given Y ⊂ PV , if Y is not contained in a hyperplane we
obtain a stratification of PV by the secant varieties of Y . Given x ∈ PV ,
define the Y -border rank of x to be the minimum k such that x ∈ σk (Y )
and the Y -rank of x to be the minimum k such that x is in the span of k
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.2. Algebraic varieties 107

points of Y . The notions of Y -rank and Y -border rank provide a geometric

generalization of the corresponding notions for tensors (and in particular
matrices).

Lemma 4.2.12 (Terracini’s Lemma). If [x] ∈ J(Y, Z)smooth with [x] = [y +

z], such that [y] ∈ Ysmooth , [z] ∈ Zsmooth , then

T̂[x] J(Y, Z) = T̂[y] Y + T̂[z] Z.

Proof. Consider the addition map add : V ×V → V given by (v, w) → v+w.

Then
Jˆ(Y, Z) = add(Ŷ × Ẑ).
Now consider the differential of add |Ŷ ×Ẑ at a general point. The result
follows. 

Note that we expect dim J(Y, Z) = dim Y + dim Z + 1, because, given

[y + z] ∈ J(Y, Z), we are free to move y in dim Y directions, z in dim Z
directions, and we may move along the line joining our points. The following
exercises discuss this expectation.
Exercises 4.2.13:
1. Show that if there exist y ∈ Y and z ∈ Z such that T̃y Y ∩ T̃z Z = ∅, then
dim J(Y, Z) = dim Y + dim Z + 1.
2. Suppose X ⊂ PV is a curve not contained in a plane. Show that
dim σ(X) = 3.
3. Calculate dim σ(Seg(P2 × P2 )).

Dual varieties.

Definition 4.2.14. Let X ⊂ PV be a variety. Define

X ∨ := {H ∈ PV ∗ | ∃x ∈ Xsmooth such that T̃x X ⊆ H} ⊂ PV ∗ ,

the dual variety of X.

Note that X ∨ = γ(X) when X is a hypersurface.

Intuitively, for X n ⊂ Pn+a , we expect X ∨ to be a hypersurface, as there
are n dimensions worth of points on X and there is an (a − 1)-dimensional
space of hyperplanes tangent to each smooth point of X.
Exercises 4.2.15:
1. Show that if X is a curve that is not contained in a hyperplane, then
X ∨ is a hypersurface.
2. Show that if X is a cone over a linear space, then X ∨ is not a hypersur-
face.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 108 4. Projective Geometry I: Basic Definitions and Examples

Linear projections and hyperplane sections. Given X ⊂ PV and p ∈ PV \X,

we consider the vector space projection V → V /p̂. The image of X̂ under this
projection is still a cone, and when we projectivize it, we obtain a subvariety
X  ⊂ P(V /p̂), called the projection of X from p. The projection of a variety
is a variety because the projection amounts to eliminating variables, an
algebraic procedure.
Given a variety X ⊂ PV and a hyperplane H ⊂ PV , we define the
hyperplane section of X by H to be X ∩ H. Hyperplane sections and linear
projections are related in Exercise 4.2.16.4 below.
Exercises 4.2.16:
1. Assume p ∈/ T̃x X. Calculate IIX  ,x in terms of IIX,x . 
2. Let H ⊂ PV be a hyperplane and assume T̃x X  H. Calculate IIX∩H,x
in terms of IIX,x . 
3. Show that the degree of X ∩ H is the same as that of X.
4. Show that if X is not a hypersurface, then (X  )∨ = X ∨ ∩ Hp , where
Hp ⊂ PV ∗ is the hyperplane corresponding to the point p ∈ PV .

Incidence correspondences. We have already seen that flag varieties can be

described as incidence varieties of Segre varieties, that is, a natural subva-
riety of a product of varieties. Many important constructions in algebraic
geometry use incidence varieties, and the auxiliary varieties defined above
are naturally defined using incidence varieties. For example, one can study
Gauss images and dual varieties using incidence varieties as follows:
Let X n ⊂ PV be a variety. Consider

(4.2) Γ(X) = {(x, T̃x X) | x ∈ Xsmooth } ⊂ F1,n+1 (V )

with its two projections:

Γ(X)
 S

/ w
S
X γ(X) ⊂ G(n + 1, V ).

Here, and it what follows, we identify Fs+1,t+1 with the space of Ps ’s con-
tained in Pt ’s contained in PV . Similarly consider

(4.3) IX,X ∨ = {(x, H) | x ∈ Xsmooth , T̃x X ⊂ H} ⊂ F1,n+a (V )

with picture
I
 S
/
 w
S
X X ∨ ⊂ PV ∗ .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.3. Varieties with degenerate Gauss mappings 109

One advantage of using incidence correspondences is that they guide us to a

natural frame bundle over X for studying the problem at hand. For example,
when studying dual varieties, we will work with a frame bundle that factors
through the first-order adapted frame bundle of IX,X ∨ . This method will
be used consistently throughout this chapter.

4.3. Varieties with degenerate Gauss mappings

In this section we begin our study of varieties X ⊂ PV such that dim γ(X) <
dim X. We show such varieties are always singular and locally ruled by the
fibers of the Gauss map, which are linear spaces. (A variety is locally ruled
by fibers isomorphic to F if an open subset U ⊂ X is a fibration, with
fibers (Zariski) open subsets of F .) We classify surfaces with degenerate
Gauss images and characterize subvarieties of the Grassmannian that occur
as Gauss images. The study of varieties with degenerate Gauss images is
continued in §12.10.
Proposition 4.3.1. Let X n ⊂ Pn+a be a variety. Let x ∈ X be a general
point, let F be the connected component of γ −1 (γ(x)) containing x and let
d = dim F . Then
1. Tx F = Singloc |IIX,x |,
2. F is a linear space, and thus X is locally ruled by d-planes.
Conversely, if X ⊂ PV is a variety, x ∈ X a general point and Singloc(IIX,x )
= 0, then γ is degenerate with fiber equal to the linear space F such that
Tx F = Singloc |IIX,x |.
This proposition is essentially due to C. Segre; see [169, p. 95]. Varieties
with degenerate Gauss maps (other than linear spaces) have singularities:
Theorem 4.3.2. Let X n ⊂ Pn+a be a subvariety. If X is not a linear
subspace of Pn+a and dim γ(X) < dim X, then X is singular. More precisely,
if F is a general fiber of γ, then Xsing contains a codimension one subset of
F which is the zero locus of a polynomial of degree n − d on F̂ .
The subset of Xsing ∩ F referred to in the theorem is called the focal
hypersurface of X along F and will be denoted by Φ.

Proof of 4.3.1. Consider the commutative diagram

FX
1

 Sρ
π
 γ S
/ w
X - G(n + 1, V )

where π(e0 , . . . , en+a ) = [e0 ] and ρ(e0 , . . . , en+a ) = [e0 ∧ . . . ∧ en ]. Here the
geometric object of interest is the incidence correspondence Γ(X) defined in
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 110 4. Projective Geometry I: Basic Definitions and Examples

Equation (4.2), but we work on F 1 in order to use the Maurer-Cartan forms

and Maurer-Cartan equation.
If f = (e0 , . . . , en+a ) ∈ F 1 is a general point, then dim γ(X) = rank ρ∗f
and moreover π∗ (ker ρ∗f ) = Tx F .
Let E = e0 , . . . , en . We calculate

ρ∗f = ωαμ ⊗eα ⊗(eμ mod E).

Recall that the ωαμ are semi-basic for π; thus assertion 1 and the “conversely”
part of Proposition 4.3.1 follow from Exercise 4.1.12.2.
We now implement a key idea in the method of the moving frame: now
that we have more geometric information, we adapt frames further to reflect
this information.
Let F γ ⊂ F 1 be the subbundle of the first-order adapted frame bundle
consisting of frames preserving the flag

x̂ ⊂ x̂, Singloc |II|x  ⊂ T̂ ⊂ V.

Let {e1 , . . . , ed } = Singloc |II|. Use additional index ranges 1 ≤ s, t ≤ d,

d + 1 ≤ i, j, k ≤ n. In indices, F γ is the set of frames preserving the flag:

{e0 } ⊂ {e0 , es } ⊂ {e0 , es , ej } ⊂ {e0 , es , ej , eμ }.

Exercise 4.3.3: Show that our adaptations have the effect that the pull-
backs of the forms ωsμ to F γ are zero and that Singloc |II| = {ω0j }⊥ = {es }.
(Recall that S ∗ (ω0α )(eβ ) = δβα for any section S : X → F 1 .)

We now show that the fibers are indeed linear spaces. The Maurer-
Cartan equation implies

dω0j ≡ −ωsj ∧ ω0s mod{ω0i }.

We claim that ωsj ≡ 0 mod{ω0i } for all j, s. This is equivalent to asserting

that the distribution Singloc |II| is integrable, which we already know by
the argument above, but derive directly as follows:
Differentiating ωsμ = 0, we obtain

(4.4) 0 = dωsμ = −ωjμ ∧ ωsj .

Since ωjμ = qjk

μ k
ω0 , applying the Cartan Lemma to (4.4) shows that, for
each j, s, we have
ωsj ≡ 0 mod{ω0i }.

By Exercise 4.1.12.2, to show that the integral manifolds of the distri-

bution (which we know to be the fibers F by our first argument) are linear
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.3. Varieties with degenerate Gauss mappings 111

spaces, it will suffice to show that IIF,[e0 ] = 0. Because ωsμ = 0 when pulled
back to F γ and we may restrict F γ to be a frame bundle over F , we have

IIF,[e0 ] = ωsj ω0s ⊗ej .

But ωsj ≡ 0 mod{ω0k } and the ω0k pulled back to F are zero; hence IIF = 0,
and the proposition is proved. 

Proof of Theorem 4.3.2. Let p = [u0 e0 + us es ] ∈ F . We calculate T̃p X.

Since F = P{e0 , es } ⊂ T̃p X for all p ∈ F , we can work modulo {e0 , es },

dp ≡ (u0 ω0j + us ωsj )ej mod{e0 , es }

≡ (u0 δjk + us Csj
k
)ω0j ek mod{e0 , es }.

Thus, p is a smooth point of X if and only if the matrix (u0 δjk + us Csj
k ) is

invertible. Moreover, we see that in this case, T̃p X = T̃[e0 ] X.

Let Matm×n (C) denote the set of m × n complex matrices. Consider the
linear map i : F̂ → Mat(n−d)×(n−d) (C) given by (u0 , us ) → (u0 δjk + us Csj k ).

Consider det, the determinant on Mat(n−d)×(n−d) (C), which is a polynomial

of degree n−d. Note that i∗ (det) is not identically zero, because i(1, 0, . . . , 0)
is the identity matrix. Thus, its zero set is a codimension one subset of F ,
and this zero set is the focal hypersurface. 

We can refine our observation about the focal hypersurface as follows:

The matrix i(u0 , us ) could be zero for some (u0 , us ). This possibility will
j
occur if, considering each Cs = Csk as a matrix, the Cs ’s fail to be linearly
independent. Consider new indices 1 ≤ ρ ≤ d , d + 1 ≤ ξ ≤ d such that
the Cρ ’s are linearly independent and the Cξ ’s are zero. Then the focal
hypersurface is a cone with vertex the linear subspace P{eξ }.
Exercises 4.3.4:
1. Show that tangential varieties have degenerate Gauss maps. 
2. Show that joins have degenerate Gauss maps. 
3. Let Mr ⊂ P(Mm×n ) denote the projectivization of the set of matrices of
rank at most r. What is the rank of the Gauss map of Mr ? 

Remark 4.3.5. Variations of the above constructions produce additional

examples of varieties with degenerate Gauss mappings, but not all of them.
It is an open, vaguely posed, problem to classify varieties of dimension
greater than two with degenerate Gauss mappings. Progress on this question
has been made by Griffiths and Harris [85], Akivis and Goldberg [4], and
by Piontkowski [160]. We discuss varieties with degenerate Gauss mappings
more in §12.10. For now we consider the case of surfaces.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 112 4. Projective Geometry I: Basic Definitions and Examples

Classification of surfaces with degenerate Gauss mappings.

Theorem 4.3.6 (C. Segre, [169, p. 105]). Let X 2 ⊂ PV be a variety with
degenerate Gauss mapping. Then X is one of the following:
i. a cone over a curve;
ii. the tangential variety to a curve.
A linearly embedded P2 is a special case of both (i) and (ii). We will see
in the proof that it is the only intersection of those cases.

Proof. Assume that X = P2 , so II = 0. Let M ⊂ Xsmooth be an open

subset where γ is of constant rank one. Let F̃ → M be the subbundle of
F 1 adapted additionally such that {e1 } = Singloc |II| and II(e2 , e2 ) = e3 .
We have
ω13 = 0,
ω23 = ω02 .
Thus
0 = dω13 = −ω23 ∧ ω12 ,
so
ω12 = λω02
for some function λ : F̃ → C. On F̃ we have the fiber motion
e1 → e1 + g10 e0
which sends
ω12 → ω12 + g10 ω02 .
Thus by choosing g10 = −λ, we may normalize λ = 0. Let F  ⊂ F̃ be the
subbundle where λ ≡ 0.
The geometric meaning of our adaptation is as follows: our adaptation
to F̃ was such that the fiber of the Gauss map through [e0 ] was P{e0 , e1 }.
In this case, by Theorem 4.3.2 there is a unique singular point (the zero set
of a degree one polynomial) in a general fiber. We have adapted frames such
that this singular point is [e1 ].
As we move [e0 ] in X transversely to the fiber of the Gauss map there
are two possibilities: First, [e1 ] could stay fixed, in which case X must be
a cone over [e1 ], because then [e1 ] ∈ T̃x M for all x ∈ M . Otherwise, [e1 ]
sweeps out a one-dimensional variety, i.e., a curve in X. We will show that,
in the second case, X is the tangential variety to this curve. On F  we have
0 = dω12 = −ω02 ∧ ω10 ,
which implies
ω10 = aω02
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 4.3. Varieties with degenerate Gauss mappings 113

for some function a : F  → C. Note that a cannot be normalized to zero.

Consider, on F 1 ,
de1 = e0 ω10 + e1 ω11 .
If a ≡ 0, then de1 ≡ 0 mod e1 , so [e1 ] ∈ PV is a fixed point independent of
x ∈ M and X is a cone over [e1 ], as described above.
If a is not identically zero, then de1 ≡ 0 mod e0 , e1 , so the points [e1 ]
sweep out a curve C such that [e0 ] ∈ T̃[e1 ] C. In this case X is the tangential
variety of C, because [e0 ] is a general point of X and dim τ (C) = 2. 

In the proof of Theorem 4.3.6 we only used the Frobenius Theorem.

With additional work one can deduce:
Theorem 4.3.7. Let M 2 ⊂ E3 be a surface with Gauss curvature K ≡ 0.
Then M minus a closed subset is the union of open subsets of the following
types of surfaces:
i. a (generalized) cylinder, i.e., a union of lines perpendicular to a plane
curve;
ii. a (generalized) cone, i.e., the union of lines connecting a fixed point
to a plane curve (minus the point);
iii. the union of tangent rays to a curve.
If M is complete, then it is a (generalized) cylinder.
See ([174], vol. III) for the proof.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/05

Chapter 5

Cartan-Kähler I:
Linear Algebra and
Constant-Coefficient
Homogeneous Systems

We have seen that differentiating the forms that generate an exterior dif-
ferential system often reveals additional conditions that integral manifolds
must satisfy (e.g., the Gauss and Codazzi equations for a surface in Eucli-
dean space). The conditions are consequences of the fact that mixed partials
must commute. What we did not see was a way of telling when one has dif-
ferentiated enough to find all hidden conditions. We do know the answer
in two cases: If a system is in Cauchy-Kowalevski form there are no extra
conditions. In the case of the Frobenius Theorem, if the system passes a
first-order test, then there are no extra conditions.

What will emerge over the next few chapters is a test, called Cartan’s
Test, that will tell us when we have differentiated enough.
The general version of Cartan’s Test is described in Chapter 8. For a
given integral element E ∈ Vn (I)x of an exterior differential system I on
a manifold Σ, it guarantees existence of an integral manifold to the system
with tangent plane E if E passes the test.
In Chapter 6, we present a version of Cartan’s Test valid for a class
of exterior differential systems with independence condition called linear
Pfaffian systems. These are systems that are generated by 1-forms and have
the additional property that the variety of integral elements through a point

115
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 116 5. Constant-Coefficient Homogeneous Systems

x ∈ Σ is an affine space. The class of linear Pfaffian systems includes all

systems of PDE expressed as exterior differential systems on jet spaces. One
way in which a linear Pfaffian system is simpler than a general EDS is that
an integral element E ∈ Vn (I, Ω)x passes Cartan’s Test if and only if all
integral elements at x do.

In this chapter we study first-order, constant-coefficient, homogeneous

systems of PDE for analytic maps f : V → W expressed in terms of tableaux.
We derive Cartan’s Test for this class of systems, which determines if the
initial data one might naı̈vely hope to specify (based on counting equations)
actually determines a solution.
We dedicate an entire chapter to such a restrictive class of EDS because
at each point of a manifold Σ with a linear Pfaffian system there is a naturally
defined tableau, and the system passes Cartan’s Test for linear Pfaffian
systems at a point x ∈ Σ if and only if its associated tableau does and the
torsion of the system (defined in Chapter 6) vanishes at x.
In analogy with the inverse function theorem, Cartan’s Test for linear
Pfaffian systems (and even in its most general form) implies that if the linear
algebra at the infinitesimal level works out right, the rest follows. What we
do in this chapter is determine what it takes to get the linear algebra to
work out right.

Throughout this chapter, V is an n-dimensional vector space, and W

is an s-dimensional vector space. We use the index ranges 1 ≤ i, j, k ≤ n,
1 ≤ a, b, c ≤ s. V has the basis v1 , . . . , vn and V ∗ the corresponding dual
basis v 1 , . . . , v n ; W has basis w1 , . . . , ws and W ∗ the dual basis w1 , . . . , ws .

5.1. Tableaux
Let x = xi vi , u = ua wa denote elements of V and W respectively. We will
consider (x1 , . . . , xn ), respectively (u1 , . . . , un ), as coordinate functions on
V and W respectively. Any first-order, constant-coefficient, homogeneous
system of PDE for maps f : V → W is given in coordinates by equations

∂ua
(5.1) Bari = 0, 1 ≤ r ≤ R,
∂xi

where the Bari are constants. For example, the Cauchy-Riemann system
u1x1 − u2x2 = 0, u1x2 + u2x1 = 0 has B111 = 1, B212 = −1, B112 = 0, B211 = 0,
B221 = 1, B122 = 1, B121 = 0 and B222 = 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.1. Tableaux 117

To phrase (5.1) in a coordinate-free manner, given a map f : V → W ,

we define a “Gauss map”
γf : V → V ∗ ⊗ W = Hom(V, W ),
∂f a (x)
wa ⊗v i , x →
∂xi
where we identify Tx V  V, Tf (x) W  W and translate the differential to
the origin. In other words, we identify Jx1 (V, W )/Jx0 (V, W ) with W ⊗ V ∗ .
More generally, we will identify Jxk (V, W )/Jxk−1 (V, W ) with W ⊗S k V ∗ , the
homogeneous W -valued polynomials of degree k on V .
Our system may be described free of coordinates as a subspace B ⊂
V ⊗W ∗ , where
B = {Bari vi ⊗ wa |1 ≤ r ≤ R}.
We think of B as the space of equations. A map f is a solution if B anni-
hilates γf (x) for all x ∈ V , i.e.,
b, γf (x) = 0 ∀x ∈ V, b ∈ B.
B is often called the space of symbol relations. (We will see that it gener-
alizes the principal symbol used in standard PDE terminology.) We think
of B ⊥ ⊂ W ⊗V ∗ as the space of admissible first derivatives (see below). In
computations we will often use B ⊥ rather than B, so it has a special name:
Definition 5.1.1. A tableau is a linear subspace A ⊆ W ⊗ V ∗ . A tableau A
determines a first-order, constant-coefficient, homogeneous system of PDE
for maps f : V → W , namely the system whose solutions are those maps f
satisfying γf (V ) ⊆ A.

Note that systems defined by a tableau always have solutions

f (x) = f0 + A0 x,
where f0 ∈ W and A0 ∈ A. We will be interested in what higher-order
terms can appear in the Taylor series of a solution.
Example 5.1.2. The equation u1x1 + u2x2 = 0 has symbol relations
B = {w1 ⊗v1 + w2 ⊗v2 } ⊂ W ∗ ⊗V
and tableau
(5.2) A = B ⊥ = {w1 ⊗ v 1 − w2 ⊗ v 2 , w1 ⊗v 2 , w2 ⊗v 1 } ⊂ W ⊗V ∗ .

Often it will be convenient to identify V with Rn , W with Rs , and to

use our fixed basis to write our expressions in matrix form. For example,
(5.2) becomes  
a c 
A= a, b, c ∈ R .
b −a 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 118 5. Constant-Coefficient Homogeneous Systems

Example 5.1.3. The Cauchy-Riemann system u1x1 = u2x2 , u1x2 = −u2x1 has
tableau
A ={a(w1 ⊗ v 1 + w2 ⊗ v 2 ) + b(−w2 ⊗ v 1 + w1 ⊗ v 2 )|a, b ∈ R}
 
a −b
 |a, b ∈ R .
b a
Example 5.1.4. The tableau A = (0) corresponds to a Frobenius system.
The equations are uaxi = 0, ∀i, a. The only solutions to this system are
constant maps. We will say solutions depend on s constants.
Example 5.1.5. When A = W ⊗ V ∗ , there are no equations and any map
is a solution. Here, solutions depend on s functions of n variables.
Example 5.1.6. Let L∗ ⊂ V ∗ be a k-dimensional linear subspace and let
A = W ⊗ L∗ . If L∗ = {v 1 , . . . , v k }, the equations are uaxρ = 0, k + 1 ≤ ρ ≤ n,
and the solutions are ua (x1 , . . . , xn ) = f a (x1 , . . . , xk ), where the f a are
arbitrary functions. Here, solutions depend on s functions of k variables.
With respect to adapted bases we may write A in block form as (∗, 0),
where the left block is free and the right block is zero.
Example 5.1.7. Let A = Y ⊗ V ∗ , where Y is a p-dimensional linear sub-
space of W . If Y is spanned by uλ , 1 ≤ λ ≤ p, the equations are uξxi = 0,
where p + 1 ≤ ξ ≤ s. Solutions are

uλ (x1 , . . . , xn ) = f λ (x1 , . . . , xn ),
uξ (x1 , . . . , xn ) = f0ξ ,
where the f λ (x1 , . . . , xn ) are arbitrary functions and the f0ξ are constants;
so, solutions depend on p functions of n variables and s − p constants. One
may think of this tableau as a combination of two disjoint systems of types
5.1.5 and 5.1.6. With respect to adapted bases we may write A in block
form as ( ∗0 ), where the top block is free and the bottom block is zero.

In the rest of this chapter, we will develop a test for a tableau such that
if it passes, we will know exactly what initial data one should specify to
determine an analytic solution to (5.1). What to do if the tableau fails the
test is taken up in Chapter 6.
Given a tableau A, suppose we try to specify the solution in terms of a
power series
(5.3) ua (x) = pa + pai xi + paij xi xj + paijk xi xj xk + . . . .
Since the system is homogeneous with constant coefficients, u is a solution
if and only if each term satisfies the system and the series converges. As we
have seen, there is no restriction on the constant terms, and the coefficients
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.1. Tableaux 119

of the linear term must satisfy pai wa ⊗v i ∈ A. Moving on to the coefficients

of the quadratic term in the series, the quadratic term must be such that
γpaij xi xj (x) ∈ A, ∀x ∈ V . Thus,

paij xi wa ⊗v j ∈ A ∀x ∈ V,
so that
(paij wa ⊗ v j ) ⊗ v i ∈ (A ⊗ V ∗ ).
Since mixed partials commute, (paij wa ⊗ v i ) ⊗ v j is also an element of W ⊗
S 2 V ∗ ; in indices this means paij = paji . Combining these two conditions gives

paij wa ⊗v i ⊗v j ∈ (A ⊗ V ∗ ) ∩ (W ⊗ S 2 V ∗ ) =: A(1) ,
where, at the risk of being redundant, the first factor is due to the equations
and the second to the commuting of derivatives. The space A(1) is called
the (first) prolongation of A.
Similarly, the condition on the next term of (5.3) is
(5.4) paijk wa ⊗v i ⊗v j ⊗v k ∈ (A ⊗ V ∗ ⊗ V ∗ ) ∩ (W ⊗ S 3 V ∗ ) =: A(2) ,
and so on for all l. In general, we define
A(l) := (A ⊗ V ∗⊗l ) ∩ (W ⊗ S (l+1) V ∗ )
to be the l-th prolongation of A. A map f : V → W , given in terms of a
convergent Taylor series, is a solution to the system defined by A if and only
if for all k, the tensor formed from the k-th term lies in A(k−1) .
Exercises 5.1.8:
1. Explicitly calculate A(1) when A is the tableau for the Cauchy-Riemann
equations.
2. Show that A(l) is naturally identified with the set of solutions to the
system defined by A that are homogeneous polynomials of degree l + 1.
3. Show that if A(l) = 0 for some l, then solutions depend on a finite number
of constants.
Definition 5.1.9. A tableau of order p is a linear subspace A ⊂ W ⊗S p V ∗ .
It determines a homogeneous constant-coefficient system of PDE of order p
for W -valued functions on V . In particular, the (p − 1)-st prolongation of a
tableau (of order one) is a tableau of order p. For a tableau A of order p,
we define its prolongations by A(l) = (A⊗V ∗⊗l ) ∩ (W ⊗S p+l V ∗ ).
Example 5.1.10. The equation uxy = 0 may be encoded as a tableau of
order two as
A = {p11 v 1 v 1 + p22 v 2 v 2 | p11 , p22 ∈ R}.
Here V = R2 , W = R.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 120 5. Constant-Coefficient Homogeneous Systems

Exercises 5.1.11:
1. Express the system uxx + vyy = 0 of one equation for two functions of
two variables as a tableau of order two. What is its dimension? 
2. Show that A(k+l) = (A(k) )(l) .

We would like to determine the space of solutions to (5.1) by performing

a finite calculation; in particular, we would like to avoid calculating A(l)
for all l. As it turns out, we will calculate A(1) , and if the system passes a
certain test, we will then know the dimension of A(l) for all l and the ‘size’
of the space of local solutions. To develop some intuition for what the test
should be, we examine some examples.

5.2. First example

Consider the tableau A = (p1 , 0, . . . , 0), where p1 is a column vector whose
entries are free. (This is Example 5.1.6 in the case dim L = 1.) This tableau
a
corresponds to equations ∂u ∂xρ = 0, where 2 ≤ ρ ≤ n. Its solutions are
ua (x1 , . . . , xn ) = f a (x1 ). In particular, specifying the values ua (x1 , 0, . . . , 0)
uniquely determines a solution, and all solutions are obtained this way.
We generalize to the tableau

A = {(pa1 v 1 + C2b
a b 2
p1 v ) ⊗ wa |pa1 ∈ R, 1 ≤ a ≤ s}
a b n
p1 v + . . . + Cnb
(5.5)
= {(p1 , C2 p1 , . . . , Cn p1 )},

where the Cρ are fixed s × s matrices and p1 = (pa1 ) is any column vector.
The symbol relations corresponding to (5.5) are

B = {wa ⊗vρ − Cρb

a b
w ⊗v1 |2 ≤ ρ ≤ n, 1 ≤ a ≤ s},

and the corresponding differential equations are

∂ua a ∂u
b
(5.6) − C ρb = 0, ρ = 2, . . . , n.
∂xρ ∂x1
We saw that when Cρ ≡ 0 ∀ρ, any convergent series with coefficients
pa1 , pa11 , pa111 , . . . determines a solution u. The hope is that under some con-
ditions the solutions of the system (5.6) are describable in a similar way. In
fact, we will determine when solutions to (5.6) can be given in terms of s
arbitrary functions of one variable, and will see that this is the largest space
of solutions one could hope for.
Assume we are given constants pa1 , pa11 , pa111 , · · · . Then we may determine
the remaining paI for any multi-index I = (i1 , . . . , ik ) as follows: To have
(pai wa ⊗ v i ) ∈ A, the terms paρ must be given by paρ = Cρb a pb . To have
1
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.2. First example 121

(paij wa ⊗v i v j ) ∈ A(1) , the other terms are determined by the pa11 :

paρ1 = Cρb
a b
p11 ,
(5.7)
paρσ = Cρb
a b a b c
pσ1 = Cρb Cσc p11 , 2 ≤ σ ≤ n.

These equations imply dim A(1) ≤ s. They may lead to conflicting equations
because it is also necessary that paρσ = paσρ , i.e., that Cρb
a C b pc = C a C b pc .
σc 11 σb ρc 11
In other words, to ensure that no choice of pa11 ’s leads to a conflict, it is
necessary that
(5.8) [Cρ , Cσ ]ac = Cρb
a b
Cσc − Cσb
a b
Cρc =0 ∀ a, c, ρ, σ.
If (5.8) holds, we are free to specify not only pa11 but in fact any convergent
power series for ua (x1 , 0, . . . , 0). More precisely,
Proposition 5.2.1. If the Cρ in tableau (5.5) are such that [Cρ , Cσ ] =
0 ∀ρ, σ, then there exists a unique solution to the initial value problem
γu (x) ∈ A = (p1 , C2 p1 , . . . , Cn p1 )
with initial condition
ua (x1 , 0, . . . , 0) = f a (x1 ),
where the f a are analytic.
Geometrically, the f a (x1 ) determine a curve in V × W , and we would
like to enlarge this curve to an n-dimensional integral manifold. By (5.7) we
see that there is at most one such, and the proposition states that if (5.8)
holds, this candidate is in fact an integral manifold. If (5.8) fails to hold,
there may still be solutions to (5.5), but there will be additional conditions
on the functions f a .

Proof of 5.2.1. We have seen that in this case the paρi are exactly deter-
mined by the pa11 , and the same computation shows the paρij are exactly
determined by the pa111 , etc. (The commutation relations take care of all
possible conflicts.) It remains to show that the formal power series for
u(x1 , . . . , xn ) determined by these coefficients converges.
Let c =max{1, maximum eigenvalue of any Cρ }. Consider the system
∂U a ∂U a
= c , ρ = 2, . . . , n.
∂xρ ∂x1
Its general solution is given by
U a (x1 , . . . , xn ) = F a (x1 + c(x2 + . . . + xn )),
where F a (t) are analytic functions of one variable. If
F a (t) = pa + pa1 t + pa11 t2 + pa111 t3 + . . . ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 122 5. Constant-Coefficient Homogeneous Systems

then we may use the functions F a to construct a formal solution to (5.5):

(5.9) ua (x1 , . . . , xn )
= pa + (pa1 x1 + Cρb
a b ρ
p1 x ) + (pa11 (x1 )2 + Cρb
a b 1 ρ a b c ρ σ
p11 x x + Cρb Cσc p11 x x )
+ (pa111 (x1 )3 + Cρb
a b
p111 (x1 )2 xρ
a b c
+ Cρb Cσc p111 x1 xρ xσ + Cρb
a b
Cσc Cτcd pd111 xρ xσ xτ ) + . . . .
On the other hand, we see that any derivative of u is bounded by the corre-
sponding derivative of U :
   
 ∂ I ua   ∂I U a 
   
 ∂xi1 . . . xid (0) ≤  ∂xi1 . . . xid (0) .
Since the series (5.9) for u is bounded in norm by a convergent series, it
must converge as well. 
Remark 5.2.2. The proof of 5.2.1 is an example of the method of majo-
rants.

In summary, the largest space of solutions one could hope for in a system
of the form (5.5) is that solutions depend on s functions of 1 variable, and
whether or not this is the case can be determined just by computing A(1) .
One always has dim A(1) ≤ s, and solutions depend on s functions of one
variable if and only if dim A(1) = s.

5.3. Second example

Now we will study a generalization of the simple tableau
 λ λ   
p1 p2 0 p1 p2 0
A= = , pλ1 , pλ2 , pξ1 ∈ R,
(5.10) pξ1 0 0 q1 0 0
1 ≤ λ, μ ≤ k, k + 1 ≤ ξ, η ≤ s.
This tableau corresponds to the equations uξx2 = 0, uax3 = 0. Its solutions
depend on k functions of 2 variables and s−k functions of 1 variable, namely
uξ = f ξ (x1 ) and uλ = f λ (x1 , x2 ). Note that here dim A(1) = s + 2k, since
one is free to specify pa11 , pλ12 , pλ22 , while all the other paij must be zero.
Our generalization is
   
p1 p2 F p1 + Gq1 + Hp2 p1 p2 p3
(5.11) A = =
q1 Cp1 + Dq1 + Ep2 Ip1 + Jq1 + Kp2 q1 q2 q3
where C, E, I, K ∈ M(s−k)×k , D, J ∈ M(s−k)×(s−k) , F, H ∈ Mk×k , G ∈
Mk×(s−k) , and Mr×s denotes the set of r × s matrices. In this tableau there
are s independent entries in the first column and k new independent entries
in the second column. So, the entries in p1 , p2 , q1 are all independent and
all other entries are linear combinations of them.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.3. Second example 123

The system of differential equations corresponding to A is

∂uξ ξ ∂u
λ
ξ ∂u
η
ξ ∂u
λ
= C λ ∂x1 + D η + Eλ ∂x2 ,
∂x2 ∂x1
∂uξ ∂uλ ∂uη ∂uλ
3
= Iλξ 1 + Jηξ 1 + Kλξ 2 ,
∂x ∂x ∂x ∂x
∂u μ λ η λ
μ ∂u μ ∂u μ ∂u
= F λ + G η + H λ .
∂x3 ∂x1 ∂x1 ∂x2
As before, for any solution we are free to specify the order zero data arbi-
trarily, and first-order data must belong to A. For the second-order terms
of a power series solution, we see that we could at most specify the vectors
p11 , p12 , p22 , and q11 , where pij = t(p1ij · · · pkij ) and qij = t(pk+1
ij · · · pij ). Then
s
a
the remainder of the pij are determined by (5.11), e.g.,
q21 = Cp11 + Dq11 + Ep21 = q12 ,
q22 = Cp12 + Dq12 + Ep22 = Cp12 + D(Cp11 + Dq11 + Ep21 ) + Ep22 ,
p31 = F p11 + Gq11 + Hp21 .
Now there are two ways to describe q32 = q23 , namely
q32 = Ip12 + Jq12 + Kp22
= Ip12 + J(Cp11 + Dq11 + Ep21 ) + Kp22 ,
q23 = Cp13 + Dq13 + Ep23
= C(F p11 + Gq11 + Hp21 ) + D(Ip11 + Jq11 + Kp21 )
+ E(F p12 + G(Cp11 + Dq11 + Ep21 ) + Hp22 ).
In order to have consistent equations for all choices of p11 , p12 , p22 , q11 , the
four matrix equations
JC = CF + DI + EGC,
I + JE = CH + DK + EF + EGE,
(5.12)
K = EH,
JD = CG + DJ + EGD
must hold. (These are obtained by setting the p11 , p12 , p22 and q11 coefficients
of q23 equal to the corresponding coefficients of q32 .) It is easy to check that
if (5.12) holds, we are able to freely specify p111 , q111 , p112 , p122 , p222 . These
exactly determine the value of the 2-jet within A(2) ; similarly, for l > 2 an
element of A(l) is determined by q1,...,1 and pi1 ,...,il with i1 , . . . , il ∈ {1, 2}.
Proposition 5.3.1. If the matrix equations (5.12) hold, then there is a
unique solution to the initial value problem
γu (x) ⊂ A,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 124 5. Constant-Coefficient Homogeneous Systems

where A is the tableau (5.11) with initial condition

uλ (x1 , x2 , 0) = f λ (x1 , x2 ), uξ (x1 , 0, 0) = f ξ (x1 ),
where f λ , f ξ are analytic.

Proof. The only thing to check is the claim about the determination of A(l)
and the convergence of the power series. These are left to the reader. 

For any tableau of the form (5.11), we have dim A(1) ≤ s + 2k. Proposi-
tion 5.3.1 says that, when equality holds, solutions depend on k functions of
2 variables and (s − k) functions of 1 variable, which is the largest possible
space of solutions.
Remark 5.3.2 (another view of 5.3.1). One could attempt to solve the
system (5.11) by solving a sequence of Cauchy problems. For example,
suppose
uξ (x1 , 0, 0) = f ξ (x1 ),
uλ (x1 , x2 , 0) = f λ (x1 , x2 )
are the functions we specify to determine a putative solution to the system
(5.11). We could first solve the Cauchy problem
(5.13)
uξx2 (x1 , x2 , 0) = Cλξ uλx1 (x1 , x2 , 0) + Dηξ uηx1 (x1 , x2 , 0) + Eλξ uλx2 (x1 , x2 , 0),
uξ (x1 , 0, 0) = f ξ (x1 )
to determine uξ (x1 , x2 , 0), then solve the Cauchy problem
(5.14)
uλx3 (x1 , x2 , x3 ) = Fμλ uμx1 (x1 , x2 , x3 ) + Gλξ uξx1 (x1 , x2 , x3 ) + Eμλ uμx2 (x1 , x2 , x3 ),
uλ (x1 , x2 , 0) = f λ (x1 , x2 )
to determine uλ (x1 , x2 , x3 ), and finally obtain a map u : V → W by solving
(5.15)
uξx3 (x1 , x2 , x3 ) = Iλξ uλx1 (x1 , x2 , x3 ) + Jηξ uηx1 (x1 , x2 , x3 ) + Kλξ uλx2 (x1 , x2 , x3 ),
uξ (x1 , x2 , 0) = solution to (5.13).
Solving (5.13) determines the constants q2 , q12 , q22 , q112 , . . . for which all sub-
scripts are 1 or 2, with at least one being 2. Solving (5.14) determines
p3 , p13 , . . . and solving (5.15) determines p23 , p33 , . . . .
Just as with the similar procedure we tried in Example 1.2.3, this pro-
cedure yields a function which may not solve our system. As with Example
1.2.3, there are other sequences of Cauchy problems—e.g., we could have
solved first for uξ (x1 , 0, x3 )—to obtain a function. However, condition (5.12)
guarantees that the function we obtain is actually a solution to our system
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.4. Third example 125

and that no matter which sequence of Cauchy problems we solve, we al-

ways obtain the same answer. (Again, the failure of (5.12) to hold does not
exclude the possibility that some sequence of Cauchy problems will yield a
solution for certain initial data; but (5.12) guarantees that this process will
work with all choices of initial data.)

5.4. Third example

To begin with, suppose the tableau A is of the form
⎛ 1 ⎞
p1 p12 . . . p1k 0 ... 0
⎜ .. .. .. .. ⎟
⎜ . . . . 0 . . . 0⎟
⎜ ⎟
⎜ .. .. .. ⎟
⎜ . . . pskk 0 . . . 0⎟
⎜ . ⎟
⎜ . .. .. ⎟
⎜ . . . 0 0 . . . 0⎟
⎜ . ⎟
⎜ . .. ⎟
⎜ . ps22 0 . 0 . . . 0⎟
(5.16) A=⎜ .⎜ ⎟.
. . ⎟
⎜ .. 0 .. .. 0 . . . 0⎟
⎜ . . .. ⎟
⎜ps1 .. .. 0 . . . 0⎟
⎜ 1 . ⎟
⎜ . . . ⎟
⎜0 .. .. .. 0 . . . 0⎟
⎜ ⎟
⎜ . .. .. .. ⎟
⎜ .. . . . 0 . . . 0⎟
⎝ ⎠
.. .. ..
0 . . . 0 ... 0

Then dim A(1) = s1 + 2s2 + 3s3 + . . . + ksk , since an element of A(1) is deter-
mined by a choice of s1 constants p111 , . . . , ps111 , then 2s2 constants p112 , . . . , ps122 ,
p122 , . . . , ps222 , then 3s3 constants p113 , . . . , ps133 , p123 , . . . , ps323 , p133 , . . . , ps333 , and so
on, up to ksk constants p11k , . . . , pskkk .
The general solution of the corresponding system of differential equations
is
ua = f a (x1 , . . . , xk ), 1 ≤ a ≤ sk ,
ua = f a (x1 , . . . , xk−1 ), 1 + sk ≤ a ≤ sk−1 ,
..
.
ua = f a (x1 ), 1 + s2 ≤ a ≤ s1 ,
u =a
ua0 (constant), 1 + s1 ≤ a ≤ s.

If we generalize tableau (5.16) by inserting linear combinations of the

entries occurring above and to the left of each zero, we obtain a tableau of
the same dimension, which satisfies the inequality
dim A(1) ≤ s1 + 2s2 + . . . + ksk
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 126 5. Constant-Coefficient Homogeneous Systems

with equality if and only if any choice of the s1 + 2s2 + . . . + ksk constants
paij for a point in A(1) determines the other paij ’s without conflict. Moreover,
in this case A(k) has its maximum possible dimension for all k as well,
i.e., commutation of all second-order derivatives implies commutation of all
higher-order derivatives.

5.5. The general case

Definition 5.5.1. Given a tableau A ⊂ W ⊗ V ∗ expressed in terms of bases
b = (v 1 , . . . , v n ) of V ∗ and q = (w1 , . . . , ws ) of W , let s1 (b), . . . , sn (b) be
defined by
s1 (b) = # of independent entries in the first column of A,
s1 (b) + s2 (b) = # of independent entries in the first 2 columns of A,
..
.
s1 (b) + . . . + sn (b) = # of independent entries in A = dim A.
In other words, sk (b) is the number of new independent entries in the k-th
column.

Note that the sk (b) do not depend on the choice of basis for W , but only
on the flag F = (F0 , F1 , . . . , Fn ) of subspaces in V ∗ induced by b, which we
denote by
Fj = {v1 , . . . , vj }⊥ = {v j+1 , . . . , v n },
with F0 = V ∗ and Fn = (0). Hence, we write sk (F ) instead of sk (b).
We define
Ak (F ) = (W ⊗Fk ) ∩ A
and observe that
(5.17) dim Ak (F ) = sk+1 (F ) + . . . + sn (F ).
One can visualize Ak (F ) as the subspace of matrices in A for which the first
k columns are zero, when we use the basis b for V . In our first example,
A1 = (0); in our second example, A1 is the k-dimensional subspace of A
obtained by setting p1 , q1 = 0.
Definition 5.5.2. Let A ⊂ W ⊗V ∗ be a tableau. Define
s1 (A) = max{s1 (F )| all flags},
s2 (A) = max{s2 (F )| flags with s1 (F ) = s1 (A)},
..
.
sn (A) = max{sn (F )| flags with s1 (F ) = s1 (A), . . . , sn−1 (F ) = sn−1 (A)}.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.5. The general case 127

The si are called the (reduced) characters of A. They are invariants of A

with respect to the action of GL(V ) × GL(W ). We will call a flag F an
A-generic flag when si (F ) = si for all i. (We will often write si instead of
si (A) when there is no risk of confusion.)

Notice that s ≥ s1 ≥ . . . ≥ sn ≥ 0. For example, tableau (5.5) has

characters s1 = s, s2 = s3 = . . . sn = 0, while (5.10) has characters s1 = s,
s2 = k, s3 = 0.
We fix an A-generic flag F induced by a basis b, and will suppress further
reference to it. Given U ⊆ W ⊗S d V ∗ , we define
Uk := U ∩ (W ⊗S d {v k+1 , . . . , v n }).
Note that
(Ak )(1) = (A(1) )k .
Proposition 5.5.3.
(5.18) dim A(1) ≤ s1 + 2s2 + 3s3 + . . . + nsn .

Proof. We have the exact sequence

0 → Ak (1) → Ak−1 (1) → Ak−1 ,
in which the last map is p → vk p. Thus
dim Ak−1 (1) − dim Ak (1) ≤ dim Ak−1 .
Summing for 1 ≤ k ≤ n (and recalling that A0 = A), we have
dim A(1) ≤ dim A + dim A1 + . . . + dim An−1 ,
which, along with (5.17), gives the inequality of the proposition. 
Definition 5.5.4. A tableau A ⊂ W ⊗V ∗ is said to be involutive if equality
holds in (5.18).

For the “model” involutive tableau (5.16), solutions are uniquely deter-
mined by specifying
uσ (x1 , . . . , xk , 0, . . . , 0) for σ ≤ sk .
This motivates the following:
Definition 5.5.5. For an integer σ between 1 and s, define the level of σ to
be the largest k such that σ ≤ sk . (If σ > s1 , its level is defined to be zero.)
Example 5.5.6. Tableau (5.10) has characters s1 = s, s2 = s − k, s3 = 0,
so
level(1) = level(2) = . . . = level(s − k) = 2
while
level(s − k + 1) = . . . = level(s) = 1.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 128 5. Constant-Coefficient Homogeneous Systems

For any tableau, the most general set of initial data one could hope to
specify is
(5.19) uσ (x1 , . . . , xk , 0, . . . , 0), where level(σ) = k,
with an integral manifold obtained by then solving a sequence of Cauchy
problems. In fact, an argument that is conceptually the same as in the
examples of the previous sections yields
Theorem 5.5.7 (Cartan-Kähler for tableaux). Let A ⊂ W⊗V ∗ be a tableau.
Choose A-generic bases of V, W which induce coordinates xi on V and ua on
W . If A is involutive, then any choice of analytic functions (5.19) uniquely
determines an integral manifold of the differential system represented by A
in some neighborhood of the origin.
Definition 5.5.8. If A is an involutive tableau such that sl = 0 but sl+1 = 0,
then sl is called the character of the system and the number l is called the
Cartan integer.

According to Theorem 5.5.7, a solution is determined by specifying

sl functions of l variables,
sl−1 − sl functions of l − 1 variables,
..
.
s1 − s2 functions of 1 variable
and s − s1 constants.
The freedom of the functions of l variables is more significant than all the
other choices in the sense that, for large p, in the pth-order term in the
Taylor series expansion of a solution there will be many more coefficients
coming from the functions of l variables than the others. Thus we usually
will say that for an involutive EDS with character sl the integral manifolds
depend on sl functions of l variables, and ignore the rest.
Exercises 5.5.9:
1. Show that a tableau A ⊂ W ⊗V ∗ having characters s1 = s2 = . . . sn−1 =
s, sn = 0 is always involutive.
2. Consider the tableau for the Cauchy-Riemann equations uxi = vyi , uyi =
−vxi for two functions u, v of 2n variables x1 , . . . , xn , y1 , . . . , yn . Calculate
the characters and show that the tableau is involutive. Describe the size of
the space of solutions based on Cartan’s Test, and say how this coincides
with how you learned how to construct solutions of this system in your first
complex analysis course.
3. Let A ⊂ W ⊗ V ∗ be any tableau and consider the enlarged tableau
 
A 0
à = ⊂ W̃ ⊗ Ṽ ∗ ,
0 0
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.6. The characteristic variety of a tableau 129

where V, W are subspaces of Ṽ , W̃ respectively. Show that A and à have

the same nonzero characters and that à is involutive if and only if A is.
4. Let dim V = n and endow V with a volume form and inner product.
Recall from §3.8 the Laplacian Δ = ∗d ∗ d + d ∗ d∗ : Ωp (V ) → Ωp (V ),
a second-order differential operator. (Here we identify V with its tangent
space at any point.)
(a) The tableau for Laplace’s system Δf = 0 for functions f : V → W is
W ⊗ S02 V ∗ ⊂ W ⊗ S 2 V ∗ , where S02 V ∗ is the traceless symmetric tensors in
V ∗ ⊗ V ∗ . (We give V an inner product and the trace is with respect to that
inner product.) What are the symbol relations?
(b) If dim V = dim W = 2, show that the Cauchy-Riemann system implies
Laplace’s system in the sense that any solution to the Cauchy-Riemann
system must also solve the Laplace system.
(c) Consider a p-form α as a Λp V -valued function on V . Write down the
tableau of the system defined by the two equations dα = 0 and d ∗ α = 0
for p = 1, and show that its prolongation is contained in W ⊗S02 V ∗ , where
W = Λp V .

5.6. The characteristic variety of a tableau

In Theorem 5.5.7 we do not necessarily obtain all solutions to an involutive
system by a choice of analytic functions as specified in the theorem. We
only obtain those that can be obtained by specifying noncharacteristic initial
data, that is, initial data based on a choice of A-generic flag for V ∗ . (The
others can be obtained by the Cartan algorithm described in Chapter 6,
but it involves specializations to nongeneric submanifolds.) If we choose
a nongeneric flag, the corresponding Cauchy problem may not have any
solutions, or it may be undetermined, with an infinite number of solutions.
Example 5.6.1. Consider the system
(5.20) ux − vy = 0, uy − vx = 0.
If we pick initial data for u, v along a line through the origin other than
x = ±y, then this extends to a unique solution to the system. But if we
specify initial data along the line y = x, then unless u − v is constant this
cannot be extended to a solution. If u − v is constant along y = x, then
there are an infinite number of extensions to a local solution.
Definition 5.6.2. For ξ ∈ V ∗ , define the symbol mapping at ξ by
σξ : W → (W ⊗V ∗ )/A,
σξ (w) = w⊗ξ mod A.
Definition 5.6.3. Define ΞA ⊂ PV ∗ , the characteristic variety of A, to be
ΞA = {[ξ] ∈ PV ∗ | ker σξ = 0}.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 130 5. Constant-Coefficient Homogeneous Systems

We can interpret ΞA as the set of hyperplanes in PV for which the

extension of an (n − 1)-dimensional integral element to an n-dimensional
integral element is not unique; see Theorem 6.7.7.
Example (5.6.1 continued). The tableau A corresponding to (5.20) consists
of matrices of the form ab ab . This has characters s1 = 2 = dim A and s2 =
0. A line in V spanned by t(x, y) is characteristic if the system ax + by = 0,
bx + ay = 0 of equations for a, b has rank less than two. Of course, this
happens for the two lines y = ±x. Thus ΞA consists of the two points [1, 1]
and [1, −1].
Example 5.6.4. Suppose dim V = 2. If s2 = 0, then ΞA = PV ∗ . If s1 = s
and s2 = 0, then ΞA is a collection of points.
Example 5.6.5. Let V and W be three-dimensional, and let A be the
five-dimensional space of matrices of the form
⎛ ⎞
a b c
⎝ b c d⎠ .
c d e
It is easy to see that A has characters s1 = 3, s2 = 2, s3 = 0. We have
⎛ ⎞
ξ1 w1 ξ2 w1 ξ3 w1
σξ (w) ≡ ⎝ξ1 w2 ξ2 w2 ξ3 w2 ⎠ mod A.
ξ1 w3 ξ2 w3 ξ3 w3
In order for this matrix to be zero modulo A, we must have ξ2 w1 = ξ1 w2 ,
ξ2 w3 = ξ3 w2 , and ξ1 w3 = ξ2 w2 = ξ3 w1 for some nonzero w. The first two
equations show we must have w1 = ξξ12 w2 and w3 = ξξ32 w2 . Plugging this into
the fourth equation, we obtain ξ1 ξ3 = ξ22 , and the last equation is redundant.
Thus the characteristic variety is the conic {ξ1 ξ3 − ξ22 = 0} in PV ∗ ∼ = RP2 .
Exercise 5.6.6: Consider the four-dimensional tableau A of matrices of the
form ⎛ ⎞
a b c
⎝b a b⎠ .
d b a
Show that ΞA consists of four points in PV ∗ .

Even when we are only interested in real-valued solutions of the under-

lying PDE, it is useful to consider the complex characteristic variety
ΞC ∗ C
A := {[ξ] ∈ PVC | ker σξ = 0},

where σξC : WC → (WC⊗VC∗ )/AC is the complexification of σξ (see Appendix

C).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.6. The characteristic variety of a tableau 131

Definition 5.6.7. We say A is determined if dim A⊥ = s, i.e., the number

of equations is the same as the number of unknowns.
Exercise 5.6.8: Show that if A is determined and ΞC ∗
A = PVC , then A is
involutive.
Definition 5.6.9. Recall from Chapter 4 that the Segre variety
Seg(PW × PV ∗ ) ⊂ P(W ⊗V ∗ )
is the set of decomposable vectors in P(W ⊗V ∗ ).

Let pV ∗ be the projection from Seg onto the first factor:

Seg(PV ∗ × PW )
pV ∗  S

/ SS
w
PV ∗ PW.

Exercise 5.6.10: Let A ⊂ W ⊗V ∗ be a tableau. Show that

(5.21) ΞA = pV ∗ (PA ∩ Seg(PV ∗ × PW )).

If U ⊆ W ⊗S p V ∗ is a tableau of order p, let vp : PV ∗ → PS p V ∗ denote

the Veronese re-embedding [v] → [v p ] (see Chapter 4), and define ΞU =
pV ∗ (PU ∩ Seg(PW ⊗vp (PV ∗ ))).
Exercises 5.6.11:
1. If in a given basis the symbol relations are spanned by R matrices Bari of
dimension n × s, and ξ = ξi v i , then σξ is given by the R × s matrix (Bari ξi ).
In particular, if R < s, then ΞA = PV ∗ .
2. Again show that if R < s, then ΞA = PV ∗ , this time using (5.21) as
the definition of ΞA , and the fact that varieties of complementary dimension
must intersect in projective space.
3. Show that if U = A(k) , then ΞU = ΞA .

Dimension and degree of the characteristic variety. In this subsec-

tion, we work over C and drop the C from the notation.
We estimate the dimension of ΞA and a modification of the degree of ΞA
in terms of the characters of A. We use the convention that the empty set
∅ has dimension −1, and denote the modified degree, which will be defined

below, by deg.
Our motivation is the following theorem:
Theorem 5.6.12. Let A ⊂ W ⊗V ∗ be a tableau. Noncharacteristic integral
 C)
manifolds of the exterior differential system induced by A depend on deg(ΞA
functions of dim ΞCA + 1 variables.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 132 5. Constant-Coefficient Homogeneous Systems

We will prove this theorem in Chapter 6. It follows from first proving

that some prolongation of A is involutive and the characteristic variety is
unchanged under prolongation.
Let A be a tableau with Cartan integer k. First note that dim ΞA ≤ k−1,
because PFk ∩ ΞA = ∅ if Fk ⊂ V is a generic codimension k subspace. Thus
the projection π : PV ∗ → P(V ∗ /Fk ) restricted to ΞA is a finite map. (If it is
surjective, its degree is deg(ΞA ).) We need to define deg precisely because
π need not be finite.
 we begin with the case where A is such
To arrive at the definition of deg
that k = 1, and hence dim(ΞA ) ≤ 0.
After choosing adapted bases, we may write A as a subspace of the space
of n × s matrices: A = (p, C2 p, . . . , Cn p), where the Cρ are s × s matrices
and p ∈ Rs1 ⊂ Rs is variable, t p = (p1 , . . . , ps1 , 0, . . . , 0). Here ΞA = ∅ iff
there is a choice of p such that the matrix A(p) is of rank one. (If s1 = s
this occurs if and only if there exists a simultaneous eigenvector for the Cρ .)
In particular, if A is involutive, then ΞA = ∅.
How many points can be in ΞA ? If the Cρ are all simultaneously diag-
onalizable in the upper s1 × s1 blocks with zeros elsewhere, then there are
s1 points, and since these s1 points span A, this is the most possible. It
may happen that one can normalize to upper s1 × s1 blocks with zeros else-
where but with simultaneous eigenspaces of dimension greater than one in
the nonzero block. This lowers the number of points, but for each eigenspace
of dimension μ, there is a vector ξ ∈ V ∗ such that there is a μ-dimensional
subspace U ⊆ W such that [ξ ⊗w] ∈ (PA ∩ Seg) for all w ∈ U .
Since the condition of primary interest is involutivity, we consider a
modification of the degree:
We decompose ΞA into irreducible components ΞA = α
α ΞA , where the
components ΞαA are of maximal dimension. We set

 A=
degΞ μα deg ΞαA ,
α

where μα is the dimension of pV ∗ −1 ([ξ]) for a general point [ξ] ∈ ΞαA .

We prove the following result, which is a special case of Theorem V.3.6
in [27]. The general case is a theorem about the characteristic sheaf of a
linear Pfaffian system, and is much more difficult to prove due to the possible
presence of nilpotents in the stalk (i.e., the characteristic scheme at a point).
The proof in [27] uses the Grothendieck-Riemann-Roch theorem.
Theorem 5.6.13. Let A ⊂ W ⊗ V ∗ be a tableau with sk (A) = 0 and
sk+1 (A) = 0. Then dim ΞA ≤ k − 1, and equality holds if and only if,
for a generic codimension k − 1 subspace Fk−1 , PFk−1 ∩ ΞA = ∅, i.e.,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 5.6. The characteristic variety of a tableau 133

PAk−1 ∩ Seg(PFk−1 × PW ) = ∅. In particular, equality holds if Ak is invo-

 A ≤ sk , with equality if and only if Ak
lutive. If dim ΞA = k − 1, then degΞ
is involutive.

Proof. We have already seen the first inequality. Equality occurs if and
only if PFk−1 ∩ ΞA = ∅, i.e., iff PAk−1 ∩ Seg(PFk−1 × PW ) = ∅. But as
a tableau, Ak−1 has characters s1 (Ak−1 ) = sk (A), s2 (Ak−1 ) = 0, so we are
reduced to the special case. Since PFk−1 is a generic linear space, the same
calculation as above is also valid for the degree. 
 of the characteristic variety may be easier to
The dimension and deg
compute than carrying out Cartan’s Test. More importantly, this theorem
is an important step towards proving Theorem 5.6.12, stated above.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/06

Chapter 6

Cartan-Kähler II: The

Cartan Algorithm for
Linear Pfaffian Systems

We now generalize the test from Chapter 5 to a test valid for a large class of
exterior differential systems called linear Pfaffian systems, which are defined
in §6.1. In §§6.2–6.4 we present three examples of linear Pfaffian systems that
lead us to Cartan’s algorithm and the definitions of torsion and prolongation,
all of which are given in §6.5. For easy reference, we give a summary and
flowchart of the algorithm in §6.6. Additional aspects of the theory, includ-
ing characteristic hyperplanes, Spencer cohomology and the Goldschmidt
version of the Cartan-Kähler Theorem, are given in §6.7. In the remainder
of the chapter we give numerous examples, beginning with elementary prob-
lems coming mostly from surface theory in §6.8, then an example motivated
by variation of Hodge structure in §6.9, then the Cartan-Janet Isometric Im-
mersion Theorem in §6.10, followed by a discussion of isometric embeddings
of space forms in §6.11 and concluding with a discussion of calibrations and
calibrated submanifolds in §6.12.

6.1. Linear Pfaffian systems

Recall that a Pfaffian system on a manifold Σ is an exterior differential
system generated by 1-forms, i.e., I = {θa }diff , θa ∈ Ω1 (Σ), 1 ≤ a ≤ s. If
Ω = ω 1 ∧ . . . ∧ ω n represents an independence condition, let J := {θa , ω i }
and I := {θa }. We will often use J to indicate the independence condition
in this chapter, and refer to the system as (I, J).

135
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 136 6. The Cartan Algorithm for Linear Pfaffian Systems

Definition 6.1.1. (I, J) is a linear Pfaffian system if dθa ≡ 0 mod J for

all 1 ≤ a ≤ s.

Exercise 6.1.2: Let (I, J) be a linear Pfaffian system as above. Let π , for
1 ≤  ≤ dim Σ − n − s, be a collection of 1-forms such that T ∗ Σ is locally
spanned by θa , ω i , π . Show that there exist functions Aai , Tija defined locally
on Σ such that

(6.1) dθa ≡ Aai π ∧ ω i + Tija ω i ∧ ω j mod I.

Example 6.1.3. The canonical contact system on J 2 (R2 , R2 ) is a linear

Pfaffian system because

d(du − p11 dx − p12 dy) = −dp11 ∧ dx − dp12 ∧ dy

≡ 0 mod{dx, dy, du − p11 dx − p12 dy, dv − p21 dx − p22 dy},
d(dv − p21 dx − p22 dy) = −dp21 ∧ dx − dp22 ∧ dy
≡ 0 mod{dx, dy, du − p11 dx − p12 dy, dv − p21 dx − p22 dy},

and the same calculation shows that the pullback of this system to any
submanifold Σ ⊂ J 2 (R2 , R2 ) is linear Pfaffian. More generally, we have

Example 6.1.4. Any system of PDE expressed as the pullback of the con-
tact system on J k (M, N ) to a subset Σ is a linear Pfaffian system. If M
has local coordinates (x1 , . . . , xn ) and N has local coordinates (u1 , . . . , us ),
then J k = J k (M, N ) has local coordinates (xi , ua , paL ), where L = (l1 , . . . , lp )
ranges over symmetric multi-indices with 1 ≤ |L| ≤ k. In these coordinates,
the contact system I is {θL̃a = dpaL̃ − paL̃j dxj }, where L̃ is a symmetric multi-
index with 1 ≤ |L̃| ≤ k − 1, θLa = dpa − pa dxj }, and J = {θ a , θ a }. On
L Lj L̃
J k (M, N ),
)
dθa = −dpaj ∧ dxj
≡ 0 mod J,
dθL̃a = −dpaL̃j ∧ dxj

and these equations continue to hold when we restrict to any subset Σ ⊂ J k .

Example 6.1.5. On R6 , let θ = y 1 dy 2 + y 3 dy 4 + y 5 dx, let I = {θ} and let

J = {θ, dx}. Then

dθ = dy 1 ∧ dy 2 + dy 3 ∧ dy 4 + dy 5 ∧ dx
y3 1
≡ (dy 3 − dy ) ∧ dy 4 mod{θ, dx}.
y1
In this case, (I, J) is not a linear Pfaffian system.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.2. First example 137

Exercises 6.1.6:
1. On ASO(n + s), with the notation from §2.5, let I = {ω a } and J =
{ω a , ω j }. The integral manifolds of (I, J) are the first-order adapted lifts of
immersed submanifolds M n ⊂ En+s . Show that (I, J) is a linear Pfaffian
system.
n+1
2. On ASO(n + s) × Rs( 2 ) , let the second factor have coordinates hajk =
hakj . Let I = {ω a , ωja − hajk ω k } and J = {ω a , ωja − hajk ω k , ω j }. Show that
the integral manifolds of (I, J) are the first-order adapted lifts of immersed
submanifolds M n ⊂ En+s together with their second fundamental forms,
and that (I, J) is a linear Pfaffian system.

In the following three sections, we will arrive at the Cartan algorithm

informally, via examples, and then state the algorithm formally in §6.5.

6.2. First example

First, let’s reconsider our system of Example 1.2.3:
θ := du − A(x, y, u)dx − B(x, y, u)dy,
Ω := dx ∧ dy.

We differentiated, and obtained

dθ ≡ (Ay − Bx + Au B − Bu A)dx ∧ dy mod I,

and saw that we had to restrict to the submanifold Σ ⊂ Σ = R3 defined

by Ay − Bx + Au B − Bu A = 0. The term Ay − Bx + Au B − Bu A, which
is an obstruction to the existence of integral manifolds, is a special case
of torsion, which will be defined invariantly in §6.5. For now, we use the
following coordinate definition: we say that apparent torsion is present if in
the expression (6.1) the functions Tija are not identically zero. We will say the
apparent torsion is absorbable if there exists a different choice of πia ’s such
that the functions Tija become zero, and otherwise that there is torsion. In
the example at hand, since θ, dx, dy already span T ∗ Σ, the apparent torsion
must be torsion. In notation that will be made invariant in §6.5, we let
T denote the apparent torsion, i.e., the collection of functions Tija , and we
write [T ] = 0 if the apparent torsion is absorbable.
Exercises 6.2.1:
1. Show that if π̃ = Mδ π δ for some invertible matrix (Mδ ), then the new
apparent torsion T̃ija is identically zero if and only if the original apparent
torsion Tija was identically zero. Conclude that the only possible way to
absorb apparent torsion is to change the complement of the subspace J ⊂
T ∗ Σ spanned by the π ’s.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 138 6. The Cartan Algorithm for Linear Pfaffian Systems

2. Show that if π̃ = π + Ma θa , then T̃ija = Tija . Conclude that the only

possible way to absorb apparent torsion is to change the complement of
J/I ⊂ T ∗ Σ/I spanned by the π ’s modulo I.
3. Show that at least in some cases one may absorb apparent torsion by
changing the complement of the flag J/I ⊂ T ∗ Σ/I by beginning with a
system without torsion (e.g., the tautological system on J 2 (R2 , R2 )) and
making a “bad” initial choice of π ’s such that there is apparent torsion.

The first step in our informal algorithm is to determine equations for

the torsion, and the next step is to restrict to a submanifold Σ ⊂ Σ where
the torsion is zero and check if the independence condition still holds.
We record these observations in the following flowchart:

Input:
linear Pfaffian system
Rename Σ as Σ - (I, J) on Σ;
calculate dI mod I
6

?
N Q
Q  Q Y
 Q  Q
  QQ Q Is [T ] = 0? Q

- ?
Q Is Σ empty? Q 
Q k
Q QQ 
Q 
Q Q N
Y Q
Q ?
? Q Restrict to Σ ⊂ Σ
Q
Done: defined by [T ] = 0
there are no and Ω |Σ = 0
integral manifolds

Remark 6.2.2. The set defined by setting the torsion to zero may have
many components and strata (as an analytic space). To find all integral
manifolds to a system, one would have to restrict to each smooth stratum
in turn.

6.3. Second example: constant coefficient homogeneous

systems
Let i : Σ → J 1 (V, W )  V + W + V ∗ ⊗W be defined by a constant coeffi-
cient homogeneous system, as in Chapter 5. Let J 1 (V, W ) have coordinates
(xi , ua , pai ) and let θa = dua − pai dxi , ω i = dxi . As usual, we omit the pull-
back i∗ in the notation. The pullback of the contact system to Σ is still
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.3. Second example: constant coefficient homogeneous systems 139

denoted
I = {θa },
J = {θa , ω i },
and we have
(6.2) dθa ≡ −dpai ∧ ω i mod I.
However, pulled back to Σ, the 1-forms dpai are not all independent. Say
we choose forms π , where dim J + 1 ≤  ≤ dim Σ, such that {θa , ω i , π }
gives a local coframing of Σ. Since Σ is defined by homogeneous equations
Bari pai = 0, we may choose the π so that
dpai = −Aai π
for some constants Aai such that Bari Aai = 0. Thus we may rewrite (6.2) as
(6.3) dθa ≡ Aai π ∧ ω i mod I.
Observe that (J/I)x  V ∗ and Ix  W ∗ . When we work with general
linear Pfaffian systems, these will be the definitions for V and W . The
tableau is recovered by taking
A := {Aai v i ⊗wa ⊆ V ∗ ⊗W | dim J + 1 ≤  ≤ dim Σ}.
We saw in Chapter 5 that if the tableau A is involutive, we have local exis-
tence of integral manifolds, roughly depending on sl functions of l variables,
where l is the Cartan integer and sl is the character of the tableau. For an
arbitrary linear Pfaffian system with no torsion, we will define a tableau at
a general point x ∈ Σ, and if this tableau is involutive, we will again have
local existence of integral manifolds.
Our informal algorithm now looks like:

Input:
linear Pfaffian system
Rename Σ as Σ - (I, J) on Σ;
calculate dI mod I
?
6
6
? N
Q Q
Q
N
 Q Y  Q
 Q  Q  Q
  Q  Is [T ] = 0? Q - Is tableau Q
 Is Σ empty?Q Q  Q involutive? 
Q  Q  Q 
Q k
Q QQ  Q 
QQ Q N Q
Y Q Y
Q ?
? Q Restrict to Σ ⊂ Σ ?
Q
Done: defined by [T ] = 0 Done:
there are no and Ω |Σ = 0 local existence of
integral manifolds integral manifolds
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 140 6. The Cartan Algorithm for Linear Pfaffian Systems

Remark 6.3.1. We can consider the more general case of a linear constant-
coefficient but not necessarily homogeneous system. Say there are constants
Cir such that the equations are of the form
Bari pai = Cir xi .
Then we must have
dpai = −Aai π + Tija ω i ,
where the Aai are as before, and Tija are constants which satisfy Cir = Barj Tija .
Now we have apparent torsion:
dθa ≡ Aai π ∧ ω i + Tija ω i ∧ ω j mod I.
As seen in Exercise 6.2.1, it might be possible to modify the π , by adding
multiples of the ω j ’s, so that the Tija all zero, i.e., so that the apparent torsion
is absorbed. We examine the absorption of apparent torsion systematically
in §6.5.

What if A is not involutive? If A is not involutive, we need to examine

third-order information. The test will be to compare a naı̈ve estimate for
the dimension of A(2) with its actual dimension. (Recall from Chapter 5
that A(2) admits the interpretation as the space of admissible third-order
terms in the Taylor series of a solution.)
We start with a new system, which is the pullback of the tautological
linear Pfaffian system on
J 2 (V, W )  V + W + V ∗ ⊗W + W ⊗S 2 V ∗
to Σ̃ := V + W + A + A(1) . This new system is called the prolongation of
(I, J) on Σ.
Recall that, in coordinates (xi , ua , pai , paij ), the contact system on
J 2 (V, W ) is generated by

θa = dua − pai dxi ,

θia = dpai − paij dxj .
We abuse notation by writing θa , θia for i∗ (θa ), i∗ (θia ), where
i : Σ̃ → V + W + V ∗ ⊗W + S 2 V ∗ ⊗W
is the inclusion. In particular, note that on Σ̃, only dim A of the forms θia
are independent.
Writing our new system as I = {θa , θia }, we observe that dθa ≡ 0 mod I,
so our (s + dim A) × n tableau, when represented by a matrix, will always
have the first s × n block zero. We now apply Cartan’s Test for involutivity
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.4. The local isometric embedding problem 141

to this tableau. If it is involutive, then local integral manifolds exist; if not,

then further prolongation is necessary.
Here is our full informal algorithm:

Input:
“Prolong”, i.e., start over
linear Pfaffian system
-  on a larger space Σ̃;
Rename Σ as Σ (I, J) on Σ;
rename Σ̃ as Σ
calculate dI mod I
and new system as (I, J)
6
6
? N
Q Q
Q
N
 Q Y  Q
 Q  Q  Q
  Q  Is [T ] = 0? Q - Is tableau QQ
 Is Σ empty?Q Q  Q involutive? 
Q  Q  Q 
Q  QQ  Q 
QQQ kQ N Q
Y Q Y
Q ?
? Q Restrict to Σ ⊂ Σ ?
Q
Done: defined by [T ] = 0 Done:
there are no and Ω |Σ = 0 local existence of
integral manifolds integral manifolds

Before formalizing this process, we illustrate it with a more substantial

example:

6.4. The local isometric embedding problem

Let (M, g) and (N, h) be Riemannian manifolds. An embedding f : M → N
is called isometric if f ∗ (h) = g.

Problem 6.4.1 (Analytic isometric embedding). Let (M n , g) be an analytic

Riemannian manifold and let r be a positive integer. What (if any) are the
local isometric embeddings M n → En+r ?

We will say that a manifold admits a local isometric embedding into Eucli-
dean space if an isometric embedding exists on some neighborhood of any
given point.
The above problem is stated vaguely—we will study it in the following
form:
Determine a function r(n) with values in Z+ ∪ ∞ such
that every analytic Riemannian manifold of dimension n
admits a local isometric embedding into En+r .
While the answer to this question is known—in fact we will find r(n) in
§6.10—the corresponding question in the C ∞ category is still open.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 142 6. The Cartan Algorithm for Linear Pfaffian Systems


In coordinates, choosing a metric corresponds to choosing n+1
2 func-
tions of n variables subject to some open (nondegeneracy) conditions, and
choosing a map Rn → Rn+r is the choice of n + r functions of n variables.
So the determined case (the
 case where there are as many equations as un-
n
knowns) is when r = 2 . We will specialize to this case when we discuss
the Cartan-Janet Theorem in §6.10.
As usual, instead of working on M × En+r , we will work on a larger
space that takes into account the group actions. Let FM and FEn+r be
the respective orthonormal frame bundles. Fix index ranges 1 ≤ i, j ≤ n,
n + 1 ≤ μ, ν ≤ n + r. We use η to denote forms on FM and ω to denote
forms on FEn+r . Recall from §2.6 the structure equations on FM :

dη j = −ηkj ∧ η k ,
dηji = −ηli ∧ ηjl + 12 Rjkl
i
ηk ∧ ηl ,
i
where the functions Rjkl : FM → R are the coefficients of the Riemann
curvature tensor. We set up an EDS whose integral manifolds are graphs
Γ ⊂ FEn+r ×FM of isometric embeddings M n → En+r . As usual, we commit
the standard abuse of notation, omitting pullbacks.
To simplify our calculations, we restrict to graphs Γ on which ω μ = 0
and ω i = η i , so Tx M is spanned by e1 , . . . , en , and the first n basis vectors
of Tx M and Tx En+r are aligned. On FM × FEn+r , define the Pfaffian system
I = {ω μ , ω j − η j }diff
with independence condition Ω = η 1 ∧ . . . ∧ η n , so that
I = {ω μ , ω j − η j },
J = {ω μ , ω j , η j }.

We calculate
dω μ = −ωjμ ∧ ω j − ωνμ ∧ ω ν ,
d(ω i − η i ) = −ωji ∧ ω j − ηji ∧ η j − ωνi ∧ ω ν .
Thus,
)
dω μ ≡ 0
mod J,
d(ω i − η i ) ≡ 0
and our system is a linear Pfaffian system, so we can begin the algorithm.
We calculate
)
dω μ ≡ −ωjμ ∧ η j
(6.4) mod I.
d(ω i − η i ) ≡ −(ωji − ηji ) ∧ η j
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.4. The local isometric embedding problem 143

There is no torsion, since the forms ωjμ , (ωji − ηji ) are linearly independent.
At a point x ∈ Σ, let V ∗ = (J/I)x , W ∗ = Ix . We can use the structure
equations (6.4), just as we used (6.3), to define the tableau
 μ
aj
A= ,
aij

where the entries are subject only to the relations aij + aji = 0. Here the π ’s
are the ωjμ and the (ωji − ηji ) for i < j. The Cartan characters (see §3.1) are
given by sj = r + (n − j).
Write an element of A(1) as (aijk wi + aμjk wμ )⊗v j v k . We see that aijk =
−ajik = aikj , so aijk = 0 for all i, j, k. The only relations on the aμjk are

aμjk = aμkj . Thus dim A(1) = r n+1
2 < s1 + 2s2 + . . . + nsn , and the tableau
is not involutive.
We continue our algorithm, starting over on a larger space where the
second derivatives are included as independent variables. In this case, we
will call these new variables hμij . (They will turn out to be the coefficients
of the second fundamental form.) We start over on the manifold
n+1
Σ̃ = FM × FEn+r × Rr( 2 ).
˜ J)
The system (I, ˜ defined by

I˜ = {ω μ , ω j − η j , ωji − ηji , ωjμ − hμjk η k },

(6.5)
J˜ = {ω μ , ω j , η j , ωji − ηji , ωjμ }
is the prolongation of (I, J).
As with the prolongation in the previous example, we have
)
dω μ ≡ 0
˜
mod I.
d(ω i − η i ) ≡ 0
We continue our computations:

d(ωji − ηji ) ≡ − 12 (Rjkl
i
− (hμik hμjl − hμil hμjk ))η k ∧ η l mod I.
˜
μ

Here we are presented with the same situation as in §6.2: there is torsion,
and so there are no integral elements except at points of Σ̃ where
 μ μ
(6.6) i
Rjkl − (hik hjl − hμil hμjk ) = 0.
μ

Thus we start over again on the submanifold Σ ⊂ Σ̃ defined by (6.6).

Since (6.6) is nothing but the Gauss equations (see Chapter 2), it is clear in
hindsight that we should have begun working on the submanifold where (6.6)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 144 6. The Cartan Algorithm for Linear Pfaffian Systems

holds. Assume for the moment that the pullback to Σ of our independence
condition is nonvanishing, so that we may go on to calculate our new tableau:
(6.7) d(ω μ − hμ ω k ) ≡ (−dhμ − hν ω μ + hμ ω l + hμ ω l ) ∧ ω k mod I.
j jk jk
˜
jk ν jl k kl j
μ
To simplify notation, let = − + πjk −dhμjk
+ hμkl ωjl . On Σ̃, hνjk ωνμ hμjl ωkl
μ
the forms πjk are all independent, but when we restrict to Σ , they satisfy
relations obtained by differentiating (6.6).
The next step is to check the torsion. First, we restrict to a simple,
special case. Other cases will be treated below and in §6.10.
Case n = 2, r = 1. The only independent curvature component is R1212 =
R, and equation (6.6) becomes
(6.8) R = h11 h22 − (h12 )2 ,
where we have suppressed the μ = 3 index on h.
Let θj = ωj3 − hjk ω k . Then (6.7) specializes to
     1
θ1 −dh11 + 2h12 η12 −dh12 − (h11 − h22 )η12 η ˜
d ≡ ∧ 2 mod I.
θ2 −dh12 − (h11 − h22 )η12 −dh22 − 2h12 η12 η
Since R : FM → R is constant on the fibers of the projection to M , we
may write dR = R1 η 1 + R2 η 2 . Differentiating (6.8) gives
(6.9) R1 η 1 + R2 η 2 − h11 dh22 − h22 dh11 + 2h12 dh12 = 0.
Upon setting
π1 = −dh11 + 2h12 η12 ,
π2 = −dh12 − (h11 − h22 )η12 ,
π3 = −dh22 − 2h12 η12 ,
the relation (6.9) becomes
(6.10) R1 η 1 + R2 η 2 + h11 π3 + h22 π1 − 2h12 π2 = 0.
Aside 6.4.2. The seeming miracle that the η12 terms all cancel is due to
the fact that the form η12 is dual to a Cauchy characteristic vector field (see
Chapter 7). Intuitively, integral curves of this vector field correspond to
spinning the frame within the tangent space of the surface, which doesn’t
change the embedding.
In fact, Theorem 7.1.21 implies that our EDS actually lives on the quo-
tient of Σ by this foliation. Since integral manifolds of the EDS exist if and
only if torsion vanishes on the quotient manifold, we expect that the torsion
upstairs will not involve η12 .
In n dimensions, the forms ηji are dual to Cauchy characteristics for
the isometric embedding system. Rather than ignoring them, we could have
incorporated them into the independence condition. Then integral manifolds
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.4. The local isometric embedding problem 145

would correspond to graphs of embeddings of FM in FEn+r which are lifts

of isometric embeddings M → En+r .

Of the forms π1 , π2 , π3 , two are independent modulo J/I. We solve for

π3 using (6.10) and plug in:
     1  
θ1 π1 π2 η 0
(6.11) d ≡ ∧ 2 − R1 1 2 .
θ2 π2 h111 (h22 π1 − 2h12 π2 ) η h11 η ∧ η

Here we have isolated the apparent torsion term.

As mentioned in Remark 6.3.1, it might be possible to eliminate the
apparent torsion by choosing a different complement to J ⊂ T ∗ Σ. Here, if
we let
R1 1
π̃1 = π1 − η ,
h22
then π̃1 is still independent of η 1 , η 2 , and the equations become
     1
θ1 π̃1 π2 η
(6.12) d ≡ ∧ 2 .
θ2 π2 h11 (h22 π̃1 − 2h12 π2 )
1
η
We have absorbed the apparent torsion by changing our choice of coframing.
We can now perform Cartan’s Test on the tableau A(1) associated to
(6.11). Here s1 = 2, s2 = 0, and s1 + 2s2 = 2 = dim A(2) , so the tableau is
involutive according to Definition 5.5.4. Invoking the Cartan-Kähler Theo-
rem 6.5.6 below, we obtain
Theorem 6.4.3. Let (M 2 , g) be an analytic Riemannian manifold of di-
mension two. Then there exist local isometric embeddings M 2 → E3 , and
such embeddings depend on two functions of one variable.
Exercise 6.4.4: Verify directly that dim A(2) = 2.

Aside 6.4.5 (Another easy case). Suppose that the codimension r is small

enough that n (n12−1) > n+1
2 2
2 r. To see if (6.6) is solvable, let V = R ,
n

W = (R ,  , ), and define

r

γ : S 2 V ∗ ⊗W → K(V ),

haij v i v j ⊗wa → (haik hajl − hail hajk )v i ⊗v j ⊗v k ⊗v l .
a

Here, K(V ) = S22 (V ) is the kernel of the skew-symmetrization map δ :

Λ2 V ⊗ Λ2 V → Λ4 V (see Appendix A), and dim K(V ) = n (n12−1) . Our
2 2

assumption implies that γ cannot be surjective. So, only curvature tensors

taking value in a subvariety of K(V ) can satisfy (6.6). Thus, there are no
isometric embeddings of a ‘generic’ metric in this range of codimensions.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 146 6. The Cartan Algorithm for Linear Pfaffian Systems

6.5. The Cartan algorithm formalized:

tableau, torsion and prolongation
Let I = {θa }, J = {θa , ω i } be a linear Pfaffian system on a manifold Σ.
Choose a complement {π } to J, where 1 ≤  ≤ dim Σ − dim J, to obtain a
local coframing of Σ. By Exercise 6.1.2 there locally exist functions Aai , Tija
such that the derivatives of forms in I may be written
dθa ≡ Aai π ∧ ω i + Tija ω i ∧ ω j mod I.
The functions Aai , Tija depend on our choices of bases for I and J. We now
construct invariants from them.

Tableaux. Fix a generic point x ∈ Σ. Let V ∗ = (J/I)x and W ∗ = Ix .

Write wa = θxa , v j = ωxj , and define the tableau of (I, J) at x by
Ax := {Aai wa ⊗v i |1 ≤  ≤ dim Σ − dim Jx } ⊆ W ⊗V ∗ .
Exercise 6.5.1: Verify that Ax is independent of any choices.
Remark 6.5.2. Let B = {Bari vi ⊗wa } ⊂ V ⊗W ∗ denote the annihilator of
A. If dθa ≡ πia ∧ ω i mod I, then
Bari πia ≡ Cjr ω j mod I ∀1 ≤ r ≤ codim A.
These equations are sometimes called the symbol relations of the tableau.
Example 6.5.3 (Second-order PDE). Recall how a second-order PDE for
one function of two variables is expressed as an EDS. We use the classical
coordinates (x, y, z, p, q, r, s, t) on J 2 (R2 , R) and contact forms
θ0 = dz − p dx − q dy,
θ1 = dp − r dx − s dy,
θ2 = dq − s dx − t dy.
(In particular, p, q, r, s, t correspond respectively to the jet coordinates p1 , p2 ,
p11 , p12 , p22 defined in §1.10.) Suppose the PDE takes the form
(6.13) F (x, y, z, p, q, r, s, t) = 0.
Assume that {F = 0} defines a smooth submanifold Σ7 ⊂ J 2 (R2 , R), and
that at least one of Fr , Fs , Ft is nonzero at each point of Σ, and say that
F is regular when these assumptions hold. Regularity ensures that θ0 , θ1 , θ2
restrict to be linearly independent on Σ, and that the projection π : Σ →
J 1 (R2 , R) is a smooth submersion at each point.
Let I be the Pfaffian system spanned by the restrictions of θ0 , θ1 , θ2 to
Σ. Since dθ0 ≡ 0 modulo I, the ideal I = {I}diff is generated algebraically
by I and the 2-forms
dθ1 = −(dr ∧ dx + ds ∧ dy), dθ2 = −(ds ∧ dx + dt ∧ dy).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.5. Tableau, torsion and prolongation 147

Near any point in Σ, one can obtain 1-forms π1 , π2 , π3 such that

)
dθ1 ≡ −(π1 ∧ dx + π2 ∧ dy)
(6.14) mod I
dθ2 ≡ −(π2 ∧ dx + π3 ∧ dy)
and such that π1 , π2 , π3 satisfy the symbol relation
(6.15) Fr π1 + Fs π2 + Ft π3 ≡ 0 mod I.
For example, if Fr = 0, we may choose
 
Fr D x F − Fs D y F Dy F
π1 = dr + 2
dx + dy,
Fr Fr
Dy F
π2 = ds + dx,
Fr
π3 = dt,
where Dx F = Fx + pFz + rFp + sFq and Dy F = Fy + qFz + sFp + tFq .

Let u ∈ Σ and let E ∈ V2 (I, Ω) be an integral 2-plane at u. If we use

the bases {θ0 , θ1 , θ2 } for W ∗ = Iu and {dx, dy} for E ∗ ∼
= V ∗ = (J/I)u , then
(6.15) shows that the tableau takes the form
⎧⎛ ⎞ ⎫
⎨ 0 0  ⎬
(6.16) Au = ⎝a b ⎠ aFr (u) + bFs (u) + cFt (u) = 0 ⊂ W ⊗ V ∗ .
⎩ ⎭
b c 

Torsion. Recall that the terms Tija ω i ∧ ω j in (6.1) are called apparent tor-
sion. If it is possible to choose new π̃ ’s such that T̃ija = 0, we say the
apparent torsion is absorbable. If it is not possible, we will say there is tor-
sion. In other words, a particular choice of complement to Jx ⊂ Tx∗ Σ may
yield apparent torsion at x, while we say there is torsion at x if all choices
yield apparent torsion. In that case there are no integral elements at x. We
now discuss these statements without reference to bases.
The change in the Tija that a change of complement π̃ = π +ei ω i effects
is
Tija = Tija + 12 (Aaj ei − Aai ej ).
a = Aa e − Aa e and let
Let Eij j i i j

δ : W ⊗V ∗ ⊗V ∗ → W ⊗Λ2 V ∗
be the skew-symmetrization map. The space of possible modifications to
apparent torsion is δ(A⊗V ∗ ), because E := Eij
a w ⊗v i ⊗v j lies in δ(A⊗V ∗ ).
a
Then the torsion of (I, J) at x is defined to be
[T ] := [Tija wa ⊗v i ∧ v j ] ∈ W ⊗Λ2 V ∗ /δ(Ax ⊗V ∗ ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 148 6. The Cartan Algorithm for Linear Pfaffian Systems

We will use the notation H 0,2 (A) = W ⊗Λ2 V ∗ /δ(Ax ⊗V ∗ ). (This is one of
the Spencer cohomology groups, which will be discussed in §6.7.)
Exercise 6.5.4: Compute the Cjr of Remark 6.5.2 in terms of the coefficients
Aai , Tija .

If [T ] ≡ 0, then any apparent torsion terms we have for a specific choice of

π ’s are absorbable and we could choose new forms π̃ to get all the apparent
torsion to disappear. For the purposes of Cartan’s Test, this choice is not
necessary because the tableau is independent of the choice of π by Exercise
6.5.1. If [T (x)] is not identically zero, we need to restrict our system to
Σ ⊂ Σ on which [T ] ≡ 0. On this subset we may obtain new symbol
relations resulting from d[T ] = 0.
It is possible that the equations d[T ] = 0 force a relation among the
ω i on an open subset of Σ . If this happens, we must restrict further to
the subset Σ ⊂ Σ where the ω i are independent. If Σ is empty, there
are no integral manifolds and we are done. In any event, restricting to Σ
and Σ will in general introduce new symbol relations, which in turn could
introduce more apparent torsion, and so we must repeat the above process.
Exercises 6.5.5:
1. Say we have an isomorphism V → W such that A  Λ2 V . Show that
H 0,2 (A) = 0. Compare this to the fundamental lemma of Riemannian ge-
ometry in Lemma 3.1.4. What about if A  S 2 V ?
2. Let (I, J) be a linear Pfaffian system on Σ. Let x ∈ Σ and suppose
dim(J/I)x = dim Ix = 2. Show that if s1 (Ax ) = 2, then [T ]x = 0.

The Cartan-Kähler Theorem for linear Pfaffian systems.

Theorem 6.5.6. Let (I, J) be an analytic linear Pfaffian system on a man-
ifold Σ, let x ∈ Σ be a point and let U be a neighborhood containing x, such
that for all y ∈ U ,
1. [T ]y = 0, and
2. the tableau Ay is involutive.
Then solving a series of Cauchy problems yields analytic integral man-
ifolds to (I, J) passing through x that depend (in the sense discussed in
Chapter 5) on sl functions of l variables, where sl is the character of the
system.
We interpret the Cartan-Kähler Theorem as saying: if there is no ob-
struction to having an integral element at a general point x ∈ Σ, then the
space of integral manifolds passing through x is the same size as the space of
integral manifolds for the corresponding linearized problem. In other words,
if the linear algebra at the infinitesimal level works out right, everything
works out right.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.6. Summary of Cartan’s algorithm for linear Pfaffian systems 149

When a linear Pfaffian system (I, J) satisfies the conditions of Theorem

6.5.6 at x, we say it is involutive at x. (Note that because x ∈ Σ is a
general point, we only need to check conditions 1 and 2 at x, and it will be
guaranteed that they are satisfied at nearby points.)
We will give a complete proof of a more general version of the Cartan-
Kähler Theorem in Chapter 8. The present version could be proved by
choosing local coordinates and doing estimates comparing a formal solution
to the system with its linearization.

Finally, we formalize the process of prolongation:

Prolongation. Let I be an arbitrary EDS defined on a manifold Σ. Con-

sider π : G(n, T Σ) → Σ, the Grassmann bundle of all n-planes in all tangent
spaces to Σ. We denote points of G(n, T Σ) by (p, E), where p ∈ Σ and
E ⊂ Tp Σ is an n-dimensional subspace. Much like the contact system on
the space of k-jets, G(n, T Σ) carries a canonical linear Pfaffian system on it,
whose integral manifolds are exactly the lifts X̃ of immersed submanifolds
X n ⊂ Σ to G(n, T Σ) defined by x̃ = (x, Tx X). The system on G(n, T Σ) is
defined by
I(p,E) := π ∗ (E ⊥ ), J(p,E) := π ∗ (Tp∗ Σ).

Now let Vn (I) ⊂ G(n, T Σ) be the subbundle (more precisely, subsheaf)

of n-dimensional integral elements to I, whose fiber over a point x ∈ Σ is
Vn (I)x as defined in §1.10. That is,
Vn (I) := {(x, E) ∈ G(n, T Σ) | Ix |E = 0}.

The prolongation of I is the restriction of the canonical system on

G(n, T Σ) to Vn (I). (We should really call this the prolongation of I for
n-dimensional integral manifolds.) If I comes equipped with an indepen-
dence condition Ω, we recall that
Vn (I, Ω) := {(x, E) ∈ G(n, T Σ) | Ωx (E) = 0 and Ix |E = 0}.
In this case we define the prolongation to be the restriction of the canonical
system to Vn (I, Ω).

6.6. Summary of Cartan’s algorithm for linear Pfaffian

systems
Remark 6.6.1. The Cartan algorithm will not necessarily yield all integral
manifolds of the original system, only the integral manifolds arising from
well-posed Cauchy problems at general points.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 150 6. The Cartan Algorithm for Linear Pfaffian Systems

Input: Prolong to a component

linear Pfaffian system of Vn (I, Ω) ⊂ Gn (T Σ);
Rename Σ as Σ - (I, J) on Σ;  rename component as Σ,
calculate dI mod I canonical system
becomes (I, J)
6
6
? N
Q Q
Q
N
 Q Y  Q
 Q  Q Is tableauQ
  Q  Is [T ] = 0? Q -  QQ
 Is Σ empty?Q Q  Q involutive? 
Q  Q  Q 
Q  QQ  Q 
QQQ kQ N Q
Y Q Y
Q ?
? Q Restrict to Σ ⊂ Σ ?
Q
Done: defined by [T ] = 0 Done:
there are no and Ω |Σ = 0 local existence of
integral manifolds integral manifolds

Figure 1. Our final flowchart.

Remark 6.6.2. Each time one prolongs, there may be many different com-
ponents of Vn (I) on which to restrict. To find all possible integral manifolds,
one must carry out the algorithm on each component. The Cartan-Kuranishi
Prolongation Theorem [27] says in effect that this process terminates even-
tually, but gives no hint of how long it will take.

Warning: Where the algorithm ends up (i.e., in which “Done” box in

the flowchart) may depend on the component one is working on.

Summary. Let (I, J) be a linear Pfaffian system on Σ. We summarize

Cartan’s algorithm:
(1) Take a local coframing of Σ adapted to the filtration I ⊂ J ⊂ T ∗ Σ.
Let x ∈ Σ be a general point. Let V ∗ = (J/I)x , W ∗ = Ix , and
let v i = ωxi , wa = θxa and vi , wa be the corresponding dual basis
vectors.
(2) Calculate dθa ; since the system is linear, these are of the form
dθa ≡ Aai π ∧ ω i + Tija ω i ∧ ω j mod I.
Define the tableau at x by
A = Ax := {Aai v i ⊗wa ⊆ V ∗ ⊗W | 1 ≤  ≤ r} ⊆ W ⊗V ∗ .
Let δ denote the natural skew-symmetrization map δ : W ⊗ V ∗ ⊗
V ∗ → W ⊗Λ2 V ∗ and let
H 0,2 (A) := W ⊗Λ2 V ∗ /δ(A⊗V ∗ ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.7. Additional remarks on the theory 151

The torsion of (I, J) at x is

[T ]x := [Tija wa ⊗v i ∧ v j ] ∈ H 0,2 (A).
(3) If [T ]x = 0, then start again on Σ ⊂ Σ defined by the equations
[T ] = 0, with the additional requirement that J/I has rank n over
Σ . Since the additional requirement is a transversality condition,
it will be generically satisfied as long as dim Σ ≥ n. In practice
one works infinitesimally, using the equations d[T ] = 0, and checks
what relations d[T ] ≡ 0 mod I imposes on the forms π used before.
(4) Assume [T ]x = 0. Let Ak := A ∩ (span{v k+1 , . . . , v n } ⊗ W ). Let
A(1) := (A⊗V ∗ ) ∩ (W ⊗S 2 V ∗ ), the prolongation of the tableau A.
Then
dim A(1) ≤ dim A + dim A1 + . . . + dim An−1
and A is involutive if equality holds.
Warning: One can fail to obtain equality even when the system
is involutive if the bases were not chosen sufficiently generically.
In practice, one does the calculation with a natural, but perhaps
nongeneric, basis and takes linear combinations of the columns of
A to obtain genericity. If the bases are generic, then equality holds
if and only if A is involutive.
When doing calculations, it is convenient to define the char-
acters sk by s1 + . . . + sk = dim A − dim Ak , in which case the
inequality becomes dim A(1) ≤ s1 + 2s2 + . . . + nsn . If sp = 0 and
sp+1 = 0, then sp is called the character of the tableau and p the
Cartan integer. If A is involutive, then the Cartan-Kähler Theo-
rem applies, and one has local integral manifolds depending on sp
functions of p variables.
(5) If A is not involutive, prolong, i.e., start over on the pullback of the
canonical system on the Grassmann bundle to the space of integral
elements. In calculations this amounts to adding the elements of
A(1) as independent variables and adding differential forms θia :=
Aai π − paij ω j to the ideal, where paij v i v j ⊗wa ∈ A(1) .

6.7. Additional remarks on the theory

Another interpretation of A(1) . We saw in Chapter 5 that for a constant-
coefficient, first-order, homogeneous system defined by a tableau A, the
prolongation A(1) is the space of admissible second-order terms paij xi xj in a
power series solution of the system. This was because the constants paij had
to satisfy paij wa ⊗v i ⊗v j ∈ A(1) .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 152 6. The Cartan Algorithm for Linear Pfaffian Systems

The following proposition, which is useful for computing A(1) , is the

generalization of this observation to linear Pfaffian systems:
Proposition 6.7.1. Let (I, J) be a linear Pfaffian system without torsion on
manifold Σ. After fixing x ∈ Σ and a particular choice of 1-forms πia mod I
satisfying dθa ≡ πia ∧ ω i mod I, A(1) may be identified with the space of
1-forms π̃ia mod I satisfying dθa ≡ π̃ia ∧ ω i mod I, as follows: any such set
(π̃ia ) may be written as π̃ia ≡ πia + paij ω j mod I, with paij wa ⊗v i v j ∈ A(1) . In
particular, A(1) is an affine space.
For example, in Chapter 5, the initial choice was πia ≡ 0 and the other
choices were paij dxj .

Proof. Let π̃ia ≡ πia +paij ω j mod I. The π̃ia must satisfy the symbol relations
Bari π̃ia ≡ 0 mod I. But Bari π̃ia ≡ Bari (πia + paij ω j ) ≡ Bari (paij ω j ) mod I, since
πia also satisfies the symbol relations. This implies Bari paij = 0 for all j,
since the ω j are independent, so paij wa ⊗v i ⊗v j ∈ A⊗V ∗ . But we also need
π̃ia ∧ ω i ≡ πia ∧ ω j + paij ω i ∧ ω j ≡ 0 mod I. We already have πia ∧ ω i ≡ 0 mod I,
so this forces paij ω i ∧ ω j = 0, which implies paij = paji . 

(1)
Remark 6.7.2. If there is no torsion, Ax is the space of integral elements
at x. After fixing an integral element, it becomes a linear space. This is the
reason for the adjective “linear” in the name linear Pfaffian systems.

Invariants of tableaux. Here we define a collection of invariants of a

tableau. We will only use the invariants H j,2 , but we might as well define
them all.

Definition 6.7.3. The Spencer cohomology groups H i,j (A) of a tableau

A ⊂ W ⊗V ∗ are defined as follows: Let

δj : (W ⊗S d V ∗ )⊗Λj V ∗ → W ⊗S d−1 V ∗ ⊗Λj+1 V ∗

be defined by
δj (f ⊗ξ) = df ∧ ξ,

where, for f ∈ W ⊗S d V ∗ , ξ ∈ Λj V ∗ , we define df by considering f as a W -

valued function on V , and extend δj by linearity. Note that δj (A(i)⊗Λj V ∗ ) ⊆
A(i−1) ⊗Λj+1 V ∗ .
Define
ker δj (A(i−1) ⊗Λj V ∗ )
H i,j (A) := .
Image δj−1 (A(i) ⊗Λj−1 V ∗ )
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.7. Additional remarks on the theory 153

Let W̃ = A ⊕ W . Then A(1) can be considered as a tableau in W̃ ⊗V ∗

with the W ⊗V ∗ block zero. Consider:
W̃ ⊗Λ2 V ∗
H 0,2 (A(1) ) =
δ(A(1) ⊗V ∗ )
W ⊗V ∗ ⊗Λ2 V ∗
=
δ(((A⊗V ∗ ) ∩ (W ⊗S 2 V ∗ ))⊗V ∗ )
= H 1,2 (A).
Exercise 6.7.4: Show that
H 0,2 (A(p) ) = H p,2 (A).

The Kuranishi Prolongation Theorem (see [27]) implies that after a finite
number of prolongations a tableau will become involutive (including the
possibility that it becomes empty). This implies
Theorem 6.7.5 (Goldschmidt version of Cartan-Kähler [74]). Let (I, J)
be a linear Pfaffian system on an analytic manifold Σ and let x ∈ Σ be a
general point. If H p,2 (Ax ) = 0 for all p, then there exist integral manifolds
passing through x.
This theorem is useful because sometimes one can show that the groups
H p,2 (Ax ) are zero without actually calculating the prolongations (see [70]).
More generally, it is sufficient to show that the corresponding vector
bundles with induced torsion sections are such that the sections are zero;
see Chapter 9 for the construction of the associated vector bundles.

Characteristic hyperplanes. Let (I, J) be a linear Pfaffian system on

a manifold Σ with no torsion, and let E ∈ Vn (I, J)x . Let H ⊂ E be a
hyperplane. We address the following question:
Under what circumstances is E the only integral n-plane that contains H?
Definition 6.7.6. A hyperplane H ⊂ E ⊂ Tx Σ is said to be characteristic
if it has more than one extension to an n-dimensional integral element.

With notation as above, we may choose the coframe so that

θa |E = πia |E = 0.
Assume moreover that E is uniquely defined by these equations. (This
amounts to assuming that {θa , ω i , πia } span the cotangent space to Σ, i.e.,
that there are no Cauchy characteristics.)
We can also choose frames such that ω n |H = 0. Let e1 , . . . , en be a basis
for E dual to the ω i . Then v ∈ Tx M , v ∈
/ H, completes H to an integral
n-plane if and only if
(6.17) (v, e1 , . . . , en−1 ) Φ = 0
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 154 6. The Cartan Algorithm for Linear Pfaffian Systems

for all Φ ∈ I n . In fact, it is sufficient to require that v θi = 0 and to verify

(6.17) for Φ = πia ∧ ω J , where J is any multi-index of length n − 1 containing
the index i. Since ei π = 0, (6.17) becomes
(6.18) v παa = 0, ∀a, α, 1 ≤ α < n.

Now recall the symbol mapping from Definition 5.6.2, defining the char-
acteristic variety ΞA for A = Ax . If ξH ∈ PE ∗ corresponds to H ⊂ E, then
ξH ∈ ΞA if and only if
Barn wa = 0
for some nonzero vector w ∈ W . In other words, ξH ∈ ΞA if and only if
the tableau contains a matrix with the first n − 1 columns zero and the
last column nonzero, or equivalently, some linear combination of the πna is
linearly independent from the forms in the first n − 1 columns, so that v is
not uniquely determined by (6.18). We have now proved
Theorem 6.7.7. A hyperplane H ⊂ E ⊂ Tx Σ is characteristic if and only
if the corresponding element ξH ∈ PE ∗ is such that ξH ∈ ΞA .
Example (6.5.3 continued). Let Σ ⊂ J 2 (R2 , R) be the submanifold defined
by a regular second-order PDE and let A be the tableau (6.16) at a point
u ∈ Σ. Then a nonzero covector ξ = ξ1 dx + ξ2 dy ∈ V ∗ belongs to ΞA if and
only if the map σξ (w) := w ⊗ ξ mod A has a nonzero kernel. Write
⎛ ⎞
w1 ξ1 w1 ξ2
w ⊗ ξ = ⎝w2 ξ1 w2 ξ2 ⎠ .
w3 ξ1 w3 ξ2
Then w ⊗ ξ ∈ A if and only if w1 = 0, w2 ξ2 = w3 ξ1 and Fr w2 ξ1 + Fs w2 ξ2 +
Ft w3 ξ2 = 0. This homogeneous system of linear equations for the wa has a
nontrivial solution if and only if
(6.19) Fr ξ12 + Fs ξ1 ξ2 + Ft ξ22 = 0.

A linear subspace of an integral plane E is said to be characteristic if it is

annihilated by some covector ξ ∈ ΞA . Thus, the characteristic hyperplanes
in E are null lines for the quadratic form defined by the symmetric matrix
 
Fr 21 Fs
1
2 Fs Ft
with respect to the basis dual to dx, dy.

6.8. Examples
The examples with Roman numerals are adapted from Cartan’s 1945 treatise
[40].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.8. Examples 155

Example 6.8.1 (The tautological system for surfaces in E3 , defined on

ASO(3) × R3 hij ). The integral manifolds of this system are the first-order
adapted lifts of surfaces in R3 (see §2.1), together with the components of
the second fundamental form.
Let I = {ω 3 , θj3 := ωj3 − hjk ω k }. Then
⎛ 3⎞ ⎛ ⎞
ω 0 0  1
⎝ 3⎠ ⎝ 2 − dh 2 − dh ⎠ ω
d θ1 ≡ 2h ω
21 1 11 (h 22 − h )ω
11 1 12 ∧ 2 mod I
ω
θ23 (h22 − h11 )ω12 − dh12 −2h21 ω12 − dh22
⎛ 0 ⎞
π1 π20  1
⎝ ⎠ ω
≡ π1 π2 ∧
1 1
2 mod I,
ω
π12 π22
where the πia ’s stand for the components of the matrix on the first line. The
symbol relations are
π10 ≡ π20 ≡ π12 − π21 ≡ 0 mod I.
There is no torsion, and s1 = 2, s2 = 1. Integral elements are given by
π10 = π20 = 0,
π11 = aω 1 + bω 2 ,
π21 = π12 = bω 1 + cω 2 ,
π22 = cω 1 + eω 2 ,

where a, b, c, e are arbitrary. Thus dim A(1) = 4 = s1 + 2s2 , and the system
is involutive. The Cartan-Kähler Theorem implies that integral manifolds
depend on one function of two variables. This should come as no surprise,
since locally any surface can be written as a graph z = f (x, y).
Example 6.8.2 (Cartan III: Linear Weingarten surfaces). Let A, B, C be
constants, with A = 0, and let AK + 2BH + C = 0 be a linear relation
between functions H and K. We set up and describe the space of general
integral manifolds for first-order adapted lifts of surfaces in R3 , together
with the components of the second fundamental form, having Gauss and
mean curvatures satisfying this relation. (Note that this class of examples
includes constant mean curvature and constant Gauss curvature surfaces.)
This is the system above restricted to a codimension one submanifold
Σ ⊂ ASO(3) × R3 hij . The submanifold is defined by the equation
A(h11 h22 − h212 ) + B(h11 + h22 ) + C = 0.
The differential of this equation implies that on Σ we may write
dh12 = αdh11 + βdh22
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 156 6. The Cartan Algorithm for Linear Pfaffian Systems

for some functions α, β, and thus

π21 = απ11 + βπ22 .
We still have no torsion and now have characters s1 = 2, s2 = 0. By Exercise
5.5.9.1 such a tableau is involutive, and solutions depend locally on two
functions of one variable. This suggests that an appropriate Cauchy problem
for determining such surfaces would be to start out with a space curve (the
two functions being the curvature and torsion of the curve) and then solve
for a unique surface containing the curve. This construction is carried out
explicitly in [40].
Exercise 6.8.3: Set up and perform Cartan’s Test for the systems analogous
to the above two examples, replacing R3 by a space form X 3 () (see §3.4 for
frames for space forms).
Example 6.8.4 (Cartan II). Let (S, g), (S, g) be surfaces with analytic Rie-
mannian metrics. We set up and solve the EDS for maps φ : S → S that
preserve the metric up to scale, i.e., for conformal maps from S to S.
Let F (S) denote the orthonormal coframe bundle of S, with coframing
= −ω21 satisfying dω i = −ωji ∧ ω j . Similarly let F (S) denote the
ω 1 , ω 2 , ω12
orthonormal coframe bundle of S with coframing ω 1 , ω 2 , ω 21 . The metric on
S is given by the pullback of (ω 1 )2 + (ω 2 )2 from any section s : S → F (S),
and that on S by the pullback of (ω 1 )2 + (ω 2 )2 . Thus a conformal map
φ : S → S induces a map Φ : F (S) → F (S) such that Φ∗ ((ω 1 )2 + (ω 2 )2 ) =
λ2 ((ω 1 )2 + (ω 2 )2 ) for some positive function λ.
Without loss of generality, we may require that Φ∗ (ω j ) = λω j for some
function λ, because we have freedom to rotate in the tangent spaces at each
point. So on Σ := F (S) × F (S) × Rλ , λ > 0, let
I = {θ1 := ω 1 − λω 1 , θ2 := ω 2 − λω 2 },
J = {θ1 , θ2 , ω 1 , ω 2 , ω12 } = {θ1 , θ2 , ω 1 , ω 2 , ω 21 }.
We compute
dθ1 = − ω21 ∧ ω 2 − dλ ∧ ω 1 + λω 12 ∧ ω 2 ,
dθ2 = − ω12 ∧ ω 1 − dλ ∧ ω 2 + λω 21 ∧ ω 1 .
Reducing mod I, we obtain
⎛ 1⎞
 1   ω
θ −λdλ
−(ω 1 − ω1 ) 0
2 2
d 2 ≡ ⎝
∧ ω2⎠ .
θ ω 21 − ω12 − dλ 0
λ ω12

λ and ω 1 − ω1 are independent mod J, there is no torsion, and we

Since − dλ 2 2

may write  
a −b 0 
A= a, b ∈ R .
b a 0 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.8. Examples 157

So s1 = 2, s2 = 0. By Definition 5.5.4, the tableau is involutive, and solutions

depend on two functions of one variable.
We may explicitly realize the dependence of conformal maps between
surfaces on two functions of one variable as follows:
Specify parametrized curves f : R → S, g : R → S (our two functions
of one variable) and look for conformal maps φ such that φ(f (t)) = g(t),
where t is the coordinate on R. The claim is that up to constants there is
a unique such map. In fact, choose an orthonormal frame e1 , e2 on S such

that e1 is tangent to f (R), and adapt similarly on S. Set λ = fg (t)
(t) ; then the
map is uniquely determined up to constants.
Conformal maps are exactly the injective holomorphic or antiholomor-
phic maps between surfaces. These are determined by their restriction to
an arc, so picking two parametrized arcs determines a holomorphic or anti-
holomorphic map. Even without knowing the coincidence of Lie groups
GL(1, C)  CO(2), one might have guessed this by observing that the
tableau A is the same as the tableau for the Cauchy-Riemann equations
(see Chapter 5 and Appendix C) augmented with zeros.
Corollary 6.8.5 (Existence of isothermal coordinates). Let S be a surface
with an analytic Riemannian metric g. Let p ∈ S. Then there exist local
coordinates (x, y) centered at p, such that g = λ2 (dx2 + dy 2 ), λ = 0, in some
neighborhood of p. Such coordinates are called isothermal coordinates.

Proof. Take S = R2 with its flat metric. 

Example 6.8.6 (Lagrangian 3-manifolds in C3 ). Let (R6 , ω,  , ) have its

standard inner product and symplectic form

ω = dx1 ∧ dx4 + dx2 ∧ dx5 + dx3 ∧ dx6 .

We may also consider it as C3 with its standard Hermitian structure (see

Appendix C). We will set up and solve an EDS for three-manifolds whose
tangent spaces are isotropic for ω. This will amount to the same system
as just looking for Lagrangian submanifolds of R6 just equipped with the
symplectic form, but studying the problem in this way will help to serve as
a warmup for studying special Lagrangian submanifolds. It is well-known
in symplectic geometry [60] that Lagrangian submanifolds are locally given
by lifts of graphs of functions to the cotangent space, so it is not surprising
that the integral manifolds will depend on one function of three variables.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 158 6. The Cartan Algorithm for Linear Pfaffian Systems

Matrices in the Lie algebra u(3) of U (3) = SO(6) ∩ Sp(R6 , ω) have the
form
⎛ ⎞
0 −α12 −α13 −ρ1 −β12 −β13
⎜α12 0 −α23 −β12 −ρ2 −β23 ⎟
⎜ 3 ⎟
⎜α1 α23 0 −β13 −β23 −ρ3 ⎟
⎜ ⎟.
⎜ ρ1 β 2 β 3 0 α12 α13 ⎟
⎜ 1 1 ⎟
⎝β 2 ρ2 β23 −α12 0 α23 ⎠
1
3
β1 β2 3 ρ3 −α13 −α23 0
Let FU (3) be the unitary frame bundle for C3 , and let I = {ω 4 , ω 5 , ω 6 } be
defined on FU (3) , with independence condition Ω = ω 1 ∧ ω 2 ∧ ω 3 . We have
structure equations
⎛ ⎞
0 0 ... 0
⎜ω 1 0 −α12 −α13 −ρ1 −β12 −β13 ⎟
⎜ 2 ⎟
⎜ω α12 0 −α23 −β12 −ρ2 −β23 ⎟
⎜ 3 ⎟
d(x, e1 , . . . , e6 ) = (x, e1 , . . . , e6 ) ⎜ 3
⎜ω α1 α2
3 0 −β13 −β23 −ρ3 ⎟
⎟.
⎜ω 4 ρ1 β 2 3
β1 0 α12 α1 ⎟
3
⎜ 1 ⎟
⎝ω 5 β 2 ρ2 β2 −α1
3 2 0 α23 ⎠
1
ω 6 β13 β23 ρ3 −α13 −α23 0

The Maurer-Cartan equation (1.19) implies that

⎛ 4⎞ ⎛ ⎞ ⎛ 1⎞
ω ρ1 β12 β13 ω
d ⎝ω 5 ⎠ ≡ − ⎝β12 ρ2 β23 ⎠ ∧ ⎝ω 2 ⎠ mod I.
ω6 β13 β23 ρ3 ω3
We see there is no torsion, and s1 = 3, s2 = 2, s3 = 1. (Note that the tableau
corresponds to the lower left hand block in u(3).)
Exercise 6.8.7: Show that this tableau is involutive.
Proposition 6.8.8. Analytic Lagrangian 3-folds in C3 depend locally on
one function of 3 variables.
Exercise 6.8.9: Do a similar computation for Lagrangian n-folds in Cn for
arbitrary n (see also Example 8.5.6).

Remark 6.8.10 (Another interpretation). The Euclidean inner product  , 

and ω determine an almost complex structure J by ω(v, w) = v, Jw. The
integral manifolds above can also be characterized as the three-manifolds
such that J(Tx M ) ⊥ Tx M .

Example 6.8.11 (Minimal surfaces in E3 with a specified curvature func-

tion). For simplicity, we assume no umbilic points and use Darboux fram-
ings. On FE3 let K(x) be a given function which will be the curvature and
let k(x) = −K(x) be the positive square root.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.8. Examples 159

Let I = {ω 3 , θ13 , θ23 } and let J = {ω 3 , θ13 , θ23 , ω 1 , ω 2 }, where we set θ13 =
ω13 − kω 1 and θ23 = ω23 + kω 2 . The structure equations are
⎛ 3⎞ ⎛ ⎞
ω 0 0  1
ω
d ⎝ θ13 ⎠ ≡ ⎝ −dk −2kω12 ⎠ ∧ 2 .
ω
θ2 3 −2kω12 dk
Write dk = k1 ω 1 + k2 ω 2 . The symbol relations are
⎫
π10 ≡ π20 ≡ 0 ⎪
⎪
⎪
⎪
⎬
π1 − π2 ≡ 0
2 1
mod I.
π11 + π22 ≡ 0 ⎪
⎪
⎪
⎪
⎭
π11 ≡ k1 ω 1 + k2 ω 2
Again we can change bases to attempt to get rid of the apparent torsion. If
we rewrite our equations as
⎛ 3⎞ ⎛ ⎞
ω 0 0  1
ω
d ⎝ θ1 ⎠ ≡ ⎝
3 0 −2kω1 + k2 ω ⎠ ∧
2 1
2 ,
ω
3
θ2 −2kω1 − k1 ω
2 2 0
the symbol relations become
⎫
π10 ≡ π20 ≡ 0 ⎪
⎪
⎬
π12 − π21 ≡ −k1 ω 1 + k2 ω 2 mod I.
⎪
⎪
π11 ≡ π22 ≡ 0 ⎭

This still has apparent torsion, but if we instead write

⎛ 3⎞ ⎛ ⎞
ω 0 0  1
⎝ ⎠ ⎝ ⎠ ω
d θ1 ≡3 0 −2kω1 + k2 ω − k1 ω
2 1 2
∧ ,
ω2
θ23 −2kω12 + k2 ω 1 − k1 ω 2 0
the symbol relations become
π10 ≡ π20 ≡ π11 ≡ π22 ≡ π12 − π21 ≡ 0 mod I.
We see that s1 = 1, s2 = 0, and dim A(1) = 0, so the system is not involutive.
We can short-cut the prolongation process by realizing that, since the
system contains both π21 ∧ω 1 and π21 ∧ω 2 , then π21 must vanish on all integral
manifolds satisfying the independence condition. Thus, we need to add
π := π21 = −2kω12 + k2 ω 1 − k1 ω 2 to the system. Call the new system I + . By
counting dimensions, we see that I + will either be Frobenius at x or have no
integral manifold passing through x. We have dω 3 ≡ dθ13 ≡ dθ23 ≡ 0 mod I + ,
and we compute
dπ = d(−2kω12 + k2 ω 1 − k1 ω 2 )
k12 + k22
≡ (−k1,1 − k2,2 + + 2kK)ω 1 ∧ ω 2 mod I + ,
Licensed to AMS.
2k
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 160 6. The Cartan Algorithm for Linear Pfaffian Systems

where we have written dkj = kj,1 ω 1 + kj,2 ω 2 . The integrability condition we

have obtained can be put more simply:
Exercise 6.8.12: Recall the Laplacian on functions defined in §3.8. Show
that
2 k 2 + k22
Δ log(−K) − 4K = (k1,1 + k2,2 − 1 − 2kK).
k 2k
In summary:
Theorem 6.8.13 (Ricci). Let (M, g) be a surface with a Riemannian met-
ric. Let K(x) denote the curvature function. Away from umbilic points,
necessary and sufficient conditions for M to be locally minimally and iso-
metrically immersed in R3 are that K < 0 and that Δ log(−K) = 4K. If K
is nonconstant, then there is a one-parameter family of such immersions, up
to congruence.
For more results along these lines, see [47].
Note that we did not need the Cartan-Kähler Theorem to prove this
theorem, since it reduced to the Frobenius Theorem. But the integrability
conditions were more easily obtained by following the Cartan algorithm.

Example 6.8.14 (Cartan IV: existence of “curvature-line coordinates”, i.e.,

isothermal coordinates along lines of curvature). Given M 2 ⊂ E3 , we ask if
there exist isothermal coordinates (x, y) on M such that the principal direc-
∂ ∂
tions are ∂x , ∂y . In this case the curves where either x or y is constant are
such that their tangent directions are principal directions. Curves whose tan-
gent directions are principal directions are (somewhat misleadingly) called
lines of curvature.
An EDS whose integral manifolds correspond to surfaces in E3 together
with isothermal coordinates along lines of curvature may be defined on
ASO(3) × R2k1 ,k2 × Ru , with u = 0, by

I = {ω 3 , ω13 − k1 ω 1 , ω23 − k2 ω 2 , d(uω 1 ), d(uω 2 )},

Ω = ω1 ∧ ω2.

Here, the forms ω13 − k1 ω 1 , ω23 − k2 ω 2 are in the ideal to ensure that (ω 1 )⊥ ,
(ω 2 )⊥ correspond to principal directions on the surface, and k1 , k2 will be
the principal curvatures. If we can multiply ω 1 , ω 2 by a common function u
to make them both closed, then the fact that closed forms are locally exact
implies local existence of the isothermal coordinates.
Note that I is not a Pfaffian system. But if we adjoin new variables
u1 , u2 such that du = u1 ω 1 + u2 ω 2 , then the generators of I of degree two
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.9. Functions whose Hessians commute 161

are
d(uω 1 ) = (−u2 ω 1 + uω12 ) ∧ ω 2 ,
d(uω 2 ) = (−u1 ω 2 − uω12 ) ∧ ω 1 .
Thus, an equivalent system is (I, Ω) defined on ASO(3)×R2k1 ,k2 ×R1u ×R2u1 ,u2
with
I := {ω 3 , ω13 − k1 ω 1 , ω23 − k2 ω 2 , du − u1 ω 1 − u2 ω 2 , uω12 − u2 ω 1 + u1 ω 2 },
which is a linear Pfaffian system.
Exercise 6.8.15: Determine the conditions on k1 , k2 such that there exist
integral manifolds of the above system. (The answer will be in the form of a
PDE that must be satisfied in addition to the Codazzi equation.) Determine
the space of integral manifolds for generic functions k1 , k2 that satisfy these
conditions.

Exercises 6.8.16:
1. (Cartan VI) Set up an EDS for pairs of surfaces S, S  ⊂ E3 together
with a local diffeomorphism φ : S → S  that is an isometry preserving a
family of asymptotic directions. What does the general solution depend on?
Geometrically characterize the surfaces S admitting a noncongruent mate
S.
2. (Cartan VII) Set up an EDS for pairs of surfaces S, S  ⊂ E3 together
with a local diffeomorphism φ : S → S  that is an isometry preserving a
family of lines of curvature. What does the general solution depend on?
Geometrically characterize the surfaces S admitting a noncongruent mate
S.
3. (Cartan XII) Set up an EDS for pairs of surfaces S, S  ⊂ E3 such
that there exists a diffeomorphism φ : S → S  that is a conformal map
preserving both families of asymptotic directions. What does the general
solution depend on? Show that, if S is not minimal, there exists at most
one such S  , up to congruence. In the special case where S is minimal, what
are the surfaces S  such that there exists such a φ?

6.9. Functions whose Hessians commute, with remarks on

singular solutions
We set up an exterior differential system for s functions f a : Rn → R,
1 ≤ a, b ≤ s, whose Hessians commute at every point. This system was
motivated by a problem in variation of Hodge structures; see [34, 145].
Let J 2 (Rn , Rs ) have coordinates (xi , ua , pai , paij ), where 1 ≤ i, j ≤ n,
with the canonical contact system generated by θa = dua − pai d xi and θia =
dpai − paij dxj .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 162 6. The Cartan Algorithm for Linear Pfaffian Systems

Let Σ ⊂ J 2 (Rn , Rs ) be the subset defined by [pa , pb ] = 0, where pa =

(paij ) and [, ] is the commutator of matrices. At each f ∈ Σ there exists
a matrix h ∈ O(n) and eigenvalues lia such that each pa = hla h−1 for a
diagonal matrix (la )ij = lia δij . We restrict attention to the open subset U of
Σ on which the common eigenspaces of the pa are all one-dimensional. Then
h is unique up to permutations, and h and the lia are smooth functions.
Functions with commuting Hessians are in one-to-one correspondence
with integral manifolds of the restriction I of the contact system on
J 2 (Rn , Rs ) to Σ. We will write the second set of generators for I in vector
form as Θa = d(pa ) − hla h−1 dx. Let ω i = (h−1 dx)i and φij = (h−1 dh)ij . We
will use matrix notation; in particular, the skew-symmetric matrix φ = (φij )
satisfies dω = −φ ∧ ω and the Maurer-Cartan equation dφ = −φ ∧ φ.
To perform Cartan’s Test, we calculate the exterior derivatives of the
forms in I. We get dθa ≡ 0 mod I and
dΘa ≡ −h([φ, la ] + dla ) ∧ ω mod I.
Since the matrix h is invertible, we can ignore it when computing characters
and dim A(1) . (Our choice of notation has the unfortunate consequence that
dim W = ns for this tableau.)
Let ψ a = [la , φ], so ψ aij = (lja − lia )φij , and in particular ψ aii = 0 and
ψ aij = ψ aji for i = j. There is no torsion, and the tableau consists of s
blocks: ⎛ a ⎞
dl1 ψ a1 2 ... ψ a1
n
⎜ ψ a2 dla ... ψ a2 ⎟
⎜ 1 2 n⎟
⎜ .. ⎟.
⎝ . ⎠
ψ an
1 ... ψ an
n−1 dlna
Because all the dlia ’s are independent,
 we see that s1 = ns (e.g., by adding
the columns together). There are n2 remaining independent entries in the
tableau. For generic values of the eigenvalues, we can recover the maximum
number of independent entries in each column. Thus
  
s2 = min{sn, n2 }, s3 = min{sn, n2 − sn, 0}, s4 = min{sn, n2 − 2sn, 0},
etc. To calculate dim A(1) , suppose
(6.20) dlj1 = ajk ω k ,
which is a choice of n2 constants. Then

ψ 1ij = aij ω i + aji ω j + bijk ω k ,
k=i,j

where the aij ’s are from (6.20) and the bijk ’s are additional constants which
must satisfy bijk = bikj = bjik .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.9. Functions whose Hessians commute 163

Let 2 ≤ ρ, σ ≤ s and let mρij = (liρ − ljρ )/(li1 − lj1 ). Then ψ ρij = mρij ψ 1ij for
each i = j, and consequently mρij bijk = mρik bikj . Since generically the values
of the mρij are all distinct, all the bijk must be zero. The only remaining
freedom in choosing an integral element are the “diagonal” elements, i.e.,
we may set

dljρ = mρij aij ω i + cρj ω j ,

where the cρj ’s are independent constants. Thus

 
n
dim A (1)
= n + (s − 1)n = 2
2
+ s1 .
2

Thus if we are in the case where n2 ≤ sn, i.e., s ≥ n−1
2 , then the system
n
is involutive and general integral manifolds depend on 2 functions of two
variables and sn functions of one variable.
There are many other families of integral manifolds of this system. For
example: a naı̈ve way to construct s functions of n variables whose Hessians
commute is to take the first function to be arbitrary, and the rest of the
form f a (x1 , . . . , xn ) = ua (x1 + . . . + xn ), so that the Hessians of f a are all
multiples of the identity matrix. Such integral manifolds are called singular
solutions.
From the perspective of EDS, these integral manifolds are obtained by
restricting to the stratum Σ ⊂ Σ where paij = C a δij for a > 1. Note that
they depend roughly on one function of n variables, and thus the space of
these integral manifolds is larger than for integral manifolds with generic
initial data.
Singular solutions. So far we have discussed integral manifolds of a system
(I, J) on Σ obtained by solving a series of Cauchy problems starting at a
generic point x ∈ Σ and a generic flag in Tx Σ among flags leading to n-
dimensional integral elements. A singular solution to an exterior differential
system is an integral manifold obtained by either starting at a nongeneric
point y ∈ Σ, or taking a nongeneric flag for the Cauchy problems.
If one is at a nongeneric point y ∈ Σ and restricts to the submanifold
Σ ⊂ Σ of points near y having the same nongenericity properties, one can
again use the Cartan algorithm restricted to Σ to find integral manifolds.
For singular solutions corresponding to a nongeneric flag, one can often
still apply Cartan’s algorithm to a modified system. In small dimensions,
singular solutions can often be treated on a case by case basis; see the
numerous examples in [40].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 164 6. The Cartan Algorithm for Linear Pfaffian Systems

6.10. The Cartan-Janet Isometric Embedding Theorem

We return to the isometric embedding
 problem. Following the proof in [13],
we will show that r(n) = n2 .
Theorem 6.10.1 (Cartan-Janet). Let (M n , g) be an analytic Riemannian
n
manifold. Then there exist local isometric embeddings of M n into En+( 2 )
that depend on n functions of n − 1 variables.
We first need to see that the variety Σ ⊂ Σ defined by (6.6) is nonempty,
i.e., that no matter what the curvature tensor of M is at a point x, there is
some second fundamental form giving rise to that curvature tensor. As in
Aside 6.4.5, we let V = Rn , W = (Rs ,  , ) and define
γ : S 2 V ∗ ⊗W → K(V ),
(6.21) 
haij v i v j ⊗wa → (haik hajl − hail hajk )v i ⊗v j ⊗v k ⊗v l ,
a

where K(V ) = S22 (V ) is the intersection of S 2 (Λ2 V ) with the kernel of

the skew-symmetrization map δ : Λ2 V ⊗ Λ2 V → Λ4 V . Let δ  denote the
restriction to K(V )⊗V of the natural map
IdΛ2 V ⊗δ : Λ2 V ⊗(Λ2 V ⊗V ) → Λ2 V ⊗Λ3 V
given by skew-symmetrization on the last two factors, and define K(1) (V ) =
ker δ  . We saw in Appendix A that dΘx ∈ K(1) (Tx∗ M ), i.e., that K(1) (V ) is
the space of derivatives of the curvature tensor.

Lemma 6.10.2. If s ≥ n2 , then γ is surjective.

Proof. Let V− = {v 1 , . . . , v n−1 } and let 1 ≤ α, β, γ,  ≤ n − 1.

Assume the lemma is true for W ⊗ S 2 V− → K(V− ); we will show it is
true for W ⊗S 2 V → K(V ). (Note that K(R1 ) = 0.)
Let R ∈ K(V ). By assumption, there are vectors hαβ = haαβ wa ∈ W
such that
R αβγ = hαγ · h β − hαβ · h γ ,
 a a
where hαγ · h β := a hαγ h β . Let W− = {hαβ }. It is sufficient to find new
vectors w0 , w1 , . . . , wn−1 ∈ W \W− such that
hαγ · wβ − hαβ · wγ = Rnαβγ ,
hαβ · w0 − wα · wβ = Rnαnβ ,
because if they exist, we may take
hαn = wα ,
hnn = w0 ,
to obtain γ(h) = R. To find appropriate vectors w0 , wα , consider the fol-
lowing:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.10. The Cartan-Janet Isometric Embedding Theorem 165

Write R = Rnαβγ v α ⊗(v β ∧ v γ ) ∈ V− ⊗Λ2 V− . Note that R ∈ S21 (V− ),

because it is in the kernel of the map δ : V− ⊗Λ2 V− → Λ3 V− . Recall from
Appendix A that S21 (V− ) may also be described as the image of the natural
map q : S 2 V− ⊗V− → V− ⊗Λ2 V− . Thus, we may write R = q(r) for some
r ∈ S 2 V− ⊗V− . In indices, Rnαβγ = rαβγ − rαγβ . We then seek vectors wα
such that hαβ · wγ = rαβγ and hαβ , wγ are linearly independent. Since the
inner product on W is nondegenerate and dim W − dim W− ≥ n, we can
always find such wα . Similarly, once we have done so we can find w0 such
that hαβ · w0 = Rnαnβ + wα · wβ , proving surjectivity. 

Proof of Cartan-Janet. Our system is (I, ˜ J)

˜ of (6.5), restricted to the
 n
submanifold Σ given by (6.6). We set r = 2 . We need to show that there
is no torsion and that the tableau is involutive.
Fix a general point y = ((x, u), (p, v), h) ∈ Σ ⊂ FM × FEn+r × W ⊗
S2V ∗. Since γ is a smooth map at general points and h is a general point
satisfying the Gauss equations, γ∗h : Th (W ⊗S 2 V ∗ ) → Tγ(h) K(V ) is defined
and surjective. We form a new map by tensoring γ∗h with V ∗ . We identify
Th (W ⊗S 2 V ∗ ) with W ⊗S 2 V ∗ , Tγ(h) K(V ) with K(V ), and symmetrize both
sides to obtain a linear map
(6.22) γ̃∗,h : W ⊗S 3 V ∗ → K(1) (V ∗ )
which is also surjective.
To show the apparent torsion is absorbable, we first show that it lives
in K(1) (V ). Note that integral elements at y are given by the equations
γ(h, dh) = dR. We may also write these equations as
γ(h, π) = ∇R,
μ
where π = πjk wμ v j v k , and we define
μ
πjk := −dhμjk − hνjk ωνμ + hμjl ωkl + hμkl ωjl
and
(∇R)ijkl = dRijkl + γ(hil , (hνjk ωνμ + hμjm ωkl + hμkm ωjm )wμ ).
Since γ is quadratic, γ∗,h (π) = γ(h, π), Thus, [T ] = 0 by the surjectivity of
γ̃∗h .
We now need to calculate dim A(1) and the characters sk . To show the
system is involutive, it is sufficient to show dim A(1) ≥ s1 + 2s2 + . . . + nsn .
We will estimate both sides to obtain the inequality (which of course then
must be an equality).
On an integral element we have πij a = ha η k for some constants ha ,
ijk ijk
and dim A(1) is the dimension of the space of choices of such constants. The
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 166 6. The Cartan Algorithm for Linear Pfaffian Systems

surjectivity of (6.22) implies

dim A(1) ≥ dim W ⊗S 3 V ∗ − dim K(1) (V )
 
n+2 n2 (n2 − 1)(n + 2)
(6.23) =r − .
3 24
We now estimate the characters. Fix p < n and let 1 ≤ s, t, u ≤ p − 1,
p ≤ x, y, z ≤ n. Consider the equations
(6.24) htu · πpx = −hpx · πtu + htx · πup + hpu · πtx ,
(6.25) htx · πpy − hty · πpx = −hpy · πtx + hpx · πty .

Equation (6.24) provides p2 (n − p + 1) equations on the πpx in terms of the
πsj . (The first factor is the number of choices of pairs t ≤ u, the second the
number of choices for x.) Equation (6.25) provides (p−1) n−p+1 2 equations.
Thus
   
p n−p+1
sp ≤ r(n − p + 1) − (n − p + 1) − (p − 1)
2 2
and sn = 0. Therefore

s1 + 2s2 + . . . + nsn
 
n−1  
p

n−p+1

≤ p r(n − p + 1) − (n − p + 1) − (p − 1) .
2 2
p=1

The summation on the right-hand side equals (6.23), so we obtain the desired
inequality. 

6.11. Isometric embeddings of space forms (mostly flat ones)

We have seen that an arbitrary analytic Riemannian n-fold locally isomet-
n
rically embeds into En+( 2 ) . However, special metrics might embed into
smaller Euclidean spaces. For example, the round sphere occurs as a hyper-
surface. Here we study embeddings of the other two types of space forms,
flat metrics and hyperbolic metrics, concentrating on the case of flat metrics.
We have seen that any flat surface in E3 must have a degenerate Gauss
map. Here we only study flat submanifolds of En+r with immersive Gauss
mappings. This is equivalent to saying that the second fundamental form II
is such that for all v ∈ Tx M , there exists w ∈ Tx M such that II(v, w) = 0.
We call such forms nondegenerate.
Lemma 6.11.1 (Cartan). Let V = Rn , W = Rr , and let W have a nonde-
generate inner product  , . Let h ∈ W ⊗S 2 V ∗ be nondegenerate and satisfy
γ(h) = 0. Then r ≥ n, and if r = n there exist a basis v 1 , . . . , v n of V ∗ and
an orthonormal basis w1 , . . . , wr of W such that h = v i ◦ v i ⊗wi .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.11. Isometric embeddings of space forms 167

Corollary 6.11.2 (Cartan). There are no local isometric embeddings of

a flat Riemannian manifold (M n , g) into En+r with nondegenerate second
fundamental form for r < n.

Proof. Assume h is as in the hypotheses of the lemma. Let v0 ∈ V . Con-

sider the linear map
v0 h : V → W,
v → h(v0 , v).
We claim that if v0 is generic, then v0 h is injective, which will prove the
first assertion. Let 1 ≤ s ≤ r and choose a basis w1 , . . . , wr of W so that
h = hs ws with hs ∈ S 2 V ∗ .
Say the claim were false; then consider U0 := {v0 hs }  V ∗ . Let
1 ≤ ξ, η ≤ dim U0 and dim U0 + 1 ≤ φ, ψ ≤ r. Adapt our basis of W so that
v0 hψ = 0 for all ψ and the v0 hξ span U0 . Let v ξ = v0 hξ . By hypothesis,
for all u, v ∈ V , we have

0 = (u ∧ v) γ(h) = (u hs ) ∧ (v hs ).
s

Consider 
0 = (v ∧ v0 ) γ(h) = (v hξ ) ∧ v ξ .
ξ
By the Cartan Lemma A.1.9 we have
(6.26) {v hξ | v ∈ V, 1 ≤ ξ ≤ dim U0 } ⊆ U0 .
This implies Singloc{hξ } = {v hξ }⊥ ⊇ U0 ⊥ and thus is nonempty. Let
v1 ∈ Singloc{hξ }. There is some index φ and v2 ∈ V such that hφ (v1 , v2 ) =
0, because Singloc(h) = 0. For all v ∈ V , v0 + v will be generic if  is
sufficiently small, because v0 is generic. Consider
U1 = {(v0 + v2 ) hs } = {(v0 + v2 ) hξ , v2 hφ }.
By (6.26), {(v0 +v2 ) hξ } ⊆ U0 , and in fact we have equality because v0 +v2
is generic. But {v2 hφ } ⊂ U0 , which implies dim U1 > dim U0 , giving a
contradiction.
To prove the normal form, consider a basis v1 , . . . , vn of V where v1 is
h-generic. Let 2 ≤ ρ, σ ≤ n. Again, since
(vρ ∧ v1 ) γ(h) = 0,
the Cartan Lemma implies that vρ hs = Cts (ρ)(v1 ht ) for some symmetric
matrices C(ρ). Using (vρ ∧ vσ ) γ(h) = 0, we see that the matrices commute
and are therefore simultaneously diagonalizable by an orthogonal matrix,
say B ∈ O(W ). Let q s = Bts ht . We have arranged things so that for each
index s, dim{vj q s | 1 ≤ j ≤ n} = 1, so rank q s = 1 for all s. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 168 6. The Cartan Algorithm for Linear Pfaffian Systems

We now study existence when r = n. (We follow the arguments in [13].)

We have seen that if we are willing to abandon orthonormal frames, there is
a nice normalization of the second fundamental form. In fact we don’t need
to abandon orthonormal frames entirely, just in the tangent space of M .
Let F be the subbundle of FGL(E2n ) consisting of frames (x, e1 , . . . , e2n )
adapted so that
ei , en+j  = 0,
en+i , en+j  = δij .
If we set
Gij = ei , ej ,
then the pullback of the Maurer-Cartan form of FGL(E2n ) to F has the rela-
tions
n+i n+j
(6.27) ωn+j + ωn+i = 0,
(6.28) ωjn+i + Gjk ωn+i
i
= 0,
while otherwise the forms are independent. (Equation (6.28) is obtained
from 0 = dej , en+j .)
Consider the Pfaffian system
I = {ω n+j , ωjn+i − δ ij ω j }
with independence condition Ω = ω 1 ∧ . . . ∧ ω n . Its integral manifolds are
lifts of n-folds whose second fundamental form is in the normal form of the
conclusion of Lemma 6.11.1. Thus the integral manifolds are flat, and all
flat n-folds in E2n with immersive Gauss maps occur as integral manifolds
of I. We calculate
dω n+j ≡ 0 mod I,
d(ωjn+i − δ ij ω j ) ≡ ωji ∧ ω i − ωn+j
n+i
∧ ω j − δ ij ωli ∧ ω l mod I.
There is no torsion (as the only symbol relations are (6.27)), so wecalculate
the characters of the tableau A. We see that s1 = n2 and s2 = n2 . (If this
is not clear, try writing out the n = 3 case.) Since dim A is the number
n+i
of independent forms among the ωji , ωn+j , we see dim A = n2 + n2 and
the remaining characters are zero. We show A is involutive by explicitly
describing the space of integral elements: let {Bji | 1 ≤ i, j ≤ n} and
{Cji | i = j} be constants, and set
n+i
ωn+j = Cji ω i − Cij ω j ,

ωii = Bki ω k ,
k
ωji = 2Bji ω i − Cji ω j .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.12. Calibrated submanifolds 169

It is easy to verify that this determines an integral element, so dim A(1) ≥
n2 + 2 n2 . But we already know that dim A(1) ≤ n2 + 2 n2 , so we have
equality, and the tableau is involutive.
In summary, we have shown:
Theorem 6.11.3 (Cartan [37]). Local isometricembeddings En → E2n
with immersive Gauss maps exist and depend on n2 functions of 2 variables.

Local isometric embeddings of hyperbolic space. We have seen that

the hyperbolic 2-space H 2 can be locally embedded into E3 . We now study
H n for n > 2.
Let g0 denote the metric on H n . Consider γ 0 : S 2 V ∗ → K obtained by
taking W = R in (6.21). The Riemann curvature tensor of H n is given by

R = −γ 0 (g0 , g0 ).

Thus in order to obtain isometric embeddings of H n , we need second funda-

mental forms h satisfying γ(h, h) = −γ 0 (g0 , g0 ). Here is a nifty trick due to
Cartan: Let Ŵ = W ⊕ R{e0 }, and ĥ = h + g0 e0 . The admissible second fun-
damental forms h are exactly those such that γ(ĥ, ĥ) = 0. The hypotheses
of Lemma 6.11.1 are satisfied for ĥ, and we conclude
Theorem 6.11.4 (Cartan [37]). Hyperbolic space H n does not admit any
local isometric embeddings into En+r for r < n − 1.
There do exist local isometric embeddings for r = n − 1; see [37] or [13].

6.12. Calibrated submanifolds

Let (X n+r , g) be a Riemannian manifold and let M n ⊂ X be a subman-
ifold (possibly with boundary). Recall that M is minimal if any x ∈ M
is contained in an open neighborhood U ⊂ M such that if V is an open
subset with V ⊂ U and V  is a submanifold of X with ∂V = ∂V  , then
vol V  ≥ vol V . M is minimal if and only if its mean curvature vector is
identically zero. M is said to be minimizing if for all x ∈ M , we may take
U = M.
How can one construct minimizing submanifolds or even prove that a
given submanifold is minimizing? There is no known general method. In this
section we will describe a method for studying special classes of minimizing
submanifolds. We begin with the simplest case:
A line is the shortest path between two points. Let p, q ∈ E2 be two points.
How does one prove that the shortest path between p and q is the line
segment connecting them?
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 170 6. The Cartan Algorithm for Linear Pfaffian Systems

Let M (t) denote the parametrized line segment from a to b and let M (t)
denote any rectifiable path from a to b. Without loss of generality, use
coordinates (x1 , x2 ) such that the line segment is along the x1 -axis.

M
~

a M b

We have
dx1 (M  (t)) = |M  (t)|,
dx1 (M̃  (t)) ≤ |M̃  (t)|.
Letting φ = dx1 , we have

(6.29) vol(M ) = φ= φ ≤ vol(M̃ ),

M M̃
where the middle equality follows from Stokes’ Theorem.
The form φ, which enables us to show that M is minimizing, is an
example of a calibration.

Calibrations. Throughout this section we use G(n, T X) to denote the

Grassmann bundle of unit volume oriented n-planes in the tangent spaces of
a Riemannian manifold X. That is, we only represent an n-plane by vectors
v1 ∧ . . . ∧ vn such that |v1 ∧ . . . ∧ vn | = 1, where the norm at each point
is induced from the Riemannian metric. We use G(n, Tx X) ⊂ Λn Tx X to
denote the fiber at a point x. We will also use G(n, m) ⊂ Λn Rm to denote
the set of unit volume n-planes in Rm .
Definition 6.12.1 (Harvey and Lawson [92]). Let (X n+r , g) be a Riemann-
ian manifold. A calibration φ ∈ Ωn (X) is a differentiable n-form such that
i. dφ = 0, and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.12. Calibrated submanifolds 171

ii. φ(E) ≤ 1 for all E ∈ G(n, T X).

A submanifold i : M → X is calibrated by φ if i∗ (φ) = dvolM , where
dvolM is the volume form induced by the immersion.
Theorem 6.12.2 (Fundamental lemma of calibrations). Calibrated sub-
manifolds are minimizing.

Proof. Use equation (6.29). 

Remark 6.12.3. The condition to be calibrated corresponds to satisfying
a first-order system of PDE, while being minimal (mean curvature vector
identically zero) corresponds to satisfying a second-order system of PDE.

Given a calibration φ, what does it calibrate? The answer can be deter-

mined by setting up and solving an exterior differential system. We will set
up two different types of systems for calibrated submanifolds of constant-
coefficient calibrations. In what follows we will use linear Pfaffian systems,
while in Chapter 8 we will use non-Pfaffian systems, once we have the general
form of the Cartan-Kähler Theorem at our disposal.
Let φ ∈ Ωn (X) be a calibration, and let x ∈ X. We define
Facex (φ) := {E ∈ G(n, Tx X) | φ(E) = 1}.
Let Face(φ) ⊂ G(n, T X) be the corresponding bundle. M n ⊂ X is cali-
brated by φ if and only if γ(M ) ⊂ Face(φ), where γ : M → G(n, T X) is the
Gauss map x → (x, Tx M ).
Remark 6.12.4. Let φ ∈ Ωn (En+r ) have constant coefficients in a linear
coordinate system. Then, after an appropriate normalization, φ is a calibra-
tion. In this situation, we may think of Face(φ) as contained in G(n, n + r)
(since Facex (φ) is independent of x), and the EDS is considerably easier to
set up.

Complex submanifolds of a Kähler manifold are minimizing. Let (X, g, J, ω)

be a Kähler manifold, that is, a complex manifold with a Riemannian met-
ric g, symplectic form ω and the compatibilities g(v, Jw) = ω(v, w) and
g(Jv, Jw) = g(v, w) for all tangent vectors at all points.
√ = Ej with standard
For example, in Cm 2m coordinates (xj , y j ), 1 ≤ j ≤
m, if we let z = x + −1y (so that J(dx ) = −dy j , J(dy j ) = dxj ), then
j j j

i
ω = (dz 1 ∧ dz 1 + . . . + dz m ∧ dz m ).
2
1 ∧k
Proposition 6.12.5. Let (X, ω) be Kähler and let φ = k! ω . Then φ is
a calibration and Facex (φ) = G(C , C ) ⊂ G(2k, Tx X). In other words,
k m

Facex (φ) = {E ∈ G(2k, Tx X) | J(E) = E}.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 172 6. The Cartan Algorithm for Linear Pfaffian Systems

Wirtinger observed that if v, w is an oriented orthonormal basis of a

plane E, then ω(v, w) ≤ 1 with equality if and only if E is a complex line.
Federer generalized this to φ, and Harvey and Lawson’s general definition
was inspired by the observation that Wirtinger’s inequality implies that
complex submanifolds are minimizing.

Proof of Wirtinger’s inequality, case k = 1. Let E have orthonormal

basis v, w. Recall the Cauchy-Schwarz inequality that x, y ≤ |x||y|, with
equality if and only if y = λx for some λ ∈ R+ :

ω(v, w) = v, J(w) ≤ |v||Jw| = |v||w|.

One has equality in the Cauchy-Schwarz inequality if and only if J(w) = v,

which implies E is complex. For the general case, see [91]. 

We now discuss four calibrations discovered by Harvey and Lawson [92],

the special Lagrangian calibration, the associative calibration, the coassocia-
tive calibration and the Cayley calibration. Each produces a large class of
minimizing submanifolds. These manifolds have recently become important
for the study of mirror symmetry and M -theory [53]. Associative and coas-
sociative calibrations exist for manifolds with G2 holonomy, and the Cayley
calibration exists for manifolds with Spin(7) holonomy (see Chapter 9 for a
discussion of holonomy).
Special Lagrangian submanifolds. Let R2m = Cm have its standard complex
structure J and flat metric  , . As explained in §A.3, we obtain a symplectic
form ω on R2m from this data. Then E ∈ G(m, R2m ) is Lagrangian, i.e.,
ω|E =0, if and only if J(E) ⊥ E (see [91]).
Using standard complex coordinates on R2m = Cm , we define a complex-
valued m-form dz = dz 1 ∧ . . . ∧ dz m .
Proposition 6.12.6. Given E ∈ G(m, 2m) (the unit volume Grassmann-
ian) we have |dz(E)| ≤ 1, with equality if and only if E is Lagrangian.

Proof. We will show that

|dz(E)| = |E ∧ J(E)|.

The result will then follow by Hadamard’s inequality (A.5).

Write R2m = V and let e1 , . . . , em , Je1 , . . . , Jem be an orthonormal basis
of V , where ej = ∂x∂ j , and let 1 , . . . , m be an orthonormal basis of E. There
exists A ∈ End(V ) such that A(ei ) = i and A(Jei ) = Ji . Such an A is
complex-linear because it commutes with J. Let E0 = e1 ∧. . .∧em . We have
ρ(A)(E0 ) = E, where ρ(A) : Λm V → Λm V is the induced linear mapping.
Therefore, |dz(E)| = |dz(ρ(A)E0 )| = | detC (A)dz(E0 )| = | detC (A)|. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 6.12. Calibrated submanifolds 173

Since dz is complex-valued, it cannot be a calibration, but we can form

calibrations from it: Let
α = Re(dz) = Re(dz 1 ∧ . . . ∧ dz m ) ∈ Ωm (Cm ).
Corollary 6.12.7. α is a calibration, called the special Lagrangian calibra-
tion.
The proof of Proposition 6.12.6 shows that in fact αθ = Re(eiθ dz) is a
calibration for all θ ∈ [0, 2π). From this proof, we also deduce Face(α):
Corollary 6.12.8. Face(α) is the SU (m)-orbit of E0 .
We now derive a linear Pfaffian system for special Lagrangian manifolds.
By Appendix A, we may write, with respect to the basis ej , Jej above,
 
κ −ψ
su(m) = | κ ∈ so(m), t ψ = ψ, trace(ψ) = 0 .
ψ κ
We set up an EDS on ASU (m) = FSU (m) → R2m which has Maurer-Cartan
form ⎛ ⎞
0 0 0
ω = ⎝ ωi ωji i
ωn+k ⎠,
n+l n+l n+l
ω ωj ωn+k
 n+i
where ωji = −ωij = ωn+j
n+i
, ωln+k = ωkn+l = −ωn+l
k , and
i ωi = 0.
Let I = {ω n+i }, J = {ω i , ω n+i }. The tableau determined by the (ωjn+l )
has characters s1 = n, s2 = n − 1, . . . , sn−1 = 2, sn = 0.
Exercise 6.12.9: Show that (I, J) is involutive.

Associative submanifolds. Let R7 be equipped with a nondegenerate positive

3-form φ ∈ Λ3 R7 , or more generally let M 7 be a Riemannian manifold
with holonomy G2 , and call the induced 3-form φ as well. Such forms are
discussed in §A.5, and holonomy is discussed in Chapter 9.
The form φ is a calibration, called the associative calibration, and its
calibrated manifolds are called associative manifolds. The name comes from
the fact that φ induces a structure of the imaginary octonions on R7 and
the three-planes calibrated by φ are exactly those isomorphic to a copy
of the imaginary quaternions, i.e., where the octonionic multiplication is
associative (see Example 8.5.8 for details).
There exists a basis e1 , . . . , e7 for R7 such that
 
ρ(B) −tA 
(6.30) g2 = B ∈ so(4), ρ(B) ∈ so(3) ,
A B 
where the first two columns of A are arbitrary but the third column is depen-
dent on the first two (see Exercises A.5.3). As in the special Lagrangian case,
φ-calibrated manifolds are the integral manifolds of a linear Pfaffian system
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 174 6. The Cartan Algorithm for Linear Pfaffian Systems

given by I = (ω μ ) and J = (ω μ , ω j ), where 1 ≤ j ≤ 3 and 4 ≤ μ ≤ 7. The

tableau of the system has the form A, and we have already seen that such
a tableau must be involutive (set n = 3 in Exercise 5.5.9.1). We conclude:
Proposition 6.12.10. The Pfaffian system for associative manifolds is in-
volutive with character 3 and Cartan integer 4.
Coassociative submanifolds. We may also consider the calibration ψ := ∗φ
on E7 .
Exercise 6.12.11: Set up and perform Cartan’s Test for coassociative sub-
manifolds.

Cayley submanifolds. One can similarly define a calibration 4-form on R8

related to the Spin7 action on R8 . The form is called the Cayley form. One
arrives at a similar EDS for Cayley manifolds as follows: Here
 
sp(1)1 + ρ(sp(1)2 ) −t A
spin7 = ,
A sp(1)3 + ρ̃(sp(1)2 )
where the columns of A satisfy J1 π1 + J2 π2 + J3 π3 − π4 = 0 and ρ, ρ̃ are
both isomorphic to the standard representations, but are given in terms of
different, nonstandard bases.
Setting up the system as for associative manifolds, the tableau is
A = (u, v, w, 1 u + 2 v + 3 w),
which is clearly involutive.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/07

Chapter 7

Applications to PDE

Consider the well-known closed-form solution of the wave equation utt −

c2 uxx = 0 due to d’Alembert:
(7.1) u(x, t) = f (x + ct) + g(x − ct),
where f and g are arbitrary C 2 -functions. It is rare that all solutions of
a given PDE are obtained by a single formula (especially one which does
not involve integration). The key to obtaining the d’Alembert solution is to
rewrite the equation in “characteristic coordinates” η = x + ct, ξ = x − ct,
yielding uηξ = 0. By integrating in η or in ξ, we get
uη = F (η), uξ = G(ξ),
where F and G are independent arbitrary functions; then (7.1) follows by
another integration. For which other PDE do such solution formulas exist?
And, how can we find them in a systematic way?

In this chapter we study invariants of exterior differential systems that

aid in constructing integral manifolds, and we apply these to the study of
first- and second-order partial differential equations, and classical surface
theory. (For second-order PDE, we will restrict attention to equations for
one function of two variables.) All functions and forms are assumed to be
smooth.
In §7.1 we define symmetry vector fields and Cauchy characteristic vector
fields for EDS. We discuss the general properties of Cauchy characteristics,
and use them to recover the classical result that any first-order PDE can be
solved using ODE methods. We also define the Cartan system of an EDS,
and establish sufficient conditions for a Pfaffian system to descend to be
well-defined on a quotient space. In §7.2 we define the Monge characteristic

175
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 176 7. Applications to PDE

systems associated to second-order PDE, and discuss hyperbolic exterior dif-

ferential systems. In §7.3 we discuss a systematic method, called Darboux’s
method, which helps uncover solution formulas like d’Alembert’s (when they
exist), and more generally determines when a given PDE is solvable by ODE
methods. We also define the derived systems associated to a Pfaffian system.
In §7.4 we treat Monge-Ampère systems, focusing on several geometric
examples. We show how solutions of the sine-Gordon equation enable us to
explicitly parametrize surfaces in E3 with constant negative Gauss curvature.
Consideration of complex characteristics for equations for minimal surfaces
and for surfaces of constant mean curvature (CMC) leads to solutions for
these equations in terms of holomorphic data. In particular, we derive the
Weierstrass representation for minimal surfaces and show that any CMC
surface has a one-parameter family of noncongruent deformations.
In §7.5 we discuss integrable extensions and Bäcklund transformations.
Examples include the Cole-Hopf and Miura transformations, the KdV equa-
tion, and Bäcklund’s original transformation for pseudospherical surfaces.
We also prove Theorem 7.5.14, relating Bäcklund transformations to
Darboux-integrability.

7.1. Symmetries and Cauchy characteristics

Infinitesimal symmetries. One of Lie’s contributions to the theory of
ordinary differential equations was to put the various solution methods for
special kinds of equations in a uniform context, based on infinitesimal sym-
metries of the equation—i.e., vector fields whose flows take solutions to
solutions. This generalizes to EDS when we let the vector fields act on
differential forms via the Lie derivative operator L (see Appendix B):
Definition 7.1.1. Let I be an EDS on Σ. A vector field v ∈ Γ(T Σ) is an
(infinitesimal) symmetry of I if Lv ψ ∈ I for all ψ ∈ I.
Exercises 7.1.2:
1. The Lie bracket of any two symmetries is a symmetry; thus, symmetries
of a given EDS form a Lie algebra. 
2. To show that v is a symmetry, it suffices to check the condition in Defi-
nition 7.1.1 on a set of forms which generate I differentially. 
3. Let I be the Pfaffian system generated by the contact form θ = dy − zdx
on R3 . Verify that
∂ ∂ ∂
(7.2) v=f +g + (gx + zgy − zfx − z 2 fy ) ,
∂x ∂y ∂z
for any functions f (x, y), g(x, y), is a symmetry. Then, find the most general
symmetry vector field for I. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.1. Symmetries and Cauchy characteristics 177

Given an integral manifold N0 of an exterior differential system I, a

symmetry vector field v can be used to obtain a one-parameter family of
integral manifolds, as follows. Let φt denote the one-parameter group of
diffeomorphisms of Σ generated by v, and let
(7.3) Nt := φt (N0 ).
Exercise 7.1.3: Show that if Nt is smooth, then it is an integral manifold
of I. 
Remark 7.1.4. Cartan [35, 44] initiated the classification of exterior dif-
ferential systems whose symmetries depend on arbitrary functions (like the
contact symmetries in Exercise 7.1.2). Determining the symmetries usually
amounts to solving an overdetermined system of differential equations, so
most of the time the Lie algebra of symmetries is empty. An interesting
exception is the Pfaffian system {dy − pdx, dp − qdx, dz − q 2 dx}diff de-
fined on R5 with coordinates x, y, z, p, q. In this case the symmetries form
a 14-dimensional Lie algebra isomorphic to the split form of the excep-
tional simple Lie algebra g2 (see Appendix A). In his famous “five variables
paper” [36] Cartan proved that this is the largest possible Lie algebra of
symmetries for a rank three Pfaffian system with generic derived flag on a
five-dimensional manifold.

Unfortunately, enumerating the symmetry vector fields for a given sys-

tem of differential equations is usually harder than solving the system itself.
For this reason, we refrain from further discussion of symmetries in general
(however, the interested reader should consult the monograph [8] or the
textbook [157]); instead we focus on a class of symmetries which are easily
calculated and allow us to reduce the dimension of the manifold on which
the system is defined:

Cauchy characteristics.
Definition 7.1.5. Let I be an EDS on Σ. A vector field v ∈ Γ(T Σ) is a
Cauchy characteristic vector field of I if v ψ ∈ I for all ψ ∈ I.

The flow lines of v are called Cauchy characteristic curves (or, simply,
Cauchy characteristics) of I.
Exercises 7.1.6:
1. If v is a Cauchy characteristic vector field for I, then it is also a symme-
try. 
2. The Lie bracket of any two Cauchy characteristic vector fields is a Cauchy
characteristic vector field. 
3. Show that v is a Cauchy characteristic vector field if and only if v ψ ∈ I
for a set of forms ψ that generate I algebraically. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 178 7. Applications to PDE

While calculating the symmetries of a given EDS involves solving dif-

ferential equations for the components of the vector field, the condition in
Definition 7.1.5 is algebraic (and linear) in the components of v with respect
to a local basis. If v is a Cauchy characteristic vector field, so is any mul-
tiple of v by a smooth nonvanishing function. (We will refer to this kind of
modification as “up to multiple”.) Thus, it would be more correct to speak
of Cauchy characteristic line fields; however, for calculations it is easier to
fix a vector field.

Example 7.1.7. On R4 with coordinates x, y, z, w, let I = {θ}diff for θ =

∂ ∂ ∂ ∂
dy − z dx. Suppose v = a +b +c +e is a Cauchy characteristic
∂x ∂y ∂z ∂w
vector field. Because dθ = dx ∧ dz, the condition v dθ ≡ 0 mod θ implies
that a = c = 0, while v θ = 0 implies that b = 0. Hence any Cauchy
characteristic vector fields for this system are multiples of ∂/∂w.

While the flow generated by a symmetry v preserves integral manifolds,

when v is a Cauchy characteristic vector field more is true:
Proposition 7.1.8. Let N0 be an n-dimensional integral manifold for an
EDS, and let v be a Cauchy characteristic vector field which is transverse to
N0 . Let Nt be as in (7.3), and let S be the (n + 1)-dimensional union of the
integral manifolds Nt . Then S, when smooth, is an integral manifold of I.

Proof (for dim N0 = 1). Suppose N0 is an integral curve of I with v trans-

verse to N0 . In order for the surface S to be smooth, we must restrict t to
the open interval of values for which v is transverse to Nt . Let Ω ∈ I 2 and
let w be tangent to one of the Nt . Then Ω(v, w) = (v Ω)(w) = 0, because
v Ω ∈ I. 

Exercise 7.1.9: Generalize this argument to the case where N0 has arbi-
trary dimension.

Example 7.1.10 (Surfaces in E3 ). Recall from §2.1 that the integral man-
ifolds of I = {ω 3 }diff on F = Fon (E3 ) are the first-order adapted lifts of
surfaces Σ2 ⊂ E3 . This system admits a Cauchy characteristic vector field
as follows:
Let (ω j )∗ , (ωji )∗ be a basis of vector fields on F dual to the entries of the
Maurer-Cartan form. Write v = aj (ω j )∗ + aji (ωji )∗ . Then v ω 3 = 0 implies
that a3 = 0, and

v dω 3 = v (−ω13 ∧ ω 1 − ω23 ∧ ω 2 )
= a1 ω13 − a13 ω 1 + a2 ω23 − a23 ω 2
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.1. Symmetries and Cauchy characteristics 179

implies that v = (ω12 )∗ up to multiple. As explained in §1.8, v corresponds to

infinitesimal rotations of the frame within the {e1 , e2 }-plane. Now, Propo-
sition 7.1.8 implies that any integral surface for I may be ‘thickened’ to an
integral 3-fold using the Cauchy characteristic.
Example 7.1.11 (A first-order PDE). Let (x1 , x2 , p1 , p2 , u) be the usual
coordinates on J 1 (R2 , R). The equation ut = ux is equivalent to the EDS
I obtained by restricting the canonical contact system (see §1.10) to the
smooth hypersurface Σ ⊂ J 1 (R2 , R) defined by p2 = p1 . We will use x =
x1 , t = x2 , u and p = p1 as coordinates on Σ, where the system is generated
algebraically by θ := du − p(dx + dt) and dθ = −dp ∧ (dx + dt).
Suppose
∂ ∂ ∂ ∂
+b +c +e v := a
∂x ∂t ∂p ∂u
is a Cauchy characteristic vector field on Σ. Then v θ = 0 implies that
e = (a + b)p. Then v dθ ≡ 0 modulo θ implies that c = 0 and a + b = 0.
∂ ∂
Hence v = − , up to multiple.
∂x ∂t
Thus, integral surfaces of I are constructed by translating integral
curves along curves where u, p and x + t are constant.

Solving first-order PDE. Example 7.1.11 generalizes, in that any first-

order PDE for one unknown function can, in theory, be solved through ODE
techniques; for, as we will show, the corresponding EDS always has a Cauchy
characteristic vector field.
Suppose u, x1 , . . . , xn , p1 , . . . , pn are the usual coordinates on J 1 (Rn , R),
and the PDE takes the form
(7.4) F (u, x1 , . . . , xn , p1 , . . . , pn ) = 0.
Let Σ ⊂ J 1 be the subset on which (7.4) holds. To ensure that Σ is a smooth
manifold, and the equation is genuinely of first order, we will assume that
∂F/∂pi = 0 for some i at each point of Σ.
Let θ be the pullback to Σ of the contact form du−p1 dx1 −. . .−pn dxn . As
explained in §1.10, solutions of the PDE (7.4) are in 1-to-1 correspondence
with n-dimensional integral manifolds i : N → Σ of the Pfaffian system
generated by θ, with the independence condition dx1 ∧ . . . ∧ dxn = 0.
Because dim Σ = 2n, θ ∧ (dθ)n = 0. Differentiating (7.4) gives
0 = dF = Fxi dxi + Fpj dpj + Fu du
(7.5)
≡ (Fxi + Fu pi )dxi + Fpj dpj mod θ.

It follows that θ ∧ (dθ)n−1 = 0 at each point (see Exercise 7.1.12 below).

So, by the Pfaff Theorem 1.10.16, there are local coordinates on Σ2n such
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 180 7. Applications to PDE

that θ is a nonzero multiple of

dy 1 + y 2 dy 3 + . . . + y 2n−2 dy 2n−1 .
Thus, v = ∂/∂y 2n gives a Cauchy characteristic. (Without reference to local
coordinates, v can be described
 as the unique vector field, up to multiple,
such that v θ ∧ (dθ) n−1 = 0.)
The Cauchy characteristic vector field enables us to reduce our PDE to
an ODE. We do this by choosing an (n − 1)-dimensional integral manifold
L of I (corresponding to initial data), and then using flow by v to get an n-
dimensional integral manifold. If F may be solved for pn , then v is transverse
to hyperplanes where xn is constant. Thus, we obtain a starting manifold
L by letting xn = C, u = f (x1 , . . . , xn−1 ) for an arbitrary function, and
pi = ∂f /∂xi . To obtain a solution to the original PDE we only need to
integrate the ODE system corresponding to flow by v.
Exercise 7.1.12: The relation (7.5) lets us locally solve for one of the dpi
in terms of the others, dxi and θ. Using this, compute dθ modulo θ on Σ2n ,
and conclude that θ ∧ (dθ)n−1 = 0.
Example 7.1.13 (The inviscid Burger’s equation ut = uux ). Using the
same coordinates on J 1 (R2 , R) as in Example 7.1.11, our equation becomes
p2 − up1 = 0, defining a smooth four-dimensional manifold Σ ⊂ J 1 (R2 , R).
On Σ, use x, t, u and p = p1 as local coordinates; then θ = du − p dx − up dt
and
(7.6) dθ ≡ −(dp − p2 dt) ∧ (dx + udt) mod θ.
Any Cauchy characteristic vector field v must be annihilated by θ and by
both factors on the right-hand side of (7.6). Thus, we may take
∂ ∂ ∂
v= −u + p2 .
∂t ∂x ∂p
The flow generated by v is obtained by solving the ODE system
dx/dt = −u, du/dt = 0, dp/dt = p2 .
This system has solution x(t) = x0 − ut, u(t) = u0 , p(t) = 1/(p−1
0 − t).
To construct an integral surface, we may start with an integral curve
of θ at t = 0 by setting x0 = s, u0 = f (s) and p0 = f  (s), where f is
an arbitrary C 1 -function. By flowing along the Cauchy characteristics we
obtain the implicit solution
(7.7) u = f (s) = f (x + ut).
Note that the implicit function theorem for (7.7) fails when tf  (s) = 1. When
this happens, the flow under v has caused p to become undefined, and the
graph of u as a function of x develops a vertical tangent. For sufficiently
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.1. Symmetries and Cauchy characteristics 181

large values of t, the surface defined by (7.7) may not be a graph over the
xt-plane.

Remark 7.1.14. In standard PDE texts, what we have discussed is called

the method of characteristics. In some treatments (e.g., [105]) the charac-
teristic curves in J 1 (R2 , R) are described as curves in J 0 (R2 , R) = R3 with a
tangent plane attached at each point, and these are known as characteristic
strips. The solution surface in R3 is then formed by the envelope of those
characteristic strips that pass through the initial data.

Exercises 7.1.15:
1. Consider the PDE u2x + u2y = 2u.
(a) Show that, when x, y, p, q are used as coordinates, the Cauchy charac-
teristic line field is generated by v = p(∂p + ∂x ) + q(∂q + ∂y ).
(b) Find all regular solutions satisfying initial condition u(x, 0) = 25 x2 . 
2. Solve 12 (u2x + u2y ) + xux + yuy = u subject to u(x, 0) = (1 − x2 )/2. 
3. [105, Ch. 1] Solve the PDE uy = u3x with the initial condition u(x, 0) =
2x3/2 . Show that the only solutions which are regular over all of R2 are
linear functions of x and y. 

Remark 7.1.16. A general first-order system of PDE does not admit a

Cauchy characteristic line field. For example, suppose a system of two
equations for two functions u(x, y), v(x, y) defines a smooth submanifold
Σ6 ⊂ J 1 (R2 , R2 ) that submerses onto J 0 (R2 , R2 ). The contact system re-
stricts to be a system I on Σ generated algebraically by a pair of 1-forms
θ1 , θ2 and a pair of linearly independent 2-forms. As we will see shortly, if
there were Cauchy characteristics, this would imply that I is locally a pull-
back to Σ of a system defined on a lower-dimensional quotient manifold—in
this case, at most dimension five. This would in turn require that the 2-
forms be decomposable, modulo the θ’s, and share a common factor. It is
easy to see that, generically, this is not the case.

Exercise 7.1.17: Show that the only quasilinear systems for u, v which have
nontrivial Cauchy characteristics are those that uncouple into two first-order
equations with the same characteristics.

Retracting spaces. Because the condition in Definition 7.1.5 is C ∞ -linear

in v, the Cauchy characteristics form a well-defined distribution A(I) ⊂ T Σ
given by
A(I)x := {v ∈ Tx Σ | v I ⊂ I}, x ∈ Σ.
The dimension of A(I) is upper semicontinuous on Σ (i.e., in an open neigh-
borhood of a point, it can only decrease), and so we may (and will) restrict
our attention to an open set in Σ on which it is constant. The dual of A(I)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 182 7. Applications to PDE

is a Pfaffian system known as the retracting space or Cartan system of I:

C(I)x := {θ ∈ Tx∗ Σ | v θ = 0, ∀ v ∈ A(I)x }.
From Exercise 7.1.6, the Lie bracket of any two sections of A(I) is
another section of A(I). If θ is a section of C(I) and v1 , v2 ∈ Γ(A(I)),
then using the “vector field definition” of the exterior derivative (see Ap-
pendix B) we get dθ(v1 , v2 ) = 0. Thus, C(I) is Frobenius. The Frobe-
nius Theorem implies that there exist local coordinates x1 , . . . , xn such that
C(I) = {dx1 , . . . , dxs }. Then I can be defined using only the variables
x1 , . . . , xs .
In what follows, we let ΛI denote the subalgebra of Ω∗ Σ spanned by
wedge products of sections of I ⊂ T ∗ Σ with each other.
Proposition 7.1.18. The retracting space C(I) is the smallest Frobenius
system such that Λ C(I) contains a set of algebraic generators for I.

Proof. Let J be a Frobenius system such that there exists a finite collection
of forms ψ1 , . . . , ψm ∈ ΛJ which generate I algebraically. In terms of these
generators, any form Φ ∈ I can be written as
m
Φ= βk ∧ ψk .
k=1

Let v be any vector field annihilated by J. Then v Φ = mk=1 (v βk ) ∧ ψk ,
which belongs in I. Hence v ∈ Γ(A(I)). Since v and Φ are arbitrary, we
conclude that C(I) ⊂ J.
Furthermore, the smallest such J cannot be larger than C(I). For, sup-
pose C(I) has codimension l > 0 within J, and let π1 , . . . , πl be linearly
independent forms in J with πk ∈ / C(I). Partition the generators for I into
ϕ1 , . . . , ϕm ∈ ΛC(I), plus generators ψ1 , . . . , ψn involving the πk . If v is any
vector in A(I), and ψ1 has lowest degree among the ψ’s, then v ψ1 must be
expressible in terms of the ϕ’s. In particular, any term in ψ1 involving one
of the πk may be omitted since that term already is generated by the ϕ’s.
It follows that we may take all the generators to be in Λ C(I). 

The above proposition implies that we may calculate the retracting space
simply by obtaining a minimal set of algebraic generators for the ideal.
Exercises 7.1.19:
1. Consider an EDS generated by a single 1-form θ ∈ Ω1 (Σ). Show that
the dimension of C(I) will be 2k + 1, where k is the largest integer such that
θ ∧ (dθ)k |Σ = 0. 
2. [76, §126] Consider the system of two second-order PDE given by uxx +
3 uxy = 0 and uxy uyy = 1. Let I be the pullback of the standard contact
1 3

system to the six-dimensional submanifold Σ ⊂ J 2 (R2 , R) defined by these

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.1. Symmetries and Cauchy characteristics 183

equations. Show that C(I) is five-dimensional, and give a Cauchy charac-

teristic vector field for I. 

Quotient spaces. Assume I is an EDS defined on Σ, and C(I) has constant

rank k. Then the maximal integral manifolds of C(I) form a foliation of Σ
by smooth submanifolds of codimension k. In some cases, the leaf space of
this foliation is a manifold and is referred to as the quotient by the Cauchy
characteristics.
Example (7.1.10 continued). For our tautological EDS for lifts of surfaces,
C(I) is the Frobenius system on F spanned by ω 1 , ω 2 , ω 3 , ω13 , ω23 . Integral
curves of C(I) form a foliation of F . The leaf space for this foliation is a
five-dimensional homogeneous space which we identify with U = U T E3 , the
unit tangent bundle of E3 , by letting x be the basepoint and e3 the unit
vector in Tx R3 .

Whenever a smooth quotient by the Cauchy characteristics exists, the

system I is generated algebraically by the pullback of a well-defined system
on the quotient manifold. This fact is a consequence of the following more
general result:
Proposition 7.1.20. Let π : Σ → Q be a smooth fibration. Let I ⊂ Λp T ∗ Σ
be a subbundle of rank k spanned by p-forms that are semi-basic for π.
Suppose that for any ψ ∈ Γ(I) and any vector field v on Σ that is tangent
to the fibers of π,
(7.8) Lv ψ ∈ Γ(I).
Then there is a well-defined subbundle I ⊂ Λp T ∗ Q such that a p-form ψ is
a section of I if and only if π ∗ ψ ∈ Γ(I).
Let J be the algebra of wedge products of forms that are semi-basic for
π. (Note that J is not an exterior ideal.) Then condition (7.8) is equivalent
to
(7.9) dψ ≡ 0 mod Jp+1 + I ∧ T ∗ Σ,
i.e., dψ is a sum of wedge products of semi-basic forms and wedge products
with sections of I.

Proof of 7.1.20. Let x ∈ Σ, and let U ⊂ Q be an open set containing π(x)

such that σ : U → Σ is a local section of the fibration and N = σ(U ) passes
through x. Let I ⊂ Λp T ∗ Σ|U be the image of I|N under σ ∗ .
To show that I is independent of choices, let v be tangent to the fibers
and let φt : Σ → Σ be the one-parameter family of diffeomorphisms gen-
erated by v. (If necessary, we restrict the domain of φt to some open set
around x, containing N , so that φt is defined for all t in an open interval
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 184 7. Applications to PDE


about t = 0.) Then (7.8) implies that φ∗t ψ ∈ Γ(I), and thus (φt ◦σ)∗ ψ ∈ Γ(I)
for any t. 

Theorem 7.1.21. If I is an EDS on Σ and the leaf space of the Cauchy

characteristics is a smooth manifold Q, then there is an EDS I on Q such
 alg , where π is the projection onto the leaf space.
that I = {π ∗ I}

Proof. Let J = C(I), and let IJp be the bundle spanned by semi-basic p-
forms in I. Then Proposition 7.1.20 implies that IJp is the pullback of the
degree p piece of a differential ideal I on Q (see the exercises below).
 alg ⊂ I. Moreover, since I has
Because Γ(IJp ) ⊂ I p , we have {π ∗ I}
a set of algebraic generators that are semi-basic (see Proposition 7.1.18),
 alg .
I ⊂ {π ∗ I} 

In practice, one exploits the existence of the quotient manifold and its
EDS to obtain a better geometric understanding of the system. However, it
may be easier to continue to calculate using the coordinates or coframing on
the original space. For example, if the original space is a submanifold of a
jet bundle, or a Lie group, then these spaces come equipped with coframes
for which the structure equations are particularly simple.
Exercises 7.1.22:
1. Show that I,  as constructed degree by degree in the proof of Theorem
7.1.21, is actually a differential ideal on Q.
2. Let Σ, π, Q and J be as above. Suppose Q has dimension 2m, and Ω is
a 2-form in J such that Ωm = 0. Under what conditions is Ω the pullback
of a well-defined symplectic form on Q? 

7.2. Second-order PDE and Monge characteristics

In this section we define the characteristic systems associated to hyperbolic
second-order PDE. Characteristic systems are connected to the characteris-
tic variety defined in Chapter 5, and will play a key role in our discussion
of Darboux’s method. (They should not be confused with Cauchy charac-
teristics, defined earlier.) We will also generalize these systems to a class of
EDS known as hyperbolic exterior differential systems.
In Example 6.5.3, we calculated the tableau (6.16) for the rank three
Pfaffian system I generated by the restrictions of the standard contact forms
θ0 = dz − p dx − q dy,
θ1 = dp − r dx − s dy,
θ2 = dq − s dx − t dy
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.2. Second-order PDE and Monge characteristics 185

to the hypersurface Σ ⊂ J 2 (R2 , R) defined by a regular second-order PDE

(7.10) F (x, y, z, p, q, r, s, t) = 0.
Applying Cartan’s Test to the tableau shows that such systems are involu-
tive with solutions depending locally on s1 = 2 functions of one variable.
Although there is usually no explicit way to describe the solutions in terms
of these functions, we will see several examples where such a description
exists when we discuss Darboux’s method.
In the continuation of Example 6.5.3 in §6.7, we saw that a nonzero
covector ξ = ξ1 dx + ξ2 dy belongs to the characteristic variety of I if and
only if
(7.11) Fr ξ12 + Fs ξ1 ξ2 + Ft ξ22 = 0.
Thus, the characteristic directions are the null lines for this quadratic form.

Definition 7.2.1. The PDE (7.10) is said to be elliptic, hyperbolic or par-

abolic according to whether the determinant Fr Ft − 14 Fs2 of the quadratic
form is positive, negative or zero, respectively (cf. [105]).

For the rest of this discussion we will assume the determinant is nonzero
at every point of Σ. In fact, we will assume the PDE is hyperbolic; however,
everything we do may be carried out in the elliptic case, by using complex-
valued forms (see §7.4).
Hyperbolic PDE and characteristic systems. Note that (7.11) implies that a
linear combination of the system 2-forms is congruent, modulo I, to a de-
composable 2-form with ξ as one of the factors. Therefore, when the PDE is
hyperbolic, there are two linearly independent decomposable 2-forms in the
ideal that are independent of the θ’s. We may choose linearly independent
forms π1 , π2 , ω1 , ω2 that are not in I 1 , such that π1 ∧ω1 and π2 ∧ω2 are these
decomposable forms in I 2 , and ω1 ∧ ω2 = 0 is the independence condition.
In order to simplify the tableau, we may also choose new forms θ̃1 , θ̃2 in I 1
so that
(7.12) dθ̃1 ≡ −π1 ∧ ω1 , dθ̃2 ≡ −π2 ∧ ω2 mod I.
(However, this is not immediately necessary.)
At each point of an integral surface for I, the tangent vectors to the
surface that are annihilated by ω1 or ω2 are characteristic in the sense defined
above. Moreover, since each πi pulls back to the surface to be a multiple of
ωi , each such vector is annihilated by all of the forms of one of the following
two characteristic systems:
M1 := {θ0 , θ1 , θ2 , π1 , ω1 }, M2 := {θ0 , θ1 , θ2 , π2 , ω2 }.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 186 7. Applications to PDE

Each integral surface is foliated by integral curves of M1 and by integral

curves of M2 . In order to distinguish them from Cauchy characteristics,
these curves are sometimes called Monge characteristics [175].

Remark 7.2.2. For parabolic equations, there is only one decomposable 2-

form and only one characteristic system. In classical terminology, one says
that in this case the Monge characteristics are confounded.

Remark 7.2.3. The characteristic systems are diffeomorphism invariants

of the exterior differential system (Σ, I). For, suppose two Pfaffian systems
(M, I) and (N, J), both arising from regular second-order PDE for one func-
tion of two variables, are contact-equivalent. (This means that there is a
diffeomorphism φ : M → N such that φ∗ J = I, and such φ respects the
independence conditions.) Then φ takes integral planes to integral planes,
and tableau to tableau.

Exercises 7.2.4:
1. Assume that Fr = 0. Using (7.11), show that ξ = dy − m dx belongs to
the characteristic variety if and only if m is a root of the quadratic equation

(7.13) Fr m2 − Fs m + Ft = 0.

Assuming that this equation has distinct roots m1 , m2 , show that the de-
composable 2-forms in I are

(dy − m1 dx) ∧ (π2 + m2 π3 ), (dy − m2 dx) ∧ (π2 + m1 π3 ),

where π1 , π2 , π3 are as in the structure equations (6.14), (6.15).

2. For the hyperbolic PDE s = pq (i.e., zxy = zx zy ), show that the charac-
teristic systems are

M1 = {θ0 , θ1 , θ2 , dr − q(r + p2 )dy, dx},

M2 = {θ0 , θ1 , θ2 , dt − p(t + q 2 )dx, dy}.

Remark 7.2.5. Note that if the PDE (6.13) is quasilinear, then the equa-
tion (7.13) defining the characteristics only involves the variables x, y, z, p, q.
More generally, we say that a second-order PDE for one function of two
variables has first-order characteristics if there exists a rank three Pfaffian
system on J 1 (R2 , R) whose pullback to Σ is contained in one of the char-
acteristic systems of I. This is the case when (6.13) is a Monge-Ampère
equation (see §7.4); for other examples, see [76, Ch. IV]. The existence of
first-order characteristics is intrinsic to the system, since J 1 (R2 , R) is the
quotient of Σ by the foliation dual to the retracting space of I’s first derived
system (see §7.3).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.2. Second-order PDE and Monge characteristics 187

Hyperbolic exterior differential systems. We now generalize the notion of

Monge characteristics to systems which have structure equations similar to
(7.12).
Definition 7.2.6 ([30]). An EDS is hyperbolic of class k if near any point it
is algebraically generated by k independent 1-forms and two decomposable
2-forms with no common factors.

Thus, the retracting space of such a system has dimension k + 4. For

this reason, a hyperbolic EDS of class k may be assumed to be defined on a
(k + 4)-dimensional manifold.
This generalization will be useful for studying hyperbolic PDE because
the prolongation of a hyperbolic system of class k is hyperbolic of class
k + 2. We show this in the special case of systems arising from second-order
hyperbolic PDE, which have class k = 3:
Example 7.2.7. Let I be the rank three Pfaffian system on Σ7 ⊂ J 2 (R2 , R)
corresponding to a second-order hyperbolic equation, as defined in the first
part of this section. Any integral element satisfying the independence con-
dition is annihilated by
θ3 := π1 − h1 ω1 , θ4 := π2 − h2 ω2
for some parameters h1 , h2 . Then these hi can be taken as coordinates along
fibers of the bundle Σ = V2 (I, Ω) of integral planes. The prolongation of I
is the Pfaffian system on Σ generated by I  = {θ0 , . . . , θ4 }. (Again, we omit
pullbacks from the notation.)
The structure equations of the prolongation are
dθ0 ≡ 0, dθ1 ≡ 0, dθ2 ≡ 0, dθ3 ≡ −π3 ∧ ω1 , dθ4 ≡ −π4 ∧ ω2 mod I  ,
where π3 := dh1 + (A1 − h1 B1 )ω2 and π4 := dh2 + (A2 − h2 B2 )ω1 , and Aj , Bj
are functions on Σ such that dπj ≡ Aj ω1 ∧ ω2 and dωj ≡ Bj ω1 ∧ ω2 modulo
I .
Notice that these structure equations have the same form as the struc-
ture equations (7.12) for I. Therefore, I  is a hyperbolic EDS of class
five. We define the characteristic systems of the prolongation as M1 =
{θ0 , . . . , θ4 , π3 , ω1 } and M2 = {θ0 , . . . , θ4 , π4 , ω2 }. Each characteristic sys-
tem of I pulls back to lie inside one of the characteristic systems of the
prolongation (see Example 7.3.16 below).

In general, for a hyperbolic EDS of class k we define two characteristic

systems of rank k + 2 in the same manner as in the example, by adjoining
to the 1-forms of the system the factors of one of the two decomposables.
Hyperbolic systems of even class arise when we consider systems of equa-
tions for two functions of two variables:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 188 7. Applications to PDE

Exercise 7.2.8: Show that a system of the form ux = f (v), vy = g(u), when
converted to an EDS on a codimension-two submanifold of J 1 (R2 , R2 ), gives
a hyperbolic EDS of class two. Show that it may also be encoded by a
hyperbolic EDS of class zero (i.e., generated by a pair of decomposable 2-
forms) defined on J 0 (R2 , R2 ). 

7.3. Derived systems and the method of Darboux

We now turn to Darboux’s method, which gives a recipe for solving certain
second-order PDE using ODE techniques. If the equation passes a certain
test, then we may specify two arbitrary functions of one variable, and then
construct solutions by solving systems of ODE. (In fact, for some equations
Darboux’s method can even lead us to a formula giving all solutions in terms
of two arbitrary functions.) The test is carried out by calculating the derived
flag for each of the characteristic systems defined in the previous section.
Accordingly, we first need to discuss such constructions:

Derived systems. One argument for using exterior differential systems to

study PDE is that it guides one to a good framing that highlights the geo-
metric aspects of the system in question. For a Pfaffian system, computing
the derived systems provides one way to adapt the coframe to the geometry
of the system.
Let I ⊂ T ∗ Σ be a Pfaffian system of constant rank. The derived system
I (1)of I is the subbundle spanned by forms in I whose exterior derivatives
are zero modulo I. More explicitly, let δ : Γ(I) → Γ(Λ2 (T ∗ Σ/I)) be de-
fined by δ(θ) = dθ mod I. Because δ is C ∞ (Σ)-linear, it is induced by a
vector bundle homomorphism from I to Λ2 (T ∗ Σ/I). Since the coefficients
of δ, when expressed in matrix form, are smooth, its rank is upper semi-
continuous, so we may restrict to an open set in Σ on which it has constant
rank. On this open set, we define I (1) := ker δ. If I (1) = I then I is Frobe-
nius (cf. Definition 1.3.10); otherwise, I (1) is smaller than I. In that case,
we define a sequence of systems I (k+1) := (I (k) )(1) , until I (N ) either is a
Frobenius system or has rank zero.
Warning. The notation I (k) is used by some authors (e.g., [27]) to denote
the kth prolongation of I. In the rest of this chapter we will use a prime to
denote prolongation when necessary.
The derived flag of I is defined as
I = I (0)  I (1)  . . .  I (N )
and N is called the derived length of I. The dimensions of the systems in the
derived flag are differential invariants and can be used to classify Pfaffian
systems into types.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.3. Derived systems and the method of Darboux 189

Remark 7.3.1. These dimensions satisfy certain inequalities. For example,

if rk = dim I (k) − dim I (k+1) and m = dim C(I) − dim I, then

r0 ≤ 12 m(m − 1),
r1 ≤ mr0 + 12 r0 (r0 − 1),
r2 ≤ r1 (r0 + m) + 12 r1 (r1 − 1),
....

For generic systems these inequalities are equalities; see [66] for further
inequalities and invariants of Pfaffian systems.

Example 7.3.2. On R4 , with coordinates u, v, x, y, let I be spanned by the

1-forms
θ1 := dy − vdx, θ2 := dv + xdu.
The above inequalities imply that the first derived system is at least one-
dimensional. In fact, since dθ1 = dx ∧ dv ≡ x du ∧ dx modulo I and dθ2 =
dx ∧ du, then I (1) is spanned by

ω := θ1 + xθ2 = dy + xdv − vdx + x2 du.

Because dω = 2dx ∧ θ2 , I (1) is not Frobenius. Hence the derived flag termi-
nates in I (2) = 0.

The generic Pfaffian system I of rank two on a four-dimensional manifold

Σ has derived length two. Such systems are said to define an Engel structure
on Σ. Like contact structures and symplectic structures, all Engel structures
are locally diffeomorphic (cf. the Engel Normal Form Theorem in [27]).
Exercise 7.3.3: On R5 , with coordinates u, v, x, y, z, let I be spanned by
the 1-forms

θ1 := dx − xdu, θ2 := dz + udx + vdy, θ3 := dy − udv.

Show that I (1) = {θ1 , θ2 } and that I (2) is Frobenius. 

Remark 7.3.4. By contrast, a generic Pfaffian system of rank three in five

variables has derived length two but contains no Frobenius system. Unlike
Engel structures, such systems are not necessarily locally diffeomorphic to
each other. Their classification is the subject of the “five variables paper”
[36].

Exercise 7.3.5: Suppose I is a hyperbolic EDS of class k, I  is its prolon-

gation, and I  is the Pfaffian system spanned by the 1-forms of I  . Show
that (I  )(1) is spanned by the pullbacks of the 1-forms of I.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 190 7. Applications to PDE

Darboux’s method. Let I be the exterior differential system constructed

in §7.2, encoding a hyperbolic second-order PDE for one function of two
variables. On any integral surface of I satisfying the independence condition,
the Monge characteristic curves form two foliations which are transverse at
each point of the surface. Suppose one characteristic system Mi happens
to contain two linearly independent 1-forms that are exact derivatives:
dW, dX ∈ Mi .
Then W and X are both constant along the integral curves of Mi . (Any such
function, whose differential lies in one of the Mi , is known as a Riemann
invariant for the PDE.) It follows that W and X are functionally related
on the integral surface. If we suppose that {dW, dX} ⊂ Mi is linearly
independent from I, we can assume that, say, dX restricts to be nonzero on
an open set in the integral surface. Then W must be some function of X
on the surface. When both characteristic systems have these properties, the
functions may be arbitrarily specified and used to construct the solution.
Example 7.3.6. For Liouville’s equation zxy = ez , the Pfaffian system I is
defined by
θ0 = dz − pdx − qdy )
dθ1 ≡ −(dr − pe z
dy) ∧ dx
θ1 = dp − rdx − ez dy with mod I.
dθ2 ≡ −(dt − qez dx) ∧ dy
θ2 = dq − e dx − tdy
z

The characteristic system M1 contains two exact derivatives, dx and

d(r− 21 p2 ), the latter obtained by adding multiples of dx and θ1 to dr−pez dy.
Similarly, dy and d(t − 12 q 2 ) are in M2 . Hence,

(7.14) r − 12 p2 = f (x), t − 12 q 2 = g(y)

for some functions that depend on the particular solution of Liouville’s equa-
tion defining our integral surface.
Again, we may use these functions to determine the solution. For, if we
choose f and g arbitrarily, and restrict I to the codimension-two submani-
fold defined by (7.14), then the restriction is a Frobenius system. This means
that we can solve for z(x, y) by solving systems of ODE. In this example,
we can solve the system
(7.15) ∂z/∂x = p, ∂p/∂x = f (x) + 12 p2
and a similar system in the y-direction, and be guaranteed by the Frobenius
condition that we have generated a solution to the original PDE.
Definition 7.3.7. A hyperbolic exterior differential system I is said to be
Darboux-integrable if each characteristic system Mi contains a Frobenius
system Δi of rank two that is independent from I.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.3. Derived systems and the method of Darboux 191

Proposition 7.3.8. Let I be a Darboux-integrable system defined on Σ.

Let W1 , W2 , X1 , X2 be functions defined on an open set U ⊂ Σ such that
dWi , dXi ∈ Δi for i = 1, 2. Let f1 , f2 be arbitrary smooth functions and let
N ⊂ U be the codimension-two submanifold defined by Wi = fi (Xi ). Then
I|N is Frobenius.

Proof. Since Mi = I ⊕ Δi , then on U the image δ(I) is spanned by dW1 ∧

dX1 and dW2 ∧ dX2 . 

Example (7.3.6 continued). We have seen that Liouville’s equation is

Darboux-integrable. Now, to obtain a solution, it seems that we must
solve Riccati differential equations (7.15) for p and q. While this is easy
to do numerically, there is no general formula for representing solutions of
a Riccati equation by quadratures. However, since f (x) is arbitrary, we
may make convenient choices for the form of f (x). In particular, if we set
f (x) = F  (x)− 12 F (x)2 , then F (x) is another solution of the Riccati equation
that p satisfies. It is standard (cf. [172]) that v = 1/(p − F (x)) satisfies
a linear differential equation. Solving that equation using an integrating
factor leads to
X  2X 
p=  − ,
X X +Y
where X and Y are arbitrary functions of x and y respectively, and F (x) =
X  /X  . Integrating gives
2X  Y 
(7.16) z = ln .
(X + Y )2
Proposition 7.3.9. The general solution to Liouville’s equation zxy = ez is
(7.16), where X(x) and Y (y) are arbitrary functions.

Exercise 7.3.10: Fill in the details in the derivation of solution formula

(7.16).

The two extra equations (7.14) which we impose on the solution of Li-
ouville’s equation are compatible with it, in the sense that the restriction of
I to the submanifold defined by them is involutive—in fact, it is Frobenius.
(In classical terminology, any extra equation that defines a submanifold to
which the system restricts to be involutive is called an “integral” for the
given PDE. If the extra equation is of lower order than the given PDE, then
it is an “intermediate integral”, and otherwise it is a “general integral”.)
The Frobenius condition means that the solution is uniquely determined,
up to a choice of constants, by the two arbitrary functions in (7.14). Thus,
we have a concrete realization of what Cartan’s Test predicts for this system,
namely that the solution depends on two functions of one variable.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 192 7. Applications to PDE

Exercise 7.3.11: Prove that every solution of Liouville’s equation arises

from integrating a Frobenius system determined by the equation and (7.14).

Example 7.3.12. For the PDE zxx + 12 zxy

2 = 0, the characteristic equation

(7.13) takes the form m2 − s m = 0. The characteristic systems are

M1 = {θ0 , θ1 , θ2 , dy, ds + sdt}, M2 = {θ0 , θ1 , θ2 , dy − s dx, ds}.
It is clear that M1 contains the rank two Frobenius system {dy, (ds/s)+dt};
so, like in (7.14), we let
t + ln s = ϕ(y)
for an arbitrary function ϕ. Meanwhile, the derived flag of M2 is
(1)
M2 = {θ0 , θ1 , dy − s dx, ds},
(2)
M2 = {dy − s dx, dp − 12 s2 dx, ds},
terminating in a Frobenius system of rank three. Using s as a characteristic
coordinate, we let
y − sx = ψ(s)
for an arbitrary function ψ. We may use these equations to determine x and
t in terms of s and y. Then, since θ0 , θ1 , θ2 span a Frobenius system, we can
determine p, q and z by integration. For example,
dp = rdx + sdy = d(rx + sy) − x dr − y ds = d(rx + sy) − ψ(s)ds
gives
s
p = rx + sy − Ψ(s) = (y + Ψ (s)) − Ψ(s),
2
where Ψ is an antiderivative of ψ.
Exercise 7.3.13: Determine q and z in terms of y and s in this example.


Remark 7.3.14. Just as the above PDE (abbreviated in classical notation

as r + 12 s2 = 0) is Darboux-integrable, so is any equation of the form r +
f (s) = 0 [76, §154].

Exercises 7.3.15:
1. Show that r − qs + pt = 0 is Darboux-integrable.
2. Determine for which functions f (v) and g(u) the system of Exercise 7.2.8
is Darboux-integrable. 
3. Let m1 = m2 be the roots of the characteristic equation (7.13) for a
PDE of the form r + f (s, t) = 0.
(a) Show that each characteristic system contains one of the 1-forms ds +
mi dt, and that both of these forms are integrable.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.3. Derived systems and the method of Darboux 193

(b) Show that the equation is Darboux-integrable if m1 and m2 both satisfy

the inviscid Burger’s equation ∂m/∂t = m ∂m/∂s. Express this condition
as an overdetermined system of PDE for f (s, t), and investigate the corre-
sponding EDS.

Note that Definition 7.3.7 applies to hyperbolic systems of arbitrary

rank. For example, a second-order PDE for z(x, y) may become Darboux-
integrable after prolonging k times. Then, as before, the Pfaffian system
becomes a Frobenius system of rank 2k + 3 when restricted to submanifolds
of the form U = f (X), V = g(Y ), where {dU, dX} and {dV, dY } are the
rank two systems described in the definition.
Example 7.3.16 (Darboux integrability after one prolongation). Consider
the PDE s = pz (i.e., zxy = zx z). The characteristic systems are
M1 = {θ0 , θ1 , θ2 , dx, dr − (p2 + rz)dy},
M2 = {θ0 , θ1 , θ2 , dy, dt − p(z 2 + q)dx}.
Since θ2 − zθ0 = d(q − 12 z 2 ) modulo dy, M2 contains a rank two Frobenius
system. However, the derived flag of M1 terminates in dx; to continue
checking for Darboux integrability, we must prolong.
Because the mixed partials of z can be expressed in terms of lower deriva-
tives, for higher prolongations we need only add variables pk = ∂ k z/∂xk and
qk = ∂ k z/∂y k for k > 2. For the first prolongation we adjoin 1-forms
θ3 = dr − (p2 + rz)dy − p3 dx,
θ4 = dt − p(z 2 + q)dx − q3 dy.
Since the dy-characteristic system contains the pullback of M2 from above,
we already know that it contains a rank two Frobenius system. Meanwhile,
the dx-characteristic system is
M1 = {θ0 , . . . , θ4 , dx, dp3 − (3rp + zp3 )dy}.
The derived flag of M1 ends, after five steps, with the Frobenius system
+    )
(5) p3 3 r 2
M1 = dx, d − .
p 2 p
Hence we may impose the equations
 2
p3 3 r
(7.17) q− 1 2
2z = ϕ(y), − = ψ(x)
p 2 p
for arbitrary functions ϕ and ψ. Then θ0 , θ1 , θ2 , θ3 restrict to give a Fro-
benius system on the submanifold defined by these equations. (The full
prolongation system θ0 , . . . , θ4 becomes integrable if we add the condition
t − zq = ϕ (y), obtained by differentiating (7.17).)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 194 7. Applications to PDE

Exercises 7.3.17:
1. Show that s = pq/(x − y)n is Darboux-integrable when n = 0, and
Darboux-integrable after one prolongation when n = 1.
2. Show that s+n(n−1)z/(x−y)2 , where n is a positive integer, is Darboux-
integrable only after n prolongations [54, IV, Ch. 3].
Remark 7.3.18. The problem of classifying which second-order PDE be-
come Darboux-integrable after a finite number of prolongations has a long
history and is still a subject of current research [7, 108].
One of the earliest results in this area is Lie’s proof that the only equa-
tions of the form s = f (u) that are Darboux-integrable after a finite number
of prolongations are those with f (u) = exp(au + b); see [76, §170] for a
proof. Another significant early result was Goursat’s classification [77] of
quasilinear equations s = f (x, y, u, p, q) that are Darboux-integrable with-
out prolongation. This classification was later extended, using symmetry
methods, by E. Vessiot; see [182] for more details.
Recently, some progress has been made on the related problem of de-
termining which systems of even class are Darboux-integrable [30]. As in
Exercise 7.2.8, such systems can arise from a system of first-order PDE for
two functions u(x, y), v(x, y) and its prolongations.
Remark 7.3.19 (Semi-integrability). If a rank-two Frobenius system is
present in only one of the two characteristic systems of a hyperbolic EDS,
it is said to be semi-integrable by the method of Darboux. In this case,
one can still construct solutions for the system using ODE techniques alone.
For simplicity, we will explain how this is done for a single second-order
hyperbolic PDE (cf. [182, Prop. 3.3]):
For dW, dX ∈ M1 , the imposed equation W = f (X) defines a six-
dimensional submanifold N ⊂ Σ on which the pullback of I has a Cauchy
characteristic vector field annihilated by M2 . Thus, we can construct an
integral surface of I by starting with an integral curve K transverse to
these Cauchy characteristics, and taking the union of the characteristics
through K. Constructing these characteristics involves solving a system
of first-order ODE for four unknowns which, along with the transversality
condition, depend on the choice of f .
Remark 7.3.20 (Darboux for parabolic PDE). Suppose that, for a para-
bolic second-order equation, the single characteristic system M contains a
rank two Frobenius system. Then M itself is Frobenius [175, §16.1]. The
contact system contains a well-defined rank two Pfaffian subsystem J, not
containing θ0 , such that M = C(J). By Theorem 7.1.21, J is well-defined
on the quotient manifold Q by the leaves of M. One can construct integral
surfaces of I by starting with an integral curve N ⊂ Q of J, and then writing
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.4. Monge-Ampère systems and Weingarten surfaces 195

θ0 as a contact form on the inverse image of N , which is three-dimensional.

Thus, integral surfaces may again be constructed by ODE methods.

We will return to the study of Darboux-integrable systems in Chapter

10, where we discuss the work of Anderson, Fels and Vassiliou establishing
the existence of superposition formulas for these systems.

7.4. Monge-Ampère systems and Weingarten surfaces

In this section we will study a class of second-order PDE’s for which one
can define an EDS on a smaller-dimensional manifold instead of the usual
seven-dimensional manifold of §7.2. These equations, called Monge-Ampère
equations, are of the form
(7.18) Azxx + 2Bzxy + Czyy + D + E(zxx zyy − zxy
2
) = 0,
where A, B, C, D, E are functions of x, y, z, zx , zy . (As in §7.2, we assume
that the partial derivatives of the left-hand side with respect to zxx , zxy and
zyy are never simultaneously zero.) Later in this section, we also will examine
a geometrically natural class of surfaces, called linear Weingarten surfaces,
which are locally equivalent to solutions of a Monge-Ampère differential
equation.
Exercise 7.4.1: Show that the PDE (7.18) is elliptic or hyperbolic, in the
sense of §7.2, if AC − B 2 − DE is positive or negative, respectively.

Let θ = dz − p dx − q dy, the contact form on J 1 (R2 , R). On an integral

surface of θ, satisfying the independence condition dx ∧ dy = 0, we have
p = ∂z/∂x and q = ∂z/∂y. Then z(x, y) satisfies (7.18) if and only if the
surface is also an integral of the 2-form

(7.19) Ψ := A dp ∧ dy + B(dq ∧ dy − dp ∧ dx)

− C dq ∧ dx + D dx ∧ dy + E dp ∧ dq.
Therefore, we may encode (7.18) by the EDS I = {θ, Ψ}diff . (Note that this
is not the only possible choice of generators; since dθ = −(dp ∧ dx + dq ∧ dy),
the B-term in (7.19) may be replaced by 2Bdq ∧ dy or by −2Bdp ∧ dx.)
The prolongation of I gives the usual rank three Pfaffian system described
in §7.2.
Remark 7.4.2. Similarly, a quasi-linear third-order evolution equation for
one function of two variables may be encoded by an EDS on a manifold of
lower dimension than J 3 (R2 , R); see Example 7.5.6.

The following definition captures the essential features of our system I

on J 1 (R2 , R):
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 196 7. Applications to PDE

Definition 7.4.3. A Monge-Ampère system on a five-dimensional manifold

Σ is an exterior differential system which is generated differentially by a
1-form θ of Pfaff rank two and a 2-form Ψ that is linearly independent from
dθ, modulo θ.
Exercises 7.4.4:
1. Show that dΨ ≡ 0 modulo θ, dθ and Ψ. 
2. Show that if (7.18) is hyperbolic in the sense of Definition 7.2.1, then the
corresponding Monge-Ampère system I is a hyperbolic EDS of class one.
Example 7.4.5. Laplace’s equation zxx + zyy = 0 is equivalent to the
Monge-Ampère system on J 1 (R2 , R) generated by θ and Ψ = dp∧dy−dq∧dx.
Example 7.4.6. The sine-Gordon equation
(7.20) uxy = sin u cos u
is equivalent to the Monge-Ampère system on J 1 (R2 , R) generated by θ
and Ψ = (dp − sin u cos u dy) ∧ dx. (The usual version of the sine-Gordon
equation, uxy = sin u, is equivalent to (7.20) by doubling u.) Note that
Ψ + dθ is also decomposable.

It is no accident that these examples of Monge-Ampère systems are

derived from Monge-Ampère equations. In fact, any Monge-Ampère system
is locally equivalent to one generated by such an equation:
Proposition 7.4.7. Let Σ5 carry a Monge-Ampère system I. In a neigh-
borhood of any point in Σ, there are local coordinates x, y, z, p, q such that
I is generated by the form θ = dz − p dx − q dy and a 2-form of the form
(7.19) for some functions A, B, C, D, E.
Exercise 7.4.8: Prove this proposition. 
Example 7.4.9. Using the coordinate expression (1.3) for the mean curva-
ture of the graph of z(x, y), we see that the Monge-Ampère system corre-
sponding to H = 0 has
Ψ = (1 + q 2 )dp ∧ dy + pq(dp ∧ dx − dq ∧ dy) − (1 + p2 )dq ∧ dx.
Similarly, the PDE (1.4) for graphs with Gauss curvature K = 1 can be
encoded by a Monge-Ampère system with
Ψ = dp ∧ dq − (1 + p2 + q 2 )2 dx ∧ dy.
These conditions on the curvatures H, K of a surface are examples of Wein-
garten equations, and are more naturally modeled by an EDS on the coframe
bundle (see below).

The characteristic systems for a Monge-Ampère equation may be defined

in a way similar to §7.2. Namely, suppose I is hyperbolic, i.e., there exist
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.4. Monge-Ampère systems and Weingarten surfaces 197

two decomposable 2-forms in I which are linearly independent modulo θ,

and such that I = {θ, ω1 ∧ π1 , ω2 ∧ π2 }alg . Then their respective factors form
the two characteristic systems Mi = {θ, ωi , πi }. As explained in §7.2, these
pull back to lie in the characteristic systems of the prolongation.
Exercise 7.4.10: Find the characteristic systems for the Monge-Ampère
system for the equation zxx zyy − zxy2 = −1. Show that, as a hyperbolic

system of class one, it is Darboux-integrable. 

Exercise 7.4.11: Show that if a hyperbolic Monge-Ampère system (M 5 , I)
is Darboux-integrable, then it is locally equivalent to the system defined by
the wave equation zxy = 0, by following these steps:
(a) On a neighborhood of any point in M there is a coframe θ, ω1 , ω2 , π1 , π2
(1) (1)
such that θ ∈ I 1 , M1 = {ω1 , π1 }, M2 = {ω2 , π2 }, and

dθ ≡ ω1 ∧ π1 + ω2 ∧ π2 mod θ.

(b) There are 1-forms β0 , β1 , β2 satisfying the structure equations

dθ = −β0 ∧ θ + ω1 ∧ π1 + ω2 ∧ π2 ,
d(ω1 ∧ π1 ) = −β1 ∧ ω1 ∧ π1 ,
d(ω2 ∧ π2 ) = −β2 ∧ ω2 ∧ π2 .

(c) By differentiating the first structure equation, show that we can assume
β1 = β0 − aθ, β2 = β0 + aθ for some function a. 
(d) By differentiating the other equations, show that a = 0 identically, and
conclude that dβ0 = 0. 
(e) On a possibly smaller neighborhood, there exist functions λ, p, q, x, y, z
such that β0 = dλ, eλ ω1 ∧ π1 = dx ∧ dp, eλ ω2 ∧ π2 = dy ∧ dq, and eλ θ =
dz − p dx − q dy.

Remark 7.4.12. The method of Darboux, when applied to a Monge-

Ampère equation which is semi-integrable (see Remark 7.3.19) is known
as Monge’s method [62, Ch. XVI]; such equations are sometimes referred
to as Monge-integrable.

Exercise 7.4.13: Determine which Monge-Ampère equations of the form

zxy = f (x, y, z, zy ) are integrable by Monge’s method. 

Linear Weingarten surfaces. Recall from Example 6.8.2 that a linear

Weingarten equation is relation AK + 2BH + C = 0 involving the Gauss
and mean curvatures of a surface in Euclidean space. A linear Weingar-
ten surface is a surface whose curvatures satisfy such a relation. (More
generally, a Weingarten surface is a surface whose curvatures satisfy some
prescribed equation F (H, K) = 0.)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 198 7. Applications to PDE

Suppose S ⊂ E3 is a smooth surface satisfying a linear Weingarten

equation, and f : S → F is a first-order adapted framing along S. Then the
framing gives an integral surface for the 1-form ω 3 on F , and for the 2-form
Ψ := Aω13 ∧ ω23 + B(ω13 ∧ ω 2 − ω23 ∧ ω 1 ) + Cω 1 ∧ ω 2 .
Exercise 7.4.14: Verify that f ∗ Ψ = 0, using the formulas (2.4) for the
pullbacks of the ωi3 .

The Weingarten equation is thus equivalent to an EDS I on F generated

by Ψ, ω 3 and
Θ := −dω 3 = ω13 ∧ ω 1 + ω23 ∧ ω 2 .
(The linear Pfaffian system described in Example 6.8.2, which encodes the
same linear Weingarten equation, is the prolongation of this EDS.)
As in Example 7.1.10, the retracting space for I is spanned by
{ω 1 , ω 2 , ω 3 , ω31 , ω32 }, and the Cauchy characteristic curves are the fibers of
the submersion from F to the five-dimensional manifold which is the unit
tangent bundle U = U T E3 :
F
@R
@
U
?
E3
According to Theorem 7.1.21, there is a well-defined EDS on U which
pulls back to I and is of Monge-Ampère type. (In fact, the structure equa-
tions of F imply that ω 3 , dω 3 and all three terms of Ψ are pullbacks of
well-defined forms on U .)
Exercise 7.4.15 (Parallel surfaces): A point in U is determined by the
basepoint x ∈ E3 and the unit vector e3 . Consider the map Fr : U → U
defined by (x, e3 ) → (x + re3 , e3 ).
(a) Show that Fr is a contact transformation, i.e., that Fr∗ ω 3 is a multiple
of ω 3 . 
(b) Suppose that M ⊂ U is an integral surface of ω 3 and of Ψ = ω13 ∧ ω23 −
r2 ω 1 ∧ ω 2 and satisfies the independence condition ω 1 ∧ ω 2 = 0. Then M is
a lift of a surface of constant Gauss curvature K = 1/r2 in E3 .
Show that Fr (M ) is a lift of a surface of constant mean curvature H =
1/(2r) in E3 . 

Thus, we recover the theorem of Bonnet, to the effect that each surface
of constant positive Gauss curvature is parallel to two surfaces of constant
mean curvature.

It follows that solutions of the PDE (1.4) for surfaces of constant positive
Gauss curvature may be obtained by a contact transformation from solutions
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.4. Monge-Ampère systems and Weingarten surfaces 199

of the Monge-Ampère system for constant mean curvature (CMC) surfaces.

Moreover, since (as we will see below) any CMC surface admits a family of
noncongruent isometric deformations, similar deformations are available for
surfaces of constant positive Gauss curvature.
Exercises 7.4.16:
1. Show that I is elliptic or hyperbolic according to whether B 2 − AC is
positive or negative, respectively. 
2. Show that linear Weingarten equations for surfaces in a three-
dimensional space form are also equivalent to Monge-Ampère systems.
Which are hyperbolic? 

We finish this section with three extended examples that show how char-
acteristics may be interpreted geometrically, and how to work with them
when the system is elliptic.
Example 7.4.17 (Pseudospherical surfaces and the sine-Gordon equation).
A pseudospherical surface is a surface S ⊂ E3 with Gauss curvature K ≡ −1.
Setting A = C = 1 and B = 0 in the above EDS shows that, in this case,
the 2-forms of I are spanned by the decomposable forms
Ψ + Θ = (ω13 − ω 2 ) ∧ (ω23 + ω 1 ), Ψ − Θ = (ω13 + ω 2 ) ∧ (ω23 − ω 1 ),
and the system is hyperbolic. Moreover, we may use the factors of the
decomposable forms to express the second fundamental form as
II = [ 12 (ω13 + ω 2 )  (ω23 + ω 1 ) − 12 (ω13 − ω 2 )  (ω23 − ω 1 )] ⊗ e3 .
This shows that each characteristic curve projects to be an asymptotic line
on S (see §2.7).
On an open subset of a pseudospherical surface that is free of umbilic
points, we may choose a Darboux framing (see §2.3). This breaks the Cauchy
characteristic symmetry of the EDS, but allows us to see how pseudospher-
ical surfaces are connected to solutions of the sine-Gordon equation (7.20).

Because we are using Darboux frames,

 3    1
ω1 tan u 0 ω
= ,
ω23 0 − cot u ω2
where u ∈ (0, π/2) is one-half of the angle between the asymptotic lines.
We adjoin u as a new variable, and enlarge our EDS to a Pfaffian system
I = {ω 3 , ω13 − (tan u)ω 1 , ω23 + (cot u)ω 2 } on F × R. Then a Darboux framing
along a pseudospherical surface corresponds to an integral surface of I =
{I}diff satisfying the independence condition ω 1 ∧ ω 2 = 0, and vice versa.
The vanishing of the exterior derivatives modulo I of the last two 1-forms
of I implies that ω 1 / cos u and ω 2 / sin u pull back to be closed 1-forms on
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 200 7. Applications to PDE

any integral surface. Hence there exist local coordinates t1 , t2 such that
ω 1 = cos u dt1 and ω 2 = sin u dt2 .
Exercise 7.4.18: Verify that ω 1 / cos u and ω 2 / sin u are closed, and show
that x = (t1 − t2 )/2 and y = (t1 + t2 )/2 are arclength coordinates along the
asymptotic lines. This implies that the asymptotic lines form what is called
a Chebyshev net [55] on the surface.

Substituting du = ux dx + uy dy into the 2-forms of the EDS shows that

ω21= ux dx − uy dy along any integral surface. Then it is easy to see that the
structure equation dω21 = ω13 ∧ ω23 implies that uxy = sin u cos u.
Conversely, we may start with a solution to the sine-Gordon equation
and produce a pseudospherical surface by integration. For, if we let F =
(e1 , e2 , e3 ), then the structure equations dei = ej ωij imply that
(7.21) ⎛ ⎞ ⎛ ⎞
0 ux − sin u 0 −uy − sin u
∂F ∂F
= F ⎝ −ux 0 − cos u⎠ = F ⎝ uy 0 cos u ⎠ .
∂x ∂y
sin u cos u 0 sin u − cos u 0
This is an overdetermined system for the matrix F (x, y), and its integrability
condition is the sine-Gordon equation for u. So, given a solution to sine-
Gordon, we may obtain the framing by solving linear systems of ODE, and
then solve for the surface X(x, y) ∈ R3 by integrating
∂X ∂X
(7.22) = e1 cos u − e2 sin u, = e1 cos u + e2 sin u.
∂x ∂y
For example, the traveling wave solution

(7.23) u = 2 arctan exp(ax + a−1 y) , a = 0,

gives the standard pseudosphere when a2 = 1 and gives Dini’s surface when
a2 = 1 (see [146] for pictures).
Hilbert’s Theorem states that pseudospherical surfaces in Euclidean
space cannot be geodesically complete. Hence, even if the sine-Gordon
solution is defined for all x and y, the corresponding surface will become
singular. In fact, ω 1 ∧ ω 2 = sin 2u dx ∧ dy shows that the surface will be sin-
gular whenever u is a multiple of π/2, i.e., the asymptotic lines are tangent
to each other.
Note that each of x + y and x − y is constant along one of the two sets
of lines of curvature. This observation is useful if we suppose that S is a
pseudospherical surface of revolution.
Exercise 7.4.19: Show that if S is a surface of revolution, the corresponding
sine-Gordon solution can be assumed to be of the form u = f (x + y). Then,
determine all sine-Gordon solutions of this form. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.4. Monge-Ampère systems and Weingarten surfaces 201

The classification of surfaces of revolution of constant Gauss curvature

is due to Minding [178].
Example 7.4.20 (Minimal surfaces and the Weierstrass representation).
Recall that a minimal surface is a surface S ⊂ E3 with mean curvature
zero. Setting A = C = 0 and B = 1 in the EDS for a general linear
Weingarten equation shows that the 2-forms of I are spanned by Θ = −dω 3
and Ψ = ω23 ∧ω 1 −ω13 ∧ω 2 . We will see below that studying the characteristic
systems for I leads us to recover the classical Weierstrass representation
(7.26) for minimal surfaces.
When we attempt to form the characteristic systems, we find that the
decomposability equation
(Θ + λΨ)2 = 0
has roots λ = ±i. So, we need to introduce complex-valued differential
forms. For example, Θ − iΨ = (ω13 − iω23 ) ∧ (ω 1 + iω 2 ), and similarly for the
complex conjugate. If we define
(7.24) τ := ω13 − iω23 , ω := ω 1 + iω 2 ,
then the characteristic systems are M = {ω 3 , τ, ω} and M = {ω 3 , τ , ω}.
When we attempt to apply Darboux’s method, we find that neither of
these systems contains a rank two integrable subsystem. In fact, dτ ≡ 0
modulo τ , but dω ≡ τ ∧ ω 3 modulo ω. However, this shows that Re{τ, τ } =
{ω13 , ω23 } =: J is a Frobenius system on F . The quotient of F by the leaves
of the distribution dual to J is a two-dimensional sphere, and the quotient
map γ : F → S 2 is given by the unit vector e3 . Thus, the restriction of γ
to an integral surface M of I is just the Gauss map of the minimal surface
π(M ).
Because τ drops to be well-defined (up to multiple) on S 2 , we can define
a complex structure on the sphere for which τ spans the (1, 0)-forms. Thus,
we recover the well-known result that the Gauss map of a minimal surface
S is holomorphic with respect to the complex structure on S defined by the
(1, 0)-form ω = ω 1 + iω 2 .
It will be useful to fix a complex coordinate on the sphere. To do this, we
will identify S 2 with the null quadric N ⊂ CP2 defined by z12 + z22 + z32 = 0,
by mapping e3 to the vector z = e1 − ie2 . (This vector is associated to e3
in a unique way, up to multiple, by the requirement that e3 × z = iz.) We
then use stereographic projection
(7.25) z = [1 − w2 , i(1 + w2 ), 2w], w ∈ C,
to define a complex local coordinate w on N . Computing dz and comparing
this with d(e1 − ie2 ) = i(e1 − ie2 )ω21 + e3 τ shows that γ ∗ dw is a multiple of
τ.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 202 7. Applications to PDE

Exercise 7.4.21: Verify that, under our identification, the point (7.25) in N
corresponds to the point in the unit sphere given by standard stereographic
projection 
(x, y) → 2x, 2y, 1 − x2 − y 2 /(1 + x2 + y 2 )
when w = x + iy. Thus, the complex structure on S 2 defined by τ is the
same as that given by the usual stereographic projection from C.

Now suppose z is a local complex coordinate on our minimal surface S.

Then ω = f (z)dz for a nonvanishing holomorphic function f , and γ ∗ w =
g(z) for some holomorphic function g. Notice that the point X on the surface
satisfies 
dX = e1 ω 1 + e2 ω 2 = Re (e1 − ie2 )(ω 1 + iω 2 ) .
By substituting the expressions for e1 − ie2 = z and ω 1 + iω 2 = ω in terms
of z, we see that

(7.26) X = Re [(1 − g 2 )f, i (1 + g 2 )f, 2f g] dz.

Remark 7.4.22. Multiplying f by a unit modulus constant eit will produce

a one-parameter family of isometric minimal surfaces parametrized by t ∈
[0, 2π). The catenoid and helicoid mentioned in §2.2 are members of one
such family.

The following exercises show how some well-known minimal surfaces can
be obtained from (7.26) for simple choices of f and g. (More generally, g(z)
can be a meromorphic function whose domain has nontrivial topology, and
this allows the resulting surface to have nontrivial topology itself.)
Exercises 7.4.23:
1. Identify the surfaces obtained from the Weierstrass representation using
f (z) = 1 and g(z) = 0, g(z) = z, g(z) = 1/z and g(z) = i/z, respectively. 
2. Show that the parametrization (7.26) is regular provided the poles of g
coincide with (and are at most half the order of) the zeros of f .
3. Conclude that the surface obtained using g(z) = tan(z/2) and f (z) =
cos2 (z/2) is regular, and identify it with one of the surfaces in question #1.
(Thus, the Weierstrass representation of a minimal surface is not unique.)


Example 7.4.24 (Surfaces of constant mean curvature). Setting A = 0,

B = 1 and C = −2H (where H is some constant) in the general EDS for
linear Weingarten surfaces leads to a Monge-Ampère system with 2-forms
Θ = −dω 3 and Ψ := ω13 ∧ ω 2 + ω 1 ∧ ω23 − 2Hω 1 ∧ ω 2 . As with the minimal
surface system, the characteristics are complex. The decomposable 2-forms
are Θ + iΨ = (τ − Hω) ∧ ω and its complex conjugate, where τ and ω are
defined by (7.24).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.4. Monge-Ampère systems and Weingarten surfaces 203

This complex coframe, adapted as it is to the characteristics, leads to

a simple basis for the EDS and a simple description for its prolongation.
Namely, we adjoin a complex variable Q and define the 1-form
(7.27) θ := τ − Hω − Qω.
The prolongation is the Pfaffian system on F × C generated by ω 3 , θ and θ.
The system 2-forms are
dθ ≡ −(dQ + 2iQω21 ) ∧ ω mod ω 3 , θ, θ
and its complex conjugate.
Note that Q gives the components of the traceless part of the second
fundamental form. For, θ = 0 implies that
 3    1
ω1 H + q1 −q2 ω
= ,
ω23 −q2 H − q1 ω2
where Q = q1 + iq2 . Thus, the zeros of Q are the umbilic points of the
surface.
Proposition 7.4.25. Let M be an oriented surface with a Riemannian met-
ric, and let Fon (M ) be its oriented orthonormal frame bundle, with canonical
forms η 1 , η 2 and connection form η21 (see §2.6). Endow M with the complex
structure which it inherits from the metric and the orientation (see Appendix
C). Suppose Q is a complex-valued function on Fon (M ) that satisfies
(7.28) (dQ + 2iQη21 ) ∧ (η 1 + iη 2 ) = 0.
Then σ = Q(η 1 + iη 2 )2 is a well-defined holomorphic differential on M .
Moreover, for any constant H, Q determines a local isometric embedding
of M as a surface of mean curvature H in R3 which is unique up to rigid
motions.
Note that σ is the Hopf differential associated to the Gauss map of
M → R3 , which turns out to be a harmonic map into S 2 [147].

Proof. To see that the differential σ is well-defined, let v be the vector field
tangent to the fibers of Fon that is dual to η21 . Then it is easy to check
that Lv σ = 0. To see that it is holomorphic, let σ = P (dz)2 , where z is a
local complex coordinate such that dz is a multiple of η 1 + iη 2 . Then (7.28)
implies that ∂P/∂ z̄ = 0.
As explained in §6.4, any isometric immersion of M into R3 will induce
an immersion from Fon (M ) to FE3 , and the graph of this map will be a
three-dimensional integral manifold of the Pfaffian system on Fon (M ) × FE3
generated by
{ω 3 , ω 1 − η 1 , ω 2 − η 2 , ω21 − η21 }.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 204 7. Applications to PDE

To this system we adjoin the form θ (defined by (7.27)) and its conjugate,
where Q is now given. Our hypothesis about Q implies that this enlarged
system is Frobenius. Thus, there exists a unique integral 3-fold through each
point. Since the left action of FE3 ∼
= ASO(3) on itself covers rigid motion
in E , any two immersions will differ only by rigid motion.
3 
Corollary 7.4.26 (Bonnet). Any constant mean curvature (CMC) surface
has a circle’s worth of isometric deformations as a CMC surface.

Proof. Given f : M → R3 , let Q be the coefficient of the Hopf differential,

as above. Then let
(7.29) Qt := eit Q
for any fixed value of t ∈ (0, 2π). Then Qt also satisfies the conditions of
Proposition 7.4.25. 

Moreover, if the surface is not totally umbilic, then at most points Q = 0

and the second fundamental form is changed by (7.29). So, the surface
constructed using Qt is, in general, not congruent to the original surface by
rigid motions.
Remark 7.4.27. These deformations are the CMC analogue of the one-
parameter families of isometric minimal surfaces mentioned earlier. Simi-
larly, any pseudospherical surface has a one-parameter family of noncongru-
ent isometric deformations in R3 . These are induced by the obvious scaling
symmetry x → λx, y → λ−1 y of the sine-Gordon equation (7.20), and are
known as Lie transformations of the given surface.
Remark 7.4.28 (Bonnet surfaces). The above deformations of CMC sur-
faces show that we can produce noncongruent surfaces with the same Gauss
and mean curvature (in this case, H is the same constant). Bonnet dis-
covered that, besides CMC surfaces, there is a finite-dimensional family
of surfaces that admit noncongruent isometric deformations preserving H.
These Bonnet surfaces may be characterized by the property that they ad-
mit isothermal coordinates (i.e., coordinates in which II is diagonalized,
or equivalently, Q is real) and in these coordinates 1/Q is harmonic, i.e.,
(1/Q)z z̄ = 0. See [17] and [46] for more information.

7.5. Integrable extensions and Bäcklund transformations

Equations which are integrable by the method of Darboux give one example
where we can construct integral manifolds of an EDS by restricting to a
submanifold where the system becomes Frobenius. Another example arises
when the integrability condition for an overdetermined system of PDE can
be reduced to finding the solution of another PDE. Then any solution of
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.5. Integrable extensions and Bäcklund transformations 205

that PDE enables us to construct (using ODE techniques) a solution of the

system.
Example 7.5.1. Recall from Example 7.4.17 that the compatibility con-
dition for the system (7.21) is that u satisfies the sine-Gordon equation.
Thus, any sine-Gordon solution enables us to construct a solution to the
Monge-Ampère system for pseudospherical surfaces by integrating a Frobe-
nius system.
Example 7.5.2 (Cole-Hopf transformation). Consider the dissipative ver-
sion of Burger’s equation:
(7.30) ut = (ux + u2 )x .
Since this evolution equation is in the form of a conservation law [157], we
can construct a potential v(x, t) such that
vx = u, vt = ux + u2 .
The compatibility condition for this overdetermined system is precisely that
u satisfy (7.30). Hence, given a solution of (7.30), we can construct v(x, t)
by integration. Moreover, if w = ev , then w(x, t) solves the heat equation
wt = wxx . In this sense, finding a solution to the heat equation is an
extension of finding a solution of Burger’s equation.

As in the preceding example, solutions of one PDE may give us solutions

of another PDE, obtained by solving an auxiliary system of ODE. When this
transformation works in both directions, the two PDE are said to be related
by a Bäcklund transformation, which we will define later in the section.
First, we formalize the relationship between the systems seen in the above
examples:
Definition 7.5.3 ([28]). Let I be an EDS on a manifold Σ, let π : B → Σ
be a submersion, and let E be an EDS defined on B. Then E is an integrable
extension of I if there exists a Pfaffian system J ⊂ T ∗ B such that
(7.31) E = {J, π ∗ I}alg
and J is transverse to the fibers of π (i.e., a basis of J pulls back to each
fiber to be a basis for the cotangent space of the fiber). The rank of an
integrable extension is the rank of J, which is the same as the dimension of
the fibers of π : B → Σ.
The condition that E, as defined by (7.31), is a differential ideal is equiv-
alent to
(7.32) dθ ≡ 0 mod J, π ∗ I
for any θ ∈ Γ(J).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 206 7. Applications to PDE

The concept of integrable extensions originated in the work of Estabrook

and Wahlquist [185, 186], who called them prolongation structures. (We
would also use that term, were it not for the possible confusion with the
notion of prolongation for an EDS.) Integrable extensions are also closely
related to coverings [117].
Example (7.5.1 continued). Let Σ = R5 and let I be the Monge-Ampère
system for sine-Gordon given in Example 7.4.6. Let B = Σ × ASO(3). The
equations (7.21) and (7.22) show that a Darboux framing for a pseudospher-
ical surface associated to a sine-Gordon solution u(x, y) is an integral surface
for the Pfaffian system
J := {ω 1 − cos u(dx + dy), ω 2 + sin u(dx − dy), ω21 − ux dx + uy dy,
ω13 − sin u(dx + dy), ω23 + cos u(dx − dy)}.
Then one can check that J satisfies (7.32).

If (B, E) is an integrable extension of rank m and N ⊂ Σ is a p-

dimensional integral manifold of I, then (7.32) implies that the pullback
of J to π −1 (N ) is Frobenius. In this way, an integral manifold of I gives an
m-parameter family of p-dimensional integral manifolds of E. (In Example
7.5.1, the six-parameter family of surfaces corresponding to a single solution
of sine-Gordon are congruent under rigid motions of E3 .)
Example (7.5.2 continued). To show that this example fits the definition,
let Σ = R4 with coordinates x, t, u, p and let
I = {(du − p dx) ∧ dt, du ∧ dx + (dp + 2up dx) ∧ dt}alg .
Let B = Σ × R, with v as the new coordinate, and let J = {dv − u dx − (p +
u2 )dt}. Again, J satisfies (7.32).
Remark 7.5.4. In general, suppose I is an EDS on a manifold Σ and
Φ ∈ I 2 is closed. For instance, in Example 7.5.2 we have
Φ = du ∧ dx + (dp + 2u du) ∧ dt.
Then on an open set U about any given point in Σ there exists a differential
form φ such that dφ = Φ. We may define a rank one integrable extension
of I by introducing a new coordinate y and letting the Pfaffian system J
on U × R be spanned by dy − φ. Because such forms Φ are examples of
conservation laws for an EDS [28], we call this construction an extension
via conservation law.
Remark 7.5.5. One could trivially satisfy Definition 7.5.3 by choosing J to
be Frobenius; such an integrable extension is said to be flat. More generally,
if the derived flag of J terminates in a Frobenius system K of rank k, then E
defines an integrable extension when pulled back to any leaf of the foliation
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.5. Integrable extensions and Bäcklund transformations 207

dual to K. In this case (B, E) is said to be a k-parameter family of integrable

extensions.

Parametric families of integrable extensions are frequently encountered

in the study of completely integrable PDE (i.e., “soliton” equations). We
now describe the integrable extensions of the Korteweg-de Vries (KdV) equa-
tion
(7.33) ut + uxxx + 6uux = 0,
which were first investigated by Estabrook and Wahlquist [185], who en-
coded the equation by an EDS on R5 .
Example 7.5.6 (KdV equation). Let Σ = R5 , with coordinates x, t, u, p, r.
Let
Θ1 := (du − p dx) ∧ dt,
Θ2 := (dp − r dx) ∧ dt,
Θ3 := (dr + 6up dx) ∧ dt − du ∧ dx,
and let I = {Θ1 , Θ2 , Θ3 }diff . Then solutions of (7.33) correspond to integral
surfaces of I satisfying the independence condition dx ∧ dt = 0. (Note that
Σ is a quotient of J 2 (R2 , R), and p, r have their classical meanings as ux and
uxx respectively. One could also use the EDS obtained by pulling back the
standard contact system on J 3 (R2 , R) to the hypersurface defined by (7.33),
but the extra variables are not needed here.)
On B = Σ × R2 , with fiber coordinates (μ, y), let J = {dμ, η}, where
η := dy − (y 2 + u − μ)dx + (r + 2py + 2(u + 2μ)(y 2 + u − μ))dt,
and let E = {J, I}alg .
Exercise 7.5.7: Verify that E is an integrable extension of I.

Since J contains a rank-one Frobenius system, we have a one-parameter

family of extensions, with μ as a parameter. Along an integral surface of E
that satisfies the independence condition, y(x, t) satisfies
(7.34) yt + yxxx + 6(μ − y 2 )yx = 0
for some constant μ. The family (7.34) of PDE may be said to be extensions
of KdV. When μ = 0, (7.34) is a version of another well-known PDE, the
modified KdV equation, and the passage from u(x, t) to y(x, t) is known as
the Miura transformation [164].
The integrable extensions of the KdV equation were classified by Es-
tabrook and Wahlquist [185]. Using a local normal form for the extension,
they showed that each extension corresponds to a finite-dimensional repre-
sentation of a certain infinite-dimensional Lie algebra, known as the KdV
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 208 7. Applications to PDE

prolongation algebra. Moreover, this algebra has a one-parameter family

of finite-dimensional quotients, each of which is isomorphic to a semidirect
product of sl(2, R) with R5 [170]. Representations of the latter algebra
allow us to construct one-parameter families of KdV extensions, the most
well-known of which is the AKNS system [164]:
   
λ u A B
(7.35) ψx = ψ, ψt = ψ,
−1 −λ C −A

where ψ(x, t) ∈ R2 and A = −4λ3 −2uλ−ux , B = −uxx −2λux −4λ2 u−2u2

and C = 4λ2 + 2u. This system, which is central to the solution of (7.33)
by inverse scattering, is sometimes regarded as defining a connection on the
trivial bundle M × R2 , for which (7.33) is the “zero-curvature condition”.
The over-determined system (7.35) of PDE may be encoded by a Pfaffian
system Ê on B̂ = M × R3 , giving an integrable extension of rank three.
Moreover, on the subset of B̂ where ψ2 = 0 we may define a bundle map
f : B̂ → B by letting μ = λ2 and y = ψ1 /ψ2 + λ, so that f ∗ E ⊂ Ê.

Exercise 7.5.8: As the above example suggests, an integrable extension

may sometimes be factored as a sequence (B̂, Ê) → (B, E) → (M, I). Given
(B̂, Ê) → (M, I), determine conditions under which such a factorization
exists. 

Bäcklund transformations. As described earlier, Bäcklund transforma-

tions are a particular form of integrable extension which allows a two-way
transformation of solutions of a PDE to solutions of another PDE or the
same PDE. One of the most well-known of these is the following transfor-
mation, which transforms “old” solutions of the sine-Gordon equation (7.20)
to “new” solutions of the same PDE:

Example 7.5.9.
ux − ūx = λ sin(u + ū),
(7.36) 1
uy + ūy = sin(u − ū).
λ
If u(x, y), ū(x, y) are smooth functions that satisfy this system for some
λ = 0, then differentiating and equating mixed partials implies that u and
ū must be solutions of sine-Gordon. Conversely, given a solution u(x, y) of
(7.20), we can determine ūx and ūy from (7.36), and integrate to get the new
solution ū(x, y) (since we are guaranteed that mixed partials commute). For
example, the traveling wave solution (7.23) can be produced in this way, by
starting with the identically zero solution of (7.20).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.5. Integrable extensions and Bäcklund transformations 209

Bäcklund transformations occur frequently in the differential geometry

of surfaces [165], and for this reason it is useful to have a coordinate-free
definition for them:

Definition 7.5.10. Let Σ1 , Σ2 carry EDS’s I1 , I2 respectively, and let

B ⊂ Σ1 × Σ2 be a submanifold such that the projections πk : B → Σk
give B the structure of a double fibration. Then (B, E) defines a Bäcklund
transformation between I1 and I2 if E is an EDS on B that is an integrable
extension of both I1 and I2 .

(This definition is a generalization of the notion of a Bäcklund map [166].)

We will often define Bäcklund transformations by giving a Pfaffian sys-
tem J ⊂ T ∗ B such that

(7.37a) {J, π1∗ I1 }alg = {J, π2∗ I2 }alg ,

(7.37b) E := {J, πk∗ Ik }alg is a differential ideal.

Note that J need not be transverse to the fibers of π1 or π2 .

Example (7.5.9 continued). Let Σ1 and Σ2 be two copies of J 1 (R2 , R),

where we denote the forms and coordinates on Σ2 with bars. For i = 1, 2 let
the EDS Ii on Σi be a copy of the Monge-Ampère system for the sine-Gordon
equation, described in Example 7.4.6. For any λ = 0, let Bλ ⊂ Σ1 × Σ2 be
defined by x̄ = x, ȳ = y and

p − p̄ = λ sin(u + ū),
1
q + q̄ = sin(u − ū).
λ
Let θ be the standard contact form on J 1 (R2 , R). On Bλ , let J = {π1∗ θ, π2∗ θ}.
Then it is easy to check that (7.37) is satisfied.

Example (7.5.6 continued). Let Σ1 , Σ2 be two copies of R5 , each carrying

a copy of the KdV exterior differential system. (Again, we use barred co-
ordinates on Σ2 .) Define B = Σ1 × R2 as before, and define a submersion
π2 : B → Σ2 by

x̄ = x,
ȳ = y,
ū = 2(μ − y 2 ) − u,
p̄ = 4y(μ − u − y 2 ) − p,
r̄ = −4(u + y 2 − μ)(u + 3y 2 − μ) − 4yp − r.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 210 7. Applications to PDE

Then J, as defined before, gives a Bäcklund transformation between the

KdV equation and itself. One can check that
⎫
π2∗ Θ1 ≡ −π1∗ Θ1 ⎪
⎬
∗ ∗ ∗
π2 Θ2 ≡ −π1 Θ2 − 4yπ1 Θ1 mod J.
∗ ∗ ∗ ∗
⎪
⎭
π2 Θ3 ≡ −π1 Θ3 − 4yπ1 Θ2 − 8(u + 2y − μ)π1 Θ1
2

The presence of an arbitrary parameter in the Bäcklund transformation

for KdV is central to its complete integrability, both in the sense of inverse
scattering and in the Hamiltonian sense. For example, it is shown in [183]
that expanding u+ū as a power series in μ−1 reveals the existence of infinitely
many independent conservation laws for (7.33).

It is not necessary that a Bäcklund transformation take solutions to

solutions for the same PDE:
Example 7.5.11. Let Σ1 ⊂ J 2 (R2 , R) be the submanifold defined by Liou-
ville’s equation zxy = ez , with the standard contact system I1 generated by
θ0 , θ1 , θ2 defined in Example 7.3.6. Use the classical notation p, q, r, s, t for
the jet coordinates restricted to Σ1 . Then (7.14) implies that (r − 12 p2 )y = 0
and (t − 12 q 2 )x = 0 for any solution. Thus, on B = Σ1 × R with co-
ordinate v on the second factor, we may define an extension by letting
J = {θ0 , θ1 , θ2 , η}, where
η := dv − (r − 12 p2 )dx − (t − 12 q 2 )dy.
Because dη ≡ 0 mod I1 , E = {J, I1 }alg is an integrable extension of I1 .
Furthermore, let Σ2 = J 2 (R2 , R) with coordinates x̄, ȳ, ū, p̄, q̄, and let
I2 be the standard Monge-Ampère system for the wave equation ūx̄ȳ = 0.
Define a submersion π2 : B → Σ2 by
x̄ = x, ȳ = y, ū = v, p̄ = r − 12 p2 , q̄ = t − 12 q 2 .
Then E is an integrable extension of I2 (with the role of the transverse
Pfaffian system played by {θ0 , θ1 , θ2 }). Thus, (B, E) defines a Bäcklund
transformation between Liouville’s equation and the wave equation.
Example 7.5.12. Let Σ1 and Σ2 be two copies of the frame bundle F of
Euclidean space E3 , each carrying a copy Ik of the Monge-Ampère system
for pseudospherical surfaces, defined in Example 7.4.17. (The frame vectors
and canonical forms on Σ2 will be distinguished from those on Σ1 by bars.)
The Bäcklund transformation for pseudospherical surfaces arises natu-
rally in a classical context, the study of line congruences in Euclidean space.
By definition, a line congruence is a two-parameter family of lines; asso-
ciated to the congruence are two focal surfaces to which each line in the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.5. Integrable extensions and Bäcklund transformations 211

family is tangent. Leaving degeneracies aside, the lines locally give a 1-to-1
correspondence between points on the focal surfaces.
Theorem (Bäcklund). If the distance λ between corresponding points on
the focal surfaces and the angle ψ between the surface normal at corre-
sponding points are both constant, then the two surfaces have constant
Gauss curvature equal to − sin2 ψ/λ2 .
(We will take λ = sin ψ, so that K = −1.)
The starting point for a proof of Bäcklund’s Theorem (cf. [48]) is adapt-
ing frames (X, e1 , e2 , e3 ) and (X̄, ē1 , ē2 , ē3 ) along the two surfaces so that
e1 = ē1 is tangent to the line connecting corresponding points X and X̄.
For any fixed λ, the graph of the Bäcklund transformation is a 6-dimensional
submanifold of F × F̄ defined by
X̄ = X + λe1 ,
ē1 = e1 ,
(7.38)
ē2 = e2 cos ψ + e3 sin ψ,
ē3 = e3 cos ψ − e2 sin ψ.
However, we will regard (7.38) as defining a 7-dimensional submanifold B ⊂
F × F̄ × R, with λ as the extra coordinate. On B, let J = {ω 3 , ω̄ 3 , dλ}.
Then (7.37) is satisfied, and Bäcklund’s Theorem follows from the fact that
{ω13 ∧ ω23 + ω 1 ∧ ω 2 , ω̄13 ∧ ω̄23 + ω̄ 1 ∧ ω̄ 2 } ≡ {dω 3 , dω̄ 3 } mod J.
Exercises 7.5.13:
1. Differentiate (7.38) to obtain the following relationships between the
pullbacks to B of the canonical forms:
ω̄ 1 = ω 1 + dλ, ω̄32 = ω32 − dψ,
 2   2 
ω̄ cos ψ sin ψ ω − λω21
= ,
ω̄ 3 − sin ψ cos ψ ω 3 − λω31
 1    1
ω̄2 cos ψ sin ψ ω2
= .
ω̄31 − sin ψ cos ψ ω31
Then, use these relationships to verify (7.37).
2. Show that asymptotic lines are taken to asymptotic lines under the
Bäcklund transformation. 
3. Show that if two pseudospherical surfaces M1 , M2 are related by a Bäck-
lund transformation and both are generated by solutions ui (x, y) of the sine-
Gordon equation, then u1 and u2 are related by a Bäcklund transformation
for the same λ-value.
4. ([192]) Consider the system defined by
u √ u 
(7.39) ux = + 2wx , uy = + 2wy .
Licensed to AMS.
x+y x+y
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 212 7. Applications to PDE

Show that this defines a Bäcklund transformation between the wave equation
uxy = 0 and
√ √
wx wy
(7.40) wxy + 2 = 0.
x+y
5. Explain why the Cole-Hopf transformation (Example 7.5.2) is not a
Bäcklund transformation. 
6. Show that existence of a Bäcklund transformation between two systems
is an equivalence relation on exterior differential systems. 

The appearance of Liouville’s equation in both Examples 7.3.6 and 7.5.11

leads one to suspect a connection between Bäcklund transformations and
Darboux integrability. The following theorem bears this out:
Theorem 7.5.14. Let I1 be a hyperbolic EDS of class k, with independence
1 , and let I2 be the Monge-Ampère system on Σ2 = R that
condition, on Σk+4 5

encodes the wave equation uxy = 0, with independence condition dx ∧ dy =

0. If I1 is Darboux-integrable, then given any point s ∈ Σ1 there exists
an open set U ⊂ Σ1 containing s, an open set V ⊂ Σ2 , and a Bäcklund
transformation (B, E) between I1 |U and I2 |V such that dim B = k + 5 and
the pullbacks of the independence conditions are equivalent modulo E.

Proof. First, suppose that I1 is Darboux-integrable and let M1 , M2 be

its characteristic systems. By the Frobenius Theorem, on an open set U
containing s there exist functions x, y, p, q such that Δ1 = {dx, dp} ⊂ M1
and Δ2 = {dy, dq} ⊂ M2 . Each Δi is independent of I1 , so
(7.41) I1 |U = {θ1 , . . . , θk , dp ∧ dx, dq ∧ dy}alg .
Moreover, we may choose x, y, p, q such that dx ∧ dy = 0 is equivalent,
modulo I1 , to the independence condition.
The graph of the four functions x, y, p, q is a submanifold U  ⊂ U × R4 .
Let B = U  × R, with u as the additional coordinate, and let V be the image
of U  ×R under the natural projection onto R5 with coordinates (x, y, p, q, u).
On R5 , let
(7.42) I2 = {du − p dx − q dy, dp ∧ dx, dq ∧ dy}alg .
On B, let the Pfaffian system J be spanned by the pullbacks of the 1-forms of
I1 and the pullback from V of θ0 = du−p dx−q dy. Because dθ0 ≡ 0 mod I1 ,
we see that E = {J, π1∗ I1 } is a differential ideal. Comparing (7.41) and (7.42)
shows that π1∗ I1 ⊂ {J, π2∗ I2 }alg and π2∗ I2 ⊂ {J, π1∗ I1 }alg . Therefore, we have
a Bäcklund transformation. 
Remark 7.5.15. For k = 1, this result is trivial, since I1 is already lo-
cally equivalent to the wave equation (see Exercise 7.4.11). If we consider
the weaker assumption that the first prolongation I1 is Darboux-integrable,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 7.5. Integrable extensions and Bäcklund transformations 213

then I1 is contact equivalent to one of the hyperbolic Monge-Ampère equa-

tions in Goursat-Vessiot classification (see Remark 7.3.18), and a Bäcklund
transformation to the wave equation locally exists for many of these. In fact,
in some cases the transformation can be solved for explicitly using an overde-
termined PDE system; see [51] for details. Obstructions to the existence
of such a Bäcklund transformation for some equations in this classification
have been pointed out by Anderson and Fels [5].
Conversely, one can prove that if a Bäcklund transformation exists be-
tween a hyperbolic Monge-Ampère system I1 and the wave equation, then
I1 is Darboux-integrable after at most one prolongation [51].
Exercises 7.5.16:
1. The above theorem implies that (7.40) is Darboux-integrable; verify this.
2. Goursat [76, §173] remarks that (7.40) is one of the few known Darboux-
integrable equations with a finite-dimensional symmetry group. Verify that
the transformations
ax + b ay − b
(7.43) w → Aw + B, x → , y → −
cx + d cy − d
are symmetries of this equation.
Remark 7.5.17. If we only knew the Bäcklund transformation (7.36) for
the sine-Gordon equation for the value λ = 1, we could still “lift” the Lie
symmetry to the total space B 6 of the transformation in a way that enables
us to recover the entire family. In the same way, the symmetries (7.43) may
be used to embed (7.39) in a one-parameter family of Bäcklund transforma-
tion for (7.40); see [50] for details.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/08

Chapter 8

Cartan-Kähler III: The

General Case

In this chapter we discuss the general Cartan-Kähler Theorem, which guar-

antees the existence of integral manifolds for arbitrary exterior differen-
tial systems in involution. This theorem is a generalization of the Cauchy-
Kowalevski Theorem (see Appendix D), which gives conditions under which
an analytic system of partial differential equations has an analytic solution
(defined in the neighborhood of a given point) satisfying a Cauchy problem,
i.e., initial data for the solution specified along a hypersurface in the domain.
Similarly, the “initial data” for the Cartan-Kähler Theorem is an integral
manifold of dimension n, which we want to extend to an integral manifold
of dimension n + 1. In Cauchy-Kowalevski, the equations are assumed to be
of a special form, which has the feature that no conflicts arise when one dif-
ferentiates them and equates mixed partials. The condition of involutivity
is a generalization of this, guaranteeing that no new integrability conditions
arise when one looks at the equations that higher jets of solutions must
satisfy.
The reader may wonder why one bothers to consider any exterior dif-
ferential systems other than linear Pfaffian systems. For, as remarked in
Chapter 6, the prolongation of any exterior differential system is a linear
Pfaffian system, so theoretically it is sufficient to work with such systems.
However, in practice it is generally better to work on a space of smaller
dimension, if possible. In fact, certain spectacular successes of the EDS
machinery were obtained by cleverly rephrasing systems that were naı̈vely
expressed as linear Pfaffian systems as systems involving generators of higher
degree on a lower-dimensional manifold. One elementary example of this is

215
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 216 8. Cartan-Kähler III: The General Case

the study of linear Weingarten surfaces (see §7.4); a more complex example
is Bryant’s proof of the existence of Riemannian manifolds with holonomy
G2 [23]. See §9.8 for a discussion.
We begin this chapter with a detailed study of the space of integral
elements of an EDS. In §8.2 we give an example that shows how one can
use the Cauchy-Kowalevski Theorem to construct triply orthogonal systems
of surfaces, and this serves as a model for the proof of the full Cartan-
Kähler Theorem in §8.3. In §8.4 we discuss Cartan’s Test, a procedure by
which one can test for involution, and which has already been described in
Chapter 6 for the special case of linear Pfaffian systems. Then in §8.5 we
give a few more examples that illustrate how one applies Cartan’s Test in
the non-Pfaffian case.

8.1. Integral elements and polar spaces

Suppose I is an exterior differential system on Σ; we will assume that I
contains no 0-forms (otherwise, we could restrict to subsets of Σ on which
the 0-forms vanish). Recall from §1.10 that Vn (I)p ⊂ G(n, Tp Σ) denotes
the space of n-dimensional integral elements in Tp Σ and Vn (I) ⊂ Gn (T Σ)
the space of all n-dimensional integral elements. (We will often abbreviate
Vn (I) as Vn when there is no chance of confusion.) In this chapter we will
obtain a criterion that guarantees that a given E ∈ Vn (I)p is tangent to an
integral manifold. We think of this as “extending” the infinitesimal solution
(p, E) to an integral manifold.

Coordinates on Gn (T Σ). To study the equations that define Vn , we use

local coordinates on the Grassmann bundle Gn (T Σ). Given E ∈ Gn (Tp Σ),
there are coordinates x1 , . . . , xn and y 1 , . . . , y s on Σ near p such that E is
spanned by the vectors ∂/∂xi . By continuity, there is a neighborhood of E
in Gn (T Σ) consisting of n-planes E  such that dx1 ∧ · · · ∧ dxn |  = 0. For each
E
 there are numbers pa such that dy a |  = pa dxi |  ; these pa , along with the
E, i E i E i
x’s and y’s, form a local coordinate system on Gn (T Σ).
Recall from Exercise 1.10.4 that E ∈ Vn (I) if and only if every ψ ∈ I n
vanishes on E. Each such ψ has some expression

ψ= fIJ dy I ∧ dxJ ,
I,J

where I and J are multi-indices with components in increasing order, such

that |I| + |J| = n, and the fIJ are smooth functions on Σ. Then ψ|E =
Fψ dx1 ∧ . . . ∧ dxn |E , where Fψ is polynomial in the pai :

Fψ = fI,J (x, y)pil11 . . . pilkk dxL ∧ dxJ ,
I,J,L
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.1. Integral elements and polar spaces 217

with I = (i1 , . . . , ik ) and L = (l1 , . . . , lk ) in increasing order. Thus, Vn (I) is

locally defined by equations that are polynomial in the pai , with coefficients
that are smooth functions on Σ.

Smooth points of Vn (I). To show existence of integral manifolds with

tangent space E at p ∈ Σ, we will need to study integral elements near
E. We will generally restrict to integral elements that are smooth points of
Vn (I), so we make the following definitions:

Definition 8.1.1. We say that k is the codimension of Vn (I) at E if k is

the maximum number of smooth functions Fψ on Gn (T Σ) that vanish on
Vn (I) and have linearly independent differentials at E.

Definition 8.1.2. An integral element E ∈ Vn (I) is Kähler-ordinary if

Vn (I) is a smooth submanifold of Gn (T Σ) near E.

If the codimension is constant on a neighborhood of E, then E is Kähler-

ordinary. If not, then by continuity of the coefficients in those polynomial
equations, there will be a neighborhood of E in which Vn has codimension
at least k. (In other words, codimension is a lower semicontinuous function
on Vn .) Since the codimension is bounded above, Kähler-ordinary elements
form an open dense set in Vn .

Example 8.1.3 (Linear Pfaffian systems). Suppose linearly independent

1-forms θa , 1 ≤ a ≤ s, generate a Pfaffian system on Σ with independence
condition ω 1 ∧ · · · ∧ ω n = 0, and satisfy structure equations
dθa ≡ πia ∧ ω i mod θ1 , . . . , θs .
Complete {θa , ω i } to a coframing with forms π , 1 ≤  ≤ r. Then πia =
Aai π + Cij
a ω j for some functions Aa , C a on Σ.
i ij
On any n-plane E satisfying the independence condition, π = pi ω i . (As
above, the pi form part of a local coordinate system on Gn (T Σ).) If E is an
integral element, then the θ’s must vanish on E and
(8.1) Aai pj − Aaj pi + Cij
a
− Cji
a
= 0.
Since these equations are linear in the pi , Vn (I) is a smooth submanifold
wherever the rank of (8.1) is locally constant. Hence, if one point in Vn is
smooth, so are all the other points in the same fiber.

Nonsmooth points of Vn (I) can occur in the fibers over points in Σ

where the generators of I vanish or become linearly dependent. However,
as Exercise 8.1.5 shows, this is not the only way in which singular points
arise.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 218 8. Cartan-Kähler III: The General Case

Example 8.1.4. On Σ = R2 , let θ = y 2 dx − x dy and let I = {θ}diff . Over

every point of Σ except the origin, the fiber of V1 (I) is a single point, and V1
is smooth. Since θ vanishes at the origin, the fiber there is G1 (R2 ) = RP1 .
We can introduce a local fiber coordinate on G1 (T Σ) such that dy − p dx|E =
0. Then V1 is defined by y 2 − xp = 0, and the integral 1-plane with p = 0 at
the origin is a singular point of V1 . Although V1 is smooth at all the other
points in the fiber above the origin, only the directions along the coordinate
axes in R2 are tangent to integral curves of I.

Exercise 8.1.5 (A degenerate 2-form): On Σ = Rm , let ω 1 , ω 2 ∈ Ω1 (Σ)

be pointwise independent and let I = {ω 1 ∧ ω 2 }alg . Complete ω 1 , ω 2 to a
coframing ω 1 , . . . , ω m with dual framing e1 , . . . , em . Fix a basepoint p and
show that e3 ∧e4 is a singular point of V2 (I)p . Determine all singular points.


More generally, we say that a form ϕ ∈ Ωk (Σ), k ≥ 2, is degenerate at p if

Ann(ϕ)p := {v ∈ Tp Σ | v ϕ = 0}

is nonzero. For I = {ϕ}alg , there will be singular points of Vk (I) in the

fiber over every point, consisting of those E ∈ Vk that, as subspaces of Tp Σ,
intersect Ann(ϕ)p nontransversely.

Polar spaces. The infinitesimal analogue of the Cauchy problem described

at the beginning of the chapter is the extension of an n-dimensional integral
element E ⊂ Tp Σ to an (n+1)-dimensional integral element E + ⊂ Tp Σ. The
space of all extensions is a (possibly empty) projective space P(H(E)/E)
where H(E) ⊆ Tp Σ is the polar space of E, defined as follows:

Definition 8.1.6. Let e1 , . . . , en be any basis for the integral element E ⊂

Tp Σ. The polar space of E is

H(E) := {v ∈ Tp Σ| ψ(v, e1 , . . . , en ) = 0 ∀ψ ∈ I n+1 }.

Exercise 8.1.7 (Properties of polar spaces): Let E ∈ Vn (I)p , E + ∈ G(n +

1, Tp Σ) and E ⊂ E + . Then
1. E ⊆ H(E).
2. E + ⊆ H(E) if and only if E + ∈ Vn+1 (I).
3. If v ∈ H(E), then (v ψ)|E = 0 for all ψ ∈ I. 
4. If E + ∈ Vn+1 (I), then H(E + ) ⊆ H(E).
5. If {ψ α } is a set of algebraic generators of I, then v ∈ H(E) if and only
if (v ψ α )|E = 0 for all generators ψ α of degree at most n + 1.

The following examples show how polar spaces are calculated.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.1. Integral elements and polar spaces 219

Example 8.1.8. Suppose the forms θ1 , θ2 , ω 1 , ω 2 and π form a coframe on

a manifold Σ5 such that
)
dθ1 ≡ ω 1 ∧ π
(8.2) mod θ1 , θ2
dθ2 ≡ ω 2 ∧ π
at some basepoint p ∈ Σ. Let I = {θ1 , θ2 }diff . Let E = {θ1 , θ2 , ω 2 , π}⊥ ⊂
Tp Σ—that is, the 1-plane in Tp Σ annihilated by θ1 , θ2 , ω 2 and π—and let
e ∈ E satisfy ω 1 (e) = 1. (For example, consider the special case where
Σ = R5 with coordinates (x, y, z, w, q), and θ1 = dz − qdx, θ2 = dw − qdy,
∂ ∂
π = dq, ω 1 = dx, ω 2 = dy. Then e = ∂x + q ∂z .)
The 2-forms of I are spanned by ω 1 ∧ π, ω 2 ∧ π and the wedge products
of θ1 and θ2 with all 1-forms. Suppose v ∈ H(E). Now, θ1 ∧ φ (v, e) = 0 for
all 1-forms φ if and only if θ1 (v) = 0. Similarly, we must have θ2 (v) = 0.
Then
ω 1 ∧ π (v, e) = −π(v),
ω 2 ∧ π (v, e) = 0
show that H(E) = {θ1 , θ2 , π}⊥ . Hence E is contained in the unique integral
2-plane E + = H(E).
Now consider E  satisfy π(ẽ) =
 = {θ1 , θ2 , ω 1 , ω 2 }⊥ ⊂ Tp Σ, and let ẽ ∈ E
∂
1. (In the special case above, this corresponds to ẽ = ∂p .) For the same
 then θ1 (v) = 0 and θ2 (v) = 0. Then
reasons, if v ∈ H(E),
ω 1 ∧ π (v, ẽ) = ω 1 (v),
ω 2 ∧ π (v, ẽ) = ω 2 (v)

show that H(E)  = E,

 so E
 is not contained in any higher-dimensional
integral element.
Example 8.1.9. Suppose the forms θ1 , θ2 , ω 1 , ω 2 , π1 and π2 form a coframe
on a manifold Σ6 such that
)
dθ1 ≡ ω 1 ∧ π1 + ω 2 ∧ π2
mod θ1 , θ2
dθ2 ≡ ω 1 ∧ π2
at some basepoint p ∈ Σ. Let I = {θ1 , θ2 }diff , let E = {θ1 , θ2 , ω 2 , π1 , π2 }⊥ ⊂
Tp Σ and let e ∈ E satisfy ω 1 (e) = 1. By polar space property 8.1.7.5 above,
v ∈ H(E) if and only θ1 (v) = θ2 (v) = 0 and v is annihilated by the interior
products of e with all generator 2-forms. Since these are
e (ω 1 ∧ π1 + ω 2 ∧ π2 ) = π1 ,
e (ω 1 ∧ π2 ) = π2 ,
we have H(E) = {θ1 , θ2 , π1 , π2 }⊥ .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 220 8. Cartan-Kähler III: The General Case

 = {θ1 , θ2 , ω 1 , π1 , π2 }⊥ , and let ẽ ∈ E

Now let E  satisfy ω 2 (ẽ) = 1. Then

ẽ (ω 1 ∧ π1 + ω 2 ∧ π2 ) = π2 ,
ẽ (ω 1 ∧ π2 ) = 0

shows that H(E)  = {θ1 , θ2 , π2 }⊥ . Note that H(E)

 is not itself an integral
3-plane, since (ω ∧ π1 + ω ∧ π2 ) |H(E)
1 2
 = 0.

Suppose E ∈ Vn−1 (I). Since every direction in H(E)/E corresponds to

an integral element E + in which E is a codimension one subspace, one might
expect the dimension of H(E) to be related to the dimensions of Vn−1 at
E and Vn at E + . This relationship is made precise in the following lemma,
which will be important for the proof of Cartan’s Test.
Lemma 8.1.10. Let E ∈ Vn−1 (I)p , E + ∈ Vn (I)p and E ⊂ E + ⊂ Tp Σ.
Then

(8.3) codimE + (Vn (I), Gn (T Σ))

≥ codimE (Vn−1 (I), Gn−1 (T Σ)) + codim(H(E), Tp Σ).

Proof. Let s = dim Σ − n, and take coordinates x1 , . . . , xn , y 1 . . . y s on Σ,

∂ ∂
centered at p, such that 1
, . . . , n span E + , dxn |E = 0, and H(E) is
∂x ∂x
annihilated by the forms dy α , where 1 ≤ α ≤ codim H(E). Then there are
linearly independent n-forms Φα in I such that
∂ ∂
(8.4) dy α (v) = Φα (v, ,..., )
∂x1 ∂xn−1
for all v ∈ Tp Σ.
For E  + ∈ Gn (T Σ) near E + , define the functions pa by requiring that
i
the vectors
∂ ∂
Xi = i
+ pai a
∂x ∂y
 + . (Here, 1 ≤ i ≤ n and 1 ≤ a ≤ s.) Similarly, for
be a basis for E
E ∈ Gn−1 (T Σ) near E, define the functions q a and uj on Gn−1 (T Σ) by
j
requiring that
∂ ∂ ∂
Zj = j
+ uj n + qja a , 1 ≤ j ≤ n − 1,
∂x ∂x ∂y
 The pa are, in fact, part of the local coordinate system on Gn (T Σ),
span E. i
in which E + is the origin, that we defined before. The qja and uj also
complete a local coordinate system on Gn−1 (T Σ) near E, in which E is the
origin.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.1. Integral elements and polar spaces 221

By the definition of codimension, there are forms φν ∈ I n−1 , where

1 ≤ ν ≤ codimE Vn−1 , such that the functions
F ν = φν (Z1 , . . . , Zn−1 )
have linearly independent differentials at E. Now let
Gν = φν ∧ dxn (X1 , . . . , Xn ),
H α = Φα (X1 , . . . , Xn ).

Certainly, Vn (I) must lie inside the common zero locus of the
codimE Vn−1 + codim H(E) functions Gν and H α near E + . We will show
that they have linearly independent differentials at E + .
Note that F ν is a polynomial in the qja and uj , with coefficients de-
pending on xi and y a , and Gν is a polynomial in the pai . In fact Gν (pai ) =
F ν (paj , 0), i.e., Gν is obtained from F ν by setting qja = paj and uj = 0.
(In particular, Gν does not involve the pan .) Furthermore, since any other
codimension one subspace of E + is also in Vn−1 (I), dF ν |E (∂/∂uj ) = 0. It
follows that the Gν have codimE Vn−1 (I) linearly independent differentials
at E + .
Let Ψα = Φα − dy α ∧ dx1 ∧ · · · ∧ dxn−1 . Then (8.4) shows that Ψα is
a sum of terms that either vanish at p, or are wedge products of degree at
least two in (dxn , dy 1 , . . . , dy s ). Thus, Ψα (X1 , . . . , Xn ) will be a polynomial
in the pai consisting of terms that either vanish at p, or are of degree two in
the pai , or, when they are of degree one, do not involve the pan . It follows
that
dH α |E + ≡ dpαn mod {dpaj , j < n}.
Thus, Gν and H α have linearly independent differentials at E + . 

Kähler-regularity. Definition 8.1.6 implies that the coefficients of the

equations defining H(E) depend continuously on E. Thus, the codimen-
sion of H(E) is a lower semicontinuous function of E—that is, on a small
enough neighborhood of E, the codimension of the polar space can only
increase. We will be interested in the case when it is locally constant, so we
make the following definition:
Definition 8.1.11. A Kähler-ordinary integral element E is Kähler-regular
 = codim H(E) for all E
if codim H(E)  in a neighborhood of E in Vn (I).

Remark 8.1.12. By convention, the polar space of 0 ∈ Tp Σ is the subspace

annihilated by the 1-forms of I at p. Hence the zero subspace in Tp Σ is
Kähler-regular if the rank of I 1 is constant on a neighborhood of p.

We would like to find conditions under which the ability to solve the
infinitesimal Cauchy problem (i.e., extending E to E + ) implies the ability
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 222 8. Cartan-Kähler III: The General Case

to solve the actual Cauchy problem. We will see that Kähler-regularity

provides such a condition. The following example shows that if there are
no Kähler-regular integral elements that can be extended, then the Cauchy
problem may not be solvable.
Example (8.1.8 continued). On any integral 2-plane E + of I, two of the
forms ω 1 , ω 2 , π must be independent, and yet ω 1 ∧ π and ω 2 ∧ π must be
zero. It follows that π|E + = 0, and E + = {θ1 , θ2 , π}⊥ .
If a 1-dimensional integral element E has π|E = 0, then E cannot be
extended to an integral 2-plane, and H(E) = E. (Since π|E = 0 is an open
condition, such E’s are Kähler-regular.) If π|E = 0, then E is contained in
a unique integral 2-plane E + . Thus, the only integral 1-planes that can be
extended are not Kähler-regular, and we now show that the extensions E +
in general are not tangent to integral surfaces.
Suppose we are trying to construct two-dimensional integral manifolds
of I. The structure equations (8.2) imply that the 1-form π vanishes on
all such surfaces and therefore must be added to the ideal. (The presence
of “extra equations” like π = 0 is typical when Kähler-regular elements
are not available, or cannot be extended.) In fact, when this is done, all
the remaining integral 1-planes are Kähler-regular, and we are looking for
integral surfaces of the Pfaffian system J = {θ1 , θ2 , π}diff ; these will exist
only where J is Frobenius.

Lemma 8.1.10 gives an upper bound for the dimension of Vn (I) at E + in

terms of the polar space H(E). The following lemma shows that this upper
bound is achieved when E is Kähler-regular.
Lemma 8.1.13. Let E, E + be as in Lemma 8.1.10. Assume E is Kähler-
regular, and let φν ∈ I n−1 , 1 ≤ ν ≤ codimE Vn−1 (I), be such that the
corresponding functions F ν on Gn−1 (T Σ) have linearly independent differ-
entials at E. Let Φα be a collection of linearly independent n-forms in I
such that
H(E) = {v ∈ Tp Σ|(v Φα )|E = 0, ∀α}.
Then there are a 1-form θ and a neighborhood U + of E + in Gn (T Σ), such
that for every E + ∈ U + , E + is a Kähler-ordinary integral n-plane if and
only if Φα |E+ = 0 and φν ∧ θ|E+ = 0 for all α and ν.

Proof. Let x1 , . . . , xn , y 1 , . . . , y s , X1 , . . . , Xn , and pai be as in the proof of

Lemma 8.1.10. Let θ = dxn , and let U + be the neighborhood of E + in
Gn (T Σ) where the pai are defined.
Since E is Kähler-ordinary, the forms φν span I n−1 in the vicinity of
p. Since E is Kähler-regular, there is a neighborhood U of E in Gn−1 (T Σ)
 ∈ U ∩ Vn−1 (I).
such that the Φα also generate the polar equations for any E
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.2. Example: triply orthogonal systems 223

If φν ∧ θ|E+ = 0, then

(8.5) φν ∧ θ(X1 , . . . , Xn ) = φν (X1 , . . . , Xn−1 )

shows that the (n−1)-plane E ⊂E  + spanned by X1 , . . . , Xn−1 is an integral

 + ⊂ H(E),
plane. If Φα |E+ = 0 as well, then E  and so E  + ∈ Vn (I).
Now it is clear that codimE + Vn (I) ≤ codimE Vn−1 (I) + codim H(E),
and this upper bound holds for all nearby integral n-planes. Then the lower
bound (8.3) and the lower semi-continuity of codimension show that E  + is
Kähler-ordinary. 

8.2. Example: triply orthogonal systems

In this section we discuss an example that uses the Cauchy-Kowalevski The-
orem to construct three-dimensional integral manifolds for an EDS. The con-
struction is relatively simple, since the system in the example is generated
algebraically by forms of degree three. The example will also serve as an
outline for our proof of Cartan-Kähler, since the setup is similar except for
details necessary when there are generators of lower degree.
A triply orthogonal system in Euclidean space consists of three foliations
of E3 by surfaces that intersect orthogonally at each point. For example,
the coordinate surfaces for Cartesian, cylindrical and spherical coordinates
all constitute triply orthogonal systems. More generally, a system of curvi-
linear coordinates in Euclidean space consists of three functions u, v, w, with
linearly independent differentials, such that the metric (du)2 + (dv)2 + (dw)2
is at each point a multiple of the Euclidean metric, and this is the case if and
only if the level surfaces of u, v, w form a triply orthogonal system. So, when
one asks for metrics on Euclidean space that are conformal to the standard
metric, one is essentially asking for a triply orthogonal system of surfaces,
from which one can recover u, v, w up to reparametrization. We will show
how “large” the family of triply orthogonal systems is.
We associate a frame field (e1 , e2 , e3 ) to the foliations of a triply or-
thogonal system by letting ei be a smoothly varying unit normal to the i-th
foliation. This frame field gives rise to a three-dimensional integral manifold
of an exterior differential system on the orthonormal frame bundle F = FEon3
(see §2.1) as follows:
Let D ⊂ E3 be the domain of the triply orthogonal system, and let
f : D → F |D be the frame field. Then f pulls back the canonical forms
(ω 1 , ω 2 , ω 3 ) on F to give the coframe field dual to (e1 , e2 , e3 ). The Frobenius
condition for the distribution defined by f ∗ (ω i ) implies that f ∗ (dω i ∧ω i ) = 0
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 224 8. Cartan-Kähler III: The General Case

for i = 1, 2, 3. Since

dω 1 ∧ ω 1 = −ω21 ∧ ω 2 ∧ ω 1 − ω31 ∧ ω 3 ∧ ω 1 ,
(8.6) dω 2 ∧ ω 2 = −ω32 ∧ ω 3 ∧ ω 2 − ω12 ∧ ω 1 ∧ ω 2 ,
dω 3 ∧ ω 3 = −ω13 ∧ ω 1 ∧ ω 3 − ω23 ∧ ω 2 ∧ ω 3 ,

we see that f (D) is a 3-dimensional integral manifold of the exterior differ-

ential system I generated by Ω1 = ω32 ∧ ω 2 ∧ ω 3 , Ω2 = ω13 ∧ ω 3 ∧ ω 1 , and
Ω3 = ω21 ∧ ω 1 ∧ ω 2 (see Exercises 8.2.1). Conversely, a 3-dimensional integral
manifold of I that satisfies the independence condition ω 1 ∧ ω 2 ∧ ω 3 = 0
defines the field of normals for a triply orthogonal system on an open set
U ⊂ E3 .
Let π : F → E3 be the bundle projection. Since I contains no forms
of degree two or lower, all 2-planes are Kähler-ordinary integral elements.
We will restrict our attention to the open subset of G(2, Tp F ) consisting
of 2-planes E such that π∗ |E is injective, since only they are contained in
3-planes that satisfy the independence condition. In fact, each such E is
contained in a unique integral 3-plane. For, π∗ |E being injective implies
that none of the 2-forms ω 1 ∧ ω 2 , ω 2 ∧ ω 3 , and ω 3 ∧ ω 1 restrict to E to be
zero. Then, one sees from the formulae for the Ωi that the polar equations
for E are rank 3, and dim H(E) = 3. Since our assumption about E is an
open condition, all such E are Kähler-regular.
We will illustrate below how a frame field for a triply orthogonal system
may be extended from a surface to an open domain in E3 using the Cauchy-
Kowalevski Theorem. Since this is a theorem in the real-analytic category,
it is necessary to check that the generators of I are analytic. First of all, the
frame bundle is diffeomorphic to the (matrix) Lie group of rigid motions,
which endows F with analytic coordinates. Then, since via this diffeomor-
phism the ω’s correspond to entries in the Maurer-Cartan form g −1 dg, the
ω’s are analytic differential forms—that is, combinations of differentials of
analytic functions with analytic functions as coefficients. Hence we have an
analytic EDS.
Assume P 2 ⊂ F is an analytic Kähler-regular integral surface of I (i.e.,
all its tangent planes are Kähler-regular), and fix a point p ∈ P . There
exist analytic coordinates x0 , x1 , x2 , y 1 , y 2 , y 3 on a neighborhood U of p in
F , such that P ∩ U is defined by x0 = y 1 = y 2 = y 3 = 0 and H(Tp P ) is
∂ ∂ ∂
spanned by 0
, 1
, and . We will construct an analytic 3-dimensional
∂x ∂x ∂x2
integral manifold N , containing P ∩ U , given by y α = F α (x0 , x1 , x2 ) for
some functions F α to be determined. If N is of this form, the vectors
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.2. Example: triply orthogonal systems 225

∂ ∂F α ∂ ∂ ∂F α ∂
+ and + will span Tπ(p) N . (Here and in what
∂x0 ∂x0 ∂y α ∂xi ∂xi ∂y α
follows, Greek indices run from 1 to 3 and Roman indices from 1 to 2.)
We will now show that the F α satisfy a system of PDE of Cauchy-
Kowalevski type. For variables q β and pβj , let
 
α ∂ β ∂ ∂ β ∂ ∂ β ∂
Ω +q , + p1 β , 2 + p2 β = Aαβ q β + B α ,
∂x0 ∂y β ∂x1 ∂y ∂x ∂y
where Aαβ and B α are polynomial in the pβj with analytic functions on U as
coefficients. Since at p, the equations
 
α ∂ β ∂ ∂ ∂
Ω +q , , =0
∂x0 ∂y β ∂x1 ∂x2
coincide with the polar equations of E, they are of rank 3 in the q β . We
may assume that Aαβ is invertible on some (possibly smaller) U ⊂ F and for
%pβj % sufficiently small. Then
 
−1 γ α ∂ β ∂ ∂ β ∂ ∂ β ∂
(A )α Ω +q , + p1 β , 2 + p2 β = q γ − C γ ,
∂x0 ∂y β ∂x1 ∂y ∂x ∂y
where C γ is an analytic function of the x’s, y’s and the pβj . So, we must
have  
∂F α α 0 1 2 1 2 3 ∂F
β
=C x ,x ,x ,F ,F ,F ;
∂x0 ∂xj
with F α (0, x1 , x2 ) = 0. Since this is now a system of PDE in Cauchy-
Kowalevski form, we conclude that a unique solution exists, on a possibly
smaller U .

Any discussion of triply orthogonal systems must mention the famous

theorem of Dupin that the surfaces in the system intersect along lines of
curvature. This is easy to prove using our system I, since one has only
to check that along a surface to which e3 is normal, e1 and e2 diagonalize
the second fundamental form. The theorem of Dupin is most striking when
applied to the triply orthogonal system formed by the confocal quadrics
x2 y2 z2
+ + = 1, a2 < b2 < c2 ,
a2 − λ b2 − λ c2 − λ
as λ varies. One sees that along the ellipsoids formed when λ < a2 , the lines
of curvature are exactly the intersections with the two kinds of hyperboloids
formed when a2 < λ < b2 and b2 < λ < c2 ([178], §2-11).
Exercises 8.2.1:
1. Use the structure equations (8.6) to show that the ideal generated by
the forms dω i ∧ ω i is the same as that generated by the Ωi . 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 226 8. Cartan-Kähler III: The General Case

2. An integral surface P ⊂ F corresponds to a surface S ⊂ E3 with a frame

field (e1 , e2 , e3 ) attached at each point. Show that if P is Kähler-regular,
then none of the vectors ei can be tangent to S.
3. Use Dupin’s theorem to show that any hyperboloid in E3 has a one-
parameter family of closed lines of curvature. 

8.3. Statement and proof of Cartan-Kähler

It is important to note that the Cartan-Kähler Theorem and Cartan’s Test
only apply to systems generated by real-analytic differential forms on an
analytic manifold Σ. (However, a version of Cartan-Kähler in the C ∞ cate-
gory is available for involutive hyperbolic systems [189].) For the rest of this
chapter our remarks are limited to the analytic category.
Theorem 8.3.1 (First version of Cartan-Kähler). Assume I is an analytic
EDS on Σ and P n ⊂ Σ is an analytic submanifold whose tangent spaces
are Kähler-regular integral elements such that, at each p ∈ P , H(Tp P ) has
dimension n + 1. Then, for each p ∈ P , there is an open neighborhood
U ⊂ Σ of p and an (n + 1)-dimensional manifold N ⊂ U which is the unique
analytic integral manifold containing P ∩ U .

Proof. As in §8.2, we will reduce our problem to solving a system using the
Cauchy-Kowalevski Theorem.
Choose analytic coordinates x0 , x1 , . . . , xn , y 1 , . . . , y s , centered at p, such
∂ ∂ ∂
that 1
, . . . , n span Tp P and 0
∈ H(Tp P ). As in Lemma 8.1.13, let
∂x ∂x ∂x
{φ } be a basis for I near p, and let the (n + 1)-forms Φα , 1 ≤ α ≤ s,
ν n

generate the polar equations in a neighborhood of Tp P .

If N n+1 is an integral manifold containing P , then Tp N = H(Tp P ), and
the x’s serve as local coordinates on N . If N is to be defined by y α =
∂ ∂F α ∂
F α (x0 , x1 , . . . , xn ), then T N will be spanned by X0 = + and
∂x0 ∂x0 ∂y α
∂ ∂F α ∂
the vectors Xi = + for 1 ≤ i ≤ n. Let Let 1 ≤ α, β, γ ≤ s and
∂xi ∂xi ∂y α
define functions Aαβ (xi , y γ ; pγj ) and B α (xi , y γ ; pγj ) by
 
∂ ∂ ∂ ∂ ∂ ∂
Φ α
0
+ q β β , 1 + pβ1 β , . . . , n + pβn β = Aαβ q β + B α .
∂x ∂y ∂x ∂y ∂x ∂y

As in §8.2, the Aαβ are the entries of a matrix which is invertible when
evaluated at the origin; so, on a possibly smaller open set (which we continue
to denote by U ) and for sufficiently small %pβj %, we can define forms Φ α =
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.3. Statement and proof of Cartan-Kähler 227

(A−1 )αβ Φβ . Let C α (xi , y β ; pβj ) be defined by

 
 ∂ ∂ ∂ β ∂ ∂ ∂
C =q −Φ
α α α
+q β
, β
+ p1 β , . . . , n + pn β .
∂x0 ∂y β ∂x1 ∂y ∂x ∂y
Then we construct N by using Cauchy-Kowalevski to solve
⎧  
⎨ ∂F α α 0 1 n 1 s ∂F β
=C x ,x ,...,x ,F ,...,F ; ,
∂x0 ∂xj
⎩ α
F (0, x1 , . . . , xn ) = 0.

It remains to be shown that the forms φν ∧ dx0 , which by Lemma 8.1.13

span the rest of the (n + 1)-forms in I, pull back to be zero on N . By (8.5),
this is equivalent to showing that φν (X1 , . . . , Xn ) = 0. On N ,
∂ ν  ∂ 
0
φ (X 1 , . . . , X n ) = dφν
(X 0 , . . . , X n ) − i
dxi ∧ φν (X0 , . . . , Xn ) .
∂x ∂x
i

However, since ≡ 0 and dφν φν ∧ dxi

≡ 0 modulo {Φα , φν ∧ dx0 }, the right-
hand side of this PDE is expressible in terms of analytic functions of the
φν (X1 , . . . , Xn ) and their xi -derivatives. Since φν (X1 , . . . , Xn ) = 0 along
x0 = 0, by the Cauchy-Kowalevski Theorem the unique solution to this
system of partial differential equations is φν (X1 , . . . , Xn ) ≡ 0 for all ν. 
Theorem 8.3.2 (Second version of Cartan-Kähler). Assume Σ, I and P n
are as in the first version, but H(Tp P ) has dimension n+r+1 for each p ∈ P .
Assume that R ⊂ Σ is an analytic submanifold of codimension r containing
P , such that Tp R intersects transversely with H(Tp P ). Then there is an
open set U ⊂ R containing P , and a submanifold N n+1 ⊂ U which is the
unique analytic integral manifold in U containing P .

Proof. The transversality assumption implies that Tp R ∩ H(Tp P ) has di-

mension n + 1. Thus, we can apply the previous version, with Σ replaced
by R, at any point p ∈ P . By applying Theorem 8.3.1 at points in P , we
obtain a collection {Uα } of open sets Uα ⊂ R such that {P ∩ Uα } is an open
cover of P , and integral submanifolds Nα ⊂ Uα with P ∩ Uα ⊂ Nα .
Suppose that P ∩ Uα ∩ Uβ is nonempty. Then Nβ ∩ Uα is an analytic
integral submanifold with a nonempty intersection with P ∩Uα , and so by the
uniqueness part of Cauchy-Kowalevski (see Theorem D.2.3) Nβ ∩ Uα ⊂ Nα .
We finish by letting U = Uα and N = Nα . 
α α

The Cartan-Kähler Theorem can be applied inductively to obtain an

integral manifold with a given integral n-plane E as its tangent space. To
do this, one needs a flag 0 ⊂ E1 ⊂ · · · ⊂ En−1 ⊂ En = E of integral elements
inside E, such that the second version of the theorem can be applied at each
step. To apply Theorem 8.3.2, one needs to choose a “restraining manifold”
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 228 8. Cartan-Kähler III: The General Case

R at each step, in order to make the Cauchy problem determined. The

proof of the next result gives an idea of how this is done.
Theorem 8.3.3 (Third version of Cartan-Kähler). Let Ek , 0 ≤ k ≤ n, be
a flag of integral elements at p for an analytic EDS, with dim Ek = k, and
such that Ek is Kähler-regular for 0 ≤ k ≤ n−1. Then there exists a smooth
n-dimensional integral manifold N whose tangent space at p is En .

Proof. Let ck = codim H(Ek ) for 0 ≤ k < n; note that the ck are nonde-
creasing. Let x1 , . . . , xn , y 1 , . . . , y s be local coordinates centered at p, defined
∂ ∂
on an open set U ⊂ Σ, and chosen so that Ek is spanned by ,..., k
∂x1 ∂x
for k ≤ n and H(Ek ) is annihilated by dy 1 , . . . , dy ck for k < n.
Let R1 ⊂ U be the submanifold given by setting x2 = · · · = xn = 0
and y a = f1a (x1 ) for a > c0 , where f1a are some analytic functions such
that f1a (0) = 0. By Theorem 8.3.2 with R = R1 , there exists a unique
1-dimensional integral manifold N1 ⊂ R1 tangent to E1 at p.
Next, let R2 ⊂ U be the manifold given by setting x3 = · · · = xn = 0
and y a = f2a (x1 , x2 ) for a > c1 . To arrange that R2 contains N1 , we require
that f2a (x1 , 0) = f1a (x1 ). By Theorem 8.3.2 with R = R2 , there exists a
unique 2-dimensional integral manifold N2 ⊂ R2 tangent to E2 at p. (Note,
however, that we may have to shrink the open set U to ensure that H(Tq N1 )
is transverse to Tq R2 at every point q ∈ N1 .)
Proceeding in this way, we eventually obtain an n-dimensional integral
manifold tangent to En at p. (At the last step, Rn is defined by y a =
fna (x1 , . . . , xn ) for a > cn−1 .) 

Remark 8.3.4. Note that, in constructing the successive restraining man-

ifolds R1 , R2 , . . ., we may choose the functions f1a (x1 ) for c0 < a ≤ c1 (sub-
ject to f1a (0) = 0), but functions f1a (x1 ) are constrained to equal f2a (x1 , 0)
for a > c1 . Similarly, functions f2a (x1 , x2 ) may be freely specified only for
c1 < a ≤ c2 . In this way, an n-dimensional integral manifold N containing
p is specified by choosing c1 − c0 functions of one variable, c2 − c1 functions
of two variables, and so on. (In the next section, these differences ck − ck−1
are defined as the characters of the flag.)
Relative to the same coordinate system, any ‘nearby’ integral n-fold N 
1 2 n a
will contain a point p̃ where x = x = . . . = x = 0 and y = f0 for some a

constants f0a . Then N  is obtained by successively applying Theorem 8.3.2

with R1 specified by x2 = . . . xn = 0 and y a = f1a (x1 ) for a > c0 (subject to
f1a (0) = f0a ), then R2 specified by x3 = . . . = xn = 0 and y a = f2a (x1 , x2 )
for a > c1 (subject to f2a (x1 , 0) = f a (x1 )), and so on. Thus, near N the
space of integral n-folds is parametrized by c0 constants, c1 − c0 functions
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.4. Cartan’s Test 229

of one variable, and so on up to s − cn−1 functions of n variables, where

s = dim Σ − n.

8.4. Cartan’s Test

In order to check that a flag of integral elements is Kähler-regular, it seems
that it would be necessary to compute not only the dimensions of the polar
spaces H(Ek ), but also the dimensions of the polar spaces of all nearby
integral elements! This is avoided by the following test, which generalizes
(5.18) and Definition 5.5.4:
Theorem 8.4.1 (Cartan’s Test for Involutivity). Let Ek , 0 ≤ k ≤ n, be a
flag of integral elements for I at p, and let ck = codim H(Ek ) within Tp Σ
for 0 ≤ k ≤ n − 1. Then the codimension of Vn within Gn (T Σ) satisfies
(8.7) codimEn Vn (I) ≥ c0 + c1 + · · · + cn−1 .
Moreover, Vn (I) is smooth of codimension exactly c0 + c1 + · · · + cn−1 at En
if and only if the Ek are all Kähler-regular for 0 ≤ k ≤ n − 1.
Definition 8.4.2. If there exists a flag for E with equality holding in (8.7),
then we say E is an ordinary integral element. If there is a neighborhood of
E of ordinary integral elements, we say that I is involutive at E.

Remark 8.4.3. If an integral element E is equipped with a flag for which

the characters fail Cartan’s Test, one must make sure the flag was chosen
generically. If the characters for a generic flag still fail the test, it is necessary
to prolong the system (see §6.5).

Proof of 8.4.1. The inequality follows by successive applications of (8.3).

Furthermore, if all the Ek are Kähler-regular, successive applications of
Lemma 8.1.13 show that Vn (I) is smooth at En , with the required codi-
mension.
Now suppose Vn (I) is smooth at En , with the above codimension. We
first show that En−1 is Kähler-ordinary. One might suspect this is the case
because Vn (I) cannot be so large without Vn−1 (I) being large. Consider
the following heuristic argument:
By the first part of this theorem, Vn−1 (I) has codimension at least c0 +
· · · + cn−2 at En−1 . On the other hand, for every E n ∈ Vn (I) near En , all

the hyperplanes inside En are in Vn−1 (I). In fact we may parametrize this
set of hyperplanes, and get a map from Vn (I) × Rn−1 to a neighborhood
of En−1 in Vn−1 (I). But for E n−1 ∈ Vn−1 (I) sufficiently close to En−1 ,
H(E n−1 ) has codimension at least cn−1 . This means the map has fiber of
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 230 8. Cartan-Kähler III: The General Case

dimension at most s − cn−1 . (Here we again set s = dim Σ − n.) Since the
dimension of Gn (T Σ) is n + s + ns, Vn (I) has dimension
d = n + s + ns − (c0 + · · · + cn−1 ).
So, we expect the image of this map to have dimension at least
d + n − 1 − (s − cn−1 − 1) = n + s + (n − 1)(s + 1) − (c0 + · · · + cn−2 ),
which is exactly the largest dimension it can be. We will now make this
argument rigorous.
Let x1 , . . . , xn , y 1 , . . . , y s be local coordinates such that Ek is spanned by
∂/∂x1 , . . . , ∂/∂xk and H(Ek ) is annihilated by dy 1 , . . . , dy ck . As in the proof
of Lemma 8.1.10, we define functions pai completing a coordinate system on
U ⊂ Gn (T Σ) centered at En , and also define Xi , qja , uj and Zj as before.
Let F : U × Rn−1 → Gn−1 (T Σ) take (En , (v1 , . . . , vn−1 )) to the subspace
n spanned by the vectors Xj + vj Xn , 1 ≤ j ≤ n − 1. In coordinates,
of E
this map is given by uj = vj and qja = paj + vj pan . Differentiating these
equations shows that the kernel of F∗ at (En , (0, . . . , 0)) is spanned by the
vectors ∂/∂pan .
Let f denote the restriction of F to (Vn (I) ∩ U ) × Rn−1 . We want to
show that the rank of f is at least d − (s − cn−1 ) at (En , (0, . . . , 0)), or
equivalently that the intersection of the kernel of F∗ with TEn Vn (I) has
dimension at most s − cn−1 . Once this is known, the image of f contains
a smooth submanifold through En−1 with maximum codimension, and it
follows that En−1 is Kähler-ordinary; in fact, the image of f fills out Vn−1 (I)
near En−1 .
We will need to look at the equations that cut out Vn (I) near En in
local coordinates. By the above inequality, codimEn−1 Vn−1 (I) ≥ c0 + · · · +
cn−2 . So, let φν , 1 ≤ ν ≤ c0 + · · · + cn−2 , and Φα , 1 ≤ α ≤ cn−1 , be
defined as in the proof of Lemma 8.1.10. We saw then that the n-forms
φν ∧ dxn and Φα give rise to functions Gν and H α on Gn (T Σ) with linearly
independent differentials at En . Since Vn (I) is smooth at En , and of exactly
this codimension, the equations Gν = 0 and H α = 0 define Vn (I) near En .
As before, Gν is a polynomial in the pai which does not involve the pan ,
and
(8.8) dH α |En ≡ dpαn
mod {dpai , i < n}, 1 ≤ α ≤ cn−1 .
s
Now, if some linear combination a=1 ta ∂/∂pan is tangent to Vn (I) at En ,
(8.8) shows that ta = 0 for 1 ≤ a ≤ cn−1 . It follows that the kernel of f∗ at
En is as small as desired, and the rank of f is as large as desired.
Now it is easy to show that En−1 is Kähler-regular. For, suppose a
nearby E n−1 ) > codim H(En−1 ). This E
n−1 ∈ Vn−1 (I) has codim H(E n−1
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.4. Cartan’s Test 231

n ∈ Vn (I) near En .
will be in the image of f , and so is contained in an E
Now (8.3) implies that the codimension of Vn (I) at E n is larger than it
really is, so we reach a contradiction.
To conclude the proof of the theorem, we apply the above argument
inductively to show that every Ek is Kähler-regular. 

Characters. In practice, the inequality of Cartan’s Test is used in a dif-

ferent form. Assume that E is a Kähler-ordinary integral n-plane with a
chosen flag, and set
s0 = c0 ,
sk = ck − ck−1 , 1 ≤ k ≤ n − 1,
sn = codim E − cn−1 = codim E − (s0 + s1 + · · · + sn−1 ).
Then dim Vn (I) − dim Σ ≤ s1 + 2s2 + 3s3 + · · · + nsn , with equality if and
only if E is ordinary and a sufficiently generic flag has been chosen.
Once an EDS has passed Cartan’s Test, the Cartan-Kähler Theorem
8.3.3 may be used to show that integral manifolds exist. Moreover, Remark
8.3.4 shows that manifolds depend on successive choices of s0 constants,
s1 functions of one variable, s2 functions of two variables, etc., just as in
Theorem 5.5.7.
Remark 8.4.4 (Linear Pfaffian systems). When we have a linear Pfaffian
system with structure equations
dθa ≡ πia ∧ ω i mod θ1 , . . . , θs ,
and E an integral element satisfying the independence condition, there will
be 1-forms π̃ia = πia − paij ω j , with paij = paji , such that π̃ia |E = 0. (In fact, we
can exchange π for π̃ in the structure equations.) Choose the flag so that
ωi |Ek = 0 for i > k, and let e1 , . . . , en be the basis for E dual to the ω i .
BecauseH(E0 ) is just the subspace of Tp Σ annihilated by the θ’s, s0
is the rank of the Pfaffian system. Applying Exercise 8.1.7.5 shows that
H(Ek ) is annihilated by the θ’s and by the forms π̃ia for i ≤ k. Thus, sk is
the number of forms in the k-th column of the tableau that are independent
of those in previous columns. So, these characters—with the exception of sn
when there are Cauchy characteristics—agree with those defined in Chapter
5.
Remark 8.4.5 (Cauchy characteristics). Note that Cauchy characteristic
vectors (see §7.1) will lie inside H(E) for every E. If the space A(I) of
Cauchy characteristics is of dimension k, it adds k to sn (but not to any of
the other characters of the flag, since they are differences of dimensions of
polar spaces). There will also be k independent 1-forms that are not involved
in the Cartan system of I, so these forms also add nk to the fiber dimension
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 232 8. Cartan-Kähler III: The General Case

of Vn (I). Thus, the Cauchy characteristics will contribute an extra nk

to each side of the inequality in Cartan’s Test. What this means is that,
if we choose to ignore Cauchy characteristics (i.e., omitting characteristic
directions from codim E and the fiber dimension of Vn ), then the result
of Cartan’s Test will be the same. (This idea is used in Examples 8.5.3
and 8.5.4 below.) When we ignore characteristics in this fashion, we are
in effect working on the manifold obtained by forming the quotient by the
characteristics, with the quotient system constructed in Theorem 7.1.21.

8.5. More examples of Cartan’s Test

We will begin with some examples that are typical of the applications of
exterior differential systems to classical surface theory found in Cartan’s
treatise [40]. As in Chapter 2, we let F = FEon3 denote the orthonormal frame
bundle of E3 . Given a surface M ⊂ E3 , recall that a section f : M → F |M is
a first-order adapted framing if Tx M is spanned by (f ∗ e1 , f ∗ e2 ). (As usual,
we suppress f and the basepoint in the notation). Along such a framing,
ω 3 = 0 and ωi3 = hij ω j , where hij is the symmetric matrix of the second
fundamental form in the e1 , e2 basis.

Example 8.5.1 (Surfaces with one of the principal curvatures constant).

(Cartan [40], Ex. III) Along such a surface, away from the umbilic points,
one can obtain a smooth Darboux framing so that e2 points in the principal
direction with constant principal curvature k0 . Since e1 points in the other
principal direction, this implies that hij is diagonal with h22 = k0 . Thus, the
framing f (M ) is an integral surface of the forms ω 3 , ω23 − k0 ω 2 and ω13 ∧ ω 1
on F .
Let θ = ω23 − k0 ω 2 . Since dθ = (ω13 − k0 ω 1 ) ∧ ω12 and dω 3 ≡ 0 modulo θ
and ω13 ∧ ω 1 , we will let π = ω13 − k0 ω 1 to simplify calculations. Then f (M )
is an integral surface of the differential ideal I = {ω 3 , θ, π ∧ ω 1 , π ∧ ω21 }alg ,
to which we now apply Cartan’s Test.
First, we investigate the space of integral 2-planes E satisfying the in-
dependence condition. While ω 3 and θ must vanish on E, suppose π = aω 1
and ω21 = bω 1 + cω 2 on E; then the generator 2-forms imply that ac = 0.
Thus, E is Kähler-ordinary when exactly one of a and c is zero; then the
fiber dimension of V2 (I) is two.

Case 1: Assume c = 0 and a = 0. Let e1 and e2 be vectors in E dual to

the restrictions of ω 1 and ω 2 to E. Then
e1 (π ∧ ω 1 ) = −π,
e1 (π ∧ ω21 ) = −bπ.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.5. More examples of Cartan’s Test 233

This shows that H({e1 }) = {ω 3 , π, θ}⊥ . Since s0 = 2, s1 = 1, and s2 = 1,

s1 + 2s2 exceeds the fiber dimension and Cartan’s Test fails to show that E
is ordinary. Since
e2 (π ∧ ω 1 ) = 0,
e2 (π ∧ ω21 ) = aπ,
we get the same result for the flag 0 ⊂ {e2 } ⊂ E, and in fact for any flag
inside E, so E is not ordinary.
If we consider the subvariety X ⊂ V2 (I) of integral elements with a = 0,
the system is not involutive at any E ∈ X. However, if we prolong at points
of X, we get a Frobenius system whose solutions are spheres of radius k10 .

Case 2: Assume a = 0 and c = 0. Let e1 and e2 be vectors in E as before.

Then
e1 (π ∧ ω 1 ) = aω 1 − π,
e1 (π ∧ ω21 ) = aω21 − bπ.
So, for the flag 0 ⊂ {e1 } ⊂ E we have s0 = 2, s1 = 2, s2 = 0; then
s1 + 2s2 = 2, and this E is ordinary.
On the other hand, suppose we had used the flag 0 ⊂ {e2 } ⊂ E. Since
e2 (π ∧ ω 1 ) = 0,
e2 (π ∧ ω21 ) = 0,
we would have had s0 = 2, s1 = 0 and s2 = 2.

We conclude that there exist many of these surfaces: in Case 2, the

integral surfaces through a given point in F depend on s1 = 2 functions of
one variable. Once we know this, we ought to find a geometric interpretation
for the two arbitrary functions.
Suppose we have one of these surfaces, with a framing adapted as above.
Let s be an arclength parameter along a line of curvature in the surface in
the e1 direction. If x represents the coordinate vector for the point on the
surface, then dx/ds = e1 along this curve. By differentiating the frame vec-
tors modulo ω 2 and I, we obtain the following system of ordinary differential
equations for the frame components along this curve:
dx de1 de2 de3
= e1 , = −be2 + (a + k0 )e3 , = be1 , = −(a + k0 )e1 .
ds ds ds ds
Conversely, if we specify functions a(s) = 0 and b(s), then we can integrate
the above system to obtain a Kähler-regular integral curve of I that has a
unique extension to an integral surface.
Exercise 8.5.2: Show that the curvature and torsion of this curve are κ =
|q| and τ = Im(q  /q), where q(s) = (a + k0 ) − ib. (Note that the real
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 234 8. Cartan-Kähler III: The General Case

and imaginary parts of q are the curvatures for a “natural” or “relatively

parallel” frame [16] along the curve.) 
Example 8.5.3 (Linear Weingarten surfaces). A surface is a Weingarten
surface if its Gauss curvature and mean curvature satisfy a given equa-
tion f (H, K) = 0. Recall from §7.4 that when this equation is linear in
H and K, the corresponding differential system is locally equivalent to a
Monge-Ampère equation. Examples include minimal surfaces (H = 0) and
pseudospherical surfaces (K = −1), both of which are discussed in detail in
§7.4.
If f (H, K) = AK + 2BH + C for constants A, B, C, then first-order
adapted framings along the surface are integrals of the 2-form
Θ = Aω13 ∧ ω23 + B(ω13 ∧ ω 2 − ω23 ∧ ω 1 ) + Cω 1 ∧ ω 2 .
Let I = {ω 3 , dω 3 , Θ}alg ; note that this is not a linear Pfaffian system. How-
ever, its prolongation was shown to be involutive in Example 6.8.2.
We now apply Cartan’s Test to I. An integral 2-plane E satisfying the
independence condition ω 1 ∧ ω 2 |E = 0 will be determined by coefficients hij
such that ωi3 |E = hij ω j , and by the values of ω12 |E . (However, since ω12 is
dual to a Cauchy characteristic for the system, according to Remark 8.4.5
we may ignore it in computing Cartan characters and the dimension of the
space of integral elements. This means we are essentially working on the
quotient five-manifold.) If h is the symmetric matrix with components hij ,
we must have
A det h + B tr h + C = 0.
Thus, V2 (I) is smooth, of fiber dimension two, except where Ah + BI = 0.
(This can only happen if AC − B 2 = 0; at any rate, we exclude points where
V2 (I) is not smooth.)
Suppose e ∈ E satisfies e ω i = xi . Let e span a line E1 ⊂ E; we will
calculate the characters for this flag. Then H(E1 ) is annihilated by ω 3 and
)
e dω 3 ≡ x1 ω13 + x2 ω23
mod ω 1 , ω 2 .
e Θ ≡ −(Ah2j xj + Bx2 )ω13 + (Ah1j xj + Bx1 )ω23
Thus s1 = 2, unless the determinant Ahij xi xj + B((x1 )2 + (x2 )2 ) vanishes.
Because we assume that Ah + BId = 0, we can always choose e ⊂ E so that
this doesn’t happen.
So, H(E1 ) = E, s2 = 0 and s1 + 2s2 = 2. We conclude that the system
for linear Weingarten surfaces is involutive wherever the space of integral 2-
planes is smooth, and that solutions depend on two functions of one variable.
For example, a minimal surface could be determined (up to rigid motion)
by specifying the curvature and torsion of a curve on the surface.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.5. More examples of Cartan’s Test 235

Example 8.5.4 (Hypersurfaces of constant scalar curvature in S 4 ). To set

up this example, we will use a trick that extracts the components of the
Ricci tensor from the curvature forms of a Riemannian manifold.
Given orthonormal vectors ω 1 , . . . , ω n , we let ω(ii ···ik ) denote the wedge
product of n − k of the ω’s such that
ω(ii ···ik ) ∧ ω i1 ∧ · · · ∧ ω ik = ω 1 ∧ · · · ∧ ω n .
For example, ω(2) = ω 3 ∧ ω 1 when n = 3 and ω(41) = −ω 2 ∧ ω 3 when n = 4.
It is easy to verify the identities
ω(j) ∧ ω i = δji ω 1 ∧ · · · ∧ ω n ,
ω(jk) ∧ ω i = δji ω(k) − δki ω(j) ,
ω(ijk) ∧ ω p = δip ω(jk) − δjp ω(ik) + δkp ω(ij) .

Recall from §3.1 that the components of the curvature tensor are deter-
mined by the curvature 2-forms:
(8.9) Θij = 12 Ri jkl ω k ∧ ω l .
Now, if we set Φij = Θik g kj , then from (8.9) and the above identities we
obtain
ω(pij) ∧ Φij = 2Rjj ω(p) − 4Rpj ω(j) ,
where Rij are the components of the Ricci tensor, and Rji = g ik Rkj . Finally,
this equation yields an expression for the scalar curvature R if we wedge with
ω p and sum on the index p:
ω(pij) ∧ Φij ∧ ω p = 2(n − 2)R ω 1 ∧ . . . ∧ ω n .
In particular, on a three-manifold this says Φ12 ∧ ω 3 + Φ23 ∧ ω 1 + Φ31 ∧
ω 2 = R ω 1 ∧ ω 2 ∧ ω 3 . We will use this identity to set up an EDS whose
3-dimensional integral manifolds correspond to hypersurfaces of constant
scalar curvature R0 in S 4 .
We will work on F = FSon4  O(5). Let (e0 , e1 , e2 , e3 , e4 ) represent a
frame with basepoint e0 . Recall that deα = eβ ωαβ , 0 ≤ α, β ≤ 4, ωβα = −ωαβ ,
and dωβα = −ωγα ∧ ωβγ on F . The forms ω i = ω0i furnish a basis of the semi-
basic forms on F , and ωji , 1 ≤ i, j ≤ 4, are the Levi-Civita connection forms
for the standard metric on S 4 .
Along a hypersurface in S 4 we can choose a framing so that e4 is always
normal to the hypersurface; this will give a section of F which is a three-
dimensional integral manifold of the 1-form ω 4 , and to which Ω = ω 1 ∧ω 2 ∧ω 3
restricts to be nonzero. (This will be our independence condition.) From
now on, let i, j, k be indices that run from 1 to 3. Since dω i ≡ −ωji ∧
ω j mod ω 4 , and the ωji are skew-symmetric, the connection forms φij for the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 236 8. Cartan-Kähler III: The General Case

Levi-Civita connection on the hypersurface coincide with ωji . The curvature

forms are given by
dωji + ωki ∧ ωjk = ωi4 ∧ ωj4 + ω i ∧ ω j .
Thus, framings along hypersurfaces of constant scalar curvature R0 will also
be integral manifolds of the 3-form
(8.10) Υ = ω14 ∧ ω24 ∧ ω 3 + ω24 ∧ ω34 ∧ ω 1 + ω34 ∧ ω14 ∧ ω 2 − (R0 − 3)ω 1 ∧ ω 2 ∧ ω 3 .
Since dω 4 ≡ −ωi4 ∧ ω i mod ω 4 , we will let I = {ω 4 , ωi4 ∧ ω i , Υ}alg . (Although
we don’t care about forms in I of degree higher than three, one can check
that dΥ ≡ 0 mod I.)
We must calculate the fiber dimension of V3 (I, Ω) before applying Car-
tan’s Test. As explained in Remark 8.4.5, we may ignore the Cauchy
characteristics, which are dual to the connection forms ωji . An integral
3-plane E will be determined by ωi4 |E = hij ω j , where the hij are the en-
 a symmetric 3 × 3 matrix h. 1Substituting
tries of this into (8.10) gives
Υ = ( i,j hii hjj − hij hji ) − (R0 − 3)) ω ∧ ω ∧ ω . So, V3 (I, Ω) is defined
2 3

by
(8.11) (tr h)2 − tr(h2 ) = R0 − 3.
Differentiation shows that V3 (I, Ω) is a smooth submanifold at points where
the entries of h are not all zero. Its fiber dimension is five.
Let E be a smooth point of V3 (I, Ω). Let v1 , v2 , v3 be the basis of E dual
to the restrictions of ω 1 , ω 2 , ω 3 to E, let Ei ⊂ E be spanned by v1 , . . . , vi ,
and of course let E0 be the zero subspace. Recall that H(E0 ) is just the
kernel of the 1-forms in I; H(E1 ) is cut out by these 1-forms and by those
obtained by feeding the vector v1 into generator 2-forms of I (cf. Exercise
8.1.7.5). Since
v1 (ωi4 ∧ ω i ) ≡ −ω14 mod ω 1 , ω 2 , ω 3 ,
we get one additional 1-form, and s1 = codim H(E1 ) − codim H(E0 ) = 1.
Similarly, v2 (ωi4 ∧ ω i ) ≡ −ω24 mod ω i , so s2 ≥ 1. But
(e1 , e2 ) Υ ≡ (h11 + h22 )ω34 mod ω 1 , ω 2 , ω 3 , ω14 , ω24 ,
so if h11 + h22 = 0, we get s2 = 2 for this flag. Following Remark 8.4.5, we
ignore Cauchy characteristics and take s3 = 0. Then s1 + 2s2 + 3s3 = 5,
and by the Cartan-Kähler Theorem we can construct an integral manifold
through E.
We now address the case when h11 +h22 = 0. If we can obtain an integral
manifold that has second fundamental form matrix h with respect to a frame
f ∈ F , then by rotation we can obtain an integral manifold through any
other point in the Cauchy characteristic leaf that contains f , and the second
fundamental form matrix there will be ghg −1 , g ∈ SO(3). So, it is enough
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.5. More examples of Cartan’s Test 237

to establish existence of integral manifolds assuming that h is diagonal. In

that case, if we cannot arrange that h11 + h22 = 0 by cyclically permuting
indices, then all three of h11 + h22 , h11 + h33 and h22 + h33 must be zero,
and this implies that h = 0.
We have proved:
Theorem 8.5.5. Given p ∈ S 4 , E ∈ G(3, Tp S 4 ), and a nonzero symmetric
3×3 matrix h that satisfies (8.11), there exists a three-dimensional subman-
ifold N ⊂ S 4 with constant scalar curvature R0 , tangent to E, with second
fundamental form h at that point.
We see by Theorem 8.3.3 that the construction of N involves choosing
s1 = 1 functions of one variable, and s2 = 2 functions of two variables.
Example 8.5.6 (Lagrangian and special Lagrangian submanifolds). Let ω
be the standard symplectic form on R2n :
ω = dx1 ∧ dy 1 + . . . + dxn ∧ dy n .
An n-dimensional submanifold is Lagrangian if it is an integral manifold of
I = {ω}alg .
Given E ∈ Vn (I), we can make a linear change of coordinates (while
keeping the form of ω) so that E is annihilated
 by dy 1 , . . . , dy n . Any nearby
integral n-planes are given by dy j = k sjk dxk for sjk = skj . Therefore, 
any such E is Kähler-ordinary and the fiber dimension of Vn is n+1 2 .
Let e1 , . . . , en ∈ E be dual to dx1 , . . . , dxn . Then
 we find that s1 = 1,
s2 = 1, . . . , sn = 1. Since s1 + 2s2 + . . . + nsn = n+1 2 , we have involutivity,
and integral manifolds depend on one function of n variables. (They can be
explicitly constructed by setting y j = ∂f /∂xj for f an arbitrary function of
x1 , . . . , xn .)
The preceding computation is a warm-up for applying Cartan’s Test to
an EDS for special Lagrangian submanifolds. Recall from §6.12 that these
are submanifolds whose tangent spaces belong to the face of the calibration

α = Re dz 1 ∧ · · · ∧ dz n ,
where z j = dxj + idy j , and the face is defined to be the set of unit volume
n-planes on which the calibration has value one.
Some calibrations α have a complementary form αc such that
|α(E)|2 + |αc (E)|2 = 1
for all unit volume planes E (see [91, Theorem 7.104]). Thus, E ∈ Face(α)
if and only if αc | E = 0. In the special Lagrangian case,

αc = Im dz 1 ∧ · · · ∧ dz n .
Thus, an EDS for special Lagrangian submanifolds is I = {ω, αc }alg .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 238 8. Cartan-Kähler III: The General Case

Given E ∈ Vn (I), we can again change coordinates so that E is annihi-

lated by dy 1 , . . . , dy n . Again taking e1 , . . . , en ∈ E to be dual to dx1 , . . . , dxn ,
we have s1 = s2 = . . . = sn−2 = 1. However,
)
ω ≡ dxn−1 ∧ dy n−1 + dxn ∧ dy n
mod dy 1 , . . . , dy n−2
αc ≡ dx1 ∧ · · · ∧ dxn−2 ∧ (dxn−1 ∧ dy n − dxn ∧ dy n−1 )

shows that sn−1 = 2 and sn = 0. Since the requirement that αc | E = 0

is one additional equation on the set of Lagrangian n-planes, the fiber of
Vn (I) has dimension n+12 − 1, and the system is again involutive, with
solutions depending on two functions of n − 1 variables. (Thus, we recover
our observations from §6.12.)
Exercise 8.5.7: Verifythat the extra equation that special Lagrangian n-
jj
planes must satisfy is j s = 0.

Example 8.5.8 (Associative submanifolds). Recall that the 14-dimensional

compact Lie group G2 arises as the automorphism group of the algebra O
of octonions (see §A.5), and leaves invariant a 3-form φ on R7 = ImO.
This φ, which is defined by (A.11) in terms of the octonionic multiplication
multiplication on the vector space and extended to a 3-form on the space,
is a calibration on R7 , and its complement is φc = 12 |Im ((xy)z − (zy)x)| for
x, y, z ∈ ImO. Since

ψ(x, y, z) := Im ((xy)z − (zy)x) = [x, y, z],

where [x, y, z] denotes the associator (see Exercises A.5.3), we see that
Face(φ) consists of 3-planes in R7 to which the octonionic multiplication
restricts to be associative.
We define an EDS I for associative submanifolds by taking the com-
ponents of the ImO-valued 3-form ψ as generators. (Since ψ is constant-
coefficient, all of these generators are closed.) However, we can avoid a
lengthy character calculation by using the fact that G2 acts transitively on
Face(φ), as follows:
Choose the basis e1 , . . . , e7 for R7 to coincide with the basis 1 , . . . , 7
for ImO in §A.5; then g2 has the form (6.30). Without loss of generality
we may assume that E = {e1 , e2 , e3 } ∈ V3 (I). Then the stabilizer of E
in G2 is six-dimensional. Thus, V3 (I) is smooth of codimension four in
G(3, R7 ). On the other hand, for any flag in E, c0 = c1 = 0 and c2 = 4 (two
independent vectors in E determine the third one by multiplication). Thus,
by Theorem 8.4.1 we have I involutive at E (hence involutive everywhere,
by homogeneity). Integral manifolds depend on s2 = 4 functions of two
variables. (Note that this system has the same character and Cartan integer
as the Pfaffian system for associative submanifolds discussed in §6.12.)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 8.5. More examples of Cartan’s Test 239

Exercises 8.5.9:
1. Recall that, if x is a point in Euclidean space, lying on a surface S with
principal curvatures k1 , k2 , then x + (1/ki )e3 , i = 1, 2, is a point on one of
the two focal surfaces of S. Show that, for the surfaces in Example 8.5.1,
one of the focal surfaces degenerates to a curve. Show that the surfaces are,
in fact, canal surfaces (see [178]) of radius 1/k0 about this curve, and use
this to interpret the “two functions of one variable” in another way. 
2. Give an example of a surface whose principal curvatures k1 , k2 are con-
stant along their respective lines of curvature. Set up and investigate an
EDS for these surfaces. (Hint: You should introduce k1 and k2 as new
variables, and define your EDS on F × R2 .) 
3. Show that, if we don’t ignore Cauchy characteristics in applying Cartan’s
Test to Example 8.5.3, the system is still involutive but now s2 = 1. Explain
how the integral surfaces depend on one function of two variables. (Hint:
How does the frame change as we move along the Cauchy characteristics?)
4. Set up and investigate an EDS for hypersurfaces in S 4 that have con-
stant scalar curvature and are minimal—that is, the trace of the second
fundamental form is zero. Can this be done without introducing the hij as
new variables? 
5. Set up and investigate a non-Pfaffian EDS for hypersurfaces in S 5 that
are Einstein manifolds [15].
6. Set up and investigate a non-Pfaffian EDS for coassociative submanifolds
of E7 .

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/09

Chapter 9

Geometric Structures
and Connections

In this chapter we study the equivalence problem for geometric structures.

That is, given two geometric structures (e.g., pairs of Riemannian manifolds,
pairs of manifolds equipped with a collection of foliations, etc.), we wish to
find differential invariants that determine existence of a local diffeomorphism
preserving the geometric structures. We begin in §9.1 with the example of
3-webs in the plane and define G-structures. In §9.2 we introduce connec-
tions on frame bundles as a step towards finding differential invariants of
G-structures. In §9.3 we continue our discussion of differential invariants,
giving an overview of Cartan’s algorithm for solving the equivalence prob-
lem. One may also view connections as a method of differentiating sections
of vector bundles. We present this approach in §9.4. We bring the two
approaches to connections together in §9.5. In §9.6 we study a natural gen-
eralization of Killing vector fields, called G-Killing vector fields. In §9.7 we
define and discuss the holonomy of a connection, paying special attention to
the Riemannian case. We conclude in §9.8 with an extended example of the
equivalence problem, finding the differential invariants of path geometries in
the plane.

9.1. G-structures
In this section we present two examples of G-structures and then give a
formal definition.

241
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 242 9. Geometric Structures and Connections

First example: 3-webs in R2 .

First formulation of the question. Let L = {L1 , L2 , L3 } be a collection of
three pairwise transverse foliations of an open subset U ⊆ R2 . Such a
structure is called a 3-web; see Figure 1.

Figure 1. U is the region inside the circle

Let L̃ = {L̃1 , L̃2 , L̃3 } be another 3-web on an open subset Ũ ⊂ R2 .

Problem 9.1.1. When does there exist a diffeomorphism φ : U → Ũ such
that φ(Lj ) = L̃j ?

If there exists such a φ, we will say the webs L, L̃ are equivalent. We

would like to find differential invariants that determine when two webs are
equivalent, as we did for local equivalence of submanifolds of homogeneous
spaces in Chapter 1 and local equivalence of Riemannian manifolds in Chap-
ter 3.
For example, let L0 be the 3-web
L01 = {y = const}, L02 = {x = const}, L03 = {y − x = const};
call this the flat case. When is a 3-web locally equivalent to the flat case?
Second formulation of the question. Let y  = F (x, y) be an ordinary differ-
ential equation in the plane. Let y  = F̃ (x, y) be another.
Problem 9.1.2. When does there exist a change of coordinates ψ : R2 → R2
of the form ψ(x, y) = (α(x), β(y)) so that solutions to one ODE are carried
to solutions of the other, i.e., such that ψ ∗ F̃ = (β  /α )F ?
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.1. G-structures 243

In particular, given y  = F , is it equivalent to y  = 1 via a change of

coordinates of the form of ψ?
Exercise 9.1.3: Determine local equivalence of first-order ordinary differ-
ential equations in the plane y  = F (x, y) under arbitrary changes of coor-
dinates. E.g., which equations are equivalent to y  = 1? 

Problems 9.1.1, 9.1.2, assuming F (x, y) is nonvanishing, are equivalent

because given any two transverse foliations, there exist local coordinates
x, y, such that the foliations are the level sets of the functions x and y,
and the space of integral curves of an ODE in coordinates provides the
third foliation. The diffeomorphisms of R2 that preserve the two coordinate
foliations are exactly those of the form of ψ in Problem 9.1.2.
In order to study the local equivalence of webs, we would like to associate
a coframe to a 3-web. For example, we could take a coframe {ω 1 , ω 2 } such
that
(a) ω 1 annihilates L1 ,
(b) ω 2 annihilates L2 , and
(c) ω 1 − ω 2 annihilates L3 .
In the case of an ODE in coordinates, we could similarly take ω 1 =
F (x, y)dx, ω2 = dy, again assuming F (x, y) is nonvanishing.
Remark 9.1.4. We are imitating the flat model L0 on the infinitesimal
level. This is what we did in Chapter 3 for Riemannian geometry when
we took a basis of the cotangent space corresponding to the standard flat
structure on the infinitesimal level. Just as any Riemannian metric looks
flat to first order, so does any 3-web in the plane.

Just as with choosing a frame for a submanifold of a homogeneous space,

we need to determine how unique our choice of adapted frame is, and we
will then work on the space of adapted frames. Any other frame satisfying
conditions (a), (b), (c) must satisfy
ω̃ 1 = λ−1 ω 1 ,
ω̃ 2 = μ−1 ω 2 ,
ω̃ 1 − ω̃ 2 = ν −1 (ω 1 − ω 2 ),
for some nonvanishing functions λ, μ, ν. Combining these three conditions,
we see that λ = μ = ν. Let FL ⊂ FGL (U ) be the space of coframes satisfying
(a), (b), (c), a principal bundle with fiber the group R∗ .
To each point of the fiber of FL at p, we can associate a dual frame
f = (p, e1 , e2 ) to ω̃ 1 , ω̃ 2 . Then e1 is tangent to L2,p , e2 is tangent to L1,p ,
and e1 + e2 is tangent to L3,p .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 244 9. Geometric Structures and Connections

Fixing a section (ω 1 , ω 2 ) of FL , we may use local coordinates (x, y, λ)

on FL . On FL we have tautological forms
 1  −1  1 
ω λ 0 ω
:= .
ω2 0 λ ω2
Exercise 9.1.5: Show that these tautological forms are the pullbacks of
the tautological forms on FGL (U ). Thus, they are independent of our initial
choice of ω 1 , ω 2 .

We have two 1-forms, but dim FL = 3 so we seek a third 1-form to obtain

a coframing of FL , imitating the situation of Riemannian geometry:
 1    1  1
ω −2 dλ 0 ω −1 dω
d = −λ ∧ +λ .
ω2 0 dλ ω2 dω 2
Since λdω j is semi-basic for the projection to R2 , we may write λdω j =
T j ω 1 ∧ ω 2 for some functions T 1 , T 2 : FL → R. Let θ = dλλ ; our equations
now have the form
 1    1  1 1 
ω θ 0 ω T ω ∧ ω2
(9.1) d =− ∧ + .
ω2 0 θ ω2 T 2ω1 ∧ ω2

In analogy to the situation in §6.5, we will refer to the terms T 1 , T 2 as

“apparent torsion”. More precisely, as we will see in §3, this is the torsion
of θ. The forms ω 1 , ω 2 , θ give a coframing of FL , but θ is not uniquely
determined. The choice of θ satisfying (9.1) is unique up to modification
by ω 1 , ω 2 : any other choice must be of the form θ̃ = θ + aω 1 + bω 2 . In
particular, if we choose θ̃ = θ − T 2 ω 1 + T 1 ω 2 , our new choice has the effect
that the apparent torsion is zero. Moreover, there is a unique such form
θ̃. So, renaming θ̃ as θ, we have an analog of the fundamental lemma of
Riemannian geometry:
Proposition 9.1.6. Let L be a 3-web on a two-dimensional manifold M .
There exists a unique form θ ∈ Ω1 (FL ) such that the equations
 1    1
ω θ 0 ω
d =− ∧
ω2 0 θ ω2
are satisfied.
Any choice of θ such that the derivative of the tautological forms is of
the form (9.1) is called a connection (or connection form), and a choice of θ
such that the torsion of θ is zero is called a torsion-free connection.
Remark 9.1.7. LetG = {l Id2 | l ∈ R∗ } ⊂ GL2 R denote the fiber group
θ 0
of FL → M . Then ∈ Ω1 (FL , g), furthering the analogy with Rie-
0 θ
mannian geometry (e.g., Lemma 3.1.4).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.1. G-structures 245

Definition 9.1.8. Let φ : U → Ũ be a diffeomorphism such that φ(Lj ) =

L̃j . Define the induced diffeomorphism of the frame bundles Φ : FGL (Ũ ) →
FGL (Ũ ). (I.e., Φ takes a frame to its pushforward under φ∗ .)

The canonical coframing (ω 1 , ω 2 , θ) that we have constructed on FL

enables us to begin to solve Problem 9.1.1:
Corollary 9.1.9. If Φ is an induced diffeomorphism of coframe bundles,
then Φ(FL ) = FL̃ and Φ∗ (ω̃ 1 , ω̃ 2 , θ̃) = (ω 1 , ω 2 , θ).
Exercise 9.1.10: Prove Corollary 9.1.9.

Necessary conditions for the existence of such a diffeomorphism will come

from differential invariants, which we obtain by differentiating the coframe
(ω 1 , ω 2 , θ). To calculate dθ, we compute

0 = d2 ω 1 = −dθ ∧ ω 1 − θ ∧ (θ ∧ ω 1 ) = −dθ ∧ ω 1 ,
0 = d2 ω 2 = −dθ ∧ ω 2 − θ ∧ (θ ∧ ω 2 ) = −dθ ∧ ω 2 .

If α ∈ Ω1 (FL ) is any 1-form, then dα = Aθ ∧ ω 1 + Bθ ∧ ω 2 + Cω 1 ∧ ω 2 for

some functions A, B, C : FL → R. Applying this to α = θ, dθ ∧ ω j = 0
implies
dθ = Kω 1 ∧ ω 2
for some function K : FL → R.
Exercise 9.1.11: Show that the form dθ is basic (i.e., well-defined on U ).
It is called the Blaschke-Chern curvature form of the web.
Exercise 9.1.12: Calculate the Blaschke-Chern curvature form for the fol-
lowing webs:
(1) {x = const, y = const, x − Cy = const}, where C is a constant.
(2) {x = const, y = const, x/y = const}.
(3) {x = const, y = const, xn + y n = const}.
(4) The web corresponding to the ODE y  = x + y.
(5) {x = const, y = const, H(x, y) = const} (the general case). 
(6) The general ODE case y  = F (x, y). 

Exercise 9.1.11 implies that while the function K is not well-defined on

U , the property K = 0 is.

Remark 9.1.13 (Hexagonality). Recall from Remark 2.5.6 that one can
interpret the Gauss curvature of a surface in E3 in terms of a limit of ratios
of areas. There is a similar geometric interpretation of the Blaschke-Chern
curvature, in terms of the following construction (see Figure 2):
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 246 9. Geometric Structures and Connections

From a fixed x ∈ U , travel along an L1 -leaf to a nearby starting point

a. Turning left, travel along the L3 -leaf through a until you reach point b
on the L2 -leaf through x, and then turn left onto the L1 -leaf through b.
Continue in this fashion, turning left from an Lk -leaf to an Lk+1 -leaf when
you encounter the Lk−1 -leaf through x (subscripts being understood modulo
3), until you return to a point g on the portion of the L1 -leaf through x
containing a. Then points a and g will coincide (and the figure a − g will be
closed) for all starting points sufficiently close to x if and only if K ≡ 0. In
the case of the standard flat model, the figure will be a hexagon. Moreover,
the failure of the figure to close up is measured by K; see [153, p. 164].

L
3 L
2
L3 L3
L2

L2
c
L
b 1
d

x L
1
g a
e L1
f

Figure 2. A nonhexagonal 3-web

Second example: Riemannian geometry. Let (M n , g) be a Riemann-

ian manifold. We have FO(V ) ⊂ F , the bundle of orthonormal coframes, and
the fundamental lemma of Riemannian geometry says there exists a unique
θ ∈ Ω1 (FO(V ) , so(V )) such that dη = −θ ∧ η, where η ∈ Ω1 (FO(V ) , V ) is the
tautological form. Our differential invariants all come from Θ = dθ + θ ∧ θ
and its derivatives.

Definition of G-structures. We seek a definition of a G-structure on a

manifold M that generalizes our two examples. First we generalize the
example of Riemannian geometry:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.1. G-structures 247

Let V != Rn with its standard basis, and GL(V ) = GL(n, R). Let
T(V ) := ⊗k ⊗ V ∗⊗l . Let ψ ∈ T(V ) be a tensor (e.g. ψ ∈ S 2 V ∗ or
k,l V
∗
ψ ∈ Λ V ) and let G = Gψ ⊂ GL(V ) be the subgroup preserving ψ. (In
2

other words, if ρ : G → GL(T(V )) is the induced action of G on T(V ), then

Gψ = {g ∈ GL(V ) | ρ(g)ψ = ψ}.) For example, taking
Q ∈ S2V ∗,
ω ∈ Λ2 V ∗ ,
J ∈ V ∗ ⊗V such that J 2 = − Id
respectively define the orthogonal group GQ = O(V, Q), the symplectic
group Gω = Sp(V, ω), and the complex linear group GJ = GL(V, J) =
GL(m, C) for m = 2n. (Q and ω are assumed to be nondegenerate.)
Recall from §2.6 that over any differentiable manifold M n we have the
general coframe bundle π : FGL (M ) → M with points f = (x, u), where
x ∈ M and u : Tx M → V is a linear isomorphism. There is a right GL(V )-
action on the fibers of FGL defined by
(9.2) Rg (x, u) := (x, g −1 ◦ u).

Given ψ ∈ T(V ), let ψ̃ ∈ Γ(M, T(T M )) be a smooth section such that

at each point x ∈ M , there exists a coframe u for which the induced map
uT : T(Tx M ) → T(V ) takes ψ̃x to ψ. (For example, if ψ ∈ S 2 V ∗ is an inner
product, then ψ̃ is a Riemannian metric and u is an isometry.) Using ψ̃, we
define a reduction of FGL to a subbundle
FG = {(x, u) ∈ FGL | uT (ψ̃) = ψ}.
In other words, the tensor field ψ̃ enables us to define a class of preferred
coframes adapted to the geometry induced by ψ̃. For example, taking ψ
to be Q, ω or J as above makes M into a semi-Riemannian manifold, an
almost symplectic manifold or an almost complex manifold respectively. (A
semi-Riemannian manifold is Riemannian if the quadratic form is positive
definite; almost complex manifolds are discussed in Example 9.2.10 and
Appendix C.)
Our first example, web geometry, is not naturally described as arising
from a tensor, but it can be described in terms of the bundle of G-frames
on M (where G is the group defined in Remark 9.1.7), so we use that as the
basis of our definition of a G-structure.
Definition 9.1.14. A bundle B → M is a right G-bundle if G has a smooth
right action on B whose restriction to the fibers is simply transitive, i.e., any
two points in the same fiber are related by a unique element of G.
Definition 9.1.15. Let G ⊂ GLn be a matrix Lie group. A G-structure on
a differentiable manifold M n is a right G-bundle FG ⊂ FGL .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 248 9. Geometric Structures and Connections

An alternative definition of a G-structure, used in defining EDS for man-

ifolds with exceptional holonomy in §9.7, is as follows: First note that any
subgroup G ⊂ GLn acts on FGL by restriction of (9.2), so we may form the
quotient bundle FGL /G.

Definition 9.1.16. A G-structure on M is a smooth section s : M →

FGL /G.

Let πG : FGL → FGL /G denote the projection. Then FG = π −1 G (s)

determines a bundle of G-frames over M , called a reduction of FGL to FG .
Given FG , we can recover the section s because πG (FG ) is a section of
FGL /G.
We generalize Definition 9.1.8:
Definition 9.1.17. Given a diffeomorphism φ : M → M  , we obtain an
induced diffeomorphism Φ : FGL (M  ) → FGL (M ); given u : Tφ(x) M  → V ,
we obtain u = Φ(u ) := u ◦ φ∗x : Tx M → V . If G ⊂ GL(V ), then we
similarly get an induced identification of FGL (M )/G with FGL (M  )/G.

Definition 9.1.18. Two G-structures FG ⊂ FGL (M ) and FG  ⊂ F (M  )

GL

are equivalent if there exists a diffeomorphism φ : M → M such that the
induced identification of FGL (M ) with FGL (M  ) takes FG to FG . In other

words, s : M → FGL (G)/G and s : M → FGL (M  )/G are equivalent if

and only if there exists a diffeomorphism φ : M → M  such that under the
induced identification Φ : FGL (M  )/G → FGL (M )/G, we have s = Φ(s ).

Definition 9.1.19. A G-structure is flat if for all x ∈ M there exist local

coordinates x1 , . . . , xn on a neighborhood of x such that (dx1 , . . . , dxn ) is a
local section of FG .

For example, an O(n)-structure (equivalently, a Riemannian metric) is

flat if and only if there exist local coordinates x1 , . . . , xn such that dxi gives
a local orthonormal coframing of M , i.e., coordinates such that g = (dx1 )2 +
· · · + (dxn )2 .
Exercise 9.1.20: Show that an Sp(2m, R)-structure determined by a non-
degenerate 2-form ω is flat if and only if there exist local coordinates (x1 , . . . ,
x2m ) such that locally ω = dx1 ∧ dxm+1 + · · · + dxm ∧ dx2m .

9.2. Connections on FG and differential invariants of

G-structures
We now begin to address the problem of finding differential invariants of
G-structures in the general case.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.2. Connections on FG and differential invariants of G-structures 249

Recall from §2.6 that the coframe bundle FGL comes equipped with—
and FG inherits—a tautological V -valued 1-form ω defined by, for w ∈
Tu FGL (M ),
ωu (w) = u(π∗ (w)).
If one fixes a basis of V , ω furnishes a basis {ω i } for the semi-basic forms
on FG .
Exercise 9.2.1: Prove that FG and FG are equivalent if and only if there

exists a diffeomorphism Φ : FG → FG such that Φ∗ (ω  ) = ω.



As in our examples, we attempt to obtain a coframing of FG in a geo-

metrically meaningful manner, so that its derivatives will furnish differential
invariants.
To obtain a coframing of FG we need to find a complementary set of
forms to the {ω i }. We want to mimic the situation of frames for subman-
ifolds of homogeneous spaces M = K/G from Chapter 1 and frames for
Riemannian geometry in Chapter 3 as much as possible. In the case of
K/G, the complement to the set of semi-basic forms was a g-valued 1-form
(the Maurer-Cartan form of g). Recall the fiber of K  FG → M is isomor-
phic to G. For an arbitrary G-structure, the fiber of π : FG → M is also
isomorphic to G, so we seek a complement to the ω i that is g-valued. To
imitate the Maurer-Cartan form as much as possible, we will ask for a form
that “pulls back” to be the Maurer-Cartan form in the sense of Definition
9.2.2 below.
Let θ ∈ Ω1 (G, g) denote the Maurer-Cartan form of G. Fix a point
(x, u0 ) in a fiber FG,x and define a map μu0 : G → FG,x by g → Rg (u0 ) =
g −1 ◦ u0 .
Definition 9.2.2. A connection form on FG is a g-valued 1-form θ ∈
Ω1 (FG , g) such that for all (x, u0 ) ∈ FG , μ∗u0 (θ) = θ.
Exercises 9.2.3:
1. Show that if θ is a connection form, then T ∗ FG is spanned by the entries
of θ and ω.
2. Show that if θ is a connection form, and u1 , u2 ∈ (FG )x , then μ∗u0 (θ) =
μ∗u1 (θ), so the definition could have been phrased in terms of existence of a
single u0 .
Proposition 9.2.4. Let θ ∈ Ω1 (FG , g). Then θ is a connection form if and
only if it satisfies the structure equation
(9.3) dω = −θ ∧ ω + 12 T (ω ∧ ω),
where T is a function taking values in Hom(Λ2 V, V ), and we use the notation
ω ∧ ω(v, w) := ω(v) ∧ ω(w), so that ω ∧ ω ∈ Ω2 (FG , Λ2 V ). In indices, where
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 250 9. Geometric Structures and Connections

v1 , . . . , vn is a basis of V with dual basis v 1 , . . . , v n , (9.3) can be written as

dω i = −θji ∧ ω j + Tjk
i j
ω ∧ ωk ,
where θji vi ⊗v j ∈ Ω1 (FG , g) and Tjk
i = −T i .
kj

Proof. It is sufficient to work locally; let (x1 , . . . , xn ) be coordinates on an

open subset U ⊂ M . We have dx ∈ Γ(M, FGL ), and there exists a map
h : U → GL(V ) such that ω = h−1 dx ∈ Γ(U, FG ). Write (x, h−1 dx) =
(x, u0 (x)). Then FG |U  U × G with (FG )x = {(x, a−1 u0 ) | a ∈ G}. We will
use a as a matrix-valued coordinate on the fibers.
Note that μ∗u0 (a−1 da) = θ, so a−1 da is a connection form. We have
ω = a−1 h−1 dx and calculate that
dω = −(a−1 da)a−1 h−1 ∧ dx − a(h−1 dh)h−1 ∧ dx
= −(a−1 da) ∧ ω − a−1 (h−1 dh)h−1 ∧ dx.
The second term is semi-basic; thus it can be written as a linear combination
of the ω i ∧ ω j , so our connection form a−1 da satisfies the structure equation
(9.3). Moreover, any other connection form is of the form a−1 da + β for
some β ∈ Ω1 (FG , g), which is semi-basic because μ∗u0 (β) = 0, and a−1 da + β
still satisfies the structure equation (9.3).
In the other direction, any g-valued form satisfying (9.3) must be of the
form a−1 da + β, and is therefore a connection form. 
i v ⊗ vj ∧ vk ∈ F 2 ∗
Definition 9.2.5. Tθ := Tjk i G → V ⊗ Λ V is called the
torsion of the connection θ, and will sometimes be denoted simply by T .
Remark 9.2.6. The analogue of Tθ in the context of linear Pfaffian systems
was called apparent torsion in Chapter 6.

We are less interested in Tθ than in the part of the torsion that is in-
dependent of our choice of θ. Recall from §6.5 that the torsion of a linear
Pfaffian system with tableau A ⊆ W ⊗V ∗ lies in
H 0,2 (A) = (W ⊗Λ2 V ∗ )/δ(A⊗V ∗ ),
where δ : (W ⊗V ∗ )⊗V ∗ → W ⊗Λ2 V ∗ is the skew symmetrization map. In
our case W = V and A = g.
Definition 9.2.7. Let FG be a G-structure on M , θ a connection form and
Tθ ∈ Γ(FG , V ⊗Λ2 V ∗ ) the torsion of θ. Let π : V ⊗Λ2 V ∗ → H 0,2 (g) denote
the quotient map and write [Tθ ] for π ◦ Tθ . Then [Tθ ] : FG → H 0,2 (g) is
called the torsion of the G-structure, or the intrinsic torsion of FG . A G-
structure whose intrinsic torsion is zero is called 1-flat, because it resembles
a flat G-structure to first order.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.2. Connections on FG and differential invariants of G-structures 251

Exercises 9.2.8:
1. If g ⊆ so(V ), then H 0,2 (g) = (so(V )/g)⊗V ∗ .
2. Verify that [Tθ ] is independent of θ, i.e., if θ is another connection form,
then [Tθ ] = [Tθ ].
3. Show that if FG → M is any G-structure, then for any fixed x ∈ M ,
there exist local coordinates (xi ) such that, (dx1 , . . . , dxn )|x ∈ (FG )x , i.e.,
there exists a section ρ of FG such that ρx = (dx1 , . . . , dxn )|x . If FG is 1-flat,
then moreover there exists a connection form θ such that, ρ∗ (θ)x = 0. (We
cannot have dθx = 0 unless FG is 2-flat at x; see Remark 9.2.14 below.)

To determine the uniqueness of a choice of connection with given torsion

(in particular, zero torsion), we need to calculate the changes of connection
that do not alter the torsion Tθ . Such changes are tensors that take value
in the prolongation:
g(1) = ker δ = (g⊗V ∗ ) ∩ (V ⊗S 2 V ∗ ).

We saw in §6.5 that the prolongation of a tableau A admits the inter-

pretation as the modifications of the forms πia that do not affect the torsion.
In this context these are the entries of β in the proof of Proposition 9.2.4.
Remark 9.2.9. We may rephrase the fundamental lemma of Riemannian
geometry as
H 0,2 (so(V )) = (V ⊗Λ2 V ∗ )/δ(so(V )⊗V ∗ ) = 0
and
so(V )(1) = (so(V )⊗V ∗ ) ∩ (V ⊗S 2 V ∗ ) = 0,
which respectively show existence and uniqueness of a torsion-free connec-
tion.
Example 9.2.10 (Almost complex manifolds). Recall the notation from
Appendix C and §A.3. Let V = R2m have an almost complex structure J
and consider the real Lie algebra of derivations of J, gl(V, J) = glm (C) 
V ∗(1,0) ⊗V (1,0) .
Let M 2m be an almost complex manifold. Let {ω 1 , . . . , ω m , ω 1 , . . . , ω m }
be a coframing of M with ω i ∈ Ω(1,0) (M ) and ω i ∈ Ω(0,1) (M ). Then the
almost complex structure defined by J is integrable, and there exist local
coordinates (z 1 , . . . , z m , z 1 , . . . , z m ) such that (dz 1 , . . . , dz m , dz 1 , . . . , dz m ) is
a G-framing (i.e., the structure is flat), if and only if
dω i ≡ 0 mod{ω 1 , . . . , ω m },
i.e., if and only if there exist forms θji such that

dω i = θji ∧ ω j .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 252 9. Geometric Structures and Connections

In general
dω i = Aijk ω j ∧ ω k + Bjk
i
ω j ∧ ω k + Cjk
i
ωj ∧ ωk
for some functions Aijk , Bjk i , C i . Taking θ i = Ai ω j + B i ω j eliminates
jk k jk jk
as much torsion as permissible. Thus the nonabsorbable torsion lives in
V (1,0) ⊗Λ2 V ∗(0,1) , so that H 0,2 (g) = V (1,0) ⊗Λ2 V ∗(0,1) .

Curvature of G structures. Let FG be a G-structure with [T ] = 0 and

g(1) = 0. Calculate 0 = d2 ω = (dθ + θ ∧ θ) ∧ ω and observe that Θ :=
dθ + θ ∧ θ ∈ Ω2 (FG , g) is a well-defined differential invariant, which is a
special case of the curvature of the G-structure.
Exercise 9.2.11: Using Exercise A.1.11, show that Θ is semi-basic.

We will see in Exercise 9.5.7 that if g(1) = 0, then Θ is basic and descends
to give a well-defined vector bundle valued 2-form on M .
If g(1) = 0 and θ is a torsion-free connection, then we may still define
the semi-basic form Θθ ∈ Ω2 (FG , g), but now it is just the curvature of θ,
rather than being intrinsic to the G-structure. We obtain the curvature of
the G-structure below in a manner similar to our passage from apparent to
intrinsic torsion.
Exercise 9.2.12: Show that the curvature of a connection Θθ,u : Λ2 Tu F →
gl(V ) on F = FGL (M ) → M satisfies the following equivariance: for g ∈
GL(V )
(9.4) Ad(g −1 )Θu (v, w) = ΘRg u (Rg∗ v, Rg∗ w),
where Rg : Fx → Fx is the right action on the fiber and Rg∗ : Tu F → TRg u F
is the induced action on the tangent spaces. Note that there is a g −1 in the
adjoint action so that the GL(V )-action on both sides of the equality is a
right action.
Exercise 9.2.13 (The first Bianchi identity): Let δ2 : (V ⊗V ∗ )⊗Λ2 V ∗ →
V ⊗Λ3 V ∗ denote the skew-symmetrization map. Show that
Θθ,u ∈ ker δ2 |g⊗Λ2 V ∗ = (g⊗Λ2 V ∗ ) ∩ (V ⊗S 2 1 V ∗ ),
3

where δ2 is the skew-symmetrization map.

If θ̃ = θ+β is another connection form, then Θθ̃ = Θθ +θ∧β +β ∧θ+dβ +

β ∧β. This modification, at the level of linear algebra, is by an element of the
image of the skew-symmetrization map δ1 : g(1)⊗V ∗ → g⊗Λ2 V ∗ . Combining
our two observations, just as we defined the torsion of a G-structure, define
the curvature of a G-structure as
ker δ2 : g⊗Λ2 V ∗ → V ⊗Λ3 V ∗
[Θθ ]u ∈ H 1,2 (g) := .
Licensed to AMS.
Image δ1 : g(1) ⊗V ∗ → g⊗Λ2 V ∗
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.3. Overview of the Cartan algorithm 253

Here H 1,2 (g) is the Spencer cohomology group (see Definition 6.7.3) where
we take A = g and W = V . As defined, we are viewing [Θθ ] as a function
on FG taking values in a fixed vector space, but we could equally view it as
a vector bundle valued 2-form on FG . Here (Θθ )u ∈ ker δ2 and the brackets
denote its equivalence class modulo the image of δ1 .
Remark 9.2.14. If s is a flat G-structure on M , then there exists a con-
nection θ with Tθ = 0 and Θθ = 0. A G-structure such that there exists a
connection θ such that Tθ = 0 and Θθ = 0 is called 2-flat, because at each
point the structure appears flat to order two.

9.3. Overview of the Cartan algorithm

Cartan developed an algorithm for solving the equivalence problem for G-
structures. In this section we mention some steps in this algorithm. We
return to the algorithm in §9.8 where, rather than attempting to give an
entire flowchart with lots of new definitions, we compute one example in
gory detail that will hopefully give the reader enough of an idea as to what
to do for any particular G-structure problem she or he might come across.
The Cartan algorithm is presented with more detail in [67] and [171].
What to do if H 0,2 (g) = 0 (or, more precisely, if [T ] = 0). Here we must
calculate how the torsion varies as one moves in the fiber, and choose some
normalization of the torsion. This normalization of torsion will reduce the
group of admissible changes of frame to a subgroup H ⊂ G, and we begin
again by studying H-frames. Problems can occur if the G-action on the
torsion is not nice, so we need a few definitions.
Consider the G-equivariant map
τ : (FG )x → H 0,2 (g),
u → [Tu ].
Definition 9.3.1. H ⊆ G is said to be of stabilizer type if there exists
v ∈ H 0,2 (g) such that H = {g ∈ G | μ(g)v = v}, where μ : G → GL(H 0,2 (g))
is the induced representation.

Let H be of stabilizer type, and define H 0,2 (g)H := {v ∈ H 0,2 (g) | Hv =

v}. Let H 0,2 (g)H• := {v ∈ H 0,2 (g)H | hv = v ⇒ h ∈ H}. A submanifold
S ⊆ H 0,2 (g)H• is a section if for all v ∈ S, Gv ∩ S = {v} and S is transverse
to Gv in H 0,2 (g)H• at v. A G-structure is of type S if τ (FG ) ⊆ S. These
conditions are sufficient for τ −1 (S) to be a smooth subbundle on which H
acts simply transitively on the fibers. In this case we can reduce to an H-
structure, i.e., there is a one-to-one correspondence between G-structures of
type S and H-structures FH := τ −1 (S).
Exercises 9.3.2:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 254 9. Geometric Structures and Connections

1. Show that WH• is an open dense subset of H 0,2 (g)H andWH• − > H 0,2 (g)H• .
2. Show that FH is smooth.

Prolongation: What to do if g(1) = 0. We eliminate the ambiguity in the

choice of connection form by working on a larger space, the enlargement be-
ing the addition of variables that parametrize the choice of connection form.
(This enlargement is analogous to our working on an adapted coframe bundle
instead of choosing a particular adapted coframe.) Then p1 : FG (1) → FG
is a principal bundle whose fiber is acted on simply transitively by the ad-
ditive group g(1) . More precisely, let p : F (FG ) → FG denote the coframe
bundle of FG , where we write a point of F (FG ) by (u, ψ). Points of FG (1)
are ordered pairs (u, ψ), with ψ = (p∗ (ω), p∗ (θ)), i.e., a coframe u at a point
x ∈ M , and a connection form. Recall that g(1) acts by adding an admissible
form β to θ.
We continue to differentiate on FG (1) in order to obtain a coframing of
it. In this case, assuming Tθ = 0, we define the curvature of the G-structure
at f ∈ FG (1) as an element Θf = [Θθ ]f ∈ H 1,2 (g) .
Eventually one hopes to normalize enough torsion such that one arrives
at a bundle with a geometrically determined coframing. In this case the
algorithm terminates and the derivatives of the coframing provide a complete
set of differential invariants.
We return to the equivalence problem in §9.8 and Chapter 11. First, as
mentioned in the introduction to this chapter, we discuss an alternative way
of arriving at connection forms, namely as a tool to obtain geometrically
defined derivatives of sections of vector bundles.

9.4. How to differentiate sections of vector bundles

In order to find differential invariants, we will need to be able to take deriva-
tives, not just of functions on M , but of sections of vector bundles over M as
well. As mentioned above, this goal will again lead us to connection forms.
Each perspective on connection forms will shed light on the other.
Given a manifold M and a vector-valued function f : M → W , where
W is an r-dimensional vector space, we can calculate df ∈ Ω1 (M, W ) as
follows: choose a basis v1 , . . . , vr of W , write f = f i vi where f i : M → R
are scalar-valued functions, and let df = vi ⊗df i . If we choose another basis
ṽi = aji vj , then in this basis f = f˜i ṽi = (f i (a−1 )ji )ṽj . In our new basis
df = ṽj ⊗(df i (a−1 )ji ) = vi ⊗df i , and thus df is well-defined.
Now let π : E → M be a rank r vector bundle. We would like to
differentiate sections s : M → E. Since E is a differentiable manifold,
we have dsx : Tx M → Ts(x) E, i.e., ds ∈ C ∞ (T M, T E). We would like a
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.4. How to differentiate sections of vector bundles 255

derivative whose values at x ∈ M are in the fiber Ex = π −1 (x), just as in

our previous example where E = M × W was the trivial bundle. To be
consistent with Riemannian geometry, use the notation ∇s ∈ Ω1 (M, E) for
the object we are looking for. Here are two ideas for defining ∇s:
First idea. Over contractible open sets U ⊂ M , E|U is trivial, so after
choosing a trivialization, our above definition, in a fixed basis, works. The
problem is that on the overlap U ∩ V of two such neighborhoods, we may
need to use a change of basis that depends on x ∈ M . Say on U we have
s = si vi and on V we have s = s̃i ṽi . On U ∩ V we will have ṽ i = aji (x)vj .
As a first try at differentiating s, write ∇U s = vi ⊗dsi and ∇V s = ṽi ⊗ds̃i .
Then
∇V s = ṽi ⊗ds̃i
= ṽj ⊗d(si (a−1 )ji )
= (a−1 )ji ṽj ⊗dsi + ṽj ⊗si d((a−1 )ji )
= vi ⊗dsi − vj ⊗si (a−1 da)ji
= ∇U s − (a−1 da)s.
The error term ∇U s − ∇V s = (a−1 da)s looks suspiciously like a Maurer-
Cartan form.
What we will do to eliminate the error term, and make the derivatives
of s agree on the overlaps, is to define ∇U s = dsi ⊗vi + stuffU in a way so
that stuffU − stuffV = −(a−1 da)s. In §9.5 we return to this approach and
explain what one needs to take for “stuff”.
Second idea. There is a distinguished subspace of Tp E, the vertical subspace
Vp E := {v ∈ Tp E | π∗p (v) = 0},
and the vertical subbundle VE = p∈E Vp E ⊂ T E. The space Vp E is
naturally isomorphic to the fiber Eπ(p) (recall that for z ∈ Ep , Tz Ep  Ep ).
So, if we could define a projection projvert : Tp E → Vp E  Eπ(p) , we could
define
(9.5) ∇s := projvert ◦ ds.
Given a vector space V and a subspace W ⊂ V , to define a projection
onto W we need a complementary subspace W c such that V = W ⊕ W c .
Then any vector v ∈ V can be written uniquely as v = v1 + v2 such that
v1 ∈ W, v2 ∈ W c , and the projection is v → v1 . But in general we don’t have
any naturally defined complement to Vp E. In the very special case of Rie-
mannian geometry, when E = T M , there is the Riemannian metric induced
on T (T M ) that lets us define such a complement (namely the orthogonal
complement), and one obtains a well-defined operator ∇.
This projection idea leads to the following definition:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 256 9. Geometric Structures and Connections

Definition 9.4.1. A connection on a vector bundle E is a choice of smooth

subbundle of T E complementary to VE. A connection induces a covariant
differential operator ∇ : Γ(E) → Γ(E ⊗T ∗ M ) as in (9.5).

If X ∈ Γ(T M ), we write ∇X : Γ(E) → Γ(E) for the induced differential

operator. We note that such a ∇ has the following properties:
1. ∇(s + t) = ∇s + ∇t for all s, t ∈ Γ(E).
2. ∇(f s) = s⊗df + f ∇s for all s ∈ Γ(E), f ∈ C ∞ (M ).
3. For X, Y ∈ Γ(T M ), with Xp = Yp at some p ∈ M , (∇X )p = (∇Y )p .
Exercises 9.4.2:
1. Let ∇ : Γ(E) → Γ(E ⊗T ∗ M ) be any operator satisfying properties 1, 2
and 3 above. Let xi be local coordinates on U ⊂ M and vα a collection of
local sections such that vα (x) is a basis of Ex for all x ∈ U , so a section
s : M → E may locally be written as s(x) = sα (x)vα . Show that ∇s =
vα ⊗dsα + vβ ⊗sα Γβαi dxi for some functions Γβαi (x). The functions Γiα,β are
called the Christoffel symbols for ∇.
2. Using the previous exercise, show that any operator ∇ satisfying 1, 2, 3
gives rise to a connection on E.

We now have a structure that enables us to differentiate sections of vector

bundles. This will be reconciled with our first attempt in §9.5 below. We
already see the needed correction term appearing as the Christoffel symbols.
It remains to see how to define such a structure in a way that is compat-
ible with a group action on the fibers of E. Before going into details, let’s
look at a familiar case:

Example 9.4.3. Let G ⊂ GLn+1 (R) be a matrix Lie group, and let M n =
G/H be a homogeneous space, with quotient map π : G → M . Assume
that if we write g ∈ G as g = (x, e1 , . . . , en ), where x and e1 , . . . , en are
the columns of the matrix representing g, the projection is π(g) = x and
we may identify e1 , . . . , en with a basis of Tx M . (This was the situation in
many of our examples in Chapters 1–3.) Assume that there exists a splitting
g = h ⊕ m such that m is H-invariant. (This will be the case, e.g., if H is
reductive.) Then we may write dej = ei ⊗ωji , where the ωji are components
of the Maurer-Cartan form of H.
Let s : M → G be a section and for X = xi s∗ (ei ) ∈ Γ(T M ) define

∇X = s∗ (ei )⊗dxi + xi s∗ (dei )

= s∗ (ej )⊗(dxj + xi s∗ (ωij )).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.5. Induced vector bundles and connections 257

To see that ∇X is well-defined, let s̃(x) = s(x) · a(x) be another section

where a : M → G, so s̃∗ (ei ) = aji s∗ (ej ). Recall from (1.21) that
s̃∗ (ωji ) = (a−1 da + a−1 s∗ (ω)a)ij .
Exercises 9.4.4:
1. Finish the calculation to see that ∇X is independent of the choice of s.
2. Generalize the above example to the case where G is not necessarily a
matrix Lie group.

Note that here the Christoffel symbols Γijk for the connection are defined
by s∗ (ωji ) = Γijk s∗ (ω k ). To relate this example to our first attempt, the
correction term we are adding on to the naı̈ve derivative dxi ⊗ ei is the
pullback of the Maurer-Cartan form of h. Moreover, the choice of different
sections is analogous to different local trivializations, and we see that we do
indeed get the desired cancellation.

9.5. Induced vector bundles and connections

We have defined and studied connections and differential invariants of G-
structures via FG . Our differential invariants were vector-valued functions
on FG . It would have been nicer to have had differential invariants that
were functions on M . We will now see how to push down our differential
invariants to M , with the price of having vector bundle-valued differential
invariants. At the same time we will see how a connection form on FG
induces a G-equivariant connection on T M and in fact on all vector bundles
that arise naturally from the G-action on T M :
Induced vector bundles. Given a G-structure with bundle FG , or more gener-
ally any G-principal bundle over M , and a representation μ : G → GL(W ),
we construct the induced vector bundle Eμ → M by
(9.6) Eμ := (FG × W )/ ∼,
where the equivalence relation on (u, w) ∈ FG ×W is (u, w) ∼ (Rg u, μ(g −1 )w).
Exercise 9.5.1: If ρ : G → GL(V ) is the representation of G used in
defining FG , then T M = Eρ and T ∗ M = Eρ∗ , where ρ∗ : G → GL(V ∗ )
denotes the dual representation.
Remark 9.5.2. A section of Eμ may be viewed as a function f : FG → W
satisfying the equivariance property f (Rg u) = μ(g)−1 f (u). Given a section
s, the associated function fs satisfies [(u(x), fs (u(x)))] = s(x) for all u ∈
(FG )x .
Aside 9.5.3. When M = G/P is homogeneous, the vector bundle E con-
structed using a P -module W by the procedure above is also called homo-
geneous. A very special case is when G is a complex simple Lie group and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 258 9. Geometric Structures and Connections

P = B is a Borel subgroup. (In the case of SL(n, C), the subgroup B con-
sists of the upper triangular matrices with determinant equal to one and
SL(n, C)/B is the space of complete flags in Cn . In general G/B is a space
of complete flags; see §4.2.) Then, taking one-dimensional representations
of B, one obtains holomorphic line bundles L → G/B.
Let Γ(L) denote the vector space of holomorphic sections of L. It is finite
dimensional because G/B is a compact complex manifold. Moreover, Γ(L)
is naturally a G-module. The Bott-Borel-Weil theorem gives an explicit one-
to-one correspondence between the irreducible finite dimensional G-modules
and the vector spaces Γ(L).

G-equivariant connections on induced vector bundles. The choice of a con-

nection θ on FG determines a horizontal distribution Δ ⊂ T FG , i.e., a
distribution complementary to the vertical tangent space Vu FG = {w ∈
Tu FG | π∗u (w) = 0}, namely Δu := ker θu .
Exercise 9.5.4: Show that, conversely, a G-equivariant horizontal distribu-
tion on FG determines a connection form θ.

We have seen that a horizontal distribution on a vector bundle E → M

determines a differential operator ∇ : Γ(M, E) → Ω1 (M, E). A connec-
tion on FG induces G-equivariant connections on all the Eμ by letting
q : FG × W → Eμ denote the quotient map (9.6) and defining the in-
duced horizontal distribution Δ[(u,w)] = q∗ (Δu ⊕ 0). Thus θ determines a
well-defined differential operator ∇ : Γ(M, Eμ ) → Ω1 (M, Eμ ).
We may also use the connection on FG to arrive at a differential operator
by following up on our first attempt to define a connection in §9.4. Consider
the commutative diagram
FG −−−−→ FG × W
(Id,s̃)
⏐ ⏐
⏐ ⏐
- -
M −−−−→ Eρ
s

which is uniquely defined by requiring that s̃ be equivariant, i.e., s̃(Rg u) =

ρ(g −1 )s̃(u). Then on FG × W define ∇s̃
˜ = ds̃ + s̃⊗(ρ∗ ◦ θ) ∈ Ω1 (FG × W, W ),
where ρ∗ : g → End(W ) is the induced Lie algebra representation.
Exercise 9.5.5: Show that ∇s̃˜ descends to an element of Ω1 (M, Eρ ) and
that this element is the same as ∇s defined by the induced horizontal dis-
tribution.
˜ is a vector-valued function on FG , to make things work out
While ∇s̃
right when we push down, we added on an extra term to the naı̈ve derivative.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.5. Induced vector bundles and connections 259

The correction term should come as no surprise, since in our example that
worked, Example 9.4.3, we had ∇X = ei ⊗(dxi + xj s∗ (θji )).
Proposition 9.5.6. Let FG → M be a G-structure with connection form
(θji ), and ∇ the induced connection on E = T M . Let s : M → FG be any
section, let ei be the corresponding frame field, and let X = xj ej ∈ Γ(T M ).
Then

∇X = ei ⊗(dxi + xj s∗ (θji )) ∈ Ω1 (M, T M ).

The proof is along the lines of Example 9.4.3.
A G-equivariant connection on E induces connections on all tensor con-
structs of E, e.g., S k E, Λk E, E ∗ , and E ⊗E ∗ , via the Leibniz rule. We will
use the same symbol ∇ for all these connections.
For each of the Spencer cohomology spaces H i,j (g), define the induced
vector bundles Hi,j (g) → M .
Exercises 9.5.7:
1. Show that if ∇ is a connection operator on E = T M , then for all
X, Y ∈ Γ(T M ), the maps T : Γ(Λ2 T M ) → Γ(T M ) and R : Γ(Λ2 T M ) →
Γ(End(T M )) defined by
i. T (X, Y ) := ∇X Y − ∇Y X − [X, Y ],
and
ii. R(X, Y ) := ∇X ∇Y − ∇Y ∇X − ∇[X,Y ]
are tensors, i.e., at x ∈ M they only depend on Xx , Yx . Thus T (v, w)x ,
R(v, w)x are well-defined for v, w ∈ Tx M .
2. Show that T is zero if and only if the torsion of θ is zero; in fact, identify
it with the torsion of θ. Show that T ∈ Γ(M, H0,2 (g)).
3. Identify R with the curvature of θ. Show that R ∈ Γ(M, H1,2 (g)).
4. Let ∇ be a connection on T M . Show that ∇ is torsion-free if and only
if sk(∇α) = −dα, where sk : Γ(T ∗ ⊗ T ∗ ) → Γ(Λ2 T ∗ ) denotes the skew-
symmetrization map.
5. Show that ∇ is torsion-free if and only if in any coordinate system the
resulting Christoffel symbols Γijk satisfy Γijk = Γikj ∀i, j, k. Because of this,
sometimes torsion-free connections are referred to as symmetric connections.

Induced forms. Given a G-structure FG → M and φ ∈ T(V ) that is G-

invariant, we obtain an induced form φ ∈ Γ(T(T M )) as follows: Let s :
M → FG be any section. Since s(x) provides an isomorphism Tx M  V , it
induces an isomorphism T(V )  T(Tx (M )) and we obtain a form φ = s∗ (φ)
on M . Any other choice of section would only change our choice by a G-
action, but G leaves φ invariant so φ is well-defined.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 260 9. Geometric Structures and Connections

Affine connections and projective structures. A connection ∇ on T M is

sometimes called an affine connection. Given an affine connection, it is
possible to define geodesics in M . Let c : R → M be a parametrized curve,
and let c∗ (∇) ∈ Ω1 (R, c∗ (T M )) denote the pullback of the connection from
M . Define the geodesics on M to be the parametrized curves such that
c∗ (∇) ∂ c∗ (c ) = 0.
∂t

Often one abuses notation and writes ∇c c = 0.

Define an equivalence relation on the space of torsion-free affine connec-
tions by saying ∇ is equivalent to ∇ ˜ if they have the same geodesics. A
choice of such an equivalence class on M is called a projective structure on
M.
Let’s look at the geodesic condition in local coordinates, as it will help
in the understanding of path geometry which we will study in §9.8. From
now on, for simplicity, assume ∇ is torsion-free.
Write c(t) = (x1 (t), . . . , xn (t)) and c (t) = (xi ) (t) ∂x
∂
i . We have

∂ ∂
c∗ (∇)c = (xi ) dt⊗ + (xi ) (xk ) Γjik dt⊗ j .
∂xi ∂x
So
(9.7) c∗ (∇)c c = (xi ) + (xj ) (xk ) Γijk ,
and we see that c∗ (∇)c c = 0 is a system of n second-order ODEs.
Now specialize to the case dim M = 2 with coordinates (x, y).
Exercise 9.5.8: Eliminate t from (9.7) to obtain a single ODE:
 3  2
d2 y dy dy dy
= Γ 1
22 + (2Γ 1
12 − Γ 2
22 ) + (Γ111 − 2Γ212 ) − Γ211 .
dx2 dx dx dx
Proposition 9.5.9 (Cartan [38]). A projective structure on a surface de-
termines a second-order ordinary differential equation on the surface of the
form
d2 y dy dy dy
2
= A( )3 + B( )2 + C( ) + D
dx dx dx dx
for some functions A, B, C, D on the surface. The solutions of this ODE are
the geodesics of the projective structure.
In Proposition 9.8.26 we will see a converse to Proposition 9.5.9.

9.6. Killing vector fields for G-structures

When does there exist a one-parameter group of diffeomorphisms ψt : M →
M with φ0 = Id that preserve the G-structure? When we asked this question
in the context of Riemannian geometry in §3.6.1 we concluded it would be
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.6. Killing vector fields for G-structures 261

best studied infinitesimally, and we looked for vector fields X ∈ Γ(T M )

whose flow is ψt . If our G-structure is defined by G-invariant tensors
φ1 , . . . , φm , then the condition would be that LX φj = 0 for all j.
Exercise 9.6.1: Show that if a G-structure is defined by G-invariant tensors
φ1 , . . . , φm , and LX φj = 0 for all j, and ∇ is a G-connection, then ∇X is
g-valued.

More generally, whenever we have a well-defined connection ∇ we would

expect the condition for X to be Killing to be that ∇X is g-valued.
Theorem 9.6.2 (Bryant, personal communication). Let FG → M be a
G-structure. Then

(1) The largest possible dimension of G-Killing vector fields on M is

the dimension of the total prolongation of g, V ⊕g⊕g(1) ⊕g(2) ⊕· · · .

(2) This dimension occurs for a flat G-structure (Definition 9.1.19).

More precisely, for a flat G-structure M and x ∈ M , the space of
G-Killing vector fields on M vanishing to order k at x modulo those
vanishing to order k + 1 is parametrized by g(k) .

(3) In particular, if some prolongation of g is zero, the space of G-

Killing vector fields on M is finite dimensional.
The full proof is a little beyond the scope of this book as one must use
the language of sheaves. It is easy to see that the flat model has at least
(k)
the prolongations’ worth of Killing vector fields. Form the bundles FG →
(k−1) (k )
FG with fiber g(k) and let FG 0 be the bundle corresponding to the
last nonzero prolongation. Then in the flat case M is locally a homogeneous
space of a Lie group whose Lie algebra is the total prolongation and the total
(k )
space of FG 0 is this group. Each element of the group gives rise to a local
diffeomorphism of M that preserves the G-structure. Differentiating curves
of such mappings at the identity shows we have at least these symmetries
in the flat case.
To get an idea of the rest of the proof, adopt the notation g(−1) = V ,
g(0)= g. Let KG ⊂ Γ(T M ) denote the space of Killing vector fields. We
!
need to show that there is a bijective map KG → ∞ (k)
k=−1 g . First note that
KG ⊂ Γ(T M ) is not just a subspace, but a Lie subalgebra and a subsheaf.
Moreover, given a point x ∈ M , KG comes equipped with a filtration, where
KG (k) is the subspace of G-Killing vector fields vanishing to order k at x, and
KG is a filtered Lie algebra. Let gr(KG ) denote the associated graded Lie
algebra. To finish the proof one observes that there is a map gr(KG ) → g(k)
and proves that it is bijective.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 262 9. Geometric Structures and Connections

9.7. Holonomy
In this section we only discuss the local theory, so throughout this section
(and exercises) assume M is connected and simply connected.
Given a G-structure FG → M , is there some proper subgroup H ⊂ G
that induces the “same geometry” as FG ? (This question will be made
precise below.)
One idea to make this question precise would be to ask if the curva-
ture Θ ∈ Ω2 (M, H1,2 (g)) takes values in some Ω2 (M, H1,2 (h)) for some
Lie subalgebra h ⊂ g. We will soon see that this is the correct idea (see
the Ambrose-Singer Theorem 9.7.14 below), but our definition of the corre-
sponding subgroup H will be different to facilitate proofs.
Our definition will be provided by the holonomy of a connection. Recall
that in a vector space, we may identify the tangent spaces at all points and
thus move vectors around from one tangent space to another. We can’t do
this on a manifold, even one equipped with a connection, but part of this
can be salvaged with a connection: we can identify tangent spaces to points
using curves that connect those points.
Parallel transport. Let π : FG → M be a G-structure with torsion-free
connection θ. (Here and below, this includes the case of G = GL(V ), i.e.,
when FG is the full frame bundle.) Let α : [a, b] → M be a smooth (or even
piecewise smooth) curve, let x0 = α(a) and let u0 ∈ π −1 (x0 ).
Exercise 9.7.1: Show that there exists a unique curve α̃ : [a, b] → FG such
that α̃(a) = u0 , π ◦ α̃ = α and α̃∗ (θ) = 0. 

We call α̃ the horizontal lift of α through u0 , and curves that are hori-
zontal lifts of some curve in the base will be referred to as horizontal curves.
Exercise 9.7.2: Show that if α̃(t) is a horizontal curve in FG , then β(t) =
Rg α̃(t) is also horizontal for any g ∈ G, and all horizontal lifts are of this
form. 

For those of you wondering about the origin of the terminology “connec-
tion”, parallel transport allows us to “connect” tangent spaces at different
points of M as follows: Consider the linear map

τα : Tα(a) M → Tα(b) M

defined by τα = α̃(b)−1 ◦ α̃(a) : Tα(a) M → V → Tα(b) M . This τα is called

a parallel translation along α with respect to the connection θ. It is inde-
pendent of the choice of u0 ∈ π −1 (x0 ), because if u ∈ π −1 (x0 ) and τα
is the corresponding linear map, then u = g −1 ◦ u0 for some g ∈ G and
α̃ (t) = Rg α̃(t) for t ∈ [a, b].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.7. Holonomy 263

~
α(a) ~(b)
α

Tα(a)M Tα(b) M

α(a) α(b)

Figure 3. Parallel translation

What is most interesting is what happens along a closed path. The

induced map Tx M → Tx M may not be the identity. In fact, the set of such
maps is the structure group of the following bundle:
Given u ∈ FG , let
Puθ = {u ∈ FG | ∃ α : [a, b] → M, α̃(a) = u, α̃(b) = u }.
In other words, Puθ is the set of points in FG which can be reached from
u by following horizontal lifts of curves in M . This is called the holonomy
bundle of θ through u, and its structure group is called the holonomy group
of θ:
Definition 9.7.3. For u ∈ FG , let Holθu = {h ∈ G|Rh u ∈ Puθ }, the holonomy
group of θ relative to u.
Proposition 9.7.4. Let M be simply connected. The holonomy groups are
connected subgroups of G. If u, u are connected by a horizontal curve in
FG , then Holθu = Holθu . If u, u are in the same fiber of FG , then Holθu and
Holθu are conjugate in G.
Exercise 9.7.5: Prove Proposition 9.7.4.

We may define the holonomy group independent of a reference coframe u,

if we are willing to forgo an explicit embedding of the group into the matrix
group G: Suppose α(a) = α(b) = x; then we obtain a map τα : Tx M → Tx M .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 264 9. Geometric Structures and Connections

Define
Holθx = {τα | ∃ α : [a, b] → M, α(a) = α(b) = x} ⊆ GL(Tx M ).
Let u ∈ π −1 (x) ⊂ FG . Then Holθx  Holθu under the isomorphism between
GL(Tx M ) and G induced by u : Tx M → V . Since we are working locally,
each of the groups Holθx is isomorphic, so we sometimes drop reference to x
in the notation and write Holθ .
We may understand Holθ as measuring the failure of parallel translation
along closed curves to take a vector to itself.
We now explore relations between holonomy and infinitesimal geometry:
Exercise 9.7.6: Let E = Eρ → M be an induced vector bundle and let
s : M → E be a section. Let X ∈ Tx M and let α be a curve in M with
α(0) = x and α (0) = X. Let τ t be parallel translation from x to α(t) along
α with respect to connection ∇ on E induced from the connection on FG .
Show that
1
∇X s = lim [(τ t )−1 (s(xt )) − s(x)].
t→0 t

Given a G-structure with torsion-free connection θ, our question posed

in the introduction to this section may be stated more precisely as: When
is Holθ isomorphic to a proper subgroup of G? For example, En equipped
with the Levi-Civita connection has Holθ = {Id}.
Proposition 9.7.7. Let G ⊆ GL(V ), let FG → M be a G-structure equipped
with a connection, and let W ⊆ V ⊗k⊗V ∗⊗l be a G-submodule. Let Eρ → M
be the vector bundle induced by ρ : G → GL(W ) with induced connection
∇. If there exists φ ∈ Γ(Eρ ) such that ∇φ = 0, then for all x ∈ M ,
Holθx ⊆ Gφ,x := {g ∈ GL(Tx M ) | ρ(g)φx = φx }.
We will say that such a φ is G-parallel.

Proof. The result is local, so fix a section s : M → FG through u which

determines a local G-framing e1 , . . . , en with dual coframing ω 1 , . . . , ω n . We
may write
φ = pij11,...,i
,...,jl ω ⊗. . .⊗ω ⊗ei1 ⊗. . .⊗eik .
k j1 jl

Then
∇φ =ω j1 ⊗. . .⊗ω jl ⊗ei1 ⊗. . .⊗eik
i ,...,i ,m,is+1 ,...,ik ∗ ∗ m
⊗[dpij11,...,i k 1 s−1
,...,jl + pj1 ,...,jl
is
s (θm ) − pij11,...,i k
,...,js−1 ,m,js+1 ,...,jl s (θjs )].
Now let α : [a, b] → M be a curve with α(a) = α(b) = x and α̃ its horizontal
lift. By definition, α̃∗ (θ) = 0, so if φ is G-parallel, then dpij11,...,i k 
,...,jl (α ) = 0, i.e.,
the functions pij11,...,i k ∗
,...,jl are constant on lifts of horizontal curves; so τ (φ) = φ,
and thus Holθx ⊆ Gφ,x . 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.7. Holonomy 265

Exercise 9.7.8: Show conversely that if at x ∈ M there exists φx ∈ Tx M⊗k⊗

Tx∗ M⊗l with Holθx ⊆ Gφ,x , then there exists a tensor field φ on M such that
φx = φ and ∇φ = 0.

In particular, given a flat G-structure we have invariant tensors φj = dxj

so we conclude:
Corollary 9.7.9. Flat G-structures have Hol = Id.
Example 9.7.10. Let (M 2 , g) be a simply connected surface with a Rie-
mannian metric. If the structure is not flat, then Holθ  SO(2), because
there are no connected Lie groups between the identity and SO(2).
Corollary 9.7.11. If Holθx is reductive and preserves a subspace L ⊂ Tx M ,
then Holθx ⊆ GL(L) × GL(Lc ).
Exercise 9.7.12: Prove Corollaries 9.7.9 and 9.7.11.

In the case of Riemannian geometry, something stronger is true:

Theorem 9.7.13 (de Rham). Let (M, g) be Riemannian, complete, and
simply connected. Say Tx M = L1 ⊕ L2 with each Lj preserved by Holθx .
Then (M, g) is isometric to a product (M1 × M2 , g1 ⊕ g2 ) with Tx Mi = Li .
For the proof, see [114, Chap. IV].
de Rham’s theorem implies that if G occurs as an irreducible holonomy
group (i.e., G = G1 × G2 ) of a Riemannian metric, it must act transitively
on the unit sphere S n−1 ⊂ Tx M .
We now state and prove the Ambrose-Singer Theorem relating holonomy
and curvature:
Theorem 9.7.14 (Ambrose-Singer). Let FG → M be a torsion-free G-
structure with connection θ. Fix u0 ∈ FG and let
W := {Θ(v, w) | v, w ∈ Tu Pu0 , for some u ∈ Pu0 },
where Pu0 denotes the corresponding holonomy bundle. Then W = hol, the
Lie algebra of Holθu0 .
The theorem may be rephrased without reference to Pu0 as:
Theorem 9.7.15 (Ambrose-Singer). Let M be equipped with a torsion-
free affine connection ∇ inducing curvature Θ = Θ∇ and fix x ∈ M . Given
c : [a, b] → M with c(a) = x, let τcgl : gl(Tx M ) → gl(Tc(b) M ) denote the
linear map induced from τc : Tx M → Tc(b) M , and define
W := {(τcgl )−1 Θc(b) (τc v, τc w) | v, w ∈ Tx M,
and c : [a, b] → M is a curve with c(a) = x} ⊂ gl(Tx M ).
Then W = hol, the Lie algebra of Holθx .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 266 9. Geometric Structures and Connections

Proof. First, note that W ⊆ hol because θ pulled back to Pu0 is hol-valued
(the fibers of Pu0 are Holθu0 ), and thus Θ = dθ + θ ∧ θ is as well. Next, we
show that W is an ideal in g. In general, U ⊆ g is an ideal (i.e., [U, g] ⊆ U )
if and only if AdG (U ) ⊆ U . Let w = Θ(v1 , v2 ). By (9.4)
Adg (w) = Θ(Rg−1 ∗ v1 , Rg−1 ∗ v2 ) ∈ W,
so W is indeed an ideal.
Now fix a splitting hol = W ⊕ W c and write θ = θW + θc , where θW is
W -valued and θc is W c -valued. We need to show that W c = 0. Let I be the
Pfaffian system on FG spanned by the components of θc . Then
Θ = (dθW + θW ∧ θW ) + (dθc + θc ∧ θc ) + (θW ∧ θc + θc ∧ θW ).
The left hand side, and the first and third terms in parentheses, all take
values in W (the third because W is an ideal), while dθc + θc ∧ θc is W c -
valued. Thus 0 = dθc +θc ∧θc ; in particular 0 ≡ dθc mod I, so I is Frobenius.
Let Q ⊆ Pu be the maximal integral manifold of I through u. Since any
horizontal curve in FG is an integral of I (as all of θ vanishes on it), we have
Q = Pu . Hence W c = 0. 

We will continue to discuss holonomy, but first we address a question

from Chapter 3: Which are the Riemannian manifolds with ∇Θ = 0?
Detour into symmetric spaces. Let M be equipped with a torsion-free con-
nection θ and curvature form Θ = Θθ . (In particular, M could be Rie-
mannian with θ the Levi-Civita connection.) Assume that ∇Θ = 0. In this
situation, we might expect M to be homogeneous. We will see that this is
the case and in fact something stronger is true.
The proof of Proposition 9.7.7 implies that on a holonomy bundle
Holθ −−−−→ Puθ
⏐
⏐
-
M,
the coefficients of Θ are constant. From this we will be able to construct a
Lie algebra as follows:
Define the vector space g := hol ⊕ V . We will give g the structure of
a Lie algebra. First, the bracket for the subalgebra hol is already defined.
Define the bracket [hol, V ] by the module action, [h, v] := h.v ∈ V . Finally,
to define the brackets [v1 , v2 ] with vj ∈ V , we use the curvature tensor
i (e ⊗ω j )⊗ω k ∧ ω l : let v = u(e ), 1 ≤ j ≤ n, be a basis of V with
Θ = Rjkl i j j
dual basis {v j } and define [vk , vl ] := Rjkl
i (v ⊗v j ) ∈ hol.
i
Exercise 9.7.16: Show that g with this bracket is a Lie algebra, i.e., that
the Jacobi identity holds.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.7. Holonomy 267

Let ψ := θ ⊕ ω ∈ Ω1 (P, g).

Exercise 9.7.17: Show that dψ = − 12 [ψ, ψ].

We conclude:
Theorem 9.7.18. Let M be an n-dimensional manifold equipped with a
torsion-free connection θ and curvature form Θ = Θθ such that ∇Θ =
0. Let H = Holθ = exp(hol) be the associated holonomy group. Then
locally M is isomorphic, as a manifold with H-structure, to G/H where
G = exp(hol ⊕ V ).

Proof. By the discussion above M is equipped with a g-valued 1-form ψ

satisfying dψ = − 12 [ψ, ψ]. We conclude by the intrinsic version of Theorem
1.6.13. 

The homogeneous space G/H has extra structure: Define a group auto-
morphism σ : G → G with σ 2 = e by exponentiating the map σ∗ : Te G →
Te G given by σ∗ = Idh ⊕(− IdV ). This descends to give an H-structure au-
tomorphism σ[e] : G/H → G/H that squares to be the identity and with [e]
an isolated fixed point. For all x ∈ G/H we may obtain such an automor-
phism σx : G/H → G/H, with x an isolated fixed point, by conjugating by
elements of g. In particular:
Theorem 9.7.19. Let (M, g) be a Riemannian manifold with curvature
tensor Θ. Then the following are equivalent:
(1) For all x ∈ M , there exists an open neighborhood U ⊂ M of x and
an isometry σx : U → U , such that σx2 = Id, with x an isolated
fixed point.
(2) ∇Θ = 0.
(3) Given x ∈ M , there exists a neighborhood U ⊂ M and an isometry
from U to an open subset U  ⊂ G/H, where H = exp(hol), hol is
the holonomy algebra of M , the group G is constructed as above,
and the Riemannian metric on G/H is the one induced from H.
A manifold satisfying the conditions in Theorem 9.7.19 is called locally
symmetric, and it is called symmetric, or a symmetric space, if U = M .

Proof. We have already seen most of the implications; it remains to show

that the existence of the isometries σx implies that M is locally homoge-
neous. But the σx ’s generate a group acting transitively on M , so M must
locally be of the form G/H for some groups G, H, where g has an involution
with eigenvalues +1 on h and −1 on g/h. Our discussion of the curvature
tensor in this situation implies ∇Θ = 0. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 268 9. Geometric Structures and Connections

Remark 9.7.20. These considerations reduce the classification of Riemann-

ian symmetric spaces to the classification of orthogonal involutive Lie alge-
bras [94], i.e. Lie algebras with an involutive automorphism, such that the
adjoint action of the +1-eigenspace on the −1-eigenspace respects an inner
product. See [94] for the classification.

Special holonomy. Symmetric spaces with irreducible holonomy have been

classified. As mentioned in Remark 9.7.20, their classification amounts to
classifying automorphisms of Lie algebras. So now we have a refined ques-
tion: which groups G arise as the holonomy of a nonsymmetric manifold
with an affine connection?
We need to have curvature Θ such that ∇Θ = 0; in particular, the space
where ∇Θ lives cannot be empty.
Exercise 9.7.21: If holθ ⊆ g, prove the second Bianchi identity that
ker δ2 : g⊗Λ2 V ∗ ⊗V ∗ → g⊗Λ3 V ∗
∇Θu ∈ H 2,2 (g) = .
Image δ1 : g(1) ⊗V ∗ → g⊗Λ2 V ∗ ⊗V ∗
In summary, we have two necessary conditions for G to be the holonomy
of a nonsymmetric space:
(1) There does not exist a proper subalgebra g ⊂ g with H 1,2 (g) =
H 1,2 (g ) (otherwise the holonomy would have to be contained in g
by the Ambrose-Singer Theorem).
(2) H 2,2 (g) = 0.
Note that for Riemannian geometry the first condition is automatic by
Exercise 9.2.8.1. Also in the Riemannian case, thanks to de Rham’s theorem,
we have the additional condition that G acts transitively on the unit sphere
S n−1 ⊂ V .
Theorem 9.7.22 (Berger, Alexseevski, Brown-Gray; see, e.g. [15]). The
Lie algebras g ⊆ so(n) with associated Lie groups acting transitively on the
unit sphere and with H 2,2 (g) = 0 are:
1. so(n),
2. u(n/2),
3. su(n/2),
4. sp(1) + sp(n/4),
5. sp(n/4),
6. g2 ⊂ so(7),
7. spin7 ⊂ so(8).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.7. Holonomy 269

Which of these occur as the holonomy of nonsymmetric spaces? They

all do.
Getting SO(n) is easy: take the standard metric on the sphere and per-
turb it a little. The holonomy group will not shrink, therefore it stays SO(n).
The same method applied naı̈vely won’t work for the other groups, as the
holonomy could jump up to SO(n) under a deformation. Nonsymmetric
Kähler metrics (case 2) exist; just take any nonhomogeneous subvariety of
complex projective space equipped with its Fubini-Study metric.
Cases 3, 4, 5 can be handled locally using a refined deformation theory;
see [15]. Cases 6 and 7 were first resolved locally by Bryant [23], and
then later complete [32] and compact [106, 107] examples were found.
We outline Bryant’s existence proof in the following subsection. It is still
not known whether or not compact examples exist for case 4. This last
problem has generated considerable interest: LeBrun and Salamon [141]
have conjectured there are no inhomogeneous compact examples. This is
only known to be true in the first possible case n = 4 [161].

Exterior differential systems for torsion-free G-structures.

A tautological system. We set up an EDS whose integral manifolds corre-
spond to torsion-free (1-flat) G-structures s : M → (FGL /G). Let π̃G :
FGL /G → M denote the projection and let H = H0,2 (g) = FGL × H 0,2 (g).
G
To calculate the torsion, we don’t need a G-structure, but merely its
1-jet. Let H denote the pullback of H to J 1 (M, FGL /G). We have a natural
map
τ : J 1 (M, FGL /G) → H
such that τ (j 1 (s)) = [Ts ]. Let Σ = τ −1 (0) ⊂ J 1 (M, FGL /G), and let (I, J)
denote the pullback of the contact system on J 1 (M, FGL /G) to Σ. Then
the integral manifolds of the system (I, J) are exactly the lifts of 1-flat G-
structures on M to J 1 (M, FGL /G).
Similarly, define a linear Pfaffian system for 2-flat G-structures by inter-
secting the prolongation of (I, J) with r−1 (0), where r : J 2 (M, FGL /G) →
H1,2 (g) is the curvature mapping.
Torsion and invariant forms.
Proposition 9.7.23. Let s : M → FGL /G be a G-structure and assume
there exists a G-invariant element φ ∈ Λk (V ∗ ), so as seen in §9.5, it induces
a form φ ∈ Ωk (M ). A necessary condition for s to be torsion-free is that
dφ = 0.

Remark 9.7.24. We already know that there exists a connection ∇ such

that ∇φ = 0 by Exercise 9.7.8.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 270 9. Geometric Structures and Connections

Proof. First observe that if s is flat, then dφ = 0. This is because, if s

is flat, take local coordinates xi such that ρ = (dxi ) is a local section of
FG . Write φ = aI dxI (using multi-index notation). We have 0 = ∇φ =
dxI ⊗(daI + aI ρ∗ (θ)), but ρ∗ (θ) = 0.
Now if a G-structure is torsion-free, i.e., 1-flat, the above is still true to
first order, since by Exercise 9.2.8, for all x ∈ M there exist local coordinates
xi centered at x, such that at x, ρx = (dx1 , . . . , dxn ) and ρ∗x (θ) = 0. Thus
dφx = 0 for all x ∈ M . 

In terms of representation theory, the above argument implies that if

φ ∈ Λk V ∗ is G-invariant, then a component of Λk+1 V ∗ must occur in the
G-module in H 0,2 (g), as dφx ∈ Λk+1 Tx M must be zero if the connection is
torsion-free, so dφx represents a component of the torsion.

Example 9.7.25. Let G = Sp(V, φ) for φ ∈ Λ2 V ∗ nondegenerate and as-

sume dim V is even. Recall from §A.2 that

sp(V )  S 2 V ∗ ,
δ(S 2 V ∗ ⊗V ∗ ) = S21 V ∗ ,

where the notation S21 V ∗ was defined in §3.5. Since V ∗ ⊗Λ2 V ∗ = Λ3 V ∗ ⊕

S21 V ∗ as a GL(V )-module, we see that H 0,2 (sp(V )) = Λ3 V ∗ as an sp(V )-
module. Since φ is a generic form, its exterior derivative will have compo-
nents in both irreducible components of Λ3 Tx∗ M , so it accounts for all of the
torsion, giving a fancy proof of Exercise 9.1.20.

Manifolds with G2 holonomy. Recall that G2 (φ) ⊂ GL(7, R) is the group

preserving a generic positive φ ∈ Λ3 R7 (see §A.5). Write V = R7 . Since
G2 ⊂ SO(7), we may identify V with V ∗ .
By Exercise 9.2.8.1, H 0,2 (g2 ) = (so(7)/g2 )⊗V = (Λ2 V /g2 )⊗V . A modest
computation or a quick click on the computer program LiE [142] shows
(Λ2 V /g2 )⊗V  V ⊗V and that as a g2 -module, V ⊗V = V20 ⊕ g2 ⊕ V ⊕ R =
V20 ⊕ V01 ⊕ V10 ⊕ V00 . (Here we use the convention that Vij is of highest
weight iω1 + jω2 ; see [65, 19].)
Moreover, Λ4 V = V2,0 ⊕ V ⊕ R, so the exterior derivative of φ can
only fill three components of H 0,2 . However, there is another G2 -invariant
differential form, coming from ∗φ ∈ Λ4 V . We have Λ5 V  Λ2 V = V ⊕ g2 .
One needs to check that the exterior derivative of ∗φ can take values in the
g2 term in Λ4 V . (Aside: once (dφ)x is known, the component of (d(∗φ))x in
the V factor in Λ5 V is determined.)
Since we have accounted for all of H 0,2 (g2 ), i.e., all the possible contri-
butions to torsion, we conclude:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 271

Proposition 9.7.26. A G2 -structure given by φ ∈ Ω3 (M ) is torsion-free if

and only if dφ = 0 and d(∗φ) = 0.
From Proposition 9.7.26, one obtains an EDS for 7-manifolds with
torsion-free connections whose holonomy is at most G2 , as follows:
Let Σ denote the open subset of Λ3 T ∗ M consisting of nondegenerate
forms. We obtain an exterior differential system generated in degrees 4 and
5 for closed and co-closed 3-forms. Bryant [23] shows that this system is
involutive. What remains to show is that at least one (in fact most) integral
manifold has holonomy equal to G2 , which he does by examining the exterior
differential system for 2-jets of torsion-free G2 -structures and showing that
its space of local integral manifolds is strictly smaller. The proof for Spin(7)
is similar.
After Bryant’s local existence results, Bryant and Salamon found com-
plete examples [32], and soon afterward D. Joyce [106, 107] found compact
examples for both G2 and Spin(7). Joyce’s constructions use methods com-
ing from algebraic geometry (Kummer construction, deformation theory)
and analysis instead of EDS, although they are inspired by Bryant’s method.
As Bryant showed, in the G2 case one needs a compact 7-manifold equipped
with a 3-form that is closed and co-closed. Joyce begins by constructing
(with the aid of deformation theory) compact 7-manifolds equipped with a
3-form φ that is closed and such that d(∗φ) is small. He then uses analytical
tools to show that φ can be perturbed into a form that is both closed and
co-closed.

9.8. Extended example: path geometry

The reader may also find discussions of path geometry in [158] and in the
appendix to [31].
Classical formulation of the problem. Given an ODE, it would be useful to
know if it is a familiar ODE disguised by a change of coordinates. Even
better would be to have a classification of ODE’s up to some notion of
‘equivalence’ under changes of variables. We already have seen one notion
of equivalence for first-order ODE’s in the case of 3-webs in R2 . The G-
structure we now study will enable us to classify second-order ODE in R2
up to the following notion of equivalence:
Definition 9.8.1. We will say that two second-order ordinary differential
equations in R2 ,
y  = F (x, y, y  ), y  = G(x, y, y  ),
are equivalent under point transformations if there exists a diffeomorphism
Φ : R2 → R2 taking solutions of one equation to solutions of the other.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 272 9. Geometric Structures and Connections

We state our problem as follows:

Problem 9.8.2. Find invariants of second-order ODE that classify them
up to point transformations. Interpret the invariants in terms of useful
properties. For example, which second-order ODE’s admit first integrals
(and are therefore solvable by quadrature)?

Geometric formulation of the problem. Let X be a surface. A path geometry

on X is a set of smooth curves C on X such that, for all x ∈ X and for all
directions l ∈ PTx X, there is a unique γ ∈ C such that x ∈ γ and Tx γ = l.
Problem 9.8.3. Locally classify path geometries in the plane by finding
differential invariants of path geometries. Interpret the invariants in terms
of familiar path geometries, such as geodesics of a projective structure on
T X or ODE in the plane.

We will see below that every path geometry is locally equivalent to the
path geometry induced by the solutions to a second-order ODE.
Example 9.8.4. Let X have a Riemannian, or more generally a reversible
Finsler metric, and take the paths to be the geodesics.
Example 9.8.5. Let X have a projective structure and take the paths to
be the geodesics.
Example 9.8.6. Let X = R2 with the origin O omitted, and let the paths
through a point x be circles and straight lines through x and O.
Definition 9.8.7. We will call a set of paths in X flat if for any x ∈ X
there is an open set U ⊂ X, x ∈ U , and a local diffeomorphism from U to an
open set in R2 that carries the paths in U to straight line segments. In other
words, the flat case locally corresponds to solutions of the ODE y  = 0.
Problem 9.8.8. Which Riemannian metrics on a surface induce a flat set
of paths?

It is clear that the Euclidean metric on R2 induces the flat set of paths,
but are there others?
Proposition 9.8.9. S 2 with its standard metric induces the flat set of paths.

Proof. Since S 2 ⊂ R3 is homogeneous, it suffices to define a local diffeo-

morphism from a neighborhood U ⊂ S 2 of the north pole N to R2 which
carries great circles to straight lines. To do this, we use a stereographic
projection mapping a point P to the point where the line OP cuts the plane
TN S 2 (where O is the origin). 
Exercise 9.8.10: Prove that the set of paths of geodesics on hyperbolic
space H 2 is flat. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 273

A path is determined by a point p ∈ X and a tangent direction [v] ∈

P(Tp X). A path geometry is a set of second-order data on X, i.e., it does
not correspond to a foliation or other kind of distribution on X. In order
to study it using a first-order G-structure, we must work on a space such
that the path data becomes first order. Let P be the projectivized tangent
bundle P(T X) over X.

Remark 9.8.11. The problem of classifying path geometries up to equiv-

alence can be viewed as an example of a generalization of the notion of a
G-structure called a G/H-structure of order two, due also to Cartan. See
[171] where it is a special case of what is called a Cartan geometry.

Later in this section, we will set up the problem of equivalence of path

geometries as a first-order G-structure problem on P. For now, we make
several observations:
We lift each path in X to a curve in P by taking its projectivized 1-jet.
This gives a two-parameter family of curves in P which foliate P. (To see
the two parameters, one has a one-parameter family of paths through each
point of X plus two dimensions worth of points on X minus one dimension
for counting a path for each point it passes through.) Call the leaves of
this foliation L1 . We will also use L1 to refer to the distribution of tangent
spaces to the leaves. Notice that the fibration
-
S 1 = RP1 P

π
?
X

gives another foliation of P by the fibers of π. Call the leaves of this foliation
L2 and let L2 ⊂ T P also denote the associated distribution.
Suppose that the quotient of P by the leaves of L1 is a smooth surface
Z. Then Z is the space of paths in X, and we have the following double
fibration:
P
 S
/
 w
S
P/L1 = Z X = P/L2
(If the path geometry results from a second-order ODE in the plane, we may
think of Z as the space of solutions to, or initial conditions for, the ODE.)
While a point in Z corresponds to a path in X, a point x ∈ X also gives
a curve in Z (the one-parameter family of paths passing through x). The
resulting set of paths in Z is called the dual path geometry.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 274 9. Geometric Structures and Connections

The space P may be thought of as an incidence correspondence

P = {(z, x)| x ∈ some path in X corresponding to z}
= {(z, x)| z ∈ some path in Z corresponding to x}.

This notion is motivated by the following important example, which will

turn out to be our “flat” model:

Example 9.8.12. Let X = RP2 (or P2 for short) have the standard set
of straight-line paths. Then Z = P2∗ , the set of lines in P2 , and P is the
flag manifold F12 := {(x, H)|x ∈ H} ⊂ P2 × P2∗ , a homogeneous space of
SL3 R. The paths in P2∗ are the linear P1∗ ’s, so the picture is completely
symmetric. The map from SL(3, R) to pairs (x, H) is given in bases by
(e1 , e2 , e3 ) → ({e3 }, {e2 , e3 }).

Remark 9.8.13. In dimensions greater than two, the space of paths is of

higher dimension than the original space. For example, if X = Pn , then
Z = G(P1 , Pn ) = G(2, n + 1) has dimension 2(n − 1). Here P is the space
of flags, x̂ ⊂ ˆl ⊂ V , and there is no naturally induced path geometry on Z.
On the other hand, one can recover the duality by restricting to a subset of
the space of all paths on X, i.e., restricting to certain submanifolds of Z of
the same dimension as X. See [88], [12] for this story and its relevance for
the Radon transform.

We define a generalized path geometry on a 3-fold M to be the following

data: two line bundles L1 , L2 ⊂ T M such that
(a) (L1 )x ∩ (L2 )x = 0, ∀x ∈ M , and
(b) the 2-plane distribution L1 + L2 is nowhere integrable.
Here M/L1 , M/L2 might only be defined locally; they play the role of
X and Z above. Condition (b) ensures that the L1 paths do not project
to M/L2 to form a foliation, but instead there is a one-parameter family of
paths through each point of M/L1 .

Example 9.8.14. Let X be a surface with a Riemannian metric; then we

may take M = Fon (X) and L1 = {ω 2 , ω12 }⊥ , L2 = {ω 1 , ω 2 }⊥ .
Exercises 9.8.15:
1. Show that the L1 -curves are liftings of geodesics in X.
2. Show that if we lift the geodesics to define a path geometry on M  =
P(T X) in the usual way, then the two path geometries are locally diffeomor-
phic via a covering p : M → M  that commutes with the fibrations.

We will use the following as a running example:

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 275

Example 9.8.16. For an ODE y  = F (x, y, y  ), let (x, y, p) be coordinates

on J 1 (R, R) = M . Let
(9.8) L1 = {dp − F dx, dy − pdx}⊥ ,
(9.9) L2 = {dx, dy}⊥ .
Then
(9.10) {dy − pdx}⊥ = L1 + L2
is a contact distribution.
Conversely, for any path geometry on a manifold M , by the Flowbox
Theorem 1.2.2, we can find locally defined functions x, y satisfying (9.9).
Since (L1 + L2 )⊥ ⊂ L⊥ ⊥
2 , there must be functions p, q such that (L1 + L2 ) =
{qdy − pdx}, and since we are working locally, perhaps after switching the
roles of x and y, we may assume q ≡ 1, implying (9.10). By Pfaff’s Theorem
1.10.16, the nonintegrability condition on L1 + L2 implies that x, y, p are
local coordinates. Then since L1 is transverse to L2 , L1 ⊥ must be spanned
by dy − pdx plus something of the form −F dx + βdy + γdp, and we can
assume β ≡ 0 by adding on a multiple of dy − pdx. Furthermore, since
we know γ = 0, we may assume γ ≡ 1. Thus any path geometry is locally
equivalent to one arising from a second-order ODE.

Setting up the G-structure. Let G ⊂ GL(3, R) be the subgroup preserv-

ing the spans {e1 } and {e3 }:
⎧⎛ ⎞ ⎫
⎨ a e 0  ⎬
G= ⎝ ⎠
0 b 0  abc = 0 .
⎩ ⎭
0 f c 
Let M 3 be equipped with a path geometry, and define a G-structure on M
by
B = FG = {u ∈ F (M ) | u(L1 ) = {e1 }, u(L2 ) = {e3 }}.
Let ω = t (ω 1 , ω 2 , ω 3 ) be the tautological form on B. We have arranged
things so that {ω 2 , ω 3 } = π ∗ (L1 ⊥ ) and {ω 1 , ω 2 } = π ∗ (L2 ⊥ ). In particular,
the pullback of ω 2 along any section is a contact form.
We now find differential invariants of such G-structures. Since the com-
putation is a little involved, we first give an overview:
(1) We find that H 0,2 (g) = 0, so we need to normalize the torsion.
(2) We reduce to a subbundle B1 with structure group G1 ⊂ G that
preserves our normalization of the torsion.
(1)
(3) We calculate that g1 (1) = 0, so we need to prolong to B1 , a
subbundle of the frame bundle over B1 that has additive structure
group g1 (1) .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 276 9. Geometric Structures and Connections

(1)
(4) On B1 , we normalize torsion and reduce to a subbundle with
smaller structure group several times, until we obtain a subbundle
(1)
B2 ⊂ B1 where the structure group G2 ⊂ g1 (1) is such that the
connection is unique.
(5) The torsion of the connection on B2 gives two functionally inde-
pendent relative differential invariants H1 , H2 . From these we con-
struct tensors on M that are diffeomorphism invariants of the path
geometry.
On a first reading of this section, the reader may now wish to skip ahead
to page 281, where we interpret the invariants geometrically. In particular,
we show that if H1 = H2 = 0, the path geometry is locally diffeomorphic to
the flat case, Example 9.8.12. Note: the flat path geometry does not induce
a flat G-structure, since it turns out that B1 always has torsion.

Normalizing torsion. Let θ = (θji ) be a connection form on B. We may

write
⎛ 1⎞ ⎛ 1 ⎞ ⎛ 1⎞ ⎛ 1 i ⎞
ω θ1 θ12 0 ω Tij ω ∧ ω j
⎜ ⎟
d ⎝ω 2 ⎠ = − ⎝ 0 θ22 0 ⎠ ∧ ⎝ω 2 ⎠ + ⎝Tij2 ω i ∧ ω j ⎠ ,
ω3 0 θ23 θ33 ω3 Tij3 ω i ∧ ω j
i ’s are functions on B.
where the Tjk
Since every term Tij1 ω i ∧ ω j involves either ω 1 or ω 2 , we can absorb these
terms into the connection matrix by adding ω’s to θ11 and θ21 . The same
is true for the third row of torsion and the parts of the second row which
involve ω 2 . So, by choosing a new connection θ̃ we can arrange for all torsion
terms except for T13 2 ω 1 ∧ ω 3 to be zero. We conclude that:

Proposition 9.8.17. dim H 0,2 (g) = 1.

2 = 0, be-
The G-structures arising from path geometry will never have T13
2
cause ω pulls back to be a contact form.
We now use the group G to normalize T13 2 . If g ∈ G, then

⎛ −1 1 ⎞
a ω
Rg∗ ω = ⎝ b−1 ω 2 ⎠ ,
c−1 ω 3
2 = b T 2 . Let B ⊂ B be the subbundle where
and it follows that g · T13 ac 13 1
T13 ≡ 1. Its structure group is
2

⎧⎛ ⎞ ⎫
⎨ a e 0  ⎬
G1 = ⎝0 ac 0⎠ ac = 0 .
⎩ ⎭
0 f c 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 277

In our usual abuse of notation, we now let θ = (θji ) denote the pullback of
the connection form to B1 . Since θ must be g1 -valued when restricted to
the fiber of B1 , we have θ22 ≡ θ11 + θ33 mod{ω j }. New apparent torsion may
arise in trying to make θ g1 -valued. Suppose that
θ22 = θ11 + θ33 + a1 ω 1 + a2 ω 2 + a3 ω 3 .
Then if we let φ1 = θ11 + a1 ω 1 and φ2 = θ33 + a3 ω 3 , our new structure
equations are
⎛ 1⎞ ⎛ 1 ⎞ ⎛ 1⎞ ⎛ ⎞
ω φ μ1 0 ω 0
(9.11) d ⎝ω 2 ⎠ = − ⎝ 0 φ1 + φ2 0 ⎠ ∧ ⎝ω 2 ⎠ + ⎝ω 1 ∧ ω 3 ⎠ ,
ω3 0 μ2 φ2 ω3 0
where μ1 = θ21 and μ2 = θ23 . Thus the apparent torsion is absorbable.
Remark 9.8.18. Our change of notation for the connection forms has an
ulterior motive. We want to distinguish the torus t ⊂ g1 which is spanned
by vectors dual to φ1 , φ2 . Recall that for a semi-simple Lie algebra, its
action on a vector space is entirely determined up to isomorphism by the
action of a maximal torus. In particular, in the case of semi-simple g, to
test if a quantity is invariant it is sufficient to check if it is annihilated by
the maximal torus t ⊂ g (or equivalently, acted upon trivially by the torus
T ⊂ G). Recall from Appendix A that such quantities are said to have weight
zero. In the case of a G-structure with G an arbitrary matrix Lie group,
it is still easy to check if a form is invariant under the action of a maximal
torus, and thus gives a necessary condition for the form to be invariant. We
will still say quantities invariant under the action of a maximal torus have
weight zero.
Example (9.8.16, continued). For the ODE y  = F (x, y, y  ), we obtain a
section of the G1 -structure by taking
ω 1 = dx,
ω 2 = dy − pdx,
ω 3 = dp − F (x, y, p)dx.
In this case, a set of connection forms which satisfy (9.11) is
(9.12) φ1 = Fp dx, φ2 = −φ1 , μ1 = 0, μ2 = −Fy dx.
Proposition 9.8.19. The nonuniqueness in the choice of connection satis-
fying (9.11) is
 1  1    1
φ̃ φ 0 s1 ω
= 1 + ,
μ̃ 1 μ s1 t1 ω2
(9.13)  2  2    2
μ̃ μ t2 s2 ω
= 2 +
φ̃2 φ s2 0 ω3
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 278 9. Geometric Structures and Connections

for some functions s1 , s2 , t1 , t2 . Hence, dim g1 (1) = 4

Since the connection is not unique, we prolong.

(1)
Prolongation. Let B1 be the prolongation of B1 , as described in §9.3.
(1)
The forms (ω j , φi , μi ) pulled back to B1 become a basis of the tautological
(1)
forms on this bundle. We would like to fix a connection on B1 and de-
termine the torsion. In order to deduce the specific form of the torsion, we
(1)
differentiate the semi-basic forms on B1 . We already know the derivatives
of the ω j , so we need the derivatives of the φi and μi , which we calculate by
computing 0 = d2 ω j :
⎛ 1⎞ ⎛ 1 ⎞ ⎛ 1⎞
ω dφ dμ1 0 ω
2 ⎝ 2⎠ ⎝ ⎠
0=d ω =− 1
0 dφ + dφ 2 0 ∧ ω2⎠
⎝
ω 3 0 dμ 2 dφ 2 ω3
⎛ 1 ⎞ ⎧ ⎛ ⎞ ⎛ 1⎞ ⎛ ⎞⎫
φ μ1 0 ⎨ φ1 μ1 0 ω 0 ⎬
+ ⎝ 0 φ1 + φ2 0 ⎠ ∧ − ⎝ 0 φ1 + φ2 0 ⎠ ∧ ⎝ω 2 ⎠ + ⎝ω 1 ∧ ω 3 ⎠
⎩ ⎭
0 μ2 φ2 0 μ2 φ2 ω3 0
⎛ ⎞
0
+ ⎝−(φ1 ∧ ω 1 + μ1 ∧ ω 2 ) ∧ ω 3 + ω 1 ∧ (μ2 ∧ ω 2 + φ2 ∧ ω 3 )⎠ .
0
Separating the rows, we have
(9.14) 0 = −d2 ω 1 = (dφ1 + μ1 ∧ ω 3 ) ∧ ω 1 + (dμ1 + μ1 ∧ φ2 ) ∧ ω 2 ,
(9.15) 0 = −d2 ω 2 = −(dφ1 + dφ2 ) ∧ ω 2 − μ1 ∧ ω 2 ∧ ω 3 − μ2 ∧ ω 1 ∧ ω 2 ,
(9.16) 0 = −d2 ω 3 = (dμ2 + μ2 ∧ φ1 ) ∧ ω 2 + (dφ2 − μ1 ∧ ω 1 ) ∧ ω 3 .

The Cartan Lemma A.1.9 applied to (9.14) and (9.16) gives

 1     1
dφ + μ1 ∧ ω 3 π1 π2 ω
= ∧ ,
dμ1 + μ1 ∧ φ2 π2 π3 ω2
(9.17)  2     2
dμ + μ2 ∧ φ1 π4 π5 ω
= ∧
dφ2 + μ2 ∧ ω 1 π5 π6 ω3
for some 1-forms πi . Substituting these equations in (9.15) and reducing
mod ω 2 gives
   
π1 −2μ2
≡ mod ω 1 , ω 2 , ω 3 .
π6 2μ1
(1)
So, π1 and π6 are semi-basic for the submersion B1 → B1 . Since, for any
c, π̃1 = π1 + cω 1 also satisfies (9.17), we may assume π1 = −2μ2 + eω 2 + f ω 3
for some functions e, f . Moreover, we may assume e ≡ 0 by adjusting π2 .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 279

Similar considerations for π6 allow us to rewrite (9.17) as

(9.18)
 1    1  1 
dφ 0 π2 ω −μ ∧ ω 3 − 2μ2 ∧ ω 1 + E1 ω 1 ∧ ω 3
= ∧ + ,
dμ1 π2 π3 ω2 φ2 ∧ μ1
 2    2  
dμ π4 π5 ω φ1 ∧ μ2
= ∧ + .
dφ2 π5 0 ω3 μ2 ∧ ω 1 − 2μ1 ∧ ω 3 + E2 ω 1 ∧ ω 3
(1)
The remaining π’s are connection forms on B1 (uniquely defined by (9.18)),
and E1 , E2 are torsion components. However, adding the first and fourth
lines and substituting into (9.15) gives E1 + E2 = 0. So, we will let E =
E1 = −E2 .
(1)
Since the structure group g1 acts additively on the fibers, it acts either
additively or trivially on the scalar E. One way of checking how torsion
components vary along the fiber is to differentiate the structure equations
and mod out by semi-basic forms. In this case, computing 0 = d2 φ1 gives
(9.19) dE ≡ E(φ2 − φ1 ) + 2(π2 − π5 ) mod ω 1 , ω 2 , ω 3 ,
showing that E varies additively when we move along the fibers in the
direction dual to π2 − π5 .
Of course, we can also compute how E varies explicitly. Given a con-
(1)
nection on B1 (i.e., a section of B1 ), change to a new connection given by
(9.13). Then
dφ̃1 = π̃2 ∧ ω 2 − μ1 ∧ ω 3 − 2μ2 ∧ ω 1 + (E + 2s1 − 2s2 )ω 1 ∧ ω 3 ,

where

π̃2 = π2 + ds1 + t1 ω 3 + 2t2 ω 1 − s1 (φ̃1 + φ̃2 ),

shows that 2s1 − 2s2 is added to E. So, we may always pick s1 , s2 so that
(1)
E becomes zero. The connections satisfying E = 0 form a subbundle B1
with structure group reduced to the subgroup where s1 = s2 .
Example (9.8.16, continued). One can compute that E = −Fpp for the
connection forms given by (9.12). We change E to zero by just using the
s1 -adjustment, passing to new connection forms
φ1 = Fp dx + 12 Fpp (dy − p dx), μ1 = 12 Fpp dx,
with φ2 and μ2 as before.
Exercise 9.8.20: For Example 9.8.14 (Riemannian surfaces), show that
ω 1 = ω 1 , ω 2 = ω 2 , ω 3 = ω12 , φ1 = 0, φ2 = 0, μ1 = −ω12 , μ2 = Kω 1
(1)
(where K is the Gauss curvature) gives a section of B1 for which E = 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 280 9. Geometric Structures and Connections

We now restrict to the subbundle on which E ≡ 0, where the fiber group

now has dimension three. (We don’t bother to name it, as we are going to
restrict to a further subbundle momentarily.) By (9.19), π2 − π5 ≡ 0 modulo
ω 1 , ω 2 , ω 3 . In fact, we may adjust the π’s, while preserving (9.18), so that
π2 − π5 is a combination of just ω 1 and ω 3 . Let
π2 = σ + (x1 ω 1 + x3 ω 3 ),
π5 = σ − (x1 ω 1 + x3 ω 3 ).
The components of π2 −π5 are new torsion, which we may normalize to zero.
For, calculating 0 = d2 φ1 − d2 φ2 gives
)
dx1 = − 32 π4 + x1 (φ1 + 2φ2 )
(9.20) mod ω 1 , ω 2 , ω 3 .
dx3 = − 2 π3 + x3 (2φ1 + φ2 )
3

Exercise 9.8.21: Show that, after making a change of connection (9.13)

with s1 = s2 = 0 and t1 = 23 x3 , t2 = 23 x1 , we may assume that π2 = π5 .
(1)
We now restrict our attention to the subbundle B2 ⊂ B1 where x1 =
x3 = 0, and we have π2 = π5 = σ. By (9.20), one sees that now π3 , π4 ≡
0 mod ω 1 , ω 2 , ω 3 .
Exercise 9.8.22: Show that by adjusting σ by a multiple of ω 2 we can
arrange for π3 ∧ ω 2 = H2 ω 3 ∧ ω 2 for some function H2 .

This normalization forces π4 ∧ ω 2 = H1 ω 1 ∧ ω 2 for some function H1 .

Note that at this point we have made so many normalizations that we have
a uniquely determined coframe on B2 . (A reduction of a G-structure to a
unique section, like we have here, is called an e-structure.) Before discussing
this coframe, we return to our example:
Example (9.8.16, continued). Substituting our connection forms into (9.18)
shows that we may take
π2 = (Fpy + 12 Dx Fpp )ω 1 + 12 Fppp ω 3 , π5 = Fpy ω 1 ,
where Dx (g(x, y, p)) = gx + p gy + g gp . These give us the values for x1 and
x3 . We make the change of connection described in Exercise 9.8.21, resulting
in

μ1 = 12 Fpp ω 1 + 16 Fppp ω 2 , μ2 = −Fy ω 1 + 16 Dx Fpp − 23 Fpy ω 3 ,
with φ1 and φ2 the same as before. Now we recalculate, obtaining

σ = 13 Fpy + 16 Dx Fpp ω 1 + aω 2 + 13 Fppp ω 3
for arbitrary a. We choose a so that π3 ≡ 0 modulo ω 2 , ω 3 , and obtain

H1 = Fyy + 1
6 Dx2 Fpp − 4Dx Fpy + Fp (4Fpy − Dx Fpp ) − 3Fy Fpp ,
(9.21)
H2 = 16 Fpppp .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 281

Our uniquely determined coframe on B2 is ω 1 , ω 2 , ω 3 , φ1 , φ2 , μ1 , μ2 , σ.

It satisfies the structure equations (9.11) and

dφ1 = −μ1 ∧ ω 3 − 2μ2 ∧ ω 1 + σ ∧ ω 2 ,

dφ2 = 2μ2 ∧ ω 3 + μ2 ∧ ω 1 + σ ∧ ω 2 ,
(9.22)
dμ1 = φ2 ∧ μ1 + σ ∧ ω 1 + H2 ω 3 ∧ ω 2 ,
dμ2 = φ1 ∧ μ2 + σ ∧ ω 3 + H1 ω 1 ∧ ω 2 .

By differentiating the above equations, we deduce that

(9.23) dσ = (φ1 + φ2 ) ∧ σ − μ1 ∧ μ2 + (I1 ω 1 + I2 ω 3 ) ∧ ω 2

for some functions I1 , I2 . Differentiating (9.22) and (9.23) also gives

dH1 ≡ −I1 ω 3 + H1 (φ2 + 3φ1 ) mod ω 1 , ω 2 ,

dH2 ≡ −I2 ω 1 + H2 (φ1 + 3φ2 ) mod ω 2 , ω 3 ,
(9.24)
dI1 ≡ I1 (3φ1 + 2φ2 ) − H1 μ1 + Jω 3 mod ω 1 , ω 2 ,
dI2 ≡ I2 (2φ1 + 3φ2 ) + H1 μ2 + Jω 1 mod ω 2 , ω 3

for some function J.

Interpreting the invariants. The equations (9.24) help us complete a

weight diagram (see [65]) which shows how our various forms and functions
scale along the fiber of B2 over M :
rH2 rI2 rJ

rI1

rμ 2 rσ rH1

rω 1 uφ1,φ2 rμ 1

rω 2 rω 3

In this diagram, weights are assigned according to how objects scale with
respect to directions in the maximal torus dual to φ1 and φ2 respectively;
the large dot represents the origin.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 282 9. Geometric Structures and Connections

The reader familiar with such diagrams will recognize that if H1 , H2 = 0

we have the weight diagram for the adjoint representation of sl3 , and will
have already guessed the following proposition:
Proposition 9.8.23. The relative invariants H1 and H2 are identically zero
if and only if the path geometry is locally diffeomorphic to the flat model of
straight lines in the plane.

Proof. For the first part, notice that for the ODE y  = 0, we obtain H1 =
H2 = 0 from the formulas (9.21). For the second part, let
⎛ 1 ⎞
− 3 (2φ1 + φ2 ) −μ2 σ
ψ=⎝ ω1 3 (φ − φ )
1 1 2 μ1 ⎠.
2 3 1 1 2
ω ω 3 (φ + 2φ )

Note that ψ takes values in the Lie algebra sl(3, R). In general,
⎛ ⎞
0 −H1 ω 1 ∧ ω 2 (I1 ω 1 + I2 ω 3 ) ∧ ω 2
dψ + ψ ∧ ψ = ⎝0 0 H2 ω 3 ∧ ω 2 ⎠.
0 0 0

If H1 and H2 are identically zero, then I1 and I2 are also zero by (9.24).
Since dψ + ψ ∧ ψ = 0, by Theorem 1.6.13 there exists, in a neighborhood
U of any point on B2 , a local diffeomorphism g : U → SL(3, R) such that
g −1 dg = ψ. Moreover, the fibers of submersion to M , which are annihilated
by the ω’s, are carried to right cosets of the subgroup P of upper triangular
matrices in SL(3, R). This means that M is locally diffeomorphic to the flag
variety F12 from Example 9.8.12. 

Since neither H1 nor H2 has weight zero, they do not descend to be

invariant functions on M . We can however pair them with quantities of
opposite weights to obtain forms of weight zero. Consider the quartic forms
h1 := H1 (ω 1 )3 ω 3 and h2 := H2 ω 1 (ω 3 )3 , which are of weight zero. These
still don’t descend to sections of S 4 T ∗ M , because of the off-diagonal part
of the action of G1 (i.e., adding ω 2 to ω 1 or ω 3 ). However, they are well-
defined sections of S 4 W ∗ , where W = ω 2⊥ ⊂ T M . Recall that M = P(T X)
and that ω 2 corresponds to the contact form, so this bundle has a natural
geometric interpretation.
Consider Example 9.8.14 of paths arising from a Riemannian metric.
Using Proposition 9.8.23, we can prove
Theorem 9.8.24 (Bonnet). The only metrics inducing the flat path geom-
etry on a surface X are metrics of constant Gauss curvature.

Proof. It is easy to check that the coframe in Exercise 9.8.20 gives a section
of B2 . Differentiating the structure equations of Fon (X) gives dK ≡ 0
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 9.8. Extended example: path geometry 283

modulo ω 1 , ω 2 . Let
(9.25) dK = K1 ω 1 + K2 ω 2 .
Using σ = −Kω 2 , we compute that H1 = K2 and H2 = 0. Furthermore,
differentiating (9.25) shows that dK2 ≡ −K1 ω12 modulo ω 1 , ω 2 . Hence, H1
is identically zero if and only if K is constant. 
Remark 9.8.25. In fact, the same result is true for path geometries of
arbitrary dimensions: the only Riemannian metrics inducing the flat path
geometries are those of constant sectional curvature.
Proposition 9.8.26. A path geometry is locally equivalent to a projective
structure if and only if H2 ≡ 0.

Proof. Since every path geometry is locally equivalent to one arising from
a second-order ODE, we can use the formulas (9.21) in that case to give
interpretations to each of H1 and H2 . For example, H2 = 0 implies that our
system of paths is locally equivalent to the integral curves of an ODE of the
form
y  = A(y  )3 + B(y  )2 + C(y  ) + D,
where A, B, C, D are functions of x and y. Now apply Proposition 9.5.9. 
Remark 9.8.27. If one really wanted to find the explicit integral curves of
an ODE giving rise to a flat set of paths, it would be rather difficult to solve,
because although one knows abstractly that one has the flat case, actually
finding the transformation of integral curves can be quite difficult.
If H1 ≡ 0 and H2 = 0, then up to a few exceptions the equation
y  = F (x, y, p) is solvable by quadrature. (In other words, the differential
equation has a solvable symmetry group.) The idea is that differentiating
H2 and I2 yields scalar invariants that are constant along the paths, yielding
first integrals for the ODE. We explain this in more detail, following [31]:
Assume that H1 ≡ 0. Then I1 ≡ 0 and J ≡ 0 from (9.24), and we have
   1  
I2 2φ + 3φ2 μ2 I2
(9.26) d ≡ mod ω 2 , ω 3 .
H2 −ω 1 φ1 + 3φ2 H2
This implies that P = (I2 ω 2 + H2 ω 3 ) ⊗ (ω 2 ∧ ω 3 ) is a well-defined tensor on
the path space Z = M/L1 . (Note that the weight diagram indicates that P
is invariant under the maximal torus.) To see this, suppose we set P2 = I2 ,
P3 = H2 . Then we may write (9.26) as
(9.27) dPa = (ψab + (tr Ψ)δab ) Pb + Pab ω b ,
for some functions Pab , where we use index ranges 2 ≤ a, b ≤ 3 and we have
set  1 
φ + φ2 μ2
Ψ = (ψab ) = .
−ω 1 φ2
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 284 9. Geometric Structures and Connections

Differentiating (9.27) allows us to extract similar tensorial invariants from

the derivatives Pab . Namely, let
Q2 = P2 (P23 − 4P32 ) + 3P3 P22 , Q3 = P3 (4P23 − P32 ) − 3P2 P33 .
Then
dQa = (ψab + 3(tr Ψ)δab ) Qb + Qab ω b
shows that Q = (Q2 ω 2 + Q3 ω 3 ) ⊗ (ω 2 ∧ ω 3 )3 gives a well-defined section of
T ∗ Z ⊗ L3 , where L = Λ2 (T ∗ Z). The wedge product P ∧ Q is therefore an
invariant taking value in the line bundle L5 .
Continuing to differentiate leads to more line-bundle invariants. For
example, if we set
R2 = Q2 (Q23 − 10Q32 ) + 9Q3 Q22 , R3 = Q3 (10Q23 − Q32 ) − 9Q2 Q33 ,
then R = (R2 + R3 ⊗ ω2 ω3) (ω 2 ∧ ω 3 )7 is well-defined on Z. If P ∧ Q is
nonzero, then the quotient
(P ∧ R)5
(P ∧ Q)9
gives a well-defined function on path space, and therefore is a first integral
for the ODE.
Exercise 9.8.28: Verify that H1 ≡ 0 for the path geometry associated to
the ODE
(y  )2 (1 + (y  )2 )
(9.28) y  =  .
y(1 + 1 + (y  )2 )
Then calculate the invariants P and Q for this geometry.
Remark 9.8.29. If one is working in the holomorphic category, one can
obtain much stronger global results. Recall that infinitesimal deformations
of submanifolds are roughly given by sections of their normal bundles. Let
Z be a complex compact surface and let C ⊂ Z be a holomorphic curve
isomorphic to CP1 . We will say C has self-intersection one if any small
deformation of C will intersect C in exactly one point. Thus its deformations
are to first order the same as those of a line in the projective plane.
Theorem 9.8.30 (Hitchin [96]). Let Z be a complex surface equipped with
a path geometry whose paths are rational curves with self-intersection one.
Then H2 ≡ 0, i.e., the path geometry comes from a projective structure.
Moreover, any path geometry in a compact complex surface Z with H2 ≡ 0
has rational curves with self-intersection one as its paths.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/10

Chapter 10

Superposition for
Darboux-Integrable
Systems

In §7.3 we introduced the notion of Darboux integrability for hyperbolic

EDS. Our earlier discussion focused on the classical results of Darboux,
Goursat and others. One might criticize the treatment of Darboux integra-
bility in the classical literature because of the relative paucity of examples,
and because there seems to be a gap between the fact of integrability and
the construction of explicit solution formulas. (For example, the solution
formula (7.16) for Liouville’s equation might seem to arise through a series
of clever tricks.) Fortunately, significant progress has recently been made in
this area through the efforts of Anderson, Fels and Vassiliou [5], [6]. Not
only did they formulate an integrability test for systems involving larger
numbers of dependent and independent variables, but they also gave a way
to systematically construct the general solutions for those systems that turn
out to be integrable. Their fundamental insight, inspired by the work of Ves-
siot [184], was that to every Darboux-integrable system I one can associate
a Lie algebra, called the Vessiot algebra, such that I arises as a quotient
of a product of simpler systems by the action of the corresponding group.
The quotient map is known as the superposition formula for I, and can
be used to construct general solutions of I from pairs of solutions of certain
simpler systems W1 and W2 defined on lower-dimensional manifolds.
In this chapter we will discuss the generalization of Darboux integrabil-
ity defined in [6], and the construction of superposition formulas. This con-
struction is carried out through a sequence of coframe adaptations. These

285
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 286 10. Superposition for Darboux-Integrable Systems

will be illustrated by a single extended example, the Darboux-integrable

PDE u uxy /uy = x + ux . (When we interrupt the example to return to the
general discussion, this will be indicated by the symbol .) This equation
is a special case of u uxy /uy = ux + f (x, uxy /uy ), which appears in Gour-
sat’s classification [77] of second-order hyperbolic Monge-Ampère equations
in the plane that are Darboux-integrable at the 2-jet level (see [51] for a
brief summary of this classification). Unlike the frequently-used example
of Liouville’s equation, our example PDE is Monge-integrable, as one of its
characteristic systems contains a rank three Frobenius system. This fea-
ture lets us illustrate the coframe adaptations in a more general (and less
symmetric) situation than that of Liouville’s equation.
In §10.1 we discuss the decomposability condition for a differential ideal
I, which generalizes the concept of the hyperbolic system and is a pre-
requisite for integrability; we also give some examples of decomposable sys-
tems, and define the singular systems which generalize the Monge character-
istic systems defined in §7.3. In §10.2 we discuss the generalized integrability
condition and the implications it has for the algebraic structure of the sin-
gular systems; we also begin our extended example, with the construction of
a 1-adapted coframe. In §10.3 we continue with these coframe adaptations
(see Propositions 10.3.1, 10.3.3, 10.3.4 and 10.3.5), eventually arriving at
the 5-adapted coframe (actually, a pair of overlapping coframes) the dual
frames of which include generators for the action of the Vessiot group. In
§10.4 we review some existence results needed to construct the action from
the generators. In §10.5 we use the group action to construct the super-
position formula (10.17), and we show that the formula makes the product
system W1 + W2 an integrable extension of I. The superposition formula is
also applied to give the general solution of the PDE in terms of two arbitrary
functions and their derivatives. We conclude the chapter with an exercise
set containing additional examples for the reader to explore.

10.1. Decomposability
Throughout this chapter I will be an EDS generated algebraically by 1-
forms and 2-forms on manifold Σ; we let I ⊂ T ∗ Σ be the bundle spanned
by the 1-forms.
Recall from §7.2 that I is hyperbolic of class k if it is generated by k
1-forms and a pair of decomposable 2-forms. In other words, at each point
of Σ there is a coframe (θ1 , . . . , θk , π11 , π12 , π21 , π22 ) such that
I = {θ1 , . . . , θk , π11 ∧ π12 , π21 ∧ π22 }alg .
In §7.2 we saw that hyperbolic systems with k = 3 occur when one makes
an EDS encoding a second-order hyperbolic PDE for one function of two
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.1. Decomposability 287

variables in the standard way (and systems with k = 1 can encode second-
order hyperbolic Monge-Ampère equations), while systems with k = 2 arise
when one encodes a quasilinear hyperbolic first-order system for a pair of
functions of two variables.
Recall that the Monge characteristic systems V1 , V2 of a hyperbolic sys-
tem are defined by
Vi = {θ1 , . . . , θk , πi1 , πi2 }.
These systems play a key role in the method of Darboux, since I is Darboux-
integrable if each Vi possesses a pair of characteristic invariants that are
independent of each other and of the 1-forms of I. (A characteristic invariant
is a first integral of Vi , i.e., a function f whose differential df is a section of
Vi ; it follows that it also must be a section of the terminal derived system
(∞)
Vi .) In seeking to extend integrability to a wider class of systems, we will
focus on the role played by V1 , V2 in the algebraic structure of the ideal.
Because I is closed under exterior differentiation,
dθj ≡ Aj π11 ∧ π12 + B j π21 ∧ π22 mod θ1 , . . . , θk
for some functions Aj , B j . A natural extension of this structure, which turns
out to encompass a variety of other kinds of systems, is based on assuming
a direct sum decomposition
T ∗ Σ = I ⊕ J1 ⊕ J2
such that I is generated algebraically by Γ(I) and 2-forms that either lie in
Γ(Λ2 J1 ) or Γ(Λ2 J2 ). This condition is unchanged if J1 or J2 is modified by
adding 1-forms belonging to I, so it makes sense to formulate it in terms of
I + J1 and I + J2 .
Definition 10.1.1. Let I be an EDS generated algebraically by 1-forms and
2-forms on Σ. Then I is decomposable if there exist subbundles V̂ , V̌ ⊂ T ∗ Σ
such that
• T ∗ Σ = V̂ + V̌ and I = V̂ ∩ V̌ ;
• I is algebraically generated by 1-forms in I and 2-forms that either
are sections of Λ2 V̂ or in Λ2 V̌ .

Such a system is said to be decomposable of type [p, q] if I has codi-

mension p within V̂ and q within V̌ . In particular, hyperbolic systems are
decomposable of type [2, 2].

Some examples.
Example 10.1.2. Consider a system of f -Gordon equations for s functions
of variables x1 , x2 :
(10.1) ua12 = F a (x, u, u1 , u2 ), a = 1, . . . , s.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 288 10. Superposition for Darboux-Integrable Systems

(On the right we group the independent and dependent variables into vectors
x and u respectively, and ui denotes the partial derivative of u with respect
to xi .) This system is encoded by an EDS generated by 1-forms θa =
dua − ua1 dx1 − ua2 dx2 and 2-forms π1a ∧ dx1 , π2a ∧ dx2 with

π1a = dua1 − F a dx2 , π2a = dua2 − F a dx1 .

This system is decomposable of type [s + 1, s + 1], with

V̂ = {θ, π 1 , dx1 }, V̌ = {θ, π 2 , dx2 }.

Here, π i indicates the s-vector of 1-forms π1a , and θ indicates the vector of
forms θa .

Remark 10.1.3. An interesting special case are the 2-dimensional Toda

lattice systems

(10.2) ua12 = Cba eub ,

where Cba is the Cartan matrix (see, e.g., [19]) of a simple Lie algebra. (For
example, the A2 Toda lattice system appears in Exercise 10.5.11.6 below.)
These systems are Darboux-integrable after sufficiently many prolongations;
see [154] for a systematic way of writing down the characteristic invariants.
If the matrix in (10.2) is replaced by certain other choices (called gener-
alized Cartan matrices), the resulting system is completely integrable, and
solutions can be used to construct harmonic maps into symmetric spaces
[87].

Remark 10.1.4. Systems of the form (10.1) are related to the defining
equations for wave maps from the Minkowski plane R1,1 into a Riemannian
manifold M , where x, y are null coordinates in the domain and the ua are
local coordinates in M . (Wave maps are exactly analogous to harmonic
maps, but with the domain being a semi-Riemannian manifold.) Two-
dimensional Riemannian manifolds M for which the wave map system is
Darboux-integrable after at most one prolongation have been classified [52];
one example appears in Exercise 10.5.11.5 below.

Example 10.1.5. Next, consider an overdetermined system of two quasi-

linear hyperbolic equations for u(x, y, z), of the form

(10.3) uxz = A(u, x, y, z, ux , uy , uz ), uyz = B(u, x, y, z, ux , uy , uz ).

In order that these equations be compatible, it is necessary and sufficient

that Dy A = Dx B, where Dx , Dy indicate total derivatives with respect to x
and y, computed using the system (10.3) to substitute for the mixed partials.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.2. Integrability 289

This system is encoded by an ideal I generated algebraically by 1-forms

θ0 = du − u1 dx − u2 dy − u3 dz,
θ1 = du1 − u11 dx1 − u12 dx2 − Adx3 ,
θ2 = du2 − u12 dx1 − u22 dx2 − Bdx3 ,
θ3 = du3 − Adx1 − Bdx2 − u33 dx3
and by 2-forms π11 ∧ dx1 + π12 ∧ dx2 , π12 ∧ dx1 + π13 ∧ dx2 and π21 ∧ dx3 , where
π11 = du11 − (Dx A)dx3 ,
π12 = du12 − (Dy A)dx3 = du12 − (Dx B)dx3 ,
π13 = du22 − (Dy B)dx3 ,
π21 = du33 − (Dz A)dx1 − (Dz B)dx2 .
Thus, this system is decomposable of type [5, 2], with V̂ = {θ, π11 , π12 , π13 , dx1 ,
dx2 } and V̌ = {θ, π21 , dx3 }.
Exercise 10.1.6: Verify that I is involutive with characters s1 = 3, s2 = 1.

10.2. Integrability
A key geometric feature of classical Darboux integrability is the double
foliation of an integral surface S by characteristic curves: since the first
integrals of each characteristic system are constant along the corresponding
curves, each set of invariants restricts to S to be functionally related, and
these relations can be used to recover S. For more general decomposable
systems, the tangent bundle of an integral submanifold S still splits as a sum
of characteristic distributions (annihilated by the pullbacks to S of V̌ and V̂ ,
respectively), but the integrability of these distributions is not automatic:
what’s necessary is that the pullbacks to S of V̂ and V̌ be Frobenius. A
condition that guarantees this, and is independent of S, is that V̂ (∞) and
V̌ (∞) are large enough that T ∗ Σ = I + V̂ (∞) + V̌ (∞) as a nondirect sum.
This is the gist of the next definition.
Definition 10.2.1. A decomposable system is Darboux-integrable if
(i) V̂ (∞) + V̌ = V̂ + V̌ (∞) = T ∗ Σ;
(ii) V̂ (∞) ∩ V̌ (∞) = 0.

Again, note that in the first condition the sum need not be direct. The
second condition is technical, ensuring that I contains no closed 1-forms.
This is not a serious restriction, for if I did contain some closed 1-forms,
these would locally be exact 1-forms dfi , and any integral submanifold would
necessarily lie in a joint level set of the fi ; but then the pullback of I to the
level set would contain no exact 1-forms.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 290 10. Superposition for Darboux-Integrable Systems

Remark 10.2.2. In fact, if V̂ and V̌ satisfy condition (i) in the definition,

and I contains no closed 1-forms (i.e., I (∞) = 0), then condition (ii) is
automatically satisfied (see Theorem 4.6 in [5]).
Remark 10.2.3. It is easy to show that the prolongation of a Darboux-
integrable hyperbolic EDS is Darboux-integrable. More generally, any in-
tegrable extension E of a decomposable system that is Darboux-integrable
(in the sense of Definition 10.2.1) is also Darboux-integrable, provided E
contains no closed 1-forms (see Theorem 5.1 in [5]).
Example 10.2.4. Throughout this chapter, we will illustrate the steps in
constructing the superposition formula by carrying them out for the Pfaffian
system associated to the second-order PDE
uxy
(10.4) u = x + ux .
uy
In the ‘classical’ coordinates on J 2 (R2 , R) introduced in Example 6.5.3, this
equation corresponds the subset defined by us = q(x + p). We restrict to the
smooth part of this subset where u = 0, and the Pfaffian system is obtained
q
by substituting s = (x + p) into the standard contact system generated by
u
θ0 = du − p dx − q dy, θ1 = dp − r dx − s dy, θ2 = dq − s dx − t dy.
However, it turns out to be easier to express the first integrals of the char-
acteristic systems if we use a coordinate w = (x + p)/u, and set s = qw in
the contact system, yielding the 1-forms
θ0 = du − (uw − x) dx − q dy, θ1 = u dw − (r + wp + 1) dx + wθ0 ,
θ2 = dq − qw dx − t dy
defined on R7 with coordinates x, y, u, w, q, r, t where u = 0. By computing
the derivatives of θ0 , θ1 , θ2 modulo themselves we obtain
q
V̂ = {θ0 , θ1 , θ2 , dx, dr − (r + xw + 1) dy}
u
and
V̌ = {θ0 , θ1 , θ2 , dy, dt − wt dx}.
By computing the derived flags (see §7.3) of each of these we obtain
   
r + xw + 1 t
V̂ (∞)
= {dx, dw, d }, V̌ (∞)
= {dy, d }.
u q
Thus, the criteria for Darboux integrability are met. Matters being so, when
we return to this example we will compute in better coordinates, eliminating
r and t in favor of
r + xw + 1 t
R= , T = . 
Licensed to AMS. u q
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.2. Integrability 291

The construction of the superposition formula for a Darboux-integrable

decomposable system I is achieved through a sequence of coframe adapta-
tions. (In what follows, we will attempt to use the terminology and notation
of [6] as much as possible, to make it easier for the reader to consult that
article for proofs that we omit.)
Definition 10.2.5. A local coframe (θ, η̂, σ̂, η̌, σ̌) is 0-adapted if
• V̂ (∞) ∩ I = {η̂} and V̌ (∞) ∩ I = {η̌};
• V̂ (∞) = {η̂, σ̂} and V̌ (∞) = {η̌, σ̌};
• I = {η̂, η̌, θ}.
(As above, boldface Greek letters stand for lists of 1-forms, e.g., η̂ = η̂ 1 , η̂ 2 ,
. . . , η̂ r where r = rk(V̂ (∞) ∩ I).)

A 0-adapted coframe can be constructed by choosing 1-forms in the order

listed in the definition, for example choosing the 1-forms in σ̂ to complete
η̂ as a basis for V̂ (∞) . (In some cases the vectors η̂ or η̌ may have length
zero.) The resulting collection of 1-forms are pointwise linearly independent
because of condition (ii) in Definition 10.2.1.
Exercise 10.2.6: Prove that a collection of pointwise linearly independent
1-forms defined on an open set U ⊂ Σ satisfying the conditions of Definition
10.2.5 span the cotangent space of Σ at every point in U . 

The following diagram illustrates how a 0-adapted coframing is adapted

to the overlapping subbundles of T ∗ Σ that make up the decomposable struc-
ture of I.
V̂
V̌
V̂ (∞) V̌ (∞)

σ̂ η̂ θ η̌ σ̌

Since the σ̂ a must span a complement to {η̂} within V̂ (∞) , they can be
chosen to be closed 1-forms, and similarly we can choose the σ̌ α to be closed.
But we can also simplify the structure equations satisfied by η̂, η̌:
Proposition 10.2.7 (Theorem 2.9 in [6]). Near any point, there exists a
0-adapted coframe satisfying
dσ̂ = 0, dη̂ = (Ê η̂ + F̂ σ̂) ∧ σ̂, dσ̌ = 0, dη̌ = (Ě η̌ + F̌ σ̌) ∧ σ̌.

Moreover, the coefficients Ê and F̂ are first integrals of V̂ (∞) , while

Ě and F̌ are first integrals of V̌ (∞) . Such a coframe is called 1-adapted.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 292 10. Superposition for Darboux-Integrable Systems

(Following [6], here we begin to use boldface capitals to indicate the ap-
propriately indexed arrays of coefficients. For example, the last structure
equation reads in full
μ α
dη̌ μ = Ěνα
μ ν
η̌ ∧ σ̌ α + 12 F̌αβ σ̌ ∧ σ̌ β ,

with summation over repeated indices, and F̌ being skew-symmetric in its

lower indices.)

Proof. By the Frobenius Theorem, the system V̂ (∞) locally has a basis of
exact differentials. We can group these differentials into two vectors dIˆ1 and
dIˆ2 (of length p and r respectively) such that the 1-forms
η̂ = dIˆ2 + R̂ dIˆ1
form a basis of V̂ (∞) ∩ I, for some r × p matrix R̂. We will take σ̂ = dIˆ1 ,
so that dσ̂ = 0. Because dη̂ involves no σ̌ ∧ σ̂ terms, dR̂ ∈ V̂ , and hence
dR̂ ∈ V̂ (∞) . So we have dR̂ = Ê η̂ + F̂ σ̂ and the desired form for dη̂ follows.
We obtain σ̌ and η̂ similarly.
Exercise 10.2.8: Show that the components of Ê and F̂ are expressible in
terms of the entries of R̂ and their partial derivatives with respect to the
functions in Iˆ1 and Iˆ2 . (This establishes the last assertion of the proposi-
tion.)

The last assertion of the proposition (which follows from the exercise)
is not part of the definition of 1-adapted given in [6], but it is an easy
consequence of the proof, and is necessary for subsequent adaptations.

Example (10.2.4 continued). To construct a 0-adapted coframe, we first
choose a basis for the intersection of V̂ (∞) with I = {θ0 , θ1 , θ2 }. This inter-
section has rank one, and is spanned by η̂ = (θ1 −wθ0 )/u = dw+(w2 −R)dx.
We complete a basis for V̂ (∞) with exact 1-forms σ̂ 1 = dx and σ̂ 2 = dw.
Similarly, we span V̌ (∞) using σ̌ 1 = dy and σ̌ 2 = dT . The intersection of
V̌ (∞) with I is zero, so there are no 1-forms η̌ in this example.
We complete the 0-adapted coframe with
θ1 = −θ0 = −du + p dx + q dy, θ2 = −θ2 = −dq + q(w dx + T dy).
Finally, we note that
dη̂ = σ̂ 1 ∧ σ̂ 2 + 2wη̂ ∧ σ̂ 1 ,
indicating that this coframe is also 1-adapted. 

The condition that I is decomposable implies that there are generator

2-forms Ω̂ and Ω̌, which are sections of Λ2 V̂ and Λ2 V̌ respectively, such
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.2. Integrability 293

that
I = {θ, η̌, η̂, Ω̂, Ω̌}alg .
Moreover, with respect to a 1-adapted coframe, we can assume that the Ω̂α
and Ω̌β are expressible in terms of pure wedge products of the σ̂ a and σ̌ b
respectively. By adjoining one or the other of these groups of 1-forms to the
generators of I, we obtain the singular systems
(10.5) V̂ = {θ, σ̂, η̂, η̌, Ω̌}alg , V̌ = {θ, σ̌, η̌, η̂, Ω̂}alg .
Exercise 10.2.9: Show that these are differential ideals. 

The next result will be needed in §10.5 to prove that the superposition
formula takes solutions to solutions. (In what follows, let 1 ≤ i, j, k ≤ s,
where s = dim Σ − rkV̂ (∞) − rkV̌ (∞) .)
Proposition 10.2.10. Let (θ, η̂, σ̂, η̌, σ̌) be a 1-adapted coframe, and let
(10.6) dθi ≡ 12 Aiab σ̂ a ∧ σ̂ b + 12 Bαβ
i
σ̌ α ∧ σ̌ β mod θ, η̂, η̌.
(Note that the decomposability condition implies that these are the only
kinds of wedge products that can occur in dθi mod I.) Then
(10.7) I = {θ, η̂, dη̂, η̌, dη̌, Aσ̂ ∧ σ̂, B σ̌ ∧ σ̌}alg ,
where Aσ̂ ∧ σ̂ and B σ̌ ∧ σ̌ denote vectors of 2-forms whose members are
the individual terms on the right-hand side of (10.6).

Proof. Let V̂  V̌ denote the set of differential forms on Σ that vanish on

the direct sum of any integral element of V̂ plus any integral element of V̌.
Exercise 10.2.11: Show that V̂  V̌ is an algebraic ideal, contained in the
differential ideal V̂ ∩ V̌. 

From this exercise and (10.5) it follows that

V̂  V̌ ⊂ V̂ ∩ V̌ = {θ, η̂, η̌, Ω̂, Ω̌, σ̂ ∧ σ̌}alg ,
where σ̂ ∧ σ̌ represents the wedge product of any of the σ̂ a with any of the
σ̌ α . However, it is easy to construct a sum Ê ⊕ Ě, of integral elements of
V̂ and V̌ respectively, on which σ̂ a ∧ σ̌ α does not vanish. Thus, V̂  V̌ can
contain no such wedge products, and
V̂  V̌ ⊂ {θ, η̂, η̌, Ω̂, Ω̌}alg = I.

Next, we claim that I ⊂ V̂  V̌. To see this, let Ê be an integral element

of V̂ and Ě an integral element of V̌. Then θ, η̂ and η̌ vanish on Ê⊕Ě. If v, v
are vectors in Ê and w, w are vectors in Ě, then (i) Ω̂(v, v ) = Ω̂(v, w) = 0
since v, v are annihilated by σ̂, (ii) Ω̂(w, w ) = 0 since the 2-forms Ω̂ are in
V̌, and similarly (iii) Ω̌(v, v ) = Ω̌(v, w) = Ω̌(w, w ) = 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 294 10. Superposition for Darboux-Integrable Systems

It follows that I = V̂  V̌. Finally, we claim that the 2-forms Aσ̂ ∧ σ̂

and B σ̌ ∧ σ̌ are in V̂  V̌. For, if v, v , w, w are as before, then we again
have σ̂ ∧ σ̂(v, v ) = σ̂ ∧ σ̂(v, v ) = 0. Moreover, since dθi ≡ 12 Aiab σ̂ a ∧ σ̂ b
modulo the 1-forms in V̌, and V̌ is a differential ideal, then the Aσ̂ ∧ σ̂ are
in V̌ and thus Aσ̂ ∧ σ̂(w, w ) = 0. The argument for B σ̂ ∧ σ̂ is similar. 

10.3. Coframe adaptations

We continue with successive coframe adaptations; the next adaptation mod-
ifies θ, η̂, η̌ while preserving previous adaptations but eliminating terms of
the form σ̂ ∧ η̌, σ̌ ∧ η̂ and η̂ ∧ η̌ from dθi .
Proposition 10.3.1 (Thm. 4.3 in [6]). Near any point, there exists a
1-adapted coframe such that
dθ ≡ Aπ̂ ∧ π̂ + B π̌ ∧ π̌ mod θ.
(Here we have grouped σ̂ and η̂ into a single vector π̂, and similarly grouped
σ̌, η̌ into π̌.) Such coframes are called 2-adapted.

Proof. Let (∂θ , ∂σ̂ , ∂η̂ , ∂σ̌ , ∂σ̂ ) be the dual frame to a 1-adapted coframe.
Using iterated Lie brackets define the vector fields ŜA and ŠB (where A and
B are multi-indices whose entries have ranges 1 . . . p and 1 . . . q respectively)
by
Ŝa = ∂σ̂a , Šα = ∂σ̌α , ŜAa = [ŜA , Ŝa ], ŠBα = [ŠB , Šα ].
Note from the structure equations that [∂σ̂ , ∂σ̌ ] = 0, and that ŜA ∈ {∂θ , ∂η̂ },
ŜB ∈ {∂θ , ∂η̌ } when |A| > 1 and |B| > 1. Furthermore, it follows by the
Jacobi identity that
(10.8) [ŜA , ŠB ] = 0.
Since the ∂σ̂ span Ĥ = (V̌ )⊥ , then a finite number of the ŜA will span
Ĥ(∞) = (V̌ (∞) )⊥ . Among these we can choose vectors ŜAm , where 1 ≤ m ≤
rk(V̂ ∞ ∩ I), so that (∂θ , ∂σ̂ , ŜAm ) are a basis for Ĥ(∞) . Similarly, there are
vectors ŠBμ , where 1 ≤ μ ≤ rk(V̌ ∞ ∩ I), such that (∂θ , ∂σ̌ , ŠBμ ) is a basis
for Ȟ(∞) .
Then (∂θ , ∂σ̂ , ŜAm , ∂σ̌ , ŠBμ ) is a local frame, and our new coframe is its dual
(θ , σ̂, η̂  , σ̌, η̌  ), where θi ≡ θi mod η̂, η̌ and
η̂  = Ĥ η̂, η̌  = Ȟ η̌
for some nonsingular matrices Ĥ and Ȟ. Thus, the new coframe is 0-
adapted.
Exercise 10.3.2: Show that the components of Ĥ and Ȟ are first integrals
of V̂ (∞) and V̌ (∞) respectively. Conclude that the new coframe is 1-adapted.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.3. Coframe adaptations 295

Moreover, from (10.8) it follows that the structure equations of the new
coframe contain no terms of the form σ̂ ∧ η̌  , σ̌ ∧ η̂  or η̂  ∧ η̌  . 
Example (10.2.4 continued). The remaining structure equations for our
1-adapted coframe are
dθ1 = uη̂ ∧ σ̂ 1 − wθ1 ∧ σ̂ 1 − θ2 ∧ σ̌ 1 ,
dθ2 = q η̂ ∧ σ̂ 1 − wθ2 ∧ σ̂ 1 − T θ2 ∧ σ̌ 1 − qσ̌ 1 ∧ σ̌ 2 .
Since these contain no η̂ ∧ σ̌ terms, this coframe is also 2-adapted. 

In what follows we will let D denote the rank s distribution on Σ anni-

hilated by the Frobenius system V̌ (∞) + V̂ (∞) = {π̂, π̂}. With a 2-adapted
coframe we have
dθ = Aπ̂ ∧ π̂ + B π̌ ∧ π̌ + Cθ ∧ θ + M π̂ ∧ θ + N π̌ ∧ θ.
The point of the next adaptation is that either M or N can be eliminated:
Proposition 10.3.3 (Thm. 4.4 in [6]). Near any point, there exists a pair
i and θ i
of 2-adapted coframes (θX , π̂, π̌) and (θ Y , π̂, π̌) such that the θX Y
have the same span, and satisfy structure equations of the form
i j
i
dθX = 12 Aiab π̂ a ∧ π̂ b + 12 Bmn
i
π̌ m ∧ π̌ n + 12 Cjk θX ∧ θX
k i
+ Mak π̂ a ∧ θX
k
,
i j
dθYi = 12 Dab
i
π̂ a ∧ π̂ b + 12 Emn
i
π̌ m ∧ π̌ n + 12 Kjk θY ∧ θYk + Nαk
i
π̌ α ∧ θk .
Moreover, if we let Xi and Yj denote the vector fields that are dual to θX i
j
and θY respectively (and annihilated by π̂ and π̌), then [Xi , Yj ] = 0. Such
coframes are called 3-adapted.

Proof. Given a 2-adapted coframe (θ, π̂, π̌), we replace this with a tempo-
rary coframe (θ, dJˆ, dJ),
ˇ where the exact 1-forms dJˆ and dJˇ form a basis
for the Frobenius systems V̂ (∞) and V̌ (∞) respectively. We let (∂θ , Ûa , Ǔα )
be the dual frame to the temporary coframe, where indices a, b range from
1 to rkV̂ (∞) while α, β range from 1 to rkV̌ (∞) . As in the previous proof
we form iterated Lie brackets ÛAa = [ÛA , Ûa ] and ǓBα = [ǓB , Ǔα ], where
A, B are multi-indices. Note that because no π̂ ∧ π̌ terms occur in the
structure equations of the 2-adapted frame, [Ûa , Ǔα ] = 0 and consequently
[ÛA , ǓB ] = 0.
Since
Ĥ = V̌ ⊥ = {θ, η̂, π̌}⊥ ⊂ {θ, dJ}
ˇ ⊥ = {Ûa },
then a finite number of the ÛA will span Ĥ(∞) . But because [Ûa , Ûb ] ∈
{∂θ } = D and [Ûa , ∂θ ] ∈ D, then the ÛA for |A| > 1 all lie in D =
Ĥ(∞) ∩ Ȟ(∞) . Therefore, there are a finite number of these ÛA which form
a basis for D; let ÛAk denote these, where k = 1, . . . , s and |Ak | > 1. Thus,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 296 10. Superposition for Darboux-Integrable Systems

i be the corresponding 1-forms

(ÛAk , Ûa , Ǔα ) is a local frame, and we let θX
of the dual coframe, i.e.,
i
θX (ÛAj ) = δji , i
θX i
(Ûa ) = θX (Ǔα ) = 0.
Since the θX i have the same span as the 1-forms θ i of the 2-adapted coframe

we started with, the coframe (θ X , π̂, π̌) is still 2-adapted. Moreover, because
[Ǔα , ÛAi ] = 0 the structure equations of the new coframe contain no π̌ ∧ θ
terms.
The 1-forms θYi are similarly derived from the ǓB for |B| > 1. 

A further adjustment can be made to reveal the Vessiot algebra:

Proposition 10.3.4 (Thm. 4.5 in [6]). Near any point, there exists a pair
of 3-adapted coframes such that the coefficients in Proposition 10.3.3 satisfy
i = −C i and the C i are structure constants for an s-dimensional Lie
Kjk jk jk
algebra g. Such coframes are called 4-adapted. Any two pairs of 4-adapted
coframes for a Darboux-integrable system lead to the same Lie algebra (up
to isomorphism), called the Vessiot algebra of I.

Example (10.2.4 continued). Following the procedure described in the proof

of Proposition 10.3.3 yields the 1-forms
dq
1
θX =u − du − x dx + (q − uT )dy,
q
du (uw − x)
θY1 = − + dx + dy,
q q
dq
2
θX = θY2 = − + w dx + T dy.
q
These satisfy
1
dθX = uσ̌ 1 ∧ σ̌ 2 − wθX
1
∧ σ̂ 1 + θX
1
∧ θX
2
− xθX 2
∧ σ̂ 1 ,
u
dθY1 = − σ̂ 1 ∧ η̂ + T θY1 ∧ σ̌ 1 − θY1 ∧ θY2 − θY2 ∧ σ̌ 1 ,
q
2
dθX = dθY = −σ̂ 1 ∧ η̂ − σ̌ 1 ∧ σ̌ 2 .
2

Because the θ ∧ θ terms have constant coefficients, this coframe is also 4-

adapted. Because C121 = 1 is the only nonzero structure constant, the Vessiot

algebra is a 2-dimensional non-Abelian Lie algebra (the unique such algebra,

up to isomorphism). 

It follows from the structure equations of the 4-adapted coframe that

the sets {Xi } and {Yj } of dual vector fields both span D at each point, and
satisfy
[Xj , Xk ] = −Cjk
i
Xi , i
[Yj , Yk ] = Cjk Yi .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.3. Coframe adaptations 297

Let G denote the simply connected Lie group whose algebra is the Vessiot
algebra g, and let {vi } be a fixed basis for Te G such that [vjL , vkL ] = −Cjk
i vL .
i
By the existence theory for group actions (see §10.4 below), the vector fields
Xi and Yj generate local right and left G-actions respectively on Σ. In other
words, there is an open neighborhood V of {e} × Σ in G × Σ, and mappings

λ : V → Σ, ρ : V → Σ,

which are left and right actions respectively, such that ρ∗ : vi → Xi and
λ∗ : vj → Yj . The actions are free, and the orbits are open subsets of
maximal integral submanifolds of D. Moreover, because [Xi , Yj ] = 0 these
actions commute. However, for reasons we now explain, we will end up using
a different set of vector fields to generate the action of G in §10.5.
The superposition formula is constructed by defining simpler Pfaffian
systems on integral manifolds of V̂ (∞) and V̌ (∞) , and using the G-action
to define a surjective mapping from the product of these manifolds to a
G-invariant open set U ⊂ Σ, in a way that carries solutions to solutions.
The mapping is based on choosing a slice S for the group action—i.e., a
submanifold in Σ which intersects each orbit in U transversely in a single
point. Given a slice, we can associate to each point x ∈ U a unique point
x ∈ S in the same orbit, and a group element γ(x) that takes x to x. (Thus,
S = γ −1 ({e}).) This enables us to identify U with the product G×S, but we
would like the identification to be equivariant with respect to the G-action.
For this to happen, we need the foliation of U by level sets of γ to be
G-invariant; in other words, the level sets must be integrals of a G-invariant
Frobenius system. One might take the 1-forms θ X as spanning a Frobenius
system whose integrals are transverse to the G-orbits; however, LXj θX i does

not belong to this system if any of the A, B or M terms is present in the

structure equations. We will need to make a few more adjustments to the
coframe to eliminate these terms.
Proposition 10.3.5 (Thm. 4.1 in [6]). Given a pair of 4-adapted coframes
(θX , π̂, π̌) and (θ Y , π̂, π̌), there are 1-forms
j
θ̂i = R̂ji θX + Ŝai π̂ a , θ̌i = Řji θYj + Šαi π̌ α

(where the components of matrices R̂, Ŝ are first integrals of V̂ (∞) , the
components of Ř, Š are first integrals of V̌ (∞) , and R̂, Ř are nonsingular)
such that
dθ̂i = 12 Cjk
i j
θ̂ ∧ θ̂k + 12 Giαβ π̌ α ∧ π̌ β ,
(10.9)
dθ̌i = − 12 Cjk
i j
θ̌ ∧ θ̌k + 12 Hab
i a
π̂ ∧ π̂ b .

The coframes (θ̂, π̂, π̌) and (θ̌, π̂, π̌) are called 5-adapted.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 298 10. Superposition for Darboux-Integrable Systems

i = Qi θ j for a nonsingular s × s
Corollary 10.3.6 (Lemma 5.8 in [6]). If θX j Y
matrix Q, and we define the matrix Λ = R̂Q(Ř)−1 , then the 1-forms

ω̂ i = θ̂i + Λij Šαj π̌ α , ω̌ i = θ̌i + (Λ−1 )ij Ŝaj π̂ a

satisfy

(10.10) dω̂ i = 12 Cjk

i
ω̂ j ∧ ω̂ k , dω̌ i = − 12 Cjk
i
ω̌ j ∧ ω̌ k .

The spans {ω̂ i } and {ω̌ i } coincide, and (by making constant-coefficient ad-
justments to R̂, Ř, Ŝ, Š) we can arrange that ω̂ i = ω̌ i at a given point x0 .
Exercise 10.3.7: Deduce the corollary from the proposition. 

Let X̂i and X̌j denote the vector fields annihilated by the π̂ and π̌
forms, and dual to θ̂i and θ̌j respectively. Note that, unlike previous steps,
5-adapted coframes are not a special case of the previous adaptation; in fact,
the 1-forms θ̂i and θ̌j are not necessarily in I. Nevertheless, both sets of
vector fields span D at each point, and they still satisfy bracket relations

[X̂j , X̂k ] = −Cjk

i
X̂i , i
[X̌j , X̌k ] = Cjk X̌i , [X̂i , X̌j ] = 0,

as well as

LX̂i π̂ a = LX̌i π̂ a = LX̂i π̌ α = LX̌i π̌ α = 0.

j k
Exercise 10.3.8: Verify these assertions, and show that LX̂i ω̂ j = Cik ω̂
j k
and LX̌ i ω̌ j = −Cik ω̌ .

Example (10.2.4 continued). In what follows, we will let U ⊂ R7 be the

open subset where q = 0. In U , we fix the reference point x0 where q = 1
and x = y = u = w = R = T = 0. (Note that, when we construct
integral surfaces of I, points where u = 0 may correspond to points where
the solution of the original PDE uxy = uy (ux + x)/u fails to be smooth.)
The construction of the 5-adapted coframes is, in general, quite deli-
cate, and depends on the structure of the Vessiot algebra. However, for our
example it is not difficult to solve a system of PDE for the components of
matrices R̂, Ŝ depending on x, w, R and Ř, Š depending on y, T , such that
θ̂i , θ̌j satisfy (10.9). This system can be simplified because, in all cases, R̂
and Ř preserve the flag in g∗ defined by the annihilators of the derived alge-
bras of g. In our case, since θX 2 annihilates the derived algebra [g, g], then θ̂ 2

is congruent (modulo π̂) to a multiple of θX 2 , and similarly θ̌ 2 is congruent

(modulo π̌) to a multiple of θY2 . So, it’s easier to solve for θ̂2 and θ̌2 first.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.4. Some results on group actions 299

Working in this way, we obtain


θ̂2 = θX
2
+ x η̂ + (R − w2 )σ̂ 1 = −q −1 dq + d(xw) + T dy,
θ̂1 = e−xw (θX
1
+ xσ̂ 1 ) = e−xw ((u/q)dq − du + (q − uT )dy) ,
θ̌2 = θY2 − T σ̌ 1 = −q −1 dq + w dx,
θ̌1 = θY1 − σ̌ 1 = q −1 (−du + (uw − x) dx) .
Using the formulas in Corollary 10.3.6, we obtain
ω̂ 1 = e−xw ((u/q)dq − du), ω̂ 2 = −q −1 dq + d(xw),
ω̌ 1 = q −1 (−du + u d(xw)) , ω̌ 2 = ω̂ 2 .
The dual vector fields are
∂ ∂ ∂ ∂ ∂
X̂ 1 = −exw , X̂ 2 = −q −u , X̌ 1 = −q , X̌ 2 = −q .
∂u ∂q ∂u ∂u ∂q
Note that X̂ 1 = X̌ 1 and X̂ 2 = X̌ 2 at the point x0 . 

10.4. Some results on group actions

Let G be a connected Lie group with Lie algebra g. In this section we
review how a set of vector fields on a manifold M satisfying the same bracket
relations as g generates a local action of G on M . It is instructive to begin
with the action of G on itself by left and right multiplication.
For a tangent vector v ∈ Te G, let vL and vR denote the corresponding
left- and right-invariant vector fields on G, coinciding with v at the identity.
Recall that the value of the exponential map at v is defined by flowing
by X = vL for one unit of time, starting at the identity. In other words, if
g → Φ(g, t) denotes the flow by X, then exp(v) = Φ(e, 1) and more generally
exp(tv) = Φ(e, t). It then follows by left-invariance of X that
(10.11) Φ(g, t) = Lg exp(tv) = Rexp(tv) g.
In this sense, the left-invariant vector field X induces the action of G on
itself by right multiplication. Similarly, if Y = vR , then flow by Y is given
by
(10.12) Ψ(g, t) = Rg exp(tv) = Lexp(tv) g.
Exercise 10.4.1: Verify that (10.11) and (10.12) give the correct flows. (In
the second case, you may need the fact that Ad(exp(tv)) fixes v.)

Let λ : G × M → M and ρ : G × M → M be left and right actions,

respectively. So, these are smooth mappings satisfying
(10.13) λ(e, m) = m, λ(ab, m) = λ(a, λ(b, m)),
(10.14) ρ(e, m) = m, ρ(ab, m) = ρ(b, ρ(a, m))
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 300 10. Superposition for Darboux-Integrable Systems

for all a, b ∈ G and m ∈ M . For v ∈ Te G, we define vector fields λ∗ v and

ρ∗ v on M by
 
d  d 
λ∗ v|m =  λ(exp(tv), m), ρ∗ v|m =  ρ(exp(tv), m)
dt t=0 dt t=0
for any m ∈ M .
Proposition 10.4.2 (cf. Prop. 3.1 in [24]). The mapping ρ∗ is a Lie algebra
homomorphism from g to X(M ), while λ∗ is an antihomomorphism, i.e.,
ρ∗ ([v1 , v2 ]) = [ρ∗ v1 , ρ∗ v2 ], λ∗ ([v1 , v2 ]) = −[λ∗ v1 , λ∗ v2 ].
(Recall that the Lie bracket on Te G is such that [v1 , v2 ]L = [v1L , v2L ].)
Exercises 10.4.3:
1. Verify the proposition for the case of G acting on itself.
2. For any m ∈ G, let λm (g) = λ(g, m) and ρm (g) = ρ(g, m). Show that
λm R m L
∗ v = λ∗ v and ρ∗ v = ρ∗ v. (The proposition follows.)

Proposition 10.4.4 (Prop. 3.2 in [24]). Let φ : g → X(M ) be an antiho-

momorphism. Then there exists a local left action λ of G on M such that
λ∗ = φ. In other words, there is an open neighborhood U of {e} × M inside
G × M , and a smooth mapping λ : U → M satisfying (10.13) for all a, b for
which both sides are defined.
It is not difficult to generalize Proposition 10.4.4 to get commuting left
and right actions:
Proposition 10.4.5. Let G be a Lie group of dimension d, and let {vi } be
a basis for Te G with [viL , vjL ] = −Cij k vL and [vR , vR ] = C k vR for constants1
k i j ij k
k = −C k . Let M be a manifold equipped with two sets of d vector fields
Cij ji
{Xi } and {Yj } which each have the same d-dimensional span at each point,
and satisfy [Xi , Xj ] = −Cij k , [Y , Y ] = C k Y and [X , Y ] = 0. Then there is
i j ij k i k
an open neighborhood U of {e} × M inside G × M and free group actions
λ : U → M and ρ : U → M such that ρ∗ vi = Xi and λ∗ vi = Yi . In addition,
the actions commute and have common orbits.

10.5. The superposition formula

We now have a pair of 5-adapted coframes on Σ, and dual vector fields
X̂i and X̌j which coincide at x0 . By Proposition 10.4.5 these vector fields
generate commuting local actions of G on Σ which are locally free; for the
sake of simplicity we will assume that there is an open subset U ⊂ Σ and
free left- and right-actions
λ : G × U → U, ρ:G×U →U
defined for all of G, such that ρ∗ : vi → X̂i and λ∗ : vj → X̌j .
Licensed to AMS.
1 The choice of sign on the structure constants is chosen to be compatible with (10.10).
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.5. The superposition formula 301

The equations (10.10) imply that {ω̂ i } = {ω̌ i } is a Frobenius system,

and by Exercise 10.3.8 its integral manifolds are G-invariant. We take S ⊂ U
to be the maximal integral manifold of this system passing through x0 . By
shrinking U if necessary, we can arrange that each G-orbit intersects S in
a single point. Then, as described above, to each x ∈ U we associate a
unique element of G that takes x to the point x where the orbit through x
intersects the slice S:
Proposition 10.5.1 (Lemma 5.3 through Corollary 5.6 in [6]). There is a
smooth mapping γ : U → G such that
(10.15) x = ρ(γ(x), x), x̄ ∈ S.
j
If ωLi and ωR denote the left- and right-invariant 1-forms on G dual to viL
R
and vj respectively, then

γ ∗ ωLi = ω̂ i and γ ∗ ωR
j
= ω̌ j .
Moreover, if we define Ψ : U → S × G by Ψ(x) = (x, γ(x)), then Ψ is
bi-equivariant, where G acts trivially on the factor S in the codomain.
Exercise 10.5.2: Show that γ(ρ(g, x)) = γ(x) · g and γ(λ(g, x)) = g · γ(x),
and use these to show the bi-equivariance of Ψ.

Example (10.2.4 continued). The flow generated by X̌1 is (u, q) → (u −

tq, q) while the flow generated by X̌2 is (u, q) → (u, e−t q), leaving all other
coordinates fixed. Accordingly, let G be the group of matrices of the form
 
1 −b
0 e−a
(where a, b are coordinates on G), so that the left action on column vectors
t(u, q) is multiplication by an element of G, i.e.,

λ(a, b, u, q) = (u − bq, e−a q).

The flow generated by X̂1 is (u, q) → (u − tewx , q) and the flow generated
by X̂2 is (u, q) → (e−t u, e−t q), so that the right action is
ρ(a, b, u, q) = (e−a u − bewx , e−a q).
Exercise 10.5.3: Verify that the left and right actions commute.

The first integrals of the Frobenius system {ω̂ 1 , ω̂ 2 } are u/q and qe−xw .
Therefore, the slice S through x0 is defined by u = 0 and q = exw . Equation
(10.15), which implicitly defines the mapping γ of Proposition 10.5.1, now
reads
(u, q) = ρ(a, b, 0, exw )
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 302 10. Superposition for Darboux-Integrable Systems

where a, b are the coordinates of γ(x). Solving this gives ea = exw /q, b =
−ue−xw , and thus
 
1 ue−xw
(10.16) γ(x) = .
0 qe−xw
Exercise 10.5.4: Compute the pullbacks  via γ of0 the left-invariant 1-forms
ωLi that are dual to the basis v1 = 00 −1
0 , v2 = 0
0 −1 for Te G, and confirm
that these pullbacks coincide with the ω̂ .i 

Let S1 denote the maximal integral manifold of V̂ (∞) |S and S2 the max-
imal integral manifold of V̌ (∞) |S passing through x0 . By the Frobenius
Theorem, on S there are functions pa such that V̂ (∞) |S = {dpa }, and func-
tions q α such that V̌ (∞) |S = {dq α }, and because T ∗ S is the direct sum of
these systems, the pa and q α form a local coordinate system on S. We can
arrange that S1 is the zero locus of the pa and S2 is the zero locus of the q α ,
and by concatenating the coordinates of a point on S1 with the coordinates
of a point on S2 , we have a diffeomorphism
χ : S1 × S2 → S
whose image is an open set S0 ⊂ S.
Let M1 and M2 be the intersections with U of maximal integral manifolds
through x0 of V̂ (∞) and V̌ (∞) respectively. Since these systems are G-
invariant, the actions λ and ρ restrict to M1 and M2 . Furthermore, if x1 ∈
M1 and x2 ∈ M2 , then xi ∈ Si = Mi ∩ S for i = 1, 2. We define the
superposition map Γ : M1 × M2 → M by
(10.17) Γ(x1 , x2 ) = ρ (γ(x1 )γ(x2 ), χ(x1 , x2 )) .
Using the results of Exercise 10.5.2, it is easy to show that Γ is invariant
under the diagonal action of G on M1 × M2 , i.e.,
Γ(ρ(g −1 , x1 ), λ(g, x2 )) = Γ(x1 , x2 ) ∀g ∈ G.
Thus, Γ covers a well-defined mapping from the quotient (M1 × M2 )/G to
U , which turns out to be a local diffeomorphism. To see this, we need to
examine how Γ relates the 5-adapted coframe on U to coframes on M1 and
M2 .
A coframe on M1 is given by the restrictions of θ̂, σ̌ and η̌, while on M2
we use the restrictions of θ̌, σ̂ and π̂. We will use a subscript i = 1 or i = 2
to denote the restriction of a 1-form to Mi , and we pull these 1-forms back
to the product M1 × M2 . Then we have:
Lemma 10.5.5 (Lemma 5.9 in [6]).
(10.18a) Γ∗ (σ̂) = σ̂ 2 , Γ∗ (η̂) = η̂ 2 , Γ∗ (σ̌) = σ̌ 1 , Γ∗ (η̌) = η̌ 1 ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.5. The superposition formula 303

and

(10.18b) Γ∗ (R̂ji θX
j
) = Λij (x2 )(θ̂1j + θ̌2j ).

Example (10.2.4 continued). Since x, w, R are the first integrals of V̂ (∞) ,

M1 ⊂ U is defined by x = w = R = 0, and similarly M2 is defined by
y = T = 0. We will use subscripts to denote the restrictions of the remaining
coordinates to each manifold, so that (y1 , u1 , q1 , T1 ) are coordinates on M1
and (x2 , u2 , q2 , w2 , R2 ) are coordinates on M2 . Within these, S1 is defined
by u1 = 0 and q1 = 1 while S2 is defined by u2 = 0 and q2 = ex2 w2 . The
mapping χ is defined by
x = x2 , y = y1 , u = 0, q = ex2 w2 , w = w2 , R = R2 , T = T2 .

For a point x1 ∈ M1 , the projection into S1 is given by

x1 = (y1 , u1 , q1 , T1 ) → x1 = (y1 , 0, 1, T1 ),
while for a point x2 ∈ M1 the projection into S2 is
x2 = (x2 , u2 , q2 , w2 , R2 ) → x2 = (x2 , 0, ex2 w2 , w2 , R2 ).
Using (10.16) and (10.17) we obtain
(10.19)
Γ : (y1 , u1 , q1 , T1 ; x2 , u2 , q2 , w2 , R2 ) → (x2 , y1 , q2 u1 + u2 , q1 q2 , w2 , R2 , T1 ).
 −xw 
e 0
Exercise 10.5.6: Using (10.19) and the values R̂ = and Λ =
0 1
 −xw 
e q −e−xw u
, verify equation (10.18b). 
0 1

The point of the superposition formula is that it relates solutions of

the Darboux-integrable system I to solutions of simpler Pfaffian systems on
lower-dimensional manifolds. The systems in question are the restriction of
V̂ to M1 and the restriction of V̌ to M2 , which from (10.5) are given by
W1 = {θ̂ 1 , η̌ 1 , Ω̌1 }alg , W2 = {θ̌ 2 , η̂ 2 , Ω̂2 }alg .
Exercise 10.5.7: Show that these are in fact Pfaffian systems. 

Let W = W1 + W2 denote the EDS on M1 × M2 generated by the

pullbacks from M1 of forms in W1 and pullbacks from M2 of forms in W2 .
Then from the above pullback formulas (10.18) we obtain:
Proposition 10.5.8. The image under Γ : M1 × M2 → Σ of any integral
submanifold of W is an integral submanifold of I.

Proof. It is equivalent to show that Γ∗ I ⊂ W. Generators of I are given

by Proposition 10.2.10 applied to the 4-adapted coframe. The pullbacks of
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 304 10. Superposition for Darboux-Integrable Systems

j
θ X , η̌, η̂ are all in W, hence so are their derivatives. On M1 , θ̂i = R̂ji θX , so
that
dθ̂ 1 ≡ R̂B σ̌ 1 ∧ σ̌ 1 mod θ̂ 1 , η̌ 1 .
Since the left-hand side belongs in W, the 2-forms B σ̌ 1 ∧ σ̌ 1 are in W, and
similarly for Aσ̂ 2 ∧ σ̂ 2 . Thus, the pullbacks of the generators of I given by
(10.7) all lie in W. 

Integrals of W = W1 + W2 are products of integrals of each system,

and the mapping Γ combines these to produce an integral of I; this is the
rationale behind the term ‘superposition’. Moreover, this mapping is locally
surjective on solutions:
Proposition 10.5.9. The system W is an integrable extension of I, relative
to the mapping Γ.

Proof. From (10.18) and the proof of the last proposition,

Γ∗ I = {θ̂ 1 + θ̌ 2 , η̌ 1 , η̂ 2 }diff ,
while by Exercise 10.5.7, W = {θ̂ 1 , θ̌ 2 , η̌ 1 , η̂ 2 }diff .
By the structure equations of the 4-adapted coframe (θX , π̂, π̌),
I = {θX , η̂, η̌, F̂ σ̂ ∧ σ̂, F̌ σ̌ ∧ σ̌, Aσ̂ ∧ σ̂, B σ̌ ∧ σ̌}alg .
The coefficients A and B are related to the coefficients in the structure
j
equations of the 5-adapted coframe by R̂ji Bαβ = Giαβ and Řji Ajab = Habi .

(These relations are obtained by differentiating the formulas and computing

the π̌ ∧ π̌ coefficients in dθ̂i , and the π̂ ∧ π̂ coefficients in dθ̌i .) This implies
in particular that the 2-forms Giαβ σ̌1α ∧ σ̌1β lie in Γ∗ I. Thus,

dθ̂ 1 ≡ 0 mod θ̂ 1 , Γ∗ I.
It follows that W = {θ̂1 , Γ∗ I}alg , and so is an integrable extension of I. 
Remark 10.5.10. We have just proved that any Darboux-integrable system
I arises (locally) as the quotient EDS of the product of two systems W1 and
W2 which have a common symmetry group G. The converse is also true:
given two such systems (generated by 1-forms and 2-forms on manifolds
M1 and M2 ) with a common symmetry group, the quotient of the product
system W1 + W2 is Darboux-integrable, under certain technical assumptions
(see Theorem 6.1 in [5]). Note that there is actually a larger group G × G
acting on M1 × M2 as symmetries of W1 + W2 , and we recover I as the
quotient by the action of the diagonal subgroup Gdiag ⊂ G × G. If we
form the quotient by a different subgroup, then under suitable regularity
assumptions the resulting quotient differential system is related to I by a
Bäcklund transformation (see Theorem 9.1 in [5]).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 10.5. The superposition formula 305

Example (10.2.4 continued). From Exercise 10.5.7 (and noting that there
are no 1-forms η̌),
W1 = {θ̂11 , θ̂12 }diff = {q1−1 dq1 − T dy, du1 − q1 dy}diff
and similarly
W2 = {θ̌21 , θ̌22 , η̂2 }diff = {dw+(w2 −R)dx, q2−1 dq2 −w dx, du2 +(x−u2 w)dx}diff .
(For the sake of convenience, we have omitted the subscripts on coordinates
x2 , y1 , w2 , R2 and T1 .) To finish the example, we will generate integral curves
for each of these systems, and use the superposition formula to produce
solutions for the PDE we started with.
Integral curves of W1 are easily obtained, by letting u1 = f (y), q1 = f  (y)
and T = f  (y)/f  (y) for an arbitrary monotone function f . For W2 , observe
that along integral curves
 
d u2 x
=− .
dx q2 q2
Thus, integral curves of W2 can be generated by letting u2 /q2 = g(x) for an
arbitrary monotone function g, and setting
x xg(x)
q2 = − , u2 = − .
g  (x) g  (x)
Formulas for w and T in terms of g and its derivatives can then be obtained
by substituting these into the system W2 . Finally, the superposition formula
gives
x
(10.20) u(x, y) = q2 u1 + u2 = −  (f (y) + g(x)).
g (x)

Exercises 10.5.11:
1. Check that (10.20) satisfies the PDE (10.4).
2. Carry out the coframe adaptations and construct the superposition for-
mula for the Darboux-integrable PDE uxy = ux uy /(u − x). In particular,
show that the Vessiot algebra is again 2-dimensional and non-Abelian (cf.
Example 6.2 in [6]).
3. Do the same for Liouville’s equation uxy = eu , and show that its Vessiot
algebra is sl(2). In particular, re-derive the solution formula (7.16) due to
Goursat (cf. Example 6.1 in [6]).
4. Show that the Born-Infeld system uy = vux , vy = uvx is Darboux-
integrable with a 1-dimensional Vessiot algebra.
5. The system
vx vy − ux uy ux vy + uy vx
uxy = u
, vxy = −
2(1 + e ) 2(1 + eu )
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 306 10. Superposition for Darboux-Integrable Systems

defines wave maps from the Minkowski plane into R2 with the confor-
mally flat metric (du2 + dv 2 )/(1 + e−u ). Show that the system is Darboux-
integrable, with Vessiot algebra R ⊕ sl(2).
6. Show that the system uxy = 2eu − ev , vxy = 2ev − eu is Darboux-
integrable after one prolongation, and its Vessiot algebra is sl(3).
7. Show that the system
(uxy + zuxx )uz (uyy + zuxy − z)uz
uxz = , uyz =
uy + zux − yz uy + zux − yz
is decomposable of type [5, 2], and is Darboux-integrable with a 3-dimension-
al nilpotent Vessiot algebra.
8. Show that the system
2u + 1 ux uz uy uz
uxy = ux uy , uxz = , uyz =
u(u + 1) u+1 u
is decomposable of type [4, 2], and is Darboux-integrable with a 2-dimension-
al non-Abelian Vessiot algebra.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/11

Chapter 11

Conformal Differential
Geometry

In this chapter we study conformal geometry from the perspective of Rie-

mannian geometry and as a G-structure problem. Conformal differential
geometry has been intensively studied because many of the differential equa-
tions from physics (e.g., Maxwell’s equations for photons, the Dirac equation
for neutrinos) are conformally invariant. One of our goals is to give an in-
troduction to, and partial synthesis of, the more advanced works [29, 56].
We begin in §11.1 by examining conformal differential geometry from a
Riemannian perspective. We determine how the curvature tensor changes
under a conformal change of metric, and set up EDS for metrics in a confor-
mal class that are Einstein and more generally of constant scalar curvature.
In §11.2 we examine conformal differential geometry as a G-structure. The
conformal analogs of Killing vector fields are discussed in §11.3. Many of
the natural differential operators in conformal geometry are not defined on
functions, but rather sections of line bundles, and language to deal with
these line bundles is developed in §11.4. With that language in place, we
then describe the conformal cousin of the Laplacian. We conclude in §11.5
by bringing several of the topics (Einstein metrics in a conformal class, the
conformal Laplacian, and a natural vector bundle associated to the princi-
pal bundle in the G-structure set-up) together to discuss the almost Einstein
manifolds defined by Gover [79].

307
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 308 11. Conformal Differential Geometry

11.1. Conformal geometry via Riemannian geometry

Change in curvature under a conformal change of metric. Before
studying conformal geometry as a G-structure, we examine it from the per-
spective of Riemannian geometry. That is, given a Riemannian manifold
(M, g), we examine how the connection and curvature change under a con-
formal change of metric. Write
ĝ = σ −2 g
for some smooth function σ : M → R+ . Fix a local orthonormal coframe η i ,
with associated connection forms ηj . We have g = i (η ) , ĝ = (σ −1 η i )2 ,
i i 2

so write η̂ i = σ −1 η i . Write dσ = σj η j for some functions σj . We compute

dη̂ i = −σ −1 dσσ −1 ∧ η i − σ −1 ηji ∧ η j
= −σk η̂ k ∧ η̂ i − ηji ∧ η̂ j ,
so the change in connection form is
(11.1) η̂ji = (σi η̂ j − σj η̂ i ) + ηji .
Now for the change in curvature:
Θ̂ij = dη̂ji + η̂ki ∧ η̂jk
= (dηji + ηki ∧ ηjk ) + (σi η̂ k − σk η̂ i ) ∧ (σk η̂ j − σj η̂ k ) + σi dη̂ j − σj dη̂ i
− dσj ∧ η̂ i + dσi ∧ η̂ j + (σi η̂ m − σm η̂ i ) ∧ ηjm + ηmi
∧ (σm η̂ j − σj η̂ m )

= Θij − ( σk2 )η̂ i ∧ η̂ j − (dσj − σm ηjm ) ∧ η̂ i + (dσi − σm ηim ) ∧ η̂ j
k
(11.2)

= Θij − σ −2 ( σk2 )η i ∧ η j − σ −1 (dσj − σm ηjm ) ∧ η i
k
−1
+σ (dσi − σm ηim ) ∧ η j .

To understand (11.2), recall from Exercise 3.1.17 that for a function on

a Riemannian manifold f : M → R, one has ∇df ∈ Γ(S 2 T ∗ M ), so writing
∇g dσ = (dσj − σk ηjk )⊗η j = sij η i η j , the last two terms become
− σ −1 (sjk η k ∧ η i − sik η k ∧ η j )
= − σ −1 (δil sjk − δik sjl − δjl sik + δjk sil )η k ∧ η l .
It will be convenient to lower the first index in Θij - we do so with respect
to g.
Recall the Kulkarni-Nomizu product from §3.5: (h  k)ijkl = hik kjl −
hil kjk − hjk kil + hjl kik . Thus the last two terms in (11.2), with first index
lowered, contribute σ −1 (∇g dσg). The second term with first index lowered,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.1. Conformal geometry via Riemannian geometry 309

contributes − 12 σ −2 |dσ|2g g g. Raising the first index back up via g : T ∗ M →

T M (and noting that êi ⊗ η̂ j = ei ⊗η j ), we conclude:
Proposition 11.1.1. Under a conformal change of metric ĝ = σ −2 g, we
have the corresponding change in curvature tensors:
1
Θ̂ = Θ + (g ⊗IdT ∗⊗3 )[σ −1 ∇g dσ − σ −2 |dσ|2g g]  g.
2
Corollary 11.1.2. The (3, 1) Weyl tensor W ∈ Γ(T M ⊗(T ∗ M )⊗3 ) is invari-
ant under a conformal change of metric ĝ = σ −2 g.

Proof. The totally traceless part of Θ̂ equals the totally traceless part of
Θ. 

Recall the Schouten tensor P ∈ Γ(S 2 T ∗ M ) defined by (3.12) satisfies

R = W + P  g.
Corollary 11.1.3. Under a conformal change of metric ĝ = σ −2 g, the
Schouten tensor P changes as follows:
1
P̂ = P + σ −1 ∇g dσ − σ −2 |dσ|2g g.
2
Recall from (3.12) that the scalar curvature can be recovered from the
1
Schouten tensor: s = 2(n−1) traceg (P).
Corollary 11.1.4. Under a conformal change of metric ĝ = σ −2 g, the scalar
curvature changes as follows:
σ n
(11.3) ŝ = σ 2 s + traceg (∇g dσ) − |dσ|2g .
2(n − 1) 2(n − 1)
Corollary 11.1.3 shows that it is possible to modify the Ricci tensor
under a conformal change of metric. How much of Ricci can be normalized
to zero?
Exercise 11.1.5: Show that for any conformal structure (M, [g]) and any
x ∈ M , there exists ĝ ∈ [g] with the Ricci curvature of ĝ vanishing at x.
(However in general one cannot make Ricci vanish in a neighborhood of x.)

When n ≤ 3, there is no Weyl curvature, which begs the question: What,

if any, are the differential invariants of a conformal structure when n ≤ 3?
This question will be answered when n = 3 by Proposition 11.2.9. We can
answer the n = 2 case immediately: CO(2)  GL(1, C) and the local dif-
ferential invariants of conformal metrics correspond to the local differential
invariants of complex manifolds (of complex dimension one), of which there
are none, i.e., all CO(2)-structures are flat.
Exercise 11.1.6: Show that if (M, g) is a space form, then locally there
exists ĝ ∈ [g] that is a flat metric. (The global version of this result is false
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 310 11. Conformal Differential Geometry

in general, e.g., consider what the Gauss-Bonnet Theorem implies for S 2 .)



Einstein metrics and metrics with constant scalar curvature in a

conformal class. As mentioned in §3.4, a Riemannian metric is Einstein
if Ric0 = 0. Corollary 11.1.3 implies:
Proposition 11.1.7. Let (M, g) be Riemannian with scalar curvature s :
M → R and Schouten tensor P ∈ Γ(M, S 2 T ∗ M ). Let σ : M → R+ be a
smooth function. Then the metric ĝ := σ −2 g has constant scalar curvature
if and only if
σ n
(11.4) σ2s + traceg (∇g dσ) − |dσ|2g
2(n − 1) 2(n − 1)
is constant. The metric ĝ is Einstein if and only if there exists l ∈ C ∞ (M )
such that
(11.5) P + σ −1 ∇g dσ + lg = 0.

We set up the EDS for metrics in a conformal class that are of constant
scalar curvature: Let F denote the orthonormal frame bundle n+1
of the Rie-
mannian manifold (M, g). Work on Σs ⊂ F × R+ × R × R n ( 2 ) where the
second, third and fourth factors have coordinates (σ, σj , sij ) and Σs is the
hypersurface given by setting (11.4) to zero. Consider the differential forms
θ := dσ − σj η j ,
θj := dσj − σm ηjm − sjm η m
and the ideal they generate. Integral manifolds of this ideal correspond to
metrics in the conformal class of constant scalar curvature.
On Σs we have the identity
. σ n /
0 = d σ2s + traceg (∇g dσ) − |dσ|2g
2(n − 1) 2(n − 1)
which will be needed in the following exercise.
Exercise 11.1.8: Show that the system on Σs is involutive, so that locally
there are many metrics in a conformal class with constant scalar curvature.


Here is an EDS for metrics

n+1
in a conformal class that are Einstein: Let
ΣE ⊂ F × R+ × R × R
n ( 2 ) × R, where the first four factors are as in the
scalar curvature EDS and and the last factor has coordinate l, and ΣE is
the submanifold given by the n+1 2 equations (11.5).
Exercise 11.1.9: Write out the EDS for Einstein metrics in the conformal
class of a fixed Riemannian metric, and show that for a generic Riemannian
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.2. Conformal differential geometry as a G-structure 311

metric there are no local integral manifolds. What are the local integral
manifolds when (M, g) is Euclidean space?

The problem of finding Einstein metrics in a given conformal class dates

back at least to [140]. We discuss it further in §11.5.

11.2. Conformal differential geometry as a G-structure

We now study conformal differential geometry from the G-structure per-
spective, where G = CO(V ) is the conformal group. The presentation in
this section is heavily influenced by [29].
Since co(V ) ⊃ so(V ), we know there exists a torsion-free connection
associated to any CO(V )-structure, and since the containment is strict, we
know such connections are not unique.
Exercise 11.2.1: Show that co(V )(1)  V ∗ as a co(V )-module, and that
co(V )(2) = 0. 

Let M n be a manifold with conformal structure [g], equivalently a

CO(V )-structure FCO(V ) → M . By Exercise 11.2.1 and the general the-
ory developed in §9.3, the proper bundle to work on is the prolongation of
FCO(V ) , which has fiber CO(V )  V ∗ , where V ∗ = co(V )(1) acts additively
on the fibers and the product is semi-direct as CO(V ) acts on V ∗ . (This
action is given explicitly in Exercise 11.2.4 below.) By Exercise 11.1.6, our
flat model of conformal geometry could be any space form: En , S n or H n .
Because S n is compact, and the inversions (see §11.3 below) have a nice
interpretation on S n , we will use it as our flat model. It will also lead to the
desired principal bundle in the general case.

A flat model. To have a model of S n equipped with a conformal structure,

let W = Rn+2 have an inner product Q of signature (n + 1, 1), and let
N ⊂ W \{0} denote the null cone. It has two components. Call one of them
N+ . (In the language of relativity theory, this choice is the choice of a time
orientation.)
Then PN+ ⊂ PW is diffeomorphic to S n . Moreover, for x ∈ PN+ , a
choice of v ∈ N+ with [v] = x determines a positive definite inner product
Qv on Tx PN+ = x̂∗ ⊗ (Tv N+ /x̂), as Q|N+ has the one-dimensional kernel
equal to x̂∗ ⊗ Tv x̂ ⊂ x̂∗ ⊗ Tv N+ . (Here x̂ ⊂ W is the line corresponding
to x.) When we quotient out by this kernel we obtain a positive definite
inner product: Qv (α⊗z mod x̂, β ⊗w mod x̂) = α(v)β(v)Q(z, w). Moreover,
different choices of v ∈ (x̂)+ re-scale the corresponding inner product, so
PN+ is indeed equipped with a conformal structure.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 312 11. Conformal Differential Geometry

Consider the projection

π : SO(n + 1, 1) → S n ,
(e0 , . . . , en+1 ) → [e0 ],
where e0 , . . . , en+1 are the columns of a matrix in SO(n + 1, 1).
Exercise 11.2.2: Show that the fibration π : SO(n + 1, 1) → S n = PN+
has fiber CO(V )  V as desired. 

Let Q be represented by the matrix

⎛ ⎞
0 0 −1
⎝ 0 Idn 0 ⎠ .
−1 0 0
Exercise 11.2.3: What are the matrix expressions of so(n + 1, 1) and
SO(n + 1, 1) in this basis?

The fiber of π over t (1, 0, . . . , 0) is the parabolic group

⎧ ⎛ 2 2t 2 ⎞ ⎫
⎨ r r vB r2 t vv B ∈ SO(n)⎬
(11.6) P1,n+1 = h = ⎝ 0 B v ⎠ | v ∈ Rn .
⎩ −2 2 ∈R ⎭
0 0 r r +

By definition, a parabolic subgroup of GL(W ) is one preserving a flag in W .

In general, a parabolic subgroup of a reductive group G is one preserving
a flag with additional structure. In our case, the additional structure on
the flag {e0 } ⊂ {e0 , e1 , . . . , en } ⊂ W is that the line is Q-isotropic and the
(n + 1)-plane is the Q-annihilator of the line. Write the Maurer-Cartan form
of SO(n + 1, 1) as
⎛ ⎞
2ρ βj 0
(11.7) ψ = ⎝ω i αji βi ⎠ ,
0 ω −2ρ
j

where 1 ≤ i, j ≤ n, αji + αij = 0, and otherwise all forms appearing are

linearly independent. Since
de0 = 2ρe0 + ω i ei
≡ ω i ei mod e0 ,

the ω i are semi-basic for the projection to S

n . Note also that if we just

project to N+ , then Q restricted to Te0 N+ is i (ω i )2 , and thus [ (ω i )2 ] is
our conformal metric.
Pulled back to a fiber of π, i.e., an integral manifold of {ω i = 0}, ψ is
the Maurer-Cartan form of P1,n+1 .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.2. Conformal differential geometry as a G-structure 313

The general case. Let M be an n-dimensional manifold equipped with a

CO(V )-structure. On FCO(V ) we have tautological forms ω i and there exists
a co(V )-valued connection form θ such that dω = −θ ∧ ω. This connection
(1)
form is not unique so we work on the bundle FCO(V ) which fits into the
following picture:
(1)
co(V )(1) −−−−→ FCO(V )
⏐
⏐
-
CO(V ) −−−−→ FCO(V )
⏐
⏐
-
M.
(1)
By pulling back ω, θ, we have a V -valued one-form ω on FCO(V ) , semi-basic
to M , and a co(V )-valued one-form θ such that dω = −θ ∧ ω.
Exercise 11.2.4: Recall that CO(V ) V ∗ is the group of pairs (A, b) where
A ∈ CO(V ), b ∈ V ∗ is a row vector, and (A1 , b1 ) · (A2 , b2 ) = (A1 A2 , b1 +
b2 A−1
1 ) is the multiplication rule.
(a) Show that

(u, θ) · (A, b) = A−1 u, A−1 (θ + ω tb − btω)A

defines a faithful right action of CO(V )  V ∗ on the fiber of FCO(V ) over

(1)

M.
(b) Show that A = r−2 B, b = tv defines a group isomorphism from P1,n+1
to CO(V )  V ∗ . Thus, we conclude that the fiber of FCO(V ) is isomorphic
(1)

to P1,n+1 .

To further the comparison with the flat case, define the ray bundle R+ ⊂
S 2 T+∗ M whose fiber corresponds to the choices of g ∈ [g], which generalizes
the choices of e0 ∈ [e0 ] that fixed a pointwise choice of metric in the flat
case. Write θ = 2ρ IdV +α where α is so(V )-valued and ρ is a scalar-valued
1-form. Then, since ρ corresponds to infinitesimal dilations, the forms semi-
(1) (1)
basic to the projection FCO(V ) → R+ , where R+ = FCO(V ) /(SO(V )  V ),
are ω i , ρ.
Following the discussion in §9.3, we want to obtain a V ∗ -valued 1-form
(1)
β to give a connection form for the bundle FCO(V ) → FCO(V ) . In particular,
(1)
we will need ω, β, ρ, α to give a coframing of FCO(V ) . Differentiating both
sides of

(11.8) dω i = 2ρ ∧ ω i − αji ∧ ω j
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 314 11. Conformal Differential Geometry

shows
(−2δij dρ + dαji + αki ∧ αjk ) ∧ ω j = 0,
which, by the generalized Cartan Lemma, implies −2δij dρ + dαji + αki ∧
αjk ≡ 0 mod{ω 1 , . . . , ω n }. In particular, setting i = j we see that dρ ≡
(1)
0 mod{ω 1 , . . . , ω n }. Thus there exist forms βi ∈ Ω1 (FCO(V ) ) such that

(11.9) 2dρ = −βi ∧ ω i .

Remark 11.2.5. In terms of our calculations on FCO(V ) , dαji + αki ∧ αjk
gives curvature with respect to one choice of torsion-free connection and
2δij dρ gives the curvature’s modification by making a different choice. At
the level of linear algebra, this freedom is the image of δ1 : co(V )(1) ⊗V ∗ →
co(V )⊗Λ2 V ∗ intersected with V ⊗S 2 1 V ∗ . This intersection is isomorphic
3
to S 2 V ∗ .

The choice of βi is not unique, as setting β̃i = sij ω j + βi with sij = sji ,
one gets the same equation with β̃i , as was predicted by Remark 11.2.5.
Writing dαji + αki ∧ αjk − βj ∧ ω i + βi ∧ ω j = 12 Rjkl
i ω k ∧ ω l , we see that under

such a change,
i
R̃jkl i
= Rjkl + (−δli sjk + δlj sik − δkj sil + δki sjl ),
which is similar to, but simpler than the expression (11.2). In particular,
since the change here is by s  g where s is an arbitrary quadratic form, it
is clear that one can choose s to make R̃ traceless. To see this explicitly, set
l = i and sum:

i
R̃jki i
= Rjki + (−nsjk + sjk − δkj sii + sjk )
i

= i
Rjki − (n − 2)sjk − δjk sii .
i
Thus if we take
1 1 
(11.10) sjk = i
(Rjki − i
δjk Rlli ),
n−2 2n − 2
l

then i
R̃jki = 0.
Remark 11.2.6. The above normalization is another example of the utility
of working on the proper bundle. With a fixed metric in a conformal class
we can have Ricci vanishing at a point by Exercise 11.1.5. Here we obtain
the analog of Ricci vanishing at all points.

Equation (11.10) is the most natural (but not canonical) choice to nor-
malize the curvature. We now have our coframe.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.2. Conformal differential geometry as a G-structure 315

(1)
Proposition 11.2.7. On FCO(V ) there is a unique coframing ω, α, ρ, β where
the forms are V -, so(V )-, R-, and V ∗ -valued respectively, satisfying
dω = (2ρ IdV −α) ∧ ω,
dρ = − 12 β ∧ ω,
dα = −α ∧ α + ω ∧ β − t (ω ∧ β) + W,
where W = 12 Wjkl i (e ⊗ ω j ) ⊗ ω k ∧ ω l satisfies the symmetries of the Weyl
i
curvature tensor. Here we write β as a row vector and ω as a column
vector, so, e.g., β ∧ ω is the scalar-valued 2-form βi ∧ ω i . With indices, these
equations read as
dω i = (2ρδji − αji ) ∧ ω j ,
1
dρ = − βi ∧ ω i ,
2
1 i k
dαji = −αki ∧ αjk + βj ∧ ω i − βi ∧ ω j + Wjkl ω ∧ ωl .
2
It remains to study dβi , which may or may not contain new differen-
tial invariants. When n = 3 we expect it to do so, because otherwise all
conformal structures in three dimensions would be flat. Differentiating the
structure equations gives
(11.11) 0 = 2d2 ρ = −dβj ∧ ω j + βi ∧ (2δji ρ − αji ) ∧ ω j ,
(11.12) 0 = d2 αji = 12 (DW )ijkl ∧ ω k ∧ ω l + dβj ∧ ω i − dβi ∧ ω j
− βk ∧ (2δki ρ − αjk ) ∧ ω i + βk ∧ (2δkj ρ − αik ) ∧ ω j .
Here
(11.13) (DW )ijkl := dWjkl
i m i
+ Wjkl θm − Wmkl
k
θjm − Wjml
k
θkm − Wjkm
k
θlm .
By the generalized Cartan Lemma A.1.9, (11.11) implies
dβj + 2ρ ∧ βj + βi ∧ αji = Cjk ∧ ω k
(1)
for some Cjk ∈ Ω1 (FCO(V ) ). Substitute this into the second equation to
obtain
[ 12 (DW )ijkl − Cik δlj + Cjk δli + Cil δkj − Cjl δki ] ∧ ω k ∧ ω l = 0,
so the term in the brackets is semi-basic. Setting i = l and summing we
obtain 
(n − 2)Cjk + δkj Cii ≡ 0 mod{ω 1 , . . . , ω n }.
i
Cjk ≡ 0 mod{ω 1 , . . . , ω n }. Examining
Examining the case j = k we see that
the term j = k gives (n − 1)Cjj + i=j Cii ≡ 0 mod{ω 1 , . . . , ω n }. Since
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 316 11. Conformal Differential Geometry

n ≥ 3, this implies the Cjj are semi-basic as well. Thus we may write

(11.14) dβi + 2ρ ∧ βi + βm ∧ αim = Cijk ω j ∧ ω k

for some functions Cijk = −Cikj . The equation d2 ρ = 0 implies a Bianchi-

type identity and we conclude C = Cijk ω i ⊗ω j ∧ ω k is S 2 1 V ∗ -valued.
3
We next examine what geometric information is in the tensor C; in
particular we determine whether or not it descends to be a section of some
bundle that is tensorial in T ∗ M and T M , as W does.
Consider (11.12) with (11.14) substituted in. We obtain

(11.15) [ 12 (DW )ijkl − Cikl ω j + Cjkl ω i ] ∧ ω k ∧ ω l = 0,

which implies that DW is semi-basic. Moreover, (11.13) shows that the

(1)
coefficients of dW vary along the fibers of FCO(V ) → M in a way that
exactly cancels the variation of ei ⊗ ω j ⊗ ω k ∧ ω l , so the quantity W :=
∗ ⊗3
2 Wjkl (ei ⊗ω )⊗ω ∧ ω descends to a section of T M ⊗(T M ) , which we
1 i j k l

already knew from our previous calculations regarding Weyl curvature.

Note that when n = 3, the equation [−Cikl ω j + Cjkl ω i ] ∧ ω k ∧ ω l = 0 is
just the identity 0 = 0, so we get no additional information from it.
Now, assuming n > 3, contract (11.15) with et , and then es , where
s, t, i, j are distinct indices. We obtain

− (DW )ijst + (DW )ijkt (es )ω k − (DW )ijks (et )ω k

− (DW )ijtl (es )ω l + (DW )ijsl (et )ω l + Cist ω j − Cjst ω i = 0.

Contract further with ei , and re-name i as i0 to emphasize there is no sum

over i0 in the following expression that we obtain:

(11.16) Cjst = − (DW )ijst

0
(ei0 ) + (DW )iji00 t (es ) − (DW )iji00 s (et )
− (DW )ijti
0
0
(es ) + (DW )ijsi
0
0
(et ).

That is, when n > 3 the tensor C can be recovered from DW .

Exercise 11.2.8: Show that when n = 3, C = Cijk ω i ⊗ω j ∧ ω k descends to
give a well-defined section of S 2 1 T ∗ M , called the Cotton-York tensor.
3

In summary:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.2. Conformal differential geometry as a G-structure 317

(1)
Proposition 11.2.9. On FCO(V ) there is a unique coframing ω, α, ρ, β where
the forms are respectively V -, so(V )-, R-, and V ∗ -valued, satisfying

dω = (2ρ IdV −α) ∧ ω,

1
dρ = − β ∧ ω,
2
dα = −α ∧ α + ω ∧ β − t (ω ∧ β) + W,
dβ = −2ρ ∧ β − β ∧ α + C,

where C is S 2 1 V ∗ -valued and W satisfies the symmetries of the (3, 1)-Weyl

3
tensor.
When n = 3, the tensor W is zero and the conformal structure is deter-
mined by the tensor field C. When n > 3, C can be recovered from the first
derivative of W and the conformal structure is determined by the tensor
field W .

Representation theory and the second Bianchi identity. In this sub-

section we give an explanation of (11.16) in terms of representation theory
and describe additional consequences of the second Bianchi identity. We
work with a fixed Riemannian metric and deal with the (4, 0)-Riemann cur-
vature and Weyl tensors, as the explanation is most easily seen from a
Riemannian perspective.
The Pieri rule (see §3.5 for the Pieri rule and the notation used in this
paragraph) implies S 1 3 V ∗ ⊗ V ∗ = S 1 3 5 V ∗ ⊕ S 1 3 V ∗ , and the second
2 4 2 4 2 4
5
Bianchi identity says that ∇R, which a priori lives in S 1 3 V ∗ ⊗V ∗ , in fact
2 4
it lives in S 1 3 5 V ∗ . As an O(V )-module, the complement to S 1 3 5 V ∗ is
2 4 2 4
S221 V ∗ = (S221 V ∗ )0 ⊕ (S21 V ∗ )0 ⊕ V ∗ , so the second Bianchi identity poten-
tially gives us three identities, one for each irreducible component of S221 V ∗ .
We have the following decompositions into irreducible O(V )-modules:

V ∗ ⊗(S 2 V ∗ )0 = (S21 V ∗ )0 ⊕ (S3 V ∗ )0 ⊕ V ∗ ,

V ∗ ⊗(S22 V ∗ )0 = (S32 V ∗ )0 ⊕ (S21 V ∗ )0 ⊕ (S221 V ∗ )0 ,
S32 V ∗ = (S32 V ∗ )0 ⊕ (S 3 V ∗ )0 ⊕ (S21 V ∗ )0 ⊕ V ∗ .

We immediately conclude that the potential (S221 V ∗ )0 component of ∇W

must be zero because it cannot be canceled by any of the other terms. In
addition, there must be relations among ∇s (where s denotes the scalar
curvature) and the V ∗ component of ∇ Ric0 because the V ∗ term in S221 V ∗
cannot appear in ∇R, and similarly, the (S21 V ∗ )0 components of ∇ Ric0 and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 318 11. Conformal Differential Geometry

∇W must satisfy a relation to ensure (S21 V ∗ )0 in S221 V ∗ does not appear

in ∇R. The following propositions make these relations explicit:
Proposition 11.2.10. For all j, in any orthonormal frame,
1 1 
(∇s)j − (∇ Ric0 )jii = 0.
2n n−2
i

Proposition 11.2.10 is proved below. It may be rephrased in terms of

the Schouten tensor P as follows:
View ∇P as taking values in S 2 V ∗ ⊗V ∗ . Then the Cotton-York tensor
C is the image of ∇P under the skew-symmetrization map S 2 V ∗ ⊗ V ∗ →
V ∗ ⊗Λ2 V ∗ ; as we have seen, the image is the module S 2 1 V ∗  S21 V ∗ .
3

Proposition 11.2.11. C ∈ (S 2 1 V ∗ )0 . In other words, j Cjjk = 0.
3

We will show that when n > 3, the Cotton-York tensor lives precisely
in the copy of (S21 V ∗ )0 that lives in S221 V ∗ , so in particular there must
be a relationship between it and the component of ∇W in (S21 V ∗ )0 . That
relation is:
Proposition 11.2.12. For all i, j, k

(∇W )ijkmm + (n − 3)Cijk = 0.
m

Proposition 11.2.12 encodes the relation between ∇ Ric0 and ∇W . When

n = 3 this identity is 0 + 0 = 0.
Remark 11.2.13. In comparing with (11.16), here we have ∇W =
(∇W )ijklm η i ⊗ η j ⊗ η k ⊗ η l ⊗ η m ∈ Γ(T ∗ M⊗5 ), whereas there (DW )ijkl is a
1-form.

Proof of Propositions 11.2.11 and 11.2.12. Cyclically sum

(∇P)ikm δjl − (∇P)ilm δjk − (∇P)jkm δil + (∇P)jlm δik + (∇W )ijklm
over m, i, j to get a quantity that is zero. Then set i = k and sum over i
(so the W terms drop out) to obtain Proposition 11.2.11, and setting l = m
and summing over l gives Proposition 11.2.12. 
Remark 11.2.14. Alternatively, one can prove Proposition 11.2.12 via
(11.16).
Exercise 11.2.15: Write out the details of the proofs above; in particular
verify that C indeed only depends on ∇ Ric0 .

Proof of Proposition 11.2.10.

1 s
Pij = Ric0ij + δij ,
n−2 2n(n − 1)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.3. Conformal Killing vector fields 319

so
(∇P)ijm − (∇P)imj
1 sm 1 sj
= (∇ Ric0 )ijm + δij − (∇ Ric0 )imj − δim .
n−2 2n(n − 1) n−2 2n(n − 1)
Setting i = j and summing over i gives the desired identity. 

11.3. Conformal Killing vector fields

We now study continuous symmetries of a conformal structure. As with the
Riemannian case treated in §3.6, this question is best studied infinitesimally.
We call X ∈ Γ(T M ) a conformal Killing vector field if the flow of X pre-
serves the conformal structure. By Theorem 9.6.2, the maximum possible
dimension of the space of conformal Killing vector fields is dim so(n + 1, 1),
and this occurs in the conformally flat case.

The flat case. On the conformally flat sphere S n , each element of

so(n + 1, 1) gives rise to a conformal Killing vector field, and if the element
is in p1,n+1 , then the origin [e0 ] = [t (1, 0, . . . , 0)] is fixed by the flow.
For example, the elements of SO(n + 1, 1) ⊂ P1,n+1 fixing [e0 ] consist of
matrices of the form
⎛ ⎞
1
⎝ B ⎠
1
where B ∈ SO(n). Under stereographic projection z → z/z0 the actions
of these matrices become rotations of Rn about the origin. Similarly, the
dilations fixing [e0 ] are
⎛ 2 ⎞
r
⎝ Idn ⎠
r −2

where r ∈ R+ . These correspond under stereographic projection from en+1

to dilations about the origin.
Now what about translations in Rn ? What is their analog on the sphere?
They should move [e0 ] and fix the “point at infinity” [en+1 ] = [t (0, . . . , 0, 1)].
These are given by elements of SO(n + 1, 1) of the form
⎛ ⎞
1
⎝ w Idn ⎠
1
2 |w| 2 tw 1
where w ∈ Rn .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 320 11. Conformal Differential Geometry

We still have to account for the conformal symmetries that infinitesi-

mally correspond to vectors in co(V )(1)  V ∗ . These are group elements
⎛ ⎞
2 |v|
1 2
1 v
⎝ Idn tv ⎠ ∈ P
1,n+1
1
where v ∈ Rn∗ . What do these do to Rn ? Note their similarity to the trans-
lations: they play the role of translations if one stereographically projects
from [e0 ] instead of [en+1 ]. So we could think of these as “translations with
the origin fixed”.
Better, as Cartan points out, both the translations and these symmetries
should be viewed on the sphere, where they are inversions, which visually
look like turning the sphere inside out, with one fixed point. The translations
are just special cases of inversions where the fixed point is the point at
infinity.
Elements of SO(n + 1, 1) may be written as:
⎛ ⎞⎛ 2 ⎞⎛ ⎞
2 |v|
1 2
1 r 1 v
(11.17) ⎝ w Idn ⎠ ⎝ B ⎠ ⎝ Idn tv ⎠
.
2 |w|
1 2 tw 1 r−2 1
Tracing through the stereographic projection map to Rn and differentiating,
we recover the following classical result:
Proposition 11.3.1. On a conformally flat Rn , with coordinates
(x1 , . . . , xn ), the general conformal Killing vector field is of the form
1
X(x) = (wi + mij xj − 2lxi − v, xxi + |x|2 v i )ei
2
i is an orthonormal frame for some  ,  ∈ [g], |x| is the norm
∂
where ei = ∂x
with respect to that inner product, and wi , mij = −mji , l, v i are constants.
It is worth looking at the inversions that fix the origin infinitesimally:
(11.18) X(x) = (−2lxi − v, xxi + 12 |x|2 v i )ei
for some v ∈ Rn .
Proposition 11.3.2. The flow φX (x, t) = (φ1 , . . . , φn ) of X in (11.18) is
given by
xj − 12 tv j |x|2
(11.19) φj (x, t) = .
1 − tv, x + 14 t2 |v|2 |x|2

Proof. While a direct proof is possible, the following proof is also instruc-
tive. The idea is to change coordinates via the antipodal map α of the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.3. Conformal Killing vector fields 321

sphere, which on Rn \0 = S n \{p, q} is α(x) = − |x|x2 =: y. Then

∂ 1 ∂ 2xj xk ∂
α∗ =− 2 j + ,
∂x j |x| ∂y |x|4 ∂y k
so in these coordinates

x
α(φX (x, t)) = φα∗ X (α(x), t) = y + 12 tv = − + 1 tv.
|x|2 2
Applying the antipodal map to both sides gives
x
x
|x|2
− 2t v
1
φX (x, t) = α(− 2 + 2 tv) = x .
|x| | |x|2 − 2t v|2

So
xj − 2t v j |x|2
xj (t) = .
1 − tv, x + 14 t2 |v|2 |x|2

Remark 11.3.3. The above proof, and the proof of Liouville’s theorem
below were communicated to us by M. Eastwood.

We may use Proposition 11.3.2 to recover the classical Liouville theorem:

Theorem 11.3.4 (Liouville). Any diffeomorphism ψ : Rn → Rn satisfying
ψ(0) = 0, dψ0 = IdT0 Rn , and ψ ∗  ,  = σ −2  ,  for some function σ, is of
the form (11.19) with t = 1.

Proof. Let ψ be a conformal change of metric as in the theorem. It can

be put in a one-parameter family of conformal changes of metric ψt , where
ψ1 = ψ and ψ0 = Id. Thus ψt is the flow of a conformal Killing field, and is
therefore of the form (11.19). 

We know from §9.6 that the largest possible space (even locally) of con-
formal Killing vector fields on anyconformal manifold (when n ≥ 3) is given
by V ⊕ co(V ) ⊕ V ∗ (i.e., a n+22 -dimensional space), and this occurs for
conformally flat manifolds. In the Riemannian case there were examples of
nonflat manifolds (namely, space forms) with the maximal space of (Rie-
mannian) Killing vector fields. This does not occur in conformal geometry:
Proposition 11.3.5. Let (M, [g]) be a manifold of dimension n ≥ 3 with
conformal structure. If the
 dimension of the space of conformal Killing
n+2
vector fields on M is 2 , then (M, [g]) is conformally flat.
(1)
Proof. The hypothesis implies that any element of FCO(V ) over a fixed
(1)
point x gives rise to a conformal Killing vector field by embedding FCO(V )
into the bundle of 2-jets of vector fields. (Recall that FGL(V ) (M ) may be
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 322 11. Conformal Differential Geometry

(1)
embedded in J 1 (M, V ) and similarly FCO(V ) may be identified with a sub-
bundle of the bundle of 2-jets sitting inside J 1 (J 1 (M, V )).) This in turn
(1)
implies, via parallel transport in FCO(V ) , that the coefficients of W, C must
be constant. In particular, (DW )ijkl , as defined in (11.13), is zero (by the
same argument as the proof of Proposition 9.7.7), but this implies C = 0 by
(11.16) (assuming n > 3).
l β
Exercise 11.3.6: Compute dCijk and show it has a term of the form Wijki l
appearing in it.

Since C = 0, one has DC = 0, which by Exercise 11.3.6 implies W = 0.

The case n = 3 is similar, one differentiates (11.14). 

11.4. Conformal densities and the Laplacian

Densities. In Riemannian geometry we needed differential invariants that
took values in vector bundles that were tensor constructs of the tangent
and cotangent bundles to distinguish between different Riemannian struc-
tures. Unlike in Riemannian geometry, the vector bundles we will need in
conformal geometry include more than tensor constructs from the tangent
and cotangent bundles. To do this, we introduce the language of conformal
densities. The ray bundle R+ ⊂ S 2 T ∗ M is not a vector bundle and our first
task will be to define a line bundle that transforms like R+ . The line bundle
(1)
will be constructed from the principal bundle FCO(V ) using the procedure
in §9.6.
Let R denote the trivial line bundle on a manifold M with fiber R, so
C ∞ (M ) = Γ(M, R). In the conformal geometry literature (e.g., [56]) the
notation E := Γ(M, R) is often used.
Recall the notation of Equation (11.6). Consider the representation
μ = μ1 : P1,n+1 → GL(1, R) given by h → r2 , where r2 is the (1, 1)-entry
of the size (n + 2) × (n + 2) matrix h. More generally, for τ ∈ R, consider
the representations μτ : P1,n+1 → GL(R1 ) given by h → r2τ . Denote the
(1)
associated line bundle by R[τ ] := (FCO(V ) × R1 )/ ∼, where the equivalence
is via μτ . (Note that the trivial bundle R is R[0].) Elements of R[τ ] are of
the form [(u, t)] with u ∈ FCO(V ) and t ∈ R1 , with [(u, t)] = [(uh, r−2τ t)].
(1)

Let E[τ ] := Γ(M, R[τ ]) denote the space of smooth sections and E[τ ]+ the
space of sections where t > 0.
τ
Remark 11.4.1. In [29] the bundle R[τ ] is denoted D n . For those familiar
with the notation from algebraic geometry, the line bundles R[τ ], when τ
is an integer and M is the sphere, are real analogs of OCPN (τ ) and, for
arbitrary M , of OX (τ ) for projective manifolds X.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.4. Conformal densities and the Laplacian 323

Recall from Remark 9.5.2 that a section of R[τ ] corresponds to a map

(1)
f : FCO(V ) → R1 (i.e., a function) with the equivariance

(11.20) f (uh) = r−2τ f (u),

where h is as above.
Exercise 11.4.2: For which τ does sections of R[τ ] transform like those of
R+ ? 

Differentiate (11.20) to obtain Rh∗ (dfuh ) = −2τ f (u)r−1 dr + r−2τ dfu , and
take h = Idn+2 to see that the derivative of f in the fiber direction is −2τ f ρ.
Therefore
(11.21) df = −2τ f ρ + fi ω i
(1)
for some functions fi : FCO(V ) → R.

Remark 11.4.3. If one prefers to work with sections, let f˜ ∈ E[τ ] denote the
section corresponding to f and let ∇R[τ ] be the connection on R[τ ] induced
(1) (1)
from the connection on FCO(V ) . Let S : M → FCO(V ) be a local section.
Then (11.21) may be written

∇R[τ ] f˜ = S ∗ (fi ω i ).

A section f˜ of any ray bundle R[τ ]+ for any τ = 0 distinguishes a metric

in the conformal class because the semi-basic form
gf := f τ ((ω 1 )2 + · · · + (ω n )2 ) ∈ Γ(FCO(V ) , S 2 T ∗ FCO(V ) )
2 (1) (1)
(11.22)

is basic.
Exercise 11.4.4: Verify that gf is indeed basic. 

The following two identities will be useful when discussing the conformal
Laplacian.
Differentiating (11.21) gives

(11.23) dfi = −τ f βi − 2(τ + 1)fi ρ + fj αij + fij ω j

(1)
for some functions fij : FCO(V ) → R with fij = fji . Differentiating (11.23)
gives
(11.24) 
dfij = −(τ +1)(fi βj +fj βi )+δij ( fk βk )−2(τ +2)fij ρ+fik αjk +fkj αik +fijk ω k
k

for some functions fijk with fijk = fjik .

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 324 11. Conformal Differential Geometry

The conformal Laplacian. Recall the Laplacian on a Riemannian mani-

fold (M, g) from §3.8. It is a second-order differential operator ∞
∞
Δ : C (M )
→ C (M ). In orthonormal frames the operator is σ → i σii , where
∇(dσ) = σij η i ⊗ η j (and σij = σji ). Let’s try to construct an analogous
operator in conformal geometry, where  the operator takes values in some
R[τ ]. A naı̈ve first guess would be f → j fjj , where f ∈ C ∞ (M ) and fjj
is defined by (11.23) with τ = 0. This does not work; in fact, even if we take

f˜ ∈ E[τ ] for τ unspecified, i fii does not in general represent a conformal
density. For, (11.24) implies
 
(11.25) d( fii ) = [(n − 2(τ + 1))(fi βi ) − 2(τ + 2)fii ρ + fiik ω k ],
i i

and when compared with (11.24) this shows that i fii does not transform
like a section of any R[τ  ] except when τ = n2 − 1, in which case τ  = n2 + 1.
In summary:
Proposition/Definition 11.4.5. The conformal Laplacian or Yamabe op-
erator is a well-defined second order differential operator
n n
Δ[g] : E[ − 1] → E[ + 1],
2 2


f˜ → fii ,
i

where fii are defined by (11.23).

The conformal Laplacian in a Riemannian frame. Let f˜ ∈ E[τ ]+ for

some τ = 0. By (11.22) it defines a Riemannian metric gf within the same
conformal class. We compute the conformal Laplacian with reference to this
Riemannian structure. Define the subbundle
(1)
Ff := {u ∈ FCO(V ) | f (u) = 1}.

Note that, on Ff , fi (u) = 0; moreover Ff = FO(V,gf ) is just the usual or-

thonormal frame bundle for the metric gf . Then ρ pulled back to Ff is zero,
and (11.8) pulls back to give the structure equations of Riemannian geom-
etry. The reduction to Ff trivializes all the R[τ ]. In other words, sections
(1)
of R[τ ], viewed as functions on FCO(V ) , when restricted to Ff descend to
functions on M .
(1)
If α is a differential form or function on FCO(V ) , let α denote its pullback
to Ff , except we denote the pullbacks of ω i , αji respectively by η i , ηji to be
consistent with earlier notation. Equation (11.23) implies β i = τ1 f ij η j , and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.4. Conformal densities and the Laplacian 325

thus

Θij = dηji + ηki ∧ ηjk

= −β i ∧ η j + β j ∧ η i + 12 W ijkl η k ∧ η l
1 1
= (δil f jk − δjl f ik − δik f jl + δjk f il )η k ∧ η l + W ijkl η k ∧ η l .
2τ 2
The first term tensored with η i ∧ η j and summed is τ −1 fij η i η j  g. Recalling
that the Schouten tensor (3.12) satisfies R = W + P  g, we see that

(11.26) P = Pf = τ −1 f jk η j η k .

We now calculate how P changes under a conformal change in metric

(1)
from this perspective: let h̃ ∈ E[−τ ]+ , giving rise to a function h on FCO(V )
and a metric gh given by (11.22). Applying (11.21) and (11.23) to h and
restricting h to Ff we obtain

dh = hi η i ,
dhi = hj ηij + (hij + hf ij )η j .

Writing ∇ = ∇gf ,

(11.27) (∇dh)ij = hij + τ hPfij .

Now consider the conformal Laplacian when we restrict to a Riemannian

frame bundle in the conformal class. Take h̃ ∈ E[ n2 − 1], set i = j in (11.27)
and sum to obtain:
n 
(11.28) Δgf h = Δ[g] h − ( − 1)h Pii ,
2
i

i.e., on Ff ,
n−2
(11.29) Δ[g] h̃ = Δgf h + sg h,
4(n − 1) f
where s = sgf is the scalar curvature.
In summary:
Proposition 11.4.6. Fix a metric g ∈ [g]. Relative to the trivializations of
R[ n2 − 1] and R[ n2 + 1], the conformal Laplacian becomes
n−2
Δ[g] = Δg + sg ,
4(n − 1)
where sg denotes the scalar curvature.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 326 11. Conformal Differential Geometry

11.5. Einstein manifolds in a conformal class and the tractor

bundle
Motivated by physics (although in physics one is primarily interested in four-
dimensional manifolds with conformal structure of signature (3, 1)), there
has been considerable research regarding existence of Einstein metrics in
a given conformal class. In this section we return to the EDS for Einstein
(1)
metrics in a given conformal class from the perspective of the bundle FCO(V ) .
This will enable us to advance our goal of uniting the perspectives of [29]
and [56].

The EDS for Einstein manifolds in a conformal class revisited. Fix

(1)
a manifold M with conformal structure [g]. Let Σ := FCO(V ) × R+ × Rn × R
and let the last three factors have coordinates (f, fi , k). Define a Pfaffian
EDS I on Σ generated by the 1-forms
θ : = df − 2f ρ − fj ω j ,
(11.30) θi : = dfi − fj αij − f βi + kω i .
As you will show in the following exercise, integral manifolds of I correspond
to Einstein manifolds in the conformal class [g] just as our earlier system in
§11.1.
Exercise 11.5.1: Verify that an integral manifold of this system yields an
Einstein metric gf given by (11.22).
Exercise 11.5.2: Show that the system is not involutive, and that when
one prolongs and computes, one needs to add the form

θ0 := dk + 2kρ + fj β j
j

to the system. Finally show that the system generated by θ, θi , θ0 is either

Frobenius or has no integral manifolds through a general point.

The global existence question, after considerable work by many authors,

was answered in [118] in all signatures. The answer is that the only cases
where there is more than a single solution that is unique up to changing vol-
ume are the space forms, plus one interesting example of a warped product
of a line with a Ricci flat manifold. The local question is still not completely
understood in dimensions greater than four; see [14, 80] and the references
therein for discussions.

(1)
The tractor bundle. The most natural vector bundle associated to FCO(V )
is the one corresponding to the standard representation of P1,n+1 on Rn+2
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.5. Einstein manifolds in a conformal class and the tractor bundle 327

given by (11.6). This bundle plays an important role in conformal geom-

etry and is called the tractor bundle and denoted T . The tractor bundle
dates back to 1925 [181], and re-appeared in modern form in [10]. Since
P1,n+1 ⊂ O(n+1, 1), the bundle T comes equipped with a Lorentzian metric
h. Since P1,n+1 preserves the flag {e0 } ⊂ {e0 , . . . , en } ⊂ Rn+2 , the tractor
bundle T is equipped with a filtration T1 ⊂ T2 ⊂ T whose associated graded
factors are
T1 = R[−1],
T2 /T1 = T M [−1]
T /T2 = R[1],

where for a vector bundle E, E[j] := E ⊗R[j].

Observe the following:
• We have the bundle isomorphism T1∗⊗(T2 /T1 )  T M because of the
bundle isomorphisms R[τ ]∗  R[−τ ] and R[τ1 ]⊗R[τ2 ]  R[τ1 + τ2 ].
• Let ∇T : Γ(T ) → Γ(T ⊗ T ∗ M ) denote the connection induced
from the coframing of FCO(V ) as in §9.5. Since ∇T is a P1,n+1 -
(1)

connection, it respects the filtration, and since P1,n+1 ⊂ SO(n +

1, 1), ∇T is a metric connection for h ∈ Γ(M, S 2 T ∗ ).
• h|T1 = 0, so h|T2 descends to a well-defined, positive definite metric
on T2 /T1 = T M [−1]. Write g := h|T2 /T1 ∈ Γ(M, S 2 T ∗ M [2]) and
call g the conformal metric. Note that g = gf ⊗f −2 .
∗
• Letting ∇S T M [2] denote the induced connection on S 2 T ∗ M [2], we
2

2 ∗
have ∇S T M [2] g = 0 since ∇T h = 0.
• The Lorentzian metric h allows us to identify T and T ∗ , and the
conformal metric g allows us to identify T M [−1]  T ∗ M [1].
• A choice of g ∈ [g], equivalently a choice of any positive section of
any R[s] with s = 0, determines a splitting of T into its graded
components because there is a reduction of structure group P1,n+2
to SO(n) ⊂ SO(n + 1, 1).
Define a second order differential operator
(11.31) D : E[1] → Γ(T ∗ ),
⎛ ⎞
h
h → ⎝ dh mod T1∗ ⎠ .
Δgf h mod T2∗
This operator is well defined because modulo T2∗ , Δgf h is independent of
the choice of f .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 328 11. Conformal Differential Geometry

Remark 11.5.3. There is a natural projection J 2 (M, R[1]) → T ∗ , where

the 2-jet part is projected to its trace. Then the operator D is just the
projection of j 2 (h) to T ∗ .
Proposition 11.5.4 ([79]). When h = 0, sgh = −h(D(h), D(h)) is the
scalar curvature of the metric g|h| defined by (11.22).
Exercise 11.5.5: Prove Proposition 11.5.4. 
Remark 11.5.6. The expression h(D(h), D(h)) makes sense even when
h = 0. Gover [79] defines triples (M, [g], h) with h(D(h), D(h)) constant to
be almost scalar constant.

The tractor bundle and the Einstein equations. Following the nota-
tion of [79], write a section of T ∗ as I = t (h, hi ω i mod T1∗ , k mod T2∗ ). Then
⎛ ⎞ ⎛ ⎞
h dh − 2hρ − hj ω j
∗
(11.32) ∇T ⎝hi ω i ⎠ = ⎝(dhi − hj αij − hβi + ω i k)⊗ω i mod T1∗ ⎠ .

k dk + 2kρ − j hj βj mod T2∗
First note that I is covariant constant if and only if I = D(h) for some
h ∈ E[1]. Comparing the EDS of Exercise 11.5.2 and (11.32), we see that a
section of T ∗ that is covariant constant is nearly the same thing as a choice
of an Einstein metric in a conformal class! The only thing that could go
wrong is if h = 0 at points. (If h < 0, simply use −h instead of h.) R.
Gover [78, 79] was the first to observe this and recognize its utility. Let
∗
I denote a section of T ∗ . Call a manifold (M, g, I) with ∇T I = 0 almost
Einstein. Note that if h is not identically zero, then we obtain an honest
Einstein metric on an open subset of M . We may use sgh to get more precise
information about the locus where the metric gh is well-defined.
Theorem 11.5.7 ([79]). Let (M, g, I = D(h)) be almost Einstein and let
Zh ⊂ M denote the zero set of h.
(1) If the constant sh is positive, then Zh = ∅ and (M, gh ) is an Einstein
manifold with positive scalar curvature.
(2) If sh = 0, then either Zh = ∅ and (M, gh ) is a Ricci flat Einstein
manifold, or Zh consists of isolated points and (M \Zh , gh ) is a Ricci
flat Einstein manifold.
(3) If sh < 0, then Zh is either empty or a smooth hypersurface and
(M \Zh , gh ) is an Einstein manifold with negative scalar curvature.

Proof. Fix a background metric gf so σ = h/f is a function on M . Then

Equation (11.3) implies
σ n
sgσ = σg2f + tracegf (∇gf dσ) − |dσ|2gf .
2(n − 1) 2(n − 1)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 11.5. Einstein manifolds in a conformal class and the tractor bundle 329

So σ(x) = 0 (i.e., h(x) = 0) implies sgσ ≤ 0, which proves the first case.
For the second case, if σ(x) = 0, we see dσx = 0 as well. Since a parallel
section of a vector bundle is nonvanishing, ∇gf (dσ)x must be nonzero to
have lx nonzero. This implies x is an isolated zero of σ, which is what we
were trying to show.
For the third case, since |dσx | = 0, the inverse function theorem guar-
antees the zero locus is smooth in the neighborhood of any x ∈ Zh , that is,
smooth everywhere. 

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/12

Chapter 12

Projective Geometry
II: Moving Frames and
Subvarieties of
Projective Space

We continue our study of submanifolds of projective space and algebraic

geometry.
We begin in §12.1 by defining the Fubini cubic form and higher-order
differential invariants. We compute these invariants for the most ubiqui-
tous homogeneous varieties in §12.2. A basic property of projective space is
the notion of lines, and in §12.3 we study the local differential geometry of
submanifolds containing lines through all its points. Dual varieties, secant
varieties and tangential varieties were introduced in Chapter 4. We return
to their study in §12.4 and §12.5. We prove the bounds of Zak and Landman
on the dual defect from a differential-geometric perspective, describe how
to calculate dim σ(X) and dim τ (X) infinitesimally, and state the Fulton-
Hansen Theorem relating tangential and secant varieties. Among homoge-
neous varieties, there is a preferred class, the cominuscule varieties, that is
particularly easy to study - we examine their differential invariants in §12.6.
We continue our study of differential invariants in §12.7, in particular making
a detailed study of lines on varieties. In §12.8, we prove the classical Fubini
theorem characterizing quadric hypersurfaces. We briefly discuss general-
izations of dual varieties introduced by Gelfand, Kapranov and Zelevinsky
[73] in §12.9. We return to the study of varieties with degenerate Gauss

331
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 332 12. Projective geometry II

maps in §12.10. In §12.11, we show that the projective second fundamen-

tal form has certain genericity properties in small codimension if X is not
too singular. We give a very detailed study of the local differential geom-
etry of varieties with degenerate secant and tangential varieties in §12.12,
stating Zak’s theorem classifying Severi varieties, the smooth varieties of
minimal codimension having secant defects, and outlining a proof of Zak’s
theorem. In §12.13, we describe results generalizing the classical Monge
equation characterizing plane conics and Fubini’s theorem. We conclude
with a differential-geometric study of complete intersections in §12.14.

12.1. The Fubini cubic and higher order differential

invariants
In this section we determine a complete set of differential invariants for deter-
mining a submanifold of projective space M ⊂ PV up to GL(V ) equivalence.
There are two types of such invariants. The first are the higher fundamen-
tal forms, which measure how M leaves its higher osculating spaces to first
order. These differential invariants are defined as sections of natural vector
bundles over M . The second type are relative differential invariants which
we call Fubini forms, named after Fubini, who first used the cubic form F3 to
characterize quadric hypersurfaces (see Theorem 12.8.1). The Fubini forms
are defined as sections of bundles over a frame bundle of M . For example,
the Fubini cubic form, a third-order relative differential invariant, measures
how M is infinitesimally leaving its embedded tangent space at a point to
second order.

Throughout this section, M ⊂ PV is a complex submanifold of dimen-

sion n and FM
1 is its first-order adapted frame bundle as in §4.1. We continue

the index ranges 1 ≤ α, β ≤ n and n + 1 ≤ μ, ν ≤ n + a.

The Fubini cubic form. Unlike Euclidean geometry, the projective funda-
mental forms do not give enough information to determine local equivalence
of complex submanifolds up to projective transformations (e.g., the case of
hypersurfaces discussed in Remark 4.1.9). Thus we differentiate further to
find higher-order differential invariants:
μ
0 = d(ωαμ − qαβ ω0β )
μ μ μ δ μ δ
= (−dqαβ − qαβ ω00 − qαβ
ν
ωνμ + qαδ ωβ + qβδ ωα ) ∧ ω0β .
μ
The Cartan Lemma implies there exist functions rαβγ , defined on F 1 , sym-
metric in their lower indices, satisfying
μ
(12.1) rαβγ ω0γ = −dqαβ
μ μ
− qαβ ω00 − qαβ
ν μ δ
ωνμ + qαδ μ δ
ωβ + qβδ ωα .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.1. The Fubini cubic and higher order differential invariants 333

For f ∈ F 1 , define F3 = (F3 )f ∈ π ∗ (S 3 T ∗ M ⊗N M ) to be

μ
(12.2) F3 = rαβγ ω0α ω0β ω0γ ⊗eμ .

Moving in the fiber by the block diagonal matrices g00 , gβα , gνμ does not
change F3 . However, F3 does not descend to be a well-defined section of
S 3 T ∗ M ⊗N M , because if (ẽ0 , ẽα , ẽμ ) is a new frame with
ẽμ = eμ + gμ0 e0 + gμα eα ,
(12.3) ẽα = eα + gα0 e0 ,
then
μ μ μ ν μ
(12.4) r̃αβγ = rαβγ + Sαβγ gα0 qβγ + Sαβγ gνδ qαβ qγδ ,
where Sαβγ is cyclic summation in the fixed indices α, β, γ.
As mentioned above, F3 ∈ Γ(F 1 , π ∗ (S 3 T ∗ M ⊗ N M )) is an example of
μ
a relative differential invariant. Adopt the notation Δrαβγ to denote the
μ
change in rαβγ by a fiber motion of the type in (12.3). By (12.4),
μ μ ν μ
Δrαβγ = Sαβγ (gα0 qβγ + gνδ qαβ qγδ ).

Let X n ⊂ Pn+1 be a hypersurface and let x ∈ Xsmooth . Then

Baseloc |IIX,x | defines a quadric hypersurface in PTx X. The space of cubic
polynomials |F3 | is not well-defined, but the ideal generated by |II|, |F3 | is
well-defined; thus there is a well-defined subvariety in PTx X that is the in-
tersection of a quadric hypersurface and a cubic hypersurface. We will see
in §12.8 that if this intersection is simply the quadric hypersurface itself at
a general point and the quadric has rank greater than one, then X must be
a quadric hypersurface.
Remark 12.1.1. One can compare F3 to the covariant derivative ∇II of
the second fundamental form of a submanifold Y of a Riemannian manifold
Z, which is a well-defined section of S 3 T ∗ Y ⊗N Y . If one works over R, the
forms s∗ (F3 ), as s : M → F 1 varies over local sections, is the set of covariant
derivatives of Riemannian metrics on Z = PV compatible with the linear
structure.
Remark 12.1.2. If one uses local coordinates, then the third-order terms
in a Taylor series expansion of X provide the coefficients of the cubic form;
see §12.7.

The third fundamental form. If II : S 2 Tx M → Nx M is not surjective,

then one can refine the flag x̂ ⊂ T̂x M ⊂ V to obtain a refined invariant
(2)
from F3 . Let T̂x M = T̂x M + II(S0 2 Tx M ). Here we regard II(S
0 2 Tx M ) ⊂
V /T̂x M , so the sum of the two spaces determines a well-defined subspace of
V.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 334 12. Projective geometry II

Let F 2 → M be the bundle of frames adapted to the flag

x̂ ⊂ T̂x M ⊂ T̂x(2) M ⊂ V,
where we restrict to an open subset of M where the dimensions in the flag
(2)
are constant. Use indices n + 1 ≤ ξ ≤ dim T̂x M − 1 =: n + a1 − 1,
(2)
n + a1 = dim T̂x M ≤ φ ≤ n + a = dim PV . Thus, frames f ∈ F 2 are of
the form f = (e0 , eα , eξ , eφ ) with T̂ (2) = {e0 , eα , eξ }.
Exercises 12.1.3:
1. Show that ωαφ = 0 when pulled back to F 2 .
2. Show that the forms ωξφ are semi-basic when pulled back to F 2 .
φ
3. Show that there exist functions qαβγ , symmetric in their lower indices,
such that
(12.5) ωξφ ωαξ = qαβγ
φ
ω0β ω0γ .
(2)
Definition 12.1.4. Let N2,x M = Tx PV /(Tx (2) M ) and eφ = (eφ mod T̂x M )
⊗e0 . The tensor ωξφ ωαξ ω0α ⊗ eφ ∈ Γ(F 2 , π ∗ (S 3 T ∗ M ⊗ N2 M ) descends to a
well-defined tensor III : S 3 Tx M → N2,x M , called the projective third fun-
damental form. The third fundamental form measures how M is moving
away from its second osculating space to first order at x.
Exercises 12.1.5:
1. Show that the third fundamental form can also be defined at general
points as the derivative of a second-order Gauss map
γ (2) : M → G(n + a1 + 1, V ),
x → T̂x(2) M.
(In general, γ (2) is only defined on a full-dimensional open subset of M .)
2. Give an interpretation of III similar to the “quadratic part of the sin-
gularity of X ∩ H at x” interpretation of II. When is the cubic part of a
singularity well-defined?

At general points, there is a severe restriction on the third fundamental

form:
Theorem 12.1.6 (Cartan [37, p. 377]). Let M n ⊂ PV be an analytic sub-
manifold. Then at general points
∗
(12.6) III(N2,x M ) ⊂ S 3 Tx∗ M ∩ (Tx∗ M ⊗II(Nx∗ M )).
That is, the third fundamental form satisfies a prolongation property (see
§5.1).
Corollary 12.1.7. If X n ⊂ PV is a variety, x ∈ X is a general point,
dim |IIX,x | = 1 and rank II(Nx∗ X) > 1, then X n is a hypersurface in some
Pn+1 .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.1. The Fubini cubic and higher order differential invariants 335

Exercises 12.1.8:
1. Prove Theorem 12.1.6. 
2. Show that the prolongation property is equivalent to the quadratic form
v P being in II(Nx∗ M ) for every v ∈ Tx M and every P ∈ III(N2,x
∗ M ).

3. Prove Corollary 12.1.7. 

The k-th fundamental form FFk . The exercises below give two definitions
of the k-th fundamental form FFkx,M ∈ S k Tx∗ M ⊗N(k−1),x M and the k-th
osculating space

T̂ (k) (k−1) 1k (S k T M ).
M + FF
x M = T̂ x x

Below the exercises we give yet another definition of the k-th fundamental
(k)
form. Let Nk,x M = Tx PV /T x M .
Exercises 12.1.9:
1. Define FFk by defining a higher-order Gauss map and taking its deriva-
tive.
2. Define FFk by refining the flag and differentiating, as we did with III =
FF3 .
3. Show the two procedures above give the same differential invariant,
called the k-th fundamental form, which measures to first order how M
is leaving its (k − 1)-st osculating space.
∗
Notation. Let |FFk |x,M = FFk (N(k−1),x M ) ⊂ S k Tx∗ M .
Exercise 12.1.10: Show that at general points x ∈ M , |FFk |M,x ⊆ (Tx∗ M ⊗
|FFk−1 |x,M ) ∩ S k Tx∗ M , i.e., that the prolongation property persists to higher
fundamental forms.

Another definition of fundamental forms. Here is a method to define

and calculate fundamental forms communicated to us by M. Green. It gener-
alizes the observation that de0 ≡ ω0α eα mod e0 and deα ≡ ωαμ eμ mod{e0 , eα }
imply II[e0 ] = ωαμ ω0α ⊗eμ .
Define a series of maps
dk e0 : (T F 1 )⊗k → V / Image(d0 e0 , . . . , dk−1 e0 )
as follows: Let d denote exterior differentiation, let d0 e0 = e0 and let d1 e0 =
de0 mod e0 . If v1 , . . . , vk ∈ Tf F 1 , extend v1 , . . . , vk to holomorphic vector
fields ṽ1 , . . . , ṽk in some neighborhood of f . Let
(12.7) dk e0 (v1 , . . . , vk ) := v1 d(ṽ2 . . . d(ṽk de0 )) mod πk −1 (Image dk−1 e0 )
where πk : V → V /(Image{d0 e0 , . . . , dk−1 e0 }) is the projection, and denotes
the contraction T × T ∗⊗l → T ∗⊗l−1 . The proof that (12.7) is independent of
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 336 12. Projective geometry II

the choice of extension to vector fields is standard (see, e.g., [174]). Then
dk e0 ⊗e0 = FFk .
For example, keeping the index ranges on F 2 that we used above to
study the third fundamental form, we have
d1 e0 ≡ ω0α eα mod{e0 },
d2 e0 = ω0α deα mod{e0 , eα }
= ω0α ωαξ eξ mod{e0 , eα },
d3 e0 = ω0α ωαξ deξ mod{e0 , eα , eξ }
= ω0α ωαξ ωξφ eφ mod{e0 , eα , eξ }.

Algebraic definition of fundamental forms via spectral sequences. For those

familiar with spectral sequences, here is yet another definition of the funda-
mental forms due to M. Green, [81]. This approach also works to define the
relative differential invariants; see [81]. The maps dk can be defined more
algebraically as follows:
The quotient map
V ∗ → V ∗ /x̂⊥ = OPV (1)x
gives rise to a spectral sequence of a filtered complex by letting
F 0K 0 = V ∗, F 0 K 1 = OX (1)x ,
F 1 K 0 = 0, F p K 1 = mpx (1),
where mx denotes the maximal ideal of functions vanishing at x, and mpx
denotes its p-th power (the functions vanishing to order p − 1 at x). We let
F p = F p K 1 for p ≥ 0 and T ∗ = Tx∗ X, etc. The maps are
d0 : V ∗ → F 0 /F 1 = OX,x (1)/mx (1)  C,
d1 : ker d0 → F 1 /F 2 = mx (1)/m2x (1)  T ∗ (1),
d2 : ker d1 → F 2 /F 3 = m2x (1)/m3x (1)  (S 2 T ∗ )(1),
..
.

12.2. Fundamental forms of Veronese, Grassmann, and

Segre varieties
Reduced frame bundles. Given X n ⊂ Pn+a , we may calculate our differential
invariants by taking any section s : X → F 1 and pulling back the relevant
forms. In particular, if there is a natural subbundle of F 1 to work on,
we may restrict our frames to the subbundle without losing any geometric
information.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.2. Fundamental forms of Veronese, Grassmann, and Segre varieties 337

Recall that we may identify GL(V )  FPV as sets. If X = G/P ⊂ PV is

homogeneous for some group G ⊂ GL(V ), we may further identify G with
a subbundle of FPV and we will often work with this subbundle.
In the next two examples G = GL(W ) and V is respectively S p W and
Λk W , and thus bases for V are obtained from those of W . In particular, if
(e0 , . . . , en ) is a basis of W , we apply our operator d repeatedly to (e0 )p :=
e0 · · · e0 and e0 ∧ e1 ∧ · · · ∧ ek−1 .
Fundamental forms of Veronese varieties. Let W = Cn+1 , let V = S p W ,
and denote bases of W by (e0 , eα ), 1 ≤ α ≤ n, so the entries of the
Maurer-Cartan form of GL(W ) are ω00 , ωβ0 , ω0α , ωβα . Let x = [(e0 )p ]. Let
ρp : GL(W ) → GL(S p W ) denote the natural inclusion. We work with the
bundle of ρp (GL(W )) ⊂ GL(S p W )-frames. The Leibniz rule d(eA eB ) =
(deA )eB + eA (deB ) implies
dep0 ≡ pω0α eα e0p−1 ,
d2 ep0 ≡ p(p − 1)ω0α ω0β eα eβ e0p−2 ,
dk ep0 ≡ p(p − 1) · · · (p − k + 1)ω0α1 · · · ω0αk eα1 · · · eαk e0p−k .
Thus
|FFkvp (PW ),[ep ] | = S k T[e∗ p ] vp (PW ), k ≤ p,
0 0

FFkvp (PW ) = 0, k > p.

Fundamental forms of Grassmannians. Let W = Cn and let V = Λk W .
Consider the inclusion GL(W ) ⊂ GL(V ) with the induced action
g(v1 ∧ . . . ∧ vk ) = (gv1 ) ∧ . . . ∧ (gvk ).
Use index ranges 1 ≤ i, j ≤ k and k + 1 ≤ s, t ≤ n.
Let E 2 W ), let E
 = e1 ∧ . . . ∧ ek ∈ G(k,  sj = (−1)j−1 e1 ∧ . . . ∧ ej−1 ∧ es ∧
ej+1 ∧ . . . ∧ ek and E  ij be E
 with es replacing ei and et replacing ej , with
st
appropriate signs, and so on. Then the E  sj11 ...s
...jp
p , with j1 < j2 < . . . < jp ,
s1 < s2 < . . . < sp and 1 ≤ p ≤ k, are a basis of Λk W .
Using the Maurer-Cartan forms for the GL(W )-frame bundle, we take
derivatives:
 ≡ ω s E i mod E,
dE 
i s
   
dEsi ≡  ij +
ωjt E ω t  ij
E − ω t  ij
E −  ij
ωjt E
st j st j st st
i<j,s<t i>j,s>t i<j,s>t i>j,s<t

mod T̂E G(k, W ).

Thus T̂E G(k, W ) is spanned by the vectors E, E  i and we recover the in-
s
∗
terpretation TE G(k, W )  E ⊗ (W/E). Similarly, II(S 2 TE G(k, W )) ⊂
 ij mod T̂E G(k, W ), and we may thus
NE G(k, W ) is spanned by the vectors E st
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 338 12. Projective geometry II

identify II(S 2 TE G(k, W )) with Λ2 E ∗ ⊗ Λ2 (W/E) as GL(E) × GL(W/E)-

modules.
It is natural (and correct!) to guess that
II : NE∗ G(k, W ) → S 2 TE∗ G(k, W )
is GL(W )-equivariant (see [136] for a proof and representation-theoretic
interpretation of the fundamental forms of rational homogeneous varieties).
Thus we expect that
Baseloc |IIG(k,W ),E | = Seg(PE ∗ × P(W/E)) ⊂ P(E ∗ ⊗W/E),
because the Segre variety is the zero set of the 2 × 2 minors, i.e., of Λ2 E ⊗
Λ2 (W/E)∗ ⊂ S 2 (E ∗ ⊗V /E)∗ .
We calculate that

IIG(k,W ),E = (ωis ωjt − ωit ωjs )⊗E ij
st ,
i<j,s<t

proving that the quadrics in |II| are spanned by the 2 × 2 minors of the
matrices in TE G(k, W ).
Exercise 12.2.1: Determine FFlG(k,W ) and its base locus. 

Since the Grassmannian is cut out by polynomials of degree two, the set
Baseloc |IIG(k,W ),E | consists of tangent directions to lines on G(k, W ) that
pass through E. To see these lines explicitly, a point [x ⊗ y] ∈
Baseloc |IIG(k,W ),E | ⊂ P(TE G(k, W )) determines a (k − 1)-plane Fx ⊂ E ⊂
V and a (k+1)-plane Ly ⊂ V containing E, where Fx = ker x and Ly = E+y.
The corresponding line in G(k, W ) is {E  | Fx ⊂ E  ⊂ Ly }.
Fundamental forms of Segre varieties. Let U = Ca+1 , W = Cb+1 , and con-
sider the Segre embedding Seg(PU × PW ) ⊂ P(U ⊗ W ). Here we use
G = GL(U )×GL(W ) ⊂ GL(U⊗W )-frames. Let (e0 , . . . , ea ) and (f0 , . . . , fb )
denote bases of U, W , and let the respective Maurer-Cartan forms be de-
noted (ω) and (η). Use index ranges 1 ≤ α ≤ a, 1 ≤ j ≤ b. Let
[e0 ⊗ f0 ] ∈ Seg(PU × PW ). Note that e0 ⊗ f0 , eα ⊗ f0 , e0 ⊗ fj , eα ⊗ fj form
a basis of U ⊗W . We compute
(12.8) d(e0 ⊗f0 ) ≡ ω0α eα ⊗f0 + η0j e0 ⊗fj mod e0 ⊗f0 ,
(12.9) d2 (e0 ⊗f0 ) ≡ ω0α η0j eα ⊗fj mod T̂[e0⊗f0 ] Seg(PU × PW ).
Equation (12.8) shows that
T̂[e0⊗f0 ] Seg(PU × PW ) = e0 ⊗W + U ⊗f0 .
Let U  = (U⊗f0 mod e0⊗f0 )⊗(e0⊗f 0 ) and W  = (e0⊗W mod e0⊗f0 )⊗(e0⊗f 0 ),
so
T[e0⊗f0 ] Seg(PU × PW ) = U  ⊕ W  .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.3. Ruled and uniruled varieties 339

Equation (12.9) shows that

∗ ∗
|II|Seg(PU ×PW ),[e0⊗f0 ] = U  ⊗W  ⊂ S 2 T[e∗ 0⊗f0 ] Seg(PU × PW )
and that IIISeg(PU ×PW ) is zero. In particular,
Baseloc |II|Seg(PU ×PW ),[e0⊗f0 ] = PU  ( PW  .
Note that here again, II is equivariant: S 2 T ∗ = S 2 (U  ⊕ W  )∗ = S 2 U  ∗ ⊕
(U  ∗ ⊗W  ∗ ) ⊕ S 2 W  ∗ and II : N ∗ → S 2 T ∗ is just the identity map on the
middle factor.
Exercises 12.2.2:
1. Consider X = Seg(Pr1 ×. . .×Prd ) = Seg(PW1 ×. . .×PWd ), dim Wi = ri .
Use the GL(W1 ) × . . . × GL(Wd ) ⊂ GL(W1 ⊗· · ·⊗Wd ) frame bundle in the
following computations:
(a) Determine the tangent space and second fundamental form at a point
of X. Interpret your results.
(b) Compute FFk for all k. In particular, show that the last nonzero fun-
damental form is FFd .
2. Let W be equipped with a nondegenerate quadratic form Q. Consider
GQ (k, W ) ⊂ G(k, W ) ⊂ P(Λk W ), the Grassmannian of k-planes isotropic
for Q. Show that TE GQ (k, W )  E ∗ ⊗(E ⊥ /E) ⊕ Λ2 (V /E ⊥ ), and use this to
calculate dim GQ (k, W ).
3. Let W be equipped with a nondegenerate symplectic form ω. Consider
Gω (k, W ), and calculate TE Gω (k, W ) and dim Gω (k, W ). 

Fundamental forms of Segre products of varieties.

Proposition 12.2.3. Let Xj ⊂ PWj , 1 ≤ j ≤ r, be varieties, let xj ∈ Xj ,
and let y = [x1 ⊗. . .⊗xr ] ∈ Y := Seg(X1 × . . . × Xr ) ⊂ P(W1 ⊗· · · ⊗Wr ).
Then
T̂y Y = Σj x1 ⊗. . .⊗xj−1 ⊗ T̂xj Xj ⊗xj+1 ⊗. . .⊗xr ,
IIY,y = Σj<k x1 ⊗. . .⊗xj−1 ⊗Tx∗j Xj ⊗xj+1 . . .⊗xk−1 ⊗Tx∗k Xk ⊗xk+1 ⊗. . .⊗xr
+ Σj x1 ⊗. . .⊗xj−1 ⊗IIXj ,xj ⊗xj+1 ⊗. . .⊗xr .
Exercise 12.2.4: Prove the proposition.
Exercise 12.2.5: Let Mr ⊂ P(Mm×n ) denote the (projectivization of the)
space of m × n matrices of rank at most r. Calculate T̂x Mr , where x has
rank exactly r. Verify that the smooth points of Mr consist of the matrices
of rank exactly r. 

12.3. Ruled and uniruled varieties

We determine conditions on the differential invariants of M ⊂ PV that imply
it contains linear spaces through a general point.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 340 12. Projective geometry II

Definition 12.3.1. Let M ⊂ PV be a submanifold or subvariety.

M is ruled by k-planes if it can be written as a fibration M → B where
each fiber is (an open subset of) a k-plane.
M is uniruled by k-planes if for all x ∈ M there exists at least one (open
subset of a) k-plane L such that x ∈ L ⊂ M . We also sometimes say M is
Pk -uniruled.
M is covered by k-planes if M is Pk -uniruled with just a finite number
of k-planes passing through a general point.

Example 12.3.2. If n ≥ 4, the smooth quadric hypersurface Qn ⊂ CPn+1

is uniruled but not ruled by  n2 -planes.

Exercise 12.3.3: Show that a generic hypersurface X n ⊂ Pn+1 of degree n

is covered by lines, with n! lines passing through a general point.

We address the following question:

How many derivatives must one take to see if a submanifold is uniruled?
We saw in Chapter 2 that on any surface with K < 0 in E3 there are two
lines osculating to order two at each point, so to see if a surface is ruled, we
need to take at least three derivatives.
Theorem 12.3.4 (Darboux). Let M 2 ⊂ P2+a (or affine space) be a smooth
(resp. analytic) surface such that at each (resp. a general) x ∈ M there is
a line lx osculating to order three at x. Then M is ruled.

Proof. We adapt frames such that e1 is tangent to the line osculating to

order three. Consider the integral curves C of the line field {e1 } (the anni-
hilator of the system {ω02 = 0}). The second fundamental form of C is

IIC,[e0 ] = ω12 ω01 ⊗(e2 mod T̂[e0 ] C)⊗e0 + ω1μ ω01 ⊗(eμ mod T̂[e0 ] C)⊗e0 .

By hypothesis ω1μ ≡ 0 mod{ω02 }, and we need to calculate ω12 . Formula

(12.2) for F3 shows
μ μ μ 2
r111 ω01 + r112 ω02 = 2q12 ω1 .
If II(e1 , e2 ) = 0 we are in the case of a degenerate Gauss map, and the
result follows by Theorem 4.3.2. Otherwise there exists a μ, say μ = 3,
where q123 = 0. We use our third-order hypothesis, which implies r 3
111 = 0,
to conclude that ω12 ≡ 0 mod{ω02 }. Thus IIC,[e0 ] = 0, so C is a line. 

Remark 12.3.5. More generally, if M n ⊂ PV is an analytic submanifold,

then a line osculating to order n + 1 at a general point must be contained
in the completion of M ; see [128].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.4. Dual varieties 341

12.4. Dual varieties

We now study varieties X ⊂ PV with degenerate dual varieties X ∨ ⊂ PV ∗ .
We begin by finding a frame bundle simultaneously adapted to the first-
order geometry on X and X ∨ . We then prove the basic theorems about
dual varieties: reflexivity, the Landman parity theorem and Zak’s bound on
the dual defect.

Frames for X and X ∨ . We use the incidence correspondence (4.3) to

find a good frame bundle. Let F ∗  → I 0 := {(x, H) | x ∈ Xsmooth , H ∈
X ∨ smooth , T̃x X ⊆ H} be the frame bundle of bases of V whose fiber over
(x, H) ∈ I 0 consists of bases adapted to the flag
0 ⊂ x̂ ⊂ T̂x X ⊂ Ĥ ⊂ V,
that is, bases (e0 , . . . , en+a ) such that
{e0 } = x̂,
{e0 , . . . , en } = T̂x X,
{e0 , . . . , en+a−1 } = Ĥ.

Let π  : F ∗  → PV and ρ : F ∗  → PV ∗ denote the projections, so

π  (e 
0 , . . . , en+a ) = [e0 ] and ρ (e0 , . . . , en+a ) = [e
n+a ], where (e0 , . . . , en+a ) is

the basis dual to (e0 , . . . , en+a ).

Then
X = π  (F ∗  ), X ∨ = ρ (F ∗  ),
and thus if f ∈ F ∗  is a general point, dim(X ∨ ) = rank ρ∗f .
To differentiate vectors in the dual frame (e0 , . . . , en+a ), we need to
determine how the Maurer-Cartan form behaves with respect to dual bases.
Let ,  denote the pairing V × V ∗ → C, with eA , eB  = δA B . We calculate

0 = deA , eB  = ωA
C
eC , eB  + eA , deB ,
which implies
deB , eA  = −ωA
B
,
i.e.,
deB = −ωC
B C
e .
Since ρ is submersive, we may calculate dim X ∨ by determining the
number of independent forms in den+a :
(12.10) den+a ≡ −ωαn+a eα − ωn+λ
n+a n+λ
e mod en+a .
Here we use the index range 1 ≤ λ, κ ≤ a − 1. The forms ωn+λ n+a
are inde-

pendent and independent of the semi-basic forms for π , and dim{ωαn+a } =
rank q n+a . We obtain an explicit description of the tangent space:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 342 12. Projective geometry II

Theorem 12.4.1 ([103]). Let X n ⊂ Pn+a be a variety, let x ∈ X be a

general point and let H ∈ X ∨ correspond to a general point H  ∈ N ∗ X.
x
Then
II(H))
T̂H X ∨ = (x̂ + Singloc  ⊥,
II(H)
where Singloc  is considered as a subspace of T̂x X/x̂ (i.e., we ignore
the twist by a line bundle), so the sum makes sense as a subspace of V .
Corollary 12.4.2 ([85]). Let X n ⊂ Pn+a be a variety, let x ∈ X be a
general point and let r be the rank of a generic quadric in |IIX,x |. Then
dim X ∨ = r + a − 1.
Exercise 12.4.3: Calculate dim X ∨ for the following varieties:
(a) X is a curve in Pa+1 not contained in a hyperplane.
(b) X = vd (PV ).
(c) X = Seg(Pa × Pb ).
(d) X = G(2, m).
(e) X is a smooth surface in Pa+2 , not contained in a hyperplane.

Which of these are self-dual, i.e., isomorphic to their dual varieties? 

Corollary 12.4.4 ([103]). Let X n ⊂ Pn+a be a variety, with dual variety
X ∨ of dimension r + a − 1, let x ∈ X be a general point and let H
 ∈ N ∗X
x
correspond to H ∈ X ∨ . If rank II(H)
 < r, then H ∈ X ∨ sing .

 < r, then T̂[en+a ] X ∨ has

Proof. Equation (12.10) shows that if rank II(H)
the wrong dimension. 
∗ X ∨ ; thus, if x ∈ X
Observe that Theorem 12.4.1 implies [e0 ] ∈ PNH smooth
∨
and H ∈ X smooth , then
T̃x X ⊆ H ⇔ T̃H X ∨ ⊆ x.
Since I 0 surjects onto Zariski open subsets of X and X ∨ , by taking closures
we conclude:
Theorem 12.4.5 (Reflexivity1 ). (X ∨ )∗ = X.
Thus, the dual variety should be viewed as a transform of X, containing
the same information as X, only reorganized. We will see (in Proposition
12.4.17 below) that in this re-organization, certain global information be-
comes local.
Definition 12.4.6. Given X n ⊂ Pn+a , define δ∗ = δ∗ (X) = n + a − 1 −
dim X ∨ , the dual defect of X.
∗ X ∨ ⊂ X, and thus:
The reflexivity theorem implies that PNH
1 The reflexivity theorem has a long history, dating back to C. Segre; see [111] for details.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.4. Dual varieties 343

Proposition 12.4.7. Let X n ⊂ Pn+a be a variety, let x ∈ X be a general

point, and let H ∈ X ∨ be a smooth point with T̃x X ⊆ H. Then

π(ρ−1 (H)) = {y ∈ X | T̃y X ⊆ H}

∗ X ∨.
is a linear Pδ∗ ⊂ X and may be identified with PNH
Reversing the roles of X and X ∨ in Corollary 12.4.4, we obtain:
Theorem 12.4.8. Let X n ⊂ Pn+a be a smooth variety, let H ∈ X ∨
be a general point, let v ∈ NH ∗ X ∨ be any nonzero vector, and let r =
∨
rank IIX ∨ ,H (v). Then dim(X ) = r + a − 1. In particular, |IIX ∨ ,H | is a
system of quadrics of projective dimension equal to the dual defect δ∗ , and
of constant rank n − δ∗ .
The first assertion of Theorem 12.4.8 is due to B. Segre [168]. It was
rediscovered by Griffiths and Harris in [85]. The second appeared in [103].
Corollary 12.4.9. Let X n ⊂ Pn+a be a smooth variety, and let H ∈ X ∨
be a general point. Then IIIX ∨ ,H = 0.
It is difficult to have a system of quadrics of constant rank. By a classical
result in linear algebra, it is impossible if r is odd, and, by [103], impossible
if the dimension of the system is one greater than the dimension of the vector
space minus the rank. This unifies and recovers two important results:
Theorem 12.4.10 (Landman parity theorem [57]). Let X n ⊂ Pn+a be a
smooth variety. Then n − δ∗ is even.
Theorem 12.4.11 (Zak [191]). Let X n ⊂ Pn+a be a smooth variety not
contained in a hyperplane. Then δ∗ < a − 1, i.e., dim X ∨ ≥ dim X.
One consequence of Zak’s theorem is the following:
Theorem 12.4.12 ([122]). Let X n ⊂ Pn+a be a smooth variety, let x ∈
X be a general point and let q ∈ |IIX,x | be a generic quadric. Then
dim(Singloc(q)) ≤ a − 1.
We will generalize Theorem 12.4.12 in §12.11.

Detour: Systems of quadrics of constant rank. Let V = Cm and let

k > 1. What are the k-dimensional linear subspaces of S 2 V of constant rank
r?
We saw above that r is necessarily even. When k = 2 there is a complete
answer:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 344 12. Projective geometry II

Normal form for pencils of quadrics of constant rank. Let

⎛ ⎞
s 0 0 ... 0
⎜t s 0 ... 0 ⎟
⎜ ⎟
⎜0 t s 0 . . .⎟
⎜ ⎟
Ap [s, t] = ⎜ . ⎟
⎜ .. ... ... ⎟
⎜ ⎟
⎝0 . . . 0 t s⎠
0 ... 0 t
p
be a space of 2 × ( p2 + 1) matrices. For all [s, t] ∈ P1 , rank Ap [s, t] = p2 . Let
 
0 Ap [s, t]
Qp [s, t] = t ,
Ap [s, t] 0
a (p+1)×(p+1) pencil, i.e., a two-dimensional linear subspace of quadrics, of
constant rank p. Any constant rank p pencil of quadrics on Cp+1 can be nor-
malized to be of this form; see, e.g., [132, §10.3]. Note that Singloc Qp [s, t]
p
is a rational normal curve of degree p2 in the P 2 spanned by the first p2 + 1
p
directions, and its linear span distinguishes a C 2 +1 ⊂ Cp+1 . Thanks to the
p
SL2 -action, there is a complementary C 2 ⊂ Cp+1 that is also distinguished.
Any constant rank r pencil of quadratic forms on Cm is of the form
⎛ ⎞
Qi1 [s, t]
⎜ .. ⎟
⎜ . ⎟
Q[s, t] = ⎜ ⎟,
⎝ Qil [s, t] ⎠
0
where if the last zero block is f ∗ ×f ∗ , we have i1 +. . .+il = r, l+f ∗ = m−r.
In our applications we will have m = r + a − 1.
The classification problem can be studied systematically via vector bun-
dles on projective space; see [103, 129].
Remark 12.4.13. For those familiar with the language of vector bundles,
a system of quadrics A ⊂ PS 2 V ∗ of constant rank r determines a rank r
vector bundle E → PA whose fiber at [q] ∈ PA is the image of the linear map
q : V → V ∗ . Muñoz [152] observed that an argument of Ein in [58] (used in
the proof of his linear fibration theorem) shows that if rank E > 1 and E is
the direct sum of line bundles, then any X with |IIX |  A at general points
is ruled by linear spaces. Sato [167] showed that if E is a uniform vector
bundle and rank E < a, then E must be the direct sum of line bundles.
Using Sato’s result, Ein [58] proved that if δ∗ ≥ n2 , then X n ⊂ Pn+a must
be P(n+δ∗ )/2 -ruled.
Exercises 12.4.14:
1. Given a linear space B of p × q matrices of constant rank s, construct a
system of quadrics of constant rank r = 2s on Cm with m = p + q. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.4. Dual varieties 345

2. Determine |IIG(2,5) |. 
3. One gets the second fundamental form of what variety when one uses
Exercise 1 in the case B = M2×q ? 

Calculation of IIX ∨ ,H . We now reduce our frame bundle to adapt to the

distinguished subspace Singloc(q n+a ): Let F ∗ → I 0 be the frame bundle of
bases of V over (x, H) ∈ I 0 adapted to the flag
0 ⊂ x̂ ⊂ {x̂, q̂sing
H
} ⊂ T̂x X ⊂ Ĥ ⊥ ⊂ V,
where qH ∈ PS 2 Tx∗ X is the quadric corresponding to H  ∈ N ∗ X. That is,
x
F ∗ is the bundle of frames (e0 , . . . , en+a ) such that
{e0 } = x̂,
{e0 , . . . , en−r } = x̂ + Singloc(qxH ),
{e0 , . . . , en } = T̂x X,
{e0 , . . . , en+a−1 } = Ĥ.

Observe that F ∗ is also first-order adapted to X ∨ because {x̂, q̂sing

H } =

(T̂H X ∨ )⊥ . The dual flag is

0 ⊂ Ĥ ⊂ (T̂x X)⊥ ⊂ T̂H X ∨ ⊂ x̂⊥ ⊂ V ∗ .
Applying Theorem 12.4.1 with the roles of X and X ∨ reversed implies
(T̂x X)⊥ = {Ĥ, IIX ∨ ,H (x)}, so F ∗ is equally adapted over X and X ∨ .
Fix further index ranges 1 ≤ s, t, u ≤ n − r, n − r + 1 ≤ i, j, k ≤ n.
Write the pullback of the Maurer-Cartan form to F ∗ as
⎛ 0 ⎞
ω0 ωt0 ωk0 0
ωn+λ 0
ωn+a
⎜ω s ωks ωts s
ωn+λ s
ωn+a ⎟
⎜ 0 ⎟
⎜ j ⎟
⎜
Ω=⎜ 0 ω ω j
k ω j
t ω j
n+λ ω j ⎟
n+a ⎟ ,
⎜ 0 ω n+κ ω n+κ ω n+κ ω n+κ ⎟
⎝ t k n+λ n+a ⎠
n+a n+a n+a
0 0 ωk ωn+λ ωn+a
where ωtn+a = 0 as we have adapted frames so that {et } = Singloc(q n+a ).
We calculate
den+a ≡ −ωkn+a ek − ωn+λ
n+a n+λ
e mod en+a .
n+a j
Note that ωkn+a = qkj n+a
ω0 with qkj n+a
invertible, and the forms ωn+λ are all
independent and independent of the semi-basic forms for π.
Exercise 12.4.15: Give a proof of Proposition 12.4.7 using a naı̈ve frame
argument as we did for varieties with degenerate Gauss maps. Namely, fix
H = [en+a ], and show that the distribution q̂ n+a |sing = {e0 , es } is integrable
and that the integral manifold Y containing [e0 ] has IIY = 0.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 346 12. Projective geometry II

Exercise 12.4.16: Show that with our adaptations, in particular assuming

μ
en+a is generic, that qst = 0 for all μ, s, t. (This is a special case of Bertini’s
Theorem 12.7.9.)

We now calculate |IIX ∨ ,H |. To compute IIX ∨ ,H we calculate

dek ≡ ωsk es + ω0k e0 mod T̂H X ∨ ,
den+λ ≡ ωsn+λ es mod T̂H X ∨ .
In our situation, (12.1) yields
n+a β
rstβ ω0 = 0,
n+a j
rsij ω0 = −qsi
n+λ n+a n+a k
ωn+λ + qkj ωs ,
which implies
n+a j
IIX ∨ ,H = (rsij n+λ n+a i s
ω0 + 2qsi ωn+λ )ω0 e + qjk ω0 ω0 e mod T̂H X ∨ .
n+a j k 0

Observe that if H = [en+a ] ∈ X ∨ smooth , then den+a has rank r + a − 1.

n+a
Since the forms ωn+λ are all independent and independent of the semi-basic
forms, we recover the fact that the forms {ωjn+a , j = n − r + 1, . . . , n} must
 has rank r.
span an r-dimensional space, i.e. II(H)
Proposition 12.4.17 (Inversion formula [103]). With respect to the bases
n+a
(ωn+λ , ω0j ) of TH∗ X ∨ and (e0 , es ) of NH X ∨ , we have
+    )
n+λ
0 0 0 qsj
IIX ∨ ,H = Q0 = n+a , Qs = n+λ n+a ,
0 qjk qsk rsjk

where the blocking is (a − 1, r) × (a − 1, r).

Remark 12.4.18. For further reading on dual varieties, see [180].

12.5. Secant and tangential varieties

Over the next few sections we outline proofs of the following results of Zak
[191]:
Theorem 12.5.1 (Zak’s theorem on linear normality). Let X n ⊂ Pn+a be
a smooth variety, not contained in a hyperplane, such that σ(X) = Pn+a .
Then a ≥ n2 + 2.
Theorem 12.5.2 (Zak’s theorem on Severi varieties). Let X n ⊂ Pn+a be
a smooth variety, not contained in a hyperplane, such that σ(X) = Pn+a ,
with a = n2 + 2. Then X is one of the four Severi varieties AP2 ⊂ PH.
As a first step, in this section we show how to compute dim σ(X) and
dim τ (X) from the differential invariants of X at a general point. We then
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.5. Secant and tangential varieties 347

state the Fulton-Hansen Theorem relating secant and tangential varieties.

Throughout this section we assume X ⊂ PV is a smooth variety.
Dimension of τ (X). Continuing the notation of §4.1, we may consider
F 1 → X as a bundle over the smooth points of τ (X) with projection
(e0 , . . . , en+a ) → [e1 ] because [e1 ] varies over all smooth points of τ (X).
In other words, F 1 sits above the following incidence correspondence:
T̃ X
 S
/
 w
S
X τ (X).

Thus, to calculate dim τ (X) we compute the differential of the projection

map onto the second factor. We fix index ranges 2 ≤ ρ ≤ n. We have
de1 ≡ ω10 e0 + ω1ρ eρ + ω1μ eμ mod{e1 },
so dim τ (X) equals the number of linearly independent forms among the
ω10 , ω1ρ , ω1μ . We are free to move e1 towards any of e0 , eρ because we are
assuming [e1 ] is a general point of τ (X). Thus, the forms ω10 , ω1ρ are all
independent and independent of the semi-basic forms. Thus we need to
consider the ω1μ = q1β μ β
ω0 . Fix v ∈ Tx X and define
IIv : Tx X → Nx X,
w → IIX,x (v, w).

We conclude:
Proposition 12.5.3 ([85]). Let x ∈ X be a general point and let v ∈ Tx X
be a generic tangent vector. Then dim τ (X) = n + rank IIv .
This agrees with our expectation that, generally, τ (X) has dimension
min{2n, n + a}. If this expectation fails, then we say τ (X) is degenerate.
For X with τ (X) degenerate, define the tangential defect to be
δτ (X) = 2n − dim τ (X).
Exercise 12.5.4: Calculate the tangential defects of the following vari-
eties:
(a) a curve X 1 ⊂ Pa+1 that is not a line;
(b) the Veronese vd (Pn );
(c) the Segre Seg(Pa × Pb );
(d) the Grassmannian G(2, 6).

Dimension of σ(X). Recall from Exercise 4.2.13.1 that if X n ⊂ Pn+a is a

variety, the expected dimension of σ(X) is min{2n + 1, n + a}. If σ(X) is
not Pn+a , define
δσ (X) = 2n + 1 − dim σ(X),
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 348 12. Projective geometry II

the secant defect of X, and we say σ(X) is degenerate if δσ (X) > 0. While
Terracini’s Lemma 4.2.12 enables us to easily compute dim σ(X) using one
derivative at two general points of X, computing dim σ(X) at one general
point of X is more subtle. We begin our calculation by first observing that
for varieties with degenerate secant varieties, part of the cubic form descends
to be a well-defined differential invariant.
For a vector space T and a subspace A ⊂ S 2 T ∗ , recall that

Singloc(A) = {v ∈ T | v q = 0, ∀q ∈ A}.

For v ∈ T , let

Ann(v) = AnnA (v) = {q ∈ A | v q = 0}.

Following [85], for varieties X n ⊂ Pn+a with degenerate tangential vari-

eties, or varieties where a > n, one can define a refinement of III at general
points x ∈ X. Namely, fixing a II-generic vector v ∈ T , the quantity

(12.11) 
ωμν ωβμ ω0β |π∗ (S 3 (Singloc(Ann(v)))) ⊗(eν mod{T̂ , II v (T )})⊗e
0

descends to be well-defined over X because by our hypotheses on the second

fundamental form, the relevant variability in F3 is forced to be zero.
Exercise 12.5.5: Prove (12.11) is indeed well defined over X. 

In what follows, we often use the abbreviations T = Tx X, N = Nx X

and suppress the reference to X and the general point x ∈ X.
Define the refined third fundamental form

III v ∈ S 3 (Singloc(Ann|II| (v)))∗ ⊗N/IIv (T )

by (12.11).
Exercises 12.5.6:
1. Show that III may be recovered from the forms III v for all v ∈ T . 
2. Show that if v ∈ T is a II-generic vector, then ker IIv ⊂ Singloc(Ann(v)).

General points of σ(X) are of the form [e0 + sf0 ], where [e0 ], [f0 ] are
points of X and s ∈ C. Label the corresponding frames for X above [e0 ], [f0 ]
with the same letters. Terracini’s Lemma 4.2.12 implies that the dimension
of σ(X) is one less than the number of linearly independent vectors among
e 0 , e α , f0 , fβ .
Now we use the assumption that X is smooth and connected. Set
v = e1 ∈ T[e0 ] X. Let (f0 (t), . . . , fn+a (t)) be a coframing over an arc f0 (t)
such that (f0 (0), . . . , fn+a (0)) = (e0 , . . . , en+a ) and f0 (0) = e1 . Expand
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.5. Secant and tangential varieties 349

f0 (t), fα (t) in the Taylor series:

t2  t3
f0 (t) = e0 + te1 + f0 (0) + f0 (0) + O(t4 ),
2 3!
t 2 t3
fα (t) = eα + tfα (0) + fα (0) + fα (0) + O(t4 ).
2 3!
Now (ignoring twists),
)
f0 (0) ≡ II(v, v)
mod T̂[e0 ] X
fα (0) ≡ II(v, eα )

and
)
f0 (0) ≡ III v (v, v, v)
mod{T̂[e0 ] X, IIv (T )}.
fα (0) ≡ III v (v, v, eα )

Higher-order terms in the series cannot contribute to the dimension of

σ(X), because either the lower-order terms (together with T̂[e0 ] X) span a
space of maximal dimension and σ(X) is nondegenerate, or III v is identi-
cally zero, and since the higher fundamental forms lie in the prolongation
of the lower fundamental forms, they are zero and the higher-order terms
cannot contribute any new directions. So, letting v ∈ T be II-generic, we
obtain the following formula, which was derived in [85]:

n + dim IIv (T ) if III v (v, v, v) = 0,

(12.12) dim σ(X) =
n + dim IIv (T ) + 1 if III v (v, v, v) = 0.

The Fulton-Hansen Theorem and consequences. Let X n ⊂ Pn+a be a variety.

Notice that by definition, τ (X) ⊆ σ(X). Recall that the expected dimension
of τ (X) is min{2n, n + a}, and that of σ(X) is min{2n + 1, n + a}. The
following theorem is proved using a version of Zariski’s main theorem valid
for open varieties (see [64]). It is an application of the celebrated Fulton-
Hansen connectedness theorem, which can be found in the same paper, [64].
Theorem 12.5.7 ([64]). Let X n ⊂ Pn+a be a variety. Then either
i. dim τ (X) = 2n and dim σ(X) = 2n + 1,
or
ii. τ (X) = σ(X).
Theorem 12.5.7, together with (12.12), implies
Proposition 12.5.8 ([123]). Let X n ⊂ Pn+a be a smooth variety such that
dim σ(X) < 2n + 1. Then for all II-generic vectors v ∈ T , we have III v = 0
(and therefore IIIX = 0). Conversely, if X n ⊂ Pn+a is a smooth variety
and IIIX is not identically zero, then σ(X) is nondegenerate.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 350 12. Projective geometry II

Corollary 12.5.9. The homogeneous varieties Seg(Pa1 ×· · ·×Pak ) for k > 2,

G(k, k + l) for k, l > 2, and vd (Pn ) for d > 2 all have nondegenerate secant
varieties.

We will use the Fulton-Hansen Theorem when proving the rank restric-
tion theorems in §12.11, and again when giving a proof of Zak’s theorem on
Severi varieties using moving frame techniques in §12.12.
The tangential variety can be generalized as follows: if Y ⊆ X ⊆ Pn+a ,
and y ∈ Y , define Ty (Y, X) to be the union of P1∗ ’s, where P1∗ is a limit
of P1xy ’s when x ∈ X and y ∈ Y , and x, y → y0 ∈ Y . Define τ (Y, X) =

y∈Y Ty (Y, X), the variety of relative tangent stars. As observed by Zak,
the Fulton-Hansen Theorem 12.5.7 generalizes to relative tangent stars and
joins as follows:
Theorem 12.5.10 ([191]). Let X n , Y y ⊂ PV be varieties, respectively of
dimensions n, y. Assume Y ⊆ X. Then either
i. dim J(Y, X) = n + y + 1 and dim τ (Y, X) = n + y,
or
ii. J(Y, X) = τ (Y, X).
Using Theorem 12.5.10, one obtains:
Theorem 12.5.11 (Zak’s theorem on tangencies [191]). Let X n ⊂ Pn+a
be a variety not contained in a hyperplane. Let dim Xsing = b, with the
convention that b = −1 if X is smooth. Let L be any Pn+k ⊂ Pn+a ; then
dim{x ∈ X | T̃x X ⊆ L} ≤ k + (b + 1).

Proof. We give the proof assuming X is smooth. Let Y = {x ∈ X | T̃x X ⊆

L} and let y = dim Y . Then τ (Y, X) ⊂ L, but since X is not contained in
a hyperplane, σ(Y, X) ⊆ L. Thus by Theorem 12.5.7, dim τ (Y, X) = y + n,
so y + n ≤ n + k, i.e., y ≤ k. 

Corollary 12.5.12 (Zak; see [191]). If X is a smooth variety that is not a

linear space, then for all y ∈ γ(X), γ −1 (y) is a finite set.

12.6. Cominuscule varieties and their differential invariants

The Severi varieties. This subsection has less detail than others. It will
be used in the proof of Zak’s theorem on Severi varieties.
Let A denote the complexification AR ⊗C of a real division algebra AR
(see §A.5), and let HR denote the AR -Hermitian forms on A3R , i.e., the 3 × 3
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.6. Cominuscule varieties and their differential invariants 351

AR -Hermitian matrices. If x ∈ HR , then we may write

⎛ ⎞
r1 ū1 ū2
x = ⎝u1 r2 ū3 ⎠ , ri ∈ R, ui ∈ AR .
u2 u3 r3
Let H = HR ⊗C.
Exercise 12.6.1: Verify that the notion of x2 and x3 make sense. In the
case of the octonions, one needs to use the Moufang identities (A.12).

Define a cubic form detA on H by

1 
(trace(x))3 + 2 trace(x3 ) − 3 trace(x) trace(x2 ) .
detA (x) :=
6
This detA is just the usual determinant of a 3 × 3 matrix when A = C.
Now, considering H as a vector space over C, let G be the subgroup of
GL(H, C) preserving detA , i.e., define
G := {g ∈ GL(H, C)| detA (gx) = detA (x) ∀x ∈ H}.
For the four division algebras, the respective groups are:
AR G
R SL(3, R)C = SL(3, C)
C SL(3, C)C = SL(3, C) × SL(3, C)
H SL(3, H)C = SL(6, C)
O ‘SL(3, O)C ’ = E6
We take the above as the definition of E6 , and have written ‘SL(3, O)C ’
merely to be suggestive.
Definition 12.6.2. The group F4 is the subgroup of E6 preserving the qua-
dratic form Q(x, x) = trace(x2 ), where x2 is the usual matrix multiplication
and one must again use the Moufang identities (A.12) to be sure that Q is
well-defined.
Exercise 12.6.3: Show that the action of F4 on H preserves the line {Id}C
and H0 := {x ∈ H | trace(x) = 0}. (In fact, F4 acts irreducibly on both
factors.)

The cubic form detA tells us which elements of H are of less than full
rank. One can also unambiguously define a notion of being rank one, either
by taking 2 × 2 minors or by noting that under the G action each x ∈ H
is diagonalizable and one can take as the rank of x the number of nonzero
elements in the diagonalization of x. Let
X := P{rank one elements of H} = P{G-orbit of any rank one matrix}.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 352 12. Projective geometry II

Then X = (AR P2 )C , that is, the complexification of the space of AR -lines

in A3R . The four varieties X ⊂ PH are called the Severi varieties. The
octonionic case is called the (complexified) Cayley plane.
Exercises 12.6.4:
1. Show that the first three Severi varieties are v2 (P2 ) ⊂ P5 , Seg(P2 × P2 ) ⊂
P8 , and G(2, 6) ⊂ P14 .
2. What are the analogous groups and varieties if one takes instead the
2 × 2 A-Hermitian forms?
3. Let OP20 = H0 ∩ OP2 , where H0 ⊂ H is the set of traceless elements.
Show that
OP20 = P{x ∈ H0 | x2 = 0}
and deduce that it is a homogeneous space of the group F4 .
4. For each of the Severi varieties, calculate dim σ(X).

Fundamental forms of the Severi varieties. Here it is easier to use local

coordinates. Choose affine coordinates based at [p], where
⎛ ⎞
1 0 0
p = ⎝ 0 0 0⎠ ,
0 0 0
and denote the affine coordinates u1 , u2 , u3 ∈ A, r2 , r3 ∈ C, where the tan-
gent space to p is {u1 , u2 } (the span is taken over C). In these coordinates:
 
1 u¯1
r2 (u1 , u2 ) = u1 u¯1 because det = 0,
u1 r2
 
1 u¯2
r3 (u1 , u2 ) = u2 u¯2 because det = 0,
u2 r3
 
1 u¯1
u3 (u1 , u2 ) = u¯2 u1 because det = 0,
u2 u3
where the last equation gives us one, two, four or eight quadratic forms. The
determinants come from the vanishing of 2 × 2 minors that must be zero
to make the Severi variety consist only of rank one elements. In division
algebra notation the second fundamental forms are
|II| = P{u1 u¯1 , u2 u¯2 , u¯2 u1 }.

The spinor variety S5 and its cousins. An important series of homogeneous

varieties is the spinor varieties described below. The first spinor variety
that does not coincide with a homogeneous variety that we have already
5 ⊂ P . It admits a description using the octonions (see §A.5)
discussed is S10 15

as follows: Let C = O2 ⊗C have (complexified) octonionic coordinates u, v.

16

Then S5 is defined by the equations uu = 0, vv = 0 and uv = 0. Note that

the last expression gives eight equations for the components of u and v.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.6. Cominuscule varieties and their differential invariants 353

Exercise 12.6.5: What are the varieties analogous to S5 for the other di-
vision algebras?

Cominuscule varieties.
Definition 12.6.6. Let X = G/P ⊂ PV be a homogeneous variety where
G is a complex semi-simple Lie group acting linearly on V and P is the
subgroup stabilizing a point of X (such a P is called a parabolic subgroup).
We say X is a generalized cominuscule variety if Tx X contains no proper
irreducible P -submodule. X is a cominuscule variety if moreover G is simple
and the embedding is minimal, which is equivalent to the embedding not
being a Veronese re-embedding.

Alternatively, the generalized cominuscule varieties are those admitting

a Hermitian symmetric metric induced from a Fubini-Study metric on the
ambient projective space, and the cominuscule varieties are those for which
the metric is irreducible and the embedding minimal.
Examples of cominuscule varieties include the quadric hypersurfaces
Qn ⊂ Pn+1 , the Grassmannians G(k, V ), the spinor varieties Sm =
Spin(2m)/P defined below, the Cayley plane OP2 , and the Lagrangian
Grassmannians Gω−null (m, 2m). In fact these are all the cominuscule va-
rieties except for Gw−null (O3 , O6 ), which is a homogeneous variety of the
exceptional group E7 (see [134] for a description of Gw−null (O3 , O6 )).

If we know the structure of the tangent space at a point of a cominuscule

variety as a P -module, the fundamental forms are easy to determine.
Proposition 12.6.7 ([136]). Let X = G/P ⊂ PV be a cominuscule variety
and let x ∈ X. Let H ⊂ P be a maximal semi-simple subgroup and con-
sider Tx X as an H-module. Let Y ⊂ PTx X be the closed H-orbit. Then
Baseloc |IIX,x | = Y , and moreover |IIX,x | = I2 (Y ).
For example, once we know TE G(k, W )  E ∗ ⊗W/E and H = SL(E) ×
SL(W/E), we may use this proposition to immediately conclude that
Baseloc |IIG(k,W ),E | = Seg(PE × P(W/E))
and the second fundamental form is spanned by the 2 × 2 minors.

Remark 12.6.8. For readers familiar with Dynkin diagrams, H and the H-
module Tx X can be determined pictorially: H is a semi-simple group with
Lie algebra whose Dynkin diagram is that of G with the node corresponding
to the root defining the maximal parabolic P removed. Moreover, if the
diagram is simply laced, the H-module Tx X is the irreducible module with
highest weight the sum of the fundamental weights corresponding to the
nodes of the Dynkin diagram adjacent to the deleted node.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 354 12. Projective geometry II

For example, in the following picture, the parabolic defining X = G(4, 7)

and its natural embedding corresponds to the darkened node in the Dynkin
diagram of A7 = SL8 , where to define the parabolic we omit the root cor-
responding to the node, and to define the embedding we take the represen-
tation with support the fundamental weight corresponding to the marked
node. The A3 × A2 = SL4 × SL3 module Tx X is pictured below it, and the
same marked diagram gives Seg(P3 × P2 ) = Baseloc |IIG(3,7) |.

Figure 1. Parabolic defining X = G(4, 7)

Figure 2. SL4 × SL3 module Tx X

Exercise 12.2.1 shows that equality in (12.6) holds for the examples of
Grassmannians. In fact, equality holds for all cominuscule varieties:
Theorem 12.6.9 ([136]). Let X = G/P ⊂ PV be a cominuscule variety
and let x ∈ X. Then for k ≥ 2,

X,x | = |FFX,x |
|FFk+1 2 (k−1)
.

This strict prolongation property fails for a general homogeneous variety,

e.g., it does not hold for GQ−null (k, m) for 2 ≤ k <  m
2 . This property has
the following geometric consequence:
Corollary 12.6.10 ([136]). Let X be a cominuscule variety, and let x ∈ X.
Then
|FFkX,x | = Ik (σk−1 (Baseloc |FF2X,x |)).
See [136] for proofs.

Spinor varieties and the spin representation. For sl(W ), the funda-
mental representations are just the exterior powers of the standard repre-
sentation (see §A.6). For so(W ), assuming for simplicity that dim W = 2m,
fundamental representations are furnished by the exterior powers of the
standard representation, up to Λm−2 W . There are two other fundamental
representations of so(W ), dual to one another, one of which we now describe.
These representations are called the spin representations. In terms of Lie
groups, the spin representations are representations of the simply connected
double cover of SO(W ), called Spin(W ), but not of SO(W ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.6. Cominuscule varieties and their differential invariants 355

We will use Corollary 12.6.10 to construct a model for the spin repre-
sentation. Readers familiar with Dynkin diagrams may skip the first step,
where the tangent space is calculated, thanks to Remark 12.6.8.
Before starting, we examine the Grassmannian G(k, W ) from the per-
spective we will use to construct the spin representation. Say we knew
Corollary 12.6.10 but were unaware of the Plücker embedding. We would
then construct V = Λk W as follows: fix E ∈ G(k, W ); then V is the sum
of the successive quotient spaces Λk E (which we identify with the point
E), E ∗ ⊗W/E = Λk−1 E ⊗W/E, Λ2 E ∗ ⊗Λ2 (W/E) = Λk−2 E ⊗Λ2 (W/E),...,
Λk E ∗ ⊗ Λk (W/E) = Λ0 E ⊗ Λk (W/E). Note that only the first space is
well-defined as a subspace of Λk W , but the sum of the first p spaces is a
well-defined subspace of Λk W for all p, so we obtain a Λk W equipped with
a filtration.
We now apply the same method to recover the spin representation. The
variety GQ−null (m, C2m ) has two isomorphic components; we define Sm to
be one of them.

Choose a basis of V = C2m so that Q = 0I I0 . With respect to this
basis SO(V, Q) has Maurer-Cartan form
 i i 
ωj ωm+k
m+k ,
ωjn+l ωm+l

where the index range is 1 ≤ i, j, k ≤ m, and ωm+l

m+k
= −ωkl , ωn+k
i = −ωm+i
k

and ωjm+l = −ωlm+j .

 = e1 ∧ . . . ∧ em ∈ Ŝm . By Remark 12.6.8 we could conclude that
Let E
TE Sm  Λ2 E ∗ , or, computing as for the Grassmannians,
 ≡ ω n+i E j mod E.
dE j n+i

This determines TE S as a linear subspace of E ∗ ⊗V /E. Since Q allows

us to identify V /E with E ∗ , we may consider TE S as a subspace of E ∗ ⊗E ∗ .
The subspace is Λ2 E ∗ because the only relations among the forms ωjn+i are

ωjn+i = −ωin+j . Thus, TE Sm = Λ2 E ∗ , and we note that dim Sm = m 2 .
At this point we may apply Proposition 12.6.7, but even without the
proposition we could calculate the second fundamental form of S using our
frame bundle:
 jl
 ≡ ω n+i ω n+k E
d2 E j l n+i,n+k mod T̂E S.
Thus
|IIS,E |  Λ4 E ∗ = I2 (G(2, E)).
Remark 12.6.11. We cheated slightly in the last calculation, as we pre-
tended to be working with the SO(W )-frame bundle, but in fact we are
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 356 12. Projective geometry II

working with its simply connected double cover Spin(W ). Since our calcu-
lation was on the level of Lie algebras, we got the correct answer.

Exercises 12.6.12:
1. Verify that Baseloc |IISm ,E | = G(2, E) by showing that II(v, v) = 0 if
and only if v = w1 ∧ w2 . 
2. Calculate the higher fundamental forms of Sm via moving frames to
verify that |FFk | = Λ2k E ∗ = Ik (σk−1 (G(2, E))).

Adding up the images of the FFk ’s, we see that Sm ⊂ P(Λeven E ∗ ). We

conclude that Spin(2m) and so2m act on Λeven Cm ; either of these actions is
called the spin representation or sometimes the half-spin representation.
In addition to the special construction for so10 in the previous section,
another model for the spin representation, using Clifford algebras, is de-
scribed in Appendix A.
Exercise 12.6.13: Show that in the case of S5 , our construction agrees with
that in 12.6.

12.7. Higher-order Fubini forms

In this section we define the higher-order relative differential invariants which
we call the Fubini forms and give some applications. We explain their use
in the study of osculating hypersurfaces to a variety. We prove the classical
Bertini Theorem for systems of quadrics and higher-order generalizations
useful for projective geometry. More extensive applications are given in
§12.3, §12.8 and §12.13.

Definition of the Fubini forms Fk . Let M n ⊂ Pn+a be a submanifold,

let π : F 1 → M be its first-order adapted frame bundle, and continue the
index ranges 1 ≤ α, β ≤ n and n + 1 ≤ μ, ν ≤ n + a.
Differentiating F3 , one obtains a fourth-order invariant
μ
F4 = rαβγδ ω0α ω0β ω0γ ω0δ ⊗eμ ∈ Γ(F 1 , π ∗ (S 4 T ∗ M ⊗N M ))
whose coefficients are defined by
μ μ μ μ μ
rαβγδ ω0δ = −drαβγ − 2rαβγ ω00 − rαβγ
ν
ωνμ + Sαβγ (rαβ ωγ + qαβ ωγ0 − qαμ qβγ
ν
ων ).
The geometric interpretation of F4 is that it measures how X leaves its
embedded tangent space to third order.
As with F3 , motion in the fiber by a block diagonal matrix will not effect
F4 . Under the fiber motion (12.3), the coefficients of F4 vary as follows:
μ μ μ
r̃αβγδ = rαβγδ + Sαβγδ [gα0 rβγδ ν
+ gν (rαβγ qδμ + qαβ
ν μ μ ν
rγδ ) + gν0 qαβ qγδ ].
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.7. Higher-order Fubini forms 357

One continues, defining forms Fk ∈ Γ(F 1 , π ∗ (S k T ∗ ⊗ N )) for all k. The

coefficients of Fl are defined by
rαμ1 ,...,αl ω0αl = − drαμ1 ,...,αl−1 − (l − 2)rαμ1 ,...,αl−1 ω00 − rαν 1 ,...,αl−1 ωνμ
+ Sα1 ,...,αl−1 [rαμ1 ,...,αl−2 β ωαβl−1 + (l − 3)rαμ1 ,...,αl−2 ωα0 l−1

l−3
− rαμ1 ,...,αp ,β rαν p+1 ,...,αl−1 ωνβ
p=1


l−3
− (p − 2 + l)rαμ1 ,...,αp rαν p+1 ,...,αl−1 ων0 ].
p=2

This formula corrects the typographical errors in (2.20) of [124] (where the
proof is given without typographical errors).

Projective differential invariants in coordinates. Let M ⊂ PV be a

submanifold and let x1 , . . . , xn+a be local coordinates such that locally M
∂f μ
is given as a graph xμ = f μ (x1 , . . . , xn ). Write fxμα = ∂x α.

1 defined by e = ∂ + f μ ∂ , e = ∂ . In
Consider the section of FM α ∂xα xα ∂xμ μ ∂xμ
this framing ω0 = dx , or, more precisely, s∗ (ω0α ) = dxα . We compute
α α

∂
. deα ≡ fxμα xβ dxβ ⊗
∂xμ
In terms of the Maurer-Cartan form, we have
deα ≡ ωαμ eμ mod{e0 , eβ },

so ωαμ = fxμα xβ ω0β . Thus

II = fxμα xβ dxα dxβ ⊗eμ .

With a little more work, one can adapt frames so that

Fk = fxμα1 ,...,xαk dxα1 . . . dxαk ⊗eμ ,
i.e., the coefficients of Fk are just the terms appearing in the k-th term of
the Taylor series expansion of the functions locally defining M as a graph
over its tangent space.
Exercise 12.7.1: Show that there exists a section s : M → F 1 such that
s∗ (F3 ) = fxμα xβ xγ dxα dxβ dxγ ⊗ ∂x∂μ .
Corollary 12.7.2. An analytic submanifold M ⊂ PV is uniquely deter-
mined up to projective equivalence by the infinite sequence of relative dif-
ferential invariants Fk at any point.

Proof. They determine the Taylor series. 

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 358 12. Projective geometry II

Remark 12.7.3. A projective variety X ⊂ PV will be uniquely determined

up to projective equivalence by a finite number of the Fk , but it is subtle
to determine exactly how many Fk are needed, even if one has the degrees
of the defining equations of X. The simplest case is described in the next
section.

Osculating hypersurfaces and Veronese re-embeddings. Given a va-

riety X ⊂ PV , and a point x ∈ X, how could we determine if X is contained
in a hyperplane H? An obvious necessary condition is that x ∈ H, and more-
over that T̃x X ⊆ H, i.e., that X does not infinitesimally move away from
H at x to first order. The next thing to check is that X does not infinites-
imally move away from H at x to second order, i.e., that H  ∈ ker IIX,x ,
where H ∈ Nx X is a vector corresponding to H. (Recall that PNx∗ X may
 ∗

be identified with the space of hyperplanes containing T̃x X.) Continuing,

X ⊆ H if and only if H  ∈ ker FFk for all k. If x is a general point and X
X,x
is a hypersurface, then we have seen (Exercise 4.1.12.2) that it is sufficient
to check that H  ∈ ker IIX,x (i.e., to check that IIX,x = 0). If the codimen-
sion of X is greater than one but still small and X is not too singular, then
Corollary 12.11.3 will imply that the same conclusion holds. In this sub-
section and §12.13 we address the corresponding question for higher degree
hypersurfaces. Our approach will be to reduce to the hyperplane case by
re-embedding X via a Veronese mapping.
We first determine, for each d and small k, a priori bounds on the
dimensions of hypersurfaces of degree d osculating to order k at a smooth
point of a variety X n ⊂ Pn+a . Recall that for degree one, the dimension of
hypersurfaces of degree one (i.e., hyperplanes) osculating to order one (i.e.,
containing T̃x X) at a smooth point x ∈ X is fixed (equal to a − 1). The
dimension of the space of hyperplanes osculating to order two depends on
the geometry of X and the point x. We show that a similar phenomenon
happens when d > 1. We begin with d = 2.
We use the notation F0 = FF0 = x̂⊗ x̂∗ and F1 = FF1 = IdTx X .

Fundamental forms of v2 (X) and osculating quadrics. Assume X ⊂

PV is such that II : S 2 Tx X → Nx X is surjective at general points x ∈ X.
Write x = [e0 ], v2 (x) = [e0 2 ], and use the Leibniz rule applied to e0 2 to
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.7. Higher-order Fubini forms 359

compute
FF1v2 (X),v2 (x) = 2F1 F0 |x̂2⊥ ,
FF2v2 (X),v2 (x) = 2(F2 F0 + F1 F1 )|(x̂T̂ )⊥ ,
FF3v2 (X),v2 (x) = 2(F3 F0 + 3F2 F1 )|ker FF2 ,
v2 (X)
(12.13)
FF4v2 (X),v2 (x) = 2(F4 F0 + 4F3 F1 + 3F2 F2 )|ker FF3 ,
v2 (X)

FF5v2 (X),v2 (x) = 2(F5 F0 + 5F4 F1 + 10F3 F2 )|ker FF4 ,

v2 (X)

..
.
Here Fj = Fj,X,x are the relative differential invariants of X ⊂ PV .
We make several observations. First, note that |FF1v2 (X),v2 (x) |  Tx∗ X,
|FF2v2 (X),v2 (x) |  S 2 Tx∗ X, and thus the dimensions of their kernels are re-
  
spectively n+a+1 2 − (n + 1) and n+a+12 − (n + 1) − n+1
2 . This says:

The dimensions of the spaces of quadrics osculating to X

to orders one and two at a smooth point x depend only on
the dimension and the codimension of X.
If we choose a splitting of our flag x̂ ⊂ T̂ ⊂ V and slightly abuse notation,
writing V = x + T + N , then
S 2 V ∗ = S 2 x∗ + x∗ ⊗T ∗ + x∗ ⊗N ∗ + S 2 T ∗ + T ∗ ⊗N ∗ + S 2 N ∗ .
The quadrics corresponding to the S 2 N ∗ factor are necessarily in the kernel
of FF3v2 (X),v2 (x) , and this observation is independent of our choices, so we see
that:

The dimensions of the spaces of quadrics osculating to X

to order three at a smooth point x is bounded below by a
function of the dimension and the codimension of X.

The dimension of ker FF3v2 (X),v2 (x) depends on the geometry of F2 =

IIX,x . Recall from Appendix A that the kernel of the symmetrization map
T ∗ ⊗ S 2 T ∗ → S 3 T ∗ , which we denoted KS , is isomorphic to the module
S21 T ∗ . An element of F1 ⊗F2 will fail to contribute to FF3v2 (X) if and only if
it lies in KS , so

The dimension of ker FF3v2 (X),v2 (x) is guaranteed to be as

small as possible if (T ∗ ⊗|II|) ∩ KS = 0.

Note that if (T ∗ ⊗ |II|) ∩ KS = 0, then there exists a relation Q1 l1 +

· · · + Qp lp = 0 with Qj ∈ |II| and lj ∈ T ∗ . Such a relation is called a linear
syzygy among the quadrics in |II|.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 360 12. Projective geometry II

Assuming ker FF3v2 (X),v2 (x) is as small as possible, ker FF4v2 (X),v2 (x) will
be assured to be as small as possible if the map |II|⊗|II| → S 4 T ∗ has no
kernel.
Exercise 12.7.4: Show that the map |II| ⊗ |II| → S 4 T ∗ has no kernel if
there are no linear syzygies among the quadrics in |II|.

vd (X) and osculating hypersurfaces of degree d. We have the following gen-

eralization of (12.13):
Proposition 12.7.5 ([124]). The fundamental forms of vd (X) are
 
FFkvd (X) = cl1 ...ld Fl1 . . . Fld mod( FFlvd (X) (S l T ))|ker FFk−1 ,
vd (X)
l1 +...+ld =k l<k
where the cl1 ...ld are nonzero constants.
For example,
FF4v3 (X) = c400 F4 F0 F0 + c310 F3 F1 F0 + c220 F2 F2 F0 + c211 F2 F1 F1 .
For the proof, see [124].
Remark 12.7.6. We need to be careful if we try to recover Fk (X) from
the FFlvd (X) , as a cancellation could occur. For example, if X has small
codimension, we will have FFkv2 (X) = 0 for k large, but Fk (X) will not be
zero in general. However it is possible to recover Fk from FFk+1
vk−1 (X) .

If Z is a hypersurface of degree d, then Z osculates to X to order p at

x ∈ X if and only if H Z ∈ ker FFk  Z ∈ N ∗ vd (X) is the
, where H
v (X),xd vd (x) d
vector associated to the equation HZ ∈ PS d V ∗ of Z. Similarly, X ⊂ Z if
 Z ∈ ker FFk
and only if H for all k.
v (X),xd d

Proposition 12.7.7 ([124]). Let X n ⊆ Pn+a be a variety and let x ∈

Xsmooth . For all p ≤ d, the dimension of the set of (not necessarily irre-
ducible) hypersurfaces of degree d osculating to order p at x is
   
n+a+d n+p
− .
d p
As with quadrics, for p > d, the dimensions of the spaces of osculating
hypersurfaces depend on the geometry of X. We also see that, independent
of X, for d + 1 ≤ k ≤ 2d − 1, there are lower bounds on the dimensions of the
space of hypersurfaces of degree d osculating to order k at x. For example:
Proposition 12.7.8 ([124]). Let X n ⊆ Pn+a be a variety, and let x ∈
Xsmooth . The dimension of the set of (not necessarily irreducible) hypersur-
faces of degree d osculating to order 2d − 1 at x is at least
 
a+d−1
− 1.
d
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.7. Higher-order Fubini forms 361

Higher-order Bertini Theorems. Let T be a vector space. The classical

Bertini Theorem is as follows:
Theorem 12.7.9. Let A ⊂ S d T ∗ be a linear subspace, let q ∈ A be generic,
and let Q ⊂ PT denote the hypersurface defined by q = 0. Then v ∈ Qsing
implies v ∈ Baseloc(A) := {v ∈ T | v ∈ Q ∀Q ∈ A}.
We will prove a generalization of this theorem that includes the original
statement.
If A now varies in a family {As }s∈S , there are restrictions on how A
moves infinitesimally in the family. We will only be concerned with the
case when d = 2 and the family is determined by the second fundamental
form of a complex submanifold. In this case the higher-order differential
invariants measure how the family varies. Although we use the language of
submanifolds of projective space, these results are valid in the more general
context of studying a subvariety Ξ ⊂ G(k, S 2 T ∗ ).
Theorem 12.7.10 (Higher-order Bertini [131]). Let M n ⊂ PV be a com-
plex submanifold and let x ∈ M be a general point.
1. Let q ∈ |II|M,x be a generic quadric. Then Qsing ⊂ Baseloc{F2 , . . . , Fk }
for all k, i.e., Qsing is tangent to a linear space on the completion of M .
2. Let h ∈ Nx∗ M and let L ⊂ II(h)sing ∩ Baseloc |II| be a linear subspace.
Then h F3 (v, w, ·) = 0 for all v, w ∈ L.
3. With h and L as in part 2, if L ⊂ (Baseloc{|II|, F3 }) ∩ L is a lin-
ear subspace, then h F4 (u, v, w, ·) = 0 for all u, v, w ∈ L , and so on for
higher orders. (Note that h F4 (u, v, w, ·) is well-defined by the lower-order
vanishing.)
4. With h and L as in part 3, if L ⊂ L ∩ (h F3 )sing is a linear space, then
for all u, v ∈ L we have h F4 (u, v, ·, ·) = 0.

Proof. In each case, the extension to linear spaces will hold by polarizing
the forms. We prove only the first assertion in each case, as an analogous
equation at each order proves the next higher order.
μ
1. Assume v = e1 and q = qαβ ω0α ω0β . Our hypotheses imply q1β
μ
= 0 for
all β. Formula (12.1) reduces to
μ
r11β ω0β = −q11
ν μ
ων .

If q is generic we are still working on F 1 , and so the ωνμ are independent of

the semi-basic forms; thus the coefficients on both sides of the equality are
zero. Moreover, the relevant equation for F4 reduces to
μ
r111β ω0β = −r111
ν
ωνμ ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 362 12. Projective geometry II

and again, both sides are zero. This and the corresponding higher-order
equations also prove both the classical Bertini Theorem and part 1 of the
theorem.
2. If h = en+a is not generic, we have to reduce to a subbundle F  ⊂ F 1
where the ωνn+a are not necessarily independent. However, the hypotheses
ν = 0 for all ν, and the required vanishing still holds.
of part 2 state that q11
n+a δ ν ω n+a + r n+a ω + q n+a q ν ω and the
For part 3, note that r111δ ω0 = r111 ν 11 1 1 11 ν
right hand side is zero under our hypotheses. Part 4 is proven similarly. 

Let Σk = Σkx ⊂ PTx X denote the variety of tangent directions to lines

osculating to order k at x. If X is a hypersurface, Σk is the zero set of
polynomials of degrees 2, 3, . . . , k, so one would expect Σn+1 to be empty.
Remark 12.3.5 above says that if it is not, it consists of tangent directions
to actual lines. More generally, the expected dimension of Σk is n − k, and
we have the following result:
Theorem 12.7.11 ([130]). Let X n ⊂ Pn+1 be a hypersurface and let x ∈ X
be a general point. If there is an irreducible component Σk0 ⊂ Σk such that
dim Σk0 > n − k, then all lines in PV through x, tangent to directions in Σk0 ,
are contained in X.

Proof. Take a basis e1 , . . . , en of Tx X such [e1 ] is a general point of Σk0 , and

T̃[e1 ] Σk0 = P{e1 , e2 , . . . , ep }, where p − 1 = dim Σk0 > n − k.
Let 1 ≤ α, β ≤ n. In this proof only, we change notation slightly and let
rα,β denote the coefficients of the second fundamental form II of X at x,
while rα1 ,...,αi denotes the coefficients of Fi . We write rαi 1 ,...,αi = rα1 ,...,αi to
keep track of i. We use index ranges: 2 ≤ s, t ≤ p and p + 1 ≤ j, l ≤ n.
λ
By our normalizations, we have r1,...,1 λ
= 0 and r1,...,1,s = 0 for 2 ≤ s ≤ p
and 2 ≤ λ ≤ k. We will show that r1,...,1 = 0 and r1,...,1,s = 0 for 2 ≤ s ≤ p
h h

and for all h, which will imply that there is a p-dimensional space of lines
passing through X at x.
We proceed by induction on h, using our lower-order equations to solve
for the connection forms ω1j in terms of the semi-basic forms ω0l , and then
plug into the higher Fh to show that Fh (e1 , . . . , e1 ) and Fh (e1 , . . . , e1 , es )
both vanish. Using (12.7) we obtain that, for λ ≤ k,
λ
r1,...,1,i ω0i = −r1,...,1,j
λ−1
ω1j .
These are k − 2 equations for the n − p 1-forms ω1j in terms of the n − p
semi-basic forms ω0j . Recall that n − p ≤ k − 2. The system is solvable,
because were the (k − 2) × (n − p) matrix (r1,...,1,j
λ ) not of maximal rank,
there would be additional directions in the tangent space to Σk0 . Thus
ω1j ≡ 0 mod{ω0p+1 , . . . , ω0n },
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.8. Varieties with vanishing Fubini cubic 363

so the equation
k+1
r1,...,1,β ω0β = −r1,...,1,j
k
ω1j
k+1 k+1 k+1
implies r1,...,1 = r1,...,1,2 = . . . = r1,...,1,p = 0. Thus the line through [e1 ] has
contact to order k + 1, and moreover T̃[e1 ] Σk+1 0 = T̃[e1 ] Σk0 . Now one can use
these equations iteratively to show the same holds to order k + 2 and all
orders, i.e., the component of Σk containing [e1 ] at a general point equals
the component of Σ∞ containing [e1 ]. 

We obtain the following corollary:

Theorem 12.7.12 ([130]). Let X n ⊂ Pn+a be covered by lines. Then there
are at most n! lines passing through a general point of X.

Proof. Without loss of generality we may assume X is a hypersurface, as

one can reduce to this case by linear projection. First note that n! is the
expected bound in the sense that in PTx X one has the ideal I generated by
F2 , F3 , . . . , Fn defining the variety Σn ⊂ PTx X of all lines having contact
with X at x to order n. (Recall that the polynomials of degree greater than
two are not well-defined individually, but the ideal I is.) Since PTx X is a
Pn−1 , if Σn is zero-dimensional, it is at most n! points and we are done. If
dim Σn > 0, then Theorem 12.7.11 applies. 

12.8. Varieties with vanishing Fubini cubic

We saw in Exercise 4.1.12.2 that if the second fundamental form of a sub-
manifold M is identically zero, then the completion of M is a linear space.
What is the geometric interpretation of F3 = 0? First of all, the condition
must be correctly interpreted–that there exists a framing of M , s : M → F 1
such that s∗ (F3 ) = 0. In codimension one we have the following theorem:
Theorem 12.8.1 (Fubini [63]). Let M n ⊂ CPn+1 , n ≥ 2, be an algebraic
or analytic hypersurface with γM nondegenerate. If there exists a framing
s : M → F 1 such that s∗ (F3 ) = 0, i.e., if for all sections s : M → F 1 we
have s∗ (F3 ) = l ◦ II with l ∈ Ω1 (M ), then M is (an open subset of) the
smooth quadric hypersurface Qn ⊂ Pn+1 .
Remark 12.8.2. Fubini’s Theorem is valid over R and in the C ∞ cate-
gory, with the conclusion that the hypersurface is an open subset of the
quadric whose second fundamental form has the same signature as M . (The
same proof as below works, with additional notation to keep track of the
signature.)

We first study SO(V, Q) → Q as an adapted frame  bundle. Use linear

0
coordinates (x , . . . , x n+1 ) on V and let Q = {x x
0 n+1 − α (xα )2 = 0}. We
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 364 12. Projective geometry II

use Q to denote both the quadratic form and the quadric hypersurface it
determines.
Exercise 12.8.3: Show that, with this choice of basis,
⎧⎛ 0 ⎞ ⎫
⎨ a0 a0β 0  0 β ⎬
 a + an+1
= 0, a n+1
− a = 0,
so(V, Q) = ⎝aα0 aαn+1 ⎠
0 n+1 β 0
aαβ .
⎩ n+1 n+1  a α
+ aβ
= 0, aα
− a0
= 0 ⎭
0 aβ an+1 β α n+1 α

With these choices, SO(V, Q) is a bundle of first-order adapted frames

to Q via the projection
SO(V, Q) → PV,
(e0 , . . . , en+1 ) → [e0 ].
Exercise 12.8.4: Show that in these frames, IIQ,[e0 ] = ((ω01 )2 +. . .+(ω0n )2 )⊗
en+1 and F3 = 0, i.e., that if i : SO(V, Q) ⊂ FQ 1 is the inclusion, then

i∗ (F3 ) = 0.

Proof of Theorem 12.8.1. We now begin with M an unknown hypersur-

face such that F3 = 0 in some framing, and show that a subbundle of the
bundle of frames f where F3f = 0 and det(f ) = 1 is SO(V, Q). We then use
Theorem 1.6.13 to conclude the proof.
We normalize the second fundamental form so that
ωαn+1 = ω0α
and reduce to the subbundle of the frame bundle where det(f ) = 1. Then
0 = rαβγ ω0γ = −δαβ (ω00 + ωn+1
n+1
) + (ωβα + ωαβ ),
which shows that
ωβα + ωαβ = 0 for α = β,
2ωαα = ω00 + n+1
ωn+1 ∀α.

The second equation implies that ωαα = ωββ for all α, β, and the condition
det(f ) = 1 implies trace(ω) = 0, so
n+1
(n + 2)ωαα = ω00 + ωn+1 + ω11 + . . . + ωnn
= 0.
Thus
ωαα = 0, ∀α,
n+1
ω00 + ωn+1 = 0.
To reduce the Maurer-Cartan form to the Maurer-Cartan form for
α
SO(V ), we need to show that ωn+1 − ωα0 = 0 and ωn+1
0 = 0. We consider
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.9. Associated varieties 365

the coefficients of F4 :
rαβγ ω0 = 0, α, β, γ distinct,
β
rααβ ω0 = (ωβ0 − ωn+1 ), α = β,
rααα ω0 = 3(ωα0 − α
ωn+1 ).
We see that for all α = β, we have rαααβ = 0, rααββ = r and rαααα = 3r for
some function r.
Exercise 12.8.5: Show that we may reduce our frame bundle further to a
subbundle where r ≡ 0.

Now that all coefficients of F3 and F4 are zero, we see that F5 and all
higher Fk must be zero as well, as, for example, the coefficients of F5 are
polynomials in the coefficients of II, F3 , F4 , but the coefficients of II only
occur multiplied by the coefficients of F3 . 

Exercise 12.8.6: Show that Fubini’s Theorem still holds for manifolds with
degenerate Gauss maps as long as the rank of the Gauss map is at least two.
In this case the closure of M is a singular quadric hypersurface.
Exercise 12.8.7: Show that, for G(k, W ) ⊂ P(Λk W ) and Seg(Pk × Pl ),
there are natural framings with F3 = 0.

Remark 12.8.8. The generalized cominuscule varieties do admit framings

where the nonfundamental Fubini forms are all zero, and they are the unique
homogeneous varieties having this property (see [136, 135]). There are
examples of inhomogeneous varieties that have this property at their smooth
points, although such varieties are locally homogeneous, e.g., the completion
to a projective variety of a paraboloid given in local coordinates as xμ =
μ α β
qαβ x x .

Exercise 12.8.9: Let X n ⊂ PW be a variety with |II| = S 2 Tx∗ X and such

that there exists a framing in which F3 = 0 for general x ∈ X. Show that
X = v2 (Pn ). 

12.9. Associated varieties

In this section we generalize the notion of dual variety and Gauss image.
Let X n ⊂ PV = Pn+a be a variety. Define the (r, s)-associated variety of X
to be

Zr,s (X)
= {E ∈ G(r + 1, V ) | ∃x ∈ Xsmooth such that dim(T̂x X ∩ E) ≥ s + 1}.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 366 12. Projective geometry II

Note that Zn,n (X) = γ(X) and Zn+a−1,n = X ∨ , and for r ≥ a we have
Zr,0 (X) = G(r + 1, V ). We will see below that Za−1,0 (X) is always a hy-
persurface, and we expect Zs+a−1,s (X) to be a hypersurface. Za−1,0 (X) is
called the associated hypersurface to X in [73] and the varieties Zs+a−1,s (X)
are called higher associated hypersurfaces, although they are not always
hypersurfaces—see below. Let δs (X) = codim Zs+a−1,s (X) − 1 and recall
the notation δ∗ (X) = δn (X).
Proposition 12.9.1. δs (X) = max{0, s + δ∗ (X) − n}.

Proof. Fix index ranges 1 ≤ σ, τ ≤ s, s + 1 ≤ j ≤ n, 1 ≤ λ ≤ a and let

f = (e0 , . . . , en+a ) be a general point of F 1 (X). Write
 = e0 ∧ e1 ∧ . . . ∧ es ∧ en+1 ∧ . . . ∧ en+a−1 .
E
Then E = [E]  is a general point of Zs+a−1,s (X). We calculate the dimension
of Zs+a−1,s (X) by calculating the dimension of its tangent space at a general
point. As before, let E A be the vector with eA removed and eB put in its
B
place in the wedge product. Then
(12.14) dE  0ωj + E
 ≡E  σ ωj + E  n+λ ω j + E
 σ ω n+a + E  n+λ ω n+a mod E.

j 0 j σ n+a σ j n+λ n+a n+λ

The forms ω0j , ωσj , ωn+λ

n+a
are all independent of each other, but ωσn+a =
n+a ω τ + q n+a k
qστ 0 σk ω0 . Zs+a−1,s (X) will be a hypersurface if and only if the ma-
n+a
trix qστ is of full rank. Since this is the restriction of a generic quadric in
the second fundamental form to a generic subspace, we fail to have full rank
if and only if δ∗ (X) (which is n minus the rank of a generic quadric in |II|)
is greater than n − s and the difference, if positive, is exactly s + δ∗ (X) − n.
In particular, we see that Za−1,0 (X) is always a hypersurface. 
Remark 12.9.2. In [73], Zs+a−1,s (X) is denoted by Zn−s+1 (X) and in
[73, Prop. 3.2.11] it is mistakenly asserted that Zn−s+1 (X) is always a
hypersurface.

Given a hypersurface Y ⊂ G(k, V ) and E ∈ Ysmooth , we have PNE∗ Y ∈

P(E⊗(V /E)∗ ). We define the rank of NE∗ Y to be the rank of a general point
of PNE∗ Y . If the rank is one, i.e., PNE∗ Y ∈ Seg(PE ⊗P(V /E)∗ ), then Y is
said to be coisotropic.
Proposition 12.9.3 ([73, 4.3.14]). Let X ⊂ PV be a variety. If Y =
Zs+a−1,s (X) is a hypersurface, then it is coisotropic.

Proof. Taking E as above (and assuming Y is a hypersurface), the only

 is E
vector that does not appear in dE  0 ω n+a , which is of rank one. 
n+a 0

Proposition 12.9.4 ([73, 4.3.14]). The coisotropic hypersurfaces in G(k, V )

are exactly the codimension one associated varieties to varieties X ⊂ P(V ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.10. More on varieties with degenerate Gauss maps 367

Exercise 12.9.5: Prove the proposition by imitating our proof for Gauss
images. Namely, take the natural variety X formed from Z and show that
Z is an associated variety to X. 

12.10. More on varieties with degenerate Gauss maps

Throughout this section, X n ⊂ PV is a variety with degenerate Gauss map
with f dimensional fibers. We now study such varieties in more detail,
taking advantage of the cubic form. We keep the index ranges of §4.3; in
μ μ
particular, we work on F γ ⊂ F 1 where qst = qsj = 0. We have
μ
rstβ =0 ∀μ, s, t, β,
(12.15) μ
rsij ω0j = qjk
μ k
ωs ∀μ, s, i.

Recall that we obtain the second fundamental form at [e0 ] by calculating

de0 ≡ ω0α eα mod e0 , deα ≡ ωαμ eμ mod T̂[e0 ] X, and, writing II = ω0α ωαμ eμ ,
where the ω0α are the coefficients of eα in de0 and the ωαμ are the coefficients
of eμ in deα . We have adapted frames such that [es ] ∈ X. We now restrict
to the open subset of F γ where each [es ] ∈ Xsmooth . We have
des ≡ ej ωsj modes ,
dej ≡ eμ ωjμ modT̂[es ] X.
Combining this with (12.15) we obtain
Proposition 12.10.1. Let x ∈ X and let F denote the fiber of γ passing
through x. Then we may recover the second fundamental form of X at all
y ∈ F ∩ Xsmooth from IIX,x and F3,X,x . More precisely,
IIX,[es ] = ωjμ ωsj (eμ mod T̂[es ] X)⊗es = rsik
μ
ω0i ω0k (eμ mod T̂[es ] X)⊗es ,
where there is no sum on s on the right hand side.

We saw several examples of varieties with degenerate Gauss mappings in

§4.3. It would be nice to be able to identify them in terms of the quantities
we have calculated. We do this in the case of cones.
Characterization of cones over Pf −1 .
Proposition 12.10.2. Assume f < n − 1. Then X is a cone over a variety
with a nondegenerate Gauss map if and only if for all x, y ∈ F ∩ Xsmooth ,
IIy is proportional to IIx as elements of S 2 Tx∗ X ⊗Nx X = S 2 Ty∗ X ⊗Ny X.

Proof. By our hypothesis and Proposition 12.10.1, there exist functions Cs

μ μ
such that rsik = Cs qik for all s, μ, i, k. Under a change of frame es → es +gs0 e0
μ μ μ
we have rsik → rsik + gs0 qik , and thus we may restrict to frames where
μ
rsik = 0. This restriction gives us a well-defined splitting F̂ = {e0 } ⊕ W ,
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 368 12. Projective geometry II

μ
where W = {es }. But rsik = 0 implies ωsj = 0, as for each j there exists a μ
μ α
with qjα ω0 = 0. We now have

des ≡ ωs0 e0 mod W.

To complete the proof in this direction we need to show that ωs0 = 0, as
then PW is a fixed linear space in each tangent space. Consider
0 = dωsj = −ω0j ∧ ωs0 .
Since there is more than one j-index by hypothesis, we conclude ωs0 = 0.
Thus X = J(PW, Y ) where Y is any subvariety of X transverse to the fibers
of X. 
Exercise 12.10.3: Prove the other direction for Proposition 12.10.2.
Remark 12.10.4. We do not know how to characterize the varieties X =
J(PW, Y ) where the Gauss map of Y is not necessarily nondegenerate. A
first guess would be that if there is a linear subspace Ŵ ⊂ T̂[e0 ] F such that
IIy is proportional to II[e0 ] for all y ∈ PŴ ∩Xsmooth , then there is a splitting
Ŵ = {e0 } ⊕ W and X is a cone with vertex PW , i.e., X = J(PW, Y ) for
some variety Y . But additional hypotheses are necessary in this case, as
pointed out in [160].

Characterization of the case where dim γ(X) = 1.

Theorem 12.10.5. If f = n−1, then X is a cone over an osculating variety
to a curve.
Exercise 12.10.6: Prove the theorem. Find a curve, as we did in Theorem
4.3.6 for the n = 2 case, and recover X from this curve.

Further invariants when δ∗ (X) = f . To simplify the calculations in what

follows, we assume there exists a quadric q ∈ |II| with rank(q) = n − f ,
i.e., that the dual defect of X equals f . As mentioned in Remark 12.10.4,
without this assumption the problem is considerably more difficult.
Our assumption implies that we may choose q n+1 to have rank n − f ,
n+1
and we further adapt bases so that qjk = δjk . (Note that under our
assumption q n+1 is a generic quadric, so the forms ωνμ are still independent
and independent of the semi-basic forms.) Equation (12.15) implies
n+1 k
ωsj = rsjk ω0 .
Furthermore, assume e1 is a generic vector in Tx F and, by using the re-
maining SO(n − f ) freedom in the changes of bases among the ej ’s, we
n+1
diagonalize r1jk . Since e1 is a generic vector, the number of distinct eigen-
n+1
values of (r1jk ) is an invariant.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.11. Rank restriction theorems 369

Discussion of the generic case. We now consider the generic case, where all
n+1 n+1
the eigenvalues of (r1jk ) are distinct. Write r1jk = λj δjk and consider

0 = dω1n+2 = ωjn+2 ∧ ω1j

ω0 ∧ (λj ω0j )
n+2 k
= qjk

= n+2
qjk (λj − λk )ω0k ∧ ω0j .
j<k
φ
We obtain n+2
= 0 for j = k and similarly qjk
qjk = 0 for all n+2 ≤ φ ≤ n + a.
This has the geometric interpretation that a generic line in F intersects the
focal hypersurface Φ ⊂ F in n − f distinct points. Recall that the degree of
Φ is n − f , so our situation means that in the defining equation of Φ (which
may be a product of polynomials of lower degree) no factor in the product
is repeated. In the language of algebraic geometry, Φ is reduced (although
not necessarily irreducible).
Theorem 12.10.7. Let X n ⊂ PV be a subvariety with degenerate Gauss
map having f < n − 1 dimensional fibers. Let F be a general fiber. If the
focal hypersurface Φ is reduced, then the first normal space is of dimension
at most n − f . Moreover, either X is a hypersurface in a Pn+1 ⊆ PV , or
IIX,x contains at least one quadric of rank one.

Proof. Our calculation above shows that all quadrics in |IIX,x | can be si-
multaneously diagonalized; thus the dimension of the system of quadrics is
at most n − f . A short calculation shows that the only way the prolongation
of such a system can be nonempty is for there to be a quadric of rank one
in the system. 

Theorem 12.10.7 overlaps a result in [4], where additional hypotheses

are made to insure X is a hypersurface. The essential idea for Theorem
12.10.7 is due to Akivis and Goldberg. For further results, see [4, 160].
In particular, a three-dimensional variety with Gauss map of rank two is
either a join of two curves, or the union of asymptotic or conjugate lines to
a surface, or obtained from the following construction: Take two curves that
are simultaneously parametrized on an open subset, say c(t), e(t). Then take
the union of planes spanned by the line c(t)e(t) and the tangent line to c(t).

12.11. Rank restriction theorems

We already saw that for smooth varieties of small codimension, there are
genericity requirements on the projective second fundamental form at gen-
eral points, namely Theorem 12.4.12. We interpret Theorem 12.4.12 as
saying that the second-order infinitesimal geometry (at a general point) can
“see” some of the global geometry. Or, to put it another way
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 370 12. Projective geometry II

In order for X to be smooth, it must “bend enough”.

In this section we discuss more general rank restriction theorems, giving
a slightly modified version of the rank restriction theorems in [122] and
[123].
We continue the convention that if b = dim Xsing , then b = −1 if X is
smooth.
Theorem 12.11.1 ([122]). Let X n ⊂ Pn+a be a variety with b = dim Xsing .
Let x ∈ X be a general point and let Ξx ⊂ PNx∗ X be a subvariety. Let
h ∈ Ξx be a general point and let r = rank II(h). Then
r ≥ n + dim Ξx − 2(a − 1) − (b + 1).

Proof. Since x ∈ X is a general point, we may extend Ξx to a neighborhood

U of x to give a local fibration Ξ → U over an open subset U ⊂ X with
the property that, as systems of quadrics, the integer-valued invariants of
Ξy are the same as those of Ξx for all y ∈ U . Let IΞ denote the incidence
correspondence with projection maps πΞ , ρΞ and let Z = ρΞ (Ξ) ⊂ PV ∗ :

IΞ = {(y, H) | T̃y X ⊂ H, H ∈ Ξy }.
 S

/ w
S
X Z

Observe that
dim πΞ (ρΞ −1 (H)) = dim ρΞ −1 (H),
dim IΞ = dim X + dim Ξx ,
dim Z = rank dρΞ,(x,H) = dim IΞ − dim ρΞ −1 (H).
Putting these together, we conclude that
rank dρΞ,(x,H) = dim X + dim Ξx − dim πΞ ρΞ −1 (H).

Let FΞ ⊂ F ∗ denote the subbundle where [en+a ] ∈ Ξ. On FΞ , the ωn+λn+a

are no longer necessarily linearly independent or linearly independent of the

semi-basic forms. Using Zak’s theorem on tangencies 12.5.11, with k = a−1,
and observing that dim ρ−1 (H) ≥ dim ρΞ −1 (H), we obtain
(12.16) n+a
dim{ωn+λ , ωαn+a } ≥ n + dim Ξx − ((a − 1) + (b + 1)).
Note that dim{ωαn+a } = r. At worst we have dim{ωn+λ
n+a
} = a − 1, which
yields the theorem. 

Taking Ξx = PNx∗ X, we recover Theorem 12.4.11. Taking Ξx to be a

point, we obtain
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.11. Rank restriction theorems 371

Corollary 12.11.2 ([122]). Let X n ⊂ Pn+a be a variety with b = dim Xsing .

Let x ∈ X be a general point and let h ∈ Nx∗ X. Then
rank II(h) ≥ n − 2(a − 1) − (b + 1).
Corollary 12.11.3 ([122]). Let X n ⊂ Pn+a be a variety with b = dim Xsing .
Let x ∈ X be a general point and assume a < n−(b+1)
2 + 1. Then IIIX,x = 0.
Exercise 12.11.4: Prove Corollary 12.11.3. 

We now refine the rank restriction theorem in the case σ(X) = Pn+a .
For simplicity, assume σ(X) is a hypersurface. (See [123] for the general
case.) In this case we may smoothly project X to Pn+a−1 from a point
not on σ(X) to obtain a variety of codimension a − 1, which gives us an
immediate improvement of the rank restriction theorem, replacing a − 1 by
a − 2 in the inequality (12.16).
We get a further improvement as follows: take Ξx to be the set of h ∈
Nx∗ X such that (q h )sing contains a generic vector. By our assumptions,
Ξx ⊂ PNx∗ X is a hypersurface.
Exercise 12.11.5: Show that Ξx is indeed a hypersurface.

We obtain
r ≥ n − dim{ωn+λ
n+a
mod ωαn+a },
so at least r ≥ n − (a − 1). In §12.12 we will define a subbundle of F 1 on
which dim{ωn+λn+a
mod ωαn+a } ≤ a − 2, which will prove
Theorem 12.11.6 (Rank restriction [123]). Let X n ⊂ Pn+a be a smooth
variety with degenerate tangential variety that is a hypersurface. Let x ∈ X
be a general point and let
Ξx = P{h ∈ Nx∗ X | II(h)sing contains a II-generic vector}.
Let r be the rank of a generic quadric in Ξx . Then r ≥ n − a + 2.

A characterization of Gauss images. Recall that if Y ⊂ G(k, V ) is a

subvariety, then for any E ∈ Y , TE Y ⊂ E ∗ ⊗V /E.
Theorem 12.11.7. Let Y ⊂ G(n + 1, V ) be 3 an r-dimensional subvariety.
Fix a general point E ∈ Y and let F (E) = P( v∈TE Y ker(v)) ⊂ PE. Then
Y is a Gauss image, i.e., Y = γ(X) for some X n ⊂ PV , if and only if
i. dim F (E) = n − r
and
ii. dim( E∈Y F (E)) = n.
In this case X = ( E∈Y F (E)).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 372 12. Projective geometry II

This result was observed independently by Piontkowski and the second

author around 1996.

Proof. It is clear that if Y = γ(X), then (i) and (ii) are satisfied, as then
F (E) is just a general fiber of γ.
Now assume (i) and (ii) are satisfied and let X = E∈Y F (E). Since
X is the transform of the variety Y under an incidence correspondence (the
image under an algebraic map of the inverse image of an algebraic map), X
is a variety. Let 1 ≤ s, t, u ≤ n − r, n − r + 1 ≤ j, k, l ≤ n, and adapt frames
over each x = [e0 ] ∈ Xgeneral such that F = P{e0 , es }, E = {e0 , es , ej }.
The set of pairs {(F (E), PE) | E ∈ Y } gives a subvariety of the flag
variety Fdim F (E),n (V ), and in particular, dF (E) ⊆ E by Exercise 4.2.9.2.
Expressed in frames,
de0 ≡ 0 mod{e0 , es , ej }.
Thus T̂x X ⊆ E. Condition ii implies dim T̂x X = dim E, so we have equality.
Since we are at a general point and X is a variety, we must have equality at
all smooth points of X, and therefore Y = γ(X). 

12.12. Local study of smooth varieties with degenerate

tangential varieties
In this section we prove the assertion necessary to prove Theorem 12.11.6,
and prove that any smooth variety with a critical secant defect resembles a
Severi variety to order two.
Fix a general point x ∈ X of a smooth variety X n ⊂ Pn+a with degen-
erate tangential variety that is a hypersurface.
Fix a II-generic vector v ∈ T . By Exercise 12.5.6.2 we have the following
flag in T :
(12.17) ker IIv ⊂ {v, ker IIv } ⊆ Singloc(Ann(v)) ⊂ T.
Recall that dim(ker IIv ) = δτ (X) and dim(Singloc(Ann(v)) = n − r, where
r is the rank of a generic quadric in |II|. We use the following index ranges
for the rest of this section: 2 ≤ s, t ≤ rank IIv − r, rank IIv − r + 1 ≤ i, j ≤
rank IIv , rank IIv + 1 ≤ , δ ≤ n and 1 ≤ λ ≤ rank IIv + 1. For future
reference, we record the way these spaces will be referred to in indices:
v = e1 ,
ker IIv = {e },
Singloc(Ann(v)) = {e1 , e , es } = {eξ },
T / ker IIv = {e1 , ej , es } = {eλ },
T = {e1 , es , ej , e } = {eα }.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.12. Smooth varieties with degenerate tangential varieties 373

Let H = [en+a ] be the unique hyperplane such that v q n+a = 0. The

vectors II(e1 , e1 ), II(e1 , es ), II(e1 , ej ) form a basis of IIv (T ) ⊂ Nx X. Let
F s → X denote the frame bundle on which the tangent bundle of X is
adapted to the flag (12.17), and furthermore en+λ = II(e1 , eλ ). Our adap-
tations have the effect that
ω1n+a = 0, ω1n+1 = ω01 , ω1n+j = ω0j , ω1n+s = ω0s .
j
We may use motions in the fiber of F s corresponding to shearing by gn+1
j
and gn+s to obtain
n+a
r11j = 0,
(12.18) n+a
r1sj = 0,
and denote the subbundle of F s where these equations hold by F s . Now
consider the equations for the cubic form (12.1) restricted to F s :
n+a β
r11β ω0 = −ωn+1
n+a
,
n+a β
(12.19) r1jβ ω0 = −ωn+j
n+a n+a k
+ qjk ω1 ,
n+a β
r1sβ ω0 = −ωn+s
n+a
.

Our assumption that III v = 0 implies

(12.20) n+a
rξηζ = 0, ∀ξ, η, ζ,
and (12.20), together with (12.18), implies
n+a
ωn+1 = 0,
n+a
(12.21) ωn+s = 0,
n+a
ωn+j = qjk ω1 − r1jk
n+a k n+a k
ω0 .
This implies
n+a
dim{ωn+λ } = dim{ωn+1
n+a n+a
, ωn+j n+a
, ωn+s } ≤ a − 2,
proving Theorem 12.11.6.

We have
n = dim(ker IIv ) + rank(Ann(v)) + 1 + S = δτ + r + 1 + S,
where S = dim Singloc(Ann(v))−(1+dim ker IIv ). Combined with the rank
restriction Theorem 12.11.6, which gives r ≥ n − a + 2, we have
n S
+2+ . a≥
2 2
This implies a ≥ n
2 + 2, proving Zak’s theorem 12.5.1 on linear normality.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 374 12. Projective geometry II

Varieties with critical tangential defects. We say a smooth variety X n ⊂

Pn+a has a critical tangential (or secant) defect if a = n2 + 2 and σ(X) is
degenerate.
We now restrict our study to critical tangential defects. This implies
τ (X) is a hypersurface and S = 0, so there are no es ’s.
We restrict to the subbundle of F s where qij
n+a
= δij .
Consider the following coefficients of F3 :
β n+j n+a
β ω0 = −qk
rkn+a ωn+j + ω k ,
r1n+j β j n+j k
β ω0 = ω + q k ω1 .

Using (12.21), we obtain


ωk ≡ qkn+j ω1j mod{ω0α },
j

ω ≡ −qjn+k ω1j mod{ω0α },

k

which imply
qkn+j + qjn+k = 0 ∀j, k, .
Phrased invariantly, we have shown:
Lemma 12.12.1. Let II ∈ S 2 T ∗ ⊗N be a critical tangential defect. Fix a
generic v ∈ T to obtain a quadratic form on T / Singloc(Ann(v)) which is
well-defined up to scale, and fix its scale. Then the mapping
ker IIv → End(T / Singloc(Ann(v)))
takes image in so(T / Singloc(Ann(v))).
The quadratic form of the lemma is q n+a |T / Singloc(Ann(v)) .
Observe that q n+1 |ker IIv is well-defined because
ker IIv ⊆ Baseloc{q n+2 , . . . , q n+a }.
Consider
n+j β n+j
r11β ω0 = −ωn+1 + 2ω1j ,
r1n+j β j n+j k
β ω0 = ω + q k ω1 ,
which imply
n+j
ωn+1 ≡ 2ω1j mod{ω0α },
ω j ≡ −q n+j
k ω1 mod{ω0 }.
k α

Computing
rn+j β n+1 n+j n+j k n+j k
δβ ω0 = −q δ ωn+1 + q k ωδ + qδk ω ,
and mod-ing out by the semi-basic forms and using (12.12), we obtain
(12.22) q n+j n+k
k qδi
n+j n+k
+ qδk q i = −2q n+1 i
δ δj ∀, δ, j, k, i.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.12. Smooth varieties with degenerate tangential varieties 375

Consider the map

ker IIv → End(T /qsing ),
w e → w qjn+κ (ej )∗ ⊗ek .
The relation (12.22) implies the fundamental Lemma A.7.5 of Clifford alge-
bras applies, thus:
Lemma 12.12.2. Let II ∈ S 2 T ∗⊗N have a critical tangential defect. Then
T / Singloc(Ann(v)) is a Cl(ker IIv , Q)-module.
Keeping in mind that dim ker IIv = a − 3 and dim T / Singloc(Ann(v)) =
a − 2, Exercise A.7.10.4 implies
Corollary 12.12.3. Let X n ⊂ Pn+a be a smooth variety with degenerate
secant variety with a critical defect, so a = n2 +2. Then the only possibilities
are dim(T / Singloc(Ann(v))) = a − 2 = 1, 2, 4, 8.
Examples of these representations have models A ⊕ A; see [91].

We now outline the proof that the second fundamental form must be
|II| = {uu, vv, uv}, i.e., the fundamental form of a Severi variety.

We normalize qjn+k to correspond to the standard multiplication by J ,

where we take the e as standard basis elements of Im(A). We obtain
n+k
qij = 0 by applying Singloc(Ann(ek )) ⊆ Baseloc(Ann(ek )). By using a
n+1 n+1
fiber motion, we may normalize qjβ = 0 and qξη = δξη . This puts the
system in the standard form of a Severi variety. For more details, see [123].

To finish the proof of Zak’s theorem on Severi varieties, we need to show

that any variety with second fundamental form at a general point isomorphic
to the second fundamental form of a Severi variety, and with III v = 0, must
be a Severi variety. The Veronese case is treated in Exercise 12.8.9. For the
other three cases something stronger is true. The general statement is
Theorem 12.12.4 ([131]). Let X n ⊂ CPn+a be a complex submanifold.
Let x ∈ X be a general point. If |II|  |IIZ,z |, where Z is a compact rank
two Hermitian symmetric space (other than a quadric hypersurface) in its
natural embedding, then X = Z.
The general technique of proof is to decompose S 3 T ∗⊗N into irreducible
R-modules, where R ⊂ GL(T ) × GL(N ) is the subgroup preserving II ∈
S 2 T ∗ ⊗ N , and to use the higher-order Bertini theorems. For details, see
[131]. For more general results with an almost calculation-free proof using
Lie algebra cohomology, see [101, 137].
Remark 12.12.5. Zak’s original proof of his theorem on Severi varieties
involved a detailed study on a potential Severi variety X of quadric sections,
i.e., linear sections X ∩ L such that X ∩ L ⊂ L is a quadric hypersurface in
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 376 12. Projective geometry II

L. A proof of Zak’s theorem by Chaput [41] uses elementary properties of

these quadric sections to show immediately that any Severi variety must be
homogeneous.

12.13. Generalized Monge systems

We present a generalization of Fubini’s Theorem 12.8.1 to characterize
smooth intersections of quadrics in small codimension.
Before addressing the question of determining if X is contained in quadric
hypersurfaces, we return to the question: How many derivatives does one
need to take to determine if X is contained in a hyperplane H? If X is a
hypersurface, then two derivatives at a general point are enough by Exer-
cise 4.1.12.2 (if X does not leave its tangent plane to first order at a general
point, it must be equal to its tangent plane). On the other hand, if the
codimension of X is not fixed, there is no fixed number of derivatives that
would guarantee that X is contained in a hyperplane (there are osculating
spaces of higher and higher order as one raises the codimension). The fol-
lowing corollary of the rank restriction theorem implies that if codim(X) is
small, the answer is the same as if X were a hypersurface:
Theorem 12.13.1 ([122]). Let X n ⊂ CPn+a be a variety with a < n−(b+1)
2 +
1 (where b = dim Xsing ). Let x ∈ X be a general point. If a hyperplane H
osculates to order two at x, then X ⊂ H.
Exercise 12.13.2: Prove the theorem. 

Thus in small codimension, two derivatives are sufficient to determine if

X is contained in a hyperplane.
In Exercise 1.7.3.2 we saw the classical Monge equation, a fifth-order
ODE characterizing plane conics. In §12.8 we saw that in higher dimensions,
quadric hypersurfaces are characterized by a third-order PDE (namely F3 =
l ◦ II). Here we derive a generalized Monge system characterizing I2 (X),
the set of quadric hypersurfaces containing X, for varieties X such that at
general points IIIX = 0 and II has no linear syzygies (see §12.7).
A variety X is locally the intersection of quadrics if Nx∗ X is spanned
by the differentials of quadratic equations. Let (x0 , xα , xμ ) be local linear
coordinates on PV adapted to the point x = (0, 0, 0) ∈ X, such that Tx X =
{ ∂x∂α }. Let (e0 , eα , eμ ) denote the dual basis. (Before, we denoted the xA
by eA .)
A basis of S 2 V ∗ is given by ((x0 )2 , xα x0 , xμ x0 , xα xβ , xα xμ , xμ xν ). A
basis of the subspace of quadrics in S 2 V ∗ whose zero locus contains x =
[e0 ] is given by (xα x0 , xμ x0 , xα xβ , xα xμ , xμ xν ). A basis of the subspace of
quadrics in S 2 V ∗ whose zero locus contains x = [e0 ] and tangent to X at x
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.13. Generalized Monge systems 377

(i.e., osculating to order one at x) is given by (xμ x0 , xα xβ , xα xμ , xμ xν ). A

basis of the subspace of quadrics in S 2 V ∗ osculating to order two at x, i.e.,
μ α β
a basis of the kernel of FF2v2 (x),v2 (X) , is (xμ x0 − qαβ x x , xα xμ , xμ xν ).
In order that Nx∗ be spanned by differentials of quadratic polynomials,
it is necessary that
(12.23) {dPx |P ∈ ker FFkv2 (X) } = Nx∗ X
for all k. (We occasionally suppress any reference to X and the basepoint
x in what follows.) For k ≤ 2, (12.23) automatically holds; for k = 3 it will
hold if and only if
μ
(12.24) rαβγ = Sαβγ aμνγ qαβ
ν

for some constants aμνγ ∈ C. Notice that if rαβγμ

= Sαβγ aμνγ qαβ
ν in some

frame, it holds in any choice of frame (with different constants aμνγ ), so the
expression (12.24) has intrinsic meaning. If (12.24) holds, then
μ α β
ker FF3v2 (X) = {xμ x0 − qαβ x x − aμνβ xν xβ , xμ xν }.
Continuing in the same fashion, we uncover the following conditions:
F3μ = 3aμνγ ω0γ II ν ,
(12.25) F4μ = 4aμνα ω0α F3ν + 3bμντ II ν II τ ,
F5μ = 5aμνγ ω0γ F4ν + 10bμντ F3ν II τ ,
where aμνα , bμντ = bμτν ∈ C. Moreover, if there are no linear syzygies among
the quadrics in |II|, as explained in §12.7, then F6 = 0; thus, Nx∗ is spanned
by the differentials of quadrics, and these quadrics are smooth along X, so
they generate I(X). In this case, we will call (12.25) the generalized Monge
system for quadrics.
In summary:
Theorem 12.13.3 ([124]). Let X ⊂ PV = Pn+a be a variety and x ∈ X a
general point. Assume IIIXx = 0 and that there are no linear syzygies in
|II|x . Then
 
a+1
dim{quadrics osculating to order three at x} ≤ a + − 1,
(12.26) 2
dim{quadrics osculating to order four at x} ≤ a − 1.
If the generalized Monge system (12.25) holds, then
I2 (X) = ker FF4v2 (X)x .
Equality occurs in the first (respectively second) line of (12.26) if and only
if the first (resp. second) line of (12.25) holds at x, in which case I(X) is
generated by I2 (X). If the generalized Monge system does not hold, then
I(X) is not generated by quadrics.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 378 12. Projective geometry II

If one assumes appropriate genericity conditions, there exist analogous

Monge equations for Id (X) of order 2d + 1 in small codimension; see [124].

12.14. Complete intersections

The least pathological algebraic varieties are the smooth hypersurfaces. For
example, the dimension is obvious and the degree is simply the degree of
the single polynomial defining X. A class of varieties that share many of
the simple properties of hypersurfaces is the class of complete intersections.
Definition 12.14.1. A variety X n ⊂ Pn+a is a complete intersection if the
ideal of X, I(X), can be generated by a elements.
Context from algebraic geometry. The classical Lefschetz theorem on hy-
perplane sections [84] states that for a smooth complete intersection, most
of the topology is inherited from the ambient projective space. Barth and
Barth-Larsen proved theorems partially generalizing Lefschetz’s results to
arbitrary smooth varieties of small codimension [11]. This led Hartshorne
to raise the question as to whether all smooth varieties of small codimension
must be complete intersections, namely his famous conjecture:
Conjecture 12.14.2 (Hartshorne’s conjecture on complete intersections,
[90]). Let X n ⊂ CPn+a be a smooth variety. If a < n2 , then X is a complete
intersection.
This conjecture has inspired an extraordinary amount of work (see, e.g.,
[156, 139, 162, 59]), although it is as open today as the day Hartshorne
made it. Zak’s theorem on linear normality (which Hartshorne conjec-
tured at the same time) can be viewed as a first-order approximation to
Hartshorne’s conjecture; see [90].
So far, all our studies have been of questions that have been primarily
local in nature. When we dealt with global properties (e.g., smoothness of
a variety), we transformed the problem to a local study (e.g., infinitesimal
study of the dual variety). Hartshorne’s conjecture is a fundamentally global
problem, yet nevertheless we will still work locally.
Singular hypersurfaces. Singular hypersurfaces are the key to understanding
the projective geometry of noncomplete intersections. To explain why, for
simplicity, assume for the moment that X is the intersection of hypersurfaces
of degree d.
Proposition 12.14.3 ([124]). Let X ⊂ PV be a variety such that I(X) is
generated by Id (X) and Id−1 (X) = (0). Then the following are equivalent:
1. X is a complete intersection.
2. Every hypersurface of degree d containing X is smooth at all x ∈
Xsmooth .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 12.14. Complete intersections 379

3. Let x ∈ Xsmooth . Every hypersurface of degree d containing X is

smooth at x.

Proof. Fix x ∈ Xsmooth . We have the surjective map

Id (X) → Nx∗ X,
P → dPx .
Since dim Nx∗ X = a, X is a complete intersection if and only if the map is
injective. 

Thus if X is a complete intersection, any hypersurface that is singular

at any x ∈ Xsmooth cannot contain X. If X is not a hypersurface, there
are always singular hypersurfaces in I(X). The proposition says that the
singularities occur away from X.
For the general case, we use terminology due to L’vovsky [143]:
Definition 12.14.4. Let X ⊂ PV be a variety. Let P ∈ Id (X) and let
Z = ZP ⊂ PV be the corresponding hypersurface. We will say Z triv-
ially contains X if P = l1 P1 + . . . + lm Pm with P1 , . . . , Pm ∈ Id−1 (X) and
l1 , . . . , lm ∈ V ∗ , and otherwise that Z essentially contains X.

Proposition 12.14.5 (A local characterization of complete intersections

[124]). Let X ⊂ PV be a variety. The following are equivalent:
1. X is a complete intersection.
2. Every hypersurface essentially containing X is smooth at all x ∈
Xsmooth .
3. Let x ∈ Xsmooth . Every hypersurface essentially containing X is
smooth at x.
Exercise 12.14.6: Prove the proposition.

Proposition 12.14.5 localizes the study of complete intersections to a

point, and furthermore filters the conormal bundle at that point to enable
us to study one degree at a time. However, to determine if a hypersurface
essentially contains X, one might need to take an arbitrarily high number of
derivatives. To have computable conditions, one could work with osculating
hypersurfaces rather than the hypersurfaces containing X. The advantage
is that one would only need to study a fixed number of derivatives for each
fixed degree of hypersurface; the disadvantage is that one will only obtain
sufficient conditions to be a complete intersection. By the results in §12.7,
at best one could prove that there are no singular hypersurfaces of degree
d osculating to order 2d + 2 at x, and that the first restrictions one could
hope for are at order d + 1.
We now specialize to the case d = 2:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 380 12. Projective geometry II

Intersections of quadrics. Looking at (12.13), we see that ker FF3v2 (X),x is as

small as possible if there are no linear syzygies among the quadrics in IIX,x .
We have the following lemma relating linear syzygies and ranks of
quadrics:
Lemma 12.14.7 ([124]). Let Ap ⊂ S 2 T ∗ be a p-dimensional system of
quadrics on an n-dimensional vector space. Say there is a linear syzygy
l1 Q1 + . . . + lp Qp = 0,
where both li ∈ T ∗ and Qi ∈ A are independent sets of vectors. Then
rank Q ≤ 2p, ∀Q ∈ A.
If one now compares Lemma 12.14.7 with the rank restriction theorem,
one sees that if a < n−(b+1)
3 , then there are no linear syzygies in |II|. Com-
bined with the generalized Monge system, we obtain
Theorem 12.14.8 ([124]). Let X n ⊂ Pn+a be a variety and x ∈ X a general
point. Let b = dim Xsing . (Set b = −1 if X is smooth.) If a < n−(b+1)
3 , then
 
a+1
dim{quadrics osculating to order three at x} ≤ a + − 1,
(12.27) 2
dim{quadrics osculating to order four at x} ≤ a − 1.
Equality occurs in the first (respectively second) line of (12.27) if and only
if the generalized Monge system holds to order three (respectively four) at
x. If the generalized Monge system holds, then X is a complete intersection
of the (a − 1)-dimensional family of quadrics osculating to order four.
Corollary 12.14.9 ([124]). Let X n ⊂ Pn+a be a variety, let x ∈ X be
a general point and let b = dim Xsing . If a < n−(b+1)
3 , then any quadric
osculating to order four at x is smooth at x and any quadric osculating to
order five at x contains X.
Corollary 12.14.10 ([124]). Let X n ⊂ Pn+a be a variety whose ideal is
generated in degree two, and let b = dim Xsing . If a < n−(b+1)
3 , then X is a
complete intersection.
There are other results concluding that a variety must be a complete
intersection, but most results that we are aware of only apply in codimension
two (see, e.g., [59]). As with Corollary 12.14.9, all such results impose
some additional hypotheses. Thus they cannot be taken as evidence for
or against Hartshorne’s conjecture, but just indicate where to look for a
potential counter-example.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/13

Appendix A

Linear Algebra and

Representation Theory

In this appendix we briefly discuss tensors (§A.1), matrix Lie groups (§A.2)
and Lie algebras (§A.4), complex vector spaces and complex structures
(§A.3), the octonions and the exceptional group G2 (§A.5), Clifford alge-
bras and spin groups (§A.7), and outline some rudiments of representation
theory (§A.6).
Unless otherwise noted, V, W are real vector spaces of dimensions n
and m, with bases v1 , . . . , vn and w1 , . . . , wm . We use the index ranges
1 ≤ i, j ≤ n and 1 ≤ s, t ≤ m.

A.1. Dual spaces and tensor products

A map f : V → W is linear if f (v + v  ) = f (v) + f (v  ) and f (kv) = kf (v)
for all v, v  ∈ V and k ∈ R.

Definition A.1.1. The dual space of V , denoted by V ∗ , is the space of all

linear maps f : V → R. It is a vector space under the operations of scalar
multiplication and addition of maps.

Exercises A.1.2:
1. Let αi ∈ V ∗ be defined by αi (vj ) = δji . Show that α1 , . . . , αn is a basis
for V ∗ , called the dual basis to v1 , . . . , vn . In particular, dim V ∗ = n.
2. Define, in a coordinate-free way, an isomorphism V → (V ∗ )∗ . (Note that
these spaces would not necessarily be isomorphic if V were replaced by an
infinite-dimensional vector space.)

381
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 382 A. Linear Algebra and Representation Theory

Geometric Aside. Given V , one can form the associated projective space
PV , which is the space of all lines thorough the origin in V . A nonzero
vector α ∈ V ∗ determines a codimension-one linear subspace ker α ⊂ V , so
that PV ∗ may be interpreted as the space of hyperplanes through the origin
in V .

Definition A.1.3. Let Hom(V, W ) denote the space of linear maps from V
to W . Like V ∗ , this is a vector space under the addition of maps. The space
of linear maps from V to V , called endomorphisms, is denoted by End(V ).
We may think of Hom(V, W ) as the space of W -valued linear functions
on V , and when doing so we denote it by V ∗ ⊗ W , the tensor product of
V ∗ and W . (Taking the tensor product with W = R is trivial because
V ∗ ⊗ R = Hom(V, R) = V ∗ .) Similarly, we define V ⊗ W as the space of
W -valued linear functions on V ∗ .

Given α ∈ V ∗ and w ∈ W , let α ⊗ w denote the element of Hom(V, W )

defined by
v → α(v)w.
We call such an element decomposable.
Exercises A.1.4:
1. Show that the decomposable elements in V ⊗ W span V ⊗ W . More
precisely, show that the nm vectors {vi ⊗ws } span V ⊗W .
2. After having fixed bases, define an explicit isomorphism between V ∗⊗W
and the space of n × m matrices.
3. Show that the decomposable elements V ∗ ⊗ W are exactly those rep-
resented by rank one matrices. More generally, show that the rank of an
element of V ⊗W is well-defined and agrees with the rank of the associated
matrix (with respect to any choices of bases).
4. Given f ∈ Hom(V, W ) we define f t ∈ Hom(W ∗ , V ∗ ), called the transpose
or adjoint of f , by f t (β)(v) = β(f (v)). (If we write basis and dual basis
vectors as columns vectors, the matrix representative of f t is the transpose
of that of f .) Show that the transpose defines a vector space isomorphism
Hom(V, W ) ∼ = Hom(W ∗ , V ∗ ).
Definition A.1.5. Let V1 , . . . , Vk be vector spaces. A function
(A.1) f : V1 × . . . × Vk → W
is multilinear if it is linear with respect to addition and scalar multiplication
in each factor V . We denote this space of multilinear functions by V1∗ ⊗
V2∗ ⊗· · ·⊗Vk∗ ⊗W . In particular, V ∗⊗2 = V ∗ ⊗V ∗ is the space of real-valued
bilinear forms on V , and V ∗⊗3 = V ∗⊗V ∗⊗V ∗ is the space of trilinear forms,
etc.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.1. Dual spaces and tensor products 383

The space V1∗ ⊗ · · · ⊗ Vk∗ is canonically isomorphic to any re-ordering of

the factors and we will make free use of this re-ordering isomorphism without
mention.
For β1 ∈ V1∗ , . . . , βk ∈ Vk∗ , define β1 ⊗· · ·⊗βk ∈ V1∗ ⊗V2∗ ⊗· · ·⊗Vk∗ by
β1 ⊗· · ·⊗βk (u1 , . . . , uk ) = β1 (u1 ) · · · βk (uk )
and call an element of V1∗ ⊗V2∗ ⊗· · ·⊗Vk∗ decomposable if it may be written
in this way.
Exercises A.1.6:
1. Show that (α, v) → α(v) is multilinear from V ∗ × V to R.
2. Show that the space of multilinear functions (A.1) is a vector space, and
determine its dimension.
3. Show that V1∗ ⊗V2∗ ⊗· · ·⊗Vk∗ is spanned by its decomposable vectors. 
4. Given α ∈ V ∗ , β ⊗ W ∗ , let α ⊗ β(v ⊗ w) = α(v)β(w). Show that this
defines an isomorphism V ∗ ⊗W ∗ ∼ = (V ⊗W )∗ . Thus, V ⊗W may be thought
of as the set of linear maps from V ∗ to W , the set of linear maps from W ∗
to V , the set of bilinear maps from V ∗ × W ∗ to R, or as the dual space of
V ∗ ⊗W ∗ .
5. Let V ⊗k denote the k-fold tensor product of V with itself. Show that
this is the dual space of V ∗⊗k .
∗ ∗
Remark A.1.7. One may define the rank of an element
∗
r X ∈ V1 ⊗ V2 ⊗
· · ·⊗Vk to be the minimal number r such that X = u=1 zu with each zu
decomposable. It turns out that the rank is quite subtle for k > 2. In fact
the maximal rank of an element of a triple tensor product is not known even
for low-dimensional vector spaces. Such open questions go under the name
Waring problems. See §12.5 for a geometric generalization.
An open question of importance to computer science is the following: Let
A, B, C be vector spaces, and consider the matrix multiplication operator
M that composes a linear map from A to B with a linear map from B to
C to obtain a linear map from A to C. Letting V1 = A∗ ⊗B, V2 = B ∗ ⊗C,
V3 = A∗⊗C, we have M ∈ V1∗⊗V2∗⊗V3 . Then the open question is, determine
the rank of M . For an overview of this problem, see [133].

Symmetric and skew-symmetric tensors. The tensor product ⊗ is not

symmetric; even in V ⊗V , v1 ⊗v2 = v2 ⊗v1 .
Consider V ⊗2 = V ⊗V with basis {vi ⊗vj |1 ≤ i, j ≤ n}. The subspaces
defined by
S 2 V := span{vi ⊗vj + vj ⊗vi , 1 ≤ i, j ≤ n}
= span{v⊗v |v ∈ V }
= {X ∈ V ⊗V | X(α, β) = X(β, α) ∀α, β ∈ V ∗ }
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 384 A. Linear Algebra and Representation Theory

and

Λ2 V := span{vi ⊗vj − vj ⊗vi , 1 ≤ i, j ≤ n}

= span{v⊗w − w⊗v |v, w ∈ V }
= {X ∈ V ⊗V | X(α, β) = −X(β, α) ∀α, β ∈ V ∗ }
are respectively the spaces of symmetric and skew-symmetric 2-tensors of
V . For arbitrary v1 , v2 ∈ V , we define v1 v2 := 12 (v1 ⊗v2 + v2 ⊗v1 ) ∈ S 2 V and
(A.2) v1 ∧ v2 := v1 ⊗v2 − v2 ⊗v1 ∈ Λ2 V.
(Sometimes a factor of 1
2 is inserted into the definitions of v1 v2 and v1 ∧ v2 .)
More generally,

(A.3) v1 ∧ · · · ∧ vk := (sgn σ)vσ(1) ⊗· · ·⊗vσ(k) ,
σ∈Sk

where Sk denotes the group of permutations on k elements and sgn σ = ±1

is the sign of σ. The product v1 ∧ · · · ∧ vk is called the wedge product of
the vectors v1 , . . . , vk . Let A ∈ End(Λk V ) be defined by linearly extending
v1 ⊗ . . . ⊗ vk → v1 ∧ . . . ∧ vk to combinations of decomposable vectors.
Exercises A.1.8:
1. Show that the two definitions for S 2 V are equivalent, as well as the two
definitions for Λ2 V . Show that
(A.4) V ⊗V = S 2 V ⊕ Λ2 V.
Note that the coordinate-free definition implies that this decomposition is
preserved under linear changes of coordinates.
2. Give similar definitions for the totally symmetric k-tensors S k V ⊂ V ⊗k
and the alternating k-tensors Λk V ⊂ V ⊗k . 
We often think of S k V ∗ as the space of homogeneous polynomials of
degree k on V .
3. Given α ∈ Λi V, β ∈ Λj V , define α ∧ β ∈ Λi+j V by composing the
inclusion Λi V ⊗ Λj V ⊂ V ⊗i+j with the linear mapping A (whose image is
Λi+j V ), divided by the factor i!j!. Show that β ∧ α = (−1) ij α ∧ β.
k n+k−1 k n
4. Show that dim S V = k and dim Λ V = k . In particular,
Λ V  R, Λ V = 0 for l > n, and S V ⊕ Λ3 V = V ⊗3 . 
n l 3

5. In this exercise, we refine the decomposition V ⊗3 = V ⊗ (V ⊗ V ) =

V ⊗S 2 V ⊕ V ⊗Λ2 V induced by (A.4). Let
ρ(v1 ⊗v2 ⊗v3 ) = 13 (v1 ⊗v2 ⊗v3 + v3 ⊗v1 ⊗v2 + v2 ⊗v3 ⊗v1 ),
and extend this to a linear map ρ : V ⊗3 → V ⊗3 .
(a) Show that ρ restricts to be a projection V ⊗S 2 V → S 3 V and let KS ⊂
V ⊗S 2 V be its kernel.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.1. Dual spaces and tensor products 385

(b) Similarly, show that ρ restricts to be a projection V ⊗Λ2 V → Λ3 V and

let KΛ be its kernel.
(c) Conclude that V ⊗3 = S 3 V ⊕ KS ⊕ Λ3 V ⊕ KΛ .
(d) Show that dim KS = dim KΛ .
6. Define
F12 = {[v⊗E] ∈ P(V ⊗Λ2 V ) | [E] ∈ G(2, V ) and v ∧ E = 0},
where G(2, V ), as defined in §1.9, is the Grassmannian of 2-planes through
the origin in V . Show that F12 ⊂ P(KΛ ) and that it admits the geometric
interpretation as the space of flags, l ⊂ E ⊂ V , where l is a line through
the origin and E is a two-plane. More generally, one can define the space
Fa1 ,...,ar of flags E1 ⊂ . . . ⊂ Er ⊂ V such that dim Ej = aj . See §4.2 for
more details.
Lemma A.1.9 (Cartan Lemma). Let v1 , . . . , vk be linearly independent
elements of V and let w1 , . . . , wk be elements of V such that w1 ∧ v1 + · · · +
wk ∧ 
vk = 0. Then there exist scalars hij = hji , 1 ≤ i, j ≤ k, such that
wi = j hij vj .
Exercise A.1.10: Prove the lemma.
Exercise A.1.11: Let v1 , . . . , vn ∈ V be a basis of V , and let R1 , . . . , Rm ∈
Λ2 V be such that Rj ∧ v j = 0. Show that Rj = 12 Rjkl v k ∧ v l for some scalars
Rjkl such that Rjlk = −Rjkl and Rjkl + Rklj + Rljk = 0. What happens
when dim V > n? 

Induced linear maps. Given α ∈ End(V ), define maps α⊗k : V ⊗k → V ⊗k

induced by α as follows. On decomposable elements, let
v1 ⊗v2 ⊗· · ·⊗vk → α(v1 )⊗α(v2 )⊗· · ·⊗α(vk ),
and extend by linearity. Note that α⊗k preserves the subspaces S k V and
Λk V .
In particular, the induced map α⊗n : Λn V → Λn V is multiplication
by some scalar. This number is called the determinant of α and denoted
det(α). Geometrically, if P ⊂ V is a parallelepiped of dimension n with one
vertex the origin, then det(α) = vol(α(P ))/ vol(P ), where vol is any volume
form compatible with the linear structure.
Exercise A.1.12: Show that det(α) equals the determinant of the square
matrix which represents α with respect to a given basis.

Interior products. For x ∈ V , define the interior product x : Λp+1 V ∗ →

Λp V ∗ by
(x φ)(v1 , . . . , vp ) := φ(x, v1 , . . . , vp ), φ ∈ Λp+1 V ∗ .
More generally, for z ∈ Λp V define z : Λq V ∗ → Λq−p V ∗ .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 386 A. Linear Algebra and Representation Theory

Exercise A.1.13: Show that x is the adjoint of the linear map x∧ : Λp V →

Λp+1 V given by wedging with x.

Induced inner products and the ∗-operator. An inner product ,  on

V induces an inner product on V ∗ , as follows: take any orthonormal basis
of V and declare the dual basis to be orthonormal. Alternatively,
 fix an
orthonormal basis e1 , . . . , en for V and define α, β = i α(e i )β(e i ) for
∗
α, β ∈ V .
Exercise A.1.14: Verify that these two definitions of induced inner prod-
ucts are both well-defined and agree.

One may induce inner products on all tensor spaces constructed from V
and V ∗ . For example, the inner product on V ⊗V is given on decomposable
vectors by a⊗b, c ⊗d = a, cb, d and extended bilinearly. However, since
our definition (A.3) for the wedge product produces k! terms, we normalize
the inner product on Λk V by dividing the inner product inherited from
V ⊗k by k!. With this normalization, the inner product satisfies Hadamard’s
inequality

(A.5) |a ∧ b| ≤ |a||b|.

Let Ω ∈ Λn V be a volume form with unit length with respect to this

renormalized inner product. For α ∈ Λk V , we define ∗α ∈ Λn−k V by
requiring that
β ∧ ∗α = α, βΩ ∀β ∈ Λk V.
Exercise A.1.15: If e1 , . . . , en is an orthonormal basis, show that Ω =
e1 ∧ · · · ∧ en has unit length, and calculate ∗ej . When n = 2, are there
vectors v such that ∗v = v?

A.2. Matrix Lie groups

Let GL(V ) ⊂ End(V ) denote the group of all invertible linear maps. Let G
be a Lie group, i.e., a C ∞ -manifold that has a group structure compatible
with its differentiable structure. A (linear) representation of G is a group
homomorphism ρ : G → GL(V ), and the vector space V is called a G-
module. If V is endowed with a basis, we call the image ρ(G) a matrix Lie
group.
A subspace V1 ⊂ V is a G-submodule if ρ(G)V1 ⊆ V1 . A G-module
is said to be irreducible if it has no proper G-submodules. For example,
V ⊗V is not irreducible as a GL(V )-module, since by (A.4) both S 2 V and
Λ2 V are G-submodules. On the other hand, S 2 V and Λ2 V are irreducible
GL(V )-modules.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.2. Matrix Lie groups 387

Suppose ρ : G → GL(V ) and ρ : G → GL(W ) are representations of

G. A linear map α : V → W is said to be a G-module homomorphism if
α(ρ(g)v) = ρ (g)α(v) for all v ∈ V and g ∈ G. Note that the images and
kernels of G-module homomorphisms are G-submodules. Two G-modules
V, W are isomorphic if there exists G-module isomorphism (i.e., a bijective
G-module homomorphism) between them.
Lemma A.2.1 (Schur’s Lemma). Let ρV : G → GL(V ), ρW : G → GL(W )
be two irreducible representations of G, and let f : V → W be a G-module
homomorphism.
If f is nonzero, then f is a G-module isomorphism and is the unique
such isomorphism up to scalar multiple.
Exercises A.2.2:
1. Prove the lemma. 
2. Show that KS , KΛ of Exercise A.1.8.5 are isomorphic GL(V )-modules.
The standard notation for this module is S21 (V ). 

Examples. Let Q ∈ S 2 V ∗ be a quadratic form that is positive definite, i.e.,

Q(v, v) > 0 for all v ∈ V \{0}. We define the following subgroups of GL(V ):

SL(V ) := {g ∈ V ⊗V ∗ | det(g) = 1},

O(V, Q) := {g ∈ V ⊗V ∗ |Q(v, w) = Q(gv, gw) ∀v, w ∈ V },
SO(V, Q) := O(V, Q) ∩ SL(V ).
These are respectively called the special linear group, the orthogonal group,
and the special orthogonal group. We often omit reference to Q when it is
understood.
Exercise A.2.3: Show that when Q is definite, SO(V ) is the connected
component of the identity of O(V ). 

When V = Rn with the standard inner product, then we write SO(n) for
SO(V ). When V = Rn or Cn , we respectively write SL(n, R) or SL(n, C)
(or SLn when we wish to remain ambiguous) for SL(V ).
When n = 2m, a 2-form ω ∈ Λ2 V ∗ is nondegenerate if ω ∧ · · · ∧ ω = 0,
where ω is wedged with itself m times (or, equivalently, if the map v → v ω
is an isomorphism from V to V ∗ ). In this case, ω is called a symplectic form
on V , and we may define the symplectic group
Sp(V, ω) := {g ∈ V ⊗V ∗ | ω(v ∧ w) = ω(gv ∧ gw) ∀v, w ∈ V }.
Since all nondegenerate 2-forms in Λ2 (R2m ) are linearly equivalent, we often
denote this group by Sp(m, R) (see Exercise A.4.9.8).
Exercises A.2.4:
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 388 A. Linear Algebra and Representation Theory

1. Show that all of the above are matrix Lie groups, i.e., that they are
smooth submanifolds of GL(V ), and they are closed under multiplication
and inverses.
2. Show that if we write the matrices of O(V ) with respect to an orthonor-
mal basis, then

O(V ) ∼
= O(n) = {A ∈ Mn×n | tAA = Id}.

3. Every matrix in SO(3) represents a rotation fixing a line through the

origin in R3 . Is the analogous statement true for matrices in SO(4)?
4. A symmetric matrix S is diagonalizable by some A ∈ SO(n), i.e., ASA−1
is diagonal. Show that if S1 , . . . , Sk are pairwise commuting symmetric ma-
trices, then they are simultaneously diagonalizable by an element of SO(n).
5. Show that GL(V ) acts simply transitively on the bases of V , and thus
as a manifold GL(V ) is isomorphic to the space of bases of V .
6. Similarly, show that O(V ) acts simply transitively on the set of orthonor-
mal bases of V . 
7. Let V have inner product Q and let |v| = Q(v, v). Define CO(V, Q) ⊂
GL(V ) (or CO(V ), if Q is understood) to be the linear transformations
preserving angles, where ∠(v, w) := Q(v, w)/|v||w|. Show that

CO(V, Q) ∼
= O(V, Q) × {λ Id | λ > 0}.

A.3. Complex vector spaces and complex structures

A complex structure on a (real) vector space V is a map J ∈ GL(V ) such that
J ◦ J = − Id. For example,√if we consider Cn as the real vector space R2n ,
then multiplication by i = −1 is not multiplication by a scalar, but it is a
linear map whose square is − Id. Define GL(V, J) := {g ∈ GL(V )|Jg = gJ}.
For any vector space V , we can define its complexification VC := V ⊗C =
V ⊕iV , which can be considered as the complex span of the vectors in V . So,
if dim V = n, then VC has dimension n as a vector space over C, while it has
dimension 2n as a real vector space. (We make this distinction in notation
by writing dimC VC = n and dimR VC = 2n.) Any f ∈ End(V ) may be
extended C-linearly to a endomorphism of VC . Note that the characteristic
polynomial of this extension is the same as that of the original, so eigenvalues
come in complex-conjugate pairs. The corresponding eigenvectors are also
conjugate under the complex conjugation that fixes V ⊂ VC .
Exercises A.3.1:
1. If (V, J) is a vector space with a complex structure, show that V must
be even-dimensional. 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.3. Complex vector spaces and complex structures 389

2. Show that J has no eigenvectors as an endomorphism of V , but VC splits

as a direct sum of +i and −i eigenspaces of J. We denote these by V (1,0)
and V (0,1) respectively.
3. Let V = R2n with basis e1 , . . . , en , f1 , . . . , fn and the standard complex
structure J defined by J(ei ) = fi and J(fi ) = −ei . Calculate V (1,0) and
V (0,1) .
4. Show that there exists a linear isomorphism φ : V → Cn such that
φ(Jv) = iφ(v) for all v ∈ V . Thus, all vector spaces with complex structures
are isomorphic to the standard example.
5. Find a decomposition of Λ2 VC into three components invariant under
the induced action of J. These components are denoted Λ(2,0) V , Λ(1,1) V
and Λ(0,2) V . More generally, decompose Λk VC into k + 1 components that
are J-invariant.

Unitary groups and conjugation. Let W be a complex vector space of

(complex) dimension n. The invertible complex-linear endomorphisms of
W form the group GL(W ) ∼ = GL(n, C). A Hermitian form h on W is a
map h : W × W → C which is C-linear in its first argument, and satisfies
h(v, w) = h(w, v). Notice that Re(h) is a symmetric bilinear form on W
(considered as a real vector space), while the imaginary part of h is skew-
symmetric. If Re(h) is an inner product, i.e., Re h(v, v) > 0 for all v = 0,
then h is said to be a Hermitian inner product. In this case we define
U (W, h) := {g ∈ GL(W ) | h(v, w) = h(gv, gw), ∀v, w ∈ W } ∼
= U (n),
SU (W, h) := {g ∈ U (W, h) | detC (g) = 1} ∼
= SU (n),
respectively the unitary and special unitary groups. Note that detC (g) is
the determinant of the n×n matrix representing g with respect to a C-basis.
Exercises A.3.2:
1. Relate detC (g) to the determinant of g when g is considered as an endo-
morphism of the underlying real vector space.
2. Let h(v, w) = v · w for v, w ∈ Cn , where v · u is the standard dot product.
Show that
U (Cn , h) =: U (n) = {U g ∈ GL(n, C) | g −1 = g t }
and deduce that | det g| = 1.

Given a vector space V with both an inner product Q and a complex

structure J, one can construct a Hermitian inner product whose real part is
Q provided that J preserves Q, i.e., Q(Jv, Jw) = Q(v, w) for all v, w ∈ V .
In this situation ω(v, w) := Q(v, Jw) is a symplectic form on V , and we may
define a Hermitian inner product by
h(v, w) = Q(v, w) + iω(v, w).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 390 A. Linear Algebra and Representation Theory

Exercises A.3.3:
1. Show that J preserves Q if and only if J ∈ so(Q), so that J ∈ SO(Q) ∩
so(Q).
2. Show that ω as defined above is also preserved by J.

Note that assuming compatibility, any two of Q, J, ω determine the third.

A.4. Lie algebras

It is usually difficult to explicitly parametrize Lie groups. Fortunately, we
rarely need to do this, but we will often need to write out explicit bases for
various Lie algebras to be defined below.
Let gl(V ) = End(V ) = V ⊗V ∗ . After a choice of basis, we may identify
gl(V ) with the set of n×n matrices. Define a skew-symmetric multiplication
[, ] on gl(V ) by
(A.6) [X, Y ] = XY − Y X,
where XY is the usual matrix multiplication (representing the composition
of linear maps). Expanding out, one can verify the Jacobi identity
(A.7) [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0
for all X, Y, Z ∈ gl(V ).
Definition A.4.1. A Lie algebra is a vector space g equipped with a skew-
symmetric bilinear map [, ] : g × g → g, called a bracket, that satisfies the
Jacobi identity (A.7).
Exercise A.4.2: Show that R3 , with bracket given by the cross product, is
a Lie algebra.
Example A.4.3. An important class of Lie algebras is Γ(T M ), the space
of smooth vector fields on a C ∞ manifold M . The bracket is [X, Y ]f :=
X(Y f )−Y (Xf ), and the algebra is infinite-dimensional (see Definition B.1).

One can make any vector space into a Lie algebra by using a bracket
which is identically zero. Such Lie algebras are called abelian Lie algebras.
Inside a Lie algebra, any subspace that is closed under the bracket is called
a subalgebra. It is a Lie algebra in its own right, since the Jacobi identity
holds by restriction.
A representation of a Lie algebra g is a linear transformation ρ : g →
gl(V ) which preserves the brackets. We say such a V is a g-module. Define
a representation ad : g → End(g) by ad(X)Y := [X, Y ], called the adjoint
representation of g. In particular, any finite-dimensional Lie algebra may
be realized as a matrix Lie algebra.
Exercise A.4.4: Show that ad is indeed a representation.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.4. Lie algebras 391

Given a matrix Lie algebra g ⊆ gl(V ), define a map expV : g → GL(V )

by expV (X) := IdV + ∞ j=1 j! X , where X is the endomorphism X ∈
1 j j

End(V ) composed with itself j times. The image is a Lie subgroup of GL(V )
(see, e.g., [174, Chap. 10] or [65, §8.3]). When V = g we drop V from the
notation, and denote the corresponding group G, the adjoint form of the Lie
group associated to g. Different representations give rise to different groups,
but all the groups will have the same universal covering group (see, e.g.
[65, §7.3, §23.1]). One can also go in the other direction, as the following
example shows:

Example A.4.5. Let G be a Lie group. A vector field X ∈ Γ(T G) is

left-invariant if La∗ (Xb ) = Xab for all a, b ∈ G, where La∗ denotes pushfor-
ward (see Appendix B) by left-multiplication by a. The reader can verify
that the Lie bracket of two left-invariant vector fields is also left-invariant.
Thus ΓL (T G), the space of left-invariant vector fields, is a Lie subalgebra of
Γ(T G).
A left-invariant vector field is determined by its value at just one point
(say, at the identity element e ∈ G), since it is given at all other points by
pushforward under left-multiplication. Thus, we may identify ΓL (T G) with
Te G. We define g = Te G ∼ = ΓL (T G) to be the Lie algebra of G.

If G ⊆ GL(V ) is a matrix Lie group, then g ∼

= TId G ⊂ gl(V ) = End(V )
is a matrix Lie algebra. Any Lie group G has a canonical representation
Ad : G → GL(g) defined by

Ad(g)X = Lg∗ Rg−1 ∗ X,

called the adjoint representation. For matrix Lie groups, Ad(g)X = gXg −1 .
Given any representation of a Lie group G, one obtains a representation
of the corresponding Lie algebra g by differentiation. In particular Ad : G →
GL(g) gives rise to a representation g → End(g).
Exercise A.4.6: Let g(t) ∈ G be a curve with g(0) = Id and g  (0) = Y ∈
d
g. Show that dt t=0
Ad(g(t))X = [Y, X]. In particular the representation
obtained by differentiating Ad is ad.

Remark A.4.7. Where does the Jacobi identity (A.7) come from? One can
interpret it as a Leibniz rule. For, if A is an algebra with product indicated
by a dot, then a map D : A → A is a derivation if D(a·b) = D(a)·b+a·D(b)
for all a, b ∈ A. Now take A = g with the product given by the bracket and
D = DX by the bracket with X ∈ g.

Exercise A.4.8: Verify that adX : Y → [X, Y ] is a derivation exactly

because the Jacobi identity holds.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 392 A. Linear Algebra and Representation Theory

If G ⊆ GL(V ) is a matrix Lie group, and g ⊂ gl(V ) is the associated

matrix Lie algebra, then g has an induced representation on the dual spaces
and tensor products constructed from V . For example, for X ∈ g and v ∈ V ,
then
ρV⊗V (X)(v⊗w) := X(v)⊗w + v⊗X(w).
For another example, for α ∈ V ∗ , ρV ∗ (X)(α) := −α ◦ X, where the circle
denotes the composition of the maps X : V → V and α : V → R.
Exercises A.4.9:
1. Consider
(A.8) sl(V ) = {X ∈ gl(V )| tr X = 0},
(A.9) so(V, Q) = {X ∈ gl(V )|Q(Xv, w) = −Q(v, Xw)},
(A.10) gl(V, J) = {X ∈ gl(V )|X ◦ J − J ◦ X = 0}.
Verify that each of these are Lie algebras, i.e., they are closed under the
bracket (A.6). Show these are isomorphic to the Lie algebras associated to
G = SL(V ), G = SO(V, Q) and G = GL(V, J) respectively.
When V is Rn with its standard basis, we write sl(V ) = sl(n, R) or sln ,
and similarly if Q is the standard quadratic form represented by the identity
matrix we write so(V, Q) = so(n, R) or so(n).
If V is a complex vector space, we similarly write sl(n, C), so(n, C) etc.
and similarly for the corresponding Lie groups.
2. Define a surjective group homomorphism SL(2, C) → SO(3, C) by ob-
serving that sl(2, C) ∼= C3 as vector spaces, and the determinant of a 2 × 2
matrix polarizes to a quadratic form. What is the kernel of this map?
3. Show that if X ∈ so(2n + 1), then det(X) = 0. Show that if X ∈
so(2n), then the eigenvalues of X are purely imaginary and come in complex
conjugate pairs, and thus det(X) is nonnegative.
4. Show that when dim V is of dimension 2n, there is a smooth positive
function Pfaff(X), called the Pfaffian, such that det(X) = Pfaff(X)2 . If xij
are the entries of X, then
1 
Pfaff(X) = n (sgn σ)xσ(1)σ(2)xσ(3)σ(4) · · · xσ(2n−1)σ(2n).
2 n!
σ∈S2n

5. A quadratic form Q ∈ S 2 V ∗ is nondegenerate if the map v → v Q is an

isomorphism V → V ∗ . Recall that Q is represented with respect to a basis
v1 , . . . , vn by a symmetric matrix with entries Qij = Q(vi , vj ). We say Q
has signature (p, q) if its matrix has p positive eigenvalues and q negative
eigenvalues. If Q is nondegenerate, show that there is a basis such that Q
is represented by
 
Idp 0
,
0 − Idq
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.4. Lie algebras 393

where Idk denotes the k × k identity matrix. We write SO(p, q) and so(p, q)
for the corresponding orthogonal groups and Lie algebras.
Show that if Q has signature (n, 0), and we choose bases such that
Q = Idn , then so(V, Q) ∼
= {X | X + X t = 0}.
6. If Q has signature (3, 1), we could take bases such that
⎛ ⎞ ⎛ ⎞
−1 0 0 0 1
⎜ 1 ⎟ ⎜0 1 0 0⎟
Q=⎜ ⎝
⎟
⎠ or Q=⎜ ⎝0 0 1 0⎠ .
⎟
1
1 1 0 0 0
If we use the first basis, we have
⎧⎛ ⎞ ⎫
⎪
⎪ 0 x21 x31 x41 ⎪
⎪
⎨⎜ 2 3 −x4 ⎟ ⎬
⎜ x 0 −x ⎟
so(3, 1)  ⎝ 3 1 2 2 | x 2 3
, x , ... ∈ R .
⎪
⎪ x1 x32 0 −x43 ⎠ 1 1 ⎪
⎪
⎩ ⎭
x41 x42 x43 0

Find the matrix presentation of so(3, 1) using the second choice for Q.
7. Similarly, given a complex structure on R4 , two choices of bases yield
⎛ ⎞ ⎛ ⎞
0 −1 0 0 0 0 −1 0
⎜1 0 0 0 ⎟ ⎜0 0 0 −1⎟
J =⎜ ⎝0 0 0 −1⎠
⎟ and J = ⎜
⎝1 0 0
⎟.
0⎠
0 0 1 0 0 1 0 0

Find the matrix presentation of gl(2, C) ⊂ gl(4, R) using these two choices
for J.
8. Two choices of symplectic form on V  R2n are
ω = dx1 ∧ dxn+1 + · · · + dxn ∧ dx2n ,
ω = dx1 ∧ dx2 + · · · + dx2n−1 ∧ dx2n .

Find the two corresponding matrix presentations of sp(n, R).

9. The symmetric bilinear form B(X, Y ) := tr(adX ◦ adY ) on a Lie algebra
is called the Killing form. Compute the signature of the Killing forms for
so(4), so(3, 1), gl(2, C) and sp(2, R).

Example A.4.10. Let su(2) denote the Lie algebra of SU (2). Taking
e1 , e2 ∈ C2 as a unitary basis gives
 
1 ix −y + iz 
su(2) =
2 y + iz −ix  x, y, z ∈ R ⊂ gl(2, C).
√
Let S 4 (C2 ) have basis v1 = e41 , v2 = 2e31 e2 , v3 = 6e21 e22 , v4 = 2e1 e32 , v5 =
e42 . Then the induced representation of gl(2, C) on S 4 (C2 ) restricts to give
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 394 A. Linear Algebra and Representation Theory

the following representation of su(2):

 
1 ix −y + iz
2 y + iz −ix
⎛ ⎞
2ix −y + iz 0 0 0
⎜ ⎟
⎜y + iz ix 3
2 (−y + iz) 0 0 ⎟
⎜ ⎟
⎜ ⎟
→ ⎜ 0 3
2 (y + iz) 0 3
(−y + iz) 0 ⎟.
⎜ 2 ⎟
⎜ ⎟
⎝ 0 0 3
2 (y + iz) −ix −y + iz ⎠
0 0 0 y + iz −2ix
Exercise A.4.11: Take an orthonormal basis for R3 , and compute the in-
duced orthonormal basis of V = S 2 (R3 ). Write out the matrix for the
representation so(3) → gl(V ) resulting from the induced representation of
gl(3, R).

A.5. Division algebras and the simple group G2

In addition to the orthogonal and symplectic groups, there is just one more
series of groups that can be defined by preserving a generic tensor of some
type on V . The reason is that most tensor spaces have dimension greater
than that of GL(V ). For example, S 3 V has dimension (n+2)(n+1)n
6 > n2 ,
where n = dim V , so the subgroup G ⊂ GL(V ) preserving a generic element
of S 3 V must be zero-dimensional, and similarly for higher symmetric powers.
In fact, the only potential examples are Λ3 V for dimensions n = 6, 7 or
8. (When n = 5, Λ3 R5 ∼ = Λ2 R5∗ , so one gets nothing new.) We expect the
corresponding groups to have dimension 16, 14 and 8 respectively.
In the case n = 8, note that V = sl(3) (over either R or C) has dimension
eight and there is a natural 3-form induced by the bracket [, ] : V × V → V .
Identifying V with V ∗ via the Killing form, we obtain an element φ ∈ V ∗⊗3 .
Exercise A.5.1: Show that φ ∈ Λ3 V ∗ .

By its invariant definition, φ is preserved by SL(3), and a little more

work shows that φ is generic. By counting dimensions, we see that sl(3)
coincides with der(φ), the Lie algebra of Aut(φ) = G.

In the case n = 7, first note that a generic φ ∈ Λ3 V ∗ and a volume form

Ω ∈ Λ7 V ∗ determine a bilinear form β(v, w) defined by
(v φ) ∧ (w φ) ∧ φ = β(v, w)Ω.
Exercise A.5.2: Show that β is symmetric, and nondegenerate, so we re-
name it Q = β. Thus G(φ) ⊆ CO(V, Q). (We obtain a conformal group,
because Q is a priori only well-defined up to the choice of volume form.) 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.5. Division algebras and the simple group G2 395

Over the reals, two types of signature are possible, one of which is def-
inite. Say we are in the case where the signature is definite; then, by com-
paring homogeneity, one can see that Q is in fact well-defined. A little more
calculation (see [91, Thm. 6.80]) shows that one obtains a new simple Lie
group of dimension 14, which is named G2 .

G2 and the octonions. The compact form of the group G2 , corresponding

to Q definite, is also the automorphism group of the octonions. (According
to [91], its presentation as the group preserving a generic 3-form in seven
variables was discovered by Bryant, long after other presentations.)
There are only four normed division algebras over R: R itself, the com-
plex numbers C, the quaternions H, and the octonions O (see, e.g., [91]).
To understand the octonions, we need to review the quaternions. Recall
that H is a normed division algebra that is a four-dimensional vector space
over R. Elements of H may be written as x = x0 + x1 e1 + x2 e2 + x3 e3 where
xj ∈ R and the symbols ej satisfy the multiplication rule ej 2 = −1 and
either of the following equivalent multiplication rules in Figure 1.

e1 e2

e1 e2 e3 e3
Figure 1. Products are positive if one multiplies with the arrow (e.g.,
e1 e2 = e3 ), negative against (e.g. e3 e2 = −e1 )

The octonions (also known as Cayley numbers) form an eight-

dimensional vector space over R, in which elements may be written as
x = x0 + x1 e1 + · · · + x7 e7 with xj ∈ R and the symbols ej satisfying
ej 2 = −1 and the multiplication rules in Figure 2.
Suppose A is one of the four normed division algebras. Define its auto-
morphism group
Aut(A) := {g ∈ GL(A) | (gu)(gv) = g(uv) ∀u, v ∈ A}.
Then Aut(A) is respectively {Id}, Z2 , SO(3, R), or the compact form of G2 .
To partially see the last assertion, take V = ImO = {e1 , . . . , e7 } and
define the form φ in terms of octonionic multiplication:
(A.11) φ(x, y, z) = − 12 Re[x(yz) − z(yx)], x, y, z ∈ ImO.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 396 A. Linear Algebra and Representation Theory

e3

e2 e4
e7

e1 e5
e6

Figure 2. Multiplication rules for the basis for ImO

The split form of G2 , corresponding to Q of signature (4, 3), is the au-

tomorphism group of the split octonions [91], and also occurs as the auto-
morphism group of a very important Pfaffian system on a 5-manifold [36].
(The action on the 5-manifold is not linear; the lowest-dimensional linear
representation of G2 has dimension seven.)
Exercises A.5.3:
1. Prove the Moufang identities: for a, b, c ∈ O,
((ab)a)c = a(b(ac)),
(A.12) c(a(ba)) = ((ca)b)a,
(ab)(ca) = a(bc)a.

2. Use the Moufang identities to verify that φ(x, y, z) of (A.11) is skew-

symmetric in x, y, z.
3. Similarly, verify that the associator

[a, b, c] := 12 [(ab)c − a(bc)], a, b, c ∈ O,

is skew-symmetric in a, b, c.
4. Calculate φ in terms of coordinates x1 , . . . , x7 on R7 = ImO. 
5. Show that the Lie algebra g2 = der(φ) has a matrix presentation
 
κ −tA 
(A.13) g2 =  κ, τ ∈ so(3) ,
A ρ(κ ⊕ τ ) 
where ρ is the isomorphism so(3) ⊕ so(3) ∼ = so(4) and a13 = a31 + a42 ,
a23 = a41 − a32 , a33 = a22 − a11 , and a43 = −a21 − a12 .
6. Determine the classical group(s) preserving a generic φ ∈ Λ3 R6 . (Over
C there is a unique generic 3-form up to GL(6, C)-equivalence, but over R
there can be several.)
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.6. A smidgen of representation theory 397

A.6. A smidgen of representation theory

This section contains a brief overview of the rudiments of representation
theory. For proofs and more complete statements of what follows, see any
of the standard texts such as [19, 65, 98, 113].
Why representation theory? In general, when a group G (or Lie algebra g)
acts on a vector space V , one would like to know the decomposition of V
into irreducible G-modules, if such a decomposition exists. The utility of
having such decompositions will become obvious as you read through this
book. For now, consider the following motivating examples:

Example A.6.1. Let (M n , g) be a Riemannian manifold. Say c : M → R+

is a smooth function and consider the new Riemannian metric cg. How does
the curvature tensor (see Chapter 3) change under such a change of metric?
Is any aspect of it unchanged? Since the group of rotations which preserve
the metric pointwise is also unchanged, one might suspect that the answer
involves the action of these rotations on the curvature.
Representation theory helps us split up the curvature tensor into irre-
ducible pieces and see how each piece changes. In particular, in §11.1 we
will show that there is a piece that doesn’t change. This was first discovered
by H. Weyl when he was investigating Einstein’s theory of general relativity.
Weyl’s study was what motivated him to make his fundamental contribu-
tions to representation theory (see [93] for more details).

Example A.6.2. Consider the rational normal curve Cd ⊂ Pd = P(S d (C2 )),
which is the image of CP1 under [x, y] → [xd , xd−1 y, xd−2 y 2 , . . . , y d ], or,
without coordinates, [v] → [v d ] for v ∈ C2 \{0} (see Chapter 4). What are
the linear changes of coordinates in Cd+1 that leave Cd invariant? The
set of all such changes forms a subgroup of GL(d + 1, C) which is iso-
morphic to GL(2, C). The action may be seen explicitly by g.[v1 · · · vd ] =
[(gv1 ) · · · (gvd )].

Exercise A.6.3: Show that the representation ρ : GL(2, C) → GL(d + 1, C)

described above is irreducible. 

We will outline how to understand representations in one easy case,

when the Lie algebra of G is simple. A Lie algebra g (or a Lie group G) is
called reductive if all g-modules decompose into a direct sum of irreducible
g-modules, simple if it has no nontrivial ideals, and semi-simple if it is the
direct sum of simple Lie algebras. Semi-simple Lie algebras are reductive,
but not all reductive Lie algebras are semi-simple. For example, sl(V ) is
simple while gl(V ) = sl(V ) ⊕ {λ Id} is reductive. However, it turns out
that all reductive Lie algebras are either semi-simple or a direct sum of a
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 398 A. Linear Algebra and Representation Theory

semi-simple and an abelian Lie algebra. (The Killing form of a Lie algebra
is nondegenerate if and only if it is semi-simple; see, e.g., [98, §5.1].)
We discuss irreducible representations of simple Lie algebras. The irre-
ducible representations of a semi-simple Lie algebra are just the tensor prod-
ucts of the irreducible representations of its simple components. Thanks to
the action of g on itself, without loss of generality, we may assume that g is
a matrix Lie algebra, i.e., g ⊆ gl(V ).
Simple Lie algebras are best studied via a certain kind of abelian sub-
algebra they contain. We decompose a given g-module V first with respect
to commuting matrices A1 , . . . , Ar that span such a subalgebra.
Exercise A.6.4: Let A1 , . . . , Ar be n × n matrices that commute. Show
that if each Aj is diagonalizable, then A1 , . . . , Ar are simultaneously diago-
nalizable. 

If a matrix A is diagonalizable, then V decomposes into eigenspaces for

A and there is an eigenvalue associated to each eigenspace. Now let t =
{A1 , . . . , Ar } ⊂ g be the subspace spanned by simultaneously diagonalizable
A1 , . . . , Ar . Then V decomposes into simultaneous eigenspaces for all A ∈ t.
For each eigenspace Vj define a function λj : t → R such that λj (A) is the
eigenvalue of A associated to the eigenspace Vj . Note that λj is a linear
map, so we may think of λj ∈ t∗ .
If there are p distinct eigenspaces of V , then these λj give p elements
of t∗ which are called the weights of V . The dimension
! of Vj is called the
multiplicity of λj in V . The decomposition V = j Vj is called the weight
space decomposition of V .
Exercise A.6.5: Show that the only irreducible representations of an abel-
ian Lie algebra t are one-dimensional. 

Now, back to our simple Lie algebra g. There always exists a maximal
simultaneously diagonalizable (and thus abelian) subalgebra t ⊂ g, unique
up to conjugation by G, called a maximal torus. We define the rank of g to be
the dimension of a maximal torus. Amazingly, irreducible representations
of g are completely determined up to equivalence by how t acts. More
precisely, suppose V is an irreducible g-module. Then as a t-module V
admits a weight space decomposition. If two irreducible g modules V, W
have the same weights (with the same multiplicities) as t-modules, then
they are isomorphic as g-modules.
Let g act on itself by the adjoint action. Then as a t-module, we have
the weight space decomposition
4
g= gα ,
α∈R
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.6. A smidgen of representation theory 399

where R ⊂ t∗ is some finite subset whose nonzero members are called the
roots of g, and gα is the eigenspace associated to each root. (Of course,
g0 = t since t is maximal.) Another amazing fact is that the eigenspaces gα
for α = 0 are all one-dimensional.
Remark A.6.6 (Why the word “root”?). Consider the adjoint representa-
tion ad : g → End(g) restricted to t. The roots of the characteristic polyno-
mial pλ (X) = det(ad(X) − λ Idg ) are the eigenvalues of X. By varying X
one obtains linear functions on t which are the roots of g.
Exercise A.6.7: In fact, weights were originally called “generalized roots”
when the theory was being developed. They too are roots of a characteristic
polynomial—which one?

Let’s look at some of our favorite simple Lie algebras:

Example A.6.8 (g = sl(n)). Here the rank of g is n − 1, and we may take
⎧ ⎛ 1 ⎞ ⎫
⎪
⎨ x ⎪
⎬
⎜ . ⎟
⎠ | x +··· + x = 0 .
1 n 1 n
t = T (x , . . . , x ) = ⎝ ..
⎪
⎩ ⎪
⎭
xn
Let eij = v i ⊗vj . Then we have T (x)(eij ) = (xi − xj )eij for i = j (no sums
here). So the roots are xi − xj and gxi −xj = {eij }.
Example A.6.9 (g = so(2n)). The rank is n, and we may take
⎛ ⎞
0 x1
⎜−x1 0 ⎟
⎜ ⎟
⎜ .. ⎟
t(x1 , . . . , xn ) = ⎜ . ⎟.
⎜ ⎟
⎝ 0 x ⎠
n

−xn 0
Exercise A.6.10: Determine the roots and the root space decomposition
for so(2n).
Remark A.6.11. The classification of complex simple Lie algebras g (due
to Killing and Cartan) is based on classifying the possible root systems, the
collection of roots for g. The rules for such systems are rather strict and
concise. For the record, we need a subset R ⊂ t∗ such that
(1) R must span t∗ , and moreover if g∗ is complex, R lies in an R-linear
(in fact Q-linear) subspace;
(2) for each α ∈ R, reflection in the hyperplane perpendicular to α
must map R to R;
(3) for all α, β ∈ R, the quantity 2β, α/α, α must be an integer;
and
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 400 A. Linear Algebra and Representation Theory

(4) for all α ∈ R, 2α ∈

/ R.
(The inner product  ,  on t is minus one times the restriction of the Killing
form.) The above holds when g is semi-simple. When g is simple, R also can-
not be split up into two separate root systems in complementary subspaces.
The classification over the reals involves a modification of the above: for
each complex simple Lie algebra, there can be several different real forms.

If g is a simple algebra, then to each irreducible representation of g

is associated a set of weights (points in t∗ ), each with some multiplicity.
Not every set of points with multiplicity in t∗ corresponds to an irreducible
representation. In the first place, the admissible points lie on a lattice,
called the weight lattice. The weight lattice is the set of  ∈ t∗ such that
, α  ∈ Z for all α ∈ LR , where LR is the lattice generated by the co-roots
α = α,α
2
α for α ∈ R.
If one fixes an appropriate order on the weight lattice, the highest weight
of an irreducible representation (which necessarily has multiplicity one) de-
termines all other weights, along with their multiplicities. (In fact, the
other weights are obtained by translating the highest weight by the negative
roots.) A vector in V is called a weight vector if it is in an eigenspace for the
torus and a highest weight vector if the corresponding weight determining
the eigenvalues is a highest weight.
The weight lattice has r = dim t∗ generators, so once one fixes a set
of generators ω1 , . . . , ωr , the irreducible representations correspond to r-
tuples of nonnegative integers (1 , . . . , r ) which determine a highest weight
 = 1 ω1 + · · · + r ωr .
Example A.6.12. When g = sl(n) = sl(V ), the weight lattice is generated
by x1 , x1 + x2 , . . . , x1 + · · · + xn−1 . These are the highest weights of the
irreducible representations V, Λ2 V, . . . , Λn−1 V respectively.
Exercise A.6.13: Verify the last statement by computing the action of t
on the highest weight vector e1 ∧ · · · ∧ ek of Λk V .
Remark A.6.14. Even when G is not simple, studying weights of a repre-
sentation (i.e., weights under the maximal torus of a maximal semi-simple
subgroup of G) can yield useful information; see §9.8.

A.7. Clifford algebras and spin groups

Let V be a real or complex vector space with a nondegenerate quadratic
form Q ∈ S 2 V ∗ . Given any linear subspace L ⊂ V , one can define its Q-
orthogonal complement L⊥ = {w ∈ V | Q(v, w) = 0 ∀v ∈ L}. If Q|L is
nondegenerate, then V = L ⊕ L⊥ . In that case, for all v ∈ V we may write
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.7. Clifford algebras and spin groups 401

v = v1 + v2 for v1 ∈ L, v2 ∈ L⊥ , and we define the reflection of v in L by

reflL (v) = v1 − v2 .

L
v2 v
v1

refl (v)
L

Recall that O(V, Q) is the subgroup of GL(V ) preserving Q, and

SO(V, Q) = O(V, Q) ∩ SL(V ).
Theorem A.7.1 (Cartan-Dieudonné [91]). The group O(V, Q) is generated
by reflections in lines, and SO(V, Q) is generated by compositions of even
numbers of reflections. More precisely, O(V, Q) = {refll1 ◦ · · · ◦ refllk | lj ∈
PV } where we may assume k ≤ n, and similarly for SO(V, Q) except k must
be even.
To define Spin(V, Q), the connected and simply connected group with
Lie algebra so(V, Q), we will need to generalize the notion of a reflection.
!
We first remark that V ⊗ := ∞ j=0 V
⊗j has a natural structure as a graded

algebra, called the tensor algebra of V . Here we use the convention V ⊗0 = R.

Earlier we defined Λk V ⊂ V ⊗k . It may also be defined as a quotient space.
More generally, define the exterior algebra of V by Λ• V := V ⊗/x⊗y+y⊗x,
where x ⊗ y + y ⊗ x denotes the ideal generated by expressions of the form
x ⊗ y + y⊗x with x, y ∈ V .
Note that Q induces a quadratic form on Λ• V which we also denote by Q.
The exterior product in Λ• V , (x, y) → x ∧ y, may be interpreted as follows:
Let Ĝ(i, V ) ⊂ Λi V denote the cone over the Grassmannian. If x ∈ Ĝ(i, V ),
y ∈ Ĝ(j, V ), then x ∧ y ∈ Ĝ(i + j, V ) ⊂ Λi+j V represents the (i + j)-plane
spanned by x and y in the following sense: If x ∈ V and y ∈ Ĝ(j, V ),
then x ∧ y is analogous to the component of x in y ⊥ . If ||y||Q = 1, then
||x ∧ y||Q = || projy⊥ (x)||Q . (Note that we do not need Q to define x ∧ y.)
Let x y be defined to be the Q-adjoint of x ∧ y, that is, Q(x y, z) =
Q(y, x∧z) for all x, y, z ∈ Λ• V . For example, if x, y ∈ V , then x y = Q(x, y).
If x ∈ V and y ∈ Ĝ(j, V ), then x y is analogous to the component of x in
y, in the sense that if y has unit length, then ||x y||Q = || projy (x)||Q .
Exercise A.7.2: Show that x (x y) = 0 for all x, y ∈ Λ• V .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 402 A. Linear Algebra and Representation Theory

Consider, for x, y ∈ Λ• V , the bilinear operation defined by

x©y := x ∧ y − x y,
which can be thought of as the “generalized reflection” of x in y.
Definition A.7.3. Let V be a vector space with a quadratic form Q. The
Clifford algebra of (V, Q) is Cl(V, Q) := (Λ• V, ©).
Exercise A.7.4: Show that we have an isomorphism of algebras
Cl(V, Q) ∼
= V ⊗ /x⊗y + y⊗x − 2Q(x, y).
Lemma A.7.5 (Fundamental Lemma of Clifford algebras). Let V be a vec-
tor space with a quadratic form Q and let A be an associative algebra with
unit. If φ : V → A is a mapping such that for all x, y ∈ V
φ(x)φ(y) + φ(y)φ(x) = 2Q(x, y) IdA ,
then φ has a unique extension to an algebra homomorphism φ̂ : Cl(V, Q) →
A.
For a proof, see, e.g., [91].
Exercise A.7.6: Show that the hypotheses of the lemma are equivalent to
φ satisfying φ(x)2 = 2||x||2Q IdA for all x ∈ V .

In Cl(V, Q) = (Λ• V, ©), the degree of a form is no longer well-defined,

but there is still a notion of parity. Let Cleven (V, Q), Clodd (V, Q) ⊂ Cl(V, Q)
denote the corresponding even and odd subspaces. We have the canonical
isomorphisms of Z2 -graded vector spaces (but not algebras) Cleven (V, Q) ∼ =
Λeven V , Clodd (V, Q) ∼
= Λodd V .
Exercise A.7.7: Verify that the parity is well-defined.
Definition A.7.8. Let Cl∗ (V, Q) ⊂ Cl(V, Q) denote the invertible elements.
Let
Pin(V, Q) := {a ∈ Cl∗ (V, Q) | a = u1 © · · · ©ur , uj ∈ V, Q(uj , uj ) = 1},
Spin(V, Q) := {a ∈ Pin(V, Q) | r is even}.
Exercises A.7.9:
Given a = u1 © · · · ©ur ∈ Cl(V, Q), let ã = (−1)r ur © · · · ©u1 .
1. Show that a → ã, extended linearly to all of CL(V, Q), is a well-defined
involution of the algebra Cl(V, Q), i.e., it is an algebra homomorphism whose
square is minus the identity.
2. For v ∈ V , show that a©v ©ã ∈ V .
3. Using the tilde involution, we define a representation ρ : Spin(V, Q) →
GL(V ) by ρ(a)v := a©v ©ã. Show that ρ is a 2-to-1 surjective group homo-
morphism Spin(V, Q) → SO(V, Q). 
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 A.7. Clifford algebras and spin groups 403

Clifford algebras as matrix algebras. From now on, assume dim V =

2m, and if we work over R, assume Q has signature (m, m).
We have defined Cl(V, Q) as Λ• V with an exotic multiplication, but in
fact, as an algebra, Cl(V, Q) is something familiar, as we now show. Fix
U, U  ⊂ V such that Q|U = Q|U  = 0, and V = U + U  . (Note that this
implies dim U = dim U  , U ∩ U  = 0, U ⊥ = U , and V = U ⊕ U  .) Thus for
all v ∈ V we may uniquely write v = x + y with x ∈ U , y ∈ U  . Define a
mapping
φ : V → End(Λ• U )
by, for v ∈ V and u ∈ Λ• U ,
√
φ(v)(u) = 2(x ∧ u − y u).
We calculate
φ(v)2 u = 2(x ∧ (x ∧ u − y u) − y (x ∧ u − y u))
= x ∧ x ∧ u − x ∧ (y u) − y (x ∧ u) − y (y u)
= 2Q(x, y)u = ||v||2Q u.
Thus the fundamental lemma applies and we obtain an algebra map φ̂ :
Cl(V, Q) → End(Λ• U ).
Exercises A.7.10:
1. φ̂ is a bijection, and thus, as an algebra, Cl(V, Q) ∼
= End(Λ• U ).
2. Moreover, show that we obtain algebra isomorphisms
Cleven (V, Q) ∼
= End(Λeven U ) ⊕ End(Λodd U ),
Clodd (V, Q) ∼
= [(Λeven U )∗ ⊗Λodd U ] ⊕ [(Λodd U )∗ ⊗Λeven U ].
3. Finally, show that Spin(V, Q) preserves End(Λeven U ) and End(Λodd U ),
which are often called the space of positive (resp. negative) spinors.
4. If V is a complex vector space with dim V > 7, show that Cl(V, Q)
cannot act nontrivially on Cn for n < 8. If dim V ≤ 7, determine the values
of n such that Cl(V ) acts nontrivially on Cn . 

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/14

Appendix B

Differential Forms

B.1. Differential forms and vector fields

Let M n be a differentiable manifold, let C ∞ (M ) denote the set of smooth
real-valued functions on M , and let x ∈ M . The cotangent space of M
at x, denoted Tx∗ M , may be defined as the set of equivalence classes of
maps f ∈ C ∞ (M ), where f ∼ g if in any coordinate system their first
derivatives agree at x. We let dfx denote the equivalence class of f . The
cotangent space can also be defined as the quotient of the maximal ideal mx
of functions vanishing at x by the ideal m2x of functions vanishing to order
two, i.e., Tx∗ X = mx /m2x .

Notation B.1.1. If π : E → M is a vector bundle, we let Γ(E) denote

the space of smooth sections of E, i.e., C ∞ maps s : M → E such that
π ◦ s = Id.

Exercises B.1.2:
1. Show that Tx∗ M is a vector space. Note that if x1 , . . . , xn are local coor-
dinates on M , then the dxi |x provide a basis of Tx∗ M , and thus dim Tx∗ M =
dim M .
2. Show that the cotangent bundle T ∗ M = x∈M Tx∗ M is a vector bundle
over M . Thus, given f : M → R, we obtain a section df ∈ Γ(T ∗ M ).
3. We define Tx M , the tangent space of M at x, as the space of equivalence
classes of differentiable mappings f : R → M with f (0) = x, where f ∼ g
if their first derivatives agree in some coordinate system near x. Show that
if f ∼ g, then their derivatives agree in every such coordinate system. The
vector bundle T M = x∈M Tx M is called the tangent bundle, and sections
X ∈ Γ(T M ) are called vector fields.
4. Show that Tx M and Tx∗ M are naturally dual vector spaces.

405
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 406 B. Differential Forms

The pairing between Tx M and Tx∗ M allows us to differentiate a function

f : M → R using a vector field X. Namely, for X ∈ Γ(T M ), define
X(f )x = dfx , Xx .
Note that a vector field is determined by how it acts on functions.
Differentiating a smooth function by a vector field produces another
function, which we may differentiate again. The failure of successive dif-
ferentiation by vector fields X and Y to commute is measured by their Lie
bracket [X, Y ], which is a vector field determined by
(B.1) [X, Y ](f ) := X(Y (f )) − Y (X(f )).
Exercises B.1.3:
∂
1. Suppose xi are local coordinates and X = ai (x1 , . . . , xn ) ∂x i . Compute
X(f ) in these coordinates, and conclude that the value of X(f )x depends
only on the function f and the value of Xx .
i , compute [X, Y ]. 
∂
2. If, in addition, Y = bi (x1 , . . . , xn ) ∂x

Notation B.1.4. The pairing between sections of T M and T ∗ M produces

a function on M which may be denoted in several ways:
α ∈ Γ(T ∗ M ), X ∈ Γ(T M ).
α, X = α(X) = X α,
! k
Let Ωk (M ) = Γ(Λk T ∗ M ). The direct sum Ω∗ (M ) := Ω (M ) denotes the
space of all differential forms on M .

If α ∈ Ωk (M ), we call α a differential form of degree k, or k-form for

short. As indicated in Exercise B.1.2.1, any 1-form can be expressed as a
C ∞ -linear combination of the dxi ’s, where x1 , . . . , xn are local coordinates
on M . In turn, any k-form can be expressed as a linear combination of
wedge products of 1-forms (see Appendix A).
Exercise B.1.5: On the circle in R2 defined by x2 + y 2 = 1, let θ ∈ (0, 2π)
be the angle from the positive x-axis and ψ ∈ (0, 2π) be the angle from the
negative x-axis, both measured counterclockwise. Show that dθ = dψ where
both angles are defined. Thus, one can define α ∈ Ω1 (S 1 ) by α = dθ where
θ is defined, and otherwise α = dψ.
Show that α = df for any function f : S 1 → R. Thus {df | f ∈
C ∞ (M, R)}  Ω1 (M ) when M = S 1 .

Pullbacks of differential forms, pushforwards of vector fields. Let φ : M → N

be a smooth mapping. For a differentiable function g : N → R, define the
pullback φ∗ g : M → R to be the composition g ◦ φ. This gives a linear
operator φ∗ : C ∞ (N ) → C ∞ (M ), which we extend to differential forms
below.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 B.2. Three definitions of the exterior derivative 407

For a vector X ∈ T M , we define φ∗ X ∈ T N by

φ∗ X(g) := X(φ∗ g).
This operation, called a pushforward, gives a vector bundle map φ∗ : T M →
TN.
We let φ∗x : Tx M → Tφ(x) N denote the pointwise map of vector spaces.
Sometimes we use the notation dφx for φ∗x . Note that φ∗x ∈ Tx∗ M ⊗Tφ(x) N
induces a transpose map, denoted φ∗ : Tφ(x)
∗ N → T ∗ M , that extends to a
x
bundle map φ∗ : Ω1 (N ) → Ω1 (M ). More generally, we get induced maps
on all tensor constructs (see §A.1), e.g. φ∗x : Λk Tφ(x)
∗ N → Λk T ∗ M , which
x
extend to bundle maps. The maps φ∗ are sometimes called pullback maps.
Exercises B.1.6:
1. Show that φ∗ α(X1 , . . . , Xk ) = α(φ∗ X1 , . . . , φ∗ Xk ).
2. Show that φ∗ [X, Y ] = [φ∗ X, φ∗ Y ].

B.2. Three definitions of the exterior derivative

We now extend the exterior derivative of a function to differential forms of
all degrees:
d : Ωk (M ) → Ωk+1 (M ).

Local coordinate definition. Suppose α = I aI dxi1 ∧ . . . ∧ dxik for
smooth coefficients aI (x), where I runs over all multi-indices with distinct
entries i1 , . . . , ik . Then
 ∂aI
dα := dx ∧ dxi1 ∧ . . . ∧ dxik .
∂x
I,

Note that if f ∈ Ω0 (M ) = C ∞ (M ), then df defined this way agrees with

our previous definition.
Exercise B.2.1: Show that d is well-defined. 

Vector field definition. If α ∈ Ω1 (M ) and X, Y ∈ Γ(T M ), then define

(B.2) dα(X, Y ) := X(α(Y )) − Y (α(X)) − α([X, Y ]).
More generally, for α ∈ Ωk (M ) and X1 , . . . , Xk+1 ∈ Γ(T M ), define
  
dα(X1 , . . . , Xk+1 ) := 2i , . . . , Xk+1 )
(−1)i+1 Xi α(X1 , . . . , X
i

+ 2i , . . . , X
(−1)i+j α([Xi , Xj ], X1 , . . . , X 2j , . . . , Xk+1 ),
i<j

where 2 denotes omission.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 408 B. Differential Forms

Exercises B.2.2:
1. Show that, according to the second definition, dα is C ∞ (M )-linear in
each Xi . Hence dα(X1 , . . . , Xk+1 )|x depends only on the values of the Xi at
x.
2. Show that the two definitions of d agree.
3. Using the second definition, show that φ∗ dα = d(φ∗ α).
4. Show that d2 = 0 (i.e., d(dα) = 0) under both definitions. In particular,
d2 = 0 is equivalent to the fact that mixed partials commute. What property
of vector fields is it equivalent to?

Fancy definition. Define d to be the unique operator such that

i. for f ∈ C ∞ (M ), df is as defined in §B.1,
ii. d(f α) = df ∧ α + f dα,
iii. d(α ∧ β) = dα ∧ β + (−1)p α ∧ dβ when α ∈ Ωp (M ), and
iv. d2 = 0.
Exercise B.2.3: Show that this definition is equivalent to the others.

Definition B.2.4. A form α ∈ Ωk (M, V ) is said to be closed if dα = 0.

Geometric interpretation of the exterior derivative of a 1-form.

Note that if α ∈ Ω1 (M ) is nowhere zero, then ker αx is a hyperplane in
Tx M . Thus, ker α is a hyperplane field in T M .
If α = df , then this hyperplane field is tangent to the level surfaces of f .
Otherwise, let E be the quotient bundle T ∗ M/{α}. Then dα mod α denotes
the image of dα under the projection Λ2 T ∗ M → Λ2 E, and measures how
much the hyperplane field defined by α fails to be tangent to hypersurfaces.

Exterior derivatives of vector-valued 1-forms. Let V be a vector

space of dimension r. A section of the bundle Λk T ∗ M ⊗ V is called a V -
valued k-form. The space of such forms is denoted by Ωk (M, V ).
Given a basis v 1 , . . . , v r ∈ V , suppose that α = αs ⊗v s for αs ∈ Ωk (M ).
We extend d to act on vector-valued forms by defining dα = dαs ⊗v s . The
reader may verify that this is independent of a choice of basis.

The Lie derivative. Let X ∈ Γ(T M ) be a vector field. While we already

know how to differentiate functions in the direction of X, the Lie derivative,
denoted by LX , enables us to differentiate vector fields, differential forms,
or sections of any of the tensor product bundles (T ∗ M )⊗k ⊗(T M )⊗l .
The Flowbox Theorem 1.2.2 implies that on a sufficiently small neigh-
borhood U of any point in M , there exists a family of local diffeomorphisms
φt : U → M such that φ0 = Id and for any x ∈ U , t → φt (x) is an integral
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 B.3. Basic and semi-basic forms 409

curve of X. For α ∈ Ωk (M ), we define LX α ∈ Ωk (M ) by

1
LX α = lim [φ∗t α − α].
t→0 t
For Y ∈ Γ(T M ), we define LX Y ∈ Γ(T M ) by
1
LX Y = lim [(φ−t )∗ Y − Y ].
t→0 t

We extend the Lie derivative to Γ((T ∗ M )⊗k ⊗(T M )⊗l ) by the Leibniz rule.
For example, for α, β ∈ Ω1 (M ),
(B.3) LX (α ⊗ β) = (LX α) ⊗ β + α ⊗ LX β.
Similarly, LX (α ∧ β) = (LX α) ∧ β + α ∧ LX β.
Exercises B.2.5:
1. Show that LX Y = [X, Y ].
2. Show that
(B.4) LX α = X dα + d(X α),
where the interior product is as defined in §A.1.

B.3. Basic and semi-basic forms

Suppose we have a fibration
i
F −−−−→ E
⏐
⏐π
-
B.
We say that X ∈ Tp E is a vertical vector if π∗p X = 0. Equivalently, X is
vertical if X = i∗q Y for some Y ∈ Tq F , where i(q) = p. A form φ ∈ Ωk (E)
is semi-basic if X φ = 0 for all vertical vectors X. The form is basic if
φ = π ∗ (φ) for some φ ∈ Ωk (B), in which case we say φ is the pullback of φ.
Exercise B.3.1: If we let {xi } be local coordinates on B and {xi , y α } be
local coordinates on E, a general 1-form on E may be written as
φ = fj (xi , y α )dxj + gβ (xi , y α )dy β .
Show that φ is semi-basic if and only if gβ ≡ 0 and basic if and only if, in
addition, the fj are functions of xi alone.
Remark B.3.2. Similar terminology applies to bundles on E. A bundle
V → E is said to be basic if V = π ∗ (V ) for some bundle V → B. If V
is basic, then a section s ∈ Γ(V ) is said to be basic if s = π ∗ s for some
s ∈ Γ(V ).
We have an easy test to determine if a form is semi-basic. The following
Licensed to AMS.
proposition gives a test to see if a form is basic:
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 410 B. Differential Forms

Proposition B.3.3. α ∈ Ωk (E) is basic if and only if α and dα are semi-

basic.
Exercise B.3.4: Prove Proposition B.3.3.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/15

Appendix C

Complex Structures
and Complex
Manifolds

Notation. If V is a real vector space, let VC = V ⊗ C (see §A.3). Likewise,

if E → M is a real vector bundle over manifold M , then EC will denote
its complexification. In particular, TC M is the complexified tangent bundle
and TC∗ M the complexified cotangent bundle of M .

C.1. Complex manifolds

In defining a complex manifold, one would like something that is modeled
locally on Cn . Two ideas come to mind:
The first idea is to imitate Riemannian manifolds. A Riemannian mani-
fold (M, g) is a space that infinitesimally “looks like” Euclidean space. That
is, for all x ∈ M , there exists a positive definite gx ∈ S 2 Tx∗ M that varies
smoothly from point to point. One possible definition of a complex mani-
fold would be to mimic this, i.e., define a “complex manifold” to be a real
manifold M such that each tangent space Tx M comes equipped with an
identification Tx M  Cn at each x ∈ M , and these identifications vary
smoothly from point to point.
The second idea is to imitate the definition of a C ∞ (real) manifold.
That is, define a “complex manifold” to be a topological manifold which
has local charts into Cn such that, where the domains of the charts overlap,
composing the charts gives invertible holomorphic functions from domains
in Cn to domains in Cn .

411
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 412 C. Complex Structures and Complex Manifolds

Do these two ideas agree? If not, which is the “better” one? The second
idea is a local definition, while the first is an infinitesimal one. It’s clear
that the second idea implies the first, since one can make identifications
using the pushforward φ∗ : Tx M → Tφ(x) Cn  Cn , and the requirement
that composition maps on the overlaps are holomorphic ensures that the
identifications are well-defined up to multiplication by a complex matrix. So
the question is whether or not the infinitesimal condition is strong enough
to imply the local one.
To answer the question, we first need to describe the infinitesimal idea
more precisely. Recall from Appendix A that a complex structure on a real
vector space V is a map J ∈ End(V ) with J ◦ J = − Id. Recall also that,
when we extend J to VC , that space splits as a sum of +i and −i eigenspaces
of J, denoted by V (1,0) and V (0,1) respectively, and that J is uniquely deter-
mined by this splitting. So, specifying a complex structure on V is equivalent
to specifying the n-dimensional complex subspace V (1,0) ⊂ VC (assuming its
intersection with its complex conjugate, which will be V (0,1) , is zero). In
∗(1,0) ∗(0,1)
turn, this is equivalent to specifying a dual splitting VC∗ = VC ⊕ VC ,
∗(1,0)
where the forms in VC vanish on the −i-eigenspace and vice versa.
Definition C.1.1. An almost complex structure on a real 2n-dimensional
manifold M is an n-dimensional smooth distribution T (1,0) M ⊂ TC M . It
determines a unique compatible J ∈ Γ(End(T M )) having T (1,0) M as its
+i-eigenspace. A manifold equipped with an almost complex structure will
be called an almost complex manifold.
A complex structure on M is an atlas of charts φi : Ui → Cn such that the
compositions φj ◦ φ−1
i , where defined, are invertible holomorphic functions
from domains in Cn to domains in Cn . A complex manifold is a real manifold
equipped with a complex structure.

Our question above may be rephrased as, are all almost complex mani-
folds complex? We will see below that if dim M = 2, the answer is yes.
Proposition C.1.2. An almost complex structure is complex if and only
if T ∗(1,0) M ⊂ Ω1C (M ) is a Frobenius ideal (equivalently, if T (1,0) M is closed
under the Lie bracket). When this is the case, the almost complex structure
is said to be integrable.
Here is the idea of the proof:
∗(1,0)
At a point p one has a subspace Tp ⊂ TC∗ M with basis, say,
αp1 , . . . , αpn . We want to find local coordinates (x1 , . . . , xn , y 1 , . . . , y n ), such
that if we define z j = xj + iy j , then dz j |p = αpj (which we can always do)
and moreover we can extend the αpj ’s to sections αj ∈ Γ(T ∗(1,0) M ) such
that in some neighborhood of p we have dz j = αj . In the real case, this
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 C.1. Complex manifolds 413

was exactly the problem we considered in discussing the Frobenius Theorem

of Chapter 1, which applies here (with some minor notational adjustments)
provided we assume that the underlying manifold is real-analytic and the
system T ∗(1,0) is also analytic. Finally we use (z 1 , . . . , z n ) as our holomor-
phic coordinate chart. What needs to be checked is that they actually give
coordinates which are compatible on the overlaps.
For more discussion of the proof, see [45] or [114, Chap. IX]. The
proof in the smooth category is considerably more difficult, and the result
in the smooth (in fact C 2 ) category is known as the Newlander-Nirenberg
Theorem.

We will use the notation Λ(p,q) V ⊂ Λp+q V from Appendix A. On an

almost complex manifold we have the vector bundles Λ(p,q) TC∗ M , spanned
by wedge products of p (1, 0)-forms and q (0, 1)-forms. These are part of a
direct sum decomposition
4
(C.1) Λk TC∗ M = Λ(p,q) TC∗ M.
p+q=k,
p≤n,q≤n

We refer to smooth sections of Λ(p,q) TC∗ M as (p, q)-forms and use the notation
Ω(p,q) (M ) = Γ(Λ(p,q) TC∗ M ).

A map f : M → N between almost complex manifolds is called holo-

morphic if df commutes with J, i.e., if for all x ∈ M , Jf (x) ◦ f∗ = f∗ ◦ Jx .
Exercise C.1.3: Show that a function f : M → C is holomorphic if and
only if df ∈ Ω(1,0) (M ).

We let Ωpholo (M ) ⊂ Ω(p,0) (M ) denote the space of holomorphic p-forms,

which are defined as follows. In local coordinates, α ∈ Ω(p,0) (M ) may be
written α = ai1 ,...,ip dz i1 ∧ . . . ∧ dz ip with ai1 ,...,ip ∈ C ∞ (M ). Then α is holo-
morphic if and only if all its coefficients ai1 ,...,ip are holomorphic functions.
Exercise C.1.4: Show that α ∈ Ω(p,0) (M ) is holomorphic if and only if

dα ∈ Ω(p+1,0) (M ),

and moreover, in this case dα ∈ Ωp+1

holo (M ).

Suppose θ1 , . . . , θn give a local framing of T ∗(1,0) M . Then any real 2n-

form on M must be a multiple of the volume form
Ω = in θ1 ∧ θ1 ∧ . . . ∧ θn ∧ θn .
This form is a nonvanishing section of Λ2n T ∗ M which gives the canonical
orientation on M associated to the almost complex structure.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 414 C. Complex Structures and Complex Manifolds

Exercise C.1.5: On R2n with coordinates xk , y k for 1 ≤ k ≤ n, let J(∂xk ) =

∂yk and J(∂yk ) = −∂xk . Show that the corresponding (1, 0)-forms are
spanned by dz k = dxk + idy k , and compute the canonical orientation.

The integrability condition for an almost complex structure is often

phrased in terms of its Nijenhuis tensor:
N (X, Y ) = [X, Y ] + J[JX, Y ] + J[X, JY ] − [JX, JY ], X, Y ∈ T M.
Through the following exercises, we will see that the vanishing of N (X, Y )
for all X, Y is equivalent to the condition in Proposition C.1.2.
Exercises C.1.6:
1. Show that N actually is a tensor, i.e., that N is C ∞ -linear in X and Y .
2. Suppose N is extended C-linearly to TC M . Compute N when X and Y
are both +i-eigenvectors for J, both −i-eigenvectors, or one of each.
3. Show that if N is identically zero, then the Lie bracket (suitably ex-
tended to vector fields with complex coefficients) of any two −i-eigenvectors
is another −i-eigenvector. Explain why this implies that T ∗(1,0) satisfies the
Frobenius condition.
4. Conversely, show that if T ∗(1,0) satisfies the Frobenius condition, then
the Nijenhuis tensor vanishes.

Examples of complex manifolds.

Example C.1.7 (Almost complex surfaces are complex). Suppose M is
a surface. Then there is just one term in the above decomposition (C.1)
for Λ2 TC∗ M . Thus, T ∗(1,0) is always integrable. A surface with a complex
structure is called a Riemann surface.
Since J defines an orientation and J 2 = − Id, then J may be thought of
as defining a 90-degree rotation in each tangent space of M . In fact, J defines
a metric, up to multiple, at each point. For, suppose θ is a nonvanishing
local section of T ∗(1,0) . Then Ω = iθ ∧ θ̄, and we may let
g(X, Y ) = Ω(X, JY ), X, Y ∈ T M.
Exercise C.1.8: Show that this defines a real-valued, symmetric, positive
definite inner product on each tangent space where θ = 0.

Of course, changing θ by a nonzero complex multiple changes g by a

positive multiple. Thus, a complex structure on a surface is equivalent to a
CO(2)-structure (see Chapter 9).
Example C.1.9. Let V be an oriented real vector space of dimension n,
with inner product  , . Let M = Gr(2, V ) ∼
= SO(n)/SO(2) × SO(n − 2) be
the Grassmannian of oriented 2-planes through the origin in V . Let E ∈ M ,
and let e1 , e2 be an oriented orthonormal basis for E ∗ with respect to the
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 C.2. The Cauchy-Riemann equations 415

inner product inherited from V . Any vector X in TE M can be written as

X = e1 ⊗ v + e2 ⊗ w for some v, w ∈ V /E. We define an almost complex
structure on M by

J(e1 ⊗ v + e2 ⊗ w) = e1 ⊗ w − e2 ⊗ v.

Exercise C.1.10: (a) Show that J is well-defined, i.e., it only depends on

the inner product and orientation of E, and not on our choice of basis for
E∗.
(b) Show that this almost complex structure is integrable. 

Almost complex manifolds are far more abundant than complex mani-
folds and have been used extensively in symplectic geometry—see [9]. On
the other hand, interesting explicit examples of almost complex structures
that are not complex are rare (just as interesting explicit ‘generic’ Riemann-
ian metrics are rare). Here is one:

Example C.1.11. Let R7 = ImO with its standard flat metric (see Ap-
pendix A for a description of the octonions O). Consider the unit sphere
S 6 ⊂ ImO. Define Ju (v) = uv, where we identify Tu S 6 with the linear
subspace u⊥ ⊂ R7 and uv is octonionic multiplication.
Exercise C.1.12: Show that this defines an almost complex structure on
S 6 . Show that the structure is not integrable.

C.2. The Cauchy-Riemann equations

If u, v are respectively the real and imaginary parts of a holomorphic function
of z = x + iy, then
∂u ∂v ∂u ∂v
(C.2) = , =− .
∂x ∂y ∂y ∂x
Since local solutions can be expressed by arbitrary convergent power series
in z, we will treat the Cauchy-Riemann equations (C.2) as “solved”. For
this reason, it is important to recognize systems that could be the Cauchy-
Riemann equations in disguise—as in Example 7.4.20. (We may also attempt
to find solutions to a given EDS by imposing additional conditions that
reduce the system to the Cauchy-Riemann equations [121].) To facilitate
this, we will examine the structure of the Pfaffian system corresponding to
(C.2).
On J 1 (R2 , R2 ) use standard coordinates xi , ua , pai (see §1.10), with all
indices running between 1 and 2. Let Σ ⊂ J 1 (R2 , R2 ) be the submanifold
defined by p11 = p22 and p12 = −p21 . Let I be the restriction to Σ of the
standard contact system generated by forms θa = dua − pai dxi . Then we
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 416 C. Complex Structures and Complex Manifolds

calculate that
 1    1
θ π1 π2 dx
(C.3) d 2 =− ∧ ,
θ −π2 π1 dx2

where π1 = dp11 and π2 = dp12 . These structure equations can be written

more compactly as

dθ = −π ∧ ω, where θ = θ1 + iθ2 , π = π1 − iπ2 , ω = dx1 + i dx2 .

The forms θ, ω, π are sections of TC∗ Σ. The structure equations show

that K = {θ, ω, π}C is the retracting space (see §7.1) of the Pfaffian system
spanned by θ. The splitting TC∗ Σ = K ⊕ K defines an almost complex
structure on Σ with K = T ∗(1,0) Σ. Since K satisfies the Frobenius condition,
in this case the almost complex structure is integrable. In this situation, θ
is said to define a complex contact structure on Σ. A complex analogue of
the Pfaff-Darboux Theorem 1.10.16 holds for such forms:
Proposition C.2.1. Let θ be a smooth nonvanishing complex-valued 1-
form on M n , and suppose θ ∧ (dθ)m = 0 but θ ∧ (dθ)m+1 = 0 for some
m such that 2m + 1 ≤ n/2. Let K ⊂ TC∗ M be the retracting space of θ,
and assume that K is integrable. Then in a neighborhood of any point on
M there are complex coordinates w, z 1 , . . . , z m , p1 , . . . , pm such that θ is a
complex multiple of dw − p1 dz 1 . . . − pm dz m .
If m = 1, then θ = dw − p dz up to multiple, and so integral manifolds of
θ may be constructed by setting w = f (z) and p = f  (z) for any holomorphic
function f (z).

Remark C.2.2. Note that, in Proposition C.2.1, it is not necessary that

M itself carry a complex structure. Instead, this result may be used to
locally construct a complex quotient manifold on which the contact structure
lives. Of course, it is preferable to identify a global quotient manifold for
the problem at hand. For example, the system for superminimal surfaces
in S 4 [22], which is defined on a ten-dimensional frame bundle, contains
a subsystem with structure equations similar to (C.3). Using Proposition
C.2.1 enables one to locally construct a quotient complex 3-fold on which
the subsystem reduces to a holomorphic contact form. However, it is only by
identifying a global quotient manifold—CP3 in the superminimal case—that
enables one to obtain results about the space of solutions.

Remark C.2.3. An important application of complex geometry to differ-

ential equations is Penrose’s twistor program, where, to quote Hitchin [96],
“the ultimate aim is ... to encode as much of mathematical physics as pos-
sible into holomorphic form and there to rely on geometry, supported only
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 C.2. The Cauchy-Riemann equations 417

by the constraints of the Cauchy-Riemann equations, to provide a descrip-

tion of the universe”. See Chapter 9 for an explicit example carried out by
Hitchin.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 https://doi.org/10.1090//gsm/175/16

Appendix D

Initial Value Problems

and the
Cauchy-Kowalevski
Theorem

D.1. Initial value problems

One of the recurring motifs you will find in this book is the attempt to
answer the following questions:
Given a system of differential equations, how large is the
space of solutions? What data is sufficient to determine a
unique solution? And, given an appropriate set of data,
how do we construct this solution?
In this appendix, we review how the classical theory of partial differential
equations answers these questions.
Traditionally, an initial value problem or Cauchy problem consists of a
differential equation or system of equations, and initial data which specify
the value of the unknown function or functions along a codimension-one set
H in the domain. The initial value problem is said to be well-posed if for all
choices of initial data (in some function space) there exists a unique solution
to the problem on some open set containing H, or a closed set with H as
boundary.

Example D.1.1. The wave equation utt − c2 uxx = 0 describes the propa-
gation of a wave in a one-dimensional medium with velocity c, where u(x, t)

419
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 420 D. Initial Value Problems

represents the height of the wave at point x and time t and damping is ne-
glected. Intuitively, since wave motion is the aggregate of the dynamics of
many small particles, we expect the solution to be determined by the initial
position u(x, 0) = h(x) and the velocity ut (x, 0) = v(x). In fact, the general
solution of this wave equation is

(D.1) u(x, t) = f (x + ct) + g(x − ct)

and f, g may be determined from h(x) and v(x). So, the solution is defined
for all x and t, and depends on two arbitrary functions of one variable,
specified along the hypersurface t = 0.

Example D.1.2. The heat equation ut = uxx models heat flow in a one-
dimensional medium, with u(x, t) representing temperature at point x and
time t. Intuitively, one expects that the temperature distribution along t = 0
exactly determines the temperature in the domain for all later times, so we
are led to pose the problem of solving the heat equation subject to u = f (x)
along the hypersurface t = 0. In fact, under reasonable assumptions on
f (x), there is a unique solution defined on R × [0, ∞) given by a convolution
formula [61].

The usual existence theorems [105] for second-order equations mod-

eled on this simple example are focused on producing solutions defined in
a neighborhood of H. However, the only general existence theorem for sys-
tems available to us, namely the Cauchy-Kowalevski Theorem stated below,
only provides for solutions in a neighborhood of a given point.
While in this sense the Cauchy-Kowalevski Theorem is not very sat-
isfying, it does help illuminate the obvious difference between the size of
the solution space for these two second-order equations. The fact that the
wave equation solution depends on two functions of x can be explained by
recasting the single equation as a system of first-order PDE, and observing
that, provided h and v are analytic, the resulting initial value problem fits
the hypotheses of Cauchy-Kowalevski. The fact that the heat equation so-
lutions seem to depend on only one function really arises from the fact that
the hypersurface t = 0, along which the data is specified, is characteristic
for the heat equation (see §7.2), and as a result Cauchy-Kowalevski cannot
be applied to that initial value problem.

D.2. The Cauchy-Kowalevski Theorem

Definition D.2.1. An initial value problem for a system of first-order PDE
for s functions ua of n + 1 variables is of Cauchy-Kowalevski form if it can
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 D.2. The Cauchy-Kowalevski Theorem 421

be written as
∂ua ∂ub
(D.2) = F a (xi , t, ub , i ),
∂t ∂x
(D.3) u |t=t0 = g (x , . . . , x ),
a a 1 n

where 1 ≤ a, b ≤ s as before, and 1 ≤ i, j ≤ n.

How general is the Cauchy-Kowalevski form? If the initial data are

specified along some hypersurface, we may always find local coordinates
xi , t so that this becomes a hyperplane where t is constant. However, we
then need to be able to solve the equations for the derivatives with respect
to this same t.
Restricting to first-order equations is no loss of generality, since any PDE
system can be turned into a first-order system by adding extra variables.
For example, we can rewrite the wave equation of Example D.1.1 by letting
u1 = u, u2 = ux and u3 = ut . We then obtain the system
⎧ 1
⎪
⎪ ∂u
⎪
⎪ = u3 ,
⎪
⎪ ∂t
⎨ 2
∂u ∂u3
⎪ = ,
⎪
⎪ ∂t ∂x
⎪
⎪
⎪ 3
⎩ ∂u = c2 ∂u ,
2

∂t ∂x
with the corresponding initial data u = h(x), u2 = h (x) and u3 = v(x).
1

Exercises D.2.2:
1. Give an explicit expression for the solution of the wave equation in terms
of initial data h(x) and v(x). Conclude that real-analyticity is not necessary.
2. Show that the initial value problem of Example D.1.2 cannot be ex-
pressed in Cauchy-Kowalevski form. (However, the heat equation by itself
can be put in the form (D.2) if we exchange x and t.)
Theorem D.2.3 (Cauchy-Kowalevski). Suppose that the system (D.2) is
quasilinear, i.e.,
∂ua ∂ub  ∂ub
= F a (xi , t, ub , i ) = Aabi (x, t, u) i + B a (x, t, u)
∂t ∂x ∂x
b,i

and that the functions Aabi and B a are real-analytic on an open set U ⊂
Rn+1 × Rs . Suppose there is an open set V ⊂ Rn on which the initial data
(D.3) is real-analytic, and such that the graph of g a |V lies in the intersection
of U with the hyperplane t = t0 . Then for any point (x10 , . . . , xn0 ) ∈ V , there
is an open set W ⊂ Rn+1 containing (x10 , . . . , xn0 , t0 ) and a solution U a (x, t)
of the initial value problem which is real-analytic on W . Moreover, if W 
is an open subset of W with a non-empty intersection with V , then any
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 422 D. Initial Value Problems

solution of the initial value problem defined on W  must coincide with the
solution ua = U a (x, t).
Remark D.2.4. The restriction of quasilinearity is not essential, since dif-
ferentiating (D.2) with respect to the xi produces equations which are linear
in the highest-order derivatives. Then adding more variables leads us to a
quasilinear system. When generalized to exterior differential systems (EDS),
this shows that the prolongation of any EDS is a linear Pfaffian system (see
Chapter 6).
Exercise D.2.5: Add more variables to the nonlinear first-order equation
u∂u/∂y = (∂u/∂x)2 to obtain a quasilinear system.

Why expect the Cauchy-Kowalevski Theorem to be true? Since we

are looking for real-analytic solutions, it makes sense to attempt to construct
them via power series in x and t. The special form of (D.2) and (D.3) means
that all “mixed” terms—involving x and t—are unambiguously determined
in terms of the initial data.
For example, consider the simple case of a first-order equation for a
single function u of two variables x, t given by
∂u ∂u
(D.4) = A(x, u) + B(x, u).
∂t ∂x
Suppose the initial condition is u(x, 0) = g(x), and assume that A, B and g
are analytic functions of their arguments on a neighborhood of the origin.
Analyticity implies there are convergent power series such that
 
A(x, u) = ajs xj us , B(x, u) = bjs xj us .
j,s≥0 j,s≥0

We want to find an analytic solution


u(x, t) = cjl xj tl
j,l≥0

for some constants cjl . Since u(x, 0) = j cj0 xj , the initial condition deter-
mines the constants cj0 . To recover the rest, we substitute in (D.4). To do
this, we need the expansions for the powers

us = csjl xj tl ,
j,l

where the csjl are linear combinations of powers of the cjl . Then (D.4) implies
that
  
lcjl xj tl−1 = iajs cskl cim xi+j+k−1 tl+m + bjs cskl xj+k tl .
j,l i,j,k,l,m,s j,k,l,s
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 D.3. Generalizations 423

In order for this equality to hold, it must hold for each power of t.
Equating the coefficients of the t0 terms on each side, we get
 
cj1 xj = iajs csk0 ci0 xi+j+k−1 + bjs csk0 xj+k .
i,j,k,s j,k,s

Observe that all terms on the right hand side are known, so if we now equate
coefficients of like powers of x, we can solve for the cj1 ’s. When we equate the
coefficients of the t1 terms, the right hand side will only involve previously
known quantities and the cj1 ’s, while the left hand side will involve the cj2 ’s.
Continuing, we can solve for all the cjl ’s.
We conclude that if there is a solution, it must be unique and given by
this procedure. To show that a solution in fact exists, it remains to show
this series converges in some neighborhood of the origin. For a proof of
Theorem D.2.3 using the method of majorants, see [105]; a special case is
given in Chapter 5.

D.3. Generalizations
When trying to discover solution formulas for PDE like the d’Alembert
formula (D.1), or when studying PDE systems that arise in geometry, we will
rewrite the differential equations as exterior differential systems—in other
words, replacing coordinates by moving frames, and equations by differential
ideals. By appropriately generalizing the notion of characteristic to these
systems, we can detect and construct explicit solution formulas for PDE
when they exist (see Chapter 7). As well, the Cauchy-Kowalevski Theorem
has a generalization which constructs solutions to initial value problems
for EDS. This theorem, called the Cartan-Kähler Theorem, is discussed in
Chapter 6 and Chapter 8.

Overdetermined systems. While the Cauchy-Kowalevski Theorem can-

not be applied to overdetermined systems (i.e., those with more equations
than unknowns), some of these systems do have solutions that depend on
more than just constants, and it is possible to solve Cauchy problems for
them. However, the equations in such a system must be compatible, in
the sense that no new higher-order equations result from differentiating the
equations in the system and equating mixed partials (see Chapter 1).
In more detail, given a first-order system of r equations for s functions
ua of n variables, there exists a change of coordinates so that the system
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 424 D. Initial Value Problems

takes the form

u1x1 = F11 (xj , ua ),
..
.
urx11 = Fr11 (xj , ua ),
u1x2 = F21 (xj , ua , uax1 ),
..
.
urx22 = F2r2 (xj , ua , uax1 ),
..
.
u1xn = Fn1 (xj , ua , uax1 , . . . , uaxn−1 ),
..
.
urxnn = Fnrn (xj , ua , uax1 , . . . , uaxn−1 ),
a
where uaxj = ∂u ∂xj
, 1 ≤ a ≤ s, 1 ≤ j ≤ n, and r1 ≤ r2 ≤ . . . ≤ rn = s with
r = r1 + . . . + rn (see [39]).
We may be able to produce solutions of this system by solving a series of
Cauchy problems. However, we need to check that equations are compatible,
i.e., that mixed partials commute:
∂ rα ∂ rα
F = F , 1 ≤ i, j ≤ n, ∀rα .
∂xi j ∂xj i
Although it would be impractical to change any given system of PDE into
the above form and perform this test, converting this system to an EDS
guides us naturally to the analogue of the above form. We can then apply
a straightforward test that signals when no further compatibility conditions
need to be checked. These topics are the focus of Chapter 5 and Chapter 6.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to
Selected Exercises

Chapter 1
1.2.4 Use Stokes’ Theorem applied to integrating about the boundary
of a rectangle.
1.4.3(f) Let A(t) be a family of circles with radius r(t) and centers c(t),
dt > | dt |, then the circles are nested.
and show that if dr dc

1.7.3.1(d) Consider the area of the region bounded by the curve and a line
parallel to the curve, and take the limit as the line approaches
the tangent line (see [83]).
1.8.4.3 For (b): Write c(t) = c(t)+r(t)e2 (t), and show that r is constant.
1.8.4.5 Consider the point c(t) + ρN + σB.
1.9.1.2 Use the skew-symmetry of the wedge product, defined in Appen-
dix A.
 
1.10.7 For part (c), n + m kj=0 n+j−1
j .

Chapter 2
2.1.2.2 Compare the speed of a curve to its curvature.
2.1.2.3 What is the characteristic polynomial of h?
2.2.2.5 Take a line in the x − z plane and rotate it about the z-axis.
v
2.2.2.6 Let r(v) = C cos(v), t(v) = 0 1 − C 2 sin2 v dv. The only com-
plete ones are the spheres.

425
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 426 Hints and Answers to Selected Exercises

2.3.1.3 Differentiate ω13 = k1 ω 1 and ω23 = k2 ω 2 .

2.3.1.4 Use (B.2).
2.3.1.5 Differentiate your answer from Exercise 2.3.1.3.
2.3.3 M compact implies there exists a point p at which k1 has a local
maximum. Then K constant implies that k2 must have a local
minimum at p as well. Deduce that p must be an umbilic point.
Let q ∈ M be any other point, and compare kj (p) with kj (q).
2.4.1 Answer: H ≡ 0, K = − a2 cosh
1
4
(v)
.

2.4.3 Locally we may write ω12 = du for some function u : F → R.

Since ω12 is not semi-basic for the projection to M , locally there
exists a vertical vector field X such that ω12 (X) > 0, or du(X) >
0, which implies u is not constant in vertical directions. Thus
we may choose a local section F : M → F such that F ∗ (u) is
constant and therefore F ∗ (ω12 ) = 0.
2.4.5 Instead of calculating the whole frame, just show that dω12 = 0.
2.5.4.2 It is sufficient to show that if we adapt frames such that e1 is
tangent to the fiber of the Gauss map, then the integral curves
to the vector field e1 are lines; see Chapter 4 for more details.
2.5.9 Use the identification in Exercise 3.7.2.4.
2.6.5 Extend c to a neighborhood and show independence of exten-
sion.
2.7.2 Use Proposition 2.5.3.
2.7.3 Use (2.18) and apply Stokes’ Theorem to dω12 .
2.8.1 Show that if one refines a triangulation by adding a new vertex
and corresponding edges, then χΔ doesn’t change.
3.2.2 If f˜ = Af for A ∈ GL(n), then (gij (f˜)) = tA−1 (gij (f ))A−1 .
2.8.5 First observe that this is true for orthonormal frames.
2.8.6 Normalize e1 × e2 to obtain N .

Chapter 3
3.1.1 For (b) differentiate the result of part (a). For (d) use the Cartan
Lemma A.1.9.
3.1.3 The fibers are isomorphic to SO(V ) without the identity labeled.
If we consider the point we are taking the differential at as the
identity, then that fiber becomes SO(V ) and thus its tangent
space at that point becomes so(V ).
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to Selected Exercises 427

3.1.8 Show that Θ̃ is unchanged under the action of O(n) on the fiber
of Fon (M ).
3.1.16 dβ is minus the skew-symmetrization of ∇β.
3.1.19 For arbitrary α = ai η i , we have ∇(α(X)) = d(ai xi ) = ai dxi +
xi dai and X ∇α = xi (dai − aj ηij ) from (3.3).
3.1.21.3 Use dα(X, Y ) = X(α(Y )) − Y (α(X)) − α(X, Y ) for α = η i , and
combine with (3.4).
3.1.21.4 Differentiating (3.4), with Z in place of X, gives a relationship
between Θ and the second derivatives of Z.
3.2.3 To do this exercise, you need to use the higher-dimensional ver-
sion of the Poincaré-Hopf Theorem; see p. 450 in Volume 1 of
[173] for further details.
3.2.7 On one hand ∇h = ω i ω j ⊗(dhij − hik ωjk − hjk ωik ) by the defini-
tion of the covariant derivative. On the other hand taking 0 =
d(ωjn+1 − hjk ω k shows (dhij − hik ωjk − hjk ωik ) ∧ ω j = 0. Now ap-
ply the Cartan Lemma to conclude (dhij − hik ωjk − hjk ωik ) = hijk
with hijk = hikj , but we also have hijk = hjik .
3.2.10 The differentials of η i − ω i are clearly zero modulo I, those
of ηji − ωji are zero thanks to the Gauss equation, and that of
ω n+1 , ωjn+1 − hjk ω k is zero thanks to the Codazzi equation.

3.3.7.2 Show that S 2 (Λ2 V ) is spanned by elements of the form (v ∧ w)2 .

3.3.10 Differentiate 0 = dηji + ηki ∧ ηjk − Θij .
3.4.2.5 Consider the orbit of a one-parameter subgroup passing through
a point x with the same tangent direction as a geodesic.
3.4.4.1 Use Proposition 11.2.10.
3.6.3.2 It is sufficient, by functoriality, to prove this is true on vector
fields Y ∈ Γ(T M ).
3.7.2.1 Use Gram-Schmidt.
3.7.7 Use Schur’s Lemma to show that for any group G and any irre-
ducible G-module W , there exists at most one G-invariant qua-
dratic form on W .

Chapter 4
4.2.2 Use the fundamental theorem of algebra.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 428 Hints and Answers to Selected Exercises

4.2.16.1 When we project, we lose a conormal direction and the tangent

space remains the same. Thus IIX  ,x is the same as IIX,x , except
μ
one quadric is lost. If we let p = [en+a ], and IIX,x = qαβ ω α ω β eμ ,
φ
then IIX  ,x = qαβ ω α ω β eφ , where n + 1 ≤ φ ≤ n + a − 1.

4.2.16.2 Taking a hyperplane section, we lose a tangent direction; say

μ
this direction is en . If IIX,x = qαβ ω α ω β eμ , then IIX∩H,x =
μ s t
qst ω ω eμ , where 1 ≤ s, t ≤ n − 1.

4.3.4.1 Calculate d(e0 + te1 ).

4.3.4.2 Use Terracini’s Lemma.

4.3.4.3 Use Exercise 4.3.4.2.

Chapter 5
5.1.11.1 A = {p111 w1 ⊗v 1 ◦ v 1 + p112 w1 ⊗v 1 ◦ v 2 + p122 w1 ⊗v 2 ◦ v 2 + p211 w2 ⊗
v 1 ◦ v 1 + p212 w2 ⊗ v 1 ◦ v 2 + p222 w2 ⊗ v 2 ◦ v 2 | p111 + p222 = 0} has
dimension 5 = 6 − 1.

Chapter 7
7.1.2.1 Use the fact that L[v,w] = Lv ◦ Lw − Lw ◦ Lv . This identity is
easily verified on 0-forms and exact 1-forms, and then follows for
all forms from the Leibniz rule (B.3) for Lie derivatives.
7.1.2.2 Suppose Lv ψ ∈ I for ψ ∈ I. Then use the Leibniz rule to show
that Lv (α ∧ ψ) ∈ I for any α, and use (B.4) to get Lv (dψ) ∈ I.
7.1.2.3 The most general v is given by (7.2) for f = −hz and g = h+zhz ,
where h is an arbitrary function of x, y, z.
7.1.3 Note that if ψ ∈ I, then φ∗t (ψ) ∈ I also.

7.1.6.1 Use (B.4).

7.1.6.2 Use the identity [v, w] ψ = Lv (w ψ) − w (Lv ψ), which may be

derived in the same way as the formula for L[v,w] above.

7.1.6.3 Use the formula v (α ∧ ψ) = (v α) ∧ ψ + (−1)deg α α ∧ (v ψ),

and let ψ be one of the algebraic generators of I.
2
7.1.15.1 u(x, y) = 25 x ± 12 y .

7.1.15.2 u(x, y) = 12 (1 − x2 ) ± y.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to Selected Exercises 429

√
7.1.15.3 u(x, y) = 2x3/2 / 1 − 27y. In general, the solution of u(s, 0) =
f (s) is u = f (s) − 2yf  (s)3 , where s is implicitly defined as a
function of x, y by s = x + 3yf  (s)2 . If f (s) isn’t linear, then
there exists an s such that f  (s) = 0, and hence there exists a y
such that ∂s/∂x is undefined. Then ∂ 2 u/∂x2 is undefined.
7.1.19.1 Use the Pfaff Theorem 1.10.16.

7.1.19.2 Using x, y, z, p, q, s as coordinates, the Cauchy characteristic is

given by
∂ ∂ ∂ 2 ∂ ∂
v= + s2 + (p + qs2 ) + s3 + 2s .
∂x ∂y ∂z 3 ∂p ∂q
7.1.22.2 See Proposition B.3.3.

7.2.8 The hyperbolic system of class zero is generated by (du −

f (v)dx) ∧ dy and (dv − g(u)dy) ∧ dx.

7.3.3 I (2) is spanned by θ1 .

 
7.3.13 q = sx − y ln s + φ(y)dy + ψ(s)/s
 ds and z = xp + qy − xys −
2 (x r − y ln s) + yφ(y)dy − 2
1 2 2 1 2
ψ(s)/s ds.
7.3.15.2 The class-zero system is integrable if f, g are constants. The
class-two system is integrable if f, g are exponential functions.
7.4.4.1 Compute dΨ, noting that dp, dq ≡ 0 modulo θ, dx, dy. Then mod
out by dθ.
7.4.8 First apply the Pfaff Theorem 1.10.16 to obtain local coordinates
in which θ is a multiple of dz − p dx − q dy. Then show that, in
these coordinates, the generator 2-form can be taken to have the
form (7.19).
7.4.10 M1,2 = {θ, dx ± dq, dy ± dp}.

7.4.11(b) Determine how β0 , β1 , β2 can be modified, by adding linear com-

binations of the coframe 1-forms, and still satisfy the structure
equations. Then, note that the derivative of the first equation
implies β1 ≡ β0 mod θ, ω1 , π1 and β2 ≡ β0 mod θ, ω2 , π2 .
7.4.11(c) The first structure equation now implies dβ0 ≡ a(ω1 ∧π1 −ω2 ∧π2 )
mod θ. Then differentiate β1 = β0 − aθ modulo θ and compare
with the derivative of the second structure equation.
7.4.13 If fz = 0, we have a first-order PDE for q = zy . If fz = 0,
then the equation is semi-integrable if and only if f (x, y, z, q) =
A(x, y, z)q + B(x, y, z) and Ay = Bz .
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 430 Hints and Answers to Selected Exercises

7.4.15 For part (a), note that ω 3 is the result of taking the dot product
of dx with e3 . For part (b), note that Fr is covered by the
transformation on F given by (x; e1 , e2 , e3 ) → (x+re3 ; e1 , e2 , e3 ).
Then use (1.26) and (1.27) to compute pullbacks of the ω’s.

7.4.16.1 Hint: when is Θ + λΨ decomposable, modulo ω 3 ?

7.4.16.2 In a space form of constant curvature , the Weingarten equation

AK+2BH+C = 0 is hyperbolic if and only if B 2 −A(C+A) < 0.
For example, the system for flat surfaces in S 3 is equivalent to
the wave equation.

7.4.19 Hint: It’s easier to use the ansatz u = 2 arctan(f (x + y)).

7.4.23 The minimal surfaces produced in question 1 are a plane, En-

neper’s surface, the catenoid, and the helicoid, respectively. The
catenoid is also the mystery surface in question 3.

7.5.8 Let V ⊂ T B be the bundle of vertical vectors for π. Then

J ⊂ T ∗ B defines a splitting T ∗ B as the direct sum J ⊕ V ⊥
of vector bundles, with dual splitting T B = V ⊕ H.
Let θi ∈ Γ(J) be a (local) basis for sections of J, with dual basis
ei ∈ Γ(V ). Then by definition
dθi ≡ Θij ∧ θj mod π ∗ I
for some forms Θij on B. Thinking of H as a kind of connection,
we define the torsion of J as
Θ = Θij ⊗ (θi ⊗ ej ) ∈ Γ((T ∗ B/(J ⊕ π ∗ I 1 )) ⊗ End(V )).
Then (B, J) splits locally if there exists a subbundle W ⊂ V
preserved by Θ.

7.5.13.2 Condition (7.37a) implies that decomposable 2-forms in one sys-

tem are congruent to decomposables for the other, modulo the
contact forms ω 3 , ω̄ 3 . The asymptotic lines are dual to the factors
of these decomposables.

7.5.13.5 Although the Cole-Hopf transformation can be described in

terms of a double fibration, with the heat equation on one side
and Burger’s equation on the other, the pullback of the heat
equation system doesn’t include the 2-forms for Burger’s equa-
tion. Equivalently, given a solution of the heat equation, there
is no Frobenius system available for us to produce multiple solu-
Licensed to AMS.
tions of Burger’s equation.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to Selected Exercises 431

7.5.13.6 The point is to show that Bäcklund-equivalence is transitive:

given (M1 , I1 ) linked to (M2 , I2 ) by a double fibration carry-
ing a Bäcklund transformation, and (M2 , I2 ) similarly linked to
(M3 , I3 ), use this data to construct a Bäcklund transformation
linking (M1 , I1 ) and (M3 , I3 ).

Chapter 8
8.1.5 Consider the bundle map i : T Σ → T ∗ Σ given by v → v (ω 1 ∧ω 2 ),
which is well-defined up to multiple. Then the singular points of
V2 (I) are those 2-planes to which this map restricts to be zero.
8.1.7.3 If ψ ∈ I i for k ≤ n, then v ψ|E = 0 implies that there exists
ψ ∧ α ∈ I n+1 such that v ψ|E = 0.
8.2.1.1 Express the right-hand side of (8.6) in terms of the Ωi and solve
for the Ωi in terms of the dω i ∧ ω i ’s.
8.2.1.3 Consider the intersection of a hyperboloid and an ellipsoid in a
system of confocal quadrics.
8.5.2 Let q = κeiθ and differentiate N = e2 sin θ + e3 cos θ.
8.5.9.1 In fact, d(x + (1/k0 )e3 ) ∧ ω 1 = 0, so each line of curvature in the
e2 -direction lies on a sphere of radius 1/k0 . The surface is the
envelope of these spheres, and the two functions may be taken
to be the curvature and torsion of the curve traced out by the
centers of the spheres.
8.5.9.2 These surfaces may also be characterized as those for which both
of the focal surfaces are degenerate. One such surface is the torus
swept out by revolving a circle about an axis. The EDS for these
surfaces becomes Frobenius after two prolongations, and the so-
lutions are the cyclides of Dupin. These are the surfaces that
are obtained by acting on circular tori of revolution by the group
O(4, 1) of conformal transformations of E3 . See [104] for more
information, including an interesting application of conformal
moving frames.
8.5.9.4 Add the 3-form ω14 ∧ ω 2 ∧ ω 3 + ω24 ∧ ω 3 ∧ ω 1 + ω34 ∧ ω 1 ∧ ω 3 to the
ideal.

Chapter 9
9.1.3 All are equivalent.
1 ∂2 Hx
9.1.12.5 (log )dx ∧ dy.
2 ∂x∂y Hy
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 432 Hints and Answers to Selected Exercises

F Fxy − Fx Fy
9.1.12.6 dx ∧ dy.
F2
9.7.1 Show that the components of θ span a codimension-one Pfaffian
system on α∗ FG .
9.7.2 Use the equivariance of θ.

9.8.10 Consider H 2 as half of a two-sheeted hyperboloid in Lorentz

space.

Chapter 10
10.2.6 Given α ∈ Γ(V̂ ), let α = α1 + α0 where α1 ∈ Γ(V̌ ) and α0 ∈
Γ(V̂ (∞) ). Conclude that α1 is a section of V̂ ∩ V̌ = I. Thus,
V̂ = {σ̂, η̂, θ, η̌}, and then use T ∗ M = V̂ + V̌ (∞) .

10.2.9 To show, e.g., that V̂ is a differential ideal, note that modulo the
1-forms in V̂ , the Ω̌ are spanned by the derivatives of θ and η̌.

10.2.11 If α ∈ V̂  V̌ and β is any k-form, consider evaluating α∧β on the

direct sum of an integral element of V̂ and an integral element
of V̌. (Take a basis for each.)

10.3.2 Show that η̂ m ([Ŝa , Ŝb ]) = −F̂ab

m . More generally, derive a recur-
m
sive formula for η̂ ([ŜA , Ŝb ]) and conclude (using the last state-
ment of Proposition 10.2.7) that η̂ m (ŜA ) is always a first integral
of V̂ (∞) .
10.3.7 It is straightforward algebra to show that ω̂ i = Λij ω̌ j . By evaluat-
ing dθX i (X , Y ) two different ways, show that X (Qi ) = C i Q .
j k j k j k
Use this and the 4-adapted structure equations to compute dQij .
Then, letting ψ̂ j = Ŝaj π̂ a and ψ̌ j = Šαj π̌ α , compare the 4-adapted
and 5-adapted structure equations to compute dR̂, dŘ, dψ̂ j and
dψ̌ j . Finally, using all of this, compute dω̂ i and dω̌ i .
10.5.7 By Proposition 10.2.10, the span of the Ω̌1 coincides with the
pullbacks of dη̌ and B σ̌ ∧ σ̌ to M1 , and these in turn are deriva-
tives of 1-forms in W1 .
10.5.11.8 When the system is encoded in the usual way on manifold Σ ⊂
J 2 (R3 , R), there is a decomposition T ∗ Σ = I ⊕ J1 ⊕ J2 ⊕ J3 such
that any generator 2-form belongs to Λ2 Ji for some i. Then one
can take V̂ to be the sum of I and any two of the Ji , while V̌ is
the sum of I and the remaining term in the decomposition.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to Selected Exercises 433

Chapter 11
11.1.6 Under a change of metric

Θ̂ij = [ + σ −2 ( σk2 )]η i ∧ η j + σ −1 (sjm η i ∧ η m − sim η j ∧ η m ).
k

In order that this is zero we see sij = 0 if i = j and sjj = 12 s is

independent of j, where

(4.5) s = −σ − σ −1 ( σk2 ).
k
(1)
Set up an EDS on × R × Rn where the last two factors
FCO(V )
have coordinates (σ, σi ) given by
I = {θ = dσ − σi ω i , θi = dσi − σj θij − μω i }.
Show the system is Frobenius, keeping in mind that the expres-
sion for ds is obtained by differentiating (4.5).
11.1.8 Since there is only one relation, the characters of the tableau
(ignoring the dθ and dλ terms) are s1 = . . . = sn−1 = n, sn =
n − 1. Any tableau with these characters cannot have torsion
and is involutive.
11.2.1 Since the map δ : so(V )⊗V ∗ → V ⊗Λ2 V ∗ is an isomorphism, the
map co(V )⊗V ∗ → V ⊗Λ2 V ∗ , whose kernel is co(V )(1) , must have
an n-dimensional kernel. Write an element of V ∗ as t = tk v k .
Map this to (δji tk − δkj ti + δki tj )(vi ⊗v j )⊗v k , and observe this is
a co(V )-module map. Note that this lies in co(V ) ⊗ V ∗ and in
V ⊗S 2 V ∗ and thus in co(V )(1) . Then to show co(V )(2) = 0, show
that this space tensored with each vm maps injectively under
δ : co(V )(1) ⊗V ∗ → V ⊗V ∗ ⊗Λ2 V ∗ .
We remark that the first part can be solved without calculation:
This map is a co(V )-module map, in particular an so(V )-module
map. As an so(V )-module map, its kernel is isomorphic to V 
V ∗ . As a co(V )-module map, its kernel is therefore of the form
V ⊗ Rμ , where the first factor is the so(V )-module V , and the
second tells how λ Id acts, namely by λ Id → λμ Id. The case
μ = 1 corresponds to the co(V )-module V and the case μ = −1
corresponds to the co(V )-module V ∗ . But on the whole space
V ⊗V ∗⊗V ∗ , the action of λ Id is with μ = −1, and so we conclude.
11.2.2 See Equation (11.6).
11.4.2 Sections of R+ correspond to sections of R[2].
11.4.4 Since it is semi-basic, it is sufficient to prove its exterior deriva-
Licensed to AMS.
tive is semi-basic.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 434 Hints and Answers to Selected Exercises

1
11.5.5 Use (11.27) and (11.3), keeping in mind that s = 2(n−1) P.

Chapter 12
12.1.8.1 Use equation (12.5).
12.1.8.3 Consider a cubic form in two variables and differentiate.
12.2.1 Use the prolongation property.
12.2.2.3 TE Gω (k, W )  E ∗ ⊗ (E ⊥ /E) ⊕ S 2 E ∗ . Here E ⊥ = {w ∈ W |
ω(v, w) = 0 ∀v ∈ E}.
12.2.5 Note that Mr is a join.
12.4.3 The cases (a), (b), (e) are always hypersurfaces, (c), (d) are hy-
persurfaces respectively when a = b and m is even. The smooth
quadric surface in P3 , v2 (P1 ), G(2, 5) and P1 × Pn are all self-
dual. The self-duality can be proven directly; or, once one knows
the dimension is correct, since the dual variety of a homogeneous
variety inherits a group action and they are the dimension of the
minimal orbit, they must be the minimal orbit.
12.4.14.1 Given A ∈ Mp×q , consider
 
0 A
tA 0 .

12.4.14.2 Any nonzero 3 × 3 skew-symmetric matrix has rank two.

12.4.14.3 Answer: |IISeg(P1 ×Pn ) |.

12.5.5 
Since we are working with eμ mod{T̂ , II v (T )}, adapt frames
further such that {eξ }  IIv (T ) and {eφ } gives a comple-
ment. You will need to show that if w ∈ Singloc(Ann(v)), then
IIw (T ) ⊆ IIv (T ).
12.5.6.1 Note that III can be recovered from III(v, v, v) for all v ∈ T .
12.6.12.1 Note that II(w1 ∧ w2 , u1 ∧ u2 ) = w1 ∧ w2 ∧ u1 ∧ u2 .
12.8.9 Adapt frames so that II = ω i ω j ⊗eij and calculate F4 , F5 . The
calculation is a little involved; see [127] for the details.
12.9.5 Use the calculation (12.14).
12.11.4 Use Corollary 12.11.2. What is ker II?
12.13.2 If x ∈ X is a general point and IIIX,x = 0, then a hyperplane
osculating to order two at x contains X.

Appendix A
A.1.6.3 Use the preceding exercise.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Hints and Answers to Selected Exercises 435

A.1.8.2 Λk V is the span of the wedge products (A.3).

A.1.8.4 ei1 ∧ · · · ∧ eik with 1 ≤ i1 < i2 < · · · < ik < n is a basis of Λk V .
A.1.11 When dim V > n, there are vectors rjk = rkj in V such that
Rj = rjk ∧ vk .
A.2.2.1 Consider the kernel of f .
A.2.2.2 Consider the map ψ : V ⊗S 2 V → Λ2 V ⊗V given by w⊗v ⊗v →
vww⊗v, and show that the image of ψ is KΛ (V ).
A.3.1.1 Consider the characteristic polynomial of J.
A.5.2 If you have trouble doing this invariantly, use the answer to Ex-
ercise A.5.3.4.
A.5.3.4 φ = dx123 + dx345 + dx156 + dx246 − dx147 − dx367 + dx257 , where
we have written dxijk for dxi ∧ dxj ∧ dxk .
A.6.3 Take a basis of Cd+1 consisting of points that project to be on
Cd .
A.6.4 First consider the case where all the eigenvalues are distinct and
let v be an eigenvector of A1 . Compute A1 A2 v = A2 A1 v to see
that A2 v is also an eigenvector of A1 with the same eigenvalue
as v.
A.6.5 Schur’s Lemma A.2.1 holds for Lie algebras, and the span of an
eigenvector gives a subrepresentation.
A.7.9.3 Calculate
A.7.10.4 Use Exercise A.7.10.2; the algebra of m × m matrices cannot act
nontrivially on a vector space of dimension less than m.

Appendix B
∂bj ∂ i ∂a
j ∂
B.1.3.2 [X, Y ] = ai − b .
∂xi ∂xj ∂xi ∂xj
B.2.1 Let y jbe a second set of coordinates and similarly let α =
(1/k!) J bJ dy j1 ∧ . . . ∧ dy jk . Using the chain rule, show that
dα is the same computed in either coordinate system.

Appendix C
C.1.10(b) Recall from Chapter 4 that the forms ωik for 3 ≤ k ≤ n and
i = 1, 2 are semi-basic for the projection from GL(V ) to G. Show
that the system spanned by the forms ω1k − iω2k is Frobenius, and
drops to G and forms a basis for T ∗(1,0) .

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Bibliography

1. L. Ahlfors, Complex Analysis, McGraw-Hill, 1966.

2. D.N. Akhiezer, Lie group actions in complex analysis, Vieweg, 1995.
3. M. Akivis, Webs and almost Grassmann structures, Soviet Math. Dokl. 21 (1980),
707–709.
4. M. Akivis, V. Goldberg, On the structure of submanifolds with degenerate Gauss maps,
Geom. Dedicata 86 (2001), 205–226.
5. I. Anderson, M. Fels, Bäcklund Transformations for Darboux Integrable Differential
Systems, Selecta Mathematica 21 (2015), 379–448.
6. I. Anderson, M. Fels, P. Vassiliou, Superposition formulas for exterior differential sys-
tems, Advances in Mathematics 221 (2009), 1910–1963.
7. I. Anderson, N. Kamran, The variational bicomplex for hyperbolic second-order scalar
partial differential equations in the plane, Duke Math. J. 87 (1997), 265–319.
8. I. Anderson, N. Kamran, P. Olver, Internal, External, and Generalized Symmetries,
Adv. Math 100 (1993), 53–100.
9. M. Audin, J. Lafontaine (eds.), Holomorphic curves in symplectic geometry,
Birkhäuser, 1994.
10. T. N. Bailey, M. G. Eastwood, and A. R. Gover, Thomas’s structure bundle for con-
formal, projective and related structures, Rocky Mountain J. Math. 24 (1994), no. 4,
1191–1217.
11. W. Barth, M. Larsen, On the homotopy groups of complex projective algebraic mani-
folds, Math. Scand. 30 (1972), 88–94.
12. R. Baston, M. Eastwood, The Penrose transform. Its interaction with representation
theory, Oxford University Press, 1989.
13. E. Berger, R. Bryant, P. Griffiths, The Gauss equations and rigidity of isometric em-
beddings, Duke Math. J. 50 (1983) 803–892.
14. Jonas Bergman, S. Brian Edgar, and Magnus Herberthson, The Bach tensor and other
divergence-free tensors, Int. J. Geom. Methods Mod. Phys. 2 (2005), no. 1, 13–21.
15. A.L. Besse, Einstein Manifolds, Springer, 1987.

437
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 438 Bibliography

16. R. Bishop, There is more than one way to frame a curve, Am. Math. Monthly 82 (1975),
246–251.
17. A. Bobenko, Exploring surfaces through methods from the theory of integrable systems:
Lectures on the Bonnet Problem, preprint (1999), arXiv:math.DG/9909003
18. R. Bott, L. Tu, Differential forms in algebraic topology, Springer, 1982.
19. N. Bourbaki, Groupes et algèbres de Lie, Chap. 4–6, Hermann, 1968.
20. Simon Brendle, Embedded minimal tori in S 3 and the Lawson conjecture, Acta Math.
211 (2013), no. 2, 177–190.
21. Simon Brendle and Fernando C. Marques, Recent progress on the Yamabe problem,
pp. 29–47 in Surveys in geometric analysis and relativity, International Press, 2011.
22. R. Bryant, Conformal and Minimal Immersions of Compact Surfaces into the 4-sphere,
J. Diff. Geom. 17 (1982), 455–473.
23. —, Metrics with exceptional holonomy, Ann. of Math. 126 (1987), 525–576.
24. —, An introduction to Lie groups and symplectic geometry, pp. 5–181 in Geometry and
quantum field theory, Amer. Math. Soc., Providence, 1995.
25. —, Rigidity and quasi-rigidity of extremal cycles in Hermitian symmetric spaces,
Annals of Mathematics Studies vol. 153, Princeton, 2002.
26. R. Bryant, S.-S. Chern, P. Griffiths, Exterior Differential Systems, pp. 219–338 in
Proceedings of the 1980 Beijing Symposium on Differential Geometry and Differential
Equations, Science Press, Beijing, 1982.
27. R. Bryant, S.-S. Chern, R.B. Gardner, H. Goldschmidt, P. Griffiths, Exterior Differ-
ential Systems, MSRI Publications, Springer, 1990.
28. R. Bryant, P. Griffiths, Characteristic Cohomology of Differential Systems (II): Conser-
vation Laws for a Class of Parabolic Equations, Duke Math. J. 78 (1995), 531–676.
29. R. Bryant, P. Griffiths, D. Grossman, Exterior differential systems and Euler-
Lagrange partial differential equations, University of Chicago Press, 2003.
30. R. Bryant, P. Griffiths, L. Hsu, Hyperbolic exterior differential systems and their con-
servation laws (I), Selecta Mathematica (N.S.) 1 (1995), 21–112.
31. —, Toward a geometry of differential equations, pp. 1–76 in Geometry, topology, and
physics, International Press, 1995.
32. R. Bryant, S. Salamon, On the construction of some complete metrics with exceptional
holonomy, Duke Math. J. 58 (1989), 829–850.
33. Andreas Čap and Jan Slovák, Parabolic geometries I: Background and general theory,
American Mathematical Society, 2009.
34. J. Carlson, D. Toledo, Generic Integral Manifolds for weight two period domains, Trans.
AMS 356 (2004), 2241-2249.
35. E. Cartan, Sur la structure des groupes infinis de transformations, Ann. Sci. Ecole
Norm. Sup. 26 (1909), 93–161.
36. —, Les systèmes de Pfaff à cinq variables et les équations aux dérivées partielles du
second ordre, Ann. Ecole Norm. Sup. 27 (1910), 109–192.
37. —, Sur les variétés de courbure constante d’un espace euclidien ou non euclidien, Bull.
Soc. Math France 47 (1919) 125–160 and 48 (1920), 132–208; see also pp. 321–432 in
Oeuvres Complètes Part 3, Gauthier-Villars, 1955.
38. —, Sur les variétés à connexion projective, Bull. Soc. Math. Fr. 52 (1924), 205–241.
39. —, Sur la théorie des systèmes en involution et ses applications à la relativité, Bull. Soc.
Math. Fr. 59 (1931), 88–118; see also pp. 1199–1230 in Oeuvres Complètes, Part 2.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Bibliography 439

40. —, Les Systèmes Extérieurs et leurs Applications Géométriques, Hermann, 1945.

41. P. Chaput, Severi varieties, Math. Z. 240 (2002), 451–459.
42. S.-S. Chern, Selected Papers Vols. 1–4, Springer, 1985/1989.
43. —, A simple intrinsic proof of the Gauss-Bonnet formula for closed Riemannian manifolds,
Ann. Math 45 (1944), 747–752.
44. —, Pseudo-groupes continus infinis, pp. 119–136 in Colloques Internationaux du Centre
National de la Recherche Scientifique: Géométrie différentielle, C.N.R.S, 1953.
45. —, Complex Manifolds without Potential Theory, Springer, 1979.
46. —, Deformations of surfaces preserving principal curvatures, pp. 155–163 in Differential
Geometry and Complex Analysis, Springer, 1984. (See also Selected Papers, vol. IV.)
47. S.-S. Chern, R. Osserman, Remarks on the Riemannian metric of a minimal submani-
fold, pp. 49–90 in Geometry Symposium, Utrecht 1980, Lecture Notes in Math. 894,
Springer, 1981.
48. S.-S. Chern, C.-L. Terng, An analogue of Bäcklund’s theorem in affine geometry, Rocky
Mountain Math J. 10 (1980), 105–124.
49. S.-S. Chern, J.G. Wolfson, A simple proof of Frobenius theorem, pp. 67–69 in Manifolds
and Lie groups (Notre Dame, Ind., 1980), Birkhäuser, 1981.
50. J. Clelland, T. Ivey, Parametric Bäcklund Transformations I: Phenomenology, Trans.
Amer. Math. Soc. 357 (2005), 1061–1093.
51. —, Bäcklund Transformations and Darboux Integrability for Nonlinear Wave Equations
(with J. Clelland), Asian J. Math. 13 (2009), 13–64.
52. J. Clelland, P. Vassiliou, A solvable string on a Lorentzian surface, Diff. Geom. Appl.
33 (2014), 177–198.
53. D. Cox, S. Katz, Mirror symmetry and algebraic geometry, AMS, 1999.
54. G. Darboux, Leçons sur la Théorie Générale des Surfaces (3rd ed.), Chelsea, 1972.
55. M. do Carmo, Differential Geometry of Curves and Surfaces, Prentice-Hall, 1976.
56. M.G. Eastwood, A.R. Gover, and K. Neusser, Conformal differential geometry, in
preparation.
57. L. Ein, Varieties with small dual varieties, I, Inventiones Math. 86 (1986), 63–74.
58. —, Varieties with small dual varieties, II, Duke Math. J. 52 (1985), 895–907.
59. Ph. Ellia, D. Franco, On codimension two subvarieties of P5 and P6 , J. Algebraic
Geom. 11 (2002), 513–533.
60. Y. Eliashberg, L. Traynor (eds.) Symplectic Geometry and Topology, AMS, 1999.
61. L. Evans, Partial Differential Equations, AMS, 1998.
62. A.R. Forsyth, Theory of Differential Equations (Part IV): Partial Differential Equa-
tions, Cambridge University Press, 1906; also, Dover Publications, 1959.
63. G. Fubini, E. Cech, Géométrie projective différentielle des surfaces, Gauthier-Villars,
1931.
64. W. Fulton, J. Hansen, A connectedness theorem for projective varieties, with applications
to intersections and singularities of mappings, Ann. of Math. (2) 110 (1979), 159–166.
65. W. Fulton, J. Harris, Representation theory – a first course, Springer, 1991.
66. R.B. Gardner, Invariants of Pfaffian Systems, Trans. A.M.S. 126 (1967), 514–533.
67. —, The Method of Equivalence and Its Applications, SIAM, 1989.
68. —, Differential geometric methods interfacing control theory, pp. 117–180 in Differential
Geometric Control Theory, Birkhäuser, 1983.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 440 Bibliography

69. R.B. Gardner, W.F. Shadwick, The GS algorithm for exact linearization, IEEE Trans.
Autom. Control 37 (1992), 224–230.
70. J. Gasqui, Sur la résolubilité locale des équations d’Einstein, Compositio Math. 47
(1982), 43–69.
71. —, Formal integrability of systems of partial differential equations, pp. 21–36 in Non-
linear equations in classical and quantum field theory, Lecture Notes in Phys. 226,
Springer, 1985.
72. I.M. Gel’fand, S. Fomin, Calculus of Variations, Prentice-Hall, 1963.
73. I.M. Gel’fand, M. Kapranov, A. Zelevinsky, Discriminants, resultants, and multidi-
mensional determinants, Birkhäuser, 1994.
74. H. Goldschmidt, Existence theorems for analytic linear partial differential equations,
Ann. of Math. 86 (1967), 246–270.
75. A.B. Goncharov, Generalized Conformal Structures on Manifolds, Selecta Math. Sov. 6
(1987), 307–340.
76. E. Goursat, Leçons sur l’intégration des équations aux dérivées partielles du second
ordre, Gauthier-Villars, 1890.
77. —, Recherches sur quelques équations aux dérivées partielles du second ordre, Annales
de la Faculté de Toulouse, deuxième serie 1 (1899), 31–78.
78. A. Rod Gover, Almost conformally Einstein manifolds and obstructions, pp. 247–260 in
Differential geometry and its applications, Matfyzpress, Prague, 2005.
79. —, Almost Einstein, Poincaré-Einstein manifolds in Riemannian signature, J. Geom.
Phys. 60 (2010), no. 2, 182–204.
80. C. Robin Graham and Kengo Hirachi, The ambient obstruction tensor and Q-curvature,
pp. 59–71 in AdS/CFT correspondence: Einstein metrics and their conformal bound-
aries, Eur. Math. Soc., Zürich, 2005.
81. Mark L. Green, Generic initial ideals, pp. 119–186 in Six lectures on commutative
algebra (J. Elias et al, eds.), Progr. Math. 166, Birkhäuser, 1998.
82. P. Griffiths, Some aspects of exterior differential systems, pp. 151–173 in Complex
geometry and Lie theory, Proc. Sympos. Pure Math. 53 (1991), AMS.
83. —, Exterior Differential Systems and the Calculus of Variations, Birkhäuser, 1983.
84. P. Griffiths, J. Harris, Principles of Algebraic Geometry, Wiley, 1978.
85. —, Algebraic geometry and local differential geometry, Ann. Sci. Ecole Norm. Sup. 12
(1979), 355–432.
86. P. Griffiths, G. Jensen, Differential systems and isometric embeddings, Princeton
University Press, 1987.
87. M. Guest, Harmonic Maps, Loop Groups and Integrable Systems, Cambridge Uni-
versity Press, 1997.
88. V. Guillemin, The integral geometry of line complexes and a theorem of Gel’fand-Graev,
pp. 135–149 in The mathematical heritage of Elie Cartan, Astérisque, Numero Hors
Serie (1985).
89. J. Harris, Algebraic geometry, a first course, Springer, 1995.
90. Robin Hartshorne, Varieties of small codimension in projective space, Bull. Amer. Math.
Soc. 80 (1974), 1017–1032.
91. F.R. Harvey, Spinors and Calibrations, Academic Press, 1990.
92. F.R. Harvey, H.B. Lawson, Calibrated geometries, Acta Math. 148 (1982), 47–157.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Bibliography 441

93. T. Hawkins, Emergence of the theory of Lie groups. An essay in the history of math-
ematics 1869–1926, Springer, 2000.
94. S. Helgason, Differential Geometry, Lie Groups and Symmetric Spaces, Academic
Press, 1978.
95. N. Hitchin, Kählerian twistor spaces, Proc. London Math. Soc. (3) 43 (1981), 133–150.
96. —, Complex manifolds and Einstein’s equations, pp. 73–99 in Twistor geometry and
nonlinear systems, Lecture Notes in Math. 970, Springer, 1982.
97. W.V.D. Hodge, D. Pedoe, Methods of Algebraic Geometry Vol. 2, Cambridge Uni-
versity Press, 1954.
98. J. Humphreys, Introduction to Lie algebras and representation theory, Springer, 1972.
99. J.M. Hwang, N. Mok, Uniruled projective manifolds with irreducible reductive G-
structures, J. reine angew. Math. 490 (1997), 55–64.
100. J.M. Hwang, N. Mok, Rigidity of irreduicible Hermitian symmetric spaces of the compact
type under Kahler deformation, Invent. Math. 131 (1998), 393–418.
101. J.M. Hwang, K. Yamaguchi, Characterization of Hermitian symmetric spaces by funda-
mental forms, Duke Math. J. 120 (2003), 621–634.
102. J. Jost, Riemannian geometry and geometric analysis, Springer, 2011.
103. Bo Ilic, J. M. Landsberg, On symmetric degeneracy loci, spaces of symmetric matrices
of constant rank and dual varieties, Math. Ann. 314 (1999), no. 1, 159–174.
104. T. Ivey, Surfaces with orthogonal families of circles, Proc. AMS 123 (1995), 865–872.
105. F. John, Partial Differential Equations (4th ed.), Springer, 1982.
106. D. Joyce, Compact Riemannian 7-manifolds with holonomy G2 , I, II, J. Diff. Geom. 43
(1996), 291–375.
107. D. Joyce, Compact 8-manifolds with holonomy Spin(7), Invent. Math. 123 (1996),
507–552.
108. M. Juráš, I. Anderson, Generalized Laplace invariants and the method of Darboux, Duke
Math. J. 89 (1997), 351–375.
109. E. Kähler, Einfürhung in die Theorie der Systeme von Differentialgleichungen, Teub-
ner, 1934.
110. N. Kamran, An elementary introduction to exterior differential systems, pp. 151–173 in
Geometric approaches to differential equations, Cambridge University Press, 2000.
111. S. Kleiman, Tangency and duality, pp. 163–225 in Proceedings of the 1984 Vancouver
conference in algebraic geometry, CMS Conf. Proc., vol. 6, Amer. Math. Soc., 1986.
112. M. Kline, Mathematical Thought from Ancient to Modern Times, Oxford University
Press, 1972.
113. A. Knapp, Lie groups beyond an introduction, Birkhäuser, 1996.
114. S. Kobayashi, K. Nomizu, Foundations of Differential Geometry, Vols. 1,2, Wiley,
1963/1969.
115. S. Kobayashi, T. Ochiai, Holomorphic structures modeled after compact Hermitian
symmetric spaces, pp. 207–222 in Manifolds and Lie groups, Birkhäuser, 1981.
116. Bertram Kostant, Holonomy and the Lie algebra of infinitesimal motions of a Riemannian
manifold, Trans. Amer. Math. Soc. 80 (1955), 528–542.
117. J.Krasil’shchick, A.M. Vinogradov (eds.), Symmetries and conservation laws for dif-
ferential equations of mathematical physics, American Math. Society, 1999.
118. Wolfgang Kühnel and Hans-Bert Rademacher, Einstein spaces with a conformal group,
Results Math. 56 (2009), no. 1-4, 421–444.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 442 Bibliography

119. M. Kuranishi, Lectures on exterior differential systems, Tata Institute, Bombay, 1962.
120. —, CR geometry and Cartan geometry, Forum Math. 7 (1995), 147–205.
121. J.M. Landsberg, Minimal submanifolds defined by first-order systems of PDE, J. Diff.
Geom. 36 (1992), 369–415.
122. —, On second fundamental forms of projective varieties, Inventiones Math. 117 (1994),
303–315.
123. —, On degenerate secant and tangential varieties and local differential geometry, Duke
Mathematical Journal 85 (1996), 605–634.
124. —, Differential-geometric characterizations of complete intersections, J. Diff. Geom. 44
(1996), 32–73.
125. —, Introduction to G-structures via three examples, pp. 133–149 in Web theory and
related topics, World Scientific, 2001,
126. —, Exterior differential systems: a geometric approach to pde, pp. 77–101 in Topology
and Geometry, Proc. Workshop Pure Math. vol. 17, Korean Academic Council, 1998.
127. —, On the infinitesimal rigidity of homogeneous varieties, Compositio Math. 118 (1999),
189–201.
128. —, Is a linear space contained in a submanifold? - On the number of derivatives needed
to tell, J. reine angew. Math. 508 (1999), 53–60.
129. —, Algebraic geometry and projective differential geometry, Lecture Notes Series, vol.
45, Seoul National University, Research Institute of Mathematics, Global Analysis
Research Center, Seoul, 1999.
130. —, Lines on projective varieties, J. Reine Angew. Math. 562 (2003), 1–3.
131. —, Griffiths-Harris rigidity of compact Hermitian symmetric spaces, J. Diff. Geom. 74
(2006), 395–405.
132. —, The border rank of the multiplication of 2 × 2 matrices is seven, J. Amer. Math.
Soc. 19 (2006), no. 2, 447–459.
133. —, Tensors: geometry and applications, American Mathematical Society, Providence,
RI, 2012.
134. J.M. Landsberg, L. Manivel, The projective geometry of Freudenthal’s magic square,
J. Algebra 239 (2001), 477–512.
135. —, Construction and classification of complex simple Lie algebras via projective geometry,
Selecta Math. (N.S.) 8 (2002), 137–159.
136. —, On the projective geometry of homogeneous varieties, Commentari Math. Helv. 78
(2003), 65–100.
137. J.M. Landsberg, C. Robles, Fubini-Griffiths-Harris rigidity and Lie algebra cohomology,
Asian J. Math. 16 (2012), 561–586.
138. H.B. Lawson, M. Michelsohn, Spin geometry, Princeton University Press, 1989.
139. R. Lazarsfeld, Positivity in algebraic geometry, Volume I, Springer, 2004.
140. Claude R. LeBrun, Ambi-twistors and Einstein’s equations, Classical and Quantum
Gravity 2 (1985), no. 4, 555–563.
141. Claude LeBrun and Simon Salamon, Strong rigidity of positive quaternion-Kähler man-
ifolds, Invent. Math. 118 (1994), no. 1, 109–132.
142. M. van Leeuwen, LiE, a computer algebra package,
available at http://young.sp2mi.univ-poitiers.fr/∼marc/LiE/.
143. S. L’vovsky, On Landsberg’s criterion for complete intersections, Manuscripta Math. 88
(1995), 185–189.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Bibliography 443

144. B. Malgrange, L’involutivité générique des systèmes différentiels analytiques, C. R.

Acad. Sci. Paris, Ser. I 326 (1998), 863–866.
145. R. Mayer, Coupled contact systems and rigidity of maximal dimensional variations of
Hodge structure, Trans. AMS 352 (2000), 2121–2144.
146. R. McLachlan, A Gallery of Constant-Negative-Curvature Surfaces, Math. Intelligencer
16 (1994), 31–37.
147. M. Melko, I. Sterling, Integrable systems, harmonic maps and the classical theory of
surfaces, pp. 129–144 in Harmonic Maps and Integrable Systems, Vieweg, 1994.
148. J. Milnor, Topology from the differentiable viewpoint, U. Virginia Press, 1965.
149. —, Morse theory, Princeton University Press, 1963.
150. J. Milnor, J.D. Stasheff, Characteristic classes, Princeton University Press, 1974.
151. D. Mumford, Algebraic Geometry I: Complex projective varieties, Springer, 1976.
152. R. Muñoz, Varieties with degenerate dual variety, Forum Math. 13 (2001), 757–779.
153. I. Nakai, Web geometry and the equivalence problem of the first order partial differential
equations, pp. 150–204 in Web theory and related topics, World Scientific, 2001.
154. Z. Nie, On characteristic integrals of Toda field theories, J. Nonlinear Math. Phys. 21
(2014), 120–131.
155. T. Ochiai, Geometry associated with semisimple flat homogeneous spaces, Trans. AMS
152 (1970), 159–193.
156. C. Okonek, M. Schneider, H. Spindler, Vector bundles on complex projective spaces,
Progress in Mathematics 3, Birkhauser, 1980.
157. P. Olver, Applications of Lie Groups to Differential Equations (2nd ed.), Springer,
1993.
158. —, Equivalence, Invariants, and Symmetry, Cambridge University Press, 1995.
159. G. Pappas, J. Lygeros, D. Tilbury, S. Sastry, Exterior differential systems in control
and robotics, pp. 271–372 in Essays on mathematical robotics, Springer, 1998.
160. J. Piontkowski, Developable varieties of Gauss rank 2, Internat. J. Math. 13 (2002),
93–110.
161. Y. S. Poon, S. M. Salamon, Quaternionic Kähler 8-manifolds with positive scalar cur-
vature, Journal of Differential Geometry 33 (1991), 363–378.
162. Z. Ran, On projective varieties of codimension 2 Invent. Math. 73 (1983), 333–336.
163. —, The (dimension +2)-secant lemma, Invent. Math. 106 (1991), no. 1, 65–71.
164. C. Rogers, Bäcklund transformations in soliton theory, pp. 97–130 in Soliton theory: a
survey of results (A. P. Fordy, ed.), St. Martin’s Press, 1990.
165. C. Rogers, W. Schief, Bäcklund and Darboux Transformations, Cambridge University
Press, 2002.
166. C. Rogers, W. Shadwick, Bäcklund Transformations and Their Applications, Aca-
demic Press, 1982.
167. E. Sato, Uniform vector bundles on a projective space, J. Math. Soc. Japan 28 (1976),
123–132.
168. B. Segre, Bertini forms and Hessian matrices, J. London Math. Soc. 26 (1951), 164–176.
169. C. Segre, Preliminari di una teoria delle varietà luoghi di spazi, Rend. Circ. Mat. Palermo
XXX (1910), 87–121.
170. W. Shadwick, The KdV Prolongation Algebra, J. Math. Phys. 21 (1980), 454–461.
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 444 Bibliography

171. R. Sharpe, Differential geometry: Cartan’s generalization of Klein’s Erlangen pro-

gram, Springer, 1997.
172. G.F. Simmons, Differential Equations, with Applications and Historical Notes,
McGraw-Hill, 1972.
173. M. Spivak, Calculus on Manifolds, Benjamin, 1965.
174. —, A Comprehensive Introduction to Differential Geometry (3rd ed.), Publish or
Perish, 1999.
175. O. Stormark, Lie’s Structural Approach to PDE Systems, Encyclopedia of Mathe-
matics and its Applications, v. 80, Cambridge University Press, 2000.
176. V. Strassen, Relative bilinear complexity and matrix multiplication, J. Reine Angew.
Math. 413 (1991), 127-180.
177. S. Sternberg, Lectures on differential geometry, Chelsea, 1983.
178. D.J. Struik, Lectures on Classical Differential Geometry (2nd ed.), Addison-Wesley,
1961; Dover, 1988.
179. A. Terracini, Alcune questioni sugli spazi tangenti e osculatori ad una varieta, I, II, III,
Atti della Societa dei Naturalisti e Matematici Torino 49 (1914), 214–247.
180. E. Tevelev, Projectively dual varieties of homogeneous spaces, pp. 183–225 in Surveys in
geometry and number thoery: reports on contemporary Russian mathematics, London
Math. Soc. Lecture Notes vol. 338, Cambridge, 2007.
181. T.Y. Thomas, Announcement of a projective theory of affinely connected manifolds,
Proc. Nat. Acad. Sci. 11 (1925), no. 11, 588–589.
182. P.J. Vassiliou, Darboux Integrability and Symmetry, Trans. AMS 353 (2001), 1705–
1739.
183. M. Wadati, H. Sanuki, and K. Konno, Relationships among inverse method, Bäcklund
transformation and an infinite number of conservation laws, Progr. Theoret. Phys. 53
(1975), 419–436.
184. E. Vessiot, Sur les équations aux dérivées partielles du second ordre intégrables par la
methode de Darboux, J. Math Pures Appl. 18 (1939), 1–61 and 21 (1942), 1–66.
185. H. Wahlquist, F. Estabrook, Prolongation structures of nonlinear evolution equations,
J. Math. Phys. 16 (1975), 1–7.
186. —, Prolongation structures, connection theory and Bäcklund transformations, pp. 64–83
in Nonlinear evolution equations solvable by the spectral transform, Res. Notes in
Math., vol. 26, Pitman, 1978.
187. F. Warner, Foundations of differentiable manifolds and Lie groups, Springer, 1983.
188. K. Yamaguchi, Differential Systems Associated with Simple Graded Lie Algebras,
pp. 413–494 in Progress in Differential Geometry, Adv. Stud. Pure Math., vol 22.,
Math. Soc. Japan, 1993.
189. D. Yang, Involutive hyperbolic differential systems, A.M.S. Memoirs # 370 (1987).
190. K. Yang, Exterior differential systems and equivalence problems, Kluwer, 1992.
191. F. Zak, Tangents and Secants of Algebraic Varieties, Translations of mathematical
monographs vol. 127, AMS, 1993.
192. M.Y. Zvyagin, Second order equations reducible to zxy = 0 by a Bäcklund transforma-
tion, Soviet Math. Dokl. 43 (1991) 30–34.

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Index

(S d V )0 , 79 III, projective third fundamental form,

1-graph, 8 333
G-structure, 241–248 III v , 348
1-flat, 250 Ln+1 , 80
2-flat, 253 Γ(E), smooth sections of E, 405
μ
curvature, 252, 254 Δrαβγ , 333
definition, 247 2
Λ V , 383
flat, 248 ΞA , characteristic variety of a tableau,
prolongation, 254 130
G(k, V ), 28 Ωk (M ), Ω∗ (M ), 406
G/H-structure of order two, 273 Ωk (M, V ), 408
I, vector subbundle spanned by 1-forms
Ω(p,q) (M ), 413
of I, 10
δσ (X), secant defect, 347
I k , vector subbundle spanned k-forms
δτ (X), tangential defect, 347
of I, 10
δ∗ , dual defect, 342
K(E), 79
φ∗ , pullback by φ, 406
MN ×N , 19
R: Riemann curvature, 69 φ∗ , pushforward by φ, 407
S k V , 384 κg , 54
T M , tangent bundle, 405 κn , 56
T ∗ M , cotangent bundle, 405 τ (X), tangential variety, 106
Tx M , tangent space, 405 τ (Y, X), 350
Γβα,i , 256 τg , 56
Λk V , 384 ASO(2), 13
E, 322 ASO(3), 25
Q, 31 as space of frames, 25
S, 31 Ann(v), 348
T tractor bundle, 327 A(1) , 119
II, Euclidean second fundamental form, A(l) , 119
50 Baseloc | IIM,x |, 100
II, projective second fundamental form, Cl(V, Q), 402
97 C ∞ (M ), 405
| IIM,x |, 100 End(V ), 382

445
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 446 Index

E6 , exceptional Lie group, 351 (p, q)-forms, 413

E7 , exceptional Lie group, 353 sk
E3 , Euclidean three-space, 2 characters of a tableau, 127
FFk , 335 characters of an EDS, 231
| FFk |, 335 [X, Y ], 406
F4 , differential invariant, 356 , 69
F4 , exceptional Lie group, 351 , interior product, 385
Fk , 357 ∇, 256
F (M ), 65 ⊗, tensor product, 382
F1 { }, linear span, 11
Euclidean, 41 { }alg , 11
projective, 97 { }diff , 11
GL(V ), 386 , 69
G(n, m), 170 , Kulkarni-Nomizu product, 85
G(k, V ), Grassmannian, 28 e-structure, 280
G2 , exceptional Lie group, 394
Gr(k, V ), orthogonal Grassmannian, 89 abuse of notation, 33, 96, 142
G(n, T Σ), 149 adjoint representation, 391
of Lie algebra, 390
Holθu , 263
affine connection, 260
Hom(V, W ), 382
affine tangent space, 96
H 0,2 (A), 148
algebraic variety, 102
H i,j (A), 152
degree of, 102
Hi,j (g), 259
dimension of, 102
(I, J), linear Pfaffian system, 135
general point of, 102
I k , k-th homogeneous component of I,
ideal of, 102
10
almost complex manifold, 247, 251, 412
I (1) , derived system, 188
almost complex structure, 412
I, differential ideal, 10
almost Einstein, 328
J(Y, Z), join of varieties, 106
almost scalar constant, 328
LX , Lie derivative, 408
almost symplectic manifold, 247
O(V, Q), orthogonal group, 387
Ambrose-Singer Theorem, 265
SL(V ), SLn , special linear group, 387
apparent torsion, 137
SO(V, Q), special orthogonal group, 387 arclength parameter, 15
SU (n), special unitary group, 389 associated hypersurface, 366
S 2 V , 383 associated varieties, 365
Singloc | IIM,x |, 100 associative submanifolds, 173, 238
Sp(V, ω), symplectic group, 387 associator, 396
Sm , spinor variety, 355 asymptotic directions, 100
T(V ), 247 asymptotic line, 56, 199, 211
Tx∗ M , cotangent space, 405
U (n), unitary group, 389 Bäcklund transformations, 208–213, 304
VC , complexification of V , 411 Bäcklund’s Theorem, 211
Xsmooth , 102 basic differential form, 409
[T Hθ ], 254 Bertini Theorem, 361
ck , codimension of polar space, 229 higher-order, 361
detA , 351 Bertrand curve, 28
det, 385 Bezout’s Theorem, 102
dk , 335 Bianchi identity
d, exterior derivative, 407 first, 252
g, Lie algebra of Lie group G, 18 first, 77
mx , functions vanishing at x, 405 second, 80
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Index 447

Bonnet surface, 48, 204 of tableau, 127

Born-Infeld system, 305 Chebyshev net, 200
Burger’s equation, 180, 205 Christoffel symbols, 256
Clifford algebras, 402
calibrated submanifold, 171 fundamental lemma of, 402
calibration, 169 Clifford torus, 82
associative, 173 co-roots, 400
Cayley form, 174 coassociative submanifold, 174
coassociative, 174 Codazzi equation
special Lagrangian, 172 for Darboux frames, 46
canonical system matrix form, 54
on Grassmann bundle, 149 codimension, 217
on space of jets, 33
coisotropic hypersurface, 366
Cartan
cominuscule variety, 353
five variables paper, 177, 189
complete intersection, 378
Cartan geometry, 273
complex characteristic variety, 130
Cartan integer, 128, 151
complex contact structure, 416
Cartan Lemma, 385
complex manifold, 411, 412
Cartan system, 182
complex structure, 388, 412
Cartan’s algorithm for linear Pfaffian
complexification
systems, 150
of a real vector space, 388
Cartan’s Test, 229
cone, 48
Cartan-Dieudonné Theorem, 401
characterization of, 367
Cartan-Hadamard theorem, 79
Cartan-Janet Theorem, 164 over a variety, 106
Cartan-Kähler Theorem, 226–229 conformal Killing vector field, 319
for linear Pfaffian systems, 148 conformal Laplacian, 324
for tableaux, 128 conformally flat, 93
Goldschmidt version, 153 connection
catenoid, 47 affine, 260
Cauchy characteristics, 177 on coframe bundle, 248–254
quotient by, 183 on induced vector bundles, 258
Cauchy problem, 419 on vector bundle, 256
Cauchy-Kowalevski form, 420 symmetric, 259
Cauchy-Kowalevski Theorem, 215, 421 torsion free, 72
Cauchy-Riemann equations, 415 connection form, 66, 249
tableau, 116, 128 conormal space, of submanifold in PN ,
Cayley plane, 352 96
Cayley submanifold, 174 contact manifold, 38
character of a tableau, 128 contact system
characteristic hyperplane, 153 on space of jets, 33
characteristic systems (Monge), 185 contact, order of, 103
characteristic variety, 129 cotangent
dimension and degree of, 131 bundle, 405
characteristics space, 405
Cauchy, 177, 231 Cotton-York tensor, 316, 318
confounded, 186 covariant derivative
first-order, 186 definition of, 70
method of, 179–181 covariant differential operator, 256
Monge, 186, 287 cubic form, 332
characters, 231 curvature
of linear Pfaffian system, 151 Gauss, 42
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 448 Index

geometric interpretation of, 51 dual basis, 381

in coordinates, 4 dual variety, 107, 341
mean, 42 defect of, 342
geometric interpretation of, 63 reflexivity, 342
in coordinates, 4 dual vector space, 381
of G-structure, 252 Dupin
of curve in E2 , 15 cyclides of, 431
of curve in E3 , 27 theorem of, 225
Ricci, 79, 235
scalar, 79, 235, 239 Einstein
sectional, 79 almost, 328
curvature-line coordinates, 160 Einstein manifold, 82
curve Einstein metric
arclength parameter, 15 in conformal class, 310
Bertrand, 28 embedded tangent space, 96
regular, 14 Engel structure, 189
speed of, 15 equivalent
curve in E2 G-structures, 248
curvature, 15 webs, 242
osculating circle, 16 Euclidean group, 25
curve in E3 Euler characteristic, 58
curvature, 27 exceptional group
differential invariants, 26–28 E6 , 351
torsion, 27 E7 , 353
cylinder, 48 F4 , 351
exterior algebra, 401
Darboux exterior derivative, 407–408
method of, 190–195 exterior differential system, 34
Darboux frame, 46 decomposable, 287
Darboux semi-integrable, 194, 197 hyperbolic, 187–188, 286
Darboux’s Theorem, 36 linear Pfaffian, 136
Darboux-integrable, 190, 212 Pfaffian, 11
decomposable EDS, 289 symmetries, 176–177
Goursat-Vessiot classification, 194, with independence condition, 31
286
de Rham Splitting Theorem, 265 face of calibration, 171
decomposable EDS, 287 first Bianchi identity, 77
decomposable tensor, 382 first fundamental form (Riemannian),
derived flag, 188 50
derived system, 188 first-order adapted frames (Euclidean),
determinant 49
of linear endomorphism, 385 flag
developable surface, 44 A-generic, 127
differential form, 406 complete, 105
basic, semi-basic, 409 partial, 105
closed, 408 flag variety, 105
homogeneous, 10 flat
left-invariant, 18 G-structure, 248
vector-valued, 408 3-web, 242
differential ideal, 10 path geometry, 272
differential invariant Riemannian manifold, 69
Euclidean, 3 isometric immersions of, 166
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Index 449

surface, 45 geodesic disk, 56

flow of a vector field, 6 geodesic torsion, 56
flowbox coordinates, 6 Goursat-Vessiot classification, 194, 286
flowchart for Cartan’s algorithm, 150 Grassmann bundle, 149
focal hypersurface, 109 canonical system on, 149
focal surface, 210, 239 Grassmannian, 28
frame isotropic, 104
Darboux, 46 tangent space of, 29
frame bundle
general, 65 half-spin representation, 356
orthonormal, 66 Hartshorne’s conjecture, 378
Frenet equations, 27 heat equation, 420
Frobenius ideal, 11, 13 helicoid, 43
Frobenius system Hermitian form, 389
tableau of, 118 Hermitian inner product, 389
Frobenius Theorem, 11–13, 34 hexagonality, 245
proof, 34 higher associated hypersurface, 366
Fubini cubic form, 332 holomorphic map, 413
Fubini forms, 332, 356 holonomy, 262–271
Fulton-Hansen Theorem, 349 holonomy bundle, 263
fundamental form holonomy group, 263
k-th, 335 homogeneous Riemannian manifold, 90
effective calculation of, 335 homogeneous space, 16
prolongation property of, 334 Hopf conjecture, 80
via spectral sequences, 336 Hopf differential, 203
fundamental lemma Hopf fibration, 91
for general frames, 73 horizontal curve, 262
horizontal lift, 262
Gauss curvature hyperbolic EDS, 187–188, 286
geometric interpretation of, 51 hyperbolic space, 80
in coordinates, 4 isometric immersions of, 169
via frames, 40–42 hyperplane section of a variety, 108
Gauss equation, 52 hypersurfaces in EN
Gauss image, 97 fundamental theorem for, 76
characterization of, 371
Gauss map ideal
algebraic, 75 algebraic, 10
Euclidean, 50 differential, 10
projective, 97 Frobenius, 11, 13
varieties with degenerate, 109 incidence correspondence, 108
Gauss’ theorema egregium, 53 independence condition, 31
Gauss-Bonnet formula, 60 index of a vector field, 57
Gauss-Bonnet Theorem, 58 index of relative nullity, 100
for compact hypersurfaces, 74 induced representation, 392
local, 56 induced vector bundle, 257
Gauss-Bonnet-Chern Theorem, 74 initial data, 419
general point, 102 initial value problem, 419
generalized Monge system, 377 integrable extension, 204–208, 286, 290
generic point, 102 via conservation law, 206
geodesic, 54 integral
of affine connection, 260 intermediate/general, 191
geodesic curvature, 54 integral curve, 5
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 450 Index

integral element, 31 line congruence, 210

Kähler-ordinary, 217 line of curvature, 56, 225
Kähler-regular, 221 isothermal coordinates along, 160
ordinary, 229 linear map, 381
integral manifold, 31, 34 transpose/adjoint of, 382
interior product, 385 linear normality
inversion in sphere, 320 Zak’s theorem on, 346
involutive linear Pfaffian systems, 136
integral element, 229 Cartan’s algorithm for, 150
linear Pfaffian system, 149 involutive, 149
tableau, 127 linear projection of variety, 108
isometric embedding, 141–145 linear syzygy, 359
isothermal coordinates, 93 Liouville’s equation, 190–191, 210, 305
existence of, 157 locally ruled variety, 109
isotropic Grassmannian, 104 locally symmetric manifold, 267
isotropy representation, 17
majorants, 122
Jacobi identity, 390 manifold
jets, 32 contact, 38
join of varieties, 106 locally symmetric, 267
restraining, 228
Kähler manifold, 171 symplectic, 36
KdV equation, 207, 209 matrix Lie groups, 386–388
prolongation algebra, 208 Maurer-Cartan equation, 20
Killing form, 393 Maurer-Cartan form, 18
Killing vector field, 86 of a matrix Lie group, 19
Kulkarni-Nomizu product, 85 maximal torus, 398
mean curvature
Laplace system geometric interpretation of, 63
tableau for, 129 in coordinates, 4
Laplace’s equation, 196 via frames, 40–42
Laplacian, 92 mean curvature vector, 64
conformal, 324 metric
Lawson conjecture, 82 conformally flat, 93
left action, 16 minimal hypersurfaces, 239
left-invariant minimal submanifold, 169
differential form, 18 minimal surface, 63, 201–202
vector field, 18, 391 Riemannian metric of, 158
level, 127 minimizing submanifold, 169
Lie algebra, 390 Minkowski space, 80
of a Lie group, 18 modified KdV equation, 207
semi-simple, 397 Monge’s method, 197
simple, 397 Monge-Ampère
Lie bracket, 406 equation, 195
Lie derivative, 408 system, 196
Lie group, 386 Monge-integrable, 197, 286
action, 299–300 moving frame, 4
exponential map, 299 adapted, 13
linear representation of, 386 multilinear, 382
matrix, 19, 386–388 multiplicity of intersection, 103
lift, 8, 17 musical isomorphism, 69
first-order adapted, 40 Myers’ Theorem, 79
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Index 451

Newlander-Nirenberg Theorem, 413 pushforward, 407

Nijenhuis tensor, 414
noncharacteristic initial data, 129 rank
nondegenerate quadratic form, 392 of a Lie algebra, 398
normal bundle, 50, 61 of a Pfaffian system, 11
normal curvature, 56 of a tensor, 383
normal space, of submanifold in PN , 96 rational homogeneous variety, 103
reduction
octonions, 395–396 of frame bundle, 248
orthogonal Grassmannian, 89 reductive
orthogonal group, 387 Lie group/Lie algebra, 397
orthogonal involutive Lie algebra, 268 refined third fundamental form, 348
osculating circle, 16 regular curve, 14
osculating hypersurface, 358, 360 regular second-order PDE, 146
osculating quadric hypersurface, 358 relative tangent star, 350
representation
parabolic group, 312 induced, 392
parabolic subgroup, 103, 353 isotropy, 17
parallel surfaces, 198 of Lie algebra, 390
parallel transport, 262 of Lie group, 386
path geometry, 271–284 restraining manifold, 228
definition of, 272, 274 retracting space, 182
dual, 273 Ricci curvature, 79, 235
flat, 272 Riemann curvature tensor, 68–69
Pfaff rank, 37 Riemann invariant, 190
Pfaff’s Theorem, 37 Riemann surface, 414
Pfaffian, 392 Riemannian geometry, 246
Pfaffian system, 11 fundamental lemma, 67–68
linear, 136 Riemannian manifold, 52
Picard’s Theorem, 5, 10 flat, 69
Pieri rule, 84 homogeneous, 90
Poincaré-Hopf Theorem, 58 Riemannian metric, 50, 52
point transformation, 271 right G-bundle, 247
polar spaces, 218–220 right action, 16
principal curvatures, 42 root, 399
principal framing, 46 root system, 399
principal symbol, 117 ruled surface, 45
projective differential invariants ruled variety, 339
in coordinates, 357
projective second fundamental form, 97 scalar curvature, 79, 235, 239
coordinate description of, 101 Schouten tensor, 85
frame definition of, 99 Schur’s Lemma, 387
projective structure, 260 Schwarzian derivative, 23
prolongation, 119, 149, 187, 193 secant defect, 348
of a G-structure, 254 secant variety, 106
prolongation property, 334 second Bianchi identity, 80
strict, 354 for a G-structure, 268
prolongation structures, 206 second fundamental form
pseudospherical surfaces, 199–201 base locus of, 100
Bäcklund transformation, 210 Euclidean, 50
of revolution, 200 projective, 97
pullback, 406 singular locus of, 100
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 452 Index

second-order PDE of revolution, 45, 200

characteristic variety, 154 parallel, 198
classical notation, 146 pseudospherical, 199
tableau, 147 ruled, 45
section warp of, 4
of vector bundle, 405 with degenerate Gauss image, 112
sectional curvature, 79 symbol mapping, 129
Segre product of varieties, 104 symbol relations, 117, 146
fundamental forms of, 339 symmetric connection, 259
Segre variety, 104, 131 symmetric space, 266, 267
fundamental forms of, 338 symplectic form, 36, 157, 171, 184, 237,
semi-basic form, 409 387
semi-Riemannian manifold, 247 symplectic group, 387
semi-simple Lie algebra, 397 symplectic manifold, 36
Severi variety, 350, 352
tableau, 117
fundamental form of, 352
determined, 131
Zak’s theorem on, 346
of linear Pfaffian system, 146
signature
of order p, 119
of quadratic form, 392
tangent
simple Lie algebra, 397
bundle, 405
sine-Gordon equation, 196, 199, 208
space, 405
singular solutions, 163
tangent star, 106
space form, 80
tangential defect, 347
isometric immersions of, 166 critical, 374
special Lagrangian submanifolds, 172, tangential surface, 44
237 tangential variety, 106
special linear group, 387 dimension of, 347
special orthogonal group, 387 tautological EDS
special unitary group, 389 for torsion-free G-structures, 269
Spencer cohomology, 152 tautological form
spin representation, 354, 356 for coframe bundle, 66
spinor variety, 104, 354 tensor algebra, 401
stabilizer type, 253 tensor product, 382
submanifold Terracini’s Lemma, 107
associative, 238 third fundamental form
Lagrangian, 157, 237 projective, 333
special Lagrangian, 172, 237 Toda lattice (2-dimensional), 288
superposition formula, 285, 300–305 torsion
surface of G-structure, 250
Bonnet, 48 of connection, 250
catenoid, 47 of curve in E3 , 27
cone, 48 of linear Pfaffian system, 137, 147
constant mean curvature, 202–204 torsion-free connection, 72
cylinder, 48 tractor bundle, 327
developable, 44 transformation
flat, 45 Bäcklund, 205, 209, 304
focal, 210, 239 Cole-Hopf, 205, 212
helicoid, 43 fractional linear, 22
isothermal coordinates on, 93 Lie, 204
linear Weingarten, 155, 197, 234 Miura, 207
minimal, 63, 201 triangulation, 57
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Index 453

triply orthogonal systems, 223–226 Yamabe operator, 324

Yamabe problem, 82
umbilic point, 42 Young symmetrizer, 83
uniruled variety, 339 Young tableau
unitary group, 389 without repetitions, 83

variation of Hodge structure, 161 Zak’s theorem

variety on linear normality, 346
algebraic, 102 on Severi varieties, 346
dual, 107, 341 on tangencies, 350
flag, 105
minuscule, 353
rational homogeneous, 103
ruled, 339
secant, 106
Segre, 104
spinor, 104, 354
tangential, 106
uniruled, 339
Veronese, 105
vector bundle
induced, 257
vector field, 405
flow of a, 6
left-invariant, 18
Veronese embedding, 105
Veronese re-embedding, 105, 358
Veronese variety, 105
fundamental forms of, 337
vertical tangent space, 67
vertical vector, 409
Vessiot algebra, 296
volume form, 50

Waring problems, 383

warp of a surface, 4
wave equation, 175, 419
wave maps, 288, 306
web, 242
hexagonality of, 245
wedge product, 384
matrix, 19
Weierstrass representation, 201–202
weight, 398
highest, 400
multiplicity of, 398
weight diagram for invariants, 281
weight lattice, 400
weight zero invariant, 277
Weingarten surface, 197
linear, 155, 197, 234
Wirtinger inequality, 172
Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 SELECTED PUBLISHED TITLES IN THIS SERIES

175 Thomas A. Ivey and Joseph M. Landsberg, Cartan for Beginners: Differential
Geometry via Moving Frames and Exterior Differential Systems, Second Edition, 2016
174 Alexander Kirillov Jr., Quiver Representations and Quiver Varieties, 2016
173 Lan Wen, Differentiable Dynamical Systems, 2016
172 Jinho Baik, Percy Deift, and Toufic Suidan, Combinatorics and Random Matrix
Theory, 2016
171 Qing Han, Nonlinear Elliptic Equations of the Second Order, 2016
170 Donald Yau, Colored Operads, 2016
169 András Vasy, Partial Differential Equations, 2015
168 Michael Aizenman and Simone Warzel, Random Operators, 2015
167 John C. Neu, Singular Perturbation in the Physical Sciences, 2015
166 Alberto Torchinsky, Problems in Real and Functional Analysis, 2015
165 Joseph J. Rotman, Advanced Modern Algebra: Third Edition, Part 1, 2015
164 Terence Tao, Expansion in Finite Simple Groups of Lie Type, 2015
163 Gérald Tenenbaum, Introduction to Analytic and Probabilistic Number Theory, Third
Edition, 2015
162 Firas Rassoul-Agha and Timo Seppäläinen, A Course on Large Deviations with an
Introduction to Gibbs Measures, 2015
161 Diane Maclagan and Bernd Sturmfels, Introduction to Tropical Geometry, 2015
160 Marius Overholt, A Course in Analytic Number Theory, 2014
159 John R. Faulkner, The Role of Nonassociative Algebra in Projective Geometry, 2014
158 Fritz Colonius and Wolfgang Kliemann, Dynamical Systems and Linear Algebra,
2014
157 Gerald Teschl, Mathematical Methods in Quantum Mechanics: With Applications to
Schrödinger Operators, Second Edition, 2014
156 Markus Haase, Functional Analysis, 2014
155 Emmanuel Kowalski, An Introduction to the Representation Theory of Groups, 2014
154 Wilhelm Schlag, A Course in Complex Analysis and Riemann Surfaces, 2014
153 Terence Tao, Hilbert’s Fifth Problem and Related Topics, 2014
152 Gábor Székelyhidi, An Introduction to Extremal Kähler Metrics, 2014
151 Jennifer Schultens, Introduction to 3-Manifolds, 2014
150 Joe Diestel and Angela Spalsbury, The Joys of Haar Measure, 2013
149 Daniel W. Stroock, Mathematics of Probability, 2013
148 Luis Barreira and Yakov Pesin, Introduction to Smooth Ergodic Theory, 2013
147 Xingzhi Zhan, Matrix Theory, 2013
146 Aaron N. Siegel, Combinatorial Game Theory, 2013
145 Charles A. Weibel, The K-book, 2013
144 Shun-Jen Cheng and Weiqiang Wang, Dualities and Representations of Lie
Superalgebras, 2012
143 Alberto Bressan, Lecture Notes on Functional Analysis, 2013
142 Terence Tao, Higher Order Fourier Analysis, 2012
141 John B. Conway, A Course in Abstract Analysis, 2012
140 Gerald Teschl, Ordinary Differential Equations and Dynamical Systems, 2012
139 John B. Walsh, Knowing the Odds, 2012
138 Maciej Zworski, Semiclassical Analysis, 2012
137 Luis Barreira and Claudia Valls, Ordinary Differential Equations, 2012

For a complete list of titles in this series, visit the

Licensed to AMS. AMS Bookstore at www.ams.org/bookstore/gsmseries/.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms
 Two central aspects of Cartan’s approach to differential geometry are the theory of
exterior differential systems (EDS) and the method of moving frames. This book pres-
ents thorough and modern treatments of both subjects, including their applications
to both classic and contemporary problems in geometry. It begins with the classical
differential geometry of surfaces and basic Riemannian geometry in the language of
moving frames, along with an elementary introduction to exterior differential systems.
Key concepts are developed incrementally, with motivating examples leading to defini-
tions, theorems, and proofs.
Once the basics of the methods are established, the authors develop applications and
advanced topics. One notable application is to complex algebraic geometry, where
they expand and update important results from projective differential geometry. As
well, the book features an introduction to G-structures and a treatment of the theory
of connections. The techniques of EDS are also applied to obtain explicit solutions of
PDEs via Darboux’s method, the method of characteristics, and Cartan’s method of
equivalence.
This text is suitable for a one-year graduate course in differential geometry, and parts
of it can be used for a one-semester course. It has numerous exercises and examples
throughout. It will also be useful to experts in areas such as geometry of PDE systems
and complex algebraic geometry who want to learn how moving frames and exterior
differential systems apply to their fields.
The second edition features three new chapters: on Riemannian geometry, empha-
sizing the use of representation theory; on the latest developments in the study of
Darboux-integrable systems; and on conformal geometry, written in a manner to
introduce readers to the related parabolic geometry perspective.

For additional information

and updates on this book, visit
www.ams.org/bookpages/gsm-175

www.ams.org
GSM/175

Licensed to AMS.
License or copyright restrictions may apply to redistribution; see https://www.ams.org/publications/ebooks/terms