Lie Groups and Geometric Aspects

of Isometric Actions
Marcos M. Alexandrino • Renato G. Bettiol

Lie Groups and Geometric

Aspects of Isometric Actions

123
Marcos M. Alexandrino Renato G. Bettiol
Departamento de Matemática Department of Mathematics
Instituto de Matemática e Estatística University of Pennsylvania
Universidade de São Paulo Philadelphia, PA, USA
São Paulo, Brazil

ISBN 978-3-319-16612-4 ISBN 978-3-319-16613-1 (eBook)

DOI 10.1007/978-3-319-16613-1

Library of Congress Control Number: 2015936906

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media (www.

springer.com)
Dedicated to our families.
Preface

This book is intended for advanced undergraduates, graduate students, and young
researchers in geometry. It was written with two main goals in mind. First, we give a
gentle introduction to the classical theory of Lie groups, using a concise geometric
approach. Second, we provide an overview of topics related to isometric actions,
exploring their relations with the research areas of the authors and giving the main
ideas of proofs. We discuss recent applications to active research areas, such as
isoparametric submanifolds, polar actions and polar foliations, cohomogeneity one
actions, and positive curvature via symmetries. In this way, the text is naturally
divided in two interrelated parts.
Let us give a more precise description of such parts. The goal of the first part
(Chaps. 1 and 2) is to introduce the concepts of Lie groups, Lie algebras and
adjoint representation, relating these objects. Moreover, we give basic results on
closed subgroups, bi-invariant metrics, Killing forms, and splitting of Lie algebras
in simple ideals. This is done concisely due to the use of Riemannian geometry,
whose fundamental techniques are also quickly reviewed.
The second part (Chaps. 3–6) is slightly more advanced. We begin with some
results on proper and isometric actions in Chap. 3, presenting a few research
comments. In Chap. 4, classical results on adjoint and conjugation actions are
presented, especially regarding maximal tori, roots of compact Lie groups, and
Dynkin diagrams. In addition, the connection with isoparametric submanifolds and
polar actions is explored. In Chap. 5, we survey on the theory of polar foliations,
which generalizes some of the objects studied in the previous chapter. Finally,
Chap. 6 briefly discusses basic aspects of homogeneous spaces and builds on all
the previous material to explore the geometry of low cohomogeneity actions and its
interplay with manifolds with positive (and nonnegative) sectional curvature.
Prerequisites expected from the reader are a good knowledge of advanced
calculus and linear algebra, together with rudiments of calculus on manifolds.
Nevertheless, a brief review of the main definitions and essential results is given
in the Appendix A.
This book can be used for a one-semester graduate course (of around 3 h per
week) or an individual study, as it was written to be as self-contained as possible.

vii
viii Preface

Part of the material in Chap. 3, as well as Chaps. 5 and 6, may be skipped by students
in a first reading. Most sections in the book are illustrated with several examples,
designed to convey a geometric intuition on the material. These are complemented
by exercises that are usually accompanied by a hint. Some exercises are labeled
with a star (), indicating that they are slightly more involved than the others. We
encourage the reader to think about them, in an effort to develop a good working
knowledge of the material and practice active reading.
The present book evolved from several lecture notes that we used to teach
graduate courses and minicourses. In 2007, 2009, and 2010, graduate courses on
Lie groups and proper actions were taught at the University of São Paulo, Brazil,
exploring mostly the first four chapters of the text. Graduate students working in
various fields followed these courses, with very positive results. During this period,
the same material was used in a graduate course at the University of Parma, Italy.
Relevant contributions also originated from short courses given by the authors
during the XV Brazilian School of Differential Geometry (Fortaleza, Brazil, July
2008), the Second São Paulo Geometry Meeting (São Carlos, Brazil, February
2009), and the Rey Pastor Seminar at the University of Murcia (Murcia, Spain, July
2009). In 2009, a preliminary draft of this text was posted on the arXiv (0901.2374),
which prompted instructors in various universities to list it as complementary
study material. Since then, we have substantially improved and updated the text,
particularly the last chapters, featuring many recent advances in the research areas
discussed.
There are several important research areas related to the content of this book that
are not treated here. We would like to point out two of these, for which we hope to
give the necessary background: first, representation theory and harmonic analysis,
for which we recommend Bröken and Tom Dieck [56], Deitmar [78], Fulton and
Harris [90], Gangolli and Varadarajan [94], Helgason [125], Katznelson [136],
Knapp [144], and Varadarajan [217], and, second, symmetries in differential equa-
tions and integrable systems, for which we recommend Bryant [57], Fehér and
Pusztai [86, 87], Guest [119], Noumi [175], and Olver [176].

Acknowledgements We acknowledge financial support from CNPq and FAPESP (Brazil) and
NSF (USA). The authors are very grateful to Alexander Lytchak, Gudlaugur Thorbergsson,
Karsten Grove, Paolo Piccione, and Wolfgang Ziller for their constant support and several valuable
discussions. It is also a pleasure to thank Augusto Ritter Stoffel, Daniel Victor Tausk, Dirk Töben,
Fábio Simas, Flausino Lucas Spindola, Francisco Carlos Junior, Ion Moutinho, Ivan Struchiner,
John Harvey, László Fehér, Leandro Lichtenfelz, Leonardo Biliotti, Martin Kerin, Martin Weilandt,
Rafael Briquet, Renée Abib, and Ricardo Mendes for their helpful comments and suggestions
during the elaboration of this book.

São Paulo Marcos M. Alexandrino

Philadelphia Renato G. Bettiol
June 2015
Contents

Part I Lie Groups

1 Basic Results on Lie Groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Lie Groups and Lie Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Lie Subgroups and Lie Homomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Exponential Map and Adjoint Representation . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Closed Subgroups and More Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2 Lie Groups with Bi-invariant Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1 Basic Facts of Riemannian Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Bi-invariant Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Killing Form and Semisimple Lie Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4 Splitting Lie Groups with Bi-invariant Metrics . . . . . . . . . . . . . . . . . . . . . . . 45

Part II Isometric Actions

3 Proper and Isometric Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Proper Actions and Fiber Bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 Slices and Tubular Neighborhoods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Isometric Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4 Principal Orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.5 Orbit Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4 Adjoint and Conjugation Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.1 Maximal Tori and Polar Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 Normal Slices of Conjugation Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Roots of a Compact Lie Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.4 Weyl Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.5 Dynkin Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5 Polar Foliations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1 Definitions and First Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 Holonomy and Orbifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

ix
x Contents

5.3 Surgery and Suspension of Homomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . 116

5.4 Differential and Geometric Aspects of Polar Foliations . . . . . . . . . . . . . . 117
5.5 Transnormal and Isoparametric Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.6 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6 Low Cohomogeneity Actions and Positive Curvature . . . . . . . . . . . . . . . . . . . 139
6.1 Cheeger Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.2 Compact Homogeneous Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.3 Cohomogeneity One Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.4 Positive and Nonnegative Curvature via Symmetries. . . . . . . . . . . . . . . . . 171

A Rudiments of Smooth Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

A.1 Smooth Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
A.2 Vector Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
A.3 Foliations and the Frobenius Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
A.4 Differential Forms, Integration, and de Rham Cohomology . . . . . . . . . 192

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
 Part I
Lie Groups
Chapter 1
Basic Results on Lie Groups

This chapter gives an introduction to Lie group theory, presenting the main concepts
and giving detailed proofs of basic results. Some knowledge of group theory, linear
algebra and advanced calculus is assumed. However, as a service to the reader, a
few facts about differentiable manifolds are recalled in Appendix A, which can be
used as preliminary reading.
The following references, which inspire our approach to the subject, can be used
for further reading material on the contents of this chapter: Bump [58], Duistermaat
and Kolk [79], Fegan [85], Gilmore [97], Gorbatsevich et al. [88, 98–100], Helga-
son [126], Hilgert and Neeb [128], Hsiang [129], Knapp [145], Onishchik [179],
Spivak [198], Varadarajan [216] and Warner [227].

1.1 Lie Groups and Lie Algebras

Definition 1.1. A smooth (respectively, analytic) manifold G is said to be a smooth

(respectively, analytic) Lie group1 if G is a group and the maps

G × G  (x, y) −→ x y ∈ G (1.1)

G  x −→ x−1 ∈ G (1.2)

are smooth (respectively, analytic).

1 Sophus Lie was a nineteenth century Norwegian mathematician, who laid the foundations of
continuous symmetry groups and transformation groups.

© Springer International Publishing Switzerland 2015 3

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_1
4 1 Basic Results on Lie Groups

Remark 1.2. The requirements that (1.1) and (1.2) be smooth, or analytic, often
appear as a requirement that the map G × G  (x, y) −→ x y−1 ∈ G be smooth, or
analytic. It is easy to prove that these conditions are equivalent. See also Exer-
cise 1.50.
In this book, we only deal with smooth Lie groups. Nevertheless, we mention the
following important result, which is explored in detail in Duistermaat and Kolk [79].
Theorem 1.3. Each C2 Lie group admits a unique analytic structure, turning G into
an analytic Lie group.
We stress that every result proved in this chapter on (smooth) Lie groups is hence
automatically valid on analytic Lie groups. Henceforth, G denotes a smooth Lie
group, and the word smooth is omitted.
Remark 1.4. Historically, an important problem was to determine if a connected
locally Euclidean topological group has a smooth structure. This problem was
known as 5th Hilbert’s problem, posed by Hilbert at the International Congress
of Mathematicians in 1900, and solved by von Neumann in 1933 in the compact
case. Only in 1952 the general case was solved, by Gleason, Montgomery and
Zippen [168].
Some of the most basic examples of Lie groups are (Rn , +), (S1 , ·) with group
operation eiθ · eiη = ei(θ +η ) , and the n-torus T n = S1 × · · · × S1 as a product group.
Example 1.5. More interesting examples are the so-called classical Lie groups,
which form four families of matrix Lie groups, closely related to symmetries of
Euclidean spaces. The term classical appeared in 1940 in Weyl’s monograph,
probably referring to the classical geometries in the spirit of Klein’s Erlangen
program.
We begin with GL(n, R), the general linear group of nonsingular (i.e., invertible)
n × n real matrices. Similarly, GL(n, C) and GL(n, H) are respectively the groups
of nonsingular n × n matrices over the complex numbers and the quaternions.2
Furthermore, the following complete the list of classical Lie groups, where I denotes
the identity matrix:
(i) SL(n, R) = {M ∈ GL(n, R) : det M = 1}, SL(n, C) and SL(n, H), the special
linear groups;
(ii) O(n) = {M ∈ GL(n, R) : M t M = I}, the orthogonal group;
(iii) SO(n) = O(n) ∩ SL(n, R), the special orthogonal group;

2 The quaternions form a real normed division algebra, which provides a noncommutative
extension of complex (and real) numbers. Quaternions were discovered by Sir William Hamilton,
after whom the usual notation is H. The algebra H is a 4-dimensional real vector space, endowed
with the quaternion multiplication. Its orthonormal basis is denoted (1, i, j, k), and the product
of basis elements is given by i2 = j2 = k2 = ijk = −1. Any elements of H are of the form
a + b i + c j + d k ∈ H, and their product is determined by the equations above and the distributive
law. Conjugation, norm and division can be defined as natural extensions of those in C.
1.1 Lie Groups and Lie Algebras 5

(iv) U(n) = {M ∈ GL(n, C) : M ∗ M = I}, the unitary group;

(v) SU(n) = U(n) ∩ SL(n, C), the special unitary group;
(vi) Sp(n) = {M ∈ GL(n, H) : M ∗ M = I}, the symplectic group.
In order to verify that those are indeed Lie groups, see Exercise 1.51. For now,
we encourage the reader to keep them in mind as important examples of Lie groups.
Remark 1.6. Another class of examples of Lie groups is constructed by quotients of
Lie groups by their normal and closed subgroups (see Corollary 3.38). In this class
of examples, there are Lie groups that are not matrix groups. In fact, consider
⎧⎛ ⎞ ⎫ ⎧⎛ ⎞ ⎫
⎨ 1ab ⎬ ⎨ 10n ⎬
G := ⎝0 1 c⎠ : a, b, c ∈ R and N := ⎝0 1 0⎠ : n ∈ Z .
⎩ ⎭ ⎩ ⎭
001 001

Then G/N is a Lie group. It is possible to prove that there are no injective
homomorphisms ϕ : G/N → Aut(V ), where Aut(V ) denotes the group of linear
automorphisms of a finite-dimensional vector space V (see Carter, Segal and
MacDonald [66]).
Definition 1.7. A Lie algebra g is a real vector space endowed with a bilinear map
[·, ·] : g × g → g, called the Lie bracket, satisfying for all X,Y, Z ∈ g,
(i) Skew-symmetry: [X,Y ] = −[Y, X];
(ii) Jacobi identity: [[X,Y ], Z] + [[Y, Z], X] + [[Z, X],Y ] = 0.
Example 1.8. Basic examples of Lie algebras are the vector spaces of n × n square
matrices over R and C, respectively denoted gl(n, R) and gl(n, C), endowed with
the Lie bracket given by the matrix commutator [A, B] = AB − BA.
Exercise 1.9. Let so(3) = {A ∈ gl(3, R) : At + A = 0}.
(i) Verify so(3) is a Lie algebra, with Lie bracket given by the matrix commutator;
⎛ ⎞
0 −x3 x2
(ii) Let AX = ⎝ x3 0 −x1 ⎠. Prove that AX v = X × v.
−x2 x1 0
(iii) Verify that AX×Y = [AX , AY ] = AX AY − AY AX . Using the fact that (R3 , ×) is
a Lie algebra endowed with the cross product of vectors, conclude that the
map (R3 , ×)  X → AX ∈ so(3) is a Lie algebra isomorphism, i.e., a linear
isomorphism that preserves Lie brackets.
We now proceed to associate to each Lie group a Lie algebra, by considering
left-invariant vector fields. For each g ∈ G, denote by Lg and Rg the left and right
translation maps on G, that is,

Lg (x) := g x and Rg (x) := x g. (1.3)

6 1 Basic Results on Lie Groups

These smooth maps are easily verified to be diffeomorphisms, since they have
simultaneous left and right smooth inverses Lg−1 and Rg−1 , respectively.
A (not necessarily smooth) vector field X on G is said to be left-invariant if X
is Lg -related to itself for all g ∈ G, i.e., dLg ◦ X = X ◦ Lg . This means that X(g h) =
d(Lg )h X(h), or shortly X = dLg X , for all g ∈ G. Similarly, a vector field is right-
invariant if it is Rg -related to itself for all g ∈ G, meaning that X = dRg X for all g ∈
G. A simultaneously left- and right-invariant vector field is said to be bi-invariant.
Lemma 1.10. Left-invariant vector fields are smooth.
Proof. Let X be a left-invariant vector field on G, and consider the group operation

μ : G × G → G, μ (g, h) = g h.

Differentiating μ , we obtain dμ : T (G×G) ∼

= TG×TG → TG, a smooth map. Define
s : G → TG×TG by s(g) := (0g , X(e)), where g → 0g is the null section of TG. Since
X = dμ ◦ s, it follows that X is smooth. 

Remark 1.11. Clearly, the above result also holds for right-invariant vector fields.
Given two Lie algebras g1 and g2 , a linear map ψ : g1 → g2 is called Lie algebra
homomorphism if

[ψ (X), ψ (Y )] = ψ ([X,Y ]), for all X,Y ∈ g1 .

Theorem 1.12. Let g be the set of left-invariant vector fields on the Lie group G.
Then the following hold:
(i) g is a Lie algebra, endowed with the Lie bracket of vector fields;
(ii) Consider the tangent space Te G with the bracket defined as follows. If X 1 , X 2 ∈
1 , X
Te G, set [X 1 , X 2 ] := [X 2 ] where X
i = d(Lg )e X i . Define ψ : g → G by
e g
ψ (X) := Xe . Then ψ is a Lie algebra isomorphism, where g is endowed with
the Lie bracket of vector fields and Te G with the bracket defined above.
Proof. First, note that g has a (real) vector space structure, by the linearity of d(Lg )e .
It is not difficult to see that the Lie bracket of vector fields is a Lie bracket, i.e., it is
skew-symmetric and satisfies the Jacobi identity. Equation (A.3) implies that the Lie
bracket of left-invariant vector fields is still left-invariant. Hence g is a Lie algebra,
proving (i).
To prove (ii), we first claim that ψ is injective. Indeed, if ψ (X) = ψ (Y ), for each
g ∈ G, X(g) = dLg (X(e)) = dLg (Y (e)) = Y (g). Furthermore, it is also surjective,
since for each v ∈ Te G, X (g) := dLg (v) is clearly left-invariant and satisfies
ψ (X) = v. Therefore, ψ is a linear bijection between two vector spaces, hence an
isomorphism. From the definition of Lie bracket on Te G, we have [ψ (X), ψ (Y )] =
[X,Y ]e = ψ ([X,Y ]). Thus ψ is a Lie algebra isomorphism. 

1.2 Lie Subgroups and Lie Homomorphisms 7

Definition 1.13. The Lie algebra of the Lie group G is the Lie algebra g of left-
invariant vector fields on G.
According to the above theorem, g could be equivalently defined as the tangent
space Te G, with the bracket defined as in (ii). In this way, a Lie group G gives rise
to a canonically determined Lie algebra g. A converse is given by the following
foundational result, see e.g., Duistermaat and Kolk [79].
Lie’s Third Theorem 1.14. Let g be a (finite-dimensional) Lie algebra. There
exists a unique connected and simply-connected Lie group G with Lie algebra
isomorphic to g.
We do not prove the existence of G, but its uniqueness follows from
Corollary 1.27 below. We conclude this section with an exercise that complements
Exercise 1.9, verifying that so(3) is the Lie algebra of SO(3).
Exercise 1.15. Assume the following result to be seen in Exercise 1.38: The tangent
space at the identity to a Lie subgroup of GL(n, R), endowed with the matrix
commutator, is isomorphic to its Lie algebra. Consider S ⊂ GL(3, R) the subset
of symmetric matrices, and define ϕ : GL(3, R) → S , ϕ (A) = A At .
(i) Verify that the kernel of dϕ (I) : gl(3, R) → S is the subspace of skew-
symmetric matrices in gl(3, R);
(ii) Prove that, for all A ∈ GL(3, R),

ker dϕ (A) = H ∈ gl(3, R) : A−1 H ∈ ker dϕ (I) .

Conclude that dim ker dϕ (A) = 3, for all A ∈ GL(3, R);

(iii) Prove that the identity matrix I is a regular value of ϕ ;
(iv) Conclude that the Lie group O(3) may be written as ϕ −1 (I), and compute its
dimension.
(v) Recall that SO(3) is the subgroup of O(3) determined by the connected
component of the identity I. Prove that TI SO(3) = ker dϕ (I) = so(3). Finally,
conclude that so(3) is the Lie algebra of SO(3).
Analogous statements hold for SO(n) and so(n), with identical proofs.

1.2 Lie Subgroups and Lie Homomorphisms

The aim of this section is to establish the basic relations between Lie algebras and
Lie groups, together with their natural maps and subobjects.
A group homomorphism between Lie groups ϕ : G1 → G2 is called a Lie group
homomorphism if it is smooth. In Corollary 1.49, we prove that continuity is actually
sufficient for a group homomorphism between Lie groups to be smooth. Recall
that, given two Lie algebras g1 , g2 , a linear map ψ : g1 → g2 is a Lie algebra
8 1 Basic Results on Lie Groups

homomorphism if ψ ([X,Y ]) = [ψ (X), ψ (Y )], for all X ,Y ∈ g1 . A Lie subgroup H of

a Lie group G is an abstract subgroup, such that H is an immersed submanifold of
G and

H × H  (x, y) −→ x y−1 ∈ H (1.4)

is smooth. In addition, if g is a Lie algebra, a subspace h is a Lie subalgebra if it is

closed with respect to the Lie bracket.
Proposition 1.16. Let G be a Lie group and H ⊂ G an embedded submanifold of
G that is also a group with respect to the group operation of G. Then H is a closed
Lie subgroup of G.
Proof. Consider the map f : H × H → G given by (1.4). Then f is smooth and
f (H × H) ⊂ H. Since H is embedded in G, from Proposition A.7, f : H × H → H
is smooth. Hence H is a Lie subgroup of G.
It remains to prove that H is a closed subgroup of G. Since H is an embedded
submanifold, there exists a neighborhood W of the identity e ∈ G, and a submanifold
chart ϕ = (x1 , . . . , xk ) : W → Rdim G , such that

H ∩W = g ∈ G ∩W : xi (g) = 0, i = 1, . . . dim H . (1.5)

Consider a sequence {hn } in H that converges to h ∈ G. Continuity of (1.4) implies

the existence of a neighborhood U of e ∈ G such that UU −1 ⊂ W . As hn converges to
h, for n sufficiently large, hn h−1 ∈ U. In particular, hn h−1
m ∈ H ∩W for n sufficiently
large. From (1.5), this means that xi (hn h−1
m ) = 0 for i = 1, . . . , dim H. Fixing m and
letting n → +∞, we have that xi (h h−1 −1
m ) = 0 for i = 1, . . . , dim H. Therefore, h hm ∈
H ∩W , hence h ∈ H, concluding the proof. 

A related result to Proposition 1.16 is proved in Sect. 1.4. We now investigate the
nature of the Lie algebra of a Lie subgroup.
Lemma 1.17. Let G1 and G2 be Lie groups and ϕ : G1 → G2 be a Lie group
homomorphism. Given any left-invariant vector field X ∈ g1 , there exists a unique
left-invariant vector field Y ∈ g2 that is ϕ -related to X.
Proof. First, if Y ∈ g2 is ϕ -related to X , since ϕ is a Lie group homomorphism,
Ye = dϕe Xe . Hence, uniqueness follows from left-invariance of Y .
Define Y := d(Lg )e (dϕe (Xe )). It remains to prove that Y is ϕ -related to X.
Observing that ϕ is a Lie group homomorphism, ϕ ◦ Lg = Lϕ (g) ◦ ϕ , for all g ∈ G1 .
Therefore, for each g ∈ G1 ,

dϕg (Xg ) = dϕg (d(Lg )e Xe )

= d(ϕ ◦ Lg )e Xe
= d(Lϕ (g) ◦ ϕ )e Xe
1.2 Lie Subgroups and Lie Homomorphisms 9

= d(Lϕ (g) )e (dϕe Xe )

= Y (ϕ (g)).

This shows that Y is ϕ -related to X, completing the proof. 

Proposition 1.18. Let G1 and G2 be Lie groups and ϕ : G1 → G2 be a Lie group

homomorphism. Then dϕe : g1 → g2 is a Lie algebra homomorphism.
Proof. We want to prove that dϕe : (Te G1 , [·, ·]) → (Te G2 , [·, ·]) is a Lie algebra
homomorphism, where the Lie bracket [·, ·] on Te Gi was defined in Theorem 1.12.
For each X 1 , X 2 ∈ Te G1 , define the vectors Y i := dϕe X i ∈ Te G2 and extend them to
left-invariant vector fields Xi ∈ X(G1 ) and Y i ∈ X(G2 ), by setting Xgi := dLg X i and
Ygi := dLg Y i .
On the one hand, it follows from the definition of the Lie bracket in Te G2 that

[dϕe X 1 , dϕe X 2 ] = [Y 1 ,Y 2 ] = [Y 1 , Y 2 ]e .

On the other hand, it follows from Lemma 1.17 that Xi and Y i are ϕ -related, and
from (A.3), [X1 , X2 ] and [Y 1 , Y 2 ] are ϕ -related. Therefore

[Y 1 , Y 2 ]e = dϕe [X1 , X2 ]e = dϕe [X 1 , X 2 ].

The above imply that [dϕe X 1 , dϕe X 2 ] = dϕe [X 1 , X 2 ], concluding the proof. 

Corollary 1.19. Let G be a Lie group and H ⊂ G a Lie subgroup. Then the
inclusion map i : H → G induces an isomorphism die between the Lie algebra h
of H and a Lie subalgebra die (h) of g.
A converse result is given below, on conditions under which a Lie subalgebra
gives rise to a Lie subgroup. Before that, however, we need a result on how an open
neighborhood of the identity generates the entire Lie group.
Proposition 1.20. Let G be a Lie group and G0 be the connected component of G
containing the identity e ∈ G. Then G0 is a normal Lie subgroup of G and connected
components of G are of the form gG0 , for some g ∈ G. Moreover, for any open

neighborhood U of e, G0 = n∈N U n , where U n := {g±1 ±1
1 · · · gn : gi ∈ U}.

Proof. Since G0 is a connected component, it is an open and closed subset of G.

In order to verify that it is also a Lie subgroup, let g0 ∈ G0 and consider g0 G0 =
Lg0 (G0 ). Note that g0 G0 is a connected component of G, as Lg0 is a diffeomorphism.
Since g0 ∈ G0 ∩ g0 G0 , it follows from the maximality of the connected component
that g0 G0 = G0 . Similarly, as the inversion map is also a diffeomorphism, the subset
G−1 −1 −1 −1
0 := {g0 : g0 ∈ G0 } is connected, with e ∈ G0 . Hence, G0 = G0 , using the
same argument. Therefore, G0 is a subgroup of G and an embedded submanifold of
G. From Proposition 1.16, it follows that G0 is a Lie subgroup of G.
10 1 Basic Results on Lie Groups

For each g ∈ G, consider the diffeomorphism given by conjugation x → g x g−1 .

Using the same argument of maximality of the connected component, one may
conclude that g G0 g−1 = G0 for all g ∈ G, hence G0 is normal. The proof that g G0
is the connected component of G containing g is completely analogous.

Finally, since G0 connected, to show that G0 = n∈N U n it suffices to check that
 n n
n∈N U is open and closed in G0 . It is clearly open, since U (hence U ) isopen. To
verify that it is also closed, let h ∈ G0 be the limit of a sequence {h j } in n∈N U n ,
i.e., lim h j = h. Since U −1 = {u−1 : u ∈ U} is an open neighborhood of e ∈ G, hU −1
is an open neighborhood of h. From the convergence of the sequence {h j }, there
exists j0 ∈ N such that h j0 ∈ hU −1 , that is, there exists u ∈ U such that h j0 = h u−1 .

Hence h = h j0 u ∈ n∈N U n . Therefore, this set is closed in G0 . 

Theorem 1.21. Let G be a Lie group with Lie algebra g, and h be a Lie subalgebra
of g. There exists a unique connected Lie subgroup H ⊂ G with Lie algebra h.
Proof. Define the distribution Dq = {Xq : Xq = dLq X for X ∈ h}. Since h is a Lie
algebra, D is involutive. It follows from the Frobenius Theorem (see Theorem A.19)
that there exists a unique foliation F = {Fq : q ∈ G} tangent to this distribution, i.e.,
Dq = Tq Fq , for all q ∈ G. Define H := Fe to be the leaf passing through the identity.
Note that Lg maps leaves to leaves, since dLg Da = Dga . Furthermore, for each
h ∈ H, Lh−1 (H) is the leaf passing through the identity. Therefore Lh−1 (H) = H,
which means that H is a group. It also follows from the Frobenius Theorem that H
is quasi-embedded. Consider ψ : H × H  (x, y) → x−1 y ∈ H. Since the inclusion
i : H → G and i ◦ ψ are smooth, ψ is also smooth. Thus, H is a connected Lie
subgroup of G with Lie algebra h. Uniqueness follows from the Frobenius Theorem
and Proposition 1.20. 

Definition 1.22. A smooth surjective map π : E → B is a covering map if, for each
p ∈ B, there exists an open neighborhood U of p such that π −1 (U ) is a disjoint
union of open sets Uα ⊂ E mapped diffeomorphically onto U by π . In other words,
π |Uα : Uα → U is a diffeomorphism for each α .
Theorem 1.23. Let G be a connected Lie group. There exist a unique simply-
 and a Lie group homomorphism π : G
connected Lie group G  → G, which is also a
covering map.
A proof of this theorem can be found in Boothby [46] or Duistermaat and
Kolk [79]. An example of covering map that is also a Lie group homomorphism is
the usual covering π : Rn → T n of the n-torus by Euclidean space. Other examples
of Lie group coverings are discussed in Exercise 1.55 and Remark 1.56.
Proposition 1.24. Let G1 and G2 be connected Lie groups and π : G1 → G2 be
a Lie group homomorphism. Then π is a covering map if and only if dπe is an
isomorphism.
Proof. Suppose that dπe1 : Te1 G1 → Te2 G2 is an isomorphism, where ei ∈ Gi denotes
the identity element of Gi . We claim that π is surjective.
1.2 Lie Subgroups and Lie Homomorphisms 11

Indeed, since dπe1 is an isomorphism, from the Inverse Function Theorem,

there exist open neighborhoods U of the identity e1 ∈ G1 and V of the identity
e2 ∈ G2 , such that π (U) = V and π |U is a diffeomorphism. Let h ∈ G2 . From
Proposition 1.20, there exist hi ∈ V such that h = h±1 ±1
1 · · · hn . Since π |U is a
diffeomorphism, for each 1 ≤ i ≤ n, there exists a unique gi ∈ U such that π (g±1
i )=
h±1
i . Thus,

π (g±1 ±1 ±1 ±1 ±1 ±1
1 · · · gn ) = π (g1 ) · · · π (gn ) = h1 · · · hn = h.

Therefore π is a surjective Lie group homomorphism.

Let {gα } = π −1 (e2 ). Using the fact that π is a Lie group homomorphism and
dπe1 is an isomorphism, one can prove that {gα } is discrete. Thus, π is a covering
map, since:

(i) π −1 (qV ) = α gα pU;
(ii) (gα pU) ∩ (pU) = 0,
/ if gα = e1 ;
(iii) π g pU : gα pU → qV is a diffeomorphism.
α

Conversely, if π is a covering map, it is locally a diffeomorphism, hence dπe1 is an

isomorphism.


Let G1 and G2 be Lie groups and θ : g1 → g2 be a Lie algebra homomorphism.
We prove that if G1 is connected and simply-connected, then θ induces a Lie group
homomorphism. We begin by proving uniqueness in the following lemma.
Lemma 1.25. Let G1 and G2 be Lie groups, with identities e1 and e2 respectively,
and θ : g1 → g2 be a fixed Lie algebra homomorphism. If G1 is connected, and
ϕ , ψ : G1 → G2 are Lie group homomorphisms with dϕe1 = dψe1 = θ , then ϕ = ψ .
Proof. It is easy to see that the direct sum g1 ⊕ g2 of Lie algebras has a natural Lie
algebra structure, and the direct product G1 × G2 of Lie groups has a natural Lie
group structure. Consider the Lie subalgebra of g1 ⊕ g2 given by

h := (X, θ (X )) : X ∈ g1 , (1.6)

i.e., the graph of θ : g1 → g2 . It follows from Theorem 1.21 that there exists a unique
connected Lie subgroup H ⊂ G1 × G2 with Lie algebra h.
Suppose that ϕ : G1 → G2 is a Lie group homomorphism with dϕe1 = θ . Then
 
σ : G1 → G1 × G2 , σ (g) := g, ϕ (g) ,

is a Lie group homomorphism, and

 
dσe1 : g1 → g1 ⊕ g2 , dσe1 X = X, θ (X)

is a Lie algebra homomorphism. Note that σ (G1 ) is the graph of ϕ , hence embedded
in G1 × G2 . From Proposition 1.16, σ (G1 ) is a Lie subgroup of G1 × G2 , with Lie
12 1 Basic Results on Lie Groups

algebra h = dσe1 (g1 ). Therefore, from Theorem 1.21, σ (G1 ) = H. In other words,
H is the graph of ϕ . If ψ : G1 → G2 is another Lie group homomorphism with
dψe1 = θ , following the same construction above, the graphs of ψ and ϕ would
both be equal to H, hence ϕ = ψ . 

Theorem 1.26. Let G1 and G2 be Lie groups and θ : g1 → g2 be a Lie algebra

homomorphism. There exist an open neighborhood V of e1 and a smooth map
ϕ : V → G2 that is a local homomorphism3 with dϕe1 = θ . In addition, if G1 is
connected and simply-connected, there exists a unique Lie group homomorphism
ϕ : G1 → G2 with dϕe1 = θ .
Proof. Let h be defined as in (1.6). From Theorem 1.21, there exists a unique
connected Lie subgroup H ⊂ G1 ×G2 with Lie algebra h, whose identity element we
denote e ∈ H. Consider the inclusion i : H → G1 × G2 and the projections π j : G1 ×
G2 → G j , j = 1, 2. The map (π1 ◦ i) : H → G1 is a Lie group homomorphism, such
that d(π1 ◦ i)e (X , θ (X )) = X, for all X ∈ Te1 G1 . It follows from the Inverse Function
Theorem that there exist open neighborhoods U of e in H and V of e1 in G1 , such
that (π1 ◦ i)|U : U → V is a diffeomorphism.
Define ϕ = π2 ◦ (π1 ◦ i)−1 : V → G2 . Then ϕ is a local homomorphism with
dϕe1 = θ . In fact, for each X ∈ Te1 G1 ,
 
dϕe1 (X) = d π2 ◦ (π1 ◦ i)−1 e (X)
1
 −1

= d(π2 )e d (π1 ◦ i) e X
1

= d(π2 )e (X, θ (X))

= θ (X).

Furthermore, (π1 ◦ i) is a Lie group homomorphism and d(π1 ◦ i)e is an isomor-

phism. From Proposition 1.24, (π1 ◦ i) : H → G1 is a covering map. Assuming G1
simply-connected, it follows that (π1 ◦ i) is a diffeomorphism, since a covering map
onto a simply-connected space is a diffeomorphism. Thus, (π1 ◦ i) can be globally
inverted, and we obtain a global homomorphism ϕ = π2 ◦ (π1 ◦ i)−1 : G1 → G2 with
dϕe1 = θ . The uniqueness of ϕ follows from Lemma 1.25, under the assumption
that G1 is simply-connected. 

Corollary 1.27. If G1 and G2 are connected and simply-connected and θ : g1 → g2

is an isomorphism, there exists a unique Lie group isomorphism ϕ : G1 → G2 with
dϕe1 = θ . In other words, if G1 and G2 are as above, then G1 and G2 are isomorphic
if and only if g1 and g2 are isomorphic.
Proof. By Theorem 1.26, there exists a unique Lie group homomorphism
ϕ : G1 → G2 with dϕe1 = θ . By Proposition 1.24, since dϕe1 = θ is an isomorphism,
ϕ is a covering map. Since G2 is simply-connected, ϕ is a diffeomorphism. Thus,
ϕ is a Lie group homomorphism and a diffeomorphism, therefore an isomorphism.


3 This simply means that ϕ (a b) = ϕ (a)ϕ (b), for all a, b ∈ V such that a b ∈ V .
1.3 Exponential Map and Adjoint Representation 13

1.3 Exponential Map and Adjoint Representation

Let G be a Lie group and g its Lie algebra. We recall that a Lie group homomorphism
ϕ : R → G is called a 1-parameter subgroup of G. Let X ∈ g and consider the Lie
algebra homomorphism

θ : R → g, θ (t) := t X.

From Theorems 1.21 and 1.26, there is a unique 1-parameter subgroup λX : R → G,

such that λX (0) = X.
Remark 1.28. Note that λX is an integral curve of the left-invariant vector field X
passing through e. In fact,

λX (t) =  = dLλX (t) λX (0) = dLλX (t) X = X(λX (t)).
ds λX (t + s) s=0
d

Definition 1.29. The (Lie) exponential map of G is defined as

exp : g → G, exp(X) := λX (1),

where λX is the unique 1-parameter subgroup of G such that λX (0) = X.

Proposition 1.30. The exponential map satisfies the following properties, for all
X ∈ g and t ∈ R,
(i) exp(tX) = λX (t);
(ii) exp(−tX) = exp(tX)−1 ;
(iii) exp(t1 X + t2 X ) = exp(t1 X) exp(t2 X);
(iv) exp : Te G → G is smooth and d(exp)0 = id, hence exp is a diffeomorphism from
an open neighborhood of the origin of Te G onto an open neighborhood of the
identity e ∈ G.
Proof. We claim that λX (t) = λtX (1). Consider the 1-parameter subgroup
λ (s) = λX (st). Differentiating at s = 0, it follows that

λ  (0) =  = t λX (0) = tX.
ds λX (st) s=0
d

Hence, from uniqueness of the 1-parameter subgroup in Definition 1.29,

λX (st) = λtX (s). Choosing s = 1, we obtain the expression in (i). Items (ii) and
(iii) are immediate consequences of (i), since λX is a Lie group homomorphism.
In order to prove item (iv), we construct a vector field V on the tangent bundle TG.
This tangent bundle can be identified with G × Te G, since G is parallelizable. The
construction is such that the projection of the integral curve of V passing through
(e, X) coincides with the curve t → exp(tX). Thus, from Theorem A.14, the flow
of V depends smoothly on the initial conditions, hence its projection (i.e., the
exponential map) is also smooth.
14 1 Basic Results on Lie Groups

Consider G × Te G  TG. Note that for all (g, X ) ∈ G × Te G, the tangent space
T(g,X) (G × Te G) can be identified with Tg G ⊕ Te G. Define the vector field

V ∈ X(G × Te G),  ⊕ {0} ∈ Tg G ⊕ Te G,

V (g, X) := X(g)


where X(g) = dLg X. It is not difficult to see that V is a smooth vector field. Since
t → exp(tX) is the unique integral curve of X for which λX (0) = e, as X is left-
invariant, Lg ◦ λX is the unique integral curve of X that takes value g at t = 0. Hence,
the integral curve of V through (g, X ) is t → (g exp(tX), X ). In other words, the
flow of V is given by ϕtV (g, X) = (g exp(tX), X) and, in particular, V is complete.
Let π1 : G × Te G → G be the projection onto G. Then exp(X) = π1 ◦ ϕ1V (e, X), so
exp is given by composition of smooth maps and is hence smooth. Finally, (i) and
Remark 1.28 imply that d(exp)0 = id. 

The exponential map, in general, may not be surjective. The classical example of
this situation is given by SL(2, R), see Duistermaat and Kolk [79].
Considering Lie groups of matrices GL(n, K), for K = C or K = R, it is natural
to inquire whether the Lie exponential map exp : gl(n, K) → GL(n, K) coincides
with the usual exponentiation of matrices, given for each A ∈ gl(n, K) by

∞
Ak
eA := ∑ . (1.7)
k=0 k!

We now prove that these notions indeed coincide.

To this aim, we recall two well-known properties of the exponentiation of
matrices. First, the right-hand side of (1.7) converges uniformly for A in a bounded
region of gl(n, K). This can be easily verified using the Weierstrass M-test. In
addition, given A, B ∈ GL(n, K), it is true that eA+B = eA eB if and only if A and
B commute (this fact is generalized in Remark 1.40).
Consider the map R  t → etA ∈ GL(n, K). Since each entry of etA is a power
series in t with infinite radius of convergence, it follows that this map is smooth.
Differentiating the power series term by term, it is easy to see that its tangent vector
at the origin of gl(n, K) is A, and from the properties above, this map is also a Lie
group homomorphism, hence a 1-parameter subgroup of GL(n, K). Since exp(A) is
the unique 1-parameter subgroup of GL(n, K) whose tangent vector at the origin is
A, it follows that eA = exp(A), for all A ∈ gl(n, K).
Remark 1.31. The above holds, more generally, for the exponential of endomor-
phisms of any real or complex vector space V . Namely, exp : End(V ) → Aut(V ) is
given by the exponentiation of endomorphisms, defined exactly as in (1.7).
Proposition 1.32. Let G1 and G2 be Lie groups and ϕ : G1 → G2 a Lie group
homomorphism. Then ϕ ◦ exp1 = exp2 ◦dϕe , i.e., the following diagram commutes:
1.3 Exponential Map and Adjoint Representation 15

Proof. Consider the 1-parameter subgroups of G2 given by α (t) = ϕ ◦ exp1 (tX)

and β (t) = exp2 ◦dϕe (tX). Then α  (0) = β  (0) = dϕe X, hence, it follows from
Theorem 1.26 that α = β , that is, the diagram above commutes. 

Remark 1.33. Using Proposition 1.32, we now show that if H is a Lie subgroup of
G, then the exponential map expH of H coincides with the restriction to H of the
exponential map expG of G. Consider the inclusion i : H → G, which is a Lie group
homomorphism, and its differential die : h → g. According to Corollary 1.19, this
is an isomorphism between the Lie algebra h and a Lie subalgebra of g. Then the
following diagram is commutative

Thus, with the appropriate identifications, expH (X) = i(expH (X)) = expG (die (X)) =
expG (X), for all X ∈ h, which proves the assertion.
We proceed by stating three important identities known as the Campbell formu-
las, or Campbell-Baker-Hausdorff formulas. A proof can be found in Spivak [198].
Campbell formulas 1.34. Let G be a Lie group and X,Y ∈ g. There exists ε > 0
such that, for all |t| < ε , the following hold:
 2 
(i) exp(tX) exp(tY) = exp t(X +Y ) + t2 [X,Y ] + O(t 3 ) ;
 
(ii) exp(tX) exp(tY) exp(−tX) = exp tY + t 2 [X,Y
 2] + O(t ) ; 3 
3

(iii) exp(−tX) exp(−tY) exp(tX) exp(tY) = exp t [X ,Y ] + O(t ) ;

O(t 3 )
where t3
is bounded.
We now pass to the second part of this section, discussing some properties of the
adjoint representation, or adjoint action.
Definition 1.35. Let G be any group and V a vector space. A (linear) representation
of G on V is a group homomorphism ϕ : G → Aut(V ), where Aut(V ) is the group
of vector space isomorphisms of V .
16 1 Basic Results on Lie Groups

Consider the action of G on itself by conjugation, i.e.,

a : G × G → G, a(g, x) := ag (x) = g x g−1 .

Definition 1.36. Let G be a Lie group, and g its Lie algebra. The representation

Ad : G → Aut(g), Ad(g) := d(ag )e = d(Lg )g−1 ◦ d(Rg−1 )e ,

is called the adjoint representation (or adjoint action) of G.

It follows from the above definition that
 
Ad(g)X = d
dt g exp(tX)g−1 t=0 . (1.8)

Applying Proposition 1.32 to the automorphism ag , it follows that

exp(tAd(g)X) = ag (exp(tX)) = g exp(tX)g−1 , (1.9)

and, in particular, for t = 1,

g exp(X)g−1 = exp(Ad(g)X ). (1.10)

The differential of the adjoint representation Ad is denoted ad, and given by

ad : g → End(g), ad(X)Y := dAde (X)(Y ).

It follows from the above definition that


ad(X )Y = d
dt (Ad (exp(tX))Y ) t=0 . (1.11)

Using Proposition 1.32 once more, we have

Ad(exp(tX)) = exp(tad(X))

i.e., the following diagram is commutative:

In particular, for t = 1 we obtain

Ad(exp(X )) = exp(ad(X)) (1.12)

1.3 Exponential Map and Adjoint Representation 17

Proposition 1.37. If X,Y ∈ g, then ad(X)Y = [X ,Y ].

Proof. From the Campbell formulas (Theorem 1.34), it follows that
 
exp(tX) exp(tY) exp(−tX) = exp tY + t 2 [X,Y ] + O(t 3 ) .

Using (1.10) with g = exp(tX), exp (Ad(exp(tX))tY) = exp(tY + t 2 [X,Y ] + O(t 3 )).
Since exp is locally injective near zero, we have Ad(exp(tX))tY = tY + t 2 [X,Y ] +
O(t 3 ), for sufficiently small t. Dividing by t, differentiating at t = 0, and apply-
ing (1.11), we conclude that ad(X)Y = [X ,Y ]. 

Exercise 1.38. Let G be a Lie subgroup of GL(n, R). For each g ∈ G and X,Y ∈ g,
verify the following properties:
(i) dLg X = g X and dRg X = X g are given by matrix multiplication;
(ii) Ad(g)Y = gY g−1 is given by matrix conjugation;
(iii) Use Proposition 1.37 to prove that [X,Y ] = XY − YX is the matrix commutator.
Let us now prove a result relating commutativity and the Lie bracket. Recall that
vector fields X,Y ∈ X(M) are said to commute if [X,Y ] = 0.
Proposition 1.39. Let G be a connected Lie group with Lie algebra g. The Lie
algebra g is abelian if and only if G is abelian.
Proof. Fix X,Y ∈ g and assume that [X,Y ] = 0. From (1.12) and Proposition 1.37,
∞
ad(X)k
Ad(exp(X))Y = exp(ad(X))Y = ∑ k!
Y = Y.
k=0

Thus, from (1.10), exp(X) exp(Y ) exp(−X) = exp(Y ), and hence exp(X) exp(Y )
= exp(Y ) exp(X). This means that there exists an open neighborhood U of e
such that if g1 , g2 ∈ U, then g1 g2 = g2 g1 . It follows from Proposition 1.20, that

G = n∈N U n , where U n = {g±1 ±1
1 · · · gn : gi ∈ U}. Therefore G is abelian.
4

Conversely, suppose G abelian. In particular, for all s,t ∈ R and X,Y ∈ g,

exp(sX) exp(tY) exp(−sX) = exp(tY). Differentiating at t = 0,


dt exp(sX) exp(tY) exp(−sX) t=0 = Y.
d

From (1.8), it follows that Ad(exp(sX))Y = Y . Hence, differentiating at s = 0, it

follows from (1.11) and Proposition 1.37 that ad(X)Y = [X,Y ] = 0, which means
that g is abelian. 

Remark 1.40. If X,Y ∈ g commute, that is, [X ,Y ] = 0, then

exp(X +Y ) = exp(X ) exp(Y ).

4 Notethat one cannot infer that G is abelian directly from the commutativity of exp, since exp
might not be surjective.
18 1 Basic Results on Lie Groups

Note that this does not hold in general. To verify this identity, set α : R → G,
α (t) := exp(tX) exp(tY). From Proposition 1.39, α is a 1-parameter subgroup, and
differentiating α at t = 0, we have α  (0) = X +Y . Hence α (t) = exp(t(X +Y )), and
we get the desired equation setting t = 1.
We conclude this section with a characterization of connected abelian Lie groups.
Theorem 1.41. Let G be a connected n-dimensional abelian Lie group. Then G is
isomorphic to T k × Rn−k , where T k = S1 × · · · × S1 is a k-torus. In particular, an
abelian connected and compact Lie group is isomorphic to a torus.
Proof. Using Proposition 1.39, since G is abelian, g is also abelian. Thus g can be
identified with Rn . Since G is connected, it follows from Remark 1.40 that exp : g →
G is a Lie group homomorphism. From Proposition 1.24, it is a covering map.
Consider the normal subgroup given by Γ = ker exp. We must now use two
results whose proof only appears later. The first (Theorem 1.42) asserts that any
closed subgroup of a Lie group is a Lie subgroup. As exp continuous and Γ is closed,
Γ is a Lie subgroup of Rn . The second (Corollary 3.38) asserts that the quotient of a
Lie group by a normal Lie subgroup is also a Lie group, hence Rn /Γ is a Lie group.
Note that G is isomorphic to Rn /Γ , because exp : Rn → G is a surjective Lie group
homomorphism and Γ = ker exp.

Using the fact that exp is a covering map, it is possible to prove that the isomorphism
between Rn /Γ and G defined above is in fact smooth, i.e., is a Lie group
isomorphism (this also follows from Corollary 1.49).
Finally, it is well-known that the only nontrivial discrete subgroups of Rn are
integral lattices. In other words, there exists a positive
integer k ≤ nand linearly
independent vectors e1 , . . . , ek ∈ Rn such that Γ = ∑ki=1 ni ei : ni ∈ Z . Therefore,
G is isomorphic to Rn /Γ = T k × Rn−k , where T k = S1 × · · · × S1 is a k-torus. 

1.4 Closed Subgroups and More Examples

The goal of this section is to prove that closed subgroups of a Lie group are Lie
subgroups. This fact is a very useful tool to prove that a subgroup is a Lie subgroup.
For instance, it applies to the subgroups of GL(n, K), for K = C and K = R, defined
in Sect. 1.1. We then briefly explore some corollaries, and discuss a few important
examples that complement the material of this chapter.
1.4 Closed Subgroups and More Examples 19

Theorem 1.42. Let G be a Lie group and H ⊂ G a closed subgroup of G. Then H

is an embedded Lie subgroup of G.
Proof. We prove this result through a sequence of five claims, following closely the
approach in Spivak [198, pp. 530–534]. The central idea of the proof is to reconstruct
the Lie algebra of H as a Lie subalgebra h ⊂ g. The natural candidate is

h := X ∈ Te G : exp(tX) ∈ H, for all t ∈ R .

Claim 1.43. Let {Xi } be a sequence in Te G with lim Xi = X, and {ti } a sequence of
real numbers, with limti = 0. If exp(ti Xi ) ∈ H, for all i ∈ N, then X ∈ h.
 −1
Since exp(−ti Xi ) = exp(ti Xi ) , without loss of generality, assume ti > 0.
Define Ri (t) to be the largest integer ≤ tti . Then tti − 1 < Ri (t) ≤ tti , so limti Ri (t) = t.
Therefore limti Ri (t)Xi = tX. On the one hand, it follows from continuity of exp
that lim exp(ti Ri (t)Xi ) = exp(tX). On the other hand, exp(ti Ri (t)Xi ) = [exp(ti Xi )]Ri (t)
∈ H. Since H is closed, exp(tX) ∈ H. Therefore X ∈ h.
Claim 1.44. h ⊂ Te G is a vector subspace of g.
Let X,Y ∈ h. It is clear that sX ∈ h for all s ∈ R. From the Campbell formulas
(Theorem 1.34),
 
t2
exp ti (X +Y ) + 2i [X,Y ] + O(ti3 ) = exp(ti X) exp(tiY ) ∈ H,
   
so exp ti (X +Y + t2i [X,Y ] + O(ti2 )) ∈ H and X +Y + t2i [X,Y ] + O(ti2 ) converges
to X +Y when ti goes to 0. From Claim 1.43, X +Y ∈ h.
Claim 1.45. Let k be a vector space such that Te G = h ⊕ k, and

ψ : h ⊕ k → G, ψ (X ,Y ) := exp(X ) exp(Y ).

There exists an open neighborhood U of the origin (0, 0) ∈ h ⊕ k, such that ψ |U is a

diffeomorphism.
Differentiating ψ with respect to each component, it follows that

dψ(0,0) (X, 0) = d(exp)0 X = X,

dψ(0,0) (0,Y ) = d(exp)0Y = Y.

Thus, dψ0 = id. From the Inverse Function Theorem, there exists U an open
neighborhood of the origin, such that ψ |U is a diffeomorphism.
Claim 1.46. There exists an open neighborhood V of the origin of k, such that
exp(Y ) ∈
/ H for Y ∈ V \ {0}.
20 1 Basic Results on Lie Groups

Suppose that there exists a sequence {Yi } with Yi ∈ k, limYi = 0 and exp(Yi ) ∈ H.
Choose an inner product on k and define ti = Yi  and Xi = t1i Yi . Since {Xi } is a
sequence in the unit sphere of k, which is compact, it converges up to passing to
a subsequence. Hence lim Xi = X , limti = 0 and exp(ti Xi ) ∈ H. Therefore, from
Claim 1.43, X ∈ h. This contradicts h ∩ k = {0}.
Claim 1.47. There exists an open neighborhood W of the origin in Te G such that
H ∩ exp(W ) = exp(h ∩W ).
It follows from the construction of h that H ∩ exp(W ) ⊃ exp(h ∩W ). According
to Claims 1.45 and 1.46, there exists a sufficiently small open neighborhood W of
the origin of Te G such that exp |W and ψ |W are diffeomorphisms, and (W ∩ k) ⊂ V .
Let a ∈ H ∩ exp(W ). As ψ |W is a diffeomorphism, there exist a unique X ∈ h
and a unique Y ∈ k such that a = exp(X) exp(Y ). Hence exp(Y ) = (exp(X))−1 a ∈
H. From Claim 1.46, Y = 0, that is, a = exp(X), with X ∈ h. Therefore H ∩
exp(W ) ⊂ exp(h ∩ W ). From Claim 1.47, H is an embedded submanifold of G in
a neighborhood of the identity e ∈ G. Thus, since H is a group, it is an embedded
submanifold. Finally, from Proposition 1.16, H is an embedded Lie subgroup of G.


In order to explore some consequences of the above result, we recall the Constant
Rank Theorem. It states that if a smooth map f : M → N is such that d fx has constant
rank, then for each x0 ∈ M there exists a neighborhood U of x0 such that:
(i) f (U) is an embedded submanifold of N;
(ii) The partition { f −1 (y) ∩U}y∈ f (U ) is a foliation of U;
 
(iii) For each y ∈ f (U), ker d fx = Tx f −1 (y) , for all x ∈ f −1 (y).
Lemma 1.48. Let G1 and G2 be Lie groups and ϕ : G1 → G2 be a Lie group
homomorphism. Then the following hold:
(i) dϕg has constant rank;
(ii) ker ϕ is a normal Lie subgroup of G1 ;
(iii) ker dϕe = Te (ker ϕ ).
Proof. Since ϕ is a Lie group homomorphism, ϕ ◦ L1g = Lϕ2 (g) ◦ ϕ , where Lgi denotes
the left multiplication by g on Gi . Hence, for all X ∈ Tg G1 ,

dϕg X = dϕg d(L1g )e d(L1g−1 )g X = dL2ϕ (g) dϕe d(Lg1−1 )g X.

Since Lϕ2 (g) is a diffeomorphism, dϕg X = 0 if and only if dϕe d(L1g−1 )X = 0. Hence
dim ker dϕg = dim ker dϕe , therefore dϕg has constant rank, proving (i). Item (ii)
follows from Theorem 1.42, since ker ϕ = ϕ −1 (e) is a closed normal subgroup.
Finally, (iii) follows directly from the Constant Rank Theorem. 

1.4 Closed Subgroups and More Examples 21

Corollary 1.49. Let G1 and G2 be Lie groups and ϕ : G1 → G2 a continuous

homomorphism. Then ϕ is smooth.
Proof. Let R = {(g, ϕ (g)) : g ∈ G1 } be the graph of ϕ . Then R is a closed subgroup
of G1 × G2 , hence, from Theorem 1.42, R is an embedded Lie subgroup of G1 × G2 .
Consider i : R → G1 × G2 the inclusion and the projections π j : G × G → G j , j =
1, 2. Then (π1 ◦ i) is a Lie group homomorphism, and from Lemma 1.48, d(π1 ◦ i)g
has constant rank. On the other hand, R is a graph, hence, by the Constant Rank
Theorem, (π1 ◦ i) is an immersion.
In addition, dim R = dim G1 , otherwise (π1 ◦ i)(R) would have measure zero,
contradicting (π1 ◦i)(R) = G1 . From the Inverse Function Theorem, (π1 ◦i) is a local
diffeomorphism. Since it is also bijective, it is a global diffeomorphism, therefore
ϕ = π2 ◦ (π1 ◦ i)−1 is smooth. 

Exercise 1.50 (). Let G be a Lie group and denote by μ : G × G → G the

multiplication map μ (g, h) = g h.
(i) Show that the differential dμ(g,h) : Tg G ⊕ Th G → Tgh G is given by

dμ(g,h) (X,Y ) = dLg Y + dRh X, X ∈ Tg G, Y ∈ Th G, (1.13)

and conclude that μ : G × G → G is a submersion (see Definition A.4);

(ii) Show that μ −1 (e) = {(g, −1
 g−1 ) :g ∈ G} is a submanifold of G × G, and describe
its tangent space T(g,h) μ (e) ⊂ Tg G ⊕ Th G;
(iii) Denote by i : μ −1 (e) → G × G the inclusion and by π j : G × G → G, j =
1, 2, the projection maps. Show that the restriction π1 : μ −1 (e) → G is a
diffeomorphism. Conclude that smoothness of the inversion map (1.2), which
can be written as π2 ◦ (π1 ◦ i)−1 : G → G, is a consequence of the smoothness
of μ .
Exercise 1.51. Prove that GL(n, R), GL(n, C), SL(n, R), SL(n, C), O(n), SO(n),
U(n), SU(n) and Sp(n) are Lie groups (recall definitions in Example 1.5). Verify
that their Lie algebras are, respectively,
(i) gl(n, R), the space of n × n square matrices over R;
(ii) gl(n, C), the space of n × n square matrices over C;
(iii) sl(n, R) = {X ∈ gl(n, R) : tr X = 0};
(iv) sl(n, C) = {X ∈ gl(n, C) : tr X = 0};
(v) o(n) = so(n) = {X ∈ gl(n, R) : X t + X = 0};
(vi) u(n) = {X ∈ gl(n, C) : X ∗ + X = 0};
(vii) su(n) = u(n) ∩ sl(n, C);
(viii) sp(n) = {X ∈ gl(n, H) : X ∗ + X = 0}.
Compare (v) with Exercise 1.15.
Hint: Prove directly that GL(n, R) and GL(n, C) are Lie subgroups and verify that the others are
closed subgroups of these. Use Remark 1.33 to prove that if X ∈ g and h ⊂ g, then X ∈ h if and
only if exp(tX) ∈ H, for all t ∈ R. Then, use Lemma 1.48, combined with the identity det eA = etr A
for any A ∈ gl(n, K), to assist with the computation of the above Lie algebras.
22 1 Basic Results on Lie Groups

Exercise 1.52. Let S3 = {(z1 , z2 ) ∈ C2 : |z1 |2 + |z2 |2 = 1} be the unit 3-sphere.

Verify that S3 is a Lie group when endowed with the multiplication
 
(z1 , z2 ) · (w1 , w2 ) := z1 w1 − z2 w2 , w1 z2 + z1 w2 .

(i) Consider the map ϕ : S3 → SU(2), given by

 
z1 −z2
ϕ (z1 , z2 ) =
z2 z1

Verify that ϕ is a continuous homomorphism, hence a Lie group homomor-

phism. Verify that ker ϕ is trivial, so ϕ is an isomorphism;
(ii) Consider the group of unit quaternions

Sp(1) = a + b i + c j + d k ∈ H : a2 + b2 + c2 + d 2 = 1

Find a continuous (hence smooth) group homomorphism ψ : Sp(1) → S3 ;

(iii) Conclude that S3 , Sp(1) and SU(2) are isomorphic Lie groups.
 → G its universal covering.
Exercise 1.53. Let G be a Lie group and consider ρ : G
Prove that:
(i) H := ρ −1 (e) is a normal discrete closed subgroup, and g h = h g, for all h ∈

H, g ∈ G;
∼ 
(ii) G = G/H;
(iii) π1 (G) is abelian.
Hint: Item (i) follows directly from ρ being a continuous homomorphism. The fact that g h = h g
can be proved defining f : G → H as f (x) := x h x−1 , noticing that {h} is an open and closed subset
of H and hence concluding that f −1 ({h}) is an open and closed subset of G.  In order to prove (iii),
i with α
consider two loops αi in G such that αi (0) = e, and their lifts α i (0) = e. Set hi = α
i (1) and
use (i) to conclude that α
2 ∗ α1 = h2 h1 = h1 h2 = α
1 ∗ α2 where ∗ denotes concatenation of curves.


Further details on the smooth structure of quotients of Lie groups, such as G/H
above, are given in Chap. 3, see Corollary 3.38.
Remark 1.54. The above exercise gives a glimpse of the structure of homotopy
groups of Lie groups. By using Morse theory on the space of paths on a compact
Lie group G, it is possible to prove that, in addition to π1 (G) being abelian, π2 (G)
is trivial and π3 (G) is torsion-free. These techniques also led to the discovery that
homotopy groups of classical Lie groups are periodic, which is a foundational result
known as Bott periodicity. For further details, we refer the reader to Milnor [158].
1.4 Closed Subgroups and More Examples 23

Exercise 1.55 (). Prove that SU(2) is the universal covering of SO(3), via the
following steps:
(i) Let g ∈ SU(2) ∼ = S3 , u ∈ S2 ⊂ R3 and θ ∈ [0, 2π ] be such that g = cos(θ ) +
sin(θ )u. Define Tg (v) = g v g−1 for all v ∈ R3 . Prove that Tg : R3 → R3 is a
linear orthogonal transformation, and Tg = eA2θ u , where AX is as in Exercise 1.9;
(ii) Prove that ϕ : S3  g → Tg ∈ SO(3) is a double covering map, concluding
π1 (SO(3)) ∼= Z2 .
Hint: Recall that S3 is isomorphic to SU(2), see Exercise 1.52. Also, recall that the product of
quaternions corresponds to (t1 + u1 ) · (t2 + u2 ) = (t1 t2 − u1 , u2 ) + (t2 u1 + t1 u2 + u1 × u2 ), and
verify that dtd Tcos(t θ )+sin(t θ )u (v)|t=0 = 2θ u × v. Note that under the identifications SU(2) ∼
= S3 and
SO(3) ∼ = RP3 , the double covering SU(2) → SO(3) is precisely the double covering S3 → RP3 .

Remark 1.56. Using results from basic topology, it is not difficult to show that
π1 (SO(n)) ∼= Z2 for all n ≥ 3. The universal covering of SO(n), n ≥ 3, is called
the spin group and denoted Spin(n), recall Theorem 1.23. Algebraic manipulations
slightly more elaborate than those in Exercise 1.55 show that Sp(1) × Sp(1) ∼ =
SU(2) × SU(2) is a double covering of SO(4), Sp(2) is a double covering of SO(5),
and SU(4) is a double covering of SO(6). In particular, we have the isomorphisms:

Spin(3) ∼
= Sp(1), Spin(4) ∼
= Sp(1) × Sp(1), Spin(5) ∼
= Sp(2), Spin(6) ∼
= SU(4).

The center of a Lie group G is the subgroup given by


Z(G) := g ∈ G : g h g−1 = h, for all h ∈ G , (1.14)

and the center of a Lie algebra g is defined as


Z(g) := X ∈ g : [X,Y ] = 0, for all Y ∈ g . (1.15)

The following result relates the centers of a Lie group and of its Lie algebra.
Corollary 1.57. Let G be a connected Lie group. Then the following hold:
(i) Z(G) = ker Ad;
(ii) Z(G) is a normal Lie subgroup of G;
(iii) Z(g) = ker ad;
(iv) Z(g) is the Lie algebra of Z(G).
Proof. First, we verify that Z(G) = ker Ad. If g ∈ Z(G), clearly Ad(g) = id.
Conversely, let g ∈ ker Ad. It follows from (1.10) that g exp(tX)g−1 = exp(tX), for
all X ∈ g. Hence g commutes with the elements of a neighborhood of the identity
e ∈ G. From Proposition 1.20, we conclude that g ∈ Z(G), proving (i).
Item (ii) follows from (i) and Lemma 1.48. Item (iii) follows directly from
Proposition 1.37. Since d(Ad)e = ad, (iv) follows directly from (iii) in Lemma 1.48.


24 1 Basic Results on Lie Groups

Remark 1.58. The above result gives an alternative proof of an assertion in Propo-
sition 1.39, that G is abelian if g is abelian. Indeed, if [X ,Y ] = 0 for all X ,Y ∈ g,
then Z(g) = g. Thus, from (iv) in Corollary 1.57, Z(G) is open in G and it follows
from Proposition 1.20 that G = Z(G), hence G is abelian.
Exercise 1.59 (). In this exercise, we generalize the concept of center of a Lie
group G and Lie algebra g, to the notion centralizer of subsets of G and g. Define
the centralizer of a subset S in G as

ZG (S) := g ∈ G : g h g−1 = h, for all h ∈ S ,

and the centralizer of a subset s in g as


Zg (s) := X ∈ g : [X ,Y ] = 0, for all Y ∈ s .

Note that ZG (G) = Z(G) and Zg (g) = Z(g).

Prove that ZG (S) is a Lie subgroup of G and Zg (s) is a Lie subalgebra of g.
Moreover, if G is connected, H is a Lie subgroup of G and h is the corresponding
Lie subalgebra of g, then the following hold (cf. Corollary 1.57):
(i) ZG (H) = {g ∈ G : Ad(g)X = X for all X ∈ h};
(ii) Zg (h) = {X ∈ g : [X,Y ] = 0 for all Y ∈ h};
(iii) Zg (h) is the Lie algebra of ZG (H).
Exercise 1.60 (). In this exercise, we discuss relations between U(n)
and SU(n).
(i) Show that the centers of U(n) and SU(n) are respectively given by

Z(U(n))={z id : z ∈ C, |z| = 1} ∼
= S1 and Z(SU(n))={z id : zn = 1} ∼
= Zn .

In particular, conclude that the Lie groups U(n) and SU(n) × S1 are not
isomorphic, since their centers are not isomorphic;
(ii) Show that the multiplication map Z(U(n)) × SU(n)  (z id, A) → z A ∈ U(n) is
a Lie group homomorphism, and also an n-fold covering S1 × SU(n) → U(n);
(iii) Find a Lie group homomorphism ϕ : U(n) → S1 such that ker ϕ = SU(n) and
ϕ ◦ ι = id, where ι : S1 → U(n) is the inclusion with image consisting of
diagonal matrices with eiθ in the upper left corner and 1 in the rest of the
diagonal. Conclude that U(n) ∼ = SU(n)  S1 is a semidirect product of SU(n)
1
by S .
Exercise 1.61 (). Let G be a Lie group and H a closed Lie subgroup. Define the
normalizer of H in G as

N(H) := g ∈ G : g H g−1 = H , (1.16)
1.4 Closed Subgroups and More Examples 25

which is also sometimes denoted NG (H). Prove that N(H) is a closed Lie subgroup
of G. Assuming that H is connected and denoting its Lie algebra by h, prove that
the following hold:
(i) N(H) = {g ∈ G : Ad(g)X ∈ h for all X ∈ h};
(ii) The Lie algebra of N(H) is n = {X ∈ g : [X ,Y ] ∈ h for all Y ∈ h}.
Hint: Use (1.8), (1.10) and (1.12). For a solution, see the proof of Proposition 2.37.
Chapter 2
Lie Groups with Bi-invariant Metrics

This chapter deals with Lie groups with special types of Riemannian metrics:
bi-invariant metrics. Every compact Lie group admits one such metric (see Propo-
sition 2.24), which plays a very important role in the study of its geometry. In what
follows, we use tools from Riemannian geometry to give concise proofs of several
classical results on compact Lie groups.
We begin by reviewing some auxiliary facts of Riemannian geometry. Basic
results on bi-invariant metrics and Killing forms are then discussed, e.g., we prove
that the exponential map of a compact Lie group is surjective (Theorem 2.27), and
we show that a semisimple Lie group is compact if and only if its Killing form is
negative-definite (Theorem 2.35). We also prove that a simply-connected Lie group
admits a bi-invariant metric if and only if it is a product of a compact Lie group
with a vector space (Theorem 2.45). Finally, we prove that if the Lie algebra of a
compact Lie group G is simple, then the bi-invariant metric on G is unique up to
rescaling (Proposition 2.48).
Further suggested readings on the contents of this chapter are Fegan [85],
Grove [106], Ise and Takeuchi [133], Milnor [158, 159], Onishchik [179] and
Petersen [183], which inspired the present text.

2.1 Basic Facts of Riemannian Geometry

The main objective of this section is to review basic results of Riemannian geometry
that are used in the following chapters. Proofs of most results in this section are
omitted, and can be found in any standard textbook on Riemannian geometry, such
as do Carmo [61], Petersen [183] or Lee [152].
A Riemannian manifold is a smooth manifold M endowed with a (Riemannian)
metric, i.e., a (0, 2)-tensor field g on M that is

© Springer International Publishing Switzerland 2015 27

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_2
28 2 Lie Groups with Bi-invariant Metrics

(i) Symmetric: g(X,Y ) = g(Y, X), for all X,Y ∈ TM;

(ii) Positive-definite: g(X, X) > 0, if X = 0.
This means that a metric determines an inner product g p on each tangent space Tp M.
For this reason, we sometimes write X ,Y  = g(X,Y ) and X2 = g(X, X), if
the metric g is evident from the context. It is not difficult to prove that every
manifold admits a Riemannian metric, using partitions of unity. As indicated below,
Riemannian metrics provide a way to measure distances, angles, curvature, and
other geometric properties.
It is possible to associate to each metric a so-called connection. Such map allows
to parallel translate vectors along curves (see Proposition 2.7), connecting tangent
spaces of M at different points. It is actually possible to define connections on any
vector bundles over a manifold.
Definition 2.1. A linear connection on a Riemannian manifold (M, g) is a map

∇ : X(M) × X(M)  (X ,Y ) −→ ∇X Y ∈ X(M),

satisfying the following properties:

(i) ∇X Y is C∞ (M)-linear in X, i.e., for all f , g ∈ C∞ (M),

∇ f X1 +g X2 Y = f ∇X1 Y + g ∇X2 Y ;

(ii) ∇X Y is R-linear in Y , i.e., for all a, b ∈ R,

∇X (aY1 + bY2 ) = a ∇X Y1 + b ∇X Y2 ;

(iii) ∇ satisfies the Leibniz rule, i.e., for all f ∈ C∞ (M),

∇X ( f Y ) = f ∇X Y + (Xf )Y.

Moreover, a linear connection is said to be compatible with the metric g if

Xg(Y, Z) = g(∇X Y, Z) + g(Y, ∇X Z), X ,Y, Z ∈ X(M).

It turns out that requiring a connection to be compatible with the metric does
not determine a unique connection on (M, g). To this purpose, define the torsion
tensor of ∇ to be the (1, 2)-tensor field given by T (X,Y ) = ∇X Y − ∇Y X − [X,Y ].
A connection ∇ is said to be symmetric if its torsion vanishes identically, that is, if
[X,Y ] = ∇X Y − ∇Y X for all X,Y ∈ X(M).
Exercise 2.2. Let M be an embedded surface in R3 with the induced metric, i.e.,
g = i∗ g0 where g0 is the Euclidean metric and i : M → R3 is the inclusion. Let ∇ be
the Euclidean derivative. Set ∇X Y := (∇X Y ) , i.e., the tangent part of ∇X Y where
X and Y are local extension of the vector field X and Y . Show that ∇ is a symmetric
connection compatible with the metric g.
Hint: One can prove that the connection is symmetric using Remark A.11.
2.1 Basic Facts of Riemannian Geometry 29

Levi-Civita Theorem 2.3. On a Riemannian manifold (M, g), there exists a unique
linear connection ∇ that is compatible with g and symmetric.
The key fact on the proof of this theorem is the equation known as connection
formula, or Koszul formula. It exhibits the natural candidate to the desired connec-
tion, and shows that it is uniquely determined by the metric:

∇Y X, Z = 1
2 XY, Z − ZX,Y  +Y Z, X
 (2.1)
− [X,Y ], Z − [X, Z],Y  − [Y, Z], X .

The unique symmetric linear connection compatible with the metric is called
the Levi-Civita connection, or Riemannian connection, and we refer to it simply
as connection. Using this connection, one can differentiate vector fields on a
Riemannian manifold (M, g) as we describe next.
Proposition 2.4. Let M be a manifold with linear connection ∇ and γ : I → M
be a smooth curve. Let Γ (γ ∗ TM) denote the set of smooth vector fields along
γ . There exists a unique correspondence that, to each X ∈ Γ (γ ∗ TM) associates
∗
dt X ∈ Γ (γ TM), called the covariant derivative of X along γ , satisfying:
D

(i) Linearity, i.e., for all X,Y ∈ Γ (γ ∗ TM),

D D D
(X +Y ) = X + Y ;
dt dt dt

(ii) Leibniz rule, i.e., for all X ∈ Γ (γ ∗ TM), f ∈ C∞ (I),

D df D
( f X) = X + f X;
dt dt dt

(iii) If X is induced from a vector field X ∈ X(M), that is X(t) = X(

 γ (t)), then
D
X = ∇ γ  
X.
dt
Note that to each linear connection on M, the proposition above gives a covariant
derivative operator for vector fields along γ . As mentioned before, we only consider
the Levi-Civita connection, hence the covariant derivative is uniquely defined.
Equipped with this notion, it is possible to define the acceleration of a curve
as the covariant derivative of its tangent vector field, and geodesics as curves with
D 
null acceleration. More precisely, γ : I → M is a geodesic if dt γ = 0. It is often
convenient to reparametrize a geodesic γ to become a unit speed geodesic γ, that is,
γ  = 1 and γ has the same image as γ .
Writing a local expression for the covariant derivative, it is easy to see that a
curve is a geodesic if and only if it satisfies a second-order system of ODEs, called
the geodesic equation. Hence, applying the classical ODE result that guarantees
existence and uniqueness of solutions, one can prove the following.
30 2 Lie Groups with Bi-invariant Metrics

Theorem 2.5. For any p ∈ M, t0 ∈ R and v ∈ Tp M, there exist an open interval I ⊂

R containing t0 and a geodesic γ : I → M satisfying the initial conditions γ (t0 ) = p
and γ  (t0 ) = v. In addition, any two geodesics with those initial conditions agree on
their common domain.
From uniqueness of the solution, it is possible to obtain the existence of a
geodesic with maximal domain for a prescribed initial data. The restriction of a
geodesic to a bounded subset of its maximal domain is called a geodesic segment.
Exercise 2.6. Let M be an embedded surface in R3 , with the induced connection
(defined in Exercise 2.2). Let V be a plane, and assume that, for all p ∈ M ∩V , the
normal vector ξ (p) is tangent to V . Show that the intersection M ∩V can be locally
parametrized by a geodesic segment on M. Conclude that the profile curve of a
surface of revolution is a geodesic (up to reparametrization). In particular, conclude
that great circles are the geodesics of the round sphere S2 ⊂ R3 .
Hint: Consider γ ∈ V ∩ M such that γ  (t) = 1. Note that γ  (t), γ  (t) = 0, and hence γ  (t) is
normal to γ and contained in V . Since ξ is also normal to γ and contained in V , conclude that
γ  = λ ξ and therefore dt
D 
γ = 0.

Another construction that involves covariant differentiation along curves is

parallel translation. A vector field X ∈ Γ (γ ∗ TM) is said to be parallel along γ if

dt X = 0. Thus, a geodesic γ can be characterized as a curve whose tangent field γ
D

is parallel along γ . A vector field is called parallel if it is parallel along every curve.
Proposition 2.7. Let γ : I → M be a curve, t0 ∈ I and v0 ∈ Tγ (t0 ) M. There exists a
unique parallel vector field X along γ such that X(t0 ) = v0 .
This vector field is called the parallel translate of v0 along γ . Once more, the
proof depends on basic ODE results.
An important question after having existence and uniqueness of geodesics is how
geodesics change under perturbations of the initial data, which leads to the definition
of the geodesic flow of a Riemannian manifold. This is the flow

ϕ G : U ⊂ R × TM → TM

defined in an open subset U of R × TM that contains {0} × TM, of the unique vector
field G on the tangent bundle, i.e., G ∈ X(TM), whose integral curves are of the
form t → (γ (t), γ  (t)), where γ is a geodesic in M. This means that:
(i) γ (t) = π ◦ ϕ G (t, (p, v)) is the geodesic with initial conditions γ (0) = p and
γ  (0) = v, where π : TM → M is the bundle projection;
(ii) ϕ G (t, (p, cv)) = ϕ G (ct, (p, v)), for all c ∈ R such that this equation makes sense.
In fact, supposing that such vector field G exists, one can obtain conditions
in local coordinates that this field must satisfy (corresponding to the geodesic
equation). Defining the vector field as its solutions, one may use Theorem A.14
to guarantee existence and smoothness of ϕ G .
2.1 Basic Facts of Riemannian Geometry 31

We now proceed to define a similar concept to the Lie exponential map, which
coincides with it when considering appropriate (bi-invariant) Riemannian metrics
on compact Lie groups, see Theorem 2.27.
Definition 2.8. The (Riemannian) exponential map at p ∈ M is the map

exp p : Bε (0) ⊂ Tp M → M, exp p (v) := π ◦ ϕ G (1, (p, v)).

Smoothness of exp p follows immediately from Theorem A.14. Using the Inverse
Function Theorem, one can verify that for any p ∈ M, there exist a neighborhood V
of the origin in Tp M and a neighborhood U of p ∈ M, such that exp p |V : V → U is
a diffeomorphism. Such neighborhood U is called a normal neighborhood of p.
Define the (Riemannian) distance dist(p, q) for any pair of points p, q ∈ M to
be the infimum of lengths of all piecewise smooth curve segments joining p and q,
where the length of a curve segment γ : [a, b] → M is defined as
 b
(γ ) := g(γ  (t), γ  (t)) dt.
a

Then (M, dist) is a metric space, and the topology induced by this distance coincides
with the original topology from the atlas of M. It is a very important fact that
geodesics locally minimize  among piecewise smooth curves. Geodesic segments
γ : [a, b] → M that globally minimize distance, i.e., such that dist(γ (t1 ), γ (t2 )) =
|t1 − t2 | for all t1 ,t2 ∈ [a, b], are called minimal.
A Riemannian manifold is called geodesically complete if the maximal domain
of definition of all geodesics is R. It is not difficult to see that compactness is a
sufficient condition for a manifold to be complete. The following important result
states that all completeness notions for a Riemannian manifold are equivalent.
Hopf-Rinow Theorem 2.9. Let (M, g) be a connected Riemannian manifold and
p ∈ M. The following statements are equivalent:
(i) exp p is globally defined, that is, exp p : Tp M → M;
(ii) M is geodesically complete;
(iii) (M, dist) is a complete metric space;
(iv) Every closed bounded set in M is compact.
If M satisfies any (hence all) of the above items, each two points of M can be joined
by a minimal geodesic segment. In particular, for each x ∈ M the exponential map
expx : Tx M → M is surjective.
Exercise 2.10. Give an example of a Riemannian manifold which is not complete,
but such that any pair of points can be joined by a minimal geodesic segment.
Consider M and N manifolds and f : M → N a smooth map. Recall that any
(0, s)-tensor τ on N may be pulled back by f , the result being a (0, s)-tensor f ∗ τ
on M, see (A.6). A diffeomorphism f : (M, gM ) → (N, gN ) satisfying f ∗ gN = gM ,
32 2 Lie Groups with Bi-invariant Metrics

that is, gM
p (X,Y ) = g f (p) (d f p X, d f pY ) for all p ∈ M and X,Y ∈ Tp M, is called a
N

(Riemannian) isometry. It is easy to show that Riemannian isometries map geodesics

to geodesics, and hence preserve distance, that is, are also metric isometries.
Conversely, metric isometries of a smooth Riemannian manifold are Riemannian
isometries. Thus, we simply refer to these maps as isometries.
From the above, it also follows that isometries commute with the exponential
map, in the sense that if f : M → M is an isometry, then:

f (exp p (v)) = exp f (p) (d f p v) (2.2)

for all p ∈ M and v ∈ Tp M such that exp p (v) is defined. This allows to prove that
isometries of a connected manifold are determined by its value and derivative at a
single point.
Exercise 2.11. Suppose f1 : M → M and f 2 : M → M are isometries of a connected
Riemannian manifold (M, g), such that there exists p ∈ M with f1 (p) = f2 (p) and
d( f1 ) p = d( f2 ) p . Prove that f1 (x) = f 2 (x) for all x ∈ M.
Hint: Consider the closed subset S = {x ∈ M : f1 (x) = f2 (x) and d( f1 )x = d( f2 )x }. Use (2.2) to
prove that S is also open, and hence S = M, since M is connected.

Isometries of a Riemannian manifold (M, g) clearly form a group, denoted

Iso(M, g), where the operation is given by composition. A fundamental result, due
to Myers and Steenrod [173], is that this is a Lie group that acts smoothly on M.
Myers-Steenrod Theorem 2.12. Let (M, g) be a Riemannian manifold. Any closed
subgroup of Iso(M, g) in the compact-open topology is a Lie group. In particular,
Iso(M, g) is a Lie group.
Remark 2.13. A subset G ⊂ Iso(M, g) is closed in the compact-open topology if
given any sequence { fn } of isometries in G that converges uniformly in compact
subsets to a continuous map f : M → M, then f ∈ G.
We now mention a special class of vector fields on M, closely related to Iso(M, g).
A Killing vector field is a vector field whose local flow is a local isometry.
Alternatively, Killing vector fields are characterized by the following property.
Proposition 2.14. A vector field X ∈ X(M) is a Killing vector field if and only if

g(∇Y X , Z) = −g(∇Z X,Y ), for all Y, Z ∈ X(M).

Theorem 2.15. The set iso(M, g) of Killing fields on (M, g) is a Lie algebra.
In addition, if (M, g) is complete, then iso(M, g) is the Lie algebra of Iso(M, g).
An essential concept in Riemannian geometry is that of curvature. The curvature
tensor on (M, g) is defined as the (1, 3)-tensor field given by the following
expression,1 for all X,Y, Z ∈ X(M).

1 We remark that some texts use the opposite sign convention for R, as there is no standard choice.
2.1 Basic Facts of Riemannian Geometry 33

R(X ,Y )Z := ∇[X,Y ] Z − ∇X ∇Y Z + ∇Y ∇X Z

It is possible to use the metric to deal with this tensor as a (0, 4)-tensor, given by

R(X,Y, Z,W ) := g(R(X ,Y )Z,W ). (2.3)

There are many important symmetries of this tensor that we state below. For each
X,Y, Z,W ∈ X(M),
(i) R is skew-symmetric in the first two and last two entries:

R(X,Y, Z,W ) = −R(Y, X , Z,W ) = R(Y, X,W, Z);

(ii) R is symmetric in the first and last pairs of entries:

R(X,Y, Z,W ) = R(Z,W, X,Y );

(iii) R satisfies a cyclic permutation property, called the first Bianchi identity:

R(X,Y )Z + R(Z, X)Y + R(Y, Z)X = 0.

Using the curvature tensor, we can define the sectional curvature of the plane
spanned by (linearly independent) vectors X and Y as

R(X,Y, X,Y )
sec(X,Y ) := . (2.4)
g(X , X )g(Y,Y ) − g(X,Y )2

It is possible to verify that sec(X,Y ) depends only on the plane σ spanned by X and
Y , and not specifically on the given basis {X,Y } of σ . For this reason, we sometimes
denote (2.4) by sec(σ ), where σ = span{X,Y } ⊂ Tp M.
There are several interpretations of curvature, the most naive being that it
measures the failure of second covariant derivatives to commute. The curvature
tensor is also part of the Jacobi equation along a geodesic γ , given by

DD  
J + R γ  (t), J(t) γ  (t) = 0. (2.5)
dt dt
This is an ODE whose solutions J are vector fields along γ , called Jacobi fields.
A Jacobi field J along γ is the velocity (or variational
 field) of a variation of γ
by geodesics. More precisely, J(t) = ∂∂s α (s,t)s=0 where α : (−ε , ε ) × [a, b] → M
is a piecewise smooth map, such that α (0,t) = γ (t) and α (s,t) is a geodesic for
each fixed s ∈ (−ε , ε ). Jacobi fields describe how quickly geodesics with the same
starting point and different directions move away from each other.
Exercise 2.16 (). Let M be a surface with constant sectional curvature k, and let γ
be a unit speed geodesic on M. Consider a variation α (s,t) by geodesics, such that
34 2 Lie Groups with Bi-invariant Metrics


α (s, 0) = γ (0) for all s. Consider the Jacobi field J(t) := ∂∂s α (s,t)s=0 , and assume
D
that dt J(0) = w is orthogonal to γ and w = 1. Let w(t) be the parallel translate of
w along γ (t). Prove that J(t) = snk (t)w(t), where snk is the function given by
⎧ √
⎪
⎪
sin(t
√ k) if k > 0,
⎨ k
snk (t) := t if k = 0,
⎪
⎪ √
⎩ sinh(t
√ −k) if k < 0.
−k

Hint: Write J(t) = f1 (t)γ  (t) + f2 (t)w(t) and use the properties of the curvature tensor R described
above, the hypothesis that  dt D
J(0), γ  (0) = 0, the fact that w(t), γ  (t) = 0 (because parallel
translation is an isometry), and unicity of solutions to ODEs, to conclude that f1 ≡ 0 and
f2 (t) = snk (t).

Remark 2.17. The above statement about Jacobi fields also holds on more gen-
eral space forms, i.e., Riemannian manifolds M(k) that have constant sectional
curvature k, since their curvature tensor satisfies R(X,Y, Z,W ) = k(X, ZY,W  −
Y, ZX,W ). If a space form M(k)√is simply-connected and has dimension n, then
√ to the sphere S (1/ k), the Euclidean space R or the hyperbolic
it is isometric n n

space H (1/ −k), according to the cases k > 0, k = 0 and k < 0 respectively.
n

As illustrated by the above exercise, on manifolds with nonnegative sectional

curvature (sec ≥ 0), the distance between geodesics starting at the same point grows
slower than in flat space; in other words, the exponential map exp p is distance
nonincreasing. Similarly, on manifolds with nonpositive sectional curvature (sec ≤
0), distance between geodesics starting at the same point grows faster than in flat
space; which means that exp p is distance nondecreasing. Curvature also describes
how parallel transport along short loops differs from the identity. More details on
these interpretations can be found in Grove [106], Jost [135] and Petersen [183].
Besides analyzing sectional curvatures (which actually fully determine the
curvature tensor R), it is convenient to summarize information contained in R by
constructing simpler tensors. The Ricci tensor is a (0, 2)-tensor field defined as a
trace of the curvature tensor R. More precisely, if {e1 , . . . , en } is an orthonormal
basis of Tp M,

  n
Ric(X,Y ) := tr R(X, · )Y = ∑ R(X, ei ,Y, ei ).
i=1

The scalar curvature is a function scal : M → R given by the trace of Ric, i.e.,

  n
scal := tr Ric = ∑ Ric(e j , e j ) = 2 ∑ R(ei , e j , ei , e j ) = 2 ∑ sec(ei , e j ).
j=1 1≤i< j≤n 1≤i< j≤n
2.1 Basic Facts of Riemannian Geometry 35

The curvatures Ric and scal respectively encode information on the volume
distortion and volume defect of small balls in (M, g), compared with a space form.
A metric is called Einstein if it is proportional to its Ricci tensor:

Ric(X ,Y ) = λ g(X,Y ). (2.6)

In this case, the number λ ∈ R is called the Einstein constant of g. It is easy

to see that if (M, g) has constant sectional curvature sec = k, then it is Einstein,
with constant λ = (n − 1)k, where n = dim M. Moreover, any Einstein metric with
constant λ has constant scalar curvature scal = n λ .
Remark 2.18. Einstein metrics originate from General  Relativity, as solutions to
the Einstein equations in vacuum, Ric = 12 scal −Λ g, where Λ is the cosmological
constant. Details on this equation and its impact in Physics can be found in [33, 160].
As a curiosity, the reason why Riemannian metrics are denoted g also comes from
General Relativity, since, in this context, g is interpreted as a gravitational field.
We now state a classical result that establishes a link between the curvature and
topology of a manifold. Recall that the diameter diam(X ) of a metric space X is the
smallest number d such that any two points in X are always at distance ≤ d.
Bonnet-Myers Theorem 2.19. Let (M, g) be a connected complete n-dimensional
Riemannian manifold, with n ≥ 2. Assume Ric ≥ (n − 1)k g for some k > 0. Then:
(i) diam(M) ≤ √πk ; in particular, M is compact;
(ii) The universal covering of M is compact, hence π1 (M) is finite.
We end this section with a quick discussion of Riemannian immersions and sub-
mersions, which are dual types of maps between Riemannian manifolds (of poten-
tially different dimensions), that generalize the notion of isometries.
Let (M, g) and (M, g) be Riemannian manifolds and i : M → M be an immersion
(recall Definition A.4). Then i is a Riemannian immersion if for all p ∈ M, di p
is a linear isometry from Tp M onto its image, which is a subspace of Ti(p) M. In
other words, g = i∗ g. A vector field X on M can always be locally extended to a
vector field X on M. Let ∇ be the Levi-Civita connection of (M, g). It is easy to
prove that the component of ∇X Y that is tangent to M, denoted (∇X Y ) , gives the
value of the Levi-Civita connection of (M, g) applied to vector fields X and Y along
M, independent of the choice of local extensions X and Y (cf. Exercise 2.2). The
difference between these connections is a bilinear symmetric map with values on
the normal space to M, called the second fundamental form II of M:

II p : Tp M × Tp M → ν p M, II(X,Y ) p := (∇X Y ) p − (∇X Y )

p,

where ν p M denotes the normal space to M at p, i.e., such that Ti(p) M = Tp M ⊕ ν p M

is a g-orthogonal direct sum. For each normal vector ξ to M at p, we can also define
a bilinear symmetric form IIξ by taking the inner product of ξ and II,

(IIξ ) p : Tp M × Tp M → R, IIξ (X,Y ) p := g p (ξ , II(X,Y )).

36 2 Lie Groups with Bi-invariant Metrics

If II vanishes identically, then M is called a totally geodesic submanifold. This is

equivalent to each geodesic of M being a geodesic of M, since the connections ∇
and ∇ agree and hence so do the corresponding geodesic equations. Being totally
geodesic is a very strong property; however, this class of submanifolds appears many
times along this text (see, e.g., Exercise 2.29 and Proposition 3.93).
The relation between curvatures on M and M is given by the Gauss equation

R(X,Y, X,Y ) = R(X,Y , X ,Y ) + g(II(X, X), II(Y,Y )) − g(II(X,Y ), II(X,Y )), (2.7)

where R and R denote the curvature tensors of M and M, respectively. Assuming X ,Y

and X,Y are pairs of orthonormal vectors, the quantities sec(X ,Y ) = R(X,Y, X,Y )
and sec(X,Y ) = R(X,Y , X,Y ) are called intrinsic and extrinsic curvatures of the
plane tangent to M spanned by these vectors. Note that both agree if M is totally
geodesic.
Since IIξ is symmetric, there exists a self-adjoint operator Sξ with respect to g,
called the shape operator, satisfying

g(Sξ X ,Y ) = IIξ (X ,Y ). (2.8)

It is not difficult to prove that Sξ (X) = (−∇X ξ ) where ξ is any normal field that
extends ξ . Eigenvalues and eigenvectors of Sξ (X) are respectively called principal
curvatures and principal directions; see Remark 2.21 for a geometric interpretation.
Exercise 2.20. Compute the principal curvatures and directions of a round sphere
and a round cylinder embedded in R3 .
Hint: Use the explicit description of the normal vectors fields (e.g., the normal to the unit sphere
is ξ (x, y, z) = (x, y, z)) and apply the chain rule, i.e., Sξ (X) = (−∇X ξ ) = −∇X ξ = dtd (ξ ◦ α ),
where α is a curve in the surface.

The second fundamental form also allows us to define the mean curvature vector
H of the submanifold M, which is a section of ν (M) given by the trace of the second
fundamental form, i.e.,

H(p) := ∑ II(ei , ei ) (2.9)

i

where {ei } is an orthonormal basis of Tp M. The length H of the mean curvature
vector is called the mean curvature of M. A submanifold whose mean curvature
is identically zero is called minimal. From the definitions, it is clear that totally
geodesic submanifolds are automatically minimal, but the converse need not be true.
Remark 2.21. Let M be an embedded surface in M = R3 with the induced metric.
According to the Gauss equation (2.7), the sectional curvature of M coincides with
Gaussian curvature of M, i.e., the product of eigenvalues λ1 λ2 (principal curvatures)
of the shape operator Sξ (·) = −∇(·) ξ , where ξ is a unitary normal vector to M.
Notice that this is precisely the right-hand side of (2.7), since the determinant of II
2.1 Basic Facts of Riemannian Geometry 37

is equal to the product of its eigenvalues. An important fact in differential geometry

is that each embedded surface with nonzero Gaussian  curvature is, up to rigid
motions, locally given by the graph of f (x1 , x2 ) = 12 λ1 x21 + λ2 x22 + O(x3 ). In
other words, if sec > 0 (respectively sec < 0) M is locally a small perturbation of
elliptic (respectively hyperbolic) paraboloid.
The dual notion to an immersion is that of a submersion. Let (M, g) and (M, g)
be Riemannian manifolds and π : M → M be a submersion (recall Definition A.4).
Then π is a Riemannian submersion if for all p ∈ M, dπ p is a linear isometry from
(ker dπ p )⊥ onto Tπ (p) M, where (ker dπ p )⊥ is the subspace of Tp M given by the g-
orthogonal complement to ker dπ p . The subspace

V p := ker dπ p (2.10)

is called the vertical space at p ∈ M, while


H p := (ker dπ p )⊥ = X ∈ Tp M : g(X,Y ) = 0 for all Y ∈ V p (2.11)

is called the horizontal space at p ∈ M. The collections of vertical and horizontal

spaces on M form two complementary distributions V and H on M, that we
accordingly call vertical and horizontal distributions. The vertical distribution is
clearly integrable, since it is tangent to the foliation Fπ = {π −1 (x) : x ∈ M}, but the
horizontal distribution may be nonintegrable. Again, denote by ∇ and ∇ the Levi-
Civita connections on M and M, respectively. The two fundamental tensors T and A
of the Riemannian submersion π : M → M are (1, 2)-tensors on M, given by
 H  V
TX Y = ∇X V Y V + ∇X V Y H ,
 V  H (2.12)
AX Y = ∇X H Y H + ∇X H Y V ,

where X V and X H respectively denote vertical and horizontal components of

a vector field X on M. The restrictions of the tensors T and A to V and H
can be interpreted respectively as the second fundamental form of the leaves of
Fπ and the integrability of the horizontal distribution H . In particular, we say
that a Riemannian submersion is integrable if the restriction of A to H vanishes
identically.
Analogously to the local extension of vector fields for Riemannian immersions,
there exists a horizontal lift property for Riemannian submersions. Namely, a vector
field X on M can always be lifted to a horizontal vector field X on M that is π -related
to X, i.e., such that dπ (X) = X ◦ π . Using this, it is not hard to see that if X and Y
are vector fields on M, then

∇X Y − ∇X Y = AX Y = 12 [X,Y ]V , (2.13)
38 2 Lie Groups with Bi-invariant Metrics

see Petersen [183, p. 82]. The relation between curvatures on M and M was first
studied by Gray [104] and O’Neill [177], who proved that

R(X ,Y, X,Y ) = R(X,Y , X ,Y ) + 3AX Y 2

(2.14)
= R(X,Y , X ,Y ) + 34 [X,Y ]V 2 ,

for all vectors X and Y on M. The above is called the Gray-O’Neill formula.
In particular, it follows that if M satisfies a certain lower sectional curvature bound,
then so does M. Formulas for the other sectional curvatures of M (containing vertical
directions), as well as the implications for Ricci and scalar curvatures, can be found
in Besse [33, Chap 9].

2.2 Bi-invariant Metrics

The main goal of this section is to study the special Riemannian structure on Lie
groups given by bi-invariant metrics. We mostly use ·, · to denote metrics on Lie
groups, to avoid any confusion between a metric g(·, ·) and an element g ∈ G. Bi-
invariant metrics on Lie groups are denoted Q(·, ·).
Definition 2.22. A Riemannian metric ·, · on a Lie group G is left-invariant if Lg
is an isometry for all g ∈ G, that is, if for all g, h ∈ G and X,Y ∈ Th G,

d(Lg )h X , d(Lg )hY gh = X ,Y h .

Similarly, right-invariant metrics are those for which the right translations Rg are
isometries. Note that, given an inner product ·, ·e in Te G, it is possible to define a
left-invariant metric on G by setting, for all g ∈ G and X,Y ∈ Tg G,

X,Y g := d(Lg−1 )g X, d(Lg−1 )gY e ,

and the right-invariant case is analogous.

Definition 2.23. A bi-invariant metric Q on a Lie group G is a Riemannian metric
that is simultaneously left- and right-invariant.
The natural extension of these concepts to k-forms is that a k-form ω ∈ Ω k (G)
is left-invariant if it coincides with its pull-back by left translations, i.e., L∗g ω = ω
for all g ∈ G. Right-invariant and bi-invariant forms are analogously defined. Once
more, given any ωe ∈ Λ k (Te G), it is possible to define a left-invariant k-form ω ∈
Ω (G) by setting, for all g ∈ G and Xi ∈ Tg G,

ωg (X1 , . . . , Xk ) := ωe (d(Lg−1 )g X1 , . . . , d(Lg−1 )g Xk ).

The right-invariant case is analogous.

2.2 Bi-invariant Metrics 39

Proposition 2.24. Every compact Lie group admits a bi-invariant metric Q.

Proof. Let ω be a right-invariant volume form2 on G and ·, · a right-invariant
metric. Define for all X,Y ∈ Tx G,

Q(X,Y )x := dLg X, dLgY gx ω .
G

First, we claim that Q is left-invariant. Fix X,Y ∈ Tx G and consider the function
f : G → R given by f (g) := dLg X, dLgY gx . Then,


Q(dLh X, dLhY )hx = dLg (dLh X ), dLg (dLhY )g(hx) ω
G
 
= dLgh X , dLghY (gh)x ω = f (gh)ω
G G
  
= R∗h ( f ω ) = fω = dLg X, dLgY gx ω = Q(X,Y )x ,
G G G

which proves that Q is left-invariant. We also have


Q(dRh X, dRhY )xh = dLg (dRh X), dLg (dRhY )g(xh) ω
G
 
= dRh dLg X , dRh dLgY (gx)h ω = dLg X, dLgY gx ω = Q(X,Y )x ,
G G

which proves that Q is right-invariant, concluding the proof. 

Exercise 2.25. Consider su(n) = {A ∈ gl(n, C) : A∗ + A = 0, tr A = 0}, the Lie

algebra of SU(n) (see Exercise 1.51). Verify that the inner product in Te SU(n)
defined by Q(X,Y ) = 12 Re tr(XY ∗ ) can be extended to a bi-invariant metric.
Proposition 2.26. Let G be a Lie group endowed with a bi-invariant metric Q, and
X,Y, Z ∈ g. Then the following hold:
(i) Q([X,Y ], Z) = −Q(Y, [X , Z]);
(ii) ∇X Y = 12 [X,Y ];
(iii) R(X,Y )Z = 14 [[X,Y ], Z];
(iv) R(X,Y, X,Y ) = 14 [X ,Y ]2 .
In particular, (G, Q) has nonnegative sectional curvature sec ≥ 0.

2 i.e., ω ∈ Ω n (G) is a nonvanishing n-form, where n = dim G, see Definition A.26.

40 2 Lie Groups with Bi-invariant Metrics

Proof. Differentiating the formula Q(Ad(exp(tX))Y, Ad(exp(tX))Z) = Q(Y, Z) it

follows from Proposition 1.37 and (1.11) that

Q([X ,Y ], Z) + Q(Y, [X, Z]) = 0,

which proves (i). Furthermore, (ii) follows from the Koszul formula (2.1) using (i)
and the fact that ∇ is symmetric.
To prove (iii), we use (ii) to compute R(X ,Y )Z as follows:

R(X,Y )Z = ∇[X,Y ] Z − ∇X ∇Y Z + ∇Y ∇X Z
= 12 [[X ,Y ], Z] − 12 ∇X [Y, Z] + 12 ∇Y [X, Z]
= 12 [[X ,Y ], Z] − 14 [X, [Y, Z]] + 14 [Y, [X , Z]]
 
= 14 [[X ,Y ], Z] + 14 [[X ,Y ], Z] + [[Z, X],Y ] + [[Y, Z], X]
= 14 [[X ,Y ], Z].

Finally, to prove (iv), we use (i) to verify that

Q(R(X ,Y )X,Y ) = 14 Q([[X,Y ], X],Y )

= − 14 Q([X, [X,Y ]],Y )
= 14 Q([X ,Y ], [X,Y ])
= 14 [X,Y ]2 .

Theorem 2.27. The Lie exponential map and the Riemannian exponential map at
the identity agree on Lie groups endowed with bi-invariant metrics. In particular,
the Lie exponential map of a connected compact Lie group is surjective.

Proof. Let G be a Lie group endowed with a bi-invariant metric and X ∈ g. To prove
that the exponential maps coincide, it suffices to prove that the 1-parameter subgroup
γ : R → G given by γ (t) = exp(tX) is the geodesic with γ (0) = e and γ  (0) = X.
First, recall that γ is the integral curve of the left-invariant vector field X passing
through e ∈ G at t = 0, that is, γ  (t) = X(γ (t)) and γ (0) = e. Furthermore, from
Proposition 2.26,

D  D
γ = X(γ (t)) = ∇γ  X = ∇X X = 12 [X, X] = 0.
dt dt
Therefore, γ is a geodesic and the Lie exponential map coincides with the
Riemannian exponential map.
From Proposition 2.24, if G is compact, it admits a bi-invariant metric Q. Using
that exponential maps coincide, and that the Lie exponential map is defined for
2.3 Killing Form and Semisimple Lie Algebras 41

all X ∈ g, it follows from the Hopf-Rinow Theorem 2.9 that (G, Q) is a complete
Riemannian manifold. Thus, exp = expe : Te G → G is surjective. 

Exercise 2.28. Use the fact that exp is not surjective in SL(2, R) to prove that
SL(2, R) does not admit a metric such that the Lie exponential map and the
Riemannian exponential map coincide in e.
Exercise 2.29. Let G be a compact Lie group endowed with a bi-invariant metric.
Prove that each closed subgroup H is a totally geodesic submanifold.
In the next result, we prove that each Lie group G with bi-invariant metric is
a symmetric space, i.e., for each a ∈ G there exists an isometry I a that reverses
geodesics through a (see also Exercise 6.32).
Theorem 2.30. Let G be a connected Lie group endowed with a bi-invariant metric.
For each g ∈ G, set

I g : G → G, I g (x) := gx−1 g.

Then I g is an isometry that fixes g and reverses geodesics through g. In other words,
I g ∈ Iso(G) and if γ is a geodesic with γ (0) = g, then I g (γ (t)) = γ (−t).
Proof. Since I e (g) = g−1 , the map d(I e )e : Te G → Te G is the multiplication by −1,
i.e., d(I e )e = − id. Hence it is an isometry of Te G. Since d(I e )g = d(Rg−1 )e ◦ d(I e )e ◦
d(Lg−1 )g , for any g ∈ G, the map d(I e )g : Tg G → Tg−1 G is also an isometry. Hence
I e is an isometry. It clearly reverses geodesics through e, and since I g = Rg I e R−1g , it
follows that I g is also an isometry that reverses geodesics through g. 

2.3 Killing Form and Semisimple Lie Algebras

To continue our study of bi-invariant metrics, we introduce the Killing form, which
allows to establish a classical algebraic condition for compactness of Lie groups.
Definition 2.31. Let G be a Lie group and X,Y ∈ g. The Killing form of g (also said
Killing form of G) is the symmetric bilinear form
 
B(X ,Y ) := tr ad(X)ad(Y ) .

If B is nondegenerate, then g is said to be semisimple.

We provide in Theorem 2.41 other equivalent definitions of semisimplicity for a
Lie algebra. A Lie group G is said to be semisimple if its Lie algebra g is semisimple.
Proposition 2.32. The Killing form is Ad-invariant, that is, for all X,Y ∈ g and
g ∈ G, it holds that B(Ad(g)X , Ad(g)Y ) = B(X,Y ).
Proof. If ϕ : g → g is a Lie algebra automorphism, then ad(ϕ (X))ϕ (Y ) = ϕ ◦
ad(X)Y . Thus, ad(ϕ (X)) = ϕ ◦ ad(X) ◦ ϕ −1 and hence
42 2 Lie Groups with Bi-invariant Metrics

 
B(ϕ (X), ϕ (Y )) = tr ad(ϕ (X))ad(ϕ (Y ))
 
= tr ϕ ◦ ad(X)ad(Y ) ◦ ϕ −1
 
= tr ad(X)ad(Y )
= B(X,Y ).

Since Ad(g) is a Lie algebra automorphism, the proof is complete. 

Corollary 2.33. Let G be a semisimple Lie group with negative-definite Killing

form B. Then −B is a bi-invariant metric.
Remark 2.34. Let G be a Lie group endowed with a bi-invariant metric Q. From
Proposition 2.26, it follows that

Ric(X,Y ) = tr R(X, ·)Y = tr 14 [[X, ·],Y ] = − 14 tr[Y, [X, ·]] = − 14 B(X,Y ). (2.15)

Therefore, the Ricci tensor of (G, Q) is independent of the bi-invariant metric Q.

Theorem 2.35. Let G be a n-dimensional semisimple connected Lie group. Then G
is compact if and only if its Killing form B is negative-definite.
Proof. First, suppose that B is negative-definite. From Corollary 2.33, −B is a
bi-invariant metric on G. Hence, the Hopf-Rinow Theorem 2.9 and Theorem 2.30
imply that (G, −B) is a complete Riemannian manifold, whose Ricci curvature
satisfies (2.15). It follows from the Bonnet-Myers Theorem 2.19 that G is compact.
Conversely, suppose G is compact. From Proposition 2.24, it admits a bi-invariant
metric Q. Hence, using item (i) of Propositions 2.26 and 1.37, it follows that, if
(e1 , . . . , en ) is an orthonormal basis of g, then
 
B(X, X) = tr ad(X )ad(X)
n
= ∑ Q(ad(X)ad(X)ei , ei )
i=1
n
= − ∑ Q(ad(X)ei , ad(X)ei )
i=1
n
= − ∑ ad(X)ei 2 ≤ 0.
i=1

Note that, if there exists X = 0 such that ad(X)ei 2 = 0 for all i, then by
definition of the Killing form, B(Y, X) = 0 for each Y . This would imply that B
is degenerate, contradicting the fact that g is semisimple. Thus, for each X = 0, we
have B(X, X) < 0. Therefore, B is negative-definite. 

The next result follows directly from Corollary 2.33, Remark 2.34 and
Theorem 2.35.
2.3 Killing Form and Semisimple Lie Algebras 43

Corollary 2.36. Let G be a semisimple compact connected Lie group with Killing
form B. Then (G, −B) is an Einstein manifold with positive Einstein constant λ = 14 .
We conclude this section with a discussion on equivalent definitions of semisim-
ple Lie algebras. The key concept in this discussion is that of ideal of a Lie algebra.
A Lie subalgebra h is an ideal of a Lie algebra g if [X,Y ] ∈ h, for all X ∈ h and
Y ∈ g. Using tools from the previous chapter, we now prove the following important
relation between ideals of a Lie algebra and normal subgroups of a Lie group.
Proposition 2.37. Let G be a Lie group with Lie algebra g. If h is an ideal of g,
then the connected Lie subgroup H with Lie algebra h is a normal subgroup of G.
Conversely, if H is a normal Lie subgroup of G, then its Lie algebra h is an ideal
of g.
Proof. Suppose h is an ideal of g, and let H be the connected Lie subgroup with
Lie algebra h given by Proposition 1.21. Then, by Proposition 1.37, we have that
ad(X)Y ∈ h for all X ∈ g and Y ∈ h. From (1.12), it follows that

∞
ad(X )k
Ad(exp(X))Y = exp(ad(X))Y = ∑ k! Y ∈ h.
k=0

On the other hand, from (1.10),

exp(Ad(exp(X))Y ) = exp(X) exp(Y ) exp(X)−1 ,

and hence exp(X) exp(Y ) exp(X)−1 ∈ H. From Proposition 1.20, it follows that H is
a normal subgroup of G.
Conversely, suppose H is a normal subgroup of G, that is gHg−1 = H for all g ∈
G. Then exp(tAd(g)X) = g exp(tX)g−1 ∈ H for all g ∈ G and X ∈ h. Differentiating
at t = 0, we have that Ad(g) maps h to itself. In particular, Ad(exp(tX))Y ∈ h for all
X ∈ g and Y ∈ h. Differentiating again at t = 0 and using Proposition 1.37, we have
that [X,Y ] ∈ h for all X ∈ g and Y ∈ h, that is, h is an ideal of g. 

If an ideal h has no ideals other than the trivial, {0} and h, it is called simple.
Following the usual convention, by simple Lie algebras we mean Lie subalgebras
that are noncommutative simple ideals. We stress that simple ideals (which may be
commutative) are not referred to as Lie algebras, but simply called simple ideals.
Given a Lie algebra g, consider the decreasing sequence of ideals

g(1) := [g, g], g(2) := [g(1) , g(1) ], ... g(k) := [g(k−1) , g(k−1) ], ...

If there exists a positive integer m such that g(m) = {0}, then g is said to be solvable.
Example 2.38. Consider the ideal of all n × n matrices (aij ) over K = R or K = C,
with aij = 0 if i > j. One can easily verify that this is a solvable Lie algebra. Other
trivial examples are nilpotent Lie algebras.
44 2 Lie Groups with Bi-invariant Metrics

It is possible to prove that every Lie algebra admits a maximal solvable ideal
τ , called its radical. We recall some results whose proof can be found in Ise and
Takeuchi [133].
Proposition 2.39. The following hold:
(i) If h is an ideal of g, then the Killing form Bh of h satisfies B(X,Y ) = Bh (X,Y ),
for all X,Y ∈ h;
(ii) If g = g1 ⊕ g2 is direct sum of ideals, then g1 is orthogonal to g2 with respect
to B. Thus B is the sum of the Killing forms B1 and B2 of g1 and g2 , respectively.
Cartan Theorem 2.40. A Lie algebra g is solvable if and only if B(g, g(1) ) = {0}.
In particular, if B vanishes identically, then g is solvable.
We are now ready to present a theorem that gives equivalent definitions of
semisimple Lie algebras.
Theorem 2.41. Let g be a Lie algebra with Killing form B. Then the following are
equivalent:
(i) g = g1 ⊕ · · · ⊕ gn is the direct sum of simple Lie algebras gi (i.e., noncommuta-
tive simple ideals);
(ii) g has trivial radical τ = {0};
(iii) g has no commutative ideal other than {0};
(iv) B is nondegenerate, i.e., g is semisimple.
Before proving this theorem, we present two important properties of semisimple
Lie algebras.
Remark 2.42. If the Lie algebra g is the direct sum of noncommutative simple ideals
g = g1 ⊕· · ·⊕gn , then [g, g] = g. Indeed, on the one hand, [gi , g j ] = {0} if i = j, since
(1)
gi and g j are ideals. On the other hand, [gi , gi ] = gi , since gi = [gi , gi ] is an ideal
and gi is a noncommutative simple ideal.
Remark 2.43. If g = g1 ⊕ · · · ⊕ gn is the direct sum of simple Lie algebras, then this
decomposition is unique up to permutations. In fact, consider another decomposition
g= gm and let X ∈ 
g1 ⊕ · · · ⊕  gk . Then X = ∑i Xi , where Xi ∈ gi . Since gi is a
noncommutative simple ideal, for each i such that Xi = 0, there exists Vi ∈ gi
different from Xi , such that [Vi , Xi ] = 0. Hence [X,Vi ] = 0 is a vector that belongs
to both gi and  gk . Therefore, the ideal gi ∩ 
gk is different from {0}. Since gi and  gk
are simple ideals, it follows that gi = gi ∩  gk = gk .
We now prove Theorem 2.41, following Ise and Takeuchi [133, pp. 71–72].
Proof. We proceed by proving the equivalences (i) ⇔ (ii), (i) ⇔ (iii) and (i) ⇔ (iv).
(i) ⇔ (ii). Assume that g = g1 ⊕ · · · ⊕ gn is the direct sum of noncommutative
simple ideals. Let πi : g → gi denote the projection onto each factor and note that
πi is a Lie algebra homomorphism. Thus, the projection τi = πi (τ ) is a solvable
ideal of gi . Since gi is simple, then either τi is equal to {0} or to gi . However,
2.4 Splitting Lie Groups with Bi-invariant Metrics 45

the solvable ideal τi cannot be equal to gi , since gi = [gi , gi ] (see Remark 2.42).
Therefore, τi = {0} and hence τ = {0}.
Conversely, assume that τ = {0}. For any ideal h of g, set h⊥ = {X ∈ g : B(X, h) =
0}. Note that h⊥ and h⊥ ∩h are ideals, and B restricted to h⊥ ∩h vanishes identically.
It then follows from Proposition 2.39 and the Cartan Theorem 2.40 that h⊥ ∩ h
is solvable. Since the radical is trivial, we conclude that h⊥ ∩ h = {0}. Therefore
g = h ⊕ h⊥ . Since g is finite-dimensional, by induction, g is the direct sum of simple
ideals. Moreover, τ = {0} implies that each simple ideal is noncommutative.
(i) ⇔ (iii). Suppose that there exists a nontrivial commutative ideal a. Then the
radical must contain a, hence it is nontrivial. It then follows from (i) ⇔ (ii) that g is
not a direct sum of noncommutative simple ideals.
Conversely, assume that g is not a direct sum of noncommutative simple ideals.
Then, from (i) ⇔ (ii), the radical τ is nontrivial, and there exists a positive integer m
such that τ (m−1) = {0}. Set a = τ (m−1) and note that a is a nontrivial commutative
ideal.
(i) ⇔ (iv). Assume that g = g1 ⊕ · · · ⊕ gn is the direct sum of noncommutative
simple ideals. From Proposition 2.39, it suffices to prove that B|gi is nondegenerate
for each i. Consider a fixed i and let h := {X ∈ gi : B(X, gi ) = 0}. Note that h is an
ideal of gi . Since gi is a simple ideal, either h = gi or h = {0}. If h = gi , then the
Cartan Theorem 2.40 implies that gi is solvable, contradicting the fact that [gi , gi ] =
gi . Therefore h = {0}, hence B|gi is nondegenerate.
Conversely, assume that B is nondegenerate. From the last equivalence, it
suffices to prove that g has no commutative ideals other than {0}. Let a be a
commutative ideal of g.  For each X ∈ a and Y ∈ g we have ad(X)ad(Y )(g) ⊂ a.
Therefore B(X,Y ) = tr ad(X)ad(Y ) |a . On the other hand, since a is commutative,
ad(X)ad(Y )|a = 0. Therefore B(X ,Y ) = 0 for each X ∈ a and Y ∈ g. Since B is
nondegenerate, it follows that a = {0}. 

Exercise 2.44. Show that if a compact Lie group G is semisimple, then its center
Z(G) is finite. Conclude from Exercise 1.60 that U(n) is not semisimple.

2.4 Splitting Lie Groups with Bi-invariant Metrics

In Proposition 2.24, we proved that each compact Lie group admits a bi-invariant
metric. In this section, we prove that the only simply-connected Lie groups that
admit bi-invariant metrics are products of compact Lie groups with vector spaces.
We also prove that, if the Lie algebra of a compact Lie group G is simple, then the
bi-invariant metric on G is unique up to multiplication by constants.
Theorem 2.45. Let g be a Lie algebra endowed with a bi-invariant metric. Then

g = g1 ⊕ . . . ⊕ gn is the direct orthogonal sum of simple ideals gi . In addition, let G
be the connected and simply-connected Lie group with Lie algebra isomorphic to g.
Then G is isomorphic to the product of normal Lie subgroups G1 × . . . × Gn , such
that Gi = R if gi is commutative, and Gi is compact if gi is noncommutative.
46 2 Lie Groups with Bi-invariant Metrics

Proof. In order to verify that g is the direct orthogonal sum of simple ideals, it
suffices to prove that, if h is an ideal, then h⊥ is also an ideal. Let X ∈ h⊥ , Y ∈ g and
Z ∈ h. Then, using Proposition 2.26, it follows that

Q([X,Y ], Z) = −Q([Y, X], Z) = Q(X, [Y, Z]) = 0.

Hence [X,Y ] ∈ h⊥ , and this proves the first assertion.

From Lie’s Third Theorem 1.14, given gi , there exists a unique connected and
simply-connected Lie group Gi with Lie algebra isomorphic to gi . In particular,
G1 × . . . × Gn is a connected and simply-connected Lie group, with Lie algebra
g = g1 ⊕ . . . ⊕ gn . From uniqueness in Lie’s Third Theorem, it follows that G =
G1 × . . . × Gn . If gi is commutative and simple, then gi = R. Hence, since Gi is
connected and simply-connected, Gi ∼ = R. Else, if gi is noncommutative, observe
that there does not exist X ∈ gi , X = 0, such that [X,Y ] = 0 for all Y ∈ gi . Indeed,
if there existed such X , then span{X} ⊂ gi would be a nontrivial ideal. From the
proof of Theorem 2.35, the Killing form of gi is negative-definite; and from the
same theorem, Gi is compact. Finally, since gi is an ideal of g, it follows from
Proposition 2.37 that Gi is a normal subgroup of G.  

The above theorem and Remark 2.42 imply the following corollary.
Corollary 2.46. Let g be a Lie algebra with a bi-invariant metric. Then g =  g⊕
Z(g) is a direct sum of ideals, where 
g is semisimple. In particular, [ g] = 
g, g.
Remark 2.47. Let G be a compact connected Lie group with a nontrivial connected
and simply-connected normal Lie subgroup H. From the proof of Theorem 2.45,
it follows that there exists a connected normal Lie subgroup L of G such that
G∼ = (H × L)/Γ , where Γ is a finite normal subgroup of H × L. In other words,
if G admits a nontrivial normal subgroup, then up to a finite covering it splits
as a product of Lie groups. For instance, SU(n) is a normal subgroup of U(n),
and U(n) ∼= (SU(n) × S1 )/Zn , see Exercise 1.60. Furthermore, the Lie algebra of
U(n) splits as a direct sum of ideals u(n) = su(n) ⊕ Z(u(n)), see Corollary 2.46.
From Remark 1.56, another example is SO(4) ∼ = (SU(2) × SU(2))/Z2 , that has two
normal subgroups isomorphic to SU(2).
Proposition 2.48. Let G be a compact simple Lie group with Killing form B and
bi-invariant metric Q. Then the bi-invariant metric is unique up to rescaling.
In addition, (G, Q) is an Einstein manifold, that is, it satisfies (2.6).
Proof. Let Q  be another bi-invariant metric on G, and let P : g → g be the positive-

definite self-adjoint operator such that Q(X,Y ) = Q(PX,Y ). We claim that P and ad
commute, that is, P ad(X) = ad(X)P for all X ∈ g. Indeed,


Q(P ad(X)Y, Z) = Q(ad(X)Y, Z)
 ad(X)Z)
= −Q(Y,
= −Q(PY, ad(X )Z)
= Q(ad(X )PY, Z).
2.4 Splitting Lie Groups with Bi-invariant Metrics 47

Furthermore, eigenspaces of P are ad(X)-invariant, that is, they are ideals. In fact,
let Y ∈ g be an eigenvector of P associated to an eigenvalue μ . Then P ad(X)Y =
ad(X)PY = μ ad(X)Y. Since g is simple, it follows that P = μ id, hence Q  = μ Q.
Since G is compact, it follows from Theorem 2.35 and Corollary 2.33 that −B is
a bi-invariant metric. Hence, there exists λ such that −B(X,Y ) = 4λ Q(X,Y ). There-
fore, from Remark 2.34, Ric(X,Y ) = − 14 B(X ,Y ) = λ Q(X,Y ), so G is Einstein.

Exercise 2.49. Let G be a compact semisimple Lie group with Lie algebra g =
g1 ⊕ · · · ⊕ gn given by the direct sum of noncommutative simple ideals. Consider a
bi-invariant metric Q on G. Prove that there exist positive numbers λ j such that

Q = ∑ −λ j B j ,
j

where B j = B|g j is the restriction of the Killing form to g j .

Exercise 2.50. In order to compute the Killing form B of SU(n), recall that its Lie
algebra is

su(n) = A ∈ gl(n, C) : A∗ + A = 0, tr A = 0 ,

see Exercise 1.51. Consider the following special diagonal matrices in su(n),
⎛ ⎞ ⎛ ⎞
iθ1 iζ1
⎜ .. ⎟ ⎜ .. ⎟
X =⎝ . ⎠ and Y =⎝ . ⎠.
iθn iζn

Use the fact that tr X = trY = 0, hence ∑ni=1 θi = ∑ni=1 ζi = 0, and Exercise 1.38 to
verify that computing B in the above X and Y gives:
n
B(X ,Y ) = tr (ad(X)ad(Y )) = −2n ∑ θi ζi .
i=1

From Exercise 2.25, the inner product Q(Z,W ) = Re tr(ZW ∗ ) in Te SU(n) can be
extended to a bi-invariant metric Q. Since SU(n) is simple, from Proposition 2.48,
there exists a constant c ∈ R such that B = c Q. Conclude that c = −4n and that the
Killing form of SU(n) is

B(Z,W ) = −2n Re tr(ZW ∗ ), for all Z,W ∈ su(n).

 Part II
Isometric Actions
Chapter 3
Proper and Isometric Actions

In this chapter, we present a concise introduction to the theory of proper and

isometric actions. We begin with basic definitions and a quick primer on fiber
bundles, leading to the Slice Theorem 3.49 and the Tubular Neighborhood
Theorem 3.57. These fundamental results are used throughout the rest of the book;
in particular, to establish a strong correspondence between proper and isometric
actions in Sect. 3.3. Finally, the stratification of a manifold by orbit types of a
proper action is studied in Sect. 3.5.
The following references were a source of inspiration for this chapter, and
could serve as further reading material: Bredon [53], Duistermaat and Kolk [79],
Kawakubo [137], Onishchik [179], Palais and Terng [182], Pedrosa [11, Part I],
Spindola [197] and Walschap [222].

3.1 Proper Actions and Fiber Bundles

In this section, proper actions are introduced together with a preliminary study of
fiber bundles, and accompanied by several examples.
Definition 3.1. Let G be a Lie group and M a smooth manifold. A smooth map
μ : G × M → M is called a (left) action of G on M, or a (left) G-action on M, if
(i) μ (e, x) = x, for all x ∈ M;
(ii) μ (g1 , μ (g2 , x)) = μ (g1 g2 , x), for all g1 , g2 ∈ G and x ∈ M.
Whenever μ is implicit, the manifold M is called a G-manifold or G-space, and it
is common to denote μ (g, x) by g · x, or even g x, but we avoid this slight abuse of
notation until later in the book.
Similarly, one can define a right action μ : M × G → M of G on M to be a smooth
map satisfying properties analogous to (i) and (ii) above, see Example 3.8. For a
right action, the short notation for μ (x, g) is x · g, or x g.

© Springer International Publishing Switzerland 2015 51

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_3
52 3 Proper and Isometric Actions

For example, evaluating matrices on vectors defines a left action of G = GL(n, R)

on M = Rn , given by μ (A, x) = A x. In Chap. 2, we have seen that every Lie group
G acts on its Lie algebra g via the adjoint action μ : G × g → g, μ (g, X) = Ad(g)X,
and this action is studied in detail in Chap. 4. Any Lie subgroup H ⊂ G of a Lie
group G determines a left (and a right) H-action on G, given by multiplication on
the left (and, respectively, on the right). In addition, conjugation by elements in H
is also an H-action on G.
Definition 3.2. Let μ : G × M → M be a left action and x ∈ M. The subgroup

Gx := g ∈ G : μ (g, x) = x ⊂ G

is called isotropy group or stabilizer of x ∈ M, and


G(x) := μ (g, x) : g ∈ G ⊂ M

is called the orbit of x ∈ M. If G(x) = {x}, then x is called a fixed point of the action.

The subgroup x∈M Gx is called the ineffective kernel of the action; if it is the
trivial group {e}, then the action is said to be effective. Moreover, if Gx = {e}, for
all x ∈ M, the action is said to be free. If for all x, y ∈ M, there exists g ∈ G with
μ (g, x) = y, then the action is said to be transitive. Notice that the isotropy group Gx
of a fixed point is the entire group G, and an action with fixed points cannot be free.
Remark 3.3. Every G-action can be reduced to an effective action of the quotient of

G by the ineffective kernel x∈M Gx , which is a normal subgroup of G.
Given a (left) G-action on M, let us also define two auxiliary maps1 :

μ g : M −→ M μx : G −→ M
(3.1)
x −→ μ (g, x) g −→ μ (g, x).

The key idea of an action is that each g ∈ G determines a transformation of M,

namely μ g : M → M. Since our actions are assumed smooth, μ g are diffeomor-
phisms, and hence μ G := {μ g : g ∈ G} can be identified with a subgroup2 of the
diffeomorphism group Diff(M). An orbit G(x) consists of all possible images μ g (x)
for g ∈ G, and the isotropy group Gx consist of all g ∈ G that fix x = μ g (x). If M is
endowed with a Riemannian metric g, an action on (M, g) is said to be isometric, or
by isometries, if μ g is an isometry of (M, g) for all g ∈ G. In this case, the metric g
is said to be G-invariant, and μ G can be identified with a subgroup of Iso(M, g).

1 Analogous maps can be defined for a right G-action, simply replacing μ (g, x) by μ (x, g).
2 Notice that μ G is isomorphic to G only if the action μ is effective.
3.1 Proper Actions and Fiber Bundles 53

Example 3.4. The following are three basic examples of actions on M = R3 .

(i) Let G = SO(3) and set μ (A, x) = A x. The orbit G(x) of a nonzero vector x ∈ R3
is the sphere of radius x centered at the origin, and the isotropy group Gx is
isomorphic to the group SO(2) of rotations around the line spanned by x. The
origin x = 0 is a fixed point of this action.
(ii) Let G = SO(2) and set μ (A, x) = (A(x1 , x2 ), x3 ). If x ∈ R3 is such that r2 =
x12 + x22 > 0, then the orbit G(x) is a circle of radius r contained in the plane
perpendicular to the z-axis at (0, 0, x3 ), and the isotropy group is Gx = {e}. If
r2 = x12 + x22 = 0, then x is a fixed point of the action.
(iii) Let G = SO(2) × R and set μ ((A, b), x) = (A(x1 , x2 ), b + x3 ). The orbit G(x)
of a vector x ∈ R3 with r2 = x21 + x22 > 0 is a round cylinder of radius r whose
rotation axis is the z-axis, and the isotropy group is Gx = {e}. The orbit G(x)
of a vector x = (0, 0, x3 ) is the z-axis, and the isotropy group Gx = SO(2) is the
group of rotations around the z-axis.
Action (ii) is called a subaction of action (i), since it is the restriction of the
action (i) to the subgroup of SO(3) consisting of block diagonal matrices whose
first 2 × 2 block is an element of SO(2). Action (ii) is also clearly a subaction of
action (iii). Both (i) and (ii) are called linear actions (or (linear) representations,
see Definition 1.35) since μ g is a linear transformation for all g ∈ G. Action (iii) is
not linear.
Example 3.5. Let us describe a few operations to construct new actions using other
actions. As mentioned in Example 3.4, restricting a G-action on M to a subgroup H
defines an H-action on M called a subaction. Given (left) actions μ1 : G1 ×M1 → M1
and μ2 : G2 × M2 → M2 , we may define a product action of G1 × G2 on M1 × M2 by

μ : (G1 × G2 ) × (M1 × M2 ) → M1 × M2 ,
    (3.2)
μ (g1 , g2 ), (x1 , x2 ) := μ1 (g1 , x1 ), μ2 (g2 , x2 ) .

If G1 = G2 = G, then restricting the product action (3.2) to the diagonal subgroup

Δ G := {(g, g) ∈ G× G} defines a subaction called diagonal action of G on M1 ×M2 .
Definition 3.6. Consider (left) G-actions μ1 : G×M1 → M1 and μ2 : G×M2 → M2 .
A map f : M1 → M2 is called G-equivariant if μ2 (g, f (x)) = f (μ1 (g, x)) for all x ∈
M1 and g ∈ G. Equivariant diffeomorphisms provide a natural notion of equivalence
among G-manifolds.
Exercise 3.7.  Let μ : G × M
 → M be a G-action.
 Verify that μ : G × TM → TM
defined by μ g, (p, v) := μ (g, p), d( μ g ) p v is a G-action on TM, and that the
projection π : TM → M, π (p, v) = p, is a G-equivariant map. Notice that μ induces a
subaction by the isotropy group G p on Tp M, which is called isotropy representation.
54 3 Proper and Isometric Actions

Example 3.8. Let μL : G × M → M be a left G-action, and define μR : M × G → M

by setting μR (x, g) := μL (g−1 , x). Then μR is a right G-action, and the identity map
on M is a G-equivariant diffeomorphism with respect to these actions. Right actions
can be analogously transformed into left actions.
Exercise 3.9. Let μ : G × M → M be a left action. Prove that the isotropy group Gx
changes by conjugation as x moves along its orbit G(x). More precisely, show that
Gμ (g,x) = g Gx g−1 .
If two orbits G(x) and G(y) have nontrivial intersection, then they coincide. This
means that orbits of a G-action on M form a partition of M; and hence we can
consider the quotient

M/G := G(x) : x ∈ M ,

called the orbit space or quotient space of the G-action on M. The natural projection
π : M → M/G, π (x) := G(x), is called the quotient map, or projection map, and the
topology on M/G is determined by declaring that U ⊂ M/G is open if its preimage
π −1 (U) ⊂ M is open. This implies that π is continuous, and it is possible to prove
that π is also an open map, i.e., maps open subsets of M to open sets of M/G.
Exercise 3.10. Find the orbit spaces M/G of each G-action on M = R3 from
Example 3.4, by identifying a subset of R3 that contains a point representing each
G-orbit.
Example 3.11. Consider the S1 -action on C by complex multiplication eiθ · z = eiθ z,
and the resulting product action3 of the torus S1 × S1 on C × C. The diagonal action
   
μ : S1 × (C × C) → (C × C), μ eiθ , (z1 , z2 ) = eiθ z1 , eiθ z2 (3.3)

is called the Hopf action. This S1 -action on C2 can be restricted to an S1 -action on

the unit sphere S3 = {(z1 , z2 ) ∈ C2 : |z1 |2 + |z2 |2 = 1}, using the same formula (3.3),
which is also called Hopf action.
In order to describe the orbit space S3 /S1 of the Hopf action, consider the
Northern hemisphere Σ := {(z1 , x3 ) ∈ C × R : |z1 |2 + x23 = 1, x3 > 0} of a 2-
dimensional subsphere in S3 . If z = (z1 , z2 ) ∈ S3 with z2 = 0, then there exists a
unique eiθ ∈ S1 such that eiθ z2 is a (real) positive number, i.e., such that μ (eiθ , z) ∈
Σ . On the other hand, the orbit of z = (z1 , 0) ∈ S3 is the equator ∂ Σ . Therefore,
topologically, S3 /S1 = Σ ∪ [(z1 , 0)] = S2 is the compactification of the 2-disk, hence
a 2-sphere. More details on its geometry are discussed later, see Exercise 3.31.
Exercise 3.12. Let θ : [0, +∞) → R be a smooth function. Verify that μ (t, z) :=
eit θ (|z|) z defines an R-action on C, and that an orbit of this action is either a circle

3 For more details on this action, see Example 6.48.

3.1 Proper Actions and Fiber Bundles 55

centered at the origin or a fixed point. Describe the corresponding orbit space, in
terms of the set θ −1 (0) = {r ∈ [0, +∞) : θ (r) = 0} where θ vanishes.
In order to study the geometry and topology of orbit spaces, we first need to
develop more tools. The next result asserts that, given a smooth action, we can
associate to each element of g a vector field on M.
Proposition 3.13. Consider a smooth action μ : G × M → M.
(i) Each X ∈ g induces a smooth vector field X ∗ on M, called action field, by

X ∗ (p) := 
dt μ (exp(tX), p) t=0 .
d

∗
(ii) The flow of X ∗ is given by ϕtX = μ exp(tX) .
Proof. Define σ X : M → TG × TM by σ X (p) := (Xe , 0 p ), and note that d μ ◦ σ X is
smooth. Item (i) follows from the fact that X ∗ (p) is given by
d   
dμ ◦ σ X (p) = dμ  
dt (exp(tX), p) t=0 = dt μ (exp(tX), p) t=0 .
d

For each X ∈ g and p ∈ M, notice that μ exp(tX) satisfies μ exp(tX) = id if t = 0, and

 
dt μ
d exp(tX)
(p)t=t = d exp((s+t0 )X)
ds μ (p)s=0
0
 
= ds μ
d
exp(sX) · exp(t0 X ), p) s=0
= X ∗ ( μ (exp(t0 X), p))
 
= X ∗ μ exp(t0 X) (p) ,
∗
proving that ϕtX = μ exp(tX) is the flow of the action field X ∗ . 

Remark 3.14. It is easy to verify that the association g  X → X ∗ ∈ X(M) defined

in (i) is a Lie anti-homomorphism, i.e., [X,Y ]∗ = −[X ∗ ,Y ∗ ].
A consequence of Proposition 3.13 is that, if all orbits of a G-action on M have
the same dimension, then the partition

F = G(p) p∈M (3.4)

is a foliation of M, see Definition A.16. The fact that orbits of a G-action are indeed
smooth submanifolds is proved in Proposition 3.41, and the fact that for each v ∈
Tp G(p) there exists X ∈ g such that Xp∗ = v follows from (3.11). If the orbits of a
G-action have varying dimension, then F is a singular foliation (see Remark 3.42
and Definition 5.1). Foliations of the form (3.4) are called homogeneous foliations.
56 3 Proper and Isometric Actions

Exercise 3.15. Consider the action μ : SO(3) × R3 → R3 in Example 3.4 (i).

(i) Verify that if X ∈ so(3) ∼
= R3 , then X ∗ (p) = X × p;
(ii) Consider AX defined in Exercise 1.9. Verify that etAX is a rotation in R3 about
the axis spanned by X with angular speed X.
For future reference, let us state the result of a simple computation regarding the
derivative of the map μx in (3.1), for both left and right actions.
Proposition 3.16. If μ is a left action, then the derivative of μx : G → M in (3.1)
satisfies ker d( μx )g0 = Tg0 (g0 Gx ). If μ is a right action, then it satisfies ker d( μx )g0 =
Tg0 (Gx g0 ).
We now introduce the concept of proper actions, originally due to Palais [180].
Properness is a key hypothesis to extend the theory of actions of compact groups to
the noncompact case.
Definition 3.17. An action μ : G × M → M is proper if the map

G × M  (g, x) −→ (μ (g, x), x) ∈ M × M (3.5)

is proper, i.e., if the preimage of any compact subset of M × M under (3.5) is a

compact subset of G × M.
It follows directly from the definition that each isotropy group of a proper action
is compact, since the preimage of (x, x) ∈ M × M under the above map is Gx × {x}.
Exercise 3.18. Prove that only proper actions on a compact manifold M are those
by compact groups.
The following characterization of proper actions is easy to prove, and implies
that all actions by compact groups are proper.
Proposition 3.19. An action μ : G × M → M is proper if and only if for any
sequence {gn } in G and any convergent sequence {xn } in M, such that {μ (gn , xn )}
converges, the sequence {gn } admits a convergent subsequence.
Exercise 3.20. Let H be a closed subgroup of the Lie group G. Prove that right
multiplication G × H  (g, h) → g h ∈ G is a free proper right H-action.
Exercise 3.21. An action G × M → M is called properly discontinuous if for all
x ∈ M, there exists a neighborhood U  x such that gU ∩U = 0/ for all g ∈ G \ {e}.
Let G be a discrete group that acts on a manifold M. Prove that this action is properly
discontinuous if and only if it is free and proper.
Exercise 3.22. Prove that the R-action on C defined in Exercise 3.12 is not proper
if θ −1 (0) is nonempty.
Proper actions are closely related to principal bundles. Before describing this
relation, we need some more definitions. We start with that of a fiber bundle.
3.1 Proper Actions and Fiber Bundles 57

Intuitively, a fiber bundle is a space that locally looks like a product space, but
globally may fail to be a product. In the local scale, such objects are products of a
base and a fiber, where we think of the fiber as a space that is “attached” to each
point on the base space to form the fiber bundle. The way in which this attachment
is made is locally coherent to make it look like a product at this level, however, it
allows certain twists in a global scale. Information about these “twists” is encoded
in a Lie group, called the structure group. Let us now give the rigorous definition:
Definition 3.23. Let E, B and F be manifolds and G a Lie group. Assume that
π : E → B a smooth submersion and there is an effective left G-action on F.
Furthermore, suppose that B admits an open covering {Uα } and that there exist
diffeomorphisms ψα : Uα × F → π −1 (Uα ) satisfying:
 
(i) π ◦ ψα (b, f ) = b, for all (b, f ) ∈ Uα × F;
/ then (ψβ−1 ◦ ψα )(b, f ) = (b, θα ,β (b) f ), where θα ,β (b) ∈ G and
(ii) If Uα ∩Uβ = 0,
θα ,β : Uα ∩Uβ → G is smooth.
Then (E, π , B, F, G, {Uα }, {ψα }) is called a coordinate bundle. Moreover,
(E, π , B, F, G, {Uα }, {ψα }) and (E, π , B, F, G, {Vβ }, {ϕβ }) are called equivalent
if (ϕβ−1 ◦ ψα )(b, f ) = (b, θα ,β (b) f ), where θα ,β : Uα ∩ Vβ → G is smooth.
An equivalence class of coordinate bundles, denoted (E, π , B, F, G), is called a
fiber bundle. In this case, E is called the total space, π the projection, B the base, F
the fiber and G the structure group. For each b ∈ B, the preimage π −1 (b) is called
the fiber over b and is often denoted Eb . Furthermore, ψα are called bundle charts
and θα ,β transition functions. A fiber bundle is usually denoted by F → E → B, or
simply by its total space E, if the underlying structure is implicitly understood. We
also say that E is an F-bundle over B.
Example 3.24. The product manifold E = B × F is the total space of the trivial
bundle, for which π : E → B is the projection onto the first factor, G = {e} and the
bundle charts are the identity.
Remark 3.25. The total space E of a fiber bundle can be reconstructed by gluing
trivial products Uα × F according to transition functions θα ,β . More precisely,

consider the disjoint union α (Uα × F). Whenever x ∈ Uα ∩ Uβ , identify (x, f ) ∈

Uα ×F with (x, θα ,β (x)( f )) ∈ Uβ ×F. Then the quotient space E = α (Uα ×F)/ ∼
and the projection onto the first factor is the bundle projection.
Examples of fiber bundles were already encountered in the previous chapters,
such as the tangent bundle TM of a manifold M. A point in TM consists of a pair
(p, v), where v ∈ Tp M. The projection π : TM → M is the map (p, v) → p, and the
fiber over p ∈ M is π −1 (p) = Tp M. The tangent bundle of a manifold is a particular
example of a special class of fiber bundles called vector bundles. Vector bundles
are, by definition, fiber bundles whose fiber is a vector space V and whose structure
group is the group GL(V ) of automorphisms of V . The dimension of V is called the
rank of the vector bundle. Another example of vector bundle is the cotangent bundle
TM ∗ of M, whose fiber at each p ∈ M is the dual space Tp M ∗ . Verifying that these
are indeed vector bundles of rank dim M is an easy exercise.
58 3 Proper and Isometric Actions

Another important example of vector bundle is the normal bundle of a submani-

fold. Let N be a submanifold of M, and consider their tangent bundles TN and TM,
respectively. There is a natural inclusion of Tp N in Tp M for all p ∈ N, and hence one
can consider the quotient space ν p (N) := Tp M/Tp N. This is the fiber of the normal
bundle of N over p. In other words, the normal bundle of N is defined as
 
ν (N) := ν p (N) = Tp M/Tp N, (3.6)
p∈N p∈N

and is a vector bundle over N with rank equal to the codimension of N in M. Notice
that the above definition does not require a Riemannian metric, i.e., in this context,
there is no canonical way of measuring the angle between the normal space ν p N
and the tangent space Tp N. However, if we consider M endowed with a Riemannian
metric g, then the above definition is equivalent to
 
ν (N) := v ∈ Tp M : g p (v, w) = 0 for all w ∈ Tp N ,
p∈N

which explains why it is called the normal bundle. For the same reason, the normal
bundle ν (N) is sometimes denoted TN ⊥ .
Remark 3.26. Certain geometric structures on a fiber bundle can be encoded as a
reduction of its structural group G to a Lie subgroup H ⊂ G. This corresponds to
the bundle admitting an atlas with H-valued transition functions. For instance, if M
is a manifold with dim M = n, the presence of a Riemannian metric on M allows to
reduce the structural group GL(n, R) of TM to the subgroup O(n), and furthermore
to SO(n), provided M is orientable.
Definition 3.27. A smooth section of a fiber bundle F → E → B is a smooth map
σ : B → E with π ◦ σ = idB , that is, σ is a smooth choice of σ (b) ∈ Eb in each fiber.
Example 3.28. Smooth functions on M are sections of the trivial bundle M × R.
Vector fields are sections of the tangent bundle TM, and differential 1-forms are
sections of the cotangent bundle TM ∗ . If N is a submanifold of M, then sections of
the normal bundle ν (N) are normal vector fields along N.
A fiber bundle (P, ρ , B, F, G) is called a principal G-bundle, or principal bundle,
if F = G and the action of G on itself is by left translations. As explained in
Proposition 3.33 and Theorem 3.34 below, a fiber bundle G → P → B is principal if
and only if its fibers are orbits of a free proper G-action (whose orbit space is hence
B).
Example 3.29. There are two very familiar examples of principal G-bundles,
related to basic structures of a smooth manifold M. First, the universal covering
ρ: M  → M is a principal G-bundle over M, where G = π1 (M) is its fundamental
group. Notice that π1 (M) determines a properly discontinuous action on M  by
deck transformations, see Exercise 3.21. Second, consider the frame bundle of M,
defined by
3.1 Proper Actions and Fiber Bundles 59


B(TM) := B(Tx M),
x∈M

where B(Tx M) is the set of all frames (i.e., ordered bases) on the vector space Tx M.
Then B(TM) is a principal G-bundle over M, with G = GL(n, R), where n = dim M.
Indeed, note that B(Tx M) is diffeomorphic to GL(n, R), since each frame ξ = {ξi }
in B(Tx M) determines a linear isomorphism ξ : Rn → Tx M given by ξ (ei ) = ξi ,
where {ei } is the canonical basis of Rn . Note also that composition of isomorphisms
μ (ξ , A) := ξ ◦ A determines a transitive right action of GL(n, R) on B(Tx M).
Example 3.30. The Hopf bundles S1 → S2n+1 → CPn and S3 → S4n+3 → HPn are
respectively principal S1 -bundles and S3 -bundles. Recall that the complex projective
space CPn := S2n+1 /S1 and the quaternionic projective space HPn := S4n+3 /S3 are,
respectively, the spaces of complex lines in Cn+1 and quaternionic lines in Hn+1 .
These are parametrized by equivalence classes [v] of unit vectors v ∈ S2n+1 ⊂ Cn+1 ,
or v ∈ S4n+3 ⊂ Hn+1 , that span the same complex or quaternionic line, respectively.
Such an equivalence class [v] is precisely the orbit of the unit vector v under the
multiplication action of the group S1 of unit complex numbers or the group S3 ∼ =
Sp(1) of unit quaternions. For more details, see Example 6.10.
Exercise 3.31 (). In this exercise, we study the Hopf bundle S1 → S3 → CP1 , see
also Example 3.11. Recall that CP1 is isometric to the sphere S2 12 ⊂ R3 of radius
 
2 . Write S = {(z1 , z2 ) ∈ C : |z1 | + |z2 | = 1} as a subset of C and S
1 3 2 2 2 2 2 1 =
2
{(z, x) ∈ C × R : |z|2 + x2 = 14 } as a subset of C × R. Consider the map
 
1 |z1 |2 − |z2 |2
π : S3 → S 2 2 , π (z1 , z2 ) := z1 z2 , . (3.7)
2

 
(i) Prove that π is well-defined, i.e., it maps S3 to S2 12 ;
(ii) Verify that π (z1 , z2 ) = π (w1 , w2 ) if and only if there exists eiθ ∈ S1 such that
(w1 , w2 ) = (eiθ z1 , eiθ z2 );
(iii) Let N = 0, 12 ∈ S2 12 be the North pole, and consider the open subsets U± =
 
S2 12 \ {±N}. Verify that π −1 (U± ) = S3 \ π −1 (±N);
 
(iv) Describe bundle charts ψ± : U± ×S1 → π −1 (U± ), to verify that π : S3 → S2 12
is a fiber bundle, and moreover, a principal S1 -bundle. Observe that π is also a
Riemannian submersion, see also Exercise 3.81.
Remark 3.32. It is possible to prove that a principal bundle admits a smooth
section if and only if it is trivial, see Walschap [222, p. 68]. This follows from
the correspondence between sections and equivariant bundle charts on a principal
bundle. In particular, the Hopf bundles in Example 3.30 do not admit any smooth
sections.
Proposition 3.33. Principal G-bundles G → P → B have an underlying free proper
right G-action μ : P × G → P whose orbits are the fibers of the bundle.
60 3 Proper and Isometric Actions

Proof. Denote by ρ : P → B the projection map. Given x ∈ M, let ψα : Uα × G →

ρ −1 (Uα ) be a bundle chart with x = ψα (b, f ) ∈ ρ −1 (Uα ), and set
 
μ (x, g) := ψα ψα−1 (x) · g , (3.8)

where (b, f ) · g := (b, f · g), for all b ∈ B and f , g ∈ G. Let us verify that this action
is well-defined, i.e., that the above definition does not depend on the choice of ψα .
This follows from the fact that the structure group acts on the left, while (3.8) is a
right action (see Exercise 3.40). More precisely, for any x ∈ ρ −1 (Uα ) ∩ ρ −1 (Uβ ),
 
ψα ψα−1 (x) · g = ψα (b, fg)
= ψβ (b, θα ,β (b)fg)
 
= ψβ (b, θα ,β (b) f ) · g
 
= ψβ ψβ−1 (x) · g .

The definition of μ and Exercise 3.20 imply that μ is a free proper right action. The
fact that its orbits coincide with the fibers is immediate from (3.8). 

The next theorem provides a converse to the above result, and a method to build
principal bundles. It also shows that, in special cases, M/G is a smooth manifold.
Theorem 3.34. Let μ : M × G → M be a free proper right action. Then the orbit
space M/G admits a smooth structure such that G → M → M/G is a principal
G-bundle, where the bundle projection map ρ : M → M/G is the quotient map.
Remark 3.35. The smooth structure on M/G is such that ρ : M → M/G is smooth,
and a map h : M/G → N is smooth if and only if h ◦ ρ is smooth. These properties
uniquely characterize the smooth structure of M/G.
Proof (Sketch). Let S be a submanifold containing x, with4 Tx M = Tx S ⊕ d(μx )e (g),
and let ϕ : S ×G → M be the restriction of the G-action to S × G, that is, ϕ := μ |S×G .
Claim 3.36. Up to shrinking S, there exists a G-invariant neighborhood U of G(x)
such that ϕ : S × G → U is a G-equivariant diffeomorphism, where the G-action on
S × G is by right multiplication on the second factor.
In order to prove this claim, we first verify that dϕ(s,e) is a linear isomorphism
and hence, by the Inverse Function Theorem, ϕ is a local diffeomorphism on a
neighborhood of S × {e}, up to shrinking S. Indeed, note that ϕ (·, e) is the identity
map on S, and the linearization of ϕ (s, ·) at e also has trivial kernel, since the action
is free, see Proposition 3.16. It then follows from
 
dϕ(s,g) (X, dRgY ) = d(μ g )s ◦ dϕ(s,e) (X,Y )

4A posteriori, this condition means that the submanifold S is transverse to the orbit G(x).
3.1 Proper Actions and Fiber Bundles 61

that ϕ is a local diffeomorphism near each point of S × G. Using that the action
is proper and Proposition 3.19, a standard argument with sequences implies that ϕ
must be injective, which completes the proof of Claim 3.36. Note that ϕ is not a
chart for M/G, since S is contained in M and not in M/G.
The next part of the proof is to identify S with an open subset of M/G. To
this aim, we recall that the quotient map ρ : M → M/G is continuous and open.
Moreover, using that the action is proper, it is possible to prove that M/G is
Hausdorff.5 The above considerations and Claim 3.36 imply that ρ (S) is an open
subset of M/G and ρ |S : S → ρ (S) is a homeomorphism. This allows to define charts
on M/G, and bundle charts on the bundle G → M → M/G. The compatibility of
these charts and bundle charts follows from the following.
Claim 3.37. Let S1 and S2 be submanifolds transverse to two different orbits,
/ Set ρi := ρ |Si : Si → ρ (Si ), Vi := ρi−1 (W ) and
with W := ρ (S1 ) ∩ ρ (S2 ) = 0.
ϕi := μ |Si ×G .
(i) (ρ1−1 ◦ ρ2 ) : V2 → V1 is a diffeomorphism;
 
(ii) (ψ2−1 ◦ ψ1 )(b, g) = (b, θ (b)g), where ψi : ρ (Si ) × G → ρ −1 ρ (Si ) is given by
 −1 
ψi (b, g) = ϕi ρi (b), g and b ∈ W .
In order to prove (i), note that ϕ2−1 (V1 ) is a graph, i.e., ϕ2−1 (V1 ) = {(s, θ(s)) :
s ∈ V2 }, which implies that θ : S → G is smooth. Thus, (ρ1−1 ◦ ρ2 )(s) = ϕ (s, θ(s))
is smooth, and also a diffeomorphism. As for (ii), defining the transition function as
θ (b) := θ(ρ2−1 (b)) guarantees that ψ1 and ψ2 satisfy (ii) in Definition 3.23. 

Corollary 3.38. Let G be a Lie group and H be a closed subgroup. Consider the
right H-action on G by multiplication on the right. The corresponding orbit space
G/H is a manifold, and the quotient map ρ : G → G/H, ρ (g) = gH, determines a
principal H-bundle H → G → G/H. In addition, if H is a normal subgroup of G,
then G/H is a Lie group and ρ is a Lie group homomorphism.
Proof. The first part of the above statement follows directly from Theorem 3.34 and
Exercise 3.20. To show that G/H is a Lie group if H is a normal subgroup, define

α : G × G  (g1 , g2 ) −→ g1 g−1

2 ∈G
(3.9)
 : G/H × G/H  (g1 H, g2 H) −→ g1 g−1
α 2 H ∈ G/H.

Note that, since H is normal, α is well-defined. The fact that ρ : G → G/H is a

projection of a bundle and ρ ◦ α = α ◦ (ρ × ρ ) imply that α
 is smooth. Therefore
G/H is a Lie group, see Remark 1.2. 

5 In fact, since (3.5) is a proper map between locally compact Hausdorff spaces, it is also closed,
i.e., maps closed subsets to closed subsets. Thus, its image R = {(x, y) ∈ M × M : G(x) = G(y)}
is closed. As ρ : M → M/G is an open map, ρ (M × M \ R) is open. This is easily seen to be the
complement of the diagonal in M/G×M/G, which is hence closed, proving that M/G is Hausdorff.
62 3 Proper and Isometric Actions

Example 3.39. Consider the free right action of a Lie group G on itself by
multiplication on the right (see Exercise 3.20) and the induced diagonal action on
G × G,
   
μ : (G × G) × Δ G → G × G, μ (g1 , g2 ), (g, g) := g1 g, g2 g , (3.10)

see Example 3.5. This action is clearly free, hence by Theorem 3.34, there is a
principal bundle Δ G → G × G → (G × G)/Δ G. The map  α : G × G → G defined
in (3.9) descends to a map α : (G × G)/Δ G → G, α [g1 , g2 ] = g1 g−1 2 , which is
well-defined since α (Δ G) = {e}. It is easy to verify that α is determines a Lie
group isomorphism (G × G)/Δ G ∼ = G. This provides a convenient way of rewriting
G as an orbit space, which is useful in several applications (see Remark 4.12).
Exercise 3.40. Suppose that M is a manifold with a free proper right G1 -action
μ1 : M × G1 → M and a left G2 -action μ2 : G2 × M → M. Suppose that μ1 and μ2 are
actions that commute,6 that is, μ1 (μ2 (g2 , x), g1 ) = μ2 (g2 , μ1 (x, g1 )) for all x ∈ M and
gi ∈ Gi . Show that μ2 descends to a left G2 -action μ 2 : G2 × M/G1 → M/G1 , given
2 (g2 , G1 (x)) := G1 ( μ2 (g2 , x)), and verify that π : M → M/G1 is G2 -equivariant.
by μ
Each isotropy group H = Gx of a G-action on M is such that G/H is a smooth
manifold, by Corollary 3.38. We now prove that the orbit G(x) is the image of an
immersion of G/H into M.
Proposition 3.41. Let μ : G × M → M be a left action and define μ x : G/H → M
by μx ◦ ρ = μx , where H = Gx is the isotropy at x ∈ M and ρ : G → G/H is the
quotient map. Then μ x is a G-equivariant7 injective immersion, with image G(x).
In particular, G(x) is an immersed submanifold of M whose tangent space at x ∈ M
x is an embedding
is Tx G(x) = d(μx )e (g). In addition, if the action is proper, then μ
and G(x) is an embedded submanifold of M.

G
µx
r

G/H M
µx

Proof. According to Corollary 3.38, H → G → G/H is a principal H-bundle. This

x is smooth. It follows from Proposition 3.16 that the derivative of μ
implies that μ x
x is an injective immersion. The fact that μ
at every point is injective, and hence μ x is
an embedding when the action is proper can be proved using Proposition 3.19.


μ1 and μ2 on M are said to commute if the induced

6 More generally, two (either left or right) actions

transformations (3.1) on M satisfy (μ1 )g1 ◦ (μ2 )g2 = ( μ2 )g2 ◦ (μ1 )g1 for all gi ∈ Gi .
7 The G-action on G/H is by left translations, that is, g · gH := ggH, see Exercise 3.40 and (6.13).
3.1 Proper Actions and Fiber Bundles 63

As observed above, the tangent space Tx G(x) to an orbit is the image of the linear
map d(μx )e : g → Tx M. Denote by gx the Lie algebra of the isotropy group Gx , and let
mx be a complement, i.e., g = gx ⊕ mx as vector spaces. Then, since ker d(μx )e = gx ,
we have that mx can be identified with Tx G(x). More precisely, there is a natural
identification

mx  X −→ X ∗ (x) ∈ Tx G(x), (3.11)

where X ∗ is the action field on M induced by X ∈ mx , see Proposition 3.13.

Remark 3.42. From Proposition 3.41, orbits G(x) of a left action μ : G × M → M
are immersed submanifolds of M. The dimension of these submanifolds G(x) is
lower semi-continuous on x, i.e., for each x0 ∈ M, the dimension of orbits G(x) for x
near x0 is greater than or equal to dim G(x0 ). This can be proved using that dim G(x)
is the rank of the linear map d(μx )e : g → Tx M, which is in fact constant along G(x).
From continuity of x → d(μx )e and lower semi-continuity of the rank of a continuous
family of linear maps, it follows that dim G(x) is lower semi-continuous.
Example 3.43. Consider the 2-torus T 2 = R2 /Z2as the orbit space of the right
Z2 -action by translations, μ1 : Z2 × R2 → R2 , μ1 (n, m), (x, y) = (x + n, y + m),
recall 1.14. Choose a real number α ∈ R and define a left R-action on R2 by
μ2 (t, (x, y)) := (x + t, y + α t). By Exercise 3.40, μ2 descends to an R-action μ 2
2 are embedded circles in T 2 , which
on T 2 . If α ∈ Q is rational, then the orbits of μ
correspond to straight line segments in R2 joining two points on the lattice Z2 . On
2 are dense8 (and uniformly
the other hand, if α ∈ Q is irrational, then the orbits of μ
2
distributed) in T ; in particular, they are not embedded submanifolds. Recall that,
by Exercise 3.18, the R-action μ 2 on T 2 cannot be proper, despite μ2 being proper.
Analogous statements can be proved for higher dimensional tori.
Exercise 3.44. Verify that the following are equivariantly diffeomorphic:
(i) Sn = SO(n + 1)/SO(n);
(ii) RPn = SO(n + 1)/S(O(n) × O(1));
(iii) CPn = SU(n + 1)/S(U(n) × U(1));
(iv) HPn = Sp(n + 1)/Sp(n) × Sp(1).
Here, S(O(n) × O(1)) is the subgroup of SO(n + 1) formed by the matrices
 
B 0
A= ,
0 ±1

where B ∈ O(n) and det A = 1, and analogously for S(U(1) × U(n)).

8 This is a consequence of the so-called Kronecker Approximation Theorem.

64 3 Proper and Isometric Actions

Hint: Use Example 3.30 and Proposition 3.41. To prove (ii), note that the SO(n + 1)-action on Sn
commutes with the antipodal Z2 -action and hence descends to an action on RPn , see Exercise 3.40.
More details on the solution to this exercise are given in Example 6.10.

Remark 3.45. The same idea in Exercise 3.44 can be used to find equivariant
diffeomorphisms between the k-Grassmannians on Rn , Cn and Hn and quotients
of classical Lie groups. Recall that if V is a finite-dimensional vector space, the
k-Grassmannian on V is defined as

Grk (V ) := W linear subspace of V : dimW = k .

In particular, Gr1 (Rn+1 ) = RPn , Gr1 (Cn+1 ) = CPn and Gr1 (Hn+1 ) = HPn . Observe
that SO(n), SU(n) and Sp(n) act on, respectively, Grk (Rn ), Grk (Cn ) and Grk (Hn ).
These transitive actions provide the equivariant diffeomorphisms:
(ii’) Grk (Rn ) = SO(n)/S(O(n − k) × O(k));
(iii’) Grk (Cn ) = SU(n)/S(U(n − k) × U(k));
(iv’) Grk (Hn ) = Sp(n)/Sp(n − k) × Sp(k).
Remark 3.46. The above also be used to prove that SU(n)/T is a complex flag
manifold, where T is the subgroup of SU(n) formed by diagonal matrices. Indeed,
let F1,...,n−1 (Cn ) be the set of complex flags on Cn , i.e.,

F1,...,n−1 (Cn ) := {0} ⊂ E1 ⊂ . . . ⊂ En−1 : Ei ∈ Gri (Cn ) .

Set F := {(l1 , . . . , ln ) : l1 . . . ln orthogonal lines of Cn }. On the one hand, there is a

natural bijection between F and F1,...,n−1 (Cn ) given by (l1 , . . . , ln ) → {0} ⊂ E1 ⊂
. . . ⊂ En−1 , where Ei = l1 ⊕ · · · ⊕ li . On the other hand, SU(n) acts naturally on F
and this action is transitive. Note that, for an orthogonal basis {e1 , . . . , en } of Cn , the
isotropy group of (Ce1 , . . . , Cen ) ∈ F is T . Therefore, SU(n)/T = F = F1,...,n−1 (Cn ).

3.2 Slices and Tubular Neighborhoods

In this section, we discuss the construction of associated bundles to a principal

bundle, and prove two foundational results for the theory discussed in the remainder
of the book: the Slice Theorem 3.49 and the Tubular Neighborhood Theorem 3.57.
Definition 3.47. Let μ : G × M → M be an action. A slice at x0 ∈ M is an embedded
submanifold Sx0 containing x0 and satisfying the following properties:
(i) Tx0 M = dμx0 g ⊕ Tx0 Sx0 and Tx M = dμx g + Tx Sx0 , for all x ∈ Sx0 ;
(ii) Sx0 is invariant under Gx0 , i.e., if x ∈ Sx0 and g ∈ Gx0 , then μ (g, x) ∈ Sx0 ;
(iii) If x ∈ Sx0 and g ∈ G are such that μ (g, x) ∈ Sx0 , then g ∈ Gx0 .
Example 3.48. Consider the action of SO(2) × R on R3 defined in Exam-
ple 3.4 (iii). A slice for this action at the point x = (x1 , x2 , x3 ) ∈ R3 is either a
3.2 Slices and Tubular Neighborhoods 65

straight line segment Sx = {(t x1 ,t x2 , x3 ) : |t − 1| < ε } if r2 = x21 + x22 > 0, or else,

a disk Sx = {(y1 , y2 , x3 ), ∈ R3 : y21 + y22 < ε } if r2 = x21 + x22 = 0. Notice that these
slices are invariant under the corresponding isotropies Gx = {e} and Gx = SO(2),
respectively.
Slice Theorem 3.49. Let μ : G × M → M be a proper action and x0 ∈ M. Then
there exists a slice Sx0 at x0 .
Proof. We start by constructing a Riemannian metric on M for which the subaction
of the isotropy group H = Gx0 is by isometries. This metric is defined by averaging
the H-action, that is,

X,Y  p := dμ g X, dμ gY μ (g,p) ω , (3.12)
H

where ω is a right-invariant volume form on H and ·, · is an arbitrary Riemannian

metric on M. Arguments similar to those in the proof of Proposition 2.24 imply that
H acts by isometries on M equipped with the metric ·, ·.
It is easy to see that, if g ∈ H, then dμ g : Tx0 M → Tx0 M leaves invariant the
tangent space Tx0 G(x0 ). Since H acts by isometries, we have that dμ g also leaves
invariant the orthogonal complement to Tx0 G(x0 ) in Tx0 M, which is the normal space
νx0 G(x0 ). Define a candidate to slice at x0 by setting
 
Sx0 := expx0 Bε (0) , (3.13)

where Bε (0) is an open ball of radius ε > 0 around the origin in the normal space
νx0 G(x0 ). Since dμ g leaves invariant νx0 G(x0 ) and isometries map geodesics to
geodesics, Sx0 is invariant under the H-action. Therefore, item (ii) of Definition 3.47
is satisfied. Item (i) is satisfied by the construction of Sx0 and continuity of dμ .
It only remains to verify item (iii), which we prove by contradiction. Suppose (iii)
is not satisfied for any ε > 0 in (3.13). Then there exists a sequence {xn } in Sx0 and
a sequence {gn } in G such that lim xn = x0 , lim μ (gn , xn ) = x0 , μ (gn , xn ) ∈ Sx0 and
gn ∈/ H. Since the action is proper, there exists a subsequence, which we also denote
{gn }, that converges to g ∈ H. Set gn := g−1 gn . Then lim gn = e, lim μ ( gn , xn ) =
x0 , μ (
gn , xn ) ∈ Sx0 and gn ∈
/ H. Using Proposition 3.16 and the Inverse Function
Theorem, one can prove the next claim, up to possibly shrinking Sx0 .
Claim 3.50. There exists a submanifold C ⊂ G containing e, such that g = h ⊕ TeC,
where h is the Lie algebra of H. Furthermore, ϕ : C × Sx0  (c, s) → μ (c, s) ∈ M is
a diffeomorphism onto its image.
It follows from the Inverse Function Theorem that for each gn , there exists a
unique cn ∈ C and hn ∈ H, such that gn = cn hn . Since gn ∈ / H, we conclude that
cn = e. On the other hand, μ (hn , xn ) ∈ Sx0 because hn ∈ H. The fact that cn = e
and Claim 3.50 imply that μ (cn , μ (hn , xn )) ∈ / Sx0 . This is a contradicts the fact that
μ (
gn , xn ) ∈ Sx0 , proving that (iii) must be satisfied. 

66 3 Proper and Isometric Actions

We now describe the construction of so-called associated bundles to a principal

bundle, which plays a fundamental role in the description of tubular neighborhoods
of orbits of proper actions.
Let H → P → B be a principal H-bundle and denote by μ1 : P × H → P the
underlying free proper right proper H-action, see Proposition 3.33. Suppose that F
is a manifold that admits a left H-action μ2 : H × F → F. Then the diagonal action9
   
μ : H × (P × F) → (P × F), μ h, (x, f ) := μ1 (x, h−1 ), μ2 (h, f ) (3.14)

is a proper left H-action on P × F, whose orbit space we denote by P ×H F.

Furthermore, given (x, f ) ∈ P × F, denote its projection to the orbit space by
[x, f ] ∈ P ×H F.
Theorem 3.51. The above orbit space P ×H F is a manifold, called twisted space,
and it is the total space of a fiber bundle

F → P ×H F → B,

called
 associated
 bundle with fiber F, whose projection π : P ×H F → B is given
by π [x, f ] = ρ (x), where ρ : P → B is the projection of the principal H-bundle P,
and whose structural group is H.
Proof (Sketch). Given an H-orbit H(x) in P, let S be a submanifold transverse to
H(x) and let U be an H-invariant neighborhood of H(x) in P, as in Claim 3.36.
Claim 3.52. The map ϕ : S × F → U ×H F, ϕ (s, f ) := [s, f ], is a diffeomorphism.
Let ϕ : H × (S × F) → U × F be the restriction of the H-action (3.14) to S × F,
that is, ϕ := μ |H×(S×F) . Similarly to Claim 3.36, it is possible to prove that ϕ is
an H-equivariant diffeomorphism, where the H-action on H × (S × F) is by left
multiplication on the first factor, and the H-action on U × F is given by (3.14).

H (S F) U F

S F U H F

Thus, dividing by the H-action, ϕ descends to the desired diffeomorphism ϕ . This

construction provides charts for the smooth manifold P ×H F.
Let us now define bundle charts for the fiber bundle F → P ×H F → B and verify
their compatibility. Let S1 and S2 be submanifolds transverse to two different orbits,
with ρ (S1 ) ∩ ρ (S2 ) = 0,
/ and set ρi := ρ |Si : Si → ρ (Si ). From Claim 3.52,
!
ψi : ρ (Si ) × F → U ×H F, ψi (b, f ) := ρi−1 (b), f

9 Recall Examples 3.5 and 3.8.

3.2 Slices and Tubular Neighborhoods 67

are diffeomorphisms. Furthermore, we claim that (ψ2−1 ◦ ψ1 )(b, f ) = (b, θ (b) f ),

where θ is as in the proof of Claim 3.37.  In order  to verify this, it suffices to show
that if [s1 , f1 ] = [s2 , f2 ], then f 2 = μ2 θ(s2 ), f 1 , where θ is also as in the proof of
Claim 3.37. The fact that [s1 , f1 ] = [s2 , f2 ] implies that there exists h such that

μ2 (h, f 1 ) = f2 , and μ1 (s2 , h) = s1 . (3.15)

From the proof of Theorem 3.34, s1 = μ1 (s2 , θ(s2 )). Thus, since the action on P is
free, it follows that θ(s2 ) = h, which proves the above claim. 

A familiar example of fiber bundle associated to a principal G-bundle is given by

the tangent bundle TM of a manifold M. This is the associated bundle with fiber Rn
to the principal G-bundle B(TM), defined in Example 3.29, where n = dim M and
G = GL(n, R) acts on Rn by evaluating matrices on vectors.
 
Exercise 3.53. Define ψ : B(TM) ×G Rn → TM by ψ [ξ , v] = ξ (v), where
G = GL(n, R) and a frame ξ ∈ B(Tx M) is interpreted as a linear isomorphism
ξ : Rn → Tx M. Prove that ψ is an identification between these two vector bundles.
 
Hint: To show that ψ is well-defined, note that ψ [ξ · g−1 , g(v)] = ξ (v) for all g ∈ G. To prove
   
that ψ is injective, verify that ψ [ξ1 , v1 ] = ψ [ξ2 , v2 ] if and only if ξ2 = ξ1 · g−1 and v2 = g(v1 )
for some g ∈ G.

Remark 3.54. More generally, if E is a vector bundle over M of rank n and B(E) is
the principal GL(n, R)-bundle of frames of E, then the associated bundle to B(E)
with fiber Rn is canonically identified with the original vector bundle E.
Example 3.55. Let us describe a family of associated bundles with fiber S2 to the
principal S1 -bundle S1 → S3 → S2 , discussed in detail in Exercise 3.31. Writing
S2 = {(z, x) ∈ C × R : |z|2 + x2 = 1}, define a family μk , k ∈ N, of S1 -actions on S2 ,
   
μk : S 1 × S 2 → S 2 , μk eiθ , (z, x) := ekiθ z, x ,

that is, μk is a rotation of speed k on S2 . By Theorem 3.51, the corresponding twisted

spaces Mk = S3 ×S1 S2 are manifolds which are the total space of an S2 -bundle over
S2 , that is, S2 → Mk → S2 . It is not difficult to show that Mk is diffeomorphic to
2
S2 × S2 if k is even, and to CP2 #CP if k is odd.10 In particular, we stress that the
notation P ×G F hides the fact that there may be different G-actions on F which give
rise to different twisted spaces.
Exercise 3.56. Reinterpret the isomorphism (G × G)/Δ G ∼ = G described in Exam-
ple 3.39 in terms of the associated bundle G ×G G to the trivial principal G-bundle
G → G → {e}, where the left G-action on G considered is multiplication on the left.

10 Recall that CP2 denotes the complex projective plane CP2 endowed with the orientation opposite
2
to the standard. It is well-known that the connected sum CP2 #CP is the only nontrivial S2 -bundle
over S , since such bundles are classified by π1 (Diff(S )) = π1 (SO(3)) ∼
2 2 ∼ = Z2 .
68 3 Proper and Isometric Actions

Let μ be a proper G-action on M. Given x0 ∈ M, let Sx0 be a slice at x0 . We define

a tubular neighborhood of the orbit G(x0 ) as the image of Sx0 under the G-action,
that is,
   
Tub G(x0 ) := μ G, Sx0 .
 
The next theorem gives Tub G(x0 ) the structure of an associated fiber bundle; in
particular, it shows that manifolds with proper G-actions are locally G-equivariant
to associated bundles.
Tubular Neighborhood Theorem 3.57. Let μ : G × M → M be a proper action.
For every x0 ∈ M, there exists a G-equivariant diffeomorphism between Tub(G(x0 ))
and the total space of the associated bundle with fiber Sx0 ,

Sx0 → G ×H Sx0 → G/H,

to the principal H-bundle H → G → G/H, where H = Gx0 is the isotropy at x0 .

Remark 3.58. The G-action considered on G ×H Sx0 is given by g · [g, s] := [gg, s].
Details on the principal H-bundle H → G → G/H are found in Corollary 3.38.
 
Proof. Let ϕ : G × Sx0 → Tub G(x0 ) be the restriction of the G-action to G × Sx0 ,
that is, ϕ := μ |G×Sx0 . Similarly to the proof of Claim 3.36, note that dϕ(e,x) is
surjective and dϕ(g,x) (dLg X,Y ) = d(μ g )x ◦ dϕ(e,x) (X,Y ), hence dϕ(g,x) is surjective
for all g ∈ G and x ∈ Sx0 . In particular, Tub(G(x0 )) is an open neighborhood of
G(x0 ), which is clearly G-invariant.
Claim 3.59. ϕ (g, x) = ϕ (h, y) if and only if h = g k−1 and y = μ (k, x), where k ∈ H.
Assume that ϕ (g, x) = ϕ (h, y). Then μ (g, x) = μ (h, y) and y = μ (k, x), where
k = h−1 g. Since x, y ∈ Sx0 , it follows from Definition 3.47 that k ∈ H. The converse
statement is clear. We now define the candidate to G-equivariant diffeomorphism,
   
ψ : G ×H Sx0 → Tub G(x0 ) , ψ [g, s] := μ (g, s). (3.16)

The fact that π : G × Sx0 → G ×H Sx0 , π (g, s) = [g, s], is the projection of a principal
bundle, and Claim 3.59, imply that ψ is well-defined, smooth, bijective and G-
equivariant. It only remains to prove that ψ is a diffeomorphism.

G Sx0

G H Sx0 Tub G(x)

3.3 Isometric Actions 69

Note that dπ and dϕ are surjective, and ϕ = ψ ◦ π , hence dψ is surjective. As

dim G ×H Sx0 = dim G/H + dim Sx0 = dim M = dim Tub(G(x0 )), (3.17)

it follows that dψ is an isomorphism. Thus, by the Inverse Function Theorem, ψ is

a local diffeomorphism, and hence a diffeomorphism since it is bijective. 

Remark 3.60. By the above result, the image of a slice at x0 under the trans-
formation μ g : M → M is a slice at μ (g, x0 ), that is, Sμ (g,x0 ) = μ g (Sx0 ). It also
follows from the Tubular Neighborhood
  Theorem 3.57 that there is a unique
G-equivariant retraction r : Tub G(x0 ) → G(x0 ) such that r ◦ ψ = μ x0 ◦ π , where
π : G ×H Sx0 → G/H is the projection map of the associated bundle, μ x0 is defined
as in Proposition 3.41, and ψ is the G-equivariant diffeomorphism (3.16).
Remark 3.61. Let μ : G × M → M be an action whose orbits have constant
dimension. The Tubular Neighborhood Theorem 3.57 implies that the holonomy
group Hol(G(x0 ), x0 ) of the leaf G(x0 ) of the foliation {G(x)}x∈M coincides with the
image of the slice representation of H = Gx0 , see Definition 3.72 and Remarks 5.13
and 5.14. In the general case of foliations with compact leaves and finite holonomy,
there is an analogous result to the Tubular Neighborhood Theorem 3.57, known as
the Reeb Local Stability Theorem (see Moerdijk and Mrčun [163]).

3.3 Isometric Actions

In this section, we study the relation between proper and isometric actions.
In Proposition 3.62, we show that actions by closed subgroups of isometries are
proper, and conversely, in Theorem 3.65, we show that every proper action can be
made isometric with respect to a certain Riemannian metric (see also Remark 3.67).
Proposition 3.62. Let (M, g) be a Riemannian manifold and G a closed subgroup
of the isometry group Iso(M, g). The action μ : G × M  (g, x) → g(x) ∈ M is proper.
Proof. For the sake of brevity, we only prove the result for the case in which (M, g)
is complete. Consider sequences {gn } in G and {xn } in M, such that lim μ (gn , xn ) =
y and lim xn = x. We have to prove that there exists a convergent subsequence of
{gn } in G. Choose n0 such that dist y, μ (gn , xn ) < ε2 and dist(x, xn ) < ε2 for n > n0 .
 
Claim 3.63. For fixed R1 > 0, we have μ gn , BR1 (x) ⊂ BR1 +ε (y) for all n > n0 .
Indeed, for z ∈ BR1 (x), we have
     
dist μ (gn , z), y ≤ dist y, μ (gn , xn ) + dist μ (gn , xn ), μ (gn , z)
< ε2 + dist(xn , z)
≤ ε2 + dist(z, x) + dist(x, xn )
< ε + R1 .
70 3 Proper and Isometric Actions

The fact that each gn is an isometry of (M, g) implies that {gn } is an equicon-
tinuous family. Thus, by the Arzelà-Ascoli Theorem,11 Claim 3.63, and the fact
that closed balls in M are compact (since M was assumed complete), there exists a
(1)
subsequence {g n } of {gn } that converges uniformly on BR1 (x) to a continuous map
(i)
from BR1 (x) to M. Using this argument inductively, define a subsequence {g n } of
(i−1)
{g n } that converges uniformly on BRi (x) to a continuous map from BRi (x) to M,
(i)
where Ri > Ri−1 . Note that the diagonal subsequence {g i } converges uniformly
on each BRi (x) to a continuous map g : M → M. It follows from the Myers-Steenrod
Theorem 2.12 that g is an isometry of (M, g) that belongs to G. 

Remark 3.64. The hypothesis that G is a closed subgroup of Iso(M, g) is crucial for
Proposition 3.62 to hold. For instance, the R-action μ2 on the Euclidean space R2
defined in Example 3.43 is clearly isometric, and hence so is the induced R-action
2 on the flat torus T 2 . However, μ
μ 2 is not a proper action, see Exercise 3.18.
Using the Slice Theorem 3.49, we now prove a converse result, following Palais
and Terng [182, Chapter 5].
Theorem 3.65. Let μ : G × M → M be a proper action. There exists a G-invariant
metric g on M such that μ G = {μ g : g ∈ G} is a closed subgroup of Iso(M, g).
Proof. The strategy to construct g is to define it locally, using an averaging
procedure similar to (3.12), and then glue these together using G-invariant partitions
of unity. Note that M/G is paracompact,12 since it is a Hausdorff locally compact
space and the union of countably many compact spaces. Thus, there exists a locally
finite open covering {π (Sxα )} of M/G, where Sxα is a slice at a point xα ∈ M. Set
Uα := π −1 (π (Sxα )) and let { fα } be a G-invariant partition of unity subordinate to
{Uα }, which exists by the following result of Palais [180].
Claim 3.66. If {Uα } is a locally finite open cover of M by G-invariant open sets,
then there exists a smooth partition of unity { fα } subordinate to {Uα }, such that
each fα : Uα → [0, 1] is G-invariant.
Define a Riemannian metric ·, ·α on M along the slice Sxα , i.e., a section of
(TM ∗ ⊗ TM ∗ )|Sxα , by setting

X,Y αp := dμ g X, d μ gY μ (g,p) ω ,
Hα

11 This theorem gives a criterion for convergence of continuous maps in the compact-open topology.

More precisely, a sequence of continuous functions {gn : K → B} between compact metric spaces
K and B admits a uniformly convergent subsequence if {gn } is equicontinuous, i.e., for every ε > 0
there exists δ > 0 such that dist(gn (x), gn (y)) < ε for all dist(x, y) < δ , x, y ∈ K, n ∈ N.
12 This means that any open cover of M/G admits a locally finite refinement.
3.3 Isometric Actions 71

where ω is a right-invariant volume form on the isotropy group Hα = Gxα and ·, ·
is an arbitrary Riemannian metric on Sxα . It is easy to see that ·, ·α is Hα -invariant.
Define a G-invariant Riemannian metric gα on Uα by setting
 
gα d μ g X, dμ gY μ (g,p) := X,Y αp . (3.18)

The above metric is well-defined, since ·, ·α is Hα -invariant on Sxα . Finally, define
the desired metric g on M as g := ∑α f α gα .
It remains only to verify that μ G is a closed subgroup of Iso(M, g). Assume that
μ gn converges uniformly in each compact of M to an isometry f : M → M. Thus, for
each x ∈ M we have that lim μ (gn , x) = f (x). Properness of the action implies that a
subsequence {gni } converges to g ∈ G. It is not difficult to check that if lim gni = g,
then μ gni converges uniformly in each compact of M to μ g , and hence μ g = f .

Remark 3.67. Theorem 3.65 clearly provides a type of converse statement to
Proposition 3.62. Nevertheless, there are two subtle issues that might prevent one
from exchanging isometric actions by proper actions when proving results about
these. First, and most important, the G-invariant metric g provided by Theorem 3.65
is not necessarily complete.13 We stress that completeness is a necessary hypothesis
in many results for isometric actions. Second, the closed subgroup μ G of Iso(M, g)
is only isomorphic to G if μ : G × M → M is effective, see Remark 3.3.
Remark 3.68. Theorem 3.65 can be generalized to the context of proper groupoids,
see Pflaum, Posthuma, and Tang [185]. These objects admit a Riemannian metric
such that the induced foliation is a singular Riemannian foliation. Similarly to
Remark 3.67, the ambient space with this invariant metric may fail to be complete,
and many results on Riemannian foliations that assume completeness do not
hold. For example, orbits of proper groupoids with trivial holonomy may not be
diffeomorphic to each other. Nevertheless, when M is compact, foliations induced
by a proper groupoid seem to be a particular case of orbit-like foliations on
compact manifolds, i.e., singular Riemannian foliations whose restriction to slices
are homogenous foliations, see [16] and Chap. 5.
Exercise 3.69 (). Let μ : G × M → M be an effective isometric proper action.
Prove that if dim M = n, then dim G ≤ n(n+1)
2 .
Hint: Use that dim G − dim Gx = dim G(x) ≤ dim M = n.

Given an isometric G-action on a Riemannian manifold (M, g), we usually refer

to the directions tangent to the orbits as vertical directions and directions normal to
the orbits as horizontal directions. In this way, we define the vertical space at x ∈ M
as

Vx := Tx G(x) = {X ∗ (x) : X ∈ g}, (3.19)

13 Note that completeness can be guaranteed if, e.g., M is compact.

72 3 Proper and Isometric Actions

see (3.11), and the horizontal space as

Hx := {v ∈ Tx M : g(v, X ∗ (x)) = 0 for all X ∈ g}, (3.20)

cf. (2.10) and (2.11). In this way, we get an orthogonal splitting Tx M = Vx ⊕ Hx .

Note that the dimensions of Vx and Hx may vary with x ∈ M, so these are not
distributions on M. However, they become distributions when restricted to each
stratum of M, see Theorem 3.102. Accordingly, we say that a curve γ : [a, b] → M
is vertical or horizontal if its tangent vector γ  is always vertical or horizontal,
respectively.
The following result for isometric actions was observed by Kleiner in his PhD
thesis [141], and is usually referred to as Kleiner’s lemma.
Lemma 3.70. Let μ : G × M → M be an isometric action and let γ : [0, 1] → M be a
geodesic segment between G(γ (0)) and G(γ (1)) that realizes the distance14 between
these orbits. There exists a subgroup H of G such that Gγ (t) = H for t ∈ (0, 1) and
H is a subgroup of Gγ (0) and Gγ (1) .

Proof. Let H := g ∈ G : μ (g, γ (t)) = γ (t) for all t ∈ [0, 1] . Suppose there exists
some t0 ∈ (0, 1) and g ∈ Gγ (t0 ) such that g ∈ / H. In particular, d(μ g )γ (t0 ) (γ̇ (t0 )) =
γ̇ (t0 ). Define a piecewise smooth path γ: [0, 1] → M by setting γ|[0,t0 ] := γ[0,t0 ] and
 
γ|[t0 ,1] := μ g γ[t0 ,1] . Then γ joins G(γ (0)) to G(γ (1)) and has the same length as γ ,
contradicting the fact that minimizing geodesic segments are smooth. 

Exercise 3.71. Use the isometric action in R3 described in Example 3.4 (iii) to
show that the conclusion of Lemma 3.70 fails if γ is only a minimal geodesic
segment.
We conclude this section with the important definition of slice representation,
which is the restriction of the isotropy representation of an isometric action to the
normal space to the orbit, see Exercise 3.7. More
 precisely,
 if μ : G × M → M is a
proper isometric action, note that Sx = expx Bε (0) is a slice at x, called normal
slice, where Bε (0) ⊂ νx G(x) is an open ball in the normal space to G(x), cf. (3.13).
Definition 3.72. Let μ : G × M → M be an isometric proper action and Sx a normal
slice at x. The slice representation of Gx is the linear (orthogonal) representation
   
Gx  g −→ d(μ g )x ∈ O νx G(x) ⊂ GL νx G(x) .

14 Notethat the assumption that γ realizes the distance between the orbits G(γ (0)) and G(γ (1)) is
stronger than γ being a minimal geodesic segment between its endpoints γ (0) and γ (1).
3.4 Principal Orbits 73

3.4 Principal Orbits

In this section, we study some geometric properties of principal orbits. In particular,

we prove the Principal Orbit Theorem 3.82, that implies that the set of points in
principal orbits is open and dense.
Definition 3.73. Let μ : G × M → M be a proper action. Then G(x) is a principal
orbit if there exists a neighborhood V of x in M such that for each y ∈ V , Gx ⊂ Gμ (g,y)
for some g ∈ G.
In rough terms, principal orbits are the ones that have the smallest isotropy group
(i.e., the largest type, see Definition 3.84) among the nearby orbits.
Proposition 3.74. Let μ : G × M → M be a proper left action. Then the following
are equivalent:
(i) G(x) is a principal orbit;
(ii) If Sx is a slice at x, then Gx = Gy for all y ∈ Sx .
Proof. First, let us prove that (i) implies (ii). From the definition of slice, for each
y ∈ Sx we have Gy ⊂ Gx . On the other hand, since G(x) is a principal orbit, there
exists g such that g Gx g−1 ⊂ Gy . Thus, g Gx g−1 ⊂ Gy ⊂ Gx . Since Gx is compact,
we conclude that the above inclusions are equalities. In particular, Gy = Gx .
To prove the converse, we have to prove that for z in a tubular neighborhood of
G(x), there exists g such that Gx ⊂ g Gz g−1 . The orbit G(z) intersects the slice Sx at
least at one point, say y. Since y and z are in the same orbit, there exists g such that
Gy = g Gz g−1 , recall Exercise 3.9. Therefore Gx = Gy = g Gz g−1 . 

Remark 3.75. As we will see in Chap. 5, the above proposition implies that each
principal orbit is a leaf with trivial holonomy of the foliation F = {G(x)}x∈M .

Exercise 3.76. Let μ : G × M → M be a proper action, and let K := x∈M Gx be
the ineffective kernel of the action. Consider the induced effective action μ  : G/K ×

M → M, see Remark 3.3. Prove that μ is smooth and proper. Check that if an orbit
 , then it is also principal with respect to μ .
is principal with respect to μ
Exercise 3.77. Let μ : G × M → M be an isometric proper action and Sx a normal
slice at x. Prove that G(x) is a principal orbit if and only if the slice representation
of Gx is trivial.
Let us explore some geometric properties of orbits of isometric actions.
Proposition 3.78. Let μ : G × M → M be an isometric proper action on a complete
Riemannian manifold M. Then the following hold:
(i) A geodesic γ orthogonal to an orbit G(γ (0)) remains orthogonal to all orbits it
intersects, i.e., γ is a horizontal geodesic.
74 3 Proper and Isometric Actions

In the following, suppose G(x) is a principal orbit:

(ii) Given ξ ∈ νx G(x), ξ"μ (g,x) := d(μ g )x ξ is a well-defined normal vector field
along G(x), called equivariant normal field;
−1
(iii) Sξ" = dμ g Sξ" d μ g , where Sξ" is the shape operator of G(x);
μ (g,x) x

(iv) Principal curvatures of G(x) with respect to an equivariant normal field ξ" are
constant along G(x);
  
(v) expy ξ"y : y ∈ G(x) is an orbit of μ .
Proof. In order to prove item (i), by Proposition 3.13 and (3.11), it suffices to prove
that if a Killing vector field X is orthogonal to γ  (0), then X is orthogonal to γ  (t) for
all t. Since ∇γ  (t) X, γ  (t) = 0 (see Proposition 2.14) and γ is a geodesic, we have
that dtd X, γ  (t) = 0. Thus, X is always orthogonal to γ  (t).
Item (ii) follows from Exercise 3.77. Item (iii) follows from:
# −1 $ # $
dμ g Sξ" d μ gW, Z x
= Sξ" dμ gW, dμ g Z μ (g,x)
μ (g,x) μ (g,x)
# $
= − ∇dμ gW d( μ g )x ξ , dμ g Z μ (g,x)
# $
= − ∇W ξ", Z x
# $
= Sξ" W, Z x .
x

−1
As for (iv), note that if Sξ" X = λ X , then dμ g Sξ" d μ g X = λ X. Hence
μ (g,x)
Sξ" dμ g X = λ d μ g X. Finally, item (v) follows from (2.2), because expμ (g,x)
μ (g,x)
" 
ξμ (g,x) = expμ (g,x) (dμ g ξx ) = μ g (expx (ξ )). 

Remark 3.79. The above proposition illustrates a few concepts and results of
Chaps. 4 and 5. Item (i) implies that the homogeneous foliation (3.4) by orbits of a
proper isometric action is a singular Riemannian foliation (see Definition 5.2). Item
(v) implies that one can reconstruct this foliation taking all parallel submanifolds
to a principal orbit. This is a consequence of equifocality, which is valid for every
singular Riemannian foliation. Item (iv) and the fact that equivariant normal fields
are parallel normal fields when the action is polar imply that principal orbits of a
polar action on Euclidean space are isoparametric (see Definitions 4.8 and 4.14).
Let us explore a little further item (i) of Proposition 3.78 to obtain some metric
information about the orbit space M/G of an isometric proper action. If a geodesic
segment γ : [0, 1] → M minimizes distance between the orbit G(γ (0)) and γ (1), then
it must be orthogonal to the submanifold G(γ (0)). Hence, by the above, γ must
be horizontal. Now, for any g ∈ G, since μ g : M → M is an isometry, the curve
(μ g ◦ γ ) : [0, 1] → M is a horizontal minimal geodesic segment joining G(γ (0)) and
μ (g, γ (1)) ∈ G(γ (1)), with the same length as γ . This means that for any g ∈ G,
the distances from G(γ (0)) to γ (1) and μ (g, γ (1)) are the same. Thus, orbits of
3.4 Principal Orbits 75

an isometric group action are equidistant, and hence there is a well-defined orbital
distance, i.e., a distance on M/G. The orbit space M/G equipped with this distance
becomes a metric space, more precisely, a length metric space.15
Remark 3.80. There are many important aspects of the geometry of the metric space
M/G that reflect the geometry of an isometric G-action on M, especially if M has a
lower curvature bound, making M/G an Alexandrov space.16 Much recent progress
on areas involving the interface of transformation groups and Riemannian geometry
was achieved through the use of Alexandrov geometry of orbit spaces; however,
we do not go in this direction in this book. For a brief glance at some of these
ideas, see Example 3.107, where the appropriate generalization of tangent space at
a nonregular point in M/G, called tangent cone, is mentioned in a specific example.
Details on how Alexandrov techniques apply to study isometric actions can be found
in the survey of Grove [107].
Exercise 3.81. Let μ : G × M → M be a free isometric action and recall Theo-
rem 3.34. Show that the quotient map π : M → M/G is Riemannian submersion,
where M/G is equipped with the natural orbit metric, whose associated distance is
the above orbital distance.17
We conclude this section proving the so-called Principal Orbit Theorem, which
guarantees that the subset of points in M on principal orbits is open and dense,
and that M/G has an open and dense subset which is a smooth manifold (see
Theorem 3.95 and Remark 3.106).
Principal Orbit Theorem 3.82. Let μ : G × M → M be a proper action, where M
is connected, and denote by Mprinc the set of points of M contained in principal
orbits. Then the following hold:
(i) Mprinc is open and dense in M;
(ii) The subset Mprinc /G of M/G is a connected manifold;
(iii) If G(x) and G(y) are principal orbits, there exists g ∈ G such that Gx = gGy g−1 .
Proof. We first prove the existence of a principal orbit. Since G has finite dimension
and the isotropy groups are compact, we can choose x ∈ M such that Gx has the
lowest dimension among isotropy groups and, for that dimension, the smallest
number of connected components. If Sx is a slice at x, by definition, Gy ⊂ Gx
for every y ∈ Sx . By construction, we conclude that Gy = Gx . Hence, from
Proposition 3.74, G(x) is a principal orbit.
In order to prove that Mprinc is open, let x ∈ Mprinc and let Sx be a slice at x.
Proposition 3.74 implies that Gy = Gx for every y ∈ Sx . We claim that each y ∈ Sx
belongs to a principal orbit. If a point z is close to y, then z is in a tubular

15 Ametric space is a length metric space if the distance between any two points is realized by a
shortest curve, called a geodesic.
16 An Alexandrov space is a finite-dimensional length metric space with a lower curvature bound,
in a comparison geometry sense. For details, see [59, 107].
17 Moregenerally, if μ is not free, then π : M → M/G is a submetry, which is a generalization of
submersions to metric spaces.
76 3 Proper and Isometric Actions

neighborhood of G(x). Note that G(z) intersects Sx in at least one point, so there
exists g such that μ (g, z) = w ∈ Sx . Thus, Gy = Gx = Gw = g Gz g−1 and hence each
y ∈ Sx belongs to a principal orbit, as claimed. Therefore, every point in the tubular
neighborhood Tub(G(x)) = μ (G, Sx ) belongs to a principal orbit, proving Mprinc is
open.
If μ is an isometric action on a complete Riemannian manifold, then the rest of
the proof follows easily from Proposition 3.74 and Kleiner’s Lemma 3.70, since
given any two orbits, there exists a geodesic segment that realizes the distance
between these orbits. In this case, one can also prove that the space Mprinc /G is
convex. We encourage the reader to verify these claims, and, in what follows, we
finish the proof in the general case.
To prove that Mprinc is dense, consider p ∈ / Mprinc and U a neighborhood of p.
Choose x ∈ U such that Gx has the lowest dimension among isotropy groups and,
for that dimension, the smallest number of connected components. Then, from the
argument above, we conclude that x ∈ Mprinc .
Proposition 3.74 and some arguments from the proof of Theorem 3.34 can
be used to prove that Mprinc /G is a manifold. In order to prove that Mprinc /G is
connected, assume that the action is proper and isometric. A set A ⊂ M/G does not
locally disconnect M/G if each p ∈ M/G has a neighborhood U such that U \ A is
path-connected. Using (i), it is easy to verify that if (M \ Mprinc )/G does not locally
disconnect M/G, then Mprinc /G is path-connected. Thus, it suffices to prove that
(M \ Mprinc )/G does not locally disconnect M/G. This can be done using the fact
that Sx /Gx = Tub(G(x))/G, the slice representation, Exercise 3.76 and the next:
Claim 3.83. Let K be a closed subgroup of O(n), acting on Rn by multiplication.
Then Rnprinc /K is path-connected.
We prove Claim 3.83 by induction. If n = 1, then K = Z2 or K = {1}. In both
cases, Rprinc /K is path-connected. For each sphere Sn−1 centered at the origin, we
can apply the induction hypothesis and the slice representation to conclude that
(Sn−1 \ Sprinc
n−1
)/K does not locally disconnect Sn−1 /K. Therefore, Sprinc
n−1
/K is path-
connected. Consider x, y ∈ Rprinc and let x be the projection of x on the sphere Sn−1
n

that contains y. The points along the straight line segment that joins x to x belong to
principal orbits. Thus, K(x) and K( x) are connected by a path in Rnprinc /K. As we
have already proved, K(y) and K( x) are connected by a path in Sn−1
princ /K and hence
in Rprinc /K. Therefore K(x) and K(y) are connected by a path in Rnprinc /K.
n

Finally, item (iii) is a direct consequence of item (ii). 

3.5 Orbit Types

The principal orbits discussed in the last section are one of many possible orbit
types. We now discuss some properties of orbit types of proper actions, e.g.,
we prove that on compact manifolds there is only a finite number of these, see
3.5 Orbit Types 77

Theorem 3.91. In addition, it is also proved that each connected component of a

set of orbits of the same type is a connected component of the total space of a
certain fiber bundle, see Theorem 3.95. Finally, it is also proved that the connected
components of sets of orbits of the same type give a stratification of M, see
Theorem 3.102.
Definition 3.84. Let μ : G × M → M be a proper action.
(i) The orbit G(x) has a larger orbit type than G(y) if there is a G-equivariant map
ϕ : G(x) → G(y), or, equivalently, if there is g ∈ G such that Gx ⊂ Gμ (g,y) ;
(ii) The orbits G(x) and G(y) have the same orbit type if there is a G-equivariant
diffeomorphism ϕ : G(x) → G(y), or, equivalently, if there is g ∈ G such that
Gx = Gμ (g,y) ;
(iii) An orbit G(x) is said to be regular if the dimension of G(x) coincides with the
dimension of principal orbits;
(iv) A nonprincipal regular orbit is called exceptional;
(v) A nonregular orbit is called singular.18
These definitions are also used referring to points, e.g., x and y have the same orbit
type if G(x) and G(y) have the same type, and x is singular if G(x) is singular.
It is not difficult to see that G(x) and G(y) have the same orbit type if and only
if G(x) has a larger orbit type than G(y) and G(y) has a larger orbit type than G(x),
see Exercise 3.9. It is also clear that having the same orbit type is an equivalence
relation (on both M/G and M). Note that Theorem 3.82 asserts not only that Mprinc
is open and dense in M, but also that there exists a unique type of principal orbit.

Exercise 3.85. Let μ : G × M → M be a proper action, and let K := x∈M Gx be the
 : G/K ×
ineffective kernel of this action. Consider the induced effective action μ
M → M, see Remark 3.3 and Exercise 3.76. Prove that orbits of the same type with
 are also of the same type with respect to μ .
respect to μ
Exercise 3.86. Let μ : G × M → M be a proper action and G(p) a principal orbit.
Prove that G(x) is an exceptional orbit if and only if dim G(x) = dim G(p) and the
number of connected components of Gx is greater than the number of connected
components of G p .
 : S1 × S2 → S2 of the circle S1 on
Exercise 3.87. Consider the isometric action μ
the round sphere S by rotations. Use Exercise 3.40 to verify that μ
2  induces an
isometric action μ : S1 × RP2 → RP2 , which has an exceptional orbit.
Exercise 3.88 (). Consider the action of SU(3) on itself by conjugation. Prove
that the orbits of this action are diffeomorphic to one of the following:

18 It
is important to notice that singular orbits are smooth submanifolds, whose singular nature is
simply to have a lower dimension.
78 3 Proper and Isometric Actions

(i) {λ I}, where I ∈ SU(3) is the identity and λ ∈ C satisfies λ 3 = 1;

(ii) SU(3)/S(U(2) × U(1));
(iii) SU(3)/T , where T is the subgroup of diagonal matrices in SU(3).
Conclude, using Exercise 3.44 and Remark 3.46, that an orbit is diffeomorphic
to either a point, or complex projective plane CP2 , or the complex flag manifold
SU(3)/T .
Hint: Use the fact that each matrix of SU(3) is conjugate to a matrix of T , by diagonalization.

Remark 3.89. As explained in the next chapter, Exercises 3.87 and 3.88 give
examples of polar actions. An isometric action is a polar action if for each regular
point x, the set expx (νx G(x)) is a (totally geodesic) submanifold that intersects
every orbit orthogonally, see Definition 4.8 for details. In Exercise 3.87, the polar
action admits an exceptional orbit, and in Exercise 3.88, the polar action admits only
principal and singular orbits. Note that RP2 is not simply-connected and SU(3) is
simply-connected. As explained in Chap. 5, polar actions do not admit exceptional
orbits if the ambient space is simply-connected (Corollary 5.35).
Remark 3.90. For an example of nonpolar isometric action, see Example 3.11.
Theorem 3.91. Let μ : G ×M → M be a proper action. For each x ∈ M, there exists
a slice Sx such that the tubular neighborhood Tub(G(x)) = μ (G, Sx ) contains only
finitely many different orbit types. In particular, if M is compact, there is only a finite
number of different orbit types in M.
Proof. Using Theorem 3.65, let g be a Riemannian metric on M such that μ G ⊂
Iso(M, g). It is not difficult to verify the following:
Claim 3.92. If y, z ∈ Sx have the same Gx -orbit type, then y and z have the same
G-orbit type.
Therefore, it suffices to prove the result for the Gx -action on Sx . By the slice
representation (see Definition 3.72) and Exercise 3.85, we can further reduce the
problem to proving the result for a linear (orthogonal) action of K ⊂ O(n) on Rn by
multiplication. With this goal, we proceed by induction on n. If n = 1, then K = Z2
or K = {1}. In both cases, there exists only a finite number of different orbit types
(at most 2). For a sphere Sn−1 , we can apply the induction hypothesis, Claim 3.92
and the slice representation, to conclude that there exists a finite number of K-orbit
types. For each p ∈ Sn−1 , the points along the straight line segment joining the
origin to p (except for the origin itself) have the same isotropy type (cf. Kleiner’s
Lemma 3.70). Thus, there exists a finite number of K-orbit types in Rn . 

In order to study how different orbit types of a proper action stratify a manifold,
we first need to develop structural results for sets of points that have the same orbit
type. The first step in this direction is the following result on fixed point sets.
3.5 Orbit Types 79

Proposition 3.93. Let μ : G × M → M be a proper action and H ⊂ G a subgroup.

The connected components of the fixed point set

M H := x ∈ M : μ (h, x) = x, for all h ∈ H

are embedded submanifolds of M (of possibly different dimensions), whose tangent

space is given by

Tx M H = v ∈ Tx M : d(μ h ) p v = v, for all h ∈ H .

In addition, if the action μ is isometric, then these submanifolds are totally geodesic.
Proof. The fixed point set of H coincides with the fixed point set of its closure H,
and hence we may assume H is a closed subgroup of G. From Theorem 3.65, there
exists an H-invariant Riemannian metric g on M. Given x ∈ M H , let ε > 0 be such
that expx : Tx M → M is a diffeomorphism from the open ball Bε (0) ⊂ Tx M onto its
image. Notice that, from (2.2), this diffeomorphism is H-equivariant, where the H-
action on Bε (0) is the isotropy representation defined in Exercise 3.7. In particular,
the exponential map provides a submanifold chart for M H near x, which is mapped
to the linear subspace Tx M H of Tx M. This proves that each connected component of
M H is an embedded submanifold. Finally, it is easy to deduce from (2.2) that each
of these submanifolds is totally geodesic if the G-action is isometric. 

Besides the notion of orbits (or points) that have the same orbit type, it is conve-
nient to define the following refinement, containing more geometric information.19
Definition 3.94. Two orbits G(x) and G(y) of a proper action  have
 the same
 local

orbit type if there is a G-equivariant diffeomorphism ϕ : Tub G(x) → Tub G(y) .
Thus, we have 2 equivalence relations: to have the same orbit type and to have
the same local orbit type, denoted respectively by ∼ and ≈, following [79]. Clearly,
having the same local orbit type is a stronger condition than having the same orbit
type (cf. Definition 3.84), so equivalence classes with respect to ∼ are partitioned
into equivalence classes with respect to ≈. These equivalence classes are denoted

Mx∼ := {y ∈ M : x ∼ y} and Mx≈ := {y ∈ M : x ≈ y},

and respectively called orbit type and local orbit type of x ∈ M. In this way,
an orbit type is the union of local orbit types. Using the Tubular Neighborhood
Theorem 3.57, one can prove that x ≈ y if and only if x ∼ y and the isotropy
representations of Gx on Tx M and Gy on Ty M descend to equivalent representations

19 To our knowledge, the notion of local orbit types was introduced in Duistermaat and Kolk [79].
80 3 Proper and Isometric Actions

on Tx M/Tx G(x) and on Ty M/Ty G(y), that is, there exists an equivariant linear
isomorphism between these spaces,20 recall Exercise 3.7 and Proposition 3.41.
In what follows, we show that Mx≈ is a (possibly disconnected) embedded
submanifold of M. As proved in [79, Theorem 2.6.7], Mx≈ is an open and closed
subset of Mx∼ . In particular, an orbit type Mx∼ is the disjoint union of submanifolds
of possibly different dimensions, and connected components of the same dimension
are arranged into the local orbit types My≈ , for y ∈ Mx∼ .
Orbit types Mx∼ are commonly denoted M(H) in the literature, in reference to the
set of points with isotropy group conjugate to H = Gx . We prefer to use the notation
Mx∼ (in association with Mx≈ ), in the same way as [79], to stress the dependence
on the point x ∈ M rather than on the conjugacy class of its isotropy group H.
Indeed, conjugates of H may have different isotropy representations at different
points. Furthermore, the notations Mx∼ and Mx≈ are also suggestive of leaves of a
foliation.
Theorem 3.95. Let μ : G × M → M be a proper action. Each local orbit type Mx≈ is
a union of G-invariant embedded submanifolds (of the same dimension). Moreover,
the restriction of the quotient map π : M → M/G to Mx≈ determines a fiber bundle

G/H → Mx≈ → Mx≈ /G,

whose structure group is N(H)/H, where H = Gx is the isotropy at x and N(H) is

the normalizer of H in G. In particular, subsets Mx≈ /G of orbits that have the same
local orbit type are smooth manifolds inside M/G.
Proof (Sketch). The proof of this result is analogous to the proof of the Tubular
Neighborhood Theorem 3.57, and details can be found in [79, pp. 109–111]. The
main idea is to prove that Mx≈ is the total space of the associated bundle with
fiber G/H,

G/H → G/H ×K P → P/K, (3.21)

to the principal K-bundle K → P → P/K, where P := Mx≈ ∩ M H and K := N(H)/H.

Using the properties in Definition 3.47 and the Tubular Neighborhood
Theorem 3.57, one can prove the following, where Tub(G(x)) = μ (G, Sx ).
Claim 3.96. Mx∼ ∩ Tub(G(x)) is a submanifold of M. Moreover, there is a
G-equivariant diffeomorphism between Mx∼ ∩ Tub(G(x)) and G/H × (Sx )H .
Note that if y ∈ Mx∼ ∩ M H , then Gy = g H g−1 for some g ∈ G and H ⊂ Gy , and
hence H = Gy and g ∈ N(H). Conversely, if Gy = H, then y ∈ Mx∼ ∩ M H . Thus, we
proved the first item in the following claim, whose other items are proved similarly.

20 If
the action is isometric, this simply means that the slice representations of Gx and Gy are
equivalent, i.e., related by an equivariant linear map φ : νx G(x) → νy G(y).
3.5 Orbit Types 81

Claim 3.97. The following hold:

(i) y ∈ Mx∼ ∩ M H if and only if Gy = H.
(ii) Let y ∈ Mx∼ ∩ M H . Then μ (g, y) ∈ Mx∼ ∩ M H if and only if g ∈ N(H);
(iii) P = Mx≈ ∩ M H is invariant under the subaction of N(H), and the ineffective
kernel of this action is H;
(iv) Each y ∈ P has a neighborhood diffeomorphic to (N(H)/H) × (Sx )H .
Altogether, we have that P is a manifold with a free proper (left) K-action, so
that K → P → P/K is a principal K-bundle, and P/K = Mx≈ /G. Moreover, it is not
difficult to see that there is also a free (right) K-action on G/H (see Exercise 6.9),
which allows us to construct the associated bundle (3.21) with fiber G/H.21
In order to prove that there exists a G-equivariant diffeomorphism between Mx≈
and the total space of this associated bundle, note that

ϕ : G/H × P → Mx≈ , ϕ (gH, s) := μ (g, s),

is a well-defined map, by Claim 3.97. Given y ∈ Mx≈ , we have Gμ (g−1 ,y) = H for
some g ∈ G. From Claim 3.97, μ (g−1 , y) ∈Mx∼ ∩ M H . Since Mx≈ is G-invariant,
μ (g−1 , y) ∈ P and hence y = ϕ g, μ (g−1 , y) . Thus, ϕ is surjective. Furthermore,
the restriction ϕ : G/H × (Sx )H → Mx∼ ∩ Tub(G(x)) is a diffeomorphism; and, since
 
T(gGx ,y) G/H × (Sx )H ⊂ T(gH,y) (G/H × P), (3.22)

we conclude that dϕ is surjective. We now observe the following (cf. Claim 3.59):
Claim 3.98. ϕ (g H, y) = ϕ (h H, z) if and only if h = g n−1 and z = μ (n, y) for some
n ∈ N(H).
Assume that ϕ (g H, y) = ϕ (h H, z). Then μ (g, y) = μ (h, z) and z = μ (n, y), where
n = h−1 g. Since z, y ∈ P, it follows from Claim 3.97 that n ∈ N(H). The converse
statement is clear. We now define the candidate to G-equivariant diffeomorphism,
 
ψ : G/H ×K P → Mx≈ , ψ [gH, z] := μ (g, z),

cf. (3.16). The fact that π : G/H ×P → G/H ×K P, π (gH, z) = [g, z], is the projection
of a principal bundle, and Claim 3.98, imply that ψ is well-defined, smooth, bijective
and G-equivariant. In only remains to prove that ψ is a diffeomorphism.

G/H P

G/H K P Mx

21 Here,it is convenient to use a free left action on the principal bundle instead of a right action, cf.
Proposition 3.33. Notice that, accordingly, the notation (3.21) for the twisted space is also reversed,
cf. Theorem 3.51.
82 3 Proper and Isometric Actions

Note that dπ and dϕ are surjective, and ϕ = ψ ◦ π , hence dψ is surjective. By a

dimension argument analogous to (3.17), it follows that dψ is an isomorphism.
Thus, by the Inverse Function Theorem, ψ is a local diffeomorphism, and hence
a diffeomorphism, since it is bijective. 

Remark 3.99. By Theorem 3.95, the result in Exercise 3.81 can be extended to the
case of isometric actions that are not necessarily free. Namely, the projection map
π : Mx≈ → Mx≈ /G is a Riemannian submersion, where Mx≈ /G is equipped with the
natural orbit metric. Note that Mx≈ /G is locally totally geodesic in M/G, since by
Claim 3.96, it is locally isometric to the fixed point set (Sx )H , where Sx is a slice at
x ∈ M, which is totally geodesic by Proposition 3.93.
We conclude this section with a discussion of the orbit type stratification.
Definition 3.100. A stratification of a manifold M is a partition of M by embedded
submanifolds {Mi }i∈I of M, called strata, such that:
(i) The partition is locally finite, i.e., each compact subset of M only intersects a
finite number of strata;
(ii) For each i ∈ I, there exists a subset Ii ⊂ I \ {i} such that the closure of Mi is

Mi = Mi ∪ j∈Ii M j ;
(iii) dim M j < dim Mi , for all j ∈ Ii .
Example 3.101. Consider the action of SO(2) on R3 described in Example 3.4 (ii).
Then {Mi }i∈I , given by

M1 := (0, 0, x3 ) ∈ R3 : x3 ∈ R , M2 := R3 \ M1 , (3.23)

is a stratification of R3 , where I = {1, 2}, I1 = 0/ and I2 = {1}. The stratum M1

consists of fixed points (which are singular orbits), and the stratum M2 consists of
regular points (which lie in principal orbits). Note that M2 = Mprinc is open an dense
and Mprinc /G is connected, as asserted by the Principal Orbit Theorem 3.82.
Theorem 3.102. Let μ : G × M → M be a proper action. The partition of M into
connected components of the orbit types Mx∼ , x ∈ M, determines a stratification.
Proof. The partition of M into connected components of orbit types is locally finite
by the same arguments from the proof of Theorem 3.91. Given y ∈ M, we denote by
(My∼ )0 the connected component of My∼ that contains y.

/ (My∼ )0 . Then:
Claim 3.103. Let x ∈ (My∼ )0 be such that x ∈

(i) (Mx∼ )0 is contained in (My∼ )0 ;

(ii) dim(Mx∼ )0 < dim(My∼ )0 .
In order to prove (i), we may assume that y ∈ Sx . From Claim 3.96, it suffices
to prove that if x ∈ (Sx )H , where H = Gx , then x ∈ (My∼ )0 . Using the slice
representation, we may also assume that (Sx )H is a linear subspace fixed by the
linear action of K = μ H on a vector space, and x = 0. Since the K-action fixes each
3.5 Orbit Types 83

point of (Sx )H , it leaves invariant the normal space ν p (Sx )H that contains y. Consider
the projection y = y− p + x of y onto the normal space νx(Sx )H . As Kx = K = Kp , we
have that Ky = Ky. Moreover, as the K-action is linear, Kz = Ky for each z different
from x along the straight line segment joining x to y. Thus, x ∈ (My∼ )0 , concluding
the proof of (i).
As for (ii), let (Sx )∼y be the set of points in Sx that have the same H-orbit type
of y ∈ Sx . From the above discussion, we infer that dim((Sx )∼ y ) ≥ dim(Sx ) +
0 H

dim μ H (y) + 1. In particular, dim((Sx )∼ y ) > dim(Sx ) . Thus, using Claim 3.96 and
0 H
∼ ∼
the fact that each point in ((Sx )y ) is also in (My ) , we have that
0 0

dim(Mx∼ )0 = dim G(x) + dim(Sx )H < dim G(x) + dim((Sx )∼ ∼ 0

y ) ≤ dim(My ) ,
0

which concludes the proof of (ii). 

Remark 3.104. Furthermore, it is possible to prove that an orbit type stratification

is a Whitney stratification, see Duistermaat and Kolk [79].
Remark 3.105. Each stratum of the orbit type stratification of an isometric action is
a minimal submanifold. In fact, the mean curvature vector of a stratum (see (2.9))
is tangent to the stratum by equivariance, but is also normal to the stratum by
definition, and hence vanishes.
Remark 3.106. The orbit type stratification {Mi }i∈I of a manifold M induces a
stratification {Mi /G}i∈I of the corresponding orbit space M/G, i.e., a partition into
manifolds satisfying (i)–(iii) in Definition 3.100. From Remark 3.99, the orbit strata
are locally totally geodesic. This plays an important role when M/G is studied from
the viewpoint of Alexandrov geometry, see Remark 3.80.
Example 3.107. Consider the conjugation action of SU(3) on itself. From
Exercise 3.88, there are 3 orbit types, since an orbit is either a fixed point, the
complex projective plane CP2 , or the complex flag manifold SU(3)/T . Clearly,
the latter are the principal orbits, and it is not difficult to show that each of the 2
nonprincipal orbit types have 3 connected components (of the same dimension).
This describes the orbit type stratification of SU(3), see Theorem 3.102. The
orbit space of this conjugation action is a flat equilateral triangle, whose induced
stratification is the obvious partition of a closed triangle into the union of its
vertices (corresponding to fixed points in SU(3)), edges (corresponding to the
orbits diffeomorphic to CP2 ), and interior (corresponding to principal orbits).
The linearization of the conjugation action of SU(3) at the identity I ∈ SU(3)
is its adjoint representation, see (1.8). Notice that this is precisely the isotropy
representation (and the slice representation) at I ∈ SU(3), since I is a fixed point
of conjugation. The orbit structure of this adjoint action is studied in Example 4.42,
where it is proved that its orbit space is a (closed) wedge of angle π /3 in the plane.
We remark that this is precisely the so-called tangent cone to the orbit space of the
conjugation action (the equilateral triangle) at the identity orbit (one of the vertices).
84 3 Proper and Isometric Actions

This is an instance of a general relation between tangent cones at points of the

orbit space of an isometric action and the orbit space of the corresponding slice
representation.
Remark 3.108. As demonstrated many times along this chapter (e.g., Theo-
rems 3.82, 3.91 and 3.102), the local study of proper (or isometric) actions can
be reduced to the local study of an isometric action in Euclidean space, by using the
slice representation. Molino [165] pointed out that the same idea can be used in the
local study of singular Riemannian foliations (see Definition 5.2). More precisely,
after a suitable change of metrics, the local study of any singular Riemannian
foliation is reduced to the study of a singular Riemannian foliation on the Euclidean
space (with the standard metric). This idea is used, for example, in [8, 20], see also
the survey [12].
Chapter 4
Adjoint and Conjugation Actions

Two actions play a central role in the theory of compact Lie groups G: the G-action
on itself by conjugation and the adjoint action of G on its Lie algebra g, which is
the linearization of the latter (see Definition 1.36 and Exercise 3.7). The goal of this
chapter is to describe classical results on such actions from a geometric viewpoint.
The adjoint action is an example of polar action (see Remark 3.89 and
Definition 4.8) and each of its regular orbits is an isoparametric submanifold (see
Definition 4.14). Remarkably, several results on adjoint actions can be extended,
not only to the theory of isoparametric submanifolds and polar actions, but also
in a general context of polar foliations (see Definition 5.2). This parallel is made
in Chap. 5.
The following references were a source of inspiration for this chapter, and could
serve as further reading material: Duistermaat and Kolk [79], Fegan [85], Fulton
and Harris [90], Hall [121], Helgason [126], Hsiang [129], Hunt [132], Onishchik
[179], Palais and Terng [182], Serre [195] and Thorbergsson [208, 209].

4.1 Maximal Tori and Polar Actions

The goal of this section is twofold. First, we prove the Maximal Torus Theorem 4.1,
that generalizes basic diagonalization results from linear algebra. Second, we
present the definitions of polar actions and isoparametric submanifolds, briefly
discussing some results from these theories, and relating them to the adjoint and
conjugation actions of compact Lie groups.
Recall that a Lie group T is a torus if it is isomorphic to the product S1 ×
· · · × S1 . Alternatively, these are the compact connected abelian Lie groups, see
Theorem 1.41. We say that a vector X ∈ t is an infinitesimal generator if the
corresponding 1-parameter subgroup {exp(tX) : t ∈ R} is dense in T . It is not
difficult to prove that every torus admits infinitesimal generators, see Example 3.43.

© Springer International Publishing Switzerland 2015 85

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_4
86 4 Adjoint and Conjugation Actions

Let T be the subgroup of diagonal matrices of SU(n). It follows from

Theorem 1.41 that T is a torus, and we know from linear algebra that every matrix
g ∈ SU(n) is conjugate to an element of T . This is a particular case of the next
classical result, see Example 4.7.
Maximal Torus Theorem 4.1. Let G be a connected, compact Lie group.
(i) There exists a maximal1 torus T in G;
(ii) Any two maximal tori in G are conjugate;
(iii) Every element of G is contained in a maximal torus;
(iv) For each bi-invariant metric Q on G, the orbits of the conjugation action
intersect every maximal torus Q-orthogonally.
Proof. Our proof is based on Hunt [132] and Thorbergsson [207].
To prove (i), let t ⊂ g be a maximal abelian subalgebra of the Lie algebra of
G. By Theorem 1.21, there exists a unique connected subgroup T ⊂ G with Lie
algebra t. Note that its closure T is an abelian, connected, compact Lie group, see
Proposition 1.39. Since t is maximal, T = T . From Theorem 1.41, T is a torus. If
T  ⊂ G is a torus with T ⊂ T  , then [X , Z] = 0 for all X ∈ t and Z ∈ t. By maximality
of t, we have X ∈ t and hence t = t. Thus, from Theorem 1.21, T  = T .
Claim 4.2. If X ∈ t is an infinitesimal generator of T , t = {Y ∈ g : [X,Y ] = 0}.
Note that t ⊂ {Y ∈ g : [X ,Y ] = 0} by Proposition 1.39. Conversely, let Y ∈ g be
such that [X,Y ] = 0. From commutativity of X and Y , by Remark 1.40, we have

exp(sX) exp(tY) = exp(tY) exp(sX), for all s,t ∈ R.

Thus, exp(tY) commutes with all elements of T = {exp(sX) : s ∈ R} and hence Y

commutes with all vectors of t. From Theorem 1.21, there is a unique connected
Lie subgroup T  of G with Lie algebra t = t ⊕ span{Y }. Note that T  is an abelian,
connected, compact Lie group, and hence a torus, by Theorem 1.41. Since T is
a maximal torus contained in T  , we have that T = T  and hence Y ∈ t, proving
Claim 4.2.
In order to prove (ii), let X1 and X2 be infinitesimal generators of maximal tori T1
and T2 respectively, and set

f : G → R, f (g) := Q(Ad(g)X1 , X2 ),

where Q is a bi-invariant metric on G (see Proposition 2.24). Since G is compact, f

attains a minimum, say at the point g0 ∈ G. Therefore, for all Y ∈ g,

d  
0= f exp(tY)g0 
dt t=0
d  
= Q Ad(exp(tY)) Ad(g0 ) X1 , X2 
dt t=0

1 i.e., such that if T  ⊂ G is a torus and T ⊂ T  , then T = T  .

4.1 Maximal Tori and Polar Actions 87

=Q([Y, Ad(g0 )X1 ], X2 )

=Q(Y, [Ad(g0 )X1 , X2 ]),

and hence [Ad(g0 )X1 , X2 ] = 0. From Claim 4.2, Ad(g0 )X1 ∈ t2 , which means that
g0 exp(tX 1 )g−1 −1
0 ∈ T2 . Thus, g0 T1 g0 ⊂ T2 and, by maximality of T1 , we conclude
that equality holds, proving (ii).
As for (iii), let g ∈ G. Since G is compact, by Theorem 2.27, there exists Y ∈ g
such that exp(Y ) = g. Thus, g is contained in the maximal torus that contains the
corresponding 1-parameter subgroup {exp(tY) : t ∈ R}. Item (iii) now follows from
item (ii). In order to prove (iv), let p ∈ T be a point in a maximal torus, and consider
its orbit G(p) = {g p g−1 : g ∈ G} by the conjugation action. It is easy to see that
 
Tp G(p) = dR p X − dL p X : X ∈ g , and Tp T = dR p Z : Z ∈ t ,

cf. (1.13). If Q is a bi-invariant metric on G, then Lg and Rg are isometries for all
g ∈ G. Thus, for all X ∈ g and Z ∈ t, we have
 
Q dR p X − dL p X, dR p Z = Q(X, Z) − Q(dL p X, dR p Z)
 
= Q(X, Z) − Q X , d(L p−1 ) p dR p Z
   
= Q X, Z − Ad p−1 Z
= 0,

since T is abelian, and hence Z ∈ t satisfies Ad(g)Z = Z for all g ∈ T , see (1.8). 

Remark 4.3. A maximal torus T in G is its own centralizer, that is, ZG (T ) = T , see
Exercise 1.59. In fact, it is clear that T ⊂ ZG (T ), and the existence of g ∈ ZG (T )
with g ∈
/ T would contradict the maximality of T as an abelian subgroup of G.
Definition 4.4. The rank of a connected, compact Lie group G is defined to be the
dimension of a maximal torus2 in G, and is denoted by rank G.
Example 4.5. Consider the Lie group S3 = {(z1 , z2 ) ∈ C2 : |z1 |2 + |z2 |2 = 1}, with
the product defined in Exercise 1.52, and recall that S3 ∼
= SU(2). We claim that the
great circle
  
T= eiθ , 0 ∈ S3 : eiθ ∈ S1 (4.1)

is a maximal torus in S3 . In fact, if (w1 , w2 ) ∈ S3 commutes with every element of

T , then eiθ w2 = e−iθ w2 for all θ ∈ R, so w2 = 0, and hence (w1 , w2 ) ∈ T . Thus, the
Lie group S3 ∼ = SU(2) ∼= Sp(1) has rank one. In particular, the maximal tori in S3

2 Bythe Maximal Torus Theorem 4.1 (ii), all maximal tori in G are conjugate and hence have the
same dimension.
88 4 Adjoint and Conjugation Actions

are its 1-parameter subgroups, which are the great circles through the origin (1, 0)
and its antipodal point (−1, 0). These are all conjugate to (4.1), cf. Maximal Torus
Theorem 4.1 (ii) and (iii).
In fact, the above is essentially the only rank one Lie group, due to the following
result, see Duistermaat and Kolk [79].
Theorem 4.6. If G is a connected compact Lie group of rank one, then G is
isomorphic to either SU(2), SO(3), or S1 .
Example 4.7. The subgroup T ⊂ U(n) of diagonal matrices given by
⎧⎛ iθ ⎞ ⎫
⎪
⎨ e
1
⎪
⎬
⎜ .. ⎟
T= ⎝ . ⎠ j
: θ ∈ R , (4.2)
⎪
⎩ ⎪
⎭
e iθn

is clearly an n-dimensional torus. It can be shown that T is a maximal torus in

U(n), and also in Sp(n) ⊃ U(n). In particular, rank U(n) = rank Sp(n) = n. Using
Exercise 1.60, it is easy to see that the group T /Z(U(n)), which can be identified
with the subgroup of T formed by matrices with θ1 + · · · + θn = 0, is a maximal
torus3 in SU(n), and hence rank SU(n) = n − 1. The image of T under the standard
embeddings U(n) ⊂ SO(2n) ⊂ SO(2n + 1) is also a maximal torus in both SO(2n)
and SO(2n + 1), which hence also have rank n.
The above geometric properties of the conjugation action on a compact Lie group
(G, Q), asserted in the Maximal Torus Theorem 4.1, generalize to the interesting
class of polar actions. The study of these actions was pioneered by Szenthe [202,
203] and independently by Palais and Terng [181], see also [182].
Definition 4.8. An isometric action of a compact Lie group G on a complete
Riemannian manifold (M, g) is called a polar action if there exists a connected,
complete, immersed submanifold Σ , called a section, that intersects all the G-orbits,
and every intersection between Σ and a G-orbit is orthogonal. Moreover, if Σ is flat,
the action is called hyperpolar.
Exercise 4.9 (). Let Σ be a section for a polar action on (M, g). Verify that:
(i) The dimension of Σ is equal to the codimension of the principal orbits;
(ii) For any x ∈ M, there exists a section Σ  with x ∈ Σ  ;
(iii) Sections are totally geodesic submanifolds.
Hint: In order to prove (iii), first verify that a section Σ is totally geodesic on regular points, using
properties of integrable Riemannian submersions, see (2.12). Since regular points are dense in Σ
by Kleiner’s Lemma 3.70, conclude that Σ is totally geodesic.

3 Notice that, under the isomorphism S3 ∼ = SU(2) in Exercise 1.52, the maximal torus (4.1) is
precisely the subgroup of (4.2) with n = 2 such that θ1 + θ2 = 0.
4.1 Maximal Tori and Polar Actions 89

Remark 4.10. A key property of polar actions is that they can be characterized as
the isometric actions whose horizontal distribution (3.20) is integrable on the regular
stratum. This is explained in Sect. 5.4, together with other aspects of polar actions
that hold in the more general context of polar foliations.
The Maximal Torus Theorem 4.1 implies that the conjugation action on a
compact Lie group (G, Q) is polar, since maximal tori are sections. Moreover, it
is hyperpolar, since tori in (G, Q) are flat, by Proposition 2.26.
A linear polar action on a vector space is called a polar representation. It is not
difficult to show that the linearization of a polar action, restricted to the normal space
to the orbits, is a polar representation [182], cf. Theorem 5.27 (iii).
Proposition 4.11. If μ : G × M → M is a polar action, then the slice representation
at each x ∈ M (see Definition 3.72) is a polar representation of Gx on νx G(x).
Moreover, if Σ is a section through x ∈ M, then Tx Σ is a section on νx G(x).
In particular, the adjoint representation Ad : G × g → g of (G, Q) is a polar
representation. Indeed, it is the slice representation obtained by linearizing at e ∈ G
the (hyperpolar) conjugation action, recall Definition 1.36. Furthermore, a section
is given by the Lie algebra t of a maximal torus T , since this is the tangent space to
a section T for the conjugation action on G. This can be proved directly from the
Maximal Torus Theorem 4.1, since the tangent space to the adjoint orbit of Z ∈ t is
    
TZ Ad(G) Z = dtd Ad(exp(tX))Z t=0 : X ∈ g = [X, Z] : X ∈ g , (4.3)

and 0 = Q(X, [Z,Y ]) = Q([X , Z],Y ) for all Y ∈ t.

Other examples of polar actions are given by isometric actions whose principal
orbits have codimension 1. In fact, by Proposition 3.78, a geodesic that intersects
a G-orbit orthogonally intersects all G-orbits orthogonally. Such actions are called
cohomogeneity one actions4 and are studied in detail in Chap. 6.
More examples of polar actions can be found among symmetric spaces, which are
Riemannian manifolds (M, g) such that, for each p ∈ M, there is an isometry I p that
fixes p and reverses geodesics through p. As explained in Exercise 6.32, symmetric
spaces are of the form M = G/H, where G is the connected component of Iso(M, g)
that contains the identity, and H is the isotropy group of a point p ∈ M; such a
pair of groups (G, H) is called a symmetric pair. In a symmetric space M = G/H,
a maximal totally geodesic flat submanifold Σ through p ∈ M is a section for the
subaction of the isotropy group H, which is the left translation h · gH := hgH on
G/H, see (6.13). Note that these polar actions are also hyperpolar, as Σ is flat. It can
be proved that if (G, H1 ) and (G, H2 ) are symmetric pairs, then the subaction of H1
on G/H2 is also hyperpolar. These actions are called Hermann actions, as they were
introduced by Hermann [127], see also Thorbergsson [209].

4 The cohomogeneity of an action is, by definition, the codimension of its principal orbits.
90 4 Adjoint and Conjugation Actions

Remark 4.12. The conjugation action on a Lie group G with bi-invariant metric Q
is a particular case of the above hyperpolar actions of isotropy groups of symmetric
spaces. In fact, (G, Q) is a symmetric space by Theorem 2.30, which can be written
as G ∼= (G × G)/Δ G, see Example 3.39. It is easy to check that conjugation on G
corresponds to the subaction of the isotropy group Δ G on (G × G)/Δ G, and that a
maximal torus is a maximal totally geodesic flat in this symmetric space.
It can be shown that the isotropy representation of a symmetric space M = G/H
is a polar representation of H. These representations are called s-representations.
Polar representations were classified by Dadok [72], who proved that any such
representation must be orbit-equivalent5 to an s-representation.
Podestà and Thorbergsson [186] classified polar actions on compact rank6
one symmetric spaces (see Example 6.10 for details on these spaces). Besides
cohomogeneity one actions, sections for these polar actions are not flat, since they
are totally geodesic (see Exercise 4.9) and these spaces have positive sectional
curvature. Kollross [146] classified hyperpolar (and cohomogeneity one) actions
on compact irreducible symmetric spaces, proving that they are orbit-equivalent to
a Hermann action if they have cohomogeneity at least two. It had been conjectured
that polar actions on compact irreducible symmetric spaces of rank greater than one
are hyperpolar. A partial answer was given by Biliotti [40], who proved it under
the extra assumption that the symmetric space is Hermitian. The full conjecture was
recently settled by Lytchak [155], see Theorem 5.38. Another recent development in
the area was the classification of polar manifolds with sec > 0 and cohomogeneity
at least two by Fang, Grove and Thorbergsson [84], see Theorem 5.39. For surveys
on polar actions, see Thorbergsson [208, 209].
We conclude this section with a discussion of isoparametric submanifolds, which
also generalize some of the geometric properties of orbits of the conjugation action.
Definition 4.13. Let L be an immersed submanifold of a Riemannian manifold
(M, g). A section ξ of the normal bundle ν (L) is said to be a parallel normal
field along L if ∇ν ξ vanishes identically, where ∇ν is the normal connection.7
The submanifold L is said to have flat normal bundle, if any normal vector can
be extended to a locally defined parallel normal field. In addition, L is said to have
globally flat normal bundle, if the holonomy of the normal bundle ν (L) is trivial,
which means that any normal vector can be extended to a globally defined parallel
normal field.
Definition 4.14. A submanifold F of a space form M(k) is called isoparametric if
its normal bundle is flat and principal curvatures along any parallel normal vector
field are constant.

5 Isometric actions of Lie groups G1 and G2 on Riemannian manifolds M1 and M2 are orbit-
equivalent if there exists an isometry between M1 and M2 that maps G1 -orbits to G2 -orbits.
6 The rank of a symmetric space is the dimension of a maximal totally geodesic flat submanifold.
7 In other words, ∇ν ξ is the component of ∇ξ that is normal to L.
4.1 Maximal Tori and Polar Actions 91

It can be shown that principal orbits of a polar action on a Riemannian

manifold have globally flat normal bundle, since they are described by an integrable
Riemannian submersion (see Remarks 3.79 and 4.10). In particular, equivariant
normal vectors turn out to be parallel normal fields. It is not difficult to prove that
principal curvatures along equivariant normal vector fields of principal orbits of a
polar action are constant (see Proposition 3.78). These two facts together imply the
next result.
Proposition 4.15. Principal orbits of a polar action on Euclidean space are
isoparametric submanifolds. In particular, principal orbits of the adjoint action of
a compact Lie group G are isoparametric submanifolds of its Lie algebra g.
In the particular case of adjoint actions, each regular orbit is a principal orbit (see
Theorem 4.36), and its principal curvatures and principal directions can be explicitly
computed (see Example 4.29).
It is possible to prove that the normal bundle of an isoparametric submanifold
is globally flat and hence that each normal vector can be extended to a parallel
normal vector field. An important property of isoparametric submanifolds is that sets
parallel to an isoparametric submanifold are submanifolds. This means that given
an isoparametric submanifold N and a parallel normal vector field ξ , the endpoint
map ηξ (x) := expx (ξ ) is such that ηξ (N) is a submanifold, with dimension possibly
lower than dim N. An isoparametric foliation F on M(k) is a partition of M(k) by
submanifolds parallel to a given isoparametric submanifold N. It is possible to prove
that a leaf L of F is isoparametric if dim L = dim N, in which case the partition into
parallel submanifolds to L coincides with F .
Remark 4.16. Isoparametric hypersurfaces in space forms have been studied since
Cartan [62–65], and remain a very active field of research. In Euclidean and
hyperbolic space, they can only be cylinders or umbilic hypersurfaces. In spheres,
there are other examples of isoparametric hypersurfaces, many of which are
inhomogeneous, see Ferus, Karcher and Münzner [89]. Münzner [170, 171] proved
that isoparametric hypersurfaces in spheres separate the ambient into two disk
bundles over focal submanifolds and used this to show that the number of principal
curvatures can only be 1, 2, 3, 4 or 6, with certain restrictions on the multiplicities.
All of these numbers are known to occur. The classification of isoparametric
hypersurfaces in spheres is almost complete, except for a last remaining case, see
Miyaoka [162] and Siffert [196], and also [69, 210].
A general concept of isoparametric submanifold in higher codimensions was
independently introduced by Harle [122], Carter and West [67, 68] and Terng [204],
motivated by analogies with orbits of polar actions in Rn . In particular, their study
is also connected with Coxeter groups. Regarding the relation of isoparametric
submanifolds with group actions, Thorbergsson [206] proved that a compact,
irreducible isoparametric submanifold Ln in Rn+k is homogeneous, provided k ≥ 3
and Ln does not lie in any affine hyperplane of Rn+k .
Details on Terng’s work about isoparametric submanifolds on space forms and
Hilbert spaces can be found in Palais and Terng [182]. Another important reference
92 4 Adjoint and Conjugation Actions

on isoparametric submanifolds is the book of Berndt, Console and Olmos [32]. More
information on the history of isoparametric hypersurfaces and submanifolds can
be found in the surveys of Thorbergsson [208, 209], see also Sect. 5.5.

4.2 Normal Slices of Conjugation Actions

In this section, we explain how normal slices for the conjugation action of a compact
Lie group (G, Q) can be explicitly described in terms of the collection of maximal
tori in G. Notice that the orbits of the conjugation action of G on itself are its
conjugacy classes, and the isotropy group of x ∈ G is its centralizer Gx = ZG ({x}).
Lemma 4.17. Let G0x be the connected component of Gx = {g ∈ G : g x g−1 = x}
that contains the identity. Then x ∈ G0x .
Proof. Since G is compact, there exists X ∈ g such that exp(X) = x, see
Theorem 2.27. Thus, x = exp(X) = exp(tX) exp(X) exp(−tX) = exp(tX) x exp(−tX),
so exp(tX) ∈ G0x for all t ∈ R and, in particular, x = exp(X) ∈ G0x . 

Proposition 4.18. Let T be the collection of maximal tori in G, and let Λ (x) be

the subcollection of maximal tori that contain x ∈ G. Then T ∈T T = Z(G) is the

center of G, and T ∈Λ (x) T = Gx is the identity component of the isotropy group Gx .
0

 Torus Theorem4.1 implies that G = T ∈T T and
Proof. Item (iii) of the Maximal

hence Z(G) = Z T ∈T T = T ∈T ZG (T ) = T ∈T T , by Remark 4.3. Further-

more, it is clear that T ∈Λ (x) T ⊂ G0x . For the converse inclusion, let g ∈ G0x and T 
be a maximal torus in G0x that contains g. By Lemma  4.17,
 we have that x ∈ G0x ,
and hence trivially x ∈ Z G0x . As proved above, Z G0x is the intersection of all
the maximal tori in G0x , hence x ∈ T  . Letting T be a maximal torus in G such that

T  ⊂ T , we have g, x ∈ T and hence g ∈ T ∈Λ (x) T . 

Slice Theorem 4.19. Let (G, Q) be a compact connected Lie group acting on itself
by conjugation, and let Sx be a normal slice at x ∈ G with radius ε > 0.

(i) Sx = σ ∈Λε (x) σ , where Λε (x) := {Bε (x) ∩ T : T ∈ Λ (x)};
(ii) Sy ⊂ Sx for all y ∈ Sx ;
(iii) Denoting by F the homogeneous foliation of G by orbits of the conjugation
% of a neighborhood
action, see (3.4), there exists an isoparametric foliation F
 
% = F ∩ Sx .
of the origin 0 ∈ Tx Sx , such that expx F

Proof. It follows from Proposition 4.18 that G0x ∩ Bε (x) = σ ∈Λε (x) σ . Given
σ ∈ Λε (x), the minimal geodesic segment γ joining y ∈ σ to x is contained in
σ , since every maximal torus is totally geodesic in G. By the Maximal Torus
Theorem 4.1, the conjugation orbit G(x) is orthogonal to the maximal torus σ ,

hence γ is orthogonal to G(x), so y ∈ Sx . Thus, σ ∈Λε (x) σ ⊂ Sx . Moreover, note
that dim Sx = dim G − dim G/Gx = dim Gx , from which it follows that
4.3 Roots of a Compact Lie Group 93

&
G0x ∩ Bε (x) = σ = Sx . (4.4)
σ ∈Λε (x)

Item (ii) follows from Gy ⊂ Gx and (4.4). In order to prove (iii), notice that (4.4)
implies that there is a neighborhood U of e ∈ G0x such that Lx (U) = xU = Sx . Set
Fx := {Gx (y) : y ∈ Sx } and F
e := {Gx (y) : y ∈ U}. Since x ∈ Z(Gx ), we have F x =
 
 
Lx Fe . By the properties of a slice of an isometric action, Fx = F ∩ Sx . Define

F%e := Ad(Gx )X : X ∈ (Bε (0) ∩ gx ) , where gx is the Lie algebra of Gx , and F %:=
 
dLx F %e . Clearly, Fe = exp(F %e ). Combining these observations and (2.2),

         
x = Lx F
F ∩ Sx = F e = Lx exp F
%e = exp dLx F
%e = exp F % .
x x

% is isoparametric, as F
Finally, F %e is isoparametric and Lx is an isometry. 

4.3 Roots of a Compact Lie Group

Roots play a fundamental role in the theory of compact Lie groups. They allow, for
instance, to classify compact simple Lie groups (see Sect. 4.5) and are also related
to principal curvatures of principal orbits of the adjoint action (see Example 4.29).
Roots are defined in Theorem 4.23.
For the sake of motivation, we first state a result that follows directly from
Theorem 4.23 and Remark 4.27, which explains the importance of roots in
understanding the adjoint representation, as it splits into rotations on appropriate
subspaces.
Proposition 4.20. Let G be a connected compact nonabelian Lie group with Lie
algebra g and T ⊂ G a maximal torus with Lie algebra t. There exist 2-dimensional
subspaces V j ⊂ g and linear functionals α j : t → R such that:
(i) g = t ⊕V1 ⊕ · · · ⊕Vz ;
(ii) If Q is a bi-invariant metric on G, considering a Q-orthonormal basis of V j ,
 
  cos α j (X) − sin α j (X)
Ad exp(X ) V = , for all X ∈ t;
j sin α j (X) cos α j (X)

(iii) The above decomposition is unique, and α j = αk if j = k.

Each linear functional α : t → R above is the root associated to the space Vα , see
the definition in Theorem 4.23. Before proving this main theorem, we need to state
the following auxiliary result from linear algebra.
Lemma 4.21. Let W ⊂ Cn be a complex subspace. Then W is invariant by complex
conjugation (i.e., W = W ) if and only if W = V ⊕ iV for a real subspace V ⊂ Rn .
94 4 Adjoint and Conjugation Actions

Remark 4.22. If {v j } is a real basis for a real subspace V ⊂ Rn , then {v j } is a

complex basis for the complex subspace W = V ⊕ iV . In particular, dimC W =
dimR V .
Notice that if G is a Lie group with Lie algebra g, the complexification of g,

gC := g ⊗R C, (4.5)

is a (complex) Lie algebra where [·, ·] : gC × gC → gC is the canonical extension of

the Lie bracket of g. As a real vector space, gC = g ⊕ i g. We are now ready to prove
the main result in this section, based on Duistermaat and Kolk [79, pp. 170–171].
Theorem 4.23. Let G be a connected compact nonabelian Lie group and T ⊂ G a
maximal torus, with Lie algebras g and t, respectively.
(i) For each X ∈ g, the linear map ad(X) : gC → gC is diagonalizable with purely
imaginary eigenvalues;
(ii) There exists a unique (up to permutations) decomposition of gC in subspaces

gα := Y ∈ gC : [X,Y ] = i α (X)Y, for all X ∈ t ,

where α : t → R is a linear functional called root. More precisely,

'
gC = g0 ⊕ gα ,
α ∈R

where R denotes the set of roots, also called root system;

(iii) g0 = t ⊕ i t and gα = g−α . In particular, α ∈ R implies −α ∈ R;
(iv) dimC gα = 1 and dimVα = 2, where Vα := (gα ⊕ g−α ) ∩ g;
(v) gλ α = 0 if λ ∈ C \ {−1, 0, 1};
(vi) If e2 + i e1 is the vector that spans gα , where e1 , e2 ∈ g, then:
(vi-a) {e1 , e2 } is a basis of Vα ;
(vi-b) [X , e1 ] = α (X )e2 and [X, e2 ] = −α (X)e1 , for all X ∈ t;
(vi-c) Q(e1 , e2 ) = 0 and e1  = e2 , with respect to any bi-invariant
metric Q; 
(vi-d) In the Q-orthonormal basis ee1  , ee2  ,
1 2

 
  cos α (X ) − sin α (X)
Ad exp(X) Vα = ;
sin α (X) cos α (X)

(vii) Let α ∨ ∈ t be the coroot of α , defined to be the vector such that α (α ∨ ) = 2

and α ∨ is Q-orthogonal to ker α .
(vii-a) α ∨ is orthogonal to ker α with respect to any bi-invariant metric;
4.3 Roots of a Compact Lie Group 95

(vii-b) g(α ) := spanR {α ∨ } ⊕ Vα is a Lie algebra isomorphic to su(2), and

the corresponding Lie group G(α ) := exp(g(α ) ) is isomorphic to either
SU(2) or SO(3); 
(vii-c) For each g ∈ G(α ) , Ad(g)ker α = id;
(vii-d) There exists w ∈ G(α ) , such that Ad(w)α ∨ = −α ∨ .
Proof. Using the Jordan normal form, there exists a decomposition of gC in
subspaces g j such that ad(X)|g j = c j I + N j where N j is a nilpotent matrix.8
Therefore
m j −1 k
  t k
exp tad(X) g = exp (tc j ) ∑ Nj . (4.6)
j
k=0 k!

On the other hand, using (1.12), compactness of Ad(G) implies that exp(tad(X)) is
bounded. Together with (4.6), this implies that c j is a purely imaginary number and
N j = 0, proving (i). Item (ii) follows from the next two claims.
Claim 4.24. There exists a decomposition gC = g0 ⊕ g1 ⊕ · · · ⊕ gz and linear
functionals α j : t → R, such that for each Y ∈ g j , [X,Y ] = i α j (X)Y for all X ∈ t. In
addition, α j = αk if j = k.
In order to verify this claim, consider a basis {X1 , . . . , Xr } of t. On the one hand,
from (i), each ad(Xl ) is diagonalizable. On the other hand, ad(Xl ) commutes with
(
ad(Xk ) for all k, l, since [Xl , Xk ] = 0. Thus, there is a decomposition gC = j g j in
common eigenspaces of the operators ad(Xl ). Define
) *
r r
αj ∑ xl Xl := ∑ xl α jl ,
l=1 l=1

where [Xl ,Y ] = i α jl Y for all Y ∈ g j . By construction, g j ⊕ gk is not an eigenspace

of all operators ad(Xl ), hence α j = αk if j = k.
Claim 4.25. If [X,Y ] = i β (X)Y for all X ∈ t, then there exists j such that Y ∈ g j
and β = α j , where g j and α j were defined in Claim 4.24.
In order to prove this claim, let Y = Y0 + Y1 + · · · + Yz , with Y j ∈ g j . Then
∑zj=1 i β (X)Y j = [X,Y ] = ∑zj=1 i α j (X)Y j . Since gC = g0 ⊕ g1 ⊕ · · · ⊕ gz , we have
that β (X) = α j (X). Thus, Claim 4.25 follows from the fact that α j = αk if j = k.
Item (iii) follows from Lemma 4.21 and the fact that t is a maximal abelian Lie
subalgebra of g.
With regard to (iv) and (v), since the set of roots R is finite, we can choose a
vector X ∈ ker α such that β (X) = 0, for all roots β that are not multiples of α .

j = 0 if m = m j , and N j = 0 if m < m j
8 This means that N m m
96 4 Adjoint and Conjugation Actions

Note that the Lie algebra of the identity connected component of the isotropy group
G0X with respect to the adjoint action is
' 
gX = t ⊕ gλ α ⊕ g−λ α ∩ g. (4.7)
λ =0

This implies that Z(gX ) = ker α , where Z(gX ) is the Lie algebra of Z(G0X ), see
Corollary 1.57. Set G := G0X /Z(G0X ) and note that the Lie algebra of the compact
connected Lie group G is gX /Z(gX ). It follows from (4.7) that t/ ker α is a maximal
abelian Lie subalgebra of gX /Z(gX ). Thus, G has rank one and hence gX /Z(gX ) ∼ =
su(2) by Theorem 4.6. The fact that dim gX /Z(gX ) = 3 implies that
' 
dim gλ α ⊕ g−λ α ∩ g = 2. (4.8)
λ =0

Items (iv) and (v) follow from (4.8) and Remark 4.22.
As for (vi), since gα ∩ g−α = {0}, e1 = λ e2 . This, and the fact that dimVα = 2,
implies (vi-a). Item (vi-b) follows directly from [X, e2 + i e1 ] = i α (X)e2 − α (X)e1 .
Item (vi-c) follows from the fact that ad is skew-symmetric with respect to every
bi-invariant metric (see Proposition 2.26). Finally, (vi-d) follows from (vi-b) and
from
  
d
dt Ad exp(tX) e j t=0 = [X, e j ].

In order to prove (vii), we first prove that g(α ) is a Lie algebra and that the
definition of α ∨ does not depend on the bi-invariant metric. Let Y be an infinitesimal
generator of t. The Jacobi identity and (vi-b) imply that [[e1 , e2 ],Y ] = 0. Thus, from
Claim 4.2, [e1 , e2 ] ∈ t. Note that [e1 , e2 ] = 0, because otherwise ker α ⊕ Vα would
be an abelian subalgebra with dimension greater than dim t. Item (vi-b) and the
fact that ad is skew-symmetric with respect to every bi-invariant metric imply that
[e1 , e2 ] is orthogonal to ker α . In particular, the definition of α ∨ does not depend on
the bi-invariant metric. To verify that g(α ) is a Lie algebra, note that
   
[[e1 , e2 ], e1 ] = α [e1 , e2 ] e2 , and [[e1 , e2 ], e2 ] = −α [e1 , e2 ] e1 .

Note also that g(α ) ⊂ gX , g(α ) ∩ Z(gX ) = 0 and dim g(α ) = 3 = dim gX /Z(gX ),
where X was defined in the proof of (iv). Therefore, g(α ) ∼= gX /Z(gX ) ∼= su(2) are
( α )
isomorphic Lie algebras. In particular, G is a compact Lie group isomorphic to

SU(2) or SO(3), see Exercise 1.55, which implies that Ad G(α ) g(α ) ∼ = SO(3).
Since spheres are homogeneous in g(α ) , there exists w ∈ G(α ) such that Ad(w)α ∨ =
−α ∨ . Finally, as Ad(exp(Y
 ))Z = exp(ad(Y ))Z = Z for Y ∈ g(α ) and Z ∈ ker α , we

conclude that Ad(g) ker α = id for all g ∈ G(α ) . 

In what follows, we use the same notation from Theorem 4.23.

4.3 Roots of a Compact Lie Group 97


Definition 4.26. A connected component of t \ α ∈R ker α is called a Weyl cham-
ber, and each hyperplane ker α is called a wall. For a fixed Weyl chamber C, we
define the set of positive roots as

P := α ∈ R : α (X ) > 0, for all X ∈ C .

Note that R = P ∪ −P and P ∩ −P = 0, / where −P := {−α : α ∈ P}. In particular,

|P| = 12 |R|, that is, P has exactly half of the elements of R, cf. Theorem 4.23 (iii).
Instead of defining the set of positive roots P for a given choice of Weyl chamber
C, one can instead define a Weyl chamber for a given choice of positive roots. More
precisely, a set of roots P satisfying:
(i) P ∪ −P = R and P ∩ −P = 0; /
(ii) If ∑α ∈P cα α = 0, with cα ≥ 0 for all α ∈ P, then cα = 0 for all α ∈ P;
is called a choice of positive roots. The associated Weyl chamber is then defined as

C := X ∈ t : α (X) > 0, for all α ∈ P .

Thus, there is a correspondence between sets of positive roots and Weyl chambers.
Remark 4.27. It follows from Theorem 4.23 that if G is a connected compact
(
nonabelian Lie group, then g = t ⊕ α ∈P Vα , where P is a choice of positive roots,
cf. Proposition 4.20. Thus, the number of roots in G is |R| = dim G − rank G. In
particular, the codimension of any maximal torus in G is even.
Remark 4.28. If X ∈ C belongs to a Weyl chamber, then Ad(G)X is a regular orbit
/ ker α for all
and its codimension is equal to the dimension of T . Indeed, since X ∈
α ∈ R, we conclude that gX = t. In Theorem 4.36, we prove that regular orbits of
the adjoint action are principal orbits.
It was proved in Proposition 4.15 that principal orbits of the adjoint action of a
compact Lie group are isoparametric (recall Definition 4.14). We now compute the
corresponding principal curvatures and principal directions, in terms of roots.
Example 4.29. Let (G, Q) be a connected compact Lie group with bi-invariant
metric and T ⊂ G a maximal torus, with Lie algebras g and  t respectively.
 Let Z ∈ t
be such that Ad(G)Z is a principal orbit, and recall that TZ Ad(G)Z = {[X , Z] : X ∈
g}, see (4.3). From the Maximal Torus Theorem 4.1, if N ∈ t, then

"
N(Ad(g)Z) := Ad(g)N

is a well-defined normal field along Ad(G)Z. Differentiating at g = e, we have

 
"Z ([X, Z]) = ∇[X,Z] N
dN " = [X , N], (4.9)
Z
98 4 Adjoint and Conjugation Actions

which proves that N" is a parallel normal field along Ad(G)Z, since Q([X, N],Y ) =
Q(X, [N,Y ]) = 0 for all Y ∈ t. Thus, from (4.9),

SN" ([X , Z]) = −[X, N], (4.10)

" see (2.8). If {e1 , e2 } is a basis of

where SN" is the shape operator in the direction N,
Vα ⊂ g as in Theorem 4.23, replacing X by e j in (4.10), it follows that

α (N) α (N)
SN" (e1 ) = − e1 , SN" (e2 ) = − e2 .
α (Z) α (Z)

Therefore, Vα are the curvature distributions of the orbit Ad(G)Z, with principal
curvature − αα(N)
(Z) . In particular, the spaces Vα are pairwise Q-orthogonal.

We conclude this section by computing the roots of U(n) and SU(n). Analogous
computations for the other classical Lie groups can be found in Fegan [85].
Example 4.30. Consider the maximal torus (4.2) of U(n), recall Example 4.7. Its
Lie algebra t ⊂ u(n) consists of the matrices
⎛ ⎞
i θ1
⎜ .. ⎟
X =⎝ . ⎠, (4.11)
i θn

where θk ∈ R. For each k = l, denote by Ekl the matrix with 1 in the (k, l)th entry
and −1 in the (l, k)th entry, and by Fkl the matrix with i in both (k, l)th and (l, k)th
entries, and zeros elsewhere. It is easy to see that Ekl and Fkl span the complement
V of t ⊂ u(n), i.e., u(n) = t ⊕V . A direct computation yields:

ad(X)Ekl = XEkl − Ekl X = (θk − θl )Fkl ,

ad(X)Fkl = XFkl − Fkl X = −(θk − θl )Ekl .

Thus, ad(X) : V → V leaves each 2-dimensional subspace Vkl = spanR {Ekl , Fkl },
invariant, where it has eigenvalues ±i(θk − θl ). By uniqueness of the decomposition
(
u(n) = t ⊕ α ∈P Vα , see Theorem 4.23 and Remark 4.27, the roots of U(n) are:

αkl : t → R, αkl := θk∗ − θl∗ , k = l, (4.12)

where θk∗ (X) := θk for an X as in (4.11). Notice that, since k = l, there are precisely
n(n − 1) roots αkl , and dim U(n) = n2 and rank U(n) = n, cf. Remark 4.27. It is easy
to see that P = {αkl : k < l} is a choice of positive roots.
The computation above also yields the roots of SU(n), since the Lie algebra of
its maximal torus is t/Z(u(n)), which can be identified with the subspace t of t
4.4 Weyl Group 99

spanned by elements X as in (4.11) that satisfy θ1 + · · · + θn = 0, see Example 4.7.

Thus, the restriction of (4.12) to t are the roots of SU(n). Again, notice that there
are n(n − 1) roots, and dim SU(n) = n2 − 1 and rank SU(n) = n − 1, cf. Remark 4.27.

4.4 Weyl Group

Let (G, Q) be a connected compact Lie group with bi-invariant metric and T
a maximal torus. In this section, we define the Weyl group and use it to study
intersections of orbits of the adjoint action with the Lie algebra t of the maximal
torus T .
Definition 4.31. The Weyl group is defined by

W := N(T )/T,

where N(T ) = {g ∈ G : g T g−1 = T } is the normalizer of the maximal torus T in G.

The Weyl group W is a finite group, since the connected component of N(T ) that
contains the identity is T . In fact, recall that N(T ) = {g ∈ G : Ad(g)t ⊂ t} and its Lie
algebra is n = {X ∈ g : ad(X )t ⊂ t}, see Exercise 1.61. Let X ∈ t be an infinitesimal
generator of T and Y ∈ n. Then Q([X,Y ], Z) = −Q(Y, [X, Z]) = 0 for all Z ∈ t, hence
[X,Y ] = 0, and so Y ∈ t by Claim 4.2. This proves that n = t and, in particular, the
above claim.
Remark 4.32. Since T = ZG (T ) is its own centralizer (see Remark 4.3), it follows
that W = NG (T )/ZG (T ) is the quotient of the normalizer by the centralizer of T ,
which is the group of automorphisms of T induced by inner automorphisms in G.
Remark 4.33. By the Maximal Torus Theorem 4.1, if T  is another maximal torus
in G, then T  = g T g−1 for some g ∈ G. Since N(g T g−1 ) = g N(T ) g−1 , it follows
that the corresponding Weyl groups N(T  )/T  and N(T )/T are isomorphic.
Exercise 4.34. Consider the maximal torus T of S3 = ∼ SU(2) given by (4.1).
Compute N(T ) and verify that the Weyl group of SU(2) is W ∼
= Z2 .
From the above discussion, we also have that

μ : W × t → t, μ (w T, X) := Ad(w)X, (4.13)

is a well-defined effective isometric W -action on t, whose orbits coincide with the

intersections of t with the orbits of the adjoint action of G on g. In order to study
in more details how these actions are related, it is convenient to simultaneously use
roots α : t → R and their duals α ∗ ∈ t determined by the bi-invariant metric Q, i.e.,
100 4 Adjoint and Conjugation Actions

α (·) = Q(α ∗ , ·). Notice that α ∗ = 2α ∨ /α ∨ 2 , where α ∨ is the coroot associated to
α , see Theorem 4.23 (vii). We denote by ϕα the reflection across the wall ker α ⊂ t,
  ∗
α∗ α
ϕα : t → t, ϕα (Z) := Z − 2 Q Z, ∗ .
α  α ∗ 

In particular, notice that ϕα (Z) = Z − α (Z)α ∨ for all Z ∈ t.

Item (vii) of Theorem 4.23 implies that the reflections ϕα are elements of the
Weyl group W . We prove in Theorem 4.37 that W is generated by these reflections.
With this aim, we now establish a preliminary result that has the interesting
consequence that the adjoint action does not have exceptional orbits.
Lemma 4.35. The isotropy group GX = {g ∈ G : Ad(g)X = X } of X ∈ t with respect
to the adjoint action is a connected Lie group.
Proof. Given x ∈ GX , we want to prove that x ∈ G0X , i.e., x is in the identity
connected component of GX . Let Gx be its isotropy group with respect to the
conjugation action. Then X is in the Lie algebra of Gx , since tAd(x)X = tX and
hence x exp(tX)x−1 = exp(tX), which is equivalent to exp(−tX) x exp(tX) = x.
Let T  be a maximal torus of G0x , tangent to X . Since [Z, X] = 0 for every Z in
the Lie algebra of T  , Z is in the Lie algebra of GX , hence T  ⊂ G0X . Since x ∈ G0x
(see Lemma 4.17), by the Maximal Torus Theorem 4.1, there exists g ∈ G0x such that
g x g−1 ∈ T  . Finally, since g ∈ G0x , we have that x = g x g−1 , and hence x ∈ G0X .


Theorem 4.36. Every regular orbit of the adjoint action is a principal orbit.
Proof. Let X0 ∈ g be a regular point. This implies that dim GX0 ≤ dim GX for all
X ∈ g. If SX0 is a slice at X0 , then GY ⊂ GX0 for each Y ∈ SX0 . Since dim GX0 ≤
dim GY and both are connected, it follows that GY = GX0 , i.e., G(X0 ) is a principal
orbit. 

Theorem 4.37. Let W be the Weyl group. Then the following hold:
(i) If C1 and C2 are two Weyl chambers such that X ∈∂ C1 ∩ ∂ C2 , then  there exist
reflections ϕα1 , . . . , ϕαn such that ϕαi (X) = X and ϕαn ◦ · · · ◦ ϕα1 (C2 ) = C1 ;
(ii) The closure C of each Weyl chamber C is a fundamental domain for the action
of W on t, i.e., each orbit of the adjoint action intersects C exactly once;
(iii) W is generated by {ϕα }α ∈R , where R is the set of roots.
Proof. Each isometry w ∈ W maps regular points to regular points, and hence w(C1 )
is a Weyl chamber. Let KX be the subgroup of W generated by the reflections ϕα
such that ϕα (X) = X. Let B2ε (X) be an open ball in the intersection of a slice at X
with t, and consider two regular points p1 and p2 such that pi ∈ Ci and pi ∈ Bε (X).
Define f : KX → R as f (ϕα ) := ϕα (p2 )− p1 , and let w0 ∈ KX be a minimum of f .
If w0 (p2 ) ∈
/ C1 , then the straight line segment joining w0 (p2 ) to p1 intersects a wall
ker β . Since this intersection is contained in Bε (X), we have X ∈ ker β and hence
ϕβ ∈ KX . Therefore f (ϕβ w0 ) < f (w0 ), contradicting the fact that w0 is a minimum
of f . Thus w0 (p2 ) ∈ C1 , concluding the proof of (i).
4.4 Weyl Group 101

As for (ii), the Maximal Torus Theorem 4.1 implies that each orbit of the adjoint
action intersects t. Composing with reflections ϕα , we conclude that each regular
orbit intersects C at least once, and every singular orbit intersects ∂ C at least once.
We claim that each principal orbit intersects C exactly once. Set

H := h ∈ G : Ad(h)X ∈ C , for all X ∈ C ,

and assume that there exists X ∈ C such that the orbit Ad(G)X meets C more than
once, i.e., there is g ∈ G such that X = Ad(g)X ∈ C. Clearly g ∈ / T , and it is not
difficult to prove that g ∈ H. Since Ad(G) is compact, g has finite order9 n. Since
g ∈ H, we have that Y := X + Ad(g)X + · · · + Ad(gn−1 )X ∈ C, and clearly Ad(g)Y =
Y . Thus, g ∈ GY . On the other hand, since Y ∈ C, it follows from Remark 4.28 and
Lemma 4.35 that GY = GY0 = T , contradicting the fact that g ∈ / T.
Similarly, we claim that each singular orbit of the adjoint action intersects ∂ C
exactly once. Suppose there are 0 = X ∈ ∂ C and g ∈ G such that X = Ad(g)X ∈ ∂ C.
Let γ : [0, 1] → C be a curve with γ (0) = X and γ (t) ∈ C for 0 < t ≤ 1. Composing
with reflections ϕα if necessary, observe that Ad(g)γ |(0,1] ⊂ C. Therefore, if t is
sufficiently small, γ (t) = Ad(g)γ (t) ∈ C, contradicting the fact proved above that
each principal orbit of the adjoint action intersects C exactly once.
It onlyremains to prove (iii). For any g ∈ W , there exist reflections ϕα1 , . . . , ϕαn
such that ϕαn ◦ · · · ◦ ϕα1 ◦ g (C) ⊂ C. From (ii), we have that, for all X ∈ C,
 
ϕαn ◦ · · · ◦ ϕα1 ◦ g (X) ∈ {Ad(G)X ∩C} = {X}
 
Therefore, ϕαn ◦ · · · ◦ ϕα1 ◦ g |t is the identity, concluding the proof. 

Remark 4.38. Let F = {Ad(G)X}X∈g be the (homogeneous) singular foliation

determined by the orbits of the adjoint action. The above result implies that F ∩ t is
invariant under reflections across the walls of Weyl chambers in t.
Remark 4.39. Notice that the orbit space g/G of the adjoint action can be identified
with the closure of a Weyl chamber, which can in turn be identified with the orbit
space t/W of the Weyl group action (4.13) on t.
Remark 4.40. The above theorem implies that the Weyl group permutes the Weyl
chambers and that the cardinality |W | is equal to the number of Weyl chambers.
Exercise 4.41. From Exercises 4.5 and 4.34, the great circle T in (4.1) is a maximal
torus in SU(2), and the Weyl group is W ∼ = Z2 . Prove that its Lie algebra t = R
has two Weyl chambers C± = R± , which are half lines, and W acts by reflection
across the origin. Conclude that the orbit space of the adjoint action of SU(2) on
su(2) ∼
= R3 is a half line, and relate this with Exercises 1.9 and 1.55.

9 This means that there exists n such that gn = e and gn−1 = e.

102 4 Adjoint and Conjugation Actions

Example 4.42. The Weyl group of SU(n) is the group Sn of permutations in n

letters. This can be proved by computing the normalizer of a maximal torus, see
Example 4.7. This group acts on an element X as in (4.11) by permuting its
coordinates θ1 , . . . , θn . By Remark 4.40, it follows that there are n! Weyl chambers.
Let us consider SU(3), which has rank 2 and Weyl group S3 . The Lie algebra
of a maximal torus is identified with t = {(θ1 , θ2 , θ3 ) ∈ R3 : θ1 + θ2 + θ3 = 0}
and P = {α12 , α13 , α23 } is a choice of positive roots, where αkl = θk∗ − θl∗ , see
Example 4.30. The restriction of αkl to t is clearly (αkl )|t = θk∗ − θl∗ = 2θk∗ + θm∗ ,
where {k, l, m} = {1, 2, 3}. Introducing orthonormal coordinates (x, y) on the plane
t ⊂ R3 , elementary computations show that the√ walls ker(αkl )|t√in t ∼ = R2 are
 1 3  1 
the lines spanned by the vectors (1, 0), − 2 , 2 and − 2 , − 2 , and hence 3

each of the 6 Weyl chambers is a wedge of angle π /3. In these coordinates, the
 ∗  3 √3   3 √3  √
duals (αkl )|t are the vectors 2 , − 2 , 2 , 2 and (0, 3). The Weyl group
S3 is clearly the symmetry group generated by reflections across these walls, cf.
Theorem 4.37. By the above discussion, the orbit space of the adjoint action of
SU(3) on its Lie algebra su(3) is a 2-dimensional (closed) wedge of angle π /3.
Remark 4.43. The definition of Weyl group can be extended to the general context
of polar actions μ : G × M → M, where it is also referred to as polar group. If Σ is
a section for such an action, then the Weyl group is defined as W := N/Z, where
 
N = g ∈ G : μ (g, x) ∈ Σ for all x ∈ Σ and Z = g ∈ G : μ (g, x) = x for all x ∈ Σ ,

cf. Remark 4.32. It can be proved that W is a discrete subgroup of N(H)/H, where
H is the isotropy group of a principal G-orbit. Similarly to the case (4.13) of
the conjugation action of a compact Lie group on itself, the Weyl group has an
effective isometric action on Σ , whose orbits coincide with the intersections of Σ
with the G-orbits on M. In particular, the orbit spaces M/G and Σ /W are isometric.
The Weyl group of a polar action is also finite, provided that the orbits and
sections are compact. In order to see this, let x ∈ M be in a principal G-orbit, and note
that a slice Sx at x locally coincides with the section Σ x through x. The isotropy Gx
coincides with the isotropy group Wx of the Weyl group action. By Exercise 3.77, the
slice representation is trivial, and hence Wx = {e}. By the compactness assumptions,
it follows that W (x) ∼= W /Wx is finite and hence W is finite.

4.5 Dynkin Diagrams

In this section, we briefly discuss Dynkin diagrams and the classification of compact
simple Lie groups, without giving proofs, which can be found in standard textbooks
such as [58, 90, 121, 126, 191, 195].
We henceforth assume that (G, Q) is a connected compact semisimple Lie group
with bi-invariant metric and, as before, we denote by T a maximal torus in G and
4.5 Dynkin Diagrams 103

by R the set of roots. Moreover, if α , β ∈ R are roots, we set Q(α , β ) := Q(α ∗ , β ∗ ),

where α ∗ , β ∗ ∈ t are their duals, i.e., α (·) = Q(α ∗ , ·) and β (·) = Q(β ∗ , ·).
Proposition 4.44. Let α1 , α2 ∈ R be roots with α1 = ±α2 , and denote by σ =
span{α1∗ , α2∗ } the plane in t spanned by their duals. There are exactly 2m duals
of roots α ∗j ∈ σ , where m ∈ {2, 3, 4, 6}, and the normalized vectors α ∗j /α ∗j  are
uniformly distributed along the unit circle of σ . Moreover, if α ∗j , αk∗ ∈ σ are such
that Q(α j , αk ) ≤ 0, then Q(α j , αk ) = α j αk  cos(π − φjk ), where:
(i) φjk = π /2, and α j  = αk ;
(ii) φjk = π /3, and α j  = √αk ; √
(iii) φjk = π /4, and α j  = √2αk  or αk  = √2α j ;
(iv) φjk = π /6, and α j  = 3αk  or αk  = 3α j .
Remark 4.45. Part of the above rigidity in the geometric arrangement of α ∗j ∈ σ can
be explained as follows. The reflections ϕα on σ generate a finite group F ⊂ O(2)
of linear isometries, whose intersection with SO(2) ∼= S1 is an index 2 subgroup,
which must be isomorphic to Zm for some m ∈ N. This subgroup is hence generated
by a rotation of angle 2π /m on the plane σ . The trace of this rotation matrix is
2 cos(2π /m), and (from representation theory) this number must be an integer, so
the only possible values for m are 2, 3, 4, and 6.
Definition 4.46. A positive root α is called simple if it cannot be written as the sum
of two other positive roots.
Theorem 4.47. Let P be a choice of positive roots in R.
(i) The subset S ⊂ P of simple roots is a basis of t∗ ;
(ii) Each α ∈ R is a linear combination of elements of S with integer coefficients,
either all nonnegative or all nonpositive;
(iii) If α , β ∈ S, then Q(α , β ) ≤ 0.
The set S is called a base of the root system R.
Definition 4.48. The Dynkin diagram of R with respect to a base S is a directed
graph with one vertex for each simple root α j ∈ S. If j = k, there are either 0, 1, 2,
or 3 edges between α j and αk , which are oriented pointing from the larger root, say
α j , to the smaller root, say αk , according to the cases in Proposition 4.44. Namely,
if Q(α j , αk ) = α j αk  cos(π − φjk ), we proceed as follows:
(i) If φjk = π /2, there are no edges;

• •
αj αk
104 4 Adjoint and Conjugation Actions

(ii) If φjk = π /3, there is 1 edge;

• •
αj αk

√
(iii) If φjk = π /4, and α j  = 2αk , there are 2 edges;

•  •
αj αk

√
(iv) If φjk = π /6, and α j  = 3αk , there are 3 edges.

•  •
αj αk

It is possible to prove that the Dynkin diagram of R does not depend on the choice
of bi-invariant metric Q, nor on the choice of positive roots P.
Example 4.49. Let us determine the Dynkin diagram of SO(7). From Example 4.7,
the Lie algebra t of a maximal torus in SO(7) is formed by matrices of the form
⎛ ⎞
0 θ1
⎜−θ 0 ⎟
⎜ 1 ⎟
⎜ 0 θ2 ⎟
⎜ ⎟
⎜ ⎟
X =⎜ − θ2 0 ⎟.
⎜ ⎟
⎜ 0 θ3 ⎟
⎜ ⎟
⎝ −θ3 0 ⎠
0

The same type of argument used in Example 4.30 implies that the roots of SO(7)
are ±θk∗ and θk∗ ± θl∗ , k = l, where θk∗ (X) = θk for X as above. A possible choice of
positive roots is θk∗ and θk∗ ± θl∗ , k < l. For this choice, the basis of simple roots is
S = {α1 , α2 , α3 }, where

α1 := θ1∗ − θ2∗ , α2 := θ2∗ − θ3∗ , α3 := θ3∗ .

Consider the bi-invariant metric Q(X,Y ) √

:= tr XY t of SO(7). It is easy to check that
∗ ∗
Q(θk , θl ) = δkl , hence α1  = α2  = 2, α3  = 1 and moreover, φ12 = π /3,
φ13 = π /2 and φ23 = π /4. Therefore, the Dynkin diagram of SO(7) is:

• •  •.
α1 α2 α3
4.5 Dynkin Diagrams 105

Exercise 4.50 (). Use the calculus of roots of SU(n) in Example 4.30 to prove
that the Dynkin diagram of SU(3) is

• •

Hint: Use Examples 4.42 and 4.49.

The classification of (connected) compact simple Lie groups, can be reduced

to the classification of Dynkin diagrams, due to the next result. This classification
of diagrams is purely combinatorial, and we simply state its outcome, relating the
diagrams with the corresponding groups. For the proofs, we refer the reader to
Fulton and Harris [90].
Theorem 4.51. If G1 and G2 are (connected) simply-connected compact semisim-
ple Lie groups with the same Dynkin diagram, then G1 and G2 are isomorphic.
Theorem 4.52. The possible Dynkin diagrams of a connected, compact simple Lie
group are:

An : • • ··· • • n≥1

Bn : • • ··· •  • n≥2

Cn : • • ··· •  • n≥3

Dn : • • ··· • n≥4
>>
>>
>>
>>
•

where n denotes the number of vertices, and also:

E6 : • • • • •

•
106 4 Adjoint and Conjugation Actions

E7 : • • • • • •

E8 : • • • • • • •

F4 : • •  • •

G2 : •  •

Each of the above is the Dynkin diagram of a connected compact simple Lie
group, whose rank is equal to the subindex used. Regarding classical Lie groups
of rank n,
(i) An is the Dynkin diagram of SU(n + 1), for n ≥ 1;
(ii) Bn is the Dynkin diagram of SO(2n + 1), for n ≥ 2;
(iii) Cn is the Dynkin diagram of Sp(n), for n ≥ 3;
(iv) Dn is the Dynkin diagram of SO(2n), for n ≥ 4.
The remaining Dynkin diagrams E6 , E7 , E8 , F4 and G2 are the Dynkin diagrams
of the so-called exceptional Lie groups, which are denoted by the same symbol.
Despite the terminology, these groups make several natural appearances in geo-
metric problems (see Example 6.10 (v) and Table 6.1), especially in connection with
Cayley numbers.10 Their exceptionality is mostly due to not being part of an infinite
family of groups with increasing dimension, as the classical Lie groups. For details
on the construction of such exceptional Lie groups, we refer the reader to [26, 123].
Remark 4.53. According to the above result, the Lie groups SO(7) and Sp(3) are
nonisomorphic, but each has 3 simple roots, {α1 , α2 , α3 } and {β1 , β2 , β3 }, with

 α j , αk  β j , βk 
= = cos(π − φjk )
α j αk  β j βk 

where φjk are the angles π /2, π /3 and π /4. Indeed, their Dynkin diagrams are
the same up to the orientation of the edges. In particular, the orbit spaces of the

10 The Cayley numbers Ca, or octonions, form one of the 4 real normed division algebras, together
with real numbers R, complex numbers C and quaternions H.
4.5 Dynkin Diagrams 107

adjoint actions of SO(7) and Sp(3) are isometric, despite these representations
being nonequivalent. This interesting phenomenon alludes to a recent research
trend that studies when a representation (or, more generally, an isometric action)
can be reconstructed from its orbit space together with some additional data, see
[18, 102, 118].
Chapter 5
Polar Foliations

In this chapter, we give a survey on some results from the theory of polar
foliations, also called singular Riemannian foliations with sections. This theory is a
generalization of the classical theory of adjoint actions (presented in Chap. 4), and
several results in this chapter are extensions of its results. Since our main goal is to
provide a flavor of this new field, to not lose sight of the big picture, we only give
some sketches of proofs without going into technical details.

5.1 Definitions and First Examples

We start by stating the definition of singular Riemannian foliation, see Molino [165].
Definition 5.1. Let F be a partition of a complete Riemannian manifold M into
connected immersed submanifolds, called leaves. We say that:
(i) F is singular if the module XF of smooth vector fields on M that are tangent
at each point to the corresponding leaf acts transitively on each leaf, in the sense
that, for each leaf L and each v ∈ Tp L, there exists X ∈ XF with X(p) = v;
(ii) F is transnormal if a geodesic that is perpendicular to a leaf at one point is
horizontal, i.e., remains perpendicular to all leaves it intersects.
If F satisfies (i), then it is called a singular foliation of M. If F satisfies (i) and (ii),
then it is called a singular Riemannian foliation of M.
Let F be a singular Riemannian foliation of a complete Riemannian manifold M.
A leaf L of F (and each point in L) is called regular if the dimension of L is
maximal; otherwise, it is called singular. If all the leaves of F are regular, then F
is called a Riemannian foliation. The codimension of a singular foliation is defined
as the codimension of its regular leaves.

© Springer International Publishing Switzerland 2015 109

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_5
110 5 Polar Foliations

Definition 5.2. Let F be a singular Riemannian foliation on a complete Rieman-

nian manifold M. Then F is a polar foliation, also called a singular Riemannian
foliation with sections, if for each regular point p, the set Σ = exp p (ν p L p ) is a
complete immersed submanifold that intersects all leaves orthogonally. In this case,
Σ is called a section.
Example 5.3. Using Killing vector fields, one can prove that the partition (3.4) of a
Riemannian manifold M into orbits of an isometric action is a singular Riemannian
foliation (recall Proposition 3.78). Moreover, the partition of M into orbits of a
polar action (see Definition 4.8) is a polar foliation. A foliation of a space form
by isoparametric submanifolds (see Definition 4.14) is also a polar foliation, and is
called an isoparametric foliation.
Terng and Thorbergsson [205] introduced the concept of equifocal submanifolds
with flat sections in symmetric spaces, generalizing the definition of isoparametric
submanifolds in Euclidean space. We now give a slightly more general definition of
equifocal submanifolds in Riemannian manifolds.
Definition 5.4. A connected immersed submanifold L of a complete Riemannian
manifold M is equifocal if it satisfies:
(i) The normal bundle ν (L) is flat;
(ii) L admits sections, i.e., for each p ∈ L, the set Σ = exp p (ν p L p ) is a complete
immersed totally geodesic submanifold;
(iii) For each p ∈ L, there exists a neighborhood U ⊂ L such that, for each parallel
normal field ξ along U, the derivative of the endpoint map ηξ : U → M,
ηξ (x) := expx (ξ ), has constant rank.
Terng and Thorbergsson [205] proved that any partition of a simply-connected
compact symmetric space M into submanifolds parallel to an equifocal submanifold
L with flat sections is a polar foliation.
As we prove in Theorem 5.26, regular leaves of polar foliations are equifocal
submanifolds. Conversely, a closed embedded equifocal submanifold is a leaf of
some polar foliation, see Remark 5.66.
We conclude this section with the following well-known fact about equifocal
submanifolds of the round sphere.
Proposition 5.5. Any equifocal submanifold of the round sphere Sn is an isopara-
metric submanifold of Sn .
Proof (Sketch). Let ξ be a unit normal parallel field along an equifocal submanifold
L of Sn . The proof is divided into two steps. We first establish a relation between
principal curvatures and the kernel of endpoint maps. Next, we use this relation
and equifocality of L to conclude that the set of principal curvatures with respect
to ξ at the point x is actually independent of x ∈ L, i.e., L is isoparametric, see
Definition 4.14.
5.2 Holonomy and Orbifolds 111

Claim 5.6. For any fixed x ∈ L, each principal curvature λi corresponds bijectively
to a radius ti (up to integer multiples of π ) by the equation λi = cotti , so that
ker dηti ξ ⊂ Tx L is the eigenspace of Sξ corresponding to the eigenvalue λi .
To prove the above, consider the geodesic γx (t) := expx (t ξ ) and a variation
α (s,t) = expβ (s) (t ξ ), where β is a curve in L with β (0) = x. We are interested in
the zeros of L-Jacobi fields J(t) = ∂∂s α (s,t)|s=0 along γx . Since the normal bundle
of L is flat, it is possible to prove that the L-Jacobi fields satisfy
+
J  (t) + J(t) = 0
(5.1)
J  (0) + Sξ J(0) = 0.

Let ei (t) be parallel normal fields along the geodesic α (0,t), such that Sξ (ei (0)) =
λi ei (0). Writing J(t) = ∑i ai (t)ei (t), we conclude that if J is a solution of (5.1), then
ai (t) = ai (0)(cost − λi sint). In particular, ai (ti ) = 0 if and only if ai (0) = 0 or if
λi = cotti .
Given one such principal curvature λi corresponding to a principal direction
ei ∈ Tx L, it follows from Claim 5.6 that ei ∈ ker dηti ξ (x). Since L is equifocal,
dim ker dηti ξ is constant. Thus, by Claim 5.6, λi is also a principal curvature (with
the same multiplicity) at any other point of L, concluding the proof. 

5.2 Holonomy and Orbifolds

In this section, we recall some preliminary facts about Riemannian foliations, most
of which can also be found in Molino [165]. In the particular case of partition of
the space into orbits of an proper isometric actions, some of the results and remarks
of this section were presented in Chap. 3; e.g., recall Proposition 3.78, Remark 3.79
and Exercise 3.81. We start with two equivalent characterizations of Riemannian
foliations.
Proposition 5.7. Let F be a foliation on a complete Riemannian manifold (M, g).
The following statements are equivalent:
(i) F is a Riemannian foliation;
(ii) For each q ∈ M, there exists a neighborhood U of q in M (called simple
open subset), a Riemannian manifold (σ , b) and a Riemannian submersion
f : (U, g) → (σ , b) such that the connected components of F ∩ U (called
plaques) are preimages of f ;
(iii) Let gT be the transverse metric, i.e., gT = N ∗ g where Np : Tp M → ν p L is the
orthogonal projection onto the normal space. Then the Lie derivative LX gT
vanishes for each X ∈ XF , see Definition 5.1.
112 5 Polar Foliations

With the above notion of plaques, we have the following:

Proposition 5.8. Let F be a (singular) foliation on a complete Riemannian
manifold (M, g). Then F is a (singular) Riemannian foliation if and only if its leaves
are locally equidistant, i.e., for each q ∈ M there exists a tubular neighborhood

Tubε (Pq ) = {x ∈ M : dist(x, Pq ) < ε }

of a plaque Pq of Lq , such that for x ∈ Tub(Pq ), each plaque Px contained in Tub(Pq )

is contained in the cylinder of radius dist(x, Pq ) and axis Pq .
An important property of Riemannian foliations is (Bott) equifocality. In order
to understand this concept, we need some preliminary definitions. A Bott or basic
connection ∇ of a foliation F is a connection on the normal bundle of the leaves,
with ∇X Y = [X,Y ]ν F whenever X ∈ XF and Y is a section of the normal bundle
ν F of the foliation. Here, (·)ν F denotes projection onto ν F . A normal foliated
vector field is a normal field parallel with respect to the Bott connection. If we
consider a local submersion f that describes the plaques of F in a neighborhood
of a point of L, then a normal foliated vector field is a normal projectable/basic
vector field with respect to f . In this context, Bott equifocality means that if ξ is
a normal parallel vector field (with respect to the Bott connection) along a curve
β : [0, 1] → L, then the curve t → expβ (t) (ξ ) is contained in the leaf Lexpβ (0) (ξ ) .
Remark 5.9. Bott equifocality also holds for singular Riemannian foliations, and
implies that one can reconstruct the (singular) foliation by taking all parallel
submanifolds of a (regular) leaf with trivial holonomy, see Alexandrino and
Töben [20].
The above allows us to introduce the concept of parallel translation (with respect
to the Bott connection) of horizontal segments of geodesic.
Definition 5.10. Let β : [a, b] → L be a piecewise smooth curve and γ : [0, 1] → M
be a segment of a horizontal geodesic such that γ (0) = β (a). Let ξ0 be a vector
of the normal space νβ (a) L, such that expγ (0) (ξ0 ) = γ (1), and let ξ : [a, b] → ν L
be the parallel translation of ξ0 with respect to the Bott connection along β . Define
β (γ ) := γ", where γ": [0, 1] → M is the geodesic segment s → γ"(s) = expβ (b) (s ξ (b)).
Due to equifocality of F , there is an alternative definition of holonomy map of
a Riemannian foliation.
Definition 5.11. Let β : [0, 1] → L be a piecewise smooth curve and let

Sβ (i) := expβ (i) (ξ ) : ξ ∈ νβ (i) L, ξ  < ε (5.2)

be a slice at β (i), for i = 0, 1, and ε > 0 small. The holonomy map is defined by

ϕ[β ] : Sβ (0) → Sβ (1) , ϕ[β ] (x) := ||β γ (r), (5.3)

5.2 Holonomy and Orbifolds 113

where γ : [0, r] → Sβ (0) is the minimal geodesic segment joining β (0) to x. Since the
Bott connection is locally flat, the parallel translation depends only on the homotopy
class fixing endpoints [β ].
Remark 5.12. The holonomy map of a foliation F is usually  defined
 as follows.
Consider a partition 0 = t0 < t1 < . . . < tk = 1 such that β (ti−1 ,ti ) is contained in
a simple open set Ui defined by a trivialization of F . Then {U1 , . . .Uk } is called
a chain of simple open sets. Sliding along the plaques in Uk , one can define a
diffeomorphism ϕk from an open neighborhood of the transverse manifold Sβ (tk−1 )
onto an open set of Sβ (tk ) . In the same way, sliding along the plaques Uk−1 , one
defines a diffeomorphism ϕk−1 from an open neighborhood of Sβ (tk−2 ) onto an open
neighborhood of Sβ (tk−1 ) . After some steps, we obtain a diffeomorphism ϕ1 from
an open neighborhood Sβ (t0 ) onto a neighborhood of Sβ (t1 ) . Finally, the germ of
the diffeomorphism ϕk ◦ ϕk−1 ◦ · · · ◦ ϕ1 is the holonomy map ϕ[β ] . It is possible to
prove that the latter does not depend on the chain of simple open sets used in its
construction, but only on the homotopy class fixing endpoints of β .
Remark 5.13. Note that, since the holonomy map depends only on the homotopy
class, there is a group homomorphism ϕ : π1 (L0 , x0 ) → Diffx0 (Sx0 ), from the funda-
mental group of the leaf L0 to the group of germs at x0 of local diffeomorphisms of
Sx0 . The image of this homomorphism is the holonomy group of L0 at x0 .
Remark 5.14. Consider a proper G-action μ on a manifold M, and the homoge-
neous foliation (3.4). Assume that G is connected and that all G-orbits are regular.
Using Remarks 3.60 and 5.12, it is possible to check that for each holonomy map
ϕ[β ] , there exists g such that μ (g, ·)|Sx0 = ϕ[β ] . In particular, the holonomy group
coincides with the image of the slice representation of Gx0 in Sx0 (see Remark 3.61).
We now mention some facts about pseudogroups and orbifolds, related to
Riemannian foliations. More details can be found in Salem [165, Appendix D] or
Moerdijk and Mrčun [163].
Definition 5.15. Let Σ be a Riemannian manifold, not necessarily connected.
A pseudogroup W of isometries of Σ is a collection of (local) isometries w : U → W ,
where U and V are open subsets of Σ , such that:
(i) If w ∈ W , then w−1 ∈ W ;  
(ii) If w : U → V and w:U  → V are in W , then w  ∩V ) → V
 ◦ w : w−1 (U  is in W ;
(iii) If w : U → V is in W , then its restriction to each open subset U ⊂ U is in W ;
(iv) If w : U → V is an isometry between open subsets of Σ that coincides in a
neighborhood of each point of U with an element of W , then w ∈ W .
Definition 5.16. Let A be a family of local isometries of Σ containing the identity.
The pseudogroup obtained by taking inverses of the elements of A, restrictions to
open subsets, compositions, and unions is called the pseudogroup generated by A.
114 5 Polar Foliations

Definition 5.17. Let W and Φ be pseudogroups of isometries of Riemannian

manifolds Σ and σ , respectively. An equivalence between W and Φ is a maximal
collection Ψ of isometries between open subsets of Σ and open subsets of σ such
that:
(i) The domains of the elements of Ψ form an open cover of Σ and the ranges of
the elements of Ψ form an open cover of σ ; 
(ii) If ψ ∈ Ψ , w ∈ W , and ϕ ∈ Φ , then ϕ ◦ ψ ◦ w ∈ Ψ .
(iii) If ψα , ψβ ∈ Ψ , w ∈ W , and ϕ ∈ Φ , then
   −1 
ψβ ◦ w ◦ ψα−1 ∈ Φ and ψα ◦ ϕ ◦ ψβ ∈ W.

Example 5.18. An important example of a Riemannian pseudogroup is the holon-

omy pseudogroup of a Riemannian foliation. Let F be a Riemannian foliation of
codimension k of a Riemannian manifold (M, g). Then F can be described by an
open cover {Ui } of M with Riemannian submersions fi : (Ui , g) → (σi , b), where σi
is a submanifold of dimension k, such that there are isometries wij : f i (Ui ∩ U j ) →

f j (U j ∩ Ui ) with f j = wij ◦ fi . The elements wij acting on Σ = σi generate a
pseudogroup of isometries of Σ , called the holonomy pseudogroup of F .
Definition 5.19. A k-dimensional Riemannian orbifold is an equivalence class
of pseudogroups W of isometries on a k-dimensional Riemannian manifold Σ ,
such that:
(i) The orbit space Σ /W is Hausdorff;
(ii) For each x ∈ Σ , there exists an open neighborhood U of x in Σ such that the
restriction of W to U is generated by a finite group of diffeomorphisms of U.
The orbit space Σ /W is the underlying topological space of the orbifold.
If, in addition, W is a group of isometries and Σ is connected, then Σ /W is called
a good Riemannian orbifold.
Remark 5.20. Alternatively, a good Riemannian orbifold Σ /W can be defined as
the orbit space of a discrete group of isometries W acting properly on a connected
Riemannian manifold Σ . It is possible to prove that if two good Riemannian
orbifolds are isometric, then they are equivalent. More generally, as suggested by
Lytchak, one can define a Riemannian orbifold as a metric space X where each point
x ∈ X has a neighborhood U isometric to a quotient M/Γ , where M is a Riemannian
manifold and Γ is a finite group of isometries, see Gorodski [101].
An important example of Riemannian orbifold is the space of leaves M/F ,
where M is a Riemannian manifold and F is a Riemannian foliation of M with
closed embedded leaves. Here, M/F is isomorphic (in fact, isometric) to Σ /W ,
where Σ and W were defined in Example 5.18. The following is a converse result.
Proposition 5.21. Every Riemannian orbifold Σ /W is the space of leaves of a
Riemannian foliation with compacts leaves.
5.2 Holonomy and Orbifolds 115

Proof (Sketch). Given a good Riemannian orbifold Σ /W , we describe a right

U(n)-action on a manifold M such that Σ /W = M/U(n). The remaining details
in the proof can be found in Moerdijk and Mrčun [163, Prop. 2.23].
The metric g(·, ·) of T Σ induces a Hermitian metric on the complex vector bundle
E := T Σ ⊗ C, given by the complexification of the tangent bundle T Σ . Let U(E p )

denote the set of unitary frames of E p . Then U(E) := p∈Σ U(E p ) is called the
unitary frame bundle of E, and is a principal bundle with fiber U(n) and base Σ .
The underlying free right action μ "1 : U(E) × U(n) → U(E), see Proposition 3.33, is
defined as μ"1 (ξ p , g) := ξ p ◦ g, where the unitary frame ξ p is identified with a unitary
linear map between Cn and E p .
Consider the action μ "2 : W × U(E) → U(E) given by μ "2 (w, ξ ) := dw · ξ for each
unitary frame ξ and w ∈ W . This is a free proper left action, hence M := U(E)/W
is a manifold. Since the actions μ "1 and μ"2 commute, the action μ "1 descends to a
proper right action μ : M × U(n) → M. One can then prove that Σ /W is equivalent
to the orbit space of this action, i.e.,

Σ /W = M/U(n). (5.4)

With the appropriate metric on M = U(E)/W , the equality in (5.4) can be interpreted
not only as an equivalence but also as an isometry between Riemannian orbifolds.
Let us briefly explain the construction of this metric. Note that if ξ p is an
orthogonal frame of Tp Σ , then ξ p is a unitary frame of E p . This fact can be used
to prove that the orthogonal frame bundle O(T Σ ) is contained in the unitary frame
bundle U(E), and, moreover, it is a subbundle of U(E). The Riemannian connection
on T Σ induces a linear connection H% on O(T Σ ).1 Using μ "1 , we can extend the
%
linear connection H on O(T Σ ) to a linear connection on U(E). Note that there
is an induced metric " g on U(E), for which the bundle projection π : U(E) →
Σ is a Riemannian submersion and the linear connection H % is orthogonal to
the orbits of μ" . Since isometries preserve parallel transport, we conclude that
  1
d(μ"2 )w H%x" = H% w and hence the distribution H% descends to a distribution
μ2 ) ("
(" x)
H on M = U(E)/W , transverse to the orbits of μ . We can define a metric on
H , i.e., a transverse metric, by setting g[p] ([X], [Y ]) := "
g p (X,Y ), where X,Y are
vectors tangent to H %p and [·] : U(E) → U(E)/W is the quotient map. Finally, one
can extend this transverse metric to a Riemannian metric on M, by requiring that H
be orthogonal to the orbits and averaging an arbitrary metric along the orbits. 

1 The linear connection on O(T Σ ) is defined as follows. For a fixed frame ξ p , set H %ξ as the
p
"  (0), where α
space of all vectors α " is the parallel transport of ξ p along a curve α : [0, 1] → Σ with
α (0) = p.
116 5 Polar Foliations

5.3 Surgery and Suspension of Homomorphisms

The examples of polar foliations presented in the first section have a symmetric
space as ambient. Simple ways to construct new polar foliations on nonsymmetric
spaces are to consider a polar foliation F with compact leaves on a manifold M
and change either the leaf metric locally, or the transverse metric on a tubular
neighborhood of a regular leaf L with trivial holonomy.2 After a generic perturbation
of its metric, M becomes nonsymmetric. In this section, we explain some techniques
that allow to construct more general examples of polar foliations, namely via surgery
and suspension of homomorphisms.
Let us first discuss how surgery can be used to construct polar foliations, see
Alexandrino and Töben [19]. Let Fi be a polar foliation with codimension k and
compact leaves on a complete Riemannian manifold Mi , for i = 1, 2. Suppose
that there exist regular leaves with trivial holonomy L1 ∈ F1 and L2 ∈ F2
that are diffeomorphic. Since Li has trivial holonomy, there exist trivializations
ψi : Tub3ε (Li ) → L1 × B3ε , where Tub3ε (Li ) is the tubular neighborhood of Li with
radius 3ε , and B3ε is a ball in the Euclidean k-dimensional space.
Define τ : (L1 × Bε ) \ (L1 × Bε /2 ) → (L1 × B2ε ) \ (L1 × Bε ) as the inversion in
the cylinder of radius ε and axis L1 . Define also φ : Tubε (L1 ) \ Tubε /2 (L1 ) →
Tub2ε (L2 ) \ Tubε (L2 ) by φ := ψ2−1 ◦ τ ◦ ψ1 . We now change the transverse metric of
Mi in the tubular neighborhoods of Li , such that the restriction of φ to each section
in Tubε (L1 ) \ Tubε /2 (L1 ) is an isometry. This can be done, for instance, by taking
1
g as the transverse metric on Tub2ε (Li ) and keeping the original transverse
x2 0
metric g outside Tub3ε (Li ), using partitions of unity with two appropriate functions.
Finally, define M as:
   
M := M \ Tubε /2 (L1 ) φ M \ Tubε (L2 ) .

This new manifold has a singular foliation F , and leaves are locally equidistant
with respect to the transverse metric that already exists. To conclude the construc-
tion, we only have to define a tangential metric to the leaves with partitions of unity.
.
Note that if Σ i is a section of Fi , then the connected sum Σ 1 #Σ 2 is a section of F
Let us now explore an example of polar foliation constructed with suspension of
homomorphisms, extracted from Alexandrino [6]. Other examples can be found in
Alexandrino [4]. We start by recalling the method of suspensions, see Molino [165,
p. 28–29, 96–97] for details. Let Q and V be Riemannian manifolds of dimension
p and n respectively, and consider a homomorphism ρ : π1 (Q, q0 ) → Iso(V ). Let
P" : Q" → Q be the projection of the universal covering of Q. There is a natural action
of π1 (Q, q0 ) on M=Q " ×V , given by
 
q, v) · [α ] := q" · [α ], ρ (α −1 ) · v ,
("

2 This is possible since the set of regular leaves is open and dense in M.
5.4 Differential and Geometric Aspects of Polar Foliations 117

where q" · [α ] denotes the deck transformation associated to [α ] applied to a point

" We denote by M the set of orbits of this action, and by Π : M
q" ∈ Q.  → M the
canonical projection. We claim that M is a smooth manifold. Indeed, given a simple
open neighborhood U j ⊂ Q, consider the bijection
 
Ψj : Π (P"−1 (U j ) ×V ) → U j ×V, Ψj (Π (" " q) × {v} .
q, v)) := P("
   
If Ui ∩ U j = 0/ is connected, we can see that Ψi ∩ Ψj−1 (q, v) = q, ρ ([α ]−1 )v for
a fixed [α ]. Hence, there exists a unique manifold structure on M for which Ψj are
local diffeomorphisms. Let

P : M → Q, P(Π (" " q).

q, v)) := P("

It follows that M is the total space of a fiber bundle, and P is its projection over the
base Q. In addition, the fiber is V and the structural group is the image of ρ .
Finally, define F := Π (Q, " v) , i.e., the projection of the trivial foliation defined
as the product of Q " with each v. It is possible to prove that this is a foliation
transverse to the fibers. Furthermore, this foliation is a Riemannian foliation whose
transverse metric coincides with the metric on V .
Example 5.22. In what follows, we construct a polar foliation such that the
intersection of a local section with the closure of a regular leaf is an orbit of an
action of a subgroup of isometries of the local section. This isometric action is not
polar. Therefore, there exists a polar foliation F such that the partition formed by
the closures of the leaves is a nonpolar singular Riemannian foliation.
Set V := R2 ×C × C and F %0 the singular foliation of codimension 5 on V , whose
leaves are products of points in C × C with circles in R2 centered at the origin. It
is easy to see that the foliation F%0 is polar. Let Q be the circle S1 and α ∈
/ Q be an
irrational number. Define the homomorphism
 
ρ : π1 (Q, q0 ) → Iso(V ), ρ (n) := (x, z1 , z2 ) → (x, einα · z1 , einα · z2 ) .
 
Finally, set F := Π Q "×F %0 . It turns out that F is a polar foliation and the
intersection of the section Π (0 × R × C × C) with the closure of a regular leaf is an
orbit of an isometric action on the section. This isometric action is not polar, since
the Hopf action is not polar (see Remark 3.90).

5.4 Differential and Geometric Aspects of Polar Foliations

In this section, we describe a few results from the theory of polar foliations. The
following was proposed by Palais and Terng [182, Rem 5.6.8]:
118 5 Polar Foliations

Palais-Terng Conjecture 5.23. Let G be an isometric action of a compact Lie

group on M. If the distribution of normal spaces to regular orbits is integrable,
then the G-action is polar.
This conjecture was proved by Heintze, Liu and Olmos [124]. In particular,
they proved that the set of regular points is dense in each section. A result in
Alexandrino [6] gives a positive answer to this conjecture in the case of singular
Riemannian foliations.
Theorem 5.24. Let F be a singular Riemannian foliation on a complete Rie-
mannian manifold M. If the distribution of normal spaces to the regular leaves is
integrable, then F is a polar foliation. In addition, the set of regular points is open
and dense in each section.
Proof (Sketch). Let p be a regular point. The first step is to verify that if γ is a
geodesic orthogonal to the regular leaf L p , then singular points are isolated along γ .
In the case where F is the partition of M into orbits of an isometric action, the
reader can check this claim using the slice representation, or Kleiner’s Lemma 3.70.
The second step in the proof is to ensure that, for each regular point x, there exists
a totally geodesic disk Dδ (x) orthogonal to the leaves. Its radius δ depends only on
the convex neighborhood centered at x, and not on the tubular neighborhood of a
plaque at x.
For the third step, note that we can extend Dδ (p) along a horizontal geodesic γ
by gluing totally geodesic disks along γ . Let Σr1 denote the extensions constructed
along every geodesic γ with length r1 such that γ  (0) ∈ Tp Dδ (p).
In the fourth step, one checks that Σ r1 is covered by a finite number of disks Di .
This implies that Σr1 is an immersed submanifold. Indeed, let D∗i be the lift of Di to
the Grassmannian bundle Grk TM. Since Di are totally geodesic, if D∗i and D∗j have a
common point, then they coincide in a neighborhood of this point. Thus, the union
Σr∗1 of D∗i ’s is an embedded submanifold in Grk TM that projects to the immersed
submanifold Σr1 .
Finally, considering rn → ∞ and applying the Hopf-Rinow Theorem 2.9, one can

check that the projection Σ of Σ ∗ := ri Σr∗i is a complete immersed totally geodesic
submanifold. 

Remark 5.25. Particular cases of the above result where proved by Molino and
Pierrot [166], Boualem [49] and Lytchak and Thorbergsson [156]. Szenthe [203]
also worked on this conjecture for isometric actions.
Henceforth, F denotes a polar foliation on a complete Riemannian manifold M.
The following result of Alexandrino [4] relates polar foliations to equifocal subman-
ifolds, see also Alexandrino and Töben [20].
Theorem 5.26. Let L be a regular leaf of a polar foliation F of a complete
Riemannian manifold M.
5.4 Differential and Geometric Aspects of Polar Foliations 119

(i) L is equifocal. In particular, the union of regular leaves that have trivial normal
holonomy is an open and dense set in M, provided that all leaves are compact;
(ii) Let β be a smooth curve on L and ξ be a parallel normal field to L along β .
Then the curve ηξ ◦ β is in a leaf of F ;
(iii) Suppose that L has trivial holonomy and let Ξ denote the set of all parallel
normal fields on L. Then F = {ηξ (L)}ξ ∈ Ξ .
Proof (Sketch). We first explain why the image ηξ (U) of the endpoint map is
contained in a leaf, for the appropriate basic vector ξ . Let x be a singular point
and x0 be a regular point in a small tubular neighborhood Tub(Px ). Let ξ be the
parallel normal vector field of the plaque Px0 with exp(ξx0 ) = x. Assume that the
geodesic γ (t) := exp(t ξx0 ) lies in Tub(Px ) for all t ∈ [0, 2], and x = γ (1) is the only
singular point along γ . Define α (s,t) := exp(t ξc(s) ) and J(t) = ∂∂s α (s,t)|s=0 , where
c(s) is a curve in Px0 with c(0) = x0 . We want to check that the Jacobi field J is
always tangential to the leaves, and hence that the curve α (·,t) is contained in Lγ (t) .
Since Σ is totally geodesic, we have that Tγ (t) Σ , as well as (Tγ (t) Σ )⊥ , are
families of parallel subspaces along γ which are invariant under R(γ̇ , ·)γ̇ . Note that
J(0), J  (0) ∈ (Tγ (0) Σ )⊥ . These facts and the Jabobi equation (2.5) imply that J is
orthogonal to Σ for 0 ≤ t ≤ 2.
Since the singular points are isolated in γ , we have that J(t) is tangent to the
leaves Lγ (t) for t = 1 and hence the curve α (·,t) is contained in Lγ (t) for t = 1.
It remains to prove that J(1) is tangent to Lγ (1) . Set
   
h(s,t) := dist α (s,t), Px − dist α (0,t), Px .

The above discussion implies that h(·,t) = 0 for t = 1 and hence, since h is
continuous, h(·, 1) = 0. This implies that J(1) is tangent to Lγ (1) , and α (·, 1) is
contained in Lγ (1) . As we know that α (·,t) ⊂ Lγ (t) , it follows that ηt ξ (U) ⊂ Lγ (t) .
Let us now explain why dηt ξ has constant rank for t close to 1. The classical
theory implies that, for t < 1, ηt ξ (U) is an open subset of Lγ (t) and ηt ξ : U →
ηt ξ (U) ⊂ Lγ (t) is a diffeomorphism. If t = 1, we have that the rank of dηξ is constant
on U, because the closest-point projection π : Tub(Px ) → Px is a submersion and
π |U = ηξ |U . Finally, consider t > 1. The fact that π is a submersion and π |U = ηξ |U
imply that ker dηt ξ is tangent to the slice. On the other hand, ηt ξ is a diffeomorphism
when restricted to the slice. In fact, one can check that the foliation restricted to a
small slice is invariant under geodesic involutions (i.e., invariant under the maps
expx ◦(− id) ◦ exp−1 x ). Thus, ηt ξ |U is a local diffeomorphism. It is not difficult to see
that it is bijective, and hence a diffeomorphism.3

one can prove that dηt ξ has constant rank for t close to 1 using ideas presented in
3 Alternatively,

the proof of Theorem 5.63.

120 5 Polar Foliations

We have sketched the proof of the result for small ξ ’s. The general case can be
proved by composing endpoint maps, i.e., ηξ := ηξn ◦ . . . ◦ ηξ0 , where ξi are the
appropriate small vectors. 

Note that this result guarantees that, given a regular leaf L with trivial holonomy,
we can reconstruct F by taking submanifolds parallel to L.
In order to state the next result, we recall the concepts of slice and local section.
Given x ∈ M, let Tub(Px ) be a tubular neighborhood of a plaque Px that contains x.
The connected component of expx (ν Px ) ∩ Tub(Px ) that contains x is called a slice
at x, and is usually denoted Sx . A local section σ (centered at x) of a section Σ is a
connected component Tub(Px ) ∩ Σ that contains x.
Let us recall some results about the local structure of F , in particular about
the structure of the set of singular points in a local section. The next result of
Alexandrino [4] is a generalization of Theorem 4.19.
Slice Theorem 5.27. Let F be a polar foliation of a complete Riemannian mani-
fold M, and let Sx be a slice at a singular point x.

(i) Sx = σ ∈Λ (x) σ , where Λ (x) is the set of local sections σ centered at x;
(ii) Sy ⊂ Sx for all y ∈ Sx ;
(iii) F |Sx is a polar foliation of Sx with the metric induced from M;
(iv) F |Sx is diffeomorphic to an isoparametric foliation of an open subset of Rn ,
where n = dim Sx .

Proof (Sketch). In order to prove (i), first note that Sx ⊃ σ ∈Λ (x) σ . In fact, let
σ ∈ Λ (x). Since σ is totally geodesic and orthogonal to Lx , we have that σ ⊂ Sx .
Let σ be a local section. In what follows, we show that:

If σ ∩ Sx = 0,
/ then x ∈ σ and σ ⊂ Sx . (5.5)

Since Sx ⊃ σ ∈Λ (x) σ , it suffices to check that x ∈ σ . We claim that if the
intersection of a local section σ with a slice Sx contains a regular point, then x ∈ σ .
Indeed, let p ∈ Sx ∩ σ be a regular point and γ be a geodesic segment orthogonal to
Lx , that joins x to p. Since F is a polar foliation, γ is orthogonal to L p and belongs
to the local section σ , hence x ∈ σ . Let us now consider the remaining case, in
which σ ∩ Sx may not contain regular points. The previous claim and the fact that
the regular points are dense in σ imply that there exists a slice Sx with x ∈ Lx , such
that x ∈ σ and σ ⊂ Sx. Since Sx ∩ Sx = 0,
/ we conclude that x = x.

In order to verify that Sx ⊂ σ ∈Λ (x) σ , let y ∈ Sx and σ be a local section that

contains y. It follows from (5.5) that x ∈ σ , and hence y ∈ σ ∈Λ (x) σ .
As for (ii), let y ∈ Sx and σ be a local section that contains y. It follows from (5.5)
that σ ⊂ Sx . Since, by (i), Sy is a union of local sections that contain y, we conclude
that Sy ⊂ Sx . Item (iii) follows easily from (ii).
Finally let us sketch the proof of item (iv). Set Fx := F |Sx . From the previous
items, Fx is a polar foliation on Sx . Consider a geodesic homothetic transformation
hλ with respect to x, i.e., hλ (γ (ε )) = γ (λ ε ) for all horizontal unit speed geodesics
γ starting at x. Due to equifocality, Fx is invariant under hλ . We also have that the
5.4 Differential and Geometric Aspects of Polar Foliations 121

sections of Fx are invariant under hλ . Therefore, Fx is a polar foliation with respect

to the metric gλ := λ12 h∗λ g. On the other hand, as λ 0, the metric gλ converges
smoothly to the metric g0 of the tangent space4 at x to the slice Sx . Thus, Fx is a
polar foliation with respect to the flat metric g0 , and hence Fx is a polar foliation
on Rn , after the appropriate identification.5
Recall that the regular leaves of Fx are equifocal. A standard computation
shows that equifocal submanifolds in Euclidean spaces are isoparametric, cf.
Proposition 5.5. Therefore, each regular leaf L is isoparametric. Note that, by
equifocality, the isoparametric foliation of L coincides with Fx , concluding the
proof. 

From Theorem 5.27 (iv), it is not difficult to derive the following consequence,
see Alexandrino [4].
Corollary 5.28. Let σ be a local section. The set of singular points of F contained
in σ is a finite union of totally geodesic hypersurfaces. These hypersurfaces are
diffeomorphic to focal hyperplanes contained in a section of an isoparametric
foliation of an open set of Euclidean space.
We call the set of singular points of F contained in σ the singular stratification
of the local section σ . Let Mreg denote the set of regular points in M. A Weyl chamber
of a local section σ is the closure in σ of a connected component of Mreg ∩ σ , cf.
Definition 4.26. It is possible to prove that a Weyl chamber of a local section is a
convex set.
Theorem 5.26 allows us to define the singular holonomy map, which is very
useful to study F . The following is also in Alexandrino [4].
Proposition 5.29. Let F be a polar foliation on a complete Riemannian manifold
M, and x0 , x1 ∈ Lx . Let β : [0, 1] → L p be a smooth curve contained in a regular leaf
L p , such that β (i) ∈ Sxi , where Sxi is the slice at xi for i = 0, 1. Let σi be a local
section contained in Sxi that contains β (i) and xi , i = 0, 1. Finally, let [β ] denote
the homotopy class of β . Then there exists an isometry ϕ[β ] : U0 → U1 , where the
domain U0 and range U1 are contained in σ0 and σ1 respectively, satisfying:
(i) x0 ∈ U0 ;
(ii) ϕ[β ] (y) ∈ Ly for each y ∈ U0 ;
(iii) dϕ[β ] ξ (0) = ξ (1), where ξ (s) is a parallel normal field along β (s).
Such an isometry is called the singular holonomy map along β . We remark that,
in the definition of the singular holonomy map, singular points can be contained in
the domain U0 . If the domain U0 and the range U1 are sufficiently small, and Lx is
regular, then the singular holonomy map coincides with the usual holonomy map
along β .

4 The metric g0 is such that expx becomes an isometry.

by composing the inverse of the normal exponential map expx : Bε (0) ∩ Tx Sx → Sx with an
5 I.e.,

isometry between Tx Sx and Rn .

122 5 Polar Foliations

Theorem 5.27 establishes a relation between polar foliations and isoparametric

foliations. Similarly, as in the theory of isoparametric submanifolds, it is natural to
ask whether a (generalized) Weyl group action can be defined on σ . The following
definitions and results deal with this question.
Definition 5.30. The pseudosubgroup W (σ ) generated by all singular holonomy
maps ϕ[β ] such that β (0) and β (1) belong to the same local section σ is called the
generalized Weyl pseudogroup of σ . Define W (Σ ) for a section Σ analogously.
Proposition 5.31. Let σ be a local section. The reflections across the hypersurfaces
of the singular stratification of the local section σ leave F |σ invariant. Moreover,
these reflections are elements of W (σ ).
Remark 5.32. Compare the above result, in Alexandrino [4], with Theorem 4.37 (i)
and Remark 4.38.
Using suspensions, one can construct an example of polar foliation such that
W (σ ) is larger than the pseudogroup generated by the reflections across the
hypersurfaces of the singular stratification of σ . On the other hand, a sufficient
condition to ensure that both pseudogroups coincide is that the leaves of F have
trivial normal holonomy and are compact. Thus, it is natural to ask under which
conditions the normal holonomy of regular leaves is trivial. The next result, in
Alexandrino and Töben [19], Alexandrino [7, 9], and Lytchak [154], deals with this
question.
Theorem 5.33. Let F be a polar foliation on a simply-connected manifold M.
(i) Each regular leaf has trivial holonomy;
(ii) The leaves of F are closed and embedded;
(iii) M/F is a simply-connected good Coxeter orbifold;6
(iv) Let Σ be a section of F and Π : M → M/F the canonical projection. Denote
by Ω a connected component of the set of regular points in Σ . Then Π : Ω →
Mreg /F and Π : Ω → M/F are homeomorphisms, where Mreg denotes the set
of regular points in M, and Ω is convex.7
Remark 5.34. Compare the above result with Theorems 4.36 and 4.37.
Proof (Sketch). In what follows we sketch the proof following Alexandrino [9].
Let πΣ : Σ → Σ be the Riemannian universal covering, p ∈ Σ a regular point,
and p ∈ Σ be such that πΣ ( p) = p. Let w ∈ W (Σ ) and consider a curve c : [0, 1] →
Σ with c(0) = p and c(1) = w(p). There exists an isometry w [c] on Σ such that
πΣ ◦ w
[c] = w ◦ πΣ and w [c] ( p) = c(1), where c is the horizontal lift of c starting at p.
The isometry w [c] defined above is called a lift of w along c, and clearly depends on

6 I.e.,
a quotient of a simply-connected Riemannian manifold by a reflection group in the classical
sense, i.e., a discrete group of isometries generated by reflections acting properly and effectively.
for any two points p and q in Ω , every minimal geodesic segment between p and q lies
7 I.e.,

entirely in Ω .
5.4 Differential and Geometric Aspects of Polar Foliations 123

the homotopy class [c]. Let W  be the group generated by all w,

 called the lifted Weyl
group. It is proved in [9] that there exists a surjective homomorphism

 /Γ,
ρ : π1 (M) → W (5.6)

where Γ is a normal subgroup, called lifted reflection group, that is constructed   as

follows. Let H be a wall of the Weyl chamber and H  ⊂ Σ be such that πΣ H  = H.
Each reflection in H  is an element of W  , and Γ is the group generated by all these
reflections. It follows from Davis [76, Lemma 1.1] that Γ is a reflection group in
the classical sense. In particular, a Weyl chamber C of Γ is a fundamental domain
of the action of Γ, see [76, Thm 4.1]. Along this proof, M is simply-connected, and
hence (5.6) implies that W  = Γ.
We first prove (i), i.e., show that each regular leaf has trivial holonomy. If this
was not true, there would be a regular point p ∈ Σ and w ∈ W (Σ ) such that w(p) = p
and w = id. Let δ : [0, 1] → Σ be an arbitrary curve starting at δ (0) = p. Define a
 
lift of w by w  δ(1) := w  ◦ δ (1), i.e., the lift of w along the constant curve c = p.
 
Note that w  p = p and w  = id. Since Γ = W  is a reflection group, the only points

that can be fixed by elements of Γ are singular, i.e., points that are mapped by πΣ to
singular points of Σ . Therefore, w(  p) = p, a contradiction.
Now we prove (ii), i.e., show that the leaves of F are closed and embedded.
Due to equifocality, it suffices to prove that the leaves are embedded. If L p is not
embedded, then there would be a sequence {wn } ⊂ W (Σ ) such that wn (p) → p.
 
Define w n δ(1) := (cn  ∗ wn ◦ δ )(1), where cn are short curves contained in a
simply-connected neighborhood of p and ∗ denotes concatenation of curves. Then
one can check that w n ( p) → p. Since Γ = W  is a reflection group in the classical
sense, {wn ( p)} cannot converge to p, giving the desired contradiction.
We continue by proving (iii), i.e, that M/F is a simply-connected good Coxeter
orbifold. For each loop c in Σ , consider the isometry id  [c] , i.e., the lift of the identity
along c. In this way, we identify π1 (Σ ) with a subgroup of W  . It is not difficult to
check that this is a normal subgroup and W (Σ ) = W /π1 (Σ ). Since Σ = Σ/π1 (Σ ),

we have that Σ /W (Σ ) = Σ/W  . As W  = Γ is a reflection group, we conclude that
M/F = Σ /W (Σ ) is the good Coxeter orbifold Σ/Γ. In order to check that Σ/Γ
is simply-connected, consider a homotopy cs in Σ between a loop c contained in
C = Σ/Γ ⊂ Σ and the constant curve c(0). Then ρ ( cs ) is an homotopy in C = Σ/Γ
  
between c and c(0), where ρ : Σ → Σ /Γ is the quotient map, so the conclusion
follows from the fact that Σ is simply-connected.
Finally, we prove item (iv), i.e., that Ω is a fundamental domain of the action of
W (Σ ) on Σ . Up to composing with reflections across the walls (that are elements
of W (Σ )), we may assume that each orbit intersects Ω at least once. Suppose, by
contradiction, that there exist two points x, y ∈ Ω that belong to the same leaf. Then
there is an isometry w ∈ W (Σ ) such that w(x) = y. Joining x and y with a curve c
contained in Ω and constructing an isometry w  that maps a point x to a point
[c] ∈ W
124 5 Polar Foliations

y, both inside the fundamental domain C of the action of Γ = W

 on Σ, we obtain a
contradiction. As we proved that Ω = Σ /W (Σ ) = C = Σ/Γ, we conclude that Ω is
convex. 

The following is an important consequence of Theorem 5.33.

Corollary 5.35. Orbits of a polar action on a simply-connected complete Rieman-
nian manifold are closed and embedded. Moreover, there are no exceptional orbits.
Remark 5.36. As remarked in [9], Theorem 5.33 also holds for a more general class
of foliations, namely infinitesimally polar foliations. In particular, a proper isometric
action on a simply-connected manifold whose orbit space is a good orbifold has no
exceptional orbits. Apart from polar actions, other examples of isometrics actions
whose orbit spaces are good orbifolds are given by variationally complete actions,
see Lytchack [154].
Remark 5.37. As remarked in [14], the results in [9] and Theorem 5.33 can be used
to provide elementary proofs of the existence of closed geodesics on leaf spaces
of polar foliations on compact simply-connected spaces. More generally, we have
the following result if F is a singular Riemannian foliation with closed embedded
leaves on a simply-connected Riemannian manifold M and M/F is a compact
orbifold, then M/F admits a nontrivial closed geodesic.
We now discuss very recent results in the theory of polar foliations and actions.
The following result of Lytchak [155] gives a positive answer to an important
conjecture in polar actions on symmetric spaces:
Theorem 5.38. Let M be a simply-connected, irreducible symmetric space with
sec ≥ 0, and let F be a polar foliation on M of codimension at least 3. Then either
all the leaves of F are points, or F is hyperpolar, or the symmetric space has rank
one.
Similar techniques to the ones used by Lytchak [155] were independently
employed by Fang, Grove and Thorbergsson [84], who recently proved the follow-
ing classification result.
Theorem 5.39. A polar action on a compact simply-connected manifold with
sec > 0 and of cohomogeneity ≥2 is equivariantly diffeomorphic to a polar action
on a compact rank one symmetric space (CROSS).8
Remark 5.40. Theorem 6.30, due to Verdiani [218] and Grove, Wilking and Ziller
[114], complements the above result with regard to manifolds with sec > 0 and
cohomogeneity 1 (which are automatically polar). In this case, there are several
manifolds different from a CROSS that support an invariant metric with sec > 0.

8 For details on compact rank one symmetric spaces, see Example 6.10.
5.4 Differential and Geometric Aspects of Polar Foliations 125

Remark 5.41. As proved in Mucha [169], there exist examples of hyperpolar

actions with cohomogeneity ≥2 on locally irreducible compact homogeneous
spaces with sec ≥ 0 which are not homeomorphic to symmetric spaces.
Grove and Ziller [118] associate to a polar foliation without exceptional orbits
(that they call Coxeter polar action), a graph of groups G(C), where C is a chamber
of a fixed section. The following fundamental reconstruction result is proved using
this notion.
Theorem 5.42. A Coxeter polar action (M, G) is determined by its polar data
(C, G(C)). More precisely, there is a canonical construction of a Coxeter polar G-
manifold M(D) from the polar data D = (C, G(C)), and if D is the polar data for a
Coxeter polar G-manifold M, then M(D) is equivariantly diffeomorphic to M.
This result was recently used by Gozzi [103] to classify polar actions in
manifolds of dimension ≤5. In Chap. 6, we discuss the above result in more details
in the particular case of cohomogeneity one actions, see Remark 6.38.
Since we have seen several examples of theorems about polar foliations and
actions in spaces with sec ≥ 0, is natural to wonder what happens in the dual case
of sec ≤ 0. The following was obtained by Töben [214].
Theorem 5.43. Let F be a polar foliation on a Hadamard manifold. Then, up to
diffeomorphisms, F is the product foliation of a compact isoparametric foliation
and the trivial foliation of Rn with only one leaf.
Remark 5.44. It is a classical fact that there are no nontrivial Killing vector
fields in compact negatively curved spaces, see Petersen [183, p. 191]. Töben
[214] proved that there are no polar foliations with closed leaves on such spaces,
and Lytchak [153] extended this result proving that there are not even singular
Riemannian foliations.
To conclude this section, we discuss a result that describes the behavior of polar
foliations whose leaves are not embedded. Molino [165] proved that if M is compact,
then the closure of the leaves of a (regular) Riemannian foliation forms a partition of
M that is a singular Riemannian foliation. He also proved that leaf closures are orbits
of a locally constant sheaf of germs of (transversal) Killing fields. Furthermore, he
proposed the following conjecture in the same book:
Molino Conjecture 5.45. Let F be a singular Riemannian foliation of a compact
Riemannian manifold M. Then the closures of the leaves of F forms a partition of
M that is also a singular Riemannian foliation.
Molino [165, Thm 6.2] proved that the closure of the leaves is a transnormal
system, however, it still remains an open problem to prove that F is a singular
foliation. The next result, in Alexandrino [6], gives a positive answer to the Molino
Conjecture 5.45 if F is polar.
126 5 Polar Foliations

Theorem 5.46. Let F be a polar foliation on a complete Riemannian mani-

fold M.

(i) The closures L L∈F of the leaves of F form a partition of M which is a
singular Riemannian foliation;
(ii) Each x ∈ M is contained in a homogeneous submanifold Ox (possibly 0-
dimensional). If we fix a local section σ that contains x, then Ox is a connected
component of an orbit of the closure of the Weyl pseudogroup of σ ;
(iii) If x ∈ L, then a neighborhood of x in L is the product of the homogeneous
submanifold Ox with plaques with the same dimension of the plaque Px ;
(iv) If x is a singular point and T is the intersection of the slice Sx with the singular
stratum that contains x, then the normal connection of T in Sx is flat;
(v) Let x be a singular point and T as in (iv). Consider v a parallel normal vector
field along T , y ∈ T and z = expy (v). Then Oz = ηv (Oy ).
Proof (Sketch). Let x and σ be as in the statement. It follows from the definition
of Weyl pseudogroups that the intersection of the leaf Lx with the local section σ
coincides with the orbit of x under the Weyl pseudogroup. This observation and
results of Salem [165, Appendix D] then imply that Ox is a connected component
of an orbit of the closure of the Weyl pseudogroup of σ . The rest of (ii), i.e., that Ox
does not depend on σ , can be proved using that the Weyl pseudogroups of different
local sections are conjugate to each other.
Note that the partition of σ into orbits of the closure of the Weyl pseudogroup is a
singular Riemannian foliation. This fact and (ii) imply that {L}L∈F is a transnormal
system. Item (iv) can be proved using that walls of the singular stratification of
the local section σ are totally geodesic in σ . In addition, the singular stratification
is diffeomorphic to a product of simplicial cones with a subspace, which can be
identified with the intersection of σ with the stratum that contains x, called the local
minimal stratum. The above fact is also the key to prove (v). Roughly speaking, the
orbit Oz must be equidistant to the walls of the singular stratification of the local
section and therefore parallel to the local minimal stratum. Finally, combining (iv)
and (v), one can prove that the partition {L}L∈F is singular. 

The above theorem was illustrated in Example 5.22.

Remark 5.47. Lytchak [154] recently proved that polar foliations on simply-con-
nected complete Riemannian manifolds have closed and embedded leaves, see also
Theorem 5.33. Therefore, the above result is of interest only in the case where the
ambient space is not simply-connected.

5.5 Transnormal and Isoparametric Maps

In the previous section, polar foliations were presented as a natural generalization

of isoparametric foliations on Euclidean spaces, in particular of orbits of adjoint
actions. In this section, we consider a different generalization, related to partitions
5.5 Transnormal and Isoparametric Maps 127

by preimages of certain special maps, called transnormal and isoparametric maps.

This approach was already considered by Cartan [62–65], Harle [122], Terng [204]
and Wang [224].
We start by discussing the definition of transnormal functions and isoparametric
functions; for historical context, see Thorbergsson [208, 209].
Definition 5.48. A smooth function f : M → R on a complete Riemannian mani-
fold M is called transnormal if

grad f 2 = b ◦ f (5.7)

for some C2 function b : f (M) → R. Moreover, if

Δ f = a◦ f (5.8)

for some continuous function a : f (M) → R, then f is called isoparametric.9

Geometrically, (5.7) implies that the level sets of f are equidistant, i.e., the
corresponding partition of M is trasnormal (recall Definition 5.1), while (5.8)
implies that the regular level sets have constant mean curvature:
Exercise 5.49. Let f : M → R be a transnormal function satisfying (5.7). Let I =
(c0 , c1 ) be an interval without singular values of f , and set N := f −1 (I).
 
(i) Show that f : N → I, 1b dx2 is a Riemannian submersion;
(ii) Prove that integral curves of grad f , parametrized to have unit speed, are
geodesics;
(iii) Let γ (t) be one such geodesic and set c(t) := f (γ (t)). Denote by Ht the mean
curvature of the hypersurface f −1 (c(t)) ⊂ N at γ (t). Prove that:

Δ f (γ (t)) = ( f ◦ γ ) (t) − ( f ◦ γ ) (t) Ht , (5.9)

with the conventions that Δ f = tr(Hess f ) is the (nonpositive) Laplacian and

the mean curvature of a hypersurface is given by (2.9). Conclude that the
hypersurfaces f −1 (c(t)) have constant mean curvature if f is isoparametric.
It is natural to wonder which of the above geometric properties of regular level
sets remain true for critical level sets. This is answered by the following result of
Wang [224].

9 According to Thorbergsson [209], the term isoparametric hypersurface was introduced by Levi-
Civita and grad f 2 and Δ f were called the differential parameters of f . We stress that, in modern
references, the term isoparametric hypersurface is used for level sets of isoparametric functions,
see Remark 5.58.
128 5 Polar Foliations

Theorem 5.50. Let M be a connected complete Riemannian manifold and f : M →

R be a transnormal function.
(i) The critical level sets of f are submanifolds of M;
(ii) The focal manifolds of regular level sets are components of critical level sets
of f ;
(iii) There are at most two critical level sets, and each regular level set is a
geometric tube over the latter.
Remark 5.51. It also follows from the proof of the above result that the regular level
sets are equifocal. Miyaoka [161] has recently improved this by showing that if M
admits a transnormal function, then either M is diffeomorphic to a vector bundle
over a submanifold, or M is diffeomorphic to a union of disk bundles over two
submanifolds. This is analogous to a property of orbits of a cohomogeneity one
action, see Proposition 6.33.
Remark 5.52. Interesting aspects of submanifold geometry of the level sets of iso-
parametric functions on compact manifolds were studied by Ni [174], Wang [224],
and Ge and Tang [95, 96]. For example, it is proved that focal submanifolds of
an isoparametric function are minimal, and that there exists at least one minimal
hypersurface level set. These results are also analogous to properties of orbits of a
cohomogeneity one action, see Remark 6.43, and can be alternatively deduced as an
application of the mean curvature flow on polar foliations, see [17].
Let us now illustrate the above discussion of isoparametric functions with some
standard examples.
Exercise 5.53. Consider the Euclidean space Rn+1 = {(x1 , . . . , xn+1 )}. Let

f (x) = x12 + · · · + x2k+1 ,

with 0 ≤ k ≤ n. Check that f : Rn+1 → R is an isoparametric function, such that:

(i) grad f 2 = 4 f 2 ;
(ii) Δ f = 2(k + 1); √
(iii) The level set f −1 (c) = Sk ( c) × Rn−k , c > 0, has principal curvatures λ1 = 1
c
and λ2 = 0, with multiplicities k and n − k respectively.
Exercise 5.54 (). Consider the unit round sphere Sn+1 = {x21 + · · · + x2n+2 = 1}.
Let

f(x) := x21 + · · · + x2k+1 ,

with 0 ≤ k ≤ n, and set f := f|Sn+1 . Check that f : Sn+1 → R is an isoparametric

function, such that:
(i) grad f 2 = 4 f − 4 f 2 ;
(ii) Δ f = 2(k + 1) − 2(n + 2) f ;
5.5 Transnormal and Isoparametric Maps 129

√ √
(iii) The level set f −1 (c) = Sk ( c) × Sn−k ( 1 − c), 0 < c < 1, has principal
curvatures λ1 = √1−c 2 and λ2 = √−c 2 , with multiplicities k and n − k
c−c c−c
respectively.
Remark 5.55. Let us make some geometric observations, in the particular case
k = 1 and n = 2, about the foliation described in Exercise 5.54. This foliation
of S3 ⊂ R4 by level sets of f (x) = x21 + x22 is studied in detail in Example 6.48,
from the cohomogeneity one viewpoint. Here, we note that the critical level sets
f −1 (1) and f −1 (0) are respectively the great circles in S3 obtained by intersecting S3
with the linear subspaces R2 = {(x1 , x2 , 0, 0)} and R2 = {(0, 0, x3 , x4 )}. The points
(1, 0, 0, 0) ∈ f −1 (1) and (0, 0, 1, 0) ∈ f −1 (0) are joined by the unit speed geodesic
segment γ (t) = (cost, 0, sint, 0), 0 ≤ t ≤ π2 , in S3 ⊂ R4 , which is perpendicular
to f −1 (c) for each 0 ≤ c ≤ 1 and is hence a section for this polar foliation. The
value of f along γ (t) is c(t) := f (γ (t)) = cos2 t, whose preimage is the torus
f −1 (c(t)) = S1 (cost) × S1 (sint), with principal curvatures λ1 = tant and λ2 =
− cott. Observe that this also allows to compute its mean curvature Ht = λ1 + λ2 =
tant − cott from (5.9), using Exercise 5.54 (ii), and Ht vanishes only when t = π4 ,
see Remark 6.49.
All of the above observations have direct analogues for general k and n, regarding
the foliation of Sn+1 given by f −1 (c(t)) = Sk (cost) × Sn−k (sint), 0 ≤ t ≤ π2 . The
singular leaves Sk and Sn−k correspond respectively to t = 0 and t = π2 , and are the
intersection of Sn+1 ⊂ Rn+2 with each of the summands in the orthogonal direct
sum decomposition Rn+2 = Rk+1 ⊕ Rn−k+1 , see Exercise 6.50.
Exercise 5.56 (). Consider the hyperbolic space H n+1 = {−x12 + · · · +x2n+2 = −1}
given by the hyperboloid model where the Riemannian metric is induced by the
Lorentz metric gL = −dx21 + · · · + dx2n+2 on Rn+2 . Let

f(x) := −x21 + · · · + xk+1

2
,

with 2 ≤ k ≤ n, and set f := f|H n+1 . Check that f : H n+1 → R is an isoparametric

function, such that:
(i) grad f 2 = 4 f + 4 f 2 ;
(ii) Δ f = 2(k + 1) + 2(n + 2) f ; √ √
(iii) The level set f −1 (−c) = H k ( c) × Sn−k ( c − 1), c > 1, has principal curva-
tures λ1 = √c−1 and λ2 = √ c , with multiplicities k and n − k respectively.
c2 −c c2 −c

Remark 5.57. Similarly to Remark 5.55, consider the particular case k = 2 and
n = 3 in Exercise 5.56, which gives a foliation of H 4 ⊂ (R5 , gL ) by level sets
of f (x) = −x21 + x22 + x23 . Note that the critical level set f −1 (−1) is the totally
geodesic H 2 in H 3 obtained by intersecting H 3 with the linear subspace R3 =
{(x1 , x2 , x3 , 0, 0)}. The unit speed geodesic γ (t) = (cosht, 0, 0, sinht, 0), t ≥ 0, in
H 4 ⊂ R5 is perpendicular to f −1 (−c) for each c ≥ 1 and is hence a section
for this polar foliation. The value of f along γ (t) is f (γ (t)) = − cosh2 t, so that
130 5 Polar Foliations

setting c(t) := cosh2 t, t ≥ 0, the preimage f −1 (−c(t)) = H 2 (cosht) × S1 (sinht)

has principal curvatures λ1 = tanht and λ2 = cotht. Observe that this also allows
to compute its mean curvature Ht = λ1 + λ2 = tanht + cotht from (5.9), using
Exercise 5.56 (ii).
All of the above observations have direct analogues for general k and n, regarding
the foliation of H n+1 given by f −1 (−c(t)) = H k (cosht) × Sn−k (sinht), t ≥ 0. The
singular leaf H k corresponds to t = 0, and is the intersection of H n+1 ⊂ (Rn+2 , gL )
with the first summand in the direct sum decomposition Rn+2 = Rk+1 ⊕ Rn−k+1 .
These are also orbits of a cohomogeneity one action on H n+1 , see Remark 6.51.
Remark 5.58. Note that regular level sets of the isoparametric functions on space
forms considered in Exercises 5.53, 5.54 and 5.56 are isoparametric hypersurfaces
in the classical sense, i.e., hypersurfaces with constant principal curvatures. Car-
tan [62] proved that hypersurfaces in space forms are isoparametric if and only if
they are regular level sets of isoparametric functions, see also Remark 4.16. On the
other hand, Wang [223] found an example of isoparametric function f on a complex
projective space, whose regular level sets do not have constant principal curvatures.
As pointed out above, in modern references, the term isoparametric submanifold is
used for this more general concept of regular level sets of an isoparametric function.
Wang [224] also remarked that level sets of transnormal functions on Rn+1 and
S n+1 are isoparametric hypersurfaces. This follows from equifocality of the regular
level sets and Proposition 5.5. In general, however, transnormal functions are not
necessarily isoparametric. Consider, for instance, f : Rn+1 → R given by f (x) =
cos(x). A direct computation gives grad f 2 = 1 − f 2 , so f is transnormal; and
Δ f = − f + (1 − n) sin(x)/x, so f is not isoparametric. Moreover, note that the
level sets of f are the same as those of g(x) = x2 , which is isoparametric (see
Example 5.53).
Remark 5.59. The above is not true for hyperbolic spaces, i.e., there exist transnor-
mal functions on H n+1 whose level sets are not isoparametric hypersurfaces. To see
this, let N be a nonisoparametric hypersurface in H n+1 , with principal curvatures
between −1 and 1. Let Nt be the parallel hypersurface to N at distance t, which
is well-defined for all t > 0, since N does not have focal points. The function
f : H n+1 → R defined by f (Nt ) = t is transnormal, giving the desired example.
We now go in the direction of generalizing the above discussion to higher
codimension, i.e., looking for a notion of map H : M → Rq whose regular level
sets are leaves of a polar foliation of codimension q. Inspired by Definition 5.48
and Exercise 5.49, we can infer some properties that this map should fulfill. If
M0 is the principal stratum (i.e., the set of leaves with trivial holonomy), then
H|M0 : M0 → (Rq , g) should be a Riemannian submersion, where  g is an appropriate
metric on Rq . Like in Exercise 5.49, it should be possible to write  g in terms of the
gradients of the components hi of the map H. Since the regular level sets should
be leaves of a polar foliations, the distribution given by their normal spaces ought
to be integrable; in particular, near the regular level sets, the set spanned by grad hi
should be integrable. Therefore, we should expect some condition on the brackets
5.5 Transnormal and Isoparametric Maps 131

[grad hi , grad h j ], by the Frobenius Theorem A.19. Altogether, it is natural to expect

that our candidate H : M → Rq to a transnormal map is an integrable Riemannian
submersion with singularities. In addition, if we want H to be isoparametric, i.e.,
the mean curvature vector of regular level sets to be basic and have constant length,
then some condition similar to (5.8) should hold. The above discussion leads to
the following definition of transnormal and isoparametric maps, introduced by
Harle [122] and Terng [204]:
Definition 5.60. Let M n+q be a complete Riemannian manifold. A smooth map
H = (h1 , . . . , hq ) : M n+q → Rq is called transnormal if
(i) H has a regular value;
(ii) For each regular value c, there exist a neighborhood V of H −1 (c) in M and
smooth functions bij on H(V ) such that, for every x ∈ V ,
# $
grad hi (x), grad h j (x) = (bij ◦ H)(x); (5.10)

(iii) There exists a sufficiently small neighborhood of each regular level set such
that [grad hi , grad h j ] is a linear combination of grad h1 , . . . , grad hq , with
coefficients that are functions of H, for all i and j.
Moreover, a transnormal map H is said to be isoparametric if V can be chosen to be
the whole manifold M and Δ hi = ai ◦ H, where ai are smooth functions.
Exercise 5.61. Show that a map H = (h1 , . . . , hq ) : M n+q → Rq is transnormal
if and only if it has a regular value, and for each regular value c, there exists
a neighborhood V of H −1 (c) in M such that H|V : V → H(V ) is an integrable
Riemannian submersion, where the metric of H(V ) is given by the inverse matrix of
B = (bij ).
Exercise 5.62. Let H : SU(3) → C ∼ = R2 be the trace map H(A) := tr(A). Consider
the foliation F given by the partition of SU(3) into the orbits of the conjugation
action. Check that leaves of F are preimages of H and that H is a transnormal map.
A submanifold in a space form is isoparametric if and only if it is the preimage
of some isoparametric map, see [182]. Therefore, it is natural to ask under which
conditions this equivalence holds in the general case of Riemannian manifolds. In
the rest of this section we deal with this question.
First, we present a generalization of Theorem 5.50, due to Alexandrino [3],
for analytic transnormal maps on analytic Riemannian manifolds (e.g., symmetric
spaces).
Theorem 5.63. Let H : M → Rq be an analytic transnormal map on a real analytic
complete Riemannian manifold M. Let c be a regular value and L ⊂ H −1 (c) be a
connected component of H −1 (c). Denote by Ξ the set of all parallel normal fields
along L.
132 5 Polar Foliations

(i) Fc,L = {ηξ (L)}ξ ∈Ξ is a polar foliation with embedded leaves;

(ii) For each regular value c", the connected components of H −1 (" c ) are equifocal
submanifolds and leaves of Fc,L ;
(iii) Fc,L is independent of the choice of c and L, i.e., for another regular value c
and connected component  L ⊂ H −1 (
c), Fc,L = Fc,L .
Proof (Sketch). Let us briefly sketch the proofs of (i) and (ii), from [3]. Consider
the submanifold L defined in the theorem. The first step is to construct a candidate
section Σ , which is a totally geodesic submanifold orthogonal to L and to the regular
level sets of H. This can be done using the analyticity of the map H and arguments
presented in the proof of Theorem 5.24.
The second step is to check that L is an equifocal submanifold. In other words,
for a parallel normal field ξ along L and p ∈ L, we have to show that the rank of
dηξ is constant on a neighborhood U of p in L. The main idea is to show that the
focal distances and the multiplicity of focal points are constant, using a Morse index
argument.
More precisely, set g(x,t) := hi ◦ ηt ξ (p) − hi ◦ ηt ξ (x). The map H is an integrable
Riemannian submersion on a neighborhood of the regular level set, and ξ is a
projectable field of this submersion. Therefore g(x,t) = 0 for every small t and
for all x ∈ L. Since g(x,t) is analytic, g(x,t) = 0 for all t. Thus, we proved
that ηt ξ (L) ⊂ H −1 (d). Using the fact that H is a Riemannian submersion with
singularities, one can also prove the following:
Claim 5.64. Let Lt := ηt ξ (L) and d be the value such that Lt ⊂ H −1 (d). Then
ηt ξ : L → Lt is a diffeomorphism when d is a regular value.
Due to Claim 5.64, we can suppose without loss of generality that p belongs to a
convex normal neighborhood V of ηξ (p). For a small neighborhood U ⊂ L p of p, we
can define, for each x ∈ U, the geodesic segment γx (t) := expx (t ξ ), t ∈ [−ε , 1 + ε ].
Note that γx (t) ∈ V if U and ε > 0 are sufficiently small.
Using the Jacobi equation and the fact that Σ is totally geodesic, it is possible to
prove that the focal points of U along γx (t) are of tangential type, i.e., come from
variations of basic vector fields. We then have

m(γx ) = ∑ dim ker dηti (x)ξ (x), (5.11)

i

where ti (x) denotes the distance between the focal points on γx and x, and m(γx )
denotes the number of focal points on γx counted with multiplicities. Since H is
analytic, the critical points of H on γx are isolated and, without loss of generality,
we may assume that γ p (1) is the only critical point of H on γ p . It follows from
the above claims that the map ηt ξ may only fail to be a diffeomorphism if t = 1.
Therefore, (5.11) simplifies to m(γx ) = dim ker dηξ (x). On the other hand, from
the Morse Index Theorem, m(γx ) ≥ m(γ p ), for all x in some neighborhood of p in
L. Thus, since dim ker dηξ (x) ≤ dim ker dηξ (p), we conclude that dim ker dηξ is
constant on U, completing the second step of the proof.
5.5 Transnormal and Isoparametric Maps 133

The third and last step is to prove that the partition F := {ηξ (L)}ξ ∈Ξ is a polar
foliation. It is interesting to note that in this part we do not need analyticity of M or
the existence of H, but only that L is equifocal and Σ has some special properties,
see Remark 5.66 below.
Since the derivative of the endpoint map has constant rank, ηξ (U) is an
embedded manifold (for small U). These small submanifolds are the plaques of
the foliation. In what follows, we claim that part of the Slice Theorem holds for
these plaques:
Claim 5.65. Let Sx be the slice at x ∈ ηξ (U). Then
&
Sx = σ,
σ ∈Λ (x)

where Λ (x) is the set of local sections that contain x.

Claim 5.65 almost completes the proof that F is a polar foliation. In fact, for
 and for x ∈ η  (U)
another ηξ (U)  ∩ Sx , it can be used to show that Sx ⊂ Sx . From
ξ
 is contained in the boundary of a tube over
this, one can concludes that η  (U) ξ
 and ηξ (U ) are equidistant, which
ηξ (U). In other words, we conclude that ηξ (U)
implies that F is a transnormal system. Finally, Sx ⊂ Sx implies that ν (Sx ) ⊂ ν (Sx),
hence F is a singular foliation. 

Remark 5.66. As remarked in [5], the above proof can be slightly adapted to give
an alternative proof of a result of Töben [213], that can be reformulated as follows.
Let L be a closed embedded equifocal submanifold, and suppose that the subset
of regular points10 of each section Σ is open and dense. Define Ξ as the set of all
parallel normal fields along L. Then F := {ηξ (L)}ξ ∈ Ξ is a polar foliation.
It is also known (see Carter and West [68] and Terng [204]) that given an
isoparametric submanifold L of Rn+k , one can construct a polynomial isoparametric
map from Rn+k into Rk that has L as a level set. Therefore, it is natural to look
for conditions under which a polar foliation can be described with preimages of
transnormal maps. It follows from Theorem 5.27 and from the classical theory of
isoparametric foliations that this is always locally true, as stated in the following
observation of Alexandrino [4].
Proposition 5.67. Let F be a polar foliation on a complete Riemannian manifold
M. Each p ∈ M has a neighborhood U such that the plaques of F ∩U are preimages
of a transnormal map.
Since there are examples of polar foliations with nontrivial holonomy (see
Alexandrino [4]), there are examples of polar foliations that are not given as

10 Here, a point of a section is called regular if there exists only one local section that contains p,
i.e., if two local sections σ and σ contain p, then they have the same germ at p.
134 5 Polar Foliations

preimages of transnormal maps. The next result, by Alexandrino and Gorodski [13]
and Alexandrino [7], gives sufficient conditions under which a polar foliation can
be described by a transnormal map. Thus, it can be considered as a converse to
Theorem 5.63.
Theorem 5.68. Let F be a polar foliation on a complete simply-connected Rie-
mannian manifold M. Assume that leaves of F are closed and embedded, and F
admits a flat section of dimension n. Then the leaves of F are given by the level sets
of a transnormal map H : M → Rn .
Remark 5.69. Heintze, Liu and Olmos [124] proved the above result with the addi-
tional assumption that M is a simply-connected symmetric space of compact type.
If we allow the image of H to be contained in a general manifold instead of Rn ,
we have the following more general result, that can be found in [10, 12].
Theorem 5.70. Let F be a singular Riemannian foliation on a simply-connected
complete Riemannian manifold M. Assume that M/F is a good orbifold Σ /W , e.g.,
if F is polar. Then there exists a smooth map H : M → Σ such that the leaves of F
coincide with the level sets of H, and such that H(M) is isometric to M/F .
Remark 5.71. The above result was first proved for the case of polar foliations in
Alexandrino [10], and as remarked in Alexandrino, Briquet and Töben [12], one can
easily use pseudogroups to obtain the above formulation.
Proof (Sketch). For the sake of brevity, we sketch the proof when F is a polar foli-
ation. In what follows, we use the notation introduced in the proof of Theorem 5.33.
The first step is to construct a Γ-equivariant map G : Σ → Σ with special properties.
More precisely, let Σ be the universal Riemannian covering space of a section Σ
and Γ be the lifted reflection group. Fix a chamber C ⊂ Σ and denote by π : Σ →
Σ/Γ = C the canonical projection. Then there exists a smooth Γ-equivariant map
G : Σ → Σ such that
(i) G|C : C → C is a homeomorphism;
(ii) If K is a connected component of a set of points with the same orbit types for
the Γ-action on Σ, then the restriction G|K is a diffeomorphism;
(iii) The composition G ◦ π : Σ → Σ is smooth.
An easy example that illustrates the above claim is given by Σ = R, Γ = {± id}, C =
[0, ∞), G(x) := h(|x|) |x|x
(x = 0) and G(0) = 0, for an appropriate choice of function
h. Let us outline the proof of this claim, without going into details. First, one defines
a new metric  g such that normal slices to the strata are flat and totally geodesic. Then
the desired map G can be defined as the composition of Γ-equivariant maps Fi that
are constructed using the exponential map with respect to  g. Due to the properties
of g, it is possible to verify that G is smooth.
We now return to the proof of the main result. If follows from (5.6) that, if π1 (M)
is trivial, then W = Γ and W (Σ ) = Γ/π1 (Σ ). This fact and the existence of the map
5.6 Perspectives 135

G imply that there exists a smooth W (Σ )-invariant map H : Σ → Σ such that the
following diagram commutes:

where πΣ : Σ → Σ is the Riemannian covering map. Due to a result of Alexandrino

and Gorodski [13] concerning extensions of basic functions and forms, H admits a
smooth extension H : M → Σ that turns out to be the desired smooth F -invariant
map. 

For the particular case of polar actions, the above theorem can be reformulated
as follows.
Corollary 5.72. Consider a polar action on a simply-connected complete Rieman-
nian manifold M. Then there exists an isoparametric map H : M → Σ , where Σ is
a section, such that the orbits of the action coincide with the level sets of H. In
addition, H(M) is a fundamental domain for the action of the Weyl group on Σ .
Remark 5.73. Recently Radeschi [189] used Clifford systems to construct infinitely
many new examples of nonhomogenous (and nonpolar) singular Riemannian
foliations on spheres S2l−1 , that are preimages of maps H : S2l−1 → Rm+1 .

5.6 Perspectives

In this last section, we survey on some other results regarding singular Riemannian
foliations, and future research perspectives in the area.
We start by briefly discussing a few others classes of singular Riemannian folia-
tions. Lytchak and Thorbergsson [156] introduced the class of singular Riemannian
foliation without horizontal conjugate points, in order to generalize variationally
complete isometric actions. A singular Riemannian foliation F is said to be without
horizontally conjugate points, or simply variationally complete, if for every leaf
L, every L-Jacobi field tangential to a leaf other than L is vertical. They proved
that a singular Riemannian foliation without horizontal conjugate points in a
complete Riemannian manifold with sec ≥ 0 admits flat sections, i.e., are hyperpolar
foliations. Subsequently, the same authors proved in [157] that the quotient space of
a variationally complete group action is a good Riemannian orbifold. This result was
also proved to hold for singular Riemannian foliations without horizontal conjugate
points.
In the same paper, the concept of infinitesimally polar foliation is introduced.
A singular Riemannian foliation F is called infinitesimally polar if the intersection
136 5 Polar Foliations

of F with each slice is diffeomorphic to a polar foliation on an Euclidean space,

i.e., an isoparametric foliation. In fact, this diffeomorphism is given as composition
of the exponential map with a linear map. It is also proved in [157] that F is an
infinitesimally polar foliation if and only if, for each point x ∈ M, there exists a
neighborhood U of x such that the leaf space of the restricted foliation F |U is an
orbifold. Since the leaf space of an isoparametric foliation is a Coxeter orbifold, they
concluded that the leaf space U/(F |U ) is an orbifold if and only if it is a Coxeter
orbifold. They also proved that being infinitesimally polar is equivalent to having
locally bounded transverse curvature.
Typical examples of infinitesimally polar foliations are polar foliations (see [4]),
singular Riemannian foliations without horizontal conjugate points (see [156, 157])
and singular Riemannian foliations of codimension 1 or 2 (see [157]). In particular,
a partition into orbits of an isometric action is infinitesimally polar if the action is
polar, or variationally complete, or has cohomogeneity ≤2.
Lytchak [154] continued to study infinitesimally foliations and generalized
the blow up introduced by Töben [213], proving that a singular Riemannian
foliation admits a resolution preserving the transverse geometry if and only if it
is infinitesimally polar. He also deduced that singular Riemannian foliations on
simply-connected manifolds that either have sections or no horizontal conjugate
points have closed leaves.
We now discuss a few results about general singular Riemannian foliations. An
important conjecture regarding open nonnegatively curved spaces has been recently
proved by Wilking [230], using Riemannian foliations. More precisely, the dual
foliation of a singular Riemannian foliation is used to show that the metric projection
onto the soul, also called Sharafutdinov projection, is smooth. In the same paper,
Wilking proved that the dual foliation to a singular Riemannian foliation F of a
manifold with sec > 0 has only one leaf. In other words, any two points can be
joined by a piecewise smooth horizontal curve. Wilking’s dualization was also used
in the above mentioned work of Lytchak [155]; recall Theorem 5.38.
An important question in the theory of singular Riemannian foliations is the
classification of all such objects in the Euclidean space or the sphere. In spheres,
Riemannian foliations were classified by Grove and Gromoll [108] when leaves
have codimension ≤3, and singular Riemannian foliations with leaves of dimension
≤3 were classified by Radeschi [188].
As mentioned above, the concept of equifocality is also valid for singular
Riemannian foliations. In fact, Alexandrino and Töben [20] proved that the regular
leaves of a singular Riemannian foliation are Bott equifocal, i.e., the endpoint map
of a normal foliated vector field has constant rank. This implies that it is possible
to reconstruct the singular foliation by taking all parallel submanifolds to a regular
leaf with trivial holonomy. In addition, the endpoint map of a normal foliated vector
field on a leaf with trivial holonomy is a covering map.
Equifocality is a very important concept. It was used in Alexandrino [8] to prove
that each singular Riemannian foliation F on a compact manifold M can be desin-
gularized to a Riemannian foliation F %of M " after successive blow ups, generalizing
a result of Molino [164]. It was remarked by Lytchak that if the desingularization
5.6 Perspectives 137

preserves transverse geometry, then the foliation should be an infinitesimal foliation.

One sees in [8] that part of the transverse geometry is preserved. Using this fact, it
was proved in [8] that if the leaves of F are compact, then for each small ε > 0,
the regular foliation F% can be chosen so that the desingularization map ρ" : M "→M
that projects leaves of F % to leaves of F induces an ε -isometry between M/ " F % and
" %
M/F . This implies that M/F is a Gromov-Hausdorff limit of a sequence Mε /F " %ε
 
of Riemannian orbifolds, that is, distGH M %ε , M/F < ε . It actually gives more
"ε /F
information, since the desingularization induces a map ρε : M "ε /F%ε → M/F that is
11
smooth, i.e., pull-back of basic smooth functions are basic smooth functions. We
recall that, in general, the ε -isometries used in Gromov-Haudsdorff convergence are
not necessarily continuous.
As proved by Alexandrino and Radeschi [16], the desingularization turns out
to be an important tool to study the smooth structure of leaf spaces. In fact,
desingularization is used in [16] to prove smoothness of isometric flows in orbit
spaces. This result implies that isometric actions on compact orbit spaces are
smooth, indicating the possibility of developing a theory of actions on quotient
spaces, in the spirit of Chap. 3. In the same paper, smoothness of isometric flows
in the quotient was used to give a positive answer to the Molino Conjecture 5.45 in
the case of orbit-like foliations, which is the class of singular Riemannian foliations
originally studied by Molino [165]. For more information on the smooth structure of
leaf spaces see Alexandrino and Lytchak [15] and Alexandrino and Radeschi [18],
where an analogue of the Myers-Steenrod Theorem 2.12 for the leaf spaces is
proved.
Finally, we mention that recent progress regarding tautness of singular foliations
was made by Wiesendorf [228]. Among other results, it is shown that tautness of a
singular Riemannian foliation is a property of the leaf space.
Further details on singular Riemannian foliations, especially regarding nonpolar
foliations, can be found in Alexandrino, Briquet and Töben [12], while more on
(regular) Riemannian foliations and Riemannian submersions can be found in the
books of Gromoll and Walschap [105] and Molino [165].

11 Here, a basic function is a function constant along the leaves.

Chapter 6
Low Cohomogeneity Actions and Positive
Curvature

In this chapter, we study special types of isometric group actions on compact

Riemannian manifolds, namely those with low cohomogeneity. The cohomogeneity
of an action is the codimension of its principal orbits, and can also be regarded as
the dimension of its orbit space. Low cohomogeneity is an indication that there are
few orbit types, and that the original space has many symmetries. In this situation,
it is possible to study many geometric features that are not at reach in the general
case. Throughout this chapter, we emphasize connections between the geometry and
topology of manifolds with low cohomogeneity stemming from curvature positivity
conditions, such as positive (sec > 0) and nonnegative (sec ≥ 0) sectional curvature.
After briefly introducing a deformation technique that is used throughout
this chapter, we start with the case of cohomogeneity zero, that corresponds to
homogeneous spaces. We give a quick introduction with examples, stating the
classification of positively curved closed simply-connected homogeneous spaces.
We then proceed to cohomogeneity one, studying the basic structure and exploring
it via many examples (ranging from compact rank one symmetric spaces to exotic
Kervaire spheres). Finally, in the last section, we give a survey of how manifolds
with low cohomogeneity can be used to better understand the notions of positive
and nonnegative sectional curvature, in the spirit of the so-called Grove program.

6.1 Cheeger Deformation

This section contains some preliminary material on a metric deformation technique

that is a very important tool in the remainder of this chapter. This deformation
consists of rescaling a Riemannian metric in the directions tangent to the orbits
of an isometric action, generalizing the construction of Berger spheres (see Exam-
ple 6.16). This method was developed by Cheeger [70] to construct metrics of
nonnegative sectional curvature on the connected sum of two compact rank one

© Springer International Publishing Switzerland 2015 139

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1_6
140 6 Low Cohomogeneity Actions and Positive Curvature

symmetric spaces, and was later systematically studied by Müter in his PhD the-
sis [172]. A summary of the results in [172] is provided by Ziller [234, 236]. In what
follows, we discuss some of these key results. Since its inception, this technique has
proved of foundational importance and continues to influence research in the area,
see [34–36, 115, 184, 192, 194].
Let (M, g) be a Riemannian manifold and G a compact Lie group acting
isometrically on M. Fix a bi-invariant metric Q on G, and endow M × G with the
product metric g ⊕ 1t Q . Note that there is a free isometric G-action on M × G
given by:

h · (p, g) := (h p, h g), p ∈ M, g, h ∈ G. (6.1)

The orbit space of this action is easily seen to be diffeomorphic to M. Namely,

consider the principal G-bundle G → G → {e} whose base is a point. Using the
G-action on M, we may construct the associated bundle M ×G G to this principal
G-bundle (see Theorem 3.51), which is precisely the orbit space of (6.1). Since
M ×G G is a fiber bundle whose fiber is M and base is a point, we have the desired
diffeomorphism, cf. Exercise 3.56. The quotient map can be explicitly computed to
be the submersion:

ρ : M × G → M, ρ (p, g) = g−1 p. (6.2)

Since the G-action (6.1) is free and isometric,by Exercise

 3.81, there is a unique
Riemannian metric gt on M such that ρ : M × G, g ⊕ 1t Q → (M, gt ) is a
Riemannian submersion. Note that the original G-action  on (M,g) is also
isometric
on (M, gt ) for t > 0. Indeed, the isometric G-action on M × G, g ⊕ 1t Q given by
 
k ∗ (p, g) := p, g k−1 , p ∈ M, g, k ∈ G (6.3)

commutes with (6.1), and hence descends to an isometric G-action on the cor-
responding orbit space (M, gt ), see Exercise 3.40. Observe that k ρ (p, g) = k g−1
p = ρ (k ∗ (p, g)), hence the G-action induced by (6.3) is the original G-action on M.
The 1-parameter family of metrics gt on M varies smoothly with t, and extends
smoothly to t = 0, with g0 = g, see Proposition 6.3. Thus, gt , t ≥ 0, is a deformation
of g by other G-invariant metrics on M, called the Cheeger deformation of g. One
of the main reasons why Cheeger deformations are useful is that, in addition to
preserving the isometric G-action, they preserve the curvature condition sec ≥ 0.
Proposition 6.1. If secg0 ≥ 0, then secgt ≥ 0, for all t ≥ 0.
Proof. From Proposition 2.26, we have  that (G,
 Q) has sec ≥ 0. Thus, if secg0 ≥ 0,
then M × G with the product metric g ⊕ 1t Q also has sec ≥ 0. From the Gray-
O’Neill formula (2.14), it follows that also secgt ≥ 0, for all t ≥ 0. 

Remark 6.2. Although Cheeger deformations preserve sec ≥ 0, they might not
preserve other lower curvature bounds such as sec ≥ k > 0, see Example 6.5.
6.1 Cheeger Deformation 141

In order to have more geometric intuition on Cheeger deformations, we represent

gt in terms of g, to analyze how the metrics evolve, following the formalism of
Müter [172] and Ziller [234]. For each p ∈ M, denote by g p the Lie algebra of the
isotropy G p . Fix the Q-orthogonal splitting g = g p ⊕ m p , and identify m p with the
tangent space Tp G(p) to the G-orbit through p via action fields, i.e., we identify
each X ∈ m p with

X p∗ = d
ds exp(sX)ps=0 ∈ Tp G(p),

recall (3.11). This determines a g-orthogonal splitting Tp M = V p ⊕ H p , as described

in (3.19) and (3.20), in vertical and horizontal spaces, respectively

V p := Tp G(p) = {X p∗ : X ∈ m p } and H p := {v ∈ Tp M : g(v, V p ) = 0}. (6.4)

It is not difficult to check that this g-orthogonal splitting is also gt -orthogonal

for all t, that is, vertical (respectively horizontal) directions of g are also vertical
(respectively horizontal) for gt , for all t > 0. For each t ≥ 0, there is a unique Q-
symmetric automorphism Pt : m p → m p such that

Q(Pt (X),Y ) = gt (X p∗ ,Yp∗ ), X,Y ∈ m p . (6.5)

In addition, let Ct : Tp M → Tp M be the g-symmetric automorphism that relates the

original metric g and its Cheeger deformation gt , i.e., such that

g(Ct (X),Y ) = gt (X,Y ), X,Y ∈ Tp M. (6.6)

We now give an explicit formula for these automorphisms in terms of P0 .

Proposition 6.3. The symmetric automorphisms Pt : m p → m p and Ct : Tp M →
Tp M for t ≥ 0 are determined by P0 in the following way:

Pt (X) = (P0−1 + t id)−1 (X) = P0 (id +tP0 )−1 (X ), X ∈ mp,

 ∗ (6.7)
Ct (X) = (id +tP0 )−1 (Xm ) p + X H , X ∈ Tp M,

where X V and X H are the vertical and horizontal parts of X ∈ Tp M respectively,

and Xm is the unique vector in m p such that (Xm )∗p = X V .
Proof. It is easy to see that the derivative dρ(p,e) : Tp M × g → Tp M of (6.2) satisfies

dρ(p,e) (X ∗ ,Y ) = X p∗ −Yp∗ , X,Y ∈ g. (6.8)

142 6 Low Cohomogeneity Actions and Positive Curvature

The vertical space1 at (p, e) for ρ is formed ∗

∗
 by the pairs (Z , Z) ∈ ∗Tp M × g. Thus,
the vector (X ,Y ) ∈ Tp M × g is g ⊕ t Q -orthogonal to all pairs (Z , Z), and hence
1

horizontal for ρ , if and only if for all Z ∈ g,

 
0 = g(X ∗ , Z ∗ ) + 1t Q(Y, Z) = Q(P0 Xm , Z) + Q( 1t Y, Z) = Q P0 Xm + 1t Y, Z ,

that is, Y = −tP0 Xm . Here, for X ∈ g, we denote by Xm the component of X in

m p . In particular, the horizontal lift with respect to the Riemannian submersion ρ of
X ∗ ∈ Tp M is given by (W ∗ , −tP0 Wm ) ∈ Tp M × g where W ∈ g satisfies the equation
X ∗ = dρ(p,e) (W ∗ , −tP0 Wm ). Using (6.8), this becomes
 ∗
X p∗ = Wp∗ + (tP0 Wm )∗p = (id +tP0 )Wm p , (6.9)

to which W = (id −1
+tP0 ) X−1∈ m
∗ provides a solution.Therefore, the horizontal lift
of X ∈ Tp M is (id +tP0 ) X , −tP0 (id +tP0 )−1 X . Notice that, since
∗

 −1
P0 (id +tP0 )−1 = P0 P0 (P0−1 + t id) = P0 (P0−1 + t id)−1 P0−1 = (P0−1 + t id)−1 ,

the horizontal lift of X ∗ ∈ Tp M can be rewritten as:

 ∗ 
P0−1 (P0−1 + t id)−1 X p
, −t(P0−1 + t id)−1 X ∈ Tp M × g. (6.10)

Rewriting it in this way is convenient for the computations in the remainder of this
proof, see (6.12). Analogously, the horizontal lift of a general vector V ∈ Tp M that
has vertical and horizontal parts V = V V +V H = (Vm )∗p +VpH ∈ Tp M, is given by
 ∗ 
P0−1 (P0−1 + t id)−1Vm p
+VpH , −t(P0−1 + t id)−1Vm ∈ Tp M × g, (6.11)

since it is horizontal for ρ and mapped to V by dρ(p,e) .

  
Since ρ : M × G, g ⊕ 1t Q → (M, gt ) is a Riemannian submersion, the square
length of X ∗ ∈ Tp M with respectto gt equals the square length of its horizontal
lift (6.10) with respect to g ⊕ 1t Q . Thus, we have that for all X ∈ m,
 ∗  ∗ 
gt (Xp∗ , X p∗ ) = g P0−1 (P0−1 + t id)−1 X p , P0−1 (P0−1 + t id)−1 X p
 
+ 1t Q t(P0−1 + t id)−1 X,t(P0−1 + t id)−1 X
 
= Q (P0−1 + t id)−1 X, P0−1 (P0−1 + t id)−1 X
 
+ tQ (P0−1 + t id)−1 X, (P0−1 + t id)−1 X

1 Notethat vertical/horizontal spaces for the submersion ρ : M × G → M are subspaces of Tp M × g,

while vertical/horizontal spaces for the G-action on M are subspaces of Tp M.
6.1 Cheeger Deformation 143

 
= Q (P0−1 + t id)−1 X, P0−1 (P0−1 + t id)−1 X + t(P0−1 + t id)−1 X
 
= Q (P0−1 + t id)−1 X, (P0−1 + t id)(P0−1 + t id)−1 X
 
= Q (P0−1 + t id)−1 X, X , (6.12)

which, according to the definition (6.5), proves that Pt = (P0−1 +t id)−1 . The formula
for Ct follows immediately from the above and (6.6). 

Remark 6.4. Notice that the horizontal lift (6.11) of V ∈ Tp M is quite complicated,
 ∗
while the horizontal lift of Ct−1V = (id +tP0 )Vm p +V H is equal to (V, −tP0Vm ),
which is much simpler to use for computations.
The above gives insight on how the geometry of gt changes with t. If P0 has
eigenvalues λi , then Ct has eigenvalues 1+t1 λi corresponding to the vertical directions
and eigenvalues 1 in the horizontal directions. In other words, as t grows, the metric
gt shrinks in the direction of the orbits and remains unchanged in the orthogonal
directions. Note that the speed in which the orbits shrink may vary with the orbit.
An easy consequence of the above observations is that (M, gt ) converges in the
Gromov-Hausdorff sense to the orbit space M/G as t ! +∞. In fact, horizontal
directions remain unchanged, while vertical directions collapse. This observation is
useful to study the geometry of orbit spaces of isometric actions, since it provides
a natural approximation of such objects by smooth Riemannian manifolds, cf.
Remark 3.80.
Let us informally describe two concrete examples of Cheeger deformations
before proceeding with the theory. All of the assertions below can be made precise
using the formulas in Exercise 6.62.
Example 6.5. Consider the round sphere (S2 , g) with the rotation action by S1 . The
orbit space is the closed interval [0, π ], and (S2 , gt ) can be visualized as spheres that
get “skinnier” as t gets large, since the parallels (orbits of the rotation action) are
being shrunk. For t " 1, (S2 , gt ) looks like a cylinder of height [0, π ] and very small
radius, with spherical caps at the ends, i.e., near the poles. All such manifolds have
positive curvature, but the curvature becomes arbitrarily large near the poles and
arbitrarily small everywhere else.
As a second example, consider the flat plane R2 with the rotation action by S1 .
The orbit space is a half line [0, +∞), and (R2 , gt ), t > 0, are positively curved
surfaces that, as t ! +∞, get asymptotically close to a cylinder with shrinking
radius. Intuitively, one can picture R2 sitting in R3 as a horizontal plane {(x, y, 0)}
and the evolution (R2 , gt ) is “folding” the plane up onto itself, getting arbitrarily
close to the positive z-axis {(0, 0, z) : z ≥ 0}.
As the last examples suggests, besides remaining nonnegative, curvature may
sometimes increase during a Cheeger deformation. Let us analyze how the curvature
evolves in more precise terms. For this, it is convenient to define a reparametrization
144 6 Low Cohomogeneity Actions and Positive Curvature

of the Grassmannian bundle Gr2 TM of planes on M, using the map Ct defined

in (6.6) and computed in (6.7). Given a plane σ = span{X,Y }, set

Ct−1 (σ ) := span Ct−1 X,Ct−1Y .

This provides a 1-parameter family of bundle automorphisms of Gr2 TM induced

by Ct−1 , called Cheeger reparametrization. As explained by Ziller [234], the crucial
observation of Müter [172] is that, to analyze
 the evolution of secgt , it is much
more convenient to study secgt Ct−1 (σ ) instead of secgt (σ ), cf. Remark 6.4. The
following result of Müter [172, Satz 4.9] (see also [234, Cor 1.4]) summarizes how
the sectional curvature of g0 -flat planes evolves for t > 0.
Proposition 6.6. Let gt be the Cheeger deformation of g0 , with secg0 ≥ 0. Given
a plane σ = span{X,Y } ⊂ Tp M with secg0 (σ ) = 0, consider the unnormalized gt -
sectional curvature of Ct−1 (σ ), given by
 
kC (t) := gt Rt (Ct−1 X,Ct−1Y )Ct−1 X,Ct−1Y
, ,2  
= ,Ct−1 X ∧Ct−1Y ,g secgt Ct−1 (σ ) ,
t

where Rt is the curvature tensor of gt . Then kC (0) ≥ 0, and if the equality kC (0) = 0
holds, we have:
(i) If [P0 Xm , P0 Ym ] = 0, then kC (0) = 0, kC (0) > 0 and kC (t) > 0 for all t > 0;
(ii) If [P0 Xm , P0 Ym ] = 0, then kC (t) = 0 for all t > 0.
In particular, if secg0 (σ ) = 0 and [P0 Xm , P0 Ym ]= 0, i.e., span{P0 Xm , P0 Ym } has
positive curvature2 in (G, Q), then secgt Ct−1 (σ ) > 0 for all t > 0.
Combining the above with the Gauss-Bonnet Theorem, it follows that if σ is
tangent to a totally geodesic flat torus in (M, g0 ) that contains a horizontal direction,
then secgt (Ct−1 (σ )) = 0 for all t ≥ 0. On the other hand, we also obtain the following
important positive result:
Corollary 6.7. Consider an isometric action of G = SO(3) or G = SU(2) on
(M, g0 ). If secg0 (σ ) = 0 and the image of the projection of σ ⊂ V p ⊕ H p onto V p is
 
2-dimensional, then secgt Ct−1 (σ ) > 0 for all t > 0.
In fact, by Proposition 2.26 (see also Example 6.25), these Lie groups (G, Q)
have sec > 0 and hence [P0 Xm , P0 Ym ] = 0 whenever Xm = 0 and Ym = 0.

2 Recall Proposition 2.26 (iv).

6.2 Compact Homogeneous Spaces 145

6.2 Compact Homogeneous Spaces

In this section, we quickly introduce the basic theory of compact homogeneous

spaces. For a more comprehensive introduction, see, for instance, [24, 33, 71, 179].
Let G be a compact connected Lie group and H a closed subgroup (hence H is an
embedded Lie subgroup by Theorem 1.42). Consider the space of (right) cosets

G/H := {gH : g ∈ G}

and the quotient map ρ : G → G/H, ρ (g) = gH. Then G/H has a unique smooth
structure induced by ρ , making it a smooth manifold with dim G/H = dim G −
dim H, see Corollary 3.38. Manifolds of this form are called homogeneous spaces.
The justification for this name comes from the natural transitive left G-action on
G/H, induced by the G-action by left translations on G (see Exercise 3.40), that is,

μ : G × G/H → G/H, μ (g, gH) = ggH. (6.13)

We refer to the above action as the left translation action, cf. (1.3). We follow the
usual slight abuse of notation and denote the diffeomorphism μ g = μ (g, ·) : G/H →
G/H simply by g : G/H → G/H. Also, we generally assume that the above action
is effective, i.e., μ g = id implies g = e. In some cases, however, it may be useful to
relax this condition to almost effective, which means that the ineffective kernel of
the action can be finite (see Definition 3.2).
Exercise 6.8. Show that the left translation action (6.13) is effective if and only if
G and H have no common normal subgroups. In particular, the ineffective kernel of
the action is the largest subgroup that is simultaneously normal in H and G.
Exercise 6.9. In this exercise, we show that every homogeneous space G/H comes
equipped with a free (right) action by the group3 N(H)/H, and is hence the total
space of a homogeneous principal N(H)/H-bundle:

N(H)/H → G/H → G/N(H), (6.14)

which already appeared in the proof of Theorem 3.95. Prove the following:
(i) The map μ : G/H × N(H) → G/H, μ (gH, n) := gHn = gnH defines a smooth
right action of N(H) on G/H, with ineffective kernel H;
(ii) Use Remark 3.3 and Theorem 3.34 to conclude that N(H)/H acts freely on
G/H, making it the total space of the principal bundle (6.14).
Instead of starting from the compact Lie groups G and H to build the homo-
geneous space G/H, a more classical viewpoint is that a homogeneous space is

3 Recall that N(H) is the largest subgroup of G where H is normal, see Exercise 1.61.
146 6 Low Cohomogeneity Actions and Positive Curvature

a Riemannian manifold (M, g) with a transitive isometric group action. Choosing

any subgroup G ⊂ Iso(M, g) that acts transitively on M, by Proposition 3.41, we
have that M is G-equivariantly diffeomorphic to G/H, where H := Gx is the
isotropy group of any point x ∈ M. Notice, however, that there may be many
subgroups G of Iso(M, g) that act transitively, hence M ∼ = G/H is generally not
a unique representation. The largest group G = Iso(M, g) that acts transitively on
the homogeneous space (M, g) is usually called its full isometry group.
Regarding examples of homogeneous spaces, note that Lie groups themselves
are trivially homogeneous spaces with H = {e}. Before going any further, let
us introduce a very important class of examples, which already made a quick
appearance in Exercise 3.44. For the sake of brevity, in the remainder of this chapter,
we use the shorthand notation G1 G2 for the product group G1 × G2 , as well as any
quotient (G1 × G2 )/Γ by a finite normal subgroup Γ ⊂ G1 × G2 .
Example 6.10. The following homogeneous spaces are called Compact Rank One
Symmetric Spaces, or CROSS for short, since they are symmetric spaces (see
Exercise 6.32) whose maximal totally geodesic flat submanifold has dimension 1,
i.e., is a geodesic. They consist of spheres and projective spaces relative to the 4 real
normed division algebras: R of reals, C of complex numbers, H of quaternions and
Ca of Cayley numbers (or octonions):
(i) Sn = O(n + 1)/O(n) = SO(n + 1)/SO(n), the unit sphere in Rn+1 . The
standard isometric action of O(n + 1) on Rn+1 is transitive on the unit sphere,
and the isotropy group O(n) of a point v ∈ Sn is formed by linear orthogonal
transformations of the hyperplane v⊥ ∼
= Rn ;
(ii) RP = SO(n + 1)/S(O(n)O(1)), the real projective space, is the quotient of Sn
n

by the properly discontinuous antipodal Z2 -action;

(iii) CPn = U(n + 1)/U(n)U(1) = SU(n + 1)/S(U(n)U(1)), the complex projective
space. The standard isometric action of U(n + 1) on Cn+1 restricted to the
unit sphere S2n+1 commutes with the unit multiplication U(1)-action (see
Exercise 3.40). Thus, it descends to an isometric action on the orbit space
CPn = S2n+1 /U(1). Alternatively, consider the U(n + 1)-action on the set of
complex lines in Cn+1 . The isotropy group U(n)U(1) of a complex line  is
formed by unitary transformations U(n) of ⊥ ∼= Cn and U(1) of  ∼ = C;
(iv) HP = Sp(n + 1)/Sp(n)Sp(1) the quaternionic projective space. The standard
n

isometric action of Sp(n + 1) on Hn+1 restricted to the unit sphere S4n+3

commutes with the unit multiplication Sp(1)-action (see Exercise 3.40). Thus,
it descends to an isometric action on the orbit space HPn = S4n+3 /Sp(1).
Alternatively, consider the Sp(n + 1)-action on the set of quaternionic lines
in Hn+1 . The isotropy group Sp(n)Sp(1) of a quaternionic line  is formed by
unitary transformations Sp(n) of ⊥ ∼
= Hn and Sp(1) of  ∼ = H;
(v) CaP = F4 /Spin(9), the Cayley plane, or the octonionic projective plane.
2

This is the homogeneous space of octonionic lines in Ca3 , and is given by

the quotient of the exceptional Lie group F4 by a subgroup isomorphic to
Spin(9), see [123, Thm 14.99]. For more details on its geometry, see [26, 123].
We remark that, differently from the above examples, CaP2 is not covered
6.2 Compact Homogeneous Spaces 147

by a sphere. Also, higher dimensional analogues do not exist, due to non-

associativity of the octonions.4
As in the case of Lie groups, one is usually interested in studying objects invariant
under left translations. In particular, a metric g on G/H is G-invariant if (6.13) is an
isometric action on the Riemannian manifold (G/H, g), i.e., every g : G/H → G/H
is an isometry. It is important to note that, in this case, the full isometry group
Iso(G/H, g) may be strictly larger5 than G.
Because of the presence of a transitive isometric group action, geometric
properties of homogeneous manifolds are determined by information at a single
point. For example, the scalar curvature of a homogeneous manifold is clearly
a constant function. In order to better describe how the geometry of G/H is
determined at a point, we look at the corresponding linearized objects.
Denote by g and h the Lie algebras of G and H, respectively. Since we are
assuming G is compact, by Proposition 2.24, G admits a bi-invariant metric, that
we denote by Q. Using this metric, define m to be the orthogonal complement of
h in g. By the bi-invariance of Q, it follows that m is invariant under the adjoint
representation restricted to H, i.e., Ad(h)m ⊂ m, for all h ∈ H. This representation
is closely related to the geometry of G/H, as explained in the next result.
Proposition 6.11. The tangent space TeH G/H is identified with the Ad(H)-
invariant complement m of h in g. Under this identification, the isotropy
representation of H on TeH G/H is the adjoint representation Ad : H × m → m.
Proof. Consider the derivative dρe : g → TeH G/H of the quotient map ρ : G →
G/H, ρ (g) = gH. This is a linear map given by

dρe X = d
dt exp(tX)H t=0 , (6.15)

so its restriction to the second factor of the decomposition g = h ⊕ m is a

linear isomorphism m ∼ = TeH G/H, giving the desired identification. To make this
identification H-equivariant, we need to prove that the restriction of the adjoint
representation H  h → Ad(h) ∈ GL(m) corresponds to the isotropy representation
H  h → dheH ∈ GL(TeH G/H). This follows at once by direct computation
using (1.10), since for all h ∈ H and X ∈ m,

4 The reason why CaP2 can be constructed as a projectivization of Ca3 is related to the fact that Ca
is 2-associative, i.e., any subalgebra generated by 2 elements is associative.
= Sp(n + 1)/Sp(n), but depending on what Sp(n +
5 For instance, one can write the sphere as S4n+3

1)-homogeneous metric we consider (e.g., the round metric), the full isometry group may be strictly
larger than Sp(n + 1), see Example 6.16 and Remark 6.17.
148 6 Low Cohomogeneity Actions and Positive Curvature

  
dρe Ad(h)X = d
dtexp(tAd(h)X)H t=0
  
= dtd h exp(tX)h−1 H t=0 (6.16)
 
= dheH dρe X . 

Exercise 6.12. Use the two equivalent ways given above of writing the isotropy
representation on G/H to show in two different ways that if the left translation G-
action on G/H is effective, then the isotropy representation is also effective.
Hint: Interpreting the isotropy representation as:
(i) H  h → dheH ∈ GL(TeH G/H), use Proposition 2.11 to show that if h(eH) = eH and dheH =
id, then h : G/H → G/H is the identity;
(ii) H  h → Ad(h) ∈ GL(m), use Exercise 6.8 to show that if Ad(h) = id, then h = e.

Using the above, we can identify the moduli space of invariant metrics on a
homogeneous space in terms of inner products on m, as follows.
Theorem 6.13. The set of G-invariant Riemannian metrics on G/H is in 1-to-1
correspondence with the set of inner products on m that are Ad(H)-invariant.
Proof. The restriction of any G-invariant metric g to TeH G/H gives an inner product
invariant under the isotropy representation. By Proposition 6.11, this is an inner
product on m that is Ad(H)-invariant. Conversely, given an Ad(H)-invariant inner
product ·, · on m ∼
= TeH G/H, define a G-invariant metric g on G/H by setting
# $
g(X,Y ) := dLg−1 X, dLg−1 Y , X,Y ∈ TgH G/H. (6.17)

Indeed, since ·, · is Ad(H)-invariant, by (6.16) it follows that (6.17) is independent

of the choice of representative in the coset gH, and hence g is well-defined. 

Remark 6.14. Provided H is connected, many of the above statements in terms

of Ad(H)-invariance have equivalent counterparts in terms of ad(H), obtained by
differentiating the Ad(H)-invariance equation. Namely, the complement m to h in g
is Ad(H)-invariant if and only if m is ad(H)-invariant, that is, [h, m] ⊂ m. Similarly,
an inner product on m is Ad(H)-invariant if and only if ad(H) is skew-symmetric
for this inner product, cf. Proposition 2.26 (i).
In particular, Theorem 6.13 implies that the moduli space of G-invariant metrics
on G/H is a finite-dimensional open cone. To further understand the structure of
this moduli space, we apply Schur’s Lemma (see [56, p. 69]) to analyze the general
form of an Ad(H)-invariant inner product on m.
Consider the decomposition of m into the direct sum of irreducible Ad(H)-
representations. Grouping the copies of irreducible summands that are equivalent
( i
into subspaces mi := nj=1 mij , we have a decomposition m = m0 ⊕ m1 ⊕ · · · ⊕ ml .
In other words, for each 0 ≤ i ≤ l, any mij and mik , with 1 ≤ j ≤ k ≤ ni , are equivalent
6.2 Compact Homogeneous Spaces 149

Ad(H)-representations, that is, there is an equivariant isomorphism mij ∼ = mik . The

subspaces mi are called isotypic components of m, and ni ∈ N is the multiplicity
of the irreducible summand mi j . It is usually conventioned that m0 is the trivial
isotypic component, i.e., m0 j , 1 ≤ j ≤ n0 , are copies of the (1-dimensional) trivial
representation. In terms of this decomposition, we have the following:
Corollary 6.15. Let G/H be a compact homogeneous space and fix a bi-invariant
metric Q on g. Consider the above decomposition m = m0 ⊕ m1 ⊕ · · · ⊕ ml into
isotypic components. For each X ∈ m, denote by Xi ∈ mi its component on mi , and
by Xij ∈ mij its component on mij , so that Xi = ∑nj=1
i
Xij . The G-invariant metrics on
G/H are those induced by inner products on m that are of the form:

l ni
X,Y  = ∑ Xi ,Yi i , ∑ jk
where Xi ,Yi i = ai Q(Xij ,Yik ) (6.18)
i=0 j,k=1

jk
and ai are the coefficients of a positive-definite symmetric matrix Ai ∈ GL(ni , R).
Fortunately, in many examples discussed below, the nontrivial isotypic compo-
nents consist of only one irreducible summand, that is, ni = 1 for all 1 ≤ i ≤ l. In
this case, the matrices Ai , 1 ≤ i ≤ l, are simply positive real numbers ai , and hence
the inner product (6.18) is given by

l
·, · = ·, ·0 ⊕ ∑ ai Q|mi , (6.19)
i=1

where ·, ·0 = Q(A0 ·, ·) is an arbitrary inner product on m0 ∼

= Rn0 and ai > 0.
Equipped with the above, one can study the moduli space of homogeneous
metrics on G/H, provided the isotropy representation is known. For instance, let
us describe all homogeneous metrics on spheres, following the classification of
Ziller [235]. All these metrics are of the form (6.19).
Example 6.16 (Homogeneous metrics on spheres). The classification of transitive
Lie group actions on spheres was obtained by Borel [47] and Montgomery and
Samelson [167]. We list in Table 6.1 all possible transitive G-actions6 on Sn , with
the corresponding isotropy H and the decomposition of m into isotypic components.

6 Here, we assume G is connected. This entails no loss of generality, since the restriction to the
identity component of a continuous transitive group action on a connected space is still transitive.
150 6 Low Cohomogeneity Actions and Positive Curvature

Table 6.1 Transitive actions on spheres, G connected

G H dim G/H Isotropy repres.
(i) SO(n + 1) SO(n) n Irreducible
(ii) SU(n + 1) SU(n) 2n + 1 m = m0 ⊕ m1
(iii) U(n + 1) U(n) 2n + 1 m = m0 ⊕ m1
(iv) Sp(n + 1) Sp(n) 4n + 3 m = m0 ⊕ m1
(v) Sp(n + 1)Sp(1) Sp(n)Sp(1) 4n + 3 m = m1 ⊕ m2
(vi) Sp(n + 1)U(1) Sp(n)U(1) 4n + 3 m = m0 ⊕ m1 ⊕ m2
(vii) Spin(9) Spin(7) 15 m = m1 ⊕ m2
(viii) Spin(7) G2 7 Irreducible
(ix) G2 SU(3) 6 Irreducible

All of the above representations satisfy ni = 1, 1 ≤ i ≤ l, that is, each isotypic

component consists of a unique irreducible summand. Carefully analyzing each of
the above cases,7 it follows from Corollary 6.15, see also (6.19), that the moduli
space of homogeneous metrics on spheres is formed by:
(i) A 1-parameter family of U(n + 1)-homogeneous metrics on S2n+1 ;
(ii) A 3-parameter family of Sp(n + 1)-homogeneous metrics on S4n+3 ;
(iii) A 1-parameter family of Spin(9)-homogeneous metrics on S15 .
Any homogeneous metric on Sn must be, up to rescaling, isometric to one of the
above. Note that the round metric is the only one at which all families intersect.
The geometric interpretation of these families is in terms of the Hopf bundles:
1
S1 → S2n+1 → CPn , S3 → S4n+3 → HPn , S7 → S15 → S8 2 . (6.20)

Namely, each family of homogeneous metrics above coincides with the family
obtained by rescaling the unit round metric on the total space of the corresponding
Hopf bundle in the vertical directions. On each of these total spaces, consider the
metric

gλ := g|H ⊕ λ g|V , λ > 0, (6.21)

7 Letus give a brief account of this analysis, see Ziller [235, Sec. 1] for details. On cases (i), (viii)
and (ix), since the H-action on m is irreducible, the unique G-homogeneous metric is the round
metric. On cases (ii) and (iii), the H-action on m1 is the standard U(n)-action on Cn . The metrics
obtained in both cases are the same, yielding a family of U(n + 1)-homogeneous metrics on S2n+1
that depends on two parameters. On case (iv), the H-action on m1 is the standard Sp(n)-action
on Hn . This gives a family of Sp(n + 1)-homogeneous metrics on S4n+3 that depends on seven
parameters, six of which determine the restriction ·, ·0 to m0 . By appropriately diagonalizing A0 ∈
GL(3, R), this family can be further reduced to one that depends on only four parameters. Cases
(v) and (vi) are contained in case (iv). Finally, on case (vii), the H-action on the 7-dimensional
space m1 and on the 8-dimensional space m2 are the only irreducible representations of Spin(7)
in those dimensions. This gives a family of Spin(9)-homogeneous metrics on S15 that depends on
two parameters. All the above can be normalized so that a1 = 1, reducing by one the number of
parameters of each family.
6.2 Compact Homogeneous Spaces 151

where H and V denote horizontal and vertical components, and g1 = gH ⊕ gV is

the unit round metric.8 The metrics gλ account for the entire 1-parameter families
of homogeneous metrics in cases (i) and (iii) above, and for a subfamily in case
(ii). The general homogeneous metrics in case (ii) can be described as follows.
Identify the vertical space V of S4n+3 , i.e., the tangent space to the fiber S3 , with
(R3 , ×) ∼= su(2), recall Exercises 1.9 and 1.55. Use 3 positive parameters λ1 , λ2
and λ3 , to respectively rescale the 3 elements of the canonical orthonormal basis of
(R3 , ×), defining a left-invariant metric on the fiber S3 ∼
= SU(2). This metric on the
vertical directions and the round metric on the horizontal directions determines a
homogeneous metric g(λ1 ,λ2 ,λ3 ) on the total space S4n+3 . This 3-parameter family is
precisely the family in (ii). Note that if λ1 = λ2 = λ3 = λ , then g(λ1 ,λ2 ,λ3 ) = gλ .
Remark  6.17. Let us analyze
 the full isometry group of the Sp(n +1)-homogeneous
space S4n+3 , g(λ1 ,λ2 ,λ3 ) , to illustrate how a G-homogeneous space G/H might have
full isometry group strictly larger than G. When two of the parameters λi are equal,
rotations in the corresponding vertical plane are isometries, so the full isometry
group is Sp(n + 1)U(1). Similarly, when all three parameters λi are equal, the full
isometry group is Sp(n + 1)Sp(1). These respectively correspond to cases (vi) and
(v) in Table 6.1. Finally, if the λi are all equal to 1, then g(1,1,1) = g1 is the round
metric, which has the largest full isometry group O(4n + 4).
Remark 6.18. For 0 < λ ≤ 1, the above family of homogeneous metrics gλ on
the total space of the Hopf bundles S1 → S2n+1 → CPn and S3 → S4n+3 → HPn
can be interpreted as a Cheeger deformation of the round metric g1 . In fact, these
are principal bundles, and S1 ∼ = U(1), respectively S3 ∼= SU(2), acts isometrically
on the round sphere S 2n+1 , respectively S4n+3 , with Hopf fibers as orbits (see
Example 3.30). An algebraic description of these actions is given in Exercise 6.19.
The Cheeger deformation of the round metric g1 is the metric gλ (t) in (6.21) with
λ (t) = 1+t 1
, t ≥ 0. In other words, the Cheeger deformation shrinks the Hopf fibers
uniformly, and the corresponding spheres converge (in Gromov-Hausdorff sense) to
CPn (respectively HPn ) as t ! +∞, which corresponds to λ (t) 0. In particular,
from Proposition 6.1 and Corollary 6.7, all such metrics have sec > 0. On the other
hand, the invariant metrics obtained by dilating the Hopf fibers, i.e., gλ for λ > 1, are
not Cheeger deformations of g1 ; and if the dilating factor is λ ≥ 4/3 these manifolds
actually fail to have sec > 0. A complete description of homogeneous metrics with
sec > 0 on spheres is given in Theorem 6.31.
 also stress that no analogous statement holds for the Hopf bundle S → S →
7 15
8
We
1
S 2 , which is not a homogeneous bundle. In fact, there are no isometric group
actions on S15 whose orbits are the 7-dimensional Hopf fibers, see [120].
Exercise 6.19. Let H ⊂ K ⊂ G be compact Lie groups, and consider the homo-
geneous bundle K/H → G/H → G/K, whose projection map π : G/H → G/K is

equipped with the metric gλ are known as Berger spheres; though, historically, this term
8 Spheres

was mostly used referring to the case of (S3 , gλ ) first studied by Berger.
152 6 Low Cohomogeneity Actions and Positive Curvature

given by π (gH) = gK. Show that if H is a normal subgroup of K, then the right
K-action on G/H given by gH · k := gHk = gkH is well-defined and its orbits are
the fibers of π . Notice that K ⊂ N(H) and compare with (6.14) in Exercise 6.9.
The result in the above exercise applies to the homogeneous Hopf bundles in
Remark 6.18, where H ⊂ K ⊂ G are respectively U(n) ⊂ U(n)U(1) ⊂ U(n + 1) and
Sp(n) ⊂ Sp(n)Sp(1) ⊂ Sp(n + 1), see Example 6.10.
Remark 6.20. Besides the round metric, there are other homogeneous Einstein
metrics on spheres. Namely, Jensen [134] found that gλ with λ = 2n+3 1
is an
Sp(n + 1)-invariant Einstein metric on S 4n+3 , and Bourguignon and Karcher [52]
found that gλ with λ = 11 3
is a Spin(9)-invariant Einstein metric on S15 . Ziller [235]
proved that these are in fact all homogeneous Einstein metrics on spheres.
Note that, in order to verify that a G-invariant metric g on a homogeneous
space G/H is Einstein, it clearly suffices to verify that Ric = λ g at a single point,
for instance eH. This viewpoint reduces the Einstein equation, which is a partial
differential equation on a manifold, to an algebraic system on m, see [44, 225, 226].
This approach also allows to study homogeneous Ricci solitons, and the Ricci flow
of homogeneous metrics, see [147, 148, 150].
Despite not necessarily being Einstein, homogeneous metrics always have con-
stant scalar curvature, since this is a function invariant under isometries. However,
there may be constant scalar curvature metrics on homogeneous spaces G/H that
are not G-invariant. Combining classical variational bifurcation techniques and
the above characterizations of the homogeneous metrics on spheres, Bettiol and
Piccione [37] recently obtained new (non-homogeneous) constant scalar curvature
metrics on S4n+3 , n ≥ 1, in the conformal classes of gλ , for infinitely many λ ∈
(0, 1), that accumulate at 0. This was later generalized in [38] to the homogeneous
bundles described in Exercise 6.19 for which K/H has positive scalar curvature.
We now compute the moduli space of invariant metrics on the other CROSS,
following Ziller [235, Sec. 3].
Example 6.21 (Homogeneous metrics on projective spaces). Analogously to
Table 6.1, we list in Table 6.2 all transitive actions by connected Lie groups on
the remaining CROSS, obtained by Oniscik [178, p. 168], see also [179].

Table 6.2 Transitive actions G H G/H Isotropy repres.

on projective spaces, G
connected (i) SU(n + 1) S(U(n)U(1)) CPn Irreducible
(ii) Sp(n + 1) Sp(n)U(1) CP2n+1 m = m1 ⊕ m2
(iii) Sp(n + 1) Sp(n)Sp(1) HPn Irreducible
(iv) F4 Spin(9) CaP2 Irreducible

In all but one case in Table 6.2, the isotropy representation is irreducible; so
(up to rescaling) there exists a unique G-invariant metric on the corresponding
homogeneous space G/H. This is the Fubini-Study metric gFS , which, in the cases
of CPn and HPn is the metric that turns the Hopf bundles S2n+1 → CPn and
6.2 Compact Homogeneous Spaces 153

S4n+3 → HPn into Riemannian submersions, where the corresponding spheres are
endowed with the round metric.9 Since the isotropy representation is irreducible, it
also follows that these metrics are Einstein, since there is (up to rescaling) only one
invariant symmetric (0, 2)-tensor on these manifolds, by Schur’s Lemma.
In the remaining case (ii), the isotropy representation has two inequivalent
irreducible factors m1 and m2 , of dimensions 2 and 4n respectively. Consequently,
there is (up to normalization) a 1-parameter family of Sp(n + 1)-invariant metrics
on CP2n+1 . These can be interpreted, similarly to the case (6.21) of spheres, as a
deformation gFS |H ⊕ λ gFS |V of the Fubini-Study metric gFS , rescaling by λ the
vertical directions of the homogeneous bundle CP1 → CP2n+1 → HPn .
Remark 6.22. Ziller [235] found that gFS |H ⊕ λ gFS |V with λ = 1
n+1 is an
Sp(n + 1)-invariant Einstein metric on CP2n+1 , cf. Remark 6.20.
Among all homogeneous metrics on G/H, there are special metrics which are
induced by a bi-invariant metric Q on G. These metrics correspond to the inner prod-
uct Q|m via Corollary 6.15, that is, to the case in which Ai ∈ GL(ni , R), 1 ≤ i ≤ l,
in (6.18) are identity matrices. Metrics of this form are called normal homogeneous
metrics. If g is one such metric on G/H, then the projection ρ : (G, Q) → (G/H, g) is
a Riemannian submersion. In particular, (G/H, g) has sec ≥ 0 by the Gray-O’Neill
formula (2.14), since (G, Q) has sec ≥ 0 by Proposition 2.26.
The following result is an interesting application of Cheeger deformations, due
to Schwachhöfer and Tapp [192, Prop 1.1], regarding the structure of the moduli
space of invariant metrics with sec ≥ 0 on a homogeneous space.
Theorem 6.23. The moduli space of invariant metrics on a compact homogeneous
space G/H with sec ≥ 0 is path-connected.
Proof. Let g∗ := Q|m be a normal homogeneous metric. Given an invariant metric
g with secg ≥ 0, consider the path of metrics (1 + t)gt obtained by rescaling
the Cheeger deformation of g by (1 + t). From (6.7), using the notation of
Proposition 6.3, we have
   
(1 + t)gt (X,Y ) = Q (1 + t)Pt (Xm ),Ym = Q (1 + t)P0 (id +tP0 )−1 Xm ,Ym .

As t ! +∞, the above clearly converges to Q(Xm ,Ym ), proving that (1 + t)gt
converges to g∗ as t ! +∞. Evidently, (1 + t)gt converges to g as t 0. By
Proposition 6.1, the entire path (1 +t)gt is through invariant metrics with sec ≥ 0.10


9 Notethat, as no sphere fibers over the Cayley plane CaP2 , there is no analogous interpretation for
the Fubini-Study metric in this case.
10 Besidesproving the above statement, this proof indicates that the moduli space of homogeneous
metrics with sec ≥ 0 is, in some sense, star-shaped with respect to normal homogeneous metrics.
154 6 Low Cohomogeneity Actions and Positive Curvature

Example 6.24. Some normal homogeneous metrics on spheres (see Example 6.16)
do not coincide with the round metric. In fact, apart from S3 ∼ = SU(2), the round
metric is only normal homogeneous in the cases (i), (viii) and (ix) in Table 6.1,
for which there is a unique invariant metric up to rescaling. In all other cases, the
normal homogeneous metric is of the form gλ for some 0 < λ < 1, see (6.21).
More precisely, up to rescaling, the normal homogeneous metric gλ on S2n+1 =
SU(n+1)/SU(n) corresponds to λ = n+1 2n , on S
4n+3 = Sp(n+1)/Sp(n) corresponds

to λ = 2 and on S = Spin(9)/Spin(7) corresponds to λ = 14 . For details, see [116,

1 15

Lemma 2.4].
Example 6.25. Compact Lie groups (G, Q) with bi-invariant metrics are not only
normal homogeneous spaces, but also symmetric spaces, see Theorem 2.30 and
Remark 4.12. From Proposition 2.26, (G, Q) can only have sec > 0 if G has rank
one. Thus, by Theorem 4.6, the only compact Lie groups that admit a bi-invariant
metric with sec > 0 are SU(2) and SO(3).
We now present the curvature formula for a general G-invariant metric on a
homogeneous space G/H, in a slightly different way from usual textbooks in the
area, e.g., [24, 33, 71]. Namely, we use a formalism introduced by Püttmann [187]
that is very convenient for later purposes. In what follows, we repeatedly use
Theorem 6.13 to write G-invariant metrics on G/H as Ad(H)-invariant inner
products on m. Also, we fix a bi-invariant metric Q on G and use the corresponding
normal homogeneous metric on G/H as a reference metric to write general G-
invariant metrics. More precisely, given a G-invariant metric g on G/H, there exists
a unique Q-symmetric Ad(H)-equivariant automorphism P : m → m such that

g(X,Y ) = Q(PX,Y ), for all X,Y ∈ m. (6.22)

We define the auxiliary symmetric and skew-symmetric bilinear maps

 
B± (X,Y ) := 1
2 [X, PY] ∓ [PX,Y ] . (6.23)

With these definitions, using the Gray-O’Neill formula (2.14) one can prove the
following, cf. [187, Lemma 3.6], [33, Chap. 7].
Proposition 6.26. The curvature tensor of the invariant metric g(·, ·) = Q(P·, ·) on
G/H is given by
 
R(X,Y, Z,W ) = 1
2 Q(B− (X,Y ), [Z,W ]) + Q([X,Y ], B− (Z,W ))

+ 14 Q([X,W ], P[Y, Z]m ) − Q([X, Z], P[Y,W ]m )

− 2Q([X,Y ], P[Z,W ]m )
+ Q(B+ (X ,W ), P−1 B+ (Y, Z))
− Q(B+ (X , Z), P−1 B+ (Y,W )),

where Xm denotes the component in m of an element X ∈ g.

6.2 Compact Homogeneous Spaces 155

Exercise 6.27. The Lie group (G, Q) endowed with a bi-invariant metric is a
particular case of an invariant metric on a homogeneous space, where H = {e}
and P = id. Compute B± in this case and substitute in the above formula to recover
formula (iii) in Proposition 2.26.
Exercise 6.28 (). Note that when P = id, the metric g on G/H is normal homoge-
neous. For one such metric:
(i) Use Proposition 6.26 to compute sec(X ,Y ) = R(X,Y, X,Y ), where X,Y ∈ m
are orthonormal, obtaining
, ,2 , ,2
sec(X ,Y ) = 14 ,[X,Y ], + 34 ,[X,Y ]h , , (6.24)

and verifying the above claim that normal homogeneous metrics have sec ≥ 0;
(ii) Characterize the vectors X such that sec(X,Y ) = 0 for any orthogonal vector Y ;
(iii) Use the above to show that Ric(X , X ) ≥ 0 for all X ∈ m and Ric(X, X) = 0 only
if X ∈ Z(g), see (1.15).
Remark 6.29. Recall that, by the Bonnet-Myers Theorem 2.19, if a compact
Riemannian manifold (M, g) has Ric > 0, then its fundamental group π1 (M) is finite.
The converse statement holds for compact homogeneous spaces, that is, if G/H is a
compact homogeneous space with π1 (G/H) finite, then G/H admits a Riemannian
metric with Ric > 0. Indeed, any normal homogeneous metric on G/H has Ric > 0
by Exercise 6.28 (iii), since it is possible to prove that Z(g) ∩ m = {0} provided
π1 (G/H) is finite, see [28].
It is sometimes possible to increase curvature from sec ≥ 0 to sec > 0 along the
submersion ρ : G → G/H. We stress this just happens in very special cases (only
in dimensions 6, 7, 12, 13 and 24), and it is typically achieved with a left-invariant
metric on G which is not bi-invariant, but rather a deformation11 of a bi-invariant
metric. The classification of simply-connected closed homogeneous manifolds with
an invariant metric of positive sectional curvature was obtained through the work of
Berger [30], Wallach [221], Aloff-Wallach [21] and Berard-Bergery [27], see also
[232].
Theorem 6.30 (Homogeneous classification of sec > 0). Apart from the CROSS
(see Example 6.10), the only simply-connected closed manifolds to admit a homo-
geneous metric with sec > 0 are:
7 = SU(3)/S1 , gcd(k, l) = 1, kl(k + l) = 0, including
• Aloff-Wallach spaces Wk,l k,l
7
W1,1 = SU(3)SO(3)/U(2);
• Berger spaces B7 = SO(5)/SO(3) and B13 = SU(5)/S1 Sp(2);

11 More precisely, a Cheeger deformation of a bi-invariant metric on G with respect to the action of
an intermediate subgroup H ⊂ K ⊂ G.
156 6 Low Cohomogeneity Actions and Positive Curvature

• Wallach flag manifolds W 6 = SU(3)/T 2 , W 12 = Sp(3)/Sp(1)Sp(1)Sp(1) and

W 24 = F4 /Spin(8).
For details on the above homogeneous spaces, we refer the reader to Ziller [236]
and Wilking and Ziller [232].
After knowing which homogeneous spaces admit invariant metrics of positive
curvature, a natural question is to study the moduli space of such invariant metrics,
and optimal pinching constants. Recall that the pinching constant of a Riemannian
manifold (M, g) is the ratio between extremal values of secg at the same point:

min secg
δg = ≤1
max secg

It is a classical fact in Riemannian geometry that if δg = 1, then (M, g) is isometric

to a spherical space form, that is, a quotient of a round sphere. A classical result in
modern Riemannian geometry is the Topological Sphere Theorem, which states that
a complete simply-connected manifold (M, g) with δg > 14 must be homeomorphic
to a sphere. This deep result was proved by Berger [29] and Klingenberg [142,
143], elaborating on a novel idea of Rauch [190]. An exciting recent development
was the proof of the Differentiable Sphere Theorem by Brendle and Schoen [55],
which improves the conclusion from homeomorphism to diffeomorphism. In part,
this was made possible by the remarkable work of Böhm and Wilking [45], who
proved that the Ricci flow evolves a metric with positive-definite curvature operator
to a limit with constant sectional curvature. Notice that for a CROSS other than
Sn (recall Example 6.10), such as CPn and HPn , the pinching of the Fubini-Study
metric is δgFS = 14 , showing that the quarter pinched condition δg > 14 is sharp in
even dimensions.
The optimal pinching constant for a homogeneous space is the largest possible
pinching δmax among invariant metrics, providing a notion of the best invariant
metric on such space. For homogeneous spaces that are not symmetric, computing
the exact value of δmax is an extremely difficult task. This was achieved by
Valiev [215] on the flag manifolds W 6 , W 12 and W 24 , for which δmax = 64 1
, by
Eliasson [80] on the Berger space B , for which δmax = 37 , and by Püttmann [187]
7 1
7 , for which δ
max = 37 .
1
on the Berger space B13 and the Aloff-Wallach space W1,1
Finally, Püttmann [187] also proved that δmax ≤ 37 1
for all remaining Aloff-Wallach
7
spaces Wk,l .
Consider now the problem of determining the moduli space of positively curved
metrics on a given homogeneous space. A complete answer in the case of spheres
(recall Example 6.16), together with their pinching constants, was recently obtained
by Verdiani and Ziller [219].
Theorem 6.31. Denote by gλ the homogeneous metric (6.21) obtained by scaling
the round metric on each Hopf bundle (6.20) by λ in the vertical direction. Then gλ
has sec > 0 if and only if 0 < λ < 4/3. Denote by g(λ1 ,λ2 ,λ3 ) the 3-parameter family
of homogeneous metric on S4n+3 , and set
6.3 Cohomogeneity One Actions 157

Vi := (λ j2 + λk2 − 3λi2 + 2λi λ j + 2λi λk − 2λ j λk )/λi and Hi := 4 − 3λi ,

where (i, j, k) denotes a cyclic permutation of (1, 2, 3). Then g(λ1 ,λ2 ,λ3 ) has sec > 0
if and only if
√
Vi > 0, Hi > 0, and 3|λ j λk − λ j − λk + λi | < λ j λk + HiVi .

We conclude this section with an exercise that provides a glimpse of the theory
of symmetric spaces, for which we recommend as further references [71, 126, 129].
Exercise 6.32 (). By definition, (M, g) is a symmetric space if for all p ∈ M there
exists an isometry I p : M → M that fixes p and reverses geodesics through p, i.e.,
I p (p) = p and d(I p ) p = − id, see Exercise 2.11 (and recall Theorem 2.30).
(i) Use the above definition to verify that a symmetric space is complete;
(ii) Use the Hopf-Rinow Theorem 2.9 to prove that a symmetric space (M, g) is
a homogeneous space G/H, where G = Iso(M, g) is its isometry group and
H := G p is the isotropy of a point p ∈ M;
(iii) If M = G/H is a symmetric space as above, show that g → I p g I p defines
an involution of G, whose fixed point set is a closed subgroup with identity
component equal to the identity component of H;
(iv) Use this involution to show that the Lie algebra g of G splits as g = h ⊕ m,
where h is the Lie algebra of H, [h, m] ⊂ m and [m, m] ⊂ h, cf. Proposition 6.11.
A pair of Lie groups (G, H) as above is called a symmetric pair.

6.3 Cohomogeneity One Actions

In this section, we describe the basic theory of cohomogeneity one manifolds,

exploring both their topological and geometric aspects. Although this material is
available in research papers and is well-known among specialists, to our knowledge,
it has not yet appeared in textbooks with a more didactic approach. A sample
of recent papers for further study, whose main results are related to the material
presented in this section, is [34, 39, 41–43, 75, 83, 91, 111, 112, 114–116].
A connected Riemannian manifold (M, g) is said to have cohomogeneity one if
it supports an isometric action of a compact Lie group G with at least one orbit of
codimension 1. This means that the orbit space M/G is one-dimensional, and since
G can be assumed closed in the full isometry group of (M, g), we must have, up to
rescaling, one of:
(i) M/G = R, the real line;
(ii) M/G = R+ = [0, +∞), a half-line;
(iii) M/G = S1 , the circle;
(iv) M/G = [−1, 1], an interval.
158 6 Low Cohomogeneity Actions and Positive Curvature

Let us analyze the orbit structure of M in each of the above cases, see Defini-
tion 3.84, by studying the fibers of the quotient map π : M → M/G. Interior points of
M/G correspond to principal orbits, and boundary points correspond to nonprincipal
orbits, see Remark 3.106. In cases (i) and (iii) above, all G-orbits are principal, and
M is a fiber bundle over M/G with fiber the principal orbit. In case (ii), there is
one nonprincipal orbit, namely G(x0 ) = π −1 (0), so M ∼ = G ×Gx0 D is equivariantly
diffeomorphic to the tubular neighborhood of G(x0 ), by the Tubular Neighborhood
Theorem 3.57.12 Finally, in case (iv), there are two nonprincipal orbits π −1 (±1).
Topologically, the latter is the most interesting case, since it is not a fiber bundle
over a simpler manifold in any obvious way.
We now specialize to this case M/G = [−1, 1] to give a more in-depth analysis.
Choose some p0 ∈ π −1 (0) and let γ : [−1, 1] → M be the unique minimal horizontal
geodesic with γ (0) = p0 . Then γ can be extended to γ : R → M, and its image
γ (R) is either an embedded circle or a one-to-one immersed line; in particular, it
does not have self-intersections [1, Thm 6.1]. Moreover, since it is horizontal, γ (R)
intersects all orbits orthogonally. This implies that the G-action on M is polar,13 see
Definition 4.8. Denote by H := Gγ (0) the isotropy group at γ (0), which is equal to the
isotropy groups Gγ (t) for all t ≡ 1 mod 2Z (Exercise 6.35); and by K± := Gγ (±1) the
isotropy groups at p± := γ (±1), respectively. Since G/H = π −1 (t), t ∈ (−1, 1), is
a principal orbit, we call H the principal isotropy group, while each K± is called
singular or exceptional isotropy group, according to G/K± = π −1 (±1) being a
singular or exceptional orbit. Denote by D± a normal slice to the nonprincipal
orbit G/K± at p± . We are now ready to describe a decomposition of such a
cohomogeneity one manifold as union of two disk bundles.
Proposition 6.33. There exists an equivariant diffeomorphism

M∼
= (G ×K− D− ) ∪G/H (G ×K+ D+ ), (6.25)

i.e., M is the union of tubular neighborhoods of the nonprincipal orbits G/K± , glued
along their common boundary G/H.
Proof. According to the Tubular Neighborhood Theorem 3.57, π −1 ([−1, 0]) and
π −1 ([0, 1]) are equivariantly diffeomorphic to the tubular neighborhoods
 
Tub G(p± ) = G ×K± D±

of the singular orbits G/K± , respectively. The common boundary of each of

these open subsets is the principal orbit π −1 (0) = G/H. Thus, the equivariant
decomposition M ∼= π −1 ([−1, 0]) ∪π −1 (0) π −1 ([0, 1]) is precisely (6.25). 

12 Slices in this section are generally denoted by D, instead of S as in Chap. 3.

13 Note that any cohomogeneity one action is polar, regardless of M/G being a closed interval.
6.3 Cohomogeneity One Actions 159

Remark 6.34. The notion of cohomogeneity one action can be extended to the
framework of Alexandrov spaces. In this context, analogous structure results to
Proposition 6.33 have been proved by Galaz-Garcia and Searle [91].
Exercise 6.35. From Exercise 3.77, we know that the slice representation at γ (0)
is trivial. Use that the principal isotropy group H = Gγ (0) preserves the horizontal
geodesic γ pointwise to verify the claim that Gγ (t) = H for all t ∈ (−1, 1). Note that
this also follows immediately from Kleiner’s Lemma 3.70.
It is often more natural to consider a closed interval orbit space M/G = [a, b],
different from [−1, 1], for geometric reasons. We discuss the abstract theory using
[−1, 1] as a standard rescaling, but in many examples the interval will be different,
as in the following:
Exercise 6.36. As a first (somewhat trivial) example of cohomogeneity one mani-
fold, consider the unit round sphere S2 ⊂ R3 with an action of S1 by rotation about
the z-axis. Identify the quotient map with the distance function from the North pole
π : S2 → [0, π ]. Describe the isotropy groups, finding the corresponding principal
and singular orbits. Write explicitly the decomposition (6.25). A generalization of
this example is given in Example 6.47.
The normal slices D± are normal disks at p± , and their boundary is called the
⊥ = ∂ D to the singular orbit G(p ). We write
normal sphere S± ± ±

⊥
± := dim S± = codim G/K± − 1.

The singular isotropy group K± acts on the slice D± via the slice representation,
⊥ is transitive,
see Definition 3.72. The restriction of this K± -representation to S±
so the normal sphere is precisely the (possibly ineffective) homogeneous space
⊥ = K /H. This allows us to define the group diagram of this cohomogeneity
S± ±
one manifold:

K− K+ (6.26)

where the arrows denote the natural inclusions. We also denote the above diagram by
H ⊂ {K− , K+ } ⊂ G. Notice that such a group diagram determines two homogeneous
bundles K± /H → G/H → G/K± , that have the principal orbit G/H as the total space
of a sphere bundle over each singular orbit G/K± .
160 6 Low Cohomogeneity Actions and Positive Curvature

Conversely, given compact Lie groups H ⊂ {K− , K+ } ⊂ G such that K± /H

are spheres,14 there exists a cohomogeneity one G-manifold whose group diagram
is (6.26). Namely, define M using (6.25). In this way, the group diagram determines
a cohomogeneity one manifold, up to certain ambiguities (see [114, p. 44]).
Proposition 6.37. Two group diagrams H ⊂ {K− , K+ } ⊂ G and H  ⊂ {K
− , K
+ } ⊂ G
determine the same cohomogeneity one manifold (up to equivariant diffeomor-
phism) if and only if, up to switching K− and K+ , there exist a ∈ N(H)0 and b ∈ G
such that K− = b K− b−1 , H = b H b−1 and K+ = a b K+ b−1 a−1 .

Remark 6.38. The above results on how a cohomogeneity one manifold is deter-
mined by its group diagram illustrate a particular case of Theorem 5.42 regarding
reconstructing polar manifolds from the polar data, due to Grove and Ziller [118].
Cohomogeneity one manifolds are polar, and its polar data is precisely its group
diagram.
Using the decomposition (6.25), we can deduce important relations between the
orbit structure and the topology of cohomogeneity one manifolds. For instance,
it follows that the inclusion G/K± → M is ∓ -connected and G/H → M is
min{− , + }-connected.15 One can also use this decomposition to apply the Mayer-
Vietoris exact sequence, relating the topology of the orbits with that of M. We now
collect some immediate consequences of the above observations, see [2, 114, 116].
Proposition 6.39. Let M be a cohomogeneity one manifold with M/G = [−1, 1].
Then the following hold:
(i) If both nonprincipal orbits G/K± are exceptional, i.e., ± = 0, then M has
infinite fundamental group;
(ii) If M is simply-connected, then there are no exceptional orbits, i.e., ± ≥ 1;
(iii) If − ≥ 2 and + ≥ 1, then K+ is connected. If both ± ≥ 2, then all isotropy
groups H and K± are connected;
(iv) The Euler characteristic of M is χ (M) = χ (G/K− ) + χ (G/K+ ) − χ (G/H).
Remark 6.40. Item (ii) of the above result is actually true for any polar manifolds,
see Corollary 5.35. In the particular case of cohomogeneity one, however, the
proof is elementary from G/K± → M being ∓ -connected. For an example of a
cohomogeneity one action with an exceptional orbit, consider RPn with a rotation
action of SO(n), see Exercise 3.87.
We now turn to the question of determining the moduli space of invariant metrics
on a cohomogeneity one manifold. An invariant smooth metric g on M is clearly
determined by its restriction to the open dense set Mprinc of regular points (recall the
Principal Orbit Theorem 3.82), which in the cohomogeneity one case is Mprinc = M \

14 The possible pairs (K , H) with K connected for which K /H is a sphere are listed in Table 6.1.
± ± ±
15 Recall that a map f : X → Y is -connected if the induced homomorphisms f : π (X) → π (Y ) is
i i i
an isomorphism for i <  and onto for i = .
6.3 Cohomogeneity One Actions 161

{G/K± }. The choice of the horizontal geodesic γ (t) on M determines an equivariant

diffeomorphism Mprinc ∼= (−1, 1) × G/H, for which γ (t) = (t, eH). Making this our
standard choice of (global) coordinates on the regular part of M, an invariant smooth
metric g can be written as

g = dt 2 + gt , t ∈ (−1, 1), (6.27)

where gt is a smooth 1-parameter family of homogeneous metrics on G/H,

satisfying appropriate smoothness conditions as t → ±1. As usual, denote by h and
g the Lie algebras of H and G. Using a choice of background bi-invariant metric
Q on G, let m be the orthogonal complement of h in g. Then for each t ∈ (−1, 1),
the tangent space Tγ (t) M splits (orthogonally) as the direct sum of the normal space
to the orbit G(γ (t)), spanned by γ̇ (t) = ∂∂t , and Tγ (t) G(γ (t)). We identify the latter
with TeH G/H via action fields, as in (3.11), and hence with m, by Proposition 6.11.
Gathering all of the above, we proved the following:
Proposition 6.41. Let M be a cohomogeneity one manifold with M/G = [−1, 1]
and γ : [−1, 1] → M be a minimizing horizontal geodesic between the nonprincipal
orbits. Every G-invariant metric g on M for which γ is a horizontal geodesic is of
the form (6.27), where gt is a 1-parameter family of G-invariant metrics on G/H,
i.e., Ad(H)-invariant inner products on m.
By the above, one can find a smooth 1-parameter family of Q-symmetric Ad(H)-
equivariant automorphisms Pt : m → m, such that
 
g Xγ∗(t) ,Yγ∗(t) = gt (X,Y ) = Q(Pt X,Y ), for all X,Y ∈ m, t ∈ (−1, 1).

The family Pt encodes the homogeneous metrics gt and also contains information
about the extrinsic geometry of principal orbits. For instance, we now compute the
shape operator of principal orbits in terms of Pt , following Grove and Ziller [116].
Proposition 6.42. For each t ∈ (−1, 1), the shape operator St of the hypersurface
G(γ (t)) at γ (t) is identified with

St : m → m, St = − 12 Pt−1 Pt , (6.28)


where Pt : m → m is the first derivative Pt (X) = 
ds Ps (X) s=t .
d

Proof. To simplify notation, we omit t whenever possible. Let X ∈ m and consider

its action field X ∗ on M. It is easy to see that, up to a change in coordinates
that locally straightens the hypersurface G(γ (t)), the normal vector γ̇ (t) is locally
constant along G(γ (t)), hence has zero derivative in the direction of X ∗ in this
162 6 Low Cohomogeneity Actions and Positive Curvature

coordinate system. This means that [X ∗ , γ̇ ] = 0, hence ∇X ∗ γ̇ = ∇γ̇ X ∗ . In particular,

using that X ∗ and Y ∗ are Killing vector fields (see Proposition 2.14), we have
 
g(∇γ̇ X ∗ ,Y ∗ ) = g(∇X ∗ γ̇ ,Y ∗ ) = X ∗ g(γ̇ ,Y ∗ ) − g(γ̇ , ∇X ∗ Y ∗ )
= g(X ∗ , ∇γ̇ ,Y ∗ ).

Thus, we conclude the proof with

gt (St (X),Y ) = −g(∇X ∗ γ̇ ,Y ∗ )

= −g(∇γ̇ X ∗ ,Y ∗ )
= − 12 dtd g(X ∗ ,Y ∗ )
= − 12 dtd Q(Pt X ,Y )
= − 12 Q(Pt X,Y )
 
= − 12 gt Pt−1 Pt X,Y .

Remark 6.43. Let us discuss some aspects of submanifold geometry of orbits of a

cohomogeneity one manifold. By G-equivariance of the mean curvature vector H,
every principal orbit G(γ (t)) is a constant mean curvature hypersurface, i.e., H =
Ht is constant on G(γ (t)), for each t. This is equivalent to saying that the sum of
eigenvalues of the shape operator St = − 12 Pt−1 Pt at γ (t) is the same in any other
point of G(γ (t)), cf. Proposition 3.78 (iii).
Consider the mean curvature H of G(γ (t)) as a function of t ∈ (−1, 1). It
can be proved that its global minimum value is zero (corresponding to a principal
orbit of critical volume, which must hence be minimal), and H diverges to +∞ as
t → ±1, provided that the nonprincipal orbits are singular. As for the singular orbits
themselves, they are isolated orbits of their type, hence a stratum of the stratification
of M by orbit types (see Theorem 3.102). In particular, singular orbits are minimal
submanifolds, see Remark 3.105. Furthermore, singular orbits are sometimes forced
to be totally geodesic, e.g., due to the isotropy representations, since the second
fundamental form is equivariant and the space of such objects may be trivial. Also, if
the dimension of a singular orbit is < 12 (dim M − 1), then it must be totally geodesic.
Using these properties and equivariant bifurcation techniques, Bettiol and
Piccione [39] recently proved the existence of infinitely many new constant mean
curvature hypersurfaces that are diffeomorphic (but not isometric) to principal
orbits on some cohomogeneity one manifolds, generalizing Delaunay surfaces
on spheres. In the case of S3 , these are classical Delaunay surfaces, which are
nonround tori (with isometry group Zm × S1 ) that bifurcate from the family (6.33)
in Example 6.48.
6.3 Cohomogeneity One Actions 163

Remark 6.44. Some of the above submanifold geometry properties of orbits of a

cohomogeneity one action were proved to remain true for general isoparametric
foliations by Ge and Tang [95, 96], see Remark 5.52.
We now describe a particularly useful type of cohomogeneity one metric, that
Grove and Ziller [116] call diagonal. Suppose m = m1 ⊕ · · · ⊕ ml is a decomposition
in mutually orthogonal Ad(H)-invariant subspaces (that are not necessarily Ad(H)-
irreducible). Consider, along γ (t), the metric

l
gt := ∑ fi (t)2 Q|mi , (6.29)
i=1

where fi : (−1, 1) → R are positive functions satisfying appropriate smoothness

conditions as t → ±1, and Q is a bi-invariant metric on G. The resulting invariant
metric g = dt 2 + gt is called a diagonal cohomogeneity one metric. The following
exercise justifies this terminology.
Exercise 6.45. Verify that the family of automorphisms Pt : m → m and the shape
operator St of principal orbits with respect to a diagonal metric as above are:

⎛ ⎞ ⎛ f ⎞
1
f 12 f1
⎜ ⎟ ⎜ ⎟
Pt = ⎝ ..
. ⎠ and St = − ⎜
⎝
..
. ⎟.
⎠ (6.30)
fl2 fl
fl

Exercise 6.46. Let g = dt 2 + gt be a diagonal cohomogeneity one metric and

consider its Cheeger deformation gs = dt 2 + gs,t with respect to the cohomogeneity
one action. Here, s denotes the parameter of the Cheeger deformation. Verify
that gs , s ≥ 0, are also diagonal metrics, and use Proposition 6.3 to check that
gs,t (X,Y ) = Q(Ps,t X,Y ), where the Q-symmetric automorphism Ps,t : m → m is
⎛ ⎞
f 1 (t)2
⎜ 1+s f1 (t)2 ⎟
⎜ .. ⎟
Ps,t = ⎜ . ⎟. (6.31)
⎝ ⎠
fl(t)2
1+s fl (t)2

Note that as s → +∞, for each t, gs,t → 0, and hence (M, gs ) converges in Gromov-
Hausdorff sense to its orbit space M/G.
Before proceeding with abstract cohomogeneity one manifolds, let us examine in
details a few examples where many of the above results can be seen more explicitly.
Example 6.47 (Surfaces of revolution). Surfaces of revolution M in R3 can be seen
as a very simple type of diagonal cohomogeneity one manifold, where the group
acting is S1 . Consider the embedding
164 6 Low Cohomogeneity Actions and Positive Curvature

 
F : [0, R] × [0, 2π ] → R3 , F(r, θ ) = φ (r) cos θ , φ (r) sin θ , ψ (r) ,

where φ : [0, R] → R and ψ : [0, R] → R are smooth functions, such that φ is positive
and only vanishes16 at the endpoints. For simplicity, we assume that the profile curve
F(r, 0) = (φ (r), 0, ψ (r)) is parametrized with unit speed, i.e., (φ  )2 + (ψ  )2 = 1. The
manifold M = F([0, R] × [0, 2π ]) is a surface of revolution in R3 , diffeomorphic
to S2 . It has a cohomogeneity one action of S1 by rotating the parameter θ , i.e.,
rotations around the z-axis. The quotient map is given by the distance function to
F(0, θ ) on M,

π : M ⊂ R3 → [0, R], π (F(r, θ )) = r.

The singular orbits are fixed points F(0, θ ) and F(R, θ ), and principal orbits are
circles {F(r, θ ) : θ ∈ [0, 2π ]}. In particular, the horizontal17 geodesic γ : [0, R] → M
can be taken as the profile curve γ (r) := F(r, 0) = (φ (r), 0, ψ (r)). The disk bundles
in (6.25) are simply disks that, glued along their common circle boundary (a
principal orbit), give the usual topological decomposition of M ∼ = S2 .
The metric induced by F on M is easily computed to be
 
g = (φ  )2 + (ψ  )2 dr2 + φ 2 dθ 2 = dr2 + φ 2 dθ 2 , r ∈ (0, R), (6.32)

i.e., g = dr2 +gr , where gr is a round metric on S1 with length φ (r). This is trivially a
diagonal metric, with Pr = φ (r)2 , see Exercise 6.45. This generalizes the example of
the round unit sphere in Exercise 6.36, where R = π , φ (r) = sin r and ψ (r) = cos r.
Example 6.48. Let S3 = {(z, w) ∈ C2 : |z|2 + |w|2 = 1} be the round sphere, and
consider the action of the torus G = T 2 = {(eiθ1 , eiθ2 ) : θ j ∈ R} by
 iθ iθ   
e 1 , e 2 · (z, w) := eiθ1 z, eiθ2 w .

The closed geodesics {(z, 0) ∈ S3 } and {(0, w) ∈ S3 } are singular orbits of this
T 2 -action, corresponding to a circle isotropy subgroup. The remaining orbits are
principal orbits that have trivial isotropy group, and are hence diffeomorphic to T 2 .
Let γ (t) = (cost, sint), 0 ≤ t ≤ π2 , be a horizontal geodesic joining the singular
orbits. Then the orbits of G = T 2 are

G(γ (t)) = (z, w) ∈ S3 : |z| = cost, |w| = sint , 0 ≤ t ≤ π2 . (6.33)

16 Inorder for M = F([0, R] × [0, 2π ]) to be a smooth closed submanifold of R3 , some conditions

on the derivatives of φ and ψ at the endpoints are needed. This illustrates the fact that the metrics
gt in (6.27) must satisfy appropriate conditions at the corresponding endpoints t → ±1.
17 This is a horizontal curve in the sense that it is orthogonal to all circle orbits. However, in R3 , it
sits in the plane y = 0, while the circle orbits are in planes z = const.
6.3 Cohomogeneity One Actions 165

The group diagram of this cohomogeneity one action is

T2

1 S1 S1 1

Notice that the decomposition (6.25) is the usual decomposition of S3 as two solid
tori glued along their boundary. We can write the round metric g on S3 in the
form (6.27), obtaining

g = dt 2 + gt = dt 2 + cos2 t dx2 + sin2 t dy2 , (6.34)

where z = eix cost and w = eiy sint. In other words, ∂∂x and ∂∂y are the vertical vector
fields along γ (t), that span the tangent space to (6.33). Thus, (6.34) is a diagonal
metric with

f1 (t) = cost and f2 (t) = sint.

In particular, the shape operator and hence the mean curvature of (6.33) can be
computed as in Exercise 6.45, and the result is

Ht = tant − cott, 0 < t < π2 . (6.35)

Thus, we have a minimal torus when t = π4 , corresponding to the principal orbit of

largest volume; and the constant mean curvature (6.35) of the tori G(γ (t)) blows up
as t → 0 and t → π2 , cf. Remark 6.43. Notice that this partition of S3 is precisely
the singular Riemannian foliation described in Exercise 5.54, in the particular case
k = 1 and n = 2, see also Remark 5.55.
 
Remark 6.49. The torus G(γ ( π4 )) = f −1 12 is called the (minimal) Clifford torus
in S3 . As a curiosity, Brendle [54] recently gave a proof of the celebrated Lawson
conjecture, which asserts that the Clifford torus is the unique embedded minimal
torus in S3 , up to isometries.
Exercise 6.50 (). The goal of this exercise is to generalize the above Example 6.48
to Sn+1 , n ≥ 2. Use the direct sum decomposition Rn+2 = Rk+1 ⊕ Rn−k+1 to find an
action of G = SO(k + 1) × SO(n − k + 1) on Sn+1 whose orbits are the level sets of
the isoparametric function f : Sn+1 → R from Exercise 5.54. Verify that the round
metric on Sn+1 is a diagonal cohomogeneity one metric with respect to this action,
with

f1 (t) = · · · = fk (t) = cost and fk+1 (t) = · · · = fn (t) = sint.

166 6 Low Cohomogeneity Actions and Positive Curvature

Using Exercise 6.45, compute the principal curvatures of the principal orbits and
see that they match the result in Exercise 5.54 (iii) and Remark 5.55; in particular,
prove that their mean curvature is

Ht = k tant − (n − k) cott, 0 < t < π2 .

-
Note that when t = arctan n−k k
, we get a minimal Sk × Sn−k in Sn+1 ; and as t → 0
and t → π2 , the mean curvature blows up, cf. Remark 6.43.
Remark 6.51. Analogously to Exercise 6.50, the singular Riemannian foliation of
hyperbolic space H n+1 ⊂ (Rn+2 , gL ) described in Exercise 5.56 can also be seen
as a partition into orbits of a cohomogeneity one G-action, via the direct sum
decomposition Rn+2 = Rk+1 ⊕ Rn−k+1 , with G = SO(k, 1) × SO(n − k + 1). Here,
SO(k, 1) is the Lorentz group, which is the (identity component of the) group of
isometries of the Lorentz space (Rk+1 , gL ) that fix the origin, i.e., the isometry group
of H k .
Example 6.52. Let  ⊂ Cn+2 be a complex line, and consider its orthogonal
complement ⊥ ∼ = Cn+1 . There is a natural linear action of SU(n + 1) on Cn+2 =
 ⊕  that fixes  and acts by multiplication on ⊥ . This SU(n + 1)-action on
⊥

Cn+2 commutes with the circle action by complex multiplication whose quotient
is CPn+1 . As a result, there is an induced SU(n + 1)-action on CPn+1 . The image
[] ∈ CPn+1 is clearly a fixed point of this action, and it is easy to see that the
orbit space is isometric to [0, π /2]. More precisely, consider a geodesic arc in the
unit sphere of Cn+2 that joins  to ⊥ . The image of this geodesic in CPn+1 is a
geodesic γ that intersects all orbits perpendicularly. Moreover, the quotient map
π : CPn+1 → [0, π /2] can be identified with the distance function from []. It is easy
to check that the group diagram for this action is:

SU(n + 1)

SU(n + 1) S(U(n) U(1))

SU(n)

cf. Exercise 3.44, so the normal spheres to the singular orbits have dimensions
− = 2n + 1 and + = 1. The singular orbits are the fixed point {[]} = π −1 (0) and
its cut locus CPn = π −1 (π /2). Principal orbits are geodesic spheres S2n+1 = π −1 (t)
of radius t ∈ (0, π /2) centered at []. Notice that the decomposition (6.25) is
homotopically the cell decomposition of CPn+1 as CPn with a 2n-cell attached to it
via the Hopf map.
6.3 Cohomogeneity One Actions 167

Using that the quotient map S2n+3 → CPn+1 is a Riemannian submersion from
the round sphere to (CPn+1 , gFS ), we can write the Fubini-Study metric as a
cohomogeneity one metric (6.27), obtaining:
 
gFS = dt 2 + gt = dt 2 + sin2 t g|H + cos2 t g|V . (6.36)

This means that the principal orbits (S2n+1 , gt ) are isometric to Berger spheres,
obtained by rescaling by sin2 t the SU(n + 1)-invariant metric gλ with λ = cos2 t,
from Example 6.16. In order to write gt as a diagonal metric in the form (6.29), we
consider the bi-invariant metric Q on SU(n + 1), recall Exercise 2.25 and Propo-
sition 2.48. There is a Q-orthogonal decomposition m = m1 ⊕ m2 in Ad(SU(n))-
invariant subspaces mi of dimensions 2n and 1 respectively, corresponding to the
horizontal H and vertical V directions of the Hopf bundle. These Ad(SU(n))-
subrepresentations are irreducible, since SU(n) acts by multiplication on m1 and
trivially on m2 . From (6.36) and Example 6.24, we have that the corresponding
functions fi are
-
f1 (t) = sint and f 2 (t) = 2n
n+1 sint cost. (6.37)

This gives a full description of the intrinsic geometry of the principal orbits.
As to their extrinsic geometry, we first note that (S2n+1 , gt ) are the distance
spheres in (CPn+1 , gFS ), of radius t centered at the fixed point []. We can compute
their shape operator using Exercise 6.45 and (6.37), obtaining
⎛ ⎞
cott
⎜ .. ⎟
⎜ . ⎟
St = − ⎜ ⎟. (6.38)
⎝ cott ⎠
2 cot 2t

In particular, the mean curvature of the principal orbit (S2n+1 , gt ) is

Ht = 2 cot 2t + 2n cott, (6.39)

√
so, we have a minimal hypersurface when t = arctan 2n + 1, corresponding to the
principal orbit of largest volume, and (6.39) blows up as t → 0 and t → π /2, cf.
Remark 6.43.
Exercise 6.53. Verify formula (iv) for Euler characteristics in Proposition 6.39 in
the case of the rotation action on the sphere (Exercise 6.36) and in the case of
the SU(n + 1)-action on CPn+1 (Example 6.52). Recall that χ (CPn ) = n + 1 and
χ (Sn ) = 2 if n is even and 0 if n is odd, see Example A.32.
Exercise 6.54 (). Describe a quaternionic version of the cohomogeneity one
action in Example 6.52, using a quaternionic line  in Hn+2 and considering the
168 6 Low Cohomogeneity Actions and Positive Curvature

action of Sp(n+1) that fixes it. Identify the orbits of the induced Sp(n+1)-action on
HPn+1 , write the group diagram of this action, and study the intrinsic and extrinsic
geometry of the principal orbits.
Exercise 6.55 (). Consider the family of cohomogeneity one manifolds Mk , k ∈
N, defined by the group diagrams Zk ⊂ {S1 , S1 } ⊂ S3 , where both embeddings of
K± = S1 inside G = S3 are given by the maximal torus (4.1).
(i) Describe an equivariant diffeomorphism between Mk and the associated bundle
S3 ×S1 S2 to the Hopf bundle S1 → S3 → S2 , where the S1 -action on S2 is by
rotations of speed k, studied in Example 3.55;
2
(ii) Conclude that S2 × S2 and CP2 #CP admit cohomogeneity one G-actions, and
verify that the G-action on S2 × S2 descends to an SO(3)-action;
2
(iii) Find cohomogeneity one metrics on S2 × S2 and CP2 #CP with sec ≥ 0, and
study the intrinsic and extrinsic geometry of the principal orbits.
Hint: For (i), verify that a curve joining the North and South poles on S2 defines a curve on
S3 × S2 that intersects each orbit of the diagonal S1 -action, and use (6.25). Item (ii) follows from
Example 3.55 by direct inspection. For (iii), consider the product of round metrics on S3 × S2 and
use the Gray-O’Neill formula (2.14), see Remark 6.80 for more details.

Example 6.56. Consider the Brieskhorn variety Md2n−1 ⊂ Cn+1 defined by

+
zd0 + z21 + · · · + z2n = 0,
|z0 |2 + |z1 |2 + · · · + |zn |2 = 1.

When n and d are odd, Md2n−1 is homeomorphic to the sphere S2n−1 . However, if
2n − 1 ≡ 1 mod 8, then Md2n−1 is not diffeomorphic to S2n−1 . Such manifolds are
exotic spheres called Kervaire spheres.
According to [111], it was first observed by Calabi in dimension 5 and later
by Hsiang and Hsiang [130] that Md2n−1 carries a cohomogeneity one action
by SO(2)SO(n), given by (eiθ , A) · (z0 , . . . , zn ) = (e2iθ z0 , eid θ A(z1 , . . . , zn )t ). The
principal isotropy of this action is
+
(ε , diag(ε , ε , A)) if d is odd
H = Z2 × SO(n − 2) = (6.40)
(ε , diag(1, 1, A)) if d is even,

where ε = ±1 and A ∈ SO(n − 2). Thus, principal orbits are diffeomorphic to S1 ×

T1 Sn−1 , where T1 Sn−1 denotes the unit tangent bundle of Sn−1 . The singular isotropy
groups are
 
K− = SO(2)SO(n − 2) = e−iθ , diag(R(d θ ), A) , (6.41)
6.3 Cohomogeneity One Actions 169

where R(θ ) is a counterclockwise rotation of angle θ , and

+
O(n − 1) = (det B, diag(det B, B)) if d is odd
K+ = (6.42)
Z2 × SO(n − 1) = (ε , diag(1, B )) if d is even,

where ε = ±1, B ∈ O(n − 1) and B ∈ SO(n − 1).

Exercise 6.57 (). The goal of this exercise is to prove a result of Back and
Hsiang [25, Thm VI.3], showing that the Brieskhorn variety Md2n−1 with the above
cohomogeneity one action does not admit an invariant metric with sec > 0, provided
n ≥ 4 and d ≥ 3. For an improvement of this result to sec ≥ 0, see Remark 6.74.
The proof is by contradiction, verifying that one of the singular orbits must be
totally geodesic and hence would have sec > 0 intrinsically, but this is impossible
since it does not appear in the classification (Theorem 6.30). More precisely:
(i) Show that the cyclic subgroup of G generated by (e−iπ /d , −I) for n even, and
by (e−iπ /d , I) for n odd is a normal subgroup in both G and K− . This is the
largest normal subgroup common to G and K− hence the ineffective kernel of
the G-action on G/K− ;
(ii) Use that eiπ ∈ H to show that the action of eiθ ∈ SO(2) ⊂ K− on the normal
slice D− is given by a rotation of angle 2θ ;
(iii) Conclude that the singular orbit G/K− is the fixed point set of a group of
isometries, hence a totally geodesic submanifold by Proposition 3.93. Now,
use Theorem 6.30 to obtain a contradiction if Md2n−1 had an invariant metric
with sec > 0.
Given that cohomogeneity one actions are polar with section γ (R), an important
object of study is the corresponding Weyl group (also called polar group), i.e.,
the group W such that M/G = γ (R)/W , see Remark 4.43. In our case, the
cohomogeneity one Weyl group W is the stabilizer of the normal geodesic γ modulo
its kernel H. More precisely, W is the dihedral subgroup of N(H)/H generated by
the unique involutions w± ∈ (N(H) ∩ K± )/H. Since γ (R)/W = M/G = [−1, 1],
the Weyl group is finite if and only if γ is a closed geodesic, in which case |W |
is precisely twice the number of points where γ intersects the singular orbits, cf.
Remark 4.40. It also follows that the nonprincipal isotropy groups along γ (R) are
conjugate to one of K± via an element of W . Although we do not carry out a longer
discussion of the Weyl group in cohomogeneity one here, it is a fundamental tool in
classification results such as Theorem 6.66.
We now study the curvature of cohomogeneity one manifolds. Any plane tangent
to a cohomogeneity one manifold at γ (t) is spanned by orthonormal vectors of the
form X and Y + α γ̇ (t), where X and Y are vertical vectors, i.e., tangent to the
hypersurface G(γ (t)). The sectional curvature of such a plane can be written as
 
sec X,Y + α γ̇ (t) =R(X,Y, X,Y ) + 2α R(X,Y, X , γ̇ (t))
+ α 2 R(X, γ̇ (t), X , γ̇ (t)).
170 6 Low Cohomogeneity Actions and Positive Curvature

By G-invariance, any plane tangent to M at a regular point can be moved to a plane

of the above form via an isometry in G. Thus, computing the sectional curvatures
of these planes tangent at γ (t) is sufficient to compute the sectional curvature of
all planes at regular points. Using Proposition 6.26 and the Gauss equation (2.7),
one can compute each of the terms in the right-hand side of the above formula. We
follow the same notation as in the previous section, where B± : m×m → g is defined
as (6.23), but is now t-dependent, just like P : m → m. For convenience, we drop the
subindex t. The outcome of this computation is as follows, see [116, Cor 1.10].
Proposition 6.58. The sectional curvatures of a cohomogeneity one metric g at
γ (t), t ∈ (−1, 1), are determined by

R(X,Y, X,Y ) = Q(B− (X,Y ), [X,Y ]) − 34 Q(P[X,Y ]m , [X ,Y ]m )

+Q(B+ (X,Y ), P−1 B+ (X ,Y ))
−Q(B+ (X, X), P−1 B+ (Y,Y ))
+ 14 Q(P X,Y )2 − 14 Q(P X , X) Q(PY,Y )
R(X,Y, γ̇ (t),Y ) = 34 Q([X,Y ], PY )
− 12 Q(P X, P−1 B+ (Y,Y ))
+ 12 Q(PY, P−1 B+ (X,Y ))
  
R(X, γ̇ (t), X, γ̇ (t)) = Q − 12 P + 14 P P−1 P X, X .

Exercise 6.59 (). Using Exercise 6.45 and Proposition 6.58, compute the sectional
curvatures of a diagonal cohomogeneity one metric, i.e., with gt as in (6.29). The
answers can be found in [116, Cor 1.13].
Hint: It suffices to compute sec(X, γ̇ (t)) for X ∈ mi and sec(X,Y ) for X,Y ∈ mi or X ∈ mi and Y ∈
m j . To simplify the computations, assume X and Y are Q-orthogonal and use that Q([mi , m j ], h) = 0
by skew-symmetry of ad(X), see Proposition 2.26 (i).

Example 6.60. Let M ⊂ R3 be a surface of revolution as in Example 6.47, with the

cohomogeneity one metric g given by (6.32). There is a unique plane tangent at each
point of M, and its sectional curvature coincides with the Gaussian curvature of M,
see Remark 2.21. Recall that Pr = φ (r)2 , so if X ∈ Tγ (r) M is a vector normal to γ̇ (r)
with Q(X, X) = 1, then by Proposition 6.58, R(X , γ̇ (r), X, γ̇ (r)) = −φ φ  . Thus, the
curvature of (M, g) at γ (r) is given by

  φ  (r)
sec X, γ̇ (r) = − , (6.43)
φ (r)

since g(X, X) = φ 2 . By the rotational symmetry, all points F(r, θ ) ∈ M have the
same curvature (6.43) as F(r, 0).
6.4 Positive and Nonnegative Curvature via Symmetries 171

Exercise 6.61. Consider a surface of revolution (M, g) as in Examples 6.47

and 6.60. Integrating (6.43), verify that the total curvature of (M, g) is
  
sec(Tp M) volg = 2π φ  (0) − φ  (R) . (6.44)
M

Since χ (M) = χ (S2 ) = 2, conclude that the smoothness conditions φ  (0) = 1 and
φ  (R) = −1 are necessary for the Gauss-Bonnet Theorem to hold in its usual smooth
formulation.
Hint: The volume form of M is volg = φ (r) dr dθ .

Exercise 6.62 (). Use Exercise 6.46 and Examples 6.47 and 6.60 to compute the
sectional curvatures of the Cheeger deformation of the round sphere (S2 , g) and the
flat plane (R2 , g) with respect to the rotation action by S1 , verifying that

  3t   1 + 2t + t cos(2r)
secgt Tγ (r) R2 = , and secgt Tγ (r) S2 = ,
(1 + tr2 )2 (1 + t sin2 r)2

where γ (r) is a radial unit speed geodesic in this surface of revolution, and γ (0) is a
fixed point. Use the above to verify the claims in Example 6.5.
As mentioned in the previous section, the converse statement to the Bonnet-
Myers Theorem 2.19 holds for compact homogeneous spaces, see Remark 6.29.
The fact that such converse statement also holds for compact cohomogeneity one
manifolds was proved by Grove and Ziller [116].
Theorem 6.63. A compact cohomogeneity one manifold M admits an invariant
metric with Ric > 0 if and only if π1 (M) is finite.
We conclude this section mentioning an important principle that allows to find
applications of cohomogeneity one actions in Geometric Analysis, based on the
fact that many geometrically motivated partial differential equations (PDEs) on
manifolds are invariant under isometries. Any such PDE on a manifold (M, g)
with a cohomogeneity one G-action reduces to a differential equation on the 1-
dimensional orbit space M/G, which is hence an ordinary differential equation
(ODE), cf. Remark 6.20. This means that cohomogeneity one is a useful Ansatz
to simplify the search for special solutions to PDEs with many symmetries. This
approach has been extensively used to study the Einstein equation Ric = λ g, see
[41–43, 74, 83], and some of its generalizations, such as Ricci solitons, see [73, 75].

6.4 Positive and Nonnegative Curvature via Symmetries

In this section, we provide a survey on manifolds with sec > 0 and sec ≥ 0 that
have many symmetries, in the sense of admitting an isometric action with low
cohomogeneity. We first give an overview of the main results in the area, and then
172 6 Low Cohomogeneity Actions and Positive Curvature

discuss some recent results in Bettiol [34] and Bettiol and Mendes [35, 36]. The
former is a construction related to the Hopf Problem, where cohomogeneity one
and conformal techniques are used to produce metrics on S2 × S2 satisfying an
intermediate curvature condition between Ric > 0 and sec > 0, called sec0+ > 0.
The latter are related to an intermediate curvature condition between sec > 0 and
positive-definiteness of the curvature operator, known as strongly positive curvature.
The study of closed manifolds with sec > 0 is one of the most challenging areas
in Riemannian geometry. Despite being a classical subject, surprisingly little is
understood about this class. In addition, some of the oldest open problems in global
Riemannian geometry regard manifolds with sec > 0, such as the celebrated:
Hopf Problem 6.64. Does S2 × S2 admit a Riemannian metric with sec > 0?
On the one hand, there are very few known topological obstructions to sec > 0.
On the other hand, it is notoriously difficult to construct examples of closed
manifolds with sec > 0 different from the compact rank one symmetric spaces
(CROSS) discussed in Example 6.10. Despite these difficulties, many advances in
the area were achieved through the use of symmetries.
Since new examples may be discovered in the process of classifying manifolds
with sec > 0 with certain symmetries, classification results such as Theorem 6.30
play a very important role. Symmetries can also be used as an aid to settle open
problems under stronger hypotheses, providing evidence for the general case. The
archetypical example is the following result of Hsiang and Kleiner [131], which
settles the above Hopf Problem 6.64 under the assumption of continuous symmetry.
Theorem 6.65. Let M 4 be an orientable compact Riemannian manifold with
sec > 0. If M has a nontrivial isometric S1 -action, then M is homeomorphic to S4
or CP2 . In particular, if S2 × S2 has a metric with sec > 0, then its isometry group
is finite.
Motivated by the above ideas, Grove [107] formulated a classification program
for manifolds with sec > 0 or sec ≥ 0 and many symmetries, which led to several
interesting developments in the last decade. A key feature in this program is
the ambiguity in what it means for a Riemannian manifold (M, g) to have many
symmetries. Among the various notions of highly symmetric manifolds, in this
text we focus exclusively on manifolds with low cohomogeneity, i.e., admitting an
isometric action by a Lie group G so that the orbit space M/G has low dimension.
We stress that this is only one particular aspect of the above program, and that we
do not address many remarkable results in the area that have been obtained under
other interpretations of what it means to have many symmetries. Among these,
we highlight manifolds with sec > 0 and large symmetry rank,18 on which there
is extensive literature, see [22, 23, 92, 93, 109, 110, 138, 139]. Further details on
these other developments can be found in the surveys of Wilking [231] and Ziller
[236], which also informed our approach to sec > 0 in low cohomogeneity.

18 By definition, the symmetry rank of a manifold (M, g) with an isometric G-action is rank G.
6.4 Positive and Nonnegative Curvature via Symmetries 173

The lowest possible cohomogeneity is zero, which means M/G is a point, i.e., the
G-action is transitive. In this case, the only simply-connected manifolds that admit
an invariant metric with sec > 0 are listed in Theorem 6.30. Furthermore, optimal
pinching constants for homogeneous metrics have also been thoroughly studied, see
Sect. 6.2.
The next case is that of manifolds with a cohomogeneity one action, described
in Sect. 6.3. In dimensions ≤ 6, Searle [193] showed that a cohomogeneity one
manifold carrying an invariant metric with sec > 0 must be diffeomorphic to Sn
or CPn/2 . The current state-of-the-art result on cohomogeneity one manifolds with
sec > 0 is the following almost19 classification, due to the efforts of Verdiani [218],
Grove, Wilking and Ziller [114] and Verdiani and Ziller [220].
Theorem 6.66 (Cohomogeneity one pre-classification of sec > 0). Apart from
the CROSS (see Example 6.10), the only simply-connected closed manifolds that
can have an invariant cohomogeneity one metric with sec > 0 are equivariantly
diffeomorphic to:
• Berger space B7 = SO(5)/SO(3);
• Eschenburg spaces E p7 = SU(3)/S1p ;
p = SU(5)/S p Sp(2);
Bazaikin spaces B13
• 1

• One of the candidates Pk7 , or Q7k .

Details on the Eschenburg spaces E p7 and Bazaikin spaces B13p can be found in
[236]. These are biquotients (another generalization of homogeneous spaces), and
were already known to admit metrics with sec > 0. Similarly, many other manifolds
in the list (CROSS and the Berger space B7 ) were known to have a homogeneous
metric with sec > 0 (i.e., they occur in Theorem 6.30). However, note that this
does not necessarily imply that they also have invariant metrics with sec > 0 and
cohomogeneity one. The remaining manifolds Pk7 and Q7k are infinite families of
candidates to have an invariant metric with sec > 0. Their group diagrams are,
respectively,
 iθ iθ    
H =Δ Q ⊂ e , e · H, e j(1−2k)θ , e j(1+2k)θ · H ⊂ Sp(1) × Sp(1), and
    
H ={(±1, ±1), (±i, ±i)} ⊂ eiθ , eiθ · H, e j(k+1)θ , e− jkθ · H
⊂ Sp(1) × Sp(1),

where Δ Q denotes a diagonal embedding of the group Q := {±1, ±i, ± j, ±k}.

Regarding constructing invariant metrics of positive curvature on Pk7 and Q7k , the
case k = 1 is trivial, since P17 = S7 and Q71 = W1,1
7 are manifolds known to have such

metrics. Recently, a construction of an invariant metric with positive curvature on

P27 was achieved independently by Dearricott [77] and Grove, Verdiani and Ziller

19 Thisis not a complete classification, as the candidate manifolds Pk7 and Q7k are not yet known to
carry cohomogeneity one metrics with sec > 0, as we explain below.
174 6 Low Cohomogeneity Actions and Positive Curvature

[112]. In the latter, the manifold P27 was also identified to be an exotic T1 S4 , i.e.,
homeomorphic (but not diffeomorphic) to the unit tangent bundle T1 S4 of the sphere
S4 . Constructing invariant metrics with positive curvature on Pk7 , k ≥ 3, and Q7k ,
k ≥ 2, – or proving that they do not exist – remains a very important and open
problem in the area. Finally, we remark that the original classification result in
[114] contained one more candidate in dimension 7, denoted R7 , that was very
recently ruled out by Verdiani and Ziller [220]. This sequence of results illustrates
a successful instance of the above mentioned Grove program, where a new example
was discovered in the process of classifying cohomogeneity one positively curved
manifolds.
Very little is known regarding the next steps for sec > 0 in low cohomogeneity,
even cohomogeneity two. One of the major obstacles is that actions of cohomo-
geneity ≥ 2 need not be polar, cf. Theorem 5.39. One of the first steps to understand
such actions is to recognize the possible orbit spaces. For example, cohomogeneity
2 is still reasonably mild, since all corresponding orbit spaces are orbifolds, see,
e.g., [157]. A complete list of possible orbit spaces in cohomogeneity 2 can be
found in Yeager [233]. However, checking whether each of these indeed occurs
and classifying the corresponding manifolds seems currently elusive in this general
framework.
The best result available regarding sec > 0 in higher cohomogeneity is due
to Wilking [229], who ingeniously showed that any new examples with large
dimension can only occur with large cohomogeneity.
Theorem 6.67. If M n admits a metric with sec > 0 invariant under an isometric
group action of cohomogeneity k ≥ 1 with n > 18(k + 1)2 , then M is homotopy
equivalent to a CROSS.
Instead of increasing the cohomogeneity, one can relax the curvature positivity
condition sec > 0 to sec ≥ 0. This substantially enlarges the class of known
examples. For instance, in cohomogeneity one, the following result was proved by
Grove and Ziller [115]:
Theorem 6.68. Any cohomogeneity one manifold with codimension two singular
orbits (i.e., ± = 1) admits an invariant metric with sec ≥ 0.
Proof (Sketch). The construction of such invariant metrics with sec ≥ 0 is through
a gluing procedure, inspired by the work of Cheeger [70], where Cheeger deforma-
tions were used to prove that the connected sum of any two CROSS admits sec ≥ 0.
The key lemma for extending this to the cohomogeneity one setup is showing that
if H ⊂ K ⊂ G are Lie groups with K/H = S1 = ∂ D2 and Q is a bi-invariant metric
on G, then the disk bundle G ×K D2 has a G-invariant metric with sec ≥ 0 that is
a product near the boundary G ×K S1 = G/H, which is endowed with the normal
homogeneous metric induced by Q. Then, if both ± = 1, one can equip each half
G ×K± D± with such a metric and glue them along G/H, following (6.25).
To prove the above key lemma, one needs the crucial fact that if G is a compact
Lie group with bi-invariant metric Q and a ⊂ g is an abelian Lie subalgebra,
6.4 Positive and Nonnegative Curvature via Symmetries 175

then Qa := a Q|a + Q|a⊥ has sec ≥ 0 as long as 0 < a < 43 . This is where the
hypotheses ± = 1 come into play, as the subalgebra corresponding to the tangent
space to K± /H = S± is abelian provided it is one-dimensional. With that in
mind, consider a product metric on G × D, where G has a metric Qa and the
2-disk D has a K-invariant metric with sec ≥ 0, e.g., g f = dt 2 + f (t)2 dθ 2 , where
f is concave. Then Qa + g f induces a metric ga, f with sec ≥ 0 on the quotient
G ×K D. A deformation argument implies that given 1 < a ≤ 43 , one can choose f
so that ga, f is product near the boundary G/H, where the induced metric is normal
homogeneous. 

Remark 6.69. Given that the above metrics on the disk bundles G ×K D are product
near the boundary, the resulting metric on the cohomogeneity one manifold M has
totally geodesic principal orbits. This is a very strong property, e.g., combined with
sec ≥ 0 it implies that each of the disk bundles G ×K± D± are convex subsets of M.
Also, it follows that tangent planes to M spanned by one direction tangent to a totally
geodesic principal orbit and one direction orthogonal to it are flat. By convexity
arguments, these flat planes must then occur at every point of M.
Although the condition ± = 1 seems somewhat restrictive, the corresponding
class of cohomogeneity one manifolds is surprisingly large. In fact, using Theo-
rem 6.68, Grove and Ziller [115] proved the following sequence of results.
Theorem 6.70. Every principal SO(k)-bundle P over S4 carries a cohomogeneity
one action by SO(3) × SO(k) with codimension two singular orbits and hence an
invariant metric with sec ≥ 0.
Thus, every associated bundle P ×SO(k) M (see Theorem 3.51) also carries a
metric with sec ≥ 0, where M is a manifold with sec ≥ 0 on which SO(k) acts
by isometries. Indeed, since secP ≥ 0, the product P × M also has sec ≥ 0 and by
the Gray-O’Neill formula (2.14), the quotient P ×SO(k) M has sec ≥ 0. This yields:
Corollary 6.71. The total space of every sphere bundle and vector bundle over S4
carries a metric with sec ≥ 0.
A particularly interesting consequence regards the Milnor spheres, which are S3 -
bundles over S4 and were the first exotic spheres to be discovered.
Corollary 6.72. At least 10 of the 14 (unoriented) exotic 7-spheres carry a metric
with sec ≥ 0. Moreover, on each of the 4 (oriented) diffeomorphism types homotopy
equivalent to RP5 there exist infinitely many non-isometric metrics with sec ≥ 0.
Although the above construction gives rise to many metrics of cohomogeneity
one and sec ≥ 0, Grove, Verdiani, Wilking and Ziller [111] proved the following
negative counterpart result when the singular orbits have larger codimension.
Theorem 6.73. For any pair of integers (− , + ) = (1, 1), with ± ≥ 1, there
is an infinite family of cohomogeneity one manifolds M with singular orbits of
codimensions ± + 1 and no invariant metric with sec ≥ 0.
176 6 Low Cohomogeneity Actions and Positive Curvature

Remark 6.74. Among the above family are the Brieskhorn varieties Md2n−1 , n ≥ 4,
d ≥ 3, including the Kervaire spheres (discussed in Example 6.56), which have
(− , + ) = (2, n) and are shown not to have SO(2)SO(n)-invariant metrics with
sec ≥ 0 (cf. Exercise 6.57).
Remark 6.75. Although we are not discussing noncompact manifolds with sec ≥ 0
here, we point out that analogous results to Corollary 6.71 regarding vector bundles
over the remaining known closed 4-manifolds with sec ≥ 0 were proved in [117].
Let us return to questions related to the Hopf Problem 6.64. Very recently, Grove
and Wilking [113] improved the conclusion of Theorem 6.65 from homeomorphism
to equivariant diffeomorphism, in an ingenious proof that uses deep aspects
of Alexandrov geometry and the solution of the Poincaré conjecture. Besides
Theorem 6.65, Kleiner (in his PhD thesis [141]) proved that the only compact
orientable 4-manifolds that admit a metric with sec ≥ 0 invariant under an isometric
2
S1 -action are homeomorphic to S4 , CP2 , S2 × S2 , CP2 #CP2 or CP2 #CP . This
conclusion was also improved from homeomorphism to equivariant diffeomorphism
in [113].
Altogether, any potential metric with sec > 0 on S2 × S2 cannot have contin-
uous symmetry, let alone cohomogeneity one. Nevertheless, cohomogeneity one
techniques, combined with conformal methods, can be used to construct metrics
on S2 × S2 that satisfy a curvature positivity condition stronger than the conditions
Ric > 0 and sec ≥ 0 satisfied by the standard product metric. We now describe
this construction, following Bettiol [34]. Let (M, g) be a Riemannian manifold and
choose a distance (inducing the standard topology) on the Grassmannian bundle
Gr2 TM of 2-planes tangent to M.20 For each θ > 0 and σ ⊂ Tp M, set
 
secθg (σ ) := min 1
2 secg (σ ) + secg (σ  ) .
σ  ⊂Tp M
dist(σ ,σ  )≥θ

The condition secθg > 0 means that, at every point p ∈ M, the average of sectional
curvatures of any two planes σ , σ  ⊂ Tp M that are at least θ > 0 apart from each
other is positive. Intuitively, one can think of θ as a lower bound for the “angle” or
“aperture” between the planes considered. Define the following curvature condition:
(sec0+ > 0) For all θ > 0 there exists a Riemannian metric gθ on M with
secθgθ > 0, and gθ converges to a metric g0 as θ 0.

Any Riemannian manifold with sec > 0 clearly satisfies sec0+ > 0. Moreover,
routine arguments show that manifolds with sec0+ > 0 admit metrics with Ric > 0,
proving that sec0+ > 0 is an intermediate condition between Ric > 0 and sec > 0:

20 Forinstance, given σ , σ  ⊂ Tp M, one can define dist(σ , σ  ) as the Hausdorff distance between
the great circles σ ∩ S and σ  ∩ S, where S is the unit sphere in Tp M.
6.4 Positive and Nonnegative Curvature via Symmetries 177

Proposition 6.76. Let M be a compact manifold satisfying sec0+ > 0. Then M

admits metrics with Ric > 0, since Ricgθ > 0 for θ > 0 sufficiently small. In
particular, if gθ → g0 also in the C2 -topology, then Ricg0 ≥ 0 and secg0 ≥ 0.
Proof (Sketch). For any metric g on M, define:
⎛ ⎛ ⎞⎞
⎜ ⎜  ⎟⎟
⎜ ⎜ ⎟⎟
θg := min ⎜ min ⎜ min dist span{v, w1 }, span{v, w2 } ⎟⎟
⎜ ⎜ ⎟
⎟.
p∈M ⎝ v∈Tp M, ⎝ w1 ,w2 ∈Tp M, ⎠⎠
g(v,v)=1 g(v,w j )=0,
g(wi ,w j )=δi j

The above defines a positive number depending continuously on the metric g, and
if secθg > 0 for some 0 < θ ≤ θg , then Ricg > 0. Given G := {gθ : θ ∈ [0, 1]}, let
θ∗ := min{θg : g ∈ G }. It then follows that θ∗ > 0 and hence for any 0 < θ ≤ θ∗ ,
we have Ricgθ > 0. 

Recall that the standard product metric g0 on S2 × S2 satisfies the Einstein

equation Ricg0 = g0 , and, in particular, has Ric > 0. However, at every point
p ∈ (S2 × S2 , g0 ) there exist planes σ , σ  ⊂ Tp (S2 × S2 ) arbitrarily close to each
other with secg0 (σ ) = secg0 (σ  ) = 0. Namely, mixed planes, i.e., planes spanned by
vectors of the form (X, 0) and (0,Y ), are g0 -flat and form a continuous family at
each point. Following Bettiol [34], we now outline a deformation gθ , θ > 0, of the
product metric g0 that perturbes these flat planes in such way that secθgθ > 0.

Theorem 6.77. The manifold S2 × S2 satisfies sec0+ > 0.

Proof (Sketch). For each θ > 0, the construction of gθ has two steps: a Cheeger
deformation and a first-order local conformal deformation. The first step of this
construction follows the work of Müter [172, Satz 4.26] (see also Ziller [234, p. 5]
and Kerin [140, Rem 4.3]).
Let g0 be the standard product metric on S2 × S2 . There is an isometric diagonal
SO(3)-action on (S2 × S2 , g0 ) given by

A · (p1 , p2 ) := (A p1 , A p2 ), p = (p1 , p2 ) ∈ S2 × S2 ⊂ R3 ⊕ R3 , A ∈ SO(3).

This is a cohomogeneity one action with trivial principal isotropy, corresponding to

p1 = ±p2 ; and singular isotropies isomorphic to SO(2), corresponding to p1 = ±p2 .
Thus, the group diagram is {1} ⊂ {SO(2), SO(2)} ⊂ SO(3), cf. Exercise 6.55. The
singular orbits of this action are the diagonal and anti-diagonal submanifolds:

±Δ S2 := (p1 , ±p1 ) : p1 ∈ S2 ⊂ S2 × S2 .

A careful study of the vertical and horizontal distributions of this cohomogeneity

one action reveals the key fact that, at a regular point on (S2 × S2 , g0 ), there is a
unique mixed plane that contains a horizontal direction; and, at a singular point,
178 6 Low Cohomogeneity Actions and Positive Curvature

there is a circle’s worth of such planes. It follows from Corollary 6.7 that the
Cheeger deformation (S2 × S2 , gt ), t > 0, has a unique zero curvature plane at
each p ∈ / ±Δ S2 and a circle’s worth of zero curvature planes at each p ∈ ±Δ S2 ,
and all other planes have positive curvature. Thus, for any θ > 0, the metric gt ,
t > 0, satisfies secθgt ≥ 0 and equality only holds for certain planes tangent at some
p = (p1 , ±p1 ) ∈ ±Δ S2 . Moreover, these planes are not tangent to ±Δ S2 .
The second step of the construction consists of perturbing the above Cheeger
deformed metric gt , t > 0, with an appropriate symmetric (0, 2)-tensor h, in a first-
order deformation inspired by the work of Strake [199, 200]. Namely, consider the
deformation gs,t := gt + s h of gt . For a fixed θ > 0, in order for secθgs,t > 0 for any
 
s > 0 sufficiently small, it suffices that ds
d
secgs,t (σ1 ) + secgs,t (σ2 ) s=0 > 0, for all
σ1 , σ2 ⊂ Tp (S2 × S2 ) with p ∈ ±Δ S2 and secgt (σ1 ) = secgt (σ2 ) = 0. Indeed, this
can be verified using the first-order Taylor expansion of secθgs,t in s and the fact that
secθgt ≥ 0 with zeros only along ±Δ S2 . Choosing h = −φ gt , we have

d
ds secgs,t (σ )s=0 = 12 Hess φ (X, X) + 12 Hess φ (Y,Y ), (6.45)

where the Hessians are with respect to gt . Since the planes σ ⊂ Tp (S2 × S2 ) with
secgt (σ ) = 0 are never tangent to ±Δ S2 , one can choose the function φ to be the
sum of the square distances to ±Δ S2 , multiplied by an appropriate cutoff function,
in order for the above sufficient condition to hold. Since s > 0 and t > 0 can be
chosen arbitrarily small, given any θ > 0 there are s > 0 and t > 0 such that gs,t is
arbitrarily close to g0 and has secθgs,t > 0, yielding the desired metrics gθ that prove
that S2 × S2 satisfies sec0+ > 0. 

Remark 6.78. The above metrics gθ on S2 × S2 with secθgθ > 0 do not have sec ≥ 0,
see Bettiol [34, Sec 4.4]. It can also be shown that such metrics can be constructed
invariant under the antipodal action of Z2 ⊕ Z2 , whose quotient is RP2 × RP2 .
Recall that from Synge’s Theorem (see Petersen [183, p. 172]), RP2 × RP2 does
not admit any metric with sec > 0.
Remark 6.79. The question of deforming the standard product metric g0 to improve
curvature has been studied in great depth by Bourguignon, Deschamps and
Sentenac [50, 51], after Berger [31] elegantly  proved that any first-order
deformation gs = g0 + s h with ds d
secgs (σ )s=0 ≥ 0 for all mixed planes σ must

d
satisfy ds secgs (σ )s=0 = 0. Despite interesting findings regarding deformations of
higher order in [50, 51] and Theorem 6.77, even this local version of the Hopf
Problem 6.64 (that asks if metrics with sec > 0 on S2 × S2 exist arbitrarily close to
g0 ) remains unanswered.
Remark 6.80. The standard product metric on S2 × S2 can be written as a cohomo-
geneity one metric (6.27), by using the embedding S2 × S2 ⊂ R3 × R3 and choosing
as horizontal geodesic, e.g.,
  !
γ (τ ) = (cos τ )e1 + (sin τ )e2 , (cos τ )e1 − (sin τ )e2 , τ ∈ 0, π2
6.4 Positive and Nonnegative Curvature via Symmetries 179

where {e1 , e2 , e3 } is the canonical basis of R3 . At each γ (τ ) ∈ S2 × S2 , 0 < τ < π2 ,

we can identify the tangent space to the orbit through γ (τ ) as m = so(3) ∼ = R3 , and
this identification is isometric when so(3) is endowed with the standard bi-invariant
metric Q and R3 with the Euclidean metric. Then the standard product metric g0 is
given by g0 = dτ 2 + gτ , and gτ is related to the Euclidean metric Q via the tensor
Pτ : m → m, given by
⎛ ⎞
2 sin τ
Pτ = ⎝ 2 cos τ ⎠ .
2
In particular, the Cheeger deformation gt of g0 can be explicitly written in the
form gt = dτ 2 + gt,τ , by using Exercise 6.46 and the above formula for Pτ .
However, verifying the curvature properties claimed above is much easier through
Proposition 6.6, rather than attempting an explicit computation.
We conclude this section with a summary of recent developments regard-
ing another intermediate curvature positivity condition, called strongly positive
curvature, between sec > 0 and positive-definiteness of the curvature operator.
A systematic study of this condition, under the light of the Grove program, was
recently initiated by Bettiol and Mendes [36], though this notion is reminiscent of
the work of Thorpe [211, 212] in the 1970s and was also used by authors including
Pütmann [187], and Grove, Verdiani and Ziller [112], who coined the term. In order
to define this condition, recall that the curvature operator R : Λ 2 Tp M → Λ 2 Tp M of
a Riemannian manifold (M, g) is the symmetric endomorphism of Λ 2 Tp M given by

R(X ∧Y ), Z ∧W  := R(X,Y, Z,W ),

where ·, · denotes the inner product on Λ 2 Tp M induced by g, and R is defined

in (2.3). Any 2-plane σ ⊂ Tp M can be seen as an element21 X ∧ Y ∈ Λ 2 Tp M,
by choosing orthonormal vectors X,Y ∈ Tp M that span σ . From this viewpoint,
sectional curvature sec(σ ) = R(σ ), σ  is the quadratic form associated to R,
restricted to the (oriented) Grassmannian Gr2 (Tp M) = {X ∧ Y ∈ Λ 2 Tp M : X ∧
Y 2 = 1}. Any 4-form ω ∈ Λ 4 Tp M also induces a symmetric operator on Λ 2 Tp M,
given by

ω : Λ 2 Tp M → Λ 2 Tp M, ω (α ), β  := ω , α ∧ β . (6.46)

Note that the quadratic form associated to ω clearly vanishes on σ ∈ Gr2 (Tp M),
since σ ∧ σ = 0. Thus, it follows that

sec(σ ) = R(σ ), σ  = (R + ω )(σ ), σ . (6.47)

21 Forthe sake of simplifying notation, here we use the metric g to identify (0, k)-tensors and k-
forms, that is, Λ k TM ∼
= Λ k TM ∗ . Furthermore, we omit the parentheses, denoting Λ kV := Λ k (V ).
180 6 Low Cohomogeneity Actions and Positive Curvature

This observation, known as Thorpe’s trick, implies that if there exists ω ∈ Λ 4 Tp M

such that the so-called modified curvature operator (R + ω ) : Λ 2 Tp M → Λ 2 Tp M is
a positive-definite operator, then sec(σ ) > 0 for all planes σ ⊂ Tp M.
Definition 6.81. A Riemannian manifold (M, g) has strongly positive curvature if
for all p ∈ M, there exists ω ∈ Λ 4 Tp M such that (R + ω ) : Λ 2 Tp M → Λ 2 Tp M is
positive-definite. Analogously, (M, g) has strongly nonnegative curvature if for all
p ∈ M, there exists ω ∈ Λ 4 Tp M such that (R + ω ) is positive-semidefinite.
Remark 6.82. In the above definition, the assignment M  p → ω p ∈ Λ 4 Tp M may,
in principle, fail to be smooth. However, if (M, g) has strongly positive curvature,
then a standard perturbation argument shows that there exists a smooth 4-form ω ∈
Ω 4 (M) such that R + ω  is positive-definite. The same argument does not work for
strongly nonnegative curvature, as it is not an open condition, see [36, Sec 6.4].
In view of (6.47), strongly positive curvature is an intermediate condition
between sec > 0 and positive-definiteness of the curvature operator. As mentioned
before, manifolds that satisfy the latter are known to be diffeomorphic to spherical
space forms by the work of Böhm and Wilking [45]. Thus, the notion of strongly
positive curvature might improve our understanding of the gap between finite
quotients of spheres and the intriguing class of manifolds that admit sec > 0.
In dimensions ≤ 3, this gap is well-understood, since sec > 0, strongly positive
curvature and positive-definiteness of the curvature operator are all equivalent.
Remarkably, strongly positive curvature and sec > 0 remain equivalent in dimension
4, providing a different viewpoint on the Hopf Problem 6.64. The proof of this
equivalence is originally due to Thorpe [211]; however, it can be greatly simplified
using an elegant argument communicated to us by Püttmann (see also [187]), as
follows.
Proposition 6.83. Let (M, g) be a Riemannian manifold with dim M ≤ 4. Then
(M, g) has strongly positive curvature if and only if it has sec > 0.
Proof. The only nontrivial implication is that if (M, g) has dim M = 4 and sec > 0,
then it also has strongly positive curvature. Fix p ∈ M and, for each modified curva-
ture operator (R + ω ) : Λ 2 Tp M → Λ 2 Tp M, denote by λ1 (R+ ω ) := min Spec(R+ ω )
its smallest eigenvalue. Consider the min-max problem:

λ := sup λ1 (R + ω ). (6.48)
ω ∈Λ 4 Tp M

It is easy to see that, since the endomorphisms ω : Λ 2 Tp M → Λ 2 Tp M are traceless,

the above supremum λ is actually achieved at some ωmax ∈ Λ 4 Tp M. Denote by Eλ
the subspace of Λ 2 Tp M formed by the eigenvectors of R + ωmax with eigenvalue
λ . If there exists σ ∈ Eλ ∩ Gr2 (Tp M), then λ = sec(σ ) > 0 and hence R + ωmax is
positive-definite. Otherwise, we would have q(α ) = 0 for all nonzero α ∈ Eλ , where

q : Λ 2 Tp M → Λ 4 Tp M ∼
= R, q(α ) := α ∧ α . (6.49)
6.4 Positive and Nonnegative Curvature via Symmetries 181

Notice that the image q(Eλ \ {0}) ⊂ R \ {0} is contained in a half-line, say R+ :=
{x > 0}. Indeed, this is clear if dim Eλ = 1, and if dim Eλ ≥ 2, then Eλ \ {0} is
connected and hence so is q(Eλ \ {0}). Therefore, for any nonzero α ∈ Eλ , we can
construct a new modified curvature operator R + ωmax + α ∧ α that satisfies
# $ # $
(R + ωmax + α ∧ α )(β ), β = (R + ωmax )(β ), β + q(α )β 2 > λ β 2 ,
(6.50)
for all nonzero β ∈ Λ 2 Tp M, contradicting the maximality (6.48) of λ . 

A crucial property of strongly positive curvature is that, similarly to sec > 0,

this condition is preserved under Riemannian submersions.22 This was proved by
Bettiol and Mendes [36], using a convenient rearrangement of the Gray-O’Neill
formula (2.14) for curvature operators, in terms of the first Bianchi identity.
Theorem 6.84. Let π : (M, g) → (M, g) be a Riemannian submersion. If (M, g) has
strongly positive curvature, then (M, g) also has strongly positive curvature.
Proof. Fix p ∈ M and p ∈ M such that π (p) = p, and identify each X ∈ Tp M as
an element X ∈ Tp M via horizontal lift. On these horizontal vectors, the tensor A is
given by AX Y = 12 [X,Y ]V , see (2.12). As such, it induces a symmetric operator

α : Λ 2 Tp M → Λ 2 Tp M, α (X ∧Y ), Z ∧W  := AX Y, AZ W , (6.51)

which is clearly positive-semidefinite. From the Gray-O’Neill formula, see (2.14)

and [33, Thm 9.28f], we have

R(X ∧Y ), Z ∧W  = R(X ∧Y ), Z ∧W  + 2AX Y , AZ W 

− AY Z, AX W  + AX Z, AY W 
= R(X ∧Y ), Z ∧W  + 3α (X ∧Y ), Z ∧W 
− α (Y ∧ Z), X ∧W  − α (Z ∧ X),Y ∧W 
− α (X ∧Y ), Z ∧W 
= R(X ∧Y ), Z ∧W  + 3α (X ∧Y ), Z ∧W 
− 3b(α )(X,Y , Z,W ),

22 Recallthat positive-definiteness of the curvature operator is not preserved under Riemannian

submersions. For instance, there are Riemannian submersions from the round sphere onto the
projective spaces CPn and HPn , but these only have positive-semidefinite curvature operator.
182 6 Low Cohomogeneity Actions and Positive Curvature

where b is the Bianchi map, that associates to each symmetric operator S : Λ 2 Tp M →

Λ 2 Tp M the 4-form b(S) ∈ Λ 4 Tp M given by23 :

b(S)(X,Y, Z,W ) := 1
3 S(X ∧Y ), Z ∧W  + S(Y ∧ Z), X ∧W 

+ S(Z ∧ X ),Y ∧W  .

Thus, if there exists ω ∈ Λ 4 Tp M such that R + ω is positive-definite, it follows that

also R + ω is positive-definite, by setting ω := ω + 3b(α ) ∈ Λ 4 Tp M. 

Notice that the above argument implies that strongly nonnegative curvature is
also preserved under Riemannian submersions. Furthermore, similar techniques can
be used to prove that strongly positive (and nonnegative) curvature are preserved
under Cheeger deformations [36, Thm B], cf. Proposition 6.1.
Propelled by the above foundational results, we can analyze the known construc-
tions of metrics with sec > 0 (and sec ≥ 0) to verify whether they yield metrics
with strongly positive (and nonnegative) curvature. Regarding the CROSS (recall
Example 6.10), it is a direct consequence of Theorem 6.84 that CPn and HPn
have strongly positive curvature, since there are Riemannian submersions from
round spheres given by the Hopf bundles (6.20). The same does not hold for
the Cayley plane CaP2 ; in fact, there are no homogeneous metrics with strongly
positive curvature on CaP2 , see [36, Prop. 3.4]. The main reason for this is that
CaP2 admits a unique F4 -invariant metric (see Example 6.21) and any 4-form ω
such that R + ω is positive-definite could be averaged to yield an F4 -invariant (and
hence harmonic) 4-form ω  such that R + ω  is positive-definite. However, CaP2
does not admit any nonzero harmonic 4-forms, since b4 (CaP2 ) = 0. In particular,
this provides an example of a Riemannian manifold that has sec > 0 but does not
have strongly positive curvature (algebraic examples were previously constructed
by Zoltek [237]).
In order to classify simply-connected homogeneous spaces with strongly positive
curvature beyond the CROSS, one may restrict to the list given in Theorem 6.30.
By extending previous results of Wallach [221] and Eschenburg [82] with the
introduction of strongly fat homogeneous bundles, Bettiol and Mendes [36, Thm C]
found homogeneous metrics with strongly positive curvature on the Aloff-Wallach
spaces Wk,l7 , the Berger spaces B7 and B13 , and the Wallach flag manifolds W 6 and

W . The only remaining case, the Wallach flag manifold W 24 , was recently settled
12

23 The Bianchi map b on symmetric operators on Λ 2 Tp M can be seen as the orthogonal projection
onto the subspace of operators induced by 4-forms via (6.46). The complement ker b of this
subspace consists of symmetric operators on Λ 2 Tp M that satisfy the first Bianchi identity, which
are called algebraic curvature operators.
6.4 Positive and Nonnegative Curvature via Symmetries 183

by Bettiol and Mendes [35], using different techniques.24 Altogether, in terms of the
homogeneous classification of sec > 0 (see Theorem 6.30), we have the following:
Theorem 6.85 (Homogeneous classification of strongly positive curvature). All
simply-connected homogeneous spaces that admit a homogeneous metric with
sec > 0 admit a homogeneous metric with strongly positive curvature, except for
CaP2 .
Besides homogeneous manifolds, there are currently only a few other known
examples of closed manifolds with sec > 0. Namely, these are biquotients (Eschen-
burg spaces E p7 and Bazaikin spaces B13 4
p ) and the exotic T1 S , which is the manifold
P27 in Theorem 6.66. An infinite family of Eschenburg spaces E p7 can be shown to
have strongly positive curvature by a limiting argument similar to Escheburg [81,
Sec 5], using the fact that all Aloff-Wallach spaces Wk,l7 with sec > 0 have strongly
4
positive curvature. Finally, the exotic T1 S was shown to have a cohomogeneity
one metric with sec > 0 in [112] by proving that this metric actually has strongly
positive curvature. Thus, the task of verifying whether all known examples of closed
manifolds with sec > 0 admit a metric with strongly positive curvature is almost
complete, and, so far, no exceptions have been found.25
Regarding which manifolds with sec ≥ 0 have strongly nonnegative curvature,
a very satisfactory start is provided by Theorem 6.84. Since a Lie group (G, Q)
with bi-invariant metric has positive-semidefinite curvature operator (recall Propo-
sition 2.26), it follows that every normal homogeneous space G/H has strongly
nonnegative curvature. Moreover, in view of [36, Thm B] and other recent work
of Bettiol and Mendes, many constructions using Cheeger deformations (including
structural results such as Theorem 6.23) carry over to the context of strongly
nonnegative curvature, including Theorem 6.68 and its Corollaries 6.71 and 6.72.

24 Moreover, a complete description of the moduli spaces of homogeneous metrics with strongly
positive curvature on all Wallach flag manifolds was achieved in [35]. Namely, it is shown that a
homogeneous metric on W 6 or W 12 has strongly positive curvature if and only if it has sec > 0,
while a homogeneous metric on W 24 has strongly positive curvature if and only if it has sec > 0
and does not submerge onto CaP2 endowed with its standard metric.
25 We stress that the only topological obstructions to strongly positive curvature currently known
are the obstructions for sec > 0.
Appendix A
Rudiments of Smooth Manifolds

In this appendix, we summarize basic definitions and results on smooth manifolds,

that are used throughout this book. A more detailed treatment of this subject can be
found in Warner [227], Spivak [198] or Lang [149].

A.1 Smooth Manifolds

In general, by smooth or differentiable we mean of class C∞ , unless otherwise

specified. In addition, a map is called a function if its counter-domain is R.
Definition A.1. A smooth n-dimensional manifold M is a second countable para-
compact Hausdorff topological space endowed with a smooth structure. More
precisely, there is a collection of pairs (Uα , ϕα ) called charts, such that Uα are
open subsets of M, and, for each α , the map ϕα : Uα ⊂ M → ϕα (Uα ) ⊂ Rn is a
homeomorphism between open sets satisfying:

(i) M = α Uα ;
(ii) If (Uα , ϕα ) and (Uβ , ϕβ ) are charts, then the transition map

ϕβ ◦ ϕα−1 : ϕα (Uα ∩Uβ ) −→ ϕβ (Uα ∩Uβ )

is smooth. In this case, the charts ϕα and ϕβ are said to be C∞ -compatible;

(iii) The collection {(Uα , ϕα )}α of charts is maximal with respect to (i) and (ii).
Such collection of charts is called an atlas of M. If (Uα , ϕα ) is a chart and p ∈ Uα ,
the open set Uα is called a coordinate neighborhood of p and ϕα is called a local
coordinate at p. Furthermore, we write n = dim M and refer to M as an n-manifold.
Similarly, it is possible to define Ck -manifolds and analytic manifolds (Cω -
manifolds), by requiring the transition maps to be respectively Ck and analytic.

© Springer International Publishing Switzerland 2015 185

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1
186 A Rudiments of Smooth Manifolds

Thus, by analytic structure, we mean a collection of coordinate systems which

overlap analytically, that is, are locally represented by convergent power series.
In this book, we are only interested in C∞ -manifolds, therefore a manifold is
always considered to be a smooth manifold, in the above sense.
Definition A.2. Let M and N be manifolds, and denote their atlases by {(Uα , ϕα )}α
and {(Vβ , ψβ )}β , respectively. A continuous map f : M → N is said to be smooth
at p ∈ M if there exist coordinate neighborhoods Uα of p and Vβ of f (p), such that
ψβ ◦ f ◦ ϕα−1 is smooth at ϕα (p), as a map between Euclidean spaces.
It is easy to see that if this condition is satisfied by a pair of charts (Uα , ϕα ) and
(Vβ , ψβ ), then it holds for any charts. The above definition clearly implies that a
function g : M → R is smooth at p ∈ M if there exists a chart (Uα , ϕα ) such that
g ◦ ϕα−1 is smooth at g(p). Finally, maps (and functions) are called smooth if they
are smooth at every point p ∈ M.
Let M be a smooth manifold and p ∈ M. The set C∞ (M) of smooth functions
f : M → R is an algebra with the usual operations. Consider the subalgebra C∞ (p)
of smooth functions whose domain of definition includes some open neighborhood
of p, called the algebra of germs of smooth functions at p. We define the tangent
space to M at p as the vector space Tp M of linear derivations at p, i.e., the set of
maps v : C∞ (p) → R satisfying for all f , g ∈ C∞ (p),
(i) Linearity, i.e., v( f + g) = v( f ) + v(g);
(ii) Leibniz rule, i.e., v(fg) = v( f )g(p) + f (p)v(g).
The vector space operations are defined for all v, w ∈ Tp M and α ∈ R by

(v + w) f := v( f ) + w( f ), (α v)( f ) := α v( f ), for all f ∈ C∞ (p).

Let (U, ϕ ), where ϕ = (x1 , . . . , xn ), be a chart around p and f := f ◦ ϕ −1 the

representation of f ∈ C∞ (p) in the coordinates given by ϕ . The coordinate vectors
 
∂  ∂ f 
 f := 
∂ xi p ∂ xi ϕ (p)
 
form a basis ∂∂xi  p of Tp M. In particular, it follows that dim Tp M = dim M. In this
context, tangent vectors can be explicitly seen as directional derivatives. Indeed, if
v = ∑ni=1 vi ∂∂xi ∈ Tp M, then

n
∂ f 
v( f ) = ∑ vi 
i=1 ∂ xi ϕ (p)

is called the directional derivative of f in the direction v. It can also be proved that
this definition is independent of the chart ϕ .
A Rudiments of Smooth Manifolds 187

The tangent
 vector to a curve α : (−ε , ε ) → M at α (0) is defined as α  (0)( f ) =

◦ α ) t=0 . Indeed, if ϕ is a local coordinate and (u1 (t), . . . , un (t)) = ϕ ◦ α , then
dt ( f
d

n
∂
α  (0) = ∑ ui (0) ∈ Tp M.
i=1 ∂ xi

Definition A.3. Let f : M → N be a smooth map and let p ∈ M. The differential, or

derivative, of f at p is the linear map d f p : Tp M → T f (p) N, such that if v ∈ Tp M, then
d f p v is a tangent vector at f (p) satisfying d f p (v)g = v(g ◦ f ), for all g ∈ C∞ ( f (p)).
In other words, the chain rule is automatically valid on manifolds, as the
differential of a map is defined to satisfy it. Equipped with the notion of differential,
we can define important classes of maps between manifolds.
Definition A.4. A smooth map f : M → N is said to be an immersion if d f p is
injective for all p ∈ M, and it is said to be a submersion if d f p is surjective for all
p ∈ M. An immersion f : M → N is called an embedding if f : M → f (M) ⊂ N is a
homeomorphism, considering f (M) endowed with the relative topology.
A smooth bijection f with smooth inverse is called a diffeomorphism. From the
chain rule, its differential on every point is an isomorphism, and a local converse is
given by the Inverse Function Theorem. More precisely, this theorem states that if
the differential d f p of a smooth map f is an isomorphism, then the restriction of f
to some open neighborhood of p is a diffeomorphism onto its image.
Definition A.5. Let P and N be manifolds, with P ⊂ N. Then P is called an
immersed submanifold of N if the inclusion i : P → N is an immersion. Moreover,
if i : P → N is an embedding, then P is called an (embedded) submanifold.
Definition A.6. An immersed submanifold P ⊂ N is quasi-embedded if for
any smooth map f : M → N with image f (M) lying in P, the induced map
f0 : M → P defined by i ◦ f0 = f is smooth, where i : P → N is the inclusion.
Proposition A.7. If P ⊂ N is an embedded submanifold, then P is quasi-embedded.

A.2 Vector Fields

Let M be a manifold and TM be the (total space of its) tangent bundle, i.e., TM :=

p∈M {p} × Tp M. Denote by π : TM → M the projection map, defined by π (v) = p
if v ∈ Tp M. It is possible to prove that TM carries a canonical smooth structure
inherited from M, such that π is smooth. In fact, as explained in Sect. 3.1, TM is a
fiber bundle; in particular, a vector bundle.
Definition A.8. A (smooth) vector field X on M is a (smooth) section of TM, that
is, a (smooth) map X : M → TM such that π ◦ X = id.
188 A Rudiments of Smooth Manifolds

Let p ∈ M and consider a coordinate neighborhood U of p with local coordinates

ϕ = (x1 , . . . , xn ). Given f ∈ C∞ (U), the directional derivative of f in the direction
of X is defined as

X( f ) : U → R, X( f )(p) := X p ( f ). (A.1)
 
∂ 
Moreover, if ∂ xi q is a basis of Tq M for all q ∈ U , then X can be written as

n
∂
X|U = ∑ ai , (A.2)
i=1 ∂ xi

where ai : U → R are functions.

Proposition A.9. Let X be a vector field on M. The following are equivalent:
(i) X is smooth;
(ii) For every chart (U, ϕ ), ϕ = (x1 , . . . , xn ), the functions ai in (A.2) are smooth;
(iii) For every open set V of M and f ∈ C∞ (V ), we have X( f ) ∈ C∞ (V ).
The set of smooth vector fields on M is denoted X(M), and is clearly a C∞ (M)-
module with operations defined pointwise. If f : M → N is a smooth map between
two manifolds, X ∈ X(M) and Y ∈ X(N), then X and Y are said to be f -related if
d f ◦ X = Y ◦ f , i.e., if the following diagram commutes.

df
TM TN

X Y

M N
f

Note that X and Y are f -related if and only if (Yg) ◦ f = X(g ◦ f ) for all g ∈ C∞ (N).
Definition A.10. If X,Y ∈ X(M), the Lie bracket of the vector fields X and Y is the
vector field [X,Y ] ∈ X(M) given by

[X,Y ] f := X(Y ( f )) −Y (X( f )), f ∈ C∞ (M).

Note that the Lie bracket of vector fields is clearly skew-symmetric and satisfies
the Jacobi identity (see Definition 1.7). If X,Y ∈ X(M) are such that [X,Y ] = 0, then
X and Y are said to commute.
If f : M → N is a smooth map such that X 1 , X 2 ∈ X(M) and Y 1 ,Y 2 ∈ X(N) are
f -related, then [X 1 , X 2 ] and [Y 1 ,Y 2 ] are also f -related. Indeed, using the above
observation, since X i is f -related to Y i , it follows that (Y i g) ◦ f = X i (g ◦ f ).
The conclusion is immediate from the following equation, using again the same
observation:
A Rudiments of Smooth Manifolds 189

     
[Y 1 ,Y 2 ]g ◦ f = Y 1 (Y 2 g) ◦ f − Y 2 (Y 1 g) ◦ f
   
= X 1 (Y 2 g) ◦ f − X 2 (Y 1 g) ◦ f
    (A.3)
= X 1 X 2 (g ◦ f ) − X 2 X 1 (g ◦ f )
= [X 1 , X 2 ](g ◦ f )

In order to obtain a local expression for [X ,Y ], fix p ∈ M and consider a chart

(U, ϕ ), ϕ = (x1 , . . . , xn ). According to Proposition A.9, there exist functions ai , b j ∈
C∞ (M) such that X = ∑ni=1 ai ∂∂xi and Y = ∑nj=1 b j ∂∂x j . Therefore
) *
∂n
XY = X ∑ b j
j=1 ∂xj
) *
n
∂ n
∂
= ∑ ai ∑ bj ∂xj
i=1 ∂ xi j=1
 
n n
∂bj ∂ ∂2
= ∑ ∑ ai + bj
i=1 j=1 ∂ xi ∂ x j ∂ xi ∂ x j
n
∂bj ∂ n
∂2
= ∑ ai + ∑ ai b j
∂ xi ∂ x j i, j=1 ∂ xi ∂ x j
,
i, j=1

and similarly,

n
∂ ai ∂ n
∂2
YX = ∑ bj + ∑ ai b j
∂ x j ∂ xi i, j=1 ∂ xi ∂ x j
.
i, j=1

Applying the definition and expressions above, we recover the well-known local
expression for the Lie bracket,
 
n
∂bj ∂aj ∂
[X,Y ]|U = ∑ ai
∂ xi
− bi
∂ xi ∂xj
. (A.4)
i, j=1

Remark A.11. If M ⊂ Rn is an open set, then X = (a1 , . . . , an ) and Y = (b1 , . . . , bn )

are such that [X,Y ] = ∇X Y − ∇Y X, where ∇ is the usual derivative of maps on Rn .
190 A Rudiments of Smooth Manifolds

A.3 Foliations and the Frobenius Theorem

In this section, we mention some facts about flows, distributions and foliations,
including the Frobenius Theorem.
Definition A.12. An integral curve of X ∈ X(M) is a smooth curve α : I → M, such
that α  (t) = X (α (t)) for all t ∈ I.
A classical result from the theory of Ordinary Differential Equations (ODEs) is
the existence and uniqueness of integral curves with prescribed initial data.
Theorem A.13. Given X ∈ X(M) and p ∈ M, there exists a unique maximal
integral curve α p : Ip → M of X, where Ip is an interval containing 0 and α p (0) = p.
Maximality in the above result is in the sense that every other integral curve
satisfying the same initial condition is a restriction of α p to an open subinterval of
the maximal interval I p . Let D(X) be the set of all points (t, x) ∈ R × M such that
t ∈ Ix , and define:

ϕ X : D(X) → M, ϕ X (t, x) := αx (t),

where αx is the maximal integral curve of X with αx (0) = x. The map ϕ X is called
the flow of X, and D(X) is its domain of definition. The following is also a classical
result in ODEs.
Theorem A.14. Given X ∈ X(M), the domain D(X) is an open subset of R × M
that contains {0} × M and ϕ X is smooth.
Sometimes, it is also convenient to consider the local flow of X, given by the
restriction of ϕ X to an open subset (−δ , δ ) × U of D(X). For a fixed t, we have
local diffeomorphisms ϕtX , given by ϕtX (p) := ϕ X (t, p). Note that, for x ∈ U such
that ϕtX (x) ∈ U, if |s| < δ , |t| < δ and |s + t| < δ , then ϕsX ◦ ϕtX = ϕs+t
X .

A smooth vector field X is complete if D(X) = R × M and, in this case, ϕtX is a

group of diffeomorphisms parameterized by t, called the 1-parameter group of X.
This is equivalent to saying that there is an R-action on M,

R × M  (t, p) −→ ϕtX (p) ∈ M.

Typical examples of complete vector fields are fields with bounded length. For
instance, if M is compact, every smooth vector field on M is complete.
Definition A.15. Let M be an (n+k)-manifold. A n-dimensional distribution1 D on
M is an assignment of an n-dimensional subspace D p ⊂ Tp M for each p ∈ M. The
distribution D is said to be smooth if, for all p ∈ M, there are n smooth vector fields
that span D at all points in a neighborhood of p. A distribution D is involutive if, for

1 An n-dimensional distribution D on M is also called a subbundle of TM of rank n.

A Rudiments of Smooth Manifolds 191

all p ∈ M, given vector fields X,Y in an open neighborhood U of p, with Xq ,Yq ∈ Dq ,

for all q ∈ U, then [X,Y ]q ∈ Dq , for all q ∈ U. In other words, a distribution is
involutive if it is closed with respect to the bracket of vector fields.
Definition A.16. Let M be an (n + k)-manifold. A n-dimensional foliation of M is
a partition F of M into connected immersed submanifolds of dimension n, called
leaves, such that, for all p ∈ M and v ∈ Tp L p , where L p ∈ F is the leaf that contains
p, there is a smooth vector field X on M with X(p) = v and X(q) ∈ Tq Lq , for all
q ∈ M.
A trivial example of foliation is Rn+k = Rn × Rk , where the leaves are subspaces
of the form Rn × {p}, with p ∈ Rk . More generally, the partition of a manifold into
preimages of a smooth submersion is a foliation. Another example of foliation is
the partition (3.4) of a manifold into orbits of a group action, provided all the orbits
have the same dimension.
Definition A.17. Let F = {L p : p ∈ M} be an n-dimensional foliation of an
(n + k)-manifold M. It is possible to prove that, for all p ∈ M, there exist an open
 V of 0 ∈ R
neighborhood U of p, an n+k and a diffeomorphism
 open neighborhood
ψ : V → U, such that ψ V ∩ (R × {y0 }) is the connected component of Lψ (0,y0 ) ∩
n

U that contains ψ (0, y0 ). The open set U is called a simple open neighborhood,
ψ is called a trivialization,
 ψ −1 is called a foliation chart, and the submanifold
ψ V ∩ (R × {y0 }) is called a plaque.
n

Remark A.18. In the general case of singular foliations, we usually do not have
trivializations. Nevertheless, plaques can be defined at least for singular Riemannian
foliations (see Definition 5.1), as follows. Let Pq be a relatively compact open subset
of Lq and z a point in a tubular neighborhood Tub(Pq ). Since F is a singular
Riemannian foliation, each leaf Lz is contained in a stratum (the collection of leaves
with the same dimension of Lz ). In this stratum, Lz is a regular leaf and hence admits
plaques as defined above. A plaque Pz of the singular leaf Lz is then defined as a
plaque of Lz in its stratum.
If F = {L p : p ∈ M} is an n-dimensional foliation, then D = {Tp L p : p ∈ M}
is clearly a smooth involutive n-dimensional distribution. The converse statement is
given by the following result:
Frobenius Theorem A.19. Let D be a smooth n-dimensional involutive distribu-
tion on a manifold M. Then D is integrable, i.e., there exists a unique n-dimensional
foliation F = {L p : p ∈ M} satisfying Dq = Tq Lq , for all q ∈ M. Furthermore, each
leaf L p is a quasi-embedded submanifold.
The above integrability criterion fails for general singular foliations, for which a
sufficient integrability condition is given in Sussmann [201]. Further references on
foliations are Camacho and Neto [60], Lawson [151] and Molino [165].
192 A Rudiments of Smooth Manifolds

A.4 Differential Forms, Integration,

and de Rham Cohomology

In this section, we recall basic notions of tensors, differential forms, integration on

manifolds, the Stokes Theorem, and de Rham cohomology.
A (0, k)-tensor τ on a vector space V is a k-linear functional τ : V × · · · ×V → R.
We denote by T k (V ), or V ⊗ · · · ⊗ V , the vector space of (0, k)-tensors on V . The
tensor product of τ1 ∈ T k (V ) and τ2 ∈ T l (V ) is the element τ1 ⊗ τ2 ∈ T k+l (V )
defined by

(τ1 ⊗ τ2 )(X1 , . . . , Xk , Xk+1 , . . . , Xk+l ) := τ1 (X1 , . . . , Xk )τ2 (Xk+1 , . . . , Xk+l ).

A tensor τ ∈ T k (V ) is called symmetric if, given Xi ∈ V ,

τ (X1 , . . . , Xk ) = τ (Xσ (1) , Xσ (2) , . . . , Xσ (k) ),

for all permutations σ in the group Sk of permutations in k letters. Similarly, τ ∈

T k (V ) is called skew-symmetric if, given Xi ∈ V ,

τ (X1 , . . . , Xk ) = sgn(σ ) τ (Xσ (1) , Xσ (2) , . . . , Xσ (k) ),

for all σ ∈ Sk , where sgn(σ ) denotes the sign of the permutation σ . We denote
by Λ k (V ) the vector space of skew-symmetric k-tensors on V . The projection
Altk : T k (V ) → Λ k (V ), called alternating projection, is given by

Altk (τ )(X1 , . . . , Xk ) = 1
k! ∑ sgn(σ ) τ (Xσ (1) , Xσ (2) , . . . , Xσ (k) ),
σ ∈Sk

for all Xi ∈ V . When k is evident from the context, it is omitted to simplify the
notation.
The wedge product of ω ∈ Λ k (V ) and η ∈ Λ l (V ) is the element ω ∧ η ∈ Λ k+l (V )
defined by

(k+l)!
ω ∧ η := k! l! Alt(k+l) (ω ⊗ η ).

Remark A.20. If ω1 , . . . , ωk ∈ Λ 1 (V ) = T 1 (V ) are (0, 1)-tensors, then

⎛ ⎞
ω1 (X1 ) ω1 (X2 ) · · · ω1 (Xk )
⎜ω2 (X1 ) ω2 (X2 ) · · · ω2 (Xk )⎟
⎜ ⎟
(ω1 ∧ · · · ∧ ωk )(X1 , X2 , . . . , Xk ) = det ⎜ . .. .. ⎟ .
⎝ .. . . ⎠
ωk (X1 ) ωk (X2 ) · · · ωk (Xk )
A Rudiments of Smooth Manifolds 193

Some properties of the wedge product are summarized in the following result:
Proposition A.21. For any ω , ξ ∈ Λ k (V ), η ∈ Λ l (V ), θ ∈ Λ m (V ) and λ ∈ R,
(i) (ω ∧ η ) ∧ θ = ω ∧ (η ∧ θ );
(ii) (ω + ξ ) ∧ η = ω ∧ η + ξ ∧ η ;
(iii) (λ ω ) ∧ η = λ (ω ∧ η );
(iv) ω ∧ η = (−1)kl η ∧ ω .
Moreover, if ω1 , . . . , ωn is a basis of T 1 (V ), then ωi1 ∧ · · ·∧ωir , 1 ≤ i1 < · · · < ir ≤
n, form a basis of Λ r (V ), which hence has dimension nr , where n = dimV . In
particular, Λ n (V ) ∼
= R.
We now make the transition from tensors and skew-symmetric tensors on a vector
space V to tensor fields and differential forms on a manifold M.
Definition A.22. A smooth (0, k)-tensor (field) on a manifold M is a map τ that
assigns to each p ∈ M an element τ p ∈ T k (Tp M) with the property that, given
smooth vector fields X1 , . . . Xk on an open subset U of M, then τ (X1 , . . . , Xk ) is a
smooth function on U defined by τ (X1 , . . . , Xk )(p) = τ X1 (p), . . . , Xk (p) . A smooth
(differential) k-form on M is a smooth (0, k)-tensor ω on M such that ω p ∈ Λ k (Tp M)
for all p ∈ M. We denote by Ω k (M) the set of all smooth k-forms on M.
Alternatively, using the language of smooth sections of vector bundles, see Def-
inition 3.27 and Example 3.28, (0, k)-tensors and k-forms over M are respectively
smooth sections of the tensor bundles TM ∗ ⊗ · · · ⊗ TM ∗ and Λ k TM ∗ over M.
Remark A.23. A k-form ω ∈ Ω k (M) is such that

ω |U = ∑ ai1 ,...,ik dxi1 ∧ dxi2 ∧ · · · ∧ dxik , (A.5)

i1 <i2 <···<ik

where (U, ϕ ), ϕ = (x1 , . . . , xn ), is a chart of M, ai1 ,...,ik : U → R are smooth functions

 
and dxi ∂∂x j = δi j , cf. (A.2). In particular, 0-forms on M are simply smooth
functions f : M → R.

The differential of a 0-form f is the 1-form d f p = ∑ni=1 ∂∂xfi  p dxi . In what follows,
we extend this notion to general k-forms on M.
Proposition A.24. For all k ∈ N, there exists a unique linear operator
d : Ω k (M) → Ω k+1 (M), called exterior derivative, such that:
(i) If f ∈ Ω 0 (M), then d f ∈ Ω 1 (M) is the differential of the function f ∈ C∞ (M),
n
∂ f 
d fp = ∑  dxi .
i=1 ∂ xi p

d d
(ii) The composition Ω k (M) −→ Ω k+1 (M) −→ Ω k+2 (M) is identically null, i.e.
d2 = 0;
194 A Rudiments of Smooth Manifolds

(iii) d(ω ∧ η ) = dω ∧ η + (−1) j ω ∧ dη , for all ω ∈ Ω j (M) and η ∈ Ω k− j (M).

The explicit formula for the exterior derivative of a k-form ω applied to k + 1
smooth vector fields X0 , . . . , Xk ∈ X(M), is:

k  
dω (X0 , . . . , Xk ) = ∑ (−1)i Xi ω (X0 , . . . , X"i , . . . , Xk )
i=0
 
+ ∑ (−1)i+ j ω [Xi , X j ], X0 , . . . , X"i , . . . , X"j , . . . , Xk ,
i< j

where the hat denotes the omission of that element. There is also an explicit formula
for d in terms of local charts. If (U, ϕ ), ϕ = (x1 , . . . , xn ), is a chart of M for which
ω has the expression (A.5), then

dω |U = ∑ dai1 ,...,ik ∧ dxi1 ∧ dxi2 ∧ · · · ∧ dxik .

i1 <i2 <···<ik

Given a (0, k)-tensor τ on N and a smooth map F : M → N, the pull-back of τ by

F is the (0, k)-tensor F ∗ τ on M defined by

(F ∗ τ ) p (X1 , . . . , Xk ) := τF(p) (dFp X1 , . . . , dFp Xk ), Xi ∈ Tp N. (A.6)

The pull-back of k-forms by F induces a homomorphism F ∗ : Ω k (N) → Ω k (M).

Proposition A.25. F ∗ and d commute, that is, F ∗ ◦ d = d ◦ F ∗ .
With the above background, we are ready to discuss integration on manifolds,
starting from integration in Euclidean space. Let ω = f dx1 ∧ dx2 ∧ · · · ∧ dxn , be an
n-form with compact support2 in Rn . Define
 
ω := f dx1 . . . dxn .
U U

It F : V → U is a diffeomorphism between two open sets of Rn that preserves

orientation (i.e., such that det dF > 0), then, by change of variables,
 
ω= F ∗ω . (A.7)
U V

Integration on (oriented) manifolds is defined via the above equation. Before giving
more details, let us recall the notion of orientation on smooth manifolds.

2 Recall that the support of a function f is defined as supp( f ) := {x ∈ M| f (x) = 0}.

A Rudiments of Smooth Manifolds 195

Definition A.26. An n-manifold M is orientable if it is possible to define a n-form

vol ∈ Ω n (M), also denoted [M], which never vanishes, that is, vol p ∈ Λ n (Tp M) ∼
=R
is nonzero for all p ∈ M. In this case, we say that M is oriented by the choice of vol.
Moreover, vol is called a volume form on M.
Proposition A.27. A manifold M is orientable if and only if M admits an atlas
{(Uα , ϕα )} of coherently oriented coordinates, i.e., such that det d(ϕα ◦ ϕβ−1 ) > 0.
Remark A.28. A manifold M of dimension n is called a Riemannian manifold if it
admits a (0, 2)-tensor g that is symmetric and positive-definite (called a Riemannian
metric), see Sect. 2.1. If M is an oriented Riemannian manifold, then there is a
uniquely determined n-form volg which gives the orientation of M and has the value
+1 on every oriented orthonormal frame. This form volg is the volume form of the
oriented Riemannian manifold (M, g). In a local chart ϕ = (x1 , . . . , xn ), we have
-
(ϕ −1 )∗ volg = det(gi j (x)) dx1 ∧ · · · ∧ dxn ,
 
∂ ∂
where gi j = g ∂ xi , ∂ x j .

The first step to define integration of n-forms on an oriented n-manifold M is to

define integration of n-forms with small support. Let {(Uα , ϕα )} be a coherent atlas
for M and let ω be an n-form with compact support contained in Uα . We define
 
ω := (ϕ −1 )∗ ω .
Uα ϕ (Uα )

Note that this is well-defined, i.e., if we consider another chart (Uβ , ϕβ ) such that
the support of ω is contained in Uβ , then we get the same number. This follows
from (A.7) and Proposition A.27, using that (ϕβ ◦ ϕα−1 )∗ = (ϕα−1 )∗ ◦ ϕβ∗ .
In order to define the integration of any n-form with compact support (not
necessarily contained in a coordinate neighborhood) we need to use partitions of
unity. A partition of unity on M is a collection of functions { fα : M → [0, 1]} such
that:
(i) {supp fα } is a locally finite cover of M;
(ii) ∑α fα (x) = 1 for every x ∈ M.
Given any manifold M (orientable or not), any open cover {Uα } of M admits a
subordinate partition of unity, that is, a partition of unity satisfying supp fα ⊂ Uα .
Let K = supp ω and consider a finite cover of K by coordinate neighborhoods Uα
such that Uα is compact. Consider { fα } a partition of unity subordinate to {Uα }.
Note that fα ω has compact support contained in Uα , and define
 
ω := ∑ fα ω .
M α Uα
196 A Rudiments of Smooth Manifolds

One can check that in fact this definition does not depend on the cover or on the
partition of unity. Here, the fact that K is compact was very important. For example,
it guarantees that the above sum is finite and hence well-defined.
With the concept of integration of an n-form with compact support, we can
generalize (A.7) to a change of variables formula for integration of forms.
Proposition A.29. If M and N are oriented n-manifolds, F : M → N is a diffeomor-
phism that preserves orientation, and ω is an n-form with compact support on N,
then
 
ω= F ∗ω .
N M

In order to state the Stokes Theorem, which is a fundamental result of integration

of differential forms, we first discuss the definition of manifolds with boundary.
A smooth n-manifold M with boundary is a smooth n-manifold whose smooth atlas
{(Uα , ϕα )} is formed by open subsets Uα that are homeomorphic either to the half
space H n = {(x1 , . . . , xn ), xn ≥ 0} or to Rn . This class extends that of n-manifolds
(without boundary), for which all Uα are homeomorphic to Rn . Compact manifolds
(without boundary) are called closed manifolds.
Consider the upper half space H n in Rn with the orientation dx1 ∧ . . . ∧ dxn .
The induced orientation on its boundary ∂ H n = {xn = 0} is (−1)n dx1 ∧ . . . ∧
dxn−1 .3 In general, if M is an oriented manifold with boundary, we define the
induced orientation [∂ M] on ∂ M as follows. If ϕ is an orientation-preserving
diffeomorphism from some open subset U of M to the upper half space H n , then
we set [∂ M]|∂ U = ϕ ∗ [∂ H n ]. We are now ready to state the Stokes Theorem:
Theorem A.30. Let ω be an (n − 1)-form with compact support on an oriented
n-manifold M, and consider ∂ M with the induced orientation. Then
 
dω = ω.
M ∂M
.
In particular, if M is an oriented closed manifold, M dω = 0 for all ω ∈ Ω n−1 (M).
We conclude this section with some comments on the de Rham cohomology of a
manifold. A k-form ω is called closed if dω = 0, and exact if there exists a (k − 1)-
form η such that ω = dη . Since d2 = 0, see Proposition A.24, every exact form is
closed. The converse is not true, i.e., there exists closed forms that are not exact.
Example A.31. Consider ω = x2 +y 1
2 (−y dx + x dy) defined on M = R \ {(0, 0)}.
2

One can check that dω = 0, i.e., ω is a closed 1-form. If there was a 0-form
.
(i.e.,
a function) f such that d f = ω , the Stokes Theorem would imply that S1 ω = 0,

3 The sign (−1)n is needed to keep the Stokes Theorem sign-free.

A Rudiments of Smooth Manifolds 197

.
however, by the definition of integration of forms, one can check that S1 ω = 2π .
Therefore, ω = x2 +y
1
2 (−y dx + x dy) is closed, but not exact, on R \ {(0, 0)}.
2

Note that on each open connected subset S of R2 \ {(0, 0)} that does not intersect
the half line4 [0, +∞) × {0}, the 1-form ω above is exact. In fact, it is given by
ω
= dθ , where θ is the angle element of S ⊂ R2 ∼ = C, such that x+iy = reiθ with r =
y
x + y , i.e., θ (x, y) = arctan x . In particular, θ can only be continuously defined
2 2

on such subsets S, because otherwise θ would assume values that differ by 2π on

points arbitrarily close to each other. This example shows the strong dependence of
the notions of closedness and exactness on the topology of the underlying manifold.
The kth de Rham cohomology group of M is defined to be the vector space

ker dk {closed k-forms on M}

H k (M, R) := = , (A.8)
Im d k−1 {exact k-forms on M}

dk−1 dk
where Ω k−1 (M) −→ Ω k (M) −→ Ω k+1 (M) are the operators from Proposition A.24.
The above is a measure of how far closed k-forms on M are from being exact.
Clearly, the de Rham cohomology groups are invariant under diffeomorphisms. This
differentiable invariant is closely related to the topology of M, since it can be proved
that H k (M, R) is isomorphic to the kth singular cohomology
 of M.
Regarding Example A.31, H 1 R2 \ {(0, 0)}, R ∼ = R is actually spanned by the
class of ω . On the other hand, if M is diffeomorphic to Rn , then H k (M, R) ∼ = R if
k = 0 and H k (M, R) ∼= 0 if k > 0, by the so-called Poincaré Lemma.
If M is a closed n-manifold, then M is orientable if and only if H n (M, R) ∼ = R,
in which case the latter is spanned by the class of a volume form of M, and the
isomorphism H n (M, R) ∼ = R is given by integration. The Betti numbers of a closed
n-manifold M are defined as bk (M) := dim H k (M, R), and the Euler characteristic
of M is given by their alternate sum:
n
χ (M) = ∑ (−1)k bk (M).
k=0

A fundamental result on the de Rham cohomology of closed oriented manifolds

is Poincaré duality, which states that if M is a closed oriented n-manifold, then
bk (M) = bn−k (M), for all 0 ≤ k ≤ n.
Example A.32. The Betti numbers of the n-sphere Sn are b0 (Sn ) = bn (Sn ) = 1 and
bk (Sn ) = 0 for all 1 ≤ k ≤ n − 1. In particular, χ (Sn ) = 2 if n is even and χ (Sn ) = 0
if n is odd.
For more about de Rham cohomology, see Bott and Tu [48] or Warner [227].

4 It suffices to require that S does not intersect a fixed unbounded curve in R2 starting at the origin.
References

1. A.V. Alekseevsky, D.V. Alekseevsky, Riemannian G-manifold with one-dimensional orbit

space. Ann. Glob. Anal. Geom. 11(3), 197–211 (1993)
2. D.V. Alekseevsky, F. Podestà, Compact cohomogeneity one Riemannian manifolds of positive
Euler characteristic and quaternionic Kähler manifolds, in Geometry, Topology and Physics,
Campinas, 1996 (de Gruyter, Berlin, 1997), pp. 1–33
3. M.M. Alexandrino, Integrable Riemannian submersion with singularities. Geom. Dedicata
108, 141–152 (2004)
4. M.M. Alexandrino, Singular Riemannian foliations with sections. Ill. J. Math. 48(4),
1163–1182 (2004)
5. M.M. Alexandrino, Generalizations of isoparametric foliations. Mat. Contemp. 28, 29–50
(2005)
6. M.M. Alexandrino, Proofs of conjectures about singular Riemannian foliations. Geom.
Dedicata 119(1), 219–234 (2006)
7. M.M. Alexandrino, Singular holonomy of singular Riemannian foliations with sections. Mat.
Contemp. 33, 23–55 (2007)
8. M.M. Alexandrino, Desingularization of singular Riemannian foliation. Geom. Dedicata 149,
397–416 (2010)
9. M.M. Alexandrino, On polar foliations and fundamental group. Results Math. 60(1–4),
213–223 (2011)
10. M.M. Alexandrino, Polar foliations and isoparametric maps. Ann. Glob. Anal. Geom. 41(2),
187–198 (2012)
11. M.M. Alexandrino, L. Biliotti, R. Pedrosa, Lectures on Isometric Actions. XV Escola de
Geometria Diferencial, Publicações matemáticas (IMPA, Fortaleza, 2008)
12. M.M. Alexandrino, R. Briquet, D. Töben, Progress in the theory of singular Riemannian
foliations. Differ. Geom. Appl. 31, 248–267 (2013)
13. M.M. Alexandrino, C. Gorodski, Singular Riemannian foliations with sections, transnormal
maps and basic forms. Ann. Glob. Anal. Geom. 32(3), 209–223 (2007)
14. M.M. Alexandrino, M.A. Javaloyes, On closed geodesics in the leaf space of singular
Riemannian foliations. Glasg. Math. J. 53, 555–568 (2011)
15. M.M. Alexandrino, A. Lytchak, On smoothness of isometries between orbit spaces, in
Riemannian Geometry and Applications, Proceedings RIGA (Ed. Univ. Bucureşti, 2011),
pp. 17–28. arXiv:1107.2563
16. M.M. Alexandrino, M. Radeschi, Smoothness of isometric flows on orbit spaces and
applications to the theory of foliations (2014, preprint). arXiv:1301.2735

© Springer International Publishing Switzerland 2015 199

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1
200 References

17. M.M. Alexandrino, M. Radeschi, Mean curvature flow of singular Riemannian foliations
(2014, preprint). arXiv:1408.543
18. M.M. Alexandrino, M. Radeschi, Isometries between leaf spaces. Geom. Dedicata 174,
193–201 (2015)
19. M.M. Alexandrino, T. Töben, Singular Riemannian foliations on simply connected spaces.
Differ. Geom. Appl. 24, 383–397 (2006)
20. M.M. Alexandrino, T. Töben, Equifocality of singular Riemannian foliation. Proc. Am. Math.
Soc. 136, 3271–3280 (2008)
21. S. Aloff, N.R. Wallach, An infinite family of distinct 7-manifolds admitting positively curved
Riemannian structures. Bull. Am. Math. Soc. 81, 93–97 (1975)
22. M. Amann, L. Kennard, On a generalized conjecture of Hopf with symmetry.
arXiv:1402.7255
23. M. Amann, L. Kennard, Topological properties of positively curved manifolds with symmetry.
Geom. Funct. Anal. 24(5), 1377–1405 (2014)
24. A. Arvanitogeorgos, An Introduction to Lie Groups and the Geometry of Homogeneous
Spaces, Providence (AMS, 1999)
25. A. Back, W.-Y. Hsiang, Equivariant geometry and Kervaire spheres. Trans. Am. Math. Soc.
304, 207–227 (1987)
26. J. Baez, The octonions. Bull. Am. Math. Soc. 39(2), 145–205 (2002). Errata: Bull. Am. Math.
Soc. (N.S.) 42(2), 213 (2005). arXiv:math/0105155
27. L. Berard-Bergery, Les variétés riemanniennes homogènes simplement connexes de dimen-
sion impaire à courbure strictement positive. J. Math. Pures Appl. 55, 47–67 (1976)
28. V.N. Berestovskii, Homogeneous Riemannian manifolds of positive Ricci curvature. Math.
Notes 58, 905–909 (1995)
29. M. Berger, Les variétés Riemanniennes 1/4-pincées. Ann. Sc. Norm. Super. Pisa Cl. Sci.
14(2), 161–170 (1960)
30. M. Berger, Les variétés riemanniennes homogènes normales simplement connexes à courbure
strictement positive. Ann. Scuola Norm. Sup. Pisa 15, 179–246 (1961)
31. M. Berger, Trois remarques sur les variétés riemanniennes à courbure positive. C. R. Acad.
Sci. Paris Sér. A-B 263, A76–A78 (1966)
32. J. Berndt, S. Console, C. Olmos, Submanifolds and Holonomy. Research Notes in Mathemat-
ics, vol. 434 (Chapman and Hall/CRC, Boca Raton, 2003)
33. A.L. Besse, Einstein Manifolds. Series of Modern Surveys in Mathematics (Springer, Berlin,
2002)
34. R.G. Bettiol, Positive biorthogonal curvature on S 2 ×S2 . Proc. Am. Math. Soc. 142(12), 4341–
4353 (2014)
35. R.G. Bettiol, R.A.E. Mendes, Flag manifolds with strongly positive curvature. Math. Z.
(to appear)
36. R.G. Bettiol, R.A.E. Mendes, Strongly positive curvature (2014, preprint). arXiv:1403.2117
37. R.G. Bettiol, P. Piccione, Bifurcation and local rigidity of homogeneous solutions to the
Yamabe problem on spheres. Calc. Var. Partial Differ. Equ. 47(3–4), 789–807 (2013)
38. R.G. Bettiol, P. Piccione, Multiplicity of solutions to the Yamabe problem on collapsing
Riemannian submersions. Pac. J. Math. 266(1), 1–21 (2013)
39. R.G. Bettiol, P. Piccione, Delaunay type hypersurfaces in cohomogeneity one manifolds.
arXiv:1306.6043
40. L. Biliotti, Coisotropic and polar actions on compact irreducible Hermitian symmetric spaces.
Trans. Am. Math. Soc. 358, 3003–3022 (2006)
41. C. Böhm, Inhomogeneous Einstein metrics on low-dimensional spheres and other low-
dimensional spaces. Invent. Math. 134(1), 145–176 (1998)
42. C. Böhm, Non-compact cohomogeneity one Einstein manifolds. Bull. Soc. Math. Fr. 127(1),
135–177 (1999)
43. C. Böhm, Non-existence of cohomogeneity one Einstein metrics. Math. Ann. 314(1),
109–125 (1999)
References 201

44. C. Böhm, M. Wang, W. Ziller, A variational approach for compact homogeneous Einstein
manifolds. Geom. Funct. Anal. 14(4), 681–733 (2004)
45. C. Böhm, B. Wilking, Manifolds with positive curvature operators are space forms. Ann.
Math. 167(2), 1079–1097 (2008)
46. W.M. Boothby, An Introduction to Differentiable Manifolds and Riemannian Geometry, 2nd
edn. (Academic, Amsterdam/New York, 2003)
47. A. Borel, Some remarks about Lie groups transitive on spheres and tori. Bull. Am. Math. Soc.
55, 580–587 (1949)
48. R. Bott, L. Tu, Differential Forms in Algebraic Topology. Graduate Texts in Mathematics,
vol. 82 (Springer, New York, 1982)
49. H. Boualem, Feuilletages riemanniens singuliers transversalement integrables. Compositio
Mathematica 95, 101–125 (1995)
50. J.-P. Bourguignon, A. Deschamps, P. Sentenac, Conjecture de H. Hopf sur les produits de
variétés. Ann. Sci. École Norm. Sup. 5(4), 277–302 (1972)
51. J.-P. Bourguignon, A. Deschamps, P. Sentenac, Quelques variations particuliéres d’un produit
de métriques. Ann. Sci. École Norm. Sup. (4) 6, 1–16 (1973)
52. J.-P. Bourguignon, H. Karcher, Curvature operators: pinching estimates and geometric
examples. Ann. Sci. École Norm. Sup. (4) 11(1), 71–92 (1978)
53. G. Bredon, Introduction to Compact Transformation Groups. Pure and Applied Mathematics,
vol. 46 (Academic, New York/London, 1972)
54. S. Brendle, Embedded minimal tori in S3 and the Lawson conjecture. Acta Math. 211(2),
177–190 (2013)
55. S. Brendle, R. Schoen, Manifolds with 1/4-pinched curvature are space forms. J. Am. Math.
Soc. 22, 287–307 (2009)
56. T. Bröken, T. Tom Dieck, Representations of Compact Lie Group. Graduate Texts in
Mathematics, vol. 98 (Springer, New York, 1995)
57. R.L. Bryant, An Introduction to Lie Groups and Symplectic Geometry. Geometry and
Quantum Field Theory, IAS/Park City, Mathematics Series, vol. 1 (American Mathematical
Society, Providence, 1995)
58. D. Bump, Lie Groups. Graduate Texts in Mathematics, vol. 225, 2nd edn. (Springer,
New York, 2013)
59. D. Burago, Y. Burago, S. Ivanov, A Course in Metric Geometry. Graduate Studies in
Mathematics, vol. 33 (American Mathematical Society, Providence, 2001)
60. C. Camacho, A.L. Neto, Teoria geométrica das folheações. Projeto Euclides (IMPA, Brasilia,
1979)
61. M.P. do Carmo, Geometria Riemanniana. Projeto Euclides, 3rd edn. (IMPA, Rio de Janeiro,
2005)
62. É. Cartan, Familles de surfaces isoparamétriques dans les espaces á courbure constante. Ann.
di Mat. 17, 177–191 (1938)
63. É. Cartan, Sur des familles remarquables d’hypersurfaces isoparamétriques dans les espaces
shériques. Math. Z. 45, 335–367 (1939)
64. É. Cartan, Sur quelques familles remarquables d’hypersurfaces. C. R. Congrès Math. Liège
30–41 (1939)
65. É. Cartan, Sur des familles d’hypersurfaces isoparamétriques des espaces sphériques à 5 et à
9 dimensions. Revista Univ. Tucumán 1, 5–22 (1940)
66. R. Carter, G. Segal, I. MacDonald, Lectures on Lie Groups and Lie Algebras. London
Mathematical Society, Student Texts, vol. 32 (Cambridge University Press, 1995)
67. S. Carter, A. West, Isoparametric systems and transnormality. Proc. Lond. Math. Soc. 51,
520–542 (1985)
68. S. Carter, A. West, Generalised Cartan polynomials. J. Lond. Math. Soc. 32, 305–316 (1985)
69. T.E. Cecil, Q.-S. Chi, G.R. Jensen, Isoparametric hypersurfaces with four principal curva-
tures. Ann. Math. (2) 166(1), 1–76 (2007)
70. J. Cheeger, Some examples of manifolds of nonnegative curvature. J. Differ. Geom. 8,
623–628 (1973)
202 References

71. J. Cheeger, D. Ebin, Comparison Theorems in Riemannian Geometry (North-Holland Pub-

lishing, Amsterdam, 1975)
72. J. Dadok, Polar coordinates induced by actions of compact Lie groups. Trans. Am. Math. Soc.
288, 125–137 (1985)
73. A. Dancer, S. Hall, M. Wang, Cohomogeneity one shrinking Ricci solitons: an analytic and
numerical study. Asian J. Math. 17(1), 33–61 (2013)
74. A. Dancer, M. Wang, The cohomogeneity one Einstein equations from the Hamiltonian
viewpoint. J. Reine Angew. Math. 524, 97–128 (2000)
75. A. Dancer, M. Wang, On Ricci solitons of cohomogeneity one. Ann. Glob. Anal. Geom. 39(3),
259–292 (2011)
76. M.W. Davis, Groups generated by reflections and aspherical manifolds not covered by
Euclidean space. Ann. Math. 117, 293–324 (1983)
77. O. Dearricott, A 7-manifold with positive curvature. Duke Math. J. 158(2), 307–346 (2011)
78. A. Deitmar, A First Course in Harmonic Analysis. Universitext, 2nd edn. (Springer, New
York, 2005)
79. J.J. Duistermaat, J.A.C. Kolk, Lie Groups. Universitext (Springer, Berlin/New York, 2000)
80. H. Elíasson, Die Krümmung des Raumes Sp(2)/SU(2) von Berger. Math. Ann. 164, 317–323
(1966)
81. J.-H. Eschenburg, New examples of manifolds with strictly positive curvature. Invent. Math.
66(3), 469–480 (1982)
82. J.-H. Eschenburg, Freie isometrische Aktionen auf kompakten Lie-Gruppen mit positiv
gekrümmten Orbiträumen. Schriftenreihe des Mathematischen Instituts der Universität
Münster, 2. Serie, 32 (1984)
83. J.-H. Eschenburg, M. Wang, The initial value problem for cohomogeneity one Einstein
metrics. J. Geom. Anal. 10(1), 109–137 (2000)
84. F. Fang, K. Grove, G. Thorbergsson, Tits Geometry and Positive Curvature (2012, preprint).
arXiv:1205.6222 [math.DG]
85. H.D. Fegan, Introduction to Compact Lie Groups. Series in Pure Mathematics, vol. 13 (World
Scientific, Singapore, 1998)
86. L. Fehér, B.G. Pusztai, Twisted spin Sutherland models from quantum Hamiltonian reduction.
Phys. A: Math. Theor. 41, 194009 (2008)
87. L. Fehér, B.G. Pusztai, Hamiltonian reductions of free particles under polar actions of
compact Lie groups. Theor. Math. Phys. 155, 646–658 (2008)
88. B.L. Feigin, D.B. Fuchs, V.V. Gorbatsevich, O.V. Schvartsman, E.B. Vinberg, Lie Groups and
Lie Algebras II. Encyclopaedia of Mathematical Sciences, vol. 21 (Springer, Moscow, 1988)
89. D. Ferus, H. Karcher, H.F. Münzner, Cliffordalgebren und neue isoparametrische Hyper-
flächen. Math. Z. 177, 479–502 (1981)
90. W. Fulton, J. Harris, Representation Theory. A First Course. Graduate Texts in Mathematics,
vol. 129, Readings in Mathematics (Springer, New York, 1991)
91. F. Galaz-Garcia, C. Searle, Cohomogeneity one Alexandrov spaces. Transform. Groups 16(1),
91–107 (2011)
92. F. Galaz-Garcia, C. Searle, Low-dimensional manifolds with non-negative curvature and
maximal symmetry rank. Proc. Am. Math. Soc. 139(7), 2559–2564 (2011)
93. F. Galaz-Garcia, C. Searle, Nonnegatively curved 5-manifolds with almost maximal symme-
try rank. Geom. Topol. 18(3), 1397–1435 (2014)
94. R. Gangolli, V.S. Varadarajan, Harmonic Analysis of Spherical Functions on Real Reductive
Groups. Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 101 (Springer, Berlin/New
York, 1988)
95. J. Ge, Z. Tang, Isoparametric functions on exotic spheres. J. Reine Angew. Math. 683,
161–180 (2013)
96. J. Ge, Z. Tang, Geometry of isoparametric hypersurfaces in Riemannian manifolds. Asian J.
Math. 18(1), 117–125 (2014)
97. Robert E. Krieger Publishing Co., Inc. (Malabar, FL, 1994)
References 203

98. V.V. Gorbatsevich, A.L. Onishchik, E.B. Vinberg, Foundations of Lie Theory and Lie
Transformation Groups (Springer, Moscow, 1988)
99. V.V. Gorbatsevich, A.L. Onishchik, E.B. Vinberg, Lie Groups and Lie Algebras I. Encyclopae-
dia of Mathematical Sciences, vol. 20 (Springer, Moscow, 1988)
100. V.V. Gorbatsevich, A.L. Onishchik, E.B. Vinberg, Lie Groups and Lie Algebras III. Ency-
clopaedia of Mathematical Sciences, vol. 41 (Springer, Moscow, 1990)
101. C. Gorodski, A Metric Approach to Representations of Compact Lie Groups (2014,
preprint), http://www.ime.usp.br/~gorodski/publications.html, or http://math.osu.edu/file/
metric-approach-representations-compact-lie-groups.pdf
102. C. Gorodski, E. Heintze, Homogenous structures and rigidity of isoparametric submanifolds
in Hilbert spaces. J. Fixed Point Theory Appl. 11(1), 93–136 (2012)
103. F. Gozzi, Low dimensional polar actions. Geom. Dedicata (to appear). arXiv:1407.0638
104. A. Gray, Pseudo-Riemannian almost product manifolds and submersions. J. Math. Mech. 16,
715–737 (1967)
105. D. Gromoll, G. Walschap, Metric Foliations and Curvatures. Progress in Mathematics,
vol. 268 (Birkhäuser, 2009)
106. K. Grove, Riemannian Geometry: A Metric Entrance. Lecture Notes Series (Aarhus), vol. 65
(University of Aarhus, Department of Mathematics, Aarhus, 1999)
107. K. Grove, Geometry of, and via, symmetries, in Conformal, Riemannian and Lagrangian
Geometry, Knoxville, 2000. University Lecture Series, vol. 27 (American Mathematical
Society, Providence, 2002), pp. 31–53
108. K. Grove, D. Gromoll, The low-dimensional metric foliations of Euclidean spheres. J. Differ.
Geom. 28(1), 143–156 (1988)
109. K. Grove, C. Searle, Positively curved manifolds with maximal symmetry-rank. J. Pure Appl.
Algebra 91(1–3), 137–142 (1994)
110. K. Grove, C. Searle, Differential topological restrictions by curvature and symmetry. J. Differ.
Geom. 47(3), 530–559 (1997)
111. K. Grove, L. Verdiani, B. Wilking, W. Ziller, Non-negative curvature obstructions in
cohomogeneity one and the Kervaire spheres. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 5(2),
159–170 (2006)
112. K. Grove, L. Verdiani, W. Ziller, An exotic T1 S4 with positive curvature. Geom. Funct. Anal.
21(3), 499–524 (2011)
113. K. Grove, B. Wilking, A knot characterization and 1-connected nonnegatively curved 4-
manifolds with circle symmetry. Geom. Topol. 18(5), 3091–3110 (2014)
114. K. Grove, B. Wilking, W. Ziller, Positively curved cohomogeneity one manifolds and 3-
Sasakian geometry. J. Differ. Geom. 78(1), 33–111 (2008)
115. K. Grove, W. Ziller, Curvature and symmetry of Milnor spheres. Ann. Math. (2) 152(1),
331–367 (2000)
116. K. Grove, W. Ziller, Cohomogeneity one manifolds with positive Ricci curvature. Invent.
Math. 149(3), 619–646 (2002)
117. K. Grove, W. Ziller, Lifting group actions and nonnegative curvature. Trans. Am. Math. Soc.
363(6), 2865–2890 (2011)
118. K. Grove, W. Ziller, Polar manifolds and actions. J. Fixed Point Theory Appl. 11(2), 279–313
(2012)
119. M.A. Guest, Harmonic Maps, Loop Groups and Integrable Systems. London Mathematical
Society, Student Texts, vol. 38 (Cambridge University Press, Cambridge/New York, 1997)
120. L. Guijarro, G. Walschap, When is a Riemannian submersion homogeneous? Geom. Dedicata
125, 47–52 (2007)
121. B.C. Hall, Lie Groups, Lie Algebras and Representations, an Elementary Introduction.
Graduate Texts in Mathematics (Springer, New York, 2004)
122. C.E. Harle, Isoparametric families of submanifolds. Bol. Soc. Brasil. Mat. 13, 491–513 (1982)
123. F.R. Harvey, Spinors and Calibrations. Vol. 9 of Perspectives in Mathematics (Academic,
Boston, 1990)
204 References

124. E. Heintze, X. Liu, C. Olmos, Isoparametric submanifolds and a Chevalley–type restriction

theorem, in Integrable Systems, Geometry, and Topology. AMS/IP Studies in Advanced
Mathematics, vol. 36 (American Mathematical Society, Providence, 2006), pp. 151–190
125. S. Helgason, Groups and Geometric Analysis. Integral Geometry, Invariant Differential
Operators, and Spherical Functions. Corrected reprint of the 1984 original. Mathematical
Surveys and Monographs, vol. 83 (American Mathematical Society, Providence, 2000)
126. S. Helgason, Differential Geometry, Lie Groups and Symmetric Spaces. Graduate Studies in
Mathematics, vol. 34 (American Mathematical Society, Providence, 2001)
127. R. Hermann, Variational completeness for compact symmetric spaces. Proc. Am. Math. Soc.
11, 544–546 (1960)
128. J. Hilgert, K.-H. Neeb, Structure and Geometry of Lie Groups. Springer Monographs in
Mathematics (Springer, New York, 2012)
129. W.-Y. Hsiang, Lectures on Lie Groups. Series on University Mathematics, vol. 2 (World
Scientific Publishing, River Edge, 2000)
130. W.C. Hsiang, W.Y. Hsiang, On compact subgroups of the diffeomorphism groups of Kervaire
spheres. Ann. Math. 85, 359–369 (1967)
131. W.Y. Hsiang, B. Kleiner, On the topology of positively curved 4-manifolds with symmetry. J.
Differ. Geom. 29(3), 615–621 (1989)
132. G.A. Hunt, A theorem of Elie Cartan. Proc. Am. Math. Soc. 7, 307–308 (1956)
133. M. Ise, M. Takeuchi, Lie Groups I. Translations of Mathematical Monographs, vol. 85
(American Mathematical Society, Providence, 1991)
134. G. Jensen, Einstein metrics on principal fibre bundles. J. Differ. Geom. 8, 599–614 (1973)
135. J. Jost, Riemannian Geometry and Geometric Analysis. Universitext, 2nd edn. (Springer,
Berlin/New York, 1998)
136. Y. Katznelson, An Introduction to Harmonic Analysis (Dover, New York, 1968)
137. K. Kawakubo, The Theory of Transformation Groups (Oxford University Press, Oxford/New
York, 1991)
138. L. Kennard, On the Hopf conjecture with symmetry. Geom. Topol. 161(1), 563–593 (2013)
139. L. Kennard, Positively curved Riemannian metrics with logarithmic symmetry rank bounds.
Comment. Math. Helv. 89, 937–962 (2014)
140. M. Kerin, On the curvature of biquotients. Math. Ann. 352(1), 155–178 (2012)
141. B. Kleiner, Riemannian four-manifolds with nonnegative curvature and continuous symmetry,
Ph.D. thesis, University of California, Berkeley (1990)
142. W. Klingenberg, Contributions to Riemannian geometry in the large. Ann. Math. 69(3),
654–666 (1959)
143. W. Klingenberg, Über Riemannsche Mannigfaltigkeiten mit positiver Krümmung. Comment.
Math. Helv. 35(1), 47–54 (1961)
144. A.W. Knapp, Representation Theory of Semisimple Groups (Princeton University Press,
Princeton, 1986)
145. A.W. Knapp, Lie Groups: Beyond an Introduction. Progress in Mathematics, vol. 140, 2nd
edn. (Birkhäuser, Boston, 2002)
146. A. Kollross, A classification of hyperpolar and cohomogeneity one actions. Trans. Am. Math.
Soc. 354(2), 571–612 (2002)
147. R. Lafuente, J. Lauret, On homogeneous Ricci solitons. Q. J. Math. 65(2), 399–419 (2014)
148. R. Lafuente, J. Lauret, Structure of homogeneous Ricci solitons and the Alekseevskii
conjecture. J. Differ. Geom. 98(2), 315–347 (2014)
149. S. Lang, Differential and Riemannian Manifolds. Graduate Texts in Mathematics (Springer,
New York, 1995)
150. J. Lauret, Ricci flow of homogeneous manifolds. Math. Z. 274(1–2), 373–403 (2013)
151. H.B. Lawson, Foliations. Bull. Am. Math. Soc. 80, 3 (1974)
152. J.M. Lee, Riemannian Manifolds: An Introduction to Curvature. Graduate Texts in Mathe-
matics (Springer, New York, 1997)
153. A. Lytchak, Singular Riemannian Foliations on Space Without Conjugate Points. Differential
Geometry (World Scientific Publishing, Hackensack, 2009), pp. 75–82
References 205

154. A. Lytchak, Geometric resolution of singular Riemannian foliations. Geom. Dedicata 149,
379–395 (2010)
155. A. Lytchak, Polar foliations on symmetric spaces. Geom. Funct. Anal. 24, 1298–1315 (2014)
156. A. Lytchak, G. Thorbergsson, Variationally complete actions on nonnegatively curved
manifolds. Ill. J. Math. 51(2), 605–615 (2007)
157. A. Lytchak, G. Thorbergsson, Curvature explosion in quotients and applications. J. Differ.
Geom. 85(1), 117–139 (2010)
158. J. Milnor, Morse Theory. Annals of Mathematics Studies, vol. 51 (Princeton University Press,
Princeton, 1969)
159. J. Milnor, Curvatures of left-invariant metrics on Lie groups. Adv. Math. 21(3), 293–329
(1976)
160. W. Misner, K.S. Thorne, J.A. Wheeler, Gravitation (Freeman and Company, New York, 1973)
161. R. Miyaoka, Transnormal functions on a Riemannian manifold. Differ. Geom. Appl. 31,
130–139 (2013)
162. R. Miyaoka, Isoparametric hypersurfaces with (g, m) = (6, 2). Ann. Math. (2) 177(1), 53–110
(2013). Errata: Ann. Math. (2) 180(3), 1221 (2014)
163. I. Moerdijk, J. Mrčun, Introduction to Foliations and Lie Groupoids. Cambridge Studies in
Advanced Mathematics (Cambridge University Press, Cambridge/New York, 2003)
164. P. Molino, Desingularisation Des Feuilletages Riemanniens. Am. J. Math. 106(5), 1091–1106
(1984)
165. P. Molino, Riemannian Foliations. Progress in Mathematics, vol. 73 (Birkhäuser, Boston,
1988)
166. P. Molino, M. Pierrot, Théorèmes de slice et holonomie des feuilletages Riemanniens
singuliers. Ann. Inst. Fourier (Grenoble) 37(4), 207–223 (1987)
167. D. Montgomery, H. Samelson, Transformation groups of spheres. Ann. Math. (2) 44, 454–470
(1943)
168. D. Montgomery, L. Zippin, Topological Transformation Groups (Interscience Publishers,
New York, 1955)
169. M. Mucha, Polar actions on certain principal bundles over symmetric spaces of compact type.
Proc. Am. Math. Soc. 139(6), 2249–2255 (2011)
170. H.F. Münzner, Isoparametrische Hyperflächen in Sphären I. Math. Ann. 251, 57–71 (1980)
171. H.F. Münzner, Isoparametrische Hyperflächen in Sphären II. Math. Ann. 256, 215–232 (1981)
172. M. Müter, Krümmungserhöhende Deformationen mittels Gruppenaktionen, PhD thesis,
University of Münster (1987)
173. S. Myers, N. Steenrod, The group of isometries of a Riemannian manifold. Ann. Math. 40(2),
400–416 (1939)
174. L. Ni, Notes on Transnormal Functions on Riemannian Manifolds, unpublished (1997),
available at: http://math.ucsd.edu/~lni/academic/isopara.pdf
175. M. Noumi, Painlevé Equations Through Symmetry. Translations of Mathematical Mono-
graphs, vol. 223 (American Mathematical Society, Providence, 2004)
176. P.J. Olver, Applications of Lie Groups to Differential Equations. Graduate Texts in Mathe-
matics, vol. 107, 2nd edn. (Springer, New York, 2000)
177. B. O’Neill, The fundamental equations of a submersion. Mich. Math. J. 13, 459–469 (1966)
178. A.L. Onishchik, Transitive compact transformation groups. Mat. Sb. 60, 447–485 (1963);
translated in English in Am. Math. Soc. Transl. 55, 153–194 (1966)
179. A.L. Onishchik, Topology of transitive transformation groups. Johann Ambrosius Barth
Verlag GmbH, Leipzig, 1994. 300 pp. ISBN: 3-335-00355-1
180. R.S. Palais, On the existence of slices for actions of noncompact groups. Ann. Math. 73,
295–323 (1961)
181. R.S. Palais, C.L. Terng, A general theory of canonical forms. Trans. Am. Math. Soc. 300,
236–238 (1987)
182. R.S. Palais, C.L. Terng, Critical Point Theory and Submanifold Geometry. Lecture Notes in
Mathematics, vol. 1353 (Springer, Berlin/New York, 1988)
206 References

183. P. Petersen, Riemannian Geometry. Graduate Texts in Mathematics, 171, 2nd edn. New York,
NY, USA (Springer, 2006)
184. P. Petersen, F. Wilhelm, An exotic sphere with positive sectional curvature. arXiv:0805.0812
185. M.J. Pflaum, H. Posthuma, X. Tang, Geometry of orbit spaces of proper Lie groupoids. J.
Reine Angew. Math. 694, 49–84 (2014)
186. F. Podestà, G. Thorbergsson, Polar actions on rank one symmetric spaces. J. Differ. Geom.
53(1), 131–175 (1999)
187. T. Püttmann, Optimal pinching constants of odd dimensional homogeneous spaces. Invent.
Math. 138, 631–684 (1999)
188. M. Radeschi, Low dimensional singular Riemannian foliations in spheres (2012, preprint).
arXiv:1203.6113
189. M. Radeschi, Clifford algebras and new singular Riemannian foliations in spheres. Geom.
Funct. Anal. 24, 1660–1682 (2014)
190. H. Rauch, A contribution to differential geometry in the large. Ann. Math. 54(1), 38–55
(1951)
191. L.A.B. San Martin, Álgebras de Lie (Editora da Unicamp, 1999)
192. L. Schwachhöfer, K. Tapp, Homogeneous metrics with nonnegative curvature. J. Geom. Anal.
19(4), 929–943 (2009)
193. C. Searle, Cohomogeneity and positive curvature in low dimensions. Math. Z. 214, 491–498
(1993): Err. ibet. 226, 165–167 (1997)
194. C. Searle, F. Wilhelm, How to lift positive Ricci curvature (2013, preprint). arXiv:1311.1809
195. J.P. Serre, Complex Semisimple Lie Algebras. Monographs in Mathematics (Springer,
Berlin/New York, 2001)
196. A. Siffert, A new structural approach to isoparametric hypersurfaces in spheres (2014,
preprint). arXiv:1410.6206
197. F.L.N. Spindola, Grupos de Lie, ações próprias e a conjectura de Palais–Terng, MSc.
dissertation, Universidade de São Paulo (2008)
198. M. Spivak, A Comprehensive Introduction to Differential Geometry, vol. 1 (Publish or Perish,
Boston, 1975)
199. M. Strake, Variationen von Metriken nichtnegativer Schnittkrümmung. PhD thesis, University
of Münster (1986)
200. M. Strake, Curvature increasing metric variations. Math. Ann. 276(4), 633–641 (1987)
201. H. Sussmann, Orbits of families of vector fields and integrability of distributions. Trans. Am.
Math. Soc. 180, 171–188 (1973)
202. J. Szenthe, A generalization of the Weyl group. Acta Math. Hungar. 41(3–4), 347–357 (1983)
203. J. Szenthe, Orthogonally transversal submanifolds and the generalizations of the Weyl group.
Periodica Mathematica Hungarica 15(4), 281–299 (1984)
204. C.L. Terng, Isoparametric submanifolds and their Coxeter groups. J. Differ. Geom. 21,
79–107 (1985)
205. C.L. Terng, G. Thorbergsson, Submanifold geometry in symmetric spaces. J. Differ. Geom.
42, 665–718 (1995)
206. G. Thorbergsson, Isoparametric foliations and their buildings. Ann. Math. 133, 429–446
(1991)
207. G. Thorbergsson, Differentialgeometrie I und II. Lecture Notes (University of Cologne, 1997)
208. G. Thorbergsson, A survey on isoparametric hypersurfaces and their generalizations, in
Handbook of Differential Geometry, ed. by F. Dillen, L. Verstraelen, vol. 1 (Elsevier,
Amsterdam/New York, 2000)
209. G. Thorbergsson, Transformation groups and submanifold geometry. Rendiconti di Matemat-
ica, Serie VII 25, 1–16 (2005)
210. G. Thorbergsson, Singular Riemannian foliations and isoparametric submanifolds. Milan J.
Math. 78(1), 355–370 (2010)
References 207

211. J.A. Thorpe, On the curvature tensor of a positively curved 4-manifold, in Proceedings of the
Thirteenth Biennial Seminar of the Canadian Mathematical Congress, Dalhousie University,
Halifax, 1971, vol. 2 (Canadian Mathematical Congress, Montreal, 1972), pp. 156–159
212. J.A. Thorpe, The zeros of nonnegative curvature operators. J. Differ. Geom. 5, 113–125
(1971). Erratum: J. Differ. Geom. 11(2), 315 (1976)
213. D. Töben, Parallel focal structure and singular Riemannian foliations. Trans. Am. Math. Soc.
358, 1677–1704 (2006)
214. D. Töben, Singular Riemannian foliations on nonpositively curved manifolds. Math. Z.
255(2), 427–436 (2007)
215. F.M. Valiev, Precise estimates for the sectional curvatures of homogeneous Riemannian
metrics on Wallach spaces. Sib. Mat. Zhurn. 20, 248–262 (1979)
216. V.S. Varadarajan, Lie Groups, Lie Algebras, and Their Representations. Graduate Texts in
Mathematics (Springer, New York, 1984)
217. V.S. Varadarajan, An Introduction to Harmonic Analysis on Semisimple Lie Groups.
Cambridge Studies in Advanced Mathematics, vol. 16 (Cambridge University Press, Cam-
bridge/New York, 1999)
218. L. Verdiani, Cohomogeneity one manifolds of even dimension with strictly positive sectional
curvature. J. Differ. Geom. 68, 31–72 (2004)
219. L. Verdiani, W. Ziller, Positively curved homogeneous metrics on spheres. Math. Z. 261(3),
473–488 (2009)
220. L. Verdiani, W. Ziller, Concavity and rigidity in non-negative curvature. J. Differ. Geom. 97,
349–375 (2014)
221. N.R. Wallach, Compact homogeneous Riemannian manifolds with strictly positive curvature.
Ann. Math. 96, 277–295 (1972)
222. G. Walschap, Metric Structures in Differential Geometry. Graduate Texts in Mathematics
(Springer, New York, 2004)
223. Q.M. Wang, Isoparametric hypersurfaces in complex projective spaces, in Proceedings of the
1980 Beijing Symposium on Differential Geometry and Differential Equations, Beijing, 1980
(Science Press, Beijing, 1982), pp. 1509–1523
224. Q.M. Wang, Isoparametric functions on Riemannian manifolds. I. Math. Ann. 277, 639–646
(1987)
225. M. Wang, W. Ziller, On normal homogeneous Einstein manifolds. Ann. Sci. École Norm.
Sup. (4) 18(4), 563–633 (1985)
226. M. Wang, W. Ziller, Existence and nonexistence of homogeneous Einstein metrics. Invent.
Math. 84(1), 177–194 (1986)
227. F.W. Warner, Foundations of Differentiable Manifolds and Lie Groups. Graduate Texts in
Mathematics (Springer, New York, 1983)
228. S. Wiesendorf, Taut submanifolds and foliations. J. Diff. Geom. (to appear). arXiv:1112.5965
229. B. Wilking, Positively curved manifolds with symmetry. Ann. Math. 163, 607–668 (2006)
230. B. Wilking, A duality theorem for Riemannian foliations in nonnegative sectional curvature.
Geom. Funct. Anal. 17(4), 1297–1320 (2007)
231. B. Wilking, Nonnegatively and positively curved manifolds, in Surveys in Differential
Geometry, ed. by J. Cheeger, K. Grove, vol. XI (International Press, Somerville, 2007),
pp. 25–62
232. B. Wilking and W. Ziller. Revisiting homogeneous spaces with positive curvature. (2015,
preprint) arXiv:1503.06256
233. J. Yeager, Geometric and topological ellipticity in cohomogeneity two. Ph.D. thesis, Univer-
sity of Maryland (2012)
234. W. Ziller, On M. Mueter’s Ph.D. thesis on Cheeger deformations. arXiv:0909.0161
235. W. Ziller, Homogeneous Einstein metrics on spheres and projective spaces. Math. Ann.
259(3), 351–358 (1982)
236. W. Ziller, Examples of Riemannian manifolds with non-negative sectional curvature, in
Surveys in Differential Geometry, ed. by J. Cheeger, K. Grove, vol. XI (International Press,
Somerville, 2007), pp. 63–102
237. S. Zoltek, Nonnegative curvature operators: some nontrivial examples. J. Differ. Geom. 14,
303–315 (1979)
Index

Symbols SU(n), 4, 21–24, 39, 63, 77

G-equivariant map, 53 Sp(n), 5, 22
G(x), 52 U(n), 4, 24, 45
Gx , 52 μ G , 52
Mprinc , 75 μ g , 52
Mx≈ , Mx∼ , 79 μx , 52
Sx , 64, 120 ∇, 28
Z(G), Z(g), 23 Ric, 34
Sξ , 36 scal, 34
Ad, 16, 52 H, 36
Iso(M, g), 32, 69 k-Grassmannian, 64
Tub, 68 XF , 109
ad, 16 SO(n), 104
exp, 13, 15 SU(n), 88, 98, 105
CPn , 59, 63, 146, 166 U(n), 88
HPn , 59, 63, 146 1-parameter subgroup, 13
II, 35
X(M), 188
gl(n, R), gl(n, C), 5, 14, 21 A
sl(n, R), sl(n, C), 14, 21 Action, 51
so(n), o(n), 21, 104 adjoint, 16, 52, 85
sp(n), 21 by isometries, 52
su(n), u(n), 21, 39, 47, 98, 105 cohomogeneity, 89, 157
GL(n, R), GL(n, C), 14 commutes, 62
SL(n, R), SL(n, C), 4, 14 diagonal, 53
SU(n), 21, 47 effective, 52
Sp(n), 21, 63 field, 55
U(n), 21 free, 52
GL(n, R), GL(n, C), 21 hyperpolar, 88
GL(n, R), GL(n, C), 4 orbit-equivalent, 90
Grk (V ), 64 polar, 78, 88, 117
O(n), 4, 21 product, 53
SL(n, R), SL(n, C), 21 proper, 56
SO(n), 4, 7, 21, 23, 56, 63 properly discontinuous, 56

© Springer International Publishing Switzerland 2015 209

M.M. Alexandrino, R.G. Bettiol, Lie Groups and Geometric Aspects
of Isometric Actions, DOI 10.1007/978-3-319-16613-1
210 Index

Action (cont.) Distribution, 190

section, 88 horizontal, 37
subaction, 53 integrable, 191
transitive, 52 involutive, 190
Adjoint vertical, 37
action, 16, 52, 85 Dynkin diagram, 103
representation, 16
representation in GL(n, R), 17
E
Einstein
B constant, 35
Berger sphere, 151 equation (in vacuum), 35
Betti number, 197 manifold, 35, 152, 153
Bi-invariant Endpoint map, 110
k-form, 38 Equifocal submanifold, 110
metric, 38 Equifocality, 112
metric of SU(n), 39 Equivariant
vector field, 6 map, 53
Bianchi identity, 33, 181 normal field, 74
Brieskhorn variety, 168 Exponential map, 13
Exterior derivative, 193

C
Campbell formulas, 15 F
Centralizer, 24 Fiber bundle, 57
Cheeger deformation, 140, 151 associated to a principal bundle, 66
cohomogeneity one, 163 base, 57
Cheeger reparametrization, 144 bundle charts, 57
Cohomogeneity, 89, 139 fiber, 57
Compact-open topology, 32 principal, 58
Complex flag manifold, 64 projection, 57
Connection, 28 structure group, 57
basic, 112 total space, 57
Bott, 112 transition function, 57
compatible with metric, 28 trivial, 57
formula, 29 vector bundle, 57
Levi-Civita, 29 Fixed point, 52
symmetric, 28 set, 79
Covariant derivative, 29 Flag, 64
Covering map, 10 Foliation, 191
CROSS, 146 homogeneous, 55
Curvature, 32 isoparametric, 91
Gaussian, 36 leaf, 191
mean, 36, 162 polar, 110
operator, 33, 179 Riemannian, 109
principal, 36 section, 110
Ricci, 34 singular, 109
scalar, 34 singular Riemannian, 109
sectional, 33 singular Riemannian, with sections, 110
strongly nonnegative, 180 Frame bundle, 58
strongly positive, 180

G
D Gauss equation, 36
Delaunay surfaces, 162 Geodesic, 29
Differential form, 193 flow, 30
Index 211

segment, 30 Kleiner’s lemma, 72

minimal, 31 Koszul formula, 29
unit speed, 29
Grassmannian, 64
Gray-O’Neill formula, 38, 181 L
Lawson conjecture, 165
Leaf, 191
H regular, 109
Holonomy singular, 109
group, 69, 113 Left-invariant
map, 112 k-form, 38
pseudogroup, 114 metric, 38
singular map, 121 vector field, 6
Homogeneous bundle, 145, 151 Length, 31
Homogeneous metric, 148, 149 Lie algebra, 5
normal, 153 bracket, 5
on projective spaces, 152 center, 23
on spheres, 149, 156 homomorphism, 6, 8
Hopf action, 54, 59, 164 ideal, 43
Hopf bundle, 54, 59, 150, 151, 156 of SO(3), 7
Hopf Problem, 172 of SU(n), 39
of a Lie group, 7
radical, 44
I semisimple, 41
Ideal, 43 simple, 43
simple, 43 solvable, 43
Ineffective kernel, 52 subalgebra, 8
Isometry, 32, 52 Lie group, 3
group, 32 action, 51
full, 146 center, 23, 45
pseudogroup, 113 classical, 4
Isoparametric exceptional, 106
function, 127 homomorphism, 7
map, 131 normal subgroup, 43
submanifold, 90, 91, 127 rank, 87
Isoparametric hypersurface, 130 root, 94
Isotropy semisimple, 41
group, 52 subgroup, 8
representation, 53, 72, 89 Local section, 120
Isotypic component, 149

M
J Manifold, 185
Jacobi equation, 33 Ck , 185
Jacobi field, 33 analytic, 185
Jacobi identity, 5 atlas, 185
chart, 185
closed, 196
K geodesically complete, 31
Kervaire sphere, 168 tangent space, 186
Killing with boundary, 196
form, 41 Metric, 27
form of SU(n), 47 G-invariant, 52
vector field, 32 bi-invariant, 38
212 Index

Metric (cont.) length, 31

Einstein, 35 manifold, 27
left-invariant, 38 metric, 27
right-invariant, 38 orbifold, 114
Molino Conjecture, 125 submersion, 37
integrable, 37
Right-invariant
N k-form, 38
Normal bundle, 58, 90 metric, 38
flat, 90 vector field, 6
globally flat, 90 Root, 94
Normal slice, 72 base, 103
Normal space, 35 coroot, 94
Normalizer, 24 diagram, 103
of SU(n), 98
O positive, 97
Orbifold, 114 simple, 103
good, 114 system, 94
Orbit, 52
exceptional, 77, 160
larger type, 77 S
local type, 79 Second fundamental form, 35
principal, 73, 75 Semisimple
regular, 77 Lie algebra, 41
same local type, 79 Lie group, 41
same type, 77 Shape operator, 36
singular, 77 Simple
space, 54 ideal, 43
type, 79 Lie algebra, 43
Simple open subset, 111, 191
Singular Riemannian foliation with sections,
P 110
Parallel translation, 30, 112 Singular stratification of local section, 121
Partition of unity, 195 Slice, 64, 120
Pinching constant, 156 normal, 72
Plaque, 111, 191 representation, 72
Polar foliation, 110 Space form, 34
Principal bundle, 58 spherical, 156, 180
Principal direction, 36 Sphere theorem, 156
Principal orbit, 73, 75 Stabilizer, 52
Strata, 82
Stratification, 82
R Strongly positive curvature, 180
Representation, 15, 53 Submanifold, 187
s-representation, 90 embedded, 187
isotropy, 53, 89 equifocal, 110
slice, 72 immersed, 187
Riemannian isoparametric, 90
curvature, 32 minimal, 36, 83, 162, 165
distance, 31 quasi-embedded, 187, 191
exponential map, 31 totally geodesic, 36, 41
good orbifold, 114 Submersion, 37
immersion, 35 Surface, 28, 30
isometry, 32 of revolution, 30
Index 213

Symmetric space, 41, 89, 157 Tubular neighborhood, 68

pair, 89, 157 Twisted space, 66
rank, 90

V
Vector bundle, 57
T cotangent, 57
Tensor product, 192 frame bundle, 58
Theorem normal bundle, 58
Bonnet-Myers, 35 tangent, 57
Cartan, 44 Vector field, 187
Constant Rank, 20 1-parameter group, 190
Frobenius, 191 f -related, 188
Hopf-Rinow, 31 action field, 55
Inverse Function, 187 flow, 190
Levi-Civita, 29 Killing, 32
Lie’s Third, 7 Lie bracket, 188
Maximal Torus, 86 normal foliated, 112
Myers-Steenrod, 32 normal parallel, 90
Principal orbit, 75 parallel, 30
Stokes, 196 Volume form, 195
Tubular Neighborhood, 68
Thorpe’s trick, 180
Torsion, 28 W
Torus, 4, 18, 85 Wedge product, 192
infinitesimal generator, 85 Weyl
maximal, 86 chamber, 97, 121
Totally geodesic submanifold, 36, 41 group, 99, 102, 169
Transnormal map, 131 pseudogroup, 122