Marcos M. Alexandrino, Renato G. Bettiol - Lie Groups and Geometric Aspects of Isometric Actions-Springer (2015)
Marcos M. Alexandrino, Renato G. Bettiol - Lie Groups and Geometric Aspects of Isometric Actions-Springer (2015)
of Isometric Actions
Marcos M. Alexandrino • Renato G. Bettiol
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
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.
ix
x Contents
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 Sophus Lie was a nineteenth century Norwegian mathematician, who laid the foundations of
continuous symmetry groups and transformation 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
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,
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.
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
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.
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) .
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
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 ,
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
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}.
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
π (g±1 ±1 ±1 ±1 ±1 ±1
1 · · · gn ) = π (g1 ) · · · π (gn ) = h1 · · · hn = h.
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 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 .
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
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.
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
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!
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
Definition 1.36. Let G be a Lie group, and g its Lie algebra. The representation
Ad(exp(tX)) = exp(tad(X))
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
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.
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
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 ).
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 ,
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.
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 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.
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 ,
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.
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
∇ f X1 +g X2 Y = f ∇X1 Y + g ∇X2 Y ;
∇X (aY1 + bY2 ) = a ∇X Y1 + b ∇X Y2 ;
∇X ( f Y ) = f ∇X Y + (Xf )Y.
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
D D D
(X +Y ) = X + Y ;
dt dt dt
D df D
( f X) = X + f X;
dt dt dt
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
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
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.
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
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:
(iii) R satisfies a cyclic permutation property, called the first Bianchi identity:
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
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:
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)
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.,
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
V p := ker dπ p (2.10)
∇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
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].
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,
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,
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
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 (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].
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.
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 ) .
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 ).
Ric(X,Y ) = tr R(X, ·)Y = tr 14 [[X, ·],Y ] = − 14 tr[Y, [X, ·]] = − 14 B(X,Y ). (2.15)
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
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.
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
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
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
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.
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).
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
μ : (G1 × G2 ) × (M1 × M2 ) → M1 × M2 ,
(3.2)
μ (g1 , g2 ), (x1 , x2 ) := μ1 (g1 , x1 ), μ2 (g2 , x2 ) .
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)
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
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
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
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
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
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
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
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 ) .
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.
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
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.
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 ,
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
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
dim G ×H Sx0 = dim G/H + dim Sx0 = dim M = dim Tub(G(x0 )), (3.17)
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]).
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.
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
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
(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
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
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
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
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
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
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
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
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)
/ (My∼ )0 . Then:
Claim 3.103. Let x ∈ (My∼ )0 be such that x ∈
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
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].
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.
f : G → R, f (g) := Q(Ad(g)X1 , X2 ),
d
0= f exp(tY)g0
dt t=0
d
= Q Ad(exp(tY)) Ad(g0 ) X1 , X2
dt t=0
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)
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
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)
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
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.
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
Fx := {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, Fe = 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.
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)
gC := g ⊗R C, (4.5)
cos α (X ) − sin α (X)
Ad exp(X) Vα = ;
sin α (X) cos α (X)
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
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 ) .
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 .
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
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),
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:
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:
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
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,
μ : W × t → t, μ (w T, X) := Ad(w)X, (4.13)
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 ∗
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.
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.
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
• •
αj αk
104 4 Adjoint and Conjugation Actions
• •
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 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
• •
An : • • ··· • • n≥1
Bn : • • ··· • • n≥2
Cn : • • ··· • • n≥3
Dn : • • ··· • n≥4
>>
>>
>>
>>
•
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.
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.
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.
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
be a slice at β (i), for i = 0, 1, and ε > 0 small. The holonomy map is defined by
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
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
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
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).
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
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,
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
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 .
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
/Γ,
ρ : π1 (M) → W (5.6)
8 For details on compact rank one symmetric spaces, see Example 6.10.
5.4 Differential and Geometric Aspects of Polar Foliations 125
grad f 2 = b ◦ f (5.7)
Δ f = a◦ f (5.8)
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
√ √
(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
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
(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
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)
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:
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
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:
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
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 ,
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)
= 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
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.
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,
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:
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
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)
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
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
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
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
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].
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
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
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 )),
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
min secg
δg = ≤1
max secg
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.
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±
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
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
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
l
gt := ∑ fi (t)2 Q|mi , (6.29)
i=1
⎛ ⎞ ⎛ f ⎞
1
f 12 f1
⎜ ⎟ ⎜ ⎟
Pt = ⎝ ..
. ⎠ and St = − ⎜
⎝
..
. ⎟.
⎠ (6.30)
fl2 fl
fl
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,
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)
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
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
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
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
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)
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
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.
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,
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).
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
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].
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
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
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.
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
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
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
λ := sup λ1 (R + ω ). (6.48)
ω ∈Λ 4 Tp M
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 .
α : Λ 2 Tp M → Λ 2 Tp M, α (X ∧Y ), Z ∧W := AX Y, AZ W , (6.51)
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
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
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
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
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
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 )
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
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:
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 .
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
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
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 ).
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
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
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
Integration on (oriented) manifolds is defined via the above equation. Before giving
more details, let us recall the notion of orientation on smooth manifolds.
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
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,
.
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
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
4 It suffices to require that S does not intersect a fixed unbounded curve in R2 starting at the origin.
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
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
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
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
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
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