CHAPTER 6

Applications

6.1 Introduction

In this chapter, we collect a few basic and important theorems of Riemannian geometry that we
prove by using the concepts introduced so far. We also introduce some other important techniques
along the way.
We start by discussing manifolds of constant curvature. If one agrees that curvature is the main
invariant of Riemannian geometry, then in some sense the spaces of constant curvature should be
the simplest models of Riemannian manifolds. It is therefore very natural to try to understand
those manifolds. Since curvature is a local invariant, one can only expect to get global results by
further imposing other topological conditions.
Next we turn to the relation between curvature and topology. This a a central and recurring
theme for research in Riemannian geometry. One of its early pioneers was Heinz Hopf in the 1920’s
who asked to what extent the existence of a Riemannian metric with particular curvature properties
restricts the topology of the underlying smooth manifold. Since then the subject has expanded so
much that the scope of this book can only afford a glimpse at it.

6.2 Space forms

A complete Riemannian manifold with constant curvature is called a space form. If M is a space
form, its universal Riemannian covering manifold M̃ is a simply-connected space form by Proposi-
tion 3.3.8. Moreover, M is isometric to M̃ /Γ with the quotient metric, where Γ is a free and proper
discontinuous subgroup of isometries of M̃ , see section 1.3. So the classification of space forms can
be accomplished in two steps, as follows:
a. Classification of the simply-connected space forms.
b. For each simply-connected space form, classification of the subgroups of isometries acting
freely and properly discontinuously.
In this section, we will prove the Killing-Hopf theorem that solves part (a) in this program. Despite
a lot being known about part (b), it is yet an unsolved problem, and we include a brief discussion
about it after the proof of the theorem.
If (M, g) is a space form of curvature k, then, for a positive real number λ, (M, λg) is a space
form of curvature λ−1 k, see Exercise 2 in chapter 4. Therefore, the metric g can be normalized so
that k becomes equal to either one of 0, 1, or −1.

6.2.1 Theorem (Killing-Hopf ) Let M be a simply-connected space form of curvature k and

dimension bigger than one. Then M is isometric to:

c Claudio Gorodski 2007

 a. the Euclidean space Rn , if k = 0;
b. the real hyperbolic space RH n , if k = −1;
c. the unit sphere S n , if k = 1.
Proof. Fix a point p ∈ M . Let v ∈ Tp M be any unit vector, and consider the corresponding
geodesic γ(t) = expp (tv) which is defined for t ∈ R. For a vector u ∈ Tp M orthogonal to v, we
denote by U the parallel vector field along γ with U (0) = u, and we denote by Y the Jacobi field
along γ with Y (0) = 0 and Y ′ (0) = u. We know from Corollary 5.4.5 that Y (t) = d(expp )tv (tu).
Since M is complete, expp is defined on all of Tp M .
Suppose first that k = 0. The Jacobi equation in M is Y ′′ = 0, so Y (t) = tU (t). It follows that
d(expp )tv (u) = U (t) for all t. Choose a linear isometry f : Tp M → Rn with f (0p ) = 0, and set
ũ = f (u), ṽ = f (v). Also, set
F = expp ◦f −1 : Rn → M.
Then dFtṽ (ũ) = d(expp )tv (u) = U (t). Note that hũ, ũi = gp (u, u) = gγ(t) (U (t), U (t)) by the
parallelism of U . The Gauss lemma 5.5.1 implies that gγ(t) (dFtṽ (ũ), dFtṽ (ṽ)) = hũ, ṽi and it is
obvious that gγ(t) (dFtṽ (ṽ), dFtṽ (ṽ)) = hṽ, ṽi. We get that F is a local isometry. Since Rn is
complete, it follows from Proposition 3.3.8 that F is a Riemannian covering, and since M is assumed
to be simply-connected, F must be an isometry. This proves (a).
Suppose now that k = −1. Choose a point p̃ ∈ RH n and a linear isometry f : Tp M → Tp̃ RH n ,
and set ũ = f (u), ṽ = f (v). The Jacobi field Ỹ along γ̃(t) = expp̃ (tṽ) satisfying Ỹ (0) = 0 and
Ỹ ′ (0) = ũ is given by Ỹ (t) = d(expp̃ )tṽ (tũ). By Corollary 5.5.4, we know that expp̃ : Tp̃ RH n →
RH n is a diffeomorhism since Cut(p̃) = ∅, so we can define

F = expp ◦f −1 ◦ expp̃−1 : RH n → M.

Note that F (γ̃(t)) = γ(t) and

dFγ̃(t) (Ỹ (t)) = d(expp )tv ◦ f −1 ◦ d(expp̃ )−1

tṽ (Ỹ (t))
= d(expp )tv ◦ f −1 (tũ)
= d(expp )tv (tu)
= Y (t).

Now the main point is that the Jacobi equations in both M and RH n are the same, namely, Y ′′ = Y
and Ỹ ′′ = Ỹ , so Y (t) = sinh tU (t) and Ỹ (t) = sinh tŨ (t), where Ũ is the parallel vector field along
γ̃ that extends u. This together with the Gauss lemma implies that F is a local isometry, and the
rest follows as in the previous paragraph.
Finally, suppose that k = 1. Choose a point p̃ ∈ S n and a linear isometry f : Tp M → Tp̃ S n . We
know that expp̃ : B(0p̃ , π) ⊂ Tp̃ S n → S n \ {−p̃} is a diffeomorhism since Cut(p̃) = {−p̃}. Therefore
an argument similar to the one in case k = −1 making use of the fact that the Jacobi equations in
both M and S n are the same shows that the map

F = expp ◦f −1 ◦ (expp̃ |B(0p̃ ,π) )−1 : S n \ {−p̃} → M

is a local isometry. Next, choose a point p̃′ ∈ S n \ {p̃, −p̃}, and construct another local isometry

F ′ = expF (p̃′ ) ◦dFp̃′ ◦ (expp̃′ |B(0p̃′ ,π) )−1 : S n \ {−p̃′ } → M.

We have that
F ′ (p̃′ ) = expF (p̃′ ) ◦dFp̃′ (0p̃′ ) = F (p̃′ ),

88
and
dFp̃′ ′ = d(expF (p̃′ ) )0F (p̃′ ) ◦ dFp̃′ ◦ d(expp̃′ )−1
0p̃′ = dFp̃′ .

Hence F and F ′ coincide on S n \{−p̃, −p̃′ }. By exercise 15 of chapter 3, they can be pasted together
to define a local isometry S n → M , and the rest of the proof is as above. This finishes the proof.

Depending on the context in which one is interested, it is possible to find in the literature other
proofs of Theorem 6.2.1 different from the above one. The argument that we chose to use, based
on Jacobi fields, works in a more general context, and will be used to prove a generalization of
this theorem in chapter ??? of part 2. Note that the main argument in the proof of that theorem
really proves the following local result: two Riemannian manifolds of the same constant curvature
are locally isometric; the other arguments therein are used to get a global result in each one of the
three particular cases.
Next, we discuss the case of non-simply-connected space forms. In the flat case, the main result
is the following theorem.

6.2.2 Theorem (Bieberbach) A compact flat manifold M is finitely covered by a torus.

Namely, Bieberbach showed that the fundamental group π1 (M ) contains a free Abelian normal
subgroup Γ of rank n = dim M and finite index, so there is a finite covering

π1 (M )/Γ → Rn /Γ → Rn /π1 (M ) = M.

(For an example, review the contents of exercise 10 of chapter 1.) The complete classification of
compact flat Riemannian manifolds is known only in the cases n = 2, 3; see [Wol84, Cha86] for
proofs of Bieberbach’s theorem and these classifications.
Next we consider non-simply-connected space forms of positive curvature. In even dimensions,
the only examples are the real projective spaces, as the following result shows.

6.2.3 Theorem A even-dimensional space form of positive curvature is isometric either to S 2n or

to RP 2n .
Proof. We know that M = S 2n /Γ, where Γ is a subgroup of O(2n + 1) acting freely and properly
discontinuously on S 2n . Since this action is free, if an element of Γ admits a +1-eigenvalue then it
must be the identity id. Recall that the eigenvalues of an orthogonal transformation are unimodular
complex numbers, and the non-real ones must occur in complex conjugate pairs.
Next, let γ ∈ Γ. Then γ 2 ∈ SO(2n + 1), and since 2n + 1 is odd, γ 2 admits an eigenvalue +1,
thus γ 2 = id. This implies that all the eigenvalues of γ are ±1. If γ 6= id, it follows that all the
eigenvalues of γ are −1, namely, γ = −id. Hence Γ = {id} or Γ = {±id}. 
The odd-dimensional space forms of positive curvature have been completely classified by J.
Wolf [Wol84]. Here we just present a very rich family of examples.

6.2.4 Example (Lens spaces) Let p, q be relatively prime integers. The lens space Lp;q is the
quotient Riemannian manifold S 3 /Γ, where we view

S 3 = { (z1 , z2 ) ∈ C2 | |z1 |2 + |z2 |2 = 1 },

and Γ is the cyclic group of order p generated by the element

tp;q (z1 , z2 ) = (ωz1 , ω q z2 ),

89
where ω is a pth root of unity. Note that L2;1 = RP 3 . More generally, let q2 , . . . , qn be integers
relatively prime to an integer p. The lens space Lp;q2,...,qn is the quotient Riemannian manifold
S 2n−1 /Γ, where we view

S 2n−1 = { (z1 , . . . , zn ) ∈ C2 | |z1 |2 + · · · + |zn |2 = 1 },

and Γ is the cyclic group of order p generated by the element

tp;q2,...,qn (z1 , z2 , . . . , zn ) = (ωz1 , ω q2 z2 , . . . , ω qn zn ).

Of course, a lens space is a non-simply-connected space form of positive curvature. The 3-

dimensional lens spaces were introduced by Tietze in 1908. In general, lens spaces are important
in topology because they provide examples of non-homeomorphic manifolds which are homotopy-
equivalent (see [Mun84, §40, §69]). ⋆

A space form of negative curvature is called a hyperbolic manifold. Of course, a hyperbolic

manifold is isometric to the quotient of RH n by a group of isometries Γ acting freely and proper
discontinously. A compact orientable surface of genus g ≥ 2 admits many hyperbolic metrics,
which are constructed as follows. It is a theorem of Radó [Rad24] that any compact surface is
homeomorphic to the identification space of a polygon whose sides are identified in pairs. In
particular, a compact orientable surface Sg of genus g is realized as a regular 4g-sided polygon P
with a certain identification of the sides. The vertices of P are all identified to one point, so in
order to get a smooth surface it is necessary that the sum of the inner angles of P be 2π. Note
that P cannot be taken to be an Euclidean polygon, for in that case the sum of the inner angles
is known to be (4g − 2)π > 2π for g ≥ 2. Instead, we construct P as a regular polygon in the
disk model D2 of RH 2 having the center at (0, 0) and with the sides being geodesic segments. In
this case, by the Gauss-Bonnet theorem the sum of the inner angles is (4g − 2)π − A, where A
denotes the area of P . It is clear that there exist such polygons in D2 with arbitrary diameter, and
that A varies continuously with the diameter, between zero (when the diameter is near zero) and
(4g − 2)π (when the angles are near zero). Since (4g − 2)π > 2π, it follows from the intermediate
value theorem that it is possible to construct P such that the sum of the inner angles is 2π. Next
one sees that the identifications between pairs of sides can be realized by isometries of D2 such
that these isometries generate a discrete subgroup Γ of the isometry group of D2 acting freely and
properly discontinuously. This shows that Sg = D2 /Γ admits a hyperbolic metric. Further, it
is known that the hyperbolic metric on Sg for g ≥ 2 is not unique. It is a classical result that
there exist natural bijections between the following sets of structures on a compact oriented surface
Sg : conformal classes of Riemannian metrics; complex structures compatible with the orientation;
hyperbolic metrics (see e.g. [Jos06]). The moduli space Mg of Sg is the space of equivalence classes
of hyperbolic metrics on Sg , where two hyperbolic metrics belong to the same class if and only
if they differ by a diffeomorphism of Sg . It turns out that Mg is not a manifold: singularities
develop exactly at the hyperbolic metrics admitting nontrivial isometry groups. For this reason,
Teichmüller introduced a weaker equivalence relation on the space of hyperbolic metrics on Sg by
requiring two of them to be equivalent if they differ by a diffeomorphism which is homotopic to the
identity; the Teichmüller space Tg of Sg is the resulting space of equivalence classes. It is known
that Tg admits the structure of a smooth manifold of dimension 6g − 6 if g ≥ 2 [EE69].
In the higher dimensional case, it is much more difficult to construct hyperbolic metrics, and
most of the progress in this direction has been made in the 3-dimensional case, see [Thu97].

90
6.3 Synge’s theorem

We will use the following lemma in the proofs of Synge’s and Preissmann’s theorems. It is easy to
see that the compactness assumption in it is essential.

6.3.1 Lemma (Cartan) Let M be a compact Riemannian manifold. Assume that M is not
simply-connected. Then every nontrivial free homotopy class C of loops contains a closed geodesic
of minimal length in C.

Proof. We first claim that since M is compact, it is possible to find ǫ > 0 such that any two
points of M within distance less than ǫ can be joined by a unique minimizing geodesic, and this
geodesic depends smoothly on its endpoints. Indeed, cover M by finitely many balls B(pi , ǫi /2)
where pi ∈ M , ǫi > 0, and B(pi , ǫi ) is a δi -totally normal ball for some δi > 0 as in Proposition 2.4.7,
for i = 1, . . . , k. Take ǫ = mini { 12 ǫi , δi }. If d(x, y) < ǫ for points x, y ∈ M , then x ∈ B(pi0 , ǫi0 /2)
for some i0 , and then
ǫi
d(y, pi0 ) ≤ d(y, x) + d(x, pi0 ) < ǫ + 0 ≤ ǫi0 .
2
Hence x, y ∈ B(pi0 , ǫi0 ) with d(x, y) < δi0 , so the claim follows from the quoted proposition.
Let ℓ be the infimum of the lengths of the curves in C, and take a minimizing sequence (ηj ) in C
such that each ηj is parametrized on [0, 1] with constant speed. Since (ηj ) is a minimizing sequence,
L = supj L(ηj ) is finite. Choose a subdivision 0 = t0 < t1 < . . . < tn = 1 with ti − ti−1 < ǫ/2L for
i = 1, . . . , n. Then
t
ǫ
Z
d(ηj (ti−1 ), ηj (t)) ≤ ||ηj′ (t)|| dt ≤ L(ti − ti−1 ) <
ti−1 2

for ti−1 ≤ t ≤ ti . This estimate allows us to replace each curve ηj by the broken geodesic γj joining
the points ηj (0), ηj (t1 ), . . . , ηj (1). For every j, γj is homotopic to ηj ; this can be seen as follows.
Owing to
ǫ ǫ
d(γj (t), ηj (t)) ≤ d(γj (t), γj (ti−1 )) + d(ηj (ti−1 ), ηj (t)) < + = ǫ
2 2
for ti−1 ≤ t ≤ ti , we can construct a smooth homotopy from ηj |[ti−1 ,ti ] into γj |[ti−1 ,ti ] by using the
shortest geodesic from ηj (t) to γj (t).
It is clear that L(γj ) ≤ L(ηj ), so (γj ) is also a minimizing sequence in C. Using again the
compactness of M , we can select a subsequence of (γj ), denoted by the same symbol, such that
(γj (ti )) converges to a point pi as j → ∞ for all i. It follows that (γj ) converges in the C 1 -topology
to the broken geodesic γ joining the pi . It is clear that γ belongs to C and has length ℓ. Since γ is
of minimal length in C, it is locally minimizing. By Theorem 3.2.6, γ is a geodesic. 

In the case of a simply connected compact Riemannian manifold, it is still true that there exists
at least one closed geodesic (Lyusternik-Fet [LF51]). More specifically, in the case of S 2 , it is
known that every Riemannian metric must admit at least 3 geometrically distinct closed geodesics
(Lyusternik-Schnirelmann [LŠ47] 1 ).

6.3.2 Theorem (Synge) An even-dimensional orientable compact Riemannian manifold M of

positive sectional curvature must be simply connected.
1
Check Klingenberg and simpleness of curves.

91
 We remark that each one of the hypotheses in the statement of Synge’s theorem is essential.
In fact, the following manifolds are not simply-connected: RP 2 is even-dimensional, compact and
positively curved, and nonorientable; RP 3 is compact, orientable and positively curved, and odd-
dimensional; and a flat 2-torus is even-dimensional, compact and orientable and flat.
Proof of Theorem 6.3.2. Suppose, on the contrary, that M is not simply-connected and let C
denote a nontrivial free homotopy class of loops. By Lemma 6.3.1, there exists a closed geodesic
γ : [0, ℓ] → M , parametrized with unit speed, such that L(γ) = ℓ = inf η∈C L(η). Let p = γ(0) =
γ(ℓ), and denote by P : Tp M → Tp M the parallel translation map along γ from 0 to ℓ. Fix an
orientation of M . Since the parallel translation maps along γ from 0 to t, for 0 ≤ t ≤ ℓ, join P to
the identity map of Tp M , we have that P is orientation-preserving. Since γ is a geodesic, γ ′ (0) is a
fixed vector of P . Now P , being an isometry, leaves the orthogonal complement hγ ′ (0)i⊥ invariant.
Since the dimension of this subspace is odd, it contains a nonzero vector y that is fixed under P .
Let Y be the parallel vector field along γ that extends y, and construct a variation {γt } of γ with
associated variational vector field given by Y . Since M is positively curved, hR(Y, γ ′ )Y, γ ′ i < 0.
Using the variation formulas (5.3.3) and (5.3.9), we get that
d d2
E(γt ) = 0 and E(γt ) < 0.
dt t=0 dt t=0

Then, for t sufficiently small, we have that E(γt ) < E(γ) and

L(γt )2 ≤ 2ℓ E(γt ) < 2ℓ E(γ) = L(γ)2 ,

and this contradicts the fact that γ is of minimal length in C. Hence C cannot exist and M is
simply-connected. 

6.3.3 Corollary An even-dimensional compact Riemannian manifold M of positive sectional cur-

vature has fundamental group of order at most two.
Proof. Let M̃ be the orientable double cover of M . Then M̃ satisfies the hypotheses of Synge’s
theorem 6.3.2, so it is simply connected. The result follows. 
It follows from Corollary 6.3.3 that there exists no Riemannian metric of positive sectional
curvature in RP m × RP n if m + n is even. Indeed, otherwise this manifold would satisfy the
hypotheses of the corollary but its fundamental group is isomorphic to Z2 ⊕ Z2 . It is interesting to
compare this example with the fact that the nonexistence of a positively curved Riemannian metric
in S 2 × S 2 is still an unsettled question (see Add. note 4).

6.4 Bonnet-Myers’ theorem

The following result is an elementary example of a comparison theorem in Riemannian geometry.

Note that the right-hand side in (6.4.2) is exactly the Ricci curvature of the sphere S n (R).

6.4.1 Theorem (Bonnet-Myers) Let M be a complete Riemannian manifold of dimension n.

Assume there exists a constant R > 0 such that
n−1
(6.4.2) Ric(v, v) ≥ g(v, v)
R2
for every v ∈ T M . Then
diam(M ) ≤ diam(S n (R)) = πR.
In particular, M is compact and has finite fundamental group π1 (M ).

92
 Proof. Recall that diam(M ) = sup{ d(x, y) | x, y ∈ M }. This being so, for every ǫ > 0, there
exist p, q ∈ M such that d(p, q) = L and diam(M ) − ǫ < L < diam(M ). Since M is complete, there
exists even a minimal geodesic γ : [0, L] → M with unit speed and such that γ(0) = p and γ(L) = q.
Because γ is minimal, I(Y, Y ) ≥ 0 for all vector fields Y along γ vanishing at the endpoints. We
will use this remark below for some suitable vector fields.
Select an orthonormal basis {e1 , . . . , en } of Tp M with e1 = γ ′ (0), and extended it to parallel
orthonormal frame {E1 , . . . , En } along γ; of course, E1 = γ ′ . Set
πs
Yi (s) = sin Ei (s)
L
for i = 2, . . . , n. Then
Z L
I(Yi , Yi ) = −hYi′′ , Yi i + hR(γ ′ , Yi )γ ′ , Yi i ds
0
Z L  2 
2 πs π ′ ′
= sin + hR(γ , Ei )γ , Ei i ds.
0 L L2
Noting that each Yi vanishes at the endpoints of γ, we have
n Z L
π2
 
2 πs
X
′ ′
0≤ I(Yi , Yi ) = sin (n − 1) 2 − Ric(γ , γ ) ds
0 L L
i=2
 2 Z L
π 1 πs
≤ (n − 1) − sin2 ds,
L2 R 2 0 L

using the assumption on the Ricci curvature. This proves that L ≤ πR. Since L < diam(M ) is
arbitrary, we conclude that diam(M ) ≤ πR.
The other assertions in the statement can now be easily verified. The manifold M is complete
and bounded, thus, in view of Corollary 3.3.7, compact. Let M̃ denote the Riemannian universal
covering manifold of M . Since M̃ is complete and satisfies the same estimate on the Ricci curvature
as M , the previous results imply that M̃ is compact, forcing π1 (M ) to be finite. This completes
the proof of the theorem. 

6.4.3 Corollary No compact nontrivial product manifold S 1 × M admits a metric of positive Ricci
curvature.

6.4.4 Remark The assumption about the Ricci curvature in the statement of the Bonnet-Myers
theorem cannot be relaxed in the sense of requiring that the Ricci curvature only be positive, as
the following example shows. The two-sheeted hyperboloid

{ (x, y, z) ∈ R3 | x2 + y 2 − z 2 = −1 }

with the metric induced from R3 is complete, non-compact, and has Gaussian curvature at a point
(x, y, z) given by (x2 + y 2 + z 2 )−2 , which, despite being positive, goes to zero as the point tends to
infinity. ⋆

6.5 Nonpositively curved manifols

One of the main features of nonpositively curved manifols is the abundance of convex functions.
Recall that a continuous function f : I → R defined on an interval I is called convex if f ((1 −

93
t)x + ty) ≤ (1 − t)f (x) + tf (y) for every t ∈ [0, 1] and x, y ∈ I. If f is of smooth, this condition is
equivalent to requiring that its second derivative f ′′ ≥ 0. In the case of a continous function f on
a complete Riemannian manifold M , we say that f is convex if its restriction f ◦ γ is convex for
every geodesic γ of M . Strict convexity is defined analogously by replacing the inequalities above
the strict inequalities. Our point of view in this section is that most of the important results about
the geometry of manifolds with nonpositive curvature can be derived by using appropriate convex
functions on the manifold.
We will use the following remark in the proof of Lemma 6.5.1. If a convex function admits two
global minima, then a geodesic connecting these two points also consists of global minima of the
function. In fact, the function restricted to the geodesic is convex, and this implies that it cannot
have bigger values on the interior of the segment than at the endpoints forcing it to be constant
along the geodesic segment. A similar argument shows that any local minimum of a convex function
must in fact be a global one.

6.5.1 Lemma Let γ be a geodesic in a Riemannian manifold M . If the sectional curvature along
γ is nonpositive, then there are no conjugate points along γ.

Proof. Let Y be a Jacobi field along γ. We claim that the fuction f = ||Y ||2 is convex. In order
to prove this, we recall the Jacobi equation −Y ′′ + R(γ ′ , Y )γ ′ = 0 and differentiate f twice to get

f ′′ = 2(hY ′′ , Y i + ||Y ′ ||2 )

= 2(hR(γ ′ , Y )γ ′ , Y ) + ||Y ′ ||2 )
≥ 0,

in view of the assumption on the curvature; this proves the claim. Now if f (t1 ) = f (t2 ) = 0 for
some t1 < t2 , then f |[t1 ,t2 ] ≡ 0, whence Y is trivial. Hence there are no conjugate points along γ. 

6.5.2 Theorem (Hadamard-Cartan) Let M be a complete Riemannian manifold with nonpos-

itive sectional curvature. Then, for every point p ∈ M , the exponential map expp : Tp M → M is a
smooth covering. In particular, M is diffeomorphic to Rn if it is simply-connected.

Proof. Fix a point p ∈ M . In view of Lemma 6.5.1, we know that expp : Tp M → M is a local
diffeomorphism. This being so, we may endow Tp M with the pull-back metric g̃ = exp∗p g. Since a
local isometry maps geodesics to geodesics, the geodesics of (Tp M, g̃) through the origin 0p are the
straight lines, thus, defined on all of R due to the completeness of M . In view of Theorem 3.3.5(c),
this implies that (Tp M, g̃) is complete. Now expp is a covering because of Proposition 3.3.8(b), and
the last asertion in the statement is obvious. 
A complete simply-connected manifold of nonpositive sectional curvature is called a Hadamard
manifold.

6.5.3 Corollary Let M be a Hadamard manifold. Then, given p, q ∈ M , there is a unique geodesic
joining p to q.

Proof. Let γ be a geodesic joining p to q. Consider the diffeomorphism expp : Tp M → M . Then

exp−1 −1
p ◦γ is the straight line in Tp M joining the origin and expp (q), as in the proof of Theorem 6.5.2,
and this proves the uniqueness of γ. 
In particular, the preceding corollary implies that the cut-lcus of an arbitrary point in a
Hadamard manifold is empty.

94
 The Hadamard-Cartan theorem says that the universal covering manifold of a complete Rie-
mannian manifold M of nonpositive sectional curvature is Rn . Since Rn is contractible, the higher
homotopy groups πi (M ), where i ≥ 2, are all trivial. Consequently, the topological information
about M is contained in its fundamental group π1 (M ). In the following, we prove some classical
results about the fundamental group of nonpositively curved manifolds. We start with a lemma.

6.5.4 Lemma Let M be a Hadamard manifold. Then, for any point p ∈ M , the function fp :
M → R given by fp (x) = 12 d(p, x)2 is smooth and strictly convex.

Proof. Fix a point p ∈ M . Denote by γ x : [0, 1] → M the unique geodesic parametrized with
constant speed joining p to x. Plainly, γ x is minimizing, so

1 1 1
fp (x) = L(γ x )2 = E(γ x ) = ||γ x′ (0)||2 = || exp−1 2
p (x)|| ,
2 2 2
showing that fp is smooth.
Next, let η be a geodesic; we intend to verify that f ◦ η is strictly convex. For that purpose, we
set γt = γ η(t) and invoke the second variation formula (5.3.9) to write:

d2 d2
(fp ◦ η)(t) = E(γt )
dt2 t=0 dt2 t=0
(6.5.5)
∂¯ 1 1
Z
= h∇ ∂ ′
,γ i + ||Y ′ ||2 + hR(γ ′ , Y )γ ′ , Y i ds.
∂t ∂t t=0 0 0

¯∂ |
Since the variational vector field Y = ∂t ¯∂ ′′
t=0 vanishes at s = 0 and ∇ ∂ ∂t s=1 = η (0) = 0, the first
∂t t=0
term in the sum is zero; the assumption on the curvature and the fact that Y is nonzero imply that
the second term there is positive. We conclude that f is strictly convex. 

6.5.6 Remark We can get more refined information about the second derivatives of fp . It im-
mediately follows from the Cauchy-Schwarz
R 1 ′ 2inequality that a smooth function f : [0, 1] → R with
f (0) = 0 must satisfy the inequality (f ) ds ≥ f (1)2 . Retaining the notation in the proof of
P 0
Lemma 6.5.4, we write Y (s) = i ai (s)Ei (s) for smooth functions ai : [0, 1] → R and an orthonor-
mal frame {Ei } of parallel vectors along γ0 . Then
Z 1 XZ 1
′ 2
||Y || ds = (ai )′2 ds
0 i 0
X
≥ ai (1)2
i
= ||Y (1)||2
= ||η ′ (0)||2 .

Together with (6.5.5), this shows that2

Hess(fp ) ≥ g

at every point of M , as bilinear symmetric forms. ⋆

2
Define Hessian.

95
 Lemma 6.5.4 allows one to generalize the notion of center of mass of a finite set of points in
Euclidean space to the context of Hadamard manifolds. For that purpose, two remarks are in order.
First, we note that a non-negative strictly convex proper function has a unique minimum. In fact,
because of properness, there must a minimum. If there were two minima, the function would be
strictly convex when restricted to a geodesic joining the two minima, and this would imply that
the function has smaller values on the interior of this segment than at the endpoints, contradicting
the fact that the endpoints are minima. The second remark is that the maximum of any number of
strictly convex functions is still strictly convex, as one sees easily. Now, given a finite set of points
p1 , . . . , pk in a Hadamard manifold, the center of mass of the set {p1 , . . . , pk } is defined to be the
uniquely defined minimum of the non-negative strictly convex proper function

x 7→ max{fp1 (x), . . . , fpk (x)}.

6.5.7 Theorem (Cartan) Let M be a Hadamard manifold. Then any isometry of finite order of
M has a fixed point.

Proof. Let ϕ be an isometry of M of order k ≥ 1. For an arbitrary point p ∈ M , set q to be the

center of mass of the finite set {p, ϕ(p), . . . , ϕk−1 (p)}. This means that q is the unique minimum
of the function
f (x) = max{fp (x), fϕ(p) (x), . . . , fϕk−1 (p) (x)}.

Since ϕk (p) = p and ϕ is distance-preserving,

1
max d(p, ϕ(q))2 , d(ϕ(p), ϕ(q))2 , . . . , d(ϕk−1 (p), ϕ(q))2

f (ϕ(q)) =
2
1
max d(ϕk−1 (p), q)2 , d(p, q)2 , . . . , d(ϕk−2 (p), q)2

=
2
= f (q),

which shows that also ϕ(q) is a minimum of f . Hence, ϕ(q) = q. 

6.5.8 Corollary Let M be a complete Riemannian manifold of nonpositive sectional curvature.

Then the fundamental group of M is torsion-free.

Proof. The Riemannian universal covering M̃ of M is a Hadamard manifold, and the elements
of π1 (M ) act on M̃ as deck transformations, thus, without fixed points; Theorem 6.5.7 implies that
they cannot have finite order. 
Before proving the next theorem, we recall some facts about the relation between the funda-
mental group π1 (M, p) and the set of free homotopy classes of loops, which we denote by [S 1 , M ],
for a connected manifold M and p ∈ M .

6.5.9 Lemma The ‘forgetful’ map F : π1 (M, p) → [S 1 , M ], which is obtained by ignoring base-
points, sets up a one-to-one correspondance between [S 1 , M ] and the set of conjugacy classes in
π1 (M, p).

Proof. If γ, η are loops based at p then F[η · γ · η −1 ] = F[η] · F[γ] · F[η −1 ] = F[η −1 ] · F[η] · F[γ] =
F[γ], where for the second equality we cyclically permite the order of concatenation by changing
the basepoint. This proves that F is constant on conjugacy classes.
Conversely, let γ0 , γ1 : [0, 1] → M be loops based at p with F[γ0 ] = F[γ1 ]. This means there
is a homotopy γt from between those curves without necessarily preserving basepoints. The curve

96
c(t) = γt (0) = γt (1) traces out the path taken by the basepoints and thus is a loop. Now the
concatenation γ̃t = c|[0,t] · γt · (c|[0,t] )−1 is a homotopy from γ0 to c · γ1 · c−1 preserving basepoints.


6.5.10 Lemma Let γ, η be loops in M based at p, q, respectively. Then the classes [γ] = [η] in
[S 1 , M ] if and only if [γ] ∈ π1 (M, p) and [η] ∈ π1 (M, q) act by the same deck transformation on the
universal cover M̃ .
Proof. Let ζ be a curve joining p to q. Then ζ · η · ζ −1 is in the same free homotopy class as
η. Using Lemma 6.5.9, by concatenating ζ with a loop at p, we may assume that ζ is such that
[γ] = [η] in [S 1 , M ] if and only if [ζ · η · ζ −1 ] = [γ] in π1 (M, p). The desired result follows from the
standard relation between the fundamental group and deck transformations. 

6.5.11 Theorem (Preissmann) Let M be a compact Riemannian manifold of negative sectional

curvature. Then every nontrivial Abelian subgroup of its fundamental group is infinite cyclic.
Proof. We can assume that M is not simply-connected. Let M̃ be the Riemannian universal
covering of M , and let ϕ ∈ π1 (M ) an element different from the identity which we view as an
isometry of M̃ . Recall that ϕ acts on M̃ without fixed points. The fundamental remark is that
the displacement function f : M̃ → R given by f (x) = d(x, ϕ(x)) is smooth and convex. For the
purpose of proving this claim, consider the function Φ : T M → M ×M , given by Φ(v) = (x, expx (v))
for v ∈ Tx M , that was introduced in Lemma 2.4.6. Since M̃ is a Hadamard manifold, we easily
see that Φ is well defined and a global diffeomorphism. Now d : M̃ × M̃ \ ∆M̃ → R is given by
d(x, y) = gx (Φ−1 (x, y), Φ−1 (x, y))1/2 , so it is also smooth; here ∆M̃ denotes the diagonal of M̃ . This
proves that f is smooth. In order to prove the convexity of f , we resort to the second variation
formula of the length given in exercise 1 of chapter 5. Let η be a geodesic; similarly to in (6.5.5),
we can write
Z 1
d2
(6.5.12) (f ◦ η)(t) = ||Y⊥′ ||2 + hR(γ ′ , Y⊥ )γ ′ , Y⊥ i ds ≥ 0,
dt2 t=0 0

where γt is the geodesic joinig η(t) to ϕ(η(t)), Y is the variational vector field along γ0 and Y⊥
denotes its normal component, and we have used that ∇ ∂ ∂t ¯∂ is equal to η ′′ (0) = 0 and (ϕ ◦
∂t t=0
η)′′ (0) = 0 for s = 0 and 1, respectively. Although f is not strictly convex, we can derive more
refined information from formula (6.5.12). Since M̃ has negative curvature, the equality holds
in (6.5.12) if and only if Y is a constant multiple of γ ′ , so at any given point x ∈ M̃ , f is stricly
convex in any direction different from the direction of the geodesic joining x to ϕ(x).
Next, we introduce a definition. An axis of ϕ is a geodesic of M̃ that is invariant under ϕ.
Note that ϕ cannot reverse the orientation of an axis γ for otherwise the midpoint of the geodesic
segment between γ(t) and ϕ(γ(t)) would be a fixed point of ϕ for any t ∈ R. Hence the restriction
of ϕ to γ must be translation along it:

ϕ(γ(t)) = γ(t + t0 )

for some t0 ∈ R and all t ∈ R. The number t0 will be called the period of ϕ along the axis γ. For
later reference, we also note that

f (ϕ(x)) = d(ϕ(x), ϕ2 (x)) = d(x, ϕ(x)) = f (x)

for every x ∈ M̃ .

97
 Now we give three important properties of axes. The first one is that f is constant along an
axis γ of ϕ. Indeed,
f (γ(t + t0 )) = f (ϕ(γ(t))) = f (γ(t))
for all t ∈ R, where t0 is the period of γ. It follows that f ◦ γ is convex and periodic, and it is
easy to see that such a function must be constant. The second one is that an axis of ϕ is a set of
minima of f . This follows immediately from the formula of the first variation of length. The last
one is that if f is constant on a geodesic segment xy for points x 6= y, then the supporting geodesic
γ of that segment is an axis of ϕ. Indeed, f is not stricly convex along xy, so γ must coincide with
the geodesic joining x and ϕ(x). It follows that ϕ(x) lies in the image of γ. Similarly, ϕ(y) lies in
the image of γ. Since a geodesic in M̃ is uniquely defined by two points on it, γ must be an axis
of ϕ.
The next step is to prove that ϕ admits one and only one axis, up to reparametrization and
reorientation. Note that the value f at a point x ∈ M̃ is the length of the unique geodesic in M̃
joining x to ϕ(x). Such geodesics project to geodesics in M all lying in the same free homotopy
class of loops in M , independent of the point x̃, according to Lemma 6.5.10. Since M is compact,
f admits a global minimum p ∈ M̃ by Lemma 6.3.1. Since f (ϕ(p)) = f (p), we have that ϕ(p) is
also a global minimum. By convexity, f is constant along the geodesic segment joining p and ϕ(p);
let γ be the unit speed geodesic that supports this segment. By the above, γ is an axis of ϕ. Now
the points in the image of γ comprise a set of minima at each point of which f is strictly convex in
any direction different from γ. It follows that there cannot be another axis.
Finally, suppose that H is an Abelian subgroup of π1 (M ), and that ϕ belongs to H and has γ
as an axis as above. Since the elements of H commute with ϕ, they map γ to a geodesic which is
invariant under ϕ; by the above uniqueness result, γ is an axis for all the elements of H. Consider
now the period map H → R. This map is clearly an injective homomorphism, thus its image is a
subgroup of R isomorphic to H. It is not difficult to see that every subgroup of R is either infinite
cyclic or dense. Since the orbits of H on M̃ are discrete, H must be infinite cyclic. 

6.5.13 Corollary No compact nontrivial product manifold M × N admits a metric with negative
sectional curvature.

Proof. Suppose, on the contrary, that M × N supports a metric of negative sectional curvature.
Notice that M and N , being compact, cannot be simply-connected by the Hadamard-Cartan theo-
rem 6.5.2. Since π1 (M ) and π1 (N ) are non-trivial, they contain non-trivial cyclic groups H and K,
respectively. But then H × K is a non-trivial Abelian subgroup of π1 (M ) × π1 (N ) ∼ = π1 (M × N )
which is not infinite cyclic, contradicting Preissmann’s theorem. This proves the corollary. 

6.5.14 Remark An isometry ϕ of a Hadamard manifold M̃ can be of three types. Let f be the
displacement function associated to ϕ as in Preissmann’s theorem 6.5.11. Then ϕ is said to be:
a. elliptic if f attains the value zero (i.e. ϕ admits a fixed point);
b. hyperbolic if f attains a positive minimum;
c. parabolic if f attains no minimum.
The argument in Preissmann’s theorem proves that a hyperbolic isometry of a Hadamard manifold
admits an axis (which is unique in the case in which the curvature of M̃ is negative).

6.6 Additional notes

§1 The Gauss-Lobatchevsky-Bolyai discovery of hyperbolic geometry in the early nineteenth century

finally pointed out the impossibility of proving Euclid’s fifth postulate from the other postulates of

98
Euclidean geometry. In 1868, Beltrami proved the consistency of hyperbolic geometry by realizing
it as the intrinsic geometry of a well known surface in Euclidean 3-space — the so-called pseudo-
sphere — which has constant negative curvature. In his Habilitationsvortrag of 1854 in which
Riemann laid the foundations of Riemannian geometry were also exhibited examples of metrics
of arbitrary constant curvature. Based on Riemann’s ideas, Beltrami published another article
in 1869 in which he discussed spaces of constant curvature in arbitrary dimensions. In this way,
the non-Euclidean geometries were for the first time incorporated into the realm of Riemannian
geometry. In 1890, Klein drew attention to Clifford’s 1873 discovery of a 2-torus — nowadays known
as the Clifford torus — sitting in S 3 with constant zero curvature and formulated the problem of
classifying Riemannian manifolds of arbitrary constant curvature in arbitrary dimensions. The
problem, referred to as the Clifford-Klein space forms problem, was extensively studied by Killing
in an article in 1891 and a book in 1893, and then again by Heinz Hopf in 1925 culminating in
Theorem 6.2.1.
§2 The argument in the proof of the Hadamard-Cartan theorem 6.5.2 shows that if there is a
point in a simply-connected Riemannian manifold possessing no conjugate points, then the manifold
is diffeomorphic to Euclidean space. Eberhard Hopf [Hop48] proved that a compact Riemannian
manifold M without conjugate points satisfies the inequality
Z
scal ≤ 0
M

where the integral is taken with respect to the canonical Riemannian measure 3 , and the equality
holds if and only if M is flat. In the 2-dimensional case, the left-hand side equals 2π times the Euler
characteristic of M by the Gauss-Bonnet theorem. It follows E. Hopf’s result that a metric without
conjugate points on T 2 must be flat. It was a long standing conjecture that the same result should
be also valid for the higher dimensional tori. In 1994, Burago and Ivanov [BI94] finally settled the
conjecture in the positive sense.
§3 Techniques from geometric analysis have been proved to be very powerful in dealing with
problems involving curvature in Riemannian manifolds. We would like to mention two spectacular
instances of this fact. In 1960, Yamabe [Yam60] tried to deform conformally a given Riemannian
metric g on a manifold M into a metric f · g of constant scalar curvature, where f is an unknown
positive smooth function on M . If n = dim M = 2, this is classical result and amounts to showing
that M admits isothermal coordinates [Jos06], so he was dealing with the case n ≥ 3. There was
a problem with Yamabe’s arguments, though, and the question became the Yamabe problem. In
order to find f , one needs to solve the nonlinear partial differential equation
n−2 n+2
∆f + scal(M, g) = f n−2 .
4(n − 1)

This is an extremely difficult question in analysis because the exponent of f is exactly the “critical
exponent” in regard to which the standard Sobolev embedding theorems do not apply. The problem
was eventually solved through the work of of Aubin [Aub76] and Schoen [Sch84]. Thanks to
contributions by other mathematicians, the Yamabe problem is today almost completely understood
and it is known that the set of metrics of constant scalar curvature in a given conformal class of
metrics is an infinite-dimensional space if n > 2. See [Aub98] for these results in book form.
Deformation techniques like that concerning the Yamabe problem are used to prove the existence
of several objects in geometry. An interesting approach is to consider deformations on the level of
3
Ref?

99
the space of Riemannian metrics on a given smooth manifold M . For instance, Hamilton [Ham82]
introduced the following normalized Ricci flow equation in the space of Riemannian metrics on a
compact n-dimensional manifold M :
d τ
g(t) = −2Ric(g(t)) + 2 g(t),
dt n
where Ric(g(t)) denotes the Ricci curvature of the metric g(t), and τ denotes the integral of the
scalar curvature of g(t). The fixed points of this equation are the metrics of constant Ricci cur-
vature. One considers t as time and studies the equation as an initial value problem for a fixed
Riemannian metric g0 = g(0) on M . Hamilton proved that if n = 3 and the Ricci curvature
of g0 is positive, then the Ricci flow converges smoothly to a metric of constant Ricci curvature.
In particular, the manifold is diffeomorphic to a spherical space form. At that time, this was a
very interesting application of Riemannian geometry to provide a partial answer to a long-standing
open problem in topology, the so called Poincaré conjecture: Is every simply-connected compact
3-dimensional manifold homeomorphic to S 3 ? The difficulty in using Hamilton’s method to prove
the full Poincaré conjecture was that if one removes the assumption that Ric(g0 ) > 0, then the Ricci
flow develops finite-time singularities that impede the convergence to a nice metric, and those sin-
gularities were not completely understood. As it turns out, Perelman was able to overcome those
analytic difficulties. He extended Hamilton’s results and in particular proved the full Poincaré
conjecture (see e.g. [MT06]).
§4 A famous, open conjecture of Heinz Hopf asserts that S 2 × S 2 does not admit a metric
of positive sectional curvature. Indeed, known examples of simply-connected compact manifolds
with positive sectional curvature are relatively rare (owing to the Bonnet-Myers theorem 6.4.1, the
non-simply-connected examples are quotients of the simply-connected ones by finite subgroups of
isometries). The standard examples are the compact rank one symmetric spaces (see Add. notes ?
of chapter ?). Apart from these, the homogeneous examples have been classified by Wallach [Wal72]
in the odd-dimensional case and by Bérard-Bergery [BB76] in the even dimensional case. These
examples occur only in dimensions 6, 7, 12, 13 and 24, and are due to Berger, Wallach and Allof-
Wallach. The only other examples known are given by biquotients G//H. Here G is a Lie group
equipped with a bi-invariant metric and H is subgroup of G×G acting on G by (h1 , h2 )·g = h1 gh−1 2 .
This action is always proper and isometric, and if it is also free, then the quotient space is a manifold
denoted by G//H. In this case, there is a unique metric on G//H making the projection G → G//H
into a Riemannian submersion and it follows from Proposition 4.5.8 that G//H has always non-
negative curvature. More generally, one can also construct bi-quotients by considering left-invariant
metrics on G more general than the bi-invariant ones. It turns out that the only known examples
of positively curved biquotients occur in dimensions 6, 7 and 13, and these are due to Eschenburg
and Bazaikin. There is no general classification of positively curved biquotients. See [Zil07] for a
recent survey on these results and related ones.

6.7 Exercises

1 Prove that an odd-dimensional compact Riemannian manifold of positive sectional curvature is

orientable.

2 Let M be a complete Riemannian manifold of nonpositive curvature. Prove that each homotopy
class of curves with given endpoints in M contains a unique geodesic.

3 Consider the disk model Dn of RH n and let ϕ be an isometry of RH n .

100
 a. Prove that ϕ uniquely extends to a homeomorphism of the closed ball Dn . (Hint: Use
exercise 4 of chapter 3.)
b. Prove that ϕ is hyperbolic if and only if its extension to Dn admits exactly two fixed points
and those lie in the boundary S n−1 .
c. Prove that ϕ is parabolic if and only if its extension to Dn admits exactly one fixed point
and that lies in the boundary S n−1 .

4 Let G be an Abelian subgroup of the fundamental group of a nonflat space form M . Prove that
G is cyclic.

5 An isometry ϕ of a Riemannian manifold M is called a Clifford translation if the associated

displacement function x 7→ d(x, ϕ(x)) is constant. Prove that:
a. The Clifford translations for Rn are just the ordinary translations.
b. The only Clifford translation of RH n is the identity transformation.
c. A linear transformation A ∈ O(n + 1) is a Clifford trsnaformation of S n+1 if and only if either
A = ±I or there is a unimodular complex number λ such that half the eigenvalues of A are
λ and the other half are λ̄.

6 Let M be a Hadamard manifold. Prove that an isometry ϕ of M is a Clifford translation

(cf. exercise 5) if and only if the vector field X on M given by expp (Xp ) = ϕ(p) is parallel.

7 Extend Preissmann’s theorem 6.5.11 to show that every solvable subgroup of the fundamental
group of a compact Riemannian manifold of negative curvature must be infinite cyclic.

8 In this exercise, we prove that a compact homogeneous Riemannian manifold M whose Ricci
tensor is negative semidefinite everywhere is isometric to a flat torus.
a. Use exercise 8 of chapter 5 to show that the identity component of the isometry group of M
is Abelian.
b. Check that M can be identified with an n-torus equipped with a left-invariant Riemannian
metric.
c. Show that an n-torus equipped with a left-invariant Riemannian metric admits a global
parallel orthonormal frame and hence is flat.

101