More differential geometry

Matthias Ludewig
Version of August 30, 2023

5.5 Riemannian coverings and the theorem of Cartan-Hadamard

Recall that a smooth map ϕ : N → M between manifolds is called a covering if for all
p ∈ M , there exists an open neighborhood U of p such that
a
ϕ−1 (U ) = Uα ,
α∈I

where I is some index set, Uα is an open subset of N for each α ∈ I and ϕ|Uα : Uα → U is
a diffeomorphism. Any connected manifold M has a universal cover, which is a connected
and simply connected cover π : M̃ → M of M , uniquely determined up to diffeomorphism.
The fundamental group π1 (M ) acts freely and transitively on the fibers π −1 (p), for each
p ∈ M.
Example 5.1. (a) If F is a countable set and N = M × F , then the projection N → M
onto the first factor is a covering map. Such coverings are called trivial.
(b) For N = Rn and M = Rn /Zn the torus, the quotient map is a covering map.

Definition 5.2 (Riemannian covering). Let (N, h) and (M, g) be Riemannian mani-
folds. A covering map ϕ : N → M is a Riemannian covering if it is a local isometry.
Exercise 5.3. Let (N, h) and (M, g) be Riemannian manifolds and let ϕ : N → M be a
covering map. Show that if (M, g) is complete, then so is (N, h).
Proposition 5.4 (Consequence of Bonnet-Myers). Let (M, g) be a compact Rieman-
nian manifold with positive Ricci curvature. Then its fundamental group π1 (M ) is finite.
Proof. Let κ be such that ric ≥ κ(n − 1)g. Let M̃ be the universal cover of M and equip if
with the Riemannian metric g̃ obtained by pulling back the metric g along the projection
π : M̃ → M . Then π is a Riemannian covering. In particular, π is a local isometry, hence
M̃ again satisfies ricg̃ ≥ κ(n − 1)g̃. Moreover, by Exercise 5.3, (M̃ , g̃) is complete.
This shows that (M̃ , g̃) satisfies the assumptions of the Bonnet-Myers theorem, hence M̃
is compact. It is now a standard fact from topology that if a covering space is compact,
then the cover must be finite. Hence π : M̃ → M is a finite cover, so π1 (M ) (which
bijectively corresponds to the fibers of this cover) must be finite. 

1
For the next result, we need the following exercise.
Exercise 5.5. Let (N, h) and (M, g) be connected and complete Riemannian manifolds.
Show that a surjective local isometry ϕ : N → M is always a covering map.
Theorem 5.6 (Cartan-Hadamard). Let (M, g) be a complete Riemannian manifold
with sectional curvature K ≤ 0. Then for all p ∈ M , expp : Tp M → M is a covering map.
Proof. Let p ∈ M . Since M is complete, the exponential map expp is defined on all of
Tp M .
We show that the differential d expp |X : TX Tp M ∼= Tp M → Texpp (X) M is an isomorphism
for each X ∈ Tp M . To this end, we check that d expp |X has trivial kernel. We have
d expp |0 = id, so let now X 6= 0 and let γ(t) = expp (tX) is the geodesic with γ̇(0) = X.
Then we have
d expp |X (Y ) = J(1),
∇
where J is the Jacobi field along γ with J(0) = 0 and dt J(0) = Y . We claim that J has
no zeros. To this end, we set f (t) = 21 kJ(t)k2 and calculate
 
00 d ∇
f (t) = J(t), J(t)
dt dt
 2  2
∇ ∇
= J(t), J(t) + J(t)
dt2 dt
2
∇
= R(γ̇(t), J(t))γ̇(t), J(t) + J(t)
dt
2

  ∇
= −K span{γ̇(t), J(t)} · kJ(t)k2 kγ̇(t)k2 − hJ(t), γ(t)i2 + J(t)
dt
Because K ≤ 0, this is non-negative. Moreover, f satisfies
 
0 ∇
f (0) = 0, f (0) = J(0), J(0) = 0, f 00 (0) ≥ kY k2 > 0.
dt
In total, this implies that f (t) > 0 for all t > 0. Hence J(t) is never zero, so d expp |X
has trivial kernel and is therefore an isomorphism. We conclude that expp is a local
diffeomorphism.
Define a Riemannian metric g̃ on Tp M by g̃ = exp∗p g. This turns expp into a local isometry.
We claim that (Tp M, g̃) is complete. Since expp is a local isometry, the geodesics through
0 ∈ Tp M are precisely the straight lines going through zero; indeed, by definition of
the exponential map, they are mapped to the geodesics γ(t) = expp (tX) under expp ,
and preimages of geodesics under local isometries are again geodesics. Hence (Tp M, g̃) is
geodesically complete at 0. By the Hopf-Rinow theorem, (Tp M, g̃) is complete, as claimed.
By Exercise 5.5 and completeness of (M, g), expp is a covering map. 
Remark 5.7. If M is simply connected, the theorem of Cartan-Hadamard implies that M
must be diffeomorphic to Rn .

2
5.6 The injectivity radius
Let (M, g) be a Riemannian manifold and let γ : (a, b) → M be a geodesic. Recall that
two points t1 , t2 ∈ (a, b) are conjugate if there exists a non-trivial Jacobi field J along γ
with J(t1 ) = 0 and J(t2 ) = 0.

Lemma 5.8. Let (M, g) be a Riemannian manifold and let γ : [0, r] → M be a geodesic.
Suppose that t1 ∈ [0, r] is conjugate to zero along γ. Then γ is not length minimizing
among piecewise C 1 -curves between γ(0) and γ(t), for any t > t1 .

The proof of this lemma will use the second variation formula for the energy functional
at a geodesic. Recall that if γ : [a, b] → M is a geodesic and (γs )s∈(−ε,ε) is a variation with
∂
fixed endpoints and variational vector field X(t) = ∂s |s=0 γs (t), then we have

2 ˆ b 2
!
∂ ∇
D2 E|γ (X, X) := 2


E[γs ] = X(t) − R X(t), γ̇(t) γ̇(t), X(t) dt.
∂s s=0 a dt
(5.1)
The same formula is true if the variational vector field is piecewise smooth, in the sense
that there exists a subdivison a = t0 < t1 < · · · < tN = b of [a, b] such that the curves
γs are still continuous, but smooth only on the subintervals [tj−1 , tj ]. In particular, (5.1)
defines a quadratic form on the space of piecewise C 1 vector fields along γ. Polarizing
this quadratic form, we obtain the bilinear form
ˆ b  2 !
∇ ∇
D2 E|γ (X, Y ) =


X(t), Y (t) − R X(t), γ̇(t) γ̇(t), Y (t) dt,
a dt dt

defined for two piecewise C 1 vector fields X, Y along γ.

Proof. Fix t2 > t1 . We will show that there exists a vector field Y along γ with Y (0) = 0
and Y (t2 ) = 0 such that the second variation D2 E|γ of the energy is negative on Y . Then
any variation with fixed end points and variational vector field Y will then produce a
curve between γ(0) and γ(t2 ) shorter than γ.
Because 0 and t1 are conjugate along γ, there exists a non-trivial Jacobi field J such that
J(0) = 0 and J(t1 ) = 0. Let J ∗ be the Jacobi field such that J ∗ (t) = J(t) for t ≤ t1
and J ∗ (t) = 0 for t > t1 . Then J ∗ is a piecewise C 1 vector field. Let moreover X be a
∇
piecewise C 1 vector field such that X(0) = 0 and X(t2 ) = 0 and X(t1 ) = − dt J(t1 ). Then
∇
X(t1 ) 6= 0, because if we had dt J(t1 ) = 0, then J would be trivial. Now for ε > 0, set
Y = J ∗ + εX. Then

D2 E|γ (Y, Y ) = D2 E|γ (J ∗ , J ∗ ) + 2εD2 E|γ (J ∗ , X) + ε2 D∗ E|γ (X, X). (5.2)

3
For the first term, we have
ˆ t1 2
!
∇
D2 E|γ (J ∗ , J ∗ ) =


J(t) − R J(t), γ̇(t) γ̇(t), J(t) dt
0 dt
ˆ t1   2  
∇

= − J(t), J(t) + R γ̇(t), J(t) γ̇(t), J(t) dt
0 dt2
= 0.

Here we used that J ∗ is non-zero only on [0, t1 ], then integrated by parts and then used
that J vanishes at t = 0 and t = t1 and satisfies the Jacobi equation. For the next term,
we calculate
ˆ t1   
2 ∗ ∇ ∇

D E|γ (J , X) = J(t), Y (t) − R J(t), γ̇(t) γ̇(t), X(t) dt
0 dt dt
 
∇
= J(t), X(t)
dt t=t1
ˆ t1   2  
∇

+ − J(t), X(t) + R γ̇(t), J(t) γ̇(t), X(t) dt
0 dt2
| {z }
=0, as before
2
∇
=− J(t1 ) < 0,
dt

by the definition of X. We conclude that the first term (5.2) is zero and the second is
negative. Since the second term is linear in ε and the third is quadratic in ε, we can
choose ε small enough to obtain D2 E|γ (Y, Y ) < 0, as desired. 

Theorem 5.9. Let (M, g) be a Riemannian manifold and let r > 0 be such that expp is
defined on Br (0) ⊆ Tp M . Then the restriction of expp to Br (0) is injective if and only if
it is a diffeomorphism onto its image.

Proof. Clearly, if expp |Br (0) is a diffeomorphism onto its image, then it is injective.
Conversely, suppose that expp |Br (0) is not a diffeomorphism onto its image. We need to
show that it is not injective. Since expp |Br (0) is not a diffeomorphism onto its image, it is
either not injective (in which case there is nothing to show) or there is a point X ∈ Br (0)
such that d expp |X is not injective. So let us assume the latter. Since d expp |0 = idTp M ,
we have X 6= 0. Let γ(t) = expp (tX) be the geodesic with γ̇(0) = X.
Since d expp |X is not injective, there exists a non-zero Y ∈ TX Tp M ∼ = Tp M such that
d expp |X (Y ) = 0. On the other hand, we have

d expp |X (Y ) = J(1),
∇
where J is the Jacobi field along γ with J(0) = 0 and dt
J(0) = Y . Hence t = 0 and t = 1
are conjugate along γ.

4
By Lemma 5.8, γ is not minimizing between γ(0) and γ(s), for any s > 1. On the other
hand, since Br (p) ⊆ M is open, there exists some s > 1 such that γ(s) ∈ Br (p). Let
γ 0 be a minimizing geodesic between p and γ(s). Since γ 0 is shorter than γ, we have
X 0 := γ̇ 0 (0) ∈ Br (0). We obtain that

expp (sX) = γ(s) = γ 0 (1) = expp (X 0 ).

Since γ 0 is shorter than γ as a geodesic between p and γ(s), they cannot be reparametriza-
tions of each other, hence sX 6= X 0 . We conclude that expp is not injective. 

Recall that the injectivity radius injrad(p) at a point p in a Riemanian manifold (M, g)
is the supremum over all r > 0 such that expp is a diffeomorphism onto its image when
restricted to Br (0) ⊂ Tp M . By the above theorem, an alternative definition is

injrad(p) = sup{r > 0 | expp |Br (0) is injective}. (5.3)

We also have the following characterization.

Lemma 5.10. If d(p, expp (X)) = kXk for all X ∈ Br (0) ⊂ Tp M , then expp is injective
on Br (0).

Proof. Suppose that there exist vectors X, X 0 ∈ Br (0) such that expp (X) = expp (X 0 ).
We have to show X = X 0 . By the assumption of the lemma,

kXk = expp (p, X) = expp (X 0 ) = kX 0 k.

Let γ(t) = expp (tX) and γ 0 (t) = expp (tX 0 ). Let s > 1 such that γ(s) is still contained
Br (p). Then since γ and γ 0 are both length minimizing between p and γ(1) = γ 0 (1), the
curve (
γ 0 (t) t ∈ [0, 1]
η(t) =
γ(t) t ∈ [1, s]
is minimizes the length between p and γ(s). But since shortest curves are geodesics, η
must be in fact smooth, hence γ̇(1) = γ̇ 0 (1), which implies γ = γ 0 and hence X = X 0 . 

Theorem 5.11. Let (M, g) be a complete Riemannian manifold. Then injrad(p) is a

continuous function of p ∈ M .

Proof. We show that injrad(p) is both upper and lower semicontinuous in p, which implies
that injrad is continuous.
(a) We first show that injrad(p) is upper semicontinuous in p, in other words,

lim sup injrad(pk ) ≤ injrad(p)

k→∞

whenever (pk )k∈N is a sequence in M converging to p. We verify the assumption of

Lemma 5.10. Write rk = injrad(pk ) and r∗ = lim supk→∞ rk . For X ∈ Br∗ (0) ⊂ Tp M ,

5
choose a sequence of vectors Xk ∈ Tpk M converging to X. Then for k ∈ N large enough,
Xk ∈ Brk (0) ⊂ Tpk M . Therefore, setting qk = exppk (Xk ), the geodesic γk (t) = exppk (tXk )
between pk and qk is minimizing, hence d(pk , qk ) = kXk k. Since Xk → X, we have
qk → expp (X), hence by continuity of the distance function,

d(p, expp (X)) = lim d(pk , qk ) = lim kXk k = kXk.

k→∞ k→∞

As this holds for any X ∈ Br∗ (0), Lemma 5.10 implies that expp is injective on Br∗ (0) ⊂
Tp M . By (5.3), we therefore obtain injrad(p) ≥ r∗ .
(b) We show that injrad(p) is lower semicontinuous in p, in other words,

lim inf injrad(pk ) ≥ injrad(p)

k→∞

whenever (pk )k∈N is a sequence in M converging to p. Suppose this is not true. Then,
setting rk = injrad(pk ) and r = injrad(p), there exists ε > 0 such that

lim inf rk < r − ε.

k→∞

Therefore, after possibly passing to a subsequence, we have rk < r−ε. By Thm. 5.9, exppk
is not injective when restricted to Br−ε (0) ⊂ Tpk M , so there exist two distinct vectors
Xk , Yk ∈ Tpk M such that exppk (Xk ) = exppk (Yk ) and such that kXk k, kYk k ≤ r − ε. By
compactness, after passing to subsequences, Xk and Yk converge to vectors X, Y ∈ Tp M
with expp (X) = expp (Y ) and kXk, kY k ≤ r − ε. If X 6= Y , we can conclude that exp is
not injective on Br−ε (0) ⊂ Tp M , a contradiction to injrad(p) = r.
It therefore remains to show that X 6= Y ∈ Tp M . To this end, we consider the smooth
map
F := π × exp : T M → M × M,
where π : T M → M is the bundle projection. We claim that the differential dF |X :
TX T M → Tp M × Texp(X) M has full rank. To see this, we choose a subspace H ⊂ TX T M
(a space of “horizontal vectors”) such that dπ|p : H → Tp M is a linear isomorphism. Then
TX T M splits as a direct sum

TX T M = H ⊕ TX Tp M,

where both summands are isomorphic to Tp M , the first via dπ|p the second via the
canonical isomorphism. With respect to this direct sum decomposition, dFX has the form
 
dπ|p 0
dF |X = . (5.4)
∗ d expp |X
As by construction, the length of X is less than the injectivity radius of p, d expp |X has
full rank. We obtain that dF |X has full rank as well. By the inverse function theorem, F
must be a diffeomorphism onto its image when restricted to a neighborhood U ⊂ T M of
X; in particular F is injective on this neighborhood. If we now assume that X = Y , then

6
both (Xk )k∈N and (Yk )k∈N converge to X. For k ∈ N large enough, Xk and Yk are both
contained in U , but we have

F (Xk ) = (pk , exppk (Xk )) = (pk , exppk (Yk )) = F (Yk ),

a contradiction to the fact that F is injective on U . 

Definition 5.12 (Injectivity radius). We define the (global) injectivity radius of (M, g)
by
injrad(M, g) = inf injrad(p).
p∈M

We generally have injrad(M, g) ≥ 0, but the injectivity radius may be zero. However, we
have the following lemma.

Corollary 5.13. If the Riemannian manifold (M, g) is compact, then injrad(M, g) > 0.

Proof. We have injrad(p) > 0 for each p ∈ M . By Thm. 5.11, the injectivity radius is
continuous. Hence because M is compect, there exists p ∈ M such that

0 < injrad(p) = inf injrad(p) = injrad(M, g). 

p∈M

5.7 Closed geodesics

Definition 5.14. Let (M, g) be a Riemannian manifold. A closed geodesic in M is a
smooth map γ : S 1 → M such that γ satisfies the geodesic equation at every point of S 1 .

Theorem 5.15. Let (M, g) be a compact Riemannian manifold and let c0 : S 1 → M be

a piecewise C 1 curve. Then there exists a closed geodesic homotopic to c that minimizes
the energy among all piecewise C 1 curves homotopic to c0 .

For the proof of the theorem, we need the following lemma.

Lemma 5.16. Let (M, g) be a complete Riemannian manifold. Whenever two piecewise
C 1 loops c1 , c2 : S 1 → M satisfy d(c1 (t), c2 (t)) < injrad(c1 (t)) for all t ∈ S 1 , then c1 and
c2 are homotopic.

Proof. Let rt = injrad(c1 (t)). Then the exponential map expc1 (t) is injective on Brt (0) ⊂
Tc1 (t) M and since d(c1 (t), c2 (t)) < rt , there exists a unique X(t) ∈ Brt (0) ⊂ Tc1 (t) M such
that c2 (t) = expc1 (t) (X(t)).
We show that X is a piecewise C 1 vector field along c1 . To this end, consider the map F =
π × exp as in the proof of Thm. 5.11. by construction, X satisfies F (X(t)) = (c1 (t), c2 (t)).
By the assumption d(c1 (t), c2 (t)) < rt , the differential d expc1 (t) |X(t) is non-singular for
each t ∈ S 1 , so from (5.4), we get that also the differential dF |X(t) is non-singular for each

7
t ∈ S 1 . Hence F is a local diffeomorphism. Now if for some t ∈ S 1 , U ⊂ T M is an open
neighborhood of X(t) such that F is a diffeomorphism when restricted to U , we have

X(t) = (F |U )−1 (c1 (t), c2 (t)).

Since c1 and c2 are piecewise C 1 and F is smooth, we obtain that X is also piecewise C 1 .
We can now define a homotopy by

H(s, t) = F (sX(t)) = expc1 (t) (sX(t)),

which satisfes H(0, t) = c1 (t), H(1, t) = c2 (t). 

Proof (of Thm. 5.15). Set

E0 = inf{E[c] | c homotopic to c0 }

and let c1 , c2 , . . . be a sequence of piecewise C 1 loops homotopic to c0 such that E[ck ] → E0

as k → ∞. As reparametrization does not change the homotopy class, we may assume
that ci are parametrized proportionally to arc length. By the estimate L[ck ]2 ≤ 2E[ck ],
we obtain that the lengths L[ck ] are bounded uniformly over k. We think of the loops ck
as maps [0, 1] → M such that ck (0) = ck (1).
Pick r ∈ R satisfying 0 < r < injrad(M, g), which is possible by (5.13). Then since the
ci are parametrized by arc length and their lengths are uniformly bounded, there exists
some N ∈ N such that, setting tj = j/N , the points pkj := ck (tj ) (j = 0, . . . , N ), satisfy
r
d(pkj−1 , pkj ) ≤
3
for any j = 1, . . . , N and any k ∈ N (observe that pk0 = pkN since the ci are loops). We
obtain N sequences (pk1 )i∈N , . . . , (pkN )k∈N in M . Since with M also the N -fold product M N
is compact, after passing to a subsequence, we may achieve that for each j = 0, . . . , N ,
the sequence (pkj )i∈N converges to some point pj ∈ M , where p0 = pN .
By continuity, we also have
r
d(pj−1 , pj ) ≤ < injrad(M, g),
3
for each j, so there exists a unique minimizing geodesic between pj−1 and pj . Let c : S 1 →
M be the piecewise geodesic loop such that c(tj ) = pj and such that on the subintervals
[tj−1 , tj ] of [0, 1] (j = 1, . . . , N ), c is a the unique minimizing geodesic between pj−1 and
pj . Let k0 ∈ N be so large that d(pkj , pj ) ≤ r/3 for all k ≥ k0 . Then for any k ≥ k0 and
any t ∈ S 1 , we have
r r r
d(c(t), ck (t)) ≤ d(c(t), pj ) + d(pj , pkj ) + d(pkj , ck (t)) ≤ + + = r < injrad(M, g).
3 3 3

8
where j is such that t ∈ [tj−1 , tj ]. From Lemma 5.16, we obtain that c and ck are
homotopic, hence c is also homotopic to c0 . Moreover, for each j = 1, . . . , N and each
k ∈ N, we get

L[c|[tj−1,tj ] ] = d(pj−1 , pj ) ≤ d(pj−1 , pkj−1 ) + d(pkj−1 , pkj ) + d(pkj , pj )

≤ d(pj−1 , pkj−1 ) + L[ck |[tj−1 ,tj ] ] + d(pkj , pj ).

Taking the sum over j, we get

n
X
L[c] ≤ L[ck ] + 2 d(pkj , pj )
j=1
n
k→∞
p X p
≤ 2E[ck ] + 2 d(pkj , pj ) −→ 2E0 .
j=1

Let γ be the reparametrization of c proportionally to arc length, which is again homotopic

to c0 . Then since γ is parametrized proportionally to arc length, we have
1 1
E[γ] = L[γ]2 = L[c]2 ≤ E0 .
2 2
On the other hand, since E0 is is infimum of all energies of loops homotopic to c0 , we
must have E[γ] = E0 . Hence γ realizes the infinimum of the energy among all piecewise
C 1 curves homotopic to c0 . It follows that all variations (γs )s∈(−ε,ε) of γ must satisfy
d
| E[γs ] = 0, so it follows from Thm. 5.3.1 that γ is a geodesic (and in particular
ds s=0
smooth, not only piecewise smooth). 

5.8 Synge’s theorem

Let M be a manifold. For a point p ∈ M , denote by Frp (M ) the set of vector space bases
of Tp M . If n the dimension of the manifold, there is a right action of GL(n, R) on Frp (M ),
given by  
a11 . . . a1n n n
!
(E1 , . . . , En ) ·  ... ... ..  = X X
ai1 Ei , . . . , ain Ei

. 
an1 . . . ann i=1 i=1

which is free and transitive. Hence the choice of a basis in Tp M gives a bijection to
GL(n, R). Frp (M ) is then given the unique manifold structure making this bijection to
the open set GL(n, R) ⊂ Rn×n is a diffeomorphism (this is independent of the choice of
basis). The manifolds Frp (M ) fit together to the frame bundle
a
Fr(M ) = Frp (M ).
p∈M

There is then a unique manifold structure on Fr(M ) such that the footpoint projection
π : Fr(M ) → M is a surjective submersion (this is constructed similarly to the smooth

9
structure on T M ). In fact, this shows that Fr(M ) is even a fiber bundle with typical fiber
GL(n, R) (in fact, it is even a principal bundle).
GL(n, R) has two different connected components: The matrices with positive and nega-
tive determinant. Since Frp (M ) is diffeomorphic to GL(n, R) for every p ∈ M , this shows
that also each Frp (M ) has two connected components. We then have the following lemma.
Lemma 5.17. If M is connected, then Fr(M ) has at most two connected components.

Proof. This follows at once from the long exact sequence for homotopy groups, applied
to the fibration
GL(n, R) −→ Fr(M ) −→ M.
Here we have that



· · · −→ π1 (M ) −→ π0 GL(n, R) −→ π0 Fr(M ) −→ π0 (M )
| {z } | {z }
Z2 =0

is a short exact sequence of pointed sets. Hence since M is connected, π0 (GL(n, R)) ∼ = Z2
surjects onto π0 (Fr(M )).
An explicit argument is the following. Fix a base point E0 ∈ Fr(M ). Denote by p0 the
foot point of E0 . Because M is connected (hence path-connected, as M is a manifold),
given any other frame E ∈ Fr(M ), say with foot point p, there exists a continuous path
c : [0, 1] → M with c(0) = p0 and c(1) = p. Now, because Fr(M ) is a fiber bundle, it has
the path-lifting property, which implies that we can lift c to a path c : [0, 1] → Fr(M ),
satisfying bc(0) = E0 and π(c(t)) = c(t) for all t. In particular, c(1) and E are two points
in the same fiber Frp0 (M ). If these two points lie in the same component of Frp0 (M ), then
we connect them by a path in Frp (M ) and, by concatenation, obtain a path connecting
E0 and E.
Now, if E and E 0 are two points in Fr(M ), we can connect the foot points of both with
p0 through paths c and c0 , and lift these to Fr(M ), obtaining two elements c(1), c0 (1) ∈
Frp0 (M ). Now since Frp0 (M ) has two connected components, at least two of the three
elements E0 , c(1), c0 (1) must lie in the same connected component. This shows that given
two frames E and E 0 , either one of them lies in the same path component of E0 or they
both lie in the same path component. Hence π0 (Fr(M )) has at most two elements. 

Definition 5.18 (Orientability). A connected manifold M is orientable if the frame

bundle Fr(M ) has two different connected components. An orientation of an orientable
connected manifold M is a choice of one of these connected components. It is denoted by
Fr+ (M ) and its elements are called positively oriented bases.
A (possibly non-connected) manifold is called orientable if each connected component is
orientable, and an orientation of such a manifold is the choice of an orientation for each
connected component.

For an oriented manifold M , the GL(n, R)-action on Fr(M ) restricts to an action of

GL+ +
n (R) (the group of matrices with positive determinant) on Fr (M ).

10
Lemma 5.19. Let (M, g) be a Riemannian manifold. For a piecewise smooth loop c :
S 1 → M , denote by Pc the parallel transport around c.
(a) If M is oriented, then every piecewise smooth loop c : S 1 → M has det(Pc ) = 1.

(b) If M is not orientable, then there exists a piecewise smooth loop c : S 1 → M such
that det(Pc ) = −1.

Proof. We think of c as a map c : [0, 1] → M such that c(0) = c(1). We know that
Pc : Tc(0) M → Tc(1) M is an orthogonal transformation. Hence det(Pc ) = ±1.
(a) Suppose that M is orientable. Let E = (E1 , . . . , En ) ∈ Fr+ c(0) (M ) be a positively
oriented basis of of Tc(0) M and let E(t) = (E1 (t), . . . , En (t)) ∈ Frc(t) (M ) be the corre-
sponding parallel transported basis along c. Then E(t) is a continuous path in Fr+ (M ),
hence E(t) must be positively oriented for every t ∈ [0, 1]. In particular, for t = 1, we
have E(1) = (Pc E1 , . . . , Pc En ) ∈ Fr+
c(1) , so

(Pc E1 , . . . , Pc En ) = (E1 , . . . , En ) · A

for some A ∈ GL+ (R). It follows from the definition of the action that A is just a matrix
representation of Pc with respect to the basis E. Hence

det(Pc ) = det(A) > 0.

(b) Suppose that M is not orientable. Then Fr(M ) is connected (hence path-connected).
So given p ∈ M and any two bases E, E 0 ∈ Frp (M ), there exists a smooth path c in Fr(M )
with c(0) = E and c(1) = E 0 . Since E and E 0 lie in the same fiber, its foot point curve
c is a closed curve. Let E(t) be the frame obtained by parallel transport of E along c.
This gives another smooth path in Fr(M ), which lies in the same fiber as c(t) for every
t ∈ [0, 1]. Hence for each t ∈ [0, 1], there exists a unique A(t) ∈ GL(n, R) such that
E(t) = c(t) · A(t). Since both E(t) and c(t) are smooth in t, the matrix A(t) must depend
smoothly on t also. Since E(0) = E = c(0), we have A(0) = id, so det(A(t)) > 0 for each
t ∈ [0, 1]. We obtain that

(Pc E1 , . . . , Pc (En )) = E(1) = c(1) · A(1) = E 0 · A(1).

Now if E 0 does not lie in the same connected component as E, we have E 0 = E · B for a
matrix B with det(B) < 0. We therefore have

det(Pc ) = det(BA(1)) = det(B) det(A(1)) < 0. 

Lemma 5.20. Let (M, g) be a Riemannian manifold and let c, c0 : S 1 → M be two

piecewise smooth loops that are homotopic. Then the parallel transports Pc and Pc0 around
c, respectively c0 , satisfy det(Pc ) = det(Pc0 ).

Remark 5.21. The point is here that the homotopy is only required to be continuous.

11
Proof. Let c : S 1 → M be a loop (which we think of as path c : [0, 1] → M with
c(0) = c(1)) and let E ∈ Frc(0) (M ) be a basis of Tc(0) M . Let moreover E(t) be the parallel
transport of E around c. As seen in the proof of Lemma 5.19, we have det(Pc ) = 1 if
and only if E(1) lies in the same connected component as E. Let now c0 be another loop
and let H be a homotopy between c and c0 . Since Fr(M ) is a fiber bundle, we can lift
H to a map H : [0, 1] × [0, 1] → Fr(M ) such that H(0, t) = E(t) for all t ∈ [0, 1]. Write
E 0 = H(1, 0) and let E 0 (t) be the parallel transport of E 0 along c0 . Then as in the proof of
Lemma 5.19, E 0 (t) = H(1, t) · A(t) for a smooth map A : [0, 1] → GL(n, R). Now, if E(1)
lies in the same connected component as E if and only if H(s, 1) and H(s, 0) lie in the
same component of FrH(s,0) (M ) for every s ∈ [0, 1], if and only if E 0 (1) lies in the same
connected component as E 0 . 

Theorem 5.22 (Synge). Let (M, g) be a compact Riemannian manifold with positive
sectional curvature.

(a) If M is even-dimensional and oriented, then it is simply connected.

(b) If M is odd-dimensional, then it is orientable.

The proof is based on the following.

Linear Algebra Lemma 5.23. Let V be an oriented Euclidean vector space and let P
be an orthogonal transformation of V . If V is odd-dimensional, assume that det(P ) = 1.
If V is even-dimensional, assume that det(P ) = −1. Then P fixes a one-dimensional
subspace.

Proof. As an orthogonal transformation, the eigenvalues of P all lie on the complex unit
circle. Since P is real, the spectrum is symmetric with respect to the real axis, hence if λ
is an eigenvalue, then so is λ. Let E ⊂ V be the direct sum of eigenspaces to eigenvalues
with non-zero imaginary part. Then E is an even-dimensional invariant subspace, since
all eigenvalues of P |E come in complex conjugate pairs. As λλ = 1, we have det(P |E ) = 1.
On the orthogonal complement V 0 of E, P has the eigenvalues +1 and −1. We have to
show that e = dim{X | P X = X} is not zero. We have
0 )−e
det(P ) = det(P |V 0 ) det(P |E ) = det(P |V 0 ) = (−1)dim(V = (−1)dim(V )−e .

since V 0 has the same parity as V . Now observe that the assumptions imply that e must
always be odd, in particular not zero. 

Proof (of Thm. 5.22). (a) Assume that M is even-dimensional and suppose that c : S 1 →
M defines an non-trivial element of π1 (M ). By Thm. 5.15, there exists a closed geodesic
γ homotopic to c that minimizes the energy in its homotopy class. We may parametrize
γ by arc length. The parallel transport Pγ around γ satisfies

Pγ γ̇(0) = γ̇(0),

12
hence it preserves the orthogonal complement V = γ̇(0)⊥ ⊂ Tγ(0) M . Since M is even-
dimensional, V is odd-dimensional. Hence Pγ |V is orthogonal transformation of an odd-
dimensional vector space. By Lemma 5.19(a), we have det(Pγ ) = 1. By Lemma 5.23,
Pγ |V fixes a line, hence there exists X ⊥ γ̇(0) with kXk = 1 and Pγ X = X. Let X(t)
be the parallel vector field around γ corresponding to this eigenvector (the point is here
that X closes up continuously after going around γ). Let (γs )s∈(−ε,ε) be a variation with
variational vector field X. Then since X is parallel, we get
ˆ 1 2
!
∂2 ∇

E[γs ] = X(t) − R X(t), γ̇(t) γ̇(t), X(t) dt
∂s2 0 dt
ˆ 1


=− R X(t), γ̇(t) γ̇(t), X(t) dt
ˆ0 1


=− K span{X(t), γ̇(t)} dt < 0.
0

Here we use that as X ⊥ γ̇(0), we also have X(t) ⊥ γ̇(t) for
all other t, hence X(t) and
γ̇(t) span a plane in Tγ(t) M and we can identify R(X(t), γ̇(t) γ̇(t), X(t) with the sectional
curvature of this plane. We obtain that for |s| small, γs is a curve homotopic to γ such
that E[γs ] < E[γ], a contradiction to the assumption that γ is energy minimizing in its
homotopy class. Hence every element of π1 (M ) must be trivial.
(b) Suppose M is not orientable. Then by Lemma 5.19(b), there exists a piecewise smooth
loop c : S 1 → M such that det(Pc ) = −1. c must define a non-trivial element of π1 (M ),
because if c is homotopic to a constant loop, then by Lemma 5.20, we have det(Pc ) =
det(id) = 1 (using that parallel transport around constant loops is the identity). Let now
γ be an energy-minimizing geodesic homotopic to c (Thm. 5.15). By Lemma 5.20, we also
have det(Pγ ) = −1. Since M is odd-dimensional, the orthogonal complement V = γ̇(0)⊥
is an even-dimensional invariant subspace for Pγ , and we still have det(Pγ |V ) = −1. Hence
by Lemma 5.23, Pγ |V fixes a line. As before, we derive a contradiction to the fact that γ
minimizes the energy in its homotopy class. Hence there cannot be a piecewise smooth
loop c : S 1 → M with det(Pc ) = −1, so M must be orientable. 
Remark 5.24. The necessity of the assumptions of Synge’s theorem can be seen via the
following examples.
(a) The odd-dimensional spheres can be realized as S 2n−1 ⊂ Cn . For any p ∈ N and
q1 , . . . , qn coprime integers to p, there is an action of the cyclic group Z/pZ on S 2n−1 ,
where the generator of Z/pZ acts via
(z1 , . . . , zn ) 7−→ (e2πq1 /p z1 , . . . , e2πqn /p zn ).
This is a free and orientation preserving action by isometries, hence taking the quo-
tient by this action, we obtain a new Riemannian manifold, the so-called lens space
L(p; q1 , . . . , qn ). The corresponding Riemannian manifold still has constant positive
sectional curvature and non-trivial fundamental group. This shows that the manifold
in Thm. 5.22 must indeed be even-dimensional for the conclusion of (a) to hold.

13
(b) The real projective spaces RP n are non-orientable if n is even, but have positive
sectional curvature. This shows that the conclusion (b) of Synge’s theorem needs the
assumption that the dimensional of M is odd.

6 Lie groups and homogeneous spaces

6.1 Lie groups and their Lie algebras
Definition 6.1. A Lie group is a group G together with a manifold structure on G
such that the group multiplication m : G × G → G and the inversion map i : G → G are
smooth. A homomorphism between Lie groups G and H is a smooth group homomorphism
ϕ : G → H.

Example 6.2. Any countable group is a Lie group when endowed with the discrete topol-
ogy. (It cannot be uncountable because then as a topological space with the discrete
topology, it would not be second-countable.)

Example 6.3. Rn with its usual smooth structure becomes a Lie group with respect to
addition of vectors. Its quotient T n = Rn /Zn , the n-dimensional torus, is also a Lie group.
In fact, every connected abelian Lie group is of the form T n × Rm , for numbers n, m ∈ N0 .

Example 6.4. It is a general (non-trivial) fact that the isometry group Isom(M, g) of a
semi-Riemannian manifold (M, g) is a Lie group.

Example 6.5. If V is a real or complex or quaternionic vector space, then the general
linear group GL(V ) is a Lie group with the manifold structure coming from viewing it as
an open subset of End(V ). For any n ∈ N, we thus obtain Lie groups GL(n, R), GL(n, C)
and GL(n, H). There are canonical inclusions GL(n, C) ⊂ GL(2n, R) and GL(n, H) ⊂
GL(2n, C) ⊂ GL(4n, R), which for n = 1 are given by
   
a b z w
a + bi 7−→ , z + wj 7−→ .
−b a −w z

The most important theorem for generating examples of Lie groups is the following. A
proof can be found in any textbook on Lie groups.
Theorem 6.6. If G is a Lie group and H ⊂ G is a closed subgroup, then it is in fact a
submanifold and a Lie group with the induced smooth structure.

Example 6.7. It follows that any closed subgroup G ⊂ GL(V ) is a Lie group. We have
the following concrete examples.
(a) The special linear group SL(V ) = det−1 (1) ⊂ GL(V ). We also use the notations
SL(n, R), SL(n, C) and SL(n, H). In the quaternionic case, one has to take the real
determinant, coming from the inclusion GL(n, H) ⊂ GL(4n, R).

14
(b) The orthogonal group O(n) = {A ∈ GL(n, R) | AAt = id}.

(c) The special orthogonal group SO(n) = O(n) ∩ SL(n, R).

t
(d) The unitary group U(n) = {A ∈ GL(n, C) | AA∗ = id}. Here A∗ = A denotes the
transpose, followed by complex conjugation.

(e) The special unitary group SU(n) = U(n) ∩ SL(n, C).

(f) The unitary groups are not to be confused with the complex orthogonal groups O(n, C) =
{A ∈ GL(n, C) | AAt = id} and SO(n, C) = O(n, C) ∩ SL(n, C). U(n) and SU(n) are
compact, while these are not.

(g) The (compact) symplectic group Sp(n) = {A ∈ GL(n, H) | AA† = id}. Here A†
denotes transpose followed by quaternionic conjugation.

Given a Lie group G, we write

λg (h) = gh, ρg (h) = hg,

for the actions of G on itself by left and right multiplication. It follows from the axioms
of a Lie group, that λg and ρg are diffeomorphisms for every g ∈ G.
Definition 6.8 (Left-invariant vector fields). A vector field X̃ on a Lie group G is
called left invariant if for all g ∈ G, we have

dλg (X̃) = X̃ ◦ λg .

Lemma 6.9. For two left-invariant vector fields X̃, Ỹ on a Lie group G, the Lie bracket
[X̃, Ỹ ] is again left invariant.

Proof. Observe here that left-invariance of a vector field X̃ on G can be reformulated

as (λg )∗ X̃ = X̃. Therefore the lemma follows from the fact that the Lie bracket of
vector fields is natural, in the sense that for any diffeomorphism f : M → N between
manifolds and any two vector fields X̃, Ỹ on M , we have f∗ [X̃, Ỹ ] = [f∗ X̃, f∗ Ỹ ], where
f∗ X̃ = df (X̃) ◦ f −1 is the pushforward of X by f . 

For any manifold M , the Lie bracket on vector fields satisfies the so-called Jacobi identity

[[X̃, Ỹ ], Z̃] + [[Ỹ , Z̃], X̃] + [[Z̃, X̃], Ỹ ] = 0. (6.1)

Abstractly, a vectorspace V together with a skew-symmetric bracket [·, ·] : V × V → V
satisfying (6.1) is called a Lie algebra. By Lemma 6.9, the space of left invariant vector
fields Xλ (G) is a Lie subalgebra of the Lie algebra X (G) of all vector fields on G.
There is an alternative description of the Lie algebra Xλ (G) of left invariant vector fields
for a Lie group G that we discuss now.

15
Lemma 6.10. Evaluation at e ∈ G gives a vector space isomorphism between the space
of left invariant vector fields on G and Te G.

Proof. For any tangent vector X ∈ Te G, there is a left invariant vector field X̃ with
X̃|e = X, which is given by
X̃|g = dλg (X).
Conversely, every left invariant vector field X̃ is determined by its value at e ∈ G, as
X̃|g = (X̃ ◦ λg )|e = dλg (X̃|e ). 

From now on, we set

g := Te G,
which by the above Lemma is isomorphic to the Lie algebra of left invariant vector fields
on G. To obtain a description of the Lie bracket of Xλ (G) in terms of g, consider the
conjugation action
α : G × G −→ G, (g, h) 7−→ ghg −1
of G on itself. Differentiating α with respect to the second variable at the unit element
e ∈ G, we obtain an action
Ad : G × g −→ g.
of G on g. This action is called the adjoint action of G on g.

Example 6.11. If G ⊂ GL(V ) is a closed subgroup, the adjoint action is just given by
Adg (X) = gXg −1 . The tangent space at an arbitrary g ∈ G is related to g by

Tg G = {Xg | X ∈ Te G} = {gX | X ∈ Te G}.

The equality of the two descriptions above follows from the invariant of g under the adjoint
action.

Further differentiating Ad with respect to the first variable, we obtain a map

ad : g × g −→ g.

Lemma 6.12. Let X, Y ∈ g and let X̃ and Ỹ be the corresponding left invariant vector
fields. Then
adX (Y ) = [X̃, Ỹ ]|e .

Proof. We show that adX (Y ) and [X̃, Ỹ ]|e coincide as derivations on g = Te G. To this
end, let ξ, η : (−ε, ε) → G be smooth curves with ξ(0) = η(0) = e and ξ(0) ˙ = X,
η̇(0) = Y . Then
∂
adX (Y ) = Adξ(t) (Y ).
∂t t=0

16
On the other hand, σ(s) = ξ(t)η(s)ξ(t)−1 is a curve such that σ̇(0) = Adξ(t) (Y ). Hence
for f ∈ C ∞ (G), we have
∂

∂Adξ(t) (Y ) f (e) = f αξ(t) (η(s)) .
∂s s=0

We therefore get
∂ ∂2

∂adX (Y ) f (e) = ∂Adξ(t) (Y ) f (e) = f αξ(t) (η(s))
∂t t=0 ∂t∂s t=s=0

Exchanging the differentiation variables and using the chain rule, we get
∂2 ∂2
f ξ(t)η(s)ξ(t)−1



f αξ(t) (η(s)) =
∂t∂s t=s=0 ∂s∂t s=t=0
∂2 ∂2
f η(s)ξ(t)−1 .



= f ξ(t)η(s) +
|∂s∂t s=t=0
{z } |∂s∂t s=t=0
{z }
1 2
We have
∂ ∂

ξ(t)η(s) = λξ(t) η(s) = dλξ(t) |e (Y |e ) = Ỹ |ξ(t) ,
∂s s=0 ∂s s=0
hence swapping derivatives again, we get
∂2
∂
1 = f ξ(t)η(s) = ∂Ỹ f (ξ(t)) = ∂X̃ ∂Ỹ f (e)
∂t∂s t=s=0 ∂t t=0

To deal with the term 2 , we observe that differentiating the constant curve t 7→ ξ(t)ξ(t)−1
using the chain rule, we obtain that t 7→ ξ(0)−1 represents the tangent vector −ξ(0) ˙ = −X.
Hence
∂ ∂
η(s)ξ(t)−1 = λη(s) ξ(t)−1 = −dλη(s) |e (X|e ) = −X̃|η(s) ,


∂t t=0 ∂t t=0
and
∂

2 =− ∂X̃ f (η(s) = −∂Ỹ ∂X̃ f (e).
∂s s=0
Putting everything together, we obtain
∂adX (Y ) f (e) = ∂X̃ ∂Ỹ f (e) − ∂Ỹ ∂X̃ f (e) = ∂[X̃,Ỹ ] f (e),
as claimed. 

Definition 6.13 (Lie algebra). Let G be a Lie group. The Lie algebra of G is its
tangent space at the identity element g = Te G, together with the Lie bracket [X, Y ] =
adX (Y ).

Lie groups are typically denoted by capital roman letters such as G, H, K, while the
corresponding Lie algebras are denoted by the corresponding fraktur letters g, h, k. If ϕ :
G → H is a Lie group homomorphism, its differential induces a Lie algebra homomorphism
ϕ∗ = dϕ|e : g → h.

17
Example 6.14. The Lie bracket of the Lie algebra of an abelian Lie group is zero.

Example 6.15. The Lie algebra of GL(V ) is the space gl(V ) = End(V ) of vector space
endomorphisms of V , with the commutator Lie bracket

[X, Y ] = XY − Y X.

Matrix Lie groups from Example 6.7 are the following, all with the bracket induced by
the one of gl(V ):
(a) sl(V ) = {X ∈ gl(V ) | tr(X) = 0}.

(b) o(n) = so(n) = {X ∈ gl(n, R) | X t = −X}.

(c) u(n) = {X ∈ gl(n, C) | X ∗ = −X}.

(d) su(n) = {X ∈ gl(n, C) | X ∗ = −X, tr(X) = 0}.

(e) o(n, C) = so(n, C) = {X ∈ gl(n, C) | X t = −X}.

(f) sp(n) = {X ∈ gl(n, H) | X † = −X}.

It is an exercise to check that the commutator bracket indeed restricts to each of these
subspaces.

6.2 Differential geometry of Lie groups

Definition 6.16 (Invariant metrics). Let G be a Lie group. A semi-Riemannian met-
ric β on G is called left invariant, if λ∗g β = β for all h ∈ G, and bi-invariant if λ∗g β =
ρ∗g β = β for all h ∈ G.

Lemma 6.17. Let G be a Lie group with Lie algebra g.

(a) Restriction at the unit element e provides an isomorphism between the space of left
invariant semi-Riemannian metrics on G and the space of non-degenerate inner prod-
ucts on g (of the same signature).

(b) This isomorphism further refines to one between the space of bi-invariant semi-Riemannian
metrics on G and the space of Ad-invariant, non-degenerate inner products on g.

Proof. (a) Any left-invariant semi-Riemannian metric β is determined by β0 := β|e be-

cause of the formula


β|g (X|g , Y |g ) = β0 dλg−1 |g (X|g ), dλg−1 |g (Y |g ) , X, Y ∈ Tg G. (6.2)

This shows that the restriction map is injective. On the other hand, given a non-
degenerate inner product b on g = Te G, it defines a semi-Riemannian metric on G by
replacing β|e by b in the above formula. Therefore, the restriction map is surjective.

18
(b) Let β0 be an Ad-invariant, non-degenerate inner product on g and let β be the cor-
responding left-invariant metric given by (6.2). Then for vector fields X, Y ∈ Tg G, we
have

β|g (X, Y ) = β0 dλg−1 |g (X), dλg−1 |g (Y )


= β0 Adh−1 dλg−1 (X), Adh−1 dλg−1 (Y )


= β0 dαh−1 |e dλg−1 |g (X), dαh−1 |e dλg−1 |g (Y )


= β0 dλh−1 |h dρh |e dλg−1 (X), dλh−1 |h dρh |e dλg−1 (Y )


= β0 dλ(gh)−1 |gh dρh |g (X), dλ(gh)−1 |gh dρh |g (Y )


= βgh dρh |g (X), dρh |g (Y ) ,
where we used that the left and right actions commute, together with the product rule.
So β is also right invariant, hence bi-invariant. Conversely, if β is bi-invariant, then we
have for β0 := β|e that



β0 Adg (X), Adg (Y ) = g|e dαg |e (X), dαg |e (Y )


= β|e dλg |g−1 dρg |e (X), dλg |g−1 dρg |e (Y )


= β|g−1 dρg |e (X), dρg |e (Y )


= β|e X, Y
= β0 (X, Y )

for all X, Y ∈ g. 

Theorem 6.18. Every compact Lie group admits a bi-invariant Riemannian metric.

Proof. We use that any compact Lie group has a finite right invariant measure, i.e., a
measure µ with µ(G) < ∞ and such that
ˆ ˆ
∗
ρg f (h)dµ(h) = f (h)dµ(h)
G G

for all f ∈ C ∞ (G) and all g ∈ G (this can be obtained, for example, by extending a
density at e ∈ G by right invariance to all of G). Now given any positive definite inner
product β0 on G, we define
ˆ


β(X, Y ) = β0 Adh (X), Adh (Y ) dµ(h), X, Y ∈ g
G

Then ˆ



β Adg (X), Adg (Y ) = β0 Adhg (X), Adhg (Y ) dµ(h),
G

which coincides with β(X, Y ) by right invariance of µ. 

19
Example 6.19. Ad-invariant, non-degenerate inner products on gl(n, R) and gl(n, C) are
given by

β(X, Y ) = −tr(XY ), respectively β(X, Y ) = −Re tr(XY ),

inducing a semi-Riemannian metric on GL(n, R) and GL(n, C). These inner products
are positive definite on so(n) ⊂ gl(n, R) and u(n) ⊂ gl(n, C), hence induce bi-invariant
Riemannian metrics O(n), SO(n) and U(n).

Lemma 6.20. Let G be a Lie group with Lie algebra g. Then any Ad-invariant, non-
degenerate inner product β on g satisfies

β([X, Y ], Z) = β(X, [Y, Z]), X, Y, Z ∈ g.

Proof. Let ξ : (−ε, ε) → G be a path with ξ(0)˙ = X. Then for Y, Z ∈ g, the number
∗
Adξ(t) β(Y, Z) is independent of t. Hence, using the chain rule and Lemma 6.12, we get

d

0= β Adξ(t) (Y ), Adξ(t) (Z)
dtt=0   
d d
=β Adξ(t) (Y ), Z + β Y, Adξ(t) (Z)
dt t=0 dt t=0



= β adX (Y ), Z + β Y, adX (Z)



= β [X, Y ], Z + β Y, [X, Z] ,

which implies the claim. 

Lemma 6.21. Let G be a Lie group and let β be a bi-invariant semi-Riemannian metric
on G, with Levi-Civita connection ∇. Then for left invariant vector fields X̃, Ỹ , we have
1
∇X̃ Ỹ = [X̃, Ỹ ].
2

Proof. By the Koszul formula, we have

2β(∇X Y, Z) = ∂X β(Y, Z) + ∂Y β(X, Z) − ∂Z β(X, Y )

(6.3)
+ β([X, Y ], Z) − β([X, Z], Y ) − β([Y, Z], X)

For all vector fields X, Y, Z on G. If X̃, Ỹ ∈ Xλ (G) are left invariant vector fields corre-
sponding to X, Y ∈ g, then

β|g (X̃|g , Ỹ |g ) = β|g (dλg |e X, dλg |e Y ) = β|e (X, Y ),

20
so β|g (X̃|g , Ỹ |g is a constant function of g. We obtain that if X̃, Ỹ , Z̃ are left invariant
vector fields, then the first three terms of (6.3) vanish. We therefore get

2β(∇X̃ Ỹ , Z̃) = β([X̃, Ỹ ], Z̃) − β([X̃, Z̃], Ỹ ) − β([Ỹ , Z̃], X̃)

= β([X̃, Ỹ ], Z̃) + β([Z̃, X̃], Ỹ ) + β([Z̃, Ỹ ], X̃)
= β([X̃, Ỹ ], Z̃) + β(Z̃, [X̃, Ỹ ]) + β(Z̃, [Ỹ , X̃])
= β([X̃, Ỹ ], Z̃).

where we used Lemma 6.20. Since β is non-degenerate and each tangent space is spanned
by the values of left invariant vector fields, this implies the result. 

Theorem 6.22. Let G be a Lie group with a bi-invariant semi-Riemannian metric. Then
for left invariant vector fields X̃, Ỹ , Z̃, we have
1
R(X̃, Ỹ )Z̃ = − [[X̃, Ỹ ], Z̃].
4
Moreover, we have ∇R = 0, hence G is a locally symmetric space.

Proof. We calculate

R(X̃, Ỹ )Z̃ = ∇X̃ ∇Ỹ Z̃ − ∇Ỹ ∇X̃ Z̃ − ∇[X̃,Ỹ ] Z̃

1 1 1
= [X̃, [Ỹ , Z̃]] − [Ỹ , [X̃, Z̃]] − [[X̃, Ỹ ], Z̃]
4 4 2
1 1 1 1
= [X̃, [Ỹ , Z̃]] + [Ỹ , [Z̃, X̃]] + [Z̃, [X̃, Ỹ ]] − [[X̃, Ỹ ], Z̃].
4 4 4 4
The first three terms vanish by the Jacobi identity (6.1).
To see that ∇R = 0, calculate

(∇X̃ R)(Ỹ , Z̃)W̃ = ∇X̃ R(Ỹ , Z̃)W̃ − R(∇X̃ Ỹ , Z̃)W̃ − R(Ỹ , ∇X̃ Z̃)W̃ − R(Ỹ , Z̃)∇X̃ W̃
1 
= −[X̃, [[Ỹ , Z̃], W̃ ]] + [[[X̃, Ỹ ], Z̃], W̃ ] + [[Ỹ , [X̃, Z̃]], W̃ ] +[[Ỹ , Z̃], [X̃, W̃ ]]
8 | {z }
=−[[[Ỹ ,Z̃],X̃],W̃ ]
1 
= − [[W̃ , [Ỹ , Z̃]], X̃] + [[[Ỹ , Z̃], X̃], W̃ ] + [[X̃, W̃ ],[Ỹ , Z̃]] = 0,
8
by the Jacobi identity for the three vector fields X̃, W̃ and [Ỹ , Z̃]. 

Corollary 6.23. Let β be a bi-invariant semi-Riemannian metric on a Lie group G.

Then the sectional curvature of a non-degenerate 2-plane E ⊂ g spanned by X, Y ∈ g is
given by
1 β([X, Y ], [X, Y ])
K(E) = .
4 β(X, X)β(Y, Y ) − β(X, Y )2
In particular, a Riemannian (i.e., positive definite) bi-invariant metric on a Lie group
has non-negative sectional curvature.

21
Proof. By Lemma 6.20, we have
K(E) β(X, X)β(Y, Y ) − β(X, Y )2 = β R(X, Y )Y, X

1
= − β([[X, Y ], Y ], X)
4
1 
= − β([X, Y ], [Y, X])
4
1
= β([X, Y ], [X, Y ]).
4

6.3 The exponential map

Theorem 6.24. Let G be a Lie group with Lie algebra g. For each X ∈ g, there exists a
unique Lie group homomorphism γ : R → G such that γ̇(0) = X.
Proof. Let X̃ be the left-invariant vector field corresponding to X.
We start with showing uniqueness. Let γ : R → G be a group homomorphism with
γ̇(0) = X. Then differentiating the equation γ(t + s) = γ(t)γ(s) at s = 0 shows that γ
satisfies the ordinary differential equation
γ̇(t) = dλγ(t) |e γ̇(0) = X̃|γ(t) . (6.4)
By uniqueness of solutions to ordinary differential equations, there is at most one solution
to this equation with γ(0) = e.
To show existence, let γ : (−a, ε) → G be the solution to the ordinary differential equation
(6.4) with initial condition γ(0) = e, where a > 0 is chosen maximal. Suppose that a < ∞.
Then for 0 < s < ε, set γ̃(t) = γ(s)−1 γ(t + s) = λγ(s)−1 (γ(t + s)). Then γ̃(0) = e and
since X̃ is left invariant,
˙
γ̃(t) = dλγ(s)−1 |γ(t+s) (X̃|γ(t+s) ) = X̃|γ(s)−1 γ(t+s) = X̃|γ̃(t) .
So γ̃ satisfies the same ordinary differential equation (6.4) as γ, with the same initial
condition. However, γ̃ is defined on the interval (−a − s, ε − s), a contradiction to the
assumption that a is maximal. We obtain that γ is defined for all negative times. Similarly,
one shows that γ is defined for all positive times.
We now show that γ is a group homomorphism. Observe that for each s ∈ R, the path
γ̃(t) := γ(t + s) satisfies (6.4) with the initial condition γ̃(0) = γ(s). But the path
˜
γ̃(t) := γ(t)γ(s) also satisfies (6.4) with the same initial condition. Hence γ̃˜ = γ̃, so
γ(t + s) = γ(t)γ(s) for all s, t ∈ R. 
By the above lemma, we can make the following definition.
Definition 6.25 (Exponential map). Let G be a Lie group with Lie algebra g. The
exponential map of G is the map
exp : g −→ G
with the property that exp(X) = γ(1), where γ : R → G is the unique Lie group
homomorphism with γ̇(0) = X.

22
Example 6.26. If G ⊂ GL(V ) is a closed subgroup, then the exponential map of G is
given by the usual matrix exponential,
∞
X gn
exp(g) =
n=0
n!

Theorem 6.27. Let G be a Lie group and let β be a bi-invariant semi-Riemannian metric
on G. Then each Lie group homomorphism γ : R → G such that γ̇(0) = X is a geodesic
for β. In particular, the Lie group exponential map coincides with the semi-Riemannian
exponential map at the unit element.

Proof. Let γ(t) = exp(tX) and let X̃ be the left invariant vector field corresponding to
X. By definition, γ is an integral curve to X̃, so the result follows from the fact that
∇ 1
γ̇(t) = ∇X̃ X̃|γ(t) = [X̃, X̃]|γ(t) = 0.
dt 2
Here we used Lemma 6.21 and the fact that the Lie bracket is skew-symmetric. 

6.4 Symmetric spaces

Definition 6.28 (Symmetric space). Let (M, g) be a connected semi-Riemannian
manifold. A (global) geodesic reflection at p ∈ M is an isometry σp : M → M such that
σp (p) = p and such that dσp |p = −idTp M . (M, g) is called symmetric space if for every
p ∈ M , there exists a global geodesic reflection at p.

Symmetric spaces are closely related to homogeneous spaces. Recall that an action of a
Lie group G on a manifold M is a smooth map G × M → M , (g, p) 7→ g · p, satisfying

g · (h · p) = gh · p and e · p = p, g, h ∈ G, p ∈ M.

Such an action is called transitive if for any two points p, q ∈ M , there exists g ∈ G such
that g·p = q. It is called free if the map G×M → M ×M is injective, and it is called proper
if the same map is proper (i.e., inverse images of compact sets are compact). The quotient
manifold theorem states that if a Lie group G acts freely and properly on a manifold M ,
then the orbit space M/G is a topological manifold of dimension dim(M ) − dim(G) and
has a unique smooth structure turning the projection map M → M/G into a surjective
submersion.
Definition 6.29 (Homogeneous space). A homogeneous space is a manifold M to-
gether with a transitive action of a Lie group G.

If G is a Lie group and K is a closed subgroup of G, then the action of K on G by

multiplication is always free and proper, so the quotient space G/K is a smooth manifold
with a free and transitive action of G.

23
Conversely, given a homogenous space M , we can pick a point p ∈ M , and denote the
corresponding stabilizer subgroup by

K := Kp := {g ∈ G | g · p = p} ⊆ G.

Then the smooth map G → M , g 7→ g · p descends to a diffeomorphism G/K → M .

Theorem 6.30. Every symmetric space is homogenous.

Proof. Let (M, g) be a semi-Riemannian symmetric space. Then

G := Isom(M, g).

is a Lie group, see Example 6.4. We claim that G acts transitively on M . To see this,
we use that, given p, q ∈ M , there exists a broken geodesic connecting p and q, i.e., a
continuous path c : [0, r] → M together with a time partition 0 = t0 < t1 < · · · < tN = r
such that c is a geodesic on each of the subintervals [tj−1 , tj ] (in the case that M is
Riemannian and complete, p and q can even be connected by a geodesic, but we do not
want to make this assumption). For j = 1, . . . , N , let σj ∈ G be the geodesic reflection at
t +t
c( j−12 j ). Then σj exchanges c(tj−1 ) and c(tj ), hence the composition σN ◦ · · · ◦ σ1 ∈ G
sends p to q. 

Fundamental Example 6.31. Let G be a Lie group and let σ : G → G be an involutive

automorphism of G, i.e., a smooth map such that σ(gh) = σ(g)σ(h) and σ(σ(g)) = g.
Let
K ⊆ Fix(σ)
be an open subgroup of the fixed point set of σ (i.e., a disjoint union of connected com-
ponents). Moreover, define τ : G → G by τ (g) = σ(g −1 ) and let

M = Fix(τ )0

be the identity component of the fixed point set of τ .

A G-action on M is defined by the formula

g • p = gpσ(g)−1 .

There is a smooth map π : G → M defined by π(g) = gσ(g)−1 . One can check that
the image of this map is open and closed in M , hence all of M since M is connected.
As σ fixes K, we have π(g) = π(gk) for k ∈ K, hence π descends to a surjective map
G/K → M . It is injective if and only if K = Fix(σ). Indeed, if

gσ(g)−1 = hσ(h)−1 ⇐⇒ gh−1 = σ(gh−1 ),

hence gh−1 ∈ Fix(σ). Conversely, if g ∈ Fix(σ) \ K, then π(g) = π(e), so the quotient
map is not injective. We obtain that if K = Fix(σ), then π descends to a diffeomorphism

24
between M and the homogeneous space G/K, while in general, it descends to a covering
map with fiber Fix(σ)/K.
Assume that G carries a bi-invariant semi-Riemannian metric β for which σ (and hence
also τ ) are isometries, and such that β is non-degenerate when restricted to M . Then M
is a totally geodesic semi-Riemannian submanifold of (G, β). We claim that for p ∈ M , a
geodesic symmetry at p is given by σp := λp ◦ i ◦ λp−1 , where i is the inversion map of G.
On elements of M , this boils down to

σp (q) = pq −1 p, q ∈ M.

Indeed, σp clearly fixes p, is an isometry (since left translations and τ are), preserves M
by the calculation

τ (pq −1 p) = σ(p−1 qp−1 ) = σ(p−1 )σ(q)σ(p−1 ) = pq −1 p

and satisfies

dσp |p = dλp |e di|e dλp−1 |p = −dλp |e dλp−1 |p = −d(λp ◦ λp−1 )|p = −idTp M .
|{z}
−1

Since dτ |e = −dσ|e , we obtain a β-orthogonal direct sum decomposition

g=k⊕m

of the Lie algebra g of G into the Lie algebra k = {X ∈ g | dσ|e (X) = X} of K and the
tangent space m = Te M to M at e. Since M is totally geodesic, the its curvature tensor
is just the restriction of the curvature tensor of G to M , in other words,
1
RM (X, Y )Z = − [[X, Y ], Z], X, Y, Z ∈ m.
4

Special Case 6.32 (Complex projective space). Set G = U(n + 1) and σ be the automor-
phism of G given by conjugation with the matrix
 
−1
 1 
. (6.5)
 

 ...  
1

Then K := Fix(σ) = U(1) × U(n) ⊂ U(n + 1) and M ∼ = U(n + 1)/(U(1) × U(n)). M is

diffeomorphic to CP n , via the map that sends a matrix U ∈ U(n + 1) onto the first row
vector. We have
m = {bx | x ∈ Cn } ,
where for z ∈ Cn , we set
 
0 −xt
x
b := ∈ u(n + 1) ⊂ Mat(n + 1, C).
x 0

25
Consider the bi-invariant semi-Riemannian metric corresponding to the bilinear form
1
β(X, Y ) = − · Re tr(XY )
8
on u(n + 1) (the conventional factor of 18 is there to make the results look nice later on).
Then the restriction g = β|m of this metric is in terms of these elements given by
 t 
1 −x y 0 1
x, yb) = − · Re tr
g(b t = · Rehx, yi.
8 0 −xy 4
To determine the curvature, we calculate
 
−2i Imhx, yi 0
[b
x, yb] = ,
0 −(xy t − yxt )

where the brackets denote the standard Hermitean form on Cn , which we take to be
anti-linear in the first component. Hence

4R(b z = −[[b
x, yb)b x, yb], zb]
   
0 2i Imhx, yi · z t 0 z t (xy t − yxt )
=− +
−(xy t − yxt )z 0 −2i Imhx, yi · z 0
 b
= hy, zi · x − hx, zi · y − 2i Imhx, yi · z
This can be written in a more invariant form (not using the Hermitean scalar product
but the bilinear form β) as follows. Denote
 
i
 1 
j=  ∈ U(n + 1)
 
. .
 . 
1

and write J = Adj . We have J(b x) = ix,

b hence J is an endomorphism of m that satisfies
J 2 = −id and leaves g invariant. In other words, J is a complex structure for the vector
space m. Using J, we can express h , i through g. The formula is

hx, yi = 4 g(b
x, yb) + 4i g(J x
b, yb),

so  b
hy, zi · x = 4 g(b
y , zb)b
x + 4 g(J yb, zb)J x
b
 b
hx, zi · y = 4 g(b
x, zb)by + 4 g(J xb, zb)J yb
 b
2i Imhx, yi · z = 8 g(J xb, yb)J zb.
Putting together, we obtain

RM (X, Y )Z = g(Y, Z)X + g(JY, Z)JX − g(X, Z)Y − g(JX, Z)JY − 2g(JX, Y )JZ.

26
If X and Y are orthonormal, spanning a plane E ⊂ m, we get for the sectional curvature

K M (E) = 1 + 3g(JX, Y )2 .

Hence the sectional curvature lies between 1 and 4. The maximal value of 4 is taken on
2-planes E that are invariant under J (hence spanned by X and Y = JX).

Special Case 6.33 (Sphere). Let G = SO(n + 1) and let σ be conjugation by the special
matrix (6.5). For K = {1} × SO(n) ⊂ Fix(σ), the quotient G/K is identified with S n ,
while for K = Fix(σ) = (O(1) × O(n)) ∩ SO(n + 1), the quotient is identified with RP n .
The manifold M from Example 6.31 is RP n in both cases.

27