Riemannian Geometry

Contents
Chapter 1. Smooth manifolds 5
1. Tangent vectors, cotangent vectors and tensors 5
2. The tangent bundle of a smooth manifold 5
3. Vector elds, covector elds, tensor elds, n-forms 5
Chapter 2. Riemannian manifolds 7
1. Riemannian metric 7
2. The three model geometries 9
3. Connections 13
4. Geodesics and parallel translation along curves 16
5. The Riemannian connection 17
6. Connections on submanifolds and pull-back connections 19
7. Geodesics in the three geometries 20
8. The exponential map and normal coordinates 21
9. The Riemann distance function 25
Chapter 3. Curvature 29
1. The Riemann curvature tensor 29
2. Ricci curvature, scalar curvature, and Einstein metrics 31
3. Riemannian submanifolds 33
4. Sectional curvature 36
5. Jacobi elds 38
6. Comparison theorems 44
Chapter 4. Space-times 47
Chapter 5. Multilinear Algebra 49
1. Tensors 49
2. Tensors of inner product spaces 51
3. Coordinate expressions 52
Chapter 6. Non-euclidean geometry 55
1. The hyperbolic plane 55
Bibliography 59
3
CHAPTER 1
Smooth manifolds
1. Tangent vectors, cotangent vectors and tensors
1.1. Lemma. Let F : M
m
N
n
be a smooth map. Suppose that (x
1
, . . . , x
m
) are local co-
ordinates on M and (y
1
, . . . , y
n
) local coordinates on N. Then
F

(

x
j
) =
(y
i
F)
x
j

y
i
, 1 j m, (1.2)
F

(dy
i
) =
(y
i
F)
x
j
dx
j
= d(y
i
F), 1 i n (1.3)
The (nm)-matrix
_
(y
i
F)
x
j
_
is the matrix for F

and the (mn)-matrix

_
(y
i
F)
x
j
_
t
is the matrix
for F

. In other words,
_
F

x
1
. . . F

x
m
_
=
_

y
1
. . .

y
n
__
(y
i
F)
x
j
_
_
_
_
F

dy
1
.
.
.
F

dy
n
_
_
_ =
_
(y
i
F)
x
j
_
_
_
_
dx
1
.
.
.
dx
m
_
_
_
Proof. F

(dy
i
)(

x
j
) = dy
i
(F

(

x
j
)) = F

_

x
j
_
(y
i
) =
y
i
F
x
j
.
2. The tangent bundle of a smooth manifold
3. Vector elds, covector elds, tensor elds, n-forms
1.4. Proposition. The dierential d: C

(M) T
1
(M) is an R-linear map satisfying the
derivational rules
d(uv) = (du)v +u(dv), d
_
u
v
_
=
du
v

udv
v
2
=
vdu udv
v
2
for all smooth functions u, v C

(M) (where v(p) ,= 0 for all p M in the last formula).

If F : M N is a smooth map, then the diagram
C

(M)
d

(N)
F

T
1
(M) T
1
(N)
F

commutes meaning that F

(du) = d(uF) for all u C

(N).
1.5. Example. Let s: R
n+1
R be the smooth map s(x) = [x[
2
=

n+1
i=1
(x
i
)
2
. Then
s
1
(1) = S
n
R
n
is the sphere of radius R. The dierential ds =

2x
i
dx
i
is the linear
map R
n+1
= T
p
R
n+1
R given by ds
p
(v) =

2p
i
v
i
= 2p, v) with kernel ker ds
p
= p

at any
point p ,= 0. Let p be any point of S
n
and

: T
p
S
n
T
p
R
n+1
= R
n+1
the linear map induced
by the inclusion, . For any tangent vector X T
p
S
n
, ds(

X) = X(s) = X(1) = 0. Hence the

tangent space at p is the kernel of ds
p
,
T
p
S
n
= p

R
n+1
= T
p
R
n+1
and the tangent bundle of S
n
,
TS
n
= (p, v) S
n
R
n+1
[ v, p) = 0 S
n
R
n+1
5
6 1. SMOOTH manifoldS
is the vector bundle whose bre over any p S
n
is p

. A smooth vector eld on S

n
is a smooth
map v : S
n
R
n+1
such that v(p) p for all p S
n
. Show that any odd sphere has a vector eld
without zeros. Does S
2
admit a smooth vector eld with no zeros? Can you describe TRP
n
? Can
you describe TM if M = f
1
(0) consists of the manifold solutions to the equation f(x) = 0 for
some smooth map f : R
n+1
R?
1.6. Definition. A smooth
_
k

_
-tensor eld on M is a smooth section of the tensor bundle
T
k

(M) M.
Particular cases are
T
0
0
(M) = C

(M)
T
0
1
(M) consists of vector elds on M
T
1
0
(M) consists of 1-forms on M
Tensor elds admit
T
k
1

1
(M) T
k
1

1
(M)

T
k
1
+k
2

1
+
2
(M) (tensor product of tensor elds)
T
k+1
+1
(M)
tr
T
k

(M) (contraction of tensor elds)

1.7. Example. Let T
1
0
(M) be a 1-form and X T
0
1
(M) a vector eld on M. Then
X T
1
1
(M) is
_
1
1
_
-tensor eld with contraction tr(X ) = (X) (5.7).
In a coordinate patch any
_
k

_
-tensor eld A is (5.5) a C

(M)-linear combination
(1.8) A = A
j
1
j

i
1
i
k

j
1

j

dx
i
1
dx
i
k
of tensor products of the basis tensor elds and basis 1-forms. The smooth functions A
j
1
j

i
1
i
k
are
called the components of the tensor eld A.
The tensor algebra of M is the graded algebra T

(M) =

k=0
T
k
(M) equipped with the tensor
product T
r
(M) T
s
(M)

T
r+s
(M). If F : M N is a smooth map, F

: T
k
(N) T
k
(M)
is the linear map given by F

(A)(X
1
, . . . , X
k
) = A(F

X
1
, . . . , F

X
k
) for all A T
k
(N) and all
smooth vector elds X
1
, . . . , X
k
on M.
1.9. Lemma. T

(M) is a graded algebra. F

: T

(N) T

(M) is a homomorphism of C

(N)-
algebras: F

(a ) = F

(a)F

() F

().
CHAPTER 2
Riemannian manifolds
Riemanns idea was that in the innitely small, on a scale much smaller than the the smallest
particle, we do not know if Euclidean geometry is still in force. Therefore we better not assume that
this is the case and instead open up for the possibility that in the innitely small there may be other
length functions, there may be other inner products on the tangent space! A Riemannian manifold
is a smooth manifold equipped with inner product, which may or may not be the Euclidean inner
product, on each tangent space.
1. Riemannian metric
2.1. Definition. A Riemannian metric on a smooth manifold M is a symmetric, positive
denite
_
2
0
_
-tensor g T
2
0
(M).
In a coordinate frame we may write
g = g
ij
dx
i
dx
j
, g
ij
= g(
i
,
j
)
This means that g(U
i

i
, V
j

j
) = g
ij
U
i
V
j
and in particular that
(2.1)
i
,
i
) = g
ii
,
i
,
j
) = g
ij
Note that there are only
1
2
n(n + 1) dierent functions here as g
ij
= g
ji
by symmetry.
2.2. Remark. Since the metric tensor is symmetric, it is traditional to write it in a basis of
symmetric tensors. The symmetrization of is the tensor
=
1
2
( + )
Note that = and that
2
= = . Observe that
g = g
ij
dx
i
dx
j
= 2
n

i=1
g
ii
(dx
i
)
2
+ 2

1i<jn
g
ij
dx
i
dx
j
2.3. Lemma. Let F : M N be an immersion and g a Riemannian metric on N.
(1) F

g is a Riemannian metric on M.
(2) If g = g
ij
dy
i
dy
j
in a coordinate frame on N, then
F

(g)[F
1
(U) = g
ij
dy
i
Fdy
j
F
Proof. It is a general fact that F

(g) is a smooth 2-form on M (1.9). F

(g) is symmetric
because g is symmetric and it is positive denite because g is positive denite and F

is injective
on each bre. F

(g
ij
dy
i
dy
j
)
(1.9)
= g
ij
F

dy
i
F

dy
j
(1.4)
= g
ij
dy
i
Fdy
j
F.
For instance if F : U M is a parameterization (an inverse chart) of an open subset of M
R
m
, then the pull-back of the induced metric on M is
(2.4) F

(g) = F

(
ij
dx
i
dx
j
) =
ij
dF
i
dF
j
=
m

i=1
(dF
i
)
2
These expressions are tensor elds living in the tensor algebra T

(M) of M.
7
8 2. RIEMANNIAN MANIFOLDS
2.5. Example. (Graphs) Let M R
n
R be the graph of the smooth function f : M R.
Then s(x) = (x, f(x)) is a dieomorphism so that the Riemannian manifold (M,

g
n+1
) is isometric
to (R
n
, s

g
n+1
) where the metric s

g
n+1
is
s

(
n+1

i=1
(dx
i
)
2
) =
n

i=1
(dx
i
)
2
+
_
f
x
i
dx
i
_
2
=
n

i=1
(dx
i
)
2
+
f
x
i
f
x
j
dx
i
dx
j
=
_

ij
+
f
x
i
f
x
j
_
dx
i
dx
j
=
n

i=1
_
1 + (
f
x
i
)
2
_
(dx
i
)
2
+ 2

1i<jn
f
x
i
f
x
j
dx
i
dx
j
2.6. Example. let S
2
+
be the upper hemisphere on S
2
R
3
considered as the graph of the
function f(x, y) =
_
1 x
2
y
2
dened on the unit ball B
2
R
2
. Then (S
2
+
,

g
3
) is isometric
to (B
2
, s

g
3
) where
s

g
3
= (dx)
2
+ (dy)
2
+
_
x
_
1 x
2
y
2
dx +
y
_
1 x
2
y
2
dy
_
2
=
1 y
2
1 x
2
y
2
(dx)
2
+
1 x
2
1 x
2
y
2
(dy)
2
+
2xy
1 x
2
y
2
dxdy
This means (2.1) that

x
,
x
) =
1 y
2
1 x
2
y
2
,
y
,
y
) =
1 x
2
1 x
2
y
2
,
x
,
y
) =
xy
1 x
2
y
2
at the point (x, y) B
2
. In this metric, the basis tangent vectors,
x
and
y
, are not orthogonal
at any point of the unit ball away from the axes. If we consider the curve (t) = (t, 0), 1 t 1,
then the tangent vector

(
d
dt
) =
d(x)
dt

x
+
d(y)
dt

y
=
x
so that the length of this curve is
L
s

(g)
() =
_
+1
1
[

(
d
dt
)[dt =
_
+1
1
_
1
1 t
2
dt =
What is the distance between (0, 1/2) and (1/2, 0)? What is the curve of shortest length between
these two points?
2.7. Example. (Surface of revolution in R
3
)
2.8. Definition. A smooth map F : (M, g) (N, h) between two Riemannian manifolds is an
isometry if g = F

h; if g(X, Y ) = h(F

X, F

Y ) for all tangent vectors X, Y T

p
M, p M.
Two Riemannian manifolds are isometric if we can deform one into the other by bending but
not stretching. Is the upper unit hemisphere S
2
+
R
3
isometric to the open unit ball B
2
R
2
?
Certainly, the dieomorphism s: B
2
S
2
+
from Example 2.6 is not an isometry as for instance

x
,
x
) ,= 1 or because the tangent vectors
x
and
y
are not orthogonal throughout any open
(nonempty) subspace of B
2
. But there are many other dieomorphisms and maybe we could
nd one that preserves the metrics? To decide if this is the case we need to nd invariants of
Riemannian metrics. Is S
2
+
curved? And what does that mean? We need to develop some theory
to answer these questions.
In order to decide if two given Riemannian manifolds are isometric we have to know have
the metric tensor transforms under change of coordinate system? (Remember that the coordinate
expression for a metric is an artefact of the coordinate system and not an intrinsic property of the
metric.)
2.9. Lemma. Let : R
n
M and : R
n
M be parameterizations of the same open subspace
of M. If

(g) = a
ij
dx
i
dx
j
,

(g) = b
ij
dy
i
dy
j
then
(a
ij
) =
_
y
i
x
j
_
t
(b
ij
)
_
y
i
x
j
_
where y =
1
.
2. THE THREE MODEL GEOMETRIES 9
Proof. Put P =
_
y
i
x
j
_
. Then (1.1)
y

_
_
_
dy
1
.
.
.
dy
n
_
_
_ = P
_
_
_
dx
1
.
.
.
dx
n
_
_
_
Therefore
a
ij
dx
i
dx
j
=

g = (y)

g = y

g) = y

(b
ij
dy
i
dy
j
)
= y

(
_
dy
1
. . . dy
n
_
(b
ij
)
_
_
_
dy
1
.
.
.
dy
n
_
_
_)
=
_
dx
1
. . . dx
n
_
P
t
(b
ij
)P
_
_
_
dx
1
.
.
.
dx
n
_
_
_
so that (a
ij
) = P
t
(b
ij
)P.
Suppose that

g = a
ij
dx
i
dx
j
for some parameterization . Is M locally at? In other
words, does there exist a re-parameterization = y of M such that

g =
ij
dy
i
dy
j
? Such a
re-parameterization exists if and only if the set of
1
2
n(n 1) PDEs
(g
ij
) =
_
y
i
x
j
_
t
_
y
i
x
j
_
,
or equivalently,
(2.10) g
ij
=
n

k=1
y
k
x
i
y
k
x
j
, 1 i j n,
has a solution y = (y
1
, . . . , y
n
). Riemann showed (in an essay that was never properly recognized)
that (2.10) is equivalent to

2
y

x
i
x
j
=
k
ij
y

x
k
, 1 i, j, n,
and thereby that (2.10) has a solution if and only if
R
m
jk
= 0, 1 j, k, , m n,
where the
k
ij
(the Christoel symbols) and the R
m
jk
are certain functions dened in terms of the
functions g
ij
. This was the birth of the Riemann curvature tensor R
m
jk
! This direct approach,
however, is not the one used today since it is conceptually simpler rst to introduce a device called
a connection that will enable us to work in a coordinate-free way on M.
2.11. Example. (Cylinders are at.) Let (s) = (x(s), y(s)), a < s < b, be a smooth curve in
R
2
such that the tangent

(
d
ds
) = x

(s)
x
+ y

(s)
y
,= 0 for all t. Let M R
3
be the cylinder
over , the surface with parameterization (s, t) = (x(s), y(s), z(t)) where z(t) = t. Then (2.3),

(g
3
) = (dx)
2
+ (dy)
2
+ (dz)
2
= (x

(s)ds)
2
+ (y

(s)ds)
2
+ (dt)
2
= [

(
d
ds
)[
2
(ds)
2
+ (dt)
2
What are Riemanns equations in this case? Is there a solution?
2. The three model geometries
The model geometries are Euclidean geometry, spherical geometry, and hyperbolic geometry.
10 2. RIEMANNIAN MANIFOLDS
2.1. Euclidean geometry. Euclidean geometry is the geometry of the Riemannian manifold
(R
n
, g
n
) where
g
n
=
ij
dx
i
dx
j
=
n

i=1
(dx
i
)
2
meaning that g
n
(U
i

i
, V
j

j
) =

n
i=1
U
i
V
i
. (The straight line over the g is to remind you of
Euclidean geometry.)
The isometry group of Euclidean geometry
Isom(R
n
, g
n
) = A(n) = R
n
O(n)
acts transitively on R
n
by the rule (v, A)(x) = v +Ax. The isotropy subgroup at 0 R
n
is O(n)
and the projection
O(n) A(n) = O(TR
n
) A(R
n
)/O(n) = R
n
is the unit n-frame bundle of Euclidean geometry R
n
.
2.2. Spherical geometry. Spherical geometry is the geometry of the Riemannian manifold
(S
n
R
, g
n
R
) where
S
n
R
= (, ) R
n
R [ [[
2
+
2
= R
2

is the n-sphere of radius R and g

n
R
=

_
n
i=1
(d
i
)
2
+ (d)
2
_
is the restriction of the Euclidean
metric on R
n
R.
The isometry group of spherical geometry
Isom(S
n
R
, g
R
) = O(n + 1)
(The smooth action of O(n + 1) A(n + 1) on R
n+1
restricts to a smooth action on S
n

and in fact these are all isometries.) O(n + 1) acts transitively on S

n
. The isotropy subgroup at
N = (0, . . . , 0, 1) S
n
is O(n) and the projection
O(n) O(TS
n
) = O(n + 1) O(n + 1)/O(n) = S
n
is the unit n-frame bundle of spherical geometry S
n
.
2.12. Proposition. Stereographic projection : S
n
R
N R
n
is given by
(, ) =
R
R

and the inverse is given by

1
(u) = ((u), (u)), (u) =
2R
2
u
[u[
2
+R
2
, (u) = R
[u[
2
R
2
[u[
2
+R
2
Stereographic projection is a dieomorphism.
Proof. Elementary.
2.13. Proposition. The Riemannian manifold (S
n
R
N, g
n
R
) is isometric to the Riemannian
manifold (R
n
, (
1
)

g
n
R
) where
(
1
)

g
n
R
=
4R
4
(R
2
+[u[
2
)
2
g
n
is conformally equivalent to the Euclidean metric.
Proof. By 2.3
(
1
)

g
R
=
n

j=1
(d
j
)
2
+ (d)
2
2. THE THREE MODEL GEOMETRIES 11
We will now use the derivation rules 1.4. The denominator of and is [u[
2
+R
2
and d([u[
2
+R
2
) =
d([u[
2
) =

2u
j
du
j
which we shall write as 2u, du). Because
d
j
= d
_
2R
2
u
[u[
2
+R
2
_
=
2R
2
du
j
[u[
2
+R
2

4R
2
u
j
u, du)
([u[
2
+R
2
)
2
d = d
_
R
[u[
2
R
2
[u[
2
+R
2
_
= R
2u, du)
([u[
2
+R
2
)
2
R
([u[
2
R
2
)2u, du)
([u[
2
+R
2
)
2
=
2Ru, du)([u[
2
+R
2
) 2Ru, du)([u[
2
R
2
)
([u[
2
+R
2
)
2
=
4R
3
u, du)
([u[
2
+R
2
)
2
we get
n

j=1
(d
j
)
2
=
n

j=1
_
4R
4
(du
j
)
2
[u[
2
+R
2
+
16R
4
(u
j
)
2
u, du)
([u[
2
+R
2
)
4

16R
4
u
j
du
j
u, du)
([u[
2
+R
2
)
3
_
=
4R
4
(du
j
)
2
([u[
2
+R
2
)
2
n

j=1
(du
j
)
2
+
16R
4
[u[
2
u, du)
2
([u[
2
+R
2
)
4

16R
4
u, du)
2
([u[
2
+R
2
)
3
=
4R
4
(du
j
)
2
([u[
2
+R
2
)
2
g
n
+
16R
4
[u[
2
u, du) 16R
4
[u[
2
u, du)
2
16R
6
u, du)
2
([u[
2
+R
2
)
4
=
4R
4
(du
j
)
2
([u[
2
+R
2
)
2
g
n

16R
6
u, du)
2
([u[
2
+R
2
)
4
=
4R
4
(du
j
)
2
([u[
2
+R
2
)
2
g
n
(d)
2

2.14. Hyperbolic geometry. Hyperbolic geometry is the geometry of the Riemannian man-
ifold (H
n
R
, h
n
R
) where
H
n
R
= (, ) R
n
R
+
[ [[
2

2
= R
2

id the hyperbolic space n-space of radius R and h

n
R
=

_
n
i=1
(d
i
)
2
(d)
2
_
is the restriction of
the Minkowski metric on R
n
R. Let N = (0, . . . , 0, R) H
n
R
be the north pole.
2.15. Remark. (Minkowski metric) Let m be the inner product on R
n+1
with matrix D =
diag(1, . . . , 1, 1). We can view m both as an inner product on the vector space R
n
(X, Y ) =
m(X, Y ) = X
t
DY ) or as a Minkowski metric on the manifold R
n+1
(X
i

i
, Y
i

i
) = m(X, Y )).
For each p H
n
R
, the tangent space
T
p
H
n
R
= p

R
n+1
= T
p
R
n+1
exactly as in 1.5.
Let
O(n, 1) = A GL(n, R) [ A
t
DA = D
be the group of linear automorphisms of R
n+1
that preserve the inner product. The columns (or
rows) of each A O(n, 1) form an orthogonal basis for R
n+1
of vectors of length 1, . . . , 1, 1.
The elements of the Lie group O(n, 1) preserve the subspace (, ) R
n+1
[ [(, )[
2
= R
2
of
vectors of square length R
2
. This subspace has two connected components. Let O
+
(n, 1) be the
subgroup of O(n, 1) consisting of the elements that take the connected component H
n
R
to itself.
The Lie group O
+
(n, 1) acts transitively on H
n
R
; given any v H
n
R
, we can nd an A O
+
(n, 1)
whose last column is
1
R
v so that AN = v where N is the north pole. The isotropy subgroup at N
consists of the A O
+
(n, 1) whose last column is (0, . . . , 0, 1). The isotropy subgroup, isomorphic
to O(n), act transitively on the tangent space T
N
H
n
R
. The projection
O(n) O(TH
n
R
) = O
+
(n, 1) O
+
(n, 1)/O(n) = H
n
R
is the unit n-frame bundle of H
n
R
.
12 2. RIEMANNIAN MANIFOLDS
2.16. Proposition. Hyperbolic stereographic projection : H
n
R
B
n
R
is given by
(, ) =
R
R +
and its inverse is given by

1
(u) = ((u), (u)), (u) =
2R
2
u
R
2
[u[
2
, (u) = R
R
2
+[u[
2
R
2
[u[
2
Hyperbolic stereographic projection is a dieomorphism.
Proof. Elementary.
2.17. Proposition. The Riemannian manifold (H
n
R
, h
n
R
) is isometric to the Riemannian man-
ifold (B
n
(R), (
1
)

h
n
R
) where
(
1
)

(h
n
R
) =
4R
4
(R
2
[u[
2
)
2
g
n
is conformally equivalent to the Euclidean metric.
Proof. By 2.3 (applied to the Minkowski metric)
(
1
)

h
R
=
n

j=1
(d
j
)
2
(d)
2
Using that d(R
2
[u[
2
) = 2u, du) and 1.4 we get
d
j
=
2R
2
du
j
R
2
[u[
2
+
2R
2
u
j
2u, du)
(R
2
[u[
2
)
2
d = R
2u, du)
R
2
[u[
2
+R
(R
2
+[u[
2
)2u, du)
(R
2
[u[
2
)
2
=
2R
3
u, du) 2R[u[
2
u, du) + 2R
3
u, du) + 2R[u[
2
u, du)
(R
2
[u[
2
)
2
=
4R
3
u, du)
(R
2
[u[
2
)
2
Therefore
n

j=1
(d
j
)
2
=
n

j=1
_
2R
2
du
j
R
2
[u[
2
+
4R
2
u
j
u, du)
(R
2
[u[
2
)
2
_
2
=
4R
4
(R
2
[u[
2
)
2
n

j=1
(du
j
)
2
+
16R
4
[u[
2
u, du)
(R
2
[u[
2
)
4
+
16R
4
u, du)
2
(R
2
[u[
2
)
3
=
4R
4
(R
2
[u[
2
)
2
g
n
+
16R
4
[u[
2
u, du) + 16R
6
u, du)
2
16R
4
[u[
2
u, du)
2
(R
2
[u[
2
)
4
=
4R
4
(R
2
[u[
2
)
2
g
n
+
16R
6
u, du)
2
(R
2
[u[
2
)
4
=
4R
4
(R
2
[u[
2
)
2
g
n
+ (d)
2

Let U
n
= (x, y) R
n1
R [ y > 0 be the upper half plane in R
n
.
2.18. Proposition. The Riemannian manifold (H
n
R
, h
n
R
) is isometric to the smooth manifold
U
n
equipped with the Riemannian metric R
2 1
y
2
g
n
.
Proof. A computation.
Does the hyperbolic plane H
2
embed isometrically in R
3
? Any Riemannian ma embeds iso-
metrically into some Euclidean space [6].
3. CONNECTIONS 13
3. Connections
Let E M be a smooth vector bundle over M and c(M) the C

(M)-module of smooth
sections. A connection on E is a recipe for how to dierentiate a section of E along a vector eld.
2.19. Definition. A connection on E is a map
T (M) c(M)

c(M)
(X, Y )
X
Y
which is C

(M)-linear in X and R-linear in Y and satises the product rule

X
(fY ) = (Xf)Y +f
X
Y
for all f C

(M).
There always are connections, for instance the 0-connection given by
X
Y = 0.
2.20. Lemma. The value
X
Y (p) at the point p M only depends on Y in a neighborhood of
p and X at p.
Proof. Let us rst focus on Y -variable. By linearity, it is enough to show that if Y = 0 in a
neighborhood U of p, then
X
Y (p) = 0. Choose a smooth bump function such that (p) = 1
and = 0 outside U. Then Y is the zero section so that
0 =
X
0 =
X
(Y ) = (X)Y +
X
Y
Evaluating at p, we get 0 =
X
Y (p) since Y (p) = 0 and (p) = 1.
Next, we focus on the X-variable. By an argument similar to the one just given, we rst show
that if X is 0 in a neighborhood of p, then 0 =
X
Y (p). Suppose now that we only know that X
vanishes at the point p, X(p) = 0. Choose a moving frame E
i
in a neighborhood of p. Extend the
locally dened vector elds E
i
to globally dened smooth vector elds. There are smooth functions
X
i
such that X = X
i
E
i
in a neighborhood of p. Then
X
Y (p) =
X
i
E
i
Y (p) since X and X
i
E
i
are equal in a neighborhood of p. By C

(M)-linearity in X,

X
i
E
i
Y = X
i

E
i
Y
which evaluated at p is 0 since X
i
(p) = 0 for all i.
We are particularly interested in connections on the tangent bundle of M.
2.21. Definition. A linear connection is a connection T (M)T (M)

T (M) on the tangent
bundle of M.
Are there any linear connections, apart from the 0-connection, on M?
2.22. Example. The Euclidean connection
X
(Y
j

j
) = X(Y
j
)
j
is a nonzero linear connec-
tion on Euclidean space M = R
n
. Note that

j
= 0, 1 i, j n. A smooth vector eld
Y = Y
i

i
on R
n
is the same thing as a smooth map Y = Y
i
E
i
: R
n
R
n
. The derivative in the
direction of (tangent) vector X of the map Y = Y
i
E
i
is X(Y ) = X(Y
i
)E
i
. For consistency sake
we better declare the derivative of the vector eld Y = Y
i

i
to be
X
Y = X(Y
i
)
i
to ensure that
the diagram
C

(R
n
, R
n
)
X

(R
n
, R
n
)
T (R
n
)

X

T (R
n
)
commutes.
If we can construct connections on R
n
, maybe we can dene a connection in each coordinate
patch on M and then put them together? What does a connection look like locally?
14 2. RIEMANNIAN MANIFOLDS
2.23. Lemma. Suppose that U is an open subspace of M that admit a moving frame E
i
. The
linear connection on M restricts to linear connection on U. If X = X
i
E
i
and Y = Y
j
E
j
are
vector elds on U then

X
Y = (XY
k
+X
i
Y
j

k
ij
)E
k
where the n
3
smooth functions
k
ij
are the Christoel symbols given by
E
i
E
j
=
k
ij
E
k
.
Proof. We compute

X
Y =
X
(Y
j
E
j
) = XY
j
E
j
+Y
j

X
E
j
XY
j
E
j
+Y
j

X
i
E
i
E
j
= XY
j
E
j
+X
i
Y
j

E
i
E
j
= XY
j
E
j
+X
i
Y
j

k
ij
E
k
= XY
k
E
k
+X
i
Y
j

k
ij
E
k
= (XY
k
+X
i
Y
j

k
ij
)E
k
where we use the Christoel symbols.
Thus we see that we can express
X
Y by means of the n
3
smooth functions
k
ij
. Conversely,
for any choice of n
3
smooth functions
k
ij
we can dene
X
Y , X, Y T (U) by the above formula
and that will be a connection. So a linear connection on U is the same thing as a collection of n
3
smooth functions on U.
In particular if M is a smooth manifold and U

a smooth atlas on M then can nd a linear

connection

on each U

. In order to construct a linear connection on M, let

be a smooth
partition of unity subordinate to U

.
Theorem 2.24. (Existence of connections) Any smooth manifold M has many linear connec-
tions:

X
Y =

X
Y
is a linear connection on M.
Armed with a linear connection we know how to dierentiate vector elds along vector elds.
But as a bonus we can even dierentiate arbitrary tensor elds along vector elds!
2.25. Lemma. (Existence and uniqueness of the covariant derivative of a tensor eld) Let
T (M) T (M)

T (M) be a linear connection on M (2.21). Then there are unique connections
T (M) T
k

(M)

T
k

(M)
on all tensor bundles T
k

(M) M such that

(1)
X
f = Xf for all f T
0
0
(M) = C

(M)
(2)
X
Y is the given linear connection for all vector elds Y T
1
0
(M)
(3)
X
(AB) =
X
AB +A
X
B
(4)
X
tr A = tr
X
A for all A T
k+1
+1
(M)
Namely, for any 1-form ,
X
is the 1-form given by
(2.26) (
X
)(Y ) = X((Y )) (
X
Y )
and, in general, for any
_
k

_
-tensor A T
k

(M),
X
A is the
_
k

_
-tensor given by
(2.27) (
X
A)(
1
, . . . ,

, Y
1
, . . . , Y
k
) = X(A(
1
, . . . ,

, Y
1
, . . . , Y
k
))

i=1
A(
1
, . . . ,

, Y
1
, . . . ,
X
Y
i
, . . . , Y
k
)

j=1
A(
1
, . . . ,
X

j
, . . . ,

, Y
1
, . . . , Y
k
)
for any choice of k vector elds Y
1
, . . . , Y
k
and 1-forms
1
, . . . ,

on M.
Proof. Lets assume that we have connections that satisfy items (1)(4). What is the cov-
ariant derivative of a 1-form ? For any two vector elds X and Y ,
X((Y ))
(1)
=
X
((Y ))
(1.7)
=
X
(tr( Y ))
(4)
= tr
X
( Y )
(3)
= tr(
X
Y +
X
Y )
(1.7)
= (
X
)(Y ) +(
X
Y )
3. CONNECTIONS 15
which is (2.26). So we are forced to dene
X
as in (2.26). But then there is at most one
possibility for
X
A since any tensor is a sum of a smooth function times tensor products of vector
elds and 1-forms (combine the local expression (1.8) for a tensor with a smooth partition of unity).
This shows uniqueness.
To show existence, use (2.26) to dene the covariant derivative of a 1-form and then use (2.27)
to dene
X
A in general. Check that this denition satises (1)(4).
2.28. Definition. The total covariant derivative : T
k

(M) T
k+1

(M) is given by
A(
1
, . . . ,

, Y
1
, . . . , Y
k
, X) = (
X
A)(
1
, . . . ,

, Y
1
, . . . , Y
k
)
for all A T
k

(M).
Note that the total covariant derivative of the tensor eld A T
k

(M) is zero if and only if

the covariant derivative of A along all vector elds is zero: A = 0 X T (M):
X
A = 0.
2.29. Example. There are total covariant derivatives
C

(M) = T
0
0
(M)

T
1
0
(M)

T
2
0
(M)

T
3
0
(M)


If u C

(M) is a smooth function and X a vector eld, then

(u)(X) =
X
(u) = X(u) = du(X)
so that u = du.
The 2-form
2
u = u is called the covariant Hessian of u. If X and Y a vector elds, then
(
2
u)(Y, X) = (u)(Y, X)
(2.28)
= (
X
u)(Y )
(2.26)
= X((u)(Y )) u(
X
Y )
u=du
= X(Y (u)) (
X
Y )(u)
Consequently,

2
u = 0 X, Y T (M): Y (X(u)) = (
Y
X)(u)
2.30. Example. If g T
2
0
(M) is the Riemannian metric, then g is the 3-form given by
(g)(X, Y, Z)
(2.28)
= (
Z
g)(X, Y )
(2.27)
= Zg(X, Y ) g(
Z
X, Y ) g(X,
Z
Y )
for any three vector elds X, Y, Z. Consequently
g = 0 X, Y, Z T (M): ZX, Y ) =
Z
X, Y ) +X,
Z
Y )
for all vector elds X, Y, Z.
2.31. Example. There are total covariant derivatives
T (M) = T
0
1
(M)

T
1
1
(M)

T
2
1
(M)


If V T
0
1
(M) is a vector eld then V is the
_
1
1
_
-tensor given by
(V )(, X) = (
X
V )() = (
X
V )
If V = V
i

i
in local coordinates, then
(V )(dx
i
,
j
) = dx
i
(

j
V ) = dx
i
(

j
V
k

k
) = dx
i
(
j
V
k

k
+V
k

jk

) =
j
V
i
+V
k

i
jk
and therefore
V = (
j
V
i
+V
k

i
jk
)
i
dx
j
We say that the vector eld V is parallel if V = 0. Since
V = 0 X T (M):
X
V = 0,
V is parallel i the covariant derivative of V along any vector eld is zero. What tensor is
2
V ?
What does it mean if
2
V = 0?
16 2. RIEMANNIAN MANIFOLDS
4. Geodesics and parallel translation along curves
Let : I M be a smooth curve on M.
2.32. Definition. For any t I, the tangent vector

(t) =

(
d
dt
(t)) T
(t)
M,

(t)f =
d(f )
dt
(t), f C

(M)
is called the velocity vector of at the point (t).
If x is a coordinate system around (t
0
) and x = (
1
, . . . ,
n
) then
(2.33)

(t) =
d
i
dt
(t)
i
((t))
as a special case of (1.2).
2.34. Remark. We say that the curve , dened in an open neighborhood of 0, represents
the tangent vector V T
p
M if (0) = p and

(0) = V . Then V f =

(0)f =
d
dt
(f)(0) for any
smooth function f C

(M) and F

V is represented by the image curve F for any smooth map

F : M N.
2.35. Definition. A vector eld along is a smooth map V : I TM such that the diagram
TM

I
V

{
{
{
{
{
{
{
{

M
commutes. A vector eld V along is extendible if V (t) = V ((t)) for some vector eld V on a
neighborhood of (I). The C

(I)-module of all vector elds along is denoted T ().

The velocity eld

(t) is an example of a vector eld along .

The formulation of the lemma below makes use of 2.20.
2.36. Lemma (Covariant dierentiation along a curve). Let be a connection on M. There
exists precisely one R-linear map D
t
: T () T () such that
D
t
(fV ) =
df
dt
V +fD
t
V, f C

(I), V T ()
D
t
V =
(t)
V V extendible
If E
i
is a local frame around (t
0
) and V = V
j
E
j
, then
D
t
V =
dV
j
dt
E
j
+V
j

(t)
E
j
for t near t
0
.
If x is a coordinate system around (t
0
) and V = V
j

j
, then
D
t
V = (
d
k
dt
+
k
ij
d
i
dt
V
j
)
k
for t near t
0
.
A curve on M is as close as possible to being a straight line if the curve at all times just
continues in the direction of

(t), if its velocity

(t) does not change. No change means zero

covariant derivative.
2.37. Definition. A smooth curve is a geodesic (with respect to the connection ) if the
covariant derivative of its velocity eld vanishes, D
t

= 0.
A geodesic is a curve that follows its own nose. A geodesic is a curve with constant velocity.
Light rays follow geodesics in space-time.
5. THE Riemannian CONNECTION 17
If x is a coordinate system around some point (t
0
) of , then is a geodesic i
(2.38)
d
2

k
dt
2
+
k
ij
d
i
dt
d
j
dt
= 0
near t
0
. In principle, we could determine geodesics by solving this equation. In practice, this is
close to impossible. Instead we try to identify some properties that a geodesic must have (2.54,
2.58).
2.39. Proposition (Existence and uniqueness of geodesics). Let p be a point on M and
V T
p
M a tangent vector at p There exists a unique maximal geodesic
V
: I M dened on an
open interval containing 0 such that
V
(0) = p and

V
(0) = V .
2.40. Lemma (Rescaling lemma). Let V T
p
M be a tangent vector at p M and let c R
be a real number. The geodesic
cV
is dened at t i the geodesic
V
is dened at ct and then

cV
(t) =
V
(ct).
Proof. Consider the maximal geodesic
V
: I M. Put (t) =
V
(ct) for all t c
1
I. Then
(t) is a geodesic because it satises (2.38) in any coordinate system and it is dened at c
1
t. In
fact, (t) =
cV
(t) as (0) =
V
(0) = p and

(0) = cV .
A vector eld along a curve is parallel if it doesnt change; no change meaning zero covariant
derivative. A curve is a geodesic if it has a parallel velocity eld.
2.41. Definition. A vector eld V along is parallel if it does not change along , D
t
V = 0.
2.42. Proposition. Let be a curve on M. Suppose that p
0
= (t
0
) is a point on and
V
0
T
p
0
M a tangent vector at that point. There exists precisely one parallel vector eld V along
such that V (t
0
) = V
0
.
5. The Riemannian connection
On a Riemannian manifold there is a preferred connection.
Theorem 2.43 (The fundamental theorem of Riemannian geometry). A Riemannian manifold
admits precisely one symmetric connection compatible with the metric.
This particular connection is called the Riemannian connection or the LeviCivitta connection.
2.44. Definition. The connection is symmetric if

X
Y
Y
X = [X, Y ]
for all vector elds X, Y T (M).
2.45. Lemma. Let E
i
be a local moving frame such that [E
i
, E
j
] = 0, 1 i, j n (for instance
E
i
=
i
could be a coordinate frame). Then is symmetric if and only if

k
ij
=
k
ji
, 1 i, j, k n
Proof.
E
i
E
j

E
j
E
i
= (
k
ij

k
ji
)E
k
.
2.46. Definition. The connection is compatible with the metric g if
XY, Z) =
X
Y, Z) +X,
X
Z)
for any three vector elds X, Y, Z T (M).
2.47. Lemma. Let E
i
be a local moving frame. Then is compatible with the metric g if and
only if
E
k
g
ij
=

ki
g
j
+

kj
g
i
, 1 i, j, k, n
Proof. E
k
E
i
, E
j
)
E
k
E
i
, E
j
) E
i
,
E
k
E
j
) = E
k
g
ij

ki
g
j

kj
g
i
.
18 2. RIEMANNIAN MANIFOLDS
Proof of Theorem 2.43. We rst show uniqueness. Assume that is a symmetric con-
nection that is compatible with g. Then
XY, Z)
2.44
=
X
Y, Z) +Y,
X
Z)
2.46
=
X
Y, Z) +Y,
Z
X) +Y, [X, Z])
for any three vector elds X, Y, Z T (M). Permute X, Y, Z cyclically, obtain
XY, Z) Z X, Y ) +Y X, Z) = 2
X
Y, Z) +Y, [X, Z]) X, [Z, Y ]) +Z, [Y, X])
and conclude that
(2.48)
X
Y, Z) =
1
2
(XY, Z) Z X, Y ) +Y X, Z) Y, [X, Z]) +X, [Z, Y ]) Z, [Y, X]))
This equation shows that , if it exists, is determined by the metric.
Next, we show existence. We will dene in any open submanifold where we have a moving
frame E
i
with [E
i
, E
j
] = 0 (for instance in a chart domain). The only possibility is to put
(2.49)
k
ij
=
1
2
g
k
(E
i
g
j
E

g
ij
+E
j
g
i
)
because
(2.50)
k
ij
g
k
=
E
i
E
j
, E

) =
1
2
(E
i
g
j
E

g
ij
+E
j
g
i
)
by equation (2.48). Since
k
ij
=
k
ji
and

ki
g
j
+

i
g
i
=

ki
g
j
+

i
g
i
2.50
=
1
2
(E
k
g
ij
E
j
g
ik
+E
i
g
jk
)+
1
2
(E
k
g
ij
E
i
g
jk
+E
j
g
ik
) = E
k
g
ij
this connection is symmetric and compatible with g (2.45, 2.47).
2.51. Example. The Euclidean connection (2.22) is the Riemannian connection on Euc-
lidean space R
n
for it is symmetric and compatible with the Euclidean metric g (2.1) since
k
ij
= 0.
What is the Riemannian connection in spherical and hyperbolic space?
2.52. Lemma. Let be a connection and g a metric on M. The following conditions are
equivalent:
(1) and g are compatible (2.46)
(2) g = 0
(3)
d
dt
V, W) = D
t
V, W) +V, D
t
W) whenever V, W are vector elds along a smooth curve
(4) V, W) is constant whenever V, W are parallel vector elds along a smooth curve
(5) Parallel transport is an isometry
Proof. (1) (2): 2.30
(3) = (4) = (5) : Obvious.
(5) =(3) : Let P
1
, . . . , P
n
be parallel vector elds that are orthonormal at one point of the curve
and hence orthonormal at any point. Write V = V
i
P
i
and W = W
i
P
i
. Then V, W) =

V
i
W
i
and (2.36) D
t
V =
dV
i
dt
P
i
and D
t
W =
dW
i
dt
P
i
. Hence
D
t
V, W) +V, D
t
W) =

dV
i
dt
W
i
+

V
i
dW
i
dt
=
d
dt

V
i
W
i
=
d
dt
V, W)
(3) = (1) : Let Y and Z be smooth vector elds on M and let be a smooth curve, (0) = p,

(0) = X(p). Then

X
p
Y, Z) =

(0) Y, Z) =
d
dt
Y, Z) = D
t
Y, Z) +Y, D
t
Z) =

X
p
Y, Z
_
+

Y,
X
p
Z
_
(1) = (3) : Let p = (0), choose a vector eld X with X(p) =

(0), and choose an orthonormal

moving frame, E
i
, around p. Then
0 = XE
i
, E
j
) =
X
E
i
, E
j
) +E
i
,
X
E
j
)
6. CONNECTIONS ON SUBMANIFOLDS AND PULL-BACK CONNECTIONS 19
since and g are compatible Write V = V
i
E
i
and W = W
j
E
j
. Then D
t
V =

V
i
E
i
+ V
i

X
E
i
,
and similarly for W, at the point p (2.36). Therefore,
D
t
V, W) +V, D
t
W) = (

V
i
W
j
+V
i

W
j
) E
i
, E
j
) +V
i
W
j
(
X
E
i
, E
j
) +E
i
,
X
E
j
))
=

V
i
W
i
+V
i

W
i
) =
d
dt

V
i
W
i
=
d
dt
V, W)

2.53. Lemma (Covariant dierentiation commutes with lowering and raising of indices (5.2)).
If the connection is compatible with the metric then the diagram
T
k+1

(M)

T
k
+1
(M)

T
k+1

(M)

T
k
+1
(M)

commutes for any vector eld X.

Proof. It will be enough to prove this for vector elds (k = 0 = )). Suppose that X, U, V
are vector elds. The claim is
X
(V

) = (
X
V )

. We compute
(
X
(V

)(U) = X(V

)V

(
X
U) = XU, V )
X
U, V ) =
X
U, V )+U,
X
V )
X
U, V )
= U,
X
V ) = (
X
V )

(U)
using (2.26) and (2.46).
2.54. Corollary. Riemannian geodesics have constant speed.
Proof.
d
dt
[

(t)[
2
=
d
dt

(t),

(t)) = 2D
t

(t),

(t)) = 0.
6. Connections on submanifolds and pull-back connections
Let (M, g) be a Riemannian manifold and M M be an embedded submanifold. Suppose
that we have a connection on M. How can we obtain a connection on the submanifold M?
For any tangent vector X
p
T
p
M, let X
T
p
T
p
M denote the orthogonal projection of X
p
.
2.55. Proposition (Existence of uniqueness of tangential connections). There exists precisely
one connection
T
on M such that

T
X
Y =
_

X
Y
_
T
whenever X, Y are vector elds on M and X, Y their restrictions to M. If is the Riemannian
connection on M, then
T
is the Riemannian connection on M.
2.56. Lemma. Let be a connection and let X, Y T (M) be vector elds on M. Then

X
Y (p) only depends on X(p) and Y along a curve tangent to X(p).
Proof.
X
Y (p) =
(0)
Y (p) = D
t
Y (0) for any curve : (, ) M with (0) = p and

(0) = X(p).
2.57. Proposition (Pull-back connections). Let : M M be a dieomorphism and a
connection on M.
Let (also) be the map that makes the diagram
T (M) T (M)

_ _ _
T (M)

T (M) T (M)


T (M)
commutative. Thus

X
Y =

Y for all X, Y T (M).

20 2. RIEMANNIAN MANIFOLDS
(1) is a connection on M.
(2) Covariant dierentiation wrt makes the diagram
T ()
D
t

T ()

T ()
D
t

T ()
commutative. Thus

D
t
V = D
t

V for any vector eld V along the curve in M.

(3) Assume that M and M are Riemannian manifolds and that is an isometry. If
the Riemannian connection on M, then is the Riemannian connection on M. Thus

X
Y =

for all X, Y T (M).

2.58. Corollary. Isometries of Riemannian manifolds take Riemannian geodesics to Rieman-
nian geodesics.
Proof. Let : (M, g) (M, g) be an isometry of Riemannian manifolds. Then the connec-
tion on M is the pull-back of the connection on M (2.57.(3)). Let
V
, V T
p
M, be a geodesic on
M. Then
D
t
((
V
)

) = D
t
(

)
2.57.(2)
=

D
t
(

) =

0 = 0
so that
V
=

V
.
7. Geodesics in the three geometries
We determine the geodesics in Euclidean, spherical, and hyperbolic geometry.
2.59. Euclidean geometry. The Riemannian connection on R
n
is the Euclidean connection
(2.22).
Let p = (0, . . . , 0) R
n
and V = (1, 0, . . . , 0) T
p
R
n
. We know that there is a unique
maximal geodesic
V
running through p with velocity V . We also know that
V
=
V
for any
isometry O(n) of R
n
preserving (p, V ). The map (
1
,
2
,
3
, . . . ,
n
) = (
1
,
2
,
3
, . . . ,
n
)
is such an isometry (it is a dieomorphism and it preserves [[). Thus
V
must have
2

V
= 0.
Similarly,
3

V
= 0, . . . ,
n

V
= 0. Thus
V
must run along the
1
-axis. Since it has constant
speed and

V
(0) = V , we must have

V
(t) = (t, 0, . . . , 0)
This was just one geodesic! But since R
n
is homogeneous and isotropic, we have in fact determined
all geodesics: The geodesics in Euclidean geometry are the straight lines. For any point not on a
geodesic there is a unique geodesic passing through that point parallel to the given geodesic.
2.60. Spherical geometry. The Riemannian connection on
S
n
R
= (, ) R
n
R [ [[
2
+
2
= R
2
R
n+1
is the tangential connection (2.55) arising from the Euclidean connection on ambient R
n+1
.
Let N = (0, . . . , R) S
n
R
be the North Pole and V = (1, 0, . . . , 0) T
N

n
R
. What is
V
, the
geodesic running through N with velocity V ? Using the isometries that change sign on
i
for
2 i n we see, as above, that
V
must run in the intersection of S
n
R
and the
1
-plane. Thus
(2.61)
V
(t) = (Rsin(t/R), 0, . . . , 0, Rcos(t/R))
We conclude that the geodesics in spherical geometry are great circles, the intersection of S
n
R
with
planes through the origin. For any point not on a geodesic there is a no geodesic passing through
that point parallel to the given geodesic.
8. THE EXPONENTIAL MAP AND NORMAL COORDINATES 21
2.62. Hyperbolic geometry. The Riemannian connection on
H
n
R
= (, ) R
n
R
+
[ [[
2

2
= R
2
R
n+1
is the tangential connection (2.55) arising from the Euclidean connection, considered as the Rieman-
nian connection of Minkowski metric [(, )[
2
= [[
2

2
, on ambient R
n+1
.
Let N = (0, . . . , R) S
n
R
be the North Pole and V = (1, 0, . . . , 0) T
N

n
R
. What is
V
, the
geodesic running through N with velocity V ? Using the isometries that change sign on
i
for
2 i n we see, as above, that
V
must run in the intersection of H
n
R
and the
1
-plane. Thus
(2.63)
V
(t) = (Rsinh(t/R), 0, . . . , 0, Rcosh(t/R))
We conclude that the geodesics in hyperbolic geometry are great hyperbolas, the intersection of
H
n
R
with planes through the origin. (The isometry group O(n, 1)
+
takes planes through the origin,
u

, to planes through the origin.) For any point not on a geodesic there are uncountably many
geodesic passing through that point parallel to the given geodesic.
8. The exponential map and normal coordinates
Let M be a Riemannian manifold with Riemannian connection (2.43). Put
c = V TM [
V
is dened at 1
where
V
, V T
p
M, is the geodesic through (0) = p with velocity

(0) = V (2.39). We dene

(2.64) exp: c M by exp(V ) =
V
(1)
meaning that exp(V ) is obtained by following the geodesic with initial velocity vector V for one
time unit. We let exp
p
denote the restriction of exp to c
p
= c T
p
M.
2.65. Proposition (Properties of the exponential map). Let exp: c M be the exponential
map on the manifold M.
(1) c is an open subspace of TM and exp: c M is a smooth map.
(2) c
p
is star-shaped around 0 for each p M.
(3) exp
p
: c
p
M takes straight lines through 0 T
p
M to geodesics through p: exp
p
(tV ) =

V
(t) (where both functions are dened for the same set of ts).
(4) (exp
p
)

: T
0
T
p
M = T
p
M T
p
M is the identity map.
(5) The exponential map commutes with isometries: The diagram
T
p
M
exp
p

T
p
M
exp
(p)

M
commutes for any isometry : M M.
Proof. We defer the proof of (1). Let V T
p
M and t R. Then
tV c
p
exp
p
is dened at tV
tV
is dened at 1
V
is dened at t
and for such a t, exp
p
(tV ) =
tV
(1) =
V
(t) by the Rescaling lemma (2.40). In particular,
V c
p

V
is dened at 1 =
V
is dened at s sV c
p
when 0 < s 1. Thus c
p
is star-shaped around 0. Assuming that c
p
is open in T
p
M and that exp
p
is smooth we now compute the dierential of exp
p
. Since the tangent vector V T
0
T
p
M = T
p
M
is (2.34) represented by the image curve t tV , (exp
p
)

is represented by curve exp(tV ) =

V
(t)
with

V
(0) = V . Thus (exp
p
)

V = V . If : M M is an isometry, then
V
=

V
(2.58) so
that exp
p
(V ) =
V
(1) =

V
(1) = exp
(p)
(

V ).
The geodesic vector eld G on TM is the dened by
G(V )f =
d
dt

t=0
(f

V
(t)), f C

(TM)
22 2. RIEMANNIAN MANIFOLDS
where

V
(t) is curve on TM obtained by taking the velocity of the geodesic
V
. Note that

V
(t
0
+t) =

V
(t
0
)
(t) so that
G(

V
(t
0
))f =
d
dt

t=0
f(

V
(t +t
0
)) =
d
dt

t=t
0
f(

V
(t))
where t
0
is an arbitrary point in the open interval of denition for
V
. This means that integral
curves for the vector eld G are velocity elds along geodesics. By a general theorem, there exists
a smooth map : O TM, dened on an open subspace O RTM containing 0 TM, such
that (t, V ) is the maximal integral curve for G through V TM at time t = 0. Hence
c = V TM [ (1, V ) O = i
1
1
O, i
1
(V ) = (1, V ),
is open and exp(V ) = (1, V ) is smooth as a composition of smooth maps (: TM M is the
projection).
2.66. Example. Let N = (0, 0, 1) be the North Pole of S
2
R
3
and let V be a unit vector
in the tangent space T
N
S
2
= N

. The exponential map exp

N
: T
N
S
2
= N

S
2
takes the unit
speed radial line tV to the unit speed geodesic whose trace is the intersection of S
2
with the plane
through 0 R
3
containing N and V .
Normal coordinates is special coordinate system determined by the metric.
Let U
p
T
p
M be an open subset, star-shaped around 0, of the tangent space such that
exp
p
: U
p
U is a dieomorphism between U
p
and an open subset U of M. Choose an orthonormal
basis E
i
for T
p
M (with inner product g
p
) and an orthonormal basis e
i
for R
n
(with standard inner
product g). Let E: (R
n
, g) (T
p
M, g
p
) be the isometry given by Ee
i
= E
i
. Then
(2.67) x = E
1
exp
1
p
: M U x(U) R
n
M
x

T
p
M
exp
p

R
n
E

are normal coordinates around p. (We will often forget to mention E so that V can stand for a
tangent vector V T
p
M as well as a vector V R
n
.) The smooth function
(2.68) r : U p R, r(q) = [x(q)[ =
_

x
i
(q)
2
= [V [
g
(exp
p
(V ) = q)
is the radial distance function and the unit radial vector eld on U p, denoted
(2.69)

r
(q), q U p,
is the vector eld formed by the velocity vectors of the unit speed radial geodesic; to any q U p
it associates the velocity of the unit speed radial geodesic through q.
If B
R
(0) U
p
, then
B
R
(p) = exp
p
(B
R
(0)) = q x(U) [ r(q) R
B
R
(p) = exp
p
(B
R
(0)) = q x(U) [ r(q) < R
S
R
(p) = exp
p
(B
R
(0)) = q x(U) [ r(q) = R
is a (closed) geodesic ball , respectively, a geodesic sphere around p. (Make a drawing of the
situation!)
2.70. Proposition (Properties of normal coordinates). Let x: U x(U) be normal coordin-
ates (2.67) around p.
(1) x(
EV
(t)) = tV for all V T
p
M and for all small t. (In normal coordinates, the geodesics
through p are straight lines through 0.)
(2) x(p) = 0,

x
i
(p) = E
i
, g
ij
(p) =
ij
,
k
ij
(p) = 0,
k
g
ij
(p) = 0.
(3)

r
=
x
i
(q)
r(q)

x
i
for all q U p.
Proof. (1) x
1
(tV ) = exp
p
(tEV ) =
EV
(t).
(2) x
1
(0) = exp
p
(0) = p. Let (t) = x
1
(te
i
) be the ith coordinate axis. Then

(t) =

x
i
((t))
in general (2.33). In this case, (t) = x
1
(te
i
) = exp
p
(tE
i
) =
E
i
(t) is the geodesic through
(0) = p with velocity

(0) = E
i
. Thus E
i
=

x
i
(p). The components of the Riemannian metric
8. THE EXPONENTIAL MAP AND NORMAL COORDINATES 23
g at p are g
ij
(p) = g(

x
i
(p),

x
j
(p)) = g(E
i
, E
j
) =
ij
. The radial curve (t) = x
1
(t(e
i
+ e
j
)) is
a geodesic so its coordinates
k
satisfy the ODEs (2.38) which in this particular case means that

k
ij
((t)) = 0 for all t. For t = 0, we get
k
ij
(p) = 0. In other words,

j
(p) = 0 for all i, j. Then
also

k
g
ij
(p) =
k

i
,
j
) (p) =

i
(p),
j
(p)) +
i
(p),

j
(p)) = 0 + 0 = 0
since the Riemannian connection is compatible with the Riemannian metric (2.46).
(3) Let q U p. Suppose that x(q) = V R
n
. Then x
i
(q) = V
i
and r(q) = [x(q)[ = [V [ (2.68).
The unit speed radial geodesic from p to q is (t) = exp
p
(tV/[V [), 0 t [V [. Its velocity vector
is

(t) =
V
i
|V |

x
i
((t)) (2.33); at q, in particular, its velocity is
V
i
|V |

x
i
(q) =
x
i
(q)
r(q)

x
i
(q).
Let : (, ) [a, b] M be a variation of the curve (t) = (0, t). We write (s, t) for points
in (, ) [a, b] (so that the interval [a, b] is placed on the vertical t-axis!). Let
s
(t) = (s, t) =

t
(s) so that
s
is a a curve in the t-direction (a main curve) and
t
is a curve in the s-direction
(a transverse curve). Let

t
(s, t) =
d
dt

s
(t) =

s
(t) =

(

t
),
s
(s, t) =
d
ds

t
(s) =

t
(s) =

(

s
),
be the velocities of the main, respectively, the transverse curves. We may view the main velocity
eld
t
as a vector eld along a transverse curve
t
and consider its covariant derivative D
s

along
t
. Similarly, we may view the transverse velocity eld
s
as a vector eld along a main
curve
s
and consider its covariant derivative D
t

s
along
s
.
2.71. Lemma (Symmetry Lemma). D
s

t
= D
t

s
or D
s

s
= D
t

t
.
Proof. This is a local question. Choose a coordinate system x around (s
0
, t
0
). In local
coordinates, x(s, t) = (
1
(s, t), . . . ,
n
(s, t)) and (2.33)
t
=

s
=

i
t

i
,
s
=

t
=

i
s

i
. In local
coordinates there is a formula for the covariant derivative along a curve (2.36). Using that formula,
we get that
D
s

t
=
_

k
st
+
k
ij

i
s

j
t
_

k
and there is a similar formula for D
t

t
except that the s and t swap places. The point is now
that the Riemannian connection is symmetric so that the Christoel symbols are symmetric in i
and j (2.45).
The variational eld
(2.72) V (t) =

t
(t, 0) =
(0,t)
(

s
)
is the restriction to (t) of the transverse vector eld

t
.
2.73. Lemma (First Variation of smooth curves).
d
ds
L(
s
)(0) =
_
V,

b
a

_
b
a
_
V, D
t

_
dt
when (t) is a unit speed curve.
Proof. We dierentiate the function s L(
s
) and then evaluate the result at s = 0. Using
that the connection is compatible with the metric (2.52) and The Symmetry lemma we get
d
ds
L(
s
) =
d
ds
_
b
a
[

s
(t)[dt =
_
b
a

s
_

s
,

s
_
1/2
dt
2.52
=
_
b
a
1
[

s
[
_
D
s

s
,

s
_
dt
2.71
=
_
b
a
1
[

s
[
_
D
t

t
,

s
_
dt
24 2. RIEMANNIAN MANIFOLDS
When s = 0,

s
=

0
=

, [

[ = 1, and

s
(0, t) = V (t) so that
d
ds
L(
s
)(0) =
_
b
a
_
D
t
V,

_
dt
2.52
=
_
b
a
_
d
dt
_
V,

_
V, D
t

_
_
dt
=
_
V (b),

(b)
_

_
V (a),

(a)
_

_
b
a
_
V, D
t

_
dt

2.74. Lemma (Gauss lemma). In a normal neighborhood of a point p we have:

(1) The geodesic spheres are orthogonal to the geodesic rays.
(2)


r
, Y
_
= Y (r) = dr(Y ) for any tangent vector Y T
q
M, q ,= p (In other words,
grad r =

r
).
Proof. Let q = exp
p
(V ) where V ,= 0. Then x(q) = V and r(q) = [V [ = R. The claim is
that
W V =(exp
p
)

V
(1)
for any W T
V
T
p
M = T
p
M (where V

= T
V
S
R
(0)). Let (s) be a curve in S
R
(0) that represents
W, (0) = V ,

(0) = W. Consider the variation (s, t) = exp
p
(t(s)) of
V
(t) = exp
p
(tV ). Put
S =
s
and T =
t
. Note that the curves in the t-direction, t exp
p
(t(s)), are geodesics of
velocity T and speed [T[ = [(s)[ = R. Hence D
t
T = 0 and [T[ = R is constant. It follows that

t
S, T)
(2.52)
= D
t
S, T) +S, D
t
T) = D
t
S, T)
(2.71)
= D
s
T, T)
(2.52)
=
1
2

s
[T[
2
= 0
where we use that the connection is compatible with the metric and symmetric. Thus
S, T)(0, 0) = S, T)(0, 1)
since S, T)(s, t) is independent of t. We know compute
S(0, 0) =

0
(0) = 0, S(0, 1) =

1
(0) = (exp
p
)

W, T(0, 1) =

V
(1)
as
0
(s) = p,
1
(s) = exp
p
((s)), and
0
(t) = exp
p
(tV ) =
V
(t). It follows that
0 = S, T)(0, 0) = S, T)(0, 1) = (exp
p
)

W,

V
(1))
which is the rst item of the lemma.
We now know that there is an orthogonal decomposition T
q
M = R

r
(q) T
q
S
R
(p) for we
have already seen that

r
(q) is the unit vector proportional to

V
(1). Any Y T
q
M therefore
admits an orthogonal decomposition of the form Y =

r
(q) +X where R and X T
q
S
R
(p)
is tangent to the geodesic sphere. Hence
_

r
(q), Y
_
= [

r
(q)[
2
=
because X(r) = 0 as r is constant on S
R
(p). On the other hand,
Y

(r) = (

r
(q) +X)(r) =

r
(q)(r)
where

r
(q)(r) =
_
x
i
r

i
_
_

(x
i
)
2
_
1/2
=

x
i
r
2x
i
2r
=
r
2
r
2
= 1
and we have proved also the second item of the lemma.
31.03.05
9. THE RIEMANN DISTANCE FUNCTION 25
9. The Riemann distance function
Let M be a connected Riemannian manifold.
2.75. Definition. A regular curve on M is a smooth map : [a, b] M such that

(t) ,= 0
for all t [a, b]. The real number
L =
_
b
a
[

(t)[dt
is the length of the regular curve .
A piecewise regular curve on M is a continuous map : [a, b] M such that [[a
i1
, a
i
] is
regular for some subdivision a = a
0
< a
1
< < a
n
= b of [a, b]. The real number
L =

L([[a
i1
, a
i
])
is the length of the piecewise regular curve .
A piecewise regular curve on M has a velocity vector

(t) at all points t which is not one of

the break points a
i
. At a break point a
i
, we let

(a
+
i
)

(a

i
)
denote the jump between the velocity from the left,

(a

i
) T
(a
i
)
, and from the right,

(a
+
i
).
2.76. Proposition. (1) The length of a (piecewise) regular curve is invariant under re-
parametrization.
(2) Any regular curve has a unit speed parameterization.
Proof. Let : [a, b] M be a regular curve.
(1) Let t : [c, d] [a, b] be a bijective smooth map with t

(s) ,= 0 for all s [c, d]. Then

( t)

(s) = t

(s)

(t(s))
so that
_
d
c
[( t)

(s)[ds =
_
d
c
[

(t(s))[t

(s)ds =
_
b
a
[

(t)[dt
where the + applies if t

> 0 and the applies if t

< 0.
(2) Let s: [a, b] [0, L()] be the smooth map s(t) =
_
t
a
[

(t)[dt. Then s

(t) = [

(t)[ by the
Fundamental theorem of Calculus. Let t be the inverse function. Then ( t)

(s) = t

(s)

(t(s)) =
1
s

(t)

(t) is a unit speed curve.

2.77. Lemma. Any two points in M can be connected by a piecewise regular curve.
Proof. Since connected and locally path-connected spaces are path-connected, M is path-
connected [5, 17]. Given any two points, p and q in M, there exists a continuous curve : [0, 1] M
connecting them. By the Lesbesgue number lemma [5, 19], there is a subdivision 0 = a
0
< a
1
<
< a
n
= 1 of [0, 1] such that ([a
i1
, a
i
]) is contained in a coordinate neighborhood x: U R
n
such that x(U) is a ball. Replace ([a
i1
, a
i
]) by a smooth curve within this coordinate neighbor-
hood between the two end-points.
2.78. Definition. The function d: M M [0, ) given by
d(p, q) = infL() [ is a piecewise regular curve from p to q
is the Riemann distance function.
We all know from general topology that all topological manifolds are metrizable. On a Rieman-
nian manifold we can construct an explicit metric.
2.79. Lemma. d is a metric on the topological space M.
Let : [, ] [a, b] M be a xed endpoint variation of the unit speed piecewise regular
curve (t) = (0, t).
For each s, the length of the piecewise regular curve
s
is L(
s
). What is the rate of change
of the length of the main curves near (hoping that this length function is smooth)?
26 2. RIEMANNIAN MANIFOLDS
Theorem 2.80 (First variation formula for piecewise smooth curves). Let be a unit speed
piecewise regular curve and any piecewise smooth variation of . Then
d
ds

s=0
L(
s
) =
_
b
a
_
V, D
t

_
dt
n1

i=1
_
V (a
i
),
i

_
where V is the variational eld along .
Proof. Since L(
s
) =

L(
s
[[a
i1
, a
i
]) is a sum, just add the contributions from each
subinterval [a
i1
, a
i
] where we are in the smooth situation (2.73). Remember that the endpoints
are xed under the variation so that the variational eld is 0 at the endpoints.
The formula shows that the length decreases when we vary in the direction of the jumps at
the points a
i
or vary is the direction of the acceleration D
t

between the points a

i
.
2.81. Lemma. Any vector eld V (which vanishes at the endpoints) along a piecewise smooth
curve is the variational eld of some (xed endpoint) variation of the curve.
Proof. Thanks to compactness, we can nd > 0 so that V (t) c TM for all t [a, b].
Let (s, t) exp(sV (t)) for (s, t) (, ) [a, b]. Then is a piecewise smooth variation of whose
transverse curves are geodesics with velocity
d
ds

s=0
(s, t) =
d
ds

s=0
exp(sV (t))
2.65
= V (t).
2.82. Corollary. (1) Piecewise regular minimizing curves of constant speed are geodesics.
(2) Geodesics are locally minimizing curves.
Proof of first part of 2.82. Let be a minimizing curve. For any vector eld V along
, the expression on the right hand side of the equation in 2.80 is 0 as L(
s
) has a minimum at
s = 0. Use this to show rst that D
t

(t
0
) = 0 at any point which is not a break point. Thus is
a piecewise geodesic. Next show that there are no break points so that is in fact a geodesic.
What we showed was in fact that geodesics are critical points of the functional L.
Is there a minimizing curve between any two points of M? Are minimizing curves unique?
No, there are uncountably many minimizing curves between the North Pole and the South Pole.
Or look at the situation where you want to go to the other shore of a lake, there are usually two
possibilities. Only if two points are suciently close, then there is in fact a unique minimizing
curve between them.
Theorem 2.83. Let p be a point of M and B
R
(p) a closed geodesic ball around p.
(1) For any point q B
R
(p) in the geodesic ball there is a unique minimizing curve from p to
q, namely the radial geodesic. Then d(p, q) = r(q) where r is the radial distance function
(2.68).
(2) For any point q , B
R
(p) outside the geodesic ball there is a point x S
R
(p) such that
d(p, q) = R +d(x, q). Then d(p, q) > R.
Proof. (1) Suppose that q B
R
(p) lies in the geodesic ball around p and that r(q) = r where
r is the radial distance function (2.68) for that ball. The radial unit speed geodesic from p to q is
(t) = exp
p
(tV ), t [0, r], where V T
p
M is the unit vector with exp
p
(rV ) = q. This curve has
length r so that d(p, q) r.
Now let : [a, b] M be any piecewise regular unit speed curve from p to q. Let a
0
be the
last point where r((t)) = 0 and b
0
the rst point after a
0
such that r((t)) = r. Then [[a
0
, b
0
] Make a drawing!
runs inside the closed geodesic ball of radius r. On the interval (a
0
, b
0
] we decompose the velocity
(2.84)

(t) = (t)

r
((t)) +X(t)
into its radial component along the unit radial vector eld (2.69) and a component X(t)

r
.
Then
(t) =
_

(t),

r
_
2.74
=

(t)(r) =
d
dt
(r), 1 = [

(t)[
2
= [(t)[
2
+[X(t)[
2
[(t)[
2
9. THE RIEMANN DISTANCE FUNCTION 27
for t (a
0
, b
0
] and therefore
b a = L() L([[a
0
, b
0
]) = b
0
a
0
lim
0
_
b
0
a
0
+
(t)dt = lim
0
_
b
0
a
0
+
d
dt
(r)dt
= lim
0
(r(q) r((a
0
+))) = r(q) = r
Since was an arbitrary curve, we have shown that the radial geodesic is minimizing and that
d(p, q) = R.
We will now show that is the only minimizing curve (up to reparametrization). Suppose
that : [0, r] M is some minimizing unit speed curve from p to q. Then the inequalities in the
above computation are in fact equalities so that (t) = 1 for all t. Then X(t) = 0 and (2.84) says
that (t) is an integral curve through p for the radial unit vector eld. So is . Uniqueness of
integral curves implies that (t) = (t) for all t [0, r].
(2) Suppose that q , B
R
(p). By compactness, there is a point x S
R
(p) such that d(x, q)
is minimal. Then d(p, q) d(p, x) + d(x, q) = R + d(x, q). Suppose that d(p, q) < d(p, x) +
d(x, q) = R + d(x, q). Then there exists a piecewise smooth curve connecting p and q of length
L() < R +d(x, q). Let
1
be the rst part of that runs entirely inside the closed geodesic ball
from p to a point on S
R
(p), and let
2
be the last part of that runs entirely outside the closed
geodesic ball from a point of S
R
(p) to q. Then
R +d(x, q) > L() L(
1
) +L(
2
) R +L(
2
)
because L(
1
) R by the rst part of this theorem. Now L(
2
) < d(x, q) so that the start-point
of
2
is a point on S
R
(p) that is closer to q than x. Contradiction!
05.04.05
A smooth curve : I M is locally minimizing if any t
0
I has a neighborhood such that
[[t
1
, t
2
] is minimizing for all t
1
< t
2
in this neighborhood.
Proof of second part of 2.82. Let be a geodesic and (t
0
) a point on . Choose a
uniformly geodesic neighborhood W around (t
0
). The preimage
1
(W) is a union of open
intervals. Let I
0
I be the interval containing t
0
. If t
1
, t
2
I
0
then [[t
1
, t
2
] is geodesic in
W B

((t
1
)) through (t
1
) so it is a radial geodesic in B

((t
1
)), hence (2.83) minimizing.
Also when we go beyond where exp
p
is injective we can sometimes nd minimizing curves,
they may no longer be unique, though. (Look at curves on S
2
from N to S.)
2.85. Lemma. Suppose that there is a point p M such that the exponential map at p is dened
on the whole tangent space T
p
M. Then there is a minimizing curve, of the form t exp
p
(tV ),
0 t d(p, q) for some unit vector V T
p
M, from p to any other point in M.
Proof. Let q be some point dierent from p and let T = d(p, q) > 0 be the distanec between
p and q.
Choose a closed geodesic ball, B
R
(p), around p. We may assume that q is outside this ball, ie
T > R, for otherwise we already know that there exists a minimizing curve from p to q (2.83). Let
x be a point on S
R
(p) that realizes the distance between S
R
(p) and q and let be the unit speed
radial geodesic from p through x. By assumption, is dened for all t 0. The miracle is that
goes through q: (T) = q.
To see this, consider the set
o = b [0, T] [ d(p, q) = d(p, (b)) +d((b), q)
By using the continuity of the distance function d one can show that o is closed (take a sequence
of points in o). From 2.83 we know that [0, R] o. Let A = sup o and put y = (A). Then
T = d(p, q) = d(p, y) + d(y, q) = A + d(y, q) as A o. Suppose that A < T. Choose a closed
geodesic ball B

(y) around y where 0 < < T A. Let z be the point on the geodesic sphere
S

(y) such that d(y, q) = +d(z, q) (2.83) and let be the unique radial unit speed geodesic from
y to z. The piecewise smooth curve [[0, A] from p to z has length A+ and as
d(p, z) d(p, q) d(z, q) = T (d(y, q) ) = T (T A) = A+
28 2. RIEMANNIAN MANIFOLDS
it is minimizing, hence a geodesic (2.82), in particular a smooth curve with no breaks. Thus
[[0, A] and must t together to form the geodesic [[0, A +] by uniqueness of geodesics. Now
z = (A+) and d(p, z) = A+ so that
d(p, z) +d(z, q) = (A+) + (T A) = T = d(p, q)
which contradicts that A is the supremum of o.
Now we know that T o so that
T = d(p, q) = d(p, (T)) +d((T), q) = L([[0, T]) +d((T), q) = T +d((T), q)
so that d((T), q) = 0 and (T) = q.
Now comes a denition that will only be used for a very short time!
2.86. Definition. A Riemannian manifold is geodesically complete if all maximal geodesics
are dened for all of R.
07.04.05
R
2
is geodesically complete, B
2
1
(0) is not.
Recall that a metric space is complete if all Cauchy sequences converge. Compact metric
spaces are complete.
Theorem 2.87 (HopfRinow). M is complete as a metric space M is geodesically
complete.
Proof. Suppose that M is complete as a metric space. The claim is that all geodesics are
dened for all time. Suppose that there is some unit speed maximal geodesic : (a, b) M that
cannot be extended beyond b. Let t
i
(a, b) be an increasing sequence of points converging to
b. Then (t
i
) is Cauchy for d((t
i
), (t
j
)) L([[t
i
, t
j
]) = [t
i
t
j
[. Let q = limq(t
i
). Choose a
uniformly geodesic neighborhood W around q and a > 0 such that W B

(p) for all p W.

This means that any unit speed geodesic through a point of W exists at least in a time span of .
Choose t
j
so that t
j
> b and q(t
j
) W. Then we can extend the geodesic near t
j
for at least
time beyond t
j
. Contradiction!
Next sssume that there is a point p M such that exp
p
is dened for all T
p
M. Let (q
i
) be a
Cauchy sequence. We can assume that p ,= q
i
for all i (if not, throw away some of the q
i
). Choose
V
i
T
p
M such that t exp
p
(tV
i
), 0 t 1, is a minimizing radial geodesic from p to q
i
(2.85).
Then d(p, q
i
) =
_
1
0
[V
i
[dt = [V
i
[. Since Cauchy sequences are bounded, the sequence (V
i
) is bounded
in T
p
M. Any bounded sequence in the inner product space T
p
M contains a convergent subsequence
by compactness. Suppose that V
i
k
V T
p
M. Then q
i
k
= exp
p
(V
i
k
) exp
p
(V ) M by
continuity of exp
p
. Any Cauchy sequence containing a convergent subsequence is itself convergent.
Thus limq
i
= exp
p
(V ).
In the future we will not bother to say geodesically complete but just say complete.
We actually proved that the conditions
(1) M is metric complete
(2) All maximal geodesics in M are dened on R
(3) All maximal geodesics through one point of M are dened on R
are equivalent and that any of these conditions imply
(4) There is a minimizing curve between any two points of M.
The HeineBorel theorem holds in any complete Riemannian manifold.
CHAPTER 3
Curvature
1. The Riemann curvature tensor
The Riemann curvature tensor is the obstruction to atness.
3.1. Definition. The Riemann curvature tensor is the (5.6) (3, 1)-tensor R(X, Y, Z, ) =
(R(X, Y )Z) corresponding to the C

(M)-multilinear map
T(M) T(M) T(M)
R
T(M), (X, Y, Z) R(X, Y )Z :=
X

Y
Z
Y

X
Z
[X,Y ]
Z
(in a somewhat unorthodox notation) called the Riemann curvature endomorphism.
3.2. Proposition. The map from 3.1 is C

(M)-multilinear.
Proof. The function is clearly R-multilinear.
Note that [fX, Y ] = f[X, Y ] Y (f)X by computing [fX, Y ](g). The computation

fX

Y
Z
Y

fX
Z
[fX,Y ]
Z = f
X

Y
Z
Y
(f
X
Z) f
[X,Y ]
Z +Y (f)
X
Z
= f
X

Y
ZY (f)
X
Zf
Y

X
Zf
[X,Y ]
Z+Y (f)
X
Z = f(
X

Y
Z
Y

X
Z
[X,Y ]
Z)
shows that the function is C

(M)-linear in the X-variable.

The function is anti-symmetric in the X and Y -variables, so it is also C

(M)-linear in the
Y -variable.
A direct computation as above shows that is also C

(M)-linear in the Z-variable.

Equivalently, the Riemann curvature tensor is the (4, 0)-tensor eld Rm = R

given by (5.8)
Rm(X, Y, Z, W) = R(W, Y, Z, W

) = W

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

In local coordinates (x
i
), the components of the curvature tensors
R = R

ijk
dx
i
dx
j
dx
k

, Rm = R
ijk
dx
i
dx
j
dx
k
dx

are given by
R(
i
,
j
)
k
= R

ijk

so that
(3.3) R
ijk
= R(
i
,
j
)
k
,

) =
_
R
m
ijk

m
,

_
= g
m
R
m
ijk
, R
m
ijk
= g
m
R
ijk
3.4. Example. In Euclidean geometry (R
n
, g),
X
(Y
k

k
) = X(Y
k
)
k
, and (2.22) shows that
R(
i
,
j
)
k
= 0 as the basis vector elds commute, [
i
,
j
] = 0. Thus R = 0 on R
n
. On S
2
with
spherical coordinates the curvature tensor is nonzero in that R(

.
3.5. Lemma (Symmetries in the Riemann curvature tensor). Let R and Rm be the curvature
tensor of a Riemannian manifold.
(1) Rm is anti-symmetric in the rst two variables: Rm(X, Y, Z, W) = Rm(Y, X, Z, W)
(2) Rm is anti-symmetric in the last two variables: Rm(X, Y, Z, W) = Rm(X, Y, W, Z)
(3) Rm is symmetric between the rst two variables and the last two variables: Rm(X, Y, Z, W) =
Rm(Z, W, X, Y )
(4) Rm satises a cyclic permutation property of the rst three variables:
Rm(X, Y, Z, W) + Rm(Z, X, Y, W) + Rm(Y, Z, X, W) = 0
known as the First Bianchi identity.
29
30 3. CURVATURE
(5) Rm satises a cyclic permutation property of the last three variables:
Rm(X, Y, Z, V, W) +Rm(X, Y, W, Z, V ) +Rm(X, Y, V, W, Z) = 0
known as the Second or Dierential Bianchi identity.
In a local coordinate frame these identities are equivalent to
(1) R
ijk
= R
jik
(2) R
ijk
= R
ijk
(3) R
ijk
= R
kij
(4) R
ijk
+R
kij
+R
jki
= 0
(5) R
ijk;m
+R
ijmk;
+R
ijm;k
= 0
Proof. Each item is a more or less clever calculation.
(1) Clear.
(2) It is enough to show that Rm(X, Y, Z, Z) = 0 as we see by expanding Rm(X, Y, Z +W, Z +W)
using multilinearity. Now
XY [Z[
2
Y X[Z[
2
= [X, Y ][Z[
2
2.46
2(
X

Y
Z, Z) +
Y
Z,
X
Z)) 2(
Y

X
Z, Z) +
X
Z,
Y
Z)) = 2

[X,Y ]
Z, Z
_

X

Y
Z, Z)
Y

X
Z, Z) =

[X,Y ]
Z, Z
_
Rm(X, Y, Z, Z) = 0

3.6. Proposition. The Riemann curvature tensor Rm is invariant under local isometries.
Proof. If : M

M is a (local) isometry then

Rm(

X,

Y,

Z,

W) = Rm(X, Y, Z, W)
by (2.57.(3)).
3.7. Example. The curvature tensor in Euclidean geometry is R = 0. The curvature tensor
of S
2
is not zero as, for instance, R(

so that Rm(

) = g
11
= R
2
sin
2

using spherical coordinates as in [4, Exercise 5.7]. Thus S

2
is not locally isometric to R
2
.
A Riemannian manifold is at if any point has a neighborhood isometric to an open subspace
of R
n
.
Theorem 3.8. M is at R = 0.
Proof. We have already established one direction. What remains is to show that if R = 0
is a neighborhood of a point p then there are coordinates (y
i
) near p so that g
ij
=
ij
in these
coordinates. Since this is a local question we may as well assume that M = R
n
and p = 0.
Put E
i
=
i
(0) so that (E
1
, . . . , E
n
) is the standard orthonormal basis for T
0
R
n
.
First, we extend E
j
to a vector eld on R
n
. Let E
j
be the unique parallel vector eld along
the x
1
axis t (t, 0, . . . , 0) with E
j
(0) = E
j
. Next, for each xed x
1
0
, let E
j
be the unique parallel
vector eld along the line t (x
1
0
, t, 0, . . . , 0) satisfying the initial condition that at t = 0 it is
E
j
(x
1
0
, t, 0, . . . , 0). The vector eld E
j
is now dened in the x
1
x
2
-plane. Continue this way. The
result is a smooth vector eld on R
n
.
R = 0 =E
j
is parallel. By construction, E
j
is parallel along the x
1
-axis. Thus

1
E
j
= 0 at
any point on the x
1
-axis. By construction, E
j
is parallel along the lines (x
1
0
, x
2
, 0 . . . , 0, . . . , 0).
Thus

2
E
j
= 0 at any point in the x
1
x
2
-plane. Also the vector eld

1
E
j
is parallel along the
lines (x
1
0
, x
2
, 0 . . . , 0, . . . , 0), < x
2
< , for

1
E
j
=

2
E
j
=

1
0 = 0
since R = 0 and [
1
,
2
] = 0. Thus

1
E
j
is the parallel vector eld along this line with value

1
E
j
(x
1
, 0, . . . , 0) = 0 at x
2
= 0. That vector eld is the zero vector eld. We conclude that

1
E
j
= 0 at all points in the x
1
x
2
-plane. Continue this way and conclude that

1
E
j
=
0, . . . ,

n
E
j
= 0 at all points in x
1
x
2
x
n
-space. Thus
X
E
j
= 0 for any vector eld X by
C

(R
n
)-linearity.
2. RICCI CURVATURE, SCALAR CURVATURE, AND EINSTEIN METRICS 31
Compatibility = (E
1
, . . . , E
n
) is an orthonormal frame. Since the Riemannian connection is com-
patible with the inner product, parallel translation preserves the inner product (2.52): For a point
in the x
1
x
2
-plane for instance

E
i
(x
1
, x
2
, 0, . . . , 0), E
j
(x
1
, x
2
, 0, . . . , 0)
_
=

E
i
(x
1
, , 0, . . . , 0), E
j
(x
1
, 0, . . . , 0)
_
= E
i
(0, . . . , 0), E
j
(0, . . . , 0)) = E
i
, E
j
) =
ij
In general, we see in this way that (E
1
, . . . , E
n
) is an orthonormal moving frame.
Symmetry = the vector elds E
1
, . . . , E
n
commute. We have [E
i
, E
j
] =
E
i
E
j

E
j
E
i
= 0
since the Riemannian connection is symmetric (2.44).
Finally, a theorem of elementary dierential geometry says that the vector elds E
i
are co-
ordinate vector elds for some coordinate system (y
i
) near 0. Thus g
ij
=
ij
for this coordinate
system.
14.04.05
2. Ricci curvature, scalar curvature, and Einstein metrics
3.9. Definition. Ricci curvature is the trace on the rst and last variable of the Riemann
curvature endomorphism: Rc = tr(R) = tr
g
(Rm) T
2
0
(M).
This simply means that Ricci curvature is the tensor given by
Rc(X, Y ) = tr (U R(U, X)Y )
Ricci curvature is a (2, 0)-tensor with components
Rc
ij
= R
kij
k
= g
k
R
kij
= g
k
R
kij
= R

ij
= R
k
ijk
The symmetries of the Riemann curvature give
Rc
ij
= R
kij
k
= R
j
k
ki
= R
jk
k
i
Rc
ij
= R
kij
k
= R
ikj
k
= R
ki
k
j
Rc
ij
= R
j
k
ki
= R
k
jik
= Rc
ji
where the last line means that the Ricci tensor is symmetric.
3.10. Definition. Scalar curvature is the trace with respect to g (5.26) of the Ricci curvature:
S = tr
g
Rc.
Scalar curvature is the smooth function on M given by
S = Rc
j
j
= R
ij
ji
3.11. Definition. The divergence operators div are the maps
T
0

(M)

T
1

(M)
tr
T
0
1
(M), T
k
0
(M)

T
k+1
0
(M)
tr
g
T
k1
0
(M)
where in the last case the trace is taken with respect to the covariant dierentiation index and some
other lower index.
For X T
0

(M), div(X)
j
1
j
1
= X
j
1
j
1
i
;i
and for X T
k
0
(M), div(X)
i
1
i
k1
= X
i
1
i
k1
j;
j
.
For example
the divergence of a vector eld X = X
j

j
T
0
1
(M) is the smooth function div(X) = X
i
;i
the divergence of the Ricci curvature tensor Rc T
2
0
(M) is the covector eld div(Rc)
T
1
0
(M) with components
div(Rc)
m
= Rc
mj;
j
= R
imj
i j
;
We compute this tensor below (3.12).
the divergence of the metric g is div(g) = tr
g
(g) = tr
g
(0) = 0 because g = 0 as the
connection and the metric are compatible (2.52).
the divergence of the product Sg = S g of the scalar curvature and the metric is
div(Sg) = tr
g
((Sg)) = tr
g
(Sg+Sg) = tr
g
(Sg) = S
;j
g
i
j
= S
;j

j
i
= S
;i
= S
because (S g)
mij
= S
;m
g
ij
.
3.12. Lemma (Contracted Bianchi identity). div(Rc) =
1
2
S or S
;m
= 2Rc
mj;
j
= 2R
imj
i j
;
.
32 3. CURVATURE
Proof. Start with the Dierential Bianchi identity
R
ijk;m
+R
ijmk;
+R
ijm;k
= 0
and take the g-trace over i and ,
R
ijk ;m
i
+R
ijmk;
i
+R
i
ij m;k
= 0
and take the g-trace over j and k,
R
ji
ij ;m
+R
ijm
j i
;
+R
i j
ij m;
= 0
The rst term we recognize as S
;m
. The second term is R
j i
jim ;
= div(Rc)
m
. The third term is
R
i j
ijm ;
= div(Rc)
m
.
19.04.05
3.13. Definition. The metric g is an Einstein metric if it is proportional to its Ricci curvature
at any point: Rc = g for some smooth function on M.
In fact, the function is implicitly given by the Einstein equation:
Rc = g =tr
g
(Rc) = tr
g
(g) S = n =
1
n
S
because tr
g
(g) = n = dimM (5.26).
3.14. Proposition. Any connected Riemannian manifold of dimension > 2 with an Einstein
metric has constant scalar curvature.
Proof. Rc =
1
n
S g =div(Rc) =
1
n
div(S g)
1
2
S =
1
n
S
n>2
=S = 0. so that S
is constant on each component of M.
3.15. Example (Curvature of surfaces). Let M be a Riemannian manifold of dimension 2, a
Riemann surface. Let K =
1
2
S denote the function that is half of the scalar cuvature (we shall
later call it the Gaussian curvature of the surface). Let (E
1
, E
2
) be an orthonormal basis for the
tangent space T
p
M at some point p of M.
Riemann curvature: The Riemann curvature tensor has 2
4
= 16 components R
ijk
= Rm(E
i
, E
j
, E
k
, E

),
1 i, j 2. However, for (anti-)symmetry reasons R
iik
= 0 = R
ijkk
, so that
R
1221
= R
2112
= R
2121
= R
1212
are the only nonzero components.
Ricci curvature: The components of Rc = tr
g
(R) are Rc
ij
= R
1ij
1
+ R
2ij
2
= R
1ij1
+ R
2ij2
. Here
we use that R
1ij
1
= g
1k
R
1ijk
=
k
1
R
1ijk
= R
1ij1
since the basis is orthonormal so that the matrix
for g and its inverse are identity matrices. Hence
Rc
11
= R
2112
= R
1221
Rc
12
= 0
Rc
21
= 0 Rc
22
= R
1221
are the components of Ricci curvature.
Scalar curvature: S = tr
g
(Rc) = Rc
1
1
+ Rc
2
2
= Rc
11
+ Rc
22
= 2R
1221
. so that R
1221
=
1
2
S = K.
We conclude that K = R
1221
= Rc
11
= Rc
22
so that scalar curvature determines Ricci and
Riemann curvature for surfaces.
Until now we have been working with an orthonormal basis. Let us now consider an arbitrary
basis (X, Y ) for T
p
M. Then
E
1
=
X
[X[
, E
2
=
Y
_
Y,
X
|X|
_
X
|X|
[Y
_
Y,
X
|X|
_
X
|X|
[
is the orthonormal basis obtained by applying the GramSchmidt process to (X, Y ). From the
above computations
Rc(X, Y ) = Rc(X
1
E
1
+X
2
E
2
, Y
1
E
1
+Y
2
E
2
) = X
1
Y
1
Rc
11
+X
2
Y
2
Rc
22
= KX
1
Y
1
+KX
2
Y
2
= KX, Y )
Rm(X, Y, Z, W) = K(X, W) Y, Z) X, Z) Y, W))
3. Riemannian SUBMANIFOLDS 33
For the last equation, note that both sides are multilinear and equal in case X, Y, Z, W E
1
, E
2
;
try for instance (X, Y, Z, W) = (E
1
, E
2
, E
2
, E
1
). In the special case where Z = Y and W = X this
equation gives that
(3.16) K =
Rm(X, Y, Y, X)
[X[
2
[Y [
2
X, Y )
2
The denominator here is the area of the parallelogram spanned by vectors X and Y .
See A compendium of Surfaces in [8] for much more information about surfaces.
3. Riemannian submanifolds
Let (

M, g) be a Riemannian manifold and M

M an embedded submanifold equipped with
the induced metric, also called g. The 2nd fundamental form of the Riemannian submanifold M
is the dierence between the Riemannian connections

and . (The 1st fundamental form is the
metric g.)
The ambient tangent bundle T

M[M M splits
T

M[M = TM NM
into the orthogonal direct sum of the tangent bundle of M with the normal bundle NM M.
Any section of T

M[M splits orthogonally into a direct sum of its tangential and normal part.
If X, Y T (M) are smooth vector elds on M, then

X
Y is a well-dened (2.56) section of
T

M[M. We already know (2.55) that

X
Y is the tangential component of

X
Y . If we write
II(X, Y ) for the normal component then the orthogonal splitting of

X
Y has the form
(3.17)

X
Y = (

X
Y )
T
+ (

X
Y )

=
X
Y
. .
T
p
M
+II(X, Y )
. .
N
p
M
(Gauss formula)
where the normal component II(X, Y ) NM is called the second fundamental form. Equivalently,
II(X, Y ) =

X
Y
X
Y
is the dierence between the extrinsic connection

and the intrinsic connection .
3.18. Lemma. Let M

M be a Riemannian submanifold and II : T (M) T (M) ^(M)
its second fundamental form.
(1) II C

(M)-bilinear and symmetric.

(2) If X, Y T (M) are vector elds and N ^(M) a normal eld on M then
(3.19)
_

X
N, Y
_
= N, II(X, Y )) (Weingarten equation)
(3) If X, Y, Z, W T (M) are vector elds on M then
(3.20)

Rm(X, Y, Z, W) = Rm(X, Y, Z, W) II(X, W), II(Y, Z)) +II(X, Z), II(Y, W))
(Gauss equation)
(4) Let : I M be a curve in M and V a vector eld in M along . Then
(3.21)

D
t
V = D
t
V +II(

, V ) (Gauss formula along a curve)

Proof. (2)
_

X
Y, N
_
+
_
Y,

X
N
_
=

X
N, Y ) =

X
(0) = 0 where
_

X
Y, N
_
= II(X, Y ), N).
(3) Riemann curvature computed in ambient space is

Rm(X, Y, Z, W) =
_

Y
Z

X
Z

[X,Y ]
Z, W
_
Observe that
_

[X,Y ]
Z, W
_
=

[X,Y ]
Z, W
_
because the normal part of the rst vector does not
contribute to the inner product with a vector tangent to M. Observe also that
_

Y
Z, W
_
=
_

X
(
Y
Z +II(Y, Z)), W
_
=
_

Y
Z, W
_
+
_

X
II(Y, Z), W
_
=
X

Y
, W) II(Y, Z), II(X, W))
34 3. CURVATURE
It now follows that

Rm(X, Y, Z, W) = Rm(X, Y, Z, W) II(Y, Z), II(X, W)) +II(X, Z), II(Y, W))
which is Gausss equation.
3.21. Remark (Gauss formula for the velocity eld along a curve). If we apply the Gauss
formula along a curve to the special case where V =

is the velocity eld then we get that

D
t

= D
t

+II(

)
and we see that
If is a geodesic in M, then

D
t

= II(

). Thus II(V, V ), V T
p
M, is the acceleration
at p in

M of the geodesic
V
in M.
If the curve in M is a geodesic in

M, then is also a geodesic in M and II(

) = 0.
Thus II(V, V ) = 0 if the geodesic
V
in

M, V T
p
M T
p

M, happens to stay inside M.

Since the second fundamental form is a symmetric bilinear form is it completely determined by its
quadratic function V II(V, V ).
3.22. Gaussian and mean curvature of codimension one Euclidean embeddings.
We consider the simplest case of a Riemannian submanifold, namely that of (an orientable) hyper-
surface in Euclidean space, M
n
R
n+1
. We shall associate curvature to the embedding.
Choose a normal eld N that is nonzero at every point of M. (This is possible if M is
orientable; in any case it is possible to choose such a normal eld locally.) Using N, we may write
the second fundamental form II : T (M) T (M) ^(M) as
II(X, Y ) = h(X, Y )N so that

X
Y, N
_
= II(X, Y ), N) = [N[
2
h(X, Y )
where h T
2
0
(M) is symmetric (2, 0)-tensor on M, the scalar second fundamental form. In this case
of a codimension one embedding into Euclidean space R
n+1
the Gauss and Weingarten formulas
specialize to

X
Y =
T
p
M
..

X
Y +
N
p
M
..
h(X, Y )N (Gauss formula) (3.23)
[N[
2
h(X, Y ) =

X
Y, N
_
=

X
N, Y
_
(Weingarten equation for N) (3.24)
Rm(X, Y, Z, W) = [N[
2
(h(X, W)h(Y, Z) h(X, Z)h(Y, W)) (Gauss equation) (3.25)
For the Gauss equation (3.25) note that

Rm = 0 in Euclidean space and that the inner product
II(X, W), II(Y, Z)) = h(X, W)N, h(Y, Z)N) = [N[
2
h(X, W)h(Y, Z).
The shape operator s T
1
1
(M) is dened by
(3.26) X, Y T (M): sX, Y ) = [N[h(X, Y ) =
1
[N[

X
Y, N
_
=
1
[N[

X
N, Y
_
meaning that s

= [N[h or s = [N[h

(5.12) is obtained from h by raising an index. The shape

operator
sX =
1
[N[

X
N
informs about the shape of M since it measures the variation of the normal eld N as it moves on
M. The shape operator is self-adjoint,
sX, Y ) = [N[h(X, Y ) = [N[h(Y, X) = sY, X) = X, sY )
because h is symmetric. Therefore T
p
M has an orthonormal basis E
1
, . . . , E
n
of eigenvectors,
sE
i
=
i
E
i
, for s.
3.27. Definition. The principal directions and the principal curvatures of the embedding
M

M are the orthonormal eigenvectors and the eigenvalues of the shape operator. The Gauss
curvature and the mean curvature of the embedding M

M is the determinant, K = det s =

i
,
and 1/nth of the trace, H =
1
n
tr s =
1
n

i
, of s.
3. Riemannian SUBMANIFOLDS 35
Principal curvatures are invariants of the embedding and not invariants of the manifold; see
[4, p 56] for examples. It therefore came as a total surprise that the product of the two principal
curvatures of an embedded surface does not depend on the embedding but only on the surface
itself.
3.28. Remark (The shape operator in a local frame). Suppose that (E
i
) is a local frame. The
components of s = [N[h

are (5.27)
s
i
j
= [N[h
i
j
= [N[g
jk
h
ik
and we obtain the formulas
H =
1
n
tr s =
1
n
[N[g
ij
h
ij
, K = det s = [N[
n
det(h
ij
)
det(g
ij
)
for the mean and Gaussian curvature. If N is a unit normal eld, the Gaussin curvature is the
determinant of the 2nd fundamental form relative to the determinant of the 1st fundamental form.
Inserting h
ij
= h(E
i
, E
j
) =
1
|N|
2

E
i
E
j
, N
_
from the Weingarten equation (3.24) we can also
write
H =
1
n[N[
g
ij

E
i
E
j
, N
_
, K =
1
[N[
n
det(

E
i
E
j
, N
_
)
det(g
ij
)
for the mean and Gaussian curvature.
Theorem 3.29 (Theorema Egregium, Gauss 1828). The Gaussian curvature of a Riemannian
embedding M
2
R
3
of a Riemannian surface in R
3
does not depend on the embedding but only
on the Riemann surface itself. (In fact, the Gaussian curvature of the embedding equals half the
scalar curvature of the surface.)
Proof. Choose an orthonormal local frame (E
1
, E
2
) and a unit normal eld N for M (to
make life a little easier). Then the Gaussian curvature of the Riemannian submanifold M
K
3.28
= det(h
ij
) = h(E
1
, E
1
)h(E
2
, E
2
) h(E
1
, E
2
)
2
(3.25)
= Rm(E
1
, E
2
, E
2
, E
1
)
3.15
=
1
2
S
equals half of the scalar curvature of the Riemannian manifold M.
In Gauss original formulation the theorem goes something like
If an area in E
3
can be developed (i.e. mapped isometrically) into another area
of E
3
, the values of the Gaussian curvatures are identical in corresponding points
Gauss received a prize from the University of Copenhagen for this theorem.
We can therefore speak of the Gaussian curvature of an (orientable) Riemann surface (as we
did in (3.15)).
3.30. Example (The shape of a parameterized surface in R
3
). Let X(u, v) be a parameteriz-
ation of a surface in M R
3
. The vectors X
u
= X

(
u
) =
X
i
u

i
and X
v
= X

(
v
) =
X
i
v

i
(1.2)
form a basis for the tangent space of the surface. In this basis, the metric, the 1st fundamental
form, is
g =
_
X
u
, X
u
) X
u
, X
v
)
X
v
, X
u
) X
v
, X
v
)
_
, g
1
=
1
det g
_
X
v
, X
v
) X
u
, X
v
)
X
v
, X
u
) X
u
, X
u
)
_
,
The cross product N = X
u
X
v
is an everywhere nonzero normal eld. The vector X
uu
=
X
u
X
u
,
X
uu
=
X
u
X
u
=
X
u
(
X
i
u

i
)
(2.22)
= X
u
(
X
i
u
)
i
(2.34)
=

2
X
i

2
u

i
is simply obtained by dierentiating each of the three coordinate functions in X twice wrt u. The
2nd fundamental form is
h =
1
2[N[
_
X
uu
, N) X
uv
, N)
X
vu
, N) X
vv
, N)
_
36 3. CURVATURE
The formulas from 3.28 take the form
H =
1
2[N[
X
v
, X
v
) X
uu
, N) 2 X
u
, X
v
) X
uv
, N) +X
u
, X
u
) X
vv
, N)
X
u
, X
u
) X
v
, X
v
) X
u
, X
v
)
2
K =
1
[N[
2
X
uu
, N) X
vv
, N) X
uv
, N)
2
X
u
, X
u
) X
v
, X
v
) X
u
, X
v
)
2
In particular, if M is the graph of the function f, then X(u, v) = (u, v, f(u, v)), X
u
= (1, 0, f
u
),
X
v
= (0, 1, f
v
), X
uu
= (0, 0, f
uu
), X
uv
= (0, 0, f
uv
), X
vv
= (0, 0, f
vv
). As N = X
u
X
v
=
(f
u
, f
v
, 1) the formula gives the equation
H =
f
uu
(1 +f
2
v
) +f
vv
(1 +f
2
v
) 2f
uv
f
u
f
v
2(1 +f
2
u
+f
2
v
)
3/2
, K =
f
uu
f
vv
f
2
uv
(1 +f
2
u
+f
2
v
)
2
for the mean and Gaussian curvature of a graph of a function of two variables. We see that a graph
is isometric to a plane if and only if f
uu
f
vv
= f
uv
f
uv
. (The next example is a generalization.)
3.31. Example (The shape of a graph). Let M R
n
be the graph of the smooth function
f : R
n
R. We shall write f
i
for the ith partial derivative f/u
i
and f
ij
for
2
f/u
i
u
j
. The
embedding X(u) = (u, f(u)), u R
n
, is a parameterization of the graph. As in (3.30) the vectors
X
i
=
i
+f
i

n
= (0, . . . , 0, 1, 0, . . . , 0, f
i
), 1 i n,
form a basis for the tangent space T
p
M of the graph. Thus g
ij
=

X
i
, X
j
_
=
ij
+f
i
f
j
and
N =

f
i

i
+
n
= (f
1
, f
2
, . . . , f
n
, 1)
is a nonzero normal eld. Now,
X
i X
j
= f
ij

n
as in (3.30) so that

X
i X
j
, N
_
= f
ij
. From
(3.28) we get the expressions
H =
f
ij
g
ij
2
_
1 +

i
f
2
i
, K =
1
[N[
n
det(f
ij
)
det(g
ij
)
=
det(f
ij
)
(1 +

i
f
2
i
)
n/2
det(
ij
+f
i
f
j
)
for the mean and the Gaussian curvature of a graph.
4. Sectional curvature
Let (M, g) be a Riemannian manifold. We shall give a geometric interpretation of the Riemann
curvature Rm tensor of M.
Let p be a point of M. For each 2-dimensional subspace of the tangent space T
p
M, let
S

M be the surface in M that is the image under exp

p
of , or rather the image of the
part of where exp
p
is a dieomorphism, S

= exp
p
(
p
). We give S

the induced metric

so that S

M is a Riemannian embedding. Note that the tangent space of S

is T
p
S

=
T
p
exp
p
(
p
) = (exp
p
)

T
0
= T
0
= T
p
M (2.34, 2.65).
3.32. Definition. Sectional curvature at p M is the function that to any tangent plane
T
p
M to the manifold associates the Gaussian curvature K() = K(S

)
p
at p of the Riemannian
surface S

.
3.33. Proposition. The sectional curvature of the tangent plane T
p
M is
K() =
Rm(X, Y, Y, X)
[X[
2
[Y [
2
X, Y )
2
where X, Y is any basis for .
Proof. For any V T
p
S

= , the geodesic
V
in M runs inside S

. Gauss formula for the

velocity of a curve (3.21) implies that II(V, V ) = 0. Since the second fundamental form vanishes
at p, Gauss equation (3.20) says that Rm(X, Y, Z, W) is the Riemann curvature tensor for the
surface S

at p for all X, Y, Z, W T
p
S

. Now (3.16) tells us about the Gaussian curvature of S

at p.
4. SECTIONAL CURVATURE 37
In particular, if the oriented Riemannian manifold M
n
R
n+1
sits as an embedded codi-
mension one manifold in R
n+1
with Euclidean or Minkowski metric then (3.33, 3.25) the sectional
curvature is
(3.34) K(X, Y ) = [N[
2
h(X, X)h(Y, Y ) h(X, Y )
2
g(X, X)g(Y, Y ) g(X, Y )
2
=
1
[N[
2

X
X, N
_

Y
Y, N
_

X
Y, N
_
2
X, X) Y, Y ) X, Y )
2
where N is a nowhere zero normal eld.
3.35. Proposition. The sectional curvatures determine the Riemann curvature tensor of M.
Exactly how you get Riemann curvature out of sectional curvature is a bit complicated [1,
(1.10)]. It is easier to explain how you get Ricci and scalar curvature.
Let E
1
T
p
M be a unit vector. Expand (E
1
) to an orthonormal basis E
1
, E
2
, . . . , E
n
for T
p
M.
Then
Rc(E
1
, E
1
) = Rc
11
=

R
k11
k
=

k
R
k11k
=

k
Rm(E
1
, E
k
, E
k
, E
1
) =

k>1
K(E
1
, E
k
)
S = Rc
k
k
=

k
Rc
kk
=

j,k
R
jkkj
=

j,k
Rm(E
j
, E
k
, E
k
, E
j
) =

j=k
K(E
j
, E
k
)
where we write K(X, Y ) for the Gaussian curvature K(spanX, Y ) of the plane spanned by the
independent vectors X and Y .
We shall next see that the three model geometries, Euclidean, spherical, and hyperbolic, are
Riemannian manifolds of constant sectional curvature, space forms.
3.36. Euclidean geometry. In Euclidean geometry (2.1, 2.59) the curvature tensor Rm = 0
(3.4) so that the sectional curvature K() = 0 for any plane T
p
R
n
at any point p R
n
.
3.37. Spherical geometry. In spherical geometry (2.2, 2.60), the tangent space T
p
S
n
R
= p

(1.5) so that the normal space N

p
S
n
R
= Rp is the line through p. Thus N
p
=
1
R
p is the outward
pointing unit normal vector eld. For any point p S
n
R
and any unit tangent vector V T
p
S
n
R
,
Gauss formula for the velocity eld of a curve (3.21) tells us that
II(V, V ) =
1
R
N and h(V, V ) =
1
R
as this is the acceleration of the geodesic through p with unit speed V , the great circle through
p tangent to V p

. Consider for instance the geodesic through (0, . . . , 0, R) with unit speed

(0) = (1, 0, . . . , 0) as in (2.61). Its acceleration is

(0) =
1
R
2
(0) =
1
R
(
1
R
(0)) =
1
R
N((0))
Since h(V, V ) =
1
R
for all unit tangent vectors V , h(U, V ) = 0 for any orthonormal pair U, V of
tangent vectors. Formula (3.34) tells us that the sectional curvature of S
n
R
K() =
1
R
2
is constant. 27.04.05
3.38. Hyperbolic geometry. In hyperbolic geometry (2.14, 2.62), the tangent space T
p
H
n
R
=
p

(2.15) so that the normal space N

p
H
n
R
= Rp is the line through p. Thus N
p
=
1
R
p is a normal
vector eld of constant square length [N[
2
= 1. For any point p H
n
R
and any unit tangent
vector X T
p
H
n
R
, Gauss formula for the velocity eld of a curve (3.21) tells us that
II(V, V ) =
1
R
N and h(V, V ) =
1
R
as this is the acceleration of the geodesic through p with unit speed V , the great hyperbola through
p tangent to V p

. Consider for instance the geodesic through (0, . . . , 0, R) with unit speed

(0) = (1, 0, . . . , 0) as in (2.63). Its acceleration is

(0) =
1
R
2
(0) =
1
R
(
1
R
(0)) =
1
R
N((0))
38 3. CURVATURE
Since h(V, V ) =
1
R
for all unit tangent vectors V , h(U, V ) = 0 for any orthonormal pair U, V of
tangent vectors. Formula (3.34) tells us that the sectional curvature of H
n
R
K() =
1
R
2
is constant.
For space forms there are explicit formulas for the curvature tensors.
3.39. Proposition. In a Riemannian manifold with constant sectional curvature C,
R(X, Y )Z = C [Y, Z) X X, Z) Y ]
Rm(X, Y, Z, W) = C [X, W) Y, Z) X, Z) Y, W)]
Rc(X, Y ) = (n 1)C X, Y )
S = n(n 1)C
Proof. The rst and the second equation are equivalent. Both sides of the second equation
are tensors with the symmetry properties of the curvature tensor. Thus they are equal if they
are equal for (X, Y, Z, W) = (X, Y, Y, X) with orthonormal X, Y (3.35). In that case the LHS is
K(X, Y ) = C and the RHS is C.
Both sides of the third equation are symmetric tensors. Thus they are equal if they are equal
for (X, Y ) = (X, X) with X a unit vector. In that case the LHS is (n 1)C by the formula below
3.35 and the RHS is also (n 1)C.
The formula below 3.35 says that constant sectional curvature C implies constant scalar
curvature of value n(n 1)C.
3.40. Example (The shape of the sphere). Since we know the 2nd fundamental form of
the sphere (3.37) we also know the shape operator. In fact, the shape operator (3.26) for the
codimension one Euclidean embedding S
n
R
R
n+1
is the symmetric isomorphism
s: T
p
S
n
R
= p

T
p
S
n
R
= p

, s(V ) =
1
R
V
for this is clearly self-adjoint and its bilinear form, or quadratic function, is given by sV, V ) =

1
R
= h(V, V ) when [V [ = 1. Thus the principal curvatures all equal
1
R
and the mean and
Gaussian curvatures of the embedding S
n
R
R
n+1
are
H =
1
n
tr(s) =
1
R
, K = det(s) =
(1)
n
R
n
The sign depends on the choice of unit normal eld.
5. Jacobi elds
Let : [a, b] M be a geodesic from (a) = p to (b) = q.
3.41. Definition. A smooth vector eld J along is a Jacobi eld if
D
t
D
t
J +R(J,

= 0 (Jacobi equation)
3.42. Proposition (Existence and uniqueness of Jacobi elds). There is an isomorphism of
vector spaces
Jacobi elds along T
p
M T
p
M: J (J(a), D
t
J(a))
Proof. The claim is that for any given X, Y T
p
M there exists a unique Jacobi eld J along
such that J(p) = X and D
t
J(p) = Y . Let (E
i
(t)) be an orthonormal frame of parallel vector
elds along . Write J = J
i
E
i
. The vector eld J is Jacobi if and only if
i : D
t
D
t
J, E
i
) =
_
R(J,

, E
i
_
The LHS is
D
t
D
t
J, E
i
) =
d
2
dt
2
J, E
i
) =
d
2
J
i
dt
2
E
i
and the RHS is

_
R(J,

, E
i
_
=
_
R(E
j
,

, E
i
_
J
j
= Rm(E
j
,

, E
i
)J
j
5. JACOBI FIELDS 39
In matrix notation the requirement is
d
2
dt
2
_
_
_
J
1
.
.
.
J
n
_
_
_ =
_
(Rm(E
j
,

, E
i
)
_
_
_
_
J
1
.
.
.
J
n
_
_
_
This 2nd order linear system of ODEs has a unique solution with the given initial conditions.
3.43. Lemma. Suppose that V is smooth vector eld along a smooth variation of through
geodesics : (, ) [a, b] M. Then
D
s
D
t
V D
t
D
s
V = R(
s
,
t
)V
Proof. Choose local coordinates x: M R
n
. Let x(s, t) = (
1
(s, t), . . . ,
n
(s, t)) and V =
V
i
(s, t)
i
be the local expressions for and V . By properties of the covariant derivative along a
curve (2.36)
D
t
V = D
t
(V
i

i
) =
V
i
t

i
+V
i
D
t

i
and
D
s
D
t
V = D
s
(
V
i
t

i
+V
i
D
t

i
) = D
s
(
V
i
t

i
) +D
s
(V
i
D
t

i
)
=

2
V
i
st

i
+
V
i
t
D
s

i
+
V
i
s
D
t

i
+V
i
D
s
D
t

i
When we compute the dierence D
s
D
t
V D
t
D
s
V many terms cancel and we are left with
D
s
D
t
V D
t
D
s
V = V
i
(D
s
D
t
D
t
D
s
)
i
Since the vector eld
i
and its covariant derivative are extendible,
V
i
(D
s
D
t
D
t
D
s
)
i
= V
i
(

i
) = V
i
(

)
i
= R(
s
,
t
)V
where the last equality uses that [
s
,
t
] = [

(
s
),

(
t
)] =

[
s
,
t
] =

(0) = 0.
3.44. Proposition.
Jacobi elds along = Variational elds of smooth variations of through geodesics
Proof. We shall only prove sup. Let : (, ) [a, b] M be a smooth variation of
through geodesics. Lemma 3.43 applied to the vector eld
t
along says that
D
s
D
t

t
D
t
D
s

t
= R(
s
,
t
)
t

In the rst term, D

t

t
= 0 as each main curve is a geodesic. In the second term, D
s

t
= D
t

by the Symmetry Lemma 2.71. Thus the equation says that

D
t
D
t

s
+R(
s
,
t
)
t
= 0
At s = 0,
s
= V is the variational eld and
t
=

is the velocity eld of so that we have the

Jacobi equation D
t
D
t
V +R(V,

= 0.
The opposite inclusion is also true [1], but we shall need it here.
For each Y T
p
M, let J
Y
be the Jacobi eld along with J
Y
(a) = 0 and D
t
J
Y
(a) = 0. Here
is an explicit description of J
Y
.
3.45. Proposition (Jacobi elds vanishing at p). Let (t) = exp
p
(tT), T T
p
M, [T[ = 1, be
the unit speed parametrization of (dened for t such that tT
p
T
p
M). For any Y T
p
M,
J
Y
(t) = t(exp
p
)
tT
(Y )
Proof. The variation (s, t) = exp
p
(t(T +sY )) is a smooth variation of through the radial
geodesics t exp
p
(t(T + sY )). Its variational eld J is therefore a Jacobi eld. The tangent
vector J(t) at (t) is represented by the curve
s exp
p
(tT +stY )
and hence J(t) = (exp
p
)
tT
(tY ) = t(exp
p
)
tT
(Y ) = tU(t) and D
t
J = D
t
(tU) = U + tD
t
U where
U(t) = (exp
p
)
tT
(Y ). In particular, J(0) = 0 and D
t
J(0) = U(0)Y = (exp
p
)

(Y ) = Y .
40 3. CURVATURE
A Jacobi eld J along is normal if
_
J(t),

(t)
_
= 0 for all t.
3.46. Lemma. A Jacobi eld J along is normal i J(0) and D
t
J(0) are orhtogonal to

(0)
i J vanishes at two points.
Proof. The point is that
_
J,

_
is an ane function, t at +b, of t because
d
2
dt
2
_
J,

_
=
d
dt
_
D
t
J,

_
=
_
D
t
D
t
J,

_
=
_
R(J,

_
= Rm(J,

) = 0
by anti-symmetry of the Riemann curvature tensor. If such a function vanishes at two points or
itself and its derivative vanishes at one point, then it is identically zero.
The vector space of normal Jacobi elds is isomorphic to

(0)

(0)

of dimension 2n 2.
The vector space of normal Jacobi elds J along with J(p) = 0 is isomorphic to

(0)

of
dimension n 1.
The Jacobi elds that are not normal are linear combinations of the two Jacobi elds J
0
=

and J
1
= tJ
0
with J
0
(0) =

(0), D
t
J
0
(0) = 0 and J
1
(0) = 0, D
t
J
1
(0) =

(0) and D
t
D
t
J
0
= 0 =
D
t
D
t
J
1
.
3.47. Conjugate points. We say that two points on a geodesic segment are conjugate if
there exists a nonzero Jacobi (necessarily normal) eld along that vanishes at p and q. The
dimension of the vector space of all such Jacobi elds is the multiplicity of conjugacy.
3.48. Corollary. Consider the smooth function exp
p
:
p
M. Let q = exp
p
(rT) ,= p be a
point on the geodesic (t) = exp
p
(tT). Then
exp
p
is not a local dieomorphism at rT p and q are conjugate points
Proof. The Jacobi elds that vanish at p are J
Y
= t(exp
p
)
tT
(Y ) (3.45). The value of such
a Jacobi eld at q = exp
p
(rT) is J
Y
(r) = t(exp
p
)
tT
(Y ). Thus there is nonzero Jacobi eld that
vanishes at p and q i there is a nonzero vector in the kernel of (exp
p
)
rT
.
3.49. Jacobi elds in constant sectional curvature manifolds. Jacobi elds have par-
ticularly simple descriptions in Riemannian manifolds with constant sectional curvature.
3.50. Proposition (Normal Jacobi elds in constant sectional curvature manifolds). Suppose
that M is a Riemannian manifold of constant sectional curvature C. Let be a unit speed geodesic
in M with (0) = p and

(0) = T where T is a unit vector in T

p
M. For any Y T

, the normal
Jacobi eld J
Y
is
J
Y
(t) = u
C
(t)Y (t)
where Y (t) is the parallel vector eld along with Y (0) = Y and
u
C
(t) =
_

_
Rsin(t/R) C = 1/R
2
t C = 0
Rsinh(t/R) C = 1/R
2
is the solution to

u
C
+Cu
C
= 0, u
C
(0) = 0,

u
C
(0) = 1.
Proof. Put J(t) = u
C
(t)Y (t). Then D
t
J = u

C
Y and D
t
D
t
J = u

C
Y = CJ as Y (t) is
parallel. Thus J(0) = 0 and D
t
J(0) = Y . By construction, J is orthogonal to (2.52). It is a
Jacobi eld because (3.39)
R(J,

= C
_

)J J,

_
= CJ = D
t
D
t
J
as has unit speed and J is normal to . We conclude that J = J
Y
.
Look at concrete Jacobi elds along geodesics in R
n
and S
n
R
obtained by variations through
geodesics.
5. JACOBI FIELDS 41
3.51. Proposition (Space form metrics in normal coordinates). Let M be a Riemannian mani-
fold of constant sectional curvature C and p a point of M. Consider the smooth map exp
p
:
p
M.
Let T T
p
M be a unit vector and r > 0 a positive scalar such that rT
p
. Put q = exp
p
(rT).
The induced isomorphism (exp
p
)
rT
: T
p
M T
q
M satisies
[(exp
p
)
rT
Y [
2
q
= [Y

[
2
p
+
u
C
(r)
2
r
2
[Y
T
[
2
p
, Y T
rT
T
p
M = T
p
M
where Y = Y

+Y
T
are the components of Y orthogonal and tangent to the sphere through rT.
Proof. Let (t) = exp
p
(tT) be the unit speed geodesic from p to q.
The Gauss lemma (2.74) says that
[(exp
p
)
rT
Y [
2
q
= [(exp
p
)
rT
(Y

+Y
T
)[
2
q
= [(exp
p
)
rT
(Y

)[
2
q
+[(exp
p
)
rT
(Y
T
)[
2
q
as the components (exp
p
)
rT
(Y

) and (exp
p
)
rT
(Y
T
) are orthogonal.
The radial component Y

equals T for some > 0. Thus (exp

p
)
rT
Y

is represented by the
radial curve
s exp
p
(rT +sT) = (r +s)
and therefore (exp
p
)
rT
(Y

) =

(r) has length [[ = [T[

p
= [Y

[
p
. Thus (exp
p
)
rT
preserves
length in the radial direction.
The normal Jacobi eld J = J1
r
Y
T (3.45) has value 0 at p and value J(r) = r(exp
p
)
rT
(
1
r
Y
T
) =
(exp
p
)
rT
(Y
T
) at q. Now,
J(t) =
1
r
u
C
(t)Y
T
(t)
according to the description of normal Jacobi elds in manifolds of constant sectional curvature
(3.50) so that we may compute the length of J at any point (t) on as
[J(t)[
2
(t)
=
u
C
(t)
2
r
2
[Y
T
[
2
p
for the length of any parallel vector eld is constant (2.52). In particular, the length of J at q is
[(exp)
rT
(Y
T
)[
2
q
= [J(r)[
2
q
=
u
C
(r)
2
r
2
[Y
T
[
2
p
as asserted.
The above proposition says that constant sectional curvature metrics expressed in normal
coordinates (2.67)
M T
p
M
exp
p

R
n

only depend on the function u

C
so only depend on C.
3.52. Corollary (Local uniqueness of manifolds of constant sectional curvature). Let M
0
and M
1
be two Riemannian manifolds of constant sectional curvature C. Any two points, p
0
M
0
and p
1
M
1
, have isometric neighborhoods.
Proof. An isometry can be constructed as the composite in
(T
p
0
M
0
, g
0
)
exp
0

(R
n
, g
n
)

=

(T
p
1
M
1
, g
1
)
exp
1

M
0
_ _ _ _ _ _ _ _ _ _ _ _ _
M
1
where the upper horizontal maps are linear isometries and the other maps are understood to be
dened only locally and to be dieomorphisms. Since a space form metric in a normal neighborhood
around a point is determined by the the inner product at the tangent space of the point, this is an
isometry.
42 3. CURVATURE
3.53. Second variation formula. For any variation of the geodesic (t) = (0, t) we know
that
is minimizing =
d
ds
L(
s
)(0) = 0 and
d
2
ds
2
L(
s
)(0) 0
But what is the second derivative?
3.54. Lemma. Let : [a, b] M be a unit speed geodesic and : (, ) [a, b] M a smooth
variation of . Then
d
2
ds
2
L(
s
)(0) =
_
b
a
_
[D
t
V

[
2
Rm(V

, V

)
_
dt +
_
D
s

s
,

b
a
where V

is the normal part of V .

Proof. In the proof of 2.73 we saw that the rst derivative of the curve length function is
d
ds
L(
s
) =
_
b
a
D
t

s
,
t
)
[
t
[
dt =
_
b
a
D
t

s
,
t
)
t
,
t
)
1/2
dt
Dierentiating once more we get that the second derivative is
d
2
ds
2
L(
s
) =
_
b
a
d
ds
_
D
t

s
,
t
)
t
,
t
)
1/2
_
dt
=
_
b
a
_
D
s
D
t

s
,
t
)
[
t
[
+
D
t

s
, D
s

t
)
[
t
[

D
t

s
,
t
) D
s

t
,
t
)
[
t
[
3
_
dt
where
D
s
D
t

s
= D
t
D
s

s
+R(
s
,
t
)
s
(3.43)
D
s

t
= D
t

s
(2.71)
At s = 0, [
t
[ = [

[ = 1, and
s
= V so that
d
2
ds
2
L(
s
)(0) =
_
b
a
_
_
D
t
D
s

s
,

_
Rm(V,

, V ) +[D
t
V [
2

_
D
t
V,

_
2
_
dt
In the rst term we still use
s
and not V . Since is a geodesic and D
t

= 0, the rst term is

_
b
a
_
D
t
D
s

s
,

_
dt =
_
b
a
d
dt
_
D
s

s
,

_
dt =
_
D
s

s
,

b
a
Now look at the last two terms. Split V = V
T
+V

into the tangential and the normal part. The

tangential component, V
T
=
_
V,

, has covariant derivative

D
t
(V
T
) = D
t
(
_
V,

) =
_
D
t
V,

= (D
t
V )
T
and therefore also D
t
(V

) = (D
t
V )

. It follows that
[D
t
V [
2
= [(D
t
V )
T
[
2
+[(D
t
V )

[
2
=
_
D
t
V,

_
2
+[D
t
V

[
2
or [D
t
V [
2

_
D
t
V,

_
2
= [D
t
V

[
2
. Finally we note that the tangential part of V does not contribute
to Rm(V,

, V ) because Rm(

, V ) = 0 = Rm(V,

).
Theorem 3.55 (Second Variational Formula). Let : [a, b] M be a unit speed geodesic and
: (, ) [a, b] M a piecewise smooth proper variation of . Then
d
2
ds
2
L(
s
)(0) =
_
b
a
_
[D
t
V

[
2
Rm(V

, V

)
_
dt
where V

is the normal part of V .

5. JACOBI FIELDS 43
Proof. We appply the formula from 3.54 to each of the rectangels (, ) [a
i1
, a
i
] where
is smooth. This gives that
d
2
ds
2
L(
s
)(0) equals
_
b
a
_
[D
t
V

[
2
Rm(V

, V

)
_
dt +
n

i=1
_
D
s

s
(a

i
),

(a
i
)
_

_
D
s

s
(a
+
i1
),

(a
i
)
_
Note that D
s

s
(a

i
) = D
s

s
(a
+
i
) because this vector eld only depends on along the line t = a
i
.
Thus the contributions from the points a
i
telescope to
_
D
s

s
(b),

(b)
_

_
D
s

s
(a),

(a)
_
which
is 0 since the variation keeps the end-points xed.
The expression in the Second Variational Formula is the quadratic function I(V

, V

) of the
bilienar form (the index form)
(3.56) I(V, W) =
_
b
a
_
D
t
V, D
t
W) Rm(V,

, W)
_
dt
dened for all piecewise smooth normal proper vector elds V and W along . If V and W happen
to be smooth then the integrand is
d
dt
D
t
V, W) D
t
D
t
V, W) Rm(V,

, W) =
_
D
t
D
t
V +R(V,

, W
_
+
d
dt
D
t
V, W)
so that the index form may be written
I(V, W) =
_
b
a
_
D
t
D
t
V +R(V,

, W
_
dt +D
t
V, W)

b
a
In general, there may be points a = a
0
< a
1
< a
k
= b where V and W are not smooth and then
(3.57) I(V, W) =
_
b
a
_
D
t
D
t
V +R(V,

, W
_
dt
k1

i=1

D
t
V (a
+
i
) D
t
V (a

i
), W(a
i
)
_
which shows the connection to the Jacobi equation.
3.58. Corollary. If is a unit speed minimizing curve then is a geodesic and I(V, V ) 0
for any picewise smooth proper normal vector eld V along .
Proof. Any such V is the variational eld for some proper variation of (2.81). Since is
minimizing,
d
ds
L(
s
)(0) = 0 so that is a unit speed geodesic (2.82) and
d
2
ds
2
L(
s
)(0) = I(V, V ) 0
by the Second Variation Formula (3.55).
Theorem 3.59. Let be a geodesic segment from p to q. If contains an interior point
conjugate to p, then is not minimizing.
Proof. It suces to nd a proper normal vector eld V along with index form I(V, V ) < 0.
By assumption there exists a nontrivial Jacobi eld J vanishing at p and at some interior point
(c). The Jacobi eld J is normal since it vanishes at two points (3.46), and D
t
J(c) ,= 0 as J is
not the zero eld (3.42). Extend J to a piecewise smooth normal vector eld by J = 0 from (c)
to q.
Now choose a smooth normal proper vector eld W along such that W(c) = D
t
J(c

)
(by using a smooth bump function). Since J is a piecewise smooth Jacobi eld I(J, J) = 0 and
I(J, W) = [D
t
J(c

)[
2
(3.57). For any R we have
I(J +W, J +W) = 2[D
t
J(c

)[
2
+
2
I(W, W)
which is negative when is close to 0.
44 3. CURVATURE
6. Comparison theorems
Theorem 3.60 (Jacobi eld Comparison). Suppose that the sectional curvature is bounded
from above by some constant C, K C, in the manifold M. For T, Y T
p
M, T Y , let be the
unit speed geodesic through p M in direction T T
p
M and J
Y
the normal Jacobi eld along
with J
Y
(p) = 0 and D
t
J
Y
(p) = Y . Then the inequality
[J
Y
(t)[ u
C
(t)[Y [
holds for t [0, R] if C = 1/R
2
> 0 and for all t 0 if C 0. (The Jacobi eld is longer than
it would have been had the sectional curvature been constant.)
Proof. We may as well assume that Y is a unit vector. The claim is then that [J(t)[ [u
C
(t)[
for certain values of t where J is short for J
Y
. The function u
C
is the solution to

u +Cu = 0 so
we want to apply the Sturm comparison theorem with u = u
C
and v = [J[. We need to verify that
v(0) = 0, that v is dierentiable at 0 with

v (0) = 1, and that
d
2
dt
2
[J(t)[ +C[J(t)[ 0 in an interval
[0, T].
v(0) = 0: This is clear because J(0) = 0.

v (0) = 0: It is not a priori clear that the function v = [J, J[ = J, J)

1/2
is dierentiable at 0. But
J(t) = tU(t) for a certain vector eld U (3.45) so that the dierence quotient
[J(t)[ [J(0)[
t 0
= [U(t)[, t > 0,
converges to [U(0)[ = [D
t
J(0)[ = [Y [ = 1 for t 0
+
. Thus v is dierentiable at 0 and

v (0) = 1.
Since v(0) = 0 and

v (0) = 1 > 0, v(t) = [J(t)[ is postive on some interval (0, T). In this
interval v(t) = [J(t)[ = J, J)
1/2
is smooth and we can compute its derivatives.

v +Cv 0: We compute (using the Schwartz inequality and the Jacobi equation)
d
2
dt
2
[J[ =
d
dt
_
D
t
J, J)
1
[J[
_
=

D
2
t
J, J
_
+D
t
J, D
t
J)
[J[

D
t
J, J)
2
[J[
3
=
Rm(J,

, J)
[J[
+
[D
t
J[
2
[J[

D
t
J, J)
2
[J[
3

Rm(J,

, J)
[J[
+
[D
t
J[
2
[J[

[D
t
J[
2
[J[
2
[J[
3
=
Rm(J,

, J)
[J[
=
Rm(J,

, J)
[J[
2
[J[ = K(J,

)[J[
C[J[
and obtain the required estimate.
We can therefore conclude from the Sturm comparison theorem that [J(t)[ [u
C
(t)[ on [0, T].
If C 0, the function u
C
has no zeros on (0, ), so [J(t)[ can not become 0 either on this interval
and we can take T = . If C = 1/R
2
> 0, then [J(t)[ can not attain a zero before R and we can
take T = R.
3.61. Corollary (Conjugate Point Comparison Theorem). Suppose that the sectional curvature
is bounded from above by some constant C, K C, in the manifold M. If C = 1/R
2
> 0, then
there are no points conjugate to p in the geodesic ball B
R
(p) of radius R. If C 0, then there
are no points at all conjugate to p.
Theorem 3.62 (Metric Comparison Theorem). Suppose that K C in the manifold M. Then
we have that
[(exp
p
)
rT
Y [
2
q
[Y

[
2
p
+
u
C
(r)
2
r
2
[Y
T
[
2
p
, Y T
rT
T
p
M = T
p
M
in the situation of 3.51. If C 0, then
[(exp
p
)
rT
Y [
2
q
[Y [
2
p
meaning that exp
p
does not decrease length.
6. COMPARISON THEOREMS 45
Proof. The proof is the same as in 3.51 execpt that at some point we must replace an equality
by the inequality from 3.60.
Theorem 3.63 (CartanHadamard). Let M be a connected, complete Riemannian manifold
of nonpositive sectional curvature. Then exp
p
: T
p
M M is the universal covering space of M
(for any point p M). In particular, M is a K(, 1), and if M is simply connected, then
exp
p
: T
p
M M is a dieomorphism.
3.64. Lemma. Let :

M M be local isometry between two connected Riemannian manifolds.
If

M is complete, then is a covering map.
Proof of Theorem 3.63. The exponential map exp
p
is dened on all of T
p
M because M
is complete (2.87); it is a local dieomorphism since p has no conjugate points (3.48, 3.61). If we
let g = exp

p
(g) be the pull-back metric, then exp
p
is a local isometry (T
p
M, g) (M, g). The
Riemannian manifold (T
p
M, g) is complete because the straight line t tY , Y T
p
M, dened
for all t, is a geodesic as its image under the local isometry exp
p
is a geodesic. Lemma 3.64 now
implies that exp
p
is a covering map.
CHAPTER 4
Space-times
Halls book on relativity [2] is in the library. Hawking and Ellis [3] or ONeill [7] are also a
possibilities.
Let now M be a 4-dimensional smooth manifold with a Lorentz metric. Einsteins gravitational
tensor is
G = Rc
1
2
Sg = Rc
1
2
tr
g
(Rc)g
The g-trace of G is
tr
g
(G) = tr
g
(Rc)
1
2
4S = S 2S = S
since tr
g
(g) = 4 in this case (5.26). Therefore
Rc = G+
1
2
Sg = G
1
2
(S)g = G
1
2
tr
g
(G)g
and we see that G and Rc determine each other, they contain the same information. The gravita-
tional tensor has zero divergence (3.11),
div(G) = div(Rc
1
2
Sg) =
1
2
S
1
2
S = 0
as div(Rc) =
1
2
S by the Contracted Bianchi identity (3.12).
Einstein says that gravitation is not a force but a geometrical property of spacetime. A particle
in free fall, under the inuence of gravity alone, will follow a geodesic in spacetime. When you
throw a stone it will trace out a geodesic in spacetime. Einsteins equation
T =
1
8
G
says that the stress-energy tensor T of matter is proportional to the gravitational tensor. This
is not a mathematical statement but (an assertion about) a law of nature just like Newtons law
of inertia. Einsteins equation tells how matter generates Ricci curvature of space-time. The
equation div(T) = 0 tells how Ricci curvature moves matter. There is no denition of matter it
is something that has stressenergy ... Check out
Baez and Bunn: The meaning of Einsteins equation
Sean M. Carroll: Lecture Notes on General Relativity
Cliord M. Will: The Confrontation between General Relativity and Experiment.
Apparently T is determined by the distribution of mass/energy and the equation then expresses
the impact of matter on the Ricci curvature of spacetime [7].
47
CHAPTER 5
Multilinear Algebra
We review some vector space constructions from linear algebra.
1. Tensors
An endomorphism of a vector space V is a special case of a tensor on V . Thus tensors are
generalized endomorphisms or, in coordinates, generalized matrices. The formal denition goes as
follows.
Let V be a nite dimensonal real vector space and V

= Hom(V, R) the dual space of linear

forms on V .
5.1. Lemma. There are natural isomorphisms between the following three vector spaces
Bilinear maps V

V
A
R
Linear maps V

B
Hom(V, R)
Linear maps V
C
Hom(V

, R)
given by A(, v) = B()(v) = C(v)(). In particular, the evaluation morphism ev: V

V R,
given by ev(, v) = (v), corresponds to the identity V

= Hom(V, R) and to the isomorphism

V Hom(V

, R) given by v ( (v)).
5.2. Definition. The
_
k

_
-tensor space of V is the vector space T
k

(V ) of all R-multilinear
homomorphisms
V

. .

V V
. .
k
R
An element of T
k

(V ) is called a
_
k

_
-tensor.
The tensor product of A T
k
1

1
(V ) and B T
k
2

2
(V ), is the multilinear map AB T
k
1
+k
2

1
+
2
(V )
given as the composite
V

. .

1
+
2
V V
. .
k
1
+k
2
= V

. .

1
V V
. .
k
1
V

. .

2
V V
. .
k
2
AB
RR

R
where is multiplication in the ring R.
A
_
k

_
-tensor A T
k

(V ) eats covectors
1
, . . . ,

and k vectors v
1
, . . . , v
k
V and
spits out a real number A(
1
, . . . ,

, v
1
, . . . , v
k
) R.
Some special cases of tensor spaces are
T
0
0
(V ) = R
T
1
0
(V ) = Hom(V, R) = V

T
0
1
(V ) = Hom(V

, R)
5.1

= V
T
1
1
(V ) = V

V
A
R = V Hom(V

, R) = V
s
V = Hom(V, V ), A(, v) =
(sv).
T
k

(V )

= T

k
(V

)
T
k
0
is a contravariant and T
0

a covariant endofunctor of the category of nite dimensional

real vector spaces.
49
50 5. MULTILINEAR ALGEBRA
There is a bilinear map
T
0
1
(V ) T
1
0
(V ) = V V

T
1
1
(V ) = V

V R
given by (u )(, v) = (u)(v). Note that if E
i
is a basis for V and
j
the dual basis for V

,
then E
i

j
is a basis for T
1
1
(V ) = Hom(V, V ).
Guided by the principle that tensors are generalized endomorphisms we dene, more generally,
the tensor product as the bilinear map
T
k
1

1
(V ) T
k
2

2
(V )

T
k
1
+k
2

1
+
2
(V )
(A, B) AB
given by
(AB)(
1
, . . . ,

1
,

1
+1
, . . . ,

1
+
2
, v
1
, . . . , v
k
1
, v
k
1
+1
, . . . , v
k
1
+k
2
)
= A(
1
, . . . ,

1
, v
1
, . . . , v
k
1
)B(

1
+1
, . . . ,

1
+
2
, v
k
1
+1
, . . . , v
k
1
+k
2
)
5.3. Lemma. dimT
k

(V ) = (dimV )
k+
Proof. Let E
1
, . . . , E
n
be a basis for V and
1
, . . . ,
n
the dual basis for V

given by
i
(E
j
) =

j
i
, n = dimV . Then the set
(5.4) E
j
1
E
j

i
1

i
k
[ 1 i
1
, . . . , i
k
, j
1
, . . . , j

n
is a basis for T
k

(V ): These multilinear maps are clearly linearly independent and since

(5.5) A =

A
j
1
j

i
1
i
k
E
j
1
E
j

i
1

i
k
, A
j
1
j

i
1
i
k
= A(
j
1
, . . . ,
j

, E
i
1
, . . . , E
i
k
)
for any multilinear map A T
k

(V ), they generate T
k

(A)
The convention is that vectors have a lower index so that components of vectors have an upper
index: V = V
j
E
j
. Similarly, covectors have an upper index and their components have a lower
index: =
i

i
.
Now comes a more natural way of considering tensors (as generalized endomorphisms). We
have already seen the special case of T
1
1
(V ) = Hom(V, V ). In general, we may view T
k
+1
(V ) as the
vector space of all multilinear maps (V

V
k
V .
5.6. Lemma. There is an isomorphismbetween the vector space T
k
+1
(V ) of all multilinear
homomorphisms
V

. .
+1
V V
. .
k
A
R
and the vector space of all multilinear homomorphisms
V

. .

V V
. .
k
B
V
given by
A(,
1
, . . . ,

, v
1
, . . . , v
k
) = B(
1
, . . . ,

, v
1
, . . . , v
k
)
Proof. As in 5.1 there is an isomorphism bewteen
the vector space of multilinear maps V

(V

V
k
A
R and
the vector space of multilinear maps (V

V
k
B
Hom(V

, R)

= V .
given by A(, , v) = B(, v).
The trace homomorphism tr: T
1
1
(V ) = Hom(V, V ) R = T
0
0
(V ) is the unique homomorphism
that makes the diagram
V V

ev

G
G
G
G
G
G
G
G
G

T
1
1
(V )
tr
.y
y
y
y
y
y
y
y
R
2. TENSORS OF INNER PRODUCT SPACES 51
commutative; it is given by tr(v ) = (v). For eample, tr(id) = dimV . More generally, we
dene the tr: T
k+1
+1
T
k

(V ) to be the unique homomorphism that makes the diagram

V T
k

(V ) V

ev

M
M
M
M
M
M
M
M
M
M
M
M

T
k+1
+1
(V )
tr
.w
w
w
w
w
w
w
w
w
R
commutative; it is given by tr(v A) = (v)A for all tensors A T
k

(V ).
5.7. Lemma (Contraction of tensors). For each k and there exits a unique R-linear map
tr: T
k+1
+1
(V ) T
k

(V )
such that tr(v A) = (v)A for all v V = T
0
1
(V ), A T
k

(V ) and V

= T
1
0
(V ).
Proof. The only possibility is the linear map whose values on the basis (5.4) is
tr(E
j
E
j
1
E
j

i
1

i
k

i
) =
j
i
E
j
1
E
j

i
1

i
k
and this map actually has the property because
tr(v A) = tr(v
i
E
i
A
j

j
) =
j
v
i

j
i
A = (v)A
for all v V and V

.
The tensor algebra of V is the graded algebra T

(V ) =

k=0
T
k
(V ) with tensor product as
product. T

is a contravariant functor.
2. Tensors of inner product spaces
If V has an inner product then there is an isomorphism
T
1
1
(V ) = Hom(V, V )

=
V V R = T
2
0
(V )
under which the endomorphism s: V V and the bilinear map h: V V R correspond to each
other if su, v) = h(u, v). Since tensors are generalized endomorphisms, there should be a similar
correspondence for all tensor spaces.
Suppose that V has an inner product g = , ) T
2
(V ). The inner product induces isomorph-
isms

v,

_
= (v) V

(u) = u, v)
or, more generally,
V

. .

V V V V
. .
k

. .

V V V
. .
k

inverse to each other. Composition with these isomorphism induces isomorphism of tensor spaces
T
k+1

(V )

T
k
+1
(V )

given by
A

(
1
, . . . ,

, v
1
, v
2
, . . . , v
k+1
) = A(
1
, . . . ,

, v

1
, v
2
, . . . , v
k+1
), A T
k
+1
(V )
B

(
1
, . . . ,

,
+1
, v
1
, . . . , v
k
) = B(
1
, . . . ,

, (
+1
)

, v
1
, . . . , v
k
), B T
k+1

(V )
and again inverse to each other.
If we consider the tensor space T
k
+1
(V ) as the vector space of all multilinear maps (V

V
k

V as in (5.6) (and now act on the last variable) then

(5.8) A

(
1
, . . . ,

, v
1
, . . . , v
k
, v
k+1
) = A(
1
, . . . ,

, v
1
, . . . , v
k
, v

k+1
)
= v

k+1
A(
1
, . . . ,

, v
1
, . . . , v
k
) =

A(
1
, . . . ,

, v
1
, . . . , v
k
), v
k+1
_
52 5. MULTILINEAR ALGEBRA
and, equivalently,
(5.9) B(
1
, . . . ,

, v
1
, . . . , v
k
, v
k+1
) = (B

(
1
, . . . ,

, v
1
, . . . , v
k
, v
k+1
)

(
1
, . . . ,

, v
1
, . . . , v
k
), v
k+1
_
for B T
k+1

(V ). In other words, A and B correspond to each other under the isomorphisms

(V

V
k+1
R = T
k+1

(V )

T
k
+1
(V ) = (V

V
k+1
V

if and only if
(5.10) A(
1
, . . . ,

, v
1
, . . . , v
k+1
) =

B(
1
, . . . ,

, v
1
, . . . , v
k
), v
k+1
_
5.11. Definition. The trace with respect to g is the linear map
tr
g
: T
k+1

(V )

T
k
+1
(V )
tr
T
k1

(V )
where we dualize the rst V -variable and take the trace with respect to the new V

and the next

V -variable.
5.12. Example. The (2, 0)-tensor h T
2
0
(V ) and the (1, 1)-tensor s Hom(V, V ) correspond
to each other, s = h

or h = s

, if and only if h(v

1
, v
2
) = s(v
1
), v
2
). The trace with respect to g
of h is tr
g
h = tr s.
3. Coordinate expressions
Let V be a real vector space with basis E
1
, . . . , E
n
and dual basis
1
, . . . ,
n
so that the set
(5.4) is a basis for the tensor space T
k

(V ). For any tensor A T

k

(V ), the coordinates of A (5.5)

with respect to this basis,
A
j
1
j

i
1
i
k
= A(
j
1
, . . . ,
j

, E
i
1
, . . . , E
i
k
),
are called the components of A.
5.13. Kroneckers . The components of the
_
1
1
_
-tensor T
1
1
(V ) given by (, v) = (v)
are
(5.14)
j
i
= (
j
, E
i
) =
j
(E
i
) =
_
1 i = j
0 i ,= j
5.15. Endomorphisms. Let A: V V be an endomorphism of V . The components of A,
viewed (5.6) as the
_
1
1
_
-tensor A T
1
1
(V ) given by A(, v) = (Av), are A
j
i
= A(
j
, E
i
) =
j
AE
i
,
the entries in the matrix for A wrt basis E
i
.
5.16. Tensor product. If A T
k
1

1
(V ) and B T
k
2

2
(V ) then
(5.17) (AB)
j
1
j

1
j

1
+1
j

1
+
2
i
1
i
k
1
i
k
1
+1
i
k
1
+k
2
= A
j
1
j

1
i
1
i
k
1
B
j

1
+1
j

1
+
2
i
k
1
+1
i
k
1
+k
2
5.18. Trace. If the tensor A T
+1
k
(V ) has components A
j
1
j
+1
i
1
i
k+1
meaning that
A =

A
j
1
j
+1
i
1
i
k+1
E
j
1
E
j
+1

i
1

i
k+1
then
tr(A) =

A
j
1
j
+1
i
1
i
k+1
tr(E
j
1
E
j
2
E
j
+1

i
1

i
k

i
k+1
)
=

A
j
1
j
+1
i
1
i
k+1

j
1
i
k+1
E
j
2
E
j
+1

i
1

i
k
=

A
mj
2
j
+1
i
1
i
k
m
E
j
2
E
j
+1

i
1

i
k
which means that the components of tr(A) T
k

(V ) are
(5.20) tr(A)
j
2
j
+1
i
1
i
k
= A
mj
2
j
+1
i
1
i
k
m
a sum of the components of A.
3. COORDINATE EXPRESSIONS 53
5.21. Inner product. The inner product on V is a symmetric (2, 0)-tensor g T
2
0
(V ) whose
components
(5.22) g
ij
= g(E
i
, E
j
) = E
i
, E
j
)
form a symmetric matrix (g
ij
).
5.23. Raising and lowering of the index. Since
(v
j
E
j
)

= v
j
E

j
= v
j
E

j
(E
i
)
i
= v
j
E
i
, E
j
)
i
= g
ij
v
j

i
we may write
(v
j
E
j
)

= v
i

i
where v
i
= g
ij
v
j
, (
i

i
)

=
j
E
j
where
j
= g
ij

i
because the linear map V

V

has matrix (g
ij
) so that the inverse map V

V

has (g
ij
)
1
=
(g
ij
) as its matrix. We get v
i
from v
j
by lowering the index and
j
from
i
by raising the index.
More generally, if A T
k
+1
(V ), then the components of A

T
k+1

(V ) are
(A

)
j
1
j

ii
1
i
k
= g
ij
A
j
1
j

j
i
1
i
k
because
A

(
j
1
, . . . ,
j

, E
i
, E
i
1
, . . . , E
i
k
) = A(
j
1
, . . . ,
j

, E

i
, E
i
1
, . . . , E
i
k
)
= g
ij
A(
j
1
, . . . ,
j

,
j
, E
i
1
, . . . , E
i
k
) = g
ij
A
j
1
j

j
i
1
i
k
and if B T
k+1

(V ), then the components of B

T
k
+1
(V ) are
(5.25) (B

)
j
1
j

j
i
1
i
k
= g
ij
B
j
1
j

ii
1
i
k
because
B

(
j
1
, . . . ,
j

,
j
, E
i
1
, . . . , E
i
k
) = B(
j
1
, . . . ,
j

, (
j
)

, E
i
1
, . . . , E
i
k
)
= g
ij
B(
j
1
, . . . ,
j

, E
i
, E
i
1
, . . . , E
i
k
) = g
ij
B
j
1
j

ii
1
i
k
Here are some special cases:
g
i
j
= g
kj
g
ik
=
j
i
.
If the basis E
i
is orthonormal so that (g
ij
) and its inverse are identity matrices, (A

)
j
1
j

ii
1
i
k
=
g
ij
A
j
1
j

j
i
1
i
k
=
i
j
A
j
1
j

j
i
1
i
k
= A
j
1
j

i
i
1
i
k
and (B

)
j
1
j

j
i
1
i
k
= B
j
1
j

ji
1
i
k
.
5.26. Trace with respect to g. If B T
k+1

(V ) has components B
j
1
j

i
1
i
k+1
then the com-
ponents of tr
g
(B) T
k1

(V ) are
tr
g
(B)
j
1
j

i
3
i
k+1
(5.20)
= (B

)
j
1
j

j
ji
3
i
k+1
(5.25)
= g
ij
B
j
1
j

iji
3
i
k+1
if we remember to take the trace at the right places.
5.27. Example. Let h T
2
0
(V ) is a (2, 0)-tensor and s T
1
1
(V ) the corresponding endo-
morphism of V as in 5.12. Then
h
ij
= h(E
i
, E
j
) = sE
i
, E
j
) =

s
i
k
E
k
, E
j
_
= s
i
k
g
kj
or, equivalently, s
i
j
= g
jk
h
ik
. Lowering an index in s gives h and rasing an index in h gives s.
The problem we are solving here (hidden behind a lot of formalistic notation) is how to construct
s: V V from knowledge of sX, Y ) = h(X, Y . Then tr
g
h = tr s = s
i
i
= g
ij
h
ij
. In particular if
h = g is the metric, then tr
g
(g) = g
i
i
= n.
CHAPTER 6
Non-euclidean geometry
Historical account of non-euclidean geometry.
6.1. Axiom (Parallel axiom of euclidean geometry). For any straight line L and any point P
outside L there is a unique line through P that does not meet L.
6.2. Axiom (Parallel axiom of spherical geometry). For any straight line L and any point P
outside L there is no line through P that does not meet L.
6.3. Axiom (Parallel axiom of hyperbolic geometry). For any straight line L and any point P
outside L there are at least two lines through P that do not meet L.
1. The hyperbolic plane
The hyperbolic plane of radius R is the Riemannian manifold
H
2
R
= (, ) R
2
R
+
[ [[
2

2
= R
2
R
2
R = R
3
with metric induced from the Minkowski metric on R
3
. Equivalently, the hyperbolic plane of
radius R is the disc of radius R
B
2
R
(0) = u R
2
[ [u[ < R
equipped with the metric induced from H
2
R
under hyperbolic stereographic projection
: H
2
R
B
2
R
(0),
1
(u) =
_
2R
2
u
R
2
[u[
2
, R
R
2
+[u[
2
R
2
[u[
2
_
which is a dieomorphism.
Stereograhic projection induces a correspondence between geodesics in the two models since
isometries preserve geodesics. Since the geodesics in the hyperboloid model H
2
R
are the curves
(, ) H
2
R
[ (, ) (, ) = 0, R
3
(, ) ,= 0,
the geodesics in the disc model are the curves
u B
2
R
(0) [
1
(u) (, ) = 0 = u B
2
R
(0) [ 2Ru + (R
2
+[u[
2
) = 0
parametrized as constant speed curves. If = 0, this is a straight line through 0. If ,= 0, we
replace (, ) by a parallel vector of the form (, 1) and get
u B
2
R
(0) [ [u R[
2
+[R[
2
= R
2
= u B
2
R
(0) [ [u R[
2
= R
2
(1 [R[
2
)
If [[ 1, this is the empty set. When [[ < 1, this is part of a circle with center R and radius
R
_
1 [[
2
. If we let u
0
denote either of the two points where the circle meets the boundary of
the disc, then we see that
[u
0
R[
2
+[R[
2
= [u
0
[
2
which means that u
0
R u
0
so that the center, R, of the circle lies on the tangent to the
boundary circle at u
0
. We conclude that the geodesics in the ball model are straight lines through
0 and circular arcs meeting the boudary circle in right angles.
The geodescis crossing the boundary of the hyperbolic plane of radius 1 at the point (1, 0) are
circular arcs with radius V > 0 and center (1, V ) on T
the straight line through (1, 0) perpendicular to T
where T is the tangent at (1, 0) to the boundary circle.
6.4. Definition. The Vorono distance between the point P and the circle c with centre C and
radius R is V (c, P) =
_
[CP[
2
R
2
.
55
56 6. NON-EUCLIDEAN GEOMETRY
Figure 1. Geodesics in the hyperbolic plane
Figure 2. Two lines through P that do not meet the green line!
If P is outside the circle, V (c, P) is the length of the tangent segments through P; V (c, P) = 0
when P is on the circle; and V (c, P) is purely imaginary when P is inside the circle.
The centers of the geodesic circular arcs through a point P
0
in the hyperbolic plane B
2
R
(0) of
radius R are the points P for which
[P
0
P[ = V (B
2
R
(0), P)
meaning that the euclidean distance to P
0
equals the Vorono distance to the boundary circle
B
2
R
(0). This locus is the perpendicular bisector of P
0
P

0
where P

0
is the point where the radius
through P
0
crosses the boundary circle.
If P
1
and P
2
are two given points in B
2
R
(0), not on the same diameter, then the center of the
geodesic through P
1
and P
2
is the point Q for which
[P
1
Q[ = V (B
2
R
(0), Q) = [P
2
Q[
meaning that Q is the intersection of the two bisecting lines associated to P
1
and P
2
as just
described.
It is known that the hyperbolic plane embeds isometrically into euclidean R
5
but not into R
3
.
It is unknown if there exists an isometric embedding of the hyperbolic plane into R
4
.
1. THE HYPERBOLIC PLANE 57
Figure 3. The unique geodesic through two given points
Bibliography
[1] Je Cheeger and David G. Ebin, Comparison theorems in Riemannian geometry, North-Holland Publishing Co.,
Amsterdam, 1975, North-Holland Mathematical Library, Vol. 9. MR MR0458335 (56 #16538)
[2] G. S. Hall, Symmetries and curvature structure in general relativity, World Scientic Lecture Notes in Physics,
vol. 46, World Scientic Publishing Co. Inc., River Edge, NJ, 2004. MR MR2109072
[3] S. W. Hawking and G. F. R. Ellis, The large scale structure of space-time, Cambridge University Press, London,
1973, Cambridge Monographs on Mathematical Physics, No. 1. MR MR0424186 (54 #12154)
[4] John M. Lee, Riemannian manifolds, Graduate Texts in Mathematics, vol. 176, Springer-Verlag, New York,
1997, An introduction to curvature. MR MR1468735 (98d:53001)
[5] Jesper M. Mller, General topology, http://www.math.ku.dk/~moller/e03/3gt/notes/gtnotes.pdf.
[6] John Nash, The imbedding problem for Riemannian manifolds, Ann. of Math. (2) 63 (1956), 2063. MR
MR0075639 (17,782b)
[7] Barrett ONeill, Semi-Riemannian geometry, Pure and Applied Mathematics, vol. 103, Academic Press Inc.
[Harcourt Brace Jovanovich Publishers], New York, 1983, With applications to relativity. MR MR719023
(85f:53002)
[8] Michael Spivak, A comprehensive introduction to dierential geometry. Vol. II, second ed., Publish or Perish
Inc., Wilmington, Del., 1979. MR MR532831 (82g:53003b)
59