Riemannian geometry – a brief introduction

(for TMA 4190 Manifolds, spring 2011)

In di¤erential topology one associates to a manifold M its tangent bundle

T M , cotangent bundle T M ( as well as other tensor bundles), and various
other invariants, but these invariants cannot be used to distinguish manifolds
which are di¤eomorphic. However, in Riemannian geometry another kind of
structure is imposed on the manifold, which is of a geometric nature, namely
a Riemannian metric. The new structure enables one to talk about distances,
angles, length, area, volume, etc., The principal example is, indeed, an old
friend, namely Euclidean geometry . Thus, in the Euclidean plane E 2 we can
distinguish between a circle and an ellipse, for example, and also between circles
of di¤erent radius. But we cannot see any di¤erence between isometric "…gures"
or isometric manifolds. Riemannian geometry is the theory which enables us
to study manifolds from a geometric viewpoint, in a way which is as close to
Euclidean geometry as possible.
In fact, in a very small neighborhood of any point the geometry is approx-
imately Euclidean. We need to be more precise about this. First of all, by an
inner product on a vector space V we mean a positive de…nite symmetric bi-
linear form hu; vi. This enables one to introduce geometric concepts in V such
as the length of a vector, distance between points, angle between rays, area and
so on, in the space V . Here is a precise de…nition of how to generalize all this
to manifolds.

De…nition 1 A Riemannian metric on a smooth manifold is an assignment of

an inner product hu; vip in each tangent space Tp M . Moreover, for any two
smooth vector …elds X; Y on M , the function M ! R de…ned by p ! hXp ;Yp ip
is also smooth.

The last condition is needed for the metric to be smooth, loosely speeking ,so
that the inner product changes smoothly from point to point. We shall say more
about this later. Before we continue, we shall have a closer look at the simplest
type of smooth manifolds, namely the Cartesian n-space Rn which serves as the
local model for any other n-dimensional manifold.

0.1 Cartesian n-space, global and local coordinates

The Cartesian n-space Rn = f(x1 ; x2 ; :::; xn ); xi 2 Rg consists of all n-tuples of
real numbers. It is naturally an n-dimensional vector space over R, with the
zero vector as a distinguished point O called the origin. There is also an obvious
basis
e1 = (1; 0; 0; ::; 0); e2 = (0; 1; 0; ::; 0); :::; en = (0; :::; 0; 1);
often referred to as the standard basis. Rn serves as the standard model for
all n-dimensional real vector spaces, since they are all isomorphic to Rn and
the latter is evidently the simplest example. More precisely, any n-dimensional

1
vector space V with a given basis b1 ; b2 ; :::; bn can be immediately identi…ed
with Rn via the following isomorphism
X
xi bi ! (x1 ; x2 ; :::; xn )
i

In di¤erential topology we study manifolds of a given dimension n, and they are

locally modelled on Rn (or any n-dimensional real vector space). In particular,
Rn is the simplest example of an n-dimensional manifold, with the di¤erentiable
structure determined by the atlas consisting of a single chart, namely the identity
map on Rn .
A special feature of a manifold di¤eomorphic to Rn is that we may simply
replace it by Rn and thus it also becomes a vector space. Assuming this has
!
been done, points P are identi…ed with their position vector OP , on which we
can perform vector algebra and calculus calculations – well known from linear
algebra and analytic geometry in n-space.
For a smooth manifold M in general, the tangent spaces Tp M are vector
spaces, and they are isomorphic because they have the same dimension. But in
general there is no natural isomorphism between the tangent spaces at di¤erent
points. However, when M = Rn each vector space Tp Rn is naturally isomorphic
to Rn itself, essentially because position vectors in Rn and tangent vectors can
be viewed as vectors in the same vector space Rn = M . This is a very special
property which only makes sense for the manifolds Rn (or open subsets of them).
As a consequence, given any basis b1 ; b2 ; :::; bn of Rn = M , these vectors also
represent global vector …elds

bi : Rn ! T Rn , p ! bi jp 2 Tp Rn ' Rn

To such a basis corresponds a system of Cartesian coordinates (y1 ; y2 ; :::; yn ) on

M , as follows X
p= yi bi ! (y1 ; y2 ; :::; yn )
In di¤erential topology these vector …elds are often denoted
@ @
b1 = ; :::; bn = (1a)
@y1 @yn
and we call them the coordinate vector …elds associated with the coordinates
y1 ; y2 ; :::yn :Clearly, any smooth vector …eld on M can be expressed as a unique
linear combination
@ @ @
X = f1 + f2 + ::: + fn
@y1 @y2 @yn
where the fi are smooth functions.
Here we point out that Cartesian coordinates on Rn is just a very special type
of coordinates, closely related to the vector space structure . But sometimes
one also need to work with more general coordinate systems on Rn , perhaps

2
only de…ned on some open subset U: Given n functions ui on U , they yield a
coordinate system (u1 ; :::; un ) i¤ the map

p ! (p) = (u1 (p); :::; un (p)) (2)

is a di¤eomorphism between U and some open subset (U ) of Rn . To this

coordinate system there is associated vector …elds on U
@ @
; ::; (3)
@u1 @un
which we also refer to as the coordinate vector …elds. Thus they provide a basis
@ @
@u1 jp ; :::; @un jp for the tangent space Tp M ; for each p 2 U . To make sense out
@
of the symbol @u i
, recall the usual de…nition of a tangent vector in a manifold.
Namely, a vector …eld (resp. tangent vector) can be interpreted as a derivation
(resp. local derivation) acting on the ring of smooth functions. So it should be
@
clear what we mean by the operator @u i
.
Still, one may …nd all this somewhat confusing, whether the given manifold
M is Rn or not. Namely, suppose we have a chart ( ; U ) on a given manifold
M , say the function
: U ! Rn
is given by (2), where (u1 ; u2 ; :::; un ) are Cartesian coordinates on the (target)
@
space Rn . Then the latter space has coordinate vector …elds @u i
as explained
before, but on the other hand we also use the same notation for the "induced"
vector …elds on U M . The reason for this is that we want the notation as
simple as possible. But in fact, there is a linear isomorphism p : Tp M !
@ @
T (p) Rn = Rn so that @u i
jp denotes the vector which is mapped to @u i
j (p) .
n
The latter acts on functions F on R (or a nbd of (p)) as follows

@F
F ! ( (p) , partial di¤erentiation with respect to ui ;
@ui
@
whereas in the manifold M the tangent vector @ui jp acts on functions f (de…ned
on a nbd of p) by
1
@(f ) @f
f! ( (p), also denoted (p) for brevity
@ui @ui

Example 2 Consider M = R3 with the Cartesian coordinates (x; y; z) on the

one hand, and spherical coordinates ( ; '; ) on the other hand, de…ned by the
relations
x = sin ' cos ; y = sin ' sin ; z = cos ' (4)
The spherical coordinates are not globally de…ned, so let’s assume working on
some open set U where the equations (4) can be solved for ; '; in terms of
x; y; z:From the equations (4) one can immediately calculate the coordinate cov-
ector …elds dx; dy; dz on U expressed as linear combinations of the coordinate

3
covector …elds d ; d'; d , namely by di¤ erentiation
dx = sin ' cos d + cos ' cos d' sin ' sin d
dy = sin ' sin d + cos ' sin d' + sin ' cos d (5)
dz = cos 'd sin 'd'
In order to express d ; d'; d in terms of dx; dy; dz, you must either invert the
relations (4) and di¤ erentiate as above, or you can (much easier?) solve the
linear system (5).
@ @ @
Problem 3 In the previous example, express @x ; @y ; @z as a linear combination
@
of @@ ; @' ; @@ ; and conversely.

What you should know is that the covector basis dx; dy; dz (at a given point
@ @ @
p) is dual to the vector basis @x ; @y ; @z (at p), and similarly with the other
coordinate system. The system (5) can be formally written in terms of matrices
as 0 1 0 1
dx d
@ dy A = A @ d' A (6)
dz d
where A is the Jacobi matrix of the transformation ( ; '; ) ! (x; y; z), having
determinant 2 sin ', namely
0 1
sin ' cos cos ' cos sin ' sin
A = @ sin ' sin cos ' sin sin ' cos A
cos ' sin ' 0
0 1
cos sin ' sin sin ' cos '
1 1 1
A 1 = @ cos cos ' cos ' sin sin ' A
1 sin 1 cos
sin ' sin ' 0

In general, it follows from simple linear algebra that if a relation like (6)
holds for two bases in a 3-dim vector space V , then there is a corresponding
relation between their dual bases in V , but in the "opposite" direction and
with the transpose matrix, namely we have
0 @ 1 0 @ 1 0 @ 1 0 @ 1
@ @x @x @
@ @ A = AT @ @ A or @ @ A = (AT ) 1 @ @ A
@' @y @y @'
@ @ @ @
@ @z @z @

In particular, we write out for later reference

@ @ 1 @ 1 sin @
= cos sin ' + cos cos '
@x @ @' sin ' @
@ @ 1 @ 1 cos @
= sin sin ' + cos ' sin + (7)
@y @ @' sin ' @
@ @ 1 @
= cos ' sin '
@z @ @'

4
0.2 The Euclidean n-space
As a smooth manifold, the Euclidean n-space E n is just an n-dimensional vector
space, so let us say Rn . One can describe the geometry by a system of axioms
(concerning points, lines, planes, angles, parallelism, etc.), and this was also
the original approach dating back to ancient times more than 2000 years ago.
But nowadays we can also say that E n is a smooth manifold with a "‡at"
Riemannian structure. In fact, E n is the simplest n-dimensional Riemannian
manifold one can imagine.
Vector algebra (or vector geometry ) is the algebraization of Euclidean geom-
etry, which makes it possible to calculate and prove geometric theorems by alge-
braic and analytic methods. The identi…cation of E n with Rn is the coordinate
approach to geometry, which dates back to Descartes and Fermat in the 17th
century, from which analytic geometry arose. So it became customary to iden-
tify E n with the Cartesian n-space Rn , namely an n-dimensional vector space
equipped with some geometric structure. From a modern viewpoint, all we need
extra is an inner product hu; vi, from which one can de…ne the length of line
segments and angles between crossing lines. In some sense, the inner product
is the algebraization of Pythagoras theorem in Euclidean geometry, which tells
us that in a triangle with a right angle, the three sides a; b; c satisfy the iden-
tity a2 + b2 = c2 where c is the length of the side opposite to the right angle.
This property is, in fact, just a reformulation of the Euclidean parallel postulate
which says that the Euclidean plane has the following property: given a line
L and a point P outside L , then there is a unique line L0 parallel to L and
passing through P . The property is also equivalent to saying that the sum of
the three angles in a triangle is equal to two right angles (i.e., 180 degrees). It
is this property that fails in hyperbolic geometry, as well as in spherical (also
called elliptic) geometry.
Now, having an inner product in E n = Rn , we have in fact the same inner
@
product hu; vip in each tangent space Tp Rn ' Rn . Therefore, if @x 1
; :::; @x@n are
the coordinate vector …elds corresponding to Cartesian coordinates (x1 ; :::; xn ),
then the functions
@ @
gij (p) = jp ; jp (8)
@xi @xj p

are by de…nition constant, that is, independent of p:This is a characteristic

property of Euclidean geometry.
In a general Riemannian manifold one may not be able to …nd any coordinate
system, in a nbd of a given point p, such that the functions gij are constant.
That happens when the curvature (or curvature tensor) is non-vanishing at p.
But otherwise, if they are constant then we say that the space is ‡at near p. But
besides E n itself and open subsets of it, are there other ‡at n-dim Riemannian
manifolds? Well, the space must be locally isometric to E n , but how can it look
like globally? It is certainly di¢ cult to imagine a compact ‡at surface imbedded
in E 3 .
We shall not give the de…nition of curvature, but any compact surface in
3-space must obviously have positive curvature somewhere in order to close

5
up to a closed and bounded set. However, Riemannian manifolds have their
intrinsic geometry, and it is for example possible to give the 2-dimensional torus
S 1 S 1 a Riemannian ‡at metric. The discovery of the "‡at torus" was a big
surprise around 1870, but in modern mathematics the construction is quite easy.
Namely, a 2-dim torus is a quotient space of R2 (show this!) and the quotient
map : R2 ! S 1 S 1 is a local di¤eomorhism. This allows us to "push" down
the metric on R2 to a metric on the torus, which makes the quotient map a
local isometry . But being ‡at is a local property, so being ‡at in a nbd of every
point, the torus is certainly ‡at overywhere.
There is the following theorem which characterize the Riemannian manifold
En :

Theorem 4 The Riemannian manifold M is isometric to the Euclidean n-space

E n if and only if it is di¤ eomorphic to Rn and, moreover, there are n vector
…elds X1 ; X2 ; :::; Xn which are linear independent at each point and such that
the functions p ! hXi (p); Xj (pip are constant for all i; j.

0.3 Riemannian manifolds

A Riemannian manifold M has an inner product hu; vip on each tangent space.
Therefore, given local coordinates (x1 ; :::; xn ) in a nbd U the metric is uniquely
determined in U by the following functions on U

@ @
p ! gij (p) = jp ; jp
@xi @xj p

Thus the matrix function (gij ) determines the metric in the nbd U . By in-
troducing the arc-length element ds the metric on U is also classically written
as X
ds2 = gij dxi dxj (9)
The reason for this can be explained by considering a parametrized smooth
curve t ! (t): Its velocity …eld along the curve is
d dx1 @ dxn @
= + ::: +
dt dt @xi dt @xn
ds
and the speed v(t) = dt is the length of the velocity vector, namely

d 2 d d X dxi dxj
v 2 = jj jj = ; = gij (10)
dt dt dt dt dt

Multiplying v 2 by dt2 we get ds2 , so equation (9) is really the identity (10)
without any speci…c parametrization of the curve.
The length of the curve from t = a to t = b is
Z Z b Z b rX
dxi dxj
L( ) = ds = vdt = gij dt
a a dt dt

6
In fact, the length does not depend on the choice of parametrization. Now we
can also introduce a distance function by saying that the distance (p; q) from
p to q is the length of the shortest piecewice smooth curve from p to q. Well,
here is a problem, since there may not be any curve with the shortest length.
So we rather de…ne

(p; q) = inf fL( ); piecewise smooth and connects p and qg

This makes M to a metric space, and fortunately the topology de…ned by the
distance function coincides with the original topology on the manifold!
We also mention that so-called geodesic curves are the generalization of
straight lines in Euclidean geometry. For two nearby points on such a curve,
the curve segment between the points is the unique shortest curve and hence
its length gives the distance. However, globally, this may not hold. That is,
when the points are moved further away from each other, the geodesic segment
between them may cease to be the shortest curve between the points. A simple
example is the round 2-sphere S 2 , where the geodesics are the great circles.
Take a look and see what happens.

De…nition 5 Two Riemannian manifolds M and N are isometric if there is

a di¤ eomorphism : M ! N such that the induced linear isomorphism :
Tp M ! T (p) N is an isometry (i.e. preserves the inner product of vectors) for
each point p:

The Riemannian metric on a manifold M enables us to construct an isomor-

phism between the tangent bundle and the cotangent bundle

TM ' T M
TM ! # # (11)
M = M

This follows from a well known result in linear algebra, namely an inner product
on a vector space V enables one to identify an element of V , that is, a linear
functional : V ! V , with a vector 2 V , as follows

(v) = h ; vi

Conversely, any vector u in V corresponds to an element u in V , namely the

functional v ! hu; vi. In physics one likes to write h ; vi = (v) since after
all, applying to v gives a bilinear pairing V V ! R , even when there is
no inner product involved. Thus, we can write hu; vi = hu; vi = u(v) etc.
Recall that any smooth function f on M gives a covector …eld df , called the
di¤erential of f . A covector …eld is also called a 1-form. Now, with reference
to the diagram (11), let us replace the projection "arrow" on the right by an
upward going "arrow" representing df . Then, composing the function df with
the di¤eomorphism T M ! T M we obtain a cross section M ! T M , that is,
a vector …eld which we shall denote by rf . This is the gradient …eld of f . In

7
other words, the gradient of f at p is the vector rf (p) in Tp M de…ned by the
property that
hrf (p); vip = hdf jp ; vi = df (v) = v(f ) (12)
holds for every vector v in Tp M:

Problem 6 Find the expression for the gradient rF of a function on Euclidean

3-space in terms of spherical coordinates. Let f be a function on the round 2-
sphere S 2 of radius 1. Calculate the gradient of f using spherical coordinates.

Let us see what to do. In E 3 the Euclidean coordinates (x; y; z) has coordi-
@ @ @
nate …elds @x ; @y ; @z which constitute an orthonormal basis at every point, and
hence
ds2 = dx2 + dy 2 + dz 2 (13)
Everybody is told that the gradient of F is given by
@F @ @F @ @F @
rF = + + (14)
@x @x @y @y @z @z
@
and this is correct. Therefore,since the vectors @x ; :: have already been expressed
in spherical coordinates in (7), we need only substitute into (14).
On the other hand, we obtain the metric (13) in spherical coordinates by
substituting the expressions (5) into the expression (13) for ds2 , which yields

ds2 = d 2
+ 2
(d'2 + sin2 'd 2 ) (15)

Then we can calculate directly the gradient of F from the metric (15) if we use
(12). To illustrate this , let us write
@ @ @
rF = a +b +c
@ @' @
@ @ @
and try to determine a; b; c. Let v = @ + @' + @ be any vector, and
calculate the inner product

hrF; vi = a + b 2 + c 2 sin2 ' = v(F )

@F @F @F
= + +
@ @' @
Since this must hold for all numbers ; ; , we conclude that
@F 1 @F 1 @F
a= ; b= 2 ; c=
@ @' 2 sin2 ' @

Finally, consider the function f on the 2-sphere S 2 = ( = 1) in E 3 . In the

metric expression (15), put = 1 and obtain the spherical metric on S 2 , as a
sub-Riemannian manifold. Its metric in spherical coordinates ('; ) is
2
ds2 = d'2 + sin2 'd (16)

8
and calculation of the gradient yields the result
@f @ 1 @f @
rf = +
@' @' sin2 ' @ @

Let us return to the identites (7). In particular , the unit vector in the
positive z-axis direction, at points outside the z-axis, decomposes as
@ @ 1 @
k= = cos ' sin '
@z @ @'

Let p be a point on the unit sphere S 2 , so we set = 1:The vector @@ at p

is a unit vector in 3-space which is normal to the sphere, so along S 2 the …eld
@ @
@ is a normal vector …eld pointing outward. However, the vectors @' and
@ @
@ are perpendicular to @ , so they are tangential to the sphere and therefore
they constitute a basis for Tp S 2 at the point p. The basis is orthogonal but
@
not orthonormal. In fact, from (16) it follows that @' has length 1 and @@ has
length sin '. This also explains why spherical coordinates are singular along the
z-axis. On S 2 we see that the coordinate vector @@ becomes zero at the north
and south pole, but away from the poles S 2 has the following two orthonormal
vector …elds
@ 1 @
f ; g
@' sin ' @