Part III: Differential geometry (Michaelmas 2004)

Alexei Kovalev ([email protected])

The line and surface integrals studied in vector calculus require a concept of a curve and a
surface in Euclidean 3-space. These latter objects are introduced via a parameterization:
a smooth map of, respectively, an interval on the real line R or a domain in the plane R2
into R3 . In fact, one often requires a so-called regular parameterization. For a curve r(t),
this means non-vanishing of the ‘velocity vector’ at any point, ṙ(t) 6= 0. On a surface a
point depends on two parameters r = r(u, v) and regular parameterization means that the
two vectors of partial derivatives in these parameters are linearly independent at every
point ru (u, v) × rv (u, v) 6= 0.
Differential Geometry develops a more general concept of a smooth n-dimensional
differentiable manifold.1 and a systematic way to do differential and integral calculus (and
more) on manifolds. Curves and surfaces are examples of manifolds of dimension d = 1
and d = 2 respectively. However, in general a manifold need not be given (or considered)
as lying in some ambient Euclidean space.

1.1 Manifolds: definitions and first examples

The basic idea of smooth manifolds, of dimension d say, is to introduce a class of spaces
which are ‘locally modelled’ (in some precise sense) on a d-dimensional Euclidean space
Rd .
A good way to start is to have a notion of open subsets (sometimes one says ‘to have
a topology’) on a given set of points (but see Remark on page 2).

Definition. 1. A topological space is a set, M say, with a specified class of open

subsets, or neighbourhoods, such that
(i) ∅ and M are open;
(ii) the intersection of any two open sets is open;
(iii) the union of any number of open sets is open.
2. A topological space M is called Hausdorff if any two points of M possess non-
intersecting neighbourhoods.
3. A topological space M is called second countable if one can find a countable
collection B of open subsets of M so that any open U ⊂ M can be written as a union
of sets from B.

The last two parts of the above definition will be needed to avoid some pathological
examples (see below).
1
More precisely, it is often useful to also consider appropriate ‘structures’ on manifolds (e.g. Riemannian
metrics), as we shall see in due course.

1
2 alexei kovalev

The knowledge of open subsets enables one to speak of continuous maps: a map between
topological spaces is continuous if the inverse image of any open set is open. Exercise: check
that for maps between open subsets (in the usual sense) of Euclidean spaces this definition
is equivalent to other definitions of a continuous map.
A homeomorphism is a bijective continuous map with continuous inverse. More ex-
plicitly, to say that ‘a bijective mapping ϕ of U onto V is a homeomorphism’ means that
‘D ⊂ U is open if and only if ϕ(D) ⊂ V is open’.
Let M be a topological space. A homeomorphism ϕ : U → V of an open set U ⊆ M
onto an open set V ⊆ Rd will be called a local coordinate chart (or just ‘a chart’) and
U is then a coordinate neighbourhood (or ‘a coordinate patch’) in M .
Definition. A C ∞ differentiable structure, or smooth structure, on M is a collection
of coordinate charts ϕα : Uα → Vα ⊆ Rd (same d for all α’s) such that
(i) M = ∪α∈A Uα ;
(ii) any two charts are ‘compatible’: for every α, β the change of local coordinates ϕβ ◦ϕ−1
α
is a smooth (C ∞ ) map on its domain of definition, i.e. on ϕα (Uβ ∩ Uα ) ⊆ Rd .
(iii) the collection of charts ϕα is maximal with respect to the property (ii): if a chart ϕ
of M is compatible with all ϕα then ϕ is included in the collection.
A bijective smooth map with a smooth inverse is called a diffeomorphism. Notice
that clause (ii) in the above definition implies that any change of local coordinates is a
diffeomorphism between open sets ϕα (Uβ ∩ Uα ) and ϕβ (Uβ ∩ Uα ) of Rd .
In practice, one only needs to worry about the first two conditions in the above defini-
tion. Given a collection of compatible charts covering M , i.e. satisfying (i) and (ii), there
is a unique way to extend it to a maximal collection to satisfy (iii). I leave this last claim
without proof but refer to Warner, p. 6.
Definition . A topological space equipped with a C ∞ differential structure is called a
smooth manifold. Then d is called the dimension of M , d = dim M .
Sometimes in the practical examples one starts with a differential structure on a set
of points M (with charts being bijective maps onto open sets in Rd ) and then defines
the open sets in M to be precisely those making the charts into homeomorphisms. More
explicitly, one then says that D ⊂ M is open if and only if for every chart ϕ : U → V ⊆ Rd ,
ϕ(D ∩ U ) is open in Rd . (For this to be well-defined, every finite intersection of coordinate
neighbourhoods must have an open image in Rd under some chart.) We shall refer to this
as the topology induced by a C ∞ structure.
Remarks . 1. Some variations of the definition of the differentiable structure are possible.
Much of the material in these lectures could be adapted to C k rather than C ∞ differentiable
manifolds, for any integer k > 0. 2
On the other hand, replacing Rd with the complex coordinate space Cn and smooth
maps with holomorphic (complex analytic) maps between domains in C, leads of an im-
portant special class of complex manifolds—but that is another story.
2
if k = 0 then the definition of differentiable structure has no content
 differential geometry 3

By a manifold I will always mean in these lectures a smooth (real) manifold, unless
explicitly stated otherwise.
2. Here is an example of what can happen if one omits a Hausdorff property. Consider
the following ‘line with a double point’
M = (−∞, 0) ∪ {00 , 000 } ∪ (0, ∞),
with two charts being the obvious ‘identity’ maps (the induced topology is assumed)
ϕ1 : U1 = (−∞, 0) ∪ {00 } ∪ (0, ∞) → R, ϕ2 : U2 = (−∞, 0) ∪ {000 } ∪ (0, ∞) → R, so
ϕ1 (00 ) = ϕ1 (000 ) = 0. It is not difficult to check that M satisfies all the conditions of a
smooth manifold, except for the Hausdorff property (00 and 000 cannot be separated).
Omitting the 2nd countable property would allow, e.g. an uncountable collection (dis-
joint union) of lines
t0<α<1 Rα ,
each line Rα equipped with the usual topology and charts being the identity maps Rα → R.
3

Examples. The Euclidean Rd is made into a manifold using the identity chart. The
complex coordinate space Cn becomes a 2n-dimensional manifold via the chart Cn → R2n
replacing every complex coordinate zjP by a pair of real coordinates Re zj , Im zj .
n
The sphere S n = {x ∈ Rn+1 : 2
i=0 xi = 1} is made into a smooth manifold of
dimension n, by means of the two stereographic projections onto Rn ∼ = {x ∈ Rn+1 :
x0 = 0}, from the North and the South poles (±1, 0, . . . , 0). The corresponding change of
coordinates is given by (x1 , . . . , xn ) 7→ (x1 /|x|2 , . . . , xn /|x|2 ).
The real projective space RP n is the set of all lines in Rn+1 passing through 0. Elements
of RP n are denoted by x0 : x1 : . . . : xn , where not all xi are zero. Charts can be given by
ϕi (x0 : x1 : . . . : xn ) = (x0 /xi , . . . î . . . , xn /xi ) ∈ Rn with changes of coordinates given by
ϕj ◦ ϕ−1 n
i : (y1 , . . . , yn ) 7→ y1 : . . . : (1 in ith place) : . . . : yn ∈ RP 7→
 
y1 yi−1 1 yi yj−1 yj+1 yn
,..., , , ,..., , ... ,
yj yj yj yj yj yj yj
smooth functions on their domains of definition (i.e. for yj 6= 0). Thus RP n is a smooth
n-dimensional manifold.
Definition . Let M, N be smooth manifolds. A continuous map f : M → N is called
smooth (C ∞ ) if for each p ∈ M , for some (hence for every) charts ϕ and ψ, of M and
N respectively, with p in the domain of ϕ and f (p) in the domain of ψ, the composition
ψ ◦ f ◦ ϕ−1 (which is a map between open sets in Rn , Rk , where n = dim M , k = dim N )
is smooth on its domain of definition.
Exercise: write out the domain of definition for ψ ◦ f ◦ ϕ−1 .
Two manifolds M and N are called diffeomorphic is there exists a smooth bijective
map M → N having smooth inverse. Informally, diffeomorphic manifolds can be thought
of as ‘the same’.
3
For a more interesting example on a ‘non 2nd countable manifold’ see Example Sheet Q1.12.
4 alexei kovalev

1.2 Matrix Lie groups

Consider the general linear group GL(n, R) consisting of all the n × n real matrices A
satisfying det A 6= 0. The function A 7→ det A is continuous and GL(n, R) is the inverse
image of the open set R \ {0}, so it is an open subset in the n2 -dimensional linear space of
all the n × n real matrices. Thus GL(n, R) is a manifold of dimension n2 . Note that the
result of multiplication or taking the inverse depends smoothly on the matrix entries.
Similarly, GL(n, C) is a manifold of dimension 2n2 (over R).

Definition . A group G is called a Lie group if it is a smooth manifold and the map
(σ, τ ) ∈ G × G → στ −1 ∈ G is smooth.

Let A be an n × n complex matrix. The norm given by |A| = n maxij |aij | has a useful
property that |AB| ≤ |A||B| for any A, B. The exponential map on the matrices is defined
by
exp(A) = I + A + A2 /2! + . . . + An /n! + . . . .
The series converge absolutely and uniformly on any set {|A| ≤ µ}, by the Weierstrass
M-test. It follows that e.g. exp(At ) = (exp(A))t and exp(C −1 AC) = C −1 exp(A)C,
for any invertible matrix C. Furthermore, the term-by-term differentiated series also con-
verge uniformly and so exp(A) is C ∞ -smooth in A. (This means smooth as a function of
2n2 real variables, the entries of A.)
The logarithmic series

log(I + A) = A − A2 /2 + . . . + (−1)n+1 An /n + . . .

converge absolutely for |A| < 1 and uniformly on any closed subset {|A| ≤ ε}, for ε < 1,
and log(A) is smooth in A.
One has
exp(log(A)) = A, when |A − I| < 1. (1.1)
This is true in the formal sense of composing the two series in the left-hand side. The
formal computations are valid in this case as the double-indexed series in the left-hand
side is absolutely convergent.
For the other composition, one has

log(exp(A)) = A when |A| < log 2. (1.2)

again by considering a composition of power series with a similar reasoning.

Remark . Handling the power series of complex matrices in (1.1) and (1.2) is quite similar
to handling 1 × 1 matrices, i.e. complex numbers. Warning: not all the usual properties
carry over wholesale, as the multiplication of matrices is not commutative. E.g., in general,
exp(A) exp(B) 6= exp(A + B). However, the identity exp(A) exp(−A) = I does hold (and
this is used in the proof of Proposition 1.3 below).
 differential geometry 5

Proposition 1.3. The orthogonal group O(n) = {A ∈ GL(n, R) : AAt = I} has a smooth
structure making it into a manifold of dimension n(n − 1)/2.
n(n−1)
The charts take values in the 2
-dimensional linear space of skew-symmetric n × n
real matrices. E.g.

ϕ : A ∈ {A ∈ O(n) : |A − I| is small } 7→ B = log(A) ∈ {B : B t = −B}

is a chart in a neighbourhood of I ∈ O(n). The desired smooth structure is generated by

a family of charts of the form ϕC (A) = log(C −1 A), where C ∈ O(n).
The method of proof of Proposition 1.3 is not specific to the orthogonal matrices and
works for many other subgroups of GL(n, R) or GL(n, C) (Example Sheet 1, Question 4).

1.3 Tangent space to a manifold

If x(t) is a smooth regular curve in Rn then the velocity vector ẋ(0) is a tangent vector to
this curve at t = 0. In a change of coordinates x0i = x0i (x) the coordinates of this vector are
transformed according to the familiar chain rule, applied to x0 (x(t)). One consequence is
that the statement “two curves pass through the same point with the same tangent vector”
is independent of the choice of coordinates, that is to say a tangent vector (understood as
velocity vector) is a geometric object.
The above observation is local (depends only on what happens in a neighbourhood
of a point of interest), therefore the definition may be extended to an arbitrary smooth
manifold.
Definition. A tangent vector to a manifold M at a point p ∈ M is a map a assigning to
each chart (U, ϕ) with p ∈ U an element in the coordinate space (a1 , a2 , . . . , an ) ∈ Rn in
such a way that if (U 0 , ϕ0 ) is another chart then
 0
0 ∂xi
ai = aj , (1.4)
∂xj p

where xi , x0i are the local coordinates on U, U 0 respectively. 4

All the tangent vectors at a
given point p form the tangent space denoted Tp M .
Remarks . It is easy to check that Tp M is naturally a vector space.
The transformation law (1.4) is the defining property of a tangent vector.
Notation. A choice local coordinates xi on a neighbourhood U ⊆ M defines a linear
isomorphism Tp M → Rn . A basis of Tp M corresponding to the standard basis of Rn via
this isomorphism is a usually denoted by ( ∂x∂ i )p . The expression of a tangent vector in
local coordinates a(U, ϕ) = (a1 , . . . , an ) then becomes ai ( ∂x∂ i )p .
4
Here and below a convention is used that if the same letter appears as an upper and lower index then
the summation is performed over the range of this index. E.g. the summation in j = 1, . . . , n in this
instance.
6 alexei kovalev

As can be seen from (1.4), the standard basis vectors of Tp M given by the local coor-
dinates xi and x0i are related by
 0 
∂xj
  
∂ ∂
= . (1.40 )
∂xi p ∂xi p ∂x0j p

This is what one would expect in view of the chain rule from the calculus. The formula (1.40 )
tells us that every tangent vector ai ( ∂x∂ i )p gives a well-defined first-order ‘derivation’

∂ ∂f
ai ( )p : f ∈ C ∞ (M ) → ai (p) ∈ R. (1.5)
∂xi ∂xi
In fact the following converse statement is true although I shall not prove it here5 . Given
p ∈ M , every linear map a : C ∞ (M ) → R satisfying Leibniz rule a(f g) = a(f )g(p) +
f (p)a(g) arises from a tangent vector as in (1.5).
Example 1.6. Let r = r(u, v), (u, v) ∈ U ⊆ R2 be a smooth regular-parameterized surface
in R3 . Examples of tangent vectors are the partial derivatives ru , rv —these correspond
∂ ∂
to just ∂u , ∂v in the above notation, as the parameterization by u, v is an instance of a
coordinate chart.
Definition. A vector space with a multiplication [·, ·], bilinear in its arguments (thus satis-
fying the distributive law), is called a Lie algebra if the multiplication is anti-commutative
[a, b] = −[b, a] and satisfies the Jacobi identity [[a, b], c] + [[b, c], a] + [[c, a], b] = 0.
Theorem 1.7 (The Lie algebra of a Lie group). Let G be a Lie group of matrices and
suppose that log defines a coordinate chart near the identity element of G. Identify the
tangent space g = TI G at the identity element with a linear subspace of matrices, via the
log chart, and then g is a Lie algebra with [B1 , B2 ] = B1 B2 − B2 B1 .
The space g is called the Lie algebra of G.
Proof. It suffices to show that for every two matrices B1 , B2 ∈ g, the [B1 , B2 ] is also an
element of g. As [B1 , B2 ] is clearly anticommutative and the Jacobi identity holds for
matrices, g will then be a Lie algebra.
The expression

A(t) = exp(B1 t) exp(B2 t) exp(−B1 t) exp(−B2 t)

defines, for |t| < ε with sufficiently small ε, a path A(t) in G such that A(0) = I. Using
for each factor the local formula exp(Bt) = I + Bt + 21 B 2 t2 + o(t2 ), as t → 0, 6 we obtain

A(t) = I + [B1 , B2 ]t2 + o(t2 ), as t → 0.

5
A proof can be found in Warner 1.14-1.20. Note that his argument requires C ∞ manifolds (and does
not extend to C k ).
6
The notation o(tk ) means a remainder term of order higher than k, i.e. r(t) such that limt→0 r(t)/tk = 0.
 differential geometry 7

Hence
B(t) = log A(t) = [B1 , B2 ]t2 + o(t2 ) and exp(B(t) = A(t)
hold for any sufficiently small |t| and so B(t) ∈ g (as B(t) is in the image of the log
chart). Hence B(t)/t2 ∈ g for every small t 6= 0 (as g is a vector space). But then
also limt→0 B(t)/t2 = limt→0 ([B1 , B2 ] + o(1)) = [B1 , B2 ] ∈ g (as g is a closed subset of
matrices).

Notice that the idea behind the above proof is that the Lie bracket [B1 , B2 ] on a Lie
algebra g is an ‘infinitesimal version’ of the commutator g1 g2 g1−1 g2−1 in the corresponding
Lie group G.

Definition. Let M be a smooth manifold. A disjoint union T M = tp∈M Tp M is called the

tangent bundle of M .

Theorem 1.8 (The ‘manifold of tangent vectors’). The tangent bundle T M has a
canonical differentiable structure making it into a smooth 2n-dimensional manifold, where
n = dim M .

The charts identify any tp∈U Tp M ⊆ T M , for an coordinate neighbourhood U ⊆ M , with

U × Rn . 7 Exercise: check that T M is Hausdorff and second countable (if M is so).

Definition. A (smooth) vector field on a manifold M is a map X : M → T M , such that

(i) X(p) ∈ Tp M for every p ∈ M , and
(ii) in every chart, X is expressed as ai (x) ∂x∂ i with coefficients ai (x) smooth functions
of the local coordinates xi .

Theorem 1.9. Suppose that on a smooth manifold M of dimension n there exist n vector
fields X (1) , X (2) , . . . , X (n) , such that X (1) (p), X (2) (p), . . . , X (n) (p) form a basis of Tp M at
every point p of M . Then T M is isomorphic to M × Rn .

Here ‘isomorphic’ means that T M and M × Rn are diffeomorphic as smooth manifolds

and for every p ∈ M , the diffeomorphism restricts to an isomorphism between the tangent
space Tp M and vector space {p} × Rn . (Later we shall make a more systematic definition
including this situation as a special case.)

Proof. Define π : ~a ∈ T M → p ∈ M if ~a ∈ Tp M ⊂ T M . On the other hand, for any

~a ∈ T M , there is a unique way to write ~a = ai X (i) , for some ai ∈ R. Now define

Φ : ~a ∈ T M → (π(~a); a1 , . . . , an ) ∈ M × Rn .

It is clear from the construction and the hypotheses of the theorem that Φ is a bijection
and Φ converts every tangent space into a copy of Rn . It remains to show that Φ and Φ−1
are smooth.
7
The topology on T M is induced from the smooth structure.
8 alexei kovalev

Using an arbitrary chart ϕ : U ⊆ M → Rn , and the corresponding chart ϕT : π −1 (U ) ⊆

T M → Rn × Rn one can locally express Φ as

(ϕ, idRn ) ◦ Φ ◦ ϕ−1 n n

T : (x, (bj )) ∈ ϕ(U ) × R → (x, (ai )) ∈ ϕ(U ) × R ,

(i)
where xi are local coordinates on U , and ~a = bj ∂x∂ j . Writing X (i) = Xj (x) ∂x∂ j , we obtain
(i) (i)
bj = ai Xj (x), which shows that Φ−1 is smooth. The matrix X(x) = (Xj (x)) expresses a
change of basis of Tp M , from ( ∂x∂ j )p to X (j) (p), and is smooth in x, so the inverse matrix
C(x) is smooth in x too. Therefore ai = bj Cij (x) verifies that Φ is smooth.

Remark . The hypothesis of Theorem 1.9 is rather restrictive. In general, a manifold need
not admit any non-vanishing smooth (or even continuous) vector fields at all (as we shall
see, this is the case for any even-dimensional sphere S 2n ) and the tangent bundle T M will
not be a product M × Rn .

Definition. The differential of a smooth map F : M → N at a point p ∈ M is a linear

map
(dF )p : Tp M → TF (p) N
∂y
given in any charts by (dF )p : ( ∂x∂ i )p 7→ ( ∂xji )(p) ( ∂y∂ j )F (p) . Here xi are local coordi-
nates on M , yj on N , defined respectively by charts ϕ around p and ψ around F (p), and
yj = Fj (x1 , . . . , xn ) for j = 1, . . . , dim N (n = dim M ) is the expression ψ ◦ F ◦ ϕ−1 for F
in these local coordinates.

It follows, by direct calculation in local coordinates using (1.40 ), that the differential is
independent of the choice of charts and that the chain rule d(F2 ◦ F1 )p = (dF2 )F1 (p) ◦ (dF1 )p
holds.
Every (smooth) vector field, say X on M , defines a linear differential operator of first
order X : C ∞ (M ) → C ∞ (M ), according to (1.5) (with p allowed to vary in M ). Suppose
that F : M → N is a diffeomorphism. Then for each vector field X on M , (dF )X is a
well-defined vector field on N . For any f ∈ C ∞ (N ), the chain rule in local coordinates
∂yj
∂xi
(p) ∂y∂ j F (p) f = ∂x∂ i p (f ◦ F ) (xi on M and yj on N ) yields a coordinate-free relation


((dF )X)f ◦ F = X(f ◦ F ), (1.10)

Let X, Y be vector fields regarded as differential operators on C ∞ (M ). Then [X, Y ] =

∂Yi
XY −Y X defines a vector field: its local expression is (Xj ∂xj
−Yj ∂X i
) ∂ . Direct calculation
∂xj ∂xi
shows that [·, ·] satisfies the Jacobi identity and so the space V (M ) of all (smooth) vector
fields on a manifold is an infinite-dimensional Lie algebra.

Left-invariant vector fields

Let G be a Lie group, e ∈ G the identity element, and denote g = Te G. The group
operations can be used to construct non-vanishing vector fields on G as follows.
 differential geometry 9

For every g ∈ G, the left translation Lg : h ∈ G → gh ∈ G is a diffeomorphism of G.

Let ξ ∈ g be a non-zero element. Define
Xξ : g ∈ G → (dLg )e ξ ∈ Tg G ⊂ T G. (1.11)
Then Xξ 6= 0 at any point g ∈ G, for any ξ 6= 0, because the linear map (dLg )e is invertible.
Furthermore, Xξ is a well-defined smooth vector field on G.
To verify the latter claim we consider the smooth map L : (g, h) ∈ G × G → gh ∈ G,
using local coordinate charts ϕg0 : Ug0 → Rm , ϕe : Ue → Rm defined near g0 , e ∈ G,
respectively. Here m = dim G. The local expression of L near (g0 , e) via these charts,
Lloc = ϕg0 ◦ L ◦ (ϕ−1 −1
g0 , ϕe ), is a smooth map Lloc : Ug0 × Ue → Ug0 . Now the local expression
for (dLg )e is just the partial derivative map D2 Lloc , linearizing Lloc in the second m-tuple
of variables. It is clearly smooth in the first m variables, and so (dLg )e ξ depends smoothly
on g.
To sum up, we have proved
Proposition 1.12. If ξ1 , . . . , ξm is a basis of the vector space g then Xξ1 (h), . . . , Xξm (h)
define m = dim G vector fields whose values at each h ∈ G give a basis of Th G.
Hence, in view of Theorem 1.9, we obtain
Theorem 1.13. The tangent bundle T G of any Lie group G is isomorphic to the product
G × Rdim G .
The smoothness of Xξ and (1.11) together imply that (dLg )h Xξ (h) = Xξ (gh), i.e.
(dLg )Xξ = Xξ ◦ Lg . (1.14)
and we shall call any vector field satisfying (1.14) left-invariant.
It follows from Proposition 1.12 that the vector space l(G) of all the left-invariant vector
fields on G form a finite-dimensional subspace (of dimension m) in the space V (G) of all
vector fields. In fact more is true.
Theorem 1.15. The space of all the left-invariant vector fields on G is a finite-dimensional
Lie algebra, hence a Lie subalgebra of V (G).
Proof. The theorem may be restated as saying that for every pair Xξ , Xη of left-invariant
vector fields, [Xξ , Xη ] is again left-invariant. To this end, we calculate using (1.10)
(dLg )[Xξ , Xη ]f ◦ Lg = [Xξ , Xη ](f ◦ Lg ) = Xξ Xη (f ◦ Lg ) − Xη Xξ (f ◦ Lg )



= Xξ (dLg )Xη f ◦ Lg − Xη (dLg )Xξ f ◦ Lg
= (dLg )Xξ ((dLg )Xη f ) ◦ Lg − (dLg )Xη ((dLg )Xξ f ) ◦ Lg
= [(dLg )Xξ , (dLg )Xη ]f ◦ Lg

and using the left-invariant property (1.14) of Xξ and Xη for the next step

= [Xξ ◦ Lg , Xη ◦ Lg ]f ◦ Lg = ([Xξ , Xη ] ◦ Lg )f ◦ Lg .
Thus (dLg )[Xξ , Xη ]f ◦ Lg = ([Xξ , Xη ] ◦ Lg )f ◦ Lg , for each g ∈ G, f ∈ C ∞ (G), so the vector
field [Xξ , Xη ] satisfies (1.14) and is left-invariant.
10 alexei kovalev

It follows that [Xξ , Xη ] = Xζ for some ζ ∈ g, thus the Lie bracket on l(G) induces one
on g. We can identify this Lie bracket more explicitly for the matrix Lie groups.
Theorem 1.16. If G is a matrix Lie group, then the map ξ ∈ g → Xξ ∈ l(G) is an
isomorphism of the Lie algebras, where the Lie bracket on g is as defined in Theorem 1.7.
We shall prove Theorem 1.16 in the next section.

1.4 Submanifolds
Suppose that M is a manifold, N ⊂ M , and N is itself a manifold, denote by ι : N → M
the inclusion map.
Definition. N is said to be an embedded submanifold8 of M if
(i) the map ι is smooth;
(ii) the differential (dι)p at any point of p ∈ N is an injective linear map;
(iii) ι is a homeomorphism onto its image, i.e. a D ⊆ N is open in the topology of
manifold N if and only if D is open in the topology induced on N from M (i.e. the open
subsets in N are precisely the intersections with N of the open subsets in M ).

Remark . Often the manifold N is not given as a subset of M but can be identified with a
subset of M by means of an injective map ψ : N → M . In this situation, the conditions in
the above definition make sense for ψ(N ) (regarded as a manifold diffeomorphic to N ). If
these conditions hold for ψ(N ) then one says that the map ψ embeds N in M and writes
ψ : N ,→ M .
Example. A basic example of embedded submanifold is a (parameterized) curve or surface
in R3 . Then the condition (i) means that the parameterization is smooth and (ii) means
that the parameterization is regular (cf. introductory remarks on p.1).
Remark . A map ι satisfying conditions (i) and (ii) is called an immersion and respectively
N is said to be an immersed submanifold. The condition (iii) eliminates e.g. the irrational
twist flow t ∈ R → [(t, αt)] ∈ R2 /Z2 = S 1 × S 1 , α ∈ R \ Q, on the torus.
A surface or curve in R3 (or more generally in Rn ) is often defined by an equation or
a system of equations, i.e. as the zero locus of a smooth map on R3 (respectively on Rn ).
E.g. x2 + y 2 − 1 = 0 (a circle) or (x2 + y 2 + b2 + z 2 − a2 )2 − 4b2 (x2 + y 2 ) = 0, b > a > 0
(a torus). However a smooth (even polynomial) map may in general have ‘bad’ points in
its zero locus (cf. Example Sheet 1, Q.8). When does a system of equations on a manifold
define a submanifold?
Definition. A value q ∈ N of a smooth map f between manifolds M and N is called a
regular value if for any p ∈ M such that f (p) = q the differential of f at p is surjective,
(df )p (Tp M ) = Tq N .
Theorem 1.17. Let f : M → N be a smooth map between manifolds and q ∈ N a regular
value of f . The inverse image of a regular value P = f −1 (q) = {p ∈ M : f (p) = q}
(if it is non-empty) is an embedded submanifold of M , of dimension dim M − dim N .
8
In these notes I will sometimes write ‘submanifold’ meaning ‘embedded submanifold’.
 differential geometry 11

We shall need the following result from advanced calculus.

Inverse Mapping Theorem. Suppose that f : U ⊆ Rn → Rn is a smooth map defined
on an open set U , 0 ∈ U and f (0) = 0. Then f has an smooth inverse g, defined on some
neighbourhood of 0 with g(0) = 0, if and only if (df )0 is an invertible linear map of Rn .
Note that the Inverse Mapping Theorem, as stated above, is a local result, valid only if
one restricts attention to a suitably chosen neighbourhood of a point. The statement will
in general no longer hold with neighbourhoods replaced by manifolds. (Consider e.g. the
map of R to the unit circle S 1 ⊂ C given by f (x) = eix .)
Proof of Theorem 1.17. Firstly, P is Hausdorff and second countable because M is so.
Let p be an arbitrary point of P . We may assume without loss of generality that
there are local coordinates xi , yj , i = 1, . . . , n, j = 1, . . . , k = dim N , n + k = dim M ,
defined in a neighbourhood of p in M and local coordinates defined in a neighbourhood
of f (p) in N such that xi (p) = yj (p) = 0 and f is expressed in these local coordinates
as a map f = (f1 , . . . , fk ) on a neighbourhood of 0 in Rn with values in Rk , f (0) = 0,
det(∂fi /∂yj )(0, 0) 6= 0.
Then  
1 0
det (0, 0) 6= 0
(∂fi /∂xi ) (∂fi /∂yj )
and so, by the Inverse Function Theorem, the xi , fj form a valid set of new local coordinates
on a (perhaps smaller) neighbourhood of p in M . The local equation for the intersection of
P with that neighbourhood takes in the new coordinates a simple form fj = 0, j = 1, . . . , k.
Furthermore the first projection (xi , fj ) 7→ (xi ) restricts to a homeomorphism from a
neighbourhood of p in P onto a neighbourhood of zero in Rn . Define this first projection
to be a coordinate chart on P with xi the local coordinates. Then the family of all such
charts covers P and it remains to verify that any two charts defined in this way are in fact
compatible.
So let xi , fj , x0i , fj0 be two sets of local coordinates near p as above. Then, by the
construction, for every p in P , we have that x0i = x0i (x, f ), fj0 = fj0 (x, f ) with fj0 (x, 0) = 0
identically in x. Therefore, (∂fi0 /∂xj )(0, 0) = 0. Then
 0
∂xi0 /∂xi ∂x0i0 /∂fj
 0
∂xi0 /∂xi ∂x0i0 /∂fj
 
det (0, 0) = det (0, 0) 6= 0,
∂fj00 /∂xi ∂fj00 /∂fj 0 ∂fj00 /∂fj

so we must have det(∂x0i0 /∂xj )(0) 6= 0 for the n × n Jacobian matrix, and the change from
xi ’s to x0i0 ’s is a diffeomorphism near 0.
Remarks. 1. It is not true that every submanifold of M is obtainable as the inverse image
of a regular value for some smooth map on M . One counterexample is RP 1 ,→ RP 2
(an exercise: check this!).
2. Sometimes in the literature one encounters a statement ‘a subset P is (or is not) a
submanifold of M ’. Every subset P of a manifold M has a topology induced from M . It
turns out that there is at most one smooth structure on the topological space P such that
P is an embedded submanifold of M , but I shall not prove it here.
12 alexei kovalev

Theorem 1.18 (Whitney embedding theorem). Every smooth n-dimensional manifold can
be embedded in R2n (i.e. is diffeomorphic to a submanifold of R2n ).
We shall assume the Whitney embedding theorem without proof here (note however,
Examples Sheet 1 Q.9). A proof of embedding in R2n+1 is e.g. in Guillemin and Pollack,
Ch.1 §9.
It is worth to remark that the possibility of embedding any manifold in some RN , with
N possibly very large, does not particularly simplify the study of manifolds in practical
terms (but is relatively easier to prove). The essence of the Whitney embedding theorem is
the minimum possible dimension of the ambient Euclidean space as way of measuring the
‘topological complexity’ of the manifold. The result is sharp, in that the dimension of the
ambient Euclidean space could not in general be lowered (as can be checked by considering
e.g. the Klein bottle).
We can now give, as promised, a proof of Theorem 1.16
Theorem 1.16. Suppose that G ⊂ GL(n, R) is a subgroup and an embedded submani-
fold of GL(n, R), and smooth structure on G is defined by the log-charts. Then the map
ξ ∈ g → Xξ ∈ l(G) is an isomorphism of the Lie algebras, where the Lie bracket on g is
the Lie bracket of matrices, as in Theorem 1.7.
Proof of Theorem 1.16. We want to show [Xξ , Xη ] = X[ξ,η] for a matrix Lie group where
the LHS is the Lie bracket of left-invariant vector fields and the RHS is defined using
Theorem 1.7. Note first that for G = GL(n, R), all the calculations can be done on an
2
open subset of Rn = Matr(n, R) with coordinates xij , i.j = 1. . . . , n. (The coordinate
chart on GL(n, R) given by the identity map is compatible, i.e. belongs to the same C ∞
structure, with the log-charts.) The map Lg , hence also dLg , is a linear map given by the
left multiplication by a fixed matrix g = (xij ). We obtain that the left-invariant vector
fields are Xξ (g) = xik ξjk ∂x∂ i . The coefficients of Xξ depend linearly on the coordinates xik
j
and [Xξ , Xη ] = X[ξ,η] follows by an elementary calculation.
In the above, we considered the coefficients of the left-invariant vector fields at I ∈
GL(n) using the basis of ‘coordinate vector fields’ corresponding to the identity chart.
On the other hand, the Lie bracket of matrices in Theorem 1.7 is defined in terms of the
basis corresponding to the log-chart. Recall from §1.2 that the log-chart around I is a
diffeomorphism
log : UI ⊂ GL(n) → V0 ⊂ Matr(n) (1.19)
between an open neighbourhood UI of I and an open neighbourhood V0 of the zero matrix.
For GL(n, R), the map (1.19) may also be considered as a change of local coordinates.
The derivative (the Jacobi matrix) of log A at A = I ∈ UI is the identity matrix and so
the bases of coordinate vector fields corresponding to the two choices of local coordinates
coincide at I.
Theorem 1.16 therefore holds for GL(n, R) (the argument also applies with some change
of notation to GL(n, C)).
Now consider a general case G ⊂ GL(n, R). Denote, as before, the inclusion map by ι.
Any left-invariant vector field Xξ on G (ξ ∈ g) may be identified by means of dι with a
 differential geometry 13

vector field defined on a subset G ⊂ GL(n, R). Further, the left translation Lg on G is
the restriction of the left translation on GL(n, R) (if g ∈ G). We find that the vector
fields (dι)Xξ (ξ ∈ g) correspond bijectively to the restrictions to G ⊂ GL(n, R) of the
left-invariant vector fields Xξ on GL(n, R), such that ξ ∈ g. Let Xξ , Xη ∈ l(GL(n, R))
with ξ, η ∈ g, where g is understood as the image of log-chart for G near I. We have
[Xξ , Xη ] = X[ξ,η] , by the above calculations on GL(n, R), and the Lie bracket of matrices
[ξ, η] ∈ g by Theorem 1.7. Therefore, X[ξ,η] restricts to give a well-defined left-invariant
vector field on G, and [Xξ , Xη ] = X[ξ,η] holds for any Xξ , Xη ∈ l(G) as claimed.

Remark. Notice that the differential dι considered in the above proof identifies g with a
2
linear subspace of n × n matrices, by considering G as a hypersurface in GL(n, R) ⊂ Rn .
In the example of G = O(n), for any path A(t) or orthogonal matrices with A(0) = I, we
may compute 0 = dtd t=0 A(t)A∗ (t) = Ȧ(0)A∗ (0) + A(0)Ȧ∗ (0) = Ȧ(0) + Ȧ∗ (0), i.e. Ȧ(0) is
skew-symmetric. Thus the log chart at the identity actually maps onto a neighbourhood
of zero in the tangent space TI G —which explains why the log-chart construction in §1.2
worked.

1.5 Exterior algebra of differential forms. De Rham cohomology

See the reference card on multilinear algebra

The differential forms

Consider a smooth manifold M of dimension n. The dual space to the tangent space Tp M ,
p ∈ M , is called the cotangent space to M at p, denoted Tp∗ M . Suppose that a chart and
hence the local coordinates are given on a neighbourhood of p. The dual basis to ( ∂x∂ i )p is
traditionally denoted by (dxi )p , i = 1, . . . , n. (Sometimes I may drop P
the subscript p from
the notation.) Thus an arbitrary element of Tp∗ M is expressed as i ai (dxi )p for some
ai ∈ R.
Recall from (1.40 ) that a change of local coordinates, say from from xi to x0i , induces
a change of basis of tangent space Tp M from ∂x∂ i to ∂x∂ 0 . By linear algebra, there is a
i
corresponding change of the dual basis of Tp∗ M , from dx0i to dxi given by the transposed
matrix. Thus the transformation law is
∂x0j
dx0j = dxi . (1.20)
∂xi

Like (1.40 ) the eq.(1.20) is a priori merely a notation which resembles familiar results from
the calculus. See however further justification in the remark on the next page.
A disjoint union of all the cotangent spaces T ∗ M = tp∈M Tp∗ M of a given manifold
M is called the cotangent bundle of M . The cotangent bundle can be given a smooth
structure making it into a manifold of dimension 2 dim M by an argument very similar
to one for the tangent bundle (but with a change of notation, replacing any occurrence
of (1.40 ) with (1.20)).
14 alexei kovalev

A smooth field of linear functionals is called a (smooth) differential 1-form (or just
1-form). More precisely, a differential 1-form is a map α : M → T ∗ M such that αp ∈ Tp∗ M
for every p ∈ M and α is expressed in any local coordinates x = (x1 , . . . , xn ) by α =
P
i ai (x)dxi where ai (x) are some smooth functions of x.
Remark . The 1-forms are of course the dual objects to the vector fields. In particular,
ai (x(p)) is obtained as the value of αp on the tangent vector ( ∂x∂ i )p . Consequently, α is
smooth on M if and only if α(X) is a smooth function for every vector field X on M .
Notice that the latter condition does not use local coordinates.
One can similarly consider a space Λr Tp∗ M of alternating multilinear functions on Tp M ×
. . . × Tp M (r factors), for any r = 0, 1, 2, . . . n, and proceed to define the r-th exterior power
Λr T ∗ M of the cotangent bundle of M and the (smooth) differential r-forms on M .
Details are left as an exercise.
The space of all the smooth differential r-forms on M is denoted by Ωr (M ) and r is
referred to as the degree of a differential form. If r = 0 then Λ0 T ∗ M = M × R and
Ω0 (M ) = C ∞ (M ). The other extreme case r = dim M is more interesting.
Theorem 1.21 (Orientation of a manifold). Let M be an n-dimensional manifold. The
following are equivalent:
(a) there exists a nowhere vanishing smooth differential n-form on M ;
(b) there exists a family of charts in the differentiable structure on M such that the re-
spective coordinate domains cover M and the Jacobian matrices have positive determinants
on every overlap of the coordinate domains;
(c) the bundle of n-forms Λn T ∗ M is isomorphic to M × R.
Proof (gist). That (a)⇔(c) is proved similarly to the proof of Theorem 1.9.
Using linear algebra, we find that the transformation of the differential forms of top
degree under a change of coordinates is given by
∂xj
0
dx1 ∧ . . . ∧ dxn = det dx1 ∧ . . . ∧ dx0n .
∂x0i
Now (a)⇒(b) is easy to see.
To obtain, (b)⇒(a) we assume the following.
Theorem 1.22 (Partition of unity). For any open cover M ⊂ ∪α∈A Uα , there exists a
countable collection of functions ρi ∈ C ∞ (M ), i = 1, 2, ..., such that the following holds:
(i) for any i, the closure of supp(ρi ) = {x ∈ M : ρi (x) 6= 0} is compact and contained in
Uα for some α = αi (i.e. depending on i);
(ii) the collection is locally finite: each x ∈ M has a neighbourhood Wx such that ρi (x) 6= 0
on Wx for only finitely many i;P and
(iii) ρi ≥ 0 on M for all i and i ρi (x) = 1 for all x ∈ M .
The collection {ρi } satisfying the above is called a partition of unity subordinate to {Uα }.
Choose a partition of unity {ρi } subordinate to the given family of coordinate neigh-
(α)
bourhoods covering M . For each i, choose local coordinates xi valid on the support
 differential geometry 15

(α) (α)
of ρi . Define ωα = dx1 ∧ . . . ∧ dxn in these local coordinates, then ρi ωαi is a well-
defined (smooth)
P n-form on all of M (extended by zero outside the coordinate domain)
and ω = i ρi ωαi is the required n-form.
A manifold M satisfying any of the conditions (a),(b),(c) of the above theorem is
called orientable. A choice of the differential form in (a), or family of charts in (b),
or diffeomorphism in (c) defines an orientation of M and a manifold endowed with an
orientation is said to be oriented.

Exterior derivative
Recall that the differential of a smooth map f : M → N between manifolds is a linear map
between respective tangent spaces. In the special case N = R, the (df )p at each p ∈ M is
a linear functional on Tp M , i.e. an element of the dual space Tp∗ M . In local coordinates
∂f
xi defined near p we have df (x) = ∂x i
(x)dxi , thus df is a well-defined differential 1-form,
whose coefficients are those of the gradient of f .
Remark . Observe that any local coordinate xi on an open domain U ⊂ M is a smooth
function on U . Then the formal symbols dxi actually make sense as the differentials of
these smooth functions (which justifies the previously introduced notation, cf.(1.20)).
Theorem 1.23 (exterior differentiation). There exists unique linear operator
k k+1
d : Ω (M ) → Ω (M ), k ≥ 0, such that
(i) if f ∈ Ω0 (M ) then df coincides with the differential of a smooth function f ;
(ii) d(ω ∧ η) = dω ∧ η + (−1)deg ω ω ∧ dη for any two differential forms ω,η;
(iii) ddω = 0 for every differential form ω.
Proof (gist). On an open set U ⊆ Rn , or in the local coordinates on a coordinate domain
on a manifold, application of conditions (ii), then (iii) and (i), yields
∂f
d(f dxi1 ∧ . . . ∧ dxir ) = df ∧ dxi1 ∧ . . . ∧ dxir = dxi ∧ dxi1 ∧ . . . ∧ dxir , (1.24)
∂xi
for any smooth function f . Extend this to arbitrary differential forms by linearity. The
conditions (i),(ii),(iii) then follow by direct calculation, in particular the last of these holds
by independence of the order of differentiation in second partial derivatives. This proves
the uniqueness, i.e. that if d exists then it must be expressed by (1.24) in local coordinates.
Observe another important consequence of (1.24): the operator d is necessarily local,
which means that the value (dω)p at a point p is determined by the values of differential
form ω on a neighbourhood of p.
To establish the existence of d one now needs to show that the defining formula (1.24)
is consistent, i.e. the result of calculation does not depend on the system of local coordi-
nates in which it is performed. So let d0 denote the exterior differentiation constructed as
in (1.24), but using different choice of local coordinates. Then, by (ii), we must have
r
X
d0 (f dxi1 ∧ . . . ∧ dxir ) = d0 f ∧ dxi1 ∧ . . . ∧ dxir + (−1)j+1 f dxi1 ∧ . . . d0 (dxij ) . . . ∧ dxir .
j=1
16 alexei kovalev

But d0f = df and d0 (dxk ) = d0 (d0 xk ) = 0, by (i) and (iii) and because we know that
the differential of a smooth function (0-form) is independent of the coordinates. Hence the
∂f
right-hand side of the above equality becomes ∂x j
dxj ∧dxi1 ∧. . .∧dxir = d(f xi1 ∧. . .∧dxir ),
and so the exterior differentiation is well-defined.

De Rham cohomology
A differential form α is said to be closed when dα = 0 and exact when α = dβ for some
differential form β. Thus exact forms are necessarily closed (but the converse is not in
general true, e.g. Example Sheet 2, Q4).
Definition. The quotient space
closed k-forms on M
k
HdR (M ) =
exact k-forms on M
is called the k-th de Rham cohomology group of the manifold M .
Any smooth map between manifolds, say f : M → N , induces a pull-back map
between exterior powers of cotangent spaces f ∗ : Λr Tf∗(p) N → Λr Tp∗ M (r = 0, 1, 2, . . .),
which is a linear map defined, for any differential r-form α on N , by
(f ∗ α)p (v1 , . . . , vr ) = αf (p) ((df )p v1 , . . . , (df )p vr ),
using the differential of f . The chain rule for differentials of smooth maps immediately
gives
(f ◦ g)∗ = g ∗ f ∗ . (1.25)
It is also straightforward to check that f ∗ preserves the ∧-product f ∗ (α∧β) = (f ∗ α)∧(f ∗ β)
and f ∗ commutes with the exterior differentiation, f ∗ (dα) = d(f ∗ α), hence f ∗ preserves the
subspaces of closed and exact differential forms. Therefore, every smooth map f : M → N
induces a linear map on the de Rham cohomology
f ∗ : H r (N ) → H r (M )
A consequence of the chain rule (1.25) is that if f is a diffeomorphism then f ∗ is a linear
isomorphism. Thus the de Rham cohomology is a diffeomorphism invariant, i.e. diffeo-
morphic manifolds have isomorphic de Rham cohomology.9
Poincaré lemma. H k (D) = 0 for any k > 0, where D denotes the open unit ball in Rn .
The proof goes by working out a way to invert the exterior derivative. More precisely,
one constructs linear maps hk : Ωk (U ) → Ωk−1 (U ) such that
hk+1 ◦ d + d ◦ hk = idΩk (U ) .
Remark . In the degree 0, one has H 0 (M ) = R for any connected manifold.
9
In fact, more is true. It can be shown, using topology, that the de Rham cohomology depends only on
the topological space underlying a smooth manifold and that homeomorphic manifolds have isomorphic de
Rham cohomology. The converse in not true: there are manifolds with isomorphic de Rham cohomology
but e.g. with different fundamental groups.
 differential geometry 17

Basic integration on manifolds

Throughout this subsection M is an oriented n-dimensional manifold. Let ω ∈ Ωn (M )
and, as before, denote supp ω = {p ∈ M : ωp 6= 0} (the support of ω). Suppose that the
closure of supp ω is compact (we then sometimes slightly inaccurately say ω has a compact
support).
Consider first the special case when the closure of supp ω is contained in the domain
of just one coordinate chart, (U, ϕ) say. Then there is a unique smooth function f (x) on
ϕ(U ) ⊂ Rn such that f (x)dx1 ∧. . .∧dxn is the local expression for ω and
R the xi ’s appear in a
‘positive’ order determined by the orientation of M . Then the integral ϕ(Uα ) f (x)dx1 . . . dxn
makes sense as in the multivariate calculus and the value of this integral is independent of
the choice of ‘positively oriented’ local coordinates. This is because the respective Jacobian
will be positive, so the local expression for ω changes precisely as required by the change
of variables formula for integrals
R of functions
R of n variables (which involves the absolute
value of the Jacobian). Thus M ω = U ω is well defined when ω supported in just one
coordinate chart. Observe also that if one uses the other orientation of M , then the integral
changes sign.
Now let ω be any n-form with compact support. Consider an oriented system of charts
(Uα , ϕα ) covering M (as in Theorem 1.21(b)). Let (ρi ) be a partition of unity subordinate
to {Uα }.

Definition. In the above situation, the integral of ω over M is given by

Z XZ
ω= ρi ω.
M i U αi

Note that the sum in the right-hand side may be assumed to have only finitelyR many
non-zero terms (by the compact support assumption). It can be checked that M ω does
not depend on the choice of the open cover {Uα } or a partition of unity on M , and is
therefore well-defined.
Stokes’ Theorem (for manifolds R without boundary)10 . Suppose that η ∈ Ωn−1 (M )
has a compact support. Then M dη = 0.
Let ρα be a partition of unity subordinate to some oriented system of coordinate neigh-
bourhoods covering M (as above). Then dη is a finite sum of the forms ρα η. Now the proof
of Stokes’ Theorem can be completed by considering compactly supported exact forms on
Rn and using calculus.

Corollary 1.26 (Integration by parts). Suppose that α and β are compactly supported
differential deg αR+ deg β = dim M − 1.
R forms on M and1+deg α
Then M α ∧ dβ = (−1) M
(dα) ∧ β.

One rather elegant application of the results discussed above is the following.

Theorem 1.27. Every (smooth) vector field on S 2m vanishes at some point.

10
Manifolds with boundary are not considered in these lectures.
18 alexei kovalev

Proof of Theorem 1.27. Notation: recall that we write S n ⊂ Rn+1 for the unit sphere about
the origin. For r > 0, let ι(r) : x ∈ S n ,→ rx ∈ Rn+1 denote the embedding of S n in the
Euclidean space as the sphere of radius r about the origin and write S n (r) = ι(r)(S n )
(thus, in particular, S n (1) = S n ).
Suppose, for a contradiction, that X(x) is a nowhere-zero vector field on S n . We
may assume, without loss of generality, that |X| = 1, identically on S n . For any real
parameter ε, define a map
f : x ∈ Rn+1 \ {0} → x + ε|x|X(x/|x|) ∈ Rn+1 \ {0}.
Here we used the inclusion S n ⊂ Rn+1 (and hence Tp S n ⊂ Tp Rn+1 ∼
= Rn+1 ) to define a
(smooth) map x ∈ Rn+1 \ {0} → X(x/|x|) ∈ Rn+1 \ {0}.
Step 1.
We claim that f is a diffeomorphism, whenever |ε| is sufficiently small. Firstly, for any
x0 6= 0,  
(df )x0 = idRn+1 +ε d(|x|X(x/|x|)) x0
and straightforward
 calculus
 shows that the norm of the linear map defined by the Jacobi
matrix d(|x|X(x/|x|)) x0 is bounded independent of x0 6= 0. Hence there is ε0 > 0, such
that (df )x0 is a linear isomorphism of Rn+1 onto itself for any x0 6= 0 and any |ε| < ε0 . But
then, by the Inverse Mapping Theorem (page 10), for any x 6= 0, f maps some open ball
B(x, δx ) of radius δx > 0 about x diffeomorphically onto its image.
Furthermore, it can be checked, by inspection of the proof of the Inverse Mapping The-
orem, that (1) δx can be taken to be continuous in x and (2) δx can be chosen independent
of ε if |ε| < ε0 . We shall assume these two latter claims without proof. Consequently, δx
can be taken to depend only on |x| (as S n (|x|) is compact).
Taking a smaller ε0 > 0√if necessary, we ensure that f is one-to-one if |ε| < ε0 . For the
latter, note that |f (x)| = 1 + ε2 |x| and so it suffices to check that that f is one-to-one
on each S n (|x|). But two points on S n (|x|) far away from each other cannot be mapped
to one because |f (x) − x| < ε|x| and two distinct points at a distance less than say 21 δ|x|
cannot be mapped to one because f restricts to a diffeomorphism (hence a bijection) on a
δ|x| -ball about each point.
A similar reasoning shows that f is surjective (onto) if ε in sufficiently small. Indeed,
f (B(x, δx )) is an open set homeomorphic to a ball and the boundary of f (B(x, δx )) is
within small distance ε(1 + δx )|x| from the boundary of B(x, δx ). Therefore, x must be
inside the boundary of f (B(x, δx )) and thus in the image of f . In all of the above, ‘small
ε0 ’ can be chosen independent of x because δx depends only on |x| and f is homogeneous
of degree 1, f (λx) = λf (x) for each positive λ.
Step 2. Now, as f is a diffeomorphism f maps the embedded submanifold S n (1)
√
diffeomorphically onto the embedded submanifold f (S n (1)) = S n ( 1 + ε2 ). Consider a
differential n-form on Rn+1
n
X
ω= (−1)i xi dx0 ∧ . . . (omit dxi ) . . . ∧ dxn ,
i=0
 differential geometry 19

We have f (S n (1)) ω = S n (1) f ∗ ω (change of variable formula) and this integral depends
R R

polynomially on the parameter ε because f ∗ ω does so (as f is linear in ε).

But on the other hand, for any r > 0,
Z Z Z Z
∗
ω− ω= ι(r) ω − ω
S n (r) S n (1) Sn Sn
Z r Z 
d ∗
= ι(s) ω ds
1 ds Sn
Z r Z 
d ∗


= ι(s) ω ds
1 S n ds

applying, on each coordinate patch, a theorem on differentiation of an integral depending

on a parameter s, from calculus
Z
= dω
1≤|x|≤r

replacing, again on each coordinate patch, a repeated integration with an (n+1)-dimensional

integral
Z
= (n + 1)dx0 ∧ . . . ∧ dxn = cn+1 (rn+1 − 1),
1≤|x|≤r

where a constant cn+1 is n + 1 times the √ volume of (n + 1)-dimensional ball (the value
of cn+1 does not matter here). Put r = 1 + ε2 and then the right-hand side is not a
polynomial in ε if n + 1 is odd (i.e. when n is even). A contradiction.

Some page references

to Warner, and Guillemin–Pollack, and Gallot–Hulin–Lafontaine.

N.B. The material in these books is sometimes covered differently from the Lectures and
may contain additional topics, thus the references are not quite ‘one-to-one’.
smooth manifolds [W] 1.2–1.6
tangent and cotangent bundles [W] 1.25
exponential map on a matrix Lie group [W] 3.35
left invariant vector fields [W] 3.6–3.7
submanifolds [W] 1.27–1.31, 1.38
differential forms [GP] 153–165,174–178
de Rham cohomology [GP] pp.178–182
Poincaré Lemma [W] 4.18
partition of unity [GP] p.52 or [W] 1.8–1.11
integration and Stokes’ Theorem [GP] 165–168, 183–185
non-existence of vector fields without zeros on S 2n (Milnor’s proof) [GHL] 1.41
Part III: Differential geometry (Michaelmas 2004)

Alexei Kovalev ([email protected])

2 Vector bundles.
Definition. Let B be a smooth manifold. A manifold E together with a smooth submer-
sion1 π : E → B, onto B, is called a vector bundle of rank k over B if the following
holds:
(i) there is a k-dimensional vector space V , called typical fibre of E, such that for any
point p ∈ B the fibre Ep = π −1 (p) of π over p is a vector space isomorphic to V ;

(ii) any point p ∈ B has a neighbourhood U , such that there is a diffeomorphism

Φ
π −1 (U ) −−−
U
→ U ×V
 

πy
pr
y 1
U U
and the diagram commutes, which means that every fibre Ep is mapped to {p} × V .
ΦU is called a local trivialization of E over U and U is a trivializing neighbour-
hood for E.

(iii) ΦU |Ep : Ep → V is an isomorphism of vector spaces.

Some more terminology: B is called the base and E the total space of this vector bundle.
π : E → B is said to be a real or complex vector bundle corresponding to the typical fibre
being a real or complex vector space. Of course, the linear isomorphisms etc. are understood
to be over the respective field R or C. In what follows vector bundles are taken to be real
vector bundles unless stated otherwise.
Definition. Any smooth map s : B → E such that π ◦ s = idB is called a section of E.
If s is only defined over a neighbourhood in B it is called a local section.

Examples. 0. A trivial, or product, bundle E = B × V with π the first projection.

Sections of this bundle are just the smooth maps C ∞ (B; V ).
1. The tangent bundle T M of a smooth manifold M has already been discussed in
Chapter 1. It is a real vector bundle of rank n = dim M which in general is not trivial.2
The sections of T M are the vector fields. In a similar way, the cotangent bundle T ∗ M and,
1
A smooth map is called a submersion if its differential is surjective at each point.
2
Theorems 1.9 and 1.26 in Chapter 1 imply that T S 2n cannot be trivial.

20
 differential geometry 21

p ∗
more

generally, the bundle of differential p-forms Λ T M are real vector bundles of rank
n
p
with sections being the differential 1-forms, respectively p-forms. Exercise: verify that
the vector bundles Λp T ∗ M (1 ≤ p ≤ dim M ) will be trivial if T M is so.
2. ‘Tautological vector bundles’ may be defined over projective spaces RP n , CP n (and,
more generally, over the Grassmannians). Let B = CP n say. Then let E be the disjoint
union of complex lines through the origin in Cn+1 , with π assigning to a point in p ∈ E
the line ` containing that point, so π(p) = ` ∈ CP n . We shall take a closer look at one
example (Hopf bundle) below and show that the tautological construction indeed gives a
well-defined (and non-trivial) complex vector bundle of rank 1 over CP 1 .

Structure group of a vector bundle.

It follows from the definition of a vector bundle E that one can define over the intersection
of two trivializing neighbourhoods Uβ , Uα a composite map

Φβ ◦ Φ−1
α (b, v) = (b, ψβα (b)v),

(b, v) ∈ (Uβ ∩ Uα ) × Rk . For every fixed b the above composition is a linear isomorphism
of Rk depending smoothly on b. The maps ψβα : Uβ ∩ Uα → GL(k, R). are called the
transition functions of E.
It is not difficult to see that transition functions ψαβ satisfy the following relations,
called ‘cocycle conditions’ 
ψαα = idRk , 
ψαβ ψβα = idRk , (2.1)
ψαβ ψβγ ψγα = idRk .


The left-hand side is defined on the intersection Uα ∩ Uβ , for the second of the above
equalities, and on Uα ∩ Uβ ∩ Uγ for the third. (Sometimes the name ‘cocycle condition’
refers to just the last of the equalities (2.1); the first two may be viewed as notation.)
Now it may happen that a vector bundle π : E → B is endowed with a system of trivi-
alizing neighbourhoods Uα covering the base and such that all the corresponding transition
functions ψβα take values in a subgroup G ⊆ GL(k, R), ψβα (b) ∈ G for all b ∈ Uβ ∩ Uα , for
all α, β, where k is the rank of E. Then this latter system {(Uα , Φα )} of local trivializations
over Uα ’s is said to define a G-structure on vector bundle E.

Examples. 0. If G consists of just one element (the identity) then E has to be a trivial
bundle E = B × Rk .
1. Let G = GL+ (k, R) be the subgroup of matrices with positive determinant. If the
typical fibre Rk is considered as an oriented vector space then the transition functions ψβα
preserve the orientation. The vector bundle E is then said to be orientable.
A basic example arises from a system of coordinate charts giving an orientation of a
manifold M . The transition functions of T M are just the Jacobians and so M is orientable
precisely when its tangent bundle is so.
2. A more interesting situation occurs when G = O(k), the subgroup of all the non-
singular linear maps in GL(k, R) which preserve the Euclidean inner product on Rk . It
22 alexei kovalev

follows that the existence of an O(k) structure on a rank k vector bundle E is equivalent
to a well-defined positive-definite inner product on the fibres Ep . This inner product
is expressed in any trivialization over U ⊆ B as a symmetric positive-definite matrix
depending smoothly on a point in U .
Conversely, one can define a vector bundle with inner product by modifying the defi-
nition on page 20: replace every occurrence of ‘vector space’ by ‘inner product space’ and
(linear) ‘isomorphism’ by (linear) ‘isometry’. This will force all the transition functions to
take value in O(k) (why?).
A variation on the theme: an orientable vector bundle with an inner product is the
same as vector bundle with an SO(k)-structure.
3. Another variant of the above: one can play the same game with rank k complex
vector bundles and consider the U (k)-structures (U (k) ⊂ GL(k, C)). Equivalently, consider
complex vector bundles with Hermitian inner product ‘varying smoothly with the fibre’.
Furthermore, complex vector bundles themselves may be regarded as rank 2k real vector
bundles with a GL(k, C)-structure (the latter is usually called a complex structure on a
vector bundle).

In the examples 2 and 3, if a trivialization Φ is ‘compatible’ with the given O(k)- or

SO(k)-structure (respectively U (k)-structure) {(Uα , Φα )} in the sense that the transition
functions Φα ◦ Φ−1 take values in the orthogonal group (respectively, unitary group) then
Φ is called an orthogonal trivialization (resp. unitary trivialization).

Principal bundles.
Let G be a Lie group. A smooth free right action of G on a manifold P is a smooth
map P × G → P , (p, h) 7→ ph, such that (1) for any p ∈ P , ph = p if and only if h is the
identity element of G; and (2) (ph1 )h2 = p(h1 h2 ) for any p ∈ P , any h1 , h2 ∈ G. (It follows
that for each h ∈ G, P × {h} → P is a diffeomorphism.)

Definition . A (smooth) principal G-bundle P over B is a smooth submersion

π : P → B onto a manifold B, together with a smooth right free action P × G → P , such
that the set of orbits of G in P is identified with B (as a set), P/G = B, and also for any
b ∈ B there exists a neighbourhood U ⊆ B of b and a diffeomorphism ΦU : π −1 (U ) → U ×G
such that pr1 ◦ΦU = π|π−1 (U ) , i.e. the following diagram is commutative

Φ
π −1 (U ) −−−
U
→ U ×G
 

πy
pr
y 1 (2.2)

U U
and ΦU commutes with the action of G, i.e. for each h ∈ G, ΦU (ph) = (b, gh), where
(b, g) = ΦU (p), π(p) = b ∈ U .
A local section of the principal bundle P is a smooth map s : U → P defined on a
neighbourhood U ⊂ B and such that π ◦ s = idU .
 differential geometry 23

For a pair of overlapping trivializing neighbourhoods Uα , Uβ one has

Φβ ◦ Φ−1
α (b, g) = (b, ψβα (b, g)),

where for each b ∈ Uα ∩ Uβ , the ψβα (b, ·) is a map G → G. Then, for each b in the domain
of ψβα , we must have ψβα (b, g)h = ψβα (b, gh) for all g, h ∈ G, in view of (2.2). It follows (by
taking g to be the unit element 1G ) that the map ψβα (b, ·) is just the multiplication on the
left by ψβα (b, 1G ) ∈ G. It is sensible to slightly simplify the notation and write g 7→ ψβα (b)g
for this left multiplication. We find that, just like vector bundles, the principal G-bundles
have transition functions ψβα : Uα ∩ Uβ → G between local trivializations. In particular,
ψβα for a principal bundle satisfy the same cocycle conditions (2.1).
A principal G-bundle over B may be obtained from a system of ψβα , corresponding
to an open cover of B and satisfying (2.1), — via the following ‘Steenrod construction’.
For each trivializing neighbourhood Uα ⊂ B for E consider Uα × G. Define an equivalence
relation between elements (b, h) ∈ Uα ×G, (b0 , h0 ) ∈ Uβ ×G, so that (b, h) ∼ (b0 , h0 ) precisely
if b0 = b and h0 = ψβα (b)h. Now let


P = tα Uα × G / ∼ (2.3)

the disjoint union of all Uα × G’s glued together according to the equivalence relation.
Theorem 2.4. P defined by (2.3) is a principal G-bundle.
Remark . The ψβα ’s can be taken from some vector bundle E over B, then P will be
‘constructed from E’. The construction can be reversed, so as to start from a principal
G-bundle P over a base manifold B and obtain the vector bundle E over B. Then E will
be automatically given a G-structure.
In either case the data of transition functions is the same for the principal G-bundle P
and the vector bundle P . The difference is in the action of the structure group G on the
typical fibre. G acts on itself by left translations in the case of the principal bundle and G
acts as a subgroup of GL(k, R) on Rk in the case of vector bundle E. 3 The vector bundle
E is then said to be associated to P via the action of G on Rk .

Example: Hopf bundle.

Hopf bundle may be defined as the ‘tautological’ (see page 21) rank 1 complex vector
bundle over CP 1 . The total space E of Hopf bundle, as a set, is the disjoint union of all
(complex) lines passing through the origin in C2 . Recall that every such line is the fibre
over the corresponding point in CP 1 . We shall verify that the Hopf bundle is well-defined
by working its transition functions, so that we can appeal to Theorem 2.4.
For a covering system of trivializing neighbourhoods in CP 1 , we can choose the coor-
dinate patches of the smooth structure of CP 1 , defined in Chapter 1. Thus

CP 1 = U1 ∪ U2 , Ui = {z1 : z2 ∈ CP 1 , zi 6= 0}, i = 1, 2,
3
G need not be explicitly a subgroup of GL(k, R), it suffices to have a representation of G on Rk .
24 alexei kovalev

with the local complex coordinate z = z2 /z1 on U1 , and ζ = z1 /z2 on U2 , and ζ = 1/z when
z 6= 0. We shall denote points in the total space E as (wz1 , wz2 ), with |z1 |2 + |z2 |2 6= 0,
w ∈ C, so as to present each point as a vector with coordinate w relative to a basis (z1 , z2 )
of a fibre. (This clarification is needed in the case when (wz1 , wz2 ) = (0, 0) ∈ C2 .) An
‘obvious’ local trivialization over Ui may be given, say over U1 , by (w, wz) ∈ π 1 (U1 ) →
(1 : z, w) ∈ U1 × C, —but in fact this is not a very good choice. Instead we define
p
Φ1 : (w, wz) ∈ π −1 (U1 ) → (1 : z, w 1 + |z|2 ) ∈ U1 × C,
p
Φ2 : (wζ, w) ∈ π −1 (U2 ) → (ζ : 1, w 1 + |ζ|2 ) ∈ U2 × C.

Calculating the inverse, we find

 
w wz
Φ−1
1 (1 : z, w) = p , p
1 + |z|2 1 + |z|2
and so
 
w wz
Φ2 ◦ Φ−1
1 (1 : z, w) = Φ2 p , p
1 + |z|2 1 + |z|2
 
|ζ|w |ζ|w |ζ|
z

= Φ2 p ζ, p = ζ : 1, w = 1 : z, w
ζ |ζ|2 + 1 ζ |ζ|2 + 1 ζ |z|

giving the transition function τ2,1 (1 : z) = (z/|z|), for 1 : z ∈ U1 ∩ U2 (i.e. z 6= 0). The
τ2,1 takes values in the unitary group U (1) = S 1 = {z ∈ C : |z| = 1}, a subgroup of
GL(1, C) = C \ {0} (it is for this reason the square root factor was useful in the local
trivialization). Theorem 2.4 now ensures that Hopf bundle E is a well-defined vector
bundle, moreover a vector bundle with a U (1)-structure. Hence there is a invariantly
defined notion of length of any vector in any fibre of E. The length of (wz1 , wz2 ) may
be calculated in the local trivializations Φ1 or Φ2 by taking the modulus of the second
component
p of Φi (in C). For each i = 1, 2, this coincides with the familiar Euclidean
length |wz1 |2 + |wz2 |2 of (wz1 , wz2 ) in C2 .
We can now use τ2,1 to construct the principal S 1 -bundle (i.e. U (1)-bundle) P → CP 1
associated to Hopf vector bundle E, cf. Theorem 2.4. If U (1) is identified with a unit
circle S 1 ⊂ C , any fibre of P may be considered as the unit circle in the respective
fibre of E. Thus P is identified as the space of all vectors in E of length 1, so P =
{(w1 , w2 ) ∈ C2 | w1 w̄1 +w2 w̄2 = 1} is the 3-dimensional sphere and the bundle projection is

π : (w1 , w2 ) ∈ S 3 → w1 : w2 ∈ S 2 , π −1 (p) ∼
= S 1,

where we used the diffeomorphism S 2 ∼ = CP 1 for the target space. (Examples 1, Q3(ii).)
1 3 2
This principal S -bundle S over S is also called Hopf bundle. It is certainly not trivial,
as S 3 is not diffeomorphic to S 2 × S 1 . (The latter claim is not difficult to verify, e.g.
by showing that the de Rham cohomology H 1 (S 3 ) is trivial, whereas H 1 (S 2 × S 1 ) is not.
Cf. Examples 2 Q5.)
 differential geometry 25

Pulling back vector bundles and principal bundles

Let P be a principal bundle over a base manifold B and E an associated vector bundle
over B. Consider a smooth map f : M → B.
The pull-back of a vector (respectively, principal) bundle is a bundle f ∗ E (f ∗ P )
over M such that there is a commutative diagram (vertical arrows are the bundle projec-
tions)
F
f ∗ E −−−→ E
 

y

y (2.5)
f
M −−−→ B,
such that the restriction of F to each fibre (f ∗ E)p over p ∈ M is an isomorphism onto a
fibre Ef (p) .
A very basic special case of the above is when f maps M to a point in B; then the pull-
back f ∗ E (and f ∗ P ) is necessarily a trivial bundle (exercise: write out a trivialization map
f ∗ E → M × (typical fibre)). As a slight generalization of this example consider the case
when M = B ×X, for some manifold X with f : B ×X → B the first projection. Then f ∗ E
(resp. f ∗ P ) may be thought of as bundles ‘trivial in the X direction’, e.g. f ∗ E ∼ = E × X,
with the projection (e, x) ∈ E × X → (π(e), x) ∈ B × X.
The construction may be extended to a general vector bundle, by working in local
trivialization. Then one has to ensure that the pull-back must be well-defined independent
of the choice of local trivialization. To this end, let {ψβα } be a system of transition
functions for E. Define
f ∗ ψβα = ψβα ◦ f
and f ∗ ψβα is a system of functions on M satisfying the cocycle condition (2.1). Therefore,
by Theorem 2.4 and a remark following this theorem, the f ∗ ψβα are transition functions
for a well-defined vector bundle and principal bundle over M . Steenrod construction shows
that these are indeed the pull-back bundles f ∗ E and f ∗ P as required by (2.5).
26 alexei kovalev

2.1 Bundle morphisms and automorphisms.

Let (E, B, π) and (E 0 , B 0 , π 0 ) be two vector bundles, and f : B → B 0 a smooth map.

Definition. A smooth map F : E → E 0 is a vector bundle morphism covering f if

for any p ∈ B F restricts to a linear map between the fibres F : Ep → Ef0 (p) for any p ∈ B,
so that
F
E −−−→ E 0
 
0

πy

πy

f
B −−−→ B 0
is a commutative diagram, π 0 ◦ F = f ◦ π.

More explicitly, suppose that the local trivializations Φ : π −1 (U ) → U × V , Φ0 :

0 −1
π (U 0 ) → U 0 × V 0 are such that f (U ) ⊆ U 0 . Then the restriction FU = Φ0 ◦ F |π−1 (U ) ◦ Φ−1
is expressed as
FU : (b, v) 7→ (f (b), h(b)v), (2.6)
for some smooth h : U → L(V, V 0 ) family of linear maps between vector spaces V ,V 0
depending on a point in the base manifold. In particular, it is easily checked that a
composition of bundle morphisms E → E 0 , E 0 → E 00 is a bundle morphism E → E 00 .

Examples. 1. If ϕ : M → N is a smooth map between manifolds M ,N then its differential

dϕ : T M → T N is a morphism of tangent bundles.
2. Recall the pull-back of a given vector bundle (E, B, π) via a smooth map f : M → B.
For every local trivialization E|U of E, the corresponding local trivialization of the pull-
back bundle is given by f ∗ E|f −1 (U ) → (f −1 (U ))×V , where V is the typical fibre of E (hence
also of f ∗ E). Thus the pull-back construction gives a well-defined map F : f ∗ E → E which
restricts to a linear isomorphism between any pair of fibres (f ∗ E)p and Ef (p) , p ∈ M . (This
isomorphism becomes just the identity map of the typical fibre V in the indicated local
trivializations.) It follows that F is a bundle morphism covering the given map f : M → B.
3. Important special case of bundle morphisms occurs when f is a diffeomorphism of B
onto B 0 . A morphism F : E → E 0 between two vector bundles over B covering f is called
an isomorphism of vector bundles if F restricts to a linear isomorphism Ep → Ef (p) ,
for every fibre of E.

An isomorphism from a vector bundle E to itself covering the identity map idB is
called a bundle automorphism of E. The set Aut E of all the bundle automorphisms
of E forms a group (by composition of maps). If E = B × V is a trivial bundle then any
automorphism of E is defined by a smooth maps B → GL(V ), so Aut E = C ∞ (B, GL(V )).
If a vector bundle E has a G-structure (G ⊆ GL(V )) then it is natural to consider the
group of G-bundle automorphisms of E, denoted AutG E and defined as follows. Recall
that a G-structure means that there is a system of local trivializations over neighbourhoods
covering the base B and with the transition functions of E taking values in G. Now a bundle
automorphism F ∈ Aut E of E → B is determined in any local trivialization over open
 differential geometry 27

U ⊆ B by a smooth map h : U → GL(V ), as in (2.6). Call F a G-bundle automorphism

if for any of the local trivializations defining the G-structure this map h takes values in the
subgroup G. It follows that AutG E is a subgroup of Aut E. (In the case of trivial bundles
the latter statement becomes C ∞ (B, G) ⊆ C ∞ (B, GL(V )).)
Remark . The group AutG E for the vector bundle E with G-structure has the same sig-
nificance as the group of all self-diffeomorphisms of M for a smooth manifold M or the
group of all linear isometries of an inner product vector space. I.e. AutG E is the ‘group
of natural symmetries of E’ and properties any objects one considers on the vector bundle
are geometrically meaningful if they are preserved by this symmetry group.
One more remark on bundle automorphisms. A map hα giving the local expression over
Uα ⊆ B for a bundle automorphism may be interchangeably viewed as a transformation
from one system of local trivializations to another. Any given local trivialization, say Φα
over Uα , is replaced by Φ0α . We have Φ0α (e) = hα (π(e))Φα (e), e ∈ E. Respectively, the
0
transition functions are replaced according to ψβα = hβ ψβα h−1
α (point-wise group multi-
plication in the right-hand side). This is quite analogous to the setting of linear algebra
where one can either rotate some vector space with respect to a fixed basis or rotate the
basis of a fixed vector space—both operations being expressed as a non-singular matrix.
In Mathematical Physics (and now also in some areas of Differential Geometry) the
group of G-bundle automorphisms is also known as the group of gauge transformations4 ,
sometimes denoted G.

4
...and informally the ‘group of gauge transformations’ is often abbreviated as the ‘gauge group’ of E,
although the ‘gauge group’ is really a different object! (It is the structure group G of vector bundle.) Alas,
there is a danger of confusion.
28 alexei kovalev

2.2 Connections.
Sections of a vector bundle generalize vector-valued functions on open domains in Rn . Is
there a suitable version of derivative for sections, corresponding to the differential in multi-
variate calculus? In order to propose such a derivative, it is necessary at least to understand
which sections are to have zero derivative, corresponding to the constant functions on Rn .
(Note that a section which is expressed as a constant in one local trivialization need not
be constant in another.)

Vertical and horizontal subspaces.

Consider a vector bundle π : E → B with typical fibre Rm and dim B = n. Let U ⊆ B be
a coordinate neighbourhood in B and also a trivializing neighbourhood for E. Write xk ,
k = 1, . . . , n for the coordinates on U and aj , j = 1, . . . , m for the standard coordinates
on Rm . Then with the help of local trivialization the tangent space Tp E for any point p,
such that π(p) ∈ U , has a basis { ∂x∂ k , ∂a∂ j }. The kernel of the differential (dπ)p : Tp E → Tb B,
b = π(p), is precisely the tangent space to the fibre Eb ⊂ E, spanned by { ∂a∂ j }.

Definition. The vector space Ker (dπ)p is called the vertical subspace of Tp E, denoted
T vp E. A subspace Sp of Tp E is called a horizontal subspace if Sp ∩ T vp E = {0} and
Sp ⊕ T vp E = Tp E.

Thus any horizontal subspace at p is isomorphic to the quotient Tp E/T vp E and has
the dimension dim(Tp E) − dim(T vp E) = dim B. Notice that, unlike the vertical tangent
space, a horizontal space can be chosen in many different ways (e.g. because there are
many choices of local trivialization near a given point in B).
It is convenient to specify a choice of a horizontal subspace at every point of E as the
kernel of a system of differential 1-forms on E, using the following

Fact from linear algebra: if θ1 , . . . , θm ∈ (Rn+m )∗ are linear functionals then one
will have dim(∩m i 1 m
i=1 Ker θ ) = n if and only if θ , . . . , θ are linearly independent
in (Rn+m )∗ .

Now let θp1 , . . . , θpm be linearly independent ‘covectors’ in Tp∗ E, p ∈ π −1 (U ), and define

Sp := {v ∈ Tp E| θpi (v) = 0, i = 1, . . . , m}.

We can write, using local coordinates on U ,

θpi = fki dxk + gji daj , i = 1, . . . , m, fki , gji ∈ R. (2.7)

and any tangent vector in Tp E as v = B k ( ∂x∂ k )p + C i ( ∂a∂ i )p , B k , C i ∈ R. The θpj cannot all
vanish on a vertical vector, i.e. on a vector having B k = 0 for all k. That is,

if gji C j = 0 for all i = 1, . . . , m then C i = 0 for all i = 1, . . . , m.

 differential geometry 29

Therefore the m × m matrix g = (gji ) must be invertible. Denote the inverse matrix by
c = g −1 , c = (cil ). Replace θi by θ̃i = cil θl = dai + eik dxk , this does not change the space Sp .
The above arrangement can be made for every p ∈ π −1 (U ), with fki (p) and gji (p)
in (2.7) becoming functions of p. Call a map p 7→ Sp a field of horizontal subspaces if
the functions fki (p) and gji (p) are smooth. To summarize,

Proposition 2.8. Let S = Sp , p ∈ E, be an arbitrary smooth field of horizontal subspaces

in T E. Let xk , aj be local coordinates on π −1 (U ) arising, as above, from some local trivial-
ization of E over a coordinate neighbourhood U . Then Sp is expressed as Sp = ∩m j
j=1 Ker θp ,
where
θj = daj + ejk (x, a)dxk , (2.9)
for some smooth functions ejk (x, a). These ejk (x, a) are uniquely determined by a local
trivialization.

Definition. A field of horizontal subspaces Sp ⊂ Tp E is called a connection on E if in

every local trivialization it can be written as Sp = Ker (θp1 , . . . , θpm ) as in (2.9), such that
the functions eik (p) = eik (x, a) are linear in the fibre variables,

eik (x, a) = Γijk (x)aj . (2.10a)

and so

θpi = dai + Γijk (x)aj dxk , (2.10b)

where Γijk : U ⊂ B → R are smooth functions called the coefficients of connection Sp

in a given local trivialization.

As we shall see below, the linearity condition in ai ensures that the horizontal sections
(which are to become the analogues of constant vector-functions) form a linear subspace of
the vector space of all sections of E. (A very reasonable thing to ask for.)
I will sometimes use an abbreviated notation

θpi = dai + Aij aj , where Aij = Γijk dxk ,

So Proposition 2.8 identifies a connection with a system of matrices A = (Aij ) of differential

1-forms, assigned to trivializing neighbourhoods U ⊂ B.

The transformation law for connections.

0
Now consider another coordinate patch U 0 ⊂ X, U 0 = {(xk )} and a local trivialization
0 0
Φ0 : π −1 (U 0 ) → U 0 × Rm with xk , ai the coordinates on U 0 × Rm .
Notation: throughout this subsection, the apostrophe 0 will refer to the local trivialization
of E over U 0 , whereas the same notation without 0 refers to similar objects in the local
trivialization of E over U . In particular, the transition (matrix-valued) functions from U
30 alexei kovalev

0 0
to U 0 are written as ψii and from U 0 to U as ψii0 , thus the matrix (ψii0 ) is inverse to (ψii ).
0 0
Likewise, (∂xk /∂xk ) denotes the inverse matrix to (∂xk /∂xk ).
Recall that on U 0 ∩ U we have
0 0 0 0
xk = xk (x), ai = ψii (x)ai . (2.11)

Then
0
∂xk k
k0 0 0 0
dx = dx , dai = dψii ai + ψii dai
∂xk
0
∂ψii i k 0
= k
a dx + ψii dai ,
∂x
so 0
i0 i0 0 0 0 0 0 ∂xk i0 j 0 k
θ = da + Γij 0 k0 aj dxk = dψii ai + ψii dai + Γ 0 0 a dx ,
∂xk j k
and 0 0
0 ∂ψji ∂xk i0 j0 j
ψii0 θi
= da + · k
+ ψi
i
i
0 ·
k
(ψii0
· Γj
k
0 k 0 · ψj )a dx .
∂x ∂x
i i0 i
But then ψi0 θ = θ and we find, by comparing with (2.10b), that
0
0 ∂x
k0 ∂ψji
0
Γijk = Γij 0 k0 ψii0 ψjj k
+ ψii0 (2.12a)
∂x ∂xk
and, using Aij = Γijk dxk ,
0 0 0
Aij 0 = ψii Aij ψjj0 + ψii dψji 0 , (2.12b)
0
Writing AΦ and AΦ for the matrix-valued 1-forms expressing the connection A in the local
trivializations respectively Φ and Φ0 and abbreviating (2.11) to Φ0 = ψΦ for the transition
function ψ we obtain from the above that

AψΦ = ψAΦ ψ −1 + ψdψ −1 = ψAΦ ψ −1 − (dψ)ψ −1 . (2.12c)

The above calculations prove.

Theorem 2.13. Any system of functions Γijk , i, j = 1, . . . , m, k = 1, . . . , n attached to the
local trivializations and satisfying the transformation law (2.12) defines on E a connection
A, whose coefficients are Γijk .
Remark . Suppose that we fix a local trivialization Φ and a connection A on E and regard
ψ as a bundle automorphism of E, ψ ∈ Aut E. With his shift of view, the formula (2.12c)
expresses the action of the group Aut E on the space of connections on E. (Cf. the remark
on bundle automorphisms and linear algebra, page 27.)
Before considering the third view on connections we need a rigorous and systematic
way to consider ‘vectors and matrices of differential forms’.
 differential geometry 31

The endomorphism bundle End E. Differential forms with values in vector

bundles.
Let (E, B, π) be a vector bundle with typical fibre V and transition functions ψβα . If G
is a linear map V → V , or endomorphism of V , G ∈ End V , in a trivialization labelled
by α, then the same endomorphism in trivialization β will be given by ψβα Gψαβ (recall
−1
that ψαβ = ψβα ).
This may be understood in the sense that the structure group of V acts linearly on
End V . Exploiting, once again, the idea of Theorem 2.4 and the accompanying remarks
one can construct from E a new vector bundle End E, with the same structure group as
for E and with a typical fibre End V . This is called the endomorphism bundle of a
vector bundle E and denoted End E.
One can further extend the above construction and define over B the vector bundle
whose fibres are linear maps Tb B → Eb , b ∈ B or, more generally, the antisymmetric
multilinear maps Tb B × . . . × Tb B → Eb on r-tuples of tangent vectors. (So the typical
fibre of the corresponding bundle is the tensor product Λr (Rn )∗ ⊗ V i.e. the space of
antisymmetric multilinear maps (Rn )r → V .) The sections of these bundles are called
differential 1-forms (respectively r-forms) with values in E and denoted ΩrB (E). In any
local trivialization, an element of ΩrB (E) may be written as a vector whose entries are
differential r-forms. (The ‘usual’ differential forms correspond in this picture to the case
when V = R and E = B × R.)
In a similar manner, one introduces the differential r-forms ΩrB (End E) with values in
the vector bundle End E. These forms are given in a local trivialization as m × m matrices
whose entries are the usual differential r-forms. The operations of products of two matrices,
or of a matrix and a vector, extend to ΩrB (E) and ΩrB (End E), in the obvious way, using
wedge product between the entries.
Now from the examination of the transformation law (2.12) we find that although a
connection is expressed by a differential form in a local trivialization, a connection is not
in general a well-defined differential form. The difference between two connections
however is a well-defined ‘matrix of 1-forms’, more precisely an element in Ω1B (End E).
Thus the space of all connections on a given vector bundle E is naturally an affine space.
Recall that an affine space of points and has a vector space assigned to it and the operation
of ‘adding a vector to a point to obtain another point’ (with certain ‘usual’ properties).
The latter vector space and affine space may be identified, but not canonically—one needs
to choose where to map the zero vector. The vector space assigned to the affine space of
connections on E is Ω1B (End E).

A remark on principal bundles

In the case of a principal G-bundle π : P → B, with G ⊆ GL(m, R) a subgroup, the
definitions of vertical and horizontal subspaces T vq P and Sq of Tq P , like for the vector
bundles above, still make sense with E now replaced by P . A connection on this principal
bundle may in fact be realized as a matrix-valued differential 1-form, θ say,—though the
entries of θ will be 1-forms on the total space P , rather than on the base. More precisely,
32 alexei kovalev

θ will determine a connection on P if and only if θ is (1) horizontal and (2) G-equivariant
with respect to the smooth free right action on P .
We may determine a choice of a field of horizontal subspaces by using the kernels of local
matrices of 1-forms θp = (θji )p defined around each p ∈ P , where (θji )p ∈ Ω1 (π −1 (U )) for
i, j = 1, . . . , m and U ⊂ B a trivializing neighbourhood (and a coordinate neighbourhood)
U ⊂ B. Denote by cij the ‘obvious’ coordinates on the space Matr(m, R) and require that
(θji )p = dcij + Γisk (x)csj dxk for some functions Γisk ∈ C ∞ (U ) satisfying the transformation
law (2.12). In a more abbreviated notation, θ = dC + AC, where C = (cij ) and A = (Aij ) =
(Γijk dxk ). One can verify, noting (2.12) and C 0 = ψC (cf. (2.11) and page 23), that the
forms C −1 θ = C −1 dC + C −1 AC agree on the intersections π −1 (U ) ∩ π −1 (U 0 ) and patch
together to define a global 1-form θ on P . Cf. Sheet 3 Q8.
Note the construction of θ does not extend to the vector bundles.

Covariant derivatives
Definition . A covariant derivative on a vector bundle E is a R-linear operator
∇E : Γ(E) → Γ(T ∗ B ⊗ E) satisfying a ‘Leibniz rule’

∇E (f s) = df ⊗ s + f ∇E s (2.14)

for any s ∈ Γ(E) and function f ∈ C ∞ (B).

Here I used Γ(·) to denote the space of sections of a vector bundle. Thus Γ(E) = Ω0B (E)
and Γ(T ∗ B ⊗ E) = Ω1B (E).
Example. Consider a connection A and put ∇E = dA defined in a local trivialization by

dA s = ds + As, s ∈ Γ(E).

More explicitly, one can write s = (s1 , . . . , sm ) with the help of a local trivialization, where
sj are smooth functions on the trivializing neighbourhood, and then
∂s1 ∂sm
dA (s1 , . . . , sm ) = ( 1 j k m j k


+ Γ jk s )dx , . . . , ( + Γ jk s )dx .
∂xk ∂xk
The operator dA is well-defined as making a transition to another trivialization we have
s = ψs0 and A = ψA0 ψ −1 − (dψ)ψ −1 , which yields the correct transformation law for
dA s = ds + As = d(ψs0 ) + (ψA0 ψ −1 − (dψ)ψ −1 )s0 = ψ(ds0 + A0 s0 ) = ψ(dA s)0 .
Theorem 2.15. Any covariant derivative ∇E arises as dA from some connection A.
Proof (gist). Firstly, any covariant derivative ∇E is a local operation, which means that
is s1 , s2 are two sections which are equal over an open neighbourhood U of b ∈ B then
(∇E s1 )|b = (∇E s2 )|b . Indeed, let U0 be a smaller neighbourhood of b with the closure
U0 ⊂ U and consider a cut-off function α ∈ C ∞ (B), so that 0 ≤ α ≤ 1, α|U0 = 1, α|B\U = 0.
Then 0 = d(α(s1 − s2 )) = (s1 − s2 ) ⊗ dα + α∇E (s1 − s2 ), whence (∇E s1 )|b = (∇E s2 )|b as
required. So, it suffices to consider ∇E in some local trivialization of E.
 differential geometry 33

The proof now simply produces the coefficients Γijk of the desired A in an arbitrary
local trivialization of E, over U say. Any local section of E defined over U may be written,
with respect to the local trivialization, as a vector valued function U → Rm (m being
the rank of E). Let ej , j = 1, . . . m denote the sections corresponding in this way to the
constant vector-valued functions on U equal to the j-th standard basis vector of Rm . Then
the coefficients of the connection A are (uniquely) determined by the formula
∂
i
Γijk = (∇E ej )( ) , (2.16)
∂xk
where ∇E (ej ) is a vector of differential 1-forms which takes the argument a vector field. We
used a local coordinate vector field ∂x∂ k (which is well-defined provided that a trivializing
neighbourhood U for E is also a coordinate neighbourhood for B) and obtained a local
section of E, expressed as a smooth map U → V . Then Γijk ∈ C ∞ (U ) is the i-th component
of this map in the basis ej of V .
It follows, from the R-linearity and Leibniz rule for ∇E , that for an arbitrary local
section we must have
∇E s = ∇E (sj ej ) = (dsi + sj Γijk dxk )ei = dA s
where as usual A = (Aij ) = (Γijk dxk ), so we recover the dA defined above. It remains to ver-
ify that Γijk ’s actually transform according to (2.12) in any change of local trivialization, so
we get a well-defined connection. The latter calculation is straightforward (and practically
equivalent to verifying that dA is well-defined independent of local trivialization.)
The definition of covariant derivative further extends to differential forms with values
in E by requiring the Leibniz rule, as follows
dA (σ ∧ ω) = (dA σ) ∧ ω + (−1)q σ ∧ (dω), σ ∈ ΩqB (E), ω ∈ Ωr (B),
with ∧ above extended in an obvious way to multiply vector-valued differential forms and
usual differential forms. It is straightforward to verify, considering local bases of sections
ej and differential forms dxk , that in any local trivialization one has dA σ = dσ + A ∧ σ,
where σ ∈ ΩqB (E).

Remark (Parallel sections). A section s of E is called parallel, or covariant constant,

if dA s = 0. In a local trivialization over coordinate neighbourhood U the section s is
expressed as s = (s1 (x), . . . , sm (x)), x ∈ U and the graph of s is respectively Σ =
{(xk , sj (x)) ∈ U × Rm : x ∈ B}, a submanifold of U × Rm . The tangent spaces to Σ are
∂sj ∂
spanned by ∂x∂ k + ∂x i i i k j
k ∂aj , for k = 1, . . . , n. We find that the 1-forms θs(x) = da + Γjk dx s ,

i = 1, . . . , m, vanish precisely on the tangent vectors to Σ.

This is just the horizontality condition for a tangent vector to E and we see that
a section s is covariant constant if and only if any tangent vector to the graph of s is
horizontal. Another form of the same statement: s : B → E defines an embedding of B
in E as the graph of s and the tangent space to the submanifold s(B) at p ∈ E is the
horizontal subspace at p (relative to A) if and only if dA s = 0.
34 alexei kovalev

To sum up, a connection on a vector bundle E can be given in three equivalent

ways:
(1) as a (smooth) field of horizontal subspaces in T E depending linearly on the fibre
coordinates, as in (2.10a);
or
(2) as a system of matrix-valued 1-forms Aij (a system of smooth functions Γijk ) assigned to
every local trivialization of E and satisfying the transformation rule (2.12) on the overlaps;
or
(3) as a covariant derivative ∇E on the sections of E and, more generally, on the differential
forms with values in E.

2.3 Curvature.
Let A be a connection on vector bundle E and consider the repeated covariant differentia-
tion of an arbitrary section (or r-form) s ∈ ΩrB (E) (assume r = 0 though). Calculation in
a local trivialization gives
dA dA s = d(ds+As)+A∧(ds+As) = (dA)s−A∧(ds)+A∧(ds)+A∧(As) = (dA+A∧A)s.
Thus dA dA is a linear algebraic operator, i.e. unlike the differential operator but dA , the
dA dA commutes with the multiplication by smooth functions.
dA dA (f s) = f dA dA s, for any f ∈ C ∞ (B). (2.17)
Notice that the formula (2.17) does not make explicit reference to any local trivialization.
We find that (dA dA s) at any point b ∈ B is determined by the value s(b) at that point. It
follows that the operator dA dA is a multiplication by an endomorphism-valued differential
2-form. (This 2-form can be recovered explicitly in coordinates similarly to (2.16), using
a basis ei say of local sections and a basis of differential 1-forms dxk in local coordinates
on B.)
Definition. The form
F (A) = dA + A ∧ A ∈ Ω2 (B; End E).
is called the curvature form of a connection A.
Definition. A connection A is said to be flat is its curvature form vanishes F (A) = 0.
Example. Consider a trivial bundle B × Rm , so the space of sections is just the vector-
functions C ∞ (B; Rm ). Then exterior derivative applied to each component of a vector-
function is a well-defined linear operator satisfying Leibniz rule (2.14). The corresponding
connection is called trivial, or product connection. It is clearly a flat connection.
The converse is only true with an additional topological condition that the base B is
simply-connected; then any flat connection on E induces a (global) trivialization E ∼ =B×
Rm (Examples 3, Q7(ii)) and will be a product connection with respect to this trivialisation.
 differential geometry 35

Bianchi identity.
Covariant derivative on a vector bundle E with respect to a connection A can be extended,
in a natural way, to any section of B of End E by requiring the following formula to hold,

(dA B)s = dA (Bs) − B(dA s),

for every section s of E. Notice that this is just a suitable form of Leibniz rule.
The definition further extends to differential forms with values End E, by setting for
every µ ∈ Ωp (B; End E) and σ ∈ Ωq (B; E),

(dA µ) ∧ σ = dA (µ ∧ σ) − (−1)p µ ∧ (dA σ).

P P
(Write µ = k Bk ωk and σ = j sj ηj .) It follows that a suitable Leibniz rule also
holds when µ1 ∈ Ωp (B; End E), µ2 ∈ Ωq (B; End E) to give dA (µ1 ∧ µ2 ) = (dA µ1 ) ∧ µ2 +
(−1)p µ1 ∧ (dA µ2 ). In the special case of trivial vector bundle, E = B × R and with dA = d
the exterior differentiation, the above formulae recover the familiar results for the usual
differential forms.
In particular, any µ ∈ Ω2 (B; End E) in a local trivialization becomes a matrix of 2-forms
and its covariant derivative is a matrix of 3-forms given by

dA µ = dµ + A ∧ µ − µ ∧ A.

Now, for any section s of E, we can write dA (dA dA )s = (dA dA )dA s, i.e. dA (F (A)s) =
F (A)dA s and comparing with the Leibniz rule above we obtain.
Proposition 2.18 (Bianchi identity). Every connection A satisfies dA F (A) = 0.

2.4 Orthogonal and unitary connections

Recall that an orthogonal structure on a (real) vector bundle E defined as a family Φα of
local trivializations (covering the base) of this bundle so that all the transition functions ψβα
between these take values in the orthogonal group O(m), m = rk E. These trivializations
Φα are then referred to as orthogonal trivializations. There is a similar concept of a unitary
structure and unitary trivializations of a complex vector bundle.
Given a choice of orthogonal (unitary) structure on E, the standard Euclidean (Hermi-
tian) inner product on the typical fibre Rm (Cm ) induces, with the help of the orthogonal
(unitary) local trivializations, a well-defined inner product on the fibres of E. Cf. Examples
3 Q2.
Definition. We say that A is an orthogonal connection relative to an orthogonal structure,
respectively a unitary connection relative to a unitary structure, on a vector bundle E if

dhs1 , s2 i = hdA s1 , s2 i + hs1 , dA s2 i

for any two sections s1 , s2 of E, where h·, ·i denotes the inner product on the fibres of E.
36 alexei kovalev

Proposition 2.19. An orthogonal connection has skew-symmetric matrix of coefficients

in any orthogonal local trivialization. A unitary connection has skew-Hermitian matrix of
coefficients in any unitary local trivialization.

Proof.

0 = hdA (si1 ei ), sj2 ej i + hsi1 ei , dA (sj2 ej )i − dhsi1 ei , sj2 ej i = (Aij + Aji )si1 sj2 for any si1 , sj2 ,

where ei is the standard basis of Rm or Cm .

Corollary 2.20. The curvature form F (A) of an orthogonal (resp. unitary) connection A
is skew-symmetric (resp. skew-Hermitian) in any orthogonal (unitary) trivialization.

2.5 Existence of connections.

Theorem 2.21. Every vector bundle E → B admits a connection.

Proof. It suffices to show that there exists a well-defined covariant derivative ∇E on sections
of E. We shall construct a example of ∇E using a partition of unity.
Let Wα be an open covering of B by trivializing neighbourhoods for E and Φα the
corresponding local trivializations. Then on each restriction E|Wα we may consider a
trivial product connection d(α) defined using Φα . Of course, the expression d(α) s will only
make sense over all of B if a section s ∈ Γ(E) is equal to zero away from Wα . Now consider
a partition of unity ρi subordinate to Wα . The expressions ρi s, ρi d(i) s make sense over all
of B as we may extend by zero away from Wi . Now define
∞
X ∞
X
E
∇ s := d(i) (ρi s) = ρi d(i) s, (2.22)
i=1 i=1

we used Leibniz rule for d(i) and the property ∞

P
where
P∞for the second equality i=1 ρi = 1
E
(so i=1 dρi = 0). The ∇ defined by (2.22) is manifestly linear in s and Leibniz rule for
∇E holds because it does for each d(i) .
Part III: Differential geometry (Michaelmas 2003, 2004, 2014, 2018)

Alexei Kovalev ([email protected])

3 Riemannian geometry
3.1 Riemannian metrics and the Levi–Civita connection
Let M be a smooth manifold.
Definition. A bilinear symmetric positive-definite form

gp : Tp M × Tp M → R

defined for every p ∈ M and smoothly depending on p is called a Riemannian metric

on M .
Positive-definite means that gp (v, v) > 0 for every v 6= 0, v ∈ Tp M . Smoothly depending
on p means that for every pair Xp ,Yp of C ∞ -smooth vector fields on M the expression
gp (Xp , Yp ) defines a C ∞ -smooth function of p ∈ M .
Alternatively, consider a coordinate neighbourhood on M containing p and let xi ,
i = 1, . . . , dim M be the local coordinates. Then any two tangent vectors u, v ∈ Tp M
may be written
as u =
ui ( ∂x∂ i )p , v = v i ( ∂x∂ i )p and gp (u, v) = gij (p)ui v j , where the functions
gij (p) = g( ∂x∂ i p , ∂x∂ j p ) express the coefficients of the metric g in local coordinates. One
often uses the following notation for a metric in local coordinates

g = gij dxi dxj .

The bilinear form (metric) g will be smooth if and only if the local coefficients gij = gij (x)
are smooth functions of local coordinates xi on each coordinate neighbourhood.
Example 3.1. Recall (from Chapter 1) that any smooth regularly parameterized surface S
in R3 ,
r : (u, v) ∈ U ⊂ R2 → r(u, v) ∈ R3 .
is a 2-dimensional manifold (more precisely, we assume here that S satisfies all the defining
conditions of an embedded submanifold). The first fundamental form1 Edu2 + 2F dudv +
Gdv 2 is a Riemannian metric on S.
The following formulae are proved in multivariate calculus.
• A curve on S may be given as γ(t) = r(u(t), v(t)), a ≤ t ≤ b. The length of γ is then
Rb Rb√
computed as a |γ̇(t)|dt = a E u̇2 + 2F u̇v̇ + Gv̇ 2 dt.
RR √
• The area of S is U EG − F 2 du dv.
1
E = (ru , ru ), F = (ru , rv ), G = (rv , rv ) using the Euclidean inner product

37
38 alexei kovalev

Theorem 3.2. Any smooth manifold M can be given a Riemannian metric.

Proof. Indeed, M may be embedded in Rm by Whitney theorem (cf. Q9 Example Sheet 1).
Then the restriction (more precisely, a pull-back) of the Euclidean metric of Rm to M
defines a Riemannian metric on M .
Remark . A metric, being a bilinear form on the tangent spaces, can be pulled back via a
smooth map, f say, in just the same way as a differential form. But a pull-back f ∗ g of a
metric g will be a well-defined metric only if f has an injective differential.
Remark . As a Riemannian metric on M is an inner product on the vector bundle T M ,
Theorem 3.2 is also a consequence of Q2 of Example Sheet 3.
Definition. A connection on a manifold M is a connection on its tangent bundle T M .
Recall that a choice of local coordinates x on M determines a choice of local trivial-
ization of T M (using the basis vector fields ∂x∂ i on coordinate patches). The transition
function ϕ for two trivializations of T M is given by the Jacobi matrices of the correspond-
∂xi
ing change of coordinates (ϕii0 ) = ( ∂x i0 ).

Let Γijk be the coefficients (Christoffel symbols) of a connection on M in local coordi-

0
nates xi . For any other choice xi of local coordinates the transition law on the overlap
becomes (cf. Chapter 2, eqn. (2.12a))
0 0 0
∂xi ∂xj ∂xk
0 ∂xi ∂ 2 xi
Γijk = Γij 0 k0 + (3.3)
∂xi0 ∂xj ∂xk ∂xi0 ∂xj ∂xk
One can see from the above formula that if Γijk are the coefficients of a connection on M
then Γikj also are the coefficients of some well-defined connection on M (in general, this
would be a different connection).
i
The difference Tjk = Γijk − Γikj is called the torsion of a connection (Γijk ). The trans-
0 i
∂x ∂x ∂x j0 k0
i i
formation law for Tjk is Tjk = Tji0 k0 ∂xi0 ∂xj ∂xk , thus the torsion of a connection is a well-
defined antisymmetric bilinear map sending a pair of vector fields X, Y to a vector field
i
T (X, Y ) = Tjk X j Y k on M .
Definition. A connection on M is symmetric if its torsion vanishes, i.e. if Γijk = Γikj .
Notation: given a connection (covariant derivative) D : Ω0M (T M ) → Ω1M (T M ) and a
smooth vector field X on M , we write DX for the composition of D and contraction of
1-forms (in Ω1M (T M )) with X. Thus DX : Ω0M (T M ) → Ω0M (T M ) is a linear differential
operator acting on vector fields on M . In local coordinates, it is expressed as (DX Y )i =
X j ∂j Y i + Γijk Y j X k .
It is not difficult to see, by comparing with the definition on page 31, that a family
of operators DX , depending on a vector field X, defines a covariant derivative precisely if
DX Y is C ∞ (M )-linear in X, R-linear in Y and satisfies the Leibniz rule
DX (hY ) = (Xh)Y + hDX Y, (3.4)
for each h ∈ C ∞ (M ) and a vector field Y (recall Xh equals the contraction of dh with X).
 differential geometry 39

Here is a way to define a symmetric connection independent of the local coordinates.

Proposition 3.5. A connection D is symmetric if and only if DX Y − DY X = [X, Y ].
The proof is an (easy) straightforward computation.
Theorem 3.6. On any Riemannian manifold (M, g) there exists a unique connection D
such that
(1) d(g(X, Y ))(Z) = g(DZ X, Y ) + g(X, DZ Y ) for any vector fields X, Y, Z on M ; and
(2) the connection D is symmetric.
D is called the Levi–Civita connection of the metric g.
The condition (1) in the above theorem is sometimes written more neatly as

dg(X, Y ) = g(DX, Y ) + g(X, DY ).

Proof. Uniqueness. The conditions (1) and (2) determine the coefficients of Levi–Civita in
local coordinates as follows. A ‘coordinate vector field’ ∂x∂ i with constant coefficients has
covariant derivative D ∂x∂ i = Γpik ∂x∂ p dxk . The condition (1) with X = ∂x∂ i , Y = ∂x∂ j , Z = ∂x∂ k
becomes
∂
gij = Γpik gpj + Γpjk gip . (3.7a)
∂xk
Cycling i, j, k in the above formula, one can write two more relations
∂
j
gki = Γpkj gpi + Γpij gkp , (3.7b)
∂x
∂
i
gjk = Γpji gpk + Γpki gjp . (3.7c)
∂x
Let (g iq ) denote the inverse matrix to (giq ), so Γpjk gqp g iq = Γijk . Adding the first two
equations of (3.7) and subtracting the third, dividing by 2, and multiplying both sides of
the resulting equation by (g iq ), one obtains the formula
 
i 1 iq ∂gqj ∂gkq ∂gjk
Γjk = g + − (3.8)
2 ∂xk ∂xj ∂xq
(also taking account of the symmetry condition (2)). Thus if the Levi–Civita connection
exists then its coefficients in local coordinates are expressed in terms of the metric by (3.8).
Exericise. By adapting the above method to arbitrary vector fields X, Y, Z on M , using
the symmetry condition (2) in the form DX Y − DY X = [X, Y ], show that the Levi–Civita
connection is uniquely determined by the identity
1

g(DX Y, Z) = Xg(Y, Z)+Y g(Z, X)−Zg(X, Y )−g(Y, [X, Z])−g(Z, [Y, X])+g(X, [Z, Y ]) .
2
(3.9)
Verify that D defined by (3.9) satisfies the conditions (1) and (2) in Theorem 3.6 (this
might be argued by essentially following your calculation of (3.9) backwards).
40 alexei kovalev

Existence. Proof 1. One way of proving the existence is to check that the Γijk computed
by the formula (3.8) are indeed the coefficients of a well-defined connection on M . This
can be done by verifying that the Γijk ’s transform in the right way, i.e. as in (3.3), under
∂xi ∂xj
a change of local coordinates. The transformation law for gij is gi0 j 0 = 0 g ij , by the
∂xi ∂xj 0
usual linear algebra. Differentiating this latter formula and using the respective formula
for the induced inner product on the dual spaces, i.e. on the cotangent spaces to M , we
can verify that the coefficients given by (3.8) indeed transform according to (3.3) and so
the Levi–Civita connection of the metric g on M is well-defined.
Proof 2. Alternatively, assuming the exercise above, we shall be done if we show that
Dx Y defined by the formula (3.9) is C ∞ (M )-linear in X and satisfies the Leibniz rule in Y .
For the first property, we note that [f X, Z] = f XZ − Z(f X) = f [X, Z] − (Zf )X, for every
f ∈ C ∞ (M ). Then 2g(Df X Y, Z) becomes
f Xg(Y, Z) + Y g(Z, f X) − Zg(f X, Y ) − g(Y, [f X, Z]) − g(Z, [Y, f X]) + g(f X, [Z, Y ])
= f Xg(Y, Z) + (Y f )g(Z, X) + f Y g(Z, X) − (Zf )g(X, Y ) − f Zg(X, Y )
−g(Y, f [X, Z] − (Zf )X) − g(Z, (Y f )X + f [Y, X]) + f g(X, [Z, Y ]),
using the Leibniz rule for vector fields. It follows that g(Df X Y, Z) = g(f DX Y, Z), thus D
is C ∞ (M )-linear in X.
For the Leibniz rule we calculate, with h a smooth function,
2g(DX (hY ), Z) = X(hg(Y, Z)) + hY g(Z, X)
− Z(hg(X, Y )) − hg(Y, [X, Z]) − g(Z, [hY, X]) + g(X, [Z, hY ])
= (Xh)g(Y, Z) + hXg(Y, Z) + hY g(Z, X) − (Zh)g(X, Y ) − hZg(X, Y )
−hg(Y, [X, Z]) − hg(Z, [Y, X]) + (Xh)g(Z, Y ) + hg(X, [Z, Y ]) + (Zh)g(X, Y )
= 2(Xh)g(Y, Z) + 2hg(DX Y, Z)
which gives (3.4) as required. Since DX Y is clearly R-linear in Y we have proved that D
is a connection on M .

3.2 Geodesics on a Riemannian manifold

Let E → M be a vector bundle endowed with a connection (Γijk ). A parameterized smooth
curve on the base M may be written in local coordinates by (xi (t). A lift of this curve
to E is locally expressed as (xi (t), aj (t)) using local trivialization of the bundle E to define
coordinates aj along the fibres. A tangent vector (ẋ(t), ȧ(t)) ∈ T(xi (t),aj (t)) E to a lifted
curve will be horizontal (recall from the chapter 2, eqn. (2.10b)) at every t precisely when
a(t) satisfies a linear ODE
ȧi + Γijk (x)aj ẋk = 0, (3.10)
where i, j = 1, . . . , rank E, k = 1, . . . , dim B.
Now if E = T M then there is also a canonical lift of any smooth curve γ(t) on the
base, as γ̇(t) ∈ Tγ(t) M .
 differential geometry 41

Definition. A curve γ(t) on a Riemannian manifold M is called a geodesic if γ̇(t) at

every t is horizontal with respect to the Levi–Civita connection.
Thus we are looking at a special case of (3.10) when a = ẋ. The condition for a path
in M to be a geodesic may be written explicitly in local coordinates as

ẍi + Γijk (x)ẋj ẋk = 0, (3.11)

a non-linear second-order ordinary differential equation for a path x(t) = (xi (t)) (here
i, j, k = 1, . . . , dim M ). By the basic existence and uniqueness theorem from the theory
of ordinary differential equations, it follows that for any choice of the initial conditions
x(0) = p, ẋ(0) = a there is a unique solution path x(t) defined for |t| < ε for some
positive ε. Thus for any p ∈ M and a ∈ Tp M there is a uniquely determined (at least
for any small |t|) geodesic with this initial data (i.e. ‘coming out of p in the direction a’).
Denote this geodesic by γp (t, a) (or γ(t, a) if this is not likely to cause confusion).
Proposition 3.12. If γ(t) is a geodesic on (M, g) then |γ̇(t)|g = const.
Proof. We shall first make a rigorous sense of the equation

Dγ̇ γ̇ = 0 (3.13)

and show that (3.13) is satisfied at each γ(t) if and only if γ is a geodesic curve. The
problem with (3.13) at the moment is that γ̇ is not a vector field defined on any open set
in M , but only along a curve γ. We define an extension, still denoted by γ̇, on a coordinate
neighbourhood U of γ(0) as follows. It may be assumed, without loss, that γ̇(0) = (ẋi (0))
has ẋ1 (0) 6= 0. We may further assume, taking a smaller U if necessary, that γ ∩ U , is
a graph of a smooth function x1 7→ (x2 (x1 ), . . . , xn (x1 )). In particular, ẋ1 (t) 6= 0 for any
small |t| and also any hyperplane x1 = x10 , such that |x10 − x1 (γ(0))| is small, meets the
curve γ ∩ U in exactly one point. Denote by π the projection along hyperplanes x1 = const
onto γ ∩ U . Define, for every p ∈ U , γ̇(p) = γ̇(π(p)) and then γ̇ is a smooth vector field
on U , such that (γ̇)p = γ̇(t) whenever p = γ(t).
Now let Γijk be the coefficients of the Levi–Civita in the coordinates on U . So DZ Y =
(Z l ∂l Y i + Γijk Y j Z k )∂i for any vector fields Z = Z l ∂l , Y = Y i ∂i on U . Let Y = Z = γ̇.
∂ ẋi
Then at any point p = γ(t) we have Z l ∂l Y i = ẋl l = ẍi by the chain rule. It follows
∂x
that the equation (3.11) is equivalent to (3.13) if the latter if restricted to the points of
the curve γ. It can also be seen, by inspection of the above construction, that Dγ̇ γ̇ at the
points of γ is independent of the choice of extension of γ̇(t) to a vector field on U .
We have γ̇(γ̇, γ̇)g = (D
γ̇ γ̇, γ̇)g + (γ̇, Dγ̇ γ̇)g on U from the defining properties of the
2
Levi–Civita. Hence γ̇ |γ̇|g = 0 at each γ(t) ∈ U , by (3.13). From the construction of the


extension γ̇ on U , we find that the directional partial derivative γ̇ |γ̇|2g at the points γ(t)
∂ d
is expressed as ẋl l |γ̇|2g = |γ̇(t)|2g by the chain rule again, whence |γ̇|g = const as we
∂x dt
had to prove.
42 alexei kovalev

Examples. 1. On Rn with the Euclidean metric (dxi )2 we have Γiik = 0, so the Levi–
P

Civita is just the exterior derivative D = d. The geodesics ẍi = 0 are straight lines
γp (t, a) = p + at parameterized with constant velocity.
2. Consider the sphere S n with the round metric (i.e. the restriction of the Euclidean
metric to S n ⊂ Rn+1 ). Then p ∈ S n and a ∈ Tp S n may be regarded as the vectors
in Rn+1 . Suppose a 6= 0, then the orthogonal reflection L in the 2-dimensional subspace
P = span{p, p + a} is an isometry of S n . Now L preserves the metric and p and a, the data
which determines the geodesic γp (·, a). As γp (·, a) is moreover uniquely determined it must
be contained in the fixed point set of L. But the fixed point set is a curve, the great circle
P ∩ S n . We find that great circles, parameterized with velocity of constant length—and
only these—are the geodesics on S n .

Observe that for any geodesic γp (t, a) and any real constant λ the path γp (λt, a) is also
a geodesic and γp (λt, a) = γp (t, λa).
By application of a general result in the theory of ordinary differential equations, a
geodesic γp (t, a) must depend smoothly on its initial conditions p, a. Furthermore, there
exist ε1 > 0 and ε2 > 0 independent of a and such that if |a| < ε1 then γp (t, a) exists for
all −2ε2 < t < 2ε2 . It follows that γp (1, a) is defined whenever |a| < ε = ε1 ε2 .

Definition. The exponential map at a point p of a Riemannian manifold (M, g) is

expp : a ∈ Ball ε (0) ⊆ Tp M → γ(1; p, a) ∈ M.

Proposition 3.14. (d expp )0 = id(Tp M )

Proof. We use the canonical identification a ∈ Tp M → dtd (ta)|t=0 to define (d expp )0 as a

linear map on Tp M (rather than on T0 (Tp M )).
Let |a| < ε, so γp (t, a) = γp (1, ta) is defined for 0 ≤ t ≤ 1. Then we have
d
(d expp )0 a = dt
expp (ta)|t=0
d
= γ (1, ta)|t=0
dt p
d
= γ (t, a)|t=0
dt p
= γ̇(0, a) = a.

Corollary 3.15. The exponential map expm defines a diffeomorphism from a neighbour-
hood of zero in Tm M to a neighbourhood of m in M .

Proof. Apply the Inverse Mapping Theorem (page 11 of these notes).

Corollary 3.15 means that the exponential map defines near every point p of a Rieman-
nian manifold a system of local coordinates—called normal (or geodesic) coordinates
at p. It is not difficult to see that the geodesics γp (t, a) are given in these coordinates by
rays emanating from the origin.
 differential geometry 43

It also makes sense to speak of geodesic polar coordinates at p ∈ M defined by the

polar coordinates on Tp M via a diffeomorphism

f : (r, v) ∈ ]0, ε[ × S n−1 → expp (r v) ∈ M. (3.16)

Here ]0, ε[×S n−1 is regarded as a subset in Tp M ∼ = Rn via the inner product g(p). If
0 < r < ε then the image Σr = f ({r} × S n−1 ⊂ Tp M ) of the metric sphere of radius r
is well-defined on M and is called a geodesic sphere about p. (So Σr is an embedded
submanifold of M .) The following remarkable result asserts that ‘the geodesic spheres are
orthogonal to their radii’.
Gauss Lemma. The geodesic γp (t, a) is orthogonal to Σr . Thus the metric g in geodesic
polar coordinates has local expression g = dr2 + h(r, v), where for any 0 < r < ε, h(r, v) is
the metric on Σr induced by restriction of g.
Proof. Let X be an arbitrary smooth vector field on the unit sphere S n−1 ⊂ Tp M . Use
polar coordinates to make sense of X as a vector field (independent of r) on the punctured
unit ball B \ {0} ⊂ Tp M . Define a vector field X̃(r, v) = rX(v) on B \ {0}. The map
expp induces a vector field Y (f (r, v)) = (d expp )rv X̃(r, v) on the punctured geodesic ball
B 0 \ {p} = expp (B \ {0}) in M .
We shall be done if we show that Y is everywhere orthogonal to the radial vector
∂
field ∂r . Note that, by construction, any geodesic from p is given in normal coordinates
∂
by γp (t, a) = at, so γ̇p (t, a)/|a| = ∂r . Here |a| means the norm in the inner product gp
on the vector space Tp M . By application of Corollary 3.15, the family γ̇p (t, a), where
|a| = 1 and 0 < |t| < ε, defines a smooth vector field on B 0 \ {p}. Recall from (3.13) that
Dγ̇ γ̇ = 0 for any geodesic γ, where D denotes the Levi–Civita covariant derivative. Also
d ∂ ∂
dt
g( ∂r , ∂r ) = dtd g(γ̇, γ̇) = 0 by Proposition 3.12, so g( ∂r
∂ ∂
, ∂r ) = 1. It remains to show that
g(Y, γ̇) = 0.
Using the diffeomorphism f in (3.16) to go to polar geodesic coordinates, we obtain

∂

d
Dγ̇ Y − DY γ̇ = (df ) D ∂ X̃ − DX̃ ∂r = (df ) X̃ = (df )(X̃/r) = Y /r,
∂r dr
with the help of Proposition 3.5. Therefore, we find
d 1 1
g(Y, γ̇) = g(Dγ̇ Y, γ̇) + g(Y, Dγ̇ γ̇) = g(Dγ̇ Y, γ̇) = g(DY γ̇ + Y, γ̇) = g(Y, γ̇).
dr r r
d
as 2g(Dγ̇, γ̇) = d g(γ̇, γ̇) = 0 by Proposition 3.13. Thus dr G = G/r, where G = g(Y, γ̇).
d d ∂
Hence G is linear in r and dr G independent of r. But limr→0 dr G = limr→0 g(X, ∂r ) = 0, as
(d expp )0 is an isometry by Proposition 3.14, and so g(Y, γ̇) = 0 and the result follows.

3.3 Curvature of a Riemannian manifold

Let g be a metric on a manifold M . The (full) Riemann curvature R = R(g) of g is, by
definition, the curvature of the Levi–Civita connection of g. Thus R ∈ Ω2M (End(T M )),
44 alexei kovalev

locally a matrix of differential 2-forms R = 12 (Rj,kl i

dxl ∧ dxk ), i, j, k, l = 1 . . . n = dim M .
i
The coefficients (Rj,kl ) form the Riemann curvature tensor of (M, g). Given two vector
fields X, Y , one can form an endomorphism field R(X, Y ) ∈ Γ(End(T M )); its matrix in
local coordinates is R(X, Y )ij = Rj,kl i
X k Y l (as usual X = X k ∂k , Y = Y l ∂l ). Denote
Rkl = R(∂k , ∂l ) ∈ End(Tp M ) (here p is any point in the coordinate neighbourhood).
Recall that in local coordinates a connection (covariant derivative) may be written as
d + A, with A = Γijk dxk = Ak dxk . We write Dk = D ∂ = ∂x∂ k + Ak . The definition of the
∂xk
curvature form of a connection (Chapter 2, p. 31) yields an expression in local coordinates
i

i
Rj,kl = Dl Dk ∂x∂ j − Dk Dl ∂x∂ j , or Rkl = −[Dk , Dl ], (3.17)

considering the coefficient at dxl ∧ dxk . Now DX = X k Dk and so we have −[DX , DY ] =

−[X k Dk , X l Dl ] = −X k (∂k Y l )Dl − X k Y l Dk Dl + Y k (∂k X l )Dl + X k Y l Dl Dk = X k Y l Rkl −
[X, Y ]l Dl . We have thus proved
Lemma 3.18. R(X, Y ) = D[X,Y ] − [DX , DY ].
One also can combine (3.17) with (3.8) and thus obtain an explicit local expression for
i
Rj,kl in terms of the coefficients of the metric g and their first and second derivatives.
q
It is convenient to consider Rij,kl = giq Rj,kl , which defines a map on 4-tuples of vector
fields (X, Y, Z, T ) 7→ g(R(X, Y )Z, T ).
Proposition 3.19.
(i) Rij,lk = −Rij,kl = Rji,kl ;
i i i
(ii) Rj,kl + Rk,lj + Rl,jk = 0 (the first Bianchi identity 2 );

(iii) Rij,kl = Rkl,ij .

Proof. (i) The first equality is clear. For the second equality, one has, from the definition
of the Levi–Civita connection, ∂g kl
∂xi
= g(Di ∂x∂ k , ∂x∂ l ) + g( ∂x∂ k , Di ∂x∂ l ), and further

∂ 2 gkl ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
j i
= g(Dj Di k , l ) + g(Di k , Dj l ) + g(Dj k , Di l ) + g( k , Dj Di l ).
∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x ∂x
2
∂ gkl ∂ gkl2
The right-hand side of the above expression is symmetric in i, j as ∂x i ∂xj = ∂xj ∂xi . The

anti-symmetric part of the right-hand side (which has to be zero) equals Rij,kl + Rji,kl .
(ii) Firstly, (Dk ∂x∂ j )i = Γijk = (Dj ∂x∂ k )i , by the symmetric property of the Levi–Civita.
The claim now follows by straightforward computation using (3.17).
We note for use in the proof of (iii) that multiplying (ii) by giq gives Rij,kl + Rik,lj +
Ril,jk = 0.
(iii) We organize the argument using the vertices and faces of an octahedron (see the
next page).
2
also known as the algebraic Bianchi identity, not to be confused with the differential Bianchi identity
in Chapter 2.
 differential geometry 45

Assign to each vertex a simulta-

neous application of the two iden-
tities in (i). Then, for each shaded
face, we may arrange the three co-
efficients of R to have the same
first index (indicated by a letter
in the middle of the face) so that
the Bianchi identity (ii) can be ap-
plied. Adding the instances of (ii)
for the two upper shaded faces and
subtracting those for the two lower
shaded faces, we obtain the re-
quired identity (iii) as all the terms
in the vertices of the equatorial
square cancel.3

Remark . Notice that the proof of (ii) shows the first Bianchi identity is valid for every
symmetric connection on M .

Corollary 3.20. The Riemann curvature tensor (Rij,kl )p defines, at any point p ∈ M a
symmetric bilinear form on the fibres of Λ2 Tp M .

There are natural ways to extract “simpler” quantities (i.e. with less components) from
the Riemann curvature tensor.

Definition. The Ricci curvature of a metric g at a point p ∈ M , Ricp = Ric(g)p , is the

trace of the endomorphism v → Rp (x, v)y of Tp M depending on a pair of tangent vectors
x, y ∈ Tp M .
P q
Thus in local coordinates Ric(p) is expressed as a matrix Ric = (Ricij ), Ricij = q Ri,jq .
That is, the Ricci curvature at p is a bilinear form on Tp M . A consequence of Proposi-
tion 3.19(iii) is that this bilinear form is symmetric, Ricij = Ricji .

Definition . The scalar curvature of a metric g at a point p ∈ M , s = scal(g)p is a

smooth function on M obtained by taking the trace of the bilinear form Ricij with respect
to the metric g.

If local coordinates
P are chosen soPthat gij (p) = δij at a point, then the latter definition
means that s(p) = Ricii (p) = i,j Rij,ji (p). For a general gij , the formula may be
P ij i ij
written as s = i g Ricij , where g is the induced inner product on the cotangent space
with respect to the dual basis, algebraically (g ij ) is the inverse matrix of (gij ).
3
I learned this argument from the lectures of M.M. Postnikov.
46 alexei kovalev

3.3.1 Some examples

(1) It makes sense to consider the condition

Ric = λg (3.21)

for some constant λ ∈ R, as both the metric and its Ricci curvature are symmetric bilinear
forms on the tangent spaces to M . When the condition (3.21) is satisfied, the Riemannian
manifold (M, g) is called Einstein manifold. In particular, if (3.21) holds with λ = 0 then
M is said to be Ricci-flat.
(2) Recall that if Σ is a surface in R3 (smooth, regularly parameterized by (u, v) in an
open set in R2 ) then there is a metric induced on Σ, expressed as the first fundamental
form Edu2 + 2F dudv + Gdv 2 . The second fundamental form Ldu2 + 2M dudv + N dv 2 is
defined by taking the inner products L = (ruu , n), M = (ruv , n), N = (rvv , n) with the
unit normal vector to Σ, n = ru × rv /|ru × rv | (the subscripts u and v denote respective
partial derivatives). The quantity
LN − M 2
K=
EG − F 2
is called the gaussian curvature of Σ. A celebrated theorema egregium, proved by Gauss,
asserts that K is determined by the coefficients of first fundamental form, i.e. by the metric
on Σ (and so K is independent of the choice of an isometric embedding of Σ in R3 ).
Taking up a general view on Σ as a 2-dimensional Riemannian manifold, one can check
that 2(EG − F 2 )−1 R12,21 = s, the scalar curvature of Σ. From the results of the next
section, we shall see (among other things) that the scalar curvature of a surface Σ is twice
its gaussian curvature s = 2K.

3.4 Riemannian submanifolds

When a manifold M n is an embedded submanifold of a Riemannian manifold, say V n+r ,
the Riemannian metric gV on V induces, by restriction, a Riemannian metric gM on M .
What is the relation between the Levi–Civita connection D̃ of gV and the Levi–Civita
connection D of gM ?
To see this relation, it is convenient to consider the vector bundle E = ι∗ (T V ) over M ,
where ι : M ,→ V is the embedding map. (Informally, E is just the restriction of T V to M
if the latter is regarded as a subset of V .)
In the next proposition, we write xk for local coordinates on M , y γ for local coordinates
on V , and α, β, γ = 1, . . . , n + r.
Proposition 3.22. Any connection ∇ ˜ on V induces in a canonical way a connection on E
∂y γ α ˜ and y = y(x) is the
α
with the coefficients Γβk = ∂xk Γβγ , where Γαβγ are the coefficients of ∇
local expression of the embedding ι.
˜ the connection on E defined by the above proposition. For
We shall still denote by ∇
p ∈ E, consider the tangent space Tp E as a subspace of Tp V and then the corresponding
 differential geometry 47

horizontal subspace of Tp E is just the intersection Sp ∩ Tp E, where Sp ⊂ Tp V is the

horizontal subspace for the connection on V .
There is also an interpretation in terms of the covariant derivatives (needed for the
proof of Gauss–Weingarten formulae below). Any local vector field X on M (respectively
local section σ of E) can be extended smoothly to a local vector field X̃ (respectively σ̃)
on V . Then (∇ ˜ σ̃)|M = ∇˜ X σ, where in the left-hand side we use the connection on E. In
X̃
particular, the right-hand side is independent of the choices of extensions X̃ and σ̃.
Thus the connection ∇ ˜ on E makes natural sense from all three points of view. Note
that we did not require any metric to define this induced connection.
Each fibre Ex of E contains Tx M as a subspace. Using now the metric on M we obtain
a direct sum decomposition
Ex = Tx M ⊕ (Tx M )⊥ . (3.23)
The disjoint union of the orthogonal complements tx∈M (Tx M )⊥ forms a vector bundle of
rank r over M called the normal bundle of M in V , denoted NM/V . Exercise: verify
that NM/V is indeed a well-defined vector bundle (recall Theorems 1.8 and 2.4).
For any two vector fields X, Y on M , we can decompose the covariant derivative
˜
(∇X Y )x = (∇X Y )x + (h(X, Y ))x , according to (3.23), where h(X, Y ) is some section
of NM/V . It turns out that ∇ is a well-defined covariant derivative (connection) on M and
h is a bilinear map Tx M × Tx M → (Tx M )⊥ (depending smoothly on x). Furthermore, in
the case when ∇ ˜ = D̃ is the Levi–Civita connection on V we obtain.

Theorem 3.24 (Gauss formula). For any vector fields X, Y on M ,

D̃X Y = DX Y + II(X, Y ),

where D is the Levi–Civita connection of the induced metric on M , and II is a symmetric

bilinear map called the second fundamental form of M in V .
Theorem 3.25 (Weingarten formula). For any vector field X on M and section ξ of the
normal bundle NM/V ,
D̃X ξ = −Sξ X + ∇0X ξ,
where for any ξ, Sξ is a endomorphism of the vector bundle T M called the shape operator
and ∇0 is a connection on NM/V . Furthermore, the shape operator is symmetric with respect
to the induced Riemannian metric M ,

gM (Sξ X, Y ) = gM (X, Sξ Y ) = gV (II(X, Y ), ξ),

for any vector field Y on M .

By direct application of the above, we can compute the Riemann curvature R = (Rij,kl )
of M in terms of the curvature of the ambient manifold and the second fundamental form.
Theorem 3.26 (Gauss).

R(X, Y, Z, T ) = R̃(X, Y, Z, T ) + gV (II(X, Z), II(Y, T )) − gV (II(X, T ), II(Y, Z)).

48 alexei kovalev

Corollary 3.27. The curvature of a submanifold M of a flat manifold is determined by

the second fundamental form of M .
When M is a surface in the Euclidean R3 , this is equivalent to theorema egregium
discussed in the previous section.

3.5 Laplace–Beltrami operator

Throughout this section M is a connected oriented Riemannian manifold of dimension n.
Let g denote a metric on M and let the orientation be given by a nowhere-zero n-form ε.
Starting from the vector fields ∂x∂ 1 , . . . ∂x∂n at a point x in a coordinate neighbourhood
U , we can apply Gram–Schmidt process with x as a parameter. Thus we obtain a new
system of (smooth) vector fields e1 , . . . , en which give an orthonormal basis of tangent
vectors on a perhaps smaller neighbourhood of p (still denote this neighbourhood by U ).
Let ω1 , . . . , ωn on U be the dual 1-forms to e1 , . . . , en , in the sense that
ωj (ei ) = δij at any point in U.
Then ωj give at every point p of U a basis of Tp∗ M , the dual basis to ej .
The metric on M induces, for every p = 0, . . . , n an inner product on the bundle Λp T ∗ M
by making {ωi1 (x) ∧ . . . ∧ ωip (x) : 1 ≤ i1 < . . . < ip ≤ n} an orthonormal basis of Λp Tx∗ M .
If ωj0 is another system of local 1-forms, on another coordinate neighbourhood U 0 say,
and ωj0 are orthonormal at every point in U 0 then
ω10 ∧ . . . ∧ ωn0 = det(Φ) ω1 ∧ . . . ∧ ωn on U 0 ∩ U,
for some orthogonal matrix Φ (depending on x ∈ U 0 ∩ U ). Assuming, as we can on
an oriented M , that all the coordinate neighbourhoods are chosen so that the Jacobians
det(Φ) are positive on the overlaps, we find that ω1 ∧ . . . ∧ ωn is a well-defined nowhere-zero
n-form ωg on all of M . We can further ensure that ωg = aε for some positive function
a ∈ C ∞ (M ). Then ωg is called the volume form of pM .
In (positively oriented) local coordinates, ωg = det gij dx1 ∧ . . . ∧ dxn .
Definition. The Hodge star on M is a linear operator on the differential forms
∗ : Λp Tx∗ M → Λn−p Tx∗ M,
such that for any two p-forms α, β ∈ Λp Tx∗ M one has α ∧ ∗β = hα, βig ωg (x), where ωg is
the volume form on M .
It follows that if ωi is an orthonormal basis of a cotangent space Tx∗ M then necessarily
∗(ω1 ∧ . . . ∧ ωp ) = ωp+1 ∧ . . . ∧ ωn . In particular, ∗1 = ωg and ∗ωg = 1. By permutations
of indices and by linearity, the Hodge star is then uniquely determined for any differential
form on M . Further, it follows that ∗∗ = (−1)p(n−p) on the p-forms.
Using the Hodge star we construct a differential operator
δ : Ωp (M ) → Ωp−1 (M )
putting δ = (−1)n(p+1)+1 ∗ d∗ if p 6= 0 and δ = 0 on Ω0 (M ) = C ∞ (M ).
 differential geometry 49

Definition. The Laplace–Beltrami operator, or Laplacian, on M is a linear differen-

tial operator ∆ : Ωp (M ) → Ωp (M ) given by

∆ = δd + dδ.

Straightforward computation shows that when M is the Euclidean Rn the definition

∂ 2f ∂ 2f
gives ∆f = − − . . . − for any smooth function f . For a general metric
(∂x1 )2 (∂xn )2  
1 ∂ p ij ∂f
g = (gij ), the local expression becomes ∆g f = − √ det g g .
det g ∂xj ∂xi
Proposition 3.28. The operator δ is the adjoint 4 of d in the sense that
Z Z
hdα, βig ωg = hα, δβig ωg ,
M M

for every compactly supported α ∈ Ωp−1 (M ), β ∈ Ωp (M ).

Using the inner product on the spaces Λp Tp∗ M , p ∈ M , we
R can define an inner product on
p 2
Ω (M ), called the L inner product, by putting hα, βiL2 = M hα, βig ωg . The inner product
makes each Ωp (M ) into a normed space, with L2 -norm defined by kαk = (hα, αiL2 )1/2 . In
particular, α = 0 if and only if kαk = 0.
Thus Proposition 3.28 says that hdα, βiL2 = hα, δβiL2 and, consequently, h∆α, βiL2 =
hα, ∆βiL2 . It follows immediately that the Laplace–Beltrami operator is self-adjoint.
A differential form α ∈ Ωp (M ) is called harmonic if ∆α = 0.
Corollary 3.29. Every harmonic differential form on a compact manifold is closed and
co-closed: ∆α = 0 if and only if both dα = 0 and δα = 0.
Proof. Integration by parts, 0 = hδdα + dδα, αiL2 = hδα, δαiL2 + hdα, dαiL2 .
It is also easily checked that ∗∆ = ∆∗ on any Ωp (M ). Therefore the Hodge star of any
harmonic form is again harmonic.
Hodge Decomposition Theorem. Let M be a compact oriented Riemannian manifold.
For every 0 ≤ p ≤ dim M , the space Hp of harmonic p-forms is finite-dimensional. Fur-
thermore, there are L2 -orthogonal direct sum decompositions

Ωp (M ) = ∆Ωp (M ) ⊕ Hp
= dδΩp (M ) ⊕ δdΩp (M ) ⊕ Hp
= dΩp−1 (M ) ⊕ δΩp+1 (M ) ⊕ Hp

(where we formally put Ω−1 (M ) = {0}).

Remark: the compactness condition on M cannot be removed.5
4
It is more correct to say that δ is the ‘formal adjoint’ of d for reasons that have to do with the Analysis.
5
The reason is that certain results in Analysis fail on non-compact sets, but this is another story.
50 alexei kovalev

Short summary of the proof. We need to introduce the concept of a weak solution of
∆ω = α. (3.30)
A weak solution of (3.30) is by definition, a linear functional l : Ωp (M ) → R which is
(i) bounded, |l(β)| ≤ Ckβk, for some C > 0 independent of β, and
(ii) satisfies l(∆ϕ) = hα, ϕiL2 .
Any solution ω of (3.30) defines a weak solution by putting lω (β) = hω, βiL2 .
The proof of Hodge Decomposition Theorem requires some results from Functional
Analysis.
Regularity Theorem. Any weak solution l of (3.30) is of the form l(β) = hω, βiL2 , for
some ω ∈ Ωp (M ) (and hence defines a solution of (3.30)).
Compactness Theorem. Assume that a sequence αn ∈ Ωp (M ) satisfies kαn k < C and
k∆αn k < C, for some C independent of n. Then αn contains a Cauchy subsequence.
We shall assume the above two theorems (and the Hahn–Banach theorem below)
without proof.
Compactness Theorem implies at once that Hp must be finite-dimensional (for,
otherwise, there would exist an infinite orthonormal sequence of forms). As Hp is finite-
dimensional, we can write an L2 -orthogonal decomposition Ωp (M ) = Hp ⊕ (Hp )⊥ .
It is easy to see that ∆Ωp (M ) ⊆ (Hp )⊥ (use Proposition 3.28). For the reverse inclusion,
suppose that α ∈ (Hp )⊥ . We want to show that the equation (3.30) has a solution.
Assuming the Regularity Theorem, we shall be done if we obtain a weak solution l :
Ωp (M ) → R of (3.30).
Define l first on a subspace ∆Ωp (M ), by putting l(∆η) = hη, αiL2 . It is not hard to
check that l is well-defined. Further, (ii) is automatically satisfied (on this subspace); we
claim that (i) holds too. To verify the latter claim, we show that l is bounded below on
∆Ωp (M ) using, once again, the Compactness Theorem.
In order to extend l to all of Ωp (M ), we appeal to
Hahn–Banach Theorem. Suppose that L is a normed vector space, and L0 a subspace
of L, and l : L0 → R a linear functional satisfying l(x0 ) < kx0 k, for all x0 ∈ L0 . Then l
extends to a linear functional on L with l(x) < kxk for all x ∈ L.
Thus we obtain a weak solution of (3.30) and deduce that Ωp (M ) = ∆Ωp (M ) ⊕ Hp
as desired. The two other versions of the L2 -orthogonal decomposition of Ωp (M ) follow
readily by application of Proposition 3.28.

Corollary 3.31. Every de Rham cohomology class a ∈ H r (M ) of a compact oriented Rie-

mannian manifold M is represented by a unique harmonic differential r-form α ∈ Ωr (M ),
[α] = a. Thus Hr ∼
= H r (M ).
Proof. Uniqueness. If α1 , α2 are harmonic p-forms and α1 − α2 = dβ then kdβk2 =
hdβ, α1 − α2 iL2 = hβ, δ(α1 − α2 )iL2 = 0.
Existence. If α is such that dδα = 0 then kδαk = 0. Hence any closed p-form must
be in dΩp−1 (M ) ⊕ Hp .
 differential geometry 51

Corollary 3.31 is a surprising result: an analytical object (harmonic forms) turns out to
be equivalent to a topological object (de Rham cohomology) via some differential geometry.
Here is a way to see ‘why such a result can be true’.
A de Rham cohomology class, a ∈ H r (M ) say, can be represented by many differential
forms; consider the (infinite-dimensional) affine space

Ba = {ξ ∈ Ωr (M ) | dξ = 0, [ξ] = a ∈ H r (M )}
= {ξ ∈ Ωr (M )|ξ = α + dβ, for some β ∈ Ωr−1 (M )}.

When does a closed form α have the smallest L2 -norm amongst all the closed forms in a
given de Rham cohomology class Ba ?
Such a form must be a critical point of the function F (α + dβ) = kα + dβ)k2 on Ba , so
the partial derivatives of F in any direction should vanish. That is, we must have

d
0= hα + t dβ, α + t dβiL2 = 2hα, dβiL2 .
dt t=0
R
Integrating by parts, we find that M hδα, βig = 0 must hold for every β ∈ Ωr−1 (M ). This
forces δα = 0, and so the extremal points of F are precisely the harmonic forms α.

Page references for Chapter 3

all to Gallot–Hulin–Lafontaine except ∗ to Warner
Riemannian metrics — pp.52–53,
Levi–Civita connection — pp.69–71,
curvature of a Riemannian manifold — pp.107–108,111–112,155–156
geodesics, Gauss Lemma — pp.80–85,89–90,
Riemannian submanifolds — pp.217–220,
∗
the Laplacian and the Hodge Decomposition Theorem — [W], pp. 140–141,220–226.
Part III: Differential geometry (Michaelmas 2010)

Some facts from multilinear algebra

Let V be a vector space over R. Assume, for simplicity, that V has finite dimension n say.

1. If W is another real finite-dimensional vector space then the tensor product V ⊗ W may
be defined as the real vector space consisting of all formal linear combinations of elements
v ⊗ w (for v ∈ V , w ∈ W ), with the relations

(av) ⊗ w = v ⊗ (aw) = a(v ⊗ w),

(v1 + v2 ) ⊗ w = v1 ⊗ w + v2 ⊗ w, and v ⊗ (w1 + w2 ) = v ⊗ w1 + u ⊗ w2 ,

for all a ∈ R; v, v1 , v2 ∈ V ; w, w1 , w2 ∈ W . If {vi } and {wj } are bases of V, W then {vi ⊗ wj }

is a basis of V ⊗ W and thus dim(V ⊗ W ) = dim V dim W . For example, Rm ⊗ Rn ∼ = Rmn .
Consider the dual space V ∗ of linear functions V → R, and similarly W ∗ , (V ⊗ W )∗ .
Assigning to each v ∗ ⊗ w∗ ∈ V ∗ ⊗ W ∗ the linear function on V ⊗ W determined by

vi ⊗ wj ∈ V ⊗ W 7→ v ∗ (vi ) w∗ (wj ) ∈ R

defines a natural isomorphism of vector spaces V ∗ ⊗ W ∗ ∼

= (V ⊗ W )∗ .
Assigning to a bilinear form ψ on V × W the linear function

v ⊗ w ∈ V ⊗ W → ψ(v, w) ∈ R

defines a natural isomorphism between the vector space of all bilinear forms on V × W and
the dual space (V ⊗ W )∗ . Recall from linear algebra that the space of bilinear forms on V × U
may be naturally identified with the space L(V, U ∗ ) of linear maps V → U ∗ . Putting U ∗ = W
and noting the above relations, one obtains a natural linear isomorphism

L(V, W ) ∼
= V ∗ ⊗ W.

In the special case W = R this recovers the definition of dual vector space V ∗ .

2. Again let v1 , . . . , vn be a basis of V . Then the dual space V ∗ of linear functions V → R

can be given the dual basis λ1 , . . . , λn , defined by the property

λj (vj ) = δij .

Here δij is the ‘Kronecker delta’, δij = 1 if i = j and is 0 otherwise. The p-th exterior power
Λp V ∗ of V ∗ (p ≥ 0) is the vector space of all the functions h : V × . . . × V → R, such that h is
(1) multilinear: h(u1 , . . . , aui + bu0i , . . . , up ) = ah(u1 , . . . , ui , . . . , up ) + bh(u1 , . . . , u0i , . . . , up );
(i.e. linear in each argument) and
(2) antisymmetric: ∀i < j ∈ {1, 2, . . . , n}, swapping the ith and j th vector changes the sign

h(u1 , . . . , ui−1 , ui , ui+1 , . . . , uj−1 , uj , uj+1 , . . . , un ) =

− h(u1 , . . . , ui−1 , uj , ui+1 , . . . , uj−1 , ui , uj+1 , . . . , un ).

It follows that for 1 ≤ p ≤ n, dim Λp V ∗ = np , a basis may be given by {λi1 ∧ . . . ∧ λip : λij ∈

V ∗ , 1 ≤ i1 < . . . < ip ≤ n}. Also Λp V ∗ = {0} when p > n. One formally defines Λ0 V ∗ = R.
 The exterior product (or wedge product) is a bilinear map

(λ, µ) ∈ Λp V ∗ × Λq V ∗ → λ ∧ µ ∈ Λp+q V ∗ .

determined for the basis vectors λi ∈ V ∗ , and inductively on p, q, by

(λi1 ∧. . .∧λip )∧(λip+1 ∧. . .∧λip+q )(u1 , . . . , up+q ) = det(λij (uk )), (uk ∈ V, j, k = 1, . . . , p+q).

(and extended by linearity). It follows, in particular, that the ∧ is associative (λ ∧ µ) ∧ ν =

λ ∧ (µ ∧ ν), and µ ∧ λ = (−1)pq λ ∧ µ (λ ∈ Λp V ∗ , µ ∈ Λq V ∗ ).

Alexei Kovalev ([email protected])