A short course on differential

geometry
Besançon - Autumn 2022

Jack Borthwick
December 9, 2022

Avant-propos
Ces notes ont été rédigées dans le cadre d’un cours doctoral de 10h donné à l’École
Doctorale Carnot-Pasteur. Elles s’adressent avant tout aux doctorants désireux d’avoir
une introduction rapide aux variétés différentielles mais également aux étudiants de
master qui n’ont jamais fait de géométrie et qui n’ont que des notions de topologie
de niveau licence.
J’aimerais remercier L. Jeanjean pour sa lecture détaillée de ces notes ainsi que ses
commentaires qui m’ont permis de les améliorer.

I. Manifolds and submanifolds

I.A. The Definition

Definition I.1
A topological manifold M modeled on a Banach space (E, k·k), or E-manifold,
is a Hausdorff topological space such that every point p ∈ M admits an open
neighbourhood that is homeomorphic to an open subset of E.

• If E = Rn , we say that M is of dimension n.

1
 • In the first version of these notes I had assumed “paracompactness” in order
to treat in a unified way infinite dimensional and finite-dimensional manifolds.
However, upon further reflexion it appealed to me that this was not the ideal
way to develop the basic theory as whilst it is a desired feature of the topol-
ogy of manifolds, paracompactness does not behave well with respect to the
usual set-theoretic operations. For instance, subsets of paracompact spaces, or
products of paracompact spaces may not be paracompact. Instead, one should
work with sufficient conditions for paracompactness that are better behaved.
For finite-dimensional manifolds, it is custom to assume that the manifold is
second-countable; and we will assume this. This means that we suppose that
the topology has a countable basis. Since finite dimensional manifolds are
automatically locally compact and [locally compact + second-countable] ⇒
[paracompact]. Subspaces of second-countable spaces are second-countable;
metric spaces are second-countable if and only if they are separable. In the
infinite dimensional case, a sufficient condition would be to suppose that the
topology is regular and second-countable.

• A local chart is a couple (U, φ) where U is an open subset of M and φ : U →

φ(U ) a homeomorphism onto an open subset φ(U ) ⊂ E.

• An atlas is a collection of charts that cover the manifold.

Example I.1 (The stupid one). Open subsets of Banach spaces.

Example I.2 (Projective space). Let R∗ act on Rn+1 \ {0} by left multiplication and
consider the quotient space RPn = Rn+1 /R∗ , equipped with the quotient topology.
This is intuitively the space of all lines in Rn+1 ; we shall show that it is a n-dimensional
topological manifold.
Recall that the quotient topology is the finest topology such that the canonical
projection: π : Rn+1 \ {0} → RPn is continuous; it is described as follows:

U is open in RPn ⇔ π −1 (U ) is open in Rn+1 \ {0}.

It satisfies the following universal property: let f : Rn+1 → X be a continuous

map, such that π(x) = π(y) ⇒ f (x) = f (y), then there is a unique continuous map
f˜ such that the following diagram is commutative:

Rn+1 \ {0} π
RP n
f
f˜
X

2
 The first step is to show that the quotient topology is Hausdorff and second-
countable; neither are guaranteed for general quotients of topological spaces hav-
ing these properties. However, our quotient comes from a continuous action of a
topological group; the following are general arguments we record here for future
reference:

Proposition I.1

Let (G, ·) be a topological group, X a topological space and φ : G × X →

X; (g, x) 7→ g · x a continuous group action. Let X/G be the quotient set
equipped with the quotient topology, then the canonical projection π : X −→
G/X is an open map.

Proof. Let U be open in X . In order to show that π(U ) is open in X/G one must
show that π −1 (π(U )) is open in X . One has:
[
π −1 (π(U )) = {x ∈ X, ∃g ∈ G, g · x ∈ U } = gU.
g∈G

For any g ∈ G, Lg : x 7→ g · x is a homeomorphism, and in particular, an open map.

Therefore, for any open set U in X , and any g ∈ G, gU ≡ Lg (U ) is open. Hence,
π(U ) is open.
One can now appeal to:

Corollary I.1. If X is second-countable, X/G is second-countable.

Proof. Let (Vn )n∈N be a countable basis for the topology of X , the claim is that
(π(Vn ))n∈N is a basis for the topology of X/G.
Let [x] = π(x) ∈ X/G and U an open neighbourhood of [x]. By definition, π −1 (U )
is an open neighbourhood of x ∈ X , therefore one can find n0 ∈ N such that
x ∈ Vn0 ⊂ π −1 (U ), it follows that: [x] = π(x) ∈ π(Vn0 ) ⊂ U .
As an immediate consequence:

RPn is second-countable.

We will now give a general argument that guarantees separation; it is slightly more
involved and requires an additional condition on the group action.

3
 Definition I.2
A continuous action of a topological group G on a locally compact topological
space X is said to be proper, if for any compact subset K ⊂ X the set:

{g ∈ G, gK ∩ K 6= ∅}

is compact.

The following proposition justifies this definition.

Proposition I.2

Let φ : G×X → X be a proper group action of a topological group on a locally

compacta topological space X , then the quotient space X/G is Hausdorff.
a
Some authors do not assume that locally compact spaces are Hausdorff, for us this hypoth-
esis will always be implicit.

Proof. We shall divide the proof into two steps; first make the following observation,
since the canonical projection π : X 7→ X/G is an open map, X/G is Hausdorff it
is sufficient to show that R = {(x, g · x), x ∈ X, g ∈ G} is closed in X × X .
To see this, assume that R is closed and suppose that π(x) 6= π(y) for some
x, y ∈ X ; then (x, y) ∈ (X × X) \ R. This set is open, therefore one can find
neighbourhoods Vx and Vy of x and y respectively such that Vx × Vy ⊂ (X × X) \ R,
but then: π(Vx ) and π(Vy ) are open neighbourhoods of π(x) and π(y) such that
π(Vx ) ∩ π(Vy ) = ∅. Indeed: if π(z) ∈ π(Vx ) ∩ π(Vy ) 6= 0, then ∃g1 , g2 ∈ G such that
(g1 z, g2 z) ∈ Vx × Vy but (g1 z, g2 z) = (g1 z, g2 g1−1 (g1 z)) ∈ R and R ∩ (Vx × Vy ) = ∅.
We shall now prove that R is closed; let (x, y) ∈ (X × X) \ R. Using that X
is locally compact, choose relatively compact neighbourhoods Vx and Vy of x and y
respectively; then Vx ×Vy is a relatively compact neighbourhood of (x, y). Introduce
the map:
G × X −→ X × X,
φ̃ :
(g, x) 7−→ (x, g · x).
Note that φ̃(G × X) = R. Moreover, the image under φ̃ of φ̃−1 (Vx × Vy ) is the set
of points in Vx × Vy that belong to R and that we want to avoid. If we can show that
φ̃−1 (Vx × Vy ) is “small”, i.e. compact, then the image will be compact too and easy
to avoid.
Set K1 = Vx × Vy and K2 = π1 (K1 ) ∪ π2 (K1 ) where π1 , π2 are respectively pro-
jections onto the first and second components of X × X . K2 is a compact subset of
X and so, since the action is proper, G0 = {g ∈ G, gK2 ∩ K2 6= ∅} is compact in G.

4
Now:

φ̃−1 (K1 ) ⊂ φ̃−1 (K2 × K2 ) = {(g, x) ∈ G × X, x ∈ K2 , g · x ∈ K2 } ⊂ G0 × K2 .

So, φ̃−1 (K1 ) is a closed subset of the compact G0 × K2 , hence compact. φ̃ being
continuous it follows that φ̃(φ̃−1 (K1 )) is a compact subset of Vx × Vy . Now Vx × Vy
and (X × X) \ φ̃(φ̃−1 (K1 )) are open neighbourhoods of (x, y) so (Vx × Vy ) ∩ (X ×
X) \ φ̃(φ̃−1 (K1 )) is an open neighbourhood of (x, y) contained in (X × X) \ R. This
proves that R is closed and consequently that X/G is Hausdorff.
The following sequential characterisation of properness is often useful.

Proposition I.3

Let X be a topological space, G a topological group, assume that both are

second-countable and X locally compact. Suppose that G acts continuously
on X , then the action is proper if and only if whenever (gn )n∈N and (xn )n∈N are
sequences of G and X respectively such that (gn ·xn )n∈N and (xn )n∈N converge
then (gn )n∈N has a convergent subsequence.

Proof. Let us prove (⇐), let K be compact, consider G0 = {g ∈ G, gK ∩ K 6= ∅}.

Since G is second-countable, we only need to show that G0 is sequentially compact1 .
Choose a sequence (gn )n∈N in G0 , then for each n ∈ N one can find xn ∈ K such
that gn · xn ∈ K since K is compact, extracting subsequences if necessary, we can
assume that (xn )n∈N and (gn · xn )n∈N converge to x ∈ K and y ∈ K respectively.
By assumption, it follows that (gn )n∈N has a convergent subsequence. Let g be the
limit of this subsequence, by continuity one has: g · x = y ∈ K so that g ∈ G0 . It
follows that G0 is sequentially compact.
To prove (⇒), let us assume that G(K) = {g ∈ G, gK ∩ K 6= ∅} is compact for
every compact K . Suppose that (gn · xn )n∈N and (xn )n∈N converge respectively to y
and x for some sequences (gn )n∈N ,(xn )n∈N . Since X is locally compact let Vx and Vy
be relatively compact neighbourhoods of x and y . For large enough n, gn · xn ∈ Vy
and xn ∈ Vx , so that, for large enough n ∈ N, gn ∈ G(K) where K = V1 ∪ V2 . By
assumption, G(K) is compact and so (gn )n∈N has a convergent subsequence.
We can now apply Proposition I.2, to show that:

RPn is Hausdorff.

Rn+1 \ {0} is certainly locally compact and second-countable and R∗ second-

countable, so we can use the sequential characterisation. Suppose that (λn )n∈N ∈
1
See Appendix A

5
(R∗ )N , (xn )n∈N ∈ (Rn+1 \ {0})N are sequences such that λn xn → y ∈ Rn+1 \ {0}
and xn → x ∈ Rn+1 \ {0}. Convergent sequences are bounded so there is M > 0
such that for all n ∈ N, |λn xn | ≤ M . Futhermore, for large enough n, ||xn || > ||x||
2
,
therefore, for large enough n, |λn | ≤ ||x|| . (λn )n∈N is therefore a bounded sequence
2M

in R and has a convergent subsequence by the Bolzano-Weirestrass theorem. 0 is

not a limit-point of the sequence since |λn xn | > ||y||
2
for large enough n; this shows
that the subsequence converges in R n+1
\ {0}.
At this point, we have worked hard2 to show that the quotient topology on RPn
satisfies the first two (global) conditions of a topological manifold. We will now
check the final – and most important ! – assumption: that we have local charts. For
every i ∈ J1, n + 1K let:

Ui = {x = (x1 , . . . , xn+1 ) ∈ Rn+1 \ {0}, xi 6= 0}

Sn+1
clearly i=1 Ui = Rn+1 \ {0}. Now set:

Ui −→ Rn
φi :
(x1 , . . . , xn+1 ) 7−→ ( xi , . . . , xi , xi , . . . , xn+1
x1 xi−1 xi+1
xi
).

The result only depends on the class of π(x1 , . . . , xn+1 ), so that the map factors to a
continuous map φ̃i : π(Ui ) → Rn ; this is easily seen to be a homeomorphism as the
inverse map can be given explicitly:

Rn −→ π(Ui )
φ̃−1
i :
(y1 , . . . , yn ) 7−→ π((y1 , . . . , yi−1 , 1, yi , . . . , yn )).

A = {(π(Ui ), φ̃i ), i ∈ J1, n+1K} is then an atlas for RPn . Hence, RPn is a topological
manifold of dimension n.

I.B. C k manifolds
Definition I.3

Let M be a topological E-manifold and A an atlas on M , k ∈ N, k ∈ [1, ∞];

A will be said to be of class C k if for any choice of charts (U, φ), (V, ψ) ∈ A
such that U ∩ V 6= ∅ the transition map: φ ◦ ψ −1 is of class C k .

In general, a full atlas will not be known explicitly and given a new chart (φ, U )
there is no way of knowing if it is in the atlas or not. We should therefore work with
atlases that contain all “reasonable” charts in the following sense:
2
More direct proofs certainly exists, but do not show the general principle.

6
 Definition I.4

Let A be a C k -atlas on a manifold M and (U, φ) an arbitrary chart. We shall

say that (U, φ) is compatible with A if its transition map with any chart in A
is of class C k .

Definition I.5

A C k -atlas will be said to be maximal if it contains all compatible charts.

This definition is useful since:

Proposition I.4

Every C k atlas A is contained a unique maximal C k -atlas; that we will write

Â.

Proof. Define an equivalence relation on atlases: A ∼ B if and only if every chart in

B is compatible with A, this means that A ∪ B is an atlas. Taking the union of all
atlases in an equivalence class of atlases yields the desired maximal atlas.
We can now define:

Definition I.6
A C k manifold is a manifold equipped with a C k -atlas. A maximal C k -atlas is
also said to be a C k -differentiable structure on M . A C ∞ -manifold is said to
be smooth.

It should be understood from the discussion in this section that to define a dif-
ferentiable structure on a manifold M , we only need to specify an atlas, and then
consider the unique maximal atlas it is contained in. For instance:
Example I.3. A = {(E, idE )}, E a Banach space, is a smooth atlas on E and the
(canonical) smooth structure on E is that of the maximal atlas  containing A.
Example I.4. On RPn , consider the atlas A = {(Ui , φ̃i ), i ∈ J1, n+1K}. The transition
map: φ̃j ◦ φ̃−1
i is given on φ̃i (Ui ∩ Uj ) by:
 
−1 y1 yi−1 1 yi+1 yj−1 yj+1 yn
φ̃j ◦ φ̃i (y1 , . . . , yn ) = ,..., , , ,..., , ,..., .
yj yj yj yj yj yj yj

This is smooth, so A is a smooth atlas for RPn and the maximal atlas  makes RPn
a smooth manifold.

7
I.C. Submanifolds
Already several pages of theory and only one non-trivial example... to remedy this
let us define submanifolds and prove a classical theorem that produces many exam-
ples. The notion of submanifold is trickier than that of a subspace in a topology and
there are, in fact, several non-equivalent definitions (that allow for different phe-
nomena); we refer to [Sha97, Chapter 1] for an interesting discussion on this topic.
Our definition corresponds to “regular” submanifolds.

Definition I.7

Let M be a smooth manifold, a subset N ⊂ M is a smooth submanifold of M

if for every q ∈ N one can find a chart (U, ψ) near q such that:

ψ(U ) = V1 × V2 , ψ(U ∩ N ) = V1 × {0},

V1 , V2 open subsets of Banach spaces E1 , E2 (that can be thought of as closed

subspaces of E ).

A smooth submanifold N has a natural atlas given by the submanifold charts. Let
q ∈ N and (U, ψ) where ψ : U → V1 × V2 is a homeomorphism, then restricting to
U ∩N and projecting onto the first factor we get a homeomorphism: ψ̃ : U ∩N → V1 .
The transition functions are easily seen to be smooth so these charts yield a smooth
structure on N that is compatible with the subspace topology.
In this definition, V1 is an open subset of a Banach space E1 and one may wonder
if E1 could change from point to point. (This was excluded in Definition I.1, but
could in principle be allowed). In fact, since the transition maps between two charts
are smooth3 diffeomorphisms, we can see that all the possible E1 are topolinear
isomorphic (i.e. there is a continuous vector space isomorphism between them) at
points in the same connected component. So we get a manifold structure according
to our definition on each connected component of the submanifold.
Remark I.1. In the finite dimensional case, this is true on topological manifolds due
to the Brouwer invariance of domain theorem, which implies the topological invariance
of dimension; this second statement is harder to prove than one might think naively
! There are nice proofs of this in homology theory and one can find a clear discussion
in [Nab10]; although it is discussed in most standard books in algebraic topology.
Example I.5. Open subsets of manifolds are submanifolds.
We prove now a (sub-optimal) version of:

3
C 1 is enough.

8
 Theorem I.1: Implicit function theorem

Let E be a Banach space, f : U ⊂ E −→ Rm , a smooth function defined on

an open subset U of E such that for every x ∈ f −1 ({0}), f ′ (x) is surjective;
then f −1 ({0}) is a submanifold of E.

Proof. Let x0 ∈ f −1 ({0}), and set F = ker f ′ (x0 ). E/F ∼ = Rm so has a basis
(e1 , . . . , em ). For each i ∈ J1, mK, choose ẽi in E such that π(ẽi ) = ei where
π : E → E/F is the canonical projection. Set G = span(e1 , . . . , em ) and note
that we have the topological direct sum decomposition E = F ⊕ G; in particular
the projections are continuous.
Indeed, F and G are closed subspaces of E which automatically implies that the
projections pF and pG are continuous. This follows from the closed graph theorem:
suppose that (xn )n∈N converges to x and (pF (xn ))n∈N converges to y , since F is closed
it follows that y ∈ F, similarly, since G is closed and (xn − pF (xn )) is a sequence in
G that converges to x − y : x − y ∈ G. But, x = x − y + y and this decomposition
is unique, thus y = pF (x) and the graph is closed. Therefore, pF and pG = IdE − pF
are continuous.
Define now f˜ : U → F × G, by: f˜(x) = (pF (x), f (x)). We claim that f˜′ (x0 ) ∈
GL(E, F × G). In virtue of the open mapping theorem, it is sufficient to show that
f˜′ (x0 ) is bijective.

• If f˜′ (x0 )h = 0, then pF (h) = 0 and f ′ (x0 )h = 0. So, h ∈ F and h = pF (h) = 0;

this implies that f˜′ (x0 ) is injective.

• Let (f, g) ∈ F×G, since f ′ (x0 ) is surjective one can find h0 such that f ′ (x0 )h0 =
g . Set h = h0 − πF (h0 ) + f , then f˜′ (x0 )h = (f, g). This proves surjectivity.

We can now apply the inverse function theorem to f˜ which provides neighbour-
hoods V and V1 × V2 ⊂ F × G of x0 and (πF (x0 ), 0) respectively, such that f˜|V :
V → V1 ×V2 is a diffeomorphism. This is the desired submanifold chart near x0 .
Remark I.2. The theorem extends to maps with values in Banach spaces, but we
need to strengthen the hypothesis on the derivative of points in x ∈ f −1 ({0}). The
correct hypothesis is that: f ′ (x) is surjective and the short exact sequence
f ′ (x)
0 −→ ker f ′ (x) −→ E −→ F −→ 0
i

splits; this means that E ∼

= ker f ′ (x) ⊕ F. As we have seen, this is automatic if F
is finite dimensional. A differentiable map that satisfies these conditions is called a
submersion.
We now have some more examples of manifolds that arise as submanifolds:

9
Example I.6. S n = {x ∈ Rn+1 , ||x||22 = 1}.
Example I.7. SLn (R) = {A ∈ Mn (R), det A = 1}.
Example I.8. On (R) = {A ∈ Mn (R), At A = In }.
Let us conclude this section by quoting without proof the following theorem:

Theorem I.2: The strong Whitney embedding theorem

Every n-dimensional manifold can be realised as a submanifold of R2n .

Remark I.3. 2n is optimal, but on specific examples one can often do better.

I.D. Appendix: The stereographic projection on the n-sphere

It feels that no course on differential geometry would be complete without men-
tioning the stereographic projection as aPway to define an atlas on the n-sphere. Let
us consider S n = {x ∈ Rn+1 , ||x||22 = n+1 i=1 xi = 1} equipped with the subspace
2

topology. Set N = (0, . . . , 0, 1), and UN = S \ {N }.

n

Let H be the hyperplane of equation xn+1 = 0 and define a map p : UN → H

such that p(x) is the intersection of the line passing through x and N . A short
computation gives:

1
p(x) = (x1 , . . . , xn , 0) , x = (x1 , . . . , xn+1 ) ∈ UN .
1 − xn+1

Identifying H with Rn in a natural way p gives a chart: φN : UN → Rn , φN (x) =

1
1−xn+1
(x1 , . . . , xn ); continuous on UN . It is in fact a homeomorphism, as:
 
2y1 2yn ||y||22 − 1
φ−1
N (y1 , . . . , yn ) = , . . . , , .
1 + ||y||22 1 + ||y||22 1 + ||y||22

Setting S = (0, . . . , 0, −1) and US = S n \ {S} one can define a chart φS in a similar
fashion. Then:
A = {(UN , φN ), (US , φS )},
is an atlas on S n . Let us prove that it is smooth by determining: φS ◦φ−1 N on R \{0},
n

one finds:
1
(φS ◦ φN )−1 (y1 , . . . , yn ) = (y1 , . . . , yn ) .
||y||22
This is smooth on Rn \ {0} so  is a smooth structure on S n .

10
II. The tangent bundle
For simplicity, from now on, all manifolds will be smooth.

II.A. Smooth functions and tangent vectors

Definition II.1
Let M and N be smooth manifolds and f : M → N , we will say that f is
smooth if for any p ∈ M one can find charts (U, φ), (V, ψ), p ∈ U , f (U ) ⊂ V
such that: ψ ◦ f ◦ φ−1 is C ∞ in the usual sense.

This definition does not depend on the charts (U, φ),(V, ψ). Indeed, suppose, for
instance, that: (Ũ , φ̃) is another chart such that p ∈ U , then (on an appropriate
intersection of the domains):
ψ ◦ f ◦ φ̃−1 = (ψ ◦ f ◦ φ−1 ) ◦ (φ ◦ φ̃−1 ),
transition functions are smooth by definition of the smooth atlas, so we can see that
ψ ◦ f ◦ φ̃−1 is C ∞ .
Example II.1. Let N ⊂ M be a submanifold of M , then the inclusion map i : N → M
is smooth. To see this let q ∈ N and (V, ψ), ψ : V → V1 × V2 a submanifold chart;
by definition of its atlas there is a corresponding chart (V1 , ψ̃), now:
ψ ◦ i ◦ ψ̃ −1 : V1 → V1 × V2 ,
is given by:
(ψ ◦ i ◦ ψ̃ −1 )(v) = (v, 0),
which is C ∞ .
Example II.2. The composition of smooth maps is smooth, in particular if f : M →
N is smooth and X is a submanifold of M , then the restriction of f to X is smooth.
Example II.3. Let M be a smooth E-manifold and (U, ψ) a chart, then ψ : U → E is
smooth; indeed: idE ◦ ψ ◦ ψ −1 is smooth.
We would now like to extend the notion of differential between functions on
Banach spaces to manifolds. One could suggest, as above, that we take local charts
and differentiate; but this would not be a chart invariant definition. Indeed let f :
M → N be a smooth map, choose, near p ∈ M , charts: (U1 , φ1 ), (U2 , φ2 ) such that
p ∈ U1 ∩ U2 and charts (V1 , ψ1 ), (V2 , ψ2 ) such that f (p) ∈ V1 ∩ V2 . Let us study how
the local representations of the function change:
(ψ2 ◦f ◦φ−1 ′ −1 ′ −1 −1 ′ −1 −1 ′
2 ) (x) = (ψ2 ◦ ψ1 ) (ψ1 ◦ f ◦ φ2 (x)) ◦(ψ1 ◦f ◦φ1 ) (φ1 ◦φ2 (x))◦(φ1 ◦ φ2 ) (x) .
| {z } | {z }
Q−1 P

11
We can see from this formula that the local chart derivative transforms under a
change of chart like a linear map when we change basis. Inspired by this analogy,
we introduce the following equivalence relation. Let p ∈ M be fixed and consider
triples (U, φ, v), (U, φ) is a chart and v ∈ E is a vector in the model space. We define
an equivalence relation on (U, φ, v) by:

(U, φ, v) ∼ (V, ψ, w) ⇔ w = (ψ ◦ φ−1 )′ (p) · v

Definition II.2: Tangent vectors and tangent space at a point p ∈ M

Let p ∈ M , a tangent vector at p is an equivalence class for the relation ∼.

We denote by Tp M the set of all tangent vectors and refer to it as the tangent
space at p; it inherits from a chart a Banachisable topological vector space
structure that does not depend on the choice of chart.

Remark II.1. Let us make the statement of Definition II.2 more explicit. Fix p ∈ M
and let [(φ, U, v)]p denote the tangent vector of which (φ, U, v) is a representative.
We define a vector space structure on Tp M by:

[(φ, U, v)]p + [(φ, U, w)]p = [(φ, U, v + w)]p ,

where (φ, U ) is an arbitrary local chart such that p ∈ U . The definition makes
sense since if (ψ, W ) is another chart with p ∈ W , then: [(φ, U, v)]p = [ψ, V, (ψ ◦
φ−1 )′ (φ(p)) · v]p for any v ∈ E, linearity of the derivative guarantees that:

[(φ, U, v + w)]p = [(ψ, V, (ψ ◦ φ−1 )′ (φ(p)) · v + (ψ ◦ φ−1 )′ (φ(p))(v) · w))]p .

Given a choice of chart (φ, U ) we have a (non-canonical) bijection between Tp M

and E, given by the maps:

φ∗p : [(φ, U, v)]p ∈ Tp M 7→ v ∈ E, (φ∗p )−1 : v ∈ E 7→ [(φ, U, v)]p ∈ Tp M.

This bijection depends on the choice of chart (φ, U ).

To get a topology on Tp M , we pull back the topology of E through such a bijec-
tion, i.e. we endow Tp M with the topology that makes the above bijection a linear
homeomorphism; this topology is uniquely defined and has the basis:

B = {(φ∗p )−1 (U ), U open in E }.

It may seem that the topology could depend on a choice of chart, but this is not the
case as by definition:

ψ∗p ◦ (φ∗p )−1 = (ψ ◦ φ−1 )′ (φ(p)),

12
which is a continuous linear isomorphism E → E, hence the topology is independent
of the choice of chart. Note that, whilst one can use the bijection φ∗p to define a
norm on Tp M , this norm depends on the choice of chart, which is why we only get a
Banachisable topological vector space structure.

Definition II.3: The tangent map at a point p ∈ M

Let f : M → N be a smooth map, and p ∈ M . Suppose that M is an E1

manifold and N and E2 manifold. The tangent map at p, written f∗p or Tp f
is the unique continuous linear map defined by the following commutative
diagram:
f∗p
Tp M Tf (p) N
ϕ∗p ψ∗f (p)
(ψ◦f ◦ϕ−1 )′ (ϕ(p))
E1 E2

where (U, φ) is a chart such that p ∈ U and (V, ψ) is a chart such that f (p) ∈
V . In the diagram we denote (abusively) by φ∗p : Tp M → E1 the canonical
continuous linear isomorphism [(U, φ, v)] → v and similarly for ψ∗f (p) .

Remark II.2. When E is a Banach space, it is clear that for any p ∈ E, there is
a canonical identification Tp E = E in the global chart given by the identity; we
will systematically make this identification. This also justifies the notation φ∗p in
Definition II.3, as one can easily verify that this is indeed the tangent map of the
chart φ at p !
The tangent map behaves like the usual differential, in particular:

Proposition II.1: Chain rule

Let f : M → N , g : N → S be smooth maps between smooth manifolds,

then:
(g ◦ f )∗p = g∗f (p) ◦ f∗p .

Proof. Introduce local charts (U, φ), (V, ψ), (W, η) such that: f (U ) ⊂ V , g(V ) ⊂ W ,
and fix p ∈ U . By definition:

η∗g(f (p)) ◦ (g ◦ f )∗p = (η ◦ g ◦ f ◦ φ−1 )(φ(p))′ ◦ φ∗p .

13
Now:
(η ◦ g ◦ f ◦ φ−1 )′ (φ(p)) = (η ◦ g ◦ ψ −1 ◦ ψ ◦ f ◦ φ−1 )′ (φ(p)),
= (η ◦ g ◦ ψ −1 )′ (ψ(f (p))) ◦ (ψ ◦ f ◦ φ−1 )′ (φ(p)),
= (η ◦ g ◦ ψ −1 ◦ ψ ◦ f ◦ φ−1 )′ (φ(p)),
= (η ◦ g ◦ ψ −1 )′ (ψ(f (p))) ◦ ψ∗p ◦ (ψ∗p )−1 ◦ (ψ ◦ f ◦ φ−1 )′ (φ(p))
= η∗g(f (p)) g∗f (p) ◦ f∗p ◦ (φ∗p )−1 .

It follows that:
η∗g(f (p)) ◦ (g ◦ f )∗p = η∗g(f (p)) g∗f (p) ◦ f∗p .
The result follows.

Corollary II.1. For any local chart (φ, U ):

(φ∗p )−1 = φ−1

∗ϕ(p) .

II.B. Derivations of the algebra of germs of smooth functions

With their current definition, tangent vectors are not all that easy to manipulate.
However, they have a very strong link with derivations on the algebra of germs of
smooth functions on M that we shall explore in this section. First we define:

Definition II.4: Derivations of an algebra

Let A be a real algebra, M an A-bimodule then a derivation on A is an R-linear

map D : A → M such that:

∀a, b ∈ A, D(ab) = (Da) · b + a · (Db).

Definition II.5: Germ of smooth function

Let p ∈ M , two smooth functions f, g on M are said to define the same germ
at p, if there is an open neighbourhood U of p, such that f |U = g|U . This
defines an equivalence class on smooth functions on M that we call the germ
of f at p, written fp . Let Cp∞ (M ) denote the set of all germs at p, pointwise
multiplication and addition gives Cp∞ (M ) the structure of an R-algebra.

Example II.4. R is a Cp∞ (M )-bimodule by defining fp λ = λfp := f (p)λ; λ ∈ R, fp ∈

Cp∞ (M ). A derivation D : Cp∞ (M ) −→ R is an R-linear map that satisfies the
Leibniz rule:
D(fp gp ) = (Dfp )g(p) + f (p)Dgp .

14
In the above formula, f (p) and g(p) denote the value at p of any representative f
(resp. g ) of the germ fp (resp. gp ); by definition it does not depend on this choice.
Note that if fp is the germ of a constant map, then Dfp = 0. Indeed, suppose
fp is the germ of a constant map equal to λ, then for any gp , fp gp = λgp so that:
λDgp = D(λgp ) = D(fp gp ) = D(fp )g(p) + λDgp . Thus D(fp )g(p) = 0 for any
function g , this implies D(fp ) = 0.
The important result is:

Proposition II.2

Let p ∈ M , vp ∈ Tp M , f : M → R a smooth function. Define Dvp f = f∗p (vp ),

then Dvp f only depends on the germ of f at p and defines a derivation Dvp at
p. The correspondance vp 7→ Dvp is injective; if M is finite-dimensional it is a
bijection (any derivation is of the form Dvp for some vp ∈ Tp M ).

Proof. Choose a chart (U, φ) near p and note that Dvp f = (f ◦ φ−1 )′ (φ(p))(φ∗p vp )
This formula shows that Dvp f only depends on the germ of f at p so factors to a
map on Cp∞ (M ). Using the product rule,

Dvp (f g) = ((f g) ◦ φ−1 )′ (φ(p))(φ∗p vp )

= (f ◦ φ−1 · g ◦ φ−1 )′ (φ(p))(φ∗p vp )


= (f ◦ φ−1 )′ (φ(p)) · φ∗p vp (g ◦ φ−1 )(φ(p)) + ((g ◦ φ−1 )′ (φ(p)) · φ∗p vp )(f ◦ φ−1 )(φ(p))
= (Dvp f )g(p) + f (p)Dvp g,

so that this map is a derivation. For the injectivity of the correspondence, suppose
that for some vp ∈ Tp M , Dvp f = 0 for any f . Let l ∈ E′ and consider f = l ◦ φ.
Then for any l ∈ E′ , l(φ∗p vp ) = 0 so φ∗p vp = 0 ⇒ vp = 0.
Now assume that M is of dimension n. We shall show that all derivations are of
the form Dvp for some vp ∈ Tp M . Let D be a derivation and (U, φ) be a local chart
near p. Let fp ∈ Cp∞ (M ) and f a representative of the germ, using a bump function,
it can be chosen such that suppf ⊂ U . For q in the support of f one can write:
Z 1
f (q) = f (p) + (f ◦ φ−1 )′ (φ(p) + t(φ(q) − φ(p)) · (φ(q) − φ(p))dt.
0

Decomposing φ(q) − φ(p) onto the canonical basis of Rn , we obtain:

X
n Z 1
f (q) = f (p) + e∗i (φ(q) − φ(p)) (f ◦ φ−1 )′ (φ(p) + t(φ(q) − φ(p)) · ei dt.
i=1 0

15
Applying D in this formula we obtain, due to the Leibniz rule and the fact that it
annihilates germs of constants:

X
n X
n
Dfp = D((e∗i ◦ φ)p )(f ◦ φ−1 )′ (φ(p))(ei ) = D((e∗i ◦ φ)p )f∗p · φ∗p (ei )
i=1 i=1
= f∗p vp ,
Pn
where: vp = φ∗p ( i=1 D((e∗i ◦ φ)p )ei ) .
This justifies:

Definition II.6: Tangent spaces on finite dimensional manifolds

Let M be a finite dimensional manifold, then the tangent space at p, Tp M , is

defined to be the set of derivations of germs of smooth functions at p.

We can now revisit the definition of the tangent map, thinking of tangent vec-
tors as derivations:

Definition II.7: Tangent map at p

Let M, N be a finite dimensional manifolds, φ : M → N a smooth map, then

we define the tangent map at p, denoted by φ∗p by:

φ∗p (X) (fϕ(p) ) = X ((f ◦ φ)p ),

|{z} X ∈ Tp M
| {z }
A derivation at ϕ(p) A derivation at p

∞
where fϕ(p) ∈ Cϕ(p) (N ) and (f ◦ φ)p denotes the germ at p of f ◦ φ, where f is
any representative of the germ.

Remark II.3. If φ : M → N , and f : N → R then the germ of f ◦ φ at p only depends

on the germ of f at φ(p). Indeed let f and g coincide on an open neighbourhood U
of φ(p), then by continuity of f , there is an open neighbourhood V of p such that
f (V ) ⊂ U , then g ◦ φ = f ◦ φ on V .
Lastly let us introduce a common notation:

16
 Definition II.8
Let M be a finite dimensional manifold, p ∈ M , (U, x) a local chart we define:
 
∂
= x−1
∗x(p) ei .
∂xi p
 
a
Any tangent vector at p can be written : X = X i ∂
∂xi
. Xi ∈ R
p

a
See the Einstein’s summation convention

To conclude this section we shall check that Definition II.7 and Definition II.3,
coincide on finite dimensional manifolds. Let φ : M → N be a smooth map between
two finite dimensional manifolds dim M = n, dim N = m, choose p ∈ M , (x, U ) a
chart at p and (y, V ) a chart on N such that f (U ) ⊂ V .
First, let us make explicit II.8. In Rn the vector ei is interpreted as the partial
derivative operator ∂ei , so, using Definition II.7, x−1 ∗x(p) ei should be interpreted as
the derivation defined by:
 
∂
(f ) = ∂ei (f ◦ x−1 )(x(p)).
∂xi p
f is the germ of some smooth function at p and, the right hand side is the ith partial
derivative (in the usual sense) of the function f ◦x−1 , defined
 on the open set x(U ) ⊂
R . Now, let us calculate φ∗p Xp , where Xp = X ∂xi again using Definition II.7,
n i ∂
p

X
n  
∂
(φ∗p Xp )(f ) = Xp (f ◦ φ) = Xi (f ◦ φ)
i=1
∂xi p
X
n
= X i ∂ei (f ◦ φ ◦ x−1 )(x(p))
i=1
X
n  
∂
= X i
(f ◦ φ)
i=1
∂xi p
Xn
= X i ∂ei (f ◦ y −1 ◦ (y ◦ φ ◦ x−1 ))(x(p))
i=1
X
n X
m
using the chain rule = X i ∂ei (y ◦ φ ◦ x−1 )j (x(p))∂ej (f ◦ y −1 )(y(φ(p)))
i=1 j=1
! 
X
m X
n
∂
−1
= X ∂ei (y ◦ f ◦ x )j (x(p))
i
(f ).
j=1 i=1
∂yj ϕ(p)

17
Where we have introduced the notation (y ◦f ◦x−1 )j to denote the j -th component
in the canonical basis of y ◦ f ◦ x−1 . Now we compare with Definition II.3, recall
that one must have:
−1
(y ◦ φ ◦ x−1 )′ (x(p)) ◦ x∗p = y∗ϕ(p) ◦ φ∗p ⇔ φ∗p = y∗y(ϕ(p)) ◦ (y ◦ φ ◦ x−1 )′ (x(p)) ◦ x∗p

Decomposing onto the canonical basis of Rm :

X
m
−1 ′
(y ◦ φ ◦ x ) (x(p)) = ej (y ◦ φ ◦ x−1 )′j (x(p))
j=1

Moreover:   !
X
n
∂ X
X= Xi = x−1
∗x(p) X i ei ,
i=1
∂xi p i=1

hence:
X
m X
n
−1 ′
(y ◦ φ ◦ x ) (x(p)) ◦ x∗p (X) = ej X i (y ◦ φ ◦ x−1 )′j (x(p)) · ei ,
j=1 i=1
X
m X
n
= ej X i ∂ei (y ◦ φ ◦ x−1 )j (x(p)).
j=1 i=1
 
−1
Since y∗y(ϕ(p)) ej = ∂
∂yj
, it follows that the two definitions are identical.
ϕ(p)
One can also use Definition II.7 to give a much more elegant (and coordinate free)
proof of the chain rule in finite dimensions.
Alternative proof of the chain rule in finite dimensions. Let φ : M → N , ψ : N → S , let
f be the germ of a smooth function at (ψ ◦ φ)(p) for fixed p ∈ M . Let Xp ∈ Tp M ,
thought of as a derivation at p, then:

(φ ◦ ψ)∗p (Xp )(f ) = Xp ((f ◦ φ) ◦ ψ),

= (ψ∗p (Xp ))(f ◦ φ)
= (φ∗ψ(p) ◦ ψ∗p )(Xp )(f ).

II.C. The tangent bundle

We would now like generalise the notion of vector fields to manifolds and make sense
of their smoothness. For this we are going to construct out of M a new manifold
T M , known as the tangent bundle. It is an object obtained by gluing together all of

18
the information contained in the tangent spaces at different points. It is the first
example that we will encounter of a fibre bundle. Furthermore, here the fibres will
also have a topological vector space structure, and hence it is an instance of a vector
bundle. We give the general definition:

Definition II.9: Vector bundles

Let E, M be smooth manifolds, E a Banach space, and π : E → M a smooth

map; suppose that there is a covering {Ui } of M with open neighbourhoods
and a family of smooth maps τi : π −1 (Ui ) → Ui × E, known as bundle or
trivialisation charts, such that:

1. π commutes with the projection on Ui , i.e. the following diagram com-

τi
π −1 (Ui ) Ui × E
mutes: π
π Ui
Ui

2. On each fibre, i.e. π −1 ({p}) ≡ Ep , p ∈ M , there is a topological vector

space structure and for each p ∈ U , the restriction: τi,p : π −1 ({p}) → E
is a toplineara isomorphism.

3. For every (Ui , τi ), (Uj , τj ) with Ui ∩ Uj 6= ∅, the transition map: p ∈

−1
Ui ∩ Uj 7→ (τj,p ◦ τi,p ) ∈ L(E) is smooth.

The family B = {(Ui , τi )} is said to be a vector bundle atlas. In the exact same
way as for atlases on manifolds, a vector bundle atlas determines a unique
maximal vector bundle atlas that we call a vector bundle structure on π (or
E ). We say that E is the total space and M the base manifold.
a
Fancy way of saying continuous linear.

Remark II.4. If M and E are finite-dimensional manifolds and E a finite dimensional

vector space then the third condition is implied by the others.
−1
The transition maps τji,p = τj,p ◦ τi,p satisfy a cocycle condition:

τkj,p ◦ τji,p = τki,p .

They contain in fact contain all the essential information of the bundle: given a
covering {Ui } and a family of transition maps {τji,p } satisfying the cocycle condition,
it is possible to reconstruct E and π .
Let us now describe how to construct T M , suppose that M`is an E-manifold.
Consider first the coproduct of sets (i.e. disjoint union) T M = p∈M Tp M . Specify

19
π : T M → M on each Tp M to be the constant map vp ∈ Tp M 7→ p.4 For every
chart (U, φ) consider a map: φ̂ : π −1 (U ) → φ(U ) × E defined by5 :

vp ∈ Tp M ,→ T M 7→ (φ(p), φ∗p vp ).

Now, equip T M with the coarsest topology6 that makes π and all the φ̂ continuous;
with this topology the φ̂ are homeomorphisms. The transition map between any
two charts is:

(ψ̂ ◦ φ̂−1 )(p, v) 7→ ((ψ ◦ φ−1 )(p), (ψ ◦ φ−1 )′ (p)v),

and hence is smooth. The charts {(π −1 (U ), φ̂), (U, φ) chart on M } therefore deter-
mine a smooth structure on T M . Define now: φ̃ : π −1 (U ) → U × E by φ(vp ) =
(p, φ∗p vp ), vp ∈ Tp M. and observe that these are smooth bundle charts; the transi-
tion maps are easily seen to be given by: p 7→ (ψ ◦ φ−1 )′ (p) and are smooth, which
shows that π is a vector bundle.
Remark II.5. Although T M is a local product, in general it is not globally diffeo-
morphic to a product of manifolds; when this is possible the manifold is said to be
parallelisable.
We quote a few examples that we can describe explicitly:
Example II.5. If U is an open subset of a Banach space E, T U = U × E.
Example II.6. Let S n ⊂ Rn+1 be the Euclidean sphere, then:

T S n = {(x, v) ∈ Rn+1 × Rn+1 , ||x||2 = 1, hx, vi = 0}.

Proposition II.3

If φ : M → N , one can define a φ∗ : T M → T N by setting for any p ∈

M, vp ∈ Tp M ,→ T M , φ∗ vp = φ∗p vp ∈ Tϕ(p) N ,→ T N . The construction of
the topologies and smooth structures of T M and T N makes this map smooth;
we will refer to it as the tangent map of φ.

4
We apply here the universal property of coproducts.
5
As
` before it is sufficient to define the function on each Tp M to define a function on T M =
p∈M Tp M
6
Let (Yi )i∈I be a family of topological spaces and fi : X → Yi a family of maps. The initial
topology of (fi )i∈I is the finest topology on X such that all the maps are continuous. A basis for
this topology are sets of the form ∩j∈J fj−1 (Uj ) where J ⊂ I is finite and Uj ⊂ Yj is open.

20
II.D. Operations on vector bundles
Certain functors of subcategories of Banach spaces extend to functors of vector
bundles, to keep this discussion brief we refer the reader to [Lan95, Chapter III,
§4], we shall simply quote a few important examples.
Example II.7. Consider first the contravariant functor that sends a Banach space E
to its dual7 E∨ ; recall that the map between morphisms is given by: u 7→ t u. There
is a corresponding functor of vector bundles of base M , π : E → M , that yields a
new vector bundle E ∨ whose fibres are the dual spaces of those of E . Applied to
the tangent bundle, this gives us the cotangent bundle.
Example II.8. If E1 and E2 are two vector bundles with the same base M one can
form their direct sum8 , E1 ⊕ E2 which is a vector bundle with base M and whose
fibre at a point x ∈ M is E1x ⊕ E2x .
Example II.9. In finite dimensions, one can also form the tensor product of vector
bundles.

II.E. Vector fields

We can now define:

Definition II.10: Vector fields

A vector field is a section of the tangent bundle π : T M → M , i.e. a map
X : M → T M such that π ◦ X = IdM .

In a local chart (U, φ) this is a map from U to U × E of the form: p 7→ (p, XU (p)).
The map XU is called the local representative of the vector field, X is smooth if and
only for every p ∈ M , one can find a chart (U, φ), such that that XU is smooth. In
finite dimensions, this translates to the fact that: X = X i ∂x∂ i is smooth if and only if
the functions (X i ) are smooth. Note that under a change of chart (U, φ) → (V, ψ),
one has: XV = (ψ◦φ−1 )′ ◦φ·XU . Conversely, if for an atlas on M one is given a family
of smooth maps, indexed by the charts, that verify this compatibility condition then
one can construct a unique vector field locally described by the family.
If X is a vector field to avoid too many parentheses it is customary to write
X(p) = Xp . We denote by Γ(T M ) the set of smooth vector fields on M ; it is a
C ∞ (M )-module; there is an injective map into the set of derivations of C ∞ (M ). A
remarkable fact is that it is also a Lie algebra.

7 |l(x)|
Banach space of continuous linear forms on E, equipped with the norm klk = supx̸=0 ||x||
8
also known as the Whitney sum

21
 Theorem II.1
Let X, Y be two vector fields on M , ∂X , ∂Y the corresponding derivations,
then there is a unique vector field [X, Y ] such that: ∂X ∂Y − ∂Y ∂X = ∂[X,Y ] .
In a local chart (U, φ), the local representative is:

[X, Y ]U = YU ∗ (φ−1 −1
∗ · XU ) − XU ∗ (φ∗ · YU ),
= (YU ◦ φ−1 )′ ◦ φ · XU − (XU ◦ φ−1 )′ ◦ φ · YU .

Proof. Uniqueness follows immediately from injectivity of the map: X 7→ ∂X . Let

(U, φ) be a chart and let XU , YU be the local representatives of X, Y in the chart. Let
f be a smooth function, then, for p ∈ U : (∂X f )(p) = (f ◦ φ−1 )′ (φ(p)) · XU , hence
on U :
(∂X ∂Y − ∂Y ∂X )(f ) =((f ◦ φ−1 )′ · YU ◦ φ−1 )′ ◦ φ · XU − ((f ◦ φ−1 )′ · XU ◦ φ−1 )′ ◦ φ · YU


=(f ◦ φ)′′ ◦ φ · (XU , YU ) + (f ◦ φ−1 )′ ◦ φ · (YU ◦ φ−1 )′ ◦ φ · XU


− (f ◦ φ)′′ ◦ φ · (YU , XU ) − (f ◦ φ−1 )′ ◦ φ · (XU ◦ φ−1 )′ ◦ φ · YU


=(f ◦ φ−1 ) ◦ φ · YU ∗ (φ−1
∗ · X U ) − X U ∗ (φ −1
∗ · Y U ).

Where in the final step we use the symmetry of the second derivative. We have
therefore shown that, on U , ∂X ∂Y − ∂Y ∂X is a derivation associated with a vector
field over U locally represented by:

[X, Y ]U = YU ∗ (φ−1 −1
∗ · XU ) − XU ∗ (φ∗ · YU ).

It remains to ensure ourselves of the fact that this can be glued together to from
a vector field on M . For this, is sufficient to check that it transforms correctly
under a change of chart. Recall that if (V, ψ) is another chart then on U ∩ V , XV =
(ψ ◦ φ−1 )′ ◦ φ · XU , to simplify notation let us write: h = (ψ ◦ φ−1 )′ ◦ φ now:

[X, Y ]V = (h · YU )∗ (ψ∗−1 h · XU ) − (h · XU )∗ (ψ∗−1 h · YU ).

Now:
ψ∗−1 h · XU = ψ∗−1 (ψ∗ ◦ φ−1 −1
∗ ) · XU = φ∗ X U ,

and:
(h · XU )∗ (φ−1 −1 ′ −1
∗ YU ) = (ψ ◦ φ ) (XU , YU ) + h · (XU )∗ (φ∗ YU ).

Again, using the symmetry of the second derivative we arrive at:

[X, Y ]V = h · [X, Y ]U .

Therefore, it is clear that this patches together to define a smooth vector field on
M , that we denote by [X, Y ]U .

22
Remark II.6. If M is of finite dimension n and (U, x) is a local chart, the local formula
becomes on U :
Xn X n  
∂Y i j ∂X i ∂
[X, Y ] = X − ·Y j
,
i=1 j=1
∂xj ∂xj ∂xi

where X = X i ∂x∂ i , Y = Y i ∂x∂ i and X i , Y j : U → R.

Example II.10. In R2 , X = x ∂x
∂
− y ∂y
∂
,Y = ∂
∂y
:

∂
[X, Y ] = .
∂y

Proposition II.4

The bracket [·, ·] of vector fields is a Lie bracket, i.e.

• It is bilinear, antisymmetric.

• For any vector fields, X, Y, Z :

[X, [Y, Z]] + [Y, [Z, X]] + [Z, [X, Y ]] = 0

Proof. It is sufficient to prove this on an open subset U of E. Identifying the vector

field with its local representative on U , i.e. a map X : U → E we have: [X, Y ] =
X ′ · Y − Y ′ · X. It is now a simple computation:

[X, [Y, Z]] = X ′ · (Y ′ · Z − Z ′ · Y ) − Y ′′ · (Z, X) − Y ′ · (Z ′ · X) + Z ′′ · (Y, X) + Z ′ · (Y ′ · X)

[Y, [Z, X]] = Y ′ · (Z ′ · X − X ′ · Z) − Z ′′ · (X, Y ) − Z ′ · (X ′ · Y ) + X ′′ · (Z, Y ) + X ′ · (Z ′ · Y )
[Z, [X, Y ]] = Z ′ · (X ′ · Y − Y ′ · X) − X ′′ · (Y, Z) − X ′ · (Y ′ · Z) + Y ′′ · (X, Z) + Y ′ · (X ′ · Z)

Summation of the three expressions shows the result, using once more symmetry of
the second derivative.
It also worth noting, for some calculations, that we have the following Leibniz
rule for [X, Y ]:
[X, f Y ] = (∂X f )Y + f [X, Y ].
Proof. Exercise.

23
II.F. Flow and bracket
Let γ : I → M be a curve on M defined on an open interval I . Note that T I = I ×R
and we have a canonical section j : t 7→ (t, 1). The canonical lift of γ is the map
γ̇ = γ∗ ◦ j , it is a curve on T M that satisfies π ◦ γ̇ = γ , where π : T M → M is the
projection of T M onto M .
Remark II.7. For each t ∈ I , γ̇(t) has the usual interpretation as the tangent (or
velocity) vector to the curve at the point γ(t). It is sometimes written:

d
γ̇(t0 ) = γ(t) .
dt t=t0

We can reinterpret tangent vectors at p as velocity vectors to curves passing through

p. Indeed if γ is a curve such that γ(0) = p then γ̇(0) ∈ Tp M . Conversely, if
vp ∈ Tp M then there is a curve such that γ(0) = p, γ̇(0) = vp . To construct it let
(U, φ) be a local chart such that p ∈ U , for small enough ε > 0, φ(p)+tφ∗p (vp ) ∈ φ(U )
for all t ∈ (−ε, ε), consider the curve: γ : (−ε, ) → M defined by:

γ(t) = φ−1 (φ(p) + tφ∗p vp ).

Clearly, γ(0) = 0 and:

γ̇(0) = φ−1
∗ϕ(p) (φ∗p vp ) = vp .

If you are not entirely convinced define: α : (−ε, ε) → φ(U ) ⊂ E by:

α(t) = φ(p) + tφ∗p vp .

Now the tangent map is given explicitly:

α∗ : T (−ε, ε) ∼
= (−ε, ε) × R → T φ(U ) ∼
= φ(U ) × E,

with: α∗ (t, λ) = (α(t), α′ (t)·λ), here: α′ (t) ∈ L(R, E) is the usual derivative, indeed,
it is naturally identified with α∗t : Tt (−ε, ε) ∼ = R → Tα(t) E ∼ = E. Now recall that
we have the canonical section j : (−ε, ε) → T (−ε, ε) = (−ε, ε) × R given by:
j(t) = (t, 1). Hence: α̇(t) = (α∗ ◦ j)(t) = (α(t), α′ (t) · 1). By the chain rule:

γ̇(0) = φ−1 −1
∗α(0) α̇(0) = φ∗ϕ(p) (φ∗p · vp ).

To understand the last inequality it is important to note that although in all rigour:
φ−1 ∼
∗ϕ(p) : Tϕ(p) φ(U ) = {φ(p)} × E → Tp M we systematically identify {φ(p)} × E with
E and write:
φ−1 −1
∗ϕ(p) (φ(p), v) ≡ φ∗ϕ(p) v.

24
 It can sometimes be useful to note that if f : M → N and vp ∈ Tp M then for any
curve γ : I → M such that γ(0) = p, γ̇(0) = vp we have:

˙
f∗p vp = (f ◦ γ)(0), (II.1)

this is just another application of the chain rule.

We now have enough tools to consider first order ordinary differential equations
on manifolds. In particular, if X is a smooth vector field on T M , we can consider
the Cauchy problem: (
γ̇ = X ◦ γ,
γ(0) = p ∈ M.
In a local chart (U, φ), recall that T U = U × E, writing: φ ◦ γ = γU : J → U , where
J ⊂ I such that γ(J) ⊂ U and γ some curve defined an open interval containing
0 then: φ∗ γ̇ = (γU , γU′ ). Hence if φ∗ X = (idU , XU ), the equation is equivalent in a
local chart to:
γU′ (t) = XU (γU (t)), t ∈ J.
Apply the classical theory of ordinary differential equations in local charts, we have
existence and uniqueness of maximal solutions to such problems; we will write the
unique maximal solution as t 7→ φX t (p); solutions are referred to as integral curves of
X.

Definition II.11: Flows

• A subset D ⊂ R × M such that for any p ∈ M , D (p) = {t ∈ R, (t, p) ∈
D} is an open interval containing 0 is known as a flow domain.

• A (local) flow is a map θ : D → M where D is a flow domain such that

for any t ∈ D (p) , s ∈ D (θ(t,p)) :

θ(s, θ(t, p)) = θ(s + t, p).

With a bit of elbow grease one can show the following fundamental theorem of
flows:

25
 Theorem II.2: Fundamental theorem of flows
Let X be a smooth vector field on a smooth manifold M , then there is a unique
maximal flow domain D such that:

1. (t, p) ∈ D 7→ φX
t (p) is a local flow,

2. For each p ∈ M , D (p) is the interval of definition of the maximal solu-

tion t 7→ φX
t (p),

3. If s ∈ D (p) then D (ϕs (p)) = D (p) − s,

X

4. For each t ∈ R the set Mt = {p ∈ M : (t, p) ∈ D} is open in M and

p 7→ φX
t (p) is a diffeomorphism of Mt onto M−t with inverse φ−t (p).
X

We refer to [Lee03] for the proof. This gives another formula for the Lie bracket
that is sometimes useful:

Proposition II.5: Lie bracket and flow

Let X, Y be two smooth vector fields on a smooth manifold then:

d  X 
[X, Y ](p) = φ−t ∗ϕX (p) · Y (φX (p)) .
dt t
t
t=0

This shows that [X, Y ] measures the infinitesimal change in Y under the effect of
the flow of X .
Proof. Exercise...

II.G. Appendix: Product manifolds

Let (M, Â) be a smooth E-manifold, (N, B̂) a smooth F-manifold. M × N can be
given naturally the structure of a smooth (E × F)-manifold. Equip M × N with its
product topology and denote by πM (resp. πN ) the projection onto the first (resp.
second) factor. Define an atlas on M by:

C = {(U × V, φ × ψ), (U, φ) ∈ Â, (V, ψ) ∈ B̂}.

where φ × ψ : (p, q) → (φ(p), ψ(q)). It is straightforward to check that φ × ψ is a

homeomorphism of U ×V onto φ(U )×ψ(V ). Note now that (φ×ψ)−1 = φ−1 ×ψ −1 ,
so:
−1
(φ × ψ) ◦ (φ′ × ψ ′ )−1 = (φ ◦ φ′−1 ) × (ψ ◦ ψ ′ ).

26
It is a smooth map; this follows from the fact that if E × F is equipped with its norm
||(x, y)||E×F = ||x|| + ||y||, then f : U ⊂ G → E × F is differentiable if and only if
π1 ◦ f and π2 ◦ f are differentiable. Equipping M × N with the maximal atlas Cˆ, this
property generalises to manifolds.

Proposition II.6

Let (M × N ) be equipped with the maximal atlas Cˆ, then:

1. the projections π1 , π2 are smooth maps,

2. a map f : L → M × N is smooth if and only if π1 ◦ f and π2 ◦ f are

smooth.

Let (p, q) ∈ M × N ; then, as one might expect, we have the toplinear isomor-
phism:
T(p,q) (M × N ) ∼
= Tp M × Tq N.
This follows directly from Definition II.2.

II.H. Appendix: Vector bundle morphisms

Definition II.12: Vector bundle morphism

Let (E, π, M ), (E ′ , π ′ , M ) two vector bundles with same base M , a vector bun-
dle morphism is a smooth map: f : E → E ′ such that

• π′ ◦ f = π,

• for each p ∈ M , the induced map between fibres: fp : π −1 (p) → π ′−1 (p)
is continuous linear

• there are bundle charts τ : π −1 (U ) → U × E, τ ′ : π ′−1 (U ) → U × E′ such

that the map from U → L(E, E′ ) determined by p 7→ πE′ ◦τ ′ ◦f ◦τ −1 ◦jp
is smooth; here, jp : v ∈ E 7→ (p, v).

Remark II.8. • The final condition is superfluous for finite dimensional mani-
folds.

• The map T : M 7→ T M is in fact a functor between smooth manifolds and

smooth vector bundles; the map between morphisms is given by the tangent
map.

27
III. Geometry on manifolds, connections

From now on, for simplicity, all manifolds will be assumed finite dimensional.

Roughly, modern differential geometry can be thought of as the study of connec-

tions on manifolds. The point of view I adopt in these notes is due to E. Cartan and
it generalises Klein’s Erlengan program to a curved setting. Klein introduced the
idea that geometry should be thought of as the study of the action of a group G on
a space X . Restricting our attention to any of the orbits of X , we can assume that
the action is transitive and hence X can be thought of as the quotient G/H , where
H is the isotropy subgroup of any point of X . Affine, Euclidean, projective and con-
formal geometry fit nicely into this framework. Our starting point will be to revisit
affine geometry from this point of view and identify the differential structures that
one would like to reproduce on a manifold.

III.A. Lie groups and their Lie algebra

Definition III.1: Lie groups

• A Lie group (G, ·) is a group equipped with the structure of a smooth

finite-dimensional manifold that is compatible with its group opera-
tions, i.e. µ : (g1 , g2 ) 7→ g1 · g2 and ι : g 7→ g −1 are smooth.

• A Lie subgroup is a subgroup that is also a smooth submanifold.

Example III.1. (R∗ , ·), GLn (R), On (R), SLn (R)

A number of important examples follow from:

Definition III.2: The closed subgroup theorem (Cartan-Von Neumann)

Let G be a Lie group and H a closed subgroup, then H is a Lie subgroup of G.

A vector field X is said to be left-invariant, if for any p, g ∈ G

Xpg = Lg ∗p Xp .
Left-invariant vector fields are completely determined by their value at the identity
element; indeed for any g ∈ G: Xg = Lg ∗e Xe . They form a dim G-dimensional
subspace of Γ(T G), and is in fact a Lie subalgebra. To prove this it is custom to use
the notion of φ-related vector fields. Let φ : M → N be a smooth map, X a vector
field on M , and Y a vector field on N . Y is said to φ-related to X , if for any p ∈ M ,
Yϕ(p) = φ∗p Xp .

28
 Proposition III.1

Let X be φ-related to X̃ , Y , φ-related to Ỹ , then [X, Y ] is φ-related to [X̃, Ỹ ].

This is sometimes written:

φ∗ [X, Y ] = [φ∗ X, φ∗ Y ].

Proof. This is not the most elegant proof, but it is valid in infinite dimensions. Let
us work in a local charts, suppose that (U, x) is a chart at p, (V, y) a chart at φ(p)
such that φ(U ) ⊂ V . Let XU and X̃V be the local representations of the fields near
in the chart then they are φ-related means that for any p ∈ U :

X̃V (φ(p)) = (y ◦ φ ◦ x−1 )′ (x(p)) · XU (p).

Now:

[X̃, Ỹ ]V (φ(p)) = (ỸV ◦ y −1 )′ (y(φ(p)) · X̃U (φ(p)) − (X̃V ◦ y −1 )′ (y(φ(p)) · ỸU (φ(p))

By the chain rule:

(ỸV ◦ y −1 )′ (y(φ(p)) · X̃U (φ(p)) = (ỸV ◦ y −1 )′ (y(φ(p)) · (y ◦ φ ◦ x−1 )′ (x(p)) · XU (p),

= (ỸV ◦ φ ◦ x−1 )′ (x(p)) · XU (p).

Furthermore:
((y ◦ φ ◦ x−1 )′ (x(p)) · YU (p))′ · XU (p) =(y ◦ φ ◦ x−1 )′′ (x(p)) · (XU (p), YU (p))
+ (y ◦ φ ◦ x−1 )(x(p)) · YU′ (p) · XU (p).

Treating the second term in exactly the same fashion and using the symmetry of the
second derivative, we find that:

[X̃, Ỹ ]V (φ(p)) = (y ◦ φ ◦ x−1 )′ (x(p)) · [X, Y ]U (p).

Which proves that [X, Y ] and [X̃, Ỹ ] are φ-related.

By definition a vector field is left-invariant if and only if it Lg -related to itself
for all g ; the above result implies directly that the Lie bracket of two left-invariant
vector fields is left-invariant so that g is a Lie subalgebra.

Definition III.3: Lie algebra of a group G

The Lie subalgebra g of Γ(T G) composed of left-invariant vector fields is

known as the Lie algebra of the group G.

29
 Since left-invariant vector fields are uniquely determined by their value at the
identity, we can identify g and Te G as vector spaces, and transfer the Lie algebra
structure to Te G, i.e. by setting:

[Xe , Ye ] = [X, Y ]e ,

where X (resp. Y ) is the unique left-invariant vector field such that X(e) = Xe (resp.
Y (e) = Ye ).

III.B. The exponential map of a Lie group

Let X ∈ g be a left-invariant vector field on a Lie group G, let us consider arbitrary
integral curves of X passing through the identity, i.e. solutions γ to the initial value
problem: (
γ̇ = X ◦ γ,
γ(0) = e.
Call u = Xe , we shall show that γ(t + s) = γ(t)γ(s). Fix s ∈ R, it suffices to show
that the curves t 7→ γ(s + t) and t 7→ γ(s)γ(t) = σ(t) satisfy the same Cauchy
problem. First, they satisfy the same initial condition and:

σ̇(t) = Lγ(s) ∗γ(t) · γ̇(t) = Lγ(s) ∗γ(t) ◦ Lγ(t) ∗e · u = Lσ(t) ∗e · u = Xσ(t) .

By uniqueness of the solution we conclude that for every t, s ∈ R, γ(t + s) =

γ(t)γ(s). It follows that:

Proposition III.2

• The integral curves of a left-invariant vector field are complete, i.e. de-
fined on all of R.

• γ(t)−1 = γ(−t) for all t ∈ R.

Corollary III.1. ι∗e : u ∈ Te G 7→ −u ∈ Te G.

Definition III.4: The exponential map

Let X ∈ g, we define the exponential map: exp : g → G, by

exp(X) = γ(1).

Note that γ(t) = exp(tX) for every t ∈ R.

30
Example III.2. Let G = GLn (R), the tangent space at In is identified with Mn (R).
The exponential map of X ∈ gln (R) = Mn (R) is given by the usual power series:
P
k∈N k! X . This enables us to determine the Lie bracket on Mn (R). Using Propo-
1 k

sition II.5 and Equation (II.1), we have:

 
d d
[X, Y ]e = exp(tX) exp(sY ) exp(−tX) ,
dt ds s,t=0
d
= (exp(tA)B exp(−tA)) ,
dt t=0
= AB − BA.
This englobes many examples as the Lie algebra of a Lie subgroup H can be iden-
tified with a Lie subalgebra h ⊂ g. One can prove this using the inclusion map
i : H ,→ G and noting that left-invariant vector fields on H are i-related to left-
invariant vector fields on G.

III.C. Affine geometry

Recall that the affine group is the semi-direct product:
G = GA(n) = Rn ⋊ GLn (R).
It is useful to see it as a Lie subgroup of GLn+1 (R). For this, identify the affine
plane with the hyperplane An = {xn+1 = 1} in Rn+1 and consider the subgroup G
of GLn+1 (R) that preserves the form (0, . . . , 0, 1). It is composed of matrices of the
form:  
A b
, A ∈ GLn (R), b ∈ Rn .
0 1
     
x A b x
G acts in the plane xn+1 = 1 and if ∈ A , g =
n
then g =
  1 0 1 1
Ax + b
. Hence G can be identified with the affine group; it is a closed subgroup
1
of GLn+1 (R) and therefore a Lie subgroup. We can also determine its Lie algebra
(as a Lie subalgebra of gln+1 (R) = Mn+1 (R)), we have:
  
M b
g= , M ∈ Mn (R), b ∈ R . n
0 0
 
0
G acts transitively on X and the isotropy subgroup of, say, O = is of course:
1
  
A 0
H= , A ∈ GLn (R) ∼ = GLn (R).
0 1

31
It is standard that we have a bijection G/H ∼ = An ; in fact G/H equipped with the
quotient topology is a topological manifold and has a unique smooth structure such
that the canonical projection π : G → G/H is a smooth submersion; it is then the
case that G/H is diffeomorphic to An . We omit this development and instead make
some further observations:
   
x In x
1. π admits a smooth global section ∈ A 7→
n
. The interpretation
1 0 1
of σ is as follows: at each point x ∈ Rn we choose a frame anchored at x and
composed of the n vectors (e1 , . . . , en ) of the canonical basis of Rn .

2. The section σ is equivalent9 to a smooth diffeomorphism: φ : G → An × H

defined by φ(σ(x)h) = (x, h).

3. H acts on the right on G and we have φ(gh) = φ(g)h, where on the rhs10 this
is the canonical right action on the product An × H i.e. (x, h1 )h2 = (x, h1 h2 )
This is our first example of a principle H -bundle that we will define below.
With respect to this structure, G is interpreted as the “frame bundle” of An :
at each point we glue the set of all possible linear frames anchored at x; the
right-action can be thought of as a (global) change of frame.

4. Our final observation is that left-multiplication Lg by an element of g is a

smooth diffeomorphism of G. The tangent map Lg−1 ∗g is a linear isomorphism
from Tg G onto Te G ∼
= g. This gives a smooth map ω : T G → g:

ωg (Xg ) = Lg−1 ∗g (Xg ), g ∈ G, Xg ∈ Tg G ,→ T G.

ω is usually interpreted as a g-valued differential form known as the Maurer-

Cartan form, i.e. a section of the bundle T ∨ G ⊗ (G × g).

To describe the properties of the Maurer-Cartan form we need to recall some alge-
bra.

9
σ(x) = ϕ−1 (x, e)
10
right-hand side

32
III.D. Differential forms and the exterior algebra

Definition III.5: k -forms

Let V be a real vector space a k -form on V is multilinear map α :

| × ·{z
V · · × V} → R that is alternating, i.e. for any σ ∈ Sk , v1 , . . . , vk ∈ V :
k times

α(vσ(1) , . . . , vσ(k) ) = ε(σ)α(v1 , . . . , vk ),

where ε(σ) ∈ {−1, 1} is the signature of σ . The set of k -forms on V is a linear

V
space denoted by k V ∨ .

Example III.3. • 1-forms are elements of the dual.

• 0-forms are constants

Let us define a product on forms:

Definition III.6: Exterior product

Let α and β be respectively a k and l form on some vector space V , we define

their exterior product to be the k + l form given by:

1 X
(α∧β)(v1 , . . . , vk ) = ε(σ)α(vσ(1) , . . . , vσ(k) )β(vσ(k+1) , . . . , vσ(k+l) )
k!l! σ∈S
k+l

Example III.4. Let u and v be one forms, then: (u ∧ v)(x, y) = u(x)v(y) − v(x)u(y).
Let V have finite dimension n ∈ N∗ , and B = (e1 , . . . , en ) a basis of V . Write the
dual basis (e∗1 , . . . , e∗n ), then by induction one can show that:

(e1∗ ∧ · · · ∧ e∗n )(v1 , . . . , vn ) = det(v1 , . . . , vn ).

B

33
 Proposition III.3
L Vk ∨
The multiplication ∧ is bilinear and associative, making R ⊕ ∞ k=1 V an
associative algebra with unit called the exterior algebra. If α and β are respec-
tively a k -form and l-form then:

(α ∧ β) = (−1)kl β ∧ α.
V
If dim V = n then m V ∨ = 0, if m > n and if (e1 , . . . , en ) is a basis for V ,
then:
{ei∗1 ∧ · · · ∧ e∗ik , 1 ≤ i1 < i2 < · · · < ik ≤ n}
V
is a basis of k V ∨

Proof. Any textbook on linear algebra.

Let U be an open subset of Rn a differential k -form on U is a smooth map α :
V
U → k V ∨ , we often write α(x) = αx . Let Ωk (U ) denote the C ∞ (U )-module of
differential forms on U .
Example III.5. • A 0-form is just a smooth function on U .
• If f is smooth f ′ can be viewed as a differential 1-form.
We will encounter the following operations on differential forms in the sequel:

Definition III.7: Pullback of differential forms

Let f : U → V , a smooth map between open subsets of Rk , α a k -form on V ,

we call the pullback of α by f , the differential form on U defined by:

(f ∗ α)x (v1 , . . . , vn ) = αf (x) (f ′ (x)v1 , . . . , f ′ (x)vn ).

Pullback distributes over the wedge product (defined in a pointwise manner),

i.e. f ∗ (α ∧ β) = f ∗ α ∧ f ∗ β .

Proposition III.4: Exterior derivative

There is a unique linear map d defined on differential forms such that:

1. If f is a smooth function df = f ′ .

2. d(α ∧ β) = dα ∧ β + (−1)k α ∧ dβ , if α is a k -form and β an l-form;

3. d2 = 0.

34
Example III.6. Let U be an open set of Rn , (e1 , . . . , en ) the canonical basis. Define,
xi : U → R which maps p to its ith coordinate of p in the canonical basis. Then for
Xn
∂f
any p ∈ U , dxip = e∗i . If f is a smooth function then we can write df = dxi .
i=1
∂x i

Example III.7. Consider on R2 , the one-form α = ydx − xdy , then dα = 2dy ∧ dx

Example III.8. Let f and g be two zero forms, then the wedge product between f
and g is simply the product f g . Point 2 in Proposition III.4 is simply the product
rule: d(f g) = df g + f dg .
Example III.9. Let us consider polar coordinates on R2 defined by x = r cos θ, y =
r sin θ. This are valid, for example, on the open set R2 \ {y = 0, x < 0}. At each
point p = (r cos θ, r sin θ) we have the corresponding basis ( ∂r
∂ ∂
, ∂r ) of Tp M , related
to the canonical basis by:

∂ ∂ ∂ ∂ ∂ ∂
= cos θ + sin θ , = −r sin θ + r cos θ
∂r ∂x ∂y ∂θ ∂x ∂y
The dual basis is
xdx + ydy xdy − ydx
dr = p , dθ =
x2 + y 2 x2 + y 2
V
To extend this to manifolds, one can construct k Tp∨ M at each point p, in the
V ` V
same way as for a vector space. Then, one can equip k T ∨ M = p∈M k Tp∨ M
V
with the structure of a vector bundle where the fibres are given by k Tp∨ M . This is
an example of the more general procedure touched upon in Section II.D.
V
The gory details are as follows. Define π : k T ∨ M → M by αp ∈ Tp∨ M 7→ p.
V
Choose a chart, (x, U ) and define a map x̂ : π −1 (U ) → U × k (Rn )∨ by:
^k
x̂(αp ) = (p, αp (x−1 −1 −1
∗x(p) , x∗x(p) , · · · , x∗x(p) )), αp ∈ Tp∨ M.

These are our natural candidates for our vector bundle charts. Composing with the
chart we get a bijective map:
^k
x̃ : π −1 (U ) → x(U ) × (Rn )∨

that is a natural candidate for a coordinate chart. As with the tangent bundle we
V
equip k T ∨ M with the coarsest topology making π and every x̃ (where x runs
through the charts on M ) continuous. The x̃ are easily seen to be homeomorphisms
with this choice of topology. If (y, V ) is another chart then:

ỹ ◦ x̃−1 : (u, α) 7→ ((y ◦ x−1 )(u), α((x ◦ y −1 )′ (y(p)), . . . , (x ◦ y −1 )′ (y(p))).

35
Since M is a smooth manifold
V these transition charts are also smooth and so define
a smooth structure on k T ∨ M . This is precisely the smooth structure that makes
the bundle charts x̂ smooth diffeomorphisms. It is straightforward to check that
Vk ∨
T M is a vector bundle.
Smooth sections of this bundle are called differential k -forms on M , they form a
C ∞ (M ) module denoted by Ωk (M ).
Example III.10. Let xi = e∗i ◦ x where (x, U ) is a local chart. Define a differential
form on U , dxi , by:
dxip = xi∗p = e∗i ◦ x∗p .
V
Then dxi is a smooth section of 1 T ∨ U ∼ = T ∨ U , indeed, in the chart (x, U ) this
∗
is locally represented by ei . For each p ∈ U , (dx1p , . . . , dxnp ) is by definition the dual
   
basis to the coordinate basis ( ∂x∂ 1 , · · · , ∂x∂n ).
p p
We can inductively define differential k -forms on U , dxi1 ∧ · · · ∧ dxik , 1 ≤ i1 <
V
· · · < ik ≤ n that at each point p give a basis of k Tp∨ M . An arbitrary differential
k -form can be represented locally:
X
α= αi1 ...ik dxi1 ∧ · · · ∧ dxik ,
0≤i1 <···<ik ≤n

where αi1 ...ik are functions on U , the differential form is smooth if and only if they
are smooth for every chart (x, U ).
This is in a nutshell how we work with forms on manifolds in a very basic fashion
and in a way that is designed to ressemble the case of an open set of Rn .
Whilst this is satisfying for computations, to develop the theory it is sometimes
nice to avoid using charts by adopting another point of view. Let α ∈ Ωk (M ), if
X1 , . . . , Xk are vector fields on M then one can define a C ∞ (M )-linear k -alternating
form A : Γ(T M ) × · · · × Γ(T M ) → C ∞ (M ) by:

A(X1 , . . . , Xk )(p) = αp ((X1 )(p), . . . , (Xk )(p)).

On finite dimensional manifolds, it turns out that the two points of view are equiva-
lent. The key to this correspondence is the following lemma:

36
 Lemme III.1
Let T : Γ(T M ) × · · · × Γ(T M ) → C ∞ (M ) be a k -multilinear map on the
C ∞ (M ) module Γ(T M ), then for each p ∈ M , and any X1 , . . . , Xk ∈ Γ(T M ),
the value of T (X1 , . . . , Xk )(p) only depends on the values of the vector fields
X1 , . . . , Xk at the point p.
Consequently, for each p, T induces a k -multilinear map: Tp : Tp M ×
· · · Tp M → R. Furthermore, if T is alternating p 7→ Tp is a smooth section of
Vk ∨
T M.

Proof. It is sufficient to treat the case k = 1 and then use induction. Fix p ∈ M and
choose a chart (x, U ) near p, let b : M → R be a smooth bump function with support
 
in U . Let X ∈ Γ(T M ) and decompose Xp on the coordinate basis as Xp = v i ∂
∂xi
.
p
Set Y = v i b ∂x∂ i . C ∞ (M )-linearity gives:
∂
α(bX − Y ) = (X i − v i )α(b )
∂xi
hence: α(bX − Y )(p) = 0, but:
∂
α(X)(p) = b(p)α(X)(p) = α(bX)(p) = α(Y )(p) = v i α(b( ))(p).
∂xi
Thus showing that α(X)(p) only depends on Xp .
This correspondence is useful for defining tensor objects on manifolds without
referring to a specific choice of chart or working in local coordinates. (Although to
actually calculate things we do have to choose coordinates at some point). For in-
stance, one can extend the exterior derivative to manifolds and we have the invariant
formula:

Proposition III.5: Invariant formula for the exterior derivative on manifolds

Let α be a k -form on M , then its exterior derivative can be defined by the same
conditions as in Proposition III.4. We have the following invariant formula:

X
k
dα(X1 , . . . , Xk+1 ) = (−1)i+1 Xi (α(X1 , . . . , X̃i , . . . , Xk+1 ))
i=1
X
+ (−1)i+j ω([Xi , Xj ], X1 , . . . , X̃i , . . . , X̃j , . . . , Xk+1 )
i<j
(III.1)
where ˜ denotes omission. Recall that vector fields are to be thought of as
derivations.

37
Example III.11. The most important case is when α is a one form then the formula
reads:
dα(X, Y ) = X(α(Y )) − Y (α(X)) − α([X, Y ]).

III.E. Lie algebra valued differential forms and the Maurer-Cartan

equation
Now we know a little bit about differential forms and what we can do with them, we
shall make one further generalisation and consider g-valued differential forms. The
difference is that we get at each point an alternating linear map, αp : Tp M × · · · ×
Tp M → g.
Whilst the notion of pullback by a diffeomorphism extends directly, Definition III.6
no longer makes sense. We can nevertheless recover a notion of exterior product by
using the bracket of the Lie algebra, we define:

1 X
[α, β](v1 , . . . , vk+l ) = ε(σ)[α(vσ(1) , . . . , vσ(k) ), β(vσ(k+1) , . . . , vσ(k+l) )].
k!l! σ∈S
k+l

Example III.12. Let α be a g-valued one form, then [α, α](v1 , v2 ) = 2[α(v1 ), α(v2 )]
The exterior derivative can be extended to g-valued forms by choosing an arbi-
trary basis E i of g and declaring that d acts on components (which are usual k -forms)
like the usual exterior derivative; one should check that this does not depend on the
choice of basis. The invariant formula in Proposition III.5 then extends without
any modification.
We can now summarise the main properties of the Maurer-Cartan form. Recall
that for every p ∈ G, Xp ∈ Tp G,

ωp (Xp ) = Lp−1 ∗p Xp ∈ Te G ∼
= g.

Now H is a closed subgroup of G and H acts on G from the right, let us denote by
Rh : G → G the map p 7→ ph. Calculating the pullback of ω under Rh , we have:

(Rh∗ ω)g = ωgh ◦ Rh∗g = Lh−1 g−1 ∗hg ◦ Rh∗g

(Proposition II.1) = (Lh−1 g−1 ◦ Rh )∗g (III.2)
= (Adh−1 ◦ Lg−1 )∗g = Adh∗e ωg

where he have defined Adh : p 7→ h−1 ph. Its tangent map at the identity element,
Adh∗e – which we will abusively also denote Adh in the sequel – defines a represen-
tation of H with representing space g, called the adjoint representation. We also say
that g is a H -module.

38
 The second property we have is that for any X ∈ h ⊂ g, one can consider a vector
in Tp G for any other p defined by:

d
Xp∗ = p exp(tA) ,
dt t=0

then:
ωp (Xp∗ ) = X. (III.3)
Note that the vector Xp∗ is in ker π∗p where π : G 7→ G/H is the canonical projec-
tion. Tangent vectors in ker π∗p are known as vertical vectors; intuitively they point
in the direction of the fibres of G above G/H . In fact all vertical vectors can be
obtained as fundamental vectors. The vector field p 7→ Xp∗ is called the fundamental
vector field associated to X . Summarising this:

Proposition III.6

Let G be a Lie group, H a closed subgroup and ω the Maurer-Cartan form of

G, then:

1. For each p ∈ G, ωp is a vector space isomorphism.

2. For each h ∈ H , Rh ∗ ω = Adh−1 ω .

3. For each X ∈ h, ω(X ∗ ) = X .

The Maurer-Cartan form satisfies one further equation; later we will interpret this
as the local expression of the flatness of affine space.

Proposition III.7: Maurer-Cartan structure equation

Let G be a Lie group, ω its Maurer-Cartan form, then:

1
dω + [ω, ω] = 0
2

Proof. Let us work with the invariant formulae. First ω is a one form so for any
vector fields X, Y :

dω(X, Y ) = X(ω(Y )) − X(ω(Y )) − ω([X, Y ]),

and
1
[ω, ω] = [ω(X), ω(Y )].
2

39
Now, recall from Lemma III.1 that the value at any given point g ∈ G of the function
dω(X, Y ) + 21 [ω, ω] only depends on the values of X and Y at g . Therefore it is
sufficient to study the case where X and Y are left-invariant vector fields, In this case:
1
dω(X, Y ) = −ω([X, Y ]) = −[Xe , Ye ] = −[ω(X), ω(Y )] = − [ω, ω].
2

III.F. Interpretation of the Maurer-Cartan form by an example

Let us consider 2-dimensional affine space, choose an arbitrary origin O and set up
an affine frame (O, ex , ey ). To see the structure at work, introduce polar coordinates
(r, θ) on U = R2 \{x < 0, y = 0} defined by: x = r cos θ, y = r sin θ, and its “moving
coordinate frame”:
∂ ∂ ∂ ∂ ∂ ∂
= cos θ + sin θ , = −r sin θ + r cos θ .
∂r ∂x ∂y ∂θ ∂x ∂y
Perhaps a more familiar moving frame is given by:
∂ 1 ∂
er = , eθ = .
∂r r ∂θ
At each point p ∈ U , choosing p as origin gives us an affine frame attached to
each point. Expressed differently, we have a section σ of π : G → G/H ∼ = R2
(G = R2 ⋊ GL2 (R)):
 
cos θ(p) − sin θ(p) r(p) cos θ(p)
p 7→  sin θ(p) cos θ(p) r(p) sin θ(p)  .
0 0 1

Remark III.1. Recall that a chart of a 2-dimensional manifold like affine space is
defined as a map from an open set of a manifold onto an open subset of R2 . Equiv-
alently this can be described by two maps: r : U → R and θ : U → R.
Let us now consider the pullback of the Maurer-Cartan form of G onto U . First,
let us determine what the Maurer-Cartan form is in matrix terms. By definition:
ωg = Lg−1 ∗g i.e. it is the derivative at g of the restriction to G of the linear map:
q 7→ g −1 q . It can be written:
ωg = g −1 dg
where dg is to be understood as the derivative at p of the injection g ,→ gln+1 (R).
The pullback along σ is given by:

σ ∗ ω p = σ −1 (p)σ∗p p∈U

40
Since:
 
cos θ(p) sin θ(p) −r(p)
σ −1 (p) = − sin θ(p) cos θ(p) 0 ,
0 0 1
 
− sin θ(p)dθ − cos θ(p)dθ cos θ(p)dr − r(p) sin θ(p)dθ
σ∗p =  cos θ(p)dθ − sin θ(p)dθ sin θ(p)dr + r(p) cos θdθ 
0 0 0

We have:  
0 −dθ dr
σ ∗ ωp = dθ 0 rdθ .
0 0 0
Analysing this result, notice that the last column is the dual basis of (er , eθ ); this can
be interpreted as the infinitesimal displacement vector, expressed in the basis er , eθ
as we move from a point p to p + δp. On the other hand, the matrix:
 
0 −dθ
,
dθ 0

is a matrix describing the infinitesimal change in the basis vectors. Indeed, differ-
entiate in the classical sense the equations defining er and eθ we find:

der = − sin θdθex + cos θdθey = dθeθ ,

deθ = − cos θdθex − sin θdθey = −dθer .

Hence, to first order, it is the change of basis matrix from (er , eθ ) to (er+δr , eθ+δθ ).
In other words the Maurer-Cartan form measures how the moving basis (section)
changes as we move from point to point.

III.G. The Frame Bundle

To extend this structure to manifolds, we need to find a way to replicate this. The
first step is to notice that the structure of π : G 7→ G/H we highlighted in Sec-
tion III.C is a (trivial) example of a principal H -bundle over a manifold:

41
 Definition III.8: Principal bundles

Let π : P → M be a smooth map between smooth manifolds. Suppose that

there is a smooth right-action of a Lie group H on P and that π satisfies the
following properties:

1. ∀r ∈ P, h ∈ H, π(r · h) = π(r),

2. Each point p ∈ M admits an open neighbourhood U , called a trivialising

neighbourhood, and a diffeomorphism φ : π −1 (U ) → U × H such that:
φ(rh) = φ(r) · h, where U × H is equipped with the canonical right-
action: (p, h1 ) · h2 = (p, h1 h2 ), and the following diagram commutes:

ϕ
π −1 (U ) U ×H
π
πU
U

Then π is a principal H -bundle with total space P , base M and structure group
H.

Remark III.2. • The right-action is free, and the orbits are the fibres:
π −1 ({π(p)}) = {p · h, h ∈ H}

• The fibres are diffeomorphic to H but do not have a natural group structure.
Every smooth n-dimensional manifold M has a natural smooth GLn (R)-principal
bundle, L(T M ) known as the frame bundle. The fibre at p ∈ M of this bundle will
be the set of linear frames of Tp M , described as follows
L(T M )p = GL(Rn , Tp M ).
`
There is a natural right-action of GL(Rn ) ∼= GLn (R) on L(T M ) = p∈M L(T M )p
given by u · g = u ◦ g , the natural projection defined by: π : up ∈ L(R, Tp M ) ,→
L(T M ) 7→ p, certainly satisfies π(ug) = π(u); i.e. preserves the fibres.
Let (x, U ) be some chart on M , then for each p ∈ U we have a natural map:
L(Rn , Tp M ) → {p} × GL(Rn ) ∼ = GLn (R) given by: up 7→ (p, x∗p ◦ up ), hence we
−1
get a bijective map φ : π (U ) → U × GLn (R) such that the following diagram
commutes:
ϕ
π −1 (U ) U × GL(Rn )
π
πU
U

42
We just need to define a topology and a smooth structure to make all of this smooth;
we proceed as for the tangent bundle. Compose φ with the chart of the manifold
and define:
φ̂ : π −1 (U ) → x(U ) × GL(Rn ),
these will be our local charts. Endow L(T M ) with the coarsest topology that makes
π and all φ̂ for any chart (x, U ) continuous; they are then homeomorphisms. If
(x, U ), (y, V ) U ∩ V 6= ∅ are two charts and φ̂, ψ̂ are the corresponding charts on
L(T M ), then: ψ̂ ◦ φ̂−1 : U ∩ V × GL(Rn ) → U ∩ V × GL(Rn ) is given by:

ψ̂ ◦ φ̂−1 (p, g) = (y ◦ x−1 (p), (y ◦ x−1 )′ (p) ◦ g).

This is smooth (because M is smooth) and so the maximal atlas determined by the
charts φ̂ is a smooth structure on L(T M ) with respect to which the maps φ are
smooth diffeomorphisms and π is smooth !

Definition III.9: Local section

Let π : P → M be a smooth principal H bundle, a local section of P is a

smooth map: σ : U → P where U is an open set of M and π ◦ σ = idU .

Example III.13. Let M be a smooth manifold and L(T M ) its frame bundle, a local
section: σ : U → L(T M ) is a smooth choice of linear frame of Tp M for each p ∈ U .
Indeed by definition for each p ∈ U , σ(p) ∈ Lp (T M ) = GL(Rn , Tp M ) hence setting
for each i ∈ {1, . . . , n}, Ei (p) = σ(p) · ei , we get n-smooth vector fields that span
Tp M for each p ∈ U .
In particular if (x, U ) is a local chart: p ∈ U 7→ x−1 ∗x(p) is a smooth section of
L(T M ) on U and the corresponding frame is the coordinate frame ( ∂x∂ i ).
Remark III.3. Smooth sections on L(T M ) are equivalent to local bundle trivialisa-
tions, indeed, define:
φ : U × H −→ π −1 (U )
(p, h) 7−→ σ(p) · h
This is a smooth bijective map since the H -action is smooth, the derivative: φ∗(p,h) =
Rh∗σ(p) σ∗p + Lσ(p) ∗h this is a bijective map at each p and so by the local inversion
theorem φ is invertible and φ−1 is a bundle chart. Conversely if φ : π −1 (U ) → U ×H ,
then σ(p) = φ−1 (p, e) is a local section on M . It follows in particular that a principal
bundle is trivial (i.e. a product) if and only if there is a global section. Moreover, in
general we should only expect there to be local sections.
The frame bundle L(T M ) on M will play the same role as G = Rn ⋊ GLn (R) in
the case of affine space, we now want an analogue of the Maurer-Cartan form ωG ...

43
III.H. Affine connections on manifolds and curvature

Throughout this section let G = Rn ⋊ GLn (R), H = GLn (R) and g, h their
Lie algebras.

We are one step away from our definition of affine connections. Since the right
action of GLn (R) preserves the fibres we have a natural generalisation of a funda-
mental vector field.

Definition III.10
Let X ∈ h, then define a vector field X ∗ on L(T M ) by:

d
Xp∗ = (p exp(Xt))
dt t=0

X ∗ is a smooth vector field on L(T M ) called the fundamental vector field

associated with X ∈ h.

We can now state:

Definition III.11: (Cartan) Affine connections

Let Rh denote right-multiplication by an element of H in L(T M ). We will
call an affine connection a g-valued one form on L(T M ) such that:

1. ωp : Tp L(T M ) → g is a linear isomorphism.

2. Rh∗ ω = Adh−1 ω , h ∈ H

3. ω(X ∗ ) = X for any X ∈ h.

Since we have no notion of left-invariant vector fields on L(T M ), we cannot ex-

pect ω to satisfy an analogue of Proposition III.7. Instead, this will be a measure of
curvature:

44
 Definition III.12: Curvature of a connection
Let ω be an affine connection on a smooth manifold, the curvature form Ω is
the g-valued 2-form defined by:

1
Ω = dω + [ω, ω].
2
We shall say that the connection is torsion-free if Ω is h-valued.

The intuitive meaning of this definition is that we are using affine geometry as
a local model for a geometry on M . The Cartan connection tells us how to “slide”
affine space along M and the curvature Ω is an infinitesimal measure of the differ-
ence between M and the model space. The curvature has the following remarkable
property:

Proposition III.8

dΩ + [ω, Ω] = 0.

Proof. This is a short computation that reduces to the Jacobi identity of the Lie
algebra g:
1
dΩ + [ω, Ω] = [dω, ω] + [ω, Ω] = [Ω − [ω, ω], ω] + [ω, Ω]
2
1
= − [[ω, ω], ω].
2
However, for any vector fields X, Y, Z ∈ Γ(T M ):
[[ω, ω], ω](X, Y, Z) = [[ω(X), ω(Y )], ω(Z)] + [[ω(Y ), ω(Z)], ω(X)]
+ [[ω(Z), ω(X)], ω(Y )]
=0

Our definition of affine connection is slightly more general than the classical def-
inition. Let us make an algebraic observation:

Proposition III.9

Let g be the Lie algebra of Rn ⋊GLn (R), h the Lie subalgebra of GLn (R), then
the GLn (R)-module g (with the adjoint representation) admits a H -module
decomposition:
g = Rn ⊕ h.

45
Proof. Recall that:
  
M b
g= , M ∈ Mn (R), b ∈ R .
n
0 0
which gives us immediately the vector space decomposition,
 we just
 need to check
A 0
that each component is indeed Ad(H)-invariant. Let h = , let us first con-
  0 1
I b
sider the Rn component and an element: x = n , then:
0 1
    −1   
−1 A 0 In b A 0 In Ab
hxh = = .
0 1 0 1 0 1 0 1

This proves that the Rn component is Ad(H)-invariant, and the adjoint representa-
tion reducesto the
standard representation of GLn (R). A similar calculation shows
B 0
that if: x = then:
0 1
 
−1 ABA−1 0
hxh = .
0 1

We conclude that the GLn (R) component of the direct sum decompostion is also
Ad(H) invariant and reduces to the standard adjoint representation of H on h =
Mn (R).
This means that any affine connection ω can be decomposed as: ω = θ + γ ,
where θ is a Rn valued one-form with Rh∗ θ = h−1 θ and γ a h-valued one form with
Rh∗ γ = Adh−1 γ .
Remark III.4. For some calculations it is nice to notice how to calculate the bracket
in terms of the semi-direct product of Lie algebras, g = Rn ⊕ h, we have:

[b + A, b′ + A′ ] = Ab′ − A′ b + [A, A′ ]. (III.4)

It turns that there is a canonical choice for the θ component on L(T M ).

Definition III.13: Solder form

Let π : L(T M ) → M be the frame bundle then for every up ∈ L(Rn , Tp M ) ,→

L(T M ), define:
θup = u−1
p ◦ π∗up .

This defines a smooth Rn -valued one-form on L(T M ) known as the canonical

one-form or solder form.

46
 Now:
(Rh∗ θ)up = (up ◦ h)−1 ◦ (π ◦ Rh )∗up = h−1 ◦ θup .
(Where we have used that π ◦ Rh = π ), it therefore satisfies the correct transform
rule (we say that it is GLn (R)-equivariant). Furthermore: θ(X ∗ ) = 0 for any X ∈ h.

Proposition III.10

Let γ be a h-valued one form on L(T M ) such that:

1. Rh∗ γ = Adh−1 α,

2. γ(X ∗ ) = X for any X ∈ h,

then ω = θ + γ , where θ is the canonical one-form, is an affine Cartan con-

nection on M .

Remark III.5. γ is known as a principal connection on L(T M ).

To specify an affine connection, it is therefore sufficient to give a principal con-
nection on L(T M ), we will sometimes refer to γ as a linear connection.

We will now assume that all the affine connections we consider arise in this
manner from a linear connection.

The curvature form Ω can also be split up into two parts:

Proposition III.11

Let ω = θ + γ be an affine connection; with respect to the H -module decom-

position g = h ⊕ Rn the curvature splits as follows:

1
Ω = dθ + [γ, θ] + dγ + [γ, γ] .
| {z } | 2
{z }
Rn -valued 2-form
gln (R) valued 2-form

where [γ, θ](X, Y ) = γ(X) · θ(Y ) − γ(Y ) · θ(X), X, Y ∈ Γ(L(T M )).

• T = dθ + [γ, θ] is called the torsion of the connection ω .

• The form R = dγ + 12 [γ, γ] is the curvature of the linear connection γ .

Proposition III.8 has the following expression in terms of T and R:

47
 Proposition III.12: Bianchi identities

1. dT + [γ, T ] = [R, θ]

2. dR + [γ, R] = 0

where:
1 X
[R, θ](X1 , X2 , X3 ) = ε(σ)R(Xσ(1) , Xσ(2) ) · θ(Xσ(3) ).
2 σ∈S
3

To understand this formula recall that R is gln (R)-valued and therefore acts
naturally on the Rn -valued form θ.

III.I. Local gauge definition

Theoretically, the definition above is very satisfying; it encodes a lot of information
into a global object. However, it is not very practical for local considerations: to cal-
culate things we need to work in charts. Besides, we are more familiar with working
with forms on T M rather than on L(T M ). We therefore encode the data of ω in
another way.
Let σ : U → L(T U ) be a local section of π : L(T M ) → M . Interpreted as a
choice of frame in Tp M at each p ∈ U . The pullback of ω by σ , σ ∗ ω = ωU is g-valued
one-form on M known as a local gauge or local connection form. This is the differential
information telling us how the frame moves. We shall now seek the consequences
of the properties of ω on ωU .
First: σ ∗ ω p : Tp M → g/h ∼ = Rn is an isomorphism, this follows directly from the
fact that ωp : Tp M → g is an isomorphism. When ω = θ + γ , where θ is the solder
form, this amounts to projecting onto this component and considering: σ ∗ θ, which
we can evaluate:
(σ ∗ θ) = θσ(p) = σ(p)−1 ◦ (π ◦ σ)∗p .
| {z }
idTp M

Recall that: σ(p) : Rn → Tp M . Define, Ei (p) = σ(p) · ei where (ei ) is the canonical
basis of Rn , and define (ω i (p)) the dual basis in Tp∨ M , one can therefore, express,
σ ∗ θ as the Rn valued one form:
 
ω1
 
σ ∗ θ =  ...  . (III.5)
n
ω

48
The pullback of the canonical one form is hence simply the map that to a vector
associates its coordinates in the basis determined by the frame σ(p) and is clearly an
isomorphism in the fibres.
The property Rh∗ ω = Adh−1 ω , tells us how to change local gauge. Suppose that
σ̃ : V → L(T M ) is another local gauge defined on V such that U ∩ V 6= ∅. Let us
set ωV = σ̃ ∗ ω and compare ωV and ωU on U ∩ V . First note that there is in fact
a smooth function h : U ∩ V 7→ GLn (R) such that σ̃ = σh = Rh ◦ σ . For every
p ∈ U ∩ V , h(p) is simply the map that changes basis of the frame at p. We claim
that:
(σ̃ ∗ ω) = Adh−1 ωU + h∗ ωH ,
where ωH is the Maurer-Cartan form of H . To prove this, we need to evaluate:
ωσ(p)h(p) ◦ (Rh ◦ σ)∗p ,
The key is to understand how to calculate the derivative of L(T M ) × H → L(T M ),
(r, h) 7→ rh, in terms of those of the maps: Rh and Lr : H → L(T M ), h 7→ rh, using
the isomorphism: T(r,h) (L(T M ) × H) = Tr L(T M ) × Th H it can be described as the
map:
(X, Y ) 7→ Rh∗r X + Lr∗h (Y ).
We therefore have:
(σ̃ω)p = ωσ(p)h(p) · Rh(p) ∗r σ∗p +ωσ(p)h(p) · Lσ(p) ∗h(p) ◦ h∗p .
| {z }
∗
σ ∗ (Rh(p) ω)=σ ∗ (Adh(p)−1 ω)=Adh(p)−1 ωU

The last term looks rather mysterious, however for any Xp ∈ Tp M .

d
Lσ(p) ∗h(p) ◦ h∗p = σ(p)γ(t) .
dt t=0

Where γ : I → H is any curve such that γ(0) = h(p) and γ̇(0) = h∗p Xp , but we
can take: γ(t) = h(p) exp(tωH h(p) h∗p Xp ) (recall that ωH h(p) = Lh(p)−1 ∗h(p) ). It then
follows that:
Lσ(p) ∗h(p) ◦ h∗p Xp = (ωH h(p) h∗p Xp )∗σ(p)h(p) .
But, ω sends fundamental vector fields to their generator, so that:
ωσ(p)h(p) · Lσ(p) ∗h(p) ◦ h∗p = h∗ ωH .

Proposition III.13

Let (Ui )i∈I be a covering of M by open sets, σi : Ui → L(T M ) a family of

local sections and (ωUi ) a family of g-valued one forms such that the previous
properties are satisfied. Then there exists a unique g-valued one form ω on
L(T M ) such that ωUi = σi∗ ω for each i ∈ I .

49
Remark III.6. Of course the θ component is known so one only needs the data that
determines γ . If σ̃ = σh for some map h : U ∩ V → GLn (R) then
∀p ∈ U ∩ V, (σ̃ ∗ θ)p = h(p)−1 (σ ∗ θ)p ,
simply because θ(X ∗ ) = 0, so it follows that:
σ̃ ∗ γ = h∗ ωH + Adh−1 σ ∗ γ.
Hence, we only need a family of h-valued one forms, γUi such that the above is
satisfied between two local sections.

IV. Covariant derivatives and geodesics

IV.A. Covariant derivatives of vector fields

In affine space, since we have a canonical identification of tangent spaces at each
point, it is clear how to define the derivative of vector fields. This amounts to choos-
ing a constant basis and differentiating the components. A connection will enable
us to extend this notion to manifolds. First a formal definition:

Definition IV.1: Covariant derivative

A covariant derivative on the tangent bundle T M is an R-bilinear operator:

∇ : Γ(T M ) × Γ(T M ) −→ Γ(T M )

(X, Y ) 7−→ ∇X Y,

such that:

1. for any X ∈ Γ(T M ), Y 7→ ∇Y X is C ∞ (M )-linear,

2. for any smooth function f ∈ C ∞ (M ), X, Y ∈ Γ(T M )

∇X (f Y ) = X(f )Y + f ∇X Y.

An affine connection θ + γ determines a covariant derivative on T M in the fol-

lowing manner.
Let σ : U → L(T M ) be a local section, i.e. a smooth choice of linear frame of
Tp M at each p ∈ U . Recall that:
σU∗ ω = σU∗ θ + σU∗ γ,
the first part tells us how the point moves (the origin of the affine frame) and the
second how the basis vectors are changing. Let us define n vector fields on U by

50
Ei (p) = σU (p) · ei . Any vector field X can be written on U , X = X i Ei where X i are
smooth functions on U . On U define:
∇Y X = Y (X i )Ei + Ei [(σU∗ γ)ij · Y ]X j . (IV.1)
Recall that (σU∗ γ) · Y ∈ gln (R) = Mn (R) is matrix, equivalently one can view σU∗ γ as
a matrix whose components are one-forms, which justifies the notation: (σU∗ γ)ij · Y .
We must check that the above formula defines a vector field on all of M . For this,
we must investigate how this transforms under a change of local section. Suppose
σV : V → L(T M ) is another local section defining vector fields Ẽi = σV · ei . Let
h : U ∩ V → GLn (R) be the smooth map such that, σV = σU h. Write X = X̃ i Ẽi
then:    
X1 X̃ 1
 ..   .. 
 .  = h . .
Xn X̃ n
So      
Y (X 1 ) X̃ 1 Y (X̃ 1 )
 ..   .   . 
 .  = h∗ (Y )  ..  + h  .. 
Y (X n ) X̃ n Y (X̃ n )
It follows that:
 
X̃ 1

 
Y (X i )Ei = Y (X̃ i )Ẽi + Ẽ1 · · · E˜n h−1 h∗ (Y )  ... 
X̃ n
Now σV∗ γ = Adh−1 σU∗ γ + h∗ ωgln (R) , or: σU∗ γ = Adh σV∗ γ − Adh h∗ ωgln (R) so:

X j (σU∗ γ)ij · Y Ei = Ei (h[(σV∗ γ) · Y ]h−1 )ij X j − Ei (h∗ (Y )h)ij X j

= Ẽi (σV∗ γ)ij · Y X̃j − Ẽi (h−1 h∗ (Y ))ij X̃ j .
Hence, the two expressions coincide on U ∩ V and can be glued together to define
a vector field on M .

Proposition IV.1

Conversely, a covariant derivative ∇ on T M determines an affine connection

on L(T M ). If (E1 , . . . , En ) is a local frame and (ω 1 , . . . , ω n ) the dual frame,
then the corresponding gln (R)-valued one form on U is given by:

γ ij = ω i (∇ej ).

Proof. See Proposition III.13.

51
IV.B. Torsion and curvature in terms of the covariant derivative ∇
Recall that ω is said to be torsion free if the torsion form T vanishes. We will now
translate this into a property of the covariant derivative:

Proposition IV.2

Let ω = θ + γ be an affine connection and ∇ the corresponding covariant

derivative. Then ω is torsion-free if and only if:

∇X Y − ∇Y X = [X, Y ], X, Y ∈ Γ(T M )

Proof. ω = θ + γ is torsion-free means that that T = dθ + [γ, θ] = 0. Let σ : U →

L(T M ) be a local section, introduce the frame Ei = σ · ei and ω i the dual frame.
Pulling back the equation T = 0 using σ , gives:

d(σ ∗ θ) + [σ ∗ γ, σ ∗ θ] = 0.

Using Equation (III.5), and writing (Γij ) = σ ∗ γ the local connection form we get the
following system of equations:
X
dω i + Γij ∧ ω j = 0.
j

Let us calculate:

dω i (X, Y ) = X(ω i (Y )) − Y (ω i (X)) − ω i ([X, Y ])

Now by definition:

ω i (∇X Y ) = X(ω i (Y )) + Γij (X)ω j (Y ).

Hence:

dω i (X, Y ) = ω i (∇X Y − ∇Y X − [X, Y ]) − (Γij (X)ω j (Y ) − Γij (Y )ω j (X)) .

| {z }
(Γij ∧ω j )(X,Y )

Therefore:

dω i (X, Y ) + (Γij ∧ ω j )(X, Y ) = ω i (∇X Y − ∇Y X − [X, Y ]).

It follows that:

T (X, Y ) = ∇X Y − ∇Y X − [X, Y ] = 0 ⇔ dθ + [γ, θ] = 0.

52
 We will now investigate how the other part of the curvature of the connection –
the curvature form of γ – translates in terms of the covariant derivative. As is ap-
parent from the above, curvature measure commutators of derivatives, let us define
the Riemann tensor of the connection:

Proposition IV.3: Riemann tensor

Let X, Y, Z ∈ Γ(T M ), define a vector field:

R(X, Y )Z = ∇X ∇Y Z − ∇Y ∇X Z − ∇[X,Y ] Z

The expression is C ∞ (R)-linear in all variables, and antisymmetric in the first

two. It is therefore a (3, 1) tensor field on M known as the Riemann tensor.

Proof. Exercice.
We shall now derive the link between R(X, Y )Z and R = dγ + 12 [γ, γ]. Choose a
local section σ : U → L(T M ) and pull back R to a h-valued one form on U that we
shall denote by (Ωij ). Again write σ ∗ γ = (Γij ). Now for two matrix-valued one-forms
α, β , define:
(α ∧ β)ij = αki ∧ βjk .
We can then write:
[σ ∗ γ, σ ∗ γ] = 2(Γ ∧ Γ).
It follows that:
Ωij = dΓij + Γki ∧ Γkj .
Proposition IV.4

Let σ be a local section of L(T M ), (E1 , . . . , En ) and (ω 1 , . . . , ω n ) the corre-

sponding frame and dual frame, then:

ω l (R(Ei , Ej )Ek ) = Ωlk (Ei , Ej ).

Proof. Exercice.

IV.C. Acceleration of a curve

Let γ : I → M be a curve, with the extra structure induced by the connection we
can now define a more familiar notion of acceleration.
A smooth vector field along a curve γ : I → M is a smooth map: V : I 7→ T M
such that π ◦ V = γ , the canonical lift (see II.F), γ̇ , is a vector field along γ . They
form a C ∞ (I) module that we denote Cγ∞ (M ).

53
Remark IV.1. An alternative way of thinking about a vector field along γ is as a section
of the pullback of the tangent bundle along γ . This is a vector bundle on I defined
by:
γ ∗ T M = {(t, v) ∈ I × M, π(v) = γ(t).}
We now state:

Proposition IV.5

Let γ : I → M be a smooth curve, a covariant derivative ∇ on M induces a

unique R-linear operator dt
D
on vector fields along curves such that:

1. ∀t ∈ I, dt
D
(f Y )(t) = f˙(t)Y (t) + f (t) dt
D
Y (t), f : I → R smooth,

2. Let t0 ∈ I , if Z is a vector field on M such that Zγ(t0 ) = γ̇(t0 ), X a

vector field on M such that on some open neighbourhood J of t0 ∈ I ,
Xγ(t) = Y (t), then:
D
Y (t) = (∇Z X)γ(t) .
dt
Sometime we write dt D
Y = ∇γ̇ Y . If σ : U −→ L(T M ) is a local section of
the tangent bundle, let (E1 , . . . , En ) denote the local frame it determines and
consider the local gauge σ ∗ γ = (Γij ), then writing Y = Y i (t)Ei (γ(t))

(∇γ̇ Y )(t) = Ẏ i (t)Ei (γ(t)) + Y j (t)(Γij · γ̇(t))Ei (γ(t)).

Remark IV.2. In essence, this is just the pullback of the connection to the pullback
bundle.

Definition IV.2
Let γ : I → M be a smooth curve, we define its acceleration to be: ∇γ̇ γ̇ . γ is
said to be a geodesic if:
∇γ̇ γ̇ = 0.
A vector field Y along the curve γ is said to be parallel along the curve if:

∇γ̇ Y = 0.

V. Riemannian geometry
Up to now we have developed a curved version of affine geometry built around the
homogeneous model G/H where G is the affine group and H the linear group. Eu-

54
clidean geometry can also be thought of as the study of a homogeneous space G/H ,
where G = Rn ⋊ O(n) and H = O(n). There is again a Maurer-Cartan form and one
can again interpret G as a frame bundle of Euclidean space. However, the relevant
frames for Euclidean geometry are orthonormal frames.
To get a curved version of this we can still base our construction on the frame
bundle L(T M ), but this is a GLn (R)-bundle and not a On (R)-bundle. We therefore
need to reduce L(T M ) to an On (R)-bundle, which is essentially specifying which
frames are orthonormal.
On the model this is achieved by the scalar product; so we can expect that to
generalise this to manifolds we should choose a scalar product on each Tp M , this
leads to the following definition:

Definition V.1: Riemannian metric

A Riemannian metric on M is a C ∞ (M )-bilinear symmetric map:

g : Γ(T M ) × Γ(T M ) → C ∞ (M ),

such that at each point the induced map gp : Tp M × Tp M → R is positive-

definite.

We call (M, g) a Riemannian manifold. Given this extra structure there is a canon-
ical choice of connection on L(T M )

Proposition V.1: Levi-Civita connection

There is a unique torsion-free covariant derivative ∇ on T M that is compatible

with the metric g in the sense that for any vector fields X, Y, Z ∈ Γ(T M ):

X(g(Y, Z)) = g(∇X Y, Z) + g(Y, ∇X Z).

Proof. Let us derive a formula for g(∇X Y, Z).

X(g(Y, Z)) = g(∇X Y, Z) + g(Y, ∇X Z) (V.1)

Y (g(X, Z)) = g(∇Y X, Z) + g(X, ∇Y Z) (V.2)
Z(g(X, Y )) = g(∇Z X, Y ) + g(X, ∇Z Y ). (V.3)

Now add the first two equations and substract the last one, and use that the con-
nection is torsion free to find:
X(g(Y, Z)) + Y (g(X, Z)) − Z(g(X, Y )) = 2g(∇X Y, Z) − g([X, Y ], Z)
+ g([X, Z], Y ) + g(X, [Y, Z]).

55
Hence:
2g(∇X Y, Z) = X(g(Y, Z)) + Y (g(X, Z)) − Z(g(X, Y ))
+ g([X, Y ], Z) + g([Z, X], Y ) + g([Z, Y ], X).

We leave it as an exercice to show that this formula determines a covariant deriva-

tive (and hence a connection) on L(T M ).
Remark V.1. The covariant derivative can be extended in a unique way to tensor fields
by imposing the Leibniz rule (with to the tensor product), i.e. define: ∇g by

X(g(Y, Z)) = (∇X g)(Y, Z) + g(∇X Y, Z) + g(Y, ∇X Z),

i.e.
(∇X g)(Y, Z) = X(g(Y, Z)) − g(∇X Y, Z) − g(Y, ∇X Z).
The compatibility condition between ∇ and g can then simply be stated as the van-
ishing of the covariant derivative of g .
Let V be a vector space S 2 V ∨ the space of symmetric bilinear forms on V . If
α, β ∈ V ∨ , define:
1
(α β)(v1 , v2 ) = (α(v1 )β(v2 ) + α(v2 )β(v1 )) .
2
If (e1 , . . . , en ) is a basis of V , then: (ei `ej )i≤j is a basis of S 2 V ∨ .
On manifolds, define S 2 (T ∨ M ) = p∈M S 2 Tp∨ M and equip it with a topology
and smooth structure as we have done for the other bundles. Any local section
σ : U → L(T M ), corresponding to vector fields (E1 , . . . , En ) that at each point
p ∈ U spans Tp M , gives rise to smooth fields: (ω i ω j )i≤j that at each p ∈ U span
S 2 Tp∨ M where (ω i ) is the dual frame. We will often use this to expression the metric
locally.
Example V.1. The standard Riemannian metric on R3 can be written, in the standard
global frame:
g = dx dx + dy dy + dz dz.
It is custom to write: dx dx = dx2 . Let us now consider the unit sphere S 2 in
R3 , and introduce spherical coordinates according to x = cos φ sin θ, y = sin φ sin θ,
z = cos θ. Then on S 2 we have the relations:
dx = − sin φ sin θdφ + cos φ cos θdθ,
dy = cos φ sin θdφ + sin φ cos θdθ, dz = − sin θdθ

Restricting g to S 2 we get the standard round metric on S 2 :

g = dθ dθ + sin2 θdφ dφ = dθ2 + sin2 θdφ2 .

56
 Let us determine the Levi-Civita connection in the local orthonormal frame:
E1 = ∂θ∂
, E2 = sin1 θ ∂ϕ
∂
. Let ω 1 = dθ, ω 2 = sin θdφ. By definition it is the unique
one-form valued matrix (Γij ) such that:

dω i + Γij ∧ ω j = 0, (torsion-free).

The compatibility condition written for Y = Ei , Z = Ej becomes, since g(Ei , Ej ) =

δij .
∀X ∈ Γ(T M ), 0 = Γji (X) + Γij (X)
i.e. (Γij (X)) ∈ so2 (R). By uniqueness if we can find such a matrix (Γij ) then it is the
Levi-Civita connection form. we have: dω 1 = d2 θ = 0 and dω 2 = cot θω 1 ∧ ω 2 .
We must have:
0 = Γ12 ∧ ω 2 ,
cot θω 1 ∧ ω 2 = −Γ21 ∧ ω 1 ,
Γ21 = cot θω 2 ,
determines a so2 (R)-valued form that solves the torsion free equation, by uniqueness
this is the Levi-Civita connection.

VI. Associated vector bundles

To conclude these notes, it seems appropriate to add some remarks that allow us
to make the bridge between the Cartan connection point of view we developed and
the more classical approach.
The most effective way to jump between invariant objects on L(T M ) and M are
the notions of equivariant horizontal vector valued forms on L(T M ) and associated
vector bundles. We start with the latter, let π : P → M be a (smooth) principal
H -bundle for some Lie group H and ρ : H → GL(V ) a linear representation of H
on a vector space V . One can then construct a vector bundle with fibre V over M ,
the associated vector bundle, written P ×H V . This is defined as the quotient set of
P × V under the equivalence relation:

(p, v) ∼ (p′ , v ′ ) ⇔ ∃h ∈ H, p′ = ph, v ′ = h−1 v.

Let us denote an equivalence class by [p, v]. The idea behind this is that the couple
(p, v) consisting of a frame p and the expression of the vector in the frame p. The
vector is then the equivalence class of all couples under a change of frame. The
projection, π̃ is given by factorisation of the map (p, v) → π(p); this is possible
because: π(ph) = π(p) and so π(p) only depends on the equivalence class of [p, v].

57
 If φ : π −1 (U ) → U × H is a local trivialisation, we construct a vector bundle chart
by factorisation of:

π̂ −1 (U ) = π −1 (U ) × V −→ U ×V
(p, v) 7−→ (π(p), πH (φ(p))v)

Where πH denotes the projection onto the H component in U × H . If we replace

(p, v) by (ph, h−1 v) then:

πH (φ(ph))h−1 v = πH (φ(p))hh−1 v = πH (φ(p))v.

Since φ(ph) = φ(p) · h. These will be our local vector bundle charts, it is then an
exercise to show that equipping P ×H V with the quotient topology, constructing
charts from these in the usual way defines a smooth structure on P ×H V that makes
these smooth vector bundle charts.
Example VI.1. Let P = L(T M ) be the frame bundle of a smooth manifold M and
ρ the standard representation of GLn (R) on V = Rn . Then P ×H L(T M ) ∼ = TM.
The isomorphism is given by [up , v] 7→ up · v ∈ Tp M ,→ T M , this is a vector space
isomorphism between the fibres and therefore a vector bundle isomorphism. All
tensor bundles arise in this way as associated bundles to the frame bundle.
We will now use associated vector bundles to relate objects on L(T M ) to objects
on M . First let us consider the h-valued one form γ , γup : Tup P → h is no longer
an isomorphism and has a kernel. Define Vup = ker π∗up , where π : P → M is
the projection . We shall refer to these vectors as vertical vectors; intuitively these
are vectors that point in the fibres of L(T M ). Since imγup = Vp by the property
γ(X ∗ ) = X , its kernel is a complementary subspace to Vp that we shall call the
horizontal subspace, Hp . Intuitively, this is a choice of subspace that is isomorphic
to Tp M in Lup (T M ). Any vector X in Tup L(T M ) therefore has a decomposition
X = XH + XV .
Now consider a representation ρ of H on some vector space V , and say that a
V -valued k -form α on L(T M ) equivariant if:

Rh∗ α = ρ(h)−1 · α.

We shall say that it is horizontal, if, α(X1 , . . . , Xk ) = 0 whenever at least one Xi is

a vertical vector field. We shall say that α is tensorial if α is both horizontal and
equivariant. It is then possible to show that:

Proposition VI.1

Let ρ be a given representation of H on V , then tensorial k -forms on L(T M )

are equivalent to sections of L(T M ) ×H V .

58
Example VI.2. • The curvature dα + 12 [α, α] is equivariant (with respect to the
adjoint representation) and horizontal; the corresponding field on M is the
Riemann tensor.

• The connection α is equivariant (with respect to the adjoint representation),

but not horizontal.

• The solder form is a horizontal and equivariant one form (with respect to the
standard representation).

59
A. Compactness in second-countable spaces
It is well known that for metric spaces compactness is equivalent to sequential com-
pactness, i.e. the following two notions of compactness coincide:

Definition: Compactness

Let X be a Hausdorff topological space. We say that:

• X is compact if satisfies the Heine-Borel property: every cover of X with

open sets has a finite subcover.

• X is sequentially compact if every sequence has a convergent subsequence.

For second-countable spaces, these notions also coincide. Let B = (Vn )n∈N be a
countable basis for the topology:
• Assume first compactness, then X is sequentially compact. Recall first that:
– A family of closed sets (Fi )i∈I in X is said to satisfy the finite intersection
property if for any finite subset J ⊂ I , ∩j∈J Fj 6= ∅.
– A Hausdorff topological space X is compact if and if for any family of
closed subsets (Fi )i∈I that satisfy the finite intersection property one has
∩i∈I Fi 6= ∅.
Indeed, suppose first that there is a family (Fi )i∈I of closed subsets that
satisfy the finite intersection property but such that ∩i∈I Fi = ∅. Then
(X \ Fi )i∈I is a family of open subsets that has no finite subcover, hence
X is not compact. Conversely, suppose that X is not compact, then
there is an open cover (Ui )i∈I of X that has no finite subcover, but then
(Fi = X \ Ui )i∈I is a family of closed sets with the finite intersection
property such that ∩i∈I Fi = X \ ∪i∈I Ui = ∅.
The decreasing family of closed subsets (Fn )n∈N , defined by Fn = {xm , m ≥ n}
for some sequence (xn )n∈N clearly satisfies the finite intersection property,
hence, X being compact ∩n∈N Fn 6= ∅. Choose x ∈ ∩n∈N Fn then one can con-
struct a subsequence converging to x in the following manner. Observe that
one can construct from B a non-increasing basis of open neighbourhoods of
x, (Vmx )m∈N ; this will play a role similar to open balls in the metric case. We
construct an increasing extraction map φ : N → N by induction. First, for
n = 0, since V0x is a neighbourhood of x there is an infinite number of n ∈ N
such that xn ∈ V0x , let φ(0) = min{n ∈ N, xn ∈ V0x }. Suppose that we have de-
fined φ(k) for k ∈ J0, nK such that φ(0) < φ(1) < · · · < φ(n) and for all k ∈ N,
xϕ (k) ∈ Vkx . To define φ(n + 1), using the fact that {n ∈ N, xn ∈ Vn+1x
} is infi-
nite one can again choose φ(n+1) > φ(n) such that xϕ(n+1) ∈ Vn+1 . (xϕ(n) )n∈N
x

60
 converges to N, because if V is any neighbourhood of x, then there is n0 ∈ N
such that Vnx0 ⊂ V , by definition of φ for any n ≥ n0 , xϕ(n) ∈ Vnx ⊂ Vnx0 ⊂ V.

• Assume now sequential compactness and let (Ui )i∈N be an arbitrary cover of
X by open sets.
1. (Ui )i∈N has a countable subcover, indeed for every x ∈ X, ∃ix ∈ I, x ∈
Uix , but one can find nx such that Vnx ⊂ Uix . However: {nx , x ∈ X} = C
is countable and ∪c∈C Vc covers X . Moreover by construction, for each
c ∈ C , one can find at least one i(c) such that Vc ⊂ Ui(c) . (Ui(c) )c∈C is
then a countable subcover.
2. Countable covers have finite subcovers. Let (Un )n∈N be a countable cover
of X by open sets. Suppose that it has no finite subcover then for every
n ∈ N there is xn ∈ X \ ∪ni=1 Ui . By sequential compactness, (xn )n∈N has
a convergent subsequence. Let us call the limit x ∈ X ; by construction:
x ∈ X \ ∪ni=1 Ui for all n ∈ N. Which contradicts the fact that (Un )n∈N
covers X . Therefore (Un )n∈N has a finite subcover.

61
B. Tensor algebra, exterior algebra, differential forms
(to be completed)
The following is a short review on the notion of tensor algebra, it gives a more ab-
stract foundation to our notion of exterior algebra. We begin with the basic propo-
sition that we shall understand as our definition for tensor products:

Proposition B.1: Tensor product

Let V1 , V2 be two vector spaces over the same field K, then up to isomorphism
there is a unique K-vector space written V1 ⊗ V2 equipped with a bilinear map:
j : V1 × V2 → V1 ⊗ V2 satisfying the following universal property: for any
bilinear map b : V1 × V2 → V3 , there is a unique linear map: l : V1 ⊗ V2 → V3
such that the following diagram commutes:

V1 × V2 b
V3
l
j

V1 ⊗ V2

Proof. See your favourite reference in general algebra; for instance: [Hun74] or [Lan05].

Remark B.1. • The same construction is valid for modules over rings.

• It is custom to write j(v1 , v2 ) = v1 ⊗ v2

• Note that elements of V1 ⊗ V2 are not all of the form v1 ⊗ v2 , they are linear
combinations of such products. Indeed, let l : V1 ⊗ V2 → K be a linear
map that vanishes on any element of the form vi ⊗ vj , then l fits into the
commutative diagram:

V1 × V2 0
R
l
j

V1 ⊗ V2

However, this diagram the null map fits into this same diagram, therefore by
uniqueness: l = 0.

• If dim V1 = n, dim V2 = m and (ei ), (fj ) are bases of V1 and V2 respectively,

then: (ei ⊗ ej ) is a basis for V1 ⊗ V2 .

62
 • K⊗V ∼
= V (map v ∈ V to 1 ⊗ v )
The tensor product of two spaces solves the problem of exchanging bilinear maps
for linear maps. As an exercise let us prove that ⊗ is associative in the following
sense: (V1 ⊗ V2 ) ⊗ V3 ∼ = V1 ⊗ (V2 ⊗ V3 ). Using the universal property there is a
unique linear map L : (V1 ⊗ V2 ) ⊗ V3 → V1 ⊗ (V2 ⊗ V3 ) such that (v1 ⊗ v2 ) ⊗ v3 7→
v1 ⊗ (v2 ⊗ v3 ) for all (v1 , v2 , v3 ) ∈ V1 × V2 × V3 . It is the required isomorphism,
indeed construct in the same way the map L̃ : V1 ⊗ (V2 ⊗ V3 ) → (V1 ⊗ V2 ) ⊗ V3 such
that v1 ⊗ (v2 ⊗ v3 ) 7→ (v1 ⊗ v2 ) ⊗ v3 then L ◦ L̃ fits into the diagram:
j
V1 ⊗ (V2 ⊗ V3 ) V1 ⊗ (V2 ⊗ V3 )
L◦L̃
j

V1 × (V2 ⊗ V3 )

But this diagram is also satisfied by id, hence, by uniqueness L̃ ◦ L = id. Inversing
the roles of the spaces we also get L ◦ L̃ = id.
We are particularly interested in the case where V1 = V2 , in this case we will write
V ⊗ V = ⊗2 V , then define by induction: ⊗n V = V ⊗ (⊗n−1 V ) and set:

M
+∞
T (V ) = ⊗k V,
k=0

where ⊗0 V = K and ⊗1 V = V . This is an associative unit algebra, equipped with

multiplication given by ⊗, the unit is 1 ∈ K known as the tensor algebra. Note that
it is not commutative !
The tensor algebra is in some sense the most general associative unit algebra one
can construct from a vector space V ; it is therefore a powerful tool in constructing
other associative algebras. To illustrate this, we will now use it to construct the
exterior algebra.
Consider I the bilateral (e.g. left and right) ideal generated by elements of the
form v ⊗ v, v ∈ V we define:
^
V = T (V )/I.
V
We will note multiplication in V by ∧, it is generated by elements of v . By con-
struction if v ∈ V , v ∧ v = 0 and 0 = (v1 + v2 ) ∧ (v1 + v2 ) = v1 ∧ vV 2 + v2L ∧ v1 . V Note
that the grading of the tensor algebra passes to ∧V and we have: = k=1 k V
n
Vk
where V is the vector space spanned by elements of the form v1 ∧ · · · ∧ vk ,
v1 , . . . , vk ∈ V . Note that by construction: v1 ∧ · · · ∧ vk vanishes if (v1 , . . . , vk ) is
a linearly dependent family, it follows immediately that if dim V = n, then the ex-
V
terior algebra is finite dimensional. k V satisfies the following universal property,

63
 | × ·{z
for any u : V · · × V} → W , k -multilinear, alternated, there is a unique linear map
k times
V
ũ : k V → W such that if j is the map: (v1 , . . . , vk ) 7→ v1 ∧ · · · ∧ vk then u = ũ ◦ j .
(This can be shown using the universal property of ⊗k V and taking the quotient).
With this in mind we can show that if (e1 , . . . , ek ) is a basis of V then,

{ei1 ∧ · · · ∧ eik , 1 ≤ i1 < i2 < · · · < ik ≤ n}

V
is a basis of k V . Suppose now that V = E ∗ for some vector space E . Recall that
a k -form on E is K-valued k -multilinear alternated map.

64
References
[Hun74] T. W. Hungerford. Algebra. Graduate Texts in Mathematics. Springer
New York, NY, 1974.
[Lan05] S. Lang. Algebra. Graduate Texts in Mathematics. Springer New York,
2005. ISBN: 9780387953854.
[Lan95] S. Lang. Differential and Riemannian Manifolds. Third Edition. Graduate
Texts in Mathematics. Springer-Verlag New York, 1995.
[Lax02] P. D. Lax. Functional Analysis. Pure and Applied Mathematics. Wiley, 2002.
[Lee03] J. M. Lee. Introduction to Smooth Manifolds. Second Edition. Graduate Texts
in Mathematics. Springer, 2003.
[Nab10] G. Naber. Topology, Geometry and Gauge fields : Foundations. Texts in Applied
Mathematics. Springer, 2010.
[Rud69] M. E. Rudin. “A new proof that metric spaces are paracompact”. In: Proc.
Amer. Math. Soc. Vol. 20. 1969, p. 603.
[Sha97] R. Sharpe. Differential Geometry: Cartan’s Generalization of Klein’s Erlangen
Program. Vol. 166. Graduate Texts in Mathematics. Springer-Verlag New
York, 1997.

65
Glossary
Basis (Topology) A collection B of open subsets of a topological space X such that
any open subset is a union of elements of B . A collection of open subsets B
is a basis if and only if for any x ∈ X and any open subset U such that x ∈ U ,
one can find B ∈ B such that x ∈ B ⊂ U . Example: open balls in metric
spaces.
Bump function A bump function is a smooth map f : M → R with compact sup-
port. For any n ∈ N∗ , one can find bump functions f : Rn → R. Indeed, first
consider: f : R → R+ defined by:
( 1
e− t t > 0
f (t) = .
0 t≤0

f is of class C ∞ ; now define, g : Rn → R by g(x) = f (1 − ||x||2 ). Where

|| · || is the Euclidean norm on Rn ; it follows that g is smooth with support
B(0, 1) = {x ∈ Rn , ||x|| ≤ 1}. See also Uryhson’s lemma.

Closed graph theorem Let E, F be Banach spaces and u ∈ L(E, F ) then u is con-
tinuous if and only if its graph:
G = {(x, f (x)), x ∈ E}
is closed in E × F .
Comparison of topologies Let T1 , T2 , be two topologies on the same set X , we say
that T1 is finer than T2 , written: T2 ⊂ T1 if U ∈ T1 ⇒ U ∈ T2 . In this case T2 is
also said to be coarser than T1 . The analogy is that of grains of sand, the finer
the topology the more ”grains of sand” i.e. open sets.

Functor This is a general notion from Category theory, where mathematicians do a

lot of abstract nonsense. Jokes aside, on a basic very level, I find it very useful
to organise general ideas about mathematics that are common to many dif-
ferent areas and extract the essential information in certain constructions, in
the form of universal properties. In brief, a category C is a collection of math-
ematical objects, a collection of disjoint sets Mor(O1 , O2 ) of morphisms (or
arrows) where O1 , O2 are two objects, and, for any three objects O1 , O2 , O3 a
composition law:
Mor(O2 , O3 ) × Mor(O1 , O2 ) −→ Mor(O1 , O3 )
(g, f ) 7−→ g ◦ f,
subject to the conditions:

66
 1. ◦ is associative and for each object O1 of C ,
2. there is a morphism: idO1 : O1 → O1 which acts as a left and right
identity under the composition law.
The class of all sets is a category where morphisms are maps between sets,
topological vector spaces form a category where morphisms are continuous
linear maps, smooth manifolds form a category where the morphisms are
smooth maps, Banachisable spaces form a category where morphisms are con-
tinuous linear maps, etc. A functor between categories C1 and C2 , is a mapping
between categories that:
• maps any object O1 of C1 to an object F (O1 ) of C2
• maps any morphism f : O1 → O2 to a morphism F (f ) : F (O1 ) →
F (O2 ), such that F (idO1 ) = idF (O1 ) and F (f ◦ g) = F (f ) ◦ F (g).
Maps between categories that satisfy all the above conditions except that they
reverse the order of composition, i.e. F (f ◦ g) = F (g) ◦ F (f ) are called
contravariant-functors. They are functors in the above sense on the opposite
category. See [Hun74, Chapter I.7, Chapter X] for a more detailed account.
The map that maps a Banach space to its dual is a contravariant functor.

Homeomorphism An isomorphism in the category of topological spaces, i.e. a con-

tinuous bijection with continuous inverse.

Inverse function theorem Let E and F be two Banach spaces and f : U → E → F

be a C k map, k ≥ 1, defined on an open subset U of E . Let x0 ∈ U , suppose
that f ′ (x0 ) ∈ GL(E, F ) then are open neighbourhoods V and W of a and f (a)
respectively, such that the restriction of f |V : V → W is a C k diffeomorphism.
It extends naturally to manifolds.

Locally compact topological space A Hausdorff topological space such that every
point has a neighbourhood with compact closure (i.e. relatively compact).

Modules and bimodules A R-module M is to a ring R, what a k -vector space is to

field k , i.e. (M, +) is an abelian group endowed with external multiplication
by elements of R, i.e. a map: (r, m) ∈ R × M 7→ r · m satisfying the same rules
as external multiplication in a vector space. When multiplication is written
on the left, M is said to be a left-module. One could also define external
multiplication on the right, i.e. a map: (r, m) ∈ R × M 7→ m · r. M is then
said to be a right-module. A priori, the two structures are different since, for
a right-module, we have:
m · (r1 r2 ) = (m · r1 ) · r2 .

67
 If we defined: r · m = m · r then this property would translate to:

(r1 r2 ) · m = r2 · (r1 · m).

If R is commutative the difference is purely notational. A bimodule is a module

equipped with both left and right multiplication..

Open map A continuous map f such that f (U ) is open whenever U is open.

Open mapping theorem Let E, F be Banach spaces and u ∈ L(E, F ) a continu-

ous surjective map, then u is an open map. Corollary: If u ∈ L(E, F ) is a
bijective continuous linear map between Banach spaces then u−1 is bounded.
See [Lax02, Chapter 15, Theorem 11].

Paracompact topological space A Hausdorff space with the property that every open
covering admits a locally finite refinement. Guarantees the existence of parti-
tions of unity. All metric spaces are in fact paracompact [Rud69].

Regular space A topological Hausdorff space X in which one can separate points
and closed sets with open sets. i.e. if x ∈ X , F ⊂ X closed and x ∈
/ F then
there are disjoint open sets U and V such x ∈ U , F ⊂ V . This is equivalent
to the statement that for any x ∈ X and any open subset U of X , one can find
an open subset V such that x ∈ V ⊂ V ⊂ U .

Separable space A topological space with a countable dense subset.

The Einstein summation convention When it is convenient and no ambiguity can

arise, I will use the Einstein summation convention which states that one
should sum
P over repeated indices, generally one “up”, one “down”. For in-
stance: i xi ei ≡ xi ei .

The subspace topology Let (X, T ) be at topological space, A ⊂ X , then the topol-
ogy of X induces a natural topology on A, defined by:

TA = {U ∩ A, U ∈ T }.

It is the initial topology of the inclusion map.

Topological group A group (G, ·) equipped with a topology compatible with its
group structure, i.e. the maps µ : G × G → G, ι : G → G defined by:

µ(g1 , g2 ) = g1 · g2 , ι(g) = g −1 , g1 , g2 , g ∈ G,

are continuous.

68
Topological space A topology on a set X is a collection T of subsets of X that
satisfies the following stability conditions:
1. ∅, X ∈ T ,
2. Arbitrary unions of elements of T are elements of T ,
3. Finite intersections of elements of T are elements of T .
Elements of T are called open sets. A topology is the minimum set of data
required to talk about limits, neighbourhoods, continuous maps, etc.

French-English Dictionary
arbitrary quelconque.

bundle fibré.

chain-rule règle de composition des dérivées.

chart carte.

closure fermeture, adhérence.

coarsest moins fin.

countable dénombrable.

data données.

decreasing strictement décroissant.

factor (to) verbe: factoriser, nom: facteur (dans un produit).

Hausdorff space espace topologique séparé.

hence d’où.

increasing strictement croissant.

induction récurrence.

isotropy subgroup stabilisateur.

manifold variété.

69
map application.

neighbourhood voisinage.

non-decreasing croissant.

non-increasing décroissant.

one to one injectif.

one to one correspondence bijection.

onto surjectif.

quote (to) citer.

record (to) verbe: enregistrer.

regular regulier.

second-countable à base dénombrable d’ouverts. On parle aussi du deuxième ax-

iome de dénombrabilité.

sequence suite.

smooth lisse.

span espace vectoriel engendré par, vect.

split scindée (en parlant d’une suite exacte courte).

submanifold sous-variété.

subsequence sous-suite.

subspace topology topologie induite.

70