GAUGE THEORY

FLORIN BELGUN

1. Fiber bundles

Definition 1.1. Let G be a Lie group, ρ : G × F → F a smooth left action of G on a

π
manifold F , and M a manifold. A fiber bundle E → M with structure (gauge) group
G and fiber F on the manifold M is a submersion π : E → M such that there exists an
atlas {(U, ψU ) | U ∈ U} of local trivializations of E, where:
(1) U is a covering of open sets U ⊂ M ;
(2) ψU : U × F → π −1 (U ) are diffeomorphisms such that π|π− 1 (U ) ◦ ψU = pr1 ,
where pr1 : U × F → U is the projection on the first factor. ψU is called a local
trivialization of E over U ⊂ M ;
(3) For any pair of intersecting open sets U, V ∈ U, there is a smooth map gU V :
U ∩ V → G (called a transitition function) such that
ψV−1 ◦ ψU (x, f ) = (x, ρ(gU V (x), f )), ∀(x, f ) ∈ (U ∩ V ) × F.
Remark 1.2. The atlas of local trivializations can be completed with any pair (U, ψU )
as above, such that, for any V ∈ U that intersects U , there exist transition functions
gU V and gV U satisfying the relations above. It is possible to construct a maximal such
an atlas. Note, however, that the fiber bundle structure does not contain a priori the
information given by a particular atlas of trivializations; it is only required that such
an atlas exists, and we can suppose, WLOG, that it is maximal. The notion of fiber
bundle over M generalizes the product M × F , which is called the trivial bundle with
fiber F and basis M (the group G is here irrelevant). The main question is whether
a given fiber bundle is trivial or not, i.e. whether there exists a global trivialization
ψM : M × F → E in the maximal atlas.
Remark 1.3. The transition functions gU V : U ∩ V → G depend actually on the
corresponding local trivializations ψU , ψV as well; in a maximal atlas it is common to
have different pairs (U, ψU ), (U, ψU0 ) corresponding to the same open set U . In order
to keep the (convenient) notation from the definition above, we need to use different
letters for (possibly identical) open sets in U, if the corresponding local trivializations
are not the same.

With this convention, the transition functions {gU V | U, V ∈ U} satisfy the following
properties:
(1) gU U : U → G is the constant map gU U = e ∈ G (e is the neutral element in G)
(2) gU V (x) is the inverse element in G of gV U (x)
(3) gU V · gV W = gU W on U ∩ V ∩ W , where the dot denotes the multiplication in G
of the values of the transition functions.
A collection of G-valued functions as above is called a 1-cocycle with values in G (the
terminology comes from sheaf theory).

1
2 FLORIN BELGUN

Example 1.4. For G := GL(k, R) and ρ : GL(k, R) × Rk → Rk the standard (linear)

action of GL(k, R) on Rk , the resulting bundle E is a vector bundle of rank k over
M . In this case the fibers Ex := π −1 (x) (which, in general, are submanifolds in E of
codimension equal to dim M ) are vector spaces isomorphic to Rk . Each local trivial-
∼
ization ψU , for x ∈ U , yields such an isomorphism ψU |EX : Ex → {x} × Rk , but as
new local trivializations can be constructed by multiplying the second factor of ψU with
some constant element of GL(k, R), it turns out that the vector bundle structure on E
does not identify canonically the fibers Ex with Rk . In fact, all such isomorphisms are
considered.

If we consider the general setting of a fiber bundle with group G and fiber F , we con-
clude that the relevant objects / properties of a fiber in such a fiber bundle correspond
to objects / properties of F that are invariant under the action ρ of G. We will come
back to this principle.
π
Definition 1.5. A (local) section in a fiber bundle E → M is a map σ : M → E
(σ : U → E, for U ⊂ M open in the local case) such that π ◦ σ is the identity of M ,
resp. of U .
Remark 1.6. For every local trivialization (U, ψU ), the inclusions σf : U → U × F
π
(f ∈ F ), defined by σf (x) := ψ −1 (x, f ), ∀x ∈ U , induce local sections in E → M .
π
Proposition 1.7. If the fiber F is contractible, then any fiber bundle E → M with fiber
F admits global sections.
Proposition 1.8. If the base manifold M is contractible, then there exists a global
trivialization ψ : E → M .

The proofs of these propositions can be found in the classical book by N. Steenrod
The topology of fiber bundles, where, in order to decide if a certain bundle admits or
not global sections, the obstruction theory is developped.

2. Isomorphic bundles, associated bundles, pull-back

π
Given a fiber bundles E → M with structure group G and fiber F , we obtain a
1-cocycle with values in G, consisting in the collection of transition functions
gU V : U ∩ V → G, U ∩ V 6= ∅.
Conversely, using such a collection of smooth maps gU V as above, we can “glue together”
the sets U × F to construct a fiber bundle with group G and fiber F , corresponding to
the action ρ of G on F :
Let E be the following set (t means disjoint union)
G
E := U × F/ ∼, (1)
U ∈U

where (x1 , f1 ) ∈ U × F and (x2 , f2 ) ∈ V × F are equivalent w.r.t. ∼ iff x1 = x2 and

f2 = ρ(gU V (x1 ), f1 ). The conditions for the collection {gU V }U,V ∈U imply then that ∼ is
an equivalence relation and it is easy to check that E is a manifold and the projections
from U × F to U induce a submersion π : E → M . Moreover, we also obtain local
trivializations ψU for all U ∈ U. We would like to say that, if we take our data {gU V }u∈U
 GAUGE THEORY 3

π
from a fiber bundle E 0 → M , that the constructed fiber bundle E is isomorphic to E 0 .
For this, we define first when two bundles are isomorphic:
π
Definition 2.1. Let Ei →i M , i = 1, 2 be two fiber bundles on M with structure group
G, and fiber F , (corresponding to the action ρ : G × F → F ). An isomorphism of fiber
bundles between E1 and E2 is a diffeomorphism Φ : E1 → E2 such that
(1) π2 ◦ Φ = π1 (Φ maps the fiber of E1 at x ∈ M to the fiber of E2 at x)
(2) for a covering U of M , fine enough such that ∀U ∈ U, there exist local trivial-
izations ψU1 of E1 , resp. ψU2 of E2 over U , the map
ψU2 ◦ Φ ◦ (ψU1 )−1 : U × F → U × F2
can be expressed, in terms of a smooth map gU : U → G as
(x, f ) 7→ (x, ρ(gU (x), f ))
Proposition 2.2. Two fiber bundles E1 , E2 with group G and fibre F are isomorphic iff,
for a fine enough covering U of M , the 1-cocycles {gUi V }U,V ∈U , i = 1, 2 are cobordant,
i.e. there exists a family gU : U → G of smooth maps such that
gU2 V = gU · gU1 V · gV−1 on U ∩ V, ∀U, V ∈ U.

The proof follows directly from the definition.

Remark 2.3. The isomorphism class of a fiber bundle with structure group G and fiber
F turns out to depend only on the transition functions. In terms of sheaf theory, the
previous proposition establishes an equivalence between an isomorphism class of fiber
bundles with group G (and any fiber F ) and a Čech cohomology class in H 1 (M, G) (see,
e.g., R. Godement, Topologie algébrique et théorie des faisceaux).
Definition 2.4. Two bundles E → M and E 0 → M with the same gauge group G and
the same transition functions {gU V | U, V ∈ U} (and different fibers F , F 0 ) are called
associated.
Remark 2.5. By (1) we can construct associated bundles for any given G-manifold F
(a G-manifold is a manifold together with a left action of G).

We introduce now some operations with fiber bundles; the first one is the pull-back,
that constructs, starting from a fiber bundle E → M as above, a fiber bundle with the
same structure group and the same fiber over another manifold:
π
Definition 2.6. If E → M is a fiber bundle with group G and fiber F (G acting by ρ
on F ), and f : N → M is a smooth map, one defines the pull-back f ∗ E of E to N the
following fiber bundle with group G and fiber F on N :
(1) E 0 := f ∗ E := {(x, y) ∈ N × E | f (x) = π(y)};
(2) π 0 : E 0 → N , π 0 (x, y) := π(y)(= f (x)), ∀(x, y) ∈ E 0 ; also denote by f¯ : E 0 → E,
f¯(x, y) := y, ∀(x, y) ∈ E 0 , the induced bundle map;
(3) the local trivializations are defined on the covering f −1 (U) := {f −1 (U ) | U ∈ U}
as
−1
ψ 0 −1 (x, y) := (x, pr2 (ψU (y))), ∀(x, y) ∈ π 0 (f −1 (U )) = f¯−1 (π −1 (U )).
f (U )

(4) the transition functions gf0 −1 (U )f −1 (V ) := gU V ◦ f , ∀U, V ∈ U.

4 FLORIN BELGUN

The pull-back f ∗ E can be characterized, up to bundle isomorphism, simply as the

manifold E 0 such that the maps π 0 and Ff below are a submersion, resp. a smooth map
such that the fibers of π 0 are mapped diffeomorphically to fibers of π, and the following
diagram commutes:
f¯ -
E0 E

π0 π
? f- ?
N M
Exercise 2.7. It is clear that over a point (seen as a manifold • of dimension 0),
there exists only one (trivial) bundle with group G and fiber F , namely F → •. Show
that a bundle E → M as above is trivial iff it is isomorphic with the pull-back of any
G, F -bundle through a constant map.
π
In the special case when f itself is the projection of a fibre bundle E2 →2 M (and
π
where we denote our initial fiber bundle on M as E1 →1 M ), we get that π2∗ E1 = π1∗ E2
and π10 = π̄2 and π20 = π̄1 are both fiber bundle projections. Moreover, if we denote by
E1 ×M E2 := π2∗ E1 = π1∗ E2 the fibered product of E1 and E2 , then the diagonal map
π := π1 ◦ π20 = π2 ◦ π10 is a fiber bundle projection, too.
π20 -
E1 ×M E2 E1

π
π10 1
-
? π2 - ?
E2 M
The gauge group is then G := G1 × G2 and the fiber is F := F1 × F2 (corresponding
to the representation ρ : G × F → F , ρ((g1 , g2 ), (f1 , f2 )) := (ρ1 (g1 , f1 ), ρ2 (g2 , f2 )),
∀(g1 , g2 ) ∈ G, (f1 , f2 ) ∈ F .
π
In the particular case where both fiber bundles Ei →i M are vector bundles, the
fibered product E1 ×M E2 has as fiber at x ∈ M the direct sum of the fibres (E1 )x and
(E2 )x . For that reason, it is usually denoted by E1 ⊕ E2 and is usually referred to as the
Whitney sum of the vector bundles E1 and E2 . Note that, by construction, the gauge
group of E1 ⊕ E2 = E1 ×M E2 is naturally a subgroup of GL(n1 , R) × GL(n2 , R) (in case
the ranks of Ei are ni and Gi act effectively of Rni , i = 1, 2). To determine whether a
given vector bundle E can be decomposed as a Whitney sum is a non-trivial topological
problem, and it amounts to reduce the structure (gauge) group of E to a subgroup of
such a product. More about gauge reduction in the next section.
π
Exercise 2.8. (1) Show that every (local) section of E → M induces a (local)
π0
section of f ∗ E → M ;
(2) Show that a (local) section in E1 ×M E2 → M is equivalent to a pair of (local)
sections in E1 → M , resp. E2 → M ;
π0
(3) Show that if f : U ,→ M is the inclusion of an open subset of M , then f ∗ E → M
is the restriction to U , resp. π −1 (U ) of the bundle projection π : E → M . f ∗ E
is sometimes denoted by E|U .
 GAUGE THEORY 5

3. Principal bundles. Associated bundles

π
Definition 3.1. A G-principal bundle over a manifold M is a fiber bundle P → M with
group G and fiber F := G, where the action of G on the fiber F is by left multiplication
of G on itself.
Remark 3.2. As we have seen before, the relevant structures that are canonical on P are
induced by the structures of the fiber that are invariant w.r.t. the action of the group.
E.g., the neutral element is not invariant w.r.t. the left multiplication, so the fibers
Px := π −1 (x) do not have a particular point, corresponding to the neutral element of G,
nor is the group multiplication on G invariant w.r.t. the left multiplication. Therefore,
the fibers of P are, in general, not Lie groups.
They do inherit, however, a special structure: the right multiplication Ra on G with
an element a ∈ G is invariant w.r.t. the left multiplication Lb by any element b ∈ G,
i.e.
Lb ◦ Ra = Ra ◦ Lb , ∀a, b ∈ G.
(This is the associativity in the group G.) Therefore, there is a canonical right action
of G on the total space P . This action is free (Ra : P → P has no fixed points if
a 6= e ∈ G) and proper (see below).

The importance of principal bundles is that, among all bundles with the same gauge
group and (up to co-boundaries) same 1-cocycle of transition maps (these bundles are
called associated to each other), it is the only one on which G acts (on the right) and such
that all other associated bundles can be retrieved from it by the following procedure:
π
Proposition 3.3. Let P → M be a G-principal bundle over M and ρ : G × F → F a
left action of G on a manifold F . Then the following manifold is the total space of a
fiber bundle over M with group G and fiber F (for the action ρ), associated to P :
E := P ×ρ F := P × F/ ∼, (p, f ) ∼ (p · g, ρ(g −1 , f )), ∀(p, f ) ∈ P × F, ∀g ∈ G.
The dot denotes the right multiplication of g ∈ G with the element p ∈ P . The projection
πE : E → M is defined, for an equivalence class [p, f ] ∈ E, by πE ([p, f ]) := π(p).

Because the equivalence relation above is induced by the right free action of G on
P × F:
(P × F ) × G 3 ((p, f ), g) 7→ (p · g, ρ(g −1 ) ∈ P × F,
the proposition above is a direct consequence of the following Theorem.
Definition 3.4. A continuous map f : A → B between topological spaces is called
proper iff f −1 (K) is compact in A, for any K ⊂ B compact.
A Lie group (right) action η : P × G → P on a smooth manifold P is said to be
proper iff the continuous map fη : P × P → P × P , defined by
fη (p, g) := (p, η(p, g)), ∀(p, g) ∈ P × G,
is proper.
Theorem 3.5. Let P be a smooth manifold and let G be a Lie group acting smoothly,
freely and properly on P . Then the orbit space of G can be given the structure of a
smooth manifold M , such that the canonical projection π : P → M is a submersion
and, moreover, a G-principal bundle.
6 FLORIN BELGUN

π
Remark 3.6. We have seen that in each trivialization domain U , a fiber bundle E → M
admits, for each y ∈ π −1 (U ), a section σy : U → π −1 (U ) passing through y.
π
Conversely, every local section σ : U → π −1 (U ) in a principal bundle P → M defines
a local trivialization of the bundle over U : indeed, the map Fσ : U × G → π −1 (U ),
Fσ (x, g) := σ(x) · g is a diffeomorphism, and for each local trivialization ψV : π −1 (V ) →
V × G, we have on (U ∩ V ) × G:
ψV (Fσ (x, g)) = ψV (σ(x) · g) = (x, gV σ (x) · g),
where gV σ (x) is the component in G of ψV (σ(x)).
For principal bundles, a (local) trivialization and a (local) section are thus equivalent
notions.

Sections in an associated bundle can also be described using the corresponding prin-
cipal bundle:
π
Proposition 3.7. Let P → M be a principal G-bundle, ρ : G × F → F a smooth action
of G on F and E := P ×ρ F the associated bundle to ρ. There is a 1–1 correspondence
between (local) sections σ : U → πE−1 U in E and G-equivariant maps f : π −1 (U ) → F
(if U = M , then σ : M → E and f : P → F ), i.e. smooth maps that satisfy the
equivariance property:
f (p · g) = ρ(g −1 , f (p)), ∀p ∈ P, g ∈ G.

In case F is a vector space V and ρ : G×V → V is a linear representation (one can see
ρ as a group homomorphism from G to GL(V )), one considers E-valued differential forms
on M as sections in the vector bundle Λ∗ M ⊗ E (as a vector bundle, it is constructed
by taking fiberwise tensor products of Λ∗ M and E: it is equally the associated bundle
to the GL(M ) ×M P with structure group GL(n, R) × G, where GL(M ) is the bundle
of frames on M (dim M = n) and P is the associated principal bundle to E, by the
tensor product representation λ ⊗ ρ : GL(Rn ) × G → GL(Λ∗ (Rn )∗ ⊗ V )). Such E-valued
differential forms can be characterized in terms of the associated principal bundle in an
analogous way as the sections are equivalent to G-equivariant maps on P :
π
Proposition 3.8. Let P → M be a principal G-bundle and ρ : G → GL(V ) be a linear
representation of G. Denote by E := P ×ρ V the associated vector bundle. A k-form on
M with values in E is equivalent to a V -valued k-form α on P such that
(1) Rg∗ α = ρg−1 (α), ∀g ∈ G
(2) The vertical distribution (the vertical tangent space, or the tangent space to the
fibers) T ∨ P := ker π∗ lies in the kernel of α, i.e. α(X, . . . ) = 0, ∀X ∈ T ∨ P .
Here Rg : P → P is the right action of g ∈ G on P and ρg−1 : V → V is the (left) linear
action of g −1 on V .

Given a fiber bundle with group G, one is usually interested in reducing the gauge
group to a subgroup H of G, in that one attemps to find a sub-atlas of local trivializa-
tions for which the transition functions take value in H. In case H = {e}, this amounts
to finding a global trivialization of E. The concept of gauge reduction can be defined in
full generality and characterized as follows:
 GAUGE THEORY 7

π
Definition 3.9. Let E → M be a fiber bundle with gauge group G and fiber F for
ρ : G×F → F the corresponding action. Let ϕ : H → G be a Lie group homomorphism.
We say that the gauge group of E reduces to H iff there exists a sub-atlas of local
trivializations of E such that the corresponding transition functions gU V : U ∩ V → G
factorize through ϕ, i.e. they can be expressed as gU V = ϕ ◦ hU V , where hU V is a
1-cocycle on M with values in H.

For E = P a principal G-bundle, this means that P is isomorphic to P red ×ϕ̄ G, where
red
P is a principal H-bundle (defined by the 1-cocycle {hU V } above) and ϕ̄ : H ×G → G
is the action of H by left multiplication (via ϕ) on G: ϕ̄(h, g) := ϕ(h)·g, ∀h ∈ H, g ∈ G.
Proposition 3.10. A reduction to H (for the group homomorphism ϕ : H → G) of
the gauge group of the G-bundle P is equivalent to a G-equivariant (ϕ-equivariant) map
F : P red → P from a principal H-bundle P red to P , i.e. F (p · h) = F (p) · ϕ(h),
∀p ∈ P red , h ∈ H. If ϕ is injective, such a reduction is also equivalent to a section in
the bundle P/ϕ(H) ' P ×L G/ϕ(H), where G acts on the right coset space G/ϕ(H) by
multiplication on the left.

A standard case of gauge reduction from GL(n, R) to O(n) and SO(n) of the tangent
bundle of an n-dimensional manifold M is given by the choice of a Riemannian metric,
resp. a metric plus an orientation. In view of the previous proposition, the reduction
from GL(n, R) to O(n) is equivalent to a section in the bundle of symmetric, positive
definite bilinear forms on T M , whose fiber is the space of positive definite symmetric
matrices. This, in turn, is shown to be a contractible space by the standard Gramm-
Schmidt orthogonalization procedure, and by the obstruction theory, a fiber bundle with
contractible fiber always admits global sections. Thus, the existence of a Riemannian
metric on any manifold is a special case of a more general phenomenon: reducing the
gauge group of a G-bundle to its maximal compact subgroup:
Theorem 3.11 (Cartan-Iwasawa-Malcev). Let G be a Lie group with finitely many
connected components. Then every compact subgroup of G is contained in a maximal
compact subgroup of G. Moreover, all such maximal compact subgroups H are conjugated
inside G, have as many connected components as G has, and the coset space G/H is
contractible.

We conclude that, from a topological viewpoint, the study of principal bundles can
be restricted to the case when the structure (gauge) group is compact.
Exercise 3.12. Define or(M ) := GL(M )/GL+ (n, R) to be the bundle of orientations
on M . Show that or(M ) → M is a two-fold covering, and also a Z2 -principal bundle.
Show that a connected manifold M is orientable iff or(M ) is not connected.
π
Exercise 3.13. Generalize the previous exercise as follows: Let P → M be a G-
principal bundle over a connected manifold M with G a discrete group (a Lie group of
dimension 0). Show that P admits a reduction to a subgroup of G iff the total space P
is non connected.

We also mention here the virtual reduction of a G-bundle E → M : here it is not

the gauge group of E itself that is reduced, but the gauge group of f ∗ E, for a suitably
chosen smooth map f : N → M .
8 FLORIN BELGUN

π π0
Example 3.14. Let P → M be a principal G-bundle. Then π ∗ P → P , as a G-principal
bundle over P , is trivial: Indeed, the diagonal map D : P → P × P induces a global
section in π ∗ P = P ×M P ⊂ P × P .

The virtual reduction will turn out to be a useful tool for computations with charac-
teristic classes, see below.
If the reduced gauge group H is not specified, it is difficult to find out whether
the gauge group G of a bundle E admits reductions (exept for the reduction to the
maximal subgroup, as seen above). If H is expected to be a closed subgroup of G, then
the reduction to H amounts to fixing some extra structure on E, which is therefore a
structure on the standard fiber F extended to the whole space E. When E is a vector
bundle, this structure may be given by a tensor, i.e. a section in some tensor power of
E. To extend such a tensor Tx ∈ ⊗∗ E from a point x ∈ M to a curve on M through x,
and then further to a neighborhood of x and even to the whole manifold M , it would be
useful to have fixed a way to lift curves (a lift of a map f : N → M is a map f˜ : N → E
such that π ◦ f˜ = f ), resp. to know that the endpoint of a lifted curve starting at x
and located in a small neighborhood of x ∈ M or, more generally, for arbitrary curves,
depends only on the endpoint of the curve on M .
The idea of curve lifting leads to the concept of a connection (it connects the fibers
by horizontal – i.e., non-vertical – curves) in a fiber bundle; the independence of the
endpoint of the chosen curve turns out to be a property of the curvature and holonomy
of the connection (for local lifts) and of its monodromy (for global lifts)

4. Connections in a principal bundle. Induced connections

Recall that a Lie group is parallelizable. More precisely, the left invariant vector fields
on G canonically identify the tangent space Tg G in a point g ∈ G to g = Te G. This
identification is in fact a 1-form on G with values in its Lie algebra g:
Definition 4.1. The Maurer-Cartan form ω M C ∈ Λ1 G⊗g on the Lie group G associates
to every vector X ∈ Tg G the element (Lg−1 )∗,g (X) ∈ Te G = g. (recall that Lh : G → G
is the left multiplication with h)

ω M C (X) is constant, for X a vector field on G, iff X is left-invariant. Recall that

the flow of a left-invariant vector field X is Rexp(tXe ) , the right multiplication with the
1-parameter subgroup {exp(tXe ) | t ∈ R} generated by Xe ∈ g, the value at e of the
vector field.
We have that ω M C is left-invariant, i.e.:
(Lg )∗ ω M C = ω M C , ∀g ∈ G.
It is important to know what impact has the right multiplication Rg with some element
g ∈ G on the Maurer-Cartan form:
Proposition 4.2. Let G be a Lie group, g ∈ G and ω M C the Maurer-Cartan form.
Then
(Rg )∗ ω M C = Adg−1 ◦ ω M C .
 GAUGE THEORY 9

Here, the left hand side is the pull-back through the diffeomorphism Rg of ω M C
and the right hand side is the composition of the linear maps Adg−1 : g → g with
ω M C : T G → g.
Because ω M C is left-invariant, it can be defined on the fibers of a G-principal bundle
π
P → M , using the trivializations (that identify the fibers with G), and the induced
forms on Px do not depend on the chosen trivialization, since the transition maps act
on G by left multiplication, which leaves ω M C invariant.
Remark 4.3. The same argument can be used to show that the left-invariant vector
fields on G induce vertical tangent fields (i.e., vector fields that are sections of the
vertical tangent space T ∨ P = ker(π∗ )), associated to every element of the Lie algebra
g. These vector fields will be called fundamental, and we denote by X̄ the fundamental
vector field on P corresponding to the element X ∈ g. The flow of X̄, at time t, is the
right multiplication on P with exp(tX) ∈ G.
Equivalently, X̄ can be defined by the formula
d
d
X̄p := t=0 Rexp(tX) p = |t=0 (p · exp(tX)) , ∀p ∈ P.
dt dt
Because ω M C is invariant to the left action of G on itself, the 1-forms ωφM C : T Px → g
defined by the local trivializations φ : Px → G are, in fact, independent of the choice of
the trivialization φ, and thus there exists a well-defined map ω M C,P : T ∨ P → g, such
that
ω M C,P (X̄) = X, ∀X ∈ g.
ω M C,P is not a 1-form on P (with values in g), but, since the right multiplications
Rg : P → P , g ∈ G, preserve the vertical distribution, one can consider the pull-back
of ω M C,P through Rg . We have then
Lemma 4.4. (Rg )∗ ω M C,P = Adg−1 (ω M C,P ), ∀g ∈ G.

Proof. It is enough to prove the identity above for the case when P is trivial; in fact it
is enough to show
(Rg )∗ ω M C = Adg−1 (ω M C ),
for the Maurer-Cartan form of G. For this, apply both sides of the identity to a left-
invariant vector field X on G. On the right hand side we obtain
Adg−1 Xe ,
and on the left hand side we obtain
ω M C ((Rg )∗ X).
We need thus to show that the vector field Y := (Rg )∗ X is also left-invariant and that
Ye = Adg−1 Xe . The flow of Y is
−1
φYt = Rg ◦ φX
t ◦ (Rg ) = Rg ◦ Rexp(tXe ) ◦ Rg−1 = Rg−1 exp(tXe )g ,
which is exactly the flow of Adg−1 Xe . 
Definition 4.5. A connection in the G-principal bundle P → M is a G-invariant
horizontal distribution on P , i.e. a subbundle H of the tangent bundle of P , such that
(Rg )∗ H = H, ∀g ∈ G.
10 FLORIN BELGUN

Equivalently, a connection is given by a connection 1-form ω ∈ C ∞ (Λ1 P ⊗ g) such

that
(1) ω|T ∨ P = ω M C,P
(2) Rg∗ ω = Adg−1 (ω), ∀g ∈ G.
Indeed, the kernel of ω is then a G-invariant horizontal distribution H ω . Conversely,
for a G-invariant horizontal distribution H there exists a unique connection 1-form ω H
such that ker ω H = H (just define ω H to be equal to ω M C,P on T ∨ P and equal to zero
on H).

4.1. Induced connections on associated bundles. Let P → M be a G-principal

bundle and ρ : G × F → F a G action. For H a connection on P , we can construct a
horizontal distribution H E on the associated bundle E := P ×ρ F as follows: H E :=
pr∗ (H × 0F ), where 0F ⊂ T F is the trivial distribution on F . H × 0F ⊂ T P × T F is
then a n-dimensional subspace of T (P × F ) which is transversal to the kernel of the
projection pr : P × F → P ×ρ F , and thus H E is a N -dimensional distribution on
P ×ρ F . This distribution is also horizontal (i.e., transversal to the vertical tangent
space of E), because H × 0F is transversal to T ∨ P × T F and T ∨ E = pr∗ (T ∨ P × T F ).
Note, however, that not every horizontal distribution on E is induced by a connection
on P . The condition that such a horizontal distribution must satisfy can be only seen
in a local gauge (trivialization):
Lemma 4.6. Let (U, ψU ) be an atlas of trivializations of a G-principal bundle P . A
connection H on P (or on any associated bundle E = P ×ρ F ) is given by a family of
1-forms with values in g, ω U ∈ C∞ (Λ1 U ⊗ g) such that
ω V = Adg−1 (ω U ) + gU∗ V ω M C . (2)
UV

Proof. A trivialization is equivalent to a local section σU : U → P . We define ω U :=

σU∗ (ω H ). We use σV = σU · gU V on U ∩ V and obtain, by differentiation,
(σV )∗,x (X) = (RgU V )∗,σU (x) ((σU )∗,x (X)) + ((LgU V (x)−1 )∗ ((gU V )∗,x (X))σU (x) , (3)

where the second term is the fundamental vector field that corresponds to the ele-
ment in g (identified with TgU V (x) G by the corresonding left multiplication) defined by
(gU V )∗,x (X).
The equation (2) follows by applying ω H to the equation above. 
Remark 4.7. We can retrieve the connection form ω out of a family (U, ωU ) as above:
Indeed, it is enough to specify the kernel H of ω along the values of a local section σU ;
then, we define H on π −1 (U ) by using the G-invariance. To define H, it is necessary
and sufficient to define the horizontal lift of every vector X on the basis M :
Proposition 4.8. Let x ∈ U ⊂ M and σU the corresponding local section in P over U .
We define
 
X̃ := (Rg )∗ (σU )∗,x (X) − ωU (X)σ(x) ∈ TσU (x)·g P, ∀X ∈ Tx M, ∀g ∈ G (4)

the horizontal lift of the vector X in the gauge (trivialization) σU .

 GAUGE THEORY 11

Then X̃ is independent of the choice of a trivialization. Therefore, the horizontal

distribution
Hσ(x) := {X̃ | X ∈ Tx M }
is a well-defined connection on P .

Proof. It suffices to check that X̃σV (x) , computed in the gauge σV , coincides with the
value at σV (x) = σU (x) · gU V (x) of the horizontal lift computet in the gauge σU . From
(3) it follows that the difference of the first (the non-vertical) terms in the expressions
of X̃ in these two gauges, cf. (4), is vertical and coincides with the value at σV (x) of
the fundamental vector field
ωV (X) − ωU (X),
which is the difference of the vertical (correction) terms in the expressions of X̃, cf.
(4), in the gauges σV , resp. σU . X̃ is thus well-defined independently of the choice of
trivialization.
H is thus well-defined as well. By definition (4), H is G-invariant, so it is a connection
on P . 

Recall that a gauge transformation φ is an automorphism of the G-principal bundle

P , i.e., a G-equivariant automorphism of P . Equivalently, it is a section in G = P ×AD G,
where AD : G×G → G is the action of G on itself by conjugation: AD(g, a) := g ·a·g −1 .
Indeed, the map
M 3 x 7→ [p, gφ(p),p ] ∈ (P ×AD G)x ,
where p ∈ Px and gp0 ,p ∈ G, for p, p0 ∈ Px , is the unique element of G such that
p0 = p · gp0 ,p (one can say, gp0 ,p is the “quotient” of p0 by p), is a well-defined section in
P ×AD G as claimed.
In fact, we have constructed above an AD-equivariant map φ̄ : P → G by the formula
φ̄(p) = gφ(p),p . The equivariance condition:
φ̄(p · g) = ADg−1 (φ̄(p)), ∀g ∈ G, ∀p ∈ P
is precisely the one that makes that the class [p, φ̄(p)] ∈ P ×AD G is independent of the
choice of p ∈ Px and thus defines a section in this bundle.
A gauge transformation can be thus expressed by one of the following equivalent
objects:
(1) A G-equivariant map φ : P → P that induces the identity on the basis M
(2) A map φ : P → P that looks in local trivializations as the left multiplication on
the fibers with a G-valued function on the base
(3) An AD-equivariant map φ̄ : P → G
(4) A section in G = P ×AD G.
(5) if G is connected, a map φ : P → P that maps each fiber Px into itself and
preserves the fundamental vector fields X̄, X ∈ g.
Note that a gauge transformation also induces a transformation on every associated
bundle: the second point above realizes directly such a transformation, but also the
point 4 (or 3): The fibers Gx , x ∈ M , of G are Lie groups, identified up to conjugation
12 FLORIN BELGUN

with G. As such, they naturally act on the corresponding fibers Ex = (P ×ρ F )x of

associated bundles E = P ×ρ F :
Gx × Ex 3 ([p, a], [p, f ]) 7→ [p, ρ(a, f )] ∈ Ex .
That the above map is well-defined action of Gx on Ex is an easy exercise.
Example 4.9. If E → M is a real vector bundle, the associated principal bundle
P is the bundle GL(E) of isomorphisms (called frames) f : Rn → Ex , x ∈ M . Here
n := dim Ex is the rank of E. The bundle G is then the bundle Aut(E) of automorphisms
of E, and the adjoint bundle Ad(P ) to GL(E) is the bundle End(E) of endomorphisms
of E. As a set, Aut(E) ⊂ End(E).
If E is a complex vector bundle of complex rank n, then the bundles P := GLC (Cn , E),
G := AutC (E) and Ad(P ) := EndC (E) are defined analogously.

Two objects are considered to be gauge equivalent iff there is a gauge transformation
that sends one object into another. A condition is called gauge invariant if it is satisfied
by a whole class of gauge equivalent objects.
Example 4.10. The zero set of a section in a vector bundle is a gauge invariant ob-
ject/condition: if a section s in E vanishes on the set A ⊂ M , then every section of E
which is gauge-equivalent to s vanishes on A.

Ther basic idea of gauge theory is to consider either gauge-invariant conditions or to

consider the classes of objects up to gauge equivalence. In the following, we investigate
the transformation law of connections w.r.t. gauge transformations:
Proposition 4.11. Let φ : P → P be a gauge transformation and let ω be a connection
form on the G-bundle P → M . The induced connection by φ on P is the connection
form
ω 0 = Adφ̄−1 (ω) + (φ̄)∗ ω M C . (5)

Proof. Let X ∈ Tx M be a vector on M and let σ be a local section in P around

x. Then Y := σ∗,x (X) ∈ Tσ(x) P is a horizontal vector (i.e., it is not vertical). Then
σ 0 := φ ◦ σ = σ · φ̄ is another local section in P around x and
ω 0 (Y ) = (σ 0 )∗ ω 0 (X) = Adφ̄(σ(x))−1 (ω(Y )) + (φ̄∗ ω M C )(Y ),
where the last equality follows from (2), where we put σU := σ, σV := σ 0 and gU V := φ̄◦σ.
But this implies the claimed equality (5) for the argument Y .
Considering now all local sections around x, such that σ(x) = p, we obtain that the
equality (5) holds for all horizontal (i.e., non-vertical) arguments Y ∈ Tp P rTp∨ P , hence
for all vectors in Tp P . 

Note that a 1-form α ∈ C∞ (P ) ⊗ g that satisfies

Rg∗ α = Adg−1 α and α|T ∨ P = 0
corresponds to a section in Λ1 M ⊗ AdP .

4.2. Pull-back of a connection. Let f : N → M be a smooth map and P → M

a G-principal bundle with a connection ω ∈ Λ1 P ⊗ g. Let F : f ∗ P → P be the (G-
equivariant) map between the pull-back of P and P . Then F ∗ ω ∈ Λ1 (f ∗ P ) ⊗ g is a
connection on f ∗ P .
 GAUGE THEORY 13

4.3. Connections on vector bundles. Covariant derivatives.

Definition 4.12. Let E → M be a vector bundle (real or complex). A covariant deriv-
ative on E is a linear map D : C ∞ (E) → C ∞ (Λ1 M ⊗R E) such that
D(f s) = f D(s) + df ⊗ s, ∀S ∈ C ∞ (E), ∀f : M → R(C). (6)
Remark 4.13. Note that a 1-form α ∈ C∞ (P ) ⊗ g that satisfies
Rg∗ α = Adg−1 α and α|T ∨ P = 0
corresponds to a section in Λ1 M ⊗ AdP . as above, by replacing Λ1 M with its com-
plexification, and taking the tensor product between Λ1 M ⊗ C and E over C; then, the
function f can be complex-valued.

The property (6) is called the Leibniz rule, and is equivalent to the fact that the
symbol of D, seen as a first order linear differential operator, is the identity (see below).
Proposition 4.14. Let P → M be a G-principal bundle and ρ : G → GL(V ) a linear
representation of G. Then every connection H on P induces a covariant derivative DH
on the vector bundle E := P ×ρ V .
Conversely, let D be a covariant derivative on the (real or complex) vector bundle
E → M , and denote by GL(E) := Hom× (V, E), for V the trivial bundle with fiber
isomorphic to the fibers of E, the set of (real or complex) isomorphisms (or invertible
homomorphisms) between V and E, in other words the bundle of frames of E. Then D
induces a connection H D on this principal GL(V )-bundle.

Here, V stands for Rn , if E is a rank n real vector bundle, and for Cn , if E is a

complex vector bundle of complex rank n. The field R or C is thus “intrinsic” in the
notations GL(V ), End(E), etc.

Proof. Let s be a section of E = P ×ρ V . Take σ : U → P a local section in P . Then s

determines a map w : U → V such that s(x) = [σ(x), w(x)], ∀x ∈ U .
For the action ρ : G → GL(V ) we define, by differentiation at e ∈ G, the derived Lie
algebra action ρ̄ : g → End(V ), that satisfies
[ρ̄(X), ρ̄(Y )] = ρ̄([X, Y ]), ∀X, Y ∈ g,
where the bracket on the left hand side is the commutator in End(V ).
We set then
D(s) := [σ, IdΛ1 M ⊗ ρ̄(ω σ )(w) + dw]. (7)
σ ∗
Here, ω : −σ ω is a 1-form on U with values in g, and ρ̄ applies only to this latter
part, so we finally get a 1-form on U with values in V , as is dw; the pairing with σ of
such a 1-form with values in V yields thus a 1-form with values in P ×ρ V = E.
We check now that the resulting 1-form with values in E is independent on the choice
of σ, using (2): Suppose σ 0 = σ · φ is another local section, with φ : U → G. Then the
corresponding V -valued function w0 is equal to ρ(φ−1 )(w) and thus
0
[σ 0 , ρ̄(ω σ )(w0 ) + dw0 ] =
= [σ · φ, ρ̄(Adφ−1 (ω σ )(ρ(φ−1 )(w) + ρ̄(φ∗ ω M C (w0 ) − ρ̄(φ∗ ω M C )(w0 ) + ρ(φ−1 )(dw)] =
= [σ, ρ̄(ω σ )(w) + dw],
thus Ds = D[σ, w] is independent on the choice of section σ in P .
14 FLORIN BELGUN

The Leibniz rule follows immediately from the one for dw.
Conversely, suppose D is a covariant derivative on a vector bundle E.
Lemma 4.15. Let E → M be vector bundle with a covariant derivative D and e ∈ Ex ,
x ∈ M . Then there exists a section s of E such that
(1) s(x) = e and
(2) (Ds)x = 0.

Proof. Let s be a section of E such that s(x) = e. Consider a local chart φ : U → Rn

around x ∈ U such that φ(x) = 0 and let x1 , ..., xn be the coordinate functions on
U associated to this chart. Then dx1 , ..., dxn is a local frame of Λ1 M around x, thus
∃e1 , ..., en ∈ Ex such that
X n
(Ds)x = dxi ⊗ ei .
i=1

Let s1 , ..., sn be local sections in E around x such that si (x) = ei , ∀i ∈ {1, ..., n} (these
sections can be extended toP global sections of E using a cut-off function on M ). Then
the local section s := S − ni=1 xi si (also extended to a global section using a cut-off
0

function to extend the functions x1 , ..., xn to M ) satisfies:

(1) s0 (x) = s(x) and P
(2) (Ds0 )x = (Ds)x − ni=1 (dxi ⊗ si )x + xi (Dsi )x = 0,
since xi all vanish at x. 

We conclude that, for every point f ∈ GL(E)x , there exists a local frame σ : U →
GL(E) such that σ(x) = f and, ∀v ∈ V , σ(v), as a local section in E, satisfies
D(σ(v))x = 0. Then we define the following horizontal distribution H E ⊂ T GL(E)
by the condition:
HfE := {σ∗,x (X) | X ∈ Tx M, σ ∈ C∞ (E) : σ(x) = f and (D(σ(q))x = 0, ∀q ∈ V }.
If σ is a local section in GL(E) such that D(σ(q))x = 0, ∀q ∈ V , then clearly σ ◦ g has
the same property, ∀g ∈ GL(V ). Since GL(E) 3 f → f ◦ g ∈ GL(E) is the right action
of GL(V ) on GL(E), the horizontal distribution H E is GL(V )-invariant, and thus a
connection on GL(E). 
Remark 4.16. The covariant derivative of a section s : M → E at x does not depend
on the values of s outside an open set U 3 x (for every such an open set; this means
that (Ds)x depends only on the germ at x of the section s): indeed, if s0 − s is a section
of E that vanishes on a neighborhood U of x, then s0 − s = f (s0 − s), where f : M → R
is a function that vanishes on a open set U0 ⊂ U , still containing x, and being equal
to 1 on M r U . Then (D(s0 − s))x = dfx ⊗ (s0 − s)(x) + f (x)(D(s0 − s))x = 0. This
argument allows us to consider the covariant derivatives of local sections in E, because
the extension of a local section to a global one does not influence the values of its
covariant derivative on the original definition domain of the section.

Terminology: A section σ of a bundle E → M with a connection (covariant deriva-

tive) is parallel at a point x ∈ M iff σ∗,x (T M ) is the horizontal space of the connection,
resp. (Dσ)x = 0.
 GAUGE THEORY 15

Like for the connections in principal bundles, we can pull-back covariant derivatives
on f ∗ E, for f : N → M a smooth map and E → M a vector bundle with a covariant
derivative D.
We can also construct covariant derivatives on tensor bundles of E, i.e. on canonical
subbundles in ⊗E, like Λ∗ E, End(E) ' E ∗ ⊗ E, starting from a covariant derivative on
E.
More generally, if E1 , E2 are vector bundles over M with covariant derivatives D1 , D2 ,
then we can define the covariant derivative D1 + D2 on E1 ⊕ E2 (we differentiate the
components) and D := 1E1 ⊗ D2 + D1 ⊗ 1E2 on E1 ⊗ E2 by the formula:
D(s1 ⊗ s2 ) := D1 s1 ⊗ s2 + s1 ⊗ D2 s2 , ∀si ∈ C∞ (Ei ), i = 1, 2.
Indeed, the decomposable sections of type s1 ⊗s2 linearly span the space of all sections of
E1 ⊗ E2 and D can be extended by linearity to define a covariant derivative on E1 ⊗ E2 .

We have seen that, for every element of a fiber bundle with connection, there exists
a local section in this bundle, extending this element, that is parallel at that point.
Asking that the section is parallel on a neighborhood is, on the other hand, much more
restrictive. The obstruction to the existence of such a parallel extension is the curvature

5. Curvature of a connection

Lemma 5.1. Let ω M C be the Maurer-Cartan form of a Lie group G. Then the following
structure equation holds:
dω M C (X, Y ) + [ω M C (X), ω M C (Y )] = 0, ∀X, Y ∈ g.

The proof is immediate.

In order to write the equation above without arguments, we introduce the following
notation
Definition 5.2. Let ω1 , ω2 be k-, resp. l-forms on M with values in a Lie algebra g.
The wedge product ω1 ∧ ω2 is a k + l-form on M with values in g, defined by
1 X  
(ω1 ∧ω2 )(X1 , ..., Xk+l ) := sign(σ) ω1 (Xσ(1) , ..., Xσ(k) ), ω2 (Xσ(k+1) , ..., Xσ(k+l) ) ,
k!l! σ∈S
k+l

for all X1 , ..., Xk+l ∈ T M .

Remark 5.3. If g is non-abelian, then ω ∧ ω may be non-zero even if ω has odd degree.
In fact, ω1 ∧ ω2 = (−1)kl+1 ω2 ∧ ω1 so the wedge product is actually precisely then
commutative when ω1 and ω2 have odd degree, and anti-commutative if one of the
factors has even degree. On the other hand, the wedge product on g-valued forms is
not necessary associative.

With this notation, [ω(X), ω(Y )] = 12 (ω ∧ ω)(X, Y ), thus the structure equation for
the Maurer-Cartan form can be written:
1
dω M C + ω M C ∧ ω M C = 0. (8)
2
16 FLORIN BELGUN

Definition 5.4. Let ω ∈ C∞ (Λ1 P ⊗ g) be a connection form on a G-principal bundle.

The curvature form of ω is the two-form Ω ∈ C∞ (Λ2 P ⊗ g) defined by
1
Ω := dω + ω ∧ ω.
2
Because ω|T ∨P coincides with the Maurer-Cartan form on the fibers, then (8) implies:
Proposition 5.5. The curvature form Ω vanishes on T ∨ P , thus Ω can be seen as a
section in Λ2 M ⊗ Ad(P ).

Proof. From (8), Ω(X̄, Ȳ ) = 0 for X̄, Ȳ fundamental vector fields. Let us consider now
one fundamental vector field X̄, X ∈ g, and a horizontal vector field Ỹ , Y vector field
on M . Then the quadratic term in ω vanishes on the pair (X̄, Ỹ ), thus
Ω(X̄, Ỹ ) = dω(X̄, Ỹ ) = X̄.(ω(Ỹ )) − Ỹ .(ω(X̄)) − ω([X̄, Ỹ ]).
the first two terms vanish, and the last term, ω applied to the Lie bracket of the
fundamental field X̄ and the horizontal vector field X̃ vanishes as well, because the flow
of X̄ acts by right multiplication on P , and thus sends Ỹ into itself, thus [X̄, Ỹ ] = 0.
So Ω vanishes on T ∨ P . Moreover, it is G-equivariant.
Lemma 5.6. Let α be a k-form on P with values in V , for ρ : G → GL(V ) a linear
representation, such that
(1) α(X, ...) = 0, ∀X ∈ T ∨ P , and
(2) Rg∗ α = ρ(g −1 )(α), ∀g ∈ G.

Then α can be identified with a section α in Λk M ⊗ P ×ρ V , such that

[σ, σ ∗ α] = α, ∀σ local section in P,
where [σ, β], for a k-form β on M with values in V and σ a local section in P , is the
local section in Λk M ⊗ P ×ρ V defined by
Tx M 3 X1 , .., Xk 7→ [σ(x), β(X1 , ..., Xk )] ∈ P ×ρ V.

Proof. The proof is similar to the case of degree 0 (that a G-equivariant function frm P
to V is equivalent to a section in the corresponding associated bundle): Let X1 , ..., Xk ∈
Tx M and p ∈ Px . Take any lifts Y1 , ..., Yk ∈ Tp P and compute α(Y1 , ..., Yk ) (the result
does not depend on the chosen lifts because T ∨ P ⊂ ker α). This defines an element
[p, α(Y1 , ..., Yk )] ∈ P ×ρ V , and, if we consider p0 := p · g ∈ Px , g ∈ G, then if we
consider Yi0 := (Rg )∗ Yi , i ∈ {1, ..., k}, as lifts of X1 , ..., Xk in p0 , then we immediately
get [p0 , α(X10 , ..., Xk0 )] = [p, α(Y1 , ..., Yk )] using the G-equivariance of the k-form α. 

It follows that Ω can be seen as a 2-form Ω on M with values in Ad(P ). 

Remark 5.7. In the case of a connection on GL(E) defined by a covariant derivative
on E, the adjoint bundle Ad(GL(E)) is just the bundle End(E) of endomorphisms of
E.

We define now the exterior covariant derivative on E-valued forms, where E is a

vector bundle with a linear connection (covariant derivative):
 GAUGE THEORY 17

Let E → M be a vector bundle (real or complex) with a connection (covariant

derivative) ∇. The operator ∇ : C∞ (L0 (E)) → C∞ (L1 (E)) induces, by the following
formula (based on the usual formula for the exterior differential), the following operators:
d∇ : C ∞ (Λk (E)) → C ∞ (Λk+1 (E)),
k
X  
∇
d (α)(X0 , . . . , Xk ) := (−1)j ∇Xj α(X0 , . . . , X̂j , . . . , Xk ) +
j=0
X (9)
+ (−1)j+l α([Xj , Xl ], X0 , . . . , X̂j , . . . , X̂l , . . . , Xk ),
0≤j<l≤k

where the hat ˆ indicates a missing term. The proof of the following proposition is
straightforward (by induction):
Proposition 5.8. Let α ∈ C ∞ (Λk (E)) and β ∈ C ∞ (Λl M ). Then
d∇ (β ∧ α) = dβ ∧ α + (−1)l β ∧ d∇ α.
Moreover, this property together with the fact that d∇ = ∇ on C∞ (E) uniquely deter-
mines the operator d∇ : C ∞ (Λ∗ (E)) −→ C ∞ (Λ∗ (E)).
Remark 5.9. If ∇ is a trivial connection (for example, given by a (local) frame), then
d∇ is the usual exterior derivative of forms with values in a (fixed) vector space, therefore
(d∇ )2 = 0.

In general, (d∇ )2 6= 0.
Definition 5.10. Let ∇ be a connection on a vector bundle E → M . Its curvature
tensor R∇ is a 2–form with values in End(E), defined by K ∇ s := (d∇ )2 s, more precisely
∇
KX,Y s = ∇X ∇Y s − ∇Y ∇X s − ∇[X,Y ] s, X, Y ∈ T M, s ∈ C ∞ (E).
Remark 5.11. Even if (d∇ )2 is not zero, it is a zero-order differential operator (i.e. a
tensor) - see below.

Using Proposition 5.8 and the fact that the space of sections (sheaf) C ∞ (Λ∗ (E)) is
isomorphic to the tensor product C ∞ (Λ∗ M ) ⊗ C ∞ (E) (the tensor product is over the
ring C ∞ (M ) of real/complex–valued functions on M ), we get
Proposition 5.12. Let α ∈ C ∞ (Λ∗ (E)). Then (d∇ )2 α = K ∇ ∧ α.

Here we have used the following notation: for a k-form α with values in E, and a
l-form η with values in End(E), we define their exterior product by taking the tensor
product of the following maps:
∧
Λk M ⊗ Λl M → Λk+l M
η ⊗ ξ 7→ η ∧ ξ,
E ⊗ End(E) → E
s ⊗ A 7→ A(s).
If we consider two forms with values in End(E), A ∈ C∞ (Λk (End(E))) and B ∈
C∞ (Λl (End(E))), we consider their wedge product
A ∧ B ∈ Ωk+l (End(E))
18 FLORIN BELGUN

by tensoring the wedge product of forms on M with the commutator (linear) map:
End(E) ⊗ End(E) → End(E)
A⊗B 7→ AB − BA.

We have the following differential Bianchi identity (sometimes called the second
Bianchi identity in the case of connections on the tangent bundle):
Proposition 5.13. The curvature tensor K ∇ ∈ C∞ (Λ2 (End(E))) of a connection ∇
on E → M is d∇ -closed.

Proof. Let us compute (d∇ )3 s, for a section s in E:

(d∇ )3 s = d∇ ((d∇ )2 s) = d∇ K ∇ s = d∇ K ∇ s + K ∇ ∧ ∇s.

On the other hand,

(d∇ )3 s = (d∇ )2 ∇s = K ∇ ∧ ∇s,
therefore d∇ K ∇ = 0. 

The differential Bianchi identity has the following formulation in terms of connection
1-forms:
Proposition 5.14. Let ω be a connection 1-form on the G-principal bundle P → M
and Ω = dω + 21 ω ∧ ω its curvature form. Then
dΩ + ω ∧ Ω = 0.

The proof follows directly by differentiating the formula defining Ω and using that
ω ∧ (ω ∧ ω) = 0 (from the Jacobi identity on g).
We need to show that the curvature tensor K is canonically determined by the cur-
vature form Ω:
Proposition 5.15. Let P → M be a G principal bundle with a connection form ω, its
curvature form Ω and let ∇ω : C∞ (E) → C∞ (Λ1 M ⊗ E) be the covariant derivative
associated to ω for an associated vector bundle E := P ×ρ V → M , for ρ : G → GL(V )
a linear representation of G on V .
Then ∀X, Y ∈ Tx M , the curvature tensor
KX,Y [p, v] = [p, ρ̄(Ω(X̃, Ỹ ))(v)], ∀[p, v] ∈ Ex .

Proof. We use the description of Ω as a 2-form on M with values in Ad(P ), as in

Proposition 5.5, and then proceed as in the proof of Proposition 4.14, from which we
also use the notations σ is a local section in P and w : M → V such that x 7→
[σ(x), w(x)] ∈ Ex is equal to a fixed section s ∈ C ∞ (E):
∇ωX [σ, w] = [σ, ρ̄(σ ∗ ω(X))(w) + dw(X)].
Then we obtain
∇ωX ∇ωY s = [σ, ρ̄(σ ∗ ω(X))(ρ̄(σ ∗ ω(Y )))(w)) + ρ̄(σ ∗ ω(X))(dw(Y )) + X.ρ̄(σ ∗ ω(Y ))(w) + X.dw(Y )].
Now we use that X.σ ∗ ω(Y ) − Y.σ ∗ ω(X) − σ ∗ ([X, Y ]) = σ ∗ dω(X, Y ) and that
ρ̄(σ ∗ ω(X)) ◦ ρ̄(σ ∗ ω(X)) − ρ̄(σ ∗ ω(X)) ◦ ρ̄(σ ∗ ω(Y )) = ρ̄(σ ∗ [ω(X), ω(Y )]) = 21 ρ̄(σ ∗ (ω ∧ ω)(X, Y )).
Also note that X.dw(Y ) = Y.dw(X) − dw([X, Y ]) = d(dw)(X, Y ) = 0. Then we get
KX,Y s = ∇ωX ∇ωY s − ∇ωY ∇ωX s − ∇ω[X,Y ] s = [σ, ρ̄(σ ∗ Ω(X, Y ))(w)].
 GAUGE THEORY 19


Remark 5.16. If P ⊂ Aut(E) (and G ⊂ GL(V )), we can simply write K = Ω.
However, it is not true that every Lie group G can be emedded in a linear group (the
universal covering of SL(2, R) is such an counterexample). For a general Lie group G,
ω
we need to distinguish between K = (d∇ )2 and Ω. Indeed, the bundle Ad(P ) where
Ω takes values, reflects the structure of E to a better extent as End(E): a U(n) (or
unitary) connection on a rank n complex vector bundle is a complex linear connection
(i.e., the covariant derivative is complex-linear) and metric, i.e., there exist a symmetric,
positive definite, Hermitian tensor h ∈ C ∞ (E ⊗R E), (Hermitian means that such that
h(iX, iY ) = h(X, Y ) for all X, Y ∈ E, where iX is the multiplication with the complex
number i on the fibers of E) that is parallel w.r.t. the covariant derivative.) In this
case, P = U(E, h), the space of unitarz frames of E, Ad(P ) can be identified with the
subbundle of End(E) of antihermitian endomorphisms of E.
Remark 5.17. Although such a statement is false for Lie groups (see above), it is in
general true that every Lie algebra admits a finite-dimensional faithful representation
(i.e., ∃ρ : G → GL(V ) s.t. ρ̄ : g → End(V ) is injective (thus ρ is an immersion)), and
thus there is no loss of information on Ω if we consider the curvature tensor K for the
bundle associated to such a representation ρ.

From now on, we will consider that E is a vector bundle with a covariant derivative
that comes from a G-connection. The difference of two such covariant derivatives comes
thus from a section θ of Λ1 M ⊗ Ad(P ), and is precisely the image of θ through ρ̄ in
Λ1 M ⊗ End(E).
Let us compute the change of the curvature tensor when the connection changes: Let
0
∇ := ∇ + θ be another connection on E, where θ ∈ C∞ (Λ1 (E)). Then d∇ = d∇ + θ ∧ ·,
0

therefore
0
(d∇ )2 s = d∇ (d∇ s + θs) + θ ∧ (d∇ s + θs) = (d∇ )2 s + d∇ θs − θ ∧ ∇s + θ ∧ ∇s + θ ∧ θs.
the last term is not zero, even if θ is a 1–form with values in End(E); this is because
End(E) is not commutative. We get
0 1
R∇ = R∇ + d∇ θ + [θ, θ]. (10)
2
Remark 5.18. The gauge transformation law of the curvature form Ω is, for a gauge
transformation φ̄ : M → AD(P )
Ω0 := φ∗ Ω = Adφ̄−1 Ω,
which is, of course, the transformation law of a section in Λ2 ⊗ Ad(P ).

Note that, if E = E1 ⊕ E2 , and the connection ∇ on E is the sum of a connection

∇1 on E1 and ∇2 on E2 , then the curvature tensor K ∇ admits a block decomposition,
i
being equal to K ∇ on End(Ei ), i = 1, 2 and zero on the “mixed blocks” E1∗ ⊗ E2 and
E2∗ ⊗ E1 .
1 2
Also, if ∇ = 1E1 ⊗∇2 +∇1 ⊗1E2 on E1 ⊗E2 , then we have K ∇ = K ∇ ⊗1E2 +1E1 ⊗K ∇ .
We will use now the curvature of a connection in E to define cohomology classes in
M , that turn out to depend only on the vector bundle E on M and not on the chosen
connection.
20 FLORIN BELGUN

6. Yang-Mills theory

Let (M, g) be a compact Riemannian manifold and let P → M be a G-principal

bundle over M , and suppose there is a fixed Ad-invariant metric h0 on g, i.e., h0 (X, Y ) =
h0 (Ada X, Ada Y ) ∀a ∈ G, X, Y ∈ g (this is possible for all compact Lie groups G). Then
h defines a metric h on the vector bundle AdP , i.e. a positive definite section h in the
vector bundle S 2 (AdP ∗ ), and for every G-connection on P , this section is parallel (or
covariant constant) for the induced covariant derivative.
The Yang-Mills theory looks for connections ∇ on P that are minimize (or, more
generally, are critical points of the Yang-Mills functional, see below) the total norm of
the curvature Z
kK ∇ kvolg .
M
Here, the pointwise norm kKx∇ k ∈ Λ Mx ⊗ AdPx is determined by the metric g on Λ2 M
2

and by h on AdP .
For non-compact groups and (M, g) pseudo-Riemannian, we need to reformulate the
above integral: on g, we replace the positive-definite scalar product h0 with a non-
degenerate Ad-invariant symmetric form B (not all Lie groups admit such a form, how-
ever a large class – including GL(n, R), GL(n, C), SO(n, 1), the latter being the group
of oiented automorphisms of the Minkowski space – does), and we replace the metric
induced by g on Λk M with a non-degenerate scalar product. For the latter, it is useful
to introduce the Hodge star operator
∗ : Λk M → Λn−k M
on an oriented, pseudo-Riemannian n-manifold. As the definition is fiberwise, it is
enough to define the ∗ operator for an oriented vector space V , on which a scalar
product g of signature (p, q) (where p + q = n) has been fixed:
Let e1 , ..., ep , ep+1 , ..., en be an oriented orthonormal basis of V ∗ , such that

 0, i 6= j
g(ei , ej ) = 1, i = j ≤ p
−1, i = j > p


The metric volume element is then voln := e1 ∧ ... ∧ en ∈ Λn V ∗ .

Then, for I = {i1 , ..., ik } ⊂ {1, ..., n}, we define eI := ei1 ∧ ... ∧ eik and we set

I J 0, I=6 J
g(e , e ) := ,
ε(i1 )...ε(ik ), I = J
where ε(i) := g(ei , ei ). We define then the Hodge ∗ operator by the property
g(∗α, β)volg = α ∧ β, ∀α ∈ Λk V ∗ , ∀β ∈ Λn−k V ∗ .
It is easy to compute
Proposition 6.1. The Hodge ∗ operator satisfies:
(1) ∗(ei1 ∧ ... ∧ eik ) := ε(i1 )...ε(ik )ε(i1 , ..., ik , j1 , ..., jn−k )ej1 ∧ ... ∧ ejn−k ,
∀1 ≤ i1 < ... < ik ≤ n, where j1 < ... < jn−k are the remaining indices, and
ε(i1 , ..., ik , j1 , ..., jn−k ) is the signum of this permutation of {1, ..., n}.
(2) ∗(∗α) = (−1)k(n−k)+p α, ∀α ∈ Λk V ∗ .
 GAUGE THEORY 21

If we have a vector bundle E → M , then we define, on the fiber at x ∈ M , the ∗

operator on Λk Mx ⊗ Ex as the tensor product of the ∗ operator defined above (with
V := Tx M ) with the identity of Ex .
Let B0 : g ⊗ g → R (or C, in case g is a complex Lie algebra) be a symmetric, non-
degenerate form (for semisimple Lie algebras g there exist such bilinear forms B0 ; for g
simple such a B0 is unique up to a factor). Then B0 defines a section B in S 2 (Adg ∗ ),
the second symmetric power of the dual AdP ∗ of the adjoint bundle. (Note that, as
B is determined by an invariant element of the representation S 2 g∗ , it is a ∇-parallel
section of S 2 (Adg ∗ ), for any G-connection ∇.)
We define then B(α ∧ β), for α ∈ C ∞ (Λk M ⊗ AdP ), β ∈ C ∞ (Λn−k M ⊗ AdP ), as
follows: consider the wedge pairing
∧ : Λk M ⊗ Λn−k M → Λn M,
and take the tensor product of this pairing with B to obtain a bilinear pairing (sym-
metric if k is even, skew-symmetric if k is odd)
(Λk M ⊗ AdP ) ⊗ (Λn−k M ⊗ AdP ) → Λn M.
Example 6.2. For G := GL(k, R), the trace of the product is a non-degenerate sym-
metric bilinear form on the Lie algebra gl(k, R). If we restrict it to the Lie subalgebra
so(k) of skew-symmetric matrices, it is negative definite
tr(A2 ) = −tr(A · A
t
) = −kAk2 ,
if we denote by k · k2 the Euclidean squared norm on End(Rn ) (the sum of the squares
of all entries of a matrix.)
Example 6.3. For G := GL(k, C), the trace of the product is a non-degenerate sym-
metric bilinear form on the complex Lie algebra gl(k, C) with values in C. If we restrict
it to the Lie subalgebra u(k) of antihermitian matrices, it is real-valued and negative
definite
tr(A2 ) = −tr(A · Ā
t
) = −kAk2 ,
if we denote by k · k2 the Euclidean squared norm on End(Cn ) (the sum of the square
norms of all entries of a matrix.)
Definition 6.4. A connection ∇ on the G-bundle P → M (where an invariant sym-
metric non-degenerate form B ∈ C ∞ (S 2 AdP ∗ ), and a pseudo-Riemannian metric on
the compact, oriented manifold M have been fixed) is said to be Yang-Mills iff it is a
critical point of the Yang-Mills functional
Z
1
Y M (∇) := B(K ∇ ∧ ∗K ∇ ),
2 M
i.e., for every family {∇t }t∈[0,1] of G-connections on P , depending smoothly on t1, and
such that ∇0 = ∇, the derivative

d 
Y M (∇t ) = 0.
dt t=0

Proposition 6.5. A connection ∇ as above is Yang-Mills iff

d∇ ∗ K ∇ = 0.
1because the space of connections is an affine space modelled on C ∞ (Λ1 M ⊗ AdP ), we can always
construct “linear” paths ∇t := ∇ + tθ, for θ some section of Λ1 M ⊗ AdP .
22 FLORIN BELGUN

Example 6.6. If n = 4 and (M, g) is a Riemannian manifold, or a manifold with

signature (2, 2), then ∗2 : Λ2 M → Λ2 M is the identity (see Proposition 6.1), so Λ2 M
splits as the direct sum of the bundles Λ2+ M ⊕ Λ2− M of self-dual (SD) and anti-self-dual
(ASD) 2-forms. The SD forms satisfy ∗α = α, and the ASD forms satisfy ∗α = −α,
so if the curvature K ∇ of ∇ is SD, resp. ASD, then the condition in the Porposition
above is equivalent to the Bianchi identity. Therefore, SD and ASD connections are
automatically Yang-Mills. We well see later that, for Riemannian (M, g) and compact
G, they are the absolute minimizers of the Yang-Mills functional.

7. Chern-Weil theory

Definition 7.1. Let g be the Lie algebra of the Lie group G. An (ad-)invariant polyno-
mial Q : g → C, resp. an (ad-)invariant symmetric multilinear form S : g ⊗ . . . ⊗ g −→
C, such that
Q(Adg A) = Q(A), ∀A ∈ g, g ∈ G,
resp. that
S(Adg A, . . . , Adg Ak ) = S(A1 , . . . , Ak ), ∀Ai ∈ g, g ∈ G.

From a symmetric multilinear form S we get a polynomial QS (A) := S(A, . . . , A) and,

conversely, from a polynomial Q on g, homogeneous of degree k, we get the symmetric
multilinear form
Q 1 dk
S (A1 , . . . , Ak ) := |t =···=tk =0 Q(t1 A1 + · · · + tk Ak ).
k! dt1 . . . dtk 1
Q
It is clear that QS = Q and that Q is Ad-invariant iff S Q is.
Example. The trace, determinant and, more generally, the homogeneous compo-
nents of A 7→ det(Id + A) are Ad-invariant polynomials on the Lie algebra End(V ), for
V a real or complex vector space.
Proposition 7.2. Let G be a connected Lie group. A symmetric multilinear form
S : g⊗k → C is invariant iff
k
X
S(A1 , . . . , [B, Aj ], . . . , Ak ) = 0, ∀B, Ai ∈ g. (11)
j=1

Proof. Let gt := exp(tB) ∈ G, for t ∈ R, B ∈ g. All elements in the neighborhood

of the identity of G can be written like this. Fix A1 , . . . , Ak ∈ g and consider F (t) :=
S(gt A1 gt−1 , . . . , gt Ak gt−1 ). Then note that
d d d
exp(tB) = B exp(tB), thus gt = Bgt and gt−1 = −gt−1 B.
dt dt dt
0
If S is invariant, then F is constant, therefore F (0) = 0, which implies (11). Conversely,
if (11) holds for every Aj and B ∈ g, in particular for gt Aj gt−1 and B, it follows that
F 0 (t) = 0 ∀t ∈ R, therefore
F (t) = F (0) = S(A1 , . . . , Ak ).
This implies that the map g 7→ S(gA1 g −1 , . . . , gAk g −1 ) is constant on a neighborhood
of Id ∈ G (for fixed A1 , . . . , Ak ∈ g). But G is connected and the considered map is
analytic (even polynomial), thus it has to be constant on whole G. 
 GAUGE THEORY 23

Let E → M be a vector bundle with fiber V and (this holds for the other examples of
Lie groups above) let S be an invariant symmetric multilinear form on End(V ). Then
S induces a multilinear bundle map
S E : End(E) ⊗ . . . ⊗ End(E) → C,
given, in a frame f ∈ GL(E), by
S f (A1 , . . . , Ak ) := S(Af1 , . . . , Afk ),
where Afj is the element in End(E) defined by Aj ∈ g and the frame f . The invariance
of S ensures that S f is independent of the frame and hence S E is well-defined.
By using again the sheaf isomorphism
C∞ (Λ∗ (End(E))) ' C∞ (Λ∗ M ) ⊗C∞ (M ) End(E)
we extend S E to a multilinear map
S E : Λ∗ (End(E))⊗k → Λ∗ M,
S E (α1 ⊗ A1 , . . . , αk ⊗ Ak ) := α1 ∧ . . . ∧ αk · S E (A1 , . . . .Ak ),
forall Aj ∈ End(E), αj ∈ Λ∗ M.
We can also define P E (α) := S E (α, . . . , α) ∈ Λkp M , α ∈ Λp (End(E)), where P is the
polynomial associated to the symmetric multilinear map S.
Theorem 7.3. (Chern-Weil) Let E → M be a vector bundle with fiber V , associated
to a G-principal bundle P . Let Q be an invariant polynomial on g of degree k and ∇ a
G-connection on E, with curvature K ∇ . Then the 2k–form QE (K ∇ ) ∈ C∞ (Λ2k M ) ⊗ C
is closed, and its class in the de Rham cohomology group H 2k (M, C) is independent of
∇.

Proof. First we show that QE (R∇ ) is d∇ –closed. This is an immediate consequence of

the following
Lemma 7.4. Let αj ∈ C∞ (Λpj (Ad(P ))) and S an invariant symmetric multilinear form
of degree k. Denote by j := p1 + · · · + pj−1 . We have then
k
X
d (S(α1 , . . . αk )) = S(α1 , . . . , (−1)j d∇ αj , . . . , αk ).
j=1

Proof. We proceed by induction over the sum  of all degrees of the forms αj , that we
suppose to be arranged such that p1 ≤ · · · ≤ pk . For all pj = 0 the claimed formula is
just the covariant derivative of
S E (A1 , . . . , Ak ),
for Aj ∈ Ad(P ), where we note that S E is induced by a constant element of Hom(g⊗k , C),
therefore its derivative vanishes.
Suppose nowPthe claim is true for any αj ∈ C ∞ (Λpj (Ad(P ))) such that the sum of
the degrees is pj =  ≥ 0 and let now αj0 := αj for 1 ≤ j < k and αk0 = αk ∧ β ∈
∞ pk +1
C (Λ (Ad(P ))), where β is a 1–form on M . If we show that the claim holds for
S and for all α1 , . . . αk−1 , αk0 as above and every 1–form β, then (again using that
C ∞ (Λp+1 (Ad(P ))) = C ∞ (Λ1 M ) ⊗ C ∞ (Λp (Ad(P ))) as a sheaf), the claim will be proven
for  + 1.
24 FLORIN BELGUN

Note that
S E (α1 , . . . , αk0 ) = S E (α1 , . . . , αk ) ∧ β,
thus
dS E (α1 , . . . , αk0 ) =
k
X
= (−1)j S E (α1 , . . . , d∇ αj , . . . , αk ) ∧ β + (−1)−1 S E (α1 , . . . , αk ) ∧ dβ =
j=1
k−1
X
= (−1)j S E (α1 , . . . , d∇ αj , . . . , αk0 ) + (−1)k S E (α1 , . . . , d∇ αk0 ).
j=1

To show that the cohomology class defined by P E (R∇ ) is independent of ∇, we

0
need to compute the difference between QE (K ∇ ) and QE (K ∇ ), for a new connection
∇0 = ∇ + θ and show that it is an exact form.
Actually, we will consider a path of connections ∇t between ∇ and ∇0 , for example
t
∇t := ∇ + tθ. Then ∇0 = ∇ and ∇1 = ∇0 . Denote by K t := K ∇ . In order to show
that
QE (K 1 ) = QE (K 0 ) + dβ, (12)
we will show that
d E t
Q (K ) = dβ t ,
dt
and conclude by integration that the class of cohomology of Q(K t ) is constant. More
precisely, (12) holds with
Z 1
β= β t dt
0

Let us compute the derivative in t at t = t0 of QE (K t ): first note that

 
d ∇t 0 1 t
|t=t0 K = lim K − K = lim d (t − t0 )θ + [(t − t0 )θ, (t − t0 )θ] = d∇ 0 θ.
t t t0
dt t→t0 t→t0 2
Therefore
k (j)
d E t
X
∇t0
|t=t Q (K ) = S (K , . . . , d θ, . . . , K t0 ),
E t0
dt 0 j=1
t
but this is, according to the previous lemma (using that d∇ 0 K t0 = 0), equal to dβ t0 ,
where
k
X (j)
to
β := S E (K t0 , . . . , θ , . . . , K t0 ).
j=1

We get thus QE (K 1 ) − QE (K 0 ) = dβ, where

Z 1 X k
!
(j)
β := S E (K t , . . . , θ , . . . , K t ) dt
0 j=1

is the integral in t of the forms β t found above. ChS(∇0 , ∇1 ) := β is called the rela-
tive Chern-Simons form of the connections ∇0 and ∇1 , corresponding to the invariant
polynomial Q of order k. 
 GAUGE THEORY 25

Corollary 7.5. (Chern classes) Let E → M be a complex vector bundle and let Ck be
the homogeneous polynomial of order k in the invariant (non-homogeneous) polynomial
A
expression A 7→ det(Id + 2πi ). Let ∇ be a connection on E. Then the cohomology
classes
cRk (E) := CkE (K ∇ )
 

are real, i.e. the above forms represent cohmology classes cRk (E) ∈ H 2k (M, R), called
the (real) Chern classes of E. Denote by c(E) := 1 + c1 (E) + c2 (E) + ... ∈ H ∗ (M, R)
the total Chern class of E → M .
Remark 7.6. The classes cRk (E) defined above are actually induced by the integer
Chern classes ck (E) ∈ H 2k (M, Z), defined in algebraic topology [2].

Proof. It suffices to consider a particular connection and show that the resulting form
CkE (K ∇ ) is a real 2k–form.
We consider a hermitian metric on E and take ∇ a hermitian connection. Then its
curvature will be a 2–form with values in End(E), actually in the subspace (or rather,
Lie subalgebra) of antihermitian endomorphisms of E. This space is isomorphic, via an
unitary frame f : U × V → E|U (a frame that is a unitary isomorphism on the fibers),
to the space of antihermitian matrices u(V ) ⊂ End(V ).
What we need to show is that Ck (iA) ∈ R, for any antihermitian matrix A. But iA
1 1
is hermitian and Id + 2πi A too, therefore det(Id + 2πi A) is real, and so must be all its
homogeneous components. 
Proposition 7.7. Let T → γ̇ 1 be the tautologic line bundle: ∀p ∈ γ̇ 1 , Tp := p ⊂ C2 . .
Then the integral over γ̇ 1 of first Chern class of T is equal to −1.

This normalization property, together with the following two:

(1) characteristic: For every complex vector bundle E → N and every smooth map
f : M → N , it holds c(f ∗ E) = f ∗ c(E).
(2) multiplicativity: For E, F → M two complex vector bundles, it holds c(E⊕F ) =
c(E) ∧ c(F )
characterize completely the total Chern class, [2].

8. Characteristic classes

Definition 8.1. A (real) characteristic class q associated to the Lie group G is a map2
from the space PG of G-principal bundles on a manifold M with values in the de Rham
cohomology H ∗ (M, R) such that, for every G-bundle P → M and any smooth (contin-
uous) map f : N → M , we have q(f ∗ P ) = f ∗ q(P ).
Remark 8.2. One can define integer or Z2 characteristic classes, by considering ap-
propriate cohomology theories [2]. In fact most characteristic classes that we construct
here are realizations of integer characteristic classes.

It is clear that the Chern-Weil theory constructs characteristic classes out of invariant
polynomials. The theory of classifying spaces shows that the converse is true: every real
characteristic class is obtained from an invariant polynomial by the Chern-Weil theory.
2functor
26 FLORIN BELGUN

We focus thus on determining the ring of invariant polynomials on a Lie algebra g:

Definition 8.3. Let A ∈ u(n) be a antihermitian matrix. Then its eigenvalues are
imaginary, thus the total Chern polynomial
  Y n
1
C(A) := det id + A = (1 + xj ),
2πi j=1

where {2πixj | j = 1, ..., n} is the set of the eigenvalues of A, is a real-valued invariant

polynomial on u(n). Its homogeneous component (as a polynomial in x1 , ..., xn ) of degree
k is X
Ck (A) = xj1 ...xjk ,
1≤j1 <...<jk ≤n
the kth Chern polynomial on u(n).
A A



Exercise 8.4. Show that C1 (A) = tr 2πi
, Cn (A) = det 2πi
and (C12 − 2C2 )(A) =
A 2


tr 2πi .
Theorem 8.5. The ring of invariant real-valued polynomials on u(n) is isomorphic to
R[u(n)]U(n) ' R[x1 , ..., xn ]sym ' R[C1 , ..., Cn ],
where R[x1 , ..., xn ]sym is the ring of symmetric polynomials in the indeterminates x1 , ..., xn .

One of the main invariant polynomial (power series) on u(n) is the Chern character
poynomial (power series)
n
! n
!
X  X 1 
CHkn := exj  = (1 + xj + xj 2 + ...)  .
 
j=1

j=1
2 
deg≤k (deg≤k)

The subscript (deg ≤ k) means that we truncate the corresponding power series at
degree k (CH n or CH∞
n
is the non truncated power series). For example:
1
CH2n = n + C1 + (C12 − 2C2 ).
2
Exercise 8.6. Show that, if A ∈ u(n), then
CH(A) = tr(exp(A)),
An
P
where the exponential is defined by the series exp(A) = k∈N n! (for n = 0 we set
A0 := Id).

When applied to the curvature of a connection on a G-bundle over M , these invariant

polynomials produce, via the Chern-Weil theory, characteristic classes, i.e. cohomology
classes on M . Here, we see that the components of degrees k > n/2 of an invariant
polynomial are irrelevant, since H 2k (M, R) = 0. This means that, for purposes of
Chern-Weil theory, truncating power series (as CH) at a sufficiently high degree is not
an issue, we can as well work with the full power series as if they were polynomials.
The Chern character power series defines, by Chern-Weil theory, the Chern character
of a complex vector bundle E over a manifold M :
ch(E) := [CH(K ∇ )] ∈ H ∗ (M, R),
for ∇ some (unitary) connection on E. (Note: the Chern character power series can
be expressed in terms of the Chern polynomials, which can be defined independently
 GAUGE THEORY 27

if ∇ is a unitary connection or not, but the resulting cohomology class is independent

of the connection, in particular we can choose ∇ to be unitary, which implies that the
corresponding cohomology class id real. It is, in fact, even integer, see [2])
The Chern character polynomial has the following properties:
Proposition 8.7. Let Aj ∈ u(nj ), j = 1, 2. Then denote by A1 ⊕ A2 ∈ u(n1 + n2 )
the “diagonal-block” matrix made of A1 and A2 , and by A1 ⊗ A2 ∈ u(n1 n2 ) the tensor
product of the two matrices. Then we have
(1) C(A1 ⊕ A2 ) = C(A1 ) · C(A2 )
(2) CH n1 +n2 (A1 ⊕ A2 ) = CH n1 (A1 ) + CH n2 (A2 )
(3) CH n1 n2 (A1 ⊗ A2 ) = CH n1 (A1 ) · CH n2 (A2 )

In other words, the total Chern polynomial is multiplicative w.r.t. direct sums, and
the Chern character polynomial is additive w.r.t. direct sums, and multiplicative w.r.t.
tensor products.
Thus implies that, for complex vector bundles E, F on M , we have
c(E) ∧ c(F ) = c(E ⊕ F ); ch(E ⊕ F ) = ch(E) + ch(F ); ch(E ⊗ F ) = ch(E) ∧ ch(F ).
For a complex line bundle L on M , we have
ch(L) = 1 + c1 (L).

For every compact Lie group G, it is possible to construct invariant polynomials on

its Lie algebra g by means of unitary representations:
Proposition 8.8. Let ρj : G → U(Vj ) be unitary (linear) representations of G on the
complex vector spaces Vj (j = 1 or 2). Let Qj ∈ R[u(Vj )]U (Vj ) be invariant polynomials.
Then a1 Q1 ◦ ρ01 + a2 Q2 ◦ ρ02 is an invariant polynomial on g. (Here ρ0j : g → u(Vj ) are
the Lie algebra actions on Vj associated to ρj , j = 1, 2.)

In particular, if we take Qj := CH Vj to be the corresponding Chern character power

series, and a1 = a2 = 1, then we simply obtain
CH V1 ◦ ρ01 + CH V2 + ρ02 = CH V1 ⊕V2 ◦ (ρ01 ⊕ ρ02 ),
the Chern character power series of the direct sum representation ρ1 ⊕ρ2 . But, using the
Proposition above, we can also define the Chern character power series of any “linear
combination of representations”, like CH(a1 ρ01 + ... + ak ρ0k ) , in particular for a “formal
difference” V1 − V2 , we can compute CH(V1 = V2 ) := CH ◦ ρ01 − CH ◦ ρ02 .
Every analytic function, symmetric in x1 , ..., xn , determines by truncation an invariant
polynomial on u(n). In particular:
Definition 8.9. The Todd power series is
n
Y xj
T D := −x
∈ R[[u(n)]]U(n) .
j=1
1−e j

The corresponding characteristic class (of a complex vector bundle E → M ) is called

the Todd class T d(E) ∈ H ∗ (M, R).
Exercise 8.10. Show that T d(E ⊕F ) = T d(E)∧T d(F ) for any complex vector bundles
on M . The Todd class is thus multiplicative w.r.t. direct sums.
28 FLORIN BELGUN

Let us determine the rings of invariant polynomials for SU(n), O(n) and SO(n):
Proposition 8.11. R[su(n)]SU(n) ' R[C2 , C3 , ..., Cn ].

Idea of proof: Clearly, all invariant polynomials on u(n) induce ones on su(n), and
C1 (essentially the trace) vanishes on su(n). One also shows that all polynomials on
su(n) can be extended to u(n). 
For the real orthogonal group, note that O(n) = GL(n, R) ∩ U(n), the intersection
taking place inside GL(n, C). As in the case of SU(n), it can be shown that all O(n)-
invariant polynomials on so(n) (the Lie algebra of O(n), but also of SO(n), the connected
component of id ∈ O(n)) can be extended to U(n)-invariant polynomials on u(n). It
1
is also easy to see that Ck |so(n) = 0 if k is odd: the eigenvalues of 2πi A, for a generic
matrix A ∈ so(n) ⊂ u(n), are pairwise opposite real numbers
{x1 , −x1 , ..., x[ n2 ] , −x[ n2 ] } for n even, or {x1 , −x1 , ..., x[ n2 ] , −x[ n2 ] , 0} for n odd,
where [ n2 ] is the integer part of n2 , thus A and −A are conjugate to each other by some
matrix in O(n). Therefore, Q(A) = Q(−A) for every O(n)-invariant polynomial on
so(n), which implies that all invariant polynomials contain only terms of even degree.
These facts can be summarized as follows:
Proposition 8.12. R[so(n)]O(n) = R[x21 , ..., x2[n ] ] = R[P1 , ..., P[ n2 ] ], where
2

j
Pj := (−1) C2j |so(n)
are the Pontryagin polynomials.
Exercise 8.13. Using Exercise 8.4 show that, for A ∈ so(n), P1 (A) = − 8π1 2 trA2 .

It can easily be shown that, for odd n, the O(n)-invariant polynomials on so(n) are
also SO(n)-invariant, thus
Proposition 8.14. Let n be odd. Then R[so(n)]SO(n) = R[x21 , ..., x2[n ] } = R[P1 , ..., P[ n2 ] }.
2

On the other hand, for n even, there is another SO(n)-invariant polynomial on so(n):
Definition 8.15. Let n := 2k and A ∈ so(n) ' Λ2 (Rn ). Then A ∧ ... ∧ A (k times)
is a 2k-form on Rn , thus a volume form. It is thus a real multiple of the standard,
SO(n)-invariant volume form on Rn voln . We define the Pfaffian Pf : so(n) → R as
A ∧ ... ∧ A
Pf (A) := .
n!(2π)n voln

The caharcteristic class induced on a real, oriented vector bundle E of rank 2k over
a manifold M by the Pfaffian on so(2k) is a cohomology class in H 2k (M, R) called the
Euler class of E.
Exercise 8.16. Show that Pf (BAB −1 ) = det B · Pf (A), for all B ∈ O(n), and if
{x1 , −x1 , ..., xk , −xk } are the eigenvalues of 2πiA, corresponding to an oriented basis
{e1 , ..., en } of Rn (more precisely, an eigenvector for the eigenvalue 2πixj is e2j−1 + ie2j ,
and an eigenvector for the eigenvalue −2πixj is e2j−1 − ie2j ), then Pf (A) = x1 · ... · xk .

It follows that Pf 2 = (−1)k Pk . One can prove

 GAUGE THEORY 29

Proposition 8.17. Let n := 2k. Then R[so(n)]SO(n) = R[x21 , ..., x2k , x1 · ... · xk } =
R[P1 , ..., Pk−1 , Pf ].
Remark 8.18. One can show in general (Th. Chevalley) that the ring of Ad-invariant
polynomials on a compact Lie algebra is a polynomial ring generated by a finite number
of such invariant polynomials.

The characteristic classes induced, by the Chern-Weil theory, by these polynomials

are
(1) on U (n)-bundles (complex vector bundles E of rank n): the Chern classes
ci (E) ∈ H 2i (M, R), induced by the polynomials Ci , i = 1, ..., n.
(2) on SU (n)-bundles (complex vector bundles E of rank n, with a nowhere-vanishing
section of Λn E): the Chern classes ci , induced by the polynomials Ci , i = 2, ..., n.
(3) on SO(2n + 1), O(2n + 1 or O(2n) bundles (on real vector bundles of any
rank, or oriented real vector bundles E of odd rank): the Pontryagin classes
pj (E) ∈ H 4j (M, R), induced by Pj , j = 1, ..., n
(4) on SO(2n) bundles (on oriented real vector bundles E of even rank): the Pon-
tryagin classes pj (E) ∈ H 4j (M, R), induced by Pj , j = 1, ..., n and e(E) ∈
H 2n (M, R) the Euler class of E, induced by Pf .
Every other real characteristic class of a bundle associated to these groups can be
written as a polynomial expression in the Chern, resp. Pontryagin classes, possibly also
the Euler class.
Remark 8.19. If E is a complex vector bundle of rank k, then its Euler class (seen
as an oriented real vector bundle of rank 2k) e(ER ) coincides with the top Chern class
ck (EC ).

We focus now on Yang-Mills theory for 4-dimensional oriented Riemannian manifolds,

and for G-bundles on them with G compact:
Proposition 8.20. Let G ⊂ U(k) be a compact Lie group and B(X, Y ) := −tr(XY ),
∀X, Y ∈ g ⊂ u(k). Let P → M be a G-bundle on the oriented Riemannian 4-manifold
(M 4 , g). Then Z 

2

Y M (∇) ≥ 4π  p1 (AdP ) ,
M
R
and,
R if ∇ satisfies the equality case, then ∇ is SD (and M 1
p (AdP ) ≥ 0), or ASD (and
p
M 1
(AdP ) ≤ 0).

Proof. If we decompose K ∇ = K+∇ + K−∇ in the SD, resp. ASD parts, we have:
B(K ∇ ∧ ∗K ∇ ) = −tr(K+∇ ∧ K+∇ ) + tr(K−∇ ∧ K−∇ ) = kK+∇ k2 + kK−∇ k2 ,
because the wedge product of a SD form with an ASD one is always zero. On the other
hand, applying the first Pontryagin polynomial to K ∇ (and using Exercise 8.13) yields:
1 1
P1 (K ∇ ) = 2 tr K ∇ ∧ K ∇ = 2 kK+∇ k2 − kK−∇ k2 .



8π 8π
The result follows by integration. (Note that ∇ is SD, resp. ASD, iff K−∇ , resp. K+∇
identically vanishes.) 
30 FLORIN BELGUN

9. Chern-Weil theory on noncompact manifolds

We can define, using the Chern-Weil Theorem, classes in some more special coho-
mology groups, for example: Let M be noncompact and consider only G-bundles with
compact support, i.e. G-bundles p : P → M , such that there exists a compact set K
and a trivialization of P on M r K. We can then consider the space of G-connections
∇ on P with compact support, i.e. ∇ on p−1 (M r K) is the trivial connection (given
by the fixed trivialization of P over M r K. A version “with compact support” of the
Chern-Weil theorem states then that, for every invariant polynomial Q on g, the in-
j
duced differential forms Q(K ∇ ), j = 1, 2, for ∇1 , ∇2 connections with compact support
as above, are
(1) with compact support contained in K (they vanish outside K)
(2) the relative Chern-Simons form ChS(∇1 , ∇2 ) also has compact support.
This implies that every invariant polynomial Q on g defines, for such G-bundles with
compact support on M , a cohomology class
[Q(K ∇ )] ∈ Hc∗ (M, R),
the cohomology with compact support on M .
Another version of the Chern-Weil Theorem will be needed for the Atiyah-Singer
Index Theorem, where cohomology classes on the total space of a vector bundle (the
tangent bundle) T → M are needed.
Definition 9.1. Let
σ
(σ) ... → Fi →i Fi+1 → ...
be a finite complex of vector bundles over T . We say that this complex has fiberwise
compact support (FCS) iff there exists a tubular neighborhood U ⊂ T of the zero
section M0 ⊂ T such that the sequence (σ) is exact when restricted to T r U .

Recall that a sequence of linear maps between vector bundles (σ) is called a complex
iff σi+1 ◦ σi = 0, ∀i ∈ Z. The complex is finite iff Fi 6= 0 only for finitely many i ∈ Z.
An open set U ⊃ M0 in T is tubular iff U ∩ Tx is relatively compact, ∀x ∈ M .
Example 9.2. If the complex only contains 2 non-zero terms, F0 and F1 , then it is
FCS iff σ0 is a vector bundle isomorphism outside a tubular neighborhood U ⊃ M0 .
Exercise 9.3. Let hi be Hermitean metrics on the vector bundles Fi (we need in the
sequel only the case when Fi are complex vector bundles). Then the linear maps σi∗ :
Fi+1 → Fi are the adjoint maps to σi :
hi (σi∗ (yi+1 ), yi )) = hi+1 (yj+1 , σi (yi )), ∀yi ∈ Fi , yi+1 ∈ Fi+1 .
Consider a sequence (σ) of linear maps between vector bundles as above. Show that:
(1) (σ) is a complex iff the adjoint sequence (σ ∗ ) is a complex, and that (σ ∗∗ ) ' (σ).
(2) (σ) is exact over U ⊃ M0 iff (σ ∗ ) is exact over U iff the map σeven ⊕ σodd ∗
:
Feven → Fodd is a isomorphism over U of the vector bundles
M M
Feven := F2j , Fodd := F2j−1 ,
j∈Z j∈Z
∗ ∗
σeven ⊕ σodd (..., y2j , ...) := (..., σ2j−2 (y2j−2 ) + σ2j−1 (y2j ), ...).
 GAUGE THEORY 31

(3) On an open set where (σ) is exact, the subspaces ker σi ⊂ Fi and σi (Fi ) ⊂ Fi+1
are vector subbundles (i.e. they have constant rank).
Definition 9.4. Let (σ) be a finite complex with FCS on T as above. A set of linear
connections (∇) := (∇i )i∈Z is called admissible or with fiberwise compact support (FCS)
iff there exists a tubular neighborhood U ⊃ M0 such that (σ) is exact over T r U and
(1) there exists a set (h) of Hermitean metrics on the bundles (Fi )i∈Z such that
∇i hi = 0 and σi (Fi−1 ) are ∇i -stable, ∀i ∈ Z

A subbundle F in a vector bundle E is said to be ∇-stable (for ∇ a linear connection

on E) iff
s ∈ C ∞ (F ) ⇒ ∇s ∈ C ∞ (Λ1 M ⊗ F ).
We also need to define the cohomology with fiberwise compact support (FCS cohomology)
on the total space T of a vector bundle over M , as the cohomology groups
Hf∗c (T ) := ker d/Im d
of the following complex
d d d
... → C ∞ (Λif c T ) → C ∞ (Λi+1
f c ) → ... ,

where Λif c T are the i-forms on T whose support is contained in a tubular neighborhood
of M0 . These forms are called differential forms with fiberwise compact support.
Remark 9.5. Note that a closed form with FCS which has vanishing FCS cohomology
class iff it is the derivative on an FCS form (being exact is also not enough, it needs a
primitive with FCS).
Theorem 9.6 (Chern-Weil theorem for complexes with FCS). Let (σ) a finite complex
of vector bundles, with FCS, on the total space T of a vector bundle over M
(1) For every admissible set of connections (∇) on (σ), the differential form
i
X
ch0 (∇) := (−1)i CH(K ∇ )
i∈Z

is a closed FCS form

(2) For every two admissible sets of connections (∇), (∇0 ) on (σ), the relative
Chern-Simons differential form ChS(∇, ∇0 ) defined as in Theorem 7.3 is a FCS
form.
Therefore, there exists a FCS cohomology class ch0 (σ) (represented by one of the ch0 (∇)
above) on T that depends on the finite FCS complex (σ) alone

10. Elliptic linear differential operators on vector bundles

We introduce now the notion of a linear differential operator.

First, define the space Fx of germs of functions (real or complex-valued) on M at
x ∈ M as the space of equivalence classes of functions that coincide on some open set
around x. This is a local ring, and its maximal ideal is mx := {f ∈ Fx | f (x) = 0}, the
space of germs vanishing at x. The space mkx is the space of germs that can be written
32 FLORIN BELGUN

as products of k germs in mx or, equivalently, the space of germs that vanish up to order
k − 1 at x, i.e.
f ∈ mx ⇔ f (x) = 0, df (x) = 0, ..., dk−1 f (x) = 0,
where note that the derivative of order l at x, dl f (x) is well-defined (without using
coordinates) as long as the previous derivatives at x vanish.
Remark 10.1. We have Fx /mx ' R (if we take real-valued functions, if we take
k−1 ∗
complex-valued functions we need to tensorize all relations with C), mk−1 k
x /mx ' S Tx M ,
1 2
in particular Λx M = mx /mx .

Quotienting out by mkx means “ignoring” the derivatives of higher order than k − 1.
Remark 10.2. One defines the jet bundle on M as the vector bundle J k M on M whose
fibers are Fx /mk+1
x . The construction of jet bundles can be made precise by glueing local
trivial parts (for chart domains) by k-order derivatives of the coordinate changes (the
formulas are complicated).

We have the following exact sequence:

0 → mkx /mk+1
x → Fx /mk+1
x → Fx /mkx →)], (13)
where the first term is S k Tx∗ M .
If we have a vector bundle E on M , we define Ex as the space of germs at x of
sections of E (which is a rank r = dim Ex free module over Fx - a basis is given, for
example, by the germ of a local frame in E around x), and we have the Fx modules
mk+1
x ⊗Fx Ex (of germs at x of sections of E that vanish up to order k at x), such that
(mkx ⊗Fx Ex )/(mk+1
x ⊗Fx Ex ) ' S k Tx∗ M ⊗ Ex (in particular Ex = Ex /(mx ⊗Fx Ex ) and
(Λ M ⊗ E)x = (mx ⊗Fx Ex )/(m2x ⊗Fx Ex )), and we have, by tensorizing over Fx with Ex ,
1

the exact sequences

0 → mkx /mk+1 ⊗Fx Ex → Fx /mk+1 ⊗Fx Ex → Fx /mkx ⊗Fx Ex → 0.




x x (14)
Definition 10.3. Let E and F be vector bundles over M . A linear differential operator
D from E to F of order at most k is a linear map D : C ∞ (E) → C ∞ (F ) such that
∀x ∈ M , D(f )x = 0, ∀f ∈ mk+1
x ⊗Fx Ex and the induced linear map
σ D : S k Tx∗ M ⊗ Ex → Fx , σ D (df1 (x) ⊗ ... ⊗ dfk (x) ⊗ e) := D(f1 ...fk s)x ,
∀f1 , ..., fk ∈ mX and s ∈ Ex , s(x) = e
is called the symbol of D. We say that D is of order k if it is of order at most k and
its symbol ist not identically zero.
Such a operator D is called elliptic iff σ D (λ, ...λ) : Ex → Fx is an isomorphism for
all x ∈ M and all λ ∈ Tx∗ M r {0}.
Example 10.4. A covariant derivative on a vector bundle E is a first order linear
differential operator from E to Λ1 ⊗ E with symbol equal to the identity of Λ1 ⊗ E. It
is elliptic iff n = 1, otrherwise dim Ex < dim(Λ1 ⊗ E)x .
Example 10.5. The exterior differential d : C ∞ (Λk M ) → C ∞ (Λk+1 M ) is a first order
linear differential operator with symbol the wedge product
σ d (λ) : Λk M → Λk+1 M, σ d (λ)(α) = λ ∧ α.
 GAUGE THEORY 33

In an analogue way, if (M, g) is a Riemannian manifold, then the codifferential δ g :

C ∞ (Λk M ) → C ∞ (Λk−1 M ),
Xn
g
δ (α) := ei y∇ei α, for e1 , ...en an o.n.b. of T M,
i=1
is a first order linear differential operator with symbol the interior product
σ δ (λ) : Λk M → Λk−1 M, σ d (λ)(α) = λyα.
None of these operators is elliptic, but
∆g := dδ g + δ g d : C ∞ (Λk M ) → C ∞ (Λk M )
is the Laplacian on forms, is a second-order linear differential operator and its symbol
σ ∆ (λ1 , λ2 ) : Λk M → Λk M, σ ∆ (λ1 , λ2 )(α) = g(λ1 , λ2 )α.
Exercise 10.6. Show that the Laplacian is elliptic iff the metric g is definite (e.g.
Riemannian).

If we compose two linear differential operators, their orders add up, if we add two
linear differential operators D1 + D2 , the order is at most the maximum of the two
orders (like the degree of polynomials). In the first case, the symbols are composed
σ D1 ◦D2 (λ) = σ D1 (λ) ◦ σ D2 (λ) (here, and from now on, we write the multiple argument
λ ∈ T ∗ M only once), and if the polynomials D1 and D2 have the same order, their
symbols add up (possibly resulting zero, i.e. D1 + D2 has smaller order) or, if the order
of D1 is larger than the order of D2 , then σ D1 +D2 = σ D1 .
It is useful to consider complexes of differential operators
Definition 10.7. A sequence of linear differential operators
D Dk+1
... → C ∞ (Ek ) →k C ∞ (Ek+1 ) → C ∞ (Ek+2 ) → ...
is a complex iff Dk+1 ◦ Dk = 0, ∀k ∈ Z. Such a complex is called elliptic iff the
corresponding sequence of symbols
σ Dk (λ) σ Dk+1 (λ) σ Dk+2 (λ)
... → Ek → Ek+1 → Ek+2 → ...
is exact for all λ ∈ T ∗ M non-zero.
Remark 10.8. If the sequence contains only two non-trivial bundles, say E0 and E1 ,
then the corresponding complex is elliptic iff D0 : C ∞ (E0 ) →: C ∞ (E1 ) is an elliptc
differential operator in the usual sense.
Example 10.9. The de Rham complex
d d d
... → C ∞ (Λk M ) → C ∞ (Λk+1 M ) → ...
is elliptic.

The main property of an elliptic operator is

Theorem 10.10. Let M be a compact manifold and D : C ∞ (E) → C ∞ (F ) an elliptic
operator. Then ker D ⊂ C ∞ (E) and cokerD = C ∞ (F )/D(C ∞ (E)) are finite dimen-
sional. Denote by ind(D) := dim ker D − dim cokerD the index of D.

For an elliptic complex, the theorem states:

34 FLORIN BELGUN

Theorem 10.11. Let M be a compact manifold and (D) = (Di )i∈Z a finite elliptic
complex of linear differential operators on vector bundles over M . Then the quotient
spaces ker Di /ImDi−1 are finite dimensional. Denote by
X
ind(D) := (−1)i dim (ker Di /ImDi−1 )
i∈Z

the index of (D).

The Atiyah-Singer Index Theorem gives a formula for the index of D uniquely in
terms of the topological properties of the symbol.
Let us first describe the relevant topological objects:
(1) Let (D) be an elliptic complex. Then the symbol determines a complex of vector
bundles
σ(Di−1 ) σ(Di ) σ(Di+1 )
(σ(D)) ... → π ∗ Ei → π ∗ Ei+1 → ...
over the total space of the cotangent bundle π : T ∗ M → M , which has FCS,
because of the ellipticity of (D). We will, however, identify T M with T ∗ M ,
because they are isomorphic bundles.
(2) Using admissible sets of connections on the bundles Fi := π ∗ Ei , the pull-backs
to T ∗ M of the bundles Ei , we can define the Chern character ch0 (σ(D)) of the
complex (σ(D)) (see Theorem 9.6).
(3) Consider T ∗ M as an oriented manifold as follows: if e1 , ..., en is a basis of Tx M ,
and e∗1 , ..., e∗n is the dual basis of Tx∗ M , then e1 , e∗1 , ..., en , e∗n defines an oriented
basis for the tangent space in α ∈ Tx∗ M of T ∗ M .
Theorem 10.12 (Atiyah-Singer Index Theorem). Let M be an even-dimensional3 man-
ifold and (D) be an elliptic complex with symbol (σ D ). Then it holds
Z
ind(D) = ch0 (σ(D)) ∧ π ∗ td(T M ⊗ C),
TM
for the above defined orientation of T M . The integration convention is that only the
homogeneous component of degree 2n in the right hand side is integrated over T M , all
other components are ignored.

11. Applications of the Index Theorem

The Index Theorem is already a spectacular result, because it states that two numbers
are equal, one coming from an analytic problem (“counting” the solutions of some
differential equations), the other from a topological one, that only involves the symbol
of the operator. Sometimes it is not obvious at all that the number from the right hand
side should be an integer, but this obviously follows from the Index Theorem.
However, there are many examples where the right hand side of the Index Theorem
can be simplified even further, so that in the end only the bundles (Ei ) matter (the
reference to the σ(D) totally vanishes)!
This happens if the symbol complex (σ(D)) is an universal complex:
3one can show that the index of a differential operator on an odd-dimensional manifold is always
zero
 GAUGE THEORY 35

Definition 11.1. Let π : T → M an oriented real vector bundle (the corresponding

group being thus SO(k), where k is the rank of T . Denote by P the SO(k)-principal
bundle on M defining T . Let Fi := π ∗ E ⊗ π ∗ Li , where E is some complex vector bundle
and Li are vector bundles associated to P via representations ρi : SO(k) → GL(Vi ).
Suppose the bundle maps from Fi to Fi+1 are the tensor products of IdE with some
maps σi0 : π ∗ Li → π ∗ Li+1 , such that:
0 0
σi−1 σ0 σi+1
(L) ... → π ∗ Li →i π ∗ Li+1 → ...
is a complex of vector bundles, and σi0 is defined by a SO(k)-equivariant map
[σi0 ] : S ki (Rk )∗ ⊗ Vi ,
where ki ∈ N. (thus the dependance of the map σi0 on the fiber elements of T is polyno-
mial, and can be expressed by a formula in terms of representations of SO(k)). Then
the complex
0 IdE ⊗s0i+1
IdE ⊗σi−1 IdE ⊗s0
(σ) → π ∗ (E ⊗ Li ) → i π ∗ (E ⊗ Li+1 ) → ...
is called universal. A complex of linear differential operator is called universal iff its
symbol complex is.
Remark 11.2. One can define universal complexes and operators by stating that they
are obtained by pull-backs from the classifying space of SO(k).
The classifying space BG of a compact Lie group is a (in general infinite-dimensional)
manifold, that admits a universal G-principal bundle P G → BG such that P G is
a contractible space. As a consequence, every G-principal bundle P → M can be
obtained as the pull-back of P G, for a suitable map f : M → BG, in fact there is a
bijective correspondence between isomorphism classes of principal G-bundles on M and
homotopy classes of maps from M to BG.
In fact, a real characteristic class for a certain Lie group G corresponds to a coho-
mology class in H ∗ (BG, R), which turns out to be isomorphic (as a ring), to the ring
of invariant polynomials over g, thus it has no zero divisors, unlike the cohomology of
a finite-dimensional manifold.
The theory of the classifying spaces is an important and classical part of the Lie group
and gauge theory and characteristic classes, but in these lectures we will not develop it
any further and “go around it”.

The universality of the symbol is satisfied by most differential operators that appear
in geometry, when they are determined by the structure of the the manifold M .
The idea of reducing an integration of a form with FCS on T M to an integral on M
is the Thom isomorphism:
Theorem 11.3 (Thom isomorphism). Let π : T → M be an oriented real vector bundle
of rank k over a manifold M . There exists a unique cohomology class U in Hfkc (T, R)
(cohomology with fiberwise compact support) such that
Z
U = 1, ∀x ∈ M.
Ex

Moreover, the following hold:

36 FLORIN BELGUN

(1) the wedge product with U of the pull-back π ∗ is an isomorphism of H ∗ (M, R)-
modules
ψ : H ∗ (M, R) → Hf∗+k
c (T, R), ψ(α) := U ∧ π ∗ α,
called the Thom isomorphism.
(2) The restriction of U to the zero section M0 ⊂ T , or its pull-back through the
inclusion i : M → T of the zero section, is the Euler class e(T ) of the real vector
bundle T → M .

The above Theorem claims that the equation

ψ(α) = β,
where the right hand side is a cohomology class in Hfk+p
c (T, R) admits a unique solution
p
in H (M, R). On the other hand, this solution satisfies
u ∧ π ∗ α = β ⇒ e(T ) ∧ α = i∗ β,
of which the right hand side equation does not, unfortunately, admit unique solutions
in the cohomology ring H ∗ (M, R).
The idea is to solve this equation (and hence find an expression for ψ −1 (β)) is that, if
β is the Chern character of an universal bundle (or an universal complex of bundles (σ)),
then β = π ∗ ch(E) ∧ β0 , where β0 is the Chern character of the complex (L) of Definition
11.1, and it can be expressed as ch0 (∇), for some admissible set of connections (∇), see
Theorem 9.6.
(0)
As e(T ) is expressed by Pf (K ∇ ), where ∇(0) is an SO(k)-connection on T , we can
construct a representative U (0) of the Thom class U by the formula
T
U (0) := Pf (K ∇ ),
where ∇T is a connection on π ∗ T , which coincides on a tubular neighborhood V0 of M0
with π ∗ ∇(0) , and satisfies
∇Tv (|v|w) = 0, ∇TX̃ w = 0, ∀v ∈ Tx , X ∈ TX M, w ∈ C ∞ (T )
outside a suitable tubular neighborhood V1 ⊃ V̄0 of M0 . Here v is seen as a vertical
vector field on T , and X̃ is the ∇(0) -horizontal lift of X to a horizontal vector on T , and
w is a section in π ∗ T that is constant on the fibers (it is the pull-back to T of a section
w on M ). Note that the pulled-back connection π ∗ ∇(0) would satisfy
∇Tv w = 0, ∇TX̃ w = 0, ∀v ∈ Tx , X ∈ TX M, w ∈ C ∞ (T ),
without the factor |v|. The presence of the factor |v| in the definition of ∇T ouside V1
implies that the tautological section of π ∗ T → T which has in v ∈ T the value v ∈ π ∗ Tv is
∇t -parallel outside this tubular neighborhood. This, in turn, implies that the associated
connections ∇i on π ∗ Li form an admissible set.
We skip the technical details concerning the choice of the tubular neighborhoods
V0 , V1 such that Z
T
Pf (K ∇ ) = 1, ∀x ∈ M,
Tx
but it is easy to see that on V0 and on T r V̄1 the curvature of ∇T (and thus of the
associated ∇i as well) has only horizontal components, in particular
T (0)
i∗ Pf (K ∇ ) = Pf (K ∇ ),
 GAUGE THEORY 37

and thus it represents the Euler class.

At this point we see that, if the polynomial
X
(−1)i CH ◦ ρ0i
i∈Z

is divisible by Pf , and the quotient is Qσ ∈ R[so(k)]SO(k) , then we can write

T
ch0 (∇) = Pf (K ∇ ) ∧ Qσ (∇),
thus
(0)
ψ −1 ch0 (∇) = Qσ (K ∇ ). (15)
In most cases, we can show that the divisibility by the Pfaffian is automatic, hence we
can reduce the computation of the index of an elliptic complex to a computation of
characteristic classes on M :
Z
(0)
ind(D) = (−1) n/2
Qσ (K ∇ ) ∧ T d(T M ⊗ C).
M

(the sign (−1)n/2 is due to the fact that the orientation of T M needed for the Index
Theorem is different from the orientation coming from the Thom isomorphism).
The details of the following examples can be found in [4]:

11.1. The de Rham complex. Consider the de Rham complex

d d d
... → C ∞ (Λk M ) → C ∞ (Λk+1 M ) → ...
on a 2n-dimensional oriented manifold M , and its symbol
v∧· v∧· v∧·
... → Λk M → Λk+1 M → ...,
for v ∈ T ∗ M . This symbol is a universal complex, thus we compute the polynomial
X n
Y
CH(Λ ·) = k
(1 − exi )(1 − e−xi ),
k∈N k=1

which is divisible by Pf = x1 ...xn , because xi are clearly zeros of the above polynomial
(power series). Denote by Qd the corresponding quotient polynomial. The formula for
Qd · T D is very simple: it is
n
Y n
Y
xi −xi
(1 − e )(1 − e ) xk (−xk )
k=1 k=1
· n = (−1)n Pf .
x1 ...xn Y
(1 − exi )(1 − e−xi )
k=1

Thus Z Z
n n ∇
χ(M ) = ind(d) = (−1) (−1) Pf (K ) = e(T M ),
M M
which is a well-known fact: the Euler characteristic χ(M ), which is the alternated sum
of the Betti numbers dim H i (M, R), is equal to the integral of the Euler class.
38 FLORIN BELGUN

11.2. The signature operator. . The de Rham complex is equivalent, via Hodge
theory, to the following operator
d + δ : C ∞ (Λeven M ) → C ∞ (Lodd M ),
where δ is the adjoint of d for a fixed Riemannian metric. If we denote by D := d + δ :
C ∞ (Λ∗ M ) → C ∞ (L∗ M ), then D is clearly self-adjoint, so its index is zero. But if we
restrict D to the bundles Λeven and Lodd as above, then its index is not trivial and is
equal to the Euler characeristic.
But we can restrict D to some other subbundles of Λ∗ M , as follows: recall that, on
an oriented, even-dimensional manifold M 2n , the Hodge star operator has the property
∗2 |Λk M = (−1)k Id,
so it is almost an involution. The following linear operator, defined on the complexified
space Λ∗ M ⊗ C is an involution (exercise!, use Proposition 6.1):
τ (α) := ik(k+1)+n ∗ α, ∀α ∈ Λk M ⊗ C.
Exercise 11.4. Check that D ◦ τ = −τ ◦ D.

We denote by Λ+ M and Λ− M the eigenspaces of τ for the eigenvalues 1, resp. −1,

and we can decompose D as D = D+ + D− , where
D+ : C ∞ (Λ+ M ) →: C ∞ (Λ− M ), D− : C ∞ (Λ− M ) →: C ∞ (Λ+ M ).
These operators are adjoint to each other, so cokerD+ = ker D− .
The bundle Λ+ M is isomorphic to the direct sum of all Λk M ⊗ C, for k < n, and of
(Λn M ⊗ C)+ , and Λ− M ' Λ1 M ⊕ ... ⊕ Λn−1 M ⊕ (Λn M ⊗ C)− . It can also be shown
that !
n−1
M
Hk (M ) ⊗ C ⊕ (Hn (M ) ⊗ C) ∩ (Λn M ⊗ C)+ ,


ker D+ '
k=0

where Hk (M ) is the space of harmonic k-forms (thus ker D ∩ Λk M ). Also

n−1
!
M
Hk (M ) ⊗ C ⊕ (Hn (M ) ⊗ C) ∩ (Λn M ⊗ C)− ,


ker D− '
k=0

thus
ind(D+ ) = dimC (Hn (M ) ⊗ C) ∩ (Λn M ⊗ C)+ −dimC (Hn (M ) ⊗ C) ∩ (Λn M ⊗ C)− .

If n (half of the dimension of M ) is odd, then τ : Λn → Λn has its square equal to –Id,
thus the eigenspaces of τ are conjugated to each other. Moreover, every real harmonic
n-form α defines harmonic forms α ± τ (α) ∈ ((Hn (M ) ⊗ C) ∩ (Λn M ⊗ C)± ), and thus
the index of D+ is zero.
If n is even (thus dim M is divisible by 4), then τ coincides with ∗ on Λn M and the
eigenspaces of τ are real, thus
n n
ind(D+ ) = dim H+ (M ) − dim H− (M ),
n
(here H± (M ) := Hn (M ) ∩ L± (M )) which is the signature of the intersection form
q : H n (M ) × H n (M ) → R, q(α, β) := ∗(α ∧ β).
 GAUGE THEORY 39

Noet that, for even n, q is bilinear and symmetric, and it induces on Hn (M ) a bilinear
n n
symmetric form which restricted to H+ (M ) is positive-definite, and restricted to H− (M )
is negative-definite.
The signature of an oriented manifold of dimension 4k (equal, by definition, to the
signature of its intersection form) is thus equal to the index of D+ .
To compute this index, we need to compute
CH(Λ+ − Λ− ),
where Λ± are the corresponding complex representations of SO(2n). It is useful, at this
point, not to restrict to even n.
Lemma 11.5. If V1 , V2 are evendimensional oriented real vector spaces, then
Λ+ (V1 ⊕ V2 ) = Λ+ (V1 ) ⊗ Λ+ (V2 ) ⊕ Λ− (V1 ) ⊗ Λ− (V2 ) ,

Λ− (V1 ⊕ V2 ) = Λ+ (V1 ) ⊗ Λ− (V2 ) ⊕ Λ− (V1 ) ⊗ Λ+ (V2 ) ,

thus
CH(Λ+ − Λ− )(V1 ⊕ V2 ) = CH(Λ+ − Λ− )(V1 ) · CH(Λ+ − Λ− )(V2 ).

Proof. The first part (the relations between Λ± of V1 , V2 and V1 ⊕V2 ) is left as an exercise.
For the second part note that block matrices A1 ⊕ A2 ∈ so(V1 ⊕ V2 ), made of the two
blocks Ai ∈ so(Vi ), cover all conjugacy classes of so(V1 ⊕ V2 ) under the adjoint action of
SO(V1 ⊕ V2 ) (Note: this only happens because Vi are both even-dimensional). Therfore,
the Chern character polynomial of the difference (Λ+ − Λ− )(V1 ⊕ V2 ) is determined by
evaluating this polynomial on such block matrices. 

The last step in computing CH(Λ+ − Λ− )(R2n ) is thus to compute it for n = 1, where
it can by done by direct computation: Take e1 , e2 an oriented ONB of R2 , such that
e1 ± ie2 ∈ R2 ⊗ C is a basis of eigenvectors of A for the eigenvalues ±2πix. Then
1 + ie1 ∧ e2 , e1 + ie2
forms a basis of eigenvectors for the action of A on Λ+ (R2 ) (the eigenvalues are zero,
resp 2πix), and
1 − ie1 ∧ e2 , e1 − ie2
forms a basis of eigenvectors of A on Λ− (R2 ), for the eigenvalues 0, resp. −2πix. The
Chern character power series, evaluated on A yields
CH(Λ+ − Λ− )(A) = 1 + ex − (1 + e−x ) = ex − e−x ,
thus the Chern character power series on Λ+ − Λ− , evaluated on a matrix A ∈ so(2n),
with eigenvalues {±2πixk }k=1,...,n , is equal to
n
Y
−
+
CH(Λ − Λ )(A) = (exk − e−xk ),
k=1

and it clearly vanishes for xk = 0, ∀k = 1, ..., n, thus it is divisible by Pf (A) = x1 ...xn

and we can compute
 x x
 x x

k − 2k k − 2k
CH(Λ+ − Λ− ) · T D Y xk e − e
n 2 e +e2 Yn xk
n n n 2
(A) = (−1) = (−1) 2 .
tanh x2k
2
Pf
 x x
k k
k=1 e 2 − e− 2 k=1
40 FLORIN BELGUN

This power series will be denoted with L ∈ R[[so(2n)]]SO(2n) (it is an invariant power
series, but we can truncate it to obtain invariant polynomials).
x
Now, the analytic function tanh x
is even, and its Taylor expansion starts with
x 1
= 1 + x2 + h.o.t.,
tanh x 3
therefore we can conclude that, for a 4-dimensional , oriented manifold, its signature is
Z
p1 (M )
sign(M ) = ,
M 3
because one-third of P1 = x21 + x22 is the only term of order 2 (that corresponds to a
4-form via Chern-Weil theory) in the integrand
x1 x2
x21 x2 x2 + x22
22 2 2
= 4(1 + + h.o.t.)(1 + 2 + h.o.t.) = 4 + 1 + h.o.t. (16)
tanh x21 tanh x22 12 12 3
In general, one defines the L-genus of a manifold to be the (characteristic) cohomology
class represented by L(K ∇ ), for ∇ an SO(2n)-connection on T M . We have obtained
Theorem 11.6 (Hirzebruch’s signature Theorem). The signature of a 4k-dimensional,
oriented, compact manifold M is equal to its L-genus integrated on M .
Remark 11.7. It is clear that the L-genus is a polynomial expression in the Pontryagin
classes of M , with rational coefficients. It is a priori not obvious that the L-genus is
an integer cohomology class; for higher dimensions the denominators in the polynomial
expressions of L in terms of the Pontryagin classes grow very fast, see [4].

12. Moduli space of self-dual connections on a compact 4-manifold

We have seen in Proposition 8.20 that SD and ASD connections on a compact oriented
Riemannian manifold M , if they exist, are absolute minimizers of the Yang-Mills func-
tional. In this section we briefly describe a famous theorem of Atiyah-Hitchin-Singer [1]
that describes, under some conditions, the moduli space of such connections:
The idea is to consider the space (in general, it is infinite-dimensional) of SD connec-
tions on a given principal G-bundle P → M , then to consider its orbit space w.r.t. the
action of the gauge group, which is the (infinite-dimensional) “Lie” group of gauge equiv-
alences. This quotient space is a topological space called the moduli space MSD (M, P )
of SD connections on P . It is a priori not clear that it Hausdorff or if it is (at least
piecewise) a smooth, finite-dimensional manifold.
Theorem 12.1. (Atiyah-Hitchin-Singer [1]) Let M be a compact Riemannain manifold
of self-dual type and with positive scalar curvature. Then the moduli space MSD
0 (M, P )
of irreducible self-dual connections on a given G-bundle (G is a compact semisimple Lie
group) P → M is either empty, or it is a smooth, Hausdorff, manifold of real dimension
Z
SD
dim M0 (M, P ) = p1 (AdP ) − (χ(M ) − sign(M )) dim G.
M

A compact semisimple Lie group is a compact Lie group with trivial, or discrete center
(e.g. SU(k), whose center is given by the k-th rooths of unity, is seminismple, U(k) is
not, because its center is U(1) acting on Ck diagonally).
 GAUGE THEORY 41

In this case, a connection is irreducible iff AdP admits no local ∇-parallel section
(if G had a non-trivial center, then g would contain non-trivial Ad-invariant elements,
which would define non-trivial global ∇-parallel sections for every G-connection).
Finally, recall that the Riemann curvature tensor R of a manifold can be seen as a
symmetric linear map from Λ2 M → Λ2 M . If dim M = 4, the 2-forms on M decompose
in SD, resp ADS parts, thus R admits a block decomposition as a symmetric map
R : Λ2+ M ⊕ Λ2− M → Λ2+ M ⊕ Λ2− M,
and the components in End(Λ2± ) are symmetric, and the trace of each of these com-
ponents is equal to Scal6
, where Scal is the scalar curvature of M . By definition, a
manifold is self-dual if the component of R in End(Λ2+ ) reduces to this diagonal part
(the algebraic components of R depend, in general, on the Ricci tensor and the Weyl
tensors of M , [1]).

Proof. The proof is involved and only few ideas will be explained here. For a well-
prepared reader, the best reference for the proof is the original paper [1]. The proof has
3 main steps:
(1) infinitesimal (where a computation of the virtual tangent space of MSD 0 (M, P )
SD
at a point ∇ ∈ M0 (M, P ) is made)
(2) local (where it is shown that the virtual vectors from the first step correspond to
local deformations (curves in MSD 0 (M, P ) starting in ∇), thus providing local
smooth charts)
(3) global (where it is shown that the topological space MSD 0 (M, P ) is Hausdorff
and thus a smooth manifold)
We only give some details of the first step: for this, we fix ∇ an SD connection and
construct an elliptic complex
∇ d∇
−
(D) 0 → C ∞ (AdP ⊗ C) → C ∞ (Λ1 M ⊗ AdP ⊗ C) → C ∞ (Λ2− M ⊗ AdP ⊗ C) → 0,
(17)
where d∇
− is the ASD-part of d∇
, the covariant exterior differential
d∇ : C ∞ (Λ1 M ⊗ AdP ) → C ∞ (Λ2 M ⊗ AdP ).
The fact that (D) is indeed a complex is due to the fact that
d∇ ◦ ∇ = K ∇ ·, thus d∇ ∇
− (∇(s)) = K− · s = 0,

because ∇ is SD, thus the ASD-part of K ∇ vanishes.

It is easy to see that this complex is elliptic (exercise!) and that it is universal, so its
index can be computed by integrating some characteristic classes on M .
Let us first see why this complex is relevant for the infinitesimal step of the proof:
note that every G-connection can be written ∇0 = ∇ + θ, where θ ∈ C ∞ (Λ1 M ⊗ AdP ).
As in the proof of the Chern-Weil theorem, we see that, if ∇t := ∇ + θt is a 1–parameter
family of connections passing with ∇0 = ∇, then
 
d  ∇t d 
K = d θ, where θ =  θt ∈ C ∞ (Λ1 M ⊗ AdP ),
∇
dt t=0 dt t=0
t
and if all ∇ are SD, then d∇
−θ = 0. So the kernel of the operator d∇
− characterizes, on
an infinitesimal level, the deformations of ∇ through SD connections.
42 FLORIN BELGUN

On the other hand, if such a deformation is induced by a family of gauge equiva-

lences φt : P → P , then the AdP -valued 1-forms (or, equivalently, the Ad-equivariant
horizontal 1-forms on P ) θt are determined by (5)
θt (X) = Adφ¯t −1 ω(X) − ω(X) + ω M C (φ̄t ∗ (X)), ∀X ∈ T P, (18)
where ω is the connection 1–form associated to ∇ and φ̄t : P → G is the AD-equivariant
function that corresponds to the gauge transformation φt . As we know that θt is a
horizontal 1-form on P which, because of its equivariance properties, induces a section
in Λ1 M ⊗ AdP , it suffices to fix a point p ∈ Px , x ∈ M , and consider X ∈ ker ωp in
(18).
We have then φ̄t (p) a curve in G, with φ¯0 = e, the neutral element of G, and θPt (X) =
MC
ω (φ̄t ∗ (X)) is a family of vectors in g. If we set

d 
X := cs , where cs ∈ P
ds  s=0
is a ∇-horizontal curve, with c0 = p, then
    
d  t d  MC d  ¯(φt )(cs ) .
θp (X) = ω
dt 
t=0 ds 
s=0 dt  t=0
0
Denote by φ (q) ∈ g the derivative in t, for t = 0, of the map
t 7→ ω M C (φ̄t (q)),
for all q ∈ P . We obtain a smooth map φ0 : P → g which induces a section in AdP
(because it is equivariant – exercise!). We conclude that

d 
θpt (X) = dφ0 (X).
dt t=0
This implies that the space of infinitesimal deformations of ∇ through gauge-equivalent
connections is the space Im(∇), for
∇ : C ∞ (AdP ) → C ∞ (Λ1 M ⊗ AdP ),
thus the virtual tangent space to MSD
0 (M, P ) is the vector space

ker d∇
− /Im∇,

i.e. the middle cohmology term of (17). In order to show that its dimension (which is
finite because the complex (17) is elliptic) is exactly equal to minus the index of (D),
we need to show that the other two cohomology terms vanish.
For the term in H 0 , which is simply ker ∇, this is implied by the fact that ∇ is
irreducible, because ker ∇ is exactly the space of ∇-parallel sections of AdP .
For the term in H 2 , which is cokerd∇
− , this is implied by the conditions on the Rie-
mannian manifold M : self-dual and with positive scalar curvature. We refer the reader
to [1] for details.
We still need to compute the index of (D); this is done with the Index Theorem: we
need to compute the quotient of the Chern character polynomial of (Λ0 − Λ1 + Λ2− ),
which is one half of the difference of
CH(Λ0 − Λ1 + (Λ2− + Λ2+ ) − Λ3 − Λ4 ) = CH(2Λ0 − 2Λ1 + (Λ2− + Λ2+ )) and
CH((Λ0 + Λ4 )+ + (Λ1 + Λ3 )+ + Λ2+ − (Λ0 + Λ4 )− − (Λ1 + Λ3 )− − Λ2− = CH(Λ2+ − Λ2− ).
 GAUGE THEORY 43

The first line above corresponds to the computation for the de Rham complex and the
second for the signature operator. Therefore, the integrand (recall that the integrand
also contains the Todd class) for the non-twisted elliptic complex
d d−
(D0 ) 0 → C ∞ (M ) → C ∞ (Λ1 M )C ∞ (M ) → C ∞ (Λ2− M ) → 0
is half of the difference of the integrands for the de Rham complex and for the signature
operator, i.e.,
1 1
(Pf − L) = −2 + (χ(M ) − sign(M ))volM ,
2 2
where we have only written the 0 degree term, which is 12 (0 − 4) = −2, and the term of
degree 4, which is the result of the R integral (hence the index of (D0 )) times the volume
element of M (chosen such that M volM = 1). Recall the formula for L from (16).
For the twisted complex (17), we need to multiply this integrand with the Chern
character of AdP ⊗ C, which is
1 1
ch(AdP ⊗C) = rank(AdP ⊗C)+c1 (AdP ⊗C)+ (c21 −c2 )(AdP ⊗C) = dim G+ p1 (AdP ).
2 2
Here again, we have only considered the terms of degree at most 4.
We conclude
Z
1
ind(D) = − dim G(χ(M ) − sign(M )).
p1 (AdP ) +
M 2
As the index of D is equal to minus the dimension of the virtal tangent space, we
conclude (assuming the steps 2 and 3 of the proof – see [1]) that MSD
0 (M, P ) is a
smooth manifold of dimension
Z
1
p1 (AdP ) − dim G(χ(M ) − sign(M )),
M 2
as claimed. 

In particular, there exist SD connections on P only if

Z
1
p1 (AdP ) ≥ dim G(χ(M ) − sign(M )).
M 2
In the case of the sphere and G := SU(2), one can show that p1 (AdP ) = 8k volM , with
k ∈ Z. The isomorphism class of a SU(2)-bundle on S 4 is, in fact, determined by this
number k. If we denote MSD SD
0 (M, P ) by M0 (M, k), we can state:

Corollary 12.2. The moduli space of irreducible self-dual connections on an SU(2)-

principal bundle of type k ∈ N over the round sphere S 4 is a smooth manifold of dimen-
sion 8k − 3.

Moreover, the round S 4 is not only seld-dual, but anti-self-dual as well, i.e. it it also
self-dual for the opposite orientation (in fact, S 4 is the only compact orientable manifold
with positive scalar curvature being at the same time self-dual and anti-self-dual), so
the conditions of the AHS theorem also apply for S 4 with opposite orientation. But SD
connections for this opposite orientation are ASD for the standard orientation, thus
Corollary 12.3. The moduli space of irreducible anti-self-dual connections on an SU(2)-
principal bundle of type k ∈ N over the round sphere S 4 is a smooth manifold of dimen-
sion −8k − 3.
44 FLORIN BELGUN

In fact, all these moduli spaces can be shown to be non-empty, if their dimension,
as given by the AHS theorem, is positive. For k = 1 there are explicit solutions of the
Yang-Mills self-dual equations covering all the moduli space.
Note that for k = 0 (the trivial bundle), all SD connections are automatically also
ASD, thus flat (the curvature vanishes identically). However, since S 4 is simply-
connected, this implies that it is a trivial connection, therefore non irreducible. So
MSD 4
0 (S , 0) = M0
ASD
(S 4 , 0) = ∅.
For k > 0, MSD 4
0 (S , k) is a smooth manifold as above and M0
ASD
(S 4 , k) = ∅ and for
k < 0, MASD
0 (S 4 , k) is a smooth manifold as above and MSD 4
0 (S , k) = ∅.

References
[1] M. Atiyah, N. Hitchin, I. Singer, Self-duality in 4-dimensional riemannian geometry. Proc. Roy.
Soc. London Series A., 362 (1978), 425-461.
[2] J. Milnor, Stasheff, Characteristic classes, Princeton Univ. Press, 1974
[3] H. Baum, Eichfeldtheorie, Springer 2009
[4] P. Shanahan, The Atiyah-Singer Index Theorem, LNM 638 (1978)