Topics in Mathematical Physics

Guo Chuan Thiang

Lecture notes for a course at Peking University, Fall 2022. Please email me
for feedback, and if you would like to use some of this material in teaching.

Contents
1 On fields and frames 4
1.1 Tangent bundle and other bundles . . . . . . . . . . . . . . . . 9

2 Geometry and gauge in quantum theory 11

2.1 “Wavefunctions” in quantum mechanics . . . . . . . . . . . . . 12
2.2 Time evolution in quantum mechanics . . . . . . . . . . . . . 18

3 Dirac operators in one dimension 20

3.1 The many faces of R . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Dirac operators on complete 1-manifolds . . . . . . . . . . . . 21
3.3 Dirac operators on incomplete 1-manifolds . . . . . . . . . . . 30
3.4 Dirac operators and SUSY quantum mechanics . . . . . . . . 35
3.5 Outlook for higher-dimensional manifolds . . . . . . . . . . . . 36

4 Smooth manifolds 38
4.1 Tangent vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Derivatives of smooth maps . . . . . . . . . . . . . . . . . . . 44
4.3 Immersions, submersions, embeddings . . . . . . . . . . . . . . 47

5 Tangent bundle and vector bundles 50

5.1 Tangent bundle . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 General vector bundles . . . . . . . . . . . . . . . . . . . . . . 52
5.3 Tautological line bundle over RP1 . . . . . . . . . . . . . . . . 53
5.4 Morphisms of vector bundles . . . . . . . . . . . . . . . . . . . 55

1
6 Principal G-bundles 58
6.1 Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2 Idea of principal bundles . . . . . . . . . . . . . . . . . . . . . 59
6.3 General principal bundle . . . . . . . . . . . . . . . . . . . . . 60
6.4 Transition functions . . . . . . . . . . . . . . . . . . . . . . . . 64
6.5 Frame bundle of a vector bundle . . . . . . . . . . . . . . . . . 67
6.6 Gauge transformations . . . . . . . . . . . . . . . . . . . . . . 68

7 Preview of Spin 73
7.1 Hopf bundle over S 2 . . . . . . . . . . . . . . . . . . . . . . . 73
7.2 Orthonormal frame bundle of S 2 . . . . . . . . . . . . . . . . . 74
7.3 Bundle of spin frames over S 2 . . . . . . . . . . . . . . . . . . 75

8 Tensors and differential forms 78

8.1 Commutator of vector fields . . . . . . . . . . . . . . . . . . . 78
8.2 Differential 1-forms . . . . . . . . . . . . . . . . . . . . . . . . 79
8.3 Pullbacks and push-forwards . . . . . . . . . . . . . . . . . . . 81
8.4 2-forms and Riemannian metrics . . . . . . . . . . . . . . . . . 82
8.5 Differential k-forms and exterior derivative . . . . . . . . . . . 83

9 Integral curves and flows 86

9.1 Integral curves of vector fields . . . . . . . . . . . . . . . . . . 86
9.2 Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9.3 Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

10 Geometry of Lie group actions 91

10.1 Lie algebra of left-invariant vector fields . . . . . . . . . . . . 91
10.2 Matrix Lie group and algebra examples . . . . . . . . . . . . . 93
10.3 Adjoint representation . . . . . . . . . . . . . . . . . . . . . . 94
10.4 Exponential map g → G . . . . . . . . . . . . . . . . . . . . . 95
10.5 Fundamental vector fields on G-manifolds . . . . . . . . . . . 99
10.6 Maurer–Cartan form . . . . . . . . . . . . . . . . . . . . . . . 102
10.7 Derivative of Lie group action . . . . . . . . . . . . . . . . . . 104

11 Connections on principal bundles and their curvature 105

11.1 Canonical vertical bundle . . . . . . . . . . . . . . . . . . . . . 105
11.2 Connection as horizontal bundle . . . . . . . . . . . . . . . . . 106
11.3 Connection as 1-form . . . . . . . . . . . . . . . . . . . . . . . 107

2
 11.4 Local description of connection: gauge potentials . . . . . . . 109
11.5 Curvature of connections . . . . . . . . . . . . . . . . . . . . . 112
11.6 Physics notation . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.7 Globally gauge-transformed connections and curvatures . . . . 116

12 Fibre bundles and structure groups 120

12.1 Vector=[frame, components] . . . . . . . . . . . . . . . . . . . 120
12.2 Associated fibre bundles . . . . . . . . . . . . . . . . . . . . . 121
12.3 Operations on vector bundles . . . . . . . . . . . . . . . . . . 124
12.4 Vector bundle metrics . . . . . . . . . . . . . . . . . . . . . . 126
12.5 Restricted frame bundles . . . . . . . . . . . . . . . . . . . . . 126
12.6 Reduction of structure group . . . . . . . . . . . . . . . . . . . 127
12.7 Scalar fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

13 Parallel transport and covariant differentiation 130

13.1 Parallel transport of frames . . . . . . . . . . . . . . . . . . . 130
13.2 Holonomy of parallel transport . . . . . . . . . . . . . . . . . . 133
13.3 Covariant derivatives on associated vector bundles . . . . . . . 134
13.4 Physics notation . . . . . . . . . . . . . . . . . . . . . . . . . 138
13.5 Curvature of covariant derivative . . . . . . . . . . . . . . . . 140
13.6 Metric connections . . . . . . . . . . . . . . . . . . . . . . . . 140
13.7 Levi-Civita connection of Riemannian manifold . . . . . . . . 141
13.8 Berry connection . . . . . . . . . . . . . . . . . . . . . . . . . 144

14 Clifford algebras 146

14.1 Algebra jargon . . . . . . . . . . . . . . . . . . . . . . . . . . 146
14.2 Clifford algebra definition . . . . . . . . . . . . . . . . . . . . 148
14.3 Clifford algebras Cln , Cln . . . . . . . . . . . . . . . . . . . . . 152

15 Spin groups and representations 155

15.1 Spin groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
15.2 Spin Lie algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 159
15.3 Spin representations . . . . . . . . . . . . . . . . . . . . . . . 161

16 Spinor geometry 167

16.1 Spin structures and spin connection . . . . . . . . . . . . . . . 167
16.2 Example: spin structure and connection on S 2 . . . . . . . . . 170
16.3 Spinor bundles and Clifford multiplication . . . . . . . . . . . 171

3
 16.4 Spinor bundle covariant derivative . . . . . . . . . . . . . . . . 174
16.5 Spin–Riemannian-curvature relations . . . . . . . . . . . . . . 176

17 Dirac operators 179

17.1 Atiyah–Singer–Dirac operator . . . . . . . . . . . . . . . . . . 179
17.2 Laplacians on vector bundles with connection . . . . . . . . . 182
17.3 Schrödinger–Peres–Lichnerowicz identity . . . . . . . . . . . . 184
17.4 Formal self-adjointness of Dirac operators and Laplacians . . . 186

18 Quantum Hall Effect 191

18.1 Landau Hamiltonian on a surface . . . . . . . . . . . . . . . . 191
18.2 Euclidean plane Landau Hamiltonian . . . . . . . . . . . . . . 193
18.3 Qualitative spectrum of general Landau Hamiltonian . . . . . 196
18.4 Some quantum Hall physics (Optional) . . . . . . . . . . . . . 198

19 Unbounded operators on Hilbert space 201

19.1 Recap of Hilbert space theory . . . . . . . . . . . . . . . . . . 201
19.2 Unbounded operators . . . . . . . . . . . . . . . . . . . . . . . 202
19.3 First-order differential operators . . . . . . . . . . . . . . . . . 206

20 Self-adjoint extensions 210

20.1 Spectrum of symmetric operators . . . . . . . . . . . . . . . . 210
20.2 von Neumann theory of self-adjoint extensions . . . . . . . . . 212
20.3 Weak derivatives . . . . . . . . . . . . . . . . . . . . . . . . . 214
20.4 Example: momentum operator on an interval . . . . . . . . . 216
20.5 Example: momentum operator on line . . . . . . . . . . . . . 221

1 On fields and frames

The complex number field C has a 1-dimensional underlying vector space (or
inner product/Hilbert space). Implicitly, it also has a distinguished basis {1}.
A general 1-dimensional Hilbert space E has no distinguished basis, and there
is a U(1)-phase freedom in choosing a normalized basis vector.
More generally, let E be an abstract n-dimensional Hilbert space. Then
each choice of orthonormal basis gives a unitary identification E ∼ = Cn . Con-
sider the set Fr(E) of orthonormal bases {e1 , . . . , en } (“frames of E”). There

4
is a right action of U(n) on Fr(E),
X
e′j = ek ukj , u = (ukj ) ∈ U(n). (1.1)
k

This action is simply transitive, so Fr(E) is a U(n)-torsor. This means that

after we fix a reference basis {ej }nj=1 , then any other basis {e′j }nj=1 can be
labelled by the unique unitary matrix satisfying (1.1). But there is no canonical
reference basis, therefore no canonical identification Fr(E) ∼
= U(n).
Similarly, if E is a real inner product space, we have Fr(E) ∼ = O(n). If
there is no inner product, and we just want to describe the set of bases, we
have Fr(E) ∼ = GL(n).

Q : Why do we care about distinguishing identifications E ∼

= Cn ?
An n-component function X → Cn defined on some initial base space X
has, conceptually, a single “external target space” Cn . All points of X get
mapped into the same Cn by the function. Also, we are allowed to start from
any other space Y and consider functions Y → Cn mapping into this same
Cn . So the target Cn has no special relation to the domain X or Y .
The concept of a field is subtly but dramatically different. At each point
x ∈ X, a field has a value localized at x, and not at some other point x′ .
That is, each point x has its own “local target space” Ex . In total, there is a
“bundle” E = ⊔x∈X Ex , equipped with a natural projection map π : E → X.
A field is a map
v : X → E, such that v(x) ∈ Ex , ∀x ∈ X,
also called a section of the bundle E → X. Equivalently, v satisfies π◦v = idX .
A priori, the target spaces Ex , Ex′ at different points are not identified. For
example, if Ex is a vector space, then a choice of basis (frame) for Ex does not
instantaneously determine basis choices for the other Ex′ .
Now, if X and the Ex are not just sets, but topological spaces, then we
have “local groupings” of the points of X coming from the topology (the
open sets). So for an open set U ⊂ X, we can at least demand that the
Ex , x ∈ U are “continuously identifiable”. Formally, we ask for the existence
of a homeomorphism
f
Ex ∼
G
π −1 (U ) ≡ = U × F,
x∈U

5
such that π(p) = πU (f (p)). Here, F is some “reference” fibre, and U × F is
a reference “trivial bundle” with πU the projection onto the first component.
This condition is called “local trivializability”.
F The observant reader might
notice that we have not yet given x∈U Ex a topology, so what we have just
described does not really make sense! Indeed, in physics, one often comes
across informal “definitions” like “. . . the tangent bundle of a manifold is the
(disjoint) union of all its tangent spaces. . . ”. However, the global topology of
E is actually extremely important, and we will spend some time understanding
how, e.g., the topology on the tangent bundle of E is defined.
Now, if the base space X is a manifold, then we not only have a notion
of “neighbourhood”, but we also have a notion of “infinitesimally approach-
ing a point” in various directions. Then there is scope for specifying how
Ex , Ex′ should be compared, or “smoothly connected” to each other. Such
relationships between the various Ex constitute an extra piece of geometric
data, called a connection on E. In physics, connections are often called gauge
fields. However, beware that this terminology can be confusing, because a
connection is not a field in the sense of “section of some bundle over X”.
Rather, we have to first fix a basis convention for the Ex , x ∈ U (i.e., refer
to a local trivialization), then the connection will be represented as a Lie-
algebra valued 1-form on X. This local representation is called a local gauge
potential/field in physics. A key point is that the basis convention is unphys-
ical. So the important information in a gauge potential does not lie in its
values (which are convention-dependent), but in those properties which are
convention-independent, and therefore intrinsic to the connection data.
Example 1.1. In Euclidean 3-space R3 , each point x ∈ R3 has a 3-dimensional
vector space of “direction vectors” based at x. A vector field is sometimes
thought of as a “three-component function” v = (v1 , v2 , v3 ) : R3 → R3 . The
components vi (x), however, only make sense when a basis choice {eP1 (x), e2 (x), e2 (x)}
3
for the vector space attached to x has been specified; then v(x) = i=1 vi (x)ei (x).
We have to do this at every x. This detail is often neglected because Euclidean
space comes with an action of the translation group R3 , so a basis choice at
one point may be “carried over” to every other point, obtaining a global par-
allelism.
For a general manifold X, a consistent identification of the tangent vector
spaces over all points of X may not even be possible! A familiar example is
X = S 2 embedded in R3 (think of the surface of the earth). Each x ∈ S 2
has a 2-dimensional tangent plane Tx S 2 ⊂ R3 attached to it. These tangent

6
planes vary with x. Altogether, the tangent bundle is the subset
G
T S2 = Tx S 2 ⊂ S 2 × R3 ,
x∈S 2

and this subset can be shown to be a manifold in its own right. Informally,
the tangent spaces Tx S 2 ⊂ R3 depend smoothly on x, and the entire collection
T S 2 is called a smooth vector bundle. However, T S 2 is not diffeomorphic to a
Cartesian product S 2 × R2 — this is a standard result in differential topology.
Consequently, we cannot reduce a tangent vector field over S 2 into a “two-
component map” v : S 2 → R2 , no matter how we try to choose bases for the
Tx S 2 . At best, we can regard v as a “three-component map” S 2 → R3 subject
to the constraint that

v(x) ∈ Tx S 2 , ∀ x ∈ S 2.

This latter description is still rather unsatisfactory, because it depends on an

extrinsic embedding of S 2 into Euclidean 3-space.
Example 1.2. Consider the unit circle S 1 in the plane, labelled by its angular
coordinate θ ∈ [0, 2π]/0∼2π . The normal space at θ ∈ S 1 is the line

Nθ S 1 = {λ(cos θ, sin θ) : λ ∈ R} ⊂ R2 .

In total, the normal bundle is the subset

G
N S1 = Nθ S 1 ⊂ S 1 × R2 .
θ∈S 1

There is no problem identifying N S 1 with S 1 × R, via

(θ, λ(cos θ, sin θ)) 7→ (θ, λ).

Essentially, we have chosen the basis vector (cos θ, sin θ) for Nθ S 1 ⊂ R2 , and
this choice is well-defined as θ is varied around all of S 1 . So we could discuss
normal vector fields over S 1 as though they were simply real-valued functions.
Now consider the quotient of S 1 by the antipodal map — this is called
the real projective line RP1 , and it is again a circle, but parametrized by
θ ∈ [0, π]/0∼π . Attached to each θ ∈ RP1 is the line Nθ S 1 ⊂ R2 as before,
but now we think of Nθ S 1 as the “tautological line” over θ. This assignment

7
of lines is consistent at θ = 0 ∼ π, since N0 S 1 and Nπ S 1 are the same line.
Altogether we have constructed the “tautological line bundle” over RP1 ,
G
LR := Nθ S 1 ⊂ RP1 × R2 .
| {z }
θ∈RP1 LR
θ

However, we no longer have a globally well-defined choice of bases for the lines
LRθ as θ is varied over RP1 . To see this, start with the basis vector (1, 0) ∈
LR0 = N0 S 1 . For continuity, we have to choose the basis vector (cos θ, sin θ) ∈
LRθ = Nθ S 1 at θ ∈ [0, π). So upon reaching θ = π, the basis vector becomes
(−1, 0) ∈ LRπ . This has a − sign mismatch compared to the starting choice
(1, 0) ∈ LR0 .
The above-mentioned sign problem is an example of holonomy. It is a
symptom of the fact that LR is not diffeomorphic to RP1 × R.

Hopefully, we have now appreciated the difference between a function X →

Cn and a section v : X → E of a vector bundle E → X. Only after an
identification E ∼
= X × Cn has been made, can a section be considered as a
function,
∼
=
X → E ↔X ×C
x 7→ v(x) ↔ (x, f (x)), f : X → Cn .

So the section becomes the graph of some function f : X → Cn . To complicate

matters further, the examples of T S 2 and LR show that such an identification
is often impossible in the first place. So in general, we can expect a section to
be describable as functions f : U → Cn only on local patches U of X.
From the physical perspective, this “promotion” of functions to fields in
not only a matter of pedantry, but has fundamental conceptual importance.
One expects physical laws to obey locality principles; for example, we do not
expect that basis conventions at Ex , Ex′ are instantaneously identified when
x, x′ are far apart. The ideas of vector bundles, sections, and connections are
consistent with such locality considerations. These allow for manifestly gauge-
invariant formulations of the physical laws. Certain quantities may depend
(either in their definitions, computation, or for convenience) on references to
gauge choices, and we must always remember to pay special attention to those
quantities which are gauge-independent.

8
Remark. In the case where Ex ∼ = R, the freedom to choose a normalized basis
vector is labelled by a discrete set O(1) ∼
= Z2 . This may be thought of as the
choice of orientation on Ex . So there are no “local gauge freedoms” in the
sense that continuity in X forces the Ex to be identified with R in the same
way for all x in a local patch of X. This rigidity is a reason why gauge-theory
ideas hardly appear when studying classical differential equations governing
real-valued scalar systems (temperature, pressure, etc.) Subtleties arise only
when the Ex happen to be globally arranged in the manner of LR . Then, for
example, one cannot give global meaning to a quantity like “height above the
central circle of a Möbius band”.
In contrast, once we have Ex being a complex vector space, then we will
naturally encounter the local gauge freedoms. This occurs for the differen-
tial equations occurring in quantum mechanics. While at first, these gauge
freedoms seem like irritating details, they are now understood to be an indis-
pensable part of modern physics!

1.1 Tangent bundle and other bundles

Historically, the development of general relativity (perhaps better described
as “Einstein’s theory of gravitation”) was motivated by an idea of satisfying
a “principle of general covariance”. Informally, the laws of physics are “the
same” in any “reference frame”. What does this mean precisely? I do not have
a good answer, but we shall see that the language of differentiable manifolds
encodes such an idea.
One of the most important roles of manifolds is that it formalizes calculus
(e.g. derivatives, integration) in a coordinate-independent way. Put in another
way, a consistent calculus “exists out there in space”, whether or not a human
∂
mind has decided to write down symbols like ∂x .
A manifold X is locally diffeomorphic to some (open subset of) Euclidean
space, and a choice of such a local diffeomorphism provides local coordinates
for points on a local patch U of the manifold. There are plenty of choices
of local coordinates, but none is canonically given as part of the manifold
data. For functions f : X → Y between two manifolds, we have a coordinate-
independent notion of the derivative of f — it is a map on the tangent bundles
df : T X → T Y , taking the tangent space Tx X linearly to the tangent space
Tf (x) Y . Local coordinates for U ∈ X also lead to the coordinate vector fields

9
 ∂
∂xi
defined over U . The ∂x∂ i , i = 1, . . . , n provide a basis for each of the
x
tangent spaces Tx X, x ∈ U , identifying each of them with Rn . The derivative
dfx is then expressed as a matrix of partial derivatives. But df itself, as well
as rules like the chain rule, is independent of such choices.
So a smooth manifold X automatically comes with its tangent bundle T X,
and tangent vector fields over X are sections of the tangent bundle T X. Tan-
gent vector fields have long been known to arise in nature, the classical electric
and magnetic vector fields being basic examples.
For a long time, geometers agonized over the problem of relating tangent
vectors attached to different points of X, and therefore the issue of differenti-
ating vector fields. One observation was that of path dependence.
Example 1.3. For example, consider on the unit 2-sphere X = S 2 , a geodesic
triangle (i.e. following great circles) with first vertex on the North pole, and the
other two vertices on the equator. Start from a horizontal tangent vector, say
(1, 0, 0) at the north pole. Intuitively, we may parallel transport it southwards
along a longitude line y 2 + z 2 = 1, until we reach the equator. As we move
eastwards along the equator, however, (1, 0, 0) no longer lies in the tangent
plane at of the base point. Rather, after moving through a polar angle of
θ along the equator, the parallel transported tangent vector would becomes
(cos θ, sin θ, 0). Then parallel transport back to the north pole results in the
final tangent vector (cos θ, sin θ, 0).
In the above example, the space of orthonormal
 frames of Ex is O(2), and
cos θ − sin θ
we acquired a holonomy of ∈ O(2) after completing a loop
sin θ cos θ
on S 2 with solid angle θ.
We observe that we have implicitly used the round metric on S 2 , inherited
from the background Euclidean metric, in the parallel transport prescription.
First, the lengths of the tangent vectors are preserved throughout the process.
Second, the parallel transported vector maintains the same angle with the
tangent to the path. Formally, we have used the parallel transport associated
to the canonical Levi–Civita connection for the round metric on S 2 . When
dealing with Riemannian manifolds, this distinguished connection is usually
implicitly used to define covariant differentiation of tangent vector fields.
General vector bundles, not canonically arising from the base manifold
alone, are harder to find in nature, which perhaps explains why it was quite
a conceptual leap to investigate them, and develop a theory of connections on

10
them. Once we entertain the idea that the valuation spaces Ex may be inde-
pendent from the smooth structure of X, we may go further and consider the
more general notion of a fibre bundle. In a fibre bundle, all the Ex are identifi-
able with a reference object F (the “typical fibre”, generally some topological
space or manifold), but not canonically so.
Historically (∼ 1918), the first “geometrically unusual” fibre bundle was
introduced by Hermann Weyl, and it had Ex ∼ = R+ = (0, ∞) as typical fibre.
A positive number in Ex was meant to be a “length scale” at x (hence the
terminology “gauge”). The multiplicative group (R>0 , ×) acts as a “change-
of-scale” on each Ex . This idea of Weyl was quite radical — he was thinking
of a generalization of (pseudo-)Riemannian geometry, where there is a bundle
of possible local gauges, without any distinguished global section/length scale.
He wanted to relate the (R>0 , ×) freedom in the local gauges with a similar
freedom in choosing electromagnetic potentials.
Weyl’s idea was rejected by Einstein, Pauli etc. on empirical grounds, and
it remained a beautiful piece of pure mathematics. As it turns out, Weyl was
ahead of his time. As quantum theory developed and increased in mathemat-
ical precision, it became apparent that it is the quantum mechanical wave-
function, rather than the metric tensor field of general relativity, which should
be coupled with the electromagnetic potential in Weyl’s theory. The “wave-
function” has a U(1)-phase indeterminacy, while the “group of length scales”
is (R>0 , ×). These two groups are “infinitesimally the same”, in the sense of
sharing the same Lie algebra. But the difference in their global topology (one
is compact, the other is simply-connected) has important consequences.

2 Geometry and gauge in quantum theory

By itself, gauge theory is a very geometrical subject. It involves the differential
geometry of fibre bundles and connections. Even in physics, its historical devel-
opment started out in relation to general relativity (Einstein’s geometrization
of gravitation), as mentioned earlier. Unfortunately, starting directly with
the formal mathematical treatment can greatly obscure the physical motiva-
tions; conversely, the usual physical presentation tends to obscure the natural
geometrical meaning.
With a lot of hindsight, we shall start the discussion of gauge theoretic
ideas from general considerations of the mathematical structure of quantum
mechanics. One reason to do this is to emphasize the importance of geometry

11
in quantum theory from the beginning. Another reason is to avoid giving the
impression that “all fundamental physics is geometry”. Rather, the worlds of
geometry–topology and functional analysis–spectral theory intersect in a very
deep way, via operators appearing in quantum theory – this is index theory.
Again, due to the historical development of index theory (Atiyah–Singer),
it appears to be an esoteric subject which the majority of physicists and math-
ematicians can safely ignore. This is quite untrue — index theoretic phenom-
ena can already be seen in fairly elementary toy examples, although such a
perspective is not easily found in existing textbooks. I particularly highlight
the phenomenon of topological phases of matter in modern condensed mat-
ter physics, as an area which can benefit greatly from a more geometric and
universal perspective afforded by index theory.

2.1 “Wavefunctions” in quantum mechanics

In quantum mechanics, a wavefunction ψ (“quantum state”) is usually taken
to be a function ψ : X → C on a measure space (X, µ) with unit L2 -norm. One
of the most important feature of general quantum theory is that a wavefunc-
tion is not itself observable. Only real-valued transition probabilities between
quantum states,
Z 2
′ 2 ′ 2
|⟨ψ|ψ ⟩| = |⟨ψ|ψ ⟩L2 (X) | = ψ(x)ψ ′ (x) dµ(x) (2.1)
X

are. This probabilistic interpretation of |⟨ψ|ψ ′ ⟩|2 is called the Born rule. The
number ⟨ψ|ψ ′ ⟩ is usually called a transition amplitude.
The absolute value in Eq. (2.1) tells us that we may independently multiply
ψ and ψ ′ by some U(1) phases,

ψ 7→ α · ψ, ψ ′ 7→ β · ψ, α, β ∈ U(1),

without affecting the transition probability. So more accurately, quantum

states are really elements of a projective Hilbert space.
This redundancy of the global U(1) phase in ψ is a general feature of quan-
tum mechanics. So far, the information of X has not entered yet. Now observe
that at each point x ∈ X, we are actually free to multiply the ψ(x) and ψ ′ (x)
by a common U(1)-valued phase function u(x), without changing the transi-
tion amplitude in Eq. (2.1). This follows from the pointwise sesquilinearity of
the L2 -inner product. It appears that the phase information contained in ψ,

12
 ψ
i.e., |ψ| : X → U(1), is not directly observable. That is, ψ is overdetermined
as a complex-valued function.
How should we understand ψ “modulo phase redundancy”? If we make
the phase information totally redundant, and work just with |ψ|, then we do
get a real-valued scalar function: the probability density of ψ,

ρψ : X → R, ρψ (x) = |ψ(x)|2 ,

which of course integrates to 1,

Z
|ψ(x)|2 dµ = ||ψ||2L2 (X) = 1,
X

as required for a probabilistic interpretation

R of ψ. Specifically, fix any measur-
able subset Z ⊂ X. Then the integral Z ρψ is the probability of finding the
particle in the set Z, when the particle is in the quantum state ψ. Indeed, for
a single state ψ, these probabilities do not require the phase information.
However, the transition amplitude between two states ψ, ψ ′ involves the
relative phase information between ψ(x) and ψ ′ (x),
Z Z
′ ′
⟨ψ|ψ ⟩L2 (X) = ψ(x)ψ (x) dµ(x) ̸= |ψ(x)||ψ ′ (x)| dµ(x) = ⟨|ψ| | |ψ ′ |⟩L2 (X) .
X X

Thus, “local U(1) phase information” is not completely redundant. In par-

ticular, relative phase information is essential in quantum mechanics. The
redundancy is, more precisely, in a simultaneous redefining all wavefunctions
ψ by any phase function u : X → U(1),

ψ ⇝ u · ψ.

This operation is an abelian gauge transformation.

The upshot is that we should not think of ψ as having complex-number
values ψ(x) ∈ C. Instead, each ψ(x) is an element of a 1-dimensional Hilbert
space Ex , where Ex is “attached to x”. The Ex does not come with any
canonical choice of basis vector, so ψ(x) does not have a canonical phase-
value. Nevertheless, at each x, the relative phase between ψ(x), ψ ′ (x) ∈ Ex is
well-defined, whatever the choice of basis for Ex .
So we are precisely in the situation discussed in Section 1. We have a
family E = ⊔x∈X Ex of 1-dimensional Hilbert spaces, and a wavefunction ψ is
really a “section” ψ : X → E such that ψ(x) ∈ Ex for all x ∈ X. Formally,

13
we write ψ ∈ L2 (X; E). To turn ψ into a complex-number-valued function,
∼
=
we need a reference global trivialization h : E → X × C, i.e. a choice of basis
for every Ex . Composing ψ with the trivialization gives the map
ψ h
X → E → X × C, x 7→ ψ(x) 7→ (x, f (x)),

where f ∈ L2 (X) is now some function, representing ψ with respect to the

trivialization h.
Let us now consider the operation of pointwise multiplication by some
phase function u : X → U(1). This operation gives another possible way of
trivializing E,
h (id,u)
E → X × C → X × C.
Basically, we are “changing the gauge”, and correspondingly, the section ψ is
now represented by the modified function u · f . Notice that the gauge trans-
formation (multiplying everything pointwise by u) is implemented unitarily on
the Hilbert space L2 (X), as any symmetry must be in quantum mechanics.
So we see that transition amplitudes ⟨ψ|ψ ′ ⟩L2 (X) are gauge-invariant, thus
observable, quantities.

2.1.1 Differentiating quantum states

Usually, X is a smooth manifold. Then we would like E = ⊔x∈X Ex to be
a smooth family of 1-dimensional Hilbert spaces. Formally, we require E to
be a manifold, which admits a diffeomorphism E ∼ = X × C that restricts
∼
to isomorphisms Ex = C. In this case, E is called a smooth, trivializable
Hermitian line bundle over X. A particular choice of identification E ∼ = X ×C
is called a trivialization of E, and it is an unphysical extra piece of data.
Now, we need to take derivatives of ψ, e.g., in the Schrödinger differential
equation governing the time-evolution of ψ. This requires a prescription for
comparing the vector spaces Ex and Ex′ at infinitesimally nearby points x, x′ —
a connection ∇ on E. This data provides a notion of “covariant differentiation”
of sections of E. To describe a connection concretely, we might decide to pick
a trivialization E ∼ = X × C, and compare ∇ with the “trivial connection” ∇0
corresponding to the usual derivatives ∂x∂ j of functions. Then ∇ becomes the
covariant derivatives ∂x∂ j − iAj , where Aj are the real-valued components of
a “gauge potential”. But we must not attach too much meaning to the Aj ,
since they require reference to a trivialization (“gauge-dependence”). As we

14
will learn, certain aspects of Aj are nevertheless intrinsic, e.g., its curvature,
holonomy, etc.
But what determines the connection on E? The section ψ is, typically,
supposed to be the “electron field”. As it turns out, Aj is identifiable with the
magnetic vector potential, and its curvature (“curl”) is identifiable with the
magnetic field that the electron is subject to. Now, in classical electromag-
netic theory by itself, one learns that the potentials Aj are “unphysical”, or
“fictitious”, or “redundant” data, and only the electric/magnetic fields they
determine are physical. But in quantum theory, the Aj have a more im-
portant role in specifying connection data for the quantum line bundle E.
More to the point, we must be careful about what precise aspect of Aj is es-
sential/physical in quantum theory. The answer is not “the potential Aj is
measurable” (literally false); rather, the connection represented by the Aj is
quantum mechanically measurable.
Here, it is not just the curvature of the connection which can be measured
(this is just the classical magnetic field). More pertinently, the holonomy of
the connection (“geometric phase”) along a path can be measured, in the so-
called Aharonov–Bohm effect, via interference experiments. This put an end
to a long confusing debate on the physical status of potentials.
Remark. In introductory courses, it is often said that electromagnetism is the
motivating and simplest example of a “U(1)-gauge theory”. This is supposed
to be due to a “gauge freedom” to modify the potentials,

Aj ⇝ Ãj := Aj + ∂j Λ,

where Λ is any smooth real-valued function on X, without changing the curl

of (A1 , A2 , A3 ), i.e., the magnetic field. In this setting, there is certainly a
“freedom”, but the reason for using the word “gauge” is completely mysterious.
For example, as far as I know, there is nothing which determines whether
the Aj are gauge potentials for a U(1) or (R>0 , ×) principal bundle (jumping
ahead quite a bit). More plainly, the “local symmetry group” could be either
one, as far as I know. As mentioned, Weyl initially thought of using (R>0 , ×).
In any case, one just obtains a reformulation of electromagnetism, and the
core gauge theory ideas do not really enter.
On the other hand, quantum theory inherently demands that E is associ-
ated to a U(1) principal bundle. The fact that the connection on E can be
identified with the magnetic potential is an empirical finding.

15
 When we change Aj to Ãj , thus ∂j − iAj is changed to ∂j − iÃj , physicists
observed that we can “compensate” by simultaneously replacing

ψ ⇝ ψ̃ := eiΛ · ψ. (2.2)

What is really happening is: when we apply eiΛ to change the gauge, this
effects the description of ψ according to Eq. (2.2), and also effects

eiΛ (∂j − iAj )e−iΛ = eiΛ e−iΛ (∂j − i∂j Λ − iAj ) = ∂j − iÃj

for the operation of taking derivatives. That is, the differentiation of the
electron field is also described in a modified way, when we change gauge.
Remark. In the seminal paper, Elektron und Gravitation. I. Z. Phys. 56, 330
(1929), H. Weyl proclaims:
. . . In my opinion the origin and necessity for the electromagnetic field is
in the following. . . From the arbitrariness of the gauge-factor in ψ appears the
necessity of introducing the electromagnetic potential. . .
To paraphrase, geometric considerations in quantum theory lie at the heart
of what is nowadays called the “gauge principle”.
Remark. There is another important setting in which connections on vector
bundles appear in quantum theory. This is when we have quantum systems
parametrized by some other manifold M (of control parameters). There is
a notion of “adiabatic parallel transport” of quantum states, which leads to
the notion of “Berry phase”, or “geometric phase” in physics. This was soon
clarified by Barry Simon to be the holonomy of a certain connection on a line
bundle over M .

All of the above discussion generalizes to vector bundles where Ex is n-

dimensional. We can also have Ex being real inner product spaces. For in-
stance, the tangent spaces of a Riemannian manifold have such a structure.
In the real case, the gauge transformations would be O(n)-valued, and there
is often occasion to restrict to SO(n)-valued ones (i.e. orientation-preserving
ones).
Finally, the global topology of vector bundles is extremely interesting and
important. This studies phenomena that arise when E is not globally trivializ-
able, i.e., no homeomorphism/diffeomorphism E ∼ = X × Cn is available. Only
local trivializations over U ⊂ X exist. Without a global trivialization, there
is no notion of globally-defined wavefunctions in the first place! Similarly, the

16
formulation of gauge transformations, connections, and so on, becomes more
subtle.
We saw such global issues in the example of T S 2 . Actually, S 2 is a complex
manifold, and the tangent spaces Tx S 2 can be thought of as 1-dimensional
complex vector spaces. The Dirac monopole involves, roughly speaking, “half
of T S 2 ”. Namely, there is another complex line bundle LC over S 2 , called
the Hopf line bundle, which is not diffeomorphic to S 2 × C. In LC , we can
never have access to genuine wavefunctions defined on all of S 2 . We only have
“wavesections”, which can only be locally turned into functions by means of
local trivializations. Similarly, a connection on LC can only be locally described
by gauge potentials. We will discuss the Dirac monopole and LC in more detail
later on.

17
2.2 Time evolution in quantum mechanics
Remark. We will discuss unbounded Hilbert space operators and their spec-
trum, and some mention of their technicalities is unavoidable. If you are un-
familiar with these (important!) analytic issues, there will be a more detailed
discussion later on.
Quantum mechanics on a Riemannian manifold (X, g) is usually done on
a Hilbert space such as L2 (X), with the measure induced by the metric g. We
suppress the line bundle E for simplicity of discussion. The classical symme-
tries of (X, g) are the isometries, which actually form a Lie group (a result of
Myers–Steenrod). You may think of a Lie group as a “smooth group”, which
is meant to act on a smooth manifold.
On L2 (X), the isometry group of X has a natural unitary representation
via pullback,

(Ts · ψ)(x) = ψ(s−1 · x), s ∈ Isom(X), ψ ∈ L2 (X).

Symmetry operations, such as the above, should preserve transition amplitude,

Eq. (2.1). This is why they must be represented unitarily. For example, we
also saw that gauge transformations of multiplication by u : X → U(1) are
unitary. (Actually, antiunitary symmetries are allowed as well.)
Next, we discuss time evolution. The equation of motion for quantum
states ψ is the Schrödinger “wave” equation
∂
i ψ = Hψ,
∂t
where H is a self-adjoint operator called the Hamiltonian, and ψ belongs to
its domain (a dense linear subspace of the Hilbert space L2 (X)). Self-adjoint
operators are infinite-dimensional versions of Hermitian matrices.
For a free particle on X, the Hamiltonian operator would be Hfree = −∇2 ,
the Laplace operator on X. This is the square of the momentum operator
d 2
−iℏ dx , just like the classical kinetic energy is momentum
2×mass
. Hereafter, we ignore
1
the 2m factor and the Planck constant ℏ for notational convenience. There
are also the Schrödinger operators H = −∇2 + V , where V is the operator of
multiplication by some (real-valued) potential function on X.
There are two major questions in the study of Schrödinger operators. First,
does H = −∇2 + V really make sense R as a self-adjoint operator? If so, then
the spectral theorem applies, H = R λ dµ(λ), where µ is the spectral measure

18
of H. This is the infinite-dimensional version of unitary diagonalization. If λ
is a (real) eigenvalue of H with eigenfunction ψ (in the domain of H), then
the Schrödinger equation simplifies to
d
i ψ = λψ ⇒ ψ(x, t) = e−iλt ψ(x).
dt
So an eigenfunction evolves in time in a very simple way — it acquires a
dynamical phase factor which rotates at a rate dependent on its eigenvalue λ
(physically, the energy).
The spectrum of the linear operator H generalizes the notion of eigenvalues
of matrices. Self-adjointness of H implies that the spectrum of H is a subset of
the real line. Furthermore, it allows one to make sense of functions of H (the
functional calculus). A self-adjoint H exponentiates to a strongly-continuous
1-parameter family of unitary time-evolution operators,
Ut = e−iHt , t ∈ R.
Here, strong continuity means that for each fixed state ψ, the map t 7→ ||Ut ψ||
is continuous. But the continuity is not generally uniform in the choice of
ψ. Conversely, any strongly-continuous time-evolution group is generated by
a unique self-adjoint operator. This correspondence between unitary time-
evolution groups and self-adjoint operators is of fundamental importance in
quantum mechanics, and is a theorem due to M. Stone.
In principle, knowing the spectral decomposition of H, thus its time-
evolution group, amounts to solving the Schrödinger equation for ψt (x) =
ψ(x, t) for any time t, when the initial t = 0 quantum state ψ0 = ψ(x, 0) is
given:
ψt = Ut ψ0 .
So the second major question is: what can we say about the spectrum of
H? The spectrum is exactly solvable only for some simple toy models (e.g.
harmonic oscillator, Landau operator).
General quantum mechanics also allows for the Hamiltonian H to be a
first-order differential operator, and this is especially important in relativistic
quantum mechanics. The key example is the Dirac operator, which is roughly
a “square root of Laplacian”. The Schrödinger equation with H being the
Dirac operator (instead of Laplacian) is sometimes called a Dirac equation.
Importantly, the Dirac operator acts on square-integrable sections of spinor
bundles (a special kind of vector bundle), rather than line bundles.

19
3 Dirac operators in one dimension
3.1 The many faces of R
Write R for the 1-dimensional vector space, and let R denote the underlying
affine space. Once an origin is chosen for R, each point of R is uniquely
labelled by the element x ∈ R which translates the origin to that point. So we
get a global coordinate function x on the manifold R.
There is also a metric on R given by the Euclidean distance |x − x′ |. Then
the additive group (R, +) acts isometrically on the Riemannian manifold R
by translations.
The 1-form dx is translation invariant, orients R, and defines a translation-
invariant measure on R (the Lebesgue measure in this case). When we talk
about integrating “functions” f on R, we are actually integrating the 1-form
f dx over the manifold R.
So G = (R, +) is a Lie group of orientation-preserving isometries on the
oriented Riemannian manifold (R, dx2 ). It is convenient to view the Lie group
(R, +) as a subgroup of invertible matrices,
  
1 s
G= : s∈R ∼ = (R, +).
0 1

Then its Lie algebra g is the vector space

  
0 s
g= : s∈R
0 0

equipped with the trivial commutator (Lie bracket). The Lie group and Lie
algebra are related by the exponential map g → G,
   
0 s 1 s
exp = .
0 0 0 1

Each Lie group element (labelled by s) is represented

 on
 R as the isometry
0 s
Ts of translation-by-s. Each Lie algebra element is represented by
0 0
the Killing vector field s∂x on R. That is, the derivative operator is the
“infinitesimal version” of translations; symbolically,

exp(s∂x ) = Ts . (3.1)

20
The classical interpretation of Eq. (3.1) is that the Killing vector field s∂x
exponentiates to an “isometric flow” Ts : x 7→ x + s of points on the manifold
R.
In quantum mechanics, we care about wavefunctions ψ ∈ L2 (R) rather
than individual points. In this setting, the Lie group of translation symme-
tries is unitarily represented by the operator (Ts ψ)(x) = ψ(x − s). This is
called the (left) regular representation of R on L2 (R). As for ∂x , it is formally
a skew-adjoint operator with respect to the L2 -inner product (verify this using
integration-by-parts). It is customary to consider the formally self-adjoint op-
erator P = −i∂x instead. The self-adjoint P is called the momentum operator
on R, and we have exp(−isP ) = Ts .
The point of this discussion is to list down some of the ingredients used in
quantum theory. Some of these are

ˆ Number line, spectrum of self-adjoint operator lies in R;

ˆ Vector/affine space, coordinates;

ˆ Riemannian manifold, isometry group;

ˆ Hilbert L2 -space with respect to Riemannian measure;

ˆ Representation of symmetry group (e.g. isometries) and/or Lie algebra.

When working in Euclidean space, the symbol R appears in all of the above
roles, which unfortunately obscures the importance of the geometric data in
the setup of a quantum mechanical problem.

3.2 Dirac operators on complete 1-manifolds

3.2.1 Dirac operator on oriented Euclidean line
Fix an origin, and write Q for the “position operator” of multiplication by the
coordinate x,
(Q · ψ)(x) = xψ(x), ψ ∈ L2 (R).
Of course, Q is an unbounded operator, which would take L2 -functions outside
of L2 . So we should actually restrict its domain to the subspace

Dom(Q) = {ψ ∈ L2 (R) : Q · ψ ∈ L2 (R)}.

21
It may be shown that Q is self-adjoint on this domain. The spectrum of Q,
denoted σ(Q), is the whole real line R. For example, an “eigenfunction” of Q
with eigenvalue λ is the delta “function” concentrated at λ.
d
A simple calculation (integrate by parts) shows that P = −i dx is formally
self-adjoint on the compactly-supported smooth functions,

⟨ψ|P ψ ′ ⟩L2 (X) = ⟨P ψ|ψ ′ ⟩L2 (X) , ψ, ψ ′ ∈ Cc∞ (R).

Actually, P is essentially self-adjoint on Cc∞ (R), meaning that its closure

(again denoted P ) is a genuine self-adjoint operator. The precise domain of
self-adjointness is a Sobolev space (usually denoted H 1 (R)), which we will
study in the future. We simply mention that non-smooth functions in L2 (R),
such as e−|x| , are allowed, and the derivative is interpreted as a weak derivative
(distributional sense).
d
From now on, we write P for the self-adjoint operator −i dx . As it turns out,
P is the precisely the (right-handed) Dirac operator on the Riemannian (spin)
manifold X = R. Informally, a Dirac operator is a first order differential
operator which squares to the Laplace operator, up to lower order terms.
Certainly, P 2 = −∇2 .
Now, think of P as the quantum mechanical Hamiltonian operator entering
the Schrödinger wave equation,
∂
i ψ = P ψ.
∂t
The time evolution group generated by P is

e−itP = e−t∂x = Tt ,

i.e., the group of translations. So the time evolution of ψ = ψ(x, t) ≡ ψt (x) is

simply
ψt = Tt ψ0 .
Thus any wavefunction simply propagates to the right at unit speed (the speed
of light, if physical units are restored). There is also a left-handed Dirac
operator, −P , which propagates quantum states to the left.
Remark. This relativistic property was one of the original motivations for
Dirac’s introduction of the concept of Dirac operators. In 3 spatial + 1 time
dimensions, the massive Dirac equation requires ψ to be a Dirac spinor field
with four components.

22
 To obtain the spectrum of P , we can perform a Fourier transform L2 (R) →
2
L (R̂). The reason for the notation R̂ is because it refers not to the abelian
translation group R, but its Pontryagin dual abelian group R̂ = Hom(R, U(1)),
also known as the momentum space. Namely, the real number p ∈ R̂ labels
the character (irreducible representation)
χp : x 7→ eipx ,
and the characters themselves obey a composition law, χp1 χp2 = χp1 +p2 . On
R̂, there is again the Lebesgue measure, defining the Hilbert space L2 (R̂).
The Fourier transform
F : L2 (R) → L2 (R̂)
Z ∞
1
ψ 7→ ψ̂ ψ̂(p) = √ e−ipx ψ(x) dx
2π −∞

is unitary (we take this for granted for now), and it conjugates P into the
operator of multiplication by the coordinate p ∈ R̂. So the spectrum of P is
the whole real line. In fact, the plane wave x 7→ eipx is an “eigenfunction”
of P corresponding to the “eigenvalue” p. If you know some spectral theory,
d
this is basically the spectral theorem applied to the operator P = −i dx . The
Fourier transform is the unitary “change of basis” which “diagonalizes” P into
the operator of multiplication by its “eigenvalues”.
Now we have a pair of self-adjoint operators (Q, P ), and they are easily
seen to satisfy the canonical commutation relation
[Q, P ] = i.
A famous theorem of Stone–von Neumann says that any pair of self-adjoint
operators Q′ , P ′ on a Hilbert space satisfying [Q′ , P ′ ] = i (more precisely, the
exponentiated version eisQ eitP = eist eitP eisQ called the Weyl relations), must
be unitarily equivalent to the standard pair (Q, P ).
Here is an interesting example: Let
Q′ = Q, P ′ = P + A, A ∈ C ∞ (R; R)
where A is (the multiplication operator by) any smooth real-valued1 “potential
function”.
1
If A is bounded, then P + A has the same domain of self-adjointness as P ; otherwise a
bit more care is needed to understand what P + A means.

23
 Clearly A commutes with Q, so that we still have

[Q′ , P ′ ] ≡ [Q, P + A] = [Q, A] = i.

(More precisely, (Q′ , P ′ ) can be shown to satisfy the Weyl relations.) There-
fore, by the Stone–von Neumann theorem, there must be a unitary operator
U on L2 (R), such that

U QU −1 = Q′ = Q, U P U −1 = P ′ = P + A.

The required U is nothing but the gauge transformation of multiplication by

the function
−i xx A
R
u : R → U(1), u(x) = e 0 .
Here, the basepoint x0 is arbitrary, and only leads to an overall phase factor in
u. You may recognize u as the integrating factor for the first-order differential
d
operator −i dx + A(x).
In the reverse direction, suppose U is the gauge transformation by a U(1)-
valued function u = e−if with smooth real-valued f . Write A = f ′ for the
d
derivative of f . Then U will conjugate P = −i dx into the operator

d d
U P U −1 = U (−i )U −1 = −i( + iA) =: −i∇A = P + A.
dx dx

Conclusion. Any of the operators P + A represents “the” Dirac operator

on the Euclidean line R. We say that A is “pure gauge” in the sense that it
can be set to zero by working in an appropriate gauge.

3.2.2 Dirac operators on a circle

The Euclidean line R is topologically trivial in the sense that it is contractible
to a point. This property allows any A dx to be expressible as A = df for
some well-defined smooth function f on R, thereby giving the u which gauge-
transforms it to zero.
Let us now work over a compact 1-dimensional Riemannian manifold X.
Parametrizing by path-length, we see that X is actually isometric to a circle of
some radius r in the Euclidean plane. Without loss of generality, we study the
d
momentum operator P = −i dθ on the unit circle X = S 1 , oriented counter-
clockwise, and parametrized by θ ∈ [0, 2π]/0∼2π .

24
 The operator P is self-adjoint on the Sobolev space H 1 (S 1 ), which we can
think of as comprising those functions ψ whose Fourier transform (ψn )n∈Z is
such that (nψn )n∈Z is square-summable. The eigenvalues and eigenfunctions
of P are easily verified to be
ψn : eiθ 7→ einθ , P ψn = nψn .
So the spectrum of P is σ(P ) = Z.
So far the story seems quite elementary and uninteresting. However, now
observe that in declaring the initial domain as the function space C ∞ (S 1 ),
we have committed to a common choice of basis for the copy of C above each
eiθ ∈ S 1 . For m ∈ Z, let us perform the gauge transformation of multiplication
by the smooth function um : eiθ 7→ e−imθ . This converts P to the operator
d −1 d
P (m) := um P u−1
m = um (−i )um = −i + m = P + m.
dθ dθ
The eigenfunctions of P (m) are still φn , but the eigenvalue of φn is now shifted
to n + m. We still have σ(P (m) ) = Z, independently of the value of m, which
is consistent with the unitarity of the gauge transformation. Therefore any
integer potential term m ∈ Z is “pure gauge”, and can be transformed away.
d
What about the operator P (ξ) := −i dθ + ξ for a potential term ξ ∈ (0, 1)?
Obviously, the spectrum becomes shifted by ξ,
σ(P (ξ) ) = Z + ξ ̸= σ(P ).
So Pξ must be unitarily inequivalent (in particular, gauge-inequivalent) to P .
In fact, the operators Pξ , ξ ∈ [0, 1) must all be mutually gauge-inequivalent!
Here, the real value of ξ is not gauge-invariant due to the possibility of modify-
ing it by an integer m through the gauge transformation um . But e2πiξ ∈ U(1)
(equivalently, ξ mod Z) is gauge-invariant (check this!); it is the holonomy of
a U(1)-connection on the line bundle S 1 × C.
So we see that there are many gauge-equivalence classes of connections on
the same (trivial) line bundle. Each such connection defines a twisted Dirac
operator P + ξ. Incidentally, the family {P + ξ}ξ∈[0,1] provides an example of
spectral flow, see Fig. 1, about which we will have much more to say.

3.2.3 Aharonov–Bohm effect

Now, ξ dθ is supposed to be the magnetic potential 1-form (The difference and
relation between vector fields and differential 1-forms will be clarified in later

25
 σ(P + ξ)

ξ
1
−1

−2

−3

Figure 1: Spectral flow of family of Dirac operators on S 1 , as the holonomy

parameter ξ ∈ [0, 1]/0∼1 is increased in a loop.

lectures). Here, we are picturing S 1 to lie on the horizontal plane of R3 , and

we work in cylindrical coordinates (r, θ, z). This gauge potential has vanishing
curvature, d(ξ dθ)) = 0, corresponding to zero magnetic field. A key point is
that dθ is only defined on R3 with the z-axis removed, so the vanishing of the
magnetic field holds only for the region r > 0.
The operator Pξ is the Hamiltonian operator for a Dirac particle on a
circle, subject to zero magnetic field. Not quite! Remember that the relevant
data for the quantum theory is not the magnetic field, but the connection, or
alternatively, the gauge potential ξ dθ. The magnetic field just happens to be
one of the intrinsic features of the connection (namely, its curvature). So Pθ
is the Dirac operator on the circle with respect to this connection data.
Still, there is another point of confusion. How is this situation different
from that of the Euclidean line, where the gauge potential A dx could always
be transformed away by an appropriate gauge transformation? After all, isn’t
dθ precisely the differential of the coordinate function θ on the circle? Why
can’t we just apply the gauge transformation uξ (θ) = e−iξθ as before?
We come to the second key point: θ is not a genuine globally defined
function on all of S 1 . The topology of S 1 , specifically the failure of simply-
connectedness, prevents us writing dθ as the derivative of a function on S 1 —

26
the gauge potential ξdθ is not an exact differential form. The candidate “gauge
transformation” uξ is necessarily discontinuous at one point. Multiplication
by uξ is a unitary transformation, so it would seem that conjugating P by uξ
should still produce “P + ξ”. This is not true: we must be extremely careful
with the domains of self-adjointness for unbounded operators on Hilbert space,
as explained in Section 3.3.4.
Remark. In cartesian coordinates, the 1-form ξ dθ is (exercise)
ξ
ξ dθ = (−y dx + x dy),
x2 + y2
ξ
which corresponds to the vector potential A(x, y, z) = x2 +y 2 (−y, x, 0). The

curl, which is the magnetic vector field, is easily seen to be B = ∇ ∧ A = 0

(away from z-axis). In practice, we consider a very thin cylinder around the
z axis, of small radius r0 ≪ 1, say. By Stokes’ theorem, the line integral of A
along any loop C around the thin cylinder is
I I Z Z
A · dx = A · dx = 2πξ = B · dS = B · dS,
C S1 B1 Br0

where the last surface integral is the magnetic flux through the cross section
of the cylinder. This is because the support of the magnetic field is contained
in this cylinder. Classically, the value of ξ ∈ R is relevant and measurable only
inside this cylinder, via the magnetic flux.
The term ξ dθ has a different status in quantum theory — as a gauge
potential representing the U(1)-phase connection data. Away from the small
cylinder, this gauge potential is non-vanishing, and intrinsically so — there is
no U(1)-valued gauge transformation which can get rid of ξ dθ.
Consider the following operator on S 1 ,
   d 
P +ξ 0 −i dθ + ξ 0
Dξ = = d .
0 −(P + ξ) 0 i dθ −ξ
This is a toy Hamiltonian operator which will illustrate the so-called Aharonov–
Bohm effect.
The self-adjoint operator Dξ governs the time-evolution of a two-component
wavefunction ψ = ψψ12 . Let the initial wave function be

 
ψ1 (θ; 0)
ψ(θ; 0) ≡ ψ(θ; t = 0) = ,
ψ2 (θ; 0)

27
where θ 7→ ψi (θ; 0) is some localized function on S 1 , with center θ0 , say. Ac-
cording to the Schrödinger equation with Hamiltonian Dξ , the first component
ψ1 propagates anticlockwise, evolving in time as (Exercise)
R θ−t
ψ1 (θ; t) = ei θ ξ dθ
ψ1 (θ − t; 0) = e−itξ ψ1 (θ − t; 0).

Similarly, ψ2 evolves clockwise as

ψ2 (θ; t) = eitξ ψ1 (θ + t; 0).

So after time t = π, the localized ψ1 , ψ2 meet again at the antipodal point

θ0 + π. The total wavefunction at t = π is
 −πiξ   
e ψ1 (θ − π; 0) −iπξ ψ1 (θ − π; 0)
ψ(θ; π) = =e ,
eπiξ ψ2 (θ + π, 0) e2πiξ ψ2 (θ − π; 0)

and is centered at θ0 + π. This final wavefunction is almost the same as the

original one, except for a relative phase factor e2πiξ between the two compo-
nents (and the propagation from θ0 to θ + π of course). Such a relative phase
is observable in interference experiments. Geometrically, this phase factor is
the holonomy of the connection 1-form ξ dθ around S 1 .
Remark. The Aharonov–Bohm phase shift can also be derived for Schrödinger
particles, but it is somewhat harder to make analytically precise without some
handwaving.
In any case, the significance of the effect is this: despite the classical mag-
netic field being zero everywhere along the path being sampled by the wave-
function, the magnetic potential, correctly understood as a gauge potential,
is physically meaningful in quantum mechanics. Furthermore, there are many
“inequivalent quantum theories” on S 1 , despite each of them being classically
the same (zero magnetic field).

28
3.2.4 Anomaly
Returning to the real line R, a simple consequence of P + A being unitar-
ily equivalent to P , is that P + A always has spectrum being the entire R.
Similarly, there is a “spectrally rigid” left-handed Dirac operator, represented
as −(P + A). These two Dirac operators are sometimes called chiral Dirac
operators, or Weyl operators.
On its own, a chiral Dirac operator becomes problematic in quantum field
theory. This is a very complicated subject, so let us take a simplified per-
spective without all that baggage. Now, we have seen that we should think
of P + A as “the same operator”, regardless of A, since A is really an artifact
of gauge choices. For simplicity of discussion, take A to be a real constant
function. Then replacing P by P + A shifts the energy by A, so we would
appear to lose any canonical meaning of “0-energy”. In quantum field theory,
or “second quantization”, this leads to the problem of defining the “vacuum”
in a gauge-invariant way.
Now consider the direct sum, D / A = (P + A) ⊕ −(P + A), of a left-handed
and a right-handed Dirac operator, which will still have full spectrum in R.
However, the spectral gaplessness is no longer rigid. Specifically, we can add
an off-diagonal “mass term”,
   
P +A m gauge P m
D/ A + M := −→
m −(P + A) m −P
 
Fourier p m
−→ , m ∈ R, p ∈ R̂.
m −p
| {z √ }
eigenvalues ± p2 +m2

/ A + M is
So the spectrum of D
[ p p
σ(D/A + M) = {− p2 + m2 , + p2 + m2 } = (−∞, −|m|] ∪ [|m|, ∞).
p∈R̂

Independently of A, a spectral gap (−|m|, |m|) opens up as long as m ̸= 0. The

“vacuum level” can now be set to 0 without ambiguity. The operator D /A + M
is called the massive Dirac operator with mass m.

29
3.3 Dirac operators on incomplete 1-manifolds
3.3.1 Momentum operator on a half-line?
Consider the half-Euclidean line R+ = (0, ∞). Note that R+ is diffeomorphic
to R, but they are not isometric. A crucial difference between R and R+ is
that the former is complete, but the latter is not. This means that geodesics
in R+ do not extend indefinitely, but will “hit the boundary” x = 0.
This incompleteness causes severe problems for the definition of a self-
adjoint momentum operator on L2 (R+ ). In a naı̈ve construction, we will start
with P = −i dx d
defined on Cc∞ (0, ∞), then extend to a suitable domain of
self-adjointness. Suppose this is possible, then we would obtain a unitary
time-evolution group Ut = e−iP t on L2 (R+ ). This time-evolution applies, in
particular, to ψ ∈ Cc∞ (0, ∞), and we know that Ut will just translate such a
ψ by a distance t. However, for t ≪ 0, we would have ψ hitting the bound-
ary and “disappearing”, violating the unitary and hence the conservation of
probability. So, the assumption that P could be made self-adjoint must be
incorrect! Let us see what the problem is.
First, a linear operator L with domain Dom(L) is formally self-adjoint (or
symmetric), if
⟨Lφ|ψ⟩ = ⟨φ|Lψ⟩, ∀ φ, ψ ∈ Dom(L).
For example, P = −i dx d
with Dom(P ) = Cc∞ (R+ ) is formally self-adjoint. As
we will learn later, this is not enough to guarantee that its spectrum is real.
Generally, the domain of the adjoint operator L∗ is defined to be the set
of φ ∈ L2 (R+ ) such that there exists some η ∈ L2 (R+ ) with
⟨η|ψ⟩ = ⟨φ|Lψ⟩, ∀ψ ∈ Cc∞ (R+ ) = Dom(L).
Then for such φ, we define L∗ φ = η. This is the precise meaning of the
equation defining the adjoint operator,
⟨L∗ φ|ψ⟩ = ⟨φ|Lψ⟩, ∀ψ ∈ Dom(L), φ ∈ Dom(L∗ ).
Notice that the larger the domain of L, the smaller the domain of L∗ .
Now take L to be the formally self-adjoint P . Then P ∗ will be an ex-
tension of P to a larger domain. For example, Dom(P ∗ ) contains functions
φ ∈ Cc ([0, ∞)) without Dirichlet condition at x = 0:
Z ∞ Z ∞
∗
⟨P φ|ψ⟩ = −iφ′ (x)ψ(x) = − iφ(x)ψ ′ (x) = ⟨φ|P ψ⟩, ∀ψ ∈ Cc∞ (R+ ).
0 0

30
No boundary term appears in the integration-by-parts because ψ ∈ Dom(P )
always has ψ(0) = 0.
On the other hand, P ∗ has no chance of being formally self-adjoint: take
any ψ ∈ Dom(P ∗ ) with ψ(0) ̸= 0. Then a short calculation shows that

⟨P ∗ ψ|ψ⟩ − ⟨ψ|P ∗ ψ⟩ = −i|ψ(0)|2 ̸= 0.

In fact, this calculation shows that we cannot enlarge Dom(P ) by any non-
Dirichlet ψ, otherwise formal self-adjointness will immediately be violated.
There is an impasse: P is forced to have Dirichlet domain, so its adjoint
will have non-Dirichlet functions. So there is simply no way of picking domains
such that P = P ∗ .
Let us stress that P ̸= P ∗ at the level of unequal domains, is not merely
a technical issue. For example, φ(x) = e−x ∈ Dom(P ∗ ), and is clearly it is an
eigenfunction,
P ∗ φ = iφ,
with imaginary eigenvalue!

3.3.2 Dirac operators on half-line?

The situation can be rectified by taking
 
P 0
/=
D , Dom(D) / = Cc∞ (R+ ) ⊕ Cc∞ (R+ ).
0 −P
∗
/ has domain without the Dirichlet condition.
As before, the adjoint operator D
This domain is too big for formal
self-adjointness to hold, so we need to cut it
down. Let us only allow ψ = ψψ12 which satisfy a boundary condition,

ψ2 (0) = αψ1 (0), α ∈ U(1).

Intuitively, α−1 is the phase-shift encountered when a left-moving ψ1 is re-

flected at x = 0 and turned into a right-moving ψ2 . Indeed, imposing this
boundary condition and completing the function space to the Sobolev space
H 1 (R+ ) (subject to boundary condition) gives a genuine self-adjoint operator.
The conclusion is that there is a family of self-adjoint D / α = P ⊕ −P ,
parametrized by α ∈ U(1). Let us mention that this relative phase shift
parameter α is gauge-invariant, so we could have replaced P by any P + A in
the above discussion.

31
 /α + M)
σ(D
σess
m
0 Arg α
π 2π
−m
σess

Figure 2: Spectral flow phenomenon for massive Dirac operators on a half-line,

parametrized by reflection boundary condition α ∈ U(1).

In a later lecture, we will learn about the rigorous and systematic theory
of self-adjoint extensions, following von Neumann. Our more immediate task
is to figure out what the spectral meaning of the parameter α is (if any).

3.3.3 Massive Dirac operators on half-line

We also have the massive Dirac operators on the half-line,
 d 
−i dx m
/ α +M =
D d , Dom(D / α +M ) = {ψ ∈ H 1 (R+ )⊕2 : ψ2 (0) = αψ1 (0)},
m i dx
/ + M on the full line R, it may be shown that the
As with the operator D
/
spectrum of Dα + M on R+ still has an “essential part” having a spectral gap,
/ α + M ) = (−∞, −|m|] ∪ [|m|, ∞).
σess (D
However, extra discrete eigenvalues may now appear within the interval (−|m|, |m|).
An explicit calculation reveals that for certain values of α, an eigenvalue
will appear, and as α is varied, this eigenvalue flows from the negative essential
spectrum to the positive essential spectrum, see Fig. 2 (Exercise). This is an
example of spectral flow of eigenvalues, as the parameter α is varied around a
loop once.
Generally speaking, non-trivial spectral flow of eigenvalues is only possible
if there is an “infinite number” of negative spectra and positive spectra as
“sources” and “sinks” for discrete eigenvalues. At the beginning and end of
the loop, there are “infinitely many negative energy states”. Yet, we saw that
one negative-energy state is “lost” during the process, yielding the mysterious
equation
∞before − ∞after = 1.

32
Much of index theory is concerned with making sense of ∞ − ∞. In fact, we
will learn that spectral flow is topological — this makes it very resistant to
perturbations.

3.3.4 Momentum/Dirac operator on a bounded interval

Let us now consider X to be a bounded interval, say X = (0, 2π). The initial
d
domain for −i dx would be Cc∞ (0, 2π), i.e., Dirichlet conditions at x = 0 and
d ∗
x = 1 are imposed. As before, the adjoint (−i dx ) will have a larger domain
comprising non-Dirichlet functions. We make a compromise and consider the
following domain

{ψ ∈ C ∞ (0, 2π) : ψ(2π) = e2πiξ ψ(0)}, e2πiξ ∈ U(1), (3.2)

labelled by a quasiperiodicity parameter e2πiξ . It can be shown that the closure

d
of −i dx on such a domain is a self-adjoint operator.
Intuitively, “what leaves x = 2π returns at x = 0”, with a specified
phase shift. This is consistent with unitary time evolution. In physics, this
sort of condition is sometimes called a twisted boundary condition, or Bloch
wave condition, and occurs in the so-called Floquet–Bloch theory of periodic
Schrödinger operators.

Relation to Dirac operators on circle. Recall the Dirac operators P (ξ) =

d
−i dθ + ξ on the circle S 1 , where ξ ∈ R. For ξ ̸∈ Z, we contemplated applying
the “singular gauge transformation”,

uξ : eiθ 7→ e−iξθ , ξ ∈ R \ Z,

to convert P to P + ξ. Here, uξ is discontinuous, but it is still well-defined

as a bounded measurable function, and it does act as a unitary operator on
L2 (S 1 ). Naı̈vely, we might expect
contradiction
?
P (ξ) := uξ P u−1
ξ = P + ξ ⇒ σ(P
(ξ)
)=ξ+Z ̸= Z = σ(P ).

What went wrong?

The key point is that the discontinuous uξ does not preserve the domain
C ∞ (S 1 ). So P ≡ P (0) is actually unitarily equivalent to the operator

P (ξ) = uξ P u−1
ξ , Dom(P (ξ) ) = uξ · C ∞ (S 1 ) ̸= C ∞ (S 1 ),

33
acting on a domain of discontinous functions. On the other hand, P + ξ is a
genuine differential operator,
d
P + ξ = −i + ξ, Dom(P + ξ) = Dom(P ) = C ∞ (S 1 ),
dθ
on the smooth functions on S 1 . So

P + ξ ̸= P (ξ) ∼
=uξ P.

That is, P + ξ and P are actually not unitarily gauge-equivalent, unless ξ

happens to be an integer.
Once we introduce a point of discontinuity, we might as well disconnect
the circle at θ = 0, and work over the interval (0, 2π). Thus we write

d
P̃ (ξ) := −i , Dom(P̃ (ξ) ) = {ψ ∈ C ∞ (0, 2π) : ψ(2π) = e2πiξ ψ(0)},
dx
These are the quasiperiodicity conditions that we saw in Eq. (3.2). Now the
gauge transformation uξ is a genuinely smooth map (0, 2π) ∋ θ 7→ e−iξθ ∈ U(1),
and it implements
d
uξ P̃ (ξ) u−1
ξ = −i +ξ, Dom(uξ P̃ (ξ) u−1 ∞
ξ ) = {ψ ∈ C (0, 2π) : ψ(2π) = ψ(0)}.
dx
(3.3)
That is,
uξ P̃ (ξ) u−1
ξ = P̃
(0)
+ ξ,
where P̃ (0) has periodic boundary conditions. Therefore,

σ(P̃ (ξ) ) = σ(P̃ (0) + ξ) = ξ + Z.

Since P̃ (0) + ξ acts on periodic functions, we may identify it with the Dirac
operator P + ξ on the circle.

3.3.5 Gauge-invariance of quasiperiodicity?

The “non-local” nature of a quasiperiodic boundary condition should make
one rather nervous. After all, it involves comparing values at two different
positions, which may be extremely far apart!

34
 Indeed, the quasiperiodicity label e2πiξ is not gauge invariant; a gauge
transformation u ≡ u(x) will modify it by an extra u(2π)
u(0)
factor. For example,
consider P̃R (ξ) + A for some smooth real-valued A on (0, 2π). Define ξA ∈ R by
2π
e2πiξA = e 0 A . Then a simple calculation shows that
Rx
ei 0 A uξ+ξ
(ξ)
P̃ + A −→ P̃ (ξ+ξA ) −→A P̃ (0) + ξ + ξA . (3.4)

The three operators in Eq. (3.4) are different representations of “the Dirac
operator on (0, 2π) coupled to a connection with holonomy e2πi(ξ+ξA ) ”, with
respect to different gauge choices. Only the combined phase e2πi(ξ+ξA ) is gauge
invariant.
Remark (Optional). What about the quasimomentum label in solid-state physics
(surely a meaningful quantity)? Is it actually the same thing as quasiperiod-
icity? (See the notion of magnetic translations in the Assignment 1.)
Remark. Altogether, there is a circle’s worth of gauge-inequivalent twisted
Dirac operators S 1 (or on (0, 2π)), labelled by e2πiξ ∈ U(1). The space R/Z
is called the the moduli space of (flat) connections on S 1 . The moduli space
and the manifold S 1 are both circles, but do not mix them up! There are
many occasions where one has to consider a family of problems parametrized
by the moduli space. (In Floquet–Bloch theory, the so-called Brillouin zone is
basically the moduli space, and S 1 is the unit cell.)

3.4 Dirac operators and SUSY quantum mechanics

The simple harmonic oscillator Hamiltonian is

d2
HSHO = − + x2 .
dx2
Let us consider the first-order operator
d
0 a†
   
0 − dx +x
DSHO = d = .
dx
+x 0 a 0

Now observe that

a† a 0
   
2 HSHO − 1 0
DSHO = = (3.5)
0 aa† 0 HSHO + 1

35
holds on C0∞ (R), or even the Schwartz functions S (R) (smooth functions
whose derivatives decay faster than any inverse power of |x|). It can be shown
[17] that DSHO and HSHO are essentially self-adjoint on S (R). We use the
same symbols for their closures to self-adjoint operators, and write a∗ instead
of a† for the genuine (rather than formal) adjoint of a.
The operators a∗ a and aa∗ are positive-definite, so their spectrum is a
subset of R≥0 . Quite generally, the non-zero spectrum of a∗ a and aa∗ coincide
(this is the key lemma in supersymmetric QM). From Eq. (3.5), we even have
aa∗ = a∗ a + 2, so aa∗ ≥ 2. So the interval (0, 2) is a spectral gap for aa∗ , thus
also for a∗ a. Then (2, 4) is a spectral gap for a∗ a + 2 = aa∗ . By induction, the
spectrum of a∗ a lies within the discrete set 2N, so it comprises eigenvalues. The
operator a∗ is a “raising operator” in the sense that if ψn is an eigenfunction
of a∗ a with eigenvalue 2n, then

(a∗ a)(a∗ ψn ) = a∗ (aa∗ ψn ) = a∗ (a∗ aψn +2ψn ) = a∗ ((2n+2)ψn ) = 2(n+1)(a∗ ψn ).

It is easy to check that a∗ ψn is nonzero. Thus a∗ ψn is an eigenfunction of a∗ a

with eigenvalue 2(n + 1).
Of course, we need an initial eigenfunction to start off a non-vacuous chain
of eigenfunctions. The lowest possible eigenvalue is 0, and it is evident that
2
the kernel of a∗ a is spanned by the Gaussian ψ0 (x) = e−x /2 . This is the well-
known ground state of HSHO . We see that the lowest eigenspace of HSHO is
the same thing as the kernel a∗ a, which is itself the kernel of the Dirac-type
operator DSHO .
In the language of supersymmetric quantum mechanics, the odd opera-
tor DSHO is a supercharge for the positive operator Eq. (3.5). The latter has
two components (a “bosonic” and a “fermionic” one), and its ground state
is intimately related to the bosonic/fermionic kernel of the supercharge. A
remarkable point is that the supersymmetric kernel, meaning the difference in
dimensions of the bosonic and fermionic kernels, is very stable — mathemati-
cally, this is the stability of the index.

3.5 Outlook for higher-dimensional manifolds

1D manifolds X have the special property that any connection (on any line/vector
bundle over X) is flat. Since there is no curvature, all connections are the same
at the classical level (where curvature ∼ magnetic field strength). Yet there

36
can be gauge-inequivalent connections when X is topologically non-trivial,
with quantum mechanical consequences (e.g. AB-effect).
For 2D manifolds X, the possibility of curved connections arises. Even
if X is a boring contractible space like R2 , curvature of the connection has
dramatic consequences for the spectrum of Dirac/Laplace operators on X.
The prime example of this is the Landau quantization of the magnetic Lapla-
cian spectrum, which is key to the famous quantum Hall effect (several Nobel
prizes).
If X = S 2 , it is possible that the line bundle over S 2 is topologically non-
trivializable (much like the tangent bundle of S 2 ), a famous example being
Dirac’s magnetic monopole. In this case, there is simply no way to present a
section ψ as a global defined wavefunction, and the gauge potentials A are at
best locally defined terms contingent on local gauge choices.
This concludes our informal discussion of gauge-theoretic matters in quan-
tum theory. To fully understand these matters rigorously, we need to develop
the theory of fibre bundles, connections and curvature.

37
4 Smooth manifolds
We start with a working definition of a smooth manifold: it is a subspace of
some Euclidean space which, near each of its points, is locally diffeomorphic
to (some open subset) of Rn . Caveat: How the manifold is embedded in a
Euclidean space, is not part of its data!
Differential calculus of maps between manifolds can be set up, by working
locally and referring to the usual multivariable calculus in Euclidean spaces.
Integral calculus also makes sense, by invoking “partition-of-unity” and sum-
ming up local results.
For functions f : Rn → Rm between Euclidean spaces, partial derivatives
∂
∂xi
are defined once we pick some coordinates (e.g. Cartesian, polar, spher-
∂f i
ical, cylindrical. . . ), and they allow us to speak about quantities like ∂xj x

representing the j-th partial derivative of the i-th component of f at the point
x. We may arrange these partial derivatives neatly into a matrix, the Jaco-
bian matrix, and apply the appropriate transformation rules when (smoothly)
changing coordinates. Here, we stress that coordinates and coordinate changes
may only be defined on some open subset of X (e.g. Cartesian to polar co-
ordinates is only defined away from a branch cut). The consistency of the
transformation rules tells us that the Jacobian matrix of partial derivatives
is simply a particular coordinate-dependent expression of the derivative of f .
The latter, denoted dfx , is meant to be the “best linear approximation” of f
at x,
f (x + v) = f (x) + dfx (v) + . . . .
But how should we make sense of this idea in a coordinate-independent way?
(Many commonly used coordinate systems are not linearly related to one an-
other. . . )
On a smooth manifold X, each local diffeomorphism of X with Rn serves
as local coordinates, and is valid only within the domain of the coordinate
chart U ⊂ X. We must at least have a fiducial collection of local coordinate
charts (called an “atlas”) that suffices to cover all points of X. These charts
must also be smoothly compatible with each other (C ∞ coordinate changes
on overlaps). However, a key conceptual point is that there are no preferred
coordinates! Rather, all local coordinates which are smoothly compatible with
the fiducial coordinate charts are allowed. The mathematical expression of this
requirement is “smooth structure ↔ maximal C ∞ -atlas”.

38
 Now for the abstract mathematical definitions:
Definition 1. A topological n-manifold is a second-countable, Hausdorff topo-
logical space X, such that every point x ∈ X has an open neighbourhood
homeomorphic to an open subset of Rn .
Thus X admits an open cover {Uα }α∈I together with homeomorphisms

φα : Uα → φα (Uα ) ⊂ Rn

called local coordinates (or a chart). So φα = (x1α , . . . xnα ) is an n-tuple of

coordinate functions

xiα : Uα → R, i = 1, . . . , n,

used to provide numerical labels for the points in Uα . The subscript α will
usually be dropped unless ambiguity arises.
Every pair of charts has a change-of-coordinates homeomorphism on the
(possibly empty) overlap,

φα ◦ φ−1
β : φβ (Uα ∩ Uβ ) → φα (Uα ∩ Uβ ).

A (smooth) atlas is a collection of charts {Uα , φα }α∈I such that {Uα }α∈I covers
X, and has all change-of-coordinates functions being C ∞ (infinitely differen-
tiable). Extra charts can be admitted to an atlas provided the change-of-
coordinates functions remain C ∞ . A maximal atlas is called a smooth struc-
ture for X. So there is no distinguished sub-atlas (no distinguished coordinate
choices), although a minimal fiducial one may be used for convenience when
we want to describe a manifold to a colleague.
In particular, translated coordinate charts can always be used (i.e. shifting
of origin), so we regularly say “coordinate chart centred at x ∈ X” to mean
local coordinates φ for a neighbourhood of x, with the property that φ(x) = 0.
Definition 2. A smooth n-manifold is a topological n-manifold with a smooth
structure.
By default, we will use “manifold” to mean “smooth manifold”.
Example 4.1. The most simple example is the vector space R, with the single
fiducial chart (R, id); similarly for Rn . These are the manifolds on which
general manifolds are modelled. The identity chart can be composed with a

39
translation by any s ∈ Rn , to obtain another compatible chart. No particular
chart is distinguished amongst these translated charts. So as a manifold, Rn
has no particular distinguished point as its origin, and we sometimes write Rn
to emphasize its role as a manifold as opposed to a vector space.
Unless otherwise specified, we will simply use “manifold” to refer to smooth
manifolds. The technical second-countable and Hausdorff conditions preclude
some pathological examples, and in practice, ensure that an abstract manifold
can be smoothly embedded into Euclidean space (Whitney embedding the-
orem), and admits partitions-of-unity subordinate to any open cover. (You
can find these technical definitions in any comprehensive textbook on differ-
entiable manifolds, such as [11].) Embedding an abstract smooth manifold
into a background Euclidean space allows for a concrete visualization of its
abstract tangent spaces as hyperplanes in the background. Partitions-of-unity
allow local constructions to be smoothly globalized.

4.1 Tangent vectors

Let f : X → Y be a map of manifolds. Then near each point x, we may
describe f in terms of coordinates (U, φ) and (V, ψ) near x and f (x), respec-
tively,
ψ ◦ f ◦ φ−1 : φ(U ∩ f −1 (V )) → ψ(V ).
| {z }
fˇ

So fˇ is the coordinate representation of f as an ordinary map between (open

subsets of) Euclidean spaces, and we know what it means for fˇ to be smooth,
from elementary calculus.
Definition 3. A map f : X → Y is smooth if there are atlases for X and Y
such that all coordinate representations of f are smooth. If f has a smooth
inverse, then we say that f implements a diffeomorphism between X and Y .
Note that once f is established to be smooth with respect to some atlases,
the C ∞ compatibility condition will automatically guarantee smoothness in
any coordinate charts in the maximal atlases. So smoothness is a coordinate-
independent concept, even if we only checked it on some (rather than arbitrary)
coordinate charts.
Example 4.2. For any chart (U, φ) in a (smooth) maximal atlas, φ is a diffeo-
morphism onto its image.

40
Example 4.3. When Y = R, the set of smooth functions X → R is denoted
C ∞ (X). Here, we actually treat the target R as a number field, not just a
manifold. So C ∞ (X) is an algebra, equipped with the compatible operations
of pointwise addition and multiplication of functions.
Let γ : (a, b) → X be a smooth map, also known as a curve in X,
parametrized by t ∈ (a, b). Let x = γ(t) be a point on the image of the curve.
Then at x, we can consider the rate at which a given function f ∈ C ∞ (X)
changes along the curve: just take the ordinary derivative of the single-variable
function f ◦ γ : (a, b) → R. We think of the curve γ as assigning the rate-of-
change (at time t) to each smooth function,

d(f ◦ γ)
γ ′ (t) : C ∞ (X) → R, γ ′ (t)(f ) := .
ds s=t

The assignment γ ′ (t) is called the velocity vector of γ at the point x = γ(t). So
the velocity vector operates on functions; the number γ ′ (t)(f ) is the derivative
of f along γ evaluated at the point x = γ(t).
Obviously, γ ′ (t) is a linear assignment (for each t). Furthermore, by ele-
mentary calculus, γ ′ (t) satisfies a Leibniz rule at x = γ(t),

d((f · g) ◦ γ)
γ ′ (t)(f · g) =
ds s=t
d((f ◦ γ) · (g ◦ γ))
=
ds s=t
d(f ◦ γ) d(g ◦ γ)
= · g(x) + f (x) ·
ds s=t ds s=t
= γ ′ (t)(f ) · g(x) + f (x) · γ ′ (t)(g), ∀ f, g ∈ C ∞ (X). (4.1)

Formally, a map v : C ∞ (X) → R is called a derivation of the algebra C ∞ (X)

at the point x, if it is linear and satisfies the Leibniz rule at x,

v(f · g) = v(f ) · g(x) + f (x) · v(g), f, g ∈ C ∞ (X).

Thus Eq. (4.1) says that the velocity vector γ ′ (t) is a derivation of C ∞ (X) at
the point x = γ(t).
If v1 , v2 are derivations at x, and λ ∈ R, then it is easy to see that
(λv1 + v2 )(f ) := λv1 (f ) + v2 (f ) defines another derivation. This motivates
the following:

41
Definition 4. The tangent space to X at a point x ∈ X, denoted Tx X, is the
vector space of derivations of C ∞ (X) at x.

This definition captures the abstract algebraic aspect (Leibniz rule) of a

derivative, and is manifestly coordinate-independent. We will see that all ab-
stract derivations at x are in fact geometric: any derivation at x is of the
form γ ′ (t) for some (non-unique) smooth curve γ through x. Evidently, many
distinct curves can give rise to the same derivation, so for the purposes of dif-
ferentiating functions, we only distinguish two curves when they define distinct
derivations.
Let x be a point in some open subset U of X. Since U is itself a manifold,
we also have Tx U . The value of a derivation applied to f depends only on
the values of f in some arbitrarily small open neighbourhood of x, so it is
actually enough to define Tx X as the space of derivations of C ∞ (U ) at x, i.e.
Tx X = Tx U . In particular, if a derivation at x is specified by a curve, we only
need to consider the part of the curve lying within a coordinate chart around
x.
For i = 1, . . . , n, let xi : U → R be the coordinate functions for a chart
(U, φ) centered at x. Consider the i-th coordinate curve through x, i.e.,

γi : (−ϵ, ϵ) → X
t 7→ φ−1 (0, . . . , 0, |{z}
t , 0, . . . , 0).
i-th

(We take any sufficiently small ϵ > 0 to ensure that the curve lies within the
domain U of the chart.) This gives a set of coordinate tangent vectors at x,

∂i |x := (γi )′ (0), i = 1, . . . , n.

Remark. If the chart is not centered at x, then we use the curves

γi : t 7→ φ−1 (φ(x) + (0, . . . , 0, t, 0, . . . , 0))

to define the coordinate tangent vectors at x.

Actually, the derivations ∂i |x′ are also defined at every other point x′ ∈ U
in the same manner. Thus a coordinate chart (U, φ) gives rise to coordinate
vector fields
∂i : x′ 7→ ∂i |x′ , x′ ∈ U, i = 1, . . . , n.

42
 Writing ιi : (−ϵ, ϵ) → Rn for the inclusion t 7→ (0, . . . , 0, |{z}
t , 0, . . . , 0), we
i-th
untangle the definition of ∂i |x using the ordinary chain rule:
n
d(f ◦ γi ) X ∂(f ◦ φ−1 ) d(ιi )j
∂i |x (f ) ≡ = ·
ds s=0
j=1
∂xj φ(x) ds s=0

∂(f ◦ φ−1 )
= , f ∈ C ∞ (U ). (4.2)
∂xi φ(x)

Thus ∂i |x (f ) simply means: take the i-th partial derivative of the coordinate
representation f ◦ φ−1 of f , at the point φ(x).
Exercise 4.1. Let v ∈ Tx X be an arbitrary derivation at x. Show that v is a
linear combination,
Xn
v= v i ∂i |x ,
i=1
i i
where v := v(x ) ∈ R. Also show that {∂i |x }i=1,...,n is a linearly independent
set of derivations, thus it is a basis for Tx X.
Exercise 4.1 shows that the coordinate vector fields ∂i actually provide, for
every point x ∈ U , a basis for the tangent space Tx X.
In fact, given v ∈ Tx X, we can easily construct an explicit geometric curve
γ whose velocity vector at x is equal to v. Assume, without loss of generality,
that the coordinates of x are (0, . . . , 0), and write v i for the components of v
with respect to the basis {∂i |x }i=1,...,n . Then the curve

γ : (−ϵ, ϵ) → U ⊂ X, t 7→ φ−1 (tv 1 , . . . , tv n ).

has velocity vector γ ′ (0) = v.

Physics conventions. Physicists usually think of tangent vectors as a list

of components with certain behaviour under change of coordinates.
Exercise 4.2. Let (U, φ) and (Ũ , φ̃) be two charts, xi , x̃j be their respective
coordinate functions, and ∂i |x , ∂˜j |x the respective coordinate tangent vectors,
i, j = 1, . . . , n. Let x ∈ U ∩ Ũ . Show that

∂ x̃j
∂i |x = ∂˜j |x . (summation convention)
∂xi φ(x)

43
 Above, we use the very handy Einstein summation convention of sum-
ming over repeated indices when they appear both as a superscript and as a
subscript.  j 
Note that ∂∂xx̃ i is precisely the Jacobian matrix for the change-of-
φ(x) ji
coordinates map φ̃ ◦ φ−1 . Since we may expand a general v ∈ Tx X as

v = v i ∂i |x or v = ṽ j ∂˜j |x ,

we obtain the contravariant transformation rule for the components of tangent

vectors,
j ∂ x̃j
ṽ = i
vi. (4.3)
∂x φ(x)
Example 4.4. Let X denote an n-dimensional real vector space, and also the
underlying manifold. Fix a point x in X. Given any vector v ∈ X, we can
consider the curve t 7→ x + tv. The velocity vector of this curve at x is some
element in Tx X, which we denote by vx . The map

X → Tx X, v 7→ vx

can be checked to be a linear isomorphism. Thus X is canonically isomor-

phic to Tx X, and this is true for all points x ∈ X. In simpler terms, for a
vector space considered as a manifold, all of the tangent spaces are canoni-
cally isomorphic to each other. This is why we can freely “parallel transport”
tangent vectors of Euclidean space from one basepoint to another, without
much problem. A general manifold, however, does not have such a privileged
identification of its various tangent spaces!

4.2 Derivatives of smooth maps

Definition 5. Let f : X → Y be a smooth map. Its derivative at x ∈ X is
the map
dfx : Tx X → Tf (x) Y
defined by
dfx (v)(g) = v(g ◦ f ), v ∈ Tx X, g ∈ C ∞ (Y ). (4.4)
Exercise 4.3. Check the following:
ˆ dfx (v) is indeed a derivation of C ∞ (Y ) at f (x).

44
 ˆ dfx is a linear map.
f g
ˆ The chain rule is satisfied: if X → Y → Z, then d(g ◦ f )x = dgf (x) ◦ dfx .

ˆ d(idX )x = idTx X .

ˆ If f : X → Y is a diffeomorphism, then dfx is a linear isomorphism

whose inverse is d(f −1 )f (x) .

Example 4.5. Let γ : (−ϵ, ϵ) → X be a smooth curve, and write x = γ(0). So

the derivation dtd t=0 is a coordinate tangent vector to the interval (−ϵ, ϵ) at
the point t = 0. We have
 
d d
(dγ)0 (g) = (g ◦ γ) ≡ γ ′ (0)(g), ∀g ∈ C ∞ (X).
dt t=0 dt t=0

Thus the velocity vector γ ′ (0) ∈ Tx X is just the derivative of the map γ applied
to dtd t=0 .
Example 4.6. Let f : X → R be a smooth function, viewed as a smooth map
X → R of manifolds. Then dfx : Tx X → Tf (x) R = R is a linear functional,
also known as a cotangent/dual vector at x.

Definition 6. Let X be a manifold. The cotangent space at x ∈ X is the

space of linear functionals Tx X → R.

Exercise 4.4. Let xi : U → R, i = 1, . . . , n be local coordinate functions for

a chart around x. Show that (dxi )x , i = 1, . . . , n are linearly independent
cotangent vectors at x.

Returning to a general smooth map f : X → Y , let us make the description

of dfx more concrete by working in a chart (U, φ) around x and a chart (V, ψ)
around f (x). Write xi and x̃i for the respective coordinates, and

fˇ := ψ ◦ f ◦ φ−1 = (fˇ1 , . . . , fˇm )

for the coordinate representation of f (here, m = dim Y ). Since dfx is linear,

45
it suffices to see what it does to the basis of coordinate tangent vectors,

dfx (∂i |x ) (g) = (∂i |x ) (g ◦ f )

∂(g ◦ f ◦ φ−1 )
= (Eq. (4.2))
∂xi φ(x)
∂(g ◦ ψ −1 ) ∂(ψ ◦ f ◦ φ−1 )j
=
∂ x̃j ψ(f (x)) ∂xi φ(x)
   ∂ fˇj
= ∂˜j |f (x) (g) ·
∂xi φ(x)
 ˇj 
∂f 
˜
= ∂j |f (x) (g), g ∈ C ∞ (Y )
∂xi φ(x)

So for a general v = v i ∂i |x , we have

∂ fˇj  
dfx (v) = v i ∂˜j |f (x) .
∂xi φ(x)

The components of the tangent vector dfx (v) with respect to the basis {∂˜j |f (x) }j=1,...,m
are

j ∂ fˇj
dfx (v) = vi, j = 1, . . . , m.
∂xi φ(x)
In other words, with respect to the bases of coordinate tangent vectors, the
linear map dfx is represented as the Jacobian matrix of fˇ,
 ˇj 
∂ f
dfx ↔ J(fˇ)|φ(x) := . (4.5)
∂xi φ(x) ji

Example 4.7. Suppose f has coordinate representation fˇ being a linear map.

Then the Jacobian of fˇ is fˇ itself.

Earlier, we explained how any tangent vector v at x may be obtained as

′
γ (0) for some curve γ with γ(0) = x. Actually, dfx (v) is simply represented
by the curve f ◦ γ in Y :

Proposition 4.1. Let f : X → Y be a smooth map, v ∈ Tx X, and γ be any

curve in X with γ(0) = x and γ ′ (0) = v. Then

dfx (v) = (f ◦ γ)′ (0).

46
Proof. By definition, for any g ∈ C ∞ (Y ), we have

d
(f ◦ γ)′ (0)(g) = (g ◦ f ◦ γ)(t) = γ ′ (0)(g ◦ f ) = v(g ◦ f ) = dfx (v)(g).
dt t=0

4.3 Immersions, submersions, embeddings

Definition 7. A smooth map f : X → Y is a immersion (resp. submersion)
at x ∈ X, if dfx is injective (resp. surjective). If f is an immersion at every
x ∈ X, it is simply called an immersion; likewise for submersions.

Remark 1. The following important facts may be found in standard texts on

smooth manifolds, e.g. [6, 11].

ˆ If dfx is an isomorphism, then the Inverse Function Theorem says that

f : X → Y is actually a local diffeomorphism at x, i.e., there is an open
neighbourhood U of x such that f : U → f (U ) is a diffeomorphism.

ˆ Write dim X = n, dim Y = m. If f is an immersion at x, then in some

coordinate charts centred at x and f (x), it is represented as a canonical
immersion (x1 , . . . , xn ) 7→ (x1 , . . . , xn , 0, . . . , 0).

ˆ Similarly, if f is a submersion at x, then it has a coordinate representa-

tion as a standard projection (x1 , . . . , xm , xm+1 , . . . , xn ) 7→ (x1 , . . . , xm ).

The following Lemma will be useful for us.

Lemma 4.2. Let π : E → X be a surjective submersion. Then a map f :

X → Y is smooth iff f ◦ π : E → Y is smooth.

Proof. Suppose f ◦π is smooth. Let x ∈ X and pick a point p ∈ π −1 (x). Since π

is a submersion at p, it has the form (x1 , . . . , xm , xm+1 , . . . , xn ) 7→ (x1 , . . . , xm )
in some coordinate charts (V, ψ) centred at p and (U, φ) centred at x. We may
then define the smooth map s : U → E,

s = ψ −1 ◦ (x1 , . . . , xm ) 7→ (x1 , . . . , xm , 0, . . . , 0) ◦ φ,

which satisfies
s(x) = p, π ◦ s = idU .

47
Then
f |U = f |U ◦ idU = f ◦ π ◦ s
is the composition of the smooth maps f ◦ π and s, thus smooth. Since x ∈ X
is arbitrary, f is smooth.
If f is smooth, obviously its composition with the smooth π is smooth.
Remark. In topology, a quotient map is a surjective map π : E → X such
that U is open in X iff π −1 (U ) is open in E. Then one shows f : X → Y
is continuous iff f ◦ π : E → Y is continuous. A surjective submersion is the
smooth analogue of a quotient map.

In the following, we consider k ≤ n and Rk ⊂ Rn as the standard slice with

last n − k coordinates being zero.

Definition 8. A k-dimensional embedded/regular submanifold (or simply sub-

manifold ) of an n-manifold X is a subspace Z ⊂ X such that every point z ∈ Z
is contained in a chart (U, φ) of X with the property that φ(U ∩Z) = φ(U )∩Rk .

As its name suggests, a submanifold Z ⊂ X is a manifold in its own right.

Specifically, we take the charts φ|U ∩Z : U ∩ Z 7→ Rk , and check that they make
Z into a k-manifold.
For example, any open subset U of a manifold X is a submanifold.

Definition 9. A smooth immersion f : Z → X which is also a homeomor-

phism onto its image f (Z) (given the subspace topology), is called a (smooth)
embedding.

Exercise 4.5. The inclusion map of a submanifold is an embedding.

Conversely, the image f (Z) of an embedding f : Z → X is a submanifold
of X, and f : Z → f (Z) is a diffeomorphism. We omit the proof (see [11] §5,
[13] §5.6), and just mention an important example.
Exercise 4.6. Let f : X → Y be any smooth map. Its graph is the subset

Γf := {(x, y) ∈ X × Y : y = f (x)}.

Then the map

X → X × Y, x 7→ (x, f (x))
is an embedding, so X is diffeomorphic to its graph.

48
 Smoothness of maps behaves well with respect to restriction to submani-
folds:
Exercise 4.7. Let f : X → Y be a smooth map. Show that if X ′ is a sub-
manifold of X, then f |X ′ remains smooth. Similarly, show that if Y ′ is a
submanifold of Y such that f (X) ⊂ Y , then f : X → Y ′ remains smooth.

Definition 10. If f : X → Y is a smooth map, then y ∈ Y is a regular value

of f if f is a submersion at all points in the (possibly empty) preimage f −1 (y).

Proposition 4.3. Let y be a regular value of a smooth map f : X → Y . Then

f −1 (y) is a submanifold of X with codimension equal to dim Y .

Proof. Exercise.
More generally:

Definition 11. A submanifold Z ⊂ Y is transverse to f : X → Y if

Image(dfx ) + Tf (x) (Z) = Tf (x) Y, ∀ x ∈ f −1 (Z).

Theorem 4.4. Let Z ⊂ Y be a submanifold transverse to a map f : X → Y .

Then its preimage f −1 (Z) is a submanifold of X, with

codim f −1 (Z) = codim Z. (4.6)

Proof. See §5 of [6], Theorem 6.30 of [11].

Such results involving transversality constitute an important part of dif-
ferential topology, you may learn more about this subject from [6, 12]. The
importance for us is the following corollary:

Proposition 4.5. Let π : E → X be a surjective submersion. Then for any

submanifold Z ⊂ X, the preimage π −1 (Z) is a submanifold of E with the same
codimension.

Proof. The stated assumptions imply that π is transverse to any submanifold

Z ⊂ X, whence the conclusion follows from (4.6).

49
5 Tangent bundle and vector bundles
5.1 Tangent bundle
As a set, the tangent bundle of a manifold X is a disjoint union
G
TX = Tx X,
x∈X

with a projection map

π : T X → X, (x, vx ) 7→ x.

We can give T X the structure of a smooth 2n-manifold, such that π : T X → X

is smooth. First, we already have coordinate charts (Uα , φα ), α ∈ I covering
X. For each α ∈ I, consider the “local tangent bundle”,
G
T Uα := π −1 (Uα ) = Tx X.
x∈Uα

For any x ∈ Uα and any vx ∈ Tx X, we have the expansion vx = vxi ∂i |x in terms

of coordinate tangent vectors. So define the injective maps

φ̃α : T Uα → φα (Uα ) × Rn ⊂ R2n

(x, vx ) 7→ (φα (x); vx1 , . . . , vxn ).

Clearly, the T Uα , α ∈ I cover T X, and we would like to use the (T Uα , φ̃α )

as charts for T X. To make sense of this, we first need to topologize T X, by
taking the basis of open sets to be

φ̃−1
α (V ), V open in R2n , α ∈ I.

One checks that this is indeed a basis for a topology; the verification of Haus-
dorffness and second-countability are omitted. The charts φ̃α are now home-
omorphisms, and T X is a topological manifold. The maps φ̃α ◦ φ̃−1 β are easily
∞
seen to be C — the base coordinates transform smoothly since X is a mani-
fold, while the vector components transform according to the smooth formula
Eq. (4.3). Therefore we have constructed a smooth atlas for T X, turning
it into a smooth (2n)-manifold. With respect to the charts (T Uα , φ̃α ) and
(Uα , φα ), the projection map π is simply projection onto the first factor, thus
smooth.

50
Remark. If we have a smooth embedding X ,→ RN , then we may construct
T X as a subset of X × RN . The advantage is that the topology is directly
inherited from X × RN . A disadvantage is that we have to check that this is
a canonical construction, independent of choice of embedding.
Observe that the maps
φ̃α φ−1
α ×id
Φα : T Uα −→ φα (Uα ) × Rn −→ Uα × Rn
are diffeomorphisms which identify T Uα as a cartesian product — these are
called local trivializations of T X, and turn it into a vector bundle in the sense
of Definition 13.

5.1.1 Tangent vector fields as derivations

Definition 12. A tangent vector field over X is a smooth map v : X → T X
such that π(v(x)) = x, ∀x ∈ X. We denote the space of tangent vector fields
by X(X).
Let v : X → T X be a not-necessarily-smooth tangent vector field over
X. So π ◦ v = idX , but v may not be smooth. Pointwise in X, we have the
derivation vx of C ∞ (X) at x. In total, we get a function,
v(f ) : X → R, v(f )(x) := vx (f ). (5.1)
Exercise 5.1. Check that the following are equivalent:
ˆ v is a smooth tangent vector field.

ˆ For all f ∈ C ∞ (X), the function v(f ) defined by Eq. (5.1) lies in C ∞ (X).
Exercise 5.1 implies that a smooth vector field v defines a derivation on
the algebra C ∞ (X), i.e.,
ˆ v : C ∞ (X) → C ∞ (X), f 7→ v(f ) is linear;

ˆ v(f g) = f v(g) + v(f )g, f, g ∈ C ∞ (X).

Remark 2. Conversely, if D is any derivation on C ∞ (X), then we obtain a
pointwise derivation Dx ∈ Tx X at each x by the formula
Dx (f ) := (Df )(x), f ∈ C ∞ (X).
The assignment x 7→ Dx is smooth, due to Exercise 5.1.

51
5.2 General vector bundles
The tangent bundle T X is a special case of the following:

Definition 13. A (real) vector bundle π : E → X of rank n is a smooth

surjective map, such that there exists an open cover {Uα }α∈I of X and diffeo-
morphisms (“local trivializations”)

Φα : E|Uα := π −1 (Uα ) → Uα × Rn

such that

ˆ πUα ◦ Φα (p) = π(p) for all p ∈ E|Uα , where πUα denotes the projection
Uα × Rn → Uα ,

ˆ For any two local trivializations (Uα , Φα ), (Uβ , Φβ ), we have

Φβ ◦ Φ−1
α (x, v) = (x, gβα · v), x ∈ Uα ∩ Uβ , v ∈ Rn , (5.2)

with smooth transition functions gβα : Uα ∩ Uβ → GL(n, R).

If there is a Φα with Uα = X, it is called a (global) trivialization of E, and E

is said to be trivializable. A section of E, also called a smooth vector field, is
a smooth map s : X → E such that π ◦ s = idX . The space of sections on E
is denoted Γ(E). If s is only defined on an open subset U ⊂ X, it is called
a local section over U . Complex vector bundles are defined in the same way,
except that Rn is replaced by Cn , and gβα takes values in GL(n, C).

If v ∈ Γ(E) is a vector field, then in a local trivialization (U, Φ), we have

Φ(v(x)) = (x, v(x)), x ∈ U,

for some smooth function v ≡ (v 1 , . . . , v n ) : U → Rn . The v i are the local

component functions of v.
Remark. For each x ∈ X, the fibre Ex := π −1 (x) inherits a vector space
structure through its identification with Rn via a(ny) local trivialization. In
particular, each Ex has a zero element, and it makes sense to ask whether a
section s has s(x) = 0 or s(x) ̸= 0. Clearly, if E is trivializable, then it admits
a nowhere-vanishing section s : X → E.

52
Remark. In an equivalent approach, the fibres Ex are a priori vector spaces,
and instead of having the smooth compatibility condition (5.2) in the defi-
nition, one requires the local trivializations Φα to restrict fibrewise to linear
∼
=
isomorphisms Ex → {x} × Rn . Then it is proved that the compatibility condi-
tion by smooth GL(n)-valued transition functions holds (details can be found
in Lemma 10.5 of [11]).
Remark. A specific choice of trivializing cover {Uα , Φα }α∈I is not part of the
data of a vector bundle, only the availability of such a cover is. In practice,
we may specify a fiducial set of local trivializations, then pass to a maximal
trivializing cover compatible with this fiducial set.
Remark. The space of sections, Γ(E), is a module over the ring C ∞ (X, K),
where K = R or K = C is the ground field of the fibres.
Exercise 5.2. Show that the projection map of a vector bundle is a submersion.

Definition 14. Let π : E → X be a rank-n vector bundle over X, and U ⊂ X

an open subset. An n-tuple of local sections (s1 , . . . , sn ) is called a local frame
for E over U , if for each x ∈ U , {si (x)}i=1,...,n is a basis for Ex .

A local trivialization Φ : π −1 (U ) → U × Rn determines a local frame for

E over U ; set si (x) = Φ−1 (x, ei ), where ei is the i-th standard basis vector for
Rn . Conversely:
Exercise 5.3. Let (s1 , . . . , sn ) be a local frame for E over U . Define the map

Ψ : U × Rn → π −1 (U )
n
X
(x, (v 1 , . . . , v n )) 7→ v i si (x).
i=1

Check that Ψ is a diffeomorphism, thus Ψ−1 is a local trivialization of E.

5.3 Tautological line bundle over RP1

The real projective line RP1 is the set of one-dimensional subspaces ℓ (i.e.
lines) in R2 . On v = (x, y) ∈ R2 \ {0}, define the equivalence relation [v] = [v ′ ]
iff v, v ′ lie in the same line. Clearly, RP1 is precisely the quotient space under
this equivalence,
q : R2 \ {0} → RP1 , v 7→ [v].

53
We equip RP1 with the quotient topology; that is, U ⊂ RP1 is open iff q −1 (U )
is open in R2 \ {0}.
Let ℓH be the line spanned by (1, 0) and ℓV be the line spanned by (0, 1).
Define the subsets
U1 = RP1 \ {ℓV }, U2 = RP1 \ {ℓH }.
Then {U1 , U2 } is an open cover for RP1 (Why?). Define the maps
φ1 : U1 → R, [(x, y)] 7→ y/x,
φ2 : U2 → R, [(x, y)] 7→ x/y,
which are well-defined homeomorphisms (why?). The overlapping region is
U12 ≡ U1 ∩ U2 = RP1 \ {ℓV , ℓH },
and clearly φi (U12 ) = R \ {0} ≡ R× for i = 1, 2. The change-of-coordinates
map is
φ2 ◦ φ−1 × ×
1 : R → R , λ 7→ λ−1 ,
which is smooth. Thus RP1 is a smooth manifold.
The tautological line bundle π : LR → RP1 sits inside the product RP1 × R2
as the following subset,
LR = {(ℓ; v) ∈ RP1 × R2 | v ∈ ℓ},
and it is equipped with the projection map,
π : LR → RP1 , (ℓ; v) 7→ ℓ.
Define the continuous bijective maps
 p 
Φ1 : π −1 (U1 ) → U1 × R, (ℓ; (x, y)) 7→ ℓ; x 1 + (φ1 (ℓ))2 ,
 p 
Φ2 : π −1 (U2 ) → U2 × R, (ℓ; (x, y)) 7→ ℓ; y 1 + (φ2 (ℓ))2 ,

which respect the projection maps to Ui . Their inverses are

Φ−1 −1
i : U1 × R → π (Ui ) ⊂ L ⊂ RP × R
R 1 2
  
φ (ℓ)
 ℓ; µ √ 1
,√ 1 , i = 1,


1+(φ1 (ℓ))2 1+(φ1 (ℓ))2
(ℓ; µ) 7→    
φ (ℓ)
 ℓ; µ √ 2 , √ 1
, i = 2.


2
1+(φ2 (ℓ)) 2
1+(φ2 (ℓ))

54
and are also continuous. We may compute (exercise)

Φ2 ◦ Φ−1
1 : U12 × R → U12 × R


(ℓ; µ) 7→ ℓ; sgn(φ2 (ℓ))µ .
φ2 sgn
Note that U12 → R× → O(1) is continuous (even smooth), so the above map
is smooth. Then (φ1 × idR ) ◦ Φ1 and (φ2 ◦ idR ) ◦ Φ2 may be regarded as
defining coordinate charts for LR as a smooth manifold. With this smooth
structure, π : LR → RP1 is smooth. Also, Φ1 , Φ2 are diffeomorphisms, and
they serve as local trivializations of LR , with smooth transition function being
sgn ◦ φ2 : U12 → O(1).

Non-trivializability. We have constructed π : LR → RP1 as a rank-1 vector

bundle. However, it does not admit any nowhere-vanishing sections (Exercise),
and it is therefore non-trivializable. So any globally-defined s ∈ Γ(LR ) will
always have some zeroes, and it is not hard to imagine some significant con-
sequences for defining differential operators on such a space.
Remark. Our local trivializations of LR appear somewhat more complicated
than necessary, but they have the nice property that the lengths of fibre vectors
are preserved. For example, if ℓ ∈ U1 , then
p p
(x, y) ∈ ℓ ⇒ |x 1 + (φ1 (ℓ))2 |2 = |x 1 + (y/x)2 |2 = x2 + y 2 = ||(x, y)||2 ;

similarly if ℓ ∈ U2 . This property is reflected in the fact that the transition

functions are O(1)-valued instead of being GL(1, R)-valued. So LR is actually
an unorientable Euclidean line bundle (this concept will be defined in a later
lecture).
Remark. It is not hard to show that RP1 is diffeomorphic to a circle S 1 . Fur-
thermore, there is a natural metric on RP1 , given on a pair of lines by the
angle between them. A little thought will show that RP1 is “half of a unit
circle”, having circumference π.

5.4 Morphisms of vector bundles

The tangent bundle T X comes canonically with the manifold X. To under-
stand this further, we must discuss morphisms between vector bundles.

55
Definition 15. A morphism of vector bundles from π : E → X to π ′ : E ′ →
X ′ is a smooth map F : E → E ′ , such that there exists a smooth map
f : X → X ′ making the following diagram commute
F
E E′
π π′ , (5.3)
f
X X′

and which restricts to linear maps F : Ex → Ef′ (x) on fibres.

Example 5.1. As a simple example, let U ⊂ X, and consider E|U := π −1 (U ),

which is still a vector bundle. Then the inclusion map F : E|U → E is a vector
bundle morphism, by taking f : U → X to be the inclusion.

5.4.1 Derivative as morphism of tangent bundles

We had already encountered the pointwise derivative of a smooth map f :
X → Y (Eq. (4.4)). This is a linear map dfx : Tx X → Tf (x) Y . Globally, the
(total) derivative of f is

df : T X → T Y,
G G G G
dfx : Tx X → Tf (x) Y ⊂ Ty Y.
x∈X x∈X x∈X y∈Y

Let fˇ = ψ ◦ f ◦ ϕ−1 be the representation of f with respect to some charts

(U, φ) and (V, ψ) around x and f (x) respectively. Recall from Eq. (4.5) that
dfx is represented as the Jacobian matrix J(fˇ)|φ(x) acting on the components
v i of a vector v ∈ Tx X. Therefore, in the corresponding charts (T U, Φ) and
(T V, Ψ) for the tangent bundles, we have df represented as

Ψ ◦ df ◦ Φ−1 : (x1 , . . . , xn ; v 1 . . . , v n ) 7→ (fˇ(x); J(fˇ)|x (v)),

| {z } | {z }
x v

which is smooth, since fˇ and J(fˇ) are both smooth. If follows that df : T X →
T Y is a smooth map.
Therefore, df is a morphism of tangent vector bundles, in the sense of
Definition 15. It also follows from Exercise 4.3 that the derivative is functorial,
in the sense that:

56
 ˆ Chain Rule: If f : X → Y and g : Y → Z are smooth maps, then
d(g ◦ f ) = dg ◦ df : T X → T Z;

ˆ d(idX ) = idT X .

In particular, if f : X → Y is a diffeomorphism, then so is df : T X → T Y .

57
6 Principal G-bundles
6.1 Lie groups
Definition 16. A Lie group is a manifold G together with a group structure
such that
G × G → G, (g1 , g2 ) 7→ g1 g2−1
is smooth.
Thus the group operations (inversion and composition) are smooth maps.
It follows easily that the operations of left and right multiplication on G by
any fixed g ∈ G,
Lg : g ′ 7→ gg ′ , Rg : g ′ → g ′ g, g ′ ∈ G,
are self-diffeomorphisms of G.
The basic example of a Lie group is GL(n) (over R or C). It is usually
considered as an open (why?) submanifold of the vector space of matrices
Mn (R) or Mn (C). The tangent spaces at each g ∈ GL(n) are the same as
those when g is considered as element of Mn (K). Thus the tangent bundle of
GL(n) is identified with
T GL(n) ∼
= GL(n) × Mn (K) .
| {z }
∼ 2
=Kn

Similarly for the orientation-preserving subgroup GL(n, R)+ (i.e. positive de-
terminant matrices). Smoothness of the group operations is deduced from
their algebraic nature in terms of the matrix entries. For an n-dimensional
vector space V , the invertible linear maps GL(V ) are identified with GL(n) by
picking a basis for V , and any two such choices are related by a conjugation.
The subgroups O(n), SO(n), U(n), SU(n) are also Lie groups, but of lower
dimension. This is usually shown by the following exercise:
Exercise 6.1. Consider Mn (R) ∼
2
= Rn and the subspace Sym(n) of symmetric
matrices. Check that 1n is a regular value for the map
f : Mn (R) → Sym(n), A 7→ At A.
Consequently, O(n) = f −1 (1n ) is actually a submanifold of Mn (R) (also of
GL(n, R)), due to Prop. 4.3. It follows (by Exercise 4.7) that O(n) is a Lie
subgroup of GL(n, R). Similarly, SO(n) and U(n), SU(n) are Lie subgroups of
GL(n, R)+ and GL(n, C) respectively.

58
6.2 Idea of principal bundles
In a vector bundle, the vector space Ex attached to x ∈ X does not come
with a canonical basis (“frame at x”). A frame, or “gauge”, at x is a choice
of isomorphism
∼
=
e : Kn → Ex
1 , 0, . . . , 0) 7→ ei ∈ Ex ,
(0, . . . , 0, |{z} i = 1, . . . , n.
i-th

We will sometimes write e = {e1 , . . . , en }.

The set Fr(Ex ) of frames at x has an important structure. Namely, it has
a right action by G = GL(n), denoted
Fr(Ex ) × G → Fr(Ex ), (e, g) 7→ e · g.
Explicitly, if e = {e1 , . . . , en }, then
( n n
)
X X
e·g = ei g i1 , . . . , ei g in , g = (g ij ) ∈ G = GL(n).
i=1 i=1

Symbolically, the frame e is written as a row vector, and it is multiplied by

the n × n matrix g to get the row vector for the new frame e · g.
The map
Fr(Ex ) × G → Fr(Ex ) × Fr(Ex )
(e, g) 7→ (e, e · g) (6.1)
is a bijection. Formally, this says that Fr(Ex ) is a G-torsor — it has a free
and transitive action of G, implementing change-of-frame. Informally, this is
“gauge freedom” of “gauge symmetry” at x.
Let eref be a reference basis/frame (or “gauge choice”). Fixing e to be eref
in Eq. (6.1), we obtain a eref -dependent bijection
G ←→ Fr(Ex )
g ←→ eref · g,
turning Fr(Ex ) into a group. Effectively, we declare eref to be “identity frame”,
and the frame space Fr(Ex ) becomes identified with Fr(Kn ) = G.
If the Ex come with an inner product, then we can restrict to orthonormal
frames. The above discussion still holds, with G replaced by O(n) or U(n).

59
6.2.1 Trivializable principal bundles
Consider a trivialized vector bundle E = X × Kn . Its (orthonormal) frame
bundle is then the product manifold
G
Fr(X × Kn ) = Fr(Kn ) = X × G, G = GL(n) (or U(n) or O(n)).
x∈X

When equipped with the projection map π : X × G → X and the right G-

action,
(x, g) · g ′ = (x, gg ′ ),
Fr(X × Kn ) is an example of a trivialized principal G-bundle.
Next, consider a trivializable vector bundle E → X. The frame bundle of
E is G
Fr(E) = Fr(Ex ), (6.2)
x∈X

which has, set-theoretically, the projection π : Fr(E) → X, as well as a right G-

∼
=
action on each fibre Fr(Ex ). Under a smooth trivializing map Φ : E → X ×Kn ,
each Ex gets identified with Kn . This is equivalent to picking a reference frame
in Fr(Ex ), and therefore an identification Fr(Ex ) ∼= G, for each x ∈ X. Thus
we get a bijective map
Φ̃ : Fr(E) → X × G,
which respects the right G-actions. Use Φ̃ to transfer the topology and smooth
structure of X × G to Fr(E).
This frame bundle π : Fr(E) → X is an example of a trivializable principal
∼
=
G-bundle. If we had used some other trivialization Φ′ : E → X × Kn of the
vector bundle E, then we would obtain another trivialization Fr(E) ∼ = X ×G
of the frame bundle. The smooth structure on Fr(E), bundle projection, and
G action on Fr(E) do not depend on which trivialization we started with.
Indeed, no preferred trivialization is specified for either E or its frame bundle
Fr(E).

6.3 General principal bundle

Definition 17. Let G be a Lie group. The trivial principal G-bundle over a
manifold X is πX : X × G → X, (x, g) 7→ x, equipped with the right G-action,

(x, h) · g = (x, hg), x ∈ X, h, g ∈ G.

60
A principal G-bundle over a manifold X is a manifold P with a smooth sur-
jection π : P → X and a smooth right G-action (by diffeomorphisms),

P × G → P, (p, g) 7→ p · g,

such that

ˆ the right G-action restricts to fibres,

π(p · g) = π(p), p ∈ P, g ∈ G;

ˆ every x ∈ X is contained in some open neighbourhood U with P |U :=

π −1 (U ) being trivializable, i.e., there exists a diffeomorphism

Φ : π −1 (U ) → U × G

such that

πU ◦ Φ = π,
Φ(p · g) = Φ(p) · g.

As with Exercise 5.2, the projection map π of a principal G-bundle can be

checked to be a submersion.
Example 6.1. Recall Example 1.2 of the real line bundle LR → RP1 . In the
fibre above each ℓ ∈ RP1 , there are two unit-length vectors. Let us consider
the subset of LR comprising the unit-length vectors. This is the (unit) sphere
bundle S(LR ) of LR .
The fibre of S(LR ) lying above ℓ ∈ RP1 comprises the two unit vectors in
the line ℓ. Think of these unit vectors as the two possible choices of orthonor-
mal frames for the line ℓ, with no preference for either one. Globally, S(LR ) is
the bundle of orthonormal frames for LR . The group O(1) ∼ = Z2 acts on S(LR )
by swapping the two frames at each ℓ ∈ RP1 . This gives S(LR ) the structure
of a principal O(1) ∼= Z2 -bundle over RP1 (Exercise).
Interestingly, S(LR ) is not simply two disconnected copies of RP1 . Rather,
the total space S(LR ) can be identified with the unit circle S 1 , as a connected
double cover of RP1 .

61
Example 6.2. Example 6.1 has a direct complex counterpart, π : LC → CP1 .
Here, CP1 is the space of one-dimensional complex subspaces in C2 , and LC is
the complex tautological line bundle over CP1 , constructed in the same way
as RP1 .
For each ℓ ∈ CP1 , we can consider only the unit complex vectors inside the
corresponding fibre of C, and this is interpreted as the space of orthonormal
frames of ℓ, which is a U(1)-torsor. The sphere bundle of LC is thus the bundle
of orthonormal frames for the line bundle LC . The group U(1) acts on this
frame bundle by “rotating frames” at each ℓ.
This frame bundle can be shown to be diffeomorphic to S 3 . So we get
an interesting principal U(1)-bundle π : S 3 → CP1 . It is well-known that
CP1 ∼= S 2 , and π is actually the Hopf fibration, which we will study in detail
in Section 7.1.

Some motivation. Now, S 3 ∼ ̸= S 2 × U(1), which is mathematically interest-

ing. But from the physical viewpoint, why would we need to consider general
abstract “twisted frame bundles”? Historically, this non-trivial principal bun-
dle arose when Dirac was considering the idea of a magnetic monopole (albeit
still not found in nature). It was not until many decades later, that Dirac’s
monopole was understood in bundle-theoretic and algebraic topology terms,
and generalizations with non-Abelian G grew in importance in physics. An-
other reason is spin, particularly its mysterious relationship with Riemannian
geometry. Dirac also played a key role in the integration of spin into quantum
theory, through his famous Dirac operator invented in 1928. Again, it was not
until many decades later, that Atiyah, Bott, Singer and others uncovered the
rich geometric-analytic structures underlying spin.
Definition 18. A morphism between two principal G-bundles P1 → X1 and
P2 → X2 , is a smooth map F : P1 → P2 which is G-equivariant, i.e.,
F (p · g) = F (p) · g, p ∈ P1 , g ∈ G.
Two principal G-bundles P1 , P2 over X are isomorphic if there is a diffeomor-
phism F : P1 → P2 covering the identity idX .
Exercise 6.2. Check that when P is a principal G-bundle (Definition 17), then
the following hold:
ˆ For any p ∈ P , its orbit under the right G-action is precisely the fibre
π −1 (π(p)).

62
 ˆ The right G-action σ is free (every non-identity element g ∈ G has no
fixed points).

If F is a morphism of principal G-bundles (Definition 18), then

ˆ For each p ∈ P , F restricts to a diffeomorphism between the fibre con-

taining p and the fibre containing F (p); thus F projects down to a well-
defined map f : X1 → X2 , and we say that F covers f :
F
P1 P2
π1 π2 . (6.3)
f
X1 X2

ˆ The induced f : X1 → X2 is smooth (recall Lemma 4.2).

ˆ If f is a diffeomorphism, then so is F , in which case F −1 : P2 → P1 is

also a morphism of principal G-bundles.

Example 6.3. Let π : P → X be a principal G-bundle, and Z ⊂ X be a

submanifold. The subset P |Z := π −1 (Z) is a submanifold of P (Prop. 4.5),
and is a principal G bundle by itself. The inclusion P |Z ,→ P is easily seen to
a principal G-bundle morphism, covering the inclusion Z ,→ X.

Definition 19. Let π : P → X be a principal G-bundle. A local gauge over

an open neighbourhood U ⊂ X is a smooth section over U , i.e., a smooth map
s : U → P such that π ◦ s = idU .

Why emphasize “local” in “local gauge”? If there exists a global trivializing

diffeomorphism Φ : P → X × G to the trivial principal G-bundle, then P is
said to be (globally) trivializable. In this case, we can indeed find a global
gauge, i.e. a section s : X → P defined on all of X. However, non-trivializable
principle bundles exist, so there may simply be no such thing as a global gauge!
Let s : U → P be a local gauge. Then at each x ∈ U , any fibre point
p ∈ π −1 (x) has the form s(x) · g for some unique label g = g(p) ∈ G. This is
because the G-action is free and transitive on each fibre π −1 (x). Therefore a
local gauge s determines a map

Φs : P |U ≡ π −1 (U ) → U × G, (s(x) · g) 7→ (x, g). (6.4)

63
The role of a local gauge s is to provide a “field of reference frames” over U ,
so that we may label any other frame p ∈ P |U by the group element g = g(p)
which “rotates” the reference frame s(π(p)) to p.
Exercise 6.3. Show that the map Φs is precisely a trivialization of P |U . The
G-equivariance and bijectivity are quite straightforward. The diffeomorphism
property is more involved, and requires some general smooth manifold and Lie
group theory (Lemma 4.2.7 of [7]).
Given a trivialization Φ : P |U → U ×G, we can transport the trivial section
x 7→ (x, e) of U × G back to P |U , by taking
sΦ : U → P |U
x 7→ Φ−1 (x, e). (6.5)
It is easily checked that sΦs = s and ΦsΦ = Φ. That is, local sec-
tions/gauges over U are in 1-1 correspondence with local trivializations over
U.

6.4 Transition functions

Definition 20. A trivializing cover for a principal G-bundle P → X, is an
open cover {Uα }α∈I of X, together with local trivializations
Φα : P |Uα → Uα × G, α ∈ I.
On a double overlap Uβα := Uβ ∩Uα , there are apparently two different ways
of viewing P |Uβα as a product bundle Uβα × G. Specifically, for each x ∈ Uαβ ,
we have two different labelling conventions for each frame p ∈ π −1 (x),
Φα (p) = (x, ϕα (p)) or Φβ (p) = (x, ϕβ (p)),
where ϕα , ϕβ are G-valued labelling maps. So we must have
(x, ϕβ (p)) = (x, gβα (p)ϕα (p)) (6.6)
for some smooth map gβα = ϕβ ϕ−1 α : P |Uβα → G (the “relabelling rule”). It
is fairly straightforward to see that gβα depends only on the basepoint x, due
to the compatibility of Φα , Φβ with the fibrewise G-actions. Thus the two
labelling conventions ϕα , ϕβ are related by
ϕβ (p) = gβα (x)ϕα (p),

64
for some transition function gβα : Uβα → G.
On triple overlaps, associativity (of group multiplication) implies that tran-
sition functions must satisfy the cocycle condition,
gαβ gβγ gγα = 1 on Uαβγ := Uα ∩ Uβ ∩ Uγ , α, β, γ ∈ I. (6.7)
Remark 3. It is important to realize that the transition functions gαβ , α, β ∈ I,
refer to a choice of trivializing cover (Uα , Φα )α∈I for P . Consider modifying
each Φα = (π, ϕα ) by picking smooth functions λα : Uα → G and defining
Φ′α = (π, λ−1 · ϕ ).
| α {z α}
ϕ′α

The transition functions obtained by using (Uα , Φ′α )α∈I are then
′ −1
gαβ = ϕ′α (ϕ′β )−1 = λ−1 −1
α · ϕα ϕβ · λβ = λα gαβ λβ .

So P only determines transition functions up to some relabellings (“local

′
changes of gauge”). The relation (6.8) says that the transition functions {gαβ }
and {gαβ } differ by a coboundary.
More generally, suppose there is another principal G-bundle P ′ → X, with
transition functions gα′ β ′ , α′ , β ′ ∈ I ′ with respect to another trivializing cover
{Uα′ , Φα′ }α′ ∈I ′ . To compare the two sets of transition functions, we consider
the open cover
{Uα ∩ Uα′ }α∈I,α′ ∈I ′ .
Every open set in this cover is contained in some open set of {Uα }α∈I ; likewise
for {Uα′ }α′ ∈I ′ . We say that the new cover is a common refinement of the
previous two covers. The unprimed transition functions can be restricted to
open sets of the new cover; likewise for the primed transition functions. Now
we can compare the two sets of restricted transition functions, since they are
both defined with respect to the same open cover of X.
Lemma 6.1. Let P → X and P ′ → X be two principal G-bundles, with
′
respective transition functions gαβ and gαβ defined with respect to the same
open cover {Uα }α∈I of X. Then P and P ′ are isomorphic iff there exist smooth
functions λα : Uα → G such that
′
gαβ (x) = λα (x)−1 gαβ (x)λβ (x), x ∈ Uαβ , α, β ∈ I. (6.8)
Proof. Exercise.

65
6.4.1 Principal bundle via transition functions
Intuitively, transition functions provide the instructions for building a global
principal G-bundle out of a collection of local trivial G-bundles Uα × G.
So suppose we are given smooth functions gαβ : Uαβ → G satisfying the
cocycle condition, Eq. (6.7), with respect to some open cover {Uα }α∈I of X.
This is called a (smooth) Čech 1-cocycle. Note that the following are also
automatically satisfied,
−1
gαα = 1, gαβ = gβα , α, β ∈ I.

With the Čech cocycle gαβ , we can construct a principal G-bundle as follows
(for full details, see §3.2 of [14], Prop. 5.2 of [9]).

ˆ Give I the discrete topology, and take the subset

T := {(x, g, α) ∈ X × G × I : x ∈ Uα }
G
= Uα × G × {α},
α∈I

which is open. (Why?)

ˆ “Gluing” via gαβ means to take equivalence classes

(x, g, α) ∼ (x′ , g ′ , β) iff x = x′ and g ′ = gβα g,

which makes sense because of the 1-cocycle condition, Eq. (6.7), (Exer-
cise). So take P to be the set of equivalence classes, with the quotient
topology, and define the projection map to be π[x, g, α] = x, and the
right G-action to be [x, g, α] · h = [x, gh, α]. Check well-definedness of π
and the G-action, that P is a manifold, and that the G-action is smooth.

ˆ Local trivializations of P are given by

Φα : π −1 (Uα ) → Uα × G, [x, g, α] 7→ (x, g), α ∈ I.

(Check diffeomorphism and G-equivariance properties).

ˆ Inspect that, by construction, the transition functions between the local

trivializations of P are precisely gαβ .

66
So if the gαβ came from some trivializing cover for Pinitial , then the principal
bundle reconstructed out of gαβ is isomorphic to Pinitial . Due to Lemma 6.1,
passing to a refinement and modifying the transition functions as in Eq. (6.8)
will not change the isomorphism class of the reconstructed bundle. So we
see that principal G-bundles over X are classified, up to isomorphism, by
equivalence classes of Čech 1-cocycles gαβ on X. This is known as the Čech
cohomology classification.

6.5 Frame bundle of a vector bundle

Let us return to the frame bundle of a general, possibly non-trivializable,
vector bundle E. As a set, Fr(E) is again Eq. (6.2), with the same projection
and fibrewise right G-action, where G = GL(n). We explain why this frame
bundle has a principal G-bundle structure.
Let (Uα , Φα )α∈I be a trivializing cover for E, with smooth transition func-
tions gαβ . Each local trivialization Φα : E|Uα ∼= Uα × Kn gives a corresponding
trivialization of the local frame bundle,

Φ̃α : Fr(E|Uα ) → Uα × G, G = GL(n),

as in Section 6.2.1. Here, the Φ̃α are G-equivariant and respect the projection
maps to the base. Each (x, g) ∈ (Uα ∩ Uβ ) × G is a basis {g1 , . . . , gn } of Kn
at x, where the gi denote the columns of g. Then Φ̃−1 α (x, g) is some frame
{e1,x , . . . , en,x } of Ex . If we switch to the β local trivialization of E, then this
frame would be mapped by Φ̃β to the basis {gβα (x)g1 , . . . , gβα (x)g1 } of Kn at
x, i.e., the element (x, gβα (x) · g) ∈ Uβ × G. Thus, we have

Φ̃β ◦ Φ̃−1
α : Uβα × G → Uβα × G
(x, g) 7→ (x, gβα (x) · g),

with the same smooth transition functions gβα .

We still lack the smooth manifold structure on Fr(E), compatible with
the smooth structures on the local frame bundles Fr(E|Uα ). For this, we first
topologize Fr(E) by declaring V ⊂ Fr(E) to be open iff V ∩ Fr(E|Uα ) is open
for all α ∈ I. The manifold charts of all the Fr(E|Uα ) ⊂ Fr(E) combine to
make Fr(E) a topological manifold, with smoothness of coordinate changes
inherited from that on E.

67
Remark. The vector bundle transition functions gαβ give a Čech 1-cocycle.
From this, we can also abstractly construct a principal G-bundle over X,
following the prescription of Section 6.4.1. This produces an isomorphic bundle
to the more concrete frame bundle described above, due to Lemma 6.1.

6.6 Gauge transformations

6.6.1 Local gauge transformations
Earlier, we saw that a local trivialization over Uα is the same thing as a local
section/gauge sα : Uα → P |Uα . So specifying a trivializing cover is the same
thing as equipping an open cover {Uα }α∈I with local sections sα . Explicitly,
the local trivialization associated with sα is

Φsα : (sα (x) · g ) 7→ (|{z}

x , g ).
| {z } |{z}
p∈P |Uα π(p) ϕα (p)

So if we have two local sections/gauges, sα : Uα → P and sβ : Uβ → P , there

are two ways of writing p ∈ π −1 (x) for points x ∈ Uαβ on the overlap,

sα (x) · ϕα (p) = p = sβ (x) · ϕβ (p).

Thus, the change of frame labelling when switching local gauges is precisely
implemented by the transition function,

ϕβ (p)ϕ−1
α (p) ≡ gβα (x).

We could also think of the transition function as implementing the change of

section/gauge,

sα (x) = sβ (x) · ϕβ (p)ϕ−1

α (p) = sβ (x) · gβα (x). (6.9)

In this capacity, the map gβα : Uβα → G is referred to as a local gauge trans-
formation (from the β to the α gauge). We will discuss this local viewpoint
further in Section 6.6.3.

6.6.2 Global gauge group

The global and gauge-independent notion of a gauge transformation is:

68
Definition 21. A gauge transformation of a principal G bundle P → X is an
automorphism of P (as a principal G-bundle) which covers the identity map
idX . The group G(P ) of gauge transformations is called the gauge group of P .
Example 6.4. Take X = pt, then P is just a G-torsor. A gauge transformation
is a self-map F : P → P which respects its structure as a G-torsor. So F
satisfies
F (p · g) = F (p) · g, ∀g ∈ G, p ∈ P. (6.10)
In other words, the G-relations between frames are preserved under the trans-
formation F .
Now, the target frames F (p) must have the form
F : p 7→ p · σF (p), (6.11)
for some uniquely-defined σF (p) ∈ G. That is, F must be implemented by
right-multiplication by some smooth map σF : P → G. Putting Eq. (6.11)
into Eq. (6.10) gives
F (p · g) = F (p) · g = (p · σF (p)) · g = (p · g) · (g −1 σF (p)g) . (6.12)
| {z }
=σF (p·g)

Therefore, the required σF : P → G is a G-equivariant map, where the action

on the target G is by conjugation.
This looks a little complicated, so consider a situation where we had a
reference frame p0 ∈ P , but wanted to switch to a shifted reference frame p0 ·h,
where h ∈ G. We do not simply shift all the other frames by h. Instead, the
correct gauge transformation Fh must be defined on a general frame p = p0 · g
by the formula
Fh (p) = F (p0 · g) = Fh (p0 ) · g = (p0 · h) · g = p · (g −1 hg).

Specifying gauge group. For P over a general base space X, a gauge

transformation F still respects each fibre as a G-torsor. The calculations in
Example 6.4 carry over in the same way, leading to Eq. (6.12). That is, the
gauge group is naturally identified with the following space of equivariant G-
valued maps on P ,
G(P ) ∼
= Map(P, G)G := {σ : P → G smooth : σ(p·g) = g −1 σ(p)g, ∀p ∈ P, g ∈ G},
acting on P by multiplication on the right.

69
Exercise 6.4. Check that Map(P, G)G is a group under pointwise multiplica-
tion, and that the correspondence

G(P ) → Map(P, G)G , F 7→ σF ,

where σF is defined by Eq. (6.11), is indeed a group isomorphism.

Abelian gauge transformations.

Example 6.5. If G is Abelian, then

Map(P, G)G = {σ : P → G smooth : σ(p · g) = σ(p) ∀p ∈ P, g ∈ G}.

Such a map is constant on the fibres, and descends to X. So we have a

canonical isomorphism

Map(X, G) → Map(P, G)G ∼

= G(P ), τ 7→ τ ◦ π.

Example 6.6. Consider the principal bundle P ∼= S 1 × U(1), with S 1 the unit
circle. For n ∈ Z, the maps gn : e 7→ e are elements of Map(S 1 , U(1)),
iθ inθ

and they implement gauge transformations of P . The gn with different val-

ues of n are not smoothly deformable to each other within Map(S 1 , U(1)).
Actually, any g ∈ Map(S 1 , U(1)) is smoothly deformable to some gn , so the
gauge transformations of P are sorted into deformation classes labelled by
n ∈ Z (which is the winding number of g). Those with winding number n ̸= 0
are called large gauge transformations since they are not deformable to the
identity transformation.

6.6.3 Local viewpoint

Let P |U be trivialized by a local section s : U → P |U . Let σ ∈ Map(P |U , G)G
be a gauge transformation of P |U (“local gauge transformation”). Then the
composition τ (σ) := σ ◦ s is an element of Map(U, G). We may then use τ (σ)
to change the local section/gauge to

s · τ (σ) : U → P |U .

We check that for σ1 , σ2 ∈ Map(P |U , G)G , we have

τ (σ1 ) · τ (σ2 )(x) = (σ1 ◦ s(x)) · (σ2 ◦ s(x)) = (σ1 · σ2 ) ◦ s(x) = τ (σ1 · σ2 )(x),

70
so
τ : Map(P |U , G)G → Map(U, G), σ 7→ τ (σ)
is a group homomorphism.
In reverse, let f ∈ Map(U, G). Define a map

η : Map(U, G) → Map(P |U , G)G

by
η(f )(s(x) · g) = g −1 f (x)g, x ∈ U, g ∈ G.
(Equivariance of η(f ) is clear from this formula.) Given f1 , f2 ∈ Map(U, G),
we check that

η(f1 · f2 )(s(x) · g) = g −1 (f1 (x)f2 (x))g = (g −1 f1 (x)g)(g −1 f2 (x)g))

= η(f1 )(s(x) · g)η(f2 )(s(x) · g)
= (η(f1 ) · η(f2 ))(s(x) · g),

therefore η(f1 · f2 ) and η(f1 ) · η(f2 ) coincide.

One may verify that τ and η are inverse maps. That is, they implement
group isomorphisms

Map(P |U , G)G ←→Map(U, G).

This is why, in physics, local gauge transformations are often simply regarded
as maps from the base U to the symmetry group G.
However, the identification of the local gauge transformations over U with
Map(U, G) uses a choice of reference local gauge s. So, for instance, gauge
transformations of non-trivializable P cannot simply be described as Map(X, G)
(in the case of non-Abelian G).

What are gauge transformations? It is hard to give an answer that

will satisfy all practitioners. Here is a possible one. As we will learn later,
the gauge group is supposed to act on some other spaces of objects, such as
connections on P , vector bundles associated to P , etc. If we “mod out” by
the action of the gauge group, we are really considering P only as an object
up to isomorphism. We had been thinking of P as an abstract frame bundle,
and each element of P as some frame located at some point x ∈ X. But
if we ultimately only care about the isomorphism class [P ] of P , then the

71
frames are not actually being regarded as things-in-themselves. Rather, only
the abstract inter-frame relations are considered important. Gauge-invariant
quantities are those which are intrinsic to this relational structure, as encoded
by [P ].

72
7 Preview of Spin
7.1 Hopf bundle over S 2
Consider S 3 as the set of unit vectors in C2 ,

S 3 = {(w, z) ∈ C2 : |w|2 + |z|2 = 1}.

The Hopf fibration is the smooth surjective map to the unit 2-sphere,

π : S 3 → S 2 ⊂ R3 = C × R
(w, z) 7→ (2wz̄, |w|2 − |z|2 ).

On S 3 , there is a smooth right action of eiθ ∈ U(1) by scalar multiplication,

(w, z) · eiθ = (weiθ , zeiθ ). It is easy to check that

π(w1 , z1 ) = π(w2 , z2 ) iff (w2 , z2 ) = (w1 , z1 ) · eiθ for some eiθ ∈ U(1).

So π is precisely the quotient map under the action of U(1).

The north and south poles of S 2 are

N = ((0, 0), 1), S = ((0, 0), −1),

and we shall consider the open cover

Uα = S 2 \ {N }, Uβ = S 2 \ {S},

with preimages

π −1 (Uα ) = {(w, z) ∈ S 3 : z ̸= 0}, π −1 (Uβ ) = {(w, z) ∈ S 3 : w ̸= 0}.

Define the maps

Φα : π −1 (Uα ) → Uα × U(1) Φβ : π −1 (Uβ ) → Uβ × U(1)

(w, z) 7→ (π(w, z); z/|z|) (w, z) 7→ (π(w, z); w/|w|)

These provide local trivializations to trivial principal U(1)-bundles. So π :

S 3 → S 2 is a principal U(1)-bundle, with transition function being

w/|w|
gβα : Uα ∩ Uβ → U(1), (2wz̄, |w|2 − |z|2 ) 7→ = exp(iArg(wz̄)).
z/|z|

73
 Note that the overlap Uα ∩Uβ ⊂ S 2 retracts to the equatorial circle S 1 ⊂ S 2 ,
and it is easy to see that the transition function gβα |S 1 : S 1 → U(1) has winding
number 1. This topologically obstructs the extension of gβα to include the
north and/or south poles. The significance of this obstruction is as follows.
Suppose there is a global trivialization Φ̃ : S 3 ∼= CP1 × U(1). Then there are
transition functions Φ̃α ◦ Φ̃−1 and Φ̃β ◦ Φ̃−1 , both of which must have winding
number zero when restricted to the equator, since they can be extended to
the south pole and the north pole respectively. This would contradict the
non-trivial winding number of gβα = Φ̃β ◦ Φ̃−1 3
α . Thus S is an example of a
non-trivializable principal U(1)-bundle.
Remark. There is also the quaternionic Hopf fibration π : S 7 → S 4 . Here, S 7
is regarded as the unit-length quaternions in H2 , and there is a right “scalar
multiplication” on S 7 by the unit quaternion group Sp(1) ∼ = S 3 . The quotient
manifold S is identified with the quaternionic projective space HP1 . As with
4

the complex Hopf fibration, there are local trivializations over the upper and
lower hemispheres of HP1 , with transition function on the equator S 3 ⊂ HP1
being a map gβα : S 3 → Sp(1) of degree 1. Thus S 7 → S 4 is a non-trivializable
principal Sp(1)-bundle. The interested reader is invited to work out the details.

7.2 Orthonormal frame bundle of S 2

The Euclidean vector space Rd has a standard inner product,
d
X
x·y = xi y i , x, y ∈ Rd .
i=1

Recall that each of its tangent spaces is itself canonically isomorphic to Rd , and
we give them the same inner product. This means that we consider Euclidean
space as a Riemannian manifold, whose tangent bundle Rd × Rd is equipped
with the standard Euclidean metric on each fibre {x}×Rd . Furthermore, there
is an orientation given by the standard basis.
Take the unit 2-sphere S 2 ⊂ R3 , whose points are labelled by a unit 3-vector
a. The tangent plane at a is
Ta R3 = {v ∈ R3 : v · a = 0}.
In total, the tangent bundle T S 2 is the subset
T S 2 = {(a, v) ∈ S 2 × R3 : v · a = 0},

74
and it is a rank-2 Euclidean vector bundle over S 2 . Actually, we can orient
each tangent space by declaring a basis {u, v} of Ta S 2 to be positively oriented
if {u, v, a} is positively oriented for R3 . So S 2 is an oriented Riemannian
manifold.
We can consider the oriented orthonormal frame bundle of S 2 , denoted
π : FrSO (S 2 ) → S 2 . By general theory to be developed later, this would be
a principal SO(2)-bundle, with the SO(2) action being rotation of oriented
frames for each tangent plane. For now, let us obtain this structure directly.
Write each element of FrSO (S 2 ) as

({u, v}; a), a ∈ S 2 , {u, v} oriented orthonormal basis for Ta S 2 .

The rotation of frames by g0 ∈ SO(2) is

({u, v}; a) · g0 = ({u, v} · g0 ; a).

Now, observe that {u, v, a} is a positively-oriented orthonormal basis for R3 ,

so the matrix (u v a) is an SO(3) matrix. Thus, we have an identification

FrSO (S 2 ) ∼
= SO(3), ({u, v}; a) ↔ (u v a), (7.1)

with π being the projection onto the last component a ∈ S 2 . Under this
identification, the action of SO(2) becomes
 

g0 0
(u v)g0 a = (u v a) ,
0 1

i.e., right multiplication by the subgroup

  
g0 0
G0 = , g0 ∈ SO(2) .
0 1

Thus the principal SO(2)-bundle structure of FrSO (S 2 ) → S 2 is identified with

the principal G0 -bundle structure of SO(3).

7.3 Bundle of spin frames over S 2

To recapitulate, we have the principal SO(2)-bundle

π : SO(3) ∼
= FrSO (S 2 ) → S 2 ,

75
as well as the principal U(1)-bundle

π : S 3 → S 2.

Actually, we shall use the dual U(1)-bundle

π̄ : S 3 → S 2 , (w, z) 7→ (2w̄z, |w|2 − |z|2 ),

which has clutching transition function S 1 → U(1) having winding −1 instead

of +1.
Now, S 3 is the manifold underlying the Lie group SU(2),
 
3 ∼ w −z̄
S = SU(2), (w, z) ←→ .
z w̄

Under this identification, the corresponding

 iθ U(1)-action
 on SU(2) is given by
e 0
right multiplication by the subgroup .
0 e−iθ
We shall make use of the famous 2 : 1 group homomorphism F : SU(2) →
SO(3) (exercise), indicated in the following commutative diagram:

Re(w2 −z̄ 2 ) Im(w2 +z̄ 2 ) 2 Re(w̄z)

   
w −z̄ F
S3 ∼
= SU(2)
F
SO(3) z w̄
−Im(w2 −z̄ 2 ) Re(w2 +z̄ 2 ) 2 Im(w̄z)
−2 Re(wz) −2 Im(wz) |w|2 −|z|2
π .
π̄ π̄ π
S2 ⊂ C × R
(2w̄z, |w|2 − |z|2 )
(7.2)
The map F restricts to a double-covering,
 
 iθ  cos(2θ) sin(2θ) 0
e 0
χ : U(1) → SO(2), −iθ 7→ − sin(2θ) cos(2θ) 0 .
0 e
0 0 1

In its capacity as a double-covering of SO(2), the group U(1) is called the spin
group, Spin(2), and we say that π̄ : S 3 ∼
= SU(2) → S 2 is a principal Spin(2)-
bundle. To stress this principal bundle structure, we write S 3 ∼
= FrSpin (S 2 ).
The commuting diagram Eq. (7.2) may be summarized as

F (S · g) = F (s) · χ(g), π̄(S) = π(F (S)).

76
The key observation is that we have “upgraded” the oriented orthonormal
frame bundle FrSO (S 2 ) to some “spin frame bundle” FrSpin (S 2 ), which has
“twice as many frames” at each point. Such an “upgrade” is mathematically
called a spin structure on S 2 (see Section 16.1).
Why do we stress the label Spin(2) instead of U(1)? Why the word “Spin”?
First, the double cover Spin(2) → SO(2) means that completing a full loop in
the orthonormal frame space SO(2) only corresponds to completing half a loop
in Spin(2). So we do not get back to where we started from, but rather acquire
a −1 mismatch, ending up with the “opposite spin frame” to the original one.
This is precisely the characteristic feature of fermionic fields in physics.
A fermion field is not like the classical geometry vector/tensor fields. On
an n-dimensional oriented Riemannian manifold X, the space of “spin frames”
at each point is labelled by the “spin rotation group” Spin(n) rather than the
classical rotation group SO(n). Globally, there is a spin frame bundle FrSpin (X)
over X.
In general, there exist topological obstructions to defining the spin frame
bundle globally, and there may not be a unique choice of spin structure. The
classification of spin structures on X is not the same thing as the classification
of principal Spin(n)-bundles over X — the doubling relation between a spin
frame bundle and the oriented orthonormal frame bundle of X is a defining
ingredient for the notion of spin and spinors. We also mention that Spin(2) is
not the same as the U(1) appearing in quantum mechanics or electromagnetism
discussed in the introduction.
Incidentally, for n = 3, the double cover SU(2) → SO(3) exhibits Spin(3) ∼ =
SU(2) as the space of “spin rotations” at a point on a 3-dimensional manifold.

77
8 Tensors and differential forms
8.1 Commutator of vector fields
Over a coordinate chart U , the coordinate tangent vector fields ∂i provide
a commuting family of derivations of C ∞ (U ). This expresses the convenient
notion that partial derivatives commute. Unfortunately, ∂i is seldom globally
defined.
For general vector fields u, v ∈ X(X), we will acquire a commutator

[u, v] : u ◦ v − v ◦ u

whenever we swap their order of operation,

u(v(f )) = v(u(f )) + [u, v](f ).

The commutator of two derivations turns out to be another derivation (Exer-

cise), so [u, v] is also a vector field. Formally, [·, ·] is a Lie bracket operation
on X(X) × X(X), satisfying the following properties

[λu + v, w] = λ[u, w] + [v, w]

[v, w] = −[w, v]
0 = [u, [v, w]] + [v, [w, u]] + [w, [u, v]], u, v, w ∈ X(X), λ ∈ R.
(8.1)

Thus X(X) is an infinite-dimensional Lie algebra.

Exercise 8.1. Verify that

[f ·u, g ·v] = f g ·[u, v]+f (u(g))·v −g(v(f ))·u, v, w ∈ X(X), f, g ∈ C ∞ (X).

So the commutator is not C ∞ -linear in either entry. Consequently, [·, ·] is

not a “tensorial” operation — it does not make sense for individual tangent
vectors. Nevertheless, the Lie bracket is natural :
Exercise 8.2. Let u, v ∈ X(X), and f : X → Y be a diffeomorphism. Show
that
f∗ [u, v] = [f∗ u, f∗ v].

78
8.2 Differential 1-forms
Recall Definition 6 of the cotangent spaces Tx∗ X := (Tx X)∗ . From doing
Exercise 4.4, we learn that the basis {(dxi )x } for Tx∗ X is dual to the basis of
coordinate tangent vectors,

(dxi )x (∂j |x ) = δji .

A general ω ∈ Tx∗ X can therefore be expanded as ω = ωi (dxi )x .

Exercise 8.3. Show that the components ωi = ω ∂i |x transform covariantly
as
∂ x̃j ∂xi
ωi = ω̃j or ω̃j = ωi (8.2)
∂xi φ(x) ∂ x̃j ψ(x)
under a change of coordinates xi → x̃j . (Cf. contravariant transformation
property for the components of tangent vectors, Eq. (4.3).)
Definition 22. A (differential) 1-form on a manifold X is an assignment of
cotangent vectors, ω : x 7→ ωx ∈ Tx∗ X, such that its component functions in
some (thus any) coordinate chart are smooth. The space of 1-forms on X is
denoted Ω1 (X).
Example 8.1. Over a coordinate chart U , we have the coordinate 1-forms
(dxi ), i = 1, . . . , n, dual to the coordinate tangent vector fields ∂i .
Example 8.2. We can cover the unit circle S 1 with two charts, with angular
coordinates θ ∈ (0, 2π) and θ̃ ∈ (−π, π). The change-of-coordinates map is
the identity map on the upper half-circle, and the shift of −2π on the lower
∂
half-circle. The coordinate vector fields ∂θ and ∂∂θ̃ coincide on the overlap, and
∂
we usually just abuse notation and write ∂θ for the globally defined tangent
1
vector field over S . Similarly the “coordinate” 1-form dθ is globally defined
over S 1 .
The pairing of 1-forms with vector fields works as follows. Let ω ∈ Ω1 (X)
and v ∈ X(X). Then we can define the function ω(v) : X → R to be

ω(v)(x) = ωx (vx ) ∈ R.

Expanding in an arbitrary coordinate basis,

n
X n
X n
X
i j j
(x)δji ωi (x)v i (x),


ωx (vx ) = ωi (x)(dx )x v (x)∂j |x = ωi (x)v =
i,j=1 i,j=1 i=1

79
which depends smoothly on x, since the components ωi and v i depend smoothly
on x. Thus we may regard a 1-form as a map ω : X(X) → C ∞ (X).
Exercise 8.4. Show that a 1-form ω is C ∞ (X)-linear, in the sense that

ω(f · v) = f · ω(v), f ∈ C ∞ (X), v ∈ X(X).

Conversely, show that any C ∞ (X)-linear map ω : X(X) → C ∞ (X) must be a

1-form.
Example 8.3. Let us revisit the derivative df of a function f ∈ C ∞ (X). So
f : X → R as a map of manifolds, and df : T X → T R = R × R. Given any
vector field v ∈ X(X), we have

df (v)(x) := dfx (vx ) : Tx X → Tf (x) R = R,

so df (v) is a R-valued function on X, which depends smoothly on x, since

dfx (vx ) = vx (f ) = v(f )(x),

and the right-hand side depends smoothly on x (Exercise 5.1). Thus we can
regard df as a 1-form,

df : X(X) → C ∞ (X)
v 7→ v(f ).

Cf. Example 4.6, where we saw that dfx was a cotangent vector at x. If you
prefer, write in local coordinates,

∂ fˇ i
 ˇ 
∂f i
v(f ) = i
v = dx (v j ∂j ) .
∂x ∂xi | {z }
| {z } v
df

∂f ˇ
Although we might be used to thinking of the n-tuple ∂x i as a “gradient vector

field”, it is perhaps better to think of it as a 1-form, which eats up various

tangent vectors to give directional derivatives of f .
Remark. In the same way as for T X, we can construct the cotangent bundle
T ∗ X as a (2n)-manifold. Alternatively, when we study T X as fibre bundles
later, we will see that T ∗ X may be viewed as the vector bundle with transition
functions being the transpose of those of T X. Then 1-forms are sections of
the cotangent bundle T ∗ X.

80
8.3 Pullbacks and push-forwards
It is usual to write Ω0 (X) := C ∞ (X), and refer to smooth functions as “0-
forms”. If g ∈ Ω0 (Y ), then it may be pre-composed with a smooth map
f : X → Y to obtain a smooth 0-form
f ∗ (g) ∈ Ω0 (X), f ∗ (g) := g ◦ f.
This is called the pullback of g by the map f .
Similarly, let ω ∈ Ω1 (Y ). Then there is a pullback 1-form f ∗ ω ∈ Ω1 (X),
defined by
(f ∗ ω)(v)(x) = ωf (x) (dfx (vx )), v ∈ X(X). (8.3)
We should verify that f ∗ ω(v) is indeed a smooth function of x. This can be
shown from the following properties of the pullback:
Exercise 8.5. Let ω ∈ Ω1 (Y ) be a 1-form, and let f : X → Y and g ∈ C ∞ (Y )
be smooth maps. Verify the following formulae:
ˆ f ∗ (g · ω) = (f ∗ (g)) · f ∗ ω.

ˆ f ∗ (dg) = d(f ∗ (g)).

Exercise 8.6. Let f1 : X → Y and f2 : Y → Z be smooth maps. Check that
(f2 ◦ f1 )∗ = f1∗ ◦ f2∗
when applied to Ω1 (Z) or to Ω0 (Z).
Given a smooth map f : X → Y , we could try to “push-forward” a vector
field on X to become a vector field on Y . Indeed, we have the derivative map
df : T X → T Y , built out of the pointwise assignments dfx : Tx X → Tf (x) Y .
However, because f may not be a bijection, df may not map sections of T X
to sections of T Y . Nevertheless, if f is a diffeomorphism, then we can use f
to push forward a vector field to another vector field.
Exercise 8.7. Let v ∈ X(X) and f : X → Y be a diffeomorphism. Define f∗ v
by
f∗ v := df ◦ v ◦ f −1 : Y → T Y
(f∗ v)y = dff −1 (y) (vf −1 (y) ) ∈ Ty Y, y ∈ Y. (8.4)
Check that f∗ v is indeed a smooth section of T Y , i.e., a smooth vector field
over Y .

81
8.4 2-forms and Riemannian metrics
The dual space V ∗ of a (real) vector space can be generalized to the space of
bilinear maps (or forms) V × V → R. A bilinear form b is non-degenerate if

b(v1 , v2 ) = 0 for all v2 ∈ V implies v1 = 0.

In this case, the bilinear form induces an isomorphism

♭ : V → V ∗, v 7→ (u 7→ b(u, v)),

with inverse denoted ♯ : V ∗ → V (this should not be confused with the same
notation used for so-called fundamental vector fields later).
A symmetric (resp. antisymmetric) bilinear form b satisfies b(v1 , v2 ) =
b(v2 , v1 ) (resp. b(v1 , v2 ) = −b(v2 , v1 )) for all v1 , v2 ∈ V . (Anti)symmetric
forms constitute a vector space.
Example 8.4. An inner product on a real vector space V is a symmetric bilinear
form g which is positive definite (g(v, v) ≥ 0 with equality iff v = 0). This
is automatically non-degenerate. There is an induced inner product on V ∗ ,
denoted with the same symbol, and given by g(η, ω) = g(η ♯ , ω ♯ ).
Example 8.5. If η, ω ∈ V ∗ , their tensor product is the bilinear form

η ⊗ ω : (u, v) 7→ η(u)ω(v).

If {e1 , . . . , en } is a basis for V and {e1 , . . . en } the dual basis for V ∗ , then ei ⊗ej
is a basis for the bilinear forms on V (Exercise).
Example 8.6. We can symmetrize η ⊗ ω by taking 21 (η ⊗ ω + ω ⊗ η). Similarly,
we can take the antisymmetrized tensor product, also called the wedge product,

η ∧ ω := η ⊗ ω − ω ⊗ η. (8.5)

The space
of skewsymmetric bilinear forms is denoted Λ2 (V ), and it has di-
mension n2 (Exercise).
On a manifold X, we consider a family of bilinear forms bx : Tx X × Tx X →
R, x ∈ X to be smooth, if in all coordinate charts (of an atlas) (Uα , φα ) of X,
we have
x 7→ bx (∂i |x , ∂j |x ), x ∈ Uα ,
being smooth, where ∂i are the coordinate tangent vector fields. We will simply
refer to a smooth family of bilinear forms as a “bilinear form on X”. Notice

82
that the wedge product of two one-forms η, ω can be defined by the pointwise
wedge product of Eq. (8.5), and this produces an antisymmetric bilinear form.
As in Exercise 8.4, we can also define a bilinear form on X as a C ∞ (X)-
bilinear map
X(X) × X(X) → C ∞ (X).
Definition 23. A (smooth) positive-definite symmetric bilinear form g on X
is called a Riemannian metric. An antisymmetric bilinear form ω on X is
called a (differential) 2-form. The space of 2-forms on X is denoted Ω2 (X).
Remark. If we think of Tx X as a linearized version of X near x, then a Rie-
mannian metric provides the notion of lengths and angles on Tx X. A 2-form
prescribes the “oriented area” of the parallelogram spanned by tangent vec-
tors u, v (in that order). There may not be any relationship between a given
Riemannian metric and a given 2-form.
Just as 1-forms can be pulled back under a map f : Y → X, we can also
pull back bilinear forms,
(f ∗ b)y (uy , vy ) := bf (y) ((df )y (uy ), (df )y (vy )), uy , vy ∈ Ty Y. (8.6)
It is easy to see that Riemannian metrics get pulled back to Riemannian
metrics, and 2-forms get pulled back to 2-forms. Furthermore, the wedge
product is compatible with pullback,
f ∗ (η ∧ ω) = f ∗ η ∧ f ∗ ω. (8.7)
Example 8.7. On RN as a manifold, all its tangent spaces are canonically RN ,
and they can be given the standard inner product. This gives the standard
Euclidean space as a Riemannian manifold. If X is a submanifold of RN , then
the Euclidean space metric restricts to a Riemannian metric on X.

8.5 Differential k-forms and exterior derivative

Now, recall from Example 8.3 that d turns 0-forms (functions) into 1-forms.
This is actually just one part of the exterior calculus on general differential
forms. Let us see how this works on 1-forms.
Definition 24. Let ω ∈ Ω1 (X) be a 1-form on X. Its exterior derivative dω
is the 2-form defined by
dω(u, v) := u(ω(v)) − v(ω(u)) − ω([u, v]). (8.8)

83
Exercise 8.8. Check that dω as defined above is indeed a 2-form. (Exercise 8.1
would be helpful.)
Exercise 8.9. Check that the exterior derivative d : Ω1 (X) → Ω2 (X) satisfies:
For λ ∈ R, η, ω ∈ Ω1 (X), f ∈ C ∞ (X),
ˆ R-linearity: d(η + λω) = dη + λdω;
ˆ Leibniz rule: d(f · ω) = f · dω + df ∧ ω;
ˆ Chain complex: d(df ) = 0;
Pn
ˆ Coordinate formula: In a coordinate chart (U, φ), so that ω|U = i=1 ωi dxi
for some ωi ∈ C ∞ (U ), we have
n n n
X
i
X
i
X ∂ωi j
d( ωi dx ) = dωi ∧ dx = j
dx ∧ dxi . (8.9)
i=1 i=1 i,j=1
∂x

ˆ Naturality: If h : Y → X is smooth, then h∗ (dω) = d(h∗ ω).

One can go beyond bilinearity, and study multilinear maps V ×. . .×V → R
with k-arguments. Taking V = Tx X and requiring smooth dependence on x,
we have the C ∞ (X)-multilinear forms,
X(X) × . . . × X(X) → C ∞ (X).
The subspace of totally antisymmetric ones (minus sign acquired upon ex-
changing any pair of arguments) are called differential k-forms, and is denoted
Ωk (X). Intuitively, at each x, a k-form assigns an “oriented volume” to each
parallelepiped spanned by k given tangent vectors.
The wedge product of a k-form η and an l-form ω may be defined as
1 X
η ∧ ω(v1 , . . . , vk+l ) = sgn(σ)η(vσ(1) , . . . , vσ(k) )ω(vσ(k+1) , . . . , vσ(k+l) ),
k! l! σ∈S
k+l

where v1 , . . . vk+l ∈ X(X), and Sr is the symmetric group (i.e. permutations)

on r objects. The pullback of k-forms is defined in the same way as Eq. (8.6),
and it is a straightforward exercise to check that pullback commutes with
wedge product, Eq. (8.7).
The exterior derivative extends to R-linear maps d : Ωk (X) → Ωk+1 (X),
and it can be axiomatically characterized as the unique R-linear maps satisfy-
ing the conditions

84
 ˆ df is the usual differential for f ∈ Ω0 (X) ≡ C ∞ (X);

ˆ d(η ∧ ω) = dη ∧ ω + (−1)p (η ∧ dω) when η is a p-form;

ˆ d2 = 0.

In particular, Definition 24 of dω for a 1-form ω coincides with this abstract

characterization, as we verified in Exercise 8.9. For a general k form, the local
coordinate formula for dη, analogous to Eq. (8.9), may be likewise derived.
Exercise 8.10. Show that the exterior derivative of a 2-form η ∈ Ω2 (X) can be
expressed as

dη(u, v, w) = u(η(v, w)) + v(η(w, u)) + w(η(u, v))

− η([u, v], w) − η([v, w]u) − η([w, u], v), u, v, w ∈ X(X).
(8.10)

85
9 Integral curves and flows
9.1 Integral curves of vector fields
Every tangent vector at x has a geometric representative as a curve through
x. For a tangent vector field, we might expect that there is a smooth family
of curves on X, whose velocity vectors reproduce the given vector field.
Definition 25. Let v be a vector field on a manifold X. An integral curve of
v is a smooth curve γ : (a, b) → X such that
γ ′ (t) = vγ(t) , t ∈ (a, b).
Example 9.1. Consider R2 with (x1 , x2 ) coordinates, and global vector field v =
∂2 . The integral curves are vertical lines. For example, γ(t) = (x0 , y0 +t), t ∈ R
has, for each f ∈ C ∞ (R2 ),
d(f ◦ γ) ∂f
γ ′ (t)(f ) = = = ∂2 |γ(t) (f ) = vγ(t) (f ).
ds s=t ∂x2 γ(t)

So the velocity vector γ ′ (t) equals the vector field vγ(t) for all t ∈ R.
Example 9.2. In standard global (x1 , x2 ) coordinates for R2 , take the vector
field v = x1 ∂2 − x2 ∂1 . It is convenient to work with (local) polar coordinates
(r, θ) on U = R2 \ {(x, 0) : x ≤ 0}. Check that on U , we have v = ∂θ
(Exercise). The curves
γ(t) = (r0 , θ0 + t), r0 > 0, θ0 + t ∈ (−π, π)
have the property that γ ′ (t) = ∂θ |(r0 ,θ0 +t) = vγ(t) . Actually, these curves can
smoothly extended to t ∈ R (write them in Cartesian coordinates).
Within a local coordinate chart, the problem of finding integral curves
reduces
Pnto that of solving first-order ODEs. In a coordinate chart (U, φ), write
i i i
v = i=1 v ∂i with v ≡ v (x) the component functions. A curve γ lying
within U has the form
γ(t) = φ−1 (γ i (t), . . . , γ n (t))
for functions γ i : (a, b) → R. For f ∈ C ∞ (X), we have
n n
d(f ◦ γ) X ∂(f ◦ φ−1 ) dγ i X dγ i
γ ′ (t)(f ) = = ≡ ∂i |γ(t) (f ).
ds s=t
i=1
∂xi γ(t) ds s=t
i=1 |ds{zs=t}
γ̇ i (t)

86
So the velocity vector of γ at time t is
n n
need
X X
′ i
γ (t) = γ̇ (t)∂i |γ(t) = v i (γ(t))∂i |γ(t) .
i=1 i=1

Comparing the coefficients, we have the ODEs

γ̇ i (t) = v i (φ−1 (γ 1 (t), . . . , γ n (t))), i = 1, . . . , n, (9.1)
for the coordinate functions γ i of the curve. Requiring the curve to pass
through a prescribed point x0 ∈ U at time t0 , is the same thing as setting the
initial conditions
(γ 1 (t0 ), . . . , γ n (t0 )) = φ(x0 ). (9.2)

ODE theory ⇒ local existence. There always exists some time interval
containing t0 , during which the initial value problem (9.1)-(9.2) has a (unique,
smooth) solution. See Theorem D.1 of [11] for details of the ODE theory
justifying this fact.
Proposition 9.1. Let v be a vector field on a manifold X. For each point
x ∈ X, there exists ϵ > 0 and an integral curve γ : (−ϵ, ϵ) → X starting from
x, i.e., γ(0) = x.
Generally, we may not be able to extend an integral curve to one which is
defined for all t ∈ R.
Exercise 9.1. Find an integral curve of the vector field ((x1 )2 + (x2 )2 )∂1 on
R2 , passing through the point (1, 0). Check that this integral curve cannot be
extended to all of R.
If γ is an integral curve of a vector field v, we have (basically by definition),
f (γ(t + t′ )) − f (γ(t))
vγ(t) (f ) = lim .
′ t →0 t′
So a vector field v differentiates functions along its integral curves.

9.2 Flows
Definition 26. A global flow on a manifold X is a smooth R-action on X,
i.e., a smooth map
θ :R×X →X

87
satisfying
θ(t, (θ(s, x)) = θ(t + s, x), θ(0, x) = x, x ∈ X, s, t ∈ R.
For each t ∈ R, a flow determines the smooth map
θt : X → X, x 7→ θ(t, x).
These θt satisfy a group law when composed,
θ0 = idX , θt ◦ θs = θs+t , s, t ∈ R,
so they are diffeomorphisms.
Starting from an initial point x ∈ X, a flow provides a smooth curve
γ (x) : R → X, t 7→ θt (x), (9.3)
and therefore a tangent vector at x,
vx = (γ (x) )′ (0) ∈ Tx X. (9.4)
Definition 27. The tangent vector assignment, Eq. (9.4), associated to a flow
on X is called the infinitesimal generator of the flow.
Proposition 9.2. Let θ be a global flow on X. Its infinitesimal generator is
a smooth vector field, and the γ (x) of Eq. (9.4) are its integral curves.
Proof. We may check that v is a derivation of C ∞ (X) (exercise), so it is a
smooth vector field (Remark 2).
Next, consider the curve γ (x) starting at a given x ∈ X. We need to check
the integral curve condition, vγ (x) (t) = (γ (x) )′ (t), t ∈ R.
Fix t ∈ R and write x′ := γ (x) (t) = θt (x). We have
′
γ (x ) (s) ≡ θs (x′ ) = θs ◦ θt (x) = θs+t (x) = γ (x) (s + t).
So for any f ∈ C ∞ (X),
′
(x′ ) ′ d(f ◦ γ (x ) (s))
vγ (x) (t) (f ) ≡ vx′ (f ) ≡ (γ ) (0)(f ) =
ds s=0
d(f ◦ γ (x) (s + t))
=
d(s + t) s+t=t
(x) ′
= (γ ) (t)(f ),
as required.

88
 On R2 , the flows given by vertical translation by t and anticlockwise rota-
tion by an angle t, respectively give rise to the vector fields of Examples 9.1,
9.2. However, the example of Exercise 9.1 shows that a general vector field
might not be associated to any globally defined flow.
So in general, we have to consider local flows θ defined only on some open
subset D ⊂ R × X called the flow domain. At time t, the flow map θt is only
defined on a space-time subset

Xt := {x ∈ X : (t, x) ∈ D}.

So from an initial point x, the curve θ(·, x) only makes sense for some open
time interval D(x) containing 0. This is enough to make sense of the tangent
vectors vx in Eq. (9.4), and the notion of infinitesimal generator.

Theorem 9.3. Let v be a vector field on X. There is a unique maximal flow

θ : D → X with infinitesimal generator being v. The curve γ (x) : D(x) → X
is the unique maximal integral curve of v starting at x. For each t ∈ R, the
maps
θt : Xt → X−t
are well-defined diffeomorphisms (of possibly empty open subsets of X), with
θt−1 = θ−t . This (maximal) flow is called the flow generated by v.

A full proof may be found in [11], Theorem 9.12. On some subset of X,

the θt satisfy θt ◦ θs = θs+t . So a vector field is sometimes said to generate a
“local 1-parameter group of local diffeomorphisms”.

Definition 28. A vector field on a manifold is complete if it generates a global

flow; equivalently, all its (maximal) integral curves are defined for all t ∈ R.

Here is a sufficient criterion for v to be complete.

Lemma 9.4 (Uniform time Lemma). Let v be a vector field on a manifold X,

and θ be its flow. Suppose there exists δ > 0 such that for all x, the curves
γ (x) are defined at least for t ∈ (−δ, δ). Then v is complete.

Proof. Suppose, for a contradiction, that v has some integral curve γ (x) which
can only be defined for t ∈ (−δ, tmax ) with δ ≤ tmax < ∞. Define x′ =

89
 ′
γ (x) (tmax − 2δ ). By hypothesis, the integral curve γ (x ) starting at x′ is at least
defined for t ∈ (−δ, δ). We try extending γ (x) as follows,
(
γ (x) (t), t ∈ (−δ, tmax ),
γ (x) (t) = ′
γ (x ) (t − tmax + 2 ), t ∈ (tmax − 3δ2 , tmax + 2δ ).
δ

This is well-defined on the overlapping time range, due to

′ δ
γ (x ) (t − tmax + ) = θt−tmax + δ (x′ ) = θt−tmax + δ ◦ θtmax − δ (x)
2 2 2 2
(x)
= θt (x) = γ (t).

The extended γ (x) remains an integral curve starting at x, but it has now been
defined until tmax + 2δ > tmax .

9.3 Lie derivative

Let v, w be vector fields, and θ be the (local) flow generated by v. It might
appear that we can simply define the “derivative of w along v”, as
(θ−t )∗ (wθt (x) ) − wx
Lv (w)(x) := lim . (9.5)
t→0 t
In other words, we use the flow of v to transport the tangent vectors wθt (x)
back to the basepoint x. After this transportation, the tangent vectors belong
to the same tangent space, so the right side of Eq. (9.5) makes sense.
In fact, one can compute (exercise) that this “derivative” is given by taking
the commutator with v,
(θ−t )∗ (wθt (x) ) − wx
[v, w]x = lim .
t→0 t
Definition 29. Let v, w ∈ X(X). The Lie derivative of w along v is the vector
field
Lv (w) := [v, w].
However, suppose we only have a curve with velocity vector vx , and we
wanted to define the rate of change of the vector field w along this curve. The
Lie derivative cannot do this for us, since a vector field is required to generate
a local flow for the transport of tangent vectors.

90
10 Geometry of Lie group actions
10.1 Lie algebra of left-invariant vector fields
Definition 30. A vector field v on a Lie group G is left-invariant if

(Lg )∗ v = v, ∀g ∈ G.

Left-invariance is preserved by taking linear combinations, and even more

is true:

Proposition 10.1. The commutator of two left-invariant vector fields on a

Lie group is left-invariant.

Proof. For left-invariant vector fields v, w on a Lie group G, we have (Lg )∗ v = v

and (Lg )∗ w = w for all g ∈ G. So using Exercise 8.2,

(Lg )∗ [v, w] = [(Lg )∗ v, (Lg )∗ w] = [v, w].

Recall from Eq. (8.1) that X(X) is a Lie algebra under the Lie bracket [·, ·].
Prop. 10.1 says that the left-invariant vector fields on X form a Lie subalgebra
of X(X).

Definition 31. The Lie algebra of a Lie group G, denoted g, is the vector space
of left-invariant vector fields on G, with Lie bracket being the commutator of
vector fields.

Theorem 10.2. Let g be the Lie algebra of a Lie group G. Evaluation at the
identity element e ∈ G,

ε : g → Te G, v 7→ ve ,

is a linear isomorphism.

Proof. The linearity and injectivity of ε is straightforward. For surjectivity,

pick ξ ∈ Te G, and look for a left-invariant vector field v such that ε(v) ≡ ve =
ξ. Left-invariance means that the only possible candidate has

vg = ((Lg )∗ v)g = (dLg )e (ve ) = (dLg )e (ξ), g ∈ G.

91
From the group property of G, it follows that this tangent vector assignment

v : g 7→ (dLg )e (ξ) (10.1)

is left-invariant, and has ve = (dLe )e (ξ) = d(idG )e (ξ) = ξ.

It remains to check that the v constructed above is a smooth tangent vector
field. By Exercise 5.1, we need to check that for any f ∈ C ∞ (G), the function

v(f ) : G → R
g 7→ vg (f )

is smooth. To this end, let γ : (−δ, δ) → G be a curve in G with γ(0) = e and

γ ′ (0) = ξ. Then we compute

d(f ◦ Lg ◦ γ)
v(f )(g) ≡ vg (f ) = ((dLg )e (ξ)) (f ) = ξ (f ◦ Lg ) = . (10.2)
ds s=0

We regard f ◦ Lg ◦ γ as a function

η : (−δ, δ) × G → R, (s, g) 7→ f ◦ Lg ◦ γ(s),

which is smooth. Then the right-hand-side of Eq. (10.2) equals the partial
derivative ∂η
∂s
, which is therefore smoothly dependent on g ∈ G. This
(s,g)=(0,g)
shows that v(f ) ∈ C ∞ (G) as required.
Thus there is only a finite-dimensional vector space of left-invariant vector
fields on G, with, dim g = dim Te G = dim G. Notice that Te G inherits a Lie
bracket operation from its identification with g.

Corollary 10.3. A Lie group is parallelizable, i.e. its tangent bundle is triv-
ializable.

Proof. Pick any basis {ξ1 , . . . , ξn } for Te G. Then we obtain left-invariant vec-
tor fields v (ξi ) ∈ g, i = 1, . . . , dim G by Theorem 10.2, Eq. (10.1), and they
are linearly independent at every g ∈ G. So we obtain a global frame for T G,
which determines a global trivialization by Exercise 5.3.
Example 10.1. GL(n, R) is an open submanifold of the vector space Mn (R) ∼
=
2
Rn . So Te (GL(n)) is naturally isomorphic to Mn (R) as a vector space. On
the one hand, the matrix commutator provides a Lie bracket for Te (GL(n)).

92
On the other hand, Theorem 10.2 tells that Te (GL(n)) is canonically identified
with the left-invariant vector fields on GL(n, R), so we get another Lie bracket
on Te (GL(n)) using the commutator of vector fields. These two natural Lie
brackets on Te (GL(n)) coincide (Exercise).
Thus it is customary to write gl(n, R) = Mn (R), and gl(n, C) = Mn (C) for
these matrix Lie algebras. If you prefer, we can also work with GL(V ) for a
general vector space V without a preferred basis. Then gl(V ) = End(V ), the
linear maps on V , with commutator as Lie bracket.

Definition 32. A Lie group homomorphism f : G → H is a smooth group ho-

momorphism, while a Lie group homomorphism g → h is a bracket-preserving
linear map.

Exercise 10.1. Let f : G → H be a Lie group homomorphism. Regard (df )e :

Te G → Te H as a linear map g → h through the identification of Theorem 10.2.
Show that (df )e is a Lie algebra homomorphism.
The above induced homomorphism is usually denoted f∗ : g → h.
Exercise 10.2. Show that if G is an abelian Lie group, then its Lie algebra g
of left-invariant vector fields has trivial Lie bracket, [u, v] = 0 for all u, v ∈ g.

10.2 Matrix Lie group and algebra examples

Example 10.2. The matrix Lie subgroups O(n), SO(n), U(n), SU(n) have their
corresponding Lie algebras being Lie subalgebras of gl(n, R) or gl(n, C), i.e.,
certain linear subspaces of matrices equipped with the matrix commutator.
Example 10.3. As an exercise, we had calculated the tangent space of O(n) at
the identity element, thus also that of SO(n), as the space of real antisymmetric
matrices, and this is identified with o(n) ∼
= so(n) via Theorem 10.2. A similar
calculation shows that (s)u(n) is the space of (traceless) antihermitian complex
n × n matrices.
Remark. Non-isomorphic Lie groups can have isomorphic Lie algebras. For
example, the Lie algebra of GL(n, R)+ (the positive determinant subgroup) is
the same as that of GL(n, R); similarly, SO(n) and O(n) have the same Lie

93
algebra. A more interesting example is su(2) ∼
= so(3). Take
      
1 0 1 1 0 −i 1 1 0
su(2) = spanR , ,
2i 1 0 2i i 0 2i 0 −1
     
 0 0 0 0 0 1 0 −1 0 
so(3) = spanR 0 0 −1 ,  0 0 0 , 1 0 0 . (10.3)
0 1 0 −1 0 0 0 0 0
 

A direct calculation shows that the linear map taking the above basis vectors
to each other is a Lie algebra isomorphism (Exercise).
Remark. Even if G is known to be a matrix Lie group, G ⊂ GL(n), it is often
represented on some other vector space V other than Cn . For example, there
is a distinguished representation of G on the vector space g (see Section 10.3).
If G is represented as a subgroup of GL(V ), then g will be represented as a
Lie subalgebra of End(V ).
One should always keep in mind the geometric meaning of g as left-invariant
vector fields on G. The description as the tangent space at the identity matrix,
is largely for convenience of calculations.

10.3 Adjoint representation

Conjugation of G by any element g ∈ G,

Cg : G → G, g ′ 7→ gg ′ g −1 ,

is a group automorphism of G. In fact, Cg is a diffeomorphism, thus a Lie

group automorphism. Therefore, there is an induced automorphism of g ∼
= Te G
(Exercise 10.1),
Adg := (Cg )∗ = (dCg )e : g → g.
So Adg ∈ GL(g) for each g ∈ G. Since

Adg1 ◦ Adg2 = Adg1 g2 ,

the map
Ad : G → GL(g), g 7→ Adg ,
is itself a Lie group homomorphism. (Verification of smoothness of Ad is
omitted.) This homomorphism is called the adjoint representation of G on its
Lie algebra g.

94
Example 10.4. Let G = GL(n, R), which has gl(n, R) ∼= Te GL(n, R) = Mn (R).
The conjugation operation Cg is linear in the matrix entries (which are serv-
ing as the coordinates). So the adjoint representation Ad : GL(n, R) →
GL(Mn (R)) is given by

Adg ≡ (dCg )e : A 7→ gAg −1 , A ∈ Mn (R). (10.4)

As for matrix Lie groups G ⊂ GL(n, K), their Lie algebras are realized as
(real) linear subspaces inside Te GL(n, K) = Mn (K). Restrict the adjoint rep-
resentation of GL(n), Eq. (10.4), to the subgroup G, and let it act only on the
corresponding tangent subspace g ⊂ Mn (K). This gives the adjoint represen-
tation of G on g.
Exercise 10.3. A basis for so(3) was given in Eq. (10.3). In this basis, the
adjoint representation of g ∈ SO(3) is given by the matrix g itself.

10.4 Exponential map g → G

Proposition 10.4. Left-invariant vector fields on Lie groups are complete.

Proof. Let θ : D → G be the flow of a left-invariant vector field v on the Lie

group G. At the identity element e ∈ G, the integral curve t 7→ γ (e) (t) = θ(t, e)
is defined for some time interval t ∈ (−δ, δ). By left-invariance, for any g ∈ G,
the curve Lg ◦ γ (e) is the integral curve γ (g) starting at g, and it is defined for
t ∈ (−δ, δ). By the uniform time Lemma 9.4, v is complete.
The completeness of left-invariant v on G has a very important conse-
quence. Let γv be the integral curve of v which starts from e at time t = 0;
this curve is defined for all times t ∈ R. Pick some other point γv (s), s ∈ R
lying on this curve. We have two ways of obtaining an integral curve of v
starting form γv (s):

ˆ t 7→ (Lγv (s) ◦ γv )(t) = γv (s) · γv (t);

ˆ t 7→ γv (s + t).

These two curves must coincide, so γv (s) · γv (t) = γ(s + t) for all s, t ∈ R.
Thus the integral curve γv : R → G is necessarily a Lie group homomorphism.

95
Definition 33. A Lie group homomorphism γ : R → G is called a one-
parameter subgroup of G. In particular, for a left-invariant vector field v ∈ g,
the integral curve γv starting at e is called the one-parameter subgroup of G
generated by v.

Theorem 10.5. One-parameter subgroups of a Lie group G are in bijection

with left-invariant vector fields on G, via taking the maximal integral curve
starting at e.

Proof. We have just seen how a left-invariant vector field generates a one-
parameter subgroup of G.
In reverse, let γ : R → G be an arbitrary one-parameter subgroup of G.
Note that γ(0) = e. Let us investigate the tangent vectors to the curve γ,
which are  
′ d
γ (s) = (dγ)s ∈ Tγ(s) G, s ∈ R.
dt t=s
Write Ls for the translation on R by s. The homomorphism property γ(s+t) =
γ(s)γ(t) can be rewritten as γ ◦ Ls = Lγ(s) ◦ γ, and this implies

dγ ◦ dLs = dLγ(s) ◦ dγ, s ∈ R.

Then
    
′ d d
γ (s) = (dγ)s = dγs (dLs )0
dt t=s dt t=0
  
d
= (dLγ(s) )e dγ0 . (10.5)
dt t=0
| {z }
≡γ ′ (0)∈Te G

By Theorem 10.2, the tangent vector γ ′ (0) ∈ Te G corresponds to a left-

invariant vector field v, whose tangent vectors along the curve are, in par-
ticular, given by
Eq. (10.5)
vγ(s) = (dLγ(s) )e (γ ′ (0)) = γ ′ (s), s ∈ R.

This says that γ is recovered as the (maximal) integral curve of the left-
invariant vector field v, starting at e.

96
Definition 34. For a Lie group G with Lie algebra g, the exponential map is
defined as
exp : g → G, v→ 7 γv (1),
where γv : R → G is the 1-parameter subgroup of G generated by v ∈ g.

Both v and sv generate the same subgroup of G, but as parametrized

subgroups, the latter has time parameter t rescaled to st. Then it follows that

exp(sv) = γsv (1) = γv (s) ⇒ exp( · v)′ (0) = γv′ (0) = v,

and the expected symbolic properties of exp hold,

exp((s + t)v) = exp(sv) exp(tv), exp(v)−1 = exp(−v), v ∈ g, s, t ∈ R.

Example 10.5. Recall that the Lie group GL(n, R) has gl(n, R) ∼ = Mn (R) ∼
=
n2 ij
R . The matrix elements X play the role of coordinates, so there are coor-
dinate vector fields ∂X∂ ij . An element A ∈ Mn (R) is regarded as the tangent
vector n
X ∂
A↔ Aij ∈ gl(n, R) = Te (GL(n, R)),
i,j=1
∂X ij e

and it also corresponds to the left-invariant vector field

∂
v (A) : g 7→ (dLg )e (A) = (gA)ij , g ∈ GL(n, R).
∂X ij g
Here we used the fact that Lg (matrix multiplication by g) is linear in the
coordinates X ij , so dLg is also represented as matrix-multiplication-by-g.
Theorem 10.5 says that the one-parameter subgroup γ : R → GL(n, R)
generated by A is the integral curve for v (A) starting at the identity matrix.
Explicitly,

(A) ∂
γ(0) = e = 1n , γ ′ (s) = vγ(s) = (γ(s)A)ij .
∂X ij γ(s)

On the right-side, the integral curve condition is exhibited as a collection of

ODEs for the matrix elements of γ(s), subject to the initial condition on the
left side. These ODEs are concisely written as the matrix equations

γ(0) = 1n , γ ′ (s) = γ(s)A.

97
The solution is provided by the matrix exponential,
∞
X 1 1
γ(s) = (sA)k = 1n + sA + (sA)2 + . . . ,
k=0
k! 2

where the series on the right converges to an element of GL(n, R), and depends
smoothly on s (Exercise). Similarly for the complex case. Setting s = 1, we
get the matrix exponential map exp : gl(n) → GL(n).
Proposition 10.6. The flow θ(v) on G generated by v ∈ g is implemented by
right-multiplication by exp(tv),
θ(v) (t, g) = g · exp(tv) ≡ Rexp(tv) (g). (10.6)
Proof. Since v ∈ g is complete (Prop. 10.4), it generates a global flow θ(v) on
G. By definition, at each g ∈ G, the curve t 7→ θ(v) (t, g) has velocity vector
being vg = dLg (ve ). Since the curve t 7→ (Lg ◦ γv )(t) has
(Lg ◦ γv )′ (0) = (dLg )e ((γv )′ (0)) = dLg (ve ),
we learn that
θ(v) (t, g) = Lg (γv (t)) = Lg (exp(tv)) = g · exp(tv).

Example 10.6. On the Lie group R, a left invariant vector field is ∂x , and it
generates the global flow θ(t, x) = x + t. In this case, Rexp(t∂x ) is translation-
by-t.
Exercise 10.4. Let f : G → H be a Lie group homomorphism, and f∗ : g → h
be the induced Lie group homomorphism. Show that the exponential map is
natural, in the sense that the following diagram commutes:
f∗
g h
exp exp .
f
G H
Proposition 10.7. Let f : G → H be a Lie group homomorphism. Then
adjoint representations of G and H are compatible with f in the sense that
f∗ ◦ Adg = Adf (g) ◦ f∗ , g ∈ G.

98
Proof. Let g ∈ G and v ∈ Te G ∼
= g. From Exercise 10.4, we have

f (g exp(tv) g −1 ) = f (g)f (exp(tv))f (g)−1 = f (g) exp(tf∗ (v))f (g)−1 ,

equivalently,
f ◦ Cg ◦ exp(tv) = Cf (g) ◦ exp(tf∗ (v)).
Differentiate at t = 0 to obtain

f∗ ◦ Adg (v) = Adf (g) ◦ f∗ (v).

Example 10.7. The “group-of-scales” considered by H. Weyl is the matrix Lie

group G = R>0 = ((0, ∞), ×) = GL(1, R)+ . The Lie algebra is g = M1 (R)
with trivial commutator. The exponential map is the usual one, t 7→ et , and
this is a diffeomorphism.
Note that G = (R>0 , ×) is isomorphic to the additive Lie group (R, +)
(by
 the  exponential,
 again). We had exhibited (R, +) as the matrix subgroup
1 x
: x ∈ R in Section 3.1, and its Lie algebra sits inside gl(2, R) =
0 1   
0 s
M2 (R) as : s ∈ R . In this embedding, the exponential map is the
0 0
matrix exponential. We have basically provided two different representations
of the same abstract Lie group (and Lie algebra).
Now consider the matrix Lie group H = U(1) ⊂ GL(1, C), which has
Lie algebra h = iR ⊂ M1 (C) with trivial commutator (Exercise). Clearly
g∼= h as Lie algebras. However, G and H are non-isomorphic, and even non-
homeomorphic. So the exponential map depends on which Lie group one is
referring to.

10.5 Fundamental vector fields on G-manifolds

Let G act smoothly on another manifold P on the right (by diffeomorphisms),
denoted by the smooth map

σ : P × G → P, (p, g) 7→ p · g.

We call P a G-space.

99
 Fix v ∈ g, and restrict attention to the action of its 1-parameter subgroup
γv : R → G,

σ :P ×R→P
(p, t) 7→ p · exp(tv) .
| {z }
γv (t)

As with the P = G case studied earlier, this right-multiplication action de-

scribes a global flow on P . Its infinitesimal generator (see Eq. (9.4)) is some
vector field on P , which we call v ♯ . Explicitly,

vp♯ = (p · exp( · v))′ (0) ∈ Tp P. (10.7)

| {z }
d
=“ dt (p·exp(tv)) ”
t=0

Notation: For a curve γ(·) in P , we shall often denote its velocity vector
γ ′ (0) ∈ Tγ(0) P at t = 0 by the suggestive expression dγ(t)
dt t=0
.
Definition 35. Let P be a G-space. The linear map

g → X(P )
v 7→ v ♯ , (10.8)

defined by Eq. (10.7), is called the infinitesimal generator of the G-action on

P . Vector fields on P induced in this way are called fundamental vector fields.
Example 10.8. Let P = G, with G acting on itself by right multiplication. If
v ∈ g is a left-invariant vector field, then at any g ′ ∈ G,
 
♯ d ′ d
vg′ = (g · exp(tv)) = (Lg′ )∗ (exp(tv)) = (Lg′ )∗ (ve ) = vg′ .
dt t=0 dt t=0

So in this case, v ♯ = v itself.

It is instructive to fix p ∈ P , and define the orbit map based at p,

Op : G → P, g 7→ p · g.

We can then rewrite vp♯ as

vp♯ = (Op ◦ γv )′ (0) = (dOp )e ((γv )′ (0)) ≡ (dOp )(ve ). (10.9)

100
Clearly Op·g = Op ◦ Lg . Then
♯
vp·g = (dOp·g )e (ve ) = (d(Op ◦ Lg ))e (ve ) = (dOp )g ((dLg )e (ve )) = (dOp )g (vg ),

where the last equality follows from left-invariance of v. Thus for all g ∈ G,
♯
dOp “pushes forward” vg to vp·g . One says that v and v ♯ are Op -related. By
the naturality of Lie brackets (see Prop. 8.30 of [11] for details), we also have
[v, w] being Op -related to [v ♯ , w♯ ]. This means, in particular, that

[v ♯ , w♯ ]p = (dOp )e ([v, w]e ) = ([v, w]♯ )p .

We obtain this equality at every other p′ ∈ P , by applying the orbit map Op′ .
So [v, w]♯ = [v ♯ , w♯ ]. We arrive at the following result:

Proposition 10.8. The infinitesimal generator of a group action of G on an

manifold P , Eq. (10.8), is a Lie algebra homomorphism.

Remark. Think of G as being represented on P as a smooth family of diffeo-

morphisms. Infinitesimally, g is represented as “infinitesimal diffeomorphisms”
of P , namely, the fundamental vector fields on P .
Example 10.9. Let S 2 ⊂ R3 be the unit sphere centered at the origin, so its
points are labelled by a unit vector (nx , ny , nz ) thought of as a row vector. The
group SO(3) acts on row vectors by right-multiplication, andthis restricts to
0 −1 0
an action on S 2 . Consider the so(3) element Jz = 1 0 0. A calculation
0 0 0
of the matrix exponential shows that Jz generates the 1-parameter subgroup
 
cos t − sin t 0
exp(tJz ) =  sin t cos t 0 ,
0 0 1

and this subgroup effects

(nx , ny , nz ) · exp(tJz ) = (nx cos t + ny sin t, −nx sin t + ny cos t, nz ).

This is a clockwise rotation by angle t in the horizontal plane. The fundamen-

tal vector field (Jz )♯ generates flow along the constant-latitude orbits, and it
vanishes at the north and south poles.

101
 Write Rg for the diffeomorphism p 7→ p · g. The following Proposition
expresses the sense in which fundamental vector fields are G-equivariant.

Proposition 10.9. Let v ♯ be the fundamental vector field on P induced by

v ∈ g. Then
(Rg )∗ (v ♯ ) = (Adg−1 (v))♯ , ∀g ∈ G.

Proof. We compute, at any p ∈ P ,

((Rg )∗ (v ♯ ))p = (d(Rg ))p·g−1 ((v ♯ )p·g−1 )

d
= (d(Rg ))p·g−1 (p · g −1 exp(tv))
dt t=0
d
= (p · g −1 exp(tv)g)
dt t=0
d
= (p · Cg−1 (exp(tv)))
dt t=0
Ex. 10.4 d
= (p · exp(tAdg−1 (v)))
dt t=0
= (Adg−1 (v))♯p .

10.6 Maurer–Cartan form

A 1-form ω ∈ Ω1 (X) is a smooth collection of linear functionals ωx : Tx X →
R, x ∈ X. More generally, we can consider 1-forms ω ∈ Ω1 (X, V ) which are
valued in a fixed vector space V . This means a smooth assignment of linear
maps, ωx : Tx X → V . So ω takes a vector field to a V -valued function, in
a C ∞ (X)-linear way. The smoothness means that upon picking any basis for
V , ω becomes an n-tuple of ordinary R-valued 1-forms. More invariantly, for
any λ ∈ V ∗ , the R-valued 1-form λ ◦ ω is smooth. We can pull back V -valued
1-forms in the same way as Eq. (8.3), and the pullback still has the properties
found in Exercise 8.5.
We are primarily interested in the case V = g.

Definition 36. The Maurer–Cartan form on a Lie group is a Te G ∼

= g-valued
1-form Θ ∈ Ω1 (G, g), defined by

Θg (vg ) := (Lg−1 )∗ (vg ), vg ∈ Tg G. (10.10)

102
 The idea is use Te G as a “reference tangent space”, and utilize the group
action L to canonically convert tangent vectors at general g to tangent vectors
at e. Doing this to an arbitrary vector field on G will turn it into a g-valued
function — this is what Θ is doing for us.
Exercise 10.5. Show that the Maurer–Cartan form Θ on a Lie group G satisfies
left-invariance,
(Lg )∗ Θ = Θ, ∀g ∈ G,
fixes fundamental vector fields on G,

Θ(v) = v, v ∈ g, (10.11)

and satisfies right-equivariance,

(Rg )∗ Θ = Adg−1 ◦ Θ. (10.12)

For G = GL(n), the Maurer–Cartan form is usually written as

Θ = g −1 dg.

What this means is that at each point g ′ ∈ GL(n, R), we have the matrix
(g ′ )−1 multiplied into the matrix of coordinate 1-forms at g ′ ,
 
(dX 11 )g′ · · · (dX 1n )g′
(g −1 dg)g′ = (g ′ )−1 ·  .. ... ..
, (10.13)
 
. .
n1 ′ nn ′
(dX )g · · · (dX )g

thus g −1 dg is a Mn (R) = gl(n, R)-valued 1-form. One checks that g −1 dg is

indeed the Maurer–Cartan form of GL(n) according to Definition 36.
For GL(n, C), the coordinate X ij is replaced by a complex coordinate Z ij ,
which should be thought of as a pair of real coordinates, Z ij = X ij + iY ij .
The Maurer–Cartan form has the same formula, with dX ij replaced by dZ ij .
For a matrix Lie group G ⊂ GL(n), the Maurer–Cartan form is still given
by the same formula, but the 1-forms dX ij are restricted to act on T G.
Exercise 10.6. Work out the Maurer–Cartan form for the matrix Lie groups
U(1) and SU(2).
Remark. When G acts on a manifold P , the condition Eq. (10.12) still makes
sense, as does Eq. (10.11) with Θ(v ♯ ) in place of Θ(v). As we will see, these
constitute the defining conditions for a connection on a principal G-bundle P .

103
10.7 Derivative of Lie group action
For later use, we record the following calculation:

Proposition 10.10. Let σ : P × G → P denote the action of a Lie group G

on a manifold P . Then

dσ(p,g) (v, ξ) = d(Rg )p (v) + (Θ(ξ))♯p·g , (v, ξ) ∈ Tp P ⊕ Tg G, (10.14)

where Rg : P → P denotes right-multiplication-by-g, and Θ is the Maurer–

Cartan form on G.

Proof. First consider (v, 0) ∈ Tp P ⊕ Tg G. We have v being represented by a

curve γ in P with γ(0) = p and γ ′ (0) = v. So (v, 0) is represented by the curve
t 7→ (γ(t), g) in P × G. By Prop. 4.1,

d(σ(γ(t), g)) d(Rg ◦ γ(t))

dσ(p,g) (v, 0) = = = d(Rg )p (γ ′ (0)) = d(Rg )p (v).
dt t=0 dt t=0

Next, consider (0, ξ) ∈ Tp ⊕ Tg G. We have

d(g exp(tΘ(ξ)))
ξ = (Lg )∗ (Lg−1 )∗ (ξ) = (Lg )∗ (Θ(ξ)) = .
dt t=0

Then
d(p · g exp(tΘ(ξ)))
dσ(p,g) (0, ξ) =
dt t=0
♯
= (Θ(ξ))p·g ,

by definition of the fundamental vector field associated to Θ(ξ) ∈ g.

104
11 Connections on principal bundles and their
curvature
We are finally ready to relate frames attached to different points.

11.1 Canonical vertical bundle

Let π : P → X be a principal G-bundle. Each p ∈ P lies in some fibre Pπ(p)
diffeomorphic to G. So we may consider the vertical tangent space at p,

Vp := Tp Pπ(p) ,

which has dimension dim G. A vertical tangent vector at p is geometrically

represented by some curve through p lying within its fibre. Such a curve
becomes a constant curve upon application of π. This means that

Vp ⊂ ker dπp .

Because π is a submersion, every dπp is surjective, so

dim ker dπp = dim P − dim X = dim G.

Thus Vp = ker dπp .

Globally, the vertical bundle of P sits inside the full tangent bundle of P
as the subset
G
V P := ker(dπ : T P → T X) = Vp ⊂ T P.
p∈P

To see that π : V P → X is indeed a subbundle of T P , we use the fundamental

vector fields v ♯ on P associated to v ∈ g ∼
= Te G, induced by the G-action on
P (Definition 35). For each p ∈ P , the orbit map

Op : G → Pπ(p) , g 7→ p · g,

pushes v to vp♯ (Eq. (10.9)),

g
z}|{
(dOp )e : Te G−→Tp Pπ(p) ≡ Vp
v 7→ vp♯ . (11.1)

105
Since P is a principal bundle, Op is a diffeomorphism, so (dOp )e is actually a
linear isomorphism. Thus Eq. (11.1) gives a canonical way to identify each Vp
with the “reference” g, generalizing the Maurer–Cartan form.
Remark 4. Any basis of g leads, via Eq. (11.1), to (dim G) smooth vector
fields which span Vp at every p ∈ P . This framing property shows that V P
is a subbundle of T P (see Lemma 10.32 of [11]). In fact, there is a canonical
vector bundle isomorphism,

P ×g→VP
(p, v) 7→ vp♯ .

So V P is trivializable, in the same way that a Lie group is (Corollary 10.3).

We also observe that a gauge transformation of P takes vertical tangent
vectors to vertical tangent vectors, and therefore leaves V P invariant.

11.2 Connection as horizontal bundle

Let us proceed to make sense of the “bundle of horizontal directions in P ”.
There is no canonical meaning of “horizontal”, so an extra piece of geometric
data must be supplied.
Definition 37. A connection on a principal G-bundle P is a choice of G-
invariant horizontal bundle/distribution HP on P , i.e., HP is a subbundle of
T X such that

Hp ⊕ Vp = Tp P,
(Rg )∗ Hp = Hpg , ∀p ∈ P, g ∈ G.

The first condition is quite an obvious one. The second condition is imposed
because all points in a fibre are supposed to be treated on an equal footing.
The term principal connection is sometimes used to stress this second property.
Once P is equipped with a connection, then

dπp : Tp P = Hp ⊕ Vp → Tπ(p) X

restricts to an isomorphism Hp → Tπ(p) X. Given v̌x ∈ Tx X and a fibre point

p ∈ π −1 (x), the corresponding vector in Hp under this isomorphism is called
the horizontal lift of v̌x at p.

106
Remark. In general, there are many vectors vp = (vpH , vpV ) ∈ Hp ⊕ Vp = Tp P
which lift v̌x in the sense that dπp (vp ) = v̌x . Nevertheless, any such lift has
the same horizontal component, namely, the horizontal lift of v̌x at p.

Definition 38. With respect to a connection on a principal G-bundle π : P →

X, the horizontal lift of v̌ ∈ X(X) is the unique vector field v H ∈ X(P ), such
that at every p ∈ P , the vector (v H )p ∈ Hp is the horizontal lift of v̌π(p) at p.

Exercise 11.1. Show that for any v̌ ∈ X(X), its horizontal lift is a G-invariant
vector field on P , i.e., (Rg )∗ v H = v H for all g ∈ G.
Exercise 11.2. Let u be a G-invariant vector field on P , and v ♯ be a funda-
mental vector field on P (thus v ♯ is vertical). Show that [u, v ♯ ] = 0.

11.3 Connection as 1-form

While Definition 37 directly captures the geometric essence of the connection
concept, it is not so convenient for calculations, so we seek a more algebraic
formulation.
Given a connection, we are able to uniquely decompose a general tangent
vector vp ∈ Tp P into its vertical and horizontal components,

vp = vpH + vpV , vpH ∈ Hp , vpV ∈ Vp .

We would like to define a map which projects each vp onto its vertical com-
ponent. Although the various Vp are not the “same” vector space, we have
seen how to identify every Vp with g, using Eq. (11.1). So we can define the
g-valued 1-form ω on P ,
(dOp )−1
e
ωp : Tp P → Vp → g
vp 7→ vpV 7→ u, (11.2)

where u is such that u♯p = vpV .

Remark. Near any p ∈ P , we can find some neighbourhood U on which T U is
the direct sum of a vertical bundle and horizontal bundle. Since ω just projects
onto the vertical part (which is trivialized to U × g according to Remark 4),
we see that ω is smooth.
The 1-form ω defined in Eq. (11.2) has the following properties:

107
 ˆ ω takes a (vertical) fundamental vector field u♯ back to u,

ωp (u♯p ) = u, ∀p ∈ P, u ∈ g. (11.3)

ˆ ω is G-equivariant,

Rg∗ ω = Adg−1 ◦ ω, ∀g ∈ G. (11.4)

The equivariance in Eq. (11.4) may be deduced by observing that, by con-

struction, ω annihilates the G-invariant horizontal bundle (Exercise).
Our second definition of a connection is:
Definition 39. A connection on a principal G bundle P is a g-valued 1-form
on P , which satisfies Eq.(11.3)-(11.4).

Relation between the two definitions of connection. From Definition

39 of a connection, we define the horizontal subspaces of Tp P to be ker(ωp ), p ∈
P . Conditions (11.3)-(11.4) ensure that ω is “vertically canonical”, and that
the kernels of ωp assemble into a G-invariant “horizontal distribution”. This
recovers the notion of a connection in the sense of Definition 37.
Remark. Intuitively, p 7→ p · g corresponds to “rotating” a reference frame by
g. Then the G-symmetry of frames with respect to p as reference, needs to be
conjugated by g −1 when p · g is used as reference. Accordingly, the labelling
of frame-wise (i.e. vertical) directions by g needs to be adjusted by applying
Adg−1 .
Example 11.1. Consider P = G as a principal G-bundle over a point. In
Example 10.8, we saw that v ♯ = v, v ∈ g. So we see from Exercise 10.5 that
the Maurer–Cartan form Θ is actually a connection on G.
Example 11.2. The trivial connection on X × G is ωtriv = pr∗2 Θ, where Θ is the
Maurer–Cartan form on G. Here, pr2 denotes projection on to the G factor.
Thus ωtriv restricts to the Maurer–Cartan form on each fibre {x} × G, while
the horizontal subspaces are specified as H(x,g) = Tx X ⊕ 0.
Let P be a trivializable principal G-bundle over X, with trivialization Ψ
written as
∼
=
Ψ:P →X ×G
p 7→ (π(p), ψ(p)).

108
Then
ψ ∗ Θ = (pr2 ◦ Ψ)∗ (Θ) = Ψ∗ (pr∗2 Θ) = Ψ∗ (ωtriv )
is a connection on P (exercise).
Without reference to a trivialization, there is no meaning to “trivial con-
nection” on P . Furthermore, if P is not trivializable, there is no notion of
“trivial connection” in the first place.
Example 11.3. Let P = S 1 × U(1), so g = iR. Let dθ, dφ be the respective
coordinate 1-forms on S 1 and U(1) (Example 8.2). The Maurer–Cartan form
∂
on U(1) is idφ. For i ∈ g, the corresponding fundamental vector field on P is ∂φ
(Exercise). Fix any ik ∈ g, and consider the g-valued 1-form ω (k) := ik dθ+idφ.
Then ω (k) is a connection 1-form on P . The horizontal subspaces defined by
ω (k) are spanned by ∂θ − k∂φ .
Example 11.4. If G is a discrete group, then g = 0, and the only available
connection 1-form is the zero-valued 1-form. As Example 6.1 of S(LR ) shows,
it is possible to have vanishing connection 1-form, with globally non-trivial
“horizontal parallel transport”.
Exercise 11.3. Let ω, ω ′ be two connections on a principal G bundle P . Show
that ω − ω ′ is a horizontal 1-form in the sense that it annihilates all vertical
vectors,
(ω − ω ′ )p (vp ) = 0, ∀vp ∈ Vp , p ∈ P.

11.4 Local description of connection: gauge potentials

It is useful to have a description of a connection as an object living on X,
rather than on P .

Definition 40. Let ω be a connection on a principal G-bundle P (Definition

39). Let s : U → P be a local section/gauge over an open subset U ⊂ X.
Then ω (s) := s∗ ω is a g-valued 1-form on U , called a local gauge potential.

Example 11.5. On the trivial principal bundle S 1 × U(1), consider the connec-
tion 1-form ω = ikdθ + idφ of Example 11.3. Pulling back by the trivializing
section s0 : eiθ 7→ (eiθ , 1), we get the gauge potential s∗0 ω = ikdθ. For n ∈ Z,
apply the large gauge transformation gn ∈ Map(S 1 , U(1)), gn (eiθ ) = einθ ,
which converts s0 to the section sn = s0 · gn . Check that ω is now represented
by the gauge potential s∗n ω = i(k + n)dθ (Exercise).

109
Theorem 11.1. Let ω ∈ Ω1 (P, g) be a connection on a principal G bundle P .
Let sα , sβ be local sections of P defined over open subsets Uα , Uβ respectively,
related by the local gauge transformation gαβ : Uαβ → G (Eq. (6.9)). Write
ω (sα ) , ω (sβ ) for the respective local gauge potentials. On the overlap Uαβ , they
are related by
∗
ω (sβ ) = Adg−1 ◦ ω (sα ) + gαβ (Θ), (11.5)
αβ

where Θ is the Maurer–Cartan form on G.

Proof. We have sβ (x) = σ(sα (x), gαβ (x)), x ∈ Uαβ . Thus sβ = σ ◦ (sα , gαβ ),
and its derivative is
d(sβ )x = dσ(sα (x),gαβ (x)) ◦ (d(sα )x , d(gαβ )x ), x ∈ Uαβ . (11.6)
We computed dσ in Prop. 10.10, with the formula Eq. (10.14) recalled below,
dσ(p,g) (v, ξ) = d(Rg )p (v) + (Θ(ξ))♯p·g .
Substitute this into Eq. (11.6) and evaluate on some η ∈ Tx X,
d(sβ )x (η) = d(Rgαβ (x) )sα (x) (d(sα )x (η)) + (Θ(d(gαβ )x (η)))♯sβ (x)
∗
= d(Rgαβ (x) )sα (x) (d(sα )x (η)) + ((gαβ Θ)x (η))♯sβ (x) . (11.7)

For convenience, we drop the reference to x. Apply ω to the first term of Eq.
(11.7),
ω d(Rgαβ )sα (d(sα )(η)) = (Rg∗αβ ω)(d(sα )(η))




= Adg−1 ω(d(sα )(η)) (Eq. (11.4))
αβ

= Adg−1 (s∗α ω(η))

αβ

= Adg−1 ◦ ω (sα ) (η).

αβ

For the second term, we get

∗ Eq. (11.3) ∗
ω((gαβ Θ)(η))♯sβ = (gαβ Θ)(η)
In total,
∗
ω(d(sβ )(η)) = (Adg−1 ◦ ω (sα ) + gαβ Θ)(η),
| {z } αβ
(sβ )
=s∗β ω(η)≡ω (η)

which is the desired transformation law, Eq. (11.5).

110
Remark. When G is abelian, the transformation law, Eq. (11.5), simplifies to
∗
ω (sβ ) = ω (sα ) + gαβ (Θ).
We may think of Eq. (11.5) as the “action” of local gauge transformations
gαβ on local gauge potentials, effecting ω (sα ) ⇝ ω (sβ ) .
Exercise 11.4. Let sγ : Uγ → P be a third local section. So we have the local
gauge transformation gβγ : Uβγ → G relating sβ to sγ , as well as gαγ : Uαγ → G
relating sα to sγ . Show that on the triple overlap Uαβγ , applying gαγ to a local
gauge potential ω (sα ) gives the same result as applying gαβ followed by gβγ .
Example 11.6. In Example 11.5, the section sn is obtained from s0 by applying
g0n : S 1 → G, eiθ 7→ einθ . The transformation law, Eq. (11.5), correctly gives
∗
ω (sn ) = ω (s0 ) + g0n (idφ) = ikdθ + indθ.

Local versus global. We emphasize that a gauge potential is a local and

gauge-dependent description of a connection. Since P is generally not trivializ-
able, we need a collection of locally defined gauge potentials ω (sα ) with respect
to some trivializing cover, in order completely capture the connection data.
Consider a trivializing cover of P specified by local sections sα , α ∈ I.
Suppose we have a collection of g-valued local 1-forms Aα ∈ Ω1 (Uα , g), such
that the transformation law, Eq. (11.5) holds on every overlap Uαβ ,
∗
Aβ = Adg−1 ◦ Aα + gαβ (Θ),
αβ

where gαβ are the local gauge transformations converting sα to sβ . Then it may
be shown that there exists a unique connection ω on P such that Aα = s∗α ω.
This “algebraic” method of defining connections via local gauge potentials and
checking (or declaring) mutual compatibility, is usually utilized in physics.
However, the global geometrical meaning of the connection becomes rather
obscured.
Finally, we remark on the terminology “gauge field”. Often, a local gauge
potential is called a “gauge field on X”. A somewhat better choice is to call
a connection on P a gauge field. Actually, it is possible to show that the
difference between two connections is a section of the so-called adjoint vector
bundle over X associated to P . So the space of connections could be identified
as a space of fields over X, but only affinely (there is no canonical “trivial” or
“zero” connection).

111
11.5 Curvature of connections
The exterior derivative of a k-form generalizes in the obvious way to V -valued
k-forms, giving rise to V -valued (k + 1)-forms. Namely, we can expand a V -
valued 1-form into R-valued k-forms with respect to any basis for V . Then
take the usual exterior derivative of each component, and sum the result up.
Definition 41. Let ω ∈ Ω1 (P, g) be a connection 1-form on a principal G-
bundle P . Its curvature is the g-valued 2-form
Ω(u, v) = dω(uH , v H ), u, v ∈ X(P ). (11.8)
Recall that (·)H denotes the horizontal part of a vector field on P , as
determined by the connection ω. The expression dω(u, v) would generally
involve three terms, as in Eq. (8.8). But on the right side of Eq. (11.8), we
only put in horizontal vector fields. Because ω annihilates horizontal vector
fields, we are left with only a single term,
Ω(u, v) = dω(uH , v H ) = −ω([uH , v H ]).

Intuitively: Starting from p, we imagine moving a small distance horizon-

tally along uH , then horizontally along v H . Now do this in the opposite order.
If the horizontal distribution defined by ω is “curved”, we might not end up
at the same point! Infinitesimally, the endpoint mismatch is an infinitesimal
vertical offset along the final fibre, i.e., a vertical tangent vector (identified
with a Lie algebra element as usual); the curvature form Ω measures such
mismatches.
Definition 42. Let ω be a connection on a principal G-bundle P . The hori-
zontal part η H of a k-form η ∈ Ωk (P, V ) is defined to be
η H (v1 , . . . , vk ) := η(v1H , . . . , vkH ), vi ∈ X(P ),
and the exterior covariant derivative of a k-form η with respect to ω is defined
to be
Dω η := (dη)H .
In particular, the curvature of ω, Eq. (11.8), can be written as
Ω = Dω (ω),
i.e., the exterior covariant derivative of ω with respect to itself.
Recall that a connection 1-form is G-equivariant (Eq. (11.4)). Similarly:

112
Proposition 11.2. The curvature Ω of a connection ω is G-equivariant,

Rg∗ Ω = Adg−1 ◦ Ω, ∀ g ∈ G.

Proof. Because (Rg )∗ respects the horizontal-vertical splitting of tangent spaces,

we have
(Rg )∗ v H = ((Rg )∗ v)H , v ∈ X(P ).
So for any u, v ∈ X(P ), we have

(Rg∗ Ω)(u, v) = Ω((Rg )∗ u, (Rg )∗ v)

= dω(((Rg )∗ u)H , ((Rg )∗ v)H )
= dω((Rg )∗ uH , (Rg )∗ v H )
= (Rg∗ (dω))(uH , v H )
= d(Rg∗ ω)(uH , v H )
= d(Adg−1 ◦ ω)(uH , v H )
= Adg−1 ◦ (dω)(uH , v H ) = Adg−1 ◦ Ω(u, v).

Definition 43. The wedge product of η, ω ∈ Ω1 (X, g) is the g-valued 2-form

defined by

[η, ω](u, v)(x) = [η(u)(x), ω(v)(x)] − [η(v)(x), ω(u)(x)], u, v ∈ X(X),

where [·, ·] on the right side denotes the Lie bracket in g.

This is like an ordinary wedge product, but instead of pointwise multiplying
η(u) and ω(v), we take the pointwise Lie bracket. In the same way, we can
take wedge products of g-valued k-forms and l-forms.
Theorem 11.3 (Cartan structure equation). The curvature of a connection
1-form ω ∈ Ω1 (P, g) can be expressed as
1
Ω = dω + [ω, ω]. (11.9)
2
Proof. Exercise. It is helpful to split the calculation of Ω(u, v) into three cases:
u, v both horizontal, u, v both vertical, u vertical and v horizontal.

113
Remark. In the particular case where P = G and ω = Θ is the Maurer–Cartan
form, there are no horizontal directions at all. So dΘ + 12 [Θ, Θ] = 0.
Exercise 11.5. Using the Cartan structure equation and Eq. (8.10), verify the
Bianchi indentity: Dω Ω ≡ (dΩ)H = 0 for the curvature Ω of any connection
ω.

11.5.1 Local field strength

As with the local gauge potentials, we define:

Definition 44. Let Ω be the curvature of a connection ω on a principal G-

bundle P . Let s : U → P be a local section of P . Then s∗ Ω ∈ Ω2 (U, g) is
called the local field strength of ω.

Because pullback is compatible with wedge product and the exterior deriva-
tive, it is straightforward to verify the local form of the structure equations,
1
s∗ Ω = dω (s) + [ω (s) , ω (s) ]. (11.10)
2
The local field strengths are gauge-dependent, so let us work out their
transformation law under local gauge transformations.

Theorem 11.4. Let Ω be the curvature of a connection ω on a principal G-

bundle. Let sα : Uα → P and sβ : Uβ → P be two local gauges/sections, related
on the overlap Uαβ by the local gauge transform gαβ : Uαβ → G. On Uαβ , the
local field strengths are related by

s∗β Ω = Adg−1 ◦ s∗α Ω. (11.11)

αβ

Proof. We need to compute

s∗β Ω(u, v) = Ω(dsβ (u), dsβ (v)), u, v ∈ X(U ).

While proving the transformation law for the local gauge potentials, Theorem
11.1, we derived Eq. (11.7),
∗
d(sβ )x (η) = d(Rgαβ (x) )sα (x) ◦ d(sα )x (η) + ((gαβ Θ)x (η))♯sβ (x) , η ∈ Tx X.

114
The second term on the right is vertical, so it does not contribute to Ω(·, ·).
Thus for any ux , vx ∈ Tx X,

(s∗β Ω)x (ux , vx ) = Ωsβ (x) (d(Rgαβ (x) )sα (x) ◦ d(sα )x (ux ), d(Rgαβ (x) )sα (x) ◦ d(sα (x))(vx ))
= (Rg∗αβ (x) Ω)sα (x) (d(sα )x (ux ), d(sα )x (vx ))
= (Adg−1 (x) ◦ Ω)sα (x) (d(sα )x (ux ), d(sα )x (vx ))
αβ

= Adg−1 (x) ◦ (s∗α Ω)x (ux , vx ),

αβ

as required for Eq. (11.11).

11.6 Physics notation

The physicists’ notation for the local gauge potentials and local field strengths
are
A = s∗ ω, F = s∗ Ω.
where reference to a local section s (or local trivialization) is usually left im-
plicit. It is seldom mentioned that these potentials arise from a globally defined
object (the connection on P ). So one only has the local structure equation,
Eq. (11.10),
1
F = dA + [A, A], (11.12)
2
which is usually taken as the definition of the (local) curvature 2-form.
Furthermore, one often works in a coordinate chart (U, φ), and expands A
in terms of coordinate 1-forms. So A (more precisely, (φ−1 )∗ A) is written as
n
X
A= Ai dxi ,
i

wherePeach component Ai is a smooth g-valued function on φ(U ). Similarly,

F = ni,j=1 Fij dxi ∧ dxj , where (Exercise)

Fij = ∂i Aj − ∂j Ai + [Ai , Aj ]

is the coordinate-dependent expression of Eq. (11.12).

Next, G is usually identified as a matrix Lie group, G ⊂ GL(n, R), so Θ =
−1
g dg (see Eq. (10.13)). Then the second term in the gauge transformation

115
law, Eq. (11.5), is
 11 1n

d(gαβ )x · · · d(gαβ )x
∗
(gαβ −1
(Θ))x = gαβ (x) ·  ... ... ..  ,

. 
n1 nn
d(gαβ )x · · · d(gαβ )x
ij
with gαβ = X ij ◦ gαβ : U → R being the ij-th matrix entry of the transition
function gαβ . It is common practice to simply write g = gαβ : Uαβ → G for the
local gauge transformation, and condense the above expression into “g −1 dg”.
The gauge transformation rules are then

A ⇝ A′ = g −1 Ag + g −1 dg
F ⇝ F ′ = g −1 Fg. (11.13)

Remark. Sometimes, g is replaced by g −1 to get a left action of local gauge

transformations, and there is a factor of i relating physicists’ “matrix Lie
algebras” and the mathematicians’ Lie algebras. The matrix identity dg =
−g(dg −1 )g may be used to write −dg −1 g instead of g −1 dg.

11.7 Globally gauge-transformed connections and cur-

vatures
We saw how the local gauge potentials and field strengths transform under a
change of local gauge (i.e. choosing a different local section), Eq. (11.5)-(11.11).
Before we discuss the global picture, we make a general observation.

Pullback connection. Let η be a G-equivariant g-valued k-form on a prin-

cipal G-bundle P2 ,
Rg∗ η = Adg−1 ◦ η, ∀g ∈ G.
If F : P1 → P2 is a morphism of principal G-bundles, then F ∗ η is a G-
equivariant k-form on P1 , because of

Rg∗ (F ∗ η) = (F ◦Rg )∗ η = (Rg ◦F )∗ η = F ∗ (Rg∗ η) = F ∗ (Adg−1 ◦η) = Adg−1 ◦(F ∗ η).

Thus, for a connection ω on P2 , the pullback F ∗ ω remains G-equivariant.

116
 Furthermore, if v ♯ is a fundamental vector field on P1 , then
 
♯ d(p · exp(tv))
(dF )p (vp ) = (dF )p
dt t=0

d(F (Rexp(tv) (p)))

=
dt t=0
d(Rexp(tv) (F (p)))
= = vF♯ (p) ,
dt t=0

so that

(F ∗ ω)p (vp♯ ) = ωF (p) ((dF )p (vp♯ )) = ωF (p) (vF♯ (p) ) = v, ∀ p ∈ P1 .

Thus F ∗ ω is a connection 1-form on P1 .

In particular, gauge transformations of P act on connections on P by pull-
back.

Theorem 11.5. Let ω ∈ Ω1 (P, g) be a connection on a principal G-bundle P ,

and let F ∈ G(P ) be a gauge transformation, implemented by right multiplica-
tion by an equivariant map σF ∈ Map(P, G)G as in Eq. (6.11). Then

F ∗ ω = Adσ−1 ◦ ω + σF∗ Θ. (11.14)

F

Proof. The argument is similar to that in Theorem 11.1, and is omitted.

Remark. A significant portion of gauge theory is concerned with understanding
the “moduli space” of connections on P , after quotienting out by the above
action of gauge transformations.
Exercise 11.6. Recall the connection 1-form ω = i(kdθ + dφ) on P = S 1 ×
U(1). For n ∈ Z, the map Fn : (eiθ , eiφ ) 7→ (eiθ , eiφ+nθ ) is a (large) gauge
transformation of P . Work out what Fn∗ ω is.
For the curvature 2-form, the global version of Theorem 11.4 is

Theorem 11.6. Let F ∈ G(P ) be a gauge transformation of a principal G-

bundle P . Let ω be a connection 1-form on P , and let F ∗ ω be its gauge
∗
transform (Eq. (11.14)). Write Ωω and ΩF ω for the respective curvature 2-
forms. Then
∗
ΩF ω = F ∗ Ωω = Adσ−1 ◦ Ωω .
F

117
Proof. For the first equality, we use the structure equations, and commutativ-
ity of pullback with wedge product,
 
∗ ω ∗ ω ∗ 1
F Ω ≡ F (D ω) = F dω + [ω, ω]
2
1
= d(F ∗ ω) + [F ∗ ω, F ∗ ω]
2
F ∗ω ∗ ∗
= D (F ω) ≡ ΩF ω .

For the second equality, let u, v ∈ X(P ). In the expression

F ∗ Ωω (u, v) = Ωω (F∗ u, F∗ v) , (11.15)

we may discard any vertical vectors occurring in the arguments on the right.
To understand F∗ = dF , we note that F : P → P may be written as the
composition
σ
p 7→ (p, σF−1 (p)) 7→ p · σF (p),
where σ : P × G → P is the group action map. The derivative of σ was found
in Eq. (10.14), and discarding vertical terms, we simply have

dσ(p,g) (vp , ξg ) “ = ” (dRg )p (vp ), vp ∈ Tp P, ξg ∈ Tg G.

Substitute this into Eq. (11.15),

(F ∗ Ωω )p (up , vp ) = Ωωp·σF (p) (dRσF (p) )p (up ), (dRσF (p )p (vp )

= Adσ−1 (p) ◦ Ωωp (up , vp )

F

where the last line is due to Ω being G-equivariant. Thus

F ∗ Ωω = Adσ−1 ◦ Ωω .
F

11.7.1 Abelian G
Notice that when G is abelian, the curvature is simply Ω = dω, which is linear
in ω. In this case, the curvature is also invariant under gauge transformations.
In particular, the local field strengths s∗ ω are independent of the choice of
local gauge/section s. This can be explicitly seen from Eq. (11.11). Thus we

118
may pick any collection of local sections, sufficient to cover all of X, and obtain
a globally well-defined field strength F ∈ Ω2 (X, g).
Furthermore, the field strength is a closed 2-form, dF = 0, because we may
compute, using any local gauge, that
1
dF = d(s∗ Ω) = d(d(ω (s) )) + d [ω (s) , ω (s) ] = d2 (ω (s) ) = 0.
2 | {z }
abelian ⇒ 0

However, it is important to remember that “F = d(ω (s) )”, or “F = dA” in

physics notation, holds only over the subset that s is defined on. One says
that F is locally exact, but not globally exact. The failure of a closed 2-form
on X (such as F) to be globally exact, is measured by the second cohomology
group of X. The latter is a topological invariant of X.

119
12 Fibre bundles and structure groups
Some motivation from physics. In classical geometry, principal G-bundles
usually arise as a bundle of frames, for some concrete vector bundle of interest,
such as a tangent bundle. In (quantum) physics, the principal G-bundle with
connection encodes an abstract notion of “local gauge symmetry”, or “inter-
frame relationships”, both at a single point of X, and between neighbouring
points. The principal bundle can be realised as the frame bundle of some as-
sociated vector bundle, and here there is considerable freedom to (unitarily)
represent G as transformations of some reference vector space. Different rep-
resentations are labelled by “charges”. The associated vector bundle is also
viewed abstractly, as a fibre bundle with structure group G, and it is determined
up to isomorphism.
Sections of the above-mentioned vector bundle are typically thought of as
matter fields (electron, proton, etc.). The principal bundle connection induces
a covariant derivative on the associated vector bundle, thereby allowing for
a gauge-independent notion of differentiation for the matter fields. The con-
nection itself could be allowed to vary — it then becomes a “dynamical gauge
field interacting with matter fields”. So we often need to consider the entire
space of connections rather than a fixed prescribed one. Although the space
of connections is (affinely) an infinite-dimensional linear space, in the end,
everything is done modulo gauge transformations. This is because we cannot
measure the section/connection as-it-is, but only the gauge-invariant aspects.
The above discussion can be taken as the “gauge principle”; compare the
situation in general quantum mechanics. It strongly constrains the variety of
admissible physical theories of matter and interactions.

12.1 Vector=[frame, components]

Let E be a vector space of dimension n. With respect to a frame/basis e =
{e1 , . . . , en }, a vector v ∈ E may be expanded as
n
X
v= ea v a ,
a=1

120
where v = (v 1 , . . . , v n ) ∈ Kn is the n-tuple of components representing v. We
could choose to expand v with respect to another frame e′ = e · g, g ∈ GL(n),
n
X n
X n
X n
X
v= e′b (v ′ )b = ea g ab (v ′ )b = ea g ab (v ′ )b .
b=1 a,b=1 a=1
|b=1 {z }
va

The new n-tuple of components v′ = (v ′ 1 , . . . , v ′ n ) is related to the old one by

v′ = g −1 · v with respect to e′ = e · g. (12.1)

Neither the choice of frame e nor the components v is intrinsic to v ∈ Ex .

Only the combination (e, v) is intrinsic,
n
X n
X
v= ea v =a
e′a (v ′ )a .
a=1 a=1

So, according to Eq. (12.1), we should make the identifications

(e, v) ∼ (e · g, g −1 · v), ∀g ∈ G. (12.2)

In words: A vector v ∈ E may be be described as “n-tuple with respect to

a frame” in many different ways. Without a preferred frame, v is simply the
equivalence class of any such description.
Formally, this operation of taking equivalence classes is denoted

E = Fr(E) ×ρ Kn := (Fr(E) × Kn )/∼Eq. (12.2) ,

where ρ : G → GL(n) makes explicit that G is to act on Kn in the defining

representation. It should be clear that the equivalence classes inherit a well-
defined vector space structure from Kn .

12.2 Associated fibre bundles

Let π : P → X be a principal G-bundle over X. Let ρ be a left G-action on a
manifold F (by diffeomorphisms). This action will be denoted

g · ξ := ρ(g)(ξ), g ∈ G, ξ ∈ F,

with ρ suppressed unless emphasis is necessary.

121
 Prompted by Eq. (12.2), we consider the set P × F , equipped with the
right G-action,
(p, ξ) · g := (p · g, g −1 · ξ).
Then pass to the set of equivalence classes modulo this G-action,

P ×ρ F := {(p, ξ) ∈ P × F : (p · g, g −1 · ξ)}. (12.3)

Notice that the projection map

πρ : P × ρ F → X
[p, ξ] 7→ π(p)

is well-defined, since π(p · g) = π(p) for all g ∈ G. Give P ×ρ F the quotient

topology, then πρ : P ×ρ F → X is continuous (exercise).
Exercise 12.1. For x ∈ X, show that

πρ−1 (x) = {[p, ξ] : ξ ∈ F }

where p can be chosen to be any point in Px = π −1 (x).

Thus every fibre πρ−1 (x) is in bijection with the typical fibre F . But such
an identification requires choosing, non-canonically, a reference frame p ∈ Px .
Next, we sketch how these fibres assemble into a smooth “fibre bundle” (See
Definition 46 later).
Let x ∈ X, and Φ : π −1 (U ) → U × G be a local trivialization of P around
x. Equivalently, Φ−1 corresponds to a local gauge

sΦ : U → P, x 7→ Φ−1 (x, 1).

We pair up this local gauge with F -valued components to get a continuous

“trivialization” of πρ−1 (U ),

(Φ(ρ) )−1 : U × F → πρ−1 (U )

(x, ξ) 7→ [sΦ (x), ξ]

Since x was arbitrary, we see that P ×ρ F is built up from “locally trivial

pieces” πρ−1 (U ) ∼
= U × F . As in the case of the tangent bundle construction,
P ×ρ F is a topological manifold, and establish its smooth structure as follows.

122
 From a trivializing cover {Uα , Φα }α∈I of P , we obtain a corresponding
(ρ)
continuous trivializing cover {Uα , Φα }α∈I of P ×ρ F . Each Uα ×F is a smooth
(ρ)
manifold, so (Φα )−1 (Uα × F ) inherits a smooth structure. Over Uαβ , the
(ρ) (ρ)
coordinate charts coming from Φα and from Φβ are compatible, due to the
following.
(ρ) (ρ)
Exercise 12.2. Let (Uα , Φα ), (Uβ , Φβ ) be two (continuous) local trivializa-
tions for P ×ρ F . Show that on the overlap Uαβ × F , we have
(ρ) −1
Φβ ◦ (Φ(ρ)
α ) : Uαβ × F → Uαβ × F
(x, ξ) 7→ (x, ρ(gβα (x))(ξ)), (12.4)

where gβα : Uαβ → G are the smooth transition functions for P .

Definition 45. πρ : P ×ρ F → X, with the above smooth structure, is called

the fibre bundle associated with π : P → X, via the action ρ on F , or simply
an associated fibre bundle.

Example 12.1. The frame bundle Fr(E) of a vector bundle E is a principal

GL(n)-bundle. Let ρ be the defining action of G as matrix multiplication on
Kn . Then Fr(E) ×ρ Kn is a vector bundle. It is isomorphic to E as follows:
if ex ∈ Fr(Ex ) is a frame at x, then [ex , ξ] ∈ Fr(E) ×ρ Kn is identified with
ex (ξ) ∈ Ex .
In the associated fibre bundle construction, observe that the transition
functions are G-valued, so not all diffeomorphisms of F will be used. So
P ×ρ F is a “bundle of F ” with extra structure:

Definition 46. A fibre bundle with typical fibre F on which the structure Lie
group G acts (on the left by diffeomorphisms), is a smooth surjective map
π : E → X, such that every x ∈ X lies in an open neighbourhood U with
E|U := π −1 (U ) locally trivializable: there is a diffeomorphism

Φ : π −1 (U ) → U × F such that π(Φ−1 (x, ξ)) = x, x ∈ U, ξ ∈ F.

Furthermore, for each pair (Uα , Φα ), (Uβ , Φβ ) of local trivializations, we have

Φβ ◦ Φ−1
α (x, ξ) = (x, gβα (x) · ξ), x ∈ Uαβ , ξ ∈ F,

with smooth gβα : Uαβ → G, called the G-valued transition functions.

123
Example 12.2. As we saw in Eq. (6.6), a principal G-bundle P is a fibre bundle
with typical fibre G, and structure group G acting on the typical fibre G by
left multiplication (“frame relabelling”). Beware that this left multiplication
only makes sense in a local trivialization, and there is no global left G-action
on P .
Example 12.3. A rank-n vector bundle is a fibre bundle, with structure group
GL(n) acting on the typical fibre Kn .
Remark. As usual, a fibre bundle is allowed the maximal compatible collec-
tion of local trivializations, in the sense of all G-valued transition functions
remaining smooth. For example, the associated fibre bundle does not inherit
preferred local trivializations from those used in its construction.

12.3 Operations on vector bundles

On vector spaces E, F , we have algebraic operations,
ˆ Duals E ∗ ;

ˆ Direct sums;

ˆ (Symmetrized/antisymmetrized) tensor products;

ˆ Hom(E, F ) ∼
= F ⊗ E ∗;
ˆ Complexification/realification;

ˆ Complex conjugate.
These algebraic operations generalize to vector bundles E, F over X.
For example, for the direct sum E ⊕F of a rank-m and a rank n-vector bun-
dle, we would take each (E ⊕ F )x to be Ex ⊕ Fx . Then use local trivializations
E|U ∼= U × Km and F |U ∼ = U × Kn to construct (E ⊕ F )|U ∼
= U × (Km ⊕ Km ),
and topologize E ⊕ F as we did for the tangent bundle in Section 5.1. The
resulting transition functions will be the direct product of those for E and
those for F , so they are smooth. More abstractly, we can also construct E ⊕ F
from the direct product GL(m) × GL(n)-valued transition functions, by the
gluing construction of Section 6.4.1.
Similarly, the cotangent bundle T ∗ X can be constructed by taking the
−1 t
transitions functions (gαβ ) , where gαβ are the ones for T X. We can con-
∗
struct T X from the tangent frame bundle by taking GL(n) to act on Rn in

124
the contragedient representation (g −1 )t . Similarly for tensor product bundles.
In particular, any manifold X automatically comes with a bundle Λk (X) of
differential k-forms, and Ωk (X) = Γ(Λk (X)).
A common procedure is to tensor Λk (X) with an auxiliary vector bundle
E, where E has structure group G′ ⊂ GL(dim E) independent from that of
T X. Elements of
Γ(Λk (X) ⊗ E) = Ωk (X, E)
are called 1-forms twisted by E.

12.3.1 Gauge transformation of vector bundles

Of particular importance is the so-called endomorphism bundle End(E) :=
E ⊗ E ∗ of E. The terminology arises because sections T ∈ Γ(End(E)) act on
sections v ∈ Γ(E) in the obvious way,

T · v(x) = T (x)(v(x)),

and T · (f v) = f T (v), for all f ∈ C ∞ (X). Elements of Γ(End(E)) are called

bundle endomorphisms. They can be added linearly, and be composed. There
is an identity bundle endomorphism, and the invertible bundle endomorphisms
form a group, denoted GL(E).
Now, if E has some extra structure (e.g. metric, orientation), and is re-
garded as a fibre bundle with reduced structure group G ⊂ GL(n), then we
can restrict to those sections T ∈ Γ(End(E)) for which each T (x) acts on Ex as
an element of G. Such a bundle endomorphism T is also called a gauge trans-
formation of E, where E regarded as a vector bundle with structure group G.
These restricted bundle endomorphisms form a subgroup, denoted G(E).
Now let F ∈ G(P ) be a gauge transformation of P , represented as right-
multiplication by the equivariant map σF : P → G. On any associated vector
bundle, there is a corresponding gauge transformation. For [p, ξ] ∈ E =
P ×ρ F , define

F · [p, ξ] := [p · σF (p), ξ] = [p, ρ(σF (p)) · ξ], (12.5)

One checks that (exercise)

ˆ Eq. (12.5) defines a bundle endomorphism of TF of E as a vector bundle

with structure group ρ(G).

125
 ˆ If ρ is a faithful representation (i.e. injective), then F 7→ TF is a group
isomorphism G(P ) → G(E).

Question: When do we need to reduce the structure group of a vector bundle

from GL(n) to some subgroup?

12.4 Vector bundle metrics

If E is a real vector bundle, then a Euclidean bundle metric is a section

⟨·, ·⟩E ∈ Γ(E ∗ ⊗ E ∗ )

which restricts at each x ∈ X to an inner product ⟨·, ·⟩x on Ex . Then E is

called a Euclidean vector bundle. The inner product of two sections is then a
real-valued function,
⟨ψ, ψ̃⟩E ∈ C ∞ (X).
For example, a Riemannian metric on X is a Euclidean bundle metric on T X.
If E is a complex vector bundle, then a Hermitian bundle metric is a section
∗
⟨·, ·⟩ ∈ Γ(E ⊗ E ∗ )

which restricts at each x ∈ X to a Hermitian inner product on Ex . Then E is

called a Hermitian vector bundle.

12.5 Restricted frame bundles

Let π : E → X be a rank-n vector bundle. The full frame bundle Fr(E) has
the structure of a principal GL(n)-bundle.

Metrics. Suppose E is a Euclidean vector bundle. Then we may restrict at-

tention to orthonormal frames. Pointwise, orthonormal frames form an O(n)-
torsor, instead of a GL(n)-torsor. If we only allow local trivializations for
E such that Ex ∼ = Kn preserve inner products, then the transition functions
will be O(n)-valued. So the structure group of E gets reduced from GL(n)
to O(n). The corresponding orthonormal frame bundle FrO (E) is a principal
O(n)-bundle inside the full frame bundle.
Similarly, if E is a Hermitian vector bundle, the orthonormal (or unitary)
frame bundle is a principal U(n)-bundle, denoted FrU (E).

126
Orientation. Two bases for a real vector space E are said to be equivalent if
the change-of-basis matrix has positive/negative determinant. An orientation
of E is a choice of equivalence classes of bases. The set of orientations is a
Z2 -torsor (there is no canonical “trivial” orientation).
For E a real vector bundle, we can assign a pointwise orientation to ev-
ery Ex . A local frame {e1 , . . . , en } over U ⊂ X is positively oriented if
{e1 (x), . . . , en (x)} has the same orientation as the given pointwise orienta-
tion for all x ∈ U . The pointwise orientation is globally smooth if every point
of X lies in the domain of an oriented local frame; in this case, E is said to
be an orientable vector bundle, and a choice of smooth pointwise orientation
is simply called an orientation on E.
Once E has been given an orientation, we can decide to use only the pos-
itively oriented local frames as the local trivializations. Then the transition
functions will be GL(n, R)+ -valued. We obtain the oriented frame bundle
Fr+ (E) as a principal GL(n, R)+ -bundle, with GL(n, R)+ -valued transition
functions.
If E has a bundle metric, there is a further reduction to (oriented) orthonor-
mal frames, and we have the (oriented) orthonormal frame bundle Fr(S)O (E)
as a principal (S)O(n)-bundle.
Remark. If E = T X is the tangent bundle of a (oriented, Riemannian) mani-
fold X, then the above frame bundles are usually referred to the various frame
bundles of X, with the vector bundle T X being implicit.

Spin? As we will learn later, there can be a further “reduction” from SO(n)
to a so-called Spin(n) group. This does not naturally take place on a vec-
tor/tensor bundle over X. Instead, the natural starting point is the principal
“spin frame bundle”, to which the vector bundle of spinors is associated.

12.6 Reduction of structure group

Definition 47. Let ϕ : H → G be a Lie group homomorphism, and π ′ : P ′ →
X be a principal G-bundle. A map F : P → P ′ , where π : P → X is a
principal H-bundle, is called a ϕ-reduction of P ′ if

ˆ π ′ ◦ F = π,

ˆ F (p · h) = F (p) · ϕ(h), h ∈ H, p ∈ P .

127
Remark. The map F restricts to a map Px → Px′ of fibres. Picking basepoints
p ∈ Px and F (p) ∈ Px′ gives identifications Px ∼
= H and Px′ ∼
= G. Then F |Px is
identified with the homomorphism ϕ : H → G.
An idG -reduction is just a principal G-bundle isomorphism. When ϕ is the
embedding of a Lie subgroup H ⊂ G, we simply call P a H-reduction of P ′ .
The image F (P ) is then a principal H-subbundle inside P ′ .
Example 12.4. When an orientable real vector bundle E is given an orientation,
the oriented frame bundle is a principal GL(n, R)+ subbundle of the full frame
bundle. We say that the former is a GL(n, R)+ -reduction of the frame bundle.
Similarly, if a bundle metric is given, the oriented orthonormal frame bundle
is a further SO(n)-reduction of the frame bundle.
Let P be a principal G-bundle, and consider an associated vector bundle
E = P ×ρ V . Suppose V is equipped with a G-invariant inner product,

⟨ξ, ζ⟩V = ⟨ρ(g)ξ, ρ(g)ζ⟩V , ξ, ζ ∈ V, g ∈ G.

In other words, the representation ρ of G on the typical fibre V is uni-

tary/orthogonal. Then the vector bundle E = P ×ρ V acquires the bundle
metric
⟨[p, ξ], [p, ζ]⟩E := ⟨ξ, ζ⟩V . (12.6)
The structure group of E may then be reduced from GL(V ) to ρ(G) ⊂ U(V )
(or ρ(G) ⊂ O(V )).

12.7 Scalar fields

What if the structure group of E is reduced to the trivial group? This
means that the allowed trivializations have trivial transition functions. A little
thought reveals that there is just a single global trivialization E ∼ = X ×V,
with all others simply being the restriction to subsets U ⊂ X. Then sections
of E are just V -valued functions that we were all familiar with before learning
about bundles. So why describe functions in such a bundle-theoretic way?
In general, we really want to distinguish functions/scalar fields (as above)
from, e.g., vector fields, k-form fields, spin- 12 , 1, 32 , . . . fields, etc. The point is
that all of these fields are sections of vector bundles associated to a common
frame bundle intrinsic to the (typically Riemannian) geometry of X. The
distinguishing factor is the representation of the structure group G on the
typical vector space.

128
 In Riemannian geometry, we would take FrSO (X), and SO(n) would be
represented in the defining way, or in some tensor representation. In quantum
theory, we have FrSpin (X) (defined in Section 16.1), and for dim X = 3, we have
the unitary representations of Spin(3) labelled by half-integers. In relativistic
(quantum) theory, the Lorentz-orthonormal frame bundle FrSO(1,3) (X), or the
Spin(1, 3) version, is used. The various irreducible representations of Spin(1, 3)
give a notion of elementary particle fields over the semi-Riemannian spacetime
X.
In any of the above situations, we may take G to be represented trivially
on V . Then the associated vector bundle is a trivial vector bundle (not just
trivializable). Its sections are scalar fields, which are now just functions since
only one global trivialization is usable. In physics, one might refer to scalar
fields as “spin-0” fields, and talk about “spinless particle” wavefunctions. So
(relativistic) frame rotation has no effect on the values/components of a scalar
field at all.
Indeed, Schrödinger’s quantum wavefunction theory (∼1926) was formu-
lated before it was established that electrons have “spin angular momentum”.
So it described “spinless electrons” — such Schrödinger operators abound in
physics models when electron spin effects can be ignored.
Later on, the geometric understanding of spin improved, triggered in part
by Dirac’s equation for the spin-ful electron field and its consistent but strange-
looking transformation behaviour under (relativistic) rotations. Over time, it
became appreciated that general quantum wavefunctions were not really scalar
fields/functions but sections of vector bundles associated to certain non-trivial
spin-representations. Then it becomes mandatory to introduce connections,
or more precisely, the induced covariant differentiation on the quantum wave-
sections.

129
13 Parallel transport and covariant differenti-
ation
13.1 Parallel transport of frames
Definition 48. Let P be a principal G-bundle over X with a connection ω.
Let γ : [0, 1] → X be a curve in X. A curve γ̃ : [0, 1] → P is a horizontal lift
of γ if π ◦ γ̃ = γ and its velocity vectors γ̃ ′ (t) are horizontal for all t ∈ [0, 1].

Note: At an end point, the velocity vector is defined by the one-sided deriva-
tive.

Theorem 13.1. Let π : P → X be a principal G-bundle with connection, and

let γ : [0, 1] → X. For each p ∈ π −1 (γ(0)) in the initial fibre, there exists a
unique horizontal lift γ̃p : [0, 1] → P , satisfying γ̃p (0) = p.

Proof. [Optional.] First, assume that γ lies within U with P |U trivializable.

Fix any section s : U → P with s(γ(0)) = p. This gives a reference lift s ◦ γ of
γ. A general (non-horizontal) lift has the form γ̃p (t) = (s ◦ γ)(t) · g(t) for some
smooth g : [0, 1] → G with g(0) = e. We need to solve for the appropriate g
such that γ̃ is horizontal.
For convenience, we leave the initial point p implicit. Writing σ : P × G →
P for the G-action, we consider γ̃ as a composition of maps,

γ̃ : t 7→ σ((s ◦ γ)(t), g(t)).

As in Eq. (11.7) we have

  ♯
′ ′ ∗ d
γ̃ (t) = d(Rg(t) )(s◦γ)(t) ((s ◦ γ) (t)) + g Θ ,
dt′ t′ =t γ̃(t)

and horizontal-ness means that it is annihilated by the connection 1-form ω,

 
need ′ ∗ ′ ∗ d
0 = ωγ̃(t) (γ̃ (t)) = (Rg(t) ω)(s◦γ)(t) ((s ◦ γ) (t)) + g Θ
dt′ t′ =t
= Adg(t)−1 ◦ ω((s ◦ γ)(t)) + (Lg(t)−1 )∗ (g ′ (t))
= (Lg(t)−1 )∗ (Rg(t) )∗ ω((s ◦ γ)′ (t)) + g ′ (t) .

130
For convenience, we write β : t 7→ ω((s ◦ γ)′ (t)), which is a fixed curve in
g = Te G. The above vanishing condition now reads

(Rg(t) )∗ (β(t)) + g ′ (t) = 0, (13.1)

which is an ODE for the function g : [0, 1] → G with initial condition g(0) = e.
To solve this ODE, we define a vector field v on G × [0, 1] by
 ∂ 
v(h,t) = − (Rh )∗ (β(t)), .
∂t′ t′ =t

Then, by construction, an integral curve of v starting at (e, 0) will have the

form t 7→ (g1 (t), t), where g1 satisfies the ODE, Eq. (13.1), and the initial
condition. However, g1 may only be defined for some small time interval
[0, δ1 ] ⊂ [0, 1]. So we consider another integral curve t 7→ (g2 (t), t + δ1 ) of v,
starting from the point (e, δ1 ); then t 7→ g2 (t − δ1 ) also also satisfies Eq. (13.1)
for t ∈ [δ1 , δ2 ]. We modify this second curve to t 7→ g1 (δ1 ) · g2 (t − δ1 ), so that
it matches up with g1 at time t = δ1 . Repeat this extension process until we
get a solution to Eq. (13.1) for t ∈ [0, 1]. Here, compactness of {e} × [0, 1]
ensures that there is a minimal δ > 0 which works no matter which t we start
an integral curve at, so the extension process terminates after a finite number
of steps.
The above procedure uniquely produces a horizontal lift in the case where
γ lies in one trivializing chart. In general, we can use compactness of the curve
to argue that there is a finite subdivision of [0, 1] into subintervals, such that
each subcurve lies within a single chart. Note that when switching chart during
the extension process, we need to convert to a different section/trivialization,
but the (partial) horizontal lift does not depend on the choice of section.

Remark. Why is lifting a curve γ in X harder than lifting a vector field? Could
we not take the velocity vectors along γ, extend it arbitrary to a vector field
on X, then use the horizontally lifted vector field on P ? The issue is whether
we can integrate the resulting vector field into a curve γ̃p above all of γ. This
is what we had to prove above.
Theorem 13.1 allows us to make the following definition:

Definition 49. Let π : P → X be a principal G-bundle with a connection ω,

and let γ : [0, 1] → X be a curve. The parallel transport along γ with respect

131
to ω is the map

τγ ≡ τγω : Pγ(0) → Pγ(1)

p 7→ γ̃p (1),

where γ̃p is the unique horizontal lift of γ which starts at p ∈ Pγ(0) .

Parallel transport is G-equivariant:

Proposition 13.2. Let ω be a connection on a principal G-bundle P → X.

Let τγω be the parallel transport map along a curve γ. Then

τγω (p · g) = (τγω (p)) · g, p ∈ Pγ(0) , g ∈ G.

Proof. Exercise.
Exercise 13.1. Show that τγ ≡ τγω has the following properties:

ˆ It is independent of the parametrization of the curve γ.

ˆ For any path γ obtained as the smooth concatenation of two paths γ1 , γ2 :

[0, 1] → X, i.e.,
(
γ1 (2t), t ∈ [0, 12 ],
γ(t) =
γ2 (2t − 1), t ∈ [ 21 , 1],

the equation τγ = τγ2 ◦ τγ1 holds.

ˆ For the flipped path γflip : t 7→ γ(1 − t), the parallel transport satisfies
τγflip = τγ−1 .

Parallel transport also transforms nicely under gauge transformations:

Proposition 13.3. Let F ∈ G(P ) be a gauge transformation, and let F ∗ ω be

the transformed connection (Eq. (11.14)). Then
∗ω
τγF = F −1 ◦ τγω ◦ F.

132
Proof. Let p ∈ P be given. Write q = F (p) = p · σF (p), and let γ̃p be the
ω-horizontal lift of γ starting at q. Our candidate (F ∗ ω)-horizontal lift is

Γ = F −1 ◦ γ̃q .

We check that Γ is indeed (F ∗ ω)-horizontal,

 
∗ ′ −1 ′
(F ω)Γ(t) (Γ (0)) = ωγ̃q (t) dFΓ(t) ◦ dFγ̃q (t) (γ̃q (t)) = ωγ̃q (t) (γ̃q (t)) = 0,

where the last equality follows from γ̃q being ω-horizontal. The initial point
of Γ is
Γ(0) = γ̃q (0) · σF−1 (γ̃q (0)) = p · σF (p) · σF−1 (p · F (p)) = p.
| {z }
−1
σF (p)

Thus
∗
τγF ω (p) = Γ(1) = F −1 ◦ γ̃q (1) = F −1 ◦ τγω ◦ F (p).

13.2 Holonomy of parallel transport

Let us restrict attention to loops in X, based at a fixed point x, i.e., γ(1) =
γ(0) = x. Then the parallel transport map along γ becomes a self-map of the
fibre at x,
τγω : Px → Px ,
called the holonomy of ω around γ. It is convenient to think of the holonomy
around γ as being implemented by right-multiplication by some map gγ : Px →
G,
τγω (p) = p · gγ (p).
By Prop. 13.2,

τγω (p · g) = τγω (p) · g = p · gγ (p) · g = (p · g) · (g −1 gγ (p)g ).

| {z }
gγ (p·g)

Thus, the holonomy map gets conjugated by g −1 when we adjust the fibre
basepoint by g.
Next, consider loops as maps γ : S 1 → X without a preferred basepoint.
The holonomy around the loop will depend on the choice of start/end point

133
for the loop (as well as the initial fibre point above it). But the ambiguity is
just a conjugation by the parallel transport along the part of the loop joining
the two choices of start/end points. So, up to conjugacy, the holonomy of ω
around a loop γ is well-defined without reference to basepoints.
Definition 50. Fix a point p ∈ Px . Given a connection ω, its holonomy group
at p is defined as

Holp (ω) := {g ∈ G : ∃ loop γ based at x with τγω (p) = p · g}

It follows from Exercise 13.1 that Holp (ω) is indeed a group, and it is easy
to see that
Holp·g (ω) = g −1 Holp (ω)g, g ∈ G.
Prop. 13.3 shows that the conjugacy class of the holonomy along a loop
is invariant under gauge transformations of the connection. To utilize this,
we pass to a representation ρ : G → GL(n), then we can take the trace of
ρ(gγ (p)) to get a gauge-invariant of the loop γ (any fibre point p above any
loop point x can be used). Such a quantity is called a Wilson loop in physics,
and sometimes simply called “the holonomy”. Note that if G is abelian, then
the conjugations are trivial, and the holonomy itself (not just the conjugacy
class) is gauge-invariant.

13.3 Covariant derivatives on associated vector bundles

We have seen that a connection ω on a principal G-bundle P → X defines
parallel transport in P along curves in X. This automatically defines a corre-
sponding parallel transport in any associated vector bundle.
Definition 51. Let P → X be a principal G-bundle with connection ω, and
E = P ×ρ V be an associated vector bundle. Given a curve γ : [0, 1] → X, the
map

τγE,ω : Eγ(0) → Eγ(1)

[p, ξ] 7→ [τγω (p), ξ]

is called the parallel transport in E, along γ, with respect to ω.

It is easy to see that τγE,ω is well-defined. Suppose [p, ξ] = [p′ , ξ ′ ] in E, so

[p′ , ξ ′ ] = [p · g, g −1 ξ]

134
for some g ∈ G. Then
Prop. 13.2
[τγω (p′ ), ξ ′ ] = [τγω (p · g), g −1 · ξ] = [τγω (p) · g, g −1 · ξ] = [τγω (p), ξ].

Basically, once we know how to parallel transport the frames, then the com-
ponents ξ just go along for the “parallel ride”.
Notice that τγE,ω provides a linear identification of the initial fibre Eγ(0)
with the final fibre Eγ(1) . Now we are finally ready to differentiate sections
of E. Let vx ∈ Tx X be a tangent vector to the base manifold at some point
x ∈ X. Let γ : (−ϵ, ϵ) → X be any curve such that γ(0) = x and γ ′ (0) = vx .
For t ∈ (−ϵ, ϵ), we write γt for the curve γ restricted to the time interval [0, t].

Definition 52. Let P → X be principal G-bundle with connection ω, and let

E be an associated vector bundle. The covariant (directional) derivative of a
section ψ ∈ Γ(E) along a tangent vector vx ∈ Tx X is

d
∇ωvx ψ := (τγE,ω
t
)−1 (ψ(γ(t))) ∈ Ex , (13.2)
dt t=0

where γ is any representative curve for vx .

In words: we parallel transport all the ψ(γ(t)) back to the same fibre Ex ,
then take the usual t-derivative. The key feature of the covariant derivative
is its manifest gauge-independence — we do not have to pick any gauge to
convert the section ψ into a V -valued function.
Let us check that Definition 52 only depends on the velocity vector vx , not
the choice of representative curve γ. Pick any local gauge s : U → P , and
write
ψ(x′ ) = [s(x′ ), ξ(x′ )], x′ ∈ U,
for some function ξ : U → F representing ψ. The local gauge gives a reference
curve s ◦ γ in P , which can be parallel transported to a curve r in the initial
fibre Px ,
r(t) := (τγωt )−1 ((s ◦ γ)(t)) ∈ Px .
The curve r has the form r(t) = s(x) · g(t) for a unique curve g ≡ g(t) in G

135
satisfying g(0) = e. With ψ locally represented as [s, ξ], we rewrite

(τγE,ω
t
)−1 (ψ(γ(t))) = (τγE,ω
t
)−1 [(s ◦ γ)(t), ξ(γ(t))]
= (τγE,ω
t
)−1 [τγωt (r(t)), ξ(γ(t))]
= [r(t), ξ(γ(t))] (Defn. 51)
= [s(x) · g(t), ξ(γ(t))]
= [s(x), ρ(g(t)) · ξ(γ(t))],

where ρ : G → GL(n) is the representation defining the associated bundle E.

Taking the t-derivative at t = 0,
∈ Te G=g
z}|{
∇ωvx ψ = [s(x), ρ(e) · dξx (vx ) + dρe ( g ′ (0) ) · ξ(x)]. (13.3)

(Here, for the V -valued function ξ, the directional derivative dξx (vx ) is ob-
tained as dξ(γ(t))
dt
|t=0 with γ a curve representing vx .)
We now calculate
horizontal
 
z }| {
connection d d
g ′ (0) ω((s(x) · g)′ (0)) = ω  ω

= τ
 dt t=0 t γ (s(x)) + (s(x) · g(t))
dt t=0 

 
d ω
=ω τ (s(x)) · g(t)
dt t=0 γt
 
d ω
=ω τ (s(x) · g(t))
dt t=0 γt
= ω((s ◦ γ)′ (0))
= s∗ ω(vx ).

So Eq. (13.3) becomes the formula

∇ωvx ψ = [s(x), dξx (vx ) + dρe (s∗ ω(vx )) · ξ(x)],

which does not refer to the curve γ.

It also makes sense to write

(∇ωv ψ)(x) := ∇ωvx ψ, ψ ∈ Γ(E), v ∈ X(X),

136
because the dependence on vx is smooth. Thus each vector field v ∈ X(X)
determines a map of sections of E,

∇ωv : Γ(E) → Γ(E).

We record the above calculations:

Proposition 13.4. With respect to a local gauge s : U → P , the covariant

derivative ∇ωv along a vector field v ∈ X(X) has the formula

∇ωv ψ = [s, dξ(v) + dρe (s∗ ω(v)) · ξ], ψ ∈ Γ(E), (13.4)

where ξ is the V -valued component function of ψ with respect to the gauge s.

For the first term in Eq. (13.4), dξ is a V -valued 1-form, so dξ(v) is the
ordinary directional derivative of the V -valued function ξ along v. The local
gauge potential appears as an extra second term s∗ ω(v) acting linearly on ξ.
Finally, it is customary to consider v ∈ X(X) itself as an input variable for
the covariant derivative:

Definition 53. Let P → X be a principal G-bundle with connection ω, and

E = P ×ρ V be an associated vector bundle. With respect to ω, the covariant
derivative is the map ∇ω : Γ(E) → Ω1 (X, E) defined by

(∇ω ψ(v))(x) ≡ (∇ωv ψ)(x) := ∇ωvx ψ, ψ ∈ Γ(E), v ∈ X(X).

Proposition 13.5. Let P → X be a principal G-bundle with connection ω.

On any associated vector bundle E (with fibre a vector space over K), the
covariant derivative ∇ω has the following properties.
For all ψ ∈ Γ(E), u, v ∈ X(X), f ∈ C ∞ (X), h ∈ C ∞ (X, K),

ˆ ∇ω is K-linear in ψ;

ˆ ∇ωu+f v ψ = ∇ωu ψ + f · ∇ωv ψ;

ˆ ∇ωv (h · ψ) = v(h) · ψ + h · ∇ωv ψ.

Proof. Exercise.
Note: When K = C, the tangent vector field v is a derivation of the function
algebra C ∞ (X, K) in the same way as in the K = R case).

137
From covariant derivative to parallel transport. Sometimes, one begins
with a vector bundle E → X, and defines a covariant derivative on E to be a
map ∇ : Γ(E) → Ω1 (X, E) satisfying the algebraic properties listed in Prop.
13.5. Such a ∇ axiomatically allows differentiation of sections in a gauge-
independent way. The notion of path-lifting, parallel transport, and holonomy
also follows from such a starting point.
However, the information of the structure group G ⊂ GL(n) (e.g. due to a
bundle metric) is not yet involved. If we want the parallel transport/covariant
differentiation to preserve this extra structure, then extra conditions must be
imposed on ∇. Some examples are discussed in Section 13.6-13.7.

13.4 Physics notation

Set ηa = (0, . . . , 0, |{z}
1 , 0, . . . , 0) where a = 1, . . . , rank(E). So the ea := [s, ηa ]
a-th
provide a basis of local sections of E, whose component functions are constant
(with respect to the local gauge s). The covariant derivative of ea simplifies to

∇ωv ea = [s, (dρ)e (s∗ ω(v)) · ηa ].

Suppose the local gauge is defined on a coordinate chart U (restrict the domain
otherwise). Then we could choose v = ∂i to be a local coordinate vector field,
and obtain
∇ωi ea := ∇ω∂i ea = [s, (dρ)e (s∗ ω(∂i )) · ηa ]
Recall the physicists’ notation for the local gauge potential s∗ ω = Aj dxj ,
where Aj = s∗ ω(∂j ) is a g-valued function over U . To simplify notation, one
suppresses the Lie algebra representation dρe on V , so
rank
XE
∇ωj ea = [s, Aj · ηa ] = [s, Ab
ja ηb ] = Abja eb
b=1

for the connection (matrix) coefficients Abja .

A general local section of E is expanded as ψ = ψ a ea . Its j-th covariant
derivative is, by the third Leibniz property in Prop. 13.5,

∇ωj ψ = ∇ωj (ψ a ea ) = (∂j ψ a )ea + ψ a ∇ωj ea

“ = ” ((∂j + Aj )ψ)a ea

138
In the last line, one pretends that ψ is a V -valued function (ψ 1 , . . . , ψ rank E ) on
which the matrix Aj = Abja acts. Finally, reference to ω and the local frame
{e1 , . . . , erank E } is usually suppressed, and there is usually an i discrepancy
between physicists’ and mathematicians’ matrix Lie algebra conventions, so

“ ∇j = ∂j − iAj ” acting on “V -valued functions” ψ.

G = U(1) Abelian example. Let P be a principal U(1)-bundle over X with

connection ω, and ρ : U(1) → U(1) be the defining representation on C. So
E = P ×ρ C is a Hermitian line bundle over X. Under dρe , we have u(1) = iR
acting on C by multiplication.
So the local gauge potential is represented on E as dρe (s∗ ω) = −i nj=1 Aj dxj
P
for some functions Aj : X → R. Then the j-th “momentum operator” becomes

−i∇j = −i∂j − Aj

Now generalize the representation to ρ(q) : U(1) → U(1), eiφ 7→ eiqφ , where
(q)
q must be some integer, called the (electric) charge. Then dρe = q · 1C , and
we will arrive at the formula

“ ∇j ψ = (∂j − iqAj )ψ ”.

Note that if U(1) is replaced by the noncompact Lie group (R, +), which
has the same Lie algebra as U(1), then ρq : R → U(1), λ 7→ eiqλ , could be used
as a unitary representation of the Lie group, whether or not q is integral. So
it is actually important to specify the Lie group and its representation, and
not just the Lie algebra.
We now understand that ψ is actually a section of a line bundle. The
∂j −iqAj is just a particular local-gauge-dependent way of writing the covariant
derivative of sections.
The local gauge potential A = (A1 , A2 , A3 ) is also called a magnetic vector
potential. That is, after using the Riemannian metric to convert A to a local
vector field, the curl of A is the magnetic vector field that we are familiar with.
Actually, in three spatial dimensions, the magnetic field is better understood
as a differential 2-form F using Hodge duality (induced by the Riemannian
metric). Then the relationship between the vector potential and the magnetic
field is simply dA = F, and we have seen that the latter is independent of
gauge because G = U(1) is abelian.

139
 Classically, A, or even the gauge-independent connection ω, is not de-
tectable. This is because it is genuinely a redundant mathematical device for
efficiently solving Maxwell’s equations for the classical electromagnetic fields.
However, we learned from the Aharonov–Bohm effect that A (or ω) does mea-
surably affect the quantum mechanical evolution of a quantum wave-section.
This is because ω is playing a fundamental role in the Schrödinger equation
of motion, by specifying what it means to take derivatives. Nevertheless, only
the gauge-invariant holonomy of ω is actually measured, in accordance with
the gauge principle.

13.5 Curvature of covariant derivative

Definition 54. Let ∇ be a covariant derivative on a vector bundle E → X.
The curvature F ∇ of ∇ is defined by the equation

F ∇ (u, v)ψ = ∇u ∇v ψ − ∇v ∇u ψ − ∇[u,v] ψ, u, v ∈ X(X), ψ ∈ Γ(E).

It may be verified that F ∇ is an element of Ω2 (X, End(E)) (Exercise). In

particular, it is C ∞ (X)-linear in its arguments.
Exercise 13.2. Let Ω be the curvature of a connection ω on a principal G-
bundle, E be an associated vector bundle, and ∇ = ∇ω be the induced covari-
ant derivative on E. In a local gauge s over U , the curvature of ∇ is related
to the local field strength s∗ Ω by the formula

F ∇ (u, v)ψ = [s, dρe (s∗ Ω(u, v))(ξ)], u, v ∈ X(X), ψ ∈ Γ(E). (13.5)

13.6 Metric connections

Let E be a vector bundle. A connection ω on its frame bundle is sometimes
called a linear connection, and it gives rise to a covariant derivative ∇ω on E.
Suppose E is a Euclidean vector bundle. By definition, a metric connection
on E is a connection ω on the orthonormal frame bundle FrO (E). Recall that
E ∼= FrO (E) ×ρstd Rn where ρstd is the defining representation of O(n) on
Rn . Furthermore, as in Eq. (12.6), the bundle metric is recovered through the
formula
⟨[p, ξ], [p, ζ]⟩E = ⟨ξ, ζ⟩Rn .
The metric connection defines parallel transport of orthonormal frames, so the
induced parallel transport on E will preserve inner products.

140
 The same is true for Hermitian vector bundles, with a metric connection
being a connection on FrU (E). In this case, the terminology unitary connection
is often adopted.

Proposition 13.6. Let E be a Euclidean or Hermitian vector bundle over X.

If ω is a metric connection, then the covariant derivative ∇ω satisfies

v(⟨ψ, φ⟩E ) = ⟨∇ωv ψ, φ⟩E + ⟨ψ, ∇ωv φ⟩E , ψ, φ ∈ Γ(E), v ∈ X(X). (13.6)

Proof. At each x, we calculate in a local gauge s, replacing ψ, φ by their

n-component functions ξ, ζ. Using Eq. (13.4), we have

⟨∇ωv ψ, φ⟩E + ⟨ψ, ∇ωv φ⟩E = ⟨dξ(v) + s∗ ω(v) · ξ, ζ⟩Kn + ⟨ξ, dζ(v) + s∗ ω(v) · ζ⟩Kn .

Here s∗ ω(v) ∈ o(n) is represented as a skew-symmetric matrix function acting

on Rn (or skew-Hermitian matrix function on Cn in the complex case). So

⟨s∗ ω(v) · ξ, ζ⟩Kn + ⟨ξ, s∗ ω(v) · ζ⟩Kn = 0.

Now we are left with

⟨∇ωv ψ, φ⟩E + ⟨ψ, ∇ωv φ⟩E = ⟨dξ(v), ζ⟩Kn + ⟨ξ, dζ(v)⟩Kn

Xn
= (dξ i (v)ζ i + ξ i dζ i (v))
i=1
Xn
= v(ξ i ζ i )
i=1
= v(⟨ξ, ζ⟩Kn ) = v(⟨ψ, φ⟩E ).

Remark. If we start with a covariant derivative ∇ on a Euclidean/Hermitian

vector bundle E, it is said to be metric-compatible if Eq. (13.6) is satisfied.

13.7 Levi-Civita connection of Riemannian manifold

As a standard important example, take the tangent bundle E = T X of a
Riemannian manifold X. The sections ψ, φ of E are tangent vector fields, and
the Riemannian metric is a bundle metric, usually denoted g(·, ·) (Apologies

141
for overusing the symbol g). Suppressing the reference to ω, Eq. (13.6) for a
metric connection on T X becomes the condition
u(g(v, w)) = g(∇u v, w) + g(v, ∇u w), u, v, w ∈ X(X). (13.7)
Furthermore, it makes sense to ask for an extra torison-free condition,
∇u v − ∇v u − [u, v] = 0, u, v ∈ X(X). (13.8)
A fundamental result in Riemannian geometry states that there is a unique
covariant derivative on T X, called the Levi–Civita connection ∇LC , which is
metric compatible and torsion-free, i.e., Eq. (13.7)-(13.8) are satisfied. You
may find a proof in any Riemannian geometry textbook. The covariant deriva-
tive on a Riemannian manifold (X, g) means ∇LC , unless otherwise stated.

13.7.1 Levi–Civita connection on submanifold

Write ∂i , i = 1, . . . , N for the standard global coordinate vector fields on the
standard Riemannian manifold RN . The Levi–Civita connection on RN is the
trivial connection,
N
X
∇triv
u (v) = u(v i )∂i , u, v ∈ X(RN ).
i=1

Let X be a submanifold of RN , and equip it with the restricted Riemannian

metric. For each x ∈ X, there is an orthogonal projection px : Tx RN = RN →
Tx X. Globally, define the map
p : X × RN → T X, (x, ξ) 7→ (x, px (ξ)).
A vector field v on the submanifold X can still be expanded as
N
X
v= v i ∂i , v i ∈ C ∞ (X).
i=1

For u, v ∈ X(X), define the connection ∇ : X(X) → Ω1 (X) ⊗ X(X) to be

N
!
X
∇u (v) = p u(v i )∂i .
i=1

It can be checked that the above ∇ is the Levi–Civita connection for X (Ex-
ercise).

142
13.7.2 Levi–Civita connection on orthonormal frame bundle
The Levi–Civita ∇LC can be viewed as the covariant derivative on T X obtained
from some connection ω LC on the orthonormal frame bundle FrO (X).
Let us work in a local orthonormal frame field e = {ei }i=1,...,n over U ⊂ X.
Thus e is a local section of FrO (X) over U . (In the physics literature, e is called
a tetrad or vielbein, sometimes written eiµ ∂µ ). So each ei is a normalized local
section of T X, and its covariant derivative along any v ∈ X(X) admits an
expansion
Xn
LC
∇v ei = ωij (v)ej ,
j=1

with the ωij being local 1-forms on U given by the formula

ωij (·) := g(∇LC

(·) ei , ej ). (13.9)

Because ∇LC is metric-compatible, Eq. (13.6) of Prop. 13.6 says that

0 = v(g(ei , ej )) = g(∇LC LC
v ei , ej ) + g(ei , ∇v ej )
| {z }
δij

= ωij (v) + ωji (v).

Thus the local 1-forms ωij are antisymmetric in the indices i, j. In fact, re-
calling that so(n) are the antisymmetric matrices, these ωij are the matrix
entries of the local gauge potential e∗ ω LC ∈ Ω1 (U, so(n)), with respect to the
local orthonormal frame field e. For historical reasons, ωij is sometimes called
the “spin connection”, but we shall reserve this terminology for a more precise
object later on.

Lemma 13.7. At any point x of a Riemannian manifold, there exists a local

orthonormal tangent frame field e = {e1 , . . . , en }, such that ∇LC ei (x) = 0 for
all i = 1, . . . , n.

Proof. Take an open subset U around x, on which the orthonormal frame

bundle is trivializable. It may be assumed that U is a coordinate chart. Use
the spherical polar coordinates to obtain a radial curve joining x to x′ for any
given x′ ∈ U . Pick any orthonormal frame ex = {e1,x , · · · , en,x } at x, and
parallel transport it to x′ along the radial curve. This gives an orthonormal
frame field e = {e1 , . . . , en } over U . For any k = 1, . . . , n, there is a radial curve

143
with initial tangent vector being ek,x . Because the ei,x′ are parallel transported
from x along such radial curves, we have

∇LC LC
ek ei (x) = ∇ek,x ei = 0, ∀i, k = 1, . . . , n.

by the parallel transport definition of ∇LC , Eq. (13.2). Then ∇LC ei = 0 holds
at x.
Remark. Note that the vanishing of ∇LC ei is only guaranteed at x, not at any
other point x′ ∈ U . Furthermore, the frame field e will not generally comprise
coordinate tangent vector fields for any choice of coordinates on U . Rather, e
is a convenient device for certain calculations that we will do later.

13.8 Berry connection

Consider a trivialized bundle of finite-dimensional Hilbert spaces, X × CN .
This has a trivial (unitary) connection ∇triv , given by regarding a section as
an N -component function, and taking the usual derivatives of each compo-
nent along tangent vectors (e.g. partial derivatives ∂i , if coordinates on X
are chosen). Let ι : E ,→ X × CN be the inclusion of a Hermitian subbun-
dle E of rank n < N . So there is a smooth family p of projection matrices,
p(x) = p(x)∗ = p(x)2 , such that each Ex is the range of p(x). The Berry
connection on the subbundle E is

∇Berry = p ◦ ∇triv ◦ ι. (13.10)

Exercise 13.3. Check that ∇Berry defines a covariant derivative on E, and is

metric-compatible in the sense of Eq. (13.6).
Remark. The Berry connection occurs very frequently in quantum mechanics.
One has a fixed ambient finite-dimensional Hilbert space CN , and a control
parameter space X for Hamiltonian operators H(x). One may follow the
lowest n-eigenvalues of H(x) as x is varied. Assuming these low eigenvalues
never cross the (n + 1)-th eigenvalue, their eigenspaces define an eigenbundle
E ⊂ X × CN . This E is a Hermitian vector bundle with structure group U(n)
(no orthonormal eigenbases are preferred) and has a Berry connection, Eq.
(13.10).
Remark (Optional). It is possible to generalize the Berry connection to ambient
bundles with infinite-dimensional Hilbert space fibres. This is routinely done

144
in Floquet–Bloch theory in solid-state physics. An important subtlety, how-
ever, is that there is no canonical trivialization on the ambient Hilbert space
bundle, and therefore no canonical “Berry connection” on an eigen-subbundle.
Rather, there is a family of Berry connections, parametrized by choices of ori-
gin (implicit in defining the Floquet–Bloch transform), as explained in the
note [arXiv:1706.01149]. Therefore, certain apparently gauge-invariant quan-
tities of a single “Berry connection”, e.g. holonomy/geometric phase, may not
actually be physical. More precisely, such quantities usually only make sense
relative to some origin (e.g. polarizations).

145
14 Clifford algebras
Motivation. The Laplace P operator on C-valued functions on Rn is ∆ =
P n 2 n
j=1 (−i∂j ) . The operator j=1 −i∂j looks like a first-order square root of
∆, but we quickly find that we cannot get rid of the mixed derivatives ∂j ∂k .
Classical differential equations did not provide demand for such a square
root. In relativistic quantum theory, a wave equation needs to be first-order in
both time and space coordinates. In 1928, Dirac proposed a completely new
class of differential operators to address this difficulty. His original construc-
tion involved C4 -valued “spinors” rather than ordinary functions.
The basic idea is already apparent from the Dirac operator on Euclidean
2
R , which can be written as
   
0 ∂x − i∂y 0 ∂
D = −i ≡ −i ¯ , (14.1)
∂x + i∂y 0 ∂ 0
and does satisfy D2 = ∆⊗12 . By using a matrix version of partial derivatives,
we have made the mixed partial derivatives disappear.
More invariantly, the ei,x = ∂i |x are orthonormal basis vectors for the
tangent spaces Tx X at each x (inner product from Riemannian metric). We
simply imposed the formal algebraic condition
ei,x ej,x = −ej,x ei,x , (14.2)
turning Tx X into a noncommutative algebra.
If one only ever works in flat Euclidean space, then this sort of algebraic
construction is more or less sufficient. Under the influence of general relativity,
in the late 1920s to early 1930s, physicists and mathematicians debated on
how to make sense of spinors and Dirac operators on curved spaces. The
problem, even locally, is very subtle, particularly the correct notion of parallel
transport/covariant derivative of spinor fields, as compared to ordinary vector
fields. Furthermore, global well-definedness of spinors is not automatic. We
will eventually learn about the intrinsic geometric meaning of spinors and
Dirac operators, using the differential geometric tools introduced earlier. But
before that, we must first become familiar with the pointwise, algebraic story,
which is already not trivial.

14.1 Algebra jargon

We restrict to K = R or K = C.

146
ˆ A (unital) K-algebra is a K-vector space A with an extra bilinear asso-
ciative product A × A → A (and a unit element 1A for the product).
So, for a, b, c ∈ A and λ ∈ K, we have e.g.,

(λa + b)c = λ(ac) + bc.

ˆ A (unital) morphism of algebras is a linear map which respects the prod-

uct (and identity elements). A (unital) subalgebra of A is a subspace of
A which is closed under the product (and contains the unit).

ˆ A left ideal in A is a subspace I ⊂ A which absorbs left multiplication,

a ∈ A , b ∈ I ⇒ ab ∈ I.

Similarly for right ideals. A (two-sided) ideal is simultaneously a left

ideal and a right ideal. The algebra structure descends to the quotient
space A/I.

ˆ A quadratic vector space (V, q) is a finite-dimensional K-vector space V

equipped with a quadratic form q : V → K, i.e.,

q(λv) = λ2 q(v), λ ∈ K, v ∈ V,

with q(u + v) − q(u) − q(v) bilinear in u, v. The associated symmetric

bilinear form (also denoted q) is
1
q(u, v) := (q(u + v) − q(u) − q(v)), u, v ∈ V, (14.3)
2
and satisfies q(v, v) = q(v). For most of our purposes, it suffices to think
of q(·, ·) as a real inner product and q(·) as the squared-norm.

ˆ The tensor algebra of V is

∞
M
T (V ) = V ⊗k = K ⊕ V ⊕ (V ⊗ V ) ⊕ . . .
k=0

Let I0 be the ideal generated by the elements v ⊗ v, where v ∈ V . The

exterior algebra of V is the quotient algebra

Λ(V ) := T (V )/I0 .

147
 Note that u ⊗ v + v ⊗ u = (u + v) ⊗ (u + v) − u ⊗ u − v ⊗ v, so symmetric
2-tensors will automatically vanish in Λ(V ). A little thought shows that
the grading on T (V ) (given by the number of tensor factors) remains
well-defined on Λ(V ), i.e.,
∞
M
Λ(V ) = Λk (V )
k=0

The tensor product descends from T (V ) to Λ(V ), whence it is denoted

ω ∧ η. One may check that if ω ∈ Λk (V ), η ∈ Λl (V ), then

ω ∧ η ∈ Λk+l (V ), ω ∧ η = (−1)kl η ∧ ω.

We have Λ0 (V ) = K and Λ1 (V ) = V . If {ei }i=1,...,n is a basis for V , then

{ei1 ∧ . . . eik : 1 ≤ i1 < . . . < ik ≤ n}

is a basis for Λk (V ), so that dim Λk (V ) = nk . In particular, Λn (V ) is

one-dimensional, and there are no higher exterior powers. So Λ(V ) has

vector space dimension 2n .
Note: For V = Tx∗ X, we recover the exterior algebra of differential forms
at x.

14.2 Clifford algebra definition

Definition 55. Let (V, q) be a quadratic space, and Jq be the ideal in T (V )
generated by elements of the form v ⊗ v + q(v)1. The Clifford algebra is the
quotient
Cl(V, q) := T (V )/Jq .
In the Clifford algebra, we will generally write uv, v 2 , instead of u⊗v, v ⊗v.
There is an inclusion ιV : V ,→ T (V ) → Cl(V, q) (Exercise). For q ≡ 0, we
recover the exterior algebra, in which the square of a vector vanishes, v ∧v = 0.
For general q ̸≡ 0, we have v 2 = −q(v) being a non-zero scalar in Cl(V, q).
In terms of the symmetric bilinear form q(·, ·), the vector elements of
Cl(V, q) satisfy the anticommutation rule

uv + vu = (u + v)2 − u2 − v 2 = −q(u + v) + q(u)2 + q(v)2

Eq. (14.3)
= −2q(u, v), u, v ∈ V. (14.4)

148
Proposition 14.1. Cl(V, q) is the unique K-algebra with the property: given
a linear map f : V → A into an associative K-algebra A satisfying f (v)2 =
−q(v)1A for all v ∈ V , there exists a unique extension of f to an algebra
homomorphism, fˆ : Cl(V, q) → A . Diagrammatically:
ιV
V Cl(V, q)
∃! fˆ
.
f
A

Sketch. There is an obvious extension of f to the tensor algebra,

F : T (V ) → A , v1 ⊗ · · · ⊗ vn 7→ f (v1 ) ⊗ · · · ⊗ f (vn ).

In particular, F (v ⊗ v + q(v)) = f (v)2 + q(v)1A = 0 for any v ∈ V , so F

annihilates the ideal Iq . So F descends to a homomorphism

fˆ : T (V )/Jq ≡ Cl(V, q) → A ,

with ⊗ replaced by ∧. Proof of uniqueness of fˆ is omitted.

f g
Suppose we have quadratic form preserving linear maps (V, q) → (V ′ , q ′ ) →
′′ ′′
(V , q ). Then Proposition 14.1 implies that we obtain a commuting diagram
ιV
V Cl(V, q)
f
V ′ ◦f
ι\
ιV ′ ′′ ◦g◦f
ιV\
g◦f V′ Cl(V ′ , q ′ ) .
g
V ′′ ◦g
ι\
ιV ′′
V ′′ Cl(V ′′ , q ′′ )

Therefore the promotion of (V, q) to Cl(V, q) is functorial — q-preserving lin-

ear maps may be consistently upgraded into Clifford algebra homomorphisms.
In particular, orthogonal transformations of (V, q) induce Bogoliubov automor-
phisms of Cl(V, q).

149
14.2.1 Clifford superalgebra
Beware that the Z-grading no longer makes sense on Cl(V, q). For example, the
2-tensor v ⊗ v and the 0-tensor −q(v) represent the same element vv = −q(v)
in the Clifford algebra. Nevertheless, Cl(V, q) does retain an important Z2 -
grading, as we now explain.
Proposition 14.1 applies, in particular, to the inversion map α : v 7→ −v.
The induced parity automorphism of Cl(V, q), again denoted α, satisfies α2 =
id. So we have a linear splitting
Cl(V, q) = Cl0 (V, q) ⊕ Cl1 (V, q),
according to the +1 (“even parity”) or −1 (“odd parity”) eigenspace of α.
Since α is a homomorphism, it follows that
Cli (V, q) · Clj (V, q) ⊂ Cli+j mod 2 (V, q).
This exhibits Cl(V, q) as a Z2 -graded algebra, or superalgebra. We often require
Cl(V, q) to be represented as linear operators on vector space W , such that:
ˆ W = W 0 ⊕ W 1 is Z2 -graded;
ˆ Cl0 (V, q) acts as grading-preserving (“even”/“diagonal”) operators, while
Cl1 (V, q) acts as grading-reversing (“odd”/“off-diagonal”) operators.

14.2.2 Linear basis for Cl(V, q)

Let (V, q) be a real quadratic space, with non-degenerate q, i.e., q(v) = 0 ⇒
v = 0. Sylvester’s law of inertia says that there exists some basis {ei }i=1,...,n
of V such that n
X
q(v) = λi (v i )2 , λi = ±1,
i=1
i
where v are the components of v in that basis. Furthermore, the number of
i with λi = −1 is independent of the basis. Such a basis is orthonormal with
respect to q, in the sense that
q(ei ) = ±1, q(ei , ej ) = 0, i ̸= j,
and provides a very convenient set of generators for the Clifford algebra. For
instance, by Eq. (14.4), the basis elements mutually anticommute,
ei ej + ej ei = 0, i ̸= j,
cf. the motivating Eq. (14.2).

150
Multi-index notation: I denotes a sequence of indices 1 ≤ i1 < . . . < ik ≤
n with length |I| = k, and
eI := ei1 . . . eik ∈ Cl(V, q), e∅ = 1.
Proposition 14.2. Cl(V, q) has dimension nk=0 nk = 2n , with a basis given
P

by eI where I runs over all multi-indices. Each of Cl0 (V, q) and Cl1 (V, q) has
dimension 2n−1 , except for the case n = 0, where Cl(V, q) = Cl0 (V, q) = K.
Proof. Exercise.
Remark. Using Prop. 14.2, we deduce that there is a linear identification
Cl(V, q) → Λ• (V ), by converting eI into ei1 ∧ . . . ∧ eik . However, this is not an
algebra isomorphism, since e.g., e1 e1 = −q(e1 ) = −1 whereas e1 ∧ e1 = 0.
Exercise 14.1. Let η ∈ Cl0 (V, q), so
X
η= η I eI ,
|I| even

with respect to some orthonormal basis {e1 , . . . , en } of (V, q). Check that the
“vacuum coefficient” η ∅ does not depend on the choice of orthonormal basis.

14.2.3 Complexification
The complex field C can also be viewed as a real 2-dimensional vector space.
Given a real vector space V , the process of extending scalar multiplication
from λ ∈ R to λ ∈ C is called complexification, and is formally written
V C = V ⊗R C.
On the above tensor product of R-vector spaces, we can scalar multiply by
complex numbers,
(vλ ⊗ µ)ν = (v ⊗ λµ)ν = v ⊗ λµν, v ∈ V, λ ∈ R, µ, ν ∈ C,
so we have produced a C-vector space. We will just write v ⊗ µ ≡ vµ.
Given a real quadratic space (V, q), the quadratic form extends to a quadratic
form q C on V C , by taking q C (vµ) = q(v)µ2 ∈ C. So (V C , q C ) is a com-
plex quadratic space. Similarly, the Clifford algebra can be complexified,
Cl(V, q) ⊗R C, with the algebra product extended complex bilinearly, e.g.,
(u ⊗ µ)(v ⊗ ν) = uv ⊗ µν.

151
It may be shown, using Prop. 14.1, that there is a C-algebra isomorphism

Cl(V C , q C ) ∼
= Cl(V, q) ⊗R C
(uµ)(vν) ↔ (uv) ⊗ µν, u, v ∈ V, µ, ν ∈ C.

14.3 Clifford algebras Cln , Cln

14.3.1 Real Clifford algebras
Let us restrict attention to real (V, q) with positive-definite q. This means that
(V, q) is a real inner product space. A choice of orthonormal basis {ei }i=1,...,n
for V identifies (V, q) with (Rn , qstandard ), and induces an isomorphism of the
Clifford algebras. So it suffies to study

Cln := Cl(Rn , qstandard ).

Cl0 . We have Cl0 = R as a real algebra.

Cl1 . A linear basis for Cl1 is {1, e1 }, where e21 = −1. That is, e1 is a square
root of −1. So Cl1 ∼
= C regarded as an R-algebra. Wecould represent this on
0 −1
the Z2 -graded vector space R ⊕ R, with e1 = .
1 0

Cl2 . A linear basis for Cl2 is {1, e1 e2 , e1 , e2 }, where e1 , e2 are odd, anticom-
muting square roots of −1. The even element e1 e2 is also a square root of
−1, anticommuting with e1 and e2 . So Cl2 is isomorphic to the quaternions
H = spanR {1, I, J, K} as R-algebras.
Exercise 14.2. Show that there is an (ungraded) algebra isomorphism Cln+1 0 ∼
=
n+1
Cln . For instance, if {ei }i=1,...,n+1 is an orthonormal basis for R , consider
the map
f : Rn → Cln+1 0
, ei 7→ en+1 ei , i = 1, . . . , n.

14.3.2 Complex Clifford algebras

We write Cln := Cln ⊗R C.

152
Volume/chirality element. Suppose Rn is oriented, and let {e1 , . . . , en }
be a positively-oriented orthonormal basis. The volume/chirality element of
Cln is
n+1
ω C := i⌊ 2 ⌋ e1 · · · en .
Exercise 14.3. Check that the chirality element ω C has the following properties:
ˆ It is independent of the choice of oriented orthonormal basis.

ˆ (ω C )2 = 1.

ˆ When n is odd, ω C is central in Cln (commutes with everything). When

n is even, only Cln0 commutes with ω C .
The chirality element provides two complementary idempotents,
1 ± ωC
π± = , π + π − = π − π + = 0.
2

Chiral subalgebras, n odd case. The idempotents π ± commute with all

of Cln . Accordingly, we get a splitting of Cln into chiral subalgebras,

Cln = Cln+ ⊕ Cln− , Cln± := π ± Cln = Cln π ± .

Furthermore, α(ω C ) = −ω C in this case, so we have a “chirality swapping”

isomorphism,
α : Cln± → Cln∓ . (14.5)

Cl0 . This is C as a C-algebra.

Cl1 . We have Cl1 = Cl1 ⊗R C. There is an isomorphism of ungraded C-

algebras, (
1 ⊗i
(1, 0) 7→ 1⊗1+e
2
= π+,
C ⊕ C → Cl1 , 1 ⊗i
(0, 1) 7→ 1⊗1−e
2
= π−.
Recall that ie1 is the chirality element. So the factors in the decomposition
C⊕C ∼ = Cl1 are the chiral subalgebras Cl1+ = π + Cl1 and Cl1− = π − Cl1
respectively. Note that
e1 ∓ i 1 ± ie1
e1 (π ± u) = u = ∓i u = ∓i(π ± u), u ∈ Cl1 .
2 2
153
So e1 is identified with ∓i when Cl1± is identified with C.
There are two inequivalent 1-dimensional (complex) representations of Cl1 ,
depending on which chiral subalgebra is represented faithfully, and which one
is represented trivially.

Cl2 . We can represent Cl2 on the vector space C2 via

   
0 −i 0 −1
e1 = , e2 =
−i 0 1 0
 
1 0
Then ie1 e2 = is the chirality element, giving the representation space
0 −1
a grading, C2 = C ⊕ C, with respect to which e1 , e2 are odd operators. We
also see that {1, e1 e2 , e1 , e2 } span M2 (C). So Cl2 ∼= M2 (C).
We do not need to work out the rest of Cln separately, due to the following
“algebraic Bott periodicity”.

Proposition 14.3. For each n ≥ 0, there is a isomorphism of Z2 -graded

= Cln ⊗Cl2 ∼
algebras Cln+2 ∼ = Cln ⊗M2 (C).

Proof. In Cln , we have odd, mutually anticommuting elements ei , i = 1, . . . , n,

each squaring to −1. Similarly, in Cl2 ∼
= M2 (C), we have ẽ1 , ẽ2 . In Cln+2 we
have odd, mutually anticommuting εi , i = i, . . . , n + 2, each squaring to −1.
Then, for example, the correspondence
√
εi ↔ ei ⊗ −1ẽ1 ẽ2 , i = 1, . . . , n,
εn+j ↔ 1⊗ẽj j = 1, 2,

gives an identification Cln+2 ∼

= Cln ⊗Cl2 .
Remark. There are other algebraic periodicity results for the real Clifford al-
gebras, including those with mixed-signature quadratic forms q. Details may
be found in, e.g. §1.4 of [10], and are important when dealing with special
involutive operations like charge-conjugation, time-reversal etc.

154
15 Spin groups and representations
15.1 Spin groups
Let Cln× denote the multiplicative group of invertible elements in Cln . This
contains, in particular, all the nonzero vectors, since
v
v· = 1, v ∈ Rn \ {0}.
−q(v)

Definition 56. The twisted adjoint representation of the group Cln× on Cln is

f : Cl× → GL(Cln )
Ad n

y 7→ Ad
f y, f y (·) := α(y)(·)y −1 ,
Ad

where α is the parity automorphism.

The case where y ∈ Rn \ {0} is instructive.

Lemma 15.1. For y ∈ Rn \ {0} ⊂ Cln× , the action Ad

f y restricts to the vector
part of Cln ,
f y (v) = v − 2 ⟨y|v⟩ y,
Ad v ∈ Rn ⊂ Cln .
||y||2
Thus Ad
f y reflects vectors through the hyperplane orthogonal to y.
y y
Proof. For y ∈ Rn \ {0}, we have α(y) = −y, y −1 = −q(y)
= − ||y|| 2 , as well as

yv + vy = −2⟨y|v⟩, v ∈ Rn ⊂ Cln .

Then
f y (v) = yv(y −1 (−||y||2 )) = yvy
||y||2 Ad
= y(−yv − 2⟨y|v⟩) = ||y||2 v − 2⟨y|v⟩y.

In Lemma 15.1, scaling y by any non-zero number leads to the same re-
flection. Also, a classical result of Cartan–Dieudonné says that any element of
O(n) is a product of reflections-through-hyperplanes. These lead us to consider
the following subgroups of Cln× .

155
Definition 57. The (S)pin group is the subgroup of Cln× generated by the
unit vectors in Rn ,

Pin(n) = {y1 · · · yk : k ∈ N, yi ∈ S n−1 ⊂ Rn },

Spin(n) = Pin(n) ∩ Cln0 = {y1 · · · y2k : k ∈ N, yi ∈ S n−1 ⊂ Rn }.

By construction, there is an orthogonal representation of Pin(n),

f : Pin(n) → O(n)
Ad
y1 · · · yk 7→ Ad
f y · · · Ad
1
fy .
k

This restricts to a special orthogonal representation of Spin(n). All elements of

Spin(n) are even-parity, so Adf reduces to the usual adjoint/conjugation action
−1
Adφ (·) = φ(·)φ for φ ∈ Spin(n). Thus we have a homomorphism

f : Spin(n) → SO(n)
χ = Ad
y1 · · · y2k 7→ Ad
f y · · · Ad
1
fy
2k
= Ady1 y2 · · · Ady2k−1 y2k .

Proposition 15.2. For n ≥ 1, there are short exact sequences of Lie groups
χ=Ad
f
0 → {±1} → Spin(n) → SO(n) → 0,
Ad
f
0 → {±1} → Pin(n) → O(n) → 0.

f ⊂ Pin(n), so α(φ)vφ−1 = v for all v ∈ Rn . Decompose

Proof. Let φ ∈ ker Ad
φ = φ0 + φ1 into its even and odd parts, so

φ0 v = vφ0 , φ1 v = −vφ1 , ∀v ∈ V. (15.1)

Let {ei }i=1,...,n be the standard orthonormal basis for Rn . Then φ0 can be
expanded as a sum of products of even numbers of the ei . Whenever e1 appears
in one of these products, it can be successively moved to the front of the
product, at the expense of introducing signs and/or lowering the degree by 2
(if e1 appears twice). So φ0 = a0 + e1 a1 for some even-parity a0 and some
odd-parity a1 , neither of which contains any e1 factor. Putting φ0 = a0 + e1 a1
and v = e1 into Eq. (15.1) gives

e1 a0 + e21 a1 = e1 φ0 = φ0 e1 = a0 e1 + e1 a1 e1 = e1 a0 − e21 a1 ⇒ a1 = 0.

156
Therefore φ0 = a0 and does not contain any e1 term. Inductively, we deduce
that φ0 does not contain e2 , . . . , en , and is simply a scalar. A similar argument
shows that φ1 does not contain any e1 , . . . , en term, so φ1 = 0.
Therefore, φ can only be a scalar, which we know is obtained as a product of
unit vectors, φ = y1 · · · yk . Now, there is a transposition map (·)t : Cln → Cln ,
given by reversing the order of products — this is well-defined because the ideal
Jq is preserved under reversal of tensor product order. So consider

φ2 = φt φ = yk · · · y1 y1 · · · yk = (−1)k q(y1 ) · · · q(yk ) = ±1.

Thus φ = ±1, and −1 is attained since −1 = e21 ∈ Spin(n) ⊂ Pin(n).

Surjectivity of Ad
f uses the Cartan–Dieudonné result. Then (S)pin(n) is
a closed subgroup of Cln× . It may be shown that Cln× is open in the vector
space Cln (general finite-dimensional algebra result), so it is a Lie group, and
therefore, so is (S)pin(n).

Proposition 15.3. For n ≥ 2, the spin group Spin(n) is connected.

Proof. Each element of Spin(n) is connected to either +1 or −1 (because

SO(n) is connected; proof omitted). So we just need to check that 1 and −1
are connected within Spin(n). For orthonormal e1 , e2 , consider

γ(t) = − cos(πt) − sin(πt)e1 e2 , t ∈ [0, 1].

This is a path in Cln joining γ(0) = −1 to γ(1) = 1. We can write

γ(t) = (cos(πt/2)e1 + sin(πt/2)e2 ) (cos(πt/2)e1 − sin(πt/2)e2 ) ,

which is a product of two unit vectors, so γ(t) ∈ Spin(n).

Remark. For n ≥ 3, it may be shown that SO(n) has fundamental group (i.e.
homotopy classes of loops) being Z2 , whereas Spin(n) is simply-connected (i.e.
it has trivial fundamental group).

15.1.1 Examples of (s)pin groups

Spin(1). This is the two element group {±1}.

157
Spin(2). It is not hard to deduce the parametrization by a circle coordinate,

Spin(2) = {a + be1 e2 : a2 + b2 = 1} = {cos(θ) + sin(θ)e1 e2 : −π ≤ θ ≤ π},

(15.2)
with group inversion given by taking θ 7→ −θ. Let φ = cos(θ) + sin(θ)e1 e2 .
Then we may calculate
f φ (e1 ) = φe1 φ−1 = cos(2θ)e1 + sin(2θ)e2
Ad
f φ (e2 ) = φe2 φ−1 = − sin(2θ)e1 + cos(2θ)e2 .
Ad
So
f : Spin(2) → SO(2)
χ = Ad
 
cos(2θ) − sin(2θ)
cos(θ) + sin(θ)e1 e2 7→ (15.3)
sin(2θ) cos(2θ)
is a double-covering of circle groups.

Spin(3). This group is contained in Cl30 . Recall from Exercise 14.2 that
Cl30 can be identified with the Clifford algebra on two generators, e3 e1 , e3 e2 .
Specifically,
Cl30 = spanR {1, e1 e2 , e3 e1 , e3 e2 }. (15.4)
|{z} |{z} |{z} |{z}
∼
=Cl2 ∼
=H I J K

Let H1 be the quaternions, viewed as a 1-dimensional “vector space” with

“scalar multiplication” by the “field” H on the right. There is an “inner
product” on H1 , given by (q1 , q2 ) = q1 q2 , where the conjugate of q = a + bI +
cJ + dK is defined as q = a − bI − cJ − dK. Next, Cl2 ∼ = H is taken to act
1
on H on the left, as an algebra of quaternion-linear operators. Within Cl2 ,
there is a group of operators which preserve the inner product on H1 , called
the symplectic group Sp(1). It is easily checked that
Sp(1) = {q ∈ H : qq = 1}
comprises the unit quaternions. We will show that Sp(1) ⊂ Cl2 coincides with
Spin(3) ⊂ ClP 0 ∼ 0
3 under the identification Cl2 = Cl3 of Eq. (15.4).
Let y = 3i=1 ai ei be a unit vector in R3 . Then e3 y = a1 e3 e1 + a2 e3 e2 − a3
is identified with the unit quaternion −a3 1 + a1 J + a2 K. Similarly for ye3 . By
definition, elements of Spin(3) have the form
y1 · · · y2k = (y1 e3 )(e3 y2 ) · · · (y2k−1 e3 )(e3 y2k ), |yi | = 1.

158
So they are identified with products of unit quaternions, i.e. Sp(1) elements.
Therefore, the identification Cl30 ∼
= Cl2 restricts to an injective homomorphism
Spin(3) ,→ Sp(1) of Lie groups. Since Sp(1) comprises the unit quaternions
in H1 ∼ = R4 , it is S 3 as a manifold, thus a connected Lie group of the same
dimension as Spin(3). It may then be shown from general Lie theory that
Spin(3) ∼ = Sp(1) (Exercise).
It is also useful to regard H1 as C2 , via (a + bI + cJ + dK) ↔ a−ib ≡ wz ,



c+id
with right-multiplication by −I corresponding to complex scalar multiplication
by i. Then left multiplication by a Cl30 ∼ = H algebra element (a + bI + cJ + dK)
is represented by the complex 2 × 2 matrix
   
a − ib −c + id w −z̄
≡ . (15.5)
c + id a + ib z w̄
Taking q 7→ q̄ corresponds to taking Hermitian conjugate; unit quaternions
correspond to SU(2) matrices. We obtain a further accidental isomorphism
Spin(3) ∼
= Sp(1) ∼
= SU(2).

15.2 Spin Lie algebra

The Lie group Cln× is open in the vector space Cln , so its Lie algebra cl×n
(tangent space at the identity) is identified with Cln , with commutator as
Lie bracket. The pin and spin groups are Lie subgroups of Cln× , so their
Lie algebras sit inside Cln . Let us understand the Lie algebra spin(n) more
explicitly.
We know that spin(n) ⊂ cl× n = Cln comprises the tangent vectors to the
×
submanifold Spin(n) ⊂ Cln at the identity element. The curves
γ(ij) : t 7→ cos(t) + sin(t)ei ej , 1 ≤ i < j ≤ n,
start at the identity, and we saw from the proof of Prop. 15.3 that these
curves lie in Spin(n). At t = 0, the velocity vector of γ(ij) is ei ej . Therefore,
spin(n) contains all the bivectors ei ej . By counting dimensions, we see that
spin(n) = spanR {ei ej : 1 ≤ i < j ≤ n}.
Next, let us compare the Lie algebras spin(n) and so(n). Recall that latter
is the vector space of skew-symmetric operators on Rn equipped with the
commutator. It will be convenient to identify Λ2 (Rn ) with so(n) as vector
spaces, by regarding each wedge product u ∧ v as the skew-symmetric operator
(u ∧ v) : w 7→ ⟨u|w⟩v − ⟨v|w⟩u, u, v, w ∈ Rn .

159
For example, e1 e2 corresponds to the matrix
 
0 −1 0 ··· 0
1 0
 0 · · · 0 
0 0
E12 =  0 · · · 0 
 .. .. .. . . . .. 
. . . .
0 ··· ··· ··· 0

The surjective Lie group homomorphism

χ = Ad : Spin(n) → SO(n) ⊂ GL(n)

induces a Lie algebra isomorphism

χ∗ : spin(n) → so(n) ⊂ End(Rn )

d(Adγ(ij) )
ei ej 7→ .
dt t=0

We wish to understand what the last expression (a skew-endomorphism) does

to w ∈ Rn . A preliminary observation is
−1
dγ(ij) (t) dγ(ij) (t)
=− = −ei ej .
dt t=0 dt t=0

Dropping the (ij)-subscripts for convenience, we have

d(Adγ (w))
χ∗ (ei ej )(w) =
dt
d
γ(t)wγ −1 (t)


=
dt t=0
= γ ′ (0)wγ −1 (0) + γ(0)w(γ −1 )′ (0)
= ei ej w − wei ej
= ei ej w + (ei w + 2⟨ei |w⟩)ej
= ei ej w − ei ej w − 2⟨ej |w⟩ei + 2⟨ei |w⟩ej
= 2(ei ∧ ej )(w).

Therefore, the preimage of ei ∧ ej is

1 1
χ−1
∗ (ei ∧ ej ) = ei ej = [ei , ej ].
2 4
160
Proposition 15.4. Under the double-covering homomorphism χ : Spin(n) →
SO(n), the induced Lie algebra isomorphism χ∗ : spin(n) → so(n) ∼
= Λ2 (Rn )
is given on basis elements ei ej ∈ spin(n) by the formula

χ∗ (ei ej ) = 2ei ∧ ej , 1 ≤ i < j ≤ n.

Consequently, for u, v ∈ Rn , we have

1
χ∗−1 (u ∧ v) = [u, v].
4
Example 15.1. Elements of Spin(2) were described in Eq. (15.2). The Lie
algebra spin(2)
 is spanned
 by e1 e2 . For SO(2), the Lie algebra so(2) is spanned
0 −1
by e1 ∧ e2 ∼ . Although spin(2) and so(2) are both 1-dimensional Lie
1 0
algebras with trivial Lie bracket, the specific isomorphism χ∗ actually takes
e1 e2 7→ 2e1 ∧ e2 .
Example 15.2. In Section 15.1.1, Eq. (15.4)-(15.5), we provided a 2×2 complex
matrix representation of Cl30 ∼
= H. In particular, the basis {e1 e2 , e2 e3 , e3 e1 } for
spin(3) ⊂ Cl30 is represented as
     
0 −i 0 −1 −i 0
e2 e3 = , e3 e1 = , e1 e2 = . (15.6)
|{z} −i 0 |{z} 1 0 |{z} 0 i
−K J I

iℏ
Up to a factor of 2
, these are the spin matrices Σx , Σy , Σz used in physics.

15.3 Spin representations

Spin groups are supposed to act as frame changes for some vector space. In
quantum theory, these vector spaces are finite-dimensional complex Hilbert
spaces, so we need Spin(n) to be represented unitarily on such a Hilbert space.
Of particular importance are “spinors”, which are constructed from from Clif-
ford algebra representations.
In Section 14.3.2, we learned that Cl0 ∼= C and Cl1 ∼
= C ⊕ C. Proposition
14.3 says that
(
M2 n2 (C), n even,
Cln ∼=
M n−1 (C) ⊕ M n−1 (C), n odd.
2 2 2 2

161
 We state a few representation-theoretic results without proof (they are not
particularly difficult.) First, any finite-dimensional representation of Mm (C)
or Mm (C) ⊕ Mm (C) is completely reducible, in the sense of splitting into a
direct sum of irreducible ones.
Next, a matrix algebra Mm (C) has very simple representation theory: up
to equivalence, its only irreducible representation (over C) is the defining one
on Cm . As for Mm (C) ⊕ Mm (C), it has two inequivalent irreducible represen-
tations: the representation space is Cm as before, except that one of the two
factors of Mm (C) acts as 0. So Cln has a unique irreducible representation
for n even, and two inequivalent irreducible representations for n odd. In the
latter case, how should we distinguish the two possibilities?

Odd n. If we have an irreducible representation of ρ : Cln → End(S) on

some vector space S, then ρ(ω C )2 = 1, so we get a splitting S = S + ⊕ S −
according to the ± eigenspaces of ρ(ω C ). Since ρ(ω C ) commutes with all the
representative operators of Cln , the representation ρ can be reduced to either
eigenspace, S + or S − . By irreducibility, either S = S + or S = S − . Thus we
deduce that the irreducible representations of Cln are distinguished by whether
ρ(ω C ) acts as +1 or −1. So we could have a “left-handed” or “right-handed”
representation space.
Example 15.3. For Cl3 , we could choose to irreducibly represent the ei as
     
0 −i 0 −1 −i 0
e1 = ± , e2 = ± , e3 = ±e1 e2 = ± . (15.7)
−i 0 1 0 0 i
This gives the left-handed (resp. right-handed) irreducible representation space
S + (resp. S − ) for Cl3 .

Even n. There is a unique (up to equivalence) irreducible representation of

Cln .
Proposition 15.5. Let n ≥ 2 be even, and let ρ : Cln → End(S) be an
irreducible representation. Write S = S + ⊕ S − according to the ± eigenspaces
of ρ(ω C ). The restriction of ρ to the even subalgebra Cln0 can be reduced to
either S + or S − , and these give inequivalent irreducible representations of
Cln0 ∼
= Cln−1 .
Proof. The action of Cln0 commutes with that of ω C . So the reduction of the
Cln0 action to S + (or to S − ) is well-defined. Recall that the isomorphism

162
Cln0 ∼
= Cln−1 can be implemented by taking, say, ẽi = en ei , i = 1, . . . , n − 1.
Then the chirality element for Cln0 ∼
= Cln−1 would be
C
ωn−1 ∼ ẽ1 · · · ẽn−1 ∼ (en e1 ) · · · (en en−1 ) ∼ (e1 · · · en−1 en )enn−2 ∼ e1 · · · en ∼ ωnC ,

where we have ignored various factors of i(·) . So, up to a possible sign, the
action of ωn−1
C
∈ Cln0 on S ± is precisely the action of ωnC on S ± , which is just
a sign ±1 by definition. As n − 1 is odd, we know from the previous case that
this sign determines which irreducible Cln−1 ∼ = Cln0 we are in.
Example 15.4. For Cl2 ∼
= M2 (C) take S = C2 , and
     
0 −i 0 −1 C 1 0
e1 = , e2 = , ω = ie1 e2 = .
−i 0 1 0 0 −1

So S = S + ⊕ S − = C ⊕ C.

Definition 58. Let Cln → End(S) be an irreducible complex representation

of Cln . By restricting this representation to Spin(n) ⊂ Cln0 , we obtain a
homomorphism
δn : Spin(n) → GL(S),
called the (complex) spin representation of Spin(n). Elements of S are called
spinors. With S = S + ⊕ S − (even n case) and S = S ± (odd n case), elements
of S ± are called left-handed spinors and right-handed spinors respectively.

Proposition 15.6. For odd n, the spin representation δn of Spin(n) is inde-

pendent of which of the two irreducible representations of Cln is used. Fur-
thermore, δn is an irreducible representation of Spin(n) in this case.
For even n, the spin representation δn is reducible, and splits into δn =
δn+ ⊕ δn− according to the ±-eigenspaces of ω C . The representations δ+ and δ−
are inequivalent irreducible representations of Spin(n).

Proof. Odd n. The parity automorphism α preserves Cln0 , but swaps Cln+
with Cln− , as we saw in Eq. (14.5). Thus Cln0 is the “diagonal” subalgebra,

Cln0 = {(φ, α(φ)) ∈ Cln+ ⊕ Cln− } ⊂ Cln .

If we choose the irreducible representation of Cln for which Cln± acts non-
trivially, we obtain a representation ρ± of Cln0 , and the latter is irreducible.

163
Observe that ρ+ and ρ− are equivalent representations of Cln0 , with α imple-
menting the equivalence. In particular, their restrictions to Spin(n) ⊂ Cln0
are equivalent, so we simply write δn for either one. We observe that Spin(n)
contains a linear basis {eI }|I| even for Cln0 , so δn is also irreducible.
Even n. We saw from Prop. 15.5 that S splits into S + ⊕ S − , and that
Cln0 acts within S + or S − separately. These two options provide inequivalent
irreducible representations of Cln0 , therefore we also get inequivalent represen-
tations δ ± of Spin(n) ⊂ Cln0 on S ± .
Example 15.5. We have Spin(1) = {±1} ⊂ Cl10 = C. The spin representation
on S ± = C is the sign representation.
Example 15.6. Recall Example 15.4. The elements cos(θ) + sin(θ)e1 e2 of
Spin(2) are represented on S ± ∼ = C by the operator of multiplication by
cos(θ) ∓ i sin(θ) = e∓iθ . So Spin(2) is represented as U(S ± ) = U(1) in two
inequivalent ways. This makes clear that we should not be too sloppy about
identifying Spin(2) and U(1).
Example 15.7. In Example 15.3, we described the irreducible representations
of Cl3 on S ± ∼
= C2 , with e1 , e2 , e3 represented as in Eq. (15.7). Because the
volume/chirality element is ω C = −e1 e2 e3 is represented as ±12 , we actually
also have
e2 e3 = ±e1 , e3 e1 = ±e2 , e1 e2 = ±e3 (15.8)
in this representation space. So Cl30 is represented on S ± by the same ma-
trices found in Eq. (15.6). In particular, we obtain the spin representation of
Spin(3) ⊂ Cl30 on S ± as the SU(2) matrices found in Eq. (15.5).
Remark. In quantum mechanics in three spatial dimensions, it is usual to
present the notion of spin as directly coming from the matrix Lie group SU(2),
acting on “spinors” in C2 . Actually, the relevant Lie group is, a priori, Spin(3)
equipped with a double-covering map onto SO(3). The isomorphism Spin(3) ∼ =
SU(2) is accidental.
Remark. The Clifford-algebraic adjoint/conjugation representation
f : Spin(3) → SO(R3 )
χ = Ad

can be computed in the basis {e1 , e2 , e3 }, so each element of Spin(3) ∼

= SU(2)
matrix becomes an SO(3) matrix, which can be verified to be given by Eq.
(7.2) (exercise). Indeed, this is where that explicit double-covering formula
came from.

164
 Because of Eq. (15.8), we have coincidentally also computed the Lie-theoretic
adjoint representation
Ad : Spin(3) → GL(spin(3)).
| {z } | {z }
SU(2) su(2)

In general, the double-cover χ : Spin(n) → SO(n) is not related to the

Lie-theoretic Ad : Spin(n) → GL(spin(n)); the representation spaces do not
have the same dimension.

15.3.1 Inner product on S ±

For n = 1, 2, irreducible representations of Cln were provided
n
in Section 14.3.2.
2⌊ 2 ⌋
With respect to the standard inner product on C the elements e1 , e2 are
explicitly unitary, therefore also skew-adjoint since they square to −1. In the
n = 2 case, we have S = S + ⊕ S − according to the chirality element ω C .
The irreducible representations of Cln , n > 2 can be iteratively obtained,
following the periodicity result of Prop. 14.3. The Cl3 case was discussed
above. Generally, there will be a Hermitian inner product on the representa-
tion space, such that the ei , i = 1, . . . , n are unitary, and are odd operators
in the even n case. More abstractly, one can use an averaging argument to
deduce the existence of such an inner product (Exercise).
Now, every vector y ∈ V is a real linear combination of the ei , so as a
Clifford algebra element, it is also represented skew-adjointly. In particular,
a unit vector y will be represented unitarily. The spin group is generated by
products of such unit vectors, so Spin(n) is unitarily represented in the spin
representation. In the even n case, the action of Spin(n) may be reduced to
each graded component S ± .
Proposition 15.7. The spin representation δn is compatible with Clifford mul-
tiplication by vectors of Rn on S, in the sense that for all g ∈ Spin(n), y ∈
Rn ⊂ Cln , ξ ∈ S,
δn (g)yδn (g −1 )(ξ) = (χ(g)y) · ξ. (15.9)
Proof. Both δn (g) and the Clifford action y·(−) come from the Clifford algebra
representation ρ : Cln → End(S). So we have
δn (g)yδn (g −1 )(ξ) = ρ(g)ρ(y)ρ(g −1 )(ξ)
= ρ(gyg −1 ) · (ξ)
= ρ(χ(g)(y))(ξ) = (χ(g)y) · ξ.

165
Remark (Physics notation). If e1 , . . . en is an oriented orthonormal basis for
Rn , then the representative matrices γi := ρ(ei ) are called Dirac or gamma
matrices. So an n-tuple (y 1 , . . . , y n ) corresponds to a linear combination y i γi
of gamma matrices. Physicists usually think of χ(g) ∼ R as an SO(n) rotation
matrix (or SO(3, 1) Lorentz) matrix Rj i , and write the equality Eq. (15.9) as
a “transformation law” for gamma matrices,

Λ(R)γi Λ(R)−1 = γj Rj i . (15.10)

Typically, the indices on γi are also raised by the bilinear form q (often a
Lorentz metric with mixed signature − + ++), and one writes {Γi , Γj } = 2g ij
for the (Clifford) anticommutation relations. Here, there is usually another
sign discrepency for the square of Γi .
Because of Eq. (15.10), the gamma matrices are sometimes said to “trans-
form as a tangent (four-)vector”. Later, we will uncover the geometric meaning
of the gamma matrices more clearly (Remark 5).

166
16 Spinor geometry
16.1 Spin structures and spin connection
Recall the double-covering homomorphism, χ : Spin(n) → SO(n), n ≥ 1.
For a Riemannian manifold, we would like to upgrade this pointwise “frame
doubling” to a “principal frame bundle doubling”.

Definition 59. Let X be an oriented Riemannian manifold. A spin structure

is a χ-reduction, F : FrSpin (X) → FrSO (X), in the sense of (Definition 47). An
oriented Riemannian manifold with a spin structure is called a spin manifold.

Some remarks:

ˆ Thus, a spin structure is the data of a principal Spin(n)-bundle which

“double covers” the SO(n)-bundle of oriented orthonormal tangent frames,
while respecting the bundle projections. The same doubling notion
makes sense for Euclidean vector bundles, but we will not require this
generalization.

ˆ [Optional information.] Not all oriented Riemannian manifolds admit

spin structures — there is a topological obstruction called the second
Stiefel–Whitney class. So sometimes one encounters the term “spinnable”
manifold. If X is spinnable, the choice of spin structure may not be
unique — it is labelled by the cohomology group H 1 (X, Z2 ).

ˆ Upon fixing a basepoint p ∈ Px , the restriction F |Px : Px → FrSO (Ex ) is

identified with χ : Spin(n) → SO(n).

ˆ At each point, an orthonormal frame lifts to two possible “spin frames”.

As usual, neither spin frame is canonically preferred, but they are related
to each other by the action of −1 ∈ Spin(n). Since there are actually
more spin frames than orthonormal frames, we sometimes call FrSpin (E)
an “equivariant lift” of FrSO (E), instead of a “reduction”.

ˆ Let e : U → FrSO (X) be a local oriented orthonormal tangent frame

field, and let U ′ be a contractible open subset of U ⊂ X. Then e(U ′ ) is a
contractible open subset in FrSO (X), and its preimage under F comprises
two disjoint copies of e(U ′ ). These are the choices of “spin-frame fields”
ẽ lifting e over U ′ . Note that a lift may not exist over all of U .

167
Example 16.1. In Section 7, we constructed a spin structure on the round
2-sphere S 2 .
Example 16.2. On S 1 , we have FrSO (S 1 ) = S 1 × {1}. There are two possible
spin structures: FrSpin (S 1 ) is either disconnected, S 1 ×Z2 , or it is the connected
double cover of S 1 (cf. S(LR )). These spin structures are respectively called
Neveu–Schwarz and Ramond, in the theoretical physics literature.

The Levi–Civita connection on FrSO has a canonical lift to FrSpin :

Definition 60. Let X be a spin manifold, so we have

F : FrSpin (X) → FrSO (X).

The spin connection on FrSpin (X) is the spin(n)-valued 1-form

ω Spin := χ−1 ∗ LC
∗ ◦F ω .

Proposition 16.1. The spin connection of Definition 60 is indeed a connec-

tion on the Spin(n)-principal bundle FrSpin (X).

Proof. Let u ∈ spin(n) and u♯ be the corresponding fundamental vector field

on FrSpin (X). The first requirement for ω Spin to be a connection on FrSpin (X)
is
(χ−1∗ ◦F ω
∗ LC
)(u♯p ) = u, p ∈ FrSpin (X). (16.1)
This condition is verified by the following computation. First,
 
♯ d
dFp (up ) = dFp p · exp(tu)
dt t=0
d
= F (p · exp(tu))
dt t=0
d
= F (p) · χ(exp(tu)) (spin structure definition)
dt t=0
d
= F (p) · exp(tχ∗ u) (Ex. 10.4)
dt t=0
= (χ∗ u)♯F (p) .

Thus
(F ∗ ω LC )p (u♯p ) = ωFLC(p) (dFp (u♯p )) = ωFLC(p) ((χ∗ u)♯F (p) ) = χ∗ u,

168
and applying χ−1
∗ gives Eq. (16.1).
Next, we check the equivariance condition, Eq. (11.4), for ω Spin . For any
g ∈ Spin(n),

Rg∗ (χ−1 ∗ LC
∗ ◦F ω ) = χ−1 ∗
∗ ◦ (F ◦ Rg ) (ω
LC
)
= χ−1 ∗
∗ ◦ (Rχ(g) ◦ F ) (ω
LC
)
= χ−1 ∗
∗ ◦ F (Rχ(g) (ω
LC
))
= χ−1 ∗
∗ ◦ F (Adχ(g)−1 (ω
LC
))
= χ−1 ∗ LC
∗ ◦ Adχ(g)−1 (F ω )
Prop. 10.7
= Adg−1 ◦ χ−1 ∗ LC
∗ (F ω ),

Local spin connection. In Section 13.7.2, we saw that a local oriented

orthonormal frame e allows us to express ω LC as a local so(n)-valued 1-form,
with (e∗ ω LC )ji = ωij . Recall the standard basis ei ∧ ej for so(n). (Note the
tricky transpositions: as a matrix, ei ∧ ej has ji-entry being +1.) So the local
form of the Levi–Civita connection is
n
∗ LC
X 1X
eω = ωij ei ∧ ej = ωij ei ∧ ej . (16.2)
1≤i<j≤n
2 i,j=1

Now let ẽ be either one of the local “spin frames” which lifts e. (Here, we
work over a small enough part of X such that the lift exists.) Then the local
gauge potential for ω Spin will be

ẽ∗ (ω Spin ) = ẽ∗ (χ−1 ∗ LC

∗ ◦F ω ) = χ−1 ∗ ∗ LC
∗ ◦ (ẽ F ω )
= χ−1 ∗ LC
∗ ◦ ((F ◦ ẽ) ω ) = χ−1 ∗ LC
∗ (e ω ). (16.3)

Recalling the formula for χ∗ from Prop. 15.4, this means that the spin con-
nection has the local expression
n
∗ Spin 1X
ẽ (ω )= ωij ei ej . (16.4)
4 i,j=1

1
The extra factor of 2
in Eq. (16.4), compared to Eq. (16.2), is very important.

169
16.2 Example: spin structure and connection on S 2
In Section 7.3, we had constructed
π̄ : S 3 → S 2 , (w, z) 7→ (2w̄z, |w|2 − |z|2 ),
as a principal U(1)-bundle with right U(1)-action,
(w, z) · exp iθ = (weiθ , zeiθ ).
Let  
iθ cos(2θ) sin(2θ)
ϕ : U(1) → SO(2), e 7→ .
− sin(2θ) cos(2θ)
Then the map F : S 3 → FrSO (S 2 ), Eq. (7.2), was shown to be a ϕ-reduction
of FrSO (S 2 ) to the U(1)-principal bundle S 3 .
Now recall the spin double covering map, Eq. (15.3),
 
cos(2θ) − sin(2θ)
χ : Spin(2) → SO(2), cos θ + sin θe1 e2 7→ .
| {z } sin(2θ) cos(2θ)
exp(θe1 e2 )

If we make the identification Spin(2) ∼ = U(1), then ϕ is identified with χ, and

we have precisely exhibited S → S 2 a spin structure for S 2 . That is, we
3

regard S 3 = FrSpin (S 2 ), with Spin(2)-action

(w, z) · exp(θe1 e2 ) = (w, z) · e−iθ = (we−iθ , ze−iθ ).
The following spin(2)-valued 1-form defines a connection on S 3 = FrSpin (S 2 ),
ω(w,z) (η, ζ) = −e1 e2 Im(w̄η + z̄ζ), (η, ζ) ∈ T(w,z) S 3 ⊂ T(w,z) C2 = C2 .
With a few identifications, ω is the magnetic monopole U(1)-connection ex-
plored in Assignment 3. We shall verify that ω is precisely the spin connection.
In polar coordinates (ϑ, φ) for S 2 ,
(2w̄z, |w|2 − |z|2 ) = (sin ϑeiφ , cos ϑ).
The following is a local spin frame field,
ẽ : S 2 → S 3 = FrSpin (S 2 )
 
ϑ −iφ/2 ϑ iφ/2
(ϑ, φ) 7→ cos e , sin e ,
2 2

170
since π̄ ◦ ẽ(ϑ, φ) = (ϑ, φ) is easily verified. The local gauge potential for ω can
be computed to be
e1 e2
ẽ∗ ω = cos ϑ dφ.
2
Now, it may be checked that F ◦ẽ = e, where e is the following orthonormal
tangent frame field,
 
cos ϑ cos φ − sin φ sin ϑ cos φ
e : (ϑ, φ) 7→  cos ϑ sin φ cos φ sin ϑ sin φ  ∈ SO(3).
− sin ϑ 0 cos ϑ

We recall that the last column of the above SO(3) matrix gives the carte-
sian coordinates (of (ϑ, φ) ∈ S 2 ), while the first two columns constitute the
orthonormal frame at that point. So e = {e1 , e2 } at (ϑ, φ) is

e1 = cos ϑ cos φ ∂x1 + cos ϑ sin φ ∂x2 − sin ϑ ∂x3 = ∂θ = eϑ ,

1
e2 = − sin φ ∂x1 + cos φ ∂x2 = ∂φ = e φ ,
sin ϑ
i.e., the normalized polar coordinate frame field on S 2 . In this frame, we may
compute the local expression for the Levi–Civita connection to be
LC
ω12 LC
(ϑ, φ) = −ω21 (ϑ, φ) = cot ϑ e∗φ = cos ϑ dφ.

So in the lifted spin frame ẽ, the spin connection has local expression
Eq. (16.4) 1 e1 e2
ẽ(ω Spin ) = LC
(e1 e2 ω12 LC
+ e2 e1 ω21 )= cos ϑ dφ,
4 2
which is the same as ẽ∗ ω found above.
To summarize: The spin frame bundle for S 2 with spin connection can
be identified with the magnetic monopole U(1)-bundle.

16.3 Spinor bundles and Clifford multiplication

We have seen that a spin manifold has a “spin-frame bundle”, equipped with
a canonical spin connection. We use quotation marks because we have not yet
described the vector bundles whose “spin frames” are organized according to
this “spin-frame bundle”.

171
Definition 61. Let X be a spin manifold. The spinor bundle S over X is the
associated Hermitian vector bundle

S = FrSpin (X) ×δn S,

where δn : Spin(n) → S is the (unitary) spin representation (Definition 58).

n ⌊n⌋
Remark. Write ηa , a = 1, . . . , 2⌊ 2 ⌋ for the a-th basis vector of S = C2 2 , and
let ex ∈ FrSpin (X)x . Then the fibre Sx is spanned by the orthonormal basis
n
ψa,x := [ex , ηa ], a = 1, . . . , 2⌊ 2 ⌋ .

Similarly, if e : U → FrSpin (X) is a local section, then

n
e = (ψ1 , . . . , ψ2⌊ n2 ⌋ ), ψa = [e, ηa ], a = 1, . . . , 2⌊ 2 ⌋ ,

is a local field of (orthonormal) spin frames for S|U .

Clifford multiplication. Recall that

TX ∼
= FrSO (X) ×ρstd Rn ,
S∼= FrSpin (X) ×δ S, n

where ρstd is the defining representation of SO(n) on Rn (which will generally

be suppressed in our notation). Furthermore, each fibre Sx of the spinor bundle
carries an irreducible Clifford algebra representation ρ, so we can promote
tangent vectors of Tx X to operators acting on Sx , as follows.
Definition 62. Clifford multiplication is the vector bundle homomorphism

µ : T X ⊗ S → S,

defined for vx ∈ Tx X, ψx ∈ Sx , x ∈ X, by the formula

vx · ψx ≡ µ([ex , v] ⊗ [ẽx , ξ]) := [ẽx , v · ξ ], (16.5)

| {z } | {z } |{z}
=vx =ψx Clifford

where ex ∈ FrSO (X) is any oriented orthonormal frame at x, v is the n-tuple

of components of vx in this frame, and ẽx ∈ FrSpin (X) is a spin frame which
lifts ex .

172
Proposition 16.2. Clifford multiplication is well-defined. For even dimen-
sional X, the spinor bundle splits canonically into an orthogonal direct sum
S = S + ⊕ S − , according to the induced Clifford action of the chirality element
ω C . Furthermore, Clifford multiplication by T X is odd in the sense of mapping
S ± to S ∓ .
Proof. For any g ∈ Spin(n), we could have changed the tangent frame and
spin frame,

vx = [ex · χ(g), χ(g −1 )v], ψx = [ẽx · g, δn (g −1 )ξ].

In this alternative representation, Clifford multiplication would be

µ(vx ⊗ ψx ) = [ex · χ(g), χ(g −1 )v] · [ẽx · g, δn (g −1 )ξ]

= [ex · χ(g), χ(g −1 )v] · [ex^
· χ(g), δn (g −1 )ξ] (spin structure)
= [ex^ · χ(g), (χ(g −1 )v) · (δn (g −1 )ξ)] (new Clifford multiplication)
= [ẽx · g, δn (g −1 )(v · ξ)] (Prop. 15.7)
= [ẽx , v · ξ], (associated spinor bundle)

which is the same as what we had in the original frames, Eq. (16.5).
The proofs of the remaining statements are left as an exercise.
Globally, we may think of Clifford multiplication as defining an action of a
tangent vector field v ∈ X(X) = Γ(T X) on spinor fields ψ ∈ Γ(S). Recall that
at the beginning of Section 14, we had motivated Clifford algebras because we
wanted to promote tangent vectors to “matrix-derivatives”. Later on, we will
complete this process by providing the (covariant) derivative aspect.
Remark 5 (Local picture of Clifford multiplication). Let e = {e1 , . . . , en } :
U → FrSO (X) be a local oriented orthonormal frame field, and ẽ be a lift
to a local spin frame field. Then v is given locally by [e, v] where v is some
Rn -valued function over U . Similarly, ψ is locally given by [±ẽ, ±ξ] where ξ is
⌊n⌋
a C2 2 -valued function over U . Then Clifford multiplication (Definition 62)
is locally implemented as

ξ(x) 7→ v(x) · ξ(x), x ∈ U.

In particular, take v(x) = ei for all x ∈ U , so that [e, ei ] is the i-th vector
field ei of the orthonormal frame field. Clifford multiplication by ei effects

173
ξ(x) 7→ ei · ξ(x); in physicists’ gamma-matrix notation,
ξ 7→ γi · ξ.
We should keep in mind that the local frames e and ẽ are implicit.
In introductory treatments, one works on flat Euclidean or Minkowski
space, and uses global orthonormal coordinate vector fields. In curved space(time),
a deeper geometric understanding is required. The gamma matrices should be
understood as the local, gauge-dependent implementers of Clifford multipli-
cation by the local orthonormal frame elements. We are free to change local
orthonormal/spin frames (“local rotation/Lorentz invariance”), and this will
cause the local gamma matrices to transform according to Eq. (15.9), or the
physicist version Eq. (15.10).
Remark (Physics terminology). In the even n case, sections of S are called
Dirac spinor fields, while sections of S ± are called Weyl spinor fields. The
“Weyl spinor subbundles” S ± could also be obtained directly as
S± ∼
= FrSpin (X) ×δn+ S ± .
In the odd n case, there are actually two possible Clifford multiplications,
depending on which irreducible Clifford representation (left or right-handed)
is used. In this case, sections of S are also called Weyl spinor fields.

16.4 Spinor bundle covariant derivative

By the general constructions of Section 13.3, the associated spinor bundle
Spin
S inherits a spin-covariant derivative ∇Spin := ∇ω from the spin connec-
tion. Furthermore, this is automatically a unitary connection, because the
spin group is represented unitarily.
The general local formula for induced covariant derivatives was given in Eq.
(13.4). For example, in a local oriented orthonormal frame
∗ LC
Pn e 1= {e1 , . . . , en },
the Levi–Civita connection is described by e ω = i,j=1 2 ωij ei ∧ ej , Eq.
(16.2), where the 1-forms ωij are given by Eq. (13.9). For the spin connection,
we have, in a lifted local spin frame ẽ, the formula
∇Spin
u ψ = ∇Spin
u [ẽ, ξ]
= [ẽ, dξ(u) + (dδn )e (ẽ∗ ω Spin (u)) · ξ]
Eq. (16.4) 1
= [ẽ, dξ(u) + ωij (u)ei ej · ξ], u ∈ X(X). (16.6)
4

174
In the last line, we use the fact that the spin representation δn comes from a
Clifford algebra representation on the spinor space S, and likewise for the the
induced Lie algebra representation of spin(n). Keeping the spin frame implicit,
and writing ωkij := ωij (∂k ), the “k-th spin-covariant derivative” is
1
“ ∇Spin k = ∂k + ωkij ei ej ”, (16.7)
4
acting on S-valued functions ξ : U → S.
Write ẽ = (ẽ1 , . . . , ẽa , . . . , ẽ2⌊ n2 ⌋ ) for the spin frame field. Each constituent
spinor field ẽa = [ẽ, ηa ] has constant component functions ηa = (0, . . . , 0, |{z} 1 , 0, . . . , 0).
a-th
The spin covariant derivative of ẽa is
1
∇Spin ẽa = [ẽ, ωij ei ej · ηa ] (sum over i, j)
4
1
= ωij [e, ei ] · [e, ej ] · [ẽ, ηa ]
4
1 Eq. (13.9)
= ωij ei · ej · ẽa , ωij = g(∇LC ei , ej ). (16.8)
4
Proposition 16.3. The connections ∇Spin , ∇LC and Clifford multiplication are
Leibniz rule compatible, in the sense that for any u, v ∈ X(X) and ψ ∈ Γ(S),
∇Spin
u (v · ψ) = (∇LC Spin
u v) · ψ + v · ∇u ψ. (16.9)
Proof. It is possible to verify this directly from the formula Eq. (16.8), by
using general properties of covariant derivatives.
We sketch a more conceptual argument. Let A ∈ spin(n) and h = exp(tA) ∈
Spin(n). For any vector v ∈ Rn ⊂ Cln and spinor ξ ∈ S,
Eq. (15.9)
h(v · ξ) = (hvh−1 ) · (hξ) = (χ(h)(v)) · (hξ).
Differentiating at t = 0 gives, with the usual Leibniz rule,
A(v · ξ) = (χ∗ (A)v) · ξ + v · (Aξ). (16.10)
This is essentially what is happening when we compute ∇Spin u (v · ψ) in a spin
frame, Eq. (16.6). The local spin(n)-valued connection acts on the local spinor
representing (v · ψ), and this splits into the sum of two Clifford multiplications
by vectors, as in Eq. (16.10). The latter is the local version of Eq. (16.9).
Proposition 16.4. When n is even, ∇Spin preserves the splitting S = S + ⊕S − .
Proof. This can be shown with the help of Prop. (16.3), and is left as an
exercise.

175
16.5 Spin–Riemannian-curvature relations
Let R ∈ Ω2 (X, End(S)) be the curvature of ∇Spin (Definition 54). Let us
work out the relationship between R and the Riemann curvature tensor R ∈
Ω2 (X, End(T X)) of ∇LC . As these are tensors, we can study their relationship
in a local orthonormal frame e over U ⊂ X, with spin frame lift ẽ.
As ω LC is a connection on a principal SO(n)-bundle, have the local Cartan
structure equation,
1
e∗ ΩLC = dω (e) + [ω (e) , ω (e )], ω (e) := e∗ ω LC ∈ Ω1 (U, so(n)).
2
For u, v ∈ X(X), the so(n)-valued function e∗ ΩLC (u, v) expresses the curvature
endomorphism R(u, v) on T X|U with respect to the local frame e. Explicitly,
X
e∗ ΩLC (u, v) = g(R(u, v)ei , ej )ei ∧ ej . (16.11)
i<j

For the spin connection, the gauge potential is (Eq. (16.3))

ω̃ (ẽ) := ẽ∗ ω Spin = χ−1 ∗ LC
∗ (e ω ) = χ−1
∗ (ω
(e)
).
Since χ∗ is a Lie algebra isomorphism, the local field strength of ω Spin is
1 1
ẽ∗ ΩSpin = dω̃ (ẽ) + [ω̃ (ẽ) , ω̃ (ẽ) ] = χ−1
∗ (dω
(e)
+ [ω (e) , ω (e) ]) = χ−1 ∗ LC
∗ (e Ω ),
2 2
and from Eq. (16.11),
1X
ẽ∗ ΩSpin (u, v) = g(R(u, v)ei , ej )ei ej , u, v ∈ X(X).
2 i<j

The above spin(n)-valued function represents the spin curvature endomor-

phism R(u, v) with respect to the spin frame ẽ. Here, ei ej acts by Clifford
multiplication. Because R(u, v) is C ∞ (X)-linear, its action on a general sec-
tion ψ ∈ Γ(S) is
n
1X
R(u, v)ψ = g(R(u, v)ei , ej )ei · ej · ψ, ψ ∈ Γ(S|U ), u, v ∈ X(X).
4 i,j=1
(16.12)
Eq. (16.12) is the basic identity relating the spin curvature endomorphism
R(·, ·) to the Riemann R(·, ·). To progress further, we first recall some general
properties of R.

176
Lemma 16.5. For any u, v, w, z ∈ X(X),
R(u, v)w + R(v, w)u + R(w, u)v = 0. (First Bianchi identity)
g(R(u, v)w, z) = g(R(w, z)u, v). (Swap symmetry)
Proof. By C ∞ (X)-linearity in all arguments, it suffices to assume that u, v, w, z
are coordinate vector fields, i.e., their Lie brackets vanish. Recall that ∇ =
∇LC is torsion-free, Eq. (13.8), so ∇u v − ∇v u = [u, v] = 0. Then
R(u, v)w + R(v, w)u + R(w, u)v = ∇u ∇v w − ∇v ∇u w + ∇v ∇w u − ∇w ∇v u = 0.
+ ∇w ∇u v − ∇u ∇w v

Next, note that R is a skew-symmetric endomorphism-valued 2 form, so g(R(u, v)w, z)

is antisymmetric in the u, v arguments as well as the w, z arguments. Then
2g(R(u, v)w, z) = −g(R(v, u)w, z) − g(R(u, v)z, w)
Bianchi
= g(R(u, w)v, z) + g(R(w, v)u, z) + g(R(v, z)u, w) + g(R(z, u)v, w)
antisymmetries
= g(R(w, u)z, v) + g(R(u, z)w, v) + g(R(z, v)w, u) + g(R(v, w)z, u)
Bianchi
= −g(R(z, w)u, v) − g(R(w, z)v, u) = 2g(R(w, z)u, v).

Ricci and scalar curvature. For each pair of vector fields v, w, we have
a tangent bundle endomorphism, (·) 7→ R(·, v)w, whose trace is denoted
Ric(v, w). If we use an orthonormal frame to compute this trace, we get
n n
local skew
X X
Ric(v, w) = g(R(ei , v)w, ei ) = g(R(v, ei )ei , w).
i=1 i=1

Ric(·, ·) is called the Ricci tensor. Contracting the Ricci tensor using the metric
gives the Riemannian scalar curvature function,
n n n
g ij =δ ij
X X X
ij
Sc = g Ric(ei , ej ) = Ric(ej , ej ) = g(R(ej , ei )ei , ej ). (16.13)
i,j=1 j=1 i,j=1

For surfaces (n = 2), Sc is twice of the classical Gaussian curvature κ.

Returning to Eq. (16.12), we would like to move the Clifford multiplications
over to the spin side, and contract the arguments. This results in the spin–
Riemann scalar curvature relationship.

177
Lemma 16.6. Let e = {e1 , . . . , en } be a local orthonormal frame for the Rie-
mannian spin manifold (X, g). Then
n n
X 1X 1
ei · ej · R(ei , ej ) = g(R(ei , ej )ej , ei ) ≡ Sc.
i,j=1
2 i,j=1 2
| {z } | {z }
“spin scalar curvature′′ Riemann scalar curvature

Proof. Let us write Rlkij := g(R(ei , ej )ek , el ), so Lemma 16.5 reads

Rlkij + Rlijk + Rljki = 0, (16.14)
Rlkij = Rijlk . (16.15)
We rewrite the sum in question in terms of Rijkl ,
n n
X Eq. (16.12) 1 X
ei · ej · R(ei , ej ) = Rlkij ei · ej · ek · el .
i,j=1
4 i,j,k,l=1
For the terms with distinct i, j, k, we may rearrange the Clifford multiplications
and relabel indices to get
X 1X
Rlkij ei · ej · ek = (Rlkij + Rlijk + Rljki )ei · ej · ek = 0.
i,j,k distinct
3 | {z
i,j,k
}
=0 by Eq. (16.14)

The i = j terms contribute 0, by antisymmetry of R(·, ·). So we are left with

n
!
X 1 X X
ei · ej · R(ei , ej ) = + Rlkij ei · ej · ek · el
i,j
4 i=k,j,l i,j=k,l
1X
= (Rliij ej · el − Rljij ei · el )
4 i,j,l
1X
=− Rljij ei · el (relabel indices)
2 i,j,l
1X
=− Rijij ei · ei (Eq. (16.15))
2 i,j
| {z }
−1
n
1X Eq. (16.13) 1
= g(R(ei , ej )ej , ei ) = Sc.
2 i,j=1 2

178
17 Dirac operators
17.1 Atiyah–Singer–Dirac operator
Definition 63. The Atiyah–Singer (spin) Dirac operator on a spin manifold
X is the first-order differential operator on the spinor bundle S defined by
/ (X) : Γ(S) → Γ(S)
/ =D
D
n
X
ψ 7→ ej · ∇Spin
ej ψ
j=1

where {e1 , . . . , en } is any local oriented orthonormal frame, ∇Spin is the spin
covariant derivative,and · denotes Clifford multiplication.
To emphasize that D / does not depend on the local frame choice, it is
sometimes formulated in a slicker way, as
Spin ♯ µ
D/ : Γ(S) ∇→ Γ(T ∗ X ⊗ S) → Γ(T X ⊗ S) → Γ(S)
where ♯ is the isomorphism T ∗ X ∼ = T X given by the Riemannian metric g,
and µ is Clifford multiplication (Definition 62).
Remark. Dirac originally used an orthonormal coordinate frame (over Minkowski
space), when he wrote down his equation in 1928. In physics, one often prefers
/ = −iγ µ ∇Spin
to use a coordinate frame {∂1 , . . . , ∂n } and writes D µ . Usually,
the ∂µ cannot be chosen to be orthonormal, so one only has an expansion in
terms of “abstract” orthonormal frame vectors, ∂µ = Eµj ej , with the coeffi-
cient functions Eµj called the vielbein. Then −iγµ = −iEµj γj acts by Clifford
multiplication, and
/ = −iγ µ ∇Spin
D µ ≡ −ig µν γν ∇Spin µν k j Spin
∂µ = g Eν Eµ ek ·∇ej = g(ek , ej )ek ·∇Spin
ej = ej ·∇Spin
ej ,

as above. Nevertheless, ∇Spin is more naturally described in terms of orthonor-

mal frames, and many calculations become simpler.

17.1.1 Even dimensional case

When n is even, the Dirac operator is an odd operator with respect to the
canonical splitting S = S + ⊕ S − , due to Prop. 16.2-16.4. So we may write
!
−
0 D /
/=
D + .
/
D 0

179
n = 2 Examples. In Example 15.4-15.6, we saw that the spin representation
of Spin(2) decomposes into δ2+ ⊕ δ2− , each of which is one-dimensional.
Example 17.1. On Euclidean R2 , the orthonormal frame bundle R2 × SO(2)
is trivialized by the global coordinate frame e1 = ∂x , e2 = ∂y . Likewise, the
spin frame bundle is trivialized, R2 × Spin(2), with trivial spin connection.
The associated spinor bundle S = S + ⊕ S − has a flat trivialization, i.e., S ± ∼ =
2 Spin Spin
R × C with ∇e1 , ∇e2 simply being ∂x , ∂y . Then the Dirac operator in this
trivialization is Eq. (14.1).
Example 17.2. We worked out the spin frame bundle and spin connection for
S 2 explicitly in Section 16.2. Use the local orthonormal frame e = {∂ϑ , sin1 ϑ ∂φ },
and its spin lift ẽ given there. We saw that ẽ∗ ω Spin = e12e2 cos ϑ dφ, and in the
spin representation, we have
   
0 −i 0 −1
e1 = , e2 = .
−i 0 1 0
So on the spinor bundle, in the local frame ẽ, the Dirac operator acts on
C ⊕ C-valued functions as
∂ϑ + cot2 ϑ − sini ϑ ∂φ
 
(S 2 ) 0
/
D = −i . (17.1)
∂ϑ + cot2 ϑ + sini ϑ ∂φ 0

A point of clarification: Although the full spinor bundle S can be shown

to be a trivializable vector bundle, the individual S + and S − are not trivial-
izable. The spin frame ẽ is only defined away from a line of longitude, and
similarly for the local trivializations of the spinor bundle. This means that Eq.
(17.1) must not be mistaken for a differential operator on smooth C ⊕ C-valued
functions over S 2 . At best, it acts on smooth functions over the coordinate
domain (0, π) × (0, 2π) ⊂ R2 , subject to certain boundary conditions. (Test
your understanding of the material by figuring out what these conditions are!)
2
Conceptually, D / (S ) : Γ(S ± ) → Γ(S ∓ ) acts on sections, and its eigenspinors
are likewise sections of S. The gauge-independent perspective of D / will allow
us to understand its spectrum directly from the geometric data (which, unlike
the gauge choice, is not a fictitious input). This type of idea will be explored
in Section 17.3.
Exercise 17.1. Equation (17.1) appears to be formally non-self-adjoint, due to
the − 2i cot ϑ term. Work out the Riemannian volume form volg (see Section
17.4.2) in the (ϑ, φ) coordinates. Verify that Eq. (17.1) is formally self-adjoint,
with respect to the L2 -inner product relative to volg .

180
17.1.2 Odd dimensional case
When n is odd, recall that there are two possible spinor bundles, the left-
handed S + or the right-handed S − .
Example 17.3. Let X ∼ = R1 be the (flat) Euclidean line, with global coordinate
x and orthonormal frame {∂x }. The spinor bundle is S ± = R1 × S ± , where
S ± = C but with e1 acting as ∓i, and the spin connection coefficients are zero
in this trivialization. So the Dirac operator is
1 d d
/ (R ) = e1 ·
D = ∓i ,
dx dx
acting on sections regarded as C-valued functions.
Similarly, on the unit circle X = S 1 with global frame {∂θ }, choose FrSpin (S 1 ) =
1
S 1 × Z2 . Then the spinor bundle is trivial, and the Dirac operator is D / (S ) =
d
∓i dθ . We had already encountered these 1D Dirac operators in Section 3.
1
Later we will see how to twist D / (S ) by tensoring with a Hermitian line bundle
with connection.
Exercise: Describe the Dirac operator on S 1 when the spin structure is in-
stead the connected double cover of S 1 .

17.1.3 Relativistic version (Optional)

Physicists say “1 + n” dimensions to indicate that they are working in a
Lorentzian spacetime manifold with n spatial dimensions (usually n = 3).
The metric g is semi-Riemannian, meaning that on each tangent space it has
mixed signature − + ++. A “timelike” coordinate tangent vector has nega-
tive length and is usually written ∂0 , while the “spacelike” coordinate tangent
vectors are written ∂i as usual. Let us illustrate the (local) Dirac operator
construction in 1+1 dimensions, with flat metric.
On V ∼ = R2 with (−, +)-signature quadratic form q, pick any orthonormal
basis {e0 , e1 }, where q(e0 ) = −1 and q(e1 ) = +1. So as Clifford algebra
elements, we have e20 = 1 while e21 = −1. When we complexify, we still have
Cl(V, q) ⊗R C ∼ = Cl2 ∼
= M2 (C), which has the usual Clifford representation on
2
C ; for example,
     
0 1 0 1 C 1 0
e0 = , e1 = , ω = e1 e0 = .
1 0 −1 0 0 −1

181
The Dirac equation with mass m ≥ 0 is

(ie0 ∂0 + ie1 ∂1 ∓ m)ψ = 0.

The sign in front of the mass term is indeterminate (another torsor!). The
main point is that the anticommutation of e0 , e1 leads to

(ie0 ∂0 + ie1 ∂1 + m)(ie0 ∂0 + ie1 ∂1 − m) = (−∂02 + ∂12 ) − m2 ,

which upon writing

E = −i∂0 , p = −i∂1 ,
is the relativistic energy-momentum-mass relation, E 2 − p2 = m2 .
To recast the relativistic Dirac equation into the (apparently) non-relativistic
Schrödinger equation form with a self-adjoint Hamiltonian operator, one mul-
tiplies throughout by e0 , to get i∂0 ψ = (ie1 e0 ∂1 + me0 )ψ, or after restoring
∂0 = ∂t , ∂1 = ∂x ,  ∂ 
∂ −i ∂x m
i = ∂ .
∂t m i ∂x
| {z }
massive Dirac Hamiltonian

Notice that m = 0 is special — the Dirac Hamiltonian splits into independent

right-handed and left-handed parts, ∓i∂x , according to the chirality operator
ω C . Physicists call these the (massless) Weyl Hamiltonians.
Remark. We have SO(1, 1) being  a noncompact Lie  group with two connected
± cosh θ sinh θ
components, coordinatized as , θ ∈ R. Similarly, SO(1, n)
sinh θ ± cosh θ
is non-compact and not connected. Usually one considers the double-cover
Spin+ (1, n) of the component connected to the identity. A more important
complication is that Spin+ (1, n) will not be unitarily represented on the (finite-
dimensional) spinor space S. Instead, one defines a certain “Dirac conjugate”
of a spinor, ψ̄, such that ψ̄ψ ∈ C is Spin+ (1, n)-invariant. Why this makes
sense physically is a tricky story in quantum field theory, which we do not
pursue further.

17.2 Laplacians on vector bundles with connection

In flat Euclidean space, we have a global orthonormal frame, and the Dirac
operator is easily seen to square to the Laplacian tensored with an identity

182
 / to curved manifolds then required a deeper
matrix. The generalization of D
/2 =
understanding of its geometric ingredients. But what is the fate of “D
Laplacian”? What does “Laplacian” mean on a general vector bundle?
Definition 64 (Connection Laplacian). Let X be a Riemannian manifold,
and E → X be a vector bundle with covariant derivative ∇. For a pair of
tangent vector fields u, v ∈ X(X), the second (covariant) derivative on Γ(E)
is defined to be
∇2u,v = ∇u ∇v − ∇∇LC
u v
.
The connection Laplacian on Γ(E) is defined to be the “contracted second
covariant derivative”,
X n
∆ ≡ ∆E := − ∇2ej ,ej , (17.2)
j=1

where {ej }j=1,...,n is a(ny) local orthonormal tangent frame. If E is a Hermi-

tian/Euclidean vector bundle and ∇ is a metric connection, we write ∇∗ ∇ in
place of ∆.
Regarding the notation ∇∗ ∇, this is meant to indicate that the connection
Laplacian can also be obtained by composing ∇ with its adjoint operator (see
Prop. 17.3).
Remark. Why does Eq. (17.2) make sense? First, note that the order of u, v
in ∇2·,· matters, and this dependence is measured by the curvature,

∇2u,v − ∇2v,u = ∇u ∇v − ∇v ∇u − ∇∇LC LC

u v−∇v u

= ∇u ∇v − ∇v ∇u − ∇[u,v] (∇LC torsion-free)

= F ∇ (u, v). (17.3)

Using this, we may deduce that ∇2u,v is tensorialP in the arguments u, v (Exer-
′
cise). So in Eq. (17.2), if we replace ej by ej = k ek hkj with h an O(n)-valued
transformation, then
n
X n
X n
X n
X
∇2e′j ,e′j = hkj hlj ∇2ek ,el = (hht
)kl ∇2ek ,el = ∇2ek ,ek .
| {z }
j=1 j,k,l=1 k,l=1 k=1
δkl

Sometimes, the connection Laplacian is defined invariantly as

∆E = −Trg (∇2·,· ),

183
where Trg is the trace with respect to the Riemannian metric g. If coordinate
tangent frames are preferred, then we can also write

∆E = −g ij (∇2∂i ,∂j ).

Example 17.4. Take E = X × K with the trivial connection, so ψ may be

regarded as an ordinary scalar field X → K. In this case, the connection
Laplacian gives one of several equivalent definitions of the Laplace–Beltrami
operator ∆ acting on functions.
Example 17.5. Let ∇ be a unitary connection on a Hermitian line bundle over
X. Then ∆ = ∇∗ ∇ is the magnetic Laplacian, or Landau Hamiltonian, with
the curvature 2-form F ∇ being the magnetic field. This describes a spinless
electron field coupled to the “magnetic vector potential” ∇. We will study
this operator in detail in Section 18.
Example 17.6. For ∇ = ∇Spin the spin covariant derivative on a spinor bundle,
∆S = ∇∗ ∇ is the spinor Laplacian. When dim X is even, Prop. 16.4 shows
that ∇Spin can be restricted to Γ(S + ) and Γ(S − ) separately. So there are
+ −
actually two “‘reduced” spinor Laplacians, ∆S and ∆S .

17.3 Schrödinger–Peres–Lichnerowicz identity

On a spin manifold, we also have the “Dirac Laplacian” D / 2 acting on Γ(S),
which does not come from the connection Laplacian construction. In particular,
it does not generally coincide with the connection Laplacian ∆S for the spin
covariant derivative.

Theorem 17.1 (Lichnerowicz, 1963). Let X be a spin manifold, and ∇ =

∇Spin be the spin covariant derivative on its spinor bundle. Then
Sc
/ 2 − ∆S =
D , (17.4)
4
where Sc is the scalar curvature function on X.

Proof. At each fixed x ∈ X, we compute in a local oriented orthonormal frame

184
{e1 , . . . , en } satisfying ∇LC ei (x) = 0 (Lemma 13.7),
n
/2 =
X
D ej · ∇ej (ek · ∇ek )
j,k=1
Xn
= ej · ek · ∇ej ∇ek (Prop. 16.3)
j,k=1
Xn X
=− ∇ej ∇ej + ej · ek · ∇2ej ,ek
j=1 1≤j̸=k≤n
1 X
= ∆S + ej · ek · (∇2ej ,ek − ∇2ek ,ej ) (antisymmetry)
2 1≤j̸=k≤n
n
1 X
S
=∆ + ej · ek · R(ej , ek ) (Eq. (17.3))
2 j,k=1
Sc
= ∆S + . (Lemma 16.6)
4

Next, let us “twist” S by tensoring it with an auxiliary Hermitian line bun-

dle L with unitary connection ∇L . Then S ⊗L is called a twisted spinor bundle,
and it is equipped with a tensor product connection/covariant derivative,

∇Spin,L (ψ ⊗ φ) = (∇Spin ψ) ⊗ φ + ψ ⊗ (∇L φ), ψ ∈ Γ(S), φ ∈ Γ(L).

As in Prop. 16.3, ∇Spin,L is compatible with ∇LC and Clifford multiplication

(on the S factor).
The twisted Dirac operator on the twisted spinor bundle is defined as
n
/ (X)
X
/L = D
D L = ej · ∇Spin,L
ej ,
j=1

with {e1 , . . . , en } a(ny) local oriented orthonormal frame.

Corollary 17.2. Let X be a spin manifold, and ∇Spin be the spin covariant
derivative on its spinor bundle S. Let L be a Hermitian line bundle with
unitary connection ∇L , and let ∇Spin,L be the tensor product connection on

185
S ⊗ L. The twisted spinor Laplacian ∆S⊗L and the twisted Dirac Laplacian
/ 2L are related by the geometric identity
D
Sc 1 ∇L
/ 2L − ∆S⊗L =
D + F , (17.5)
4 2
L
where F ∇ is the curvature of ∇L acting by Clifford multiplication,
n
L L
X
F ∇ (ψ ⊗ φ) = F ∇ (ej , ek )ej · ek · ψ ⊗ φ.
j,k=1

Proof. The calculation is mostly identical to that in Theorem 17.1,

n
1 X
/ 2L (ψ ⊗ φ) = ∆S⊗L (ψ ⊗ φ) +
L
D ej · ek · (R(ej , ek )ψ ⊗ φ + ψ ⊗ F ∇ (ej , ek )φ)
2 j,k=1
n
!
S⊗L Sc 1 X ∇L
= ∆ + + F (ej , ek )ej · ek (ψ ⊗ φ),
4 2 j,k=1

L
where in the last line, we used the fact that F ∇ (ej , ek ) is a C-valued function,
so it can be commuted through the tensor product.
Historical note. Schrödinger had in fact already derived a Lorentzian signa-
ture version of Eq. (17.5) with a mass term in 1932. In 1962, A. Peres rederived
a Lorentzian version, with no mass term, as did Lichnerowicz independently.
Remark (Optional). The operator D / L is a special case of a so-called Spinc -
Dirac operator. The latter requires X to be a Spinc -manifold, which is a slight
weakening of the Spin condition.

17.4 Formal self-adjointness of Dirac operators and Lapla-

cians
17.4.1 Integration on manifolds
To integrate a function f : X → K on an n-manifold X, we might try doing
so in a coordinate chart (U, φα ),
Z Z Z
“ f ”= −1 1 n
(f ◦ φα ) dx . . . dx ≡ fˇα dx1 . . . dxn ,
U φα (U ) φα (U )

186
and integrate the Euclidean space function fˇα = f ◦ φ−1 α . However, if we
switch to another coordinate chart (U, φβ ) with coordinates (y 1 , . . . , y n ), and
write T = φα ◦ φ−1 β for the change-of-coordinate function, then we get an
inconsistency,
Z Z
ˇ 1 n
(f ◦ φ−1 1 n
fα dx . . . dx = α ◦ T ) · | det(dT )| · dy . . . dy
φα (U ) φβ (U ) | {z }
f ◦φ−1 ˇ
β =fβ
Z
̸= fˇβ dy 1 . . . dy n .
φβ (U )

For coordinate n-forms, we have the transformation law (exercise)

dx1 ∧ . . . ∧ dxn |x = det(dT )|φβ (x) dy 1 ∧ . . . ∧ dy n |x .
If X is oriented, we define integration of a compactly-supported n-form η =
f dx1 ∧ . . . ∧ dxn as Z Z
η := fˇα dx1 . . . dxn , (17.6)
U φα (U )

with respect to any positively-oriented coordinate chart (U, φα ). Because

η|x = f (x) dx1 ∧ . . . ∧ dxn |x = f (x) det(dT )|φβ (x) dy 1 ∧ . . . ∧ dy n |x ,
if we switch to the (positively-oriented) y j coordinates instead, we would get
Z Z
η≡ f det(dT ) dy 1 ∧ . . . ∧ dy n
U Z U

= fˇβ det(dT ) dy 1 . . . dy n
φ (U )
Z β
= fˇα det(dT ) det(dT )−1 dx1 . . . dxn
φ (U )
Z α
= fˇα dx1 . . . dxn ,
φα (U )

consistent with Eq. (17.6). Thus local integration of n-forms is coordinate-

independent.
To integrate a compactly-supported n-form over all of X, we pick a partition-
of-unity {ρα }α∈I subordinate to any open cover {Uα }α∈I by coordinate charts,
then define Z XZ
η := ρα · η.
X α∈I Uα

187
17.4.2 Riemannian volume form
To make sense Rof the function space L2 (X), we need an L2 -type inner product,
⟨f |h⟩L2 (X) = “ X f h”. But we had just seen that the right side does not make
coordinate-invariant sense, unless we have a prescription for converting func-
tions to n-forms. For example, weR can pick some nowhere-vanishing n-form,
called vol, and define ⟨f |h⟩L2 (X) = X f h vol (this requires X to be orientable).
Such a “volume form” is canonically provided when X is a Riemannian man-
ifold.

Definition 65. Let (X, g) be an oriented Riemannian n-manifold. Its (Rie-

mannian) volume form is given in a positively-oriented coordinate chart as
p
vol ≡ volX,g := det g dx1 ∧ . . . ∧ dxn ,
Pn i j
where g = ij=1 gij dx ⊗ dx and det g := det(gij )ij . The Hilbert space
L2 (X, g) is obtained by completing the space of compactly-supported smooth
functions with respect to the L2 -inner product
Z
⟨f |h⟩L2 (X,g) := f h vol.
X

It is straightforward to check that vol is independent of the coordinate

chart, so it is indeed a well-defined nowhere-vanishing n-form on X (exercise).
We may also regard vol as defining a measure µg on X, so that we may integrate
in the measure-theoretic sense. In this latter sense, L2 (X, g) comprises the
(equivalence classes of) µg -square-integrable measurable functions. We often
drop explicit reference to g and just write L2 (X).
Now suppose E is a Hermitian vector bundle over (X, g). Then for two
smooth sections ψ, φ, we have ⟨ψ, φ⟩E being a C ∞ (X) function given by the
pointwise inner products. In particular, for compactly-supported sections, we
can integrate to obtain the C-valued inner product
Z
⟨ψ|φ⟩L2 (X;E) = ⟨ψ, φ⟩E vol,
X

and complete this to a Hilbert space L2 (X; E) of square-integrable sections.

188
17.4.3 Formal self-adjointness
Proposition 17.3. Let ∇ be a unitary connection on a Hermitian vector
bundle E over an oriented Riemannian manifold X. The connection Laplacian
∆E is formally self-adjoint on the compactly-supported smooth sections of E.

Proof. Fix any x ∈ X, and work in a local orthonormal tangent frame with
∇LC ei = 0 at x. For compactly supported ψ, ψ̃ ∈ Γ(E), we calculate that at
x,
n
X
⟨∆E ψ, ψ̃⟩E = − ⟨∇ei ∇ei ψ, ψ̃⟩E (∇LC ei (x) = 0)
i=1
n 
X 
= −ei ⟨∇ei ψ, ψ̃⟩E + ⟨∇ei ψ, ∇ei ψ̃⟩E (metric connection (13.6))
i=1

The first sum is the divergence of a certain tangent vector field (this is inde-
pendent of the choice of frame), so it integrates to zero due to ∂X = ∅. [We
omit the Riemannian geometry treatment of div, grad, curl and Stokes’ the-
orem.] The second sum can be written as the inner product of the E-valued
1-forms ∇ψ, ∇ψ̃ at x. Altogether, we have
Z Z
E
⟨∆ ψ, ψ̃⟩E vol = ⟨∇ψ, ∇ψ̃⟩T ∗ X⊗E vol.
X X

A similar argument gives

Z Z
E
⟨ψ, ∆ ψ̃⟩E vol = ⟨∇ψ, ∇ψ̃⟩T ∗ X⊗E vol,
X X

showing that ∆E is formally self-adjoint.

The above calculation also shows that ∆E is formally the composition
of ∇ : Γ(E) → Ω1 (X; E) with its formal adjoint ∇∗ : Ω1 (X; E) → Γ(E),
justifying the notation ∆E = ∇∗ ∇.
Notice that classical geometric derivatives like d and ∇ turn sections into
section-valued 1-forms. We need to compose it with the adjoint operator to
return to the original space of sections. Upon doing so, we obtain a formally
self-adjoint operator, and after making it genuinely self-adjoint, the spectral
problem makes sense. This is why Laplace-type operators are very common in

189
geometric analysis and classical physics. From this viewpoint, the first-order
Dirac operator achieves the remarkable feat of being formally self-adjoint on
its own.

Proposition 17.4. The twisted Dirac operator D/ L is formally self-adjoint on

compactly supported smooth sections of S ⊗ L.

Proof. Write ∇ = ∇Spin,L for convenience. Similar to the proof of Prop. 17.3,
we calculate at x,
n
X
/ L ψ, ψ̃⟩S⊗L =
⟨D ⟨ei · ∇ei ψ, ψ̃⟩S⊗L
i=1
Xn
= −⟨∇ei ψ, ei · ψ̃⟩S⊗L (ei · skew)
i=1
Xn  
= −ei ⟨ψ, ei · ψ̃⟩S⊗L + ⟨ψ, ∇ei ei · ψ̃⟩S⊗L (metric connection)
i=1
n
X
= −ei ⟨ψ, ei · ψ̃⟩S⊗L
i=1
n
X
+ ⟨ψ, ei · ∇ei ψ̃⟩S⊗L . (compatibility and ∇LC ei = 0)
|i=1 {z }
/ L ψ̃⟩S⊗L
⟨ψ,D

The first sum is a divergence of some compactly-supported vector field (details

omitted), so its integral over X vanishes.

Remark. We will often encounter manifolds such as the half-line (0, ∞). When
considered as a subset of R, the manifold (0, ∞) has a “boundary” at 0, and
we often talk about “boundary conditions”. For the formal self-adjointness
results, the compact-support condition for sections over (0, ∞) means, in par-
ticular, that the sections are “supported away from the boundary”, thus they
vanish there.

190
18 Quantum Hall Effect
The quantum-mechanical spinless electron on a two-dimensional sample (em-
bedded in a laboratory Euclidean R3 ), and subjected to a magnetic field,
exhibits astonishing behaviour known as the quantum Hall effect (QHE). In
the recent literature, it has become quite popular to call the QHE a “topo-
logical phase” phenomenon. Unfortunately, this terminology is not accurate,
and often leads a newcomer (especially the enthusiastic topologist) to imagine
that the effect has something to do with, e.g., the genus of the sample. In-
stead, it has much more to do with geometry and analysis, and stability under
deformation of this data. The latter deformation-invariance is loosely thought
of as “topology” in physics.

18.1 Landau Hamiltonian on a surface

On an oriented Riemannian 3-manifold (such as R3 ), there is a (Hodge) du-
ality between tangent vector fields and 2-forms. It is customary to think of a
classical “magnetic field B” as a 2-form B, and the divergence-free condition,
∇ · B = 0, corresponds to dB = 0.
Let X be a connected, embedded 2-submanifold in R3 , assumed to be
oriented. Let g be the restricted Riemannian metric on X, so (X, g) has a
Riemannian volume 2-form vol ≡ volX,g . The restriction of the magnetic 2-
form B to X, which we still denote by B, can therefore be written as

B = B · vol, B ∈ C ∞ (X; R).

Note that B is automatically closed, since X is two-dimensional. Also, the

function B represents the intrinsic magnetic field strength felt by an electron
confined to X (and unaware of the embedding of X in R3 ). On a 2-sphere, for
instance, an intrinsically constant magnetic vector field would point outwards
perpendicularly to the sphere, and this is not extrinsically constant.
By itself, the free spinless electron dynamics on X is governed by the
Laplace–Beltrami operator ∆ on (scalar) functions X → C. As we have
learned, the magnetic field is (i-times) the curvature of a unitary connection
∇ on a Hermitian line bundle L → X,

F ∇ = −iB = −iB · vol.

191
In physics, the formally self-adjoint magnetic Laplacian ∆L = ∇∗ ∇ on L is
called the Landau Hamiltonian, and it governs the spinless quantum-mechanical
electron subjected to the magnetic field B.
The Landau Hamiltonian depends not only on the classical magnetic field
B, but on the connection ∇. Two simplifications often occur, rendering the
choice of ∇ unimportant:
1. The line bundle L is trivializable, so we can find some global U(1)-frame
such that ∇ is represented as a global u(1)-valued 1-form A satisfying

dA = F ∇ = −iB.

This happens, e.g., if the 2-manifold X is non-compact so its second

cohomology group vanishes (Omitted; this group classifies the complex
line bundles over X.)

2. Suppose ∇˜ is another connection on the trivializable L with the same

˜
curvature, F ∇ = F ∇ . Let à be the representative u(1)-valued 1-form
(with respect to the global gauge from earlier). Then
˜
d(Ã − A) = F ∇ − F ∇ = 0.

Thus η = −i(Ã − A) is a closed 1-form. If X is simply-connected, then

its first cohomology group vanishes, and this implies that η = dΛ for
some 0-form Λ (i.e. function). So in this case,

à = A + iη = A + e−iΛ d(eiΛ ),

showing that there is a globally-defined gauge transformation which con-

˜ Thus all connections on L with the same curvature are
verts ∇ into ∇.
actually gauge equivalent. Up to unitary equivalence, the Landau Hamil-
tonian depends only on the curvature form.
If the above simplifications occur, we will denote the Landau Hamiltonian
and the line bundle L by

HB := ∆L , L ≡ LB ,

which, up to unitary gauge equivalence, is unambiguous.

Remark. Recall that condition 2 fails in the Aharonov–Bohm effect.

192
 Experimentally, the quantum Hall effect refers to the quantization of the
measured values of a macroscopic transport quantity, called the transverse
Hall conductance. A priori, this conductance is real-valued, and “quantiza-
tion” means that only integer multiples of some “indivisible” conductance are
observed. Furthermore, the quantization effect persists under various kinds of
perturbations.
In order to non-vacuously explain how this quantization occurs, we cannot
assume from the beginning that it is correlated with an integer-valued topolog-
ical invariant. Instead, the spectrum of the Landau Hamiltonian is supposed
to determine the aforementioned Hall conductance. The non-trivial work is
to explain and prove how the quantization occurs, and why it is stable under
a large variety of perturbations. One aspect of the stability concerns some-
thing called Anderson localization of the spectrum under the introduction of
disorder, but we will not discuss this.

18.2 Euclidean plane Landau Hamiltonian

Consider the Euclidean plane X = R2 with coordinates (x, y), Riemannian
volume form vol = dx ∧ dy, and let B = B(x, y) = b be a nonzero constant
function (b ∈ R). This describes a spinless electron on the plane, subject to
an external magnetic field of uniform strength b perpendicular to the plane.
This is the highly-simplified geometry for a typical quantum Hall sample.
Since the plane is contractible, there is a global gauge such that

A = −ibx dy ⇒ dA = F ∇ = −ib · vol.

This is called Landau gauge in physics. In this gauge, sections ψ of Lb are

identified as functions X → C, and the covariant derivatives ∇∂x , ∇∂y act as

∇ ∂x = ∂ x , ∇∂y = ∂y − ibx.

The Landau Hamiltonian in Landau gauge is the operator

Hb = −∇2∂x ,∂x − ∇2∂y ,∂y = −∇∂x ∇∂x − ∇∂y ∇∂y = −∂x2 − (∂y − ibx)2 . (18.1)

18.2.1 Spectrum of Euclidean Landau Hamiltonian.

Direct method. The operator Hb can be shown to be essentially self-adjoint
on compactly-supported smooth functions, and we close it to a self-adjoint

193
operator (see Section 19.3.1). This operator is a remarkable example of an
exactly solvable Hamiltonian. In fact, its spectrum was already computed by
Landau in 1930, long before the discovery of the quantum Hall effect. Let us
outline the calculation.
First, observe that Hb is translation invariant in the y-direction. So we
may perform a Fourier transform, converting ∂y into ipy ,
Z ⊕
d2
Hb = Hb (py ), Hb (py ) = − 2 + (bx − py )2 .
py ∈ R
b dx

For each fixed py ∈ R,b we have a simple harmonic oscillator centered at x = py .

b
So all the Hb (py ) are isospectral, and it suffices to work out the spectrum for
the py = 0 case. By a scaling substitution, x̃ = |b|1/2 x, we get

d2
 
2
Hb (0) = |b| − 2 + x̃ = |b|HSHO ,
dx̃
where HSHO is the simple harmonic oscillator Hamiltonian. In Section 3.4, we
had already found the spectrum of HSHO to be 2N + 1.
Therefore, we conclude that

σ(Hb ) = (2N + 1)|b|.

The eigenvalue (2n + 1)|b| is called the n-th Landau level, and it is infinitely
degenerate. Contrast this with the b = 0 case, where the ordinary Laplacian
H0 = −(∂x2 + ∂y2 ) has continuous spectrum [0, ∞). The dramatic qualitative
change of the spectrum induced by the mangetic field is called Landau quan-
tization, and was first calculated in 1930 by Landau, half a century before the
unexpected experimental discovery of the quantum Hall effect by von Klitzing!

Geometric, gauge-invariant method. On the (flat) Euclidean plane, the

spinor bundle S = S + ⊕ S − has S ± being trivializable line bundles with flat
Spin
connection, F ∇ = 0. Therefore, each twisted spinor line bundle S ± ⊗Lb has
connection with curvature −ib·vol coming purely from Lb . In other words, the
connection Laplacian on S ⊗ Lb is (unitarily gauge-equivalent to) two copies
of a Landau Hamiltonian,
 S + ⊗L   L   
S⊗Lb ∆ b 0 ∼ ∆ b 0 Hb 0
∆ = − = = .
0 ∆S ⊗Lb 0 ∆Lb 0 Hb

194
 On S ⊗ Lb , we also have the formally self-adjoint twisted Dirac operator,
 
0 D /−
D/ Lb = , D/− = D/ ∗+ .
D/+ 0

Apply the Lichnerowicz identity, Corollary 17.2, with e1 = ∂x , e2 = ∂y , and

Sc = 0,
 ∗ 
2 / +D
D /+ 0
D/ Lb = = ∆S⊗Lb − ib vol(e1 , e2 )e1 · e2 ·
0 D/ +D/ ∗+
 
Hb − b 0
= . (18.2)
0 Hb + b

Here, we recall that ie1 · e2 · acts as ±1 on S ± .

Before proceeding, we mention that D / 2Lb and ∆S⊗Lb are essentially
/ Lb , D
self-adjoint on the compactly-supported smooth sections (see Section 19.3.1),
so that Eq. (18.2) holds at the level of self-adjoint operators. We may now run
the same inductive supersymmetric argument used in Section 3.4.
Suppose b > 0. We have Hb − b ≥ 0, so Hb + b ≥ 2b, and (0, 2b) is a spectral
gap for the bottom right operator. By spectral supersymmetry, the same gap
appears for the top left operator Hb −b. So Hb +b has spectral gap (2b, 4b), and
again this gap appears for Hb − b. Inductively, we see that σ(Hb − b) ⊂ 2bN.
If ker(Hb − b) = 0, then Hb + b does not have 2b in its spectrum, and neither
does Hb − b, and inductively, we would deduce that Hb − b has empty spectrum
(which is not possible). So ker(Hb − b) must be nontrivial, and inductively, we
get
σ(Hb − b) = 2bN ⇔ σ(Hb ) = (2N + 1)b.
The b < 0 case is similarly handled, by rewriting Hb ± b as Hb ∓ |b|.
Remark (Optional.). The zero-eigenspace ker(Hb − b) plays a crucial role. We
have
2
/ Lb )
ker(Hb − b) = ker((Hb − b) ⊕ (Hb + b)) = ker(D
(
/ + ), b > 0,
ker(D
/ Lb ) =
= ker(D
/ − ), b < 0.
ker(D

/ Lb to be the formal difference

We may define the “index” of D
/ Lb ) = dim ker(D
“ Index(D / + ) − dim ker(D
/ − )”.

195
This index does not immediately make sense as an integer number, because the
kernel dimensions are infinite. However, it is possible to understand this “Dirac
index” as an object in operator K-theory. Furthermore, this K-theoretic in-
dex is extremely stable against geometric perturbations. Substantially more
mathematical background is needed to understand this, but we will attempt
to illustrate the idea by considering the non-geometrically idealized Landau
Hamiltonian.

18.3 Qualitative spectrum of general Landau Hamilto-

nian
As mentioned, the real interest in the quantum Hall effect is its stability un-
der geometric perturbations. The idealized, homogeneous Hb allows one to
bring in Fourier transform methods and define Chern number “topological in-
variants”. But this approach only makes sense for a fictitious setup, whereas
we are actually interested in the non-idealized setup. For example, it is im-
possible to confine electrons to a perfectly flat 2D surface, or to generate a
perfectly uniform magnetic field. A posteriori, we learn from experiments that
the quantum Hall effect is somehow able to ignore such imperfections (up to
a certain point). This does not mean that we should a priori throw away
imperfections in the theory. Indeed, imperfections, or disorder, are actually
needed to explain important aspects of the effect.
At first glance, the spectral problem for HB on a general surface (X, g) and
a general magnetic 2-form −iB·volX,g looks impossibly complicated. Certainly,
we cannot exactly solve for its spectrum with present-day mathematics. The
non-trivial work is justify rigorously, why Hb is enough to predict the qualita-
tively relevant part of the true HB spectrum.

18.3.1 Lichnerowicz identity for Landau Hamiltonians

Consider a noncompact, simply-connected, oriented Riemannian surface (X, g).
Under these conditions, a Hermitian line bundle L → X is trivializable, and
the curvature of its connection ∇ is a complete gauge invariant. Furthermore,
a unique spin structure exists.
So let the curvature F ∇ = −iB · vol of L = LB be given. We also need the
curvature of the spinor line bundles S ± . For this, recall Eq. (16.12) for the
spin curvature 2-form R in terms of the Riemann tensor. In two dimensions,

196
we only need to understand the endomorphism R(e1 , e2 ) for {e1 , e2 } a local
oriented orthonormal frame,
n
1X
R(e1 , e2 ) = g(R(e1 , e2 )ei , ej )ei · ej ·
4 i,j=1
1
= (g(R(e1 , e2 )e1 , e2 )e1 · e2 · +g(R(e1 , e2 )e2 , e1 )e2 · e1 ·)
4
1
= (−g(R(e1 , e2 )e2 , e1 )e1 · e2 · −g(R(e2 , e1 )e1 , e2 )e1 · e2 ·)
4
n
1X Sc
=− g(R(ej , ei )ei , ej )e1 · e2 · = − e1 · e2 · .
4 i,j=1 4

Since e1 · e2 · acts as ∓i on S ± , we deduce that R(e1 , e2 ) = ± i 4Sc on S ± , and

this calculation is independent of the orthonormal frame. So the curvature of
S ± is ± i 4Sc · vol, and that of S ± ⊗ LB is −i(∓ Sc
4
+ B)vol.
Putting these calculations into the Lichnerowicz identity, Eq. (17.5),

(X,g) Sc
/ LB )2 = ∆S⊗LB +
(D − iBe1 · e2 ·
4 !
HB− Sc + Sc
4
−B 0
= 4
Sc
0 HB+ Sc + 4
+B
4

Since we are typically given a magnetic field strength B rather than B − Sc

4
,
the following equivalent identities are useful,
 
(X,g) 2 HB − B 0
(D/ L Sc ) = ,
B+ 4 0 HB+ Sc + Sc
2
+B
2

HB− Sc + Sc
 
(X,g) 2 2
−B 0
/ L Sc ) =
(D 2 . (18.3)
B− 4 0 HB + B

Now, the key point is the following. We can consider the true magnetic
field strength B to be a constant b, up to some perturbation Bpert with
supx∈X |Bpert (x)| ≪ |b|. If the scalar curvature of X is also much smaller than
|b|, then the spectral gap argument of Section 18.2.1 still works, although the
deduced gap will be smaller. For instance, HB+ Sc + Sc 2
+ B will remain strictly
2
positive, so that HB − B retains some spectral gap above 0, e.g., (0, c), where

197
c is somewhat smaller than |b|. Adding back the function B, we deduce that
HB has an isolated band of spectrum around |b|.
Generally, the effect of non-constant curvatures B and Sc is to spread out
the Landau levels into “Landau bands”.
Remark (Optional). In general, the spectral subspace of the lowest Landau
band is identified with that of the Dirac operator around 0 energy. Its “large-
scale dimension” is measured by a certain K-theoretic Dirac index, which
occurs in quantized units. This index is a modern generalization of the classical
Fredholm index (which is a genuine integer counting kernel dimensions). The
Dirac index is stable (in the appropriate sense) against geometric perturbations
(in B and Sc). Unlike the standard Fredholm index of elliptic (e.g. Dirac)
operators on compact manifolds, the Dirac index on noncompact manifolds is
not a topological invariant. It is in fact possible to puncture many holes in X,
drastically changing its topology, without changing the index.
Remark (Optional). The idealized Landau Hamiltonian with constant B = b
on a hyperbolic plane (constant negative scalar curvature) had been studied
previously by Comtet–Houston. There, the scalar curvature needs to be over-
come by a sufficiently large field b before any isolated Landau levels can form
in the spectrum. This is already very different from the (idealized) Euclidean
plane case.

18.4 Some quantum Hall physics (Optional)

We have explained the geometric origin of the formation of Landau levels/bands
in the spectrum of magnetic Laplacians. As we outlined, each spectrally iso-
lated Landau band has a Dirac index. In the idealized Euclidean plane setting,
a completely filled Landau level is understood by physicists to contribute one
unit of Hall conductance σxy ,

n filled Landau levels ↔ n units of Hall conductance. (18.4)

Regarding the “filling” of Landau levels: we are not actually studying a

single electron, but a whole density of independent electrons occupying X. It
suffices to understand the spectrum of Hb (describing a single electron), and
then construct the many-electron wavefunction by a certain “second quanti-
zation” procedure. The electrons will “fill up” the eigenspaces of Hb one by
one, starting from the lowest energy one.

198
 Each Landau level is able to admit a certain density of electron states. So
the total electron density divided by this Landau level capacity gives the filling
factor ν. When physical units are restored to the discussion, it turns out that
the Landau level capacity is proportional to the magnetic field strength |b|. So
by controlling b, the filling factor ν may be continuously varied.
Eq. (18.4) is rather tautological (integer = integer), and possibly useless,
since it only describes very specific filling factors. The key experimental sur-
prise is that σxy remains an integer even when ν is significantly varied around
integer values,

(≈ n filled Landau levels) ↔ n units of Hall conductance

The key to understanding the “forced quantization” of σxy is realizing that

the true experimental setting is “dirty”. The electron actually experiences
some electrostatic potential V coming, e.g., from the atomic lattice constitut-
ing the material sample, and more generally, impurities. So we really need
to study the spectrum of the magnetic Schrödinger operator HB + V , where
again, V is assumed to be small but nonzero, and may even be random (i.e.
basically indeterminate to the experimentalist). The effect of this “dirt” is
to modify the spectrum of HB slightly, by introducing lots of new localized
eigenstates with energies near the idealized Landau levels. In this regard, the
Bpert term in B = b + Bpert can also be thought of as “geometrical dirt”. In
contrast to the delocalized “eigenstates” constituting the clean Landau bands,
the localized eigenstates do not contribute to the Hall conductance.
Now we can understand what happens when ν is not exactly integral —
the “fractional remainder” occupies the part of the spectrum contributed by
localized eigenstates. Nevertheless, the core of the Landau band is protected
from becoming completely localized, because of its non-trivial index (the index
cannot abruptly jump to zero).

Edge states. Recall that the “geometric” calculation of σ(HB ) only works
under the assumption of essential self-adjointness. If X has a boundary, and
appropriate boundary conditions (e.g. Dirichlet) are imposed on the Landau
Hamiltonian, then the formal Dirac square root will not actually exist as a
self-adjoint operator. So we can no longer deduce the spectral gaps by the
geometric-algebraic method. In fact, the Dirichlet Landau Hamiltonian Hb
on the Euclidean half-plane can be shown to have unbroken spectrum [|b|, ∞).
The “bulk spectral gaps” ((2n + 1)|b|, (2n + 3)|b|), n ∈ N for Hb get completely

199
filled up upon the introduction of a boundary, and this continuous spectrum
is actually “unbreakable” in some sense! Here, we find another example where
issues of self-adjointness cannot be ignored, and are instead central to the
question at hand.

Conclusion. The quantum Hall effect is a remarkable spectral phenomenon

for quantum mechanical electrons subject to the combined geometry of a 2D
sample and a gauge field. The qualitative phenomenon is turns out to be in-
sensitive to the fine details of this geometry. Furthermore, a dramatic spectral
gap-filling phenomenon occurs when a boundary is introduced. To explain this
“topological phase phenomenon”, one cannot begin with a theory with only
topological ingredients.

200
19 Unbounded operators on Hilbert space
19.1 Recap of Hilbert space theory
We take for granted the Lp spaces of complex-valued functions on measure
R
1/p
spaces X, e.g., that the p-norm ||ψ||p = X |ψ(x)|p defines a complete
R
2
metric, as does the L -inner product, ⟨ψ|φ⟩ = X ψ(x)φ(x). So Lp (X) is a
Banach space, and L2 (X) is a Hilbert space.
On a general Hilbert space H, each ψ ∈ H determines a functional ⟨ψ| :
φ 7→ ⟨ψ|φ⟩, whose norm makes sense, ||⟨ψ| || := sup||φ||=1 |⟨ψ|φ⟩|. The Riesz
representation theorem says that the correspondence ψ ↔ ⟨ψ| is antilinear,
norm-preserving, and exhausts all continuous linear functionals on H. So the
dual space H∗ can be turned into a Hilbert space in its own right, and is anti-
isomorphic to H. This is the justification behind the ket-bra duality in Dirac’s
notation for the inner product.
A Hilbert space admits orthonormal bases, whose cardinality is an invari-
ant. (Note: convergent infinite linear combinations are allowed!) A linear
operator T : H → H is continuous iff it is bounded in the operator norm,

||T ||op := sup ||T ψ||,

||ψ||=1

and we write B(H) for the normed linear space of bounded linear operators
on H. For T ∈ B(H), its adjoint T ∗ ∈ B(H) is well-defined by the condition

⟨T ∗ ψ|φ⟩ = ⟨ψ|T φ⟩, ∀ ψ, φ ∈ H, (19.1)

and (T ∗ )∗ = T holds. If T : H1 → H2 is a bounded linear operator between

different Hilbert spaces, the adjoint T ∗ : H2 → H1 is defined similarly.
If a linear subspace H0 ⊂ H happens to be closed, then it is itself a Hilbert
space. The orthogonal complement H0⊥ is automatically closed. In fact, the
closure of any linear subspace H0 ⊂ H is obtained by taking the double-
complement, H0 = (H0⊥ )⊥ .
For closed H0 , there is a uniquely defined projection operator, p = p2 = p∗ ,
whose range is H0 ; the complementary projection 1 − p has range H0⊥ .
The spectrum of T ∈ B(H) is defined to be

σ(T ) := {λ ∈ C : T − λ not invertible in B(H)}.

201
Of particular interest are unitary operators, U ∗ = U −1 , and self-adjoint opera-
tors H = H ∗ . These are examples of normal operators, T ∗ T = T T ∗ , for which
the spectral theorem holds: there exists a unitary transformation V such that
Z
−1
V TV = λ dµ(λ), (19.2)
C

where µ is the projection operator-valued spectral measure of T . This “uni-

tary diagonalizability” is a deep result requiring core theorems of functional
analysis, and the precise meaning of µ is quite involved. In the special case
where H admits a basis of eigenvectors for T , the spectral theorem reduces to
the “discrete spectral decomposition”
∞
X
T = λi |ψi ⟩⟨ψi |, (19.3)
i=i

where λi are the eigenvalues and ψi the corresponding normalized L2 -eigenvectors.

However, many bounded self-adjoint operators (such as multiplication by
bounded real-valued function on L2 ) do not have any L2 -eigenvectors at all,
so Eq. (19.3) does not make sense! The spectrum generally has “continu-
ous” parts which are not accounted for by normalizable eigenfunctions, so Eq.
(19.3) is needed. This is a significant departure from linear algebra, and the
validity of spectral theory hinges on the implicit operator topologies in the
Hilbert/Banach space. In fact, in quantum mechanics practice, one wants to
distinguish true eigenvectors (“bound states”) and approximate ones coming
from the “continuous” part of the spectrum (“scattering states”).

19.2 Unbounded operators

Many realistic and important operators coming from quantum mechanics are
not bounded/continuous. For example, dθ d
can be defined on C ∞ (S 1 ), but the
derivative can blow up (e.g. apply it to ψn (θ) = exp inθ, and take n → ∞).
A bounded operator T can completely specified by how it acts on any
orthonormal basis; the extension to arbitrary Hilbert space elements is then
given by linearity and continuity. This prescription fails for an unbounded
(equivalently, discontinuous) operator T .

202
Closure. For example, dθ d
: C ∞ (S 1 ) → C ∞ (S 1 ) ⊂ L2 (S 1 ) initially makes
d
sense, but we cannot try to “continuously extend” the meaning of dθ to all L2
functions. So derivative operators T are always unbounded, and can only be
defined on some dense linear subspace (such a subspace is not closed in H!),

T : Dom(T ) → H.
| {z }
⊊H

Usually, we start by defining a differential operator T on some nice space

of smooth L2 functions. We may attempt to enlarge the domain of T , thereby
obtaining an extension of T (the range must still land in H, of course). For
example, we could consider the graph of T ,

ΓT = {(ψ, T ψ) : ψ ∈ Dom(T )} ⊂ H ⊕ H,

and take its closure ΓT in H ⊕ H. We say that T is closed if its graph is a

closed subset of H ⊕ H. There is a complication: there is no guarantee that
ΓT will still be the graph of some linear operator (extending T ).

Adjoint. For a linear operator to make sense as a quantum mechanical

Hamiltonian operator, it must be self-adjoint (reality of the spectrum fol-
lows from spectral theory of such operators). For unbounded operators, the
bounded operator definition of adjoint, Eq. (19.1), does not make sense. The
correct generalization which takes Dom(T ) into account is the following.
Definition 66. Let T : Dom(T ) → H be a densely-defined linear operator.
Define Dom(T ∗ ) to be the set of φ ∈ H such that there exists (a unique)
ηφ ∈ H obeying the condition

⟨ηφ |ψ⟩ = ⟨φ|T ψ⟩, ∀ψ ∈ Dom(T ).

Then the adjoint operator T ∗ : Dom(T ∗ ) → H is defined as

T ∗ φ := ηφ , φ ∈ Dom(T ∗ ).

By design, the adjoint’s domain is the maximal one for which the adjoint-
ness condition makes sense on the respective domains of T and T ∗ ,

⟨T ∗ φ|ψ⟩ = ⟨ηφ |ψ⟩ = ⟨φ|T ψ⟩, ∀ψ ∈ Dom(T ), φ ∈ Dom(T ∗ ). (19.4)

Notice that the larger Dom(T ) is, the smaller Dom(T ∗ ) will be.

203
Lemma 19.1. Let T be a densely-defined operator on a Hilbert space H. Then
Range(T )⊥ = ker(T ∗ ), and thus
H = ker(T ∗ ) ⊕ Range(T ). (19.5)
Proof. For any φ ∈ ker(T ∗ ) and f = T ψ ∈ Range(T ),
⟨φ|f ⟩ = ⟨φ|T ψ⟩ = ⟨T ∗ φ|ψ⟩ = 0.
Thus
ker(T ∗ ) ⊂ Range(T )⊥ .
Conversely, suppose f ∈ Range(T )⊥ , so ⟨f |T ψ⟩ = 0 = ⟨0|ψ⟩ for all ψ ∈
Dom(T ). By definition, f ∈ Dom(T ∗ ) with T ∗ f = 0, i.e., f ∈ ker(T ∗ ).
Thus Range(T )⊥ = ker(T ∗ ), and the orthogonal decomposition, Eq. (19.5),
follows.

19.2.1 Self-adjointness
Now that we know what adjoint operators are, we can ponder about self-
adjointness. By default, we only consider densely-defined operators from now
on. First, we say that H is symmetric, or formally self-adjoint, if
⟨Hφ|ψ⟩ = ⟨φ|Hψ⟩, ∀φ, ψ ∈ Dom(H).
By definition, the true adjoint H ∗ will generally have a larger domain than
H. But note that H ∗ is generally not symmetric, because its domain is too
large. What we seek is a genuine equality H = H ∗ , including equality of their
domains, in which case, we say that H is self-adjoint.
So the issue of self-adjointness is the issue of finding an appropriate en-
largement of Dom(H) to Dom(H̃),
Dom(H) ⊂ Dom(H̃) = Dom(H̃ ∗ ) ⊂ Dom(H ∗ ).
| {z }
?

Of course we should also specify how the extended operator H̃ acts on the
extended domain.
As a preliminary step, we state, without proof, some general facts:
ˆ For a symmetric H, the graph-closure ΓH remains the graph of another
symmetric operator, denoted H. The latter operator is called the closure
of H, and ΓH = ΓH .

204
 ˆ H ∗ is automatically closed (i.e. its graph is closed), and (H ∗ )∗ = H.

ˆ (H)∗ = (H ∗ )∗∗ = H ∗ = H ∗ , so we may as well assume that H is a closed

symmetric operator to begin with.

If we are fortunate, H is already defined on a natural domain such that the

closure H̄ is self-adjoint. In this case, H is said to be essentially self-adjoint.
Unfortunately, H̄ will generally fail the following self-adjointness criterion.

Proposition 19.2. A closed symmetric operator H on H is self-adjoint iff

ker(H ∗ ± i) = 0 iff Range(H ± i) = H.

Proof. Evidently, a self-adjoint H = H ∗ cannot have ±i eigenvalues, otherwise

H ∗ ψ = Hψ = iψ ⇒ i⟨ψ|ψ⟩ = ⟨ψ|Hψ⟩ = ⟨Hψ|ψ⟩ = −i⟨ψ|ψ⟩.

Now assume H ∗ has no ±i-eigenvalue. If 0 ̸= ψ ∈ Range(H ± i)⊥ , then for

all φ ∈ Dom(H),

⟨ψ|(H ± i)φ⟩ = 0 ⇒ ψ ∈ Dom(H ∗ ) with 0 = (H ∗ ∓ i)ψ,

contradicting the assumption. Thus Range(H ±i) must be dense in H. In fact,

Range(H ±i) is closed (this is proved in Eq. (20.1) later), so Range(H ±i) = H.
Finally, assume Range(H ± i) = H. So for any φ ∈ Dom(H ∗ ), we can write
(H ∓ i)φ = (H ∓ i)ψ± for some ψ∓ ∈ Dom(H). Since Dom(H) ⊂ Dom(H ∗ ),
∗

we actually have

(H ∗ ∓ i)(φ − ψ∓ ) = 0 ⇔ φ − ψ∓ ∈ ker(H ∗ ∓ i)
⇔ φ − ψ∓ ∈ Range(H ± i)⊥ = 0 (Lemma 19.1).

Thus φ = ψ∓ ∈ Dom(H), i.e., Dom(H ∗ ) ⊂ Dom(H). So H and H ∗ actually

have the same domains, meaning that H is self-adjoint.
In Section 20, we will learn how to find the self-adjoint extensions for H.
But we first discuss the motivating setting of differential operators.

205
19.3 First-order differential operators
Let us begin with first-order differential operators D acting on C ∞ (U ; CN ),
where U is some open domain U ⊂ Rn . So
n
!
X
Dζ = Aj ∂j + B ζ, ζ ∈ C ∞ (U ; CN ), (19.6)
j=1

where Aj , B are some smooth N × N matrix-valued functions on U . The

Fourier transform of D entails the replacement of −i∂j by a momentum-space
variable ηj . If Aj , B were constant over U = Rn , then we could take the
Fourier transform,
n
!
X
Db η ζ̂η = i Aj ηj + B ζ̂η , ζ̂η ∈ CN , η = (η1 , . . . , ηn ).
j=1

That is, Db becomes a family of N ×N matrices parametrized by η ∈ R b n . Here,

j
it is useful to think of η as the cotangent vector ηj dx , and the first-order part
of Db as a “symbol map”,
n
X n
X
j
σD : η 7→ iη( A ∂j ) = i Aj ηj ∈ End(CN ).
j=1 j=1

Of course, we often do not have constant coefficients Aj , B. Moreover,

we are interested in first-order differential operators D acting on sections of
Hermitian vector bundles E over Riemannian manifolds X. This is a linear
map
D : Γ(E) → Γ(E)
such that each point x ∈ X lies in a neighbourhood U on which D has a
local representation as a first-order operator, as in Eq. (19.6). So the function
ζ in Eq. (19.6) is to be regarded as a local representation of ψ, while the
Aj , B are local matrix representations of sections Ãj , B̃ of End(E). Changing
the coordinate chart and/or local trivialization of E will change the matrix
functions Aj , B, so it does not really make sense to say that the latter are
“constant”.

206
Symbol and ellipticity. In particular, we have to remember basepoints
x ∈ X, and consider ∂j |x ∈ Tx X, as well as ηx ∈ Tx∗ X. This means that the
“Fourier transform” of D should really be an object parametrized by T ∗ X.
Therefore, we define the (principal) symbol of D to be the bundle map
σD : T ∗ X → End(E)
n
X
η 7→ i Ãj (x)ηj , η = ηj dxj |x ∈ Tx∗ X.
j=1

We say that D is elliptic if its (principal) symbol σD (·) is invertible for all
non-zero cotangent vectors.
Remark. To see that σD is independent of the choice of coordinate cotangent
basis {dxj }j=1,...,n , write η = df |x for some smooth function f . Then the
calculation
 
n
 X j
([iD, f ]ψ)(x) = i à (x) ∂j |x f  ψ(x) = σD (η)ψ(x), ψ ∈ Γ(E),

| {z }
j=1
ηj

provides an alternative coordinate-free description of σD (η).

Domains. The compactly-supported smooth sections, Γ∞ c (E), serves as a

natural initial domain for D, which is dense in L2 (X; E). Suppose D = H is
a symmetric (first-order) differential operator on Γ∞
c (E).

ˆ The closure H is a closed symmetric operator, and Dom(H) is called the

minimal domain for D.
ˆ At the other extreme, the maximal domain for H is Dom(H ∗ ). Con-
cretely, Dom(H ∗ ) comprises those ψ ∈ L2 (X; E) such that Dψ, taken in
the distributional sense (see Section 20.3), is square-integrable. There are
no further constraints on this maximal domain, because the integration-
by-parts and compact support of Dom(H) ensure that the adjoint rela-
tion, Eq. (19.4), holds.
Example 19.1. By construction, the (twisted) Dirac operator has (principal)
symbol
n
X
σD/ (η) = i ηj ej · = iη · ,
j=1

207
where η = ηj ej is expanded in a local orthonormal cotangent frame; the or-
thonormal frame elements ej act as endomorphisms of the spinors by Clifford
multiplication. Whenever ||η|| = ̸ 0, the endomorphism iη · squares to ||η||2 , so
/ is elliptic. Our calculation in Prop. 17.4 shows that D
it is invertible; thus D /
is symmetric on the initial domain.
Remark. There is a notion of higher-order differential operators acting from
sections of one bundle to another, and the principal symbol is similarly defined.
For example, a Laplacian has principal symbol being scalar, σ∆ (η) = ||η||2 ,
and is thus elliptic.

19.3.1 Essential self-adjointness

In the theory of elliptic PDEs, there is a general principle of elliptic regularity,
whereby weak (distributional) solutions ψ to Dψ = f , for smooth given f ,
boundary conditions on ψ, coefficients of the differential operator D, etc.,
are actually smooth (“strong solutions”). See Example 20.5, for a concrete
instance of this.
To illustrate an application of this principle, consider a closed manifold
X (meaning compact and without boundary), and a symmetric differential
operator H of any order. On the minimal domain, H cannot have any ±i-
eigenvalues, otherwise we would have

∓i||ψ||2 = ⟨Hψ|ψ⟩ = ⟨ψ|Hψ⟩ = ±i||ψ||2 .

On the maximal domain, H ∗ might have some ±i-eigenvalues. A priori, a

±i-eigenfunction in the enlarged domain, Dom(H ∗ ), would not be smooth.
But if H is elliptic, then such an eigenfunction must actually be smooth, so
it already belongs to the minimal domain, contradicting our earlier assertion.
Thus ker(H ∗ ± i) = 0, and we conclude from Prop. 19.2 that H is self-adjoint.
That is, elliptic symmetric operators (e.g., Dirac, Laplace) on closed manifolds
are essentially self-adjoint on the domain of smooth sections.
In general, the question of essential self-adjointness is very difficult. One
sufficent criterion for first-order symmetric D is:

ˆ The Riemannian metric on X is complete. (This is automatic for closed

manifolds.)

208
In this case, all powers of D are also essentially self-adjoint (on Γ∞
c (E)). This
result requires substantial analytic techniques (“mollifiers”), which we will not
develop. A reference for these facts is [8].
There are also criteria for essential self-adjointness of symmetric H which
are of second-order and/or elliptic [17, 3], and these usually depend on the be-
haviour of lower-order parts of H (e.g. potential terms), not just the principal
symbol of H or geometric properties of X alone.
One approach to dealing with the self-adjointness issue is to avoid it al-
together, by always working in situations where essential self-adjointness is
automatic (e.g. closed manifolds). Unfortunately, this “black box” attitude is
not enough to handle many situations of interest to us.
Instead, we shall investigate some “bad” examples, involving non-complete
manifolds, where essential self-adjointness fails. For example, X may have
“boundaries” that compactly-supported interior signals propagated by H can
reach in finite time. Then different self-adjoint extensions of H are obtained
by imposing different boundary conditions on Dom(H ∗ ). It is also possible
that no self-adjoint boundary conditions exist at all.

209
20 Self-adjoint extensions
20.1 Spectrum of symmetric operators
Definition 67. Let T be a closed operator on H. Its resolvent set is

ρ(T ) := {λ ∈ C : (T − λ)−1 : H → Dom(T ) exists and is bounded}.

The bounded operator

Rλ (T ) := (T − λ)−1 , λ ∈ ρ(T ),

is called the resolvent of T at λ. The spectrum of T is

σ(T ) = C \ ρ(T ).

The simplest example of a spectral value is an eigenvalue: If (T − λ)ψ = 0

for some 0 ̸= ψ ∈ Dom(T ), then T − λ is not injective thus (T − λ)−1 cannot
exist. But not all spectral values are eigenvalues!
The spectrum/resolvent set has some basic topological features. For exam-
ple, it may be shown that ρ(T ) is an open subset of C (possibly empty), and
that Rλ (T ) depends analytically on λ ∈ ρ(T ) (see [16] §8 for a proof). Fur-
thermore, from the closed graph theorem in functional analysis, the spectrum
of a non-closed T is all of C. This is why we investigate spectral questions for
closed operators (or implicitly take the operator closure otherwise).
Proposition 20.1. Let H be a closed symmetric operator. Then,
ˆ dim ker(H ∗ − λ) is constant for λ in the open upper half-plane of C;

ˆ dim ker(H ∗ − λ) is constant for λ in the open lower half-plane of C;

ˆ σ(H) is either the closed upper-half plane, closed lower-half plane, the
entire complex plane, or a subset of R. The last case occurs iff H is
genuinely self-adjoint, iff dim ker(H ∗ ± i) = 0.
Proof. (Optional.) Let λ = a + ib ∈ C \ R, so b ̸= 0. Since H is symmetric,
for all ψ ∈ Dom(H),

||(H − λ̄)ψ||2 = ||(H − a)ψ||2 − ib⟨ψ|(H − a)⟩ψ + ib⟨(H − a)ψ|ψ⟩ + b2 ||ψ||2

≥ b2 ||ψ||2 . (20.1)

210
If limn→∞ (H − λ̄)ψn exists (with ψn ∈ Dom(H)), then the above inequality
means that ψn → ψ converges as well. Since H is assumed to be closed,
(ψ, (H − λ̄)ψ) lies in the graph of H − λ̄, with (H − λ̄)ψ = limn→∞ (H − λ̄)ψn .
Therefore, Range(H − λ̄) is a closed subspace of H. By Eq. (19.5),
ker(H ∗ − λ)⊥ = Range(H − λ̄). (20.2)
The next step is to show that a sufficiently small change of λ does not modify
dim ker(H ∗ − λ).
So consider a small perturbation λ+η, with |η| ≪ |b| so that λ+η remains in
the same upper/lower half-plane. Take any unit vector φ ∈ ker(H ∗ − (λ + η)),
and suppose φ ∈ ker(H ∗ − λ)⊥ as well. By Eq. (20.2), we may write φ =
(H − λ)ψ for some ψ ∈ Dom(H), and we have
0 = ⟨(H ∗ − (λ + η))φ |ψ⟩ = ⟨φ| (H − λ)ψ ⟩ − η⟨φ|ψ⟩ = ||φ||2 −η⟨φ|ψ⟩. (20.3)
| {z } | {z } | {z }
0 φ 1

By Eq. (20.1), we also have

||ψ|| ≤ ||(H − λ)ψ||/|b| = ||φ||/|b| = 1/|b| ≪ /|η|,
which is too small for the right-side of Eq. (20.3) to vanish (Cauchy–Schwarz).
This contradiction shows that
ker(H ∗ − (λ + η)) ∩ ker(H ∗ − λ)⊥ = 0, ∀ |η| ≪ |b|.
It follows (exercise) that
dim ker(H ∗ − (λ + η)) ≤ dim ker(H ∗ − λ).
So by taking sufficiently small |η| and swapping the roles of λ + η and λ, we
see that the dimension of ker(H ∗ − λ) is locally constant as a function of λ in
the upper/lower half-plane. By connectedness, ker(H ∗ − λ) is constant on the
upper/lower half-planes.
Note that Eq. (20.1) implies that H − λ̄ is injective, with (H − λ̄)−1 :
Range(H − λ̄) → Dom(H) bounded. By Eq. (20.2), this inverse is defined on
all of H iff 0 = dim ker(H ∗ − λ). So either σ(H) or ρ(H) contains the open
upper half-plane; similarly for the lower half-plane. Since σ(H) is closed, it
has to be one of the four given options.
The fourth option, σ(H) ⊂ R, occurs iff dim ker(H ∗ − λ) = 0 whenever
λ ∈ C \ R, iff dim ker(H ∗ ± i) = 0, iff H is self-adjoint, by Prop. 19.2.

211
 Proposition 20.1 shows that the obstructions to self-adjointness are the
following:
Definition 68. Let H be a (closed) symmetric operator on H. Its deficiency
subspaces and (possibly infinite) deficiency indices are
H± := ker(H ∗ ∓ i), n± := dim H± .

20.2 von Neumann theory of self-adjoint extensions

Let H be closed symmetric, then a self-adjoint extension of H is, in particular,
a closed symmetric extension. Let us first look for closed symmetric extensions
of H.
Introduce the graph inner product,
(φ, ψ)H := ⟨φ|ψ⟩ + ⟨H ∗ φ|H ∗ ψ⟩, φ, ψ ∈ Dom(H ∗ ).
Note that this defines a stronger topology and different notion of orthogonality
than those coming from the Hilbert space inner product; but these notions can
only be applied to Dom(H ∗ ) ⊂ H. The terminology “H-closed, H-orthogonal”
etc., will refer to the graph inner product. A linear subspace L ⊂ Dom(H ∗ ) is
H-symmetric if
[ψ, φ]H := ⟨H ∗ φ|ψ⟩ − ⟨φ|H ∗ ψ⟩ = 0, ∀φ, ψ ∈ L.
For example, Dom(H) is H-closed-and-symmetric.
The importance of these definitions is the following. Suppose H̃ is a ex-
tension of H. Then in terms of domains, we have the hierarchy
H ⊂ H̃ ⊂ H̃ ∗ ⊂ H ∗ .
It is not hard to check that H̃ is closed symmetric iff Dom(H̃) is H-closed-
and-symmetric. So closed symmetric extensions H̃ are obtained by restricting
H ∗ to H-closed-and-symmetric subspaces L ⊂ Dom(H ∗ ).
Lemma 20.2. Let H be a closed symmetric operator. There is an H-orthogonal
decomposition
Dom(H ∗ ) = Dom(H) ⊕H H+ ⊕H H− .
Furthermore, the H-closed-and-symmetric subspaces of H+ ⊕H H− are in one-
to-one correspondence with the H-closed-and-symmetric subspaces lying be-
tween Dom(H) and Dom(H ∗ ), via
L ↔ Dom(H) ⊕H L.

212
Proof. This involves a tedious direct check, which we omit (A full proof can
be found in pp. 138 of [17]).

Theorem 20.3. Let H be a closed symmetric operator. Its closed symmetric

extensions are in one-to-one correspondence with partial isometries U : H+ →
H− .

Proof. Let H̃ be a closed symmetric extension of H. By Lemma 20.2 and

the discussion above it, we have Dom(H̃) = Dom(H) ⊕H L for some uniquely
determined H-closed-and-symmetric subspace L ⊂ H+ ⊕H H− . Consider any
φ = φ+ + φ− ∈ L, then
iφ+ −iφ−
z}|{
0 = [φ, φ]H ≡ ⟨ H ∗ φ |φ⟩ − ⟨φ|H ∗ φ⟩ (H-symmetry of L)
= −2i⟨φ+ |φ+ ⟩ + 2i⟨φ− |φ− ⟩,

so ||φ+ || = ||φ− ||. Thus there is a well-defined isometry L ∩ H+ to L ∩ H− ,

taking φ+ 7→ φ− . Extending-by-zero gives a partial isometry U : H+ → H−
with initial space I(U ) = L ∩ H+ .
Conversely, starting form a partial isometry U : H+ → H− with initial
space I(U ), define

Dom(H̃U ) := {ψ + φ+ + U φ+ : ψ ∈ Dom(H), φ+ ∈ I(U )}, (20.4)

H̃U (ψ + φ+ + U φ+ ) := H ∗ (ψ + φ+ + U φ+ ) = Hψ + iφ+ − iU φ+ . (20.5)

A little thought shows that Dom(H̃U ) is an H-closed-and-symmetric subspace
of Dom(H ∗ ), so H̃U is a closed symmetric extension of H by Lemma 20.2.
In the construction of H̃U , only pairs (φ+ , U φ+ ) ∈ H+ ⊕ H− appear in the
domain. Since φ+ is an i-eigenvector of H ∗ while U φ+ is a (−i)-eigenvector
of H ∗ , the pair (φ+ , U φ+ ) no longer contributes to the ±i-eigenspaces. In
this way, the deficiency subspaces of H ∗ are “killed off in pairs”. It should
therefore be clear that we can completely kill the deficiency subspaces by
choosing a unitary U : H+ → H− . Of course, such a complete pairing can
only be achieved iff n+ = n− (possibly infinite).

Corollary 20.4 (von Neumann theory of self-adjoint extensions). Let H be a

closed symmetric operator with deficiency subspaces H± and deficiency indices
(n+ , n− ).

213
 ˆ H is self-adjoint iff n+ = n− = 0.

ˆ H admits self-adjoint extensions iff n+ = n− (including the case where

both are infinite); the self-adjoint extensions are parametrized by unitary
maps U : H+ → H− , according to the formulae in Eq. (20.4)-(20.5).

Remark. In case n+ = n− = n, the self-adjoint extensions of H are parametrized

by U(n), but not canonically so. This is because the identification of unitary
maps H+ → H− with U(n) requires choosing orthonormal bases for H± .

20.3 Weak derivatives

Let f : Ω → C be a differentiable function, with Ω an open subset of Rn . We
usually think of ∂k f as another function. There are many situations where we
have to deal with functions f which are not differentiable, but only approxi-
mated (in some specified sense) by some sequence of differentiable functions
fn . It is then useful to have a weaker notion of “derivative”, which can be
applied to such non-differentiable f .
The idea is to consider “∂k f ” not as a function, but as a distribution.
Generally speaking, a distribution is a linear functional on a space of test
functions. We will take Cc∞ (Ω) to be the test functions.
Some distributions are represented by actual functions. For instance, con-
sider the space of locally integrable functions L1loc (Ω). This means integrability
over any compact subset of Ω. If f ∈ L1loc (Ω), then it defines a linear functional
on Cc∞ (Ω) via the formula
Z
f : φ 7→ f φ, φ ∈ Cc∞ (Ω).
Ω

Notice that when a function f is serving as a distribution, we only need its

values almost-everywhere.
Some distributions are not representable by any (almost-everywhere de-
fined) function, the most famous one being the delta distribution δx0 , x0 ∈ Ω,
Z
δx0 : φ 7→ φ(x0 ) = “ δx0 (x)φ(x) ”, φ ∈ Cc∞ (Ω).
Ω

The quotation marks indicate that δx0 is not actually a function, and its “in-
tegration against φ” is just a formal expression.

214
 If f is classically continuously differentiable so that ∂k f is a genuine con-
tinuous function, then ∂k f would be the following distribution,
Z Z
∂k f : φ 7→ (∂k f )φ = − f · ∂k φ.
Ω Ω

In the last equality, we used integration-by-parts and compact support of φ.

If f ∈ L1loc (Ω), then “∂k f ” may not exist as a function, but it makes sense as
a distribution,
Z
“ ∂k f ” : φ 7→ − f · ∂k φ, φ ∈ Cc∞ (Ω). (20.6)
Ω

By construction, Eq. (20.6) generalizes the classical derivative.

Example 20.1. Let f : (−1, 1) → C be the discontinuous function,
(
1, x ≥ 0
f (x) =
0, x < 0.
For any φ ∈ Cc∞ (−1, 1), we have
Z 1 Z 1
′
− fφ = − φ′ = φ(0).
−1 0

Thus, the distributional derivative of f is the delta distribution δ0 at x = 0.

Example 20.2. Let f : (−1, 1) → C be the function f (x) = |x|, which is
only almost-everywhere differentiable. It is straightforward to check that the
distributional derivative of f is the sign function.
Example 20.2 is slightly better-behaved than Example 20.1, in the sense
that the distributional derivative is represented by a (almost-everywhere de-
fined) function in L1loc . In such cases, we use the term weak derivative for
emphasis.

20.3.1 Absolutely continuous functions

A function ψ : [a, b] → C is absolutely continuous if, given any ϵ > 0, there ex-
ists δ > 0 such that for any finite collection
P of disjoint intervals (xi , x′i ) ⊂ (a, b)
with total length less than δ, one has i |ψ(x′i ) − ψ(xi )| < ϵ. Absolutely con-
tinuous functions are, in particular, uniformly continuous. We write AC[a, b]
for the set of absolutely continuous functions on [a, b].
The Fundamental Theorem of (Lebesgue integral) Calculus is:

215
Theorem 20.5. Let ψ ∈ AC[a, b]. Then ψ is almost everywhere differentiable,
and its almost-everywhere defined derivative ψ ′ satisfies
Z x
′ 1
ψ ∈ L (a, b), ψ(x) = ψ ′ (x̃) dx̃.
a
Rx
Conversely, let f ∈ L1 (a, b). Then F (x) := a
f (x̃) dx̃ is absolutely continu-
ous, with F ′ = f almost everywhere.
Suppose ψ, φ ∈ AC[a, b], then ψφ ∈ AC[a, b] as well, and (ψφ)′ = ψ ′ φ+ψφ′
almost everywhere, and is L1 . Integrating over [a, b] gives the integration-by-
parts rule, Z b Z b
ψφ′ = ψ(b)φ(b) − ψ(a)φ(a) − ψ ′ φ.
a a
Remark. If ψ ∈ AC[a, b], is its almost-everywhere defined (classical) derivative
the same thing as its distributional/weak derivative? By definition,
Z b Z b
′,dist ′
ψ : φ 7→ − ψφ = ψ ′ φ, φ ∈ Cc∞ (a, b),
a a

where we used integration-by-parts (φ is AC, and vanishes at x = a, b). Thus

ψ ′ , considered as a distribution, coincides with ψ ′,dist . Because of this, we
refer to the almost-everywhere defined derivative ψ ′ ∈ L1 (a, b) as the weak
derivative of ψ ∈ AC[0, 1].
Now suppose ψ ∈ L1 (a, b) has weak derivative ψ ′ ∈ L1 (a, b). Then the
indefinite integral Ψ of ψ ′ is absolutely continuous. Because Ψ and ψ have the
same weak derivative, they coincide (almost everywhere) up to a constant. So
as an L1 equivalence class, ψ may be identified with its absolutely continuous
representative.

20.4 Example: momentum operator on an interval

d
For convenience, we will also refer to −i dx as a derivative operator.
d
Example 20.3. Let us consider the weak derivative operator T = −i dx on

Dom(T ) = {ψ ∈ AC[a, b] : ψ ′ ∈ L2 (a, b)}.

We have
AC[a, b] ⊂ L2 (a, b) ⊂ L1 (a, b),

216
(e.g. by Hölder’s inequality), so by Remark 20.3.1, we can actually write
Dom(T ) = {ψ ∈ L2 (a, b) : ψ ′ ∈ L2 (a, b)} =: W 1,2 (a, b),
Here, W 1,2 (a, b) is an example of a Sobolev space, equipped with the Sobolev
norm,

1/2
||ψ||1,2 = ||ψ||2L2 + ||ψ ′ ||2L2 .
Sometimes, W 1,2 is written H 1 , but we use the former notation, since H is
already being used for “Hamiltonian”.
As an exercise in analysis, we may check that T is a closed operator. In
d
fact, it is the closure of the classical derivative operator −i dx defined initially
∞
on C [a, b].
Let us compute the spectrum of T . For each λ ∈ C, the smooth func-
tion eiλ(·) belongs to Dom(T ). Furthermore, it is an eigenfunction of T with
eigenvalue λ,
T (eiλx ) = λ eiλx .
Thus the spectrum of T is the entire complex plane, and T is definitely not
self-adjoint.
d
Example 20.4. Let H be the classical derivative operator −i dx on the initial
domain
Dom(H) = Cc∞ (a, b).
So functions in Dom(H) vanish at the end points. Similar to Example 20.3,
the closure H is the weak derivative operator on the domain
Dom(H) = {ψ ∈ W 1,2 (a, b) : ψ(a) = 0 = ψ(b)}.
Integration-by-parts shows that H is a symmetric operator: for ψ, φ ∈ Dom(H) ⊂
AC[a, b],
Z b Z b
′
⟨ψ|Hφ⟩ = ψ(−iφ ) = −iψ ′ φ = ⟨Hψ|φ⟩.
a a
∗
Now consider the adjoint operator H = H . Suppose ψ ∈ Dom(H ∗ ). By
∗

definition, there exists some H ∗ ψ ∈ L2 (a, b) such that

⟨H ∗ ψ|φ⟩ = ⟨ψ|Hφ⟩, ∀φ ∈ Dom(H).
Equivalently,
Z b Z b Z b
′
H ∗ ψφ = −i ψφ = − −iψφ′ , ∀φ ∈ Cc∞ (a, b).
a a a

217
The right side is the distributional derivative of −iψ applied to φ. The left
side is the distribution represented by the L2 -function H ∗ ψ. Thus the distri-
butional/weak derivative of ψ is precisely H ∗ ψ ∈ L2 (a, b), and we see that
ψ ∈ W 1,2 (a, b).
Conversely, suppose ψ ∈ W 1,2 (a, b), so ψ is AC with L2 weak derivative.
When integrating by parts, the condition ψ(a) = 0 = ψ(b) for ψ ∈ Dom(H)
ensures that no boundary terms will appear. So ψ is allowed in Dom(H ∗ ) with
H ∗ ψ being its weak derivative.
In summary, H ∗ is the weak derivative −i dxd
on the maximal domain
Dom(H ∗ ) = W 1,2 (a, b),
which is the operator T from Example 20.3.
Example 20.5. For simplicity, we shall consider the interval (−π, π). Example
d
20.4 shows that H = −i dx is symmetric but not self-adjoint on

Dom(H) = {ψ ∈ W 1,2 (−π, π) : ψ(−π) = 0 = ψ(π)} =: W01,2 (−π, π).

Let us use the von Neumann theory to find the self-adjoint extensions of H.
First, we work out the deficiency subspaces H± = ker(H ∗ ∓ i). Suppose
ψ ∈ Dom(H ∗ ) lies in H+ . So ψ is absolutely continuous, and satisfies
−iψ ′ − iψ = 0. (20.7)
Then ψ ′ is also absolutely continuous. Differentiate Eq. (20.7) again, and
deduce similarly that ψ ′′ is absolutely continuous. Inductively, ψ is actually
smooth, and satisfies ψ ′ = −ψ (ordinary derivative). Similarly for H− . There-
fore,
H± = span{x 7→ e∓x },
and the deficiency indices are (1, 1). So there is a family of self-adjoint exten-
sions of H, parametrized by the maps H+ → H− . Noting that φ+ : x 7→ e−x
and φ− : x 7→ ex have the same norm on L2 (−π, π), we see that unitary maps
Uα : H+ → H− are simply labelled by a phase factor α ∈ U(1),
Uα : φ+ 7→ αφ− .
Correspondingly, the self-adjoint extensions Hα of H are the restrictions of
H ∗ = −i dx
d
(the weak derivative) to

Dom(Hα ) = {ψ + β(φ+ + αφ− ) : ψ ∈ W01,2 (−π, π), β ∈ C}.

218
 Observe that if f = ψ + β(φ+ + αφ− ) ∈ Dom(Hα ), then

f (−π) = β(eπ + αe−π ),

f (π) = β(e−π + αeπ ),

so f (π) = γα f (−π) where

e−π + αeπ 1 + αe2π

γα = = α ∈ U(1).
eπ + αe−π 1 + αe2π
Thus the parameter α corresponds to the quasi-periodicity phase parameter
γα .
We had already found the operators Hα of Example 20.5, in Section 3.3.4,
by less rigorous methods. The precise domain comprises not just the smooth
functions obeying a quasiperiodicity condition, but also absolutely continuous
ones (with square-integrable weak derivative).

20.4.1 Momentum operator on a circle

The self-adjoint operator Hα=1 acts on the domain of periodic W 1,2 functions.
Writing S 1 = [−π, π]/−π∼π , this domain is equivalently

{ψ ∈ L2 (S 1 ) : ψ ′ ∈ L2 (S 1 )} ≡ W 1,2 (S 1 ),

a Sobolev space on the circle.

Note that although L2 (S 1 ) ∼= L2 (−π, π) as Hilbert spaces, to make sense of
the weak derivative on S 1 , we need to work with overlapping coordinate charts,
not just on (−π, π). For example, the initial domain Cc∞ (S 1 ) = C ∞ (S 1 ) is not
the same as the initial domain Cc∞ (−π, π). The adjoint of the classical deriva-
tive operator −i dxd
on C ∞ (S 1 ) is the weak derivative operator on W 1,2 (S 1 ).
On S , the existence of the weak derivative over S 1 already implies continuity
1

at θ = π.
d
To summarize, the momentum operator −i dθ on the circle is already es-
sentially self-adjoint on C ∞ (S 1 ). Its closure D is the weak derivative on the
domain
Dom(D) = W 1,2 (S 1 ).
The function space W 1,2 (S 1 ) has a concrete description in terms of the Fourier
transform.

219
 There is a Hilbert space isomorphism

ℓ2 (Z) → L2 (S 1 )
√
δm 7→ wm , wm (θ) = eimθ / 2π.

1 , 0 . . .), and the assignment δm 7→ wm is a map of

Here, δm = (. . . , 0, |{z}
m-th
orthonormal basis elements. A general element ψ ∈ L2 (S 1 ) is expandable as
(an L2 -convergent sum) X
ψ= ψ(m)w
b m,
m∈Z

where the Fourier coefficients are computed as

Z
1
ψ(m) = ⟨wm |ψ⟩L2 (S 1 ) = √
b e−imθ ψ(θ) dθ.
2π S 1

If ψ ∈ Dom(D) = W 1,2 (S 1 ), then the Fourier coefficients of Dψ = −iψ ′

are

Dψ(m)
d = ⟨wm | − iψ ′ ⟩L2 (S 1 ) = ⟨−iwm
′
|ψ⟩L2 (S 1 ) = m⟨wm |ψ⟩L2 (S 1 ) = m · ψ(m).
b

Here, we used integration-by-parts and the periodicity of wm , ψ when regarded

as functions on (−π, π). So the Fourier transform of D is the “position oper-
ator” Q on ℓ2 (Z), with domain
X
Dom(Q) = {ψb ∈ ℓ2 (Z) : (1 + m2 )|ψ(m)|
b 2
< ∞}, (20.8)
m∈Z

and action
(Qψ)(m)
b = m · ψ(m),
b ψb ∈ Dom(Q).
Note that Eq. (20.8) is the largest domain on which Q can make sense as
an operator mapping into ℓ2 (Z). Eq. (20.8) is often used as the definition of
the Sobolev space W 1,2 (S 1 ).
We have found that D is unitarily equivalent to a (self-adjoint, unbounded)
multiplication operator. The spectrum of a multiplication operator is the
(essential) range, and this is Z in this case.

220
20.5 Example: momentum operator on line
The analysis is similar with (a, b) replaced by the unbounded real line R. We
say that ψ : R → C is absolutely continuous, denoted ψ ∈ AC(R), if it is abso-
lutely continuous on any compact interval. (This is sometimes called “locally
absolutely continuous”.) In this case, ψ is almost-everywhere differentiable,
with ψ ′ ∈ L1loc (R). If ψ ∈ L1loc (R) has a weak derivative (in L1loc (R)), then
we may consider ψ to be represented by an absolutely continuous function.
Integration-by-parts works as in the bounded interval case, with these modifi-
cations.
d
The classical derivative H = −i dx is symmetric on the initial domain
∞
Cc (R). The maximal domain is the Sobolev space

Dom(H ∗ ) = {ψ ∈ L2 (R) : ψ ′ ∈ L2 (R)} =: W 1,2 (R).

Note that if ψ ∈ W 1,2 (R), then integration-by-parts shows that

Z x
(ψψ ′ + ψ ′ ψ) + |ψ(0)|2 = |ψ(x)|2 ,
0

and the left side has a finite limit as x → ∞, by square-integrability of ψ, ψ ′ ;

similarly for x → −∞. So limx→±∞ |ψ(x)| exists and it must be zero for
square-integrability of ψ to hold.
The fact that ψ ∈ Dom(H ∗ ) has limx→±∞ ψ(x) = 0 implies that H ∗ is
symmetric — the “boundary terms at infinity” vanish. Therefore H ∗ is self-
adjoint, i.e., H ∗ = H. In other words, H is essentially self-adjoint on Cc∞ (R).

20.5.1 Momentum operator on half-line

d
Now consider the half-line R+ = (0, ∞). The maximal domain of −i dx is the
Sobolev space W 1,2 (R+ ) (defined in a similar way), but this is no longer a
domain of self-adjointness. This is because ψ ∈ W 1,2 (R+ ) may have ψ(0) ̸= 0,
and contribute a boundary term violating the symmetry condition.
The deficiency subspace H+ is spanned by the function x 7→ e−x , while
H− is trivial (because ex is not in L2 (R+ )). Since the deficiency indices are
unequal, the symmetric operator H = −i dx d
on the initial domain Cc∞ (R+ ) has
no self-adjoint extensions.
d d
/ = −i dx d
Similarly, the deficiency indices of +i dx are (0, 1). Thus D ⊕i dx has
deficiency indices (1, 1), and therefore a U(1)-family of self-adjoint extensions.

221
It is easy to see that the domains of self-adjointness are
/ α ) = {(ψ1 , ψ2 ) ∈ W 1,2 (R+ ; C2 ) : ψ2 (0) = αψ1 (0)},
Dom(D α ∈ U(1).

As a general fact (exercise), if H is self-adjoint on Dom(H), then for any

bounded self-adjoint V , the operator H + V remains self-adjoint onDom(H). 
0 m
So, for example, we can add to D / α an off-diagonal mass term M = ,
m 0
as in Section 3.3.3, obtaining the self-adjoint massive Dirac Hamiltonians on
R+ . We had seen that the spectrum D / α + M depends on the parameter α,
and in fact exhibits spectral flow (Fig. 2).

20.5.2 Index of domain-wall Dirac operator

In Section 17.1.3, we mentioned that the massive Dirac Hamiltonian on the
real line,  d 
−i dx m
D/ +M = d , m ̸= 0,
m i dx
has an ambiguity in the sign of m. This operator is self-adjoint on W 1,2 (R; C2 ),
and the spectrum is easily found by Fourier transform to be

σ(Dm ) = (−∞, |m|] ∪ [|m|, ∞).

Let us replace m by the function m · sgn(x). This means that each half-
line has a different sign for the mass term. The corresponding massive Dirac
Hamiltonian is called a domain-wall Dirac Hamiltonian,
d
 
−i dx m · sgn
Ddw = d , Dom(Ddw ) = W 1,2 (R; C2 ).
m · sgn i dx

Naı̈vely, it would seem that the spectrum of Dm and Ddw should coincide.
This is almost true, except for a remarkable 0-eigenfunction (exercise),
 
i
ψJR (x) = e−|mx| ,
sgn(m)
called the Jackiw–Rebbi solution. Notice that ψJR has a cusp at x = 0, so it
does not belong to the initial domain of smooth functions. Nevertheless, it is
square-integrable, and has square-integrable weak derivative, so it belongs to
W 1,2 (R; C2 ), and is admissible as an eigenfunction for Ddw .

222
 This is an example of an index phenomenon.
 The
 operator Ddw is odd
0 i
with respect to the grading operator ie1 = . In exactly the same
−i 0
way as the Dirac operators on graded spinor bundles, the operator Ddw can
be rewritten as  
0 (Ddw )−
Ddw =
(Ddw )+ 0
with respect to a basis in which ie1 is diagonal. The usual spectral supersym-
metry argument holds, except for the 0-eigenspace: the kernels of (Ddw )+ and
(Ddw )− may not coincide, and this asymmetry is measured by

Index(Ddw ) = dim ker(Ddw )+ − dim ker(Ddw )− ∈ Z.

We have just seen that the mass function m · sgn leads to an index of ±1 for
the operator Ddw .
Of course, the 0-eigenvalue could actually be a spurious spectral phe-
nomenon, which disappears if the mass function is modified slightly. An index
theorem says that it is stable — the index is actually predicted by the sign
change in the asymptotic values (at x = ±∞) of the mass function, indepen-
dently of the precise form of the function. This is not too hard to prove for
the 1D case of Ddw . As an exercise, convince yourself that the replacement
of m · sgn(x) with a smooth version, such as m · tanh(x), still exhibits a ±1
index.

223
References
[1] Baez, J., Muniain, J.P.: Gauge Fields, Knots, and Gravity. World Sci-
entific, 1994

[2] Bleecker, D., Booß-Bavnbek, B.: Index theory with applications to math-
ematics and physics. International Press, 2013

[3] Braverman, M., Milatovic, O., Shubin, M.: Essential self-adjointness of

Schrödinger type operators on manifolds. Russian Math. Surveys 57(4)
641 (2002)

[4] Choquet-Bruhat, Y., Dewitt-Morette, C.: Analysis, manifolds, and

physics. Part I: Basics. North Holland (1982)

[5] Eguchi, T., Gilkey, P.B., Hanson, A.J.: Gravitation, gauge theories and
differential geometry. Physics Reports 66(6) 213–393 (1980)

[6] Guillemin, V., Pollack, A.: Differential Topology. Prentice-Hall, 1974

[7] Hamilton, M.: Mathematical Gauge Theory. With Applications to the

Standard Model of Particle Physics. Springer, 2017

[8] Higson, N., Roe, J.: Analytic K-homology. Oxford Univ. Press, 2000

[9] Kobayashi, S., Nomizu, K.: Foundations of Differential Geometry. Vol 1.

Wiley, 1963

[10] Lawson, H.B., Michelsohn, M.-L.: Spin Geometry. Princeton Univ.

Press, 1989

[11] Lee, J.M.: Introduction to smooth manifolds. Graduate Texts in Math.

Vol 218, Springer 2013

[12] Milnor, J.: Topology from the differentiable viewpoint. Univ. Press Vir-
ginia, 1965

[13] Naber, G.: Topology, Geometry, and Gauge Fields. Foundations. Texts
in Appl. Math., Springer, 2011

[14] Naber, G.: Topology, Geometry, and Gauge Fields. Interactions. Texts
in Appl. Math., Springer, 2011

224
[15] Nash, C.: Differential Topology and Quantum Field Theory. Acad. Press,
1992

[16] Reed, M., Simon, B.: Methods of Modern Mathematical Physics, Vol. 1.
Functional Analysis. Acad. Press, 1980

[17] Reed, M., Simon, B.: Methods of Modern Mathematical Physics, Vol.
II. Fourier analysis, Self-adjointness. Acad. Press, 1975

[18] Steenrod, N.: The Topology of Fibre Bundles. Princeton Math. Series,
1960

225