Quantum Field Theory

Umut Gürsoy

ITP, Utrecht University

Contents

1 Need for QFT 7

1.1 Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.2 Special Relativity . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.3 Free particle as an example . . . . . . . . . . . . . . . . . . . . . 10

1.3.1 Klein-Gordon equation . . . . . . . . . . . . . . . . . . . . 10

1.3.2 Dirac equation . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4 Fundamental clash between QM and SR . . . . . . . . . . . . . . 13

2 Quantum fields and canonical quantization 15

2.0.1 Second quantization . . . . . . . . . . . . . . . . . . . . . 15

2.1 Scalar field as a prototype . . . . . . . . . . . . . . . . . . . . . . 17

2.1.1 Lagrangian density, lagrangian and action . . . . . . . . . 18

2.1.2 Solution to the Klein-Gordon Equation . . . . . . . . . . 19

2.1.3 Classical Hamiltonian . . . . . . . . . . . . . . . . . . . . 20

2.2 Canonical quantization . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2.1 Real scalar as an example . . . . . . . . . . . . . . . . . . 21

2.2.2 Quantum Hamiltonian . . . . . . . . . . . . . . . . . . . . 22

3 Path integral quantization 25

3.1 Path integrals in quantum mechanics . . . . . . . . . . . . . . . . 25

3.1.1 Path integrals and time ordering . . . . . . . . . . . . . . 28

1
2 CONTENTS

3.1.2 Sources and n-point functions . . . . . . . . . . . . . . . . 29

3.1.3 Transition amplitude between arbitrary states . . . . . . . 30

3.2 Path integrals in QFT . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2.1 Definition by discretization . . . . . . . . . . . . . . . . . 33

3.2.2 Generating function: free case . . . . . . . . . . . . . . . . 36

3.2.3 Wick’s theorem . . . . . . . . . . . . . . . . . . . . . . . . 38

3.2.4 Effective potential . . . . . . . . . . . . . . . . . . . . . . 39

4 Interactions in QFT 42

4.1 Generating function: interacting case . . . . . . . . . . . . . . . . 42

4.2 Perturbative expansion and Feynman diagrams . . . . . . . . . . 43

4.2.1 Feynman rules . . . . . . . . . . . . . . . . . . . . . . . . 43

4.2.2 Example: One point function . . . . . . . . . . . . . . . . 44

4.2.3 Symmetry factor of diagrams . . . . . . . . . . . . . . . . 47

4.2.4 Connected vs. disconnected diagrams . . . . . . . . . . . 49

4.3 Feynman rules in the momentum space . . . . . . . . . . . . . . . 51

5 From theory to experiment: scattering in high energy physics 55

5.1 LSZ reduction formula . . . . . . . . . . . . . . . . . . . . . . . . 56

5.2 Feynman rules for scattering amplitudes . . . . . . . . . . . . . . 58

5.3 Cross sections and decay rates . . . . . . . . . . . . . . . . . . . 60

5.3.1 Example: Real scalar with cubic interaction . . . . . . . . 63

6 An introduction to renormalization in QFT 65

6.1 Loop divergences . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6.2 Counterterms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.3 Renormalization procedure . . . . . . . . . . . . . . . . . . . . . 68

6.4 Renormalization conditions . . . . . . . . . . . . . . . . . . . . . 69

CONTENTS 3

7 Representations of the Lorentz group and particles 73

7.1 Group representations . . . . . . . . . . . . . . . . . . . . . . . . 73

7.2 Lorentz generators and their algebra . . . . . . . . . . . . . . . . 75

7.3 Finite dimensional representations of the Lorentz group . . . . . 76

7.4 Representing the Poincare group in quantum mechanics . . . . . 78

7.5 Wigner’s construction of unitary representations . . . . . . . . . 80

8 Spinors 83

8.1 Weyl and Dirac Lagrangians . . . . . . . . . . . . . . . . . . . . . 83

8.2 Dirac algebra and Lorentz transformations . . . . . . . . . . . . . 85

8.3 Chirality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

8.4 Solutions to the Dirac Lagrangian . . . . . . . . . . . . . . . . . 86

8.5 Spin and Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 89

8.6 Canonical Quantization of Spinors . . . . . . . . . . . . . . . . . 89

8.7 Spinor propagator . . . . . . . . . . . . . . . . . . . . . . . . . . 91

8.8 Path integral Quantization of Spinors . . . . . . . . . . . . . . . 92

9 Vector fields and gauge invariance 94

9.1 Lagrangian for the vector field . . . . . . . . . . . . . . . . . . . 94

9.2 Classical massive vector . . . . . . . . . . . . . . . . . . . . . . . 96

9.3 Lagrangian for the massless vector . . . . . . . . . . . . . . . . . 97

9.4 Gauge invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

9.5 Gauge fixing in the Coulomb gauge . . . . . . . . . . . . . . . . . 99

9.6 Quantum Gauge Field: Canonical Quantization . . . . . . . . . . 100

9.7 Quantum Gauge Field: Path Integral Quantization . . . . . . . . 102

9.8 Photon propagator . . . . . . . . . . . . . . . . . . . . . . . . . . 103

10 Quantum electrodynamics 105

10.1 Coupling photon to matter . . . . . . . . . . . . . . . . . . . . . 105

4 CONTENTS

10.2 Covariant derivatives and ‘gauging a symmetry’ . . . . . . . . . . 107

10.3 Feynman rules for QED . . . . . . . . . . . . . . . . . . . . . . . 108

10.4 Pair annihilation . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

Preliminaries

Content: We will follow these notes for a first introduction to QFT. Our plan
will be to cover the following topics:
• QFT basics (Clash b/w QM and SR)

• Scalar Field (Klein-Gordon Field) as a basic example

• Canonical Quantization

• Particles as unitary irreducible representations of the Lorentz algebra

• Spinors

• Vector fields

• Path integrals

• Correlation functions

• Scattering

• Gauge invariance and QED

Reading material: In addition to my own collection over the years, these notes
are inspired mainly by two QFT textbooks by

• Srednicki, M.

• Schwartz, M

There are other excellent material which you can benefit from:
• Peskin, Schröder

• A. Zee

• Landau & Lifshitz

• Weinberg

• Brown, Ryder, Ramond

5
6 CONTENTS

• D. Tong’s lectures

In general it is a good idea to use multiple sources and compare different ap-
proaches in studying an advanced subject like QFT. There are various different
but equivalent ways to approach an idea in physics. For example you can study
QM using wavefunctions, operators, path integrals etc. and a concept such as
uncertainty in QM has different manifestations in different approaches. It is
essential to know all for a full grasp of the subject. Even though the devel-
opment of QFT from early last century on has always been inspired by high
energy experiments, here I do not cover much experimental material. For these
you may want to follow Particle Physics I and II in the experimental physics
master’s program.

I try to be concise in these notes as we have to move quickly cover this vast
subject in a total of 20 hours. Therefore, I include a paragraph at the end of
each chapter with references for more detailed reading.
Chapter 1

Need for QFT

It is a phenomenological fact that our world is described in terms of ele-

mentary particles at least at the highest energy scales available to us today.
This is for example what we observe everyday in high energy experiments at
CERN. Two of the fundamental theories developed in the 20th century become
applicable at these energy scales: quantum mechanics because we probe ex-
tremely short distance scales, and special relativity because these experiments
reach velocities close to the velocity of light.

The need to upgrade quantum mechanics to quantum field theory where

the basic dynamical variables are relativistic fields as opposed to single particle
Schrödinger wavefunctions stemmed from a profound clash between quantum
mechanics and special relativity. There are various different manifestations of
this clash. Let me first describe it in a heuristic way1 . We know from QM that
at extremely short time scales energy can fluctuate wildly because of the Heisen-
berg’s uncertainty principle ∆E∆t ≥ ~. Then, according to Einstein’s special
relativity these fluctuations in energy can be converted to create new particles
essentially because energy and mass are the same thing: E = mc2 . Therefore
even if you start with a single particle like an electron, when you probe it at very
short distances you will end up with seeing multiple electrons and photons. In
other words, in the quantum theory, the probability of an electron degenerating
into another electron and a photon, and then that photon degenerating into
an electron and an anti-electron, so on so forth is non-vanishing. If probability
of a certain process is non-vanishing in the quantum world and you choose to
ignore this, you end up in a terrible situation where the total probability of all
processes at a given time do not add up to 1, that is to say probability is not
conserved. Mathematically put, time evolution in quantum mechanics will be
non-unitary. Loss of unitarity is the end of the story, it is like non-conservation
of energy. So we are forced to include multi-particle states in our quantum the-
ory in a manner consistent with relativity. As it turns out, these multi-particle
states are well captured in a form consistent with special relativity and unitarity
in terms of dynamical variables called quantum fields which are infinite sums

1 See Landau & Lifshitz and A. Zee’s books for a detailed account.

7
8 CHAPTER 1. NEED FOR QFT

of single-particle creation and annihilation operators weighed by corresponding

wavefunctions.

As I said there are different manifestations of this clash. A more technical

way to phrase it is to compare the way time and space are treated in QM.
Whereas the position of a particle correspond to the eigenvalues of the position
operator X̂i there is no such an operator corresponding to time. Time enters in
QM merely as a label that labels the particular state of the system |ψ, ti in the
Hilbert space. This asymmetry in the way time and space are treated in QM
is in direct clash with special relativity where they form different components
of the same continuum, which is called the space-time. In other words Lorentz
transformations rotate space and time into one another.

1.1 Quantum Mechanics

To explain this conflict in mathematical terms we first state briefly what

we mean by quantum mechanics and special relativity. Axioms of quantum
mechanics are,

(i) State of the system at time t is represented by a vector in the Hilbert space
|ψ, ti ∈ H.
(ii) Observables are described by Hermitian operators Ô .
(iii) Measurement of an observable yields one of the (real) eigenvalues of the
operator Ô.
(iv) Time evolution of a state |ψ, ti is governed by the Schrödinger equation:

i~∂t |ψ, ti = H |ψ, ti (1.1)

In particular for a time-independent Hamiltonian the last equation is solved by

i
|ψ, ti = e− ~ Ht |ψ, 0i , (1.2)

showing that time evolution in QM is unitary. In other words hψ, t1 |ψ, t1 i =

hψ, t2 |ψ, t2 i for any t1 and t2 . Notice that this fact is a direct consequence of
having a single time derivative in the Schrödinger equation.

1.2 Special Relativity

Physics laws should be independent of the change in the inertial reference

¯). For simplicity, we define x0 = ct and xµ = (ct, ~x) where
frame: (ct, ~x) → (ct̄, ~x
µ = 0, 1, 2, 3 defines a space-time point in the Minkowski Spacetime. We have

x0 = −x0 , xi = xi .
1.2. SPECIAL RELATIVITY 9

More generally, we define the metric tensor to raise and lower indices:
xµ = ηµν xν , xµ = η µν xν (1.3)
where we employ the Einstein’s summation convention where repeated indices
are summed. From this equation you can infer that the metric with the upper
indices is inverse of the one with lower indices:


ηµν η νβ = δµβ → η µν = η −1 µν . (1.4)

In Minkowski Spacetime, the metric tensor is given by:

 
−1 0 0 0
0 1 0 0
ηµν = 
0 0
 (1.5)
1 0
0 0 0 1
The inverse metric is defined as:
η µν ηνα = δαµ (1.6)
For an inertial frame, we have the transformation:
xµ → x̄µ
which is achieved through the Poincare transformation:
x̄µ = Λµ ν xν + aµ
where a is a constant translation in space-time and Λ is the Lorentz matrix.

Demanding that the space-time interval is invariant under Poincare trans-

formation
0 0
(x − x0 )2 = (xµ − x µ )ηµν (xν − x ν )
= −c2 (t − t0 )2 + |~x − x~0 |2
~
= (x̄ − x̄0 )2 = −c2 (t̄ − t¯0 )2 + |~x̄ − x̄0 |2
we arrive at the condition that defines the Lorentz transformation matrices.
Λµ ρ Λν σ ηµν = ηρσ (1.7)
A symmetric 4 by 4 matrix Λ can have 10 different entries. Satisfying the
condition above, it can only have 6 different parameters. These are
(i) 3 rotation angles θi
(ii) 3 boosts βi .

One can put equation (1.7) in the matrix form as follows:

ΛT · η · Λ = η , (1.8)
where T denotes transpose. This shows that the inverse Lorentz matrix is given
by
Λ−1 = η · ΛT · η , (1.9)


−1 α α
which in the index notation reads Λ β = Λβ .
10 CHAPTER 1. NEED FOR QFT

1.3 Free particle as an example

To demonstrate the technical conflict between QM and SP let us consider

the hamiltonian for a free non-relativistic particle with mass m,

p~ˆ2
Ĥ = , p~ˆ = −i~∇ .
2m
then the position space Schrödinger equation can be written as
~2 2
i~∂t ψ(~x, t) = − ∇ ψ(~x, t)
2m
where we define the position wavefunction as ψ(~x, t) = h~x|ψ, ti. This equation
is obviously not invariant under Lorentz transformations because there are two
space derivatives and only one time derivative. A Lorentz transformation would
mix space and time and necessitates treatment of time and space on equal
footing.

Well, you may say this is what you should expect because you considered a
non-relativistic Hamiltonian. Then, let us try a special relativistic Hamiltonian,
again for a free particle:
p p~2
H= p~2 c2 + m2 c4 ' mc2 + + ...
2m
which contains the correct physics in the non-relativistic limit. Then, we have
the equation in position space as:
p
i~∂t ψ(~x, t) = −~2 c2 ∇2 + m2 c4 ψ(~x, t) (1.10)
= mc ψ(~x, t) + O(∇2 ) ,
which can be Taylor expanded for a wavefunction varying over space slower than
the scale set by 1/m.

There are two problems with this equation:

(i) It is obviously still not Lorentz invariant as the number of time and space
derivatives are different.
(ii) This is a non-local equation of motion. This is because of the appearance
of infinitely many space derivatives obtained by expanding the square root. If
you are unfamiliar with what this means, think of a case you know better:
ei~a∇ ψ(~x, t) = ψ(~x + ~a, t)
where the infinite number of derivatives shifts the wavefunction in all space
non-locally.

1.3.1 Klein-Gordon equation

Now, let us square the Schrodinger equation to circumvent (ii) (at least):
−~2 ∂t2 ψ(~x, t) = (−~2 c2 ∇2 + m2 c4 )ψ(~x, t) . (1.11)
1.3. FREE PARTICLE AS AN EXAMPLE 11

Let us denote a four-derivative as

 
∂ 1
∂µ = = ∂t , ∇ ,
∂xµ c

also define ∂ 2 = η µν ∂µ ∂ν . Then eq. (1.11) becomes:

 
2 m2 c2
−∂ + 2 ψ(~x, t) = 0 (1.12)
~

This is called the Klein-Gordon equation. We can easily see that it is Lorentz
invariant by changing the frame of reference by xµ → x̄µ = Λµ ν xν . The deriva-
tives ∂µ transform as ∂ µ → ∂¯µ = Λµ ν ∂ ν . Then using (1.7) ∂ 2 = ∂¯2 and the
Klein-Gordon equation remain the same if we identify

ψ̄(x̄) = ψ(x) . (1.13)

A field that transforms with this rule is called a scalar field. Note that this
transformation is trivial in the sense that the value of the field in any given
location remains the same. So equation (1.12) indeed seems to solve our problem
of finding a Lorentz invariant equation of motion.

It is not at all clear, however, whether the field ψ(x), when regarded as
a quantum mechanical wave-function ψ(x) = hx|ψi evolves unitarily in time.
Not clear because the Klein-Gordon equation, containing two time derivatives,
is certainly not of the Schrodinger form i~∂t |ψi = H |ψi which guarantees
unitary evolution for any hermitean H as the solution reads
i
|ψ(t0 )i = U (t0 − t) |ψ(t)i , U (t) ≡ e− ~ Ht . (1.14)

Thus, we still seem to find a clash between Lorentz invariance and axiom (iv)
of QM reviewed above. Inspecting (1.11), you see that a promising solution
to this problem would be to somehow find a Hermitean differential operator
O, different than the Hamiltonian with the square-root in (1.10), that satisfies
O2 = −~2 ∇2 + m2 c4 so that a solution to i~∂t ψ = Oψ would also solve (1.10).
This would be both Lorentz invariant and unitary by construction. No such
O exists, as it should involve a ∇ not contracted with another vector which
results in LHS being a scalar but RHS being a vector. A more sophisticated
thing to do would be to write O = ai ∂i + · · · and impose the coefficients ai
satisfy ai aj = δij and so on. Of course no such collection of ordinary numbers
exist, but a collection of matrices does. This is what we discuss next.

1.3.2 Dirac equation

The problem of constructing such a first order in time, yet unitary, equation
was first addressed successfully by Dirac for the spin 1/2 particles (spinors).
Spinors can be in two distinct quantum states ψ1 = |↑i and ψ2 = |↓i. This allows
the hamiltonian carry an additional spinor index Hab such that the Schrödinger
equation s to i~∂t |ψa , ti = Hab |ψb , ti. Therefore H is now a matrix and could
solve the problem above. The most general Shrödinger equation first order in
12 CHAPTER 1. NEED FOR QFT

the derivatives and all vector indices contracted (to assure rotational invariance)
is then of the form (written in the position basis)


i~∂t ψa (~x, t) = −i~c(αj )ab ∂j + mc2 (β)ab ψb (~x, t) ,

where αi i = 1, 2, 3 and β are matrices with spinor indices. αi are much like
the Pauli matrices σ i but, as we see below, one should allow for a more general
possibility. The Hamiltonian written in a basis independent way is


Hab = cPj αj ab + mc2 (β)ab , (1.15)

and leads to unitary evolution by construction

 i 
i~∂t |ψa , ti = Hab |ψb , ti ⇒ |ψa , ti = e− ~ H t |ψb , ti , (1.16)
ab

if we demand αi and β be Hermitean. To determine these matrices we require

square of this Hamiltonian to give the Klein-Gordon Hamiltonian, which we
know is Lorentz invariant2

(H 2 )ab = c2 Pj Pk (αj αk )ab + mc2 Pj (αj β + βαj )ab + m2 c4 (β 2 )ab

= (p2 c2 + m2 c4 )δab .

The solution is

{αj , αk }ab = 2δ jk δab (1.17a)

j
{α , β} = 0 (1.17b)
2
(β )ab = δab (1.17c)

where {A, B} = AB + BA. In Problem set I, exercise 2 you show that αj , β

should at least be 4 × 4. These are called Dirac γ−matrices.

Now the equation in the position basis can be written succinctly as,


i~γ µ ∂µ − mc2 ψ(~x, t) = 0 , (1.18)

where we suppress the spin indices a for simplicity. This is called the Dirac
equation. It describes propagation of a particle with two spin states, called a
spin-1/2 particle.

In exercise 2 you will also show that the gamma matrices are traceless.
Therefore Hab is a 4×4 matrix with the eigenvalues {E↑ (p), E↓ (p), −E↑ (p), −E↓ (p)}.
This means that along with the desired energy eigenstates for the up and down
spins we get two more negative energy eigenvalues. This leads to instability of
the system as follows. If you now couple the system to external medium, that
is to say, if the Dirac particles can interact with the surroundings e.g. with
electromagnetic waves, then they can lose energy indefinitely.

Dirac sea: To solve this issue Dirac supposed that all the negative eigen-
states are filled. Pauli exclusion principle allows this assumption because, as we
2 This is necessary but not sufficient for Lorentz invariance. We will see later when we

study the representations of Lorentz algebra, that indeed the equation before squaring it is
already Lorentz invariant.
1.4. FUNDAMENTAL CLASH BETWEEN QM AND SR 13

will see later, spin-1/2 particles are necessarily fermions and as Pauli showed
a state of a fermion can be occupied only once. Dirac argues that we do not
observe the Dirac sea, as we only observe the difference in energy in an experi-
ment. As the Dirac sea is always there, before and after experiment, we cannot
observe it. But, it has the following consequence.

If we send in a strong photon, then we can create a hole in the Dirac sea,
as some particles inside the sea will be excited. So, Dirac postulated that for
every fermion with charge e there should be another fermion with charge −e.
For every electron there should be a positron. Dirac’s prediction of the positron
in 1927 was confirmed experimentally in 1932 leading to his Nobel prize.

1.4 Fundamental clash between QM and SR

To recap, we started with the observation that QM and SR are incompatible

with the single particle picture because many particles can be created from a
single one due to uncertainty in energy fluctuations in QM and energy-mass
equivalence in SR. What we managed to obtain above is an example of a single
particle propagation that is Lorentz invariant. But this evolution is unitary
only if you include the negative energy eigenstates! Dirac solved this issue by
assuming an infinite sea of particles, which brings us back to the original issue:
there seems no way you can avoid the presence of many particle states once you
include interactions. In fact, as we will show below, the right solution to the
problem is to find a many particle representation of particle evolution in QM
and SR. That is to say the eigenstates should be combination of many particle
states. This will lead us to the description in terms of quantum fields.

The easiest way to see the clash between QM and SR mathematically, and
the easiest explanation, thereof, for the need for fields, is to recall the second
axiom of QM: Observables are represented by Hermitian operators: ~x, p~ etc.
However, there is no such operator for the time eigenvalue t which is only treated
as a label in the wave-function in QM. Thus, in QM, time (a label) and space (an
operator) is inherently separate and they transform differently under Lorentz
transformatoins making the entire formulation unsuitable.

There are two ways to solve this problem:

(i) Promote time to an operator

(ii) Demote ~x to a label

Let us start with the first and consider time in quantum mechanics also as
an eigenvalue of an operator. Now all space and time are treated democratically
and one can introduce a space-time basis: h~x, t|ψi = ψ(t, ~x). But then how do
we characterize the evolution in time, if time is a basis vector. The answer
14 CHAPTER 1. NEED FOR QFT

t t

τ τ

x y
x y σ

Figure 2: “worldsheet” of an open string parametrized

Figure 1.1: (Left) Worldline of a particlebyparametrized by the
proper time τ and proper proper
distance σ. time τ .
(Right) Worldsheet of an open string parametrized τ and σ.

is provided by special relativity: in terms of the proper time τ . Indeed the

history of a classical special relativistic particle is given by a curve xµ (τ ) in
space-time, that is labelled by the proper time (see left figure 1.1). This is
called the worldline of the particle. Thus the states in QM are labelled as
|ψ, τ i and the Schrodinger equation should be written with respect to τ . The
interactions in this description is given by splitting and joining of the worldlines.
This description is first introduced by Schwinger who used this idea to obtain
the first quantum correction to electron magnetic moment µ = 0.5 + .... This
idea is called the proper time formulation. It turns out to be hard to work with
in the presence of interactions. But this formulation had a great spin-off: by
including another variable σ to parametrize propagation one s the worldline to
a worldsheet (see right figure 1.1), that is, one arrives at the string theory!

Material for further reading: You can read in more detail about the notes
until here in Srednicki, part I sections 1 and 2, albeit presented in a different
order. I also recommend very much the introductory sections of A. Zee’s QFT
book discussing the motivation behind introducing quantum fields.
Chapter 2

Quantum fields and

canonical quantization

In this course we will instead focus on the second method above, that is, we
will demote the position ~x from an operator (or eigenvalue of an operator) to a
label of the quantum state of the system. This second method is literally the
quantum field theory where we demote the position ~x to a label. Hence both
position and the time become merely labels of the quantum state of the system:

|ψ, ~x, ti = |ψ(~x), ti (2.1)

It will be more useful for us to work in the Heisenberg picture where the opera-
tors become time dependent instead of the states. We then define the quantum
field operator in space-time as ψ̂(~x, t) and its time evolution is given by

ψ̂(~x, t) = eiHt/~ ψ̂(~x, 0)e−iHt/~ . (2.2)

I stress that the space dependence, either in (2.1) or in (2.2) is not the same as
the space dependence of the wave function a single particle in the position basis.
Here the quantum state of the system depends on the position irrespective of
the basis it is evaluated in. Equivalently the quantum operators themselves
depends on the position rather than their expectation value as in the single
particle case: h~x| Ô |~xi = O(~x) where the RHS is a function, not an operator.
In quantum field theory we have Ô(x) as an operator instead. In other words,
the Hilbert space of the system is extended infinitely many times; it is now the
infinite tensor product of the single particle Hilbert space at every space point
~x. This point will become more clear in the example below.

2.0.1 Second quantization

In fact, this is not the first time you encounter the use of fields in quantum
mechanics. Non-relativistic quantum mechanics of N identical particles also

15
16CHAPTER 2. QUANTUM FIELDS AND CANONICAL QUANTIZATION

leads to a quantum field, and this formulation of many body quantum mechanics
is called the second quantization.

The Schrödinger equation in the position basis for N particles with equal
mass, moving in potential U (~x) and with interparticle potential V (~x − ~y ) reads,
 
X N  2
 X N XN
~
i~∂t ψ =  − ∇2j + U (~xj ) + V (~xj − ~xk ) ψ
j=1
2m j=1 k=1,j6=k

where the wave function depends on particle positions and time ψ = ψ(~x1 , ..., ~xN , t).
We can write down an abstract Schrödinger equation, cf. eq. (1.1) by defining
creation and annihilation operators at every space point ~x. We introduce an
harmonic oscillator at every ~x: a(~x) annihilates a particle at position ~x, whereas
a† (~x) creates a particle at this position. These creation and annihilation oper-
ators are called quantum fields. Then the commutation relations between the
quantum fields follow from generalisation of the single harmonic oscillator to a
field of harmonic oscillators:

[a(~x), a(x~0 )] = [a† (~x), a† (x~0 )] = 0 (2.3a)

[a(~x), a† (x~0 )] = δ 3 (~x − x~0 ) (2.3b)

Using these creation and annihilation operators, the Hamiltonian can be written
as
Z  
~2 2
H = d3 xa† (~x) − ∇ + U (~x) a(~x) (2.4)
2m
Z
1
+ d3 x d3 yV (x − y)a† (x)a† (y)a(x)a(y) (2.5)
2

and the state as

Z
|ψ, ti = d3 x1 ... d3 xN ψ(x1 , ..., xN , t)a† (x1 )..a† (xN ) |0i . (2.6)

Note that the particle number N only enters in the state of system, eq. (2.6),
neither in the Hamiltonian (2.4) nor in the commutation relations (2.3b). This
means that this formalism is capable of describing any number of states — the
only thing you need to change is the number of creation operators (and the
number of integrals) in (2.6). Obviously the formalism can also be used when
the number N is not fixed but allowed to change in a process. We learn that the
use of quantum fields potentially solves the main problem in combining special
relativity with quantum mechanics, that I alluded to in the first page of chapter
2, namely the possibility to create many particles in sufficiently short increments
of time.

To understand the meaning of the expression (2.4) let us consider a simpler

system of a free non-relativistic scalar field in the second quantized picture:
Z  
∇2
H = d3 x a† (x) − a(x) .
2m
2.1. SCALAR FIELD AS A PROTOTYPE 17

The Fourier transform of the quantum fields a(x) and a† (x) gives the annihila-
tion and creation operators in the momentum space:
Z
d3 x −i~p·~x
ã(~
p) = e a(~x) , (2.7)
(2π)3/2

similarly for a† (x). Then the hamiltonian above can be written in the momen-
tum space as
Z
p|2 †
|~
H = d3 p ã (~
p)ã(~
p) . (2.8)
2m

The meaning is now clear. If we define the number operator N̂ (~ p) = ㆠ(~

p)ã(~
p)
that counts the number of particles carrying momentum p~, then, the hamiltonian
can be written as
Z
H = d3 pE(|~ p|)N̂ (~
p) ,

which is nothing but the operator that measures the total energy in the sys-
tem. To obtain
p a relativistic Hamiltonian, we can take the energy as relativistic
p|) = |~
E(|~ p|2 c2 + m2 c4 ).

In particular, the state with no particles is called the vacuum state and
denoted by |0i. It is defined as

p) |0i = 0,
ã(~ ∀~
p

Other states in the Hilbert space are obtained from the vacuum by acting with
creation operators:

ㆠ(~
p) |0i : one particle state
† †
ã (~ p2 ) |0i :
p1 )ã (~ two particle state ,

etc.

In passing we note that one can infer the statistics of the particle, whether
it is a fermion or a boson from the commutation relations. Commutators
a† (x)a† (x0 ) = a† (x0 )a† (x) result in Bose-Einstein statistics as they require
a symmetric wave-function in (2.6) whereas anti-commutators a† (x)a† (x0 ) =
−a† (x0 )a† (x)) lead to Fermi statistics.

Also, from now on I will be using the natural units in relativistic quantum
fields by setting ~ = c = 1.

2.1 Scalar field as a prototype

The simplest example of a relativistic quantum field is the scalar field we

discussed in section 1.3.1. We will derive most of the basic results of QFT using
this simple example and generalize to spin-1/2 and spin-1 fields later on. Also,
18CHAPTER 2. QUANTUM FIELDS AND CANONICAL QUANTIZATION

from here on I will be denoting the scalar field by φ(x). As discussed above it
transforms trivially under Lorentz transformations: φ̄(x̄) = φ(x) where x̄ = Λx
is the transformed frame. In the absence of interactions the relativistic scalar
field obeys the Klein-Gordon equation (1.12) which we reproduce here:


−∂ 2 + m2 φ(x) = 0 . (2.9)

So far there is nothing quantum about this equation. It just describes propa-
gation of a classical scalar field. One strategy to obtain a quantum relativistic
scalar field is to promote this classical field to a quantum field φ(x) → φ̂(x).
But how do we do this?

Actually, what we need is precisely the relativistic version of the quantum

fields a(x), a† (x) discussed in the previous section. Obviously, we should first
ask the question, what is the Hamiltonian associated with this field φ. To obtain
it we will first determine the lagrangian that leads to equation (2.9) and then
make a Legendre transformation to arrive at the hamiltonian.

2.1.1 Lagrangian density, lagrangian and action

In classical mechanics the lagrangian is the the function whose minimization

leads to the equation of motion, and the action is defined as the time integral of
the lagrangian. Here we are dealing with a field φ(x) defined at every space (and
time) point, therefore the relevant quantity to minimize is lagrangian density L.
The total lagrangian is given by the integral over L over the whole space
Z
L(t) = d3 x L(~x, t) , (2.10)

and the action is given by the time integral of the lagrangian

Z Z
S = dt L = d4 x L(~x, t) . (2.11)

As in classical mechanics the lagrangian depends on the field and its deriva-
tivesIn addition, we need a theory that preserves Lorentz transformations, thus
the dependence on the derivatives should be in the relativistic form ∂µ φ(x).

More generally, the action should be invariant under Lorentz transforma-

tions if it leads to Lorentz invariant equations of motion. This means that
the lagrangian density should transform as a scalar under the Lorentz trans-
formations. You can easily see this: the space-time volume element d4 x is
itself invariant because the jacobian associated with Lorentz transformations is
|detΛ| = 1, which follows from (1.8) and (1.5). Then, for S to be invariant
under x → x̄ = Λ · x we need L̄(x̄) = L(x). This is the definition of a scalar.

We can now use the facts i) it is a scalar, and ii) its variation should lead
to (2.9) to write down the most general langangian density as follows:
1 1
LKG (x) = − ∂µ φ∂ µ φ − m2 φ2 + Ω0 , (2.12)
2 2
2.1. SCALAR FIELD AS A PROTOTYPE 19

where Ω0 is a constant. The EOM can be obtained by the Euler-Lagrange

equations. The action is:
Z  
1 1
SKG = d4 x − ∂µ φ∂ µ φ − m2 φ2 + Ω0
2 2
Changing φ(x) → φ(x) + δφ(x), the change in action is:
Z


δSKG = d4 x −∂µ (δφ)∂ µ φ − m2 φ(δφ)
Z Z


= − d x∂µ (δφ ∂ φ) + d4 x (∂µ ∂ µ − m2 )φ (δφ) = 0 .
4 µ

The first total derivative term gives no contribution as we require the fields
and its derivatives go to zero sufficiently facts as |x| → ∞. The second term
then gives us the Klein-Gordon equation, since it should vanish for an arbitrary
variation δφ(x).

2.1.2 Solution to the Klein-Gordon Equation

Most general solution to the Klein-Gordon equation is of the superposition

of plane waves of the form
Z
d3 k  ~ i~k·~x−iwt ~

φ(x) = a(k)e + b(~k)eik·~x+iwt
f (|~k|)
where a and b are, yet arbitrary functions that determine the magnitude of
each individual plane wave and f is a function that we will determine shortly
by demanding Lorentz invariance of the integral measure. This solves equation
(2.9) for the choice q
w(~k) = |~k|2 + m2 . (2.13)
This is called the “mass shell condition” or the “dispersion relation” depending
on the context. Note that there are tho linearly independent solutions with ±w
in the exponent. The one that goes with −w would correspond to a negative
energy particle if we quantize this field by replacing the coefficients a and b by
creation operators. As we see below, the solution of QFT of this problem is to
replace a with an annihilation operator instead.

A specific case is a real scalar field which means that φ(x) ∈ R for all x. This
is what we will assume in the rest of this chapter. The more general, complex
scalar field case will be discussed in the exercises. Demanding φ(x) real, we find
that b∗ (−~k) = a(~k) and f (~k) ∈ R. This yields
Z
d3 k  ~ i~k·~x−iwt ~

φ(x) = a(k)e + a∗ (~k)e−ik·~x+iwt , (2.14)
f (|~k|)
after a simple change of variables. As mentioned above we determine f (|~k|)
by requiring that the integral measure in (2.14) is Lorentz invariant. We will
starting from a measure that we already know to be Lorentz invariant:
kµ →k̄µ
d4 k = dk 0 d3 k −−−−−→ dk̄ 0 d3 k̄ = |detΛ| dk 0 d3 k = dk 0 d3 k,
20CHAPTER 2. QUANTUM FIELDS AND CANONICAL QUANTIZATION

because |detΛ| = 1 by the definition of Lorentz transformations (ΛηΛT = η =⇒

|detΛ|2 = 1). Here we defined k 0 = w for a compact notation. We should now
make sure to impose the on-shell condition (2.13). This is done by placing a
delta function δ(k 2 + m2 ) in the integral. Note that this delta-function itself
is Lorentz invariant so the measure together with it, d2 k δ(k 2 + m2 ) continues
to be Lorentz invariant. But we should also include a Heavyside theta function
Θ(k 0 ) in the measure because w is positive definite by definition (2.13). This
would pick up one of the two solutions of k 2 + m2 = 0 with positive energy.
This theta function is also invariant by itself under the orthochronous subgroup
of Lorentz transformations, see Appendix 10.4. All in all we can change the
integral into:
Z Z Z
d3~k
d3 k dk0 δ(k 2 + m2 )Θ(k 0 ) = ,
2w

where k 2 + m2 = −k02 + |~k|2 + m2 = w2 − k02 . This means that the measure in

(2.14) will be invariant if we choose f to be proportional to w. Our convention
is to chose

˜ = d3 k
dk , (2.15)
(2π)3 2w
as the Lorentz invariant integral measure. The resulting scalar field becomes:
Z  
φ(x) = dk ˜ a(~k)ei~k·~x−iwt + a∗ (~k)e−i~k·~x+iwt
Z  
= dk ˜ a(~k)eik·x + a∗ (~k)e−ik·x . (2.16)

where k · x = k µ xµ = k 0 x0 + ~k · ~x = −wt + ~k · ~x.

2.1.3 Classical Hamiltonian

Having found the lagrangian, we now obtain the hamiltonian for the classical
scalar field via the Legendre transformation. Recall that in classical mechanics
the Legendre transform from the Lagrangian L(qi , q̇i ) of a collection
P of degree
of freedom labelled by i to the hamiltonian is H(qi , pi ) = i pi q̇i − L(qi , q̇i ).
The generalised momenta are defined as pi = δL/δ q̇i . In classical field theory,
these notions are immediately generalised to fields1 :

qi → φ(x) : Canonical Position

δL
pi → Π(x) = : Canonical Momentum
δ φ̇(x)
H(qi , pi ) → H[φ(x), Π(x)] = Π(x)φ̇(x) − L(x) : Hamiltonian density

Using (2.12) we find the canonical momentum for the Klein-Gordon field as
Π(x) = φ̇(x). Then the hamiltonian, that is the integral of the hamiltonian
1 You should think of discretising space by dividing it into little boxes and labelling these

boxes by i.
2.2. CANONICAL QUANTIZATION 21

density over space, is

Z
1 2 
H = d3 x φ̇ (x) + |∇φ(x)|2 + m2 φ2 (x) − 2Ω0 . (2.17)
2
We can now substitute the solution (2.16) and obtain
Z   Z
1 ˜ w a∗ (~k)a(~k) + a(~k)a∗ (~k) − d3 xΩ0 .
H= dk (2.18)
2
In the next section we use canonical methods to quantize this system.

2.2 Canonical quantization

The basic idea of canonical quantization of a field is the same as in quantum

mechanics. In quantum mechanics of a N particles, to quantize canonically
we start with the Hamiltonian formulation of the classical system where the
system is described in terms of the canonical degrees of freedom are the canonical
positions qi and canonical momenta pi . Then we promote these position and
momenta to operators q̂i and p̂i and promote their Poisson brackets

{qi , qj } = 0 = {pi , pj }; {qi , pj } = δij , (2.19)

to commutation relations2 :

[q̂i , q̂j ] = 0 = [p̂i , p̂j ]; [q̂i , p̂j ] = i~ δij . (2.20)

Note that they are equal time commutators. That is, once they are defined at
a given time, say t = 0, their form do not change in any later time t. This is
obvious from the time evolution q̂(t) = exp(iHt)q̂(0) exp(−iHt), etc.

Generalization to canonical quantization of fields is completely straightfor-

ward. Again, one first promotes the canonical degrees of freedom to quantum
operators and the Poisson brackets to (anti-)commutators.

2.2.1 Real scalar as an example

Easiest example is the real scalar field we discussed above. The canonical
position is given by the field φ(x) itself, see eq. (2.16). To promote this to
a quantum operator we need to promote the coefficients to operators acting
on a Hilbert space: a(~k) → â(~k) and a∗ (k) → a† (~k). These are precisely the
creation/annihilation operators defined in section 2.0.1, they destroy and create
a particle of momentum ~k out of the vacuum state. So the generalised position
is the quantum scalar field:
Z  
φ̂(x) = dk ˜ â(~k)eik·x + ↠(~k)e−ik·x . (2.21)

2 or to anti-commutation relations if these degrees of freedom are fermionic.

22CHAPTER 2. QUANTUM FIELDS AND CANONICAL QUANTIZATION

Then one imposes the equal-time canonical commutation relations that gener-
alise (2.20) to fields3 :

[φ(~x, t), φ(~x0 , t)] = 0 = [Π(~x, t), Π(~x0 , t)]

[φ(~x, t), Π(~x0 , t)] = iδ 3 (~x − ~x0 ) . (2.22)

It is a straightforward exercise to substitute (2.21) in the commutators (2.22)

to obtain the commutators between the creation/annihilation operators:

[a(~k), a(~k 0 )] = 0 = [a† (~k), a† (~k 0 )] = 0

[a(~k), a† (~k 0 )] = (2π)3 2wδ 3 (~k − ~k 0 ) . (2.23)

Note that if we had a spin 1/2 system, commutators would become anticommu-
tators when we apply the canonical quantization {a(~k), a† (~k 0 )} = 2w(2π)3 δ 3 (~k−
~k 0 ).

To understand better what φ(x) does as an operator let us act it on the

vacuum state. Defining the one-particle states as

|ki = a† (~k) |0i , (2.24)

we find Z
φ(x) |0i = ˜ −ik·x |ki .
dke (2.25)

This means that the quantum field φ(x) creates q

a superposition of one-particle
states with momentum ~k and energy ω(~k) = |~k|2 + m2 at the space-time
point x with the (free) one-particle wave-function exp{−ikx}. This suggests
that φ(x) |0i behaves as an analog of the position eigenstate |~xi. We can check
this interpretation by calculating the overlap of this state with the one-particle
eigenstate (2.25). One finds

h0| φ(x) |pi = e−ipx , (2.26)

where we used the commutation relations (2.23). This shows that indeed φ(x) |0i
represents position on the Fock space of QFT. We conclude that φ(x) acting on
the vacuum state creates a particle at the space-time point x. Since its position
is fixed, its momentum is arbitrary.

2.2.2 Quantum Hamiltonian

The Hamiltonian of the quantum field is now obtained by promoting the

coefficients a(~k) → â(~k) and a∗ (k) → a† (~k) in the classical Hamiltonian (2.18)
and using the commutation relations (2.23) as
Z
H = dk ˜ ω(~k) a† (~k)a(~k) + (0 − Ω0 )V (2.27)

3 From now on I will drop the hats on the fields as it will be clear from the context whether

they are classical or quantum fields.

2.2. CANONICAL QUANTIZATION 23

where V is the volume of the space and

Z q
1 ∞ d3 k
0 = |~k|2 + m2 , (2.28)
2 −∞ (2π)3

is the energy density of the vacuum. This term arises from using the commuta-
tion relations (2.23)
q in passing from (2.18) to (2.27). Recalling the dispersion
~
relation ω(k) = |~k|2 + m2 , we see that 0 is just the sum over the ground
states energy of harmonic oscillators. The system describes a collection of har-
monic oscillators of every ~k at every space point. Translation invariance then
gives rise to the volume factor in (2.27).

Note that the additive constant in (2.27) depends on the order as and a∗ s we
choose in (2.18) before quantization. This is called the ordering ambiguity and
it shows that the passage from classical to quantum mechanics is not unique.
However, this ambiguity does not show up in energy differences which is what
we typically measure e.g. the energy of an excited state with respect to the
vacuum energy4 . To fix the ordering ambiguity we will use the extra constant
Ω0 and set the ground state energy of the system to zero by choosing Ω0 = 0 .
Then the Hamiltonian becomes:
Z
H = dk ˜ ω(~k) a† (~k)a(~k). (2.29)

The creation/annihilation operators in this Hamiltonian satisfy the commuta-

tion relations (2.23). Defining the more conventional operators by
√ √
a(~k) = 2ω(2π)3/2 ã(~k) , a† (~k) = 2ω(2π)3/2 ㆠ(~k) ⇒ [ã(~k), ㆠ(~k 0 )] = δ 3 (~k−~k 0 ) ,
(2.30)
and recalling the definition dk ˜ = d3 k/2ω(2π)3 we discover that the Hamiltonian
of the Klein-Gordon field is
Z q
H = d3 k |~k|2 + m2 ㆠ(~k)ã(~k). (2.31)

that is precisely the total energy of a free relativistic system. Compare this
with the non-relativistic counterpart (2.8). This is the answer we sought in the
beginning of this chapter: the relativistic Klein-Gordon field upon quantization
indeed describes a collection of relativistic free bosons. Furthermore, not only
the theory is Lorentz invariant but also it describes unitary evolution given by
the Hamiltonian (2.31). The issues associated with a relativistic Schrodinger
equation of a single particle that we discussed below equation (1.10) are all
resolved by enlarging the Hilbert space from a single particle to a field.

Let’s pause for a moment and discuss in more detail how the QFT solved
for us the problem of negative energy eigenstates of a relativistic quantum-
mechanical hamiltonian. If we were to interpret the modes in (2.21) as quantum-
mechanical wave-functions, we would again run into trouble with the negative
energy modes, but now we instead have “positive and negative frequency modes”
whose coefficients are creation and annihilation operators. This is essentially
4 It is important however then quantum fields are coupled to gravity, as the gravitational

force does depend on the energy of the fields they couple to.
24CHAPTER 2. QUANTUM FIELDS AND CANONICAL QUANTIZATION

saying that the would-be negative energy eigenstates are interpreted in QFT as
the absence of positive energy eigenstates. Think of holes in the Dirac see for
an analogy in condensed matter systems. Another way to see this is to compute
the commutator of the hamiltonian in (2.29) with the annihilation and creation
operators a, a† . One finds,

[H, a† (~k)] = ω(~k)a† (~k) , [H, a(~k)] = −ω(~k)a(~k) . (2.32)

So QFT associates frequencies with the wrong sign with annihilation operators.
This extra minus sign is not coming from having negative energies but having
an annihilation operator. We will see later that this is not only a mathematical
trick to hide unwanted minus signs but also has observable consequences. In
particular when we assign an electric charge to particles, the terms with a† and
a in (2.21) will be associated with opposite charged particles. Thus given a type
of particle with a given mass and spin, QFT predicts existence of particles with
same mass and spin but with opposite electric charge. In this simple case of
real scalar field the particle is its own antiparticle. It is important to underline
this fact: when we act on the vacuum with the operator φ̂(x) we annihilate an
infinite collection of particles with all possible momenta. But this should not be
allowed in a theory that conserves charge unless one also creates a particle with
the same charge corresponding to each particle that is annihilated with specific
momentum, or alternatively, one also annihilates an anti-particle with charge
opposite to each particle that is annihilated with specific momentum.
Chapter 3

Path integral quantization

We have discussed the following two important points in the previous chap-
ters:

1. The need to pass from single particle to field to retain both Lorentz in-
variance and unitarity.
2. The method of canonical quantization of a field

In this chapter we introduce another basic tool in QFT, namely its description in
terms of path integrals. This is completely equivalent to the canonical methods,
yet has a great advantage over the former which will become clear when we
discuss interactions in the next chapter. Below, we will first introduce the
concept of path integrals in quantum mechanics, then generalize this to fields.

3.1 Path integrals in quantum mechanics

Consider a quantum mechanical Hamiltonian of a non-relativistic single par-

ticle with mass m in a potential well V in one dimension (for simplicity) with
position q and momentum p. A generic Hamiltonian reads,

P2
H(P, Q) = + V (Q) , (3.1)
2m
where P and Q are the momentum and position operators respectively: P |pi =
p |pi, Q |qi = q |qi. A basic object we want to know is the probability amplitude
for this particle, a.k.a the propagator, to travel from point qi at time ti to qf at
time tf :
hqf , tf |qi , ti i = hqf | e−iH(tf −ti ) |qi i . (3.2)
Now divide the time interval T = tf − ti into N + 1 equal length pieces δt =
T /(N +1). In the end we will take the limit δt → 0, N → ∞ with T = (N +1)δt

25
26 CHAPTER 3. PATH INTEGRAL QUANTIZATION

= constant. Then we can write

Z ∞ N
Y
hqf , tf |qi , ti i = dqj hqf | e−iHδt |qN i hqN | e−iHδt |qN −1 i · · · hq1 | e−iHδt |qi i ,
−∞ j=1
(3.3)
by inserting complete sets of position eigenstates at N points. Consider one of
these pieces now, hq2 | e−iHδt |q1 i and apply the Campell-Baker-Hausdorf formula
1
eA+B = eA eB e− 2 [A,B]+··· .

We have
δt 2 2
e−iHδt = e−i 2m P e−iδtV (Q) eO(δt ) .
We can ignore the O(δt2 ) terms in the δt → 0 limit. Then we can evaluate this
piece by inserting a complete set of momentum eigenstates in between the two
exponentials:
Z ∞
δt 2
hq2 | e−iHδt |q1 i = dp1 hq2 | e−i 2m P |p1 i hp1 | e−iδtV (Q) |q1 i
−∞
Z ∞
dp1 −i δt p21 −iδtV (q1 ) ip1 (q2 −q1 )
= e 2m e e
−∞ 2π
Z ∞
dp1 −iH(p1 ,q1 )δt ip1 (q2 −q1 )
= e e
−∞ 2π
√
where we used hq|pi = exp(ipq)/ 2π. Gathering all, we have,
Z ∞ N
Y YN
dpj ipj (qj+1 −qj ) −iH(pj ,qj )δt
hqf , tf |qi , ti i = dqj e e (3.4)
−∞ j=1 j=0
2π

where we defined q0 = qi and qN +1 = qf . We now take the δt → 0 limit. First,

we can use the definition of the derivative
qj+1 − qj
q̇j = lim ,
δt→0 δt
in this limit. Similarly, the product of exponentials can be written as the expo-
nent of an integral
QN over t. The last crucial step is to realise that the product of
the q integrals j=1 dqj altogether can be understood as a sum over all possible
paths between points qi and qf , see figure 3.1. The same is true for the product
of p integrals. We call this limit as a path integral and denote it by a curly D
symbol:
YN Z ∞
Dq(t) ∝ lim dqj , (3.5)
δt→0,N →∞ −∞
j=1

where the proportionality factor will be fixed in the exercise. It can be rigorously
proven that this limit exists and results in a smooth integral over paths between
two points in the case of a single variable e.g. q(t). In the next section we will
generalise this to a path integral over fields, which can also be proven to exist for
many QFTs (certainly for the theories we consider in these lectures). For more
complicated QFTs such as gauge theories e.g. the theory of strong interactions,
3.1. PATH INTEGRALS IN QUANTUM MECHANICS 27

QCD, validity of the path integral description is not doubted but this is not yet
rigorously proven.

Our final answer for the propagator is then,

Z q(tf )=qf Z R tf
hqf , tf |qi , ti i = Dq(t) Dp(t)ei ti dt{p(t)q̇(t)−H((p(t),q(t))} . (3.6)
q(ti )=qi

Some comments:

• The boundary conditions on the q integral means that we are summing

over all paths in space that starts at qi at time ti and end at qf at time
tf , see figure 3.1.
• There is no boundary condition on the p integral, which means that we
should also be integrating over the initial and final momenta. This, of
course, could not be otherwise, since the Heisenberg uncertainty principle
implies that the initial and final momenta are unfixed if the position are.
• Notice that all reference to operators, Hilbert space etc. all those notions
that we are familiar form canonical quantization disappeared: there are no
operators inside the path integral (3.6). All terms are ordinary functions.

Equation (3.6) is our first path integral formula. We will now derive a simpler
expression by performing the p-path integral in the particular case when H is
at most quadratic in p. To do this we go back to the discrete expression (3.4)
and we perform each of the pj integrals using Gaussian integration. Denote the
exponent by F (p) = f2 p2 + f1 p + f0 and rewrite it as F (p) = f2 (p − p∗ )2 + F (p∗ )
where p∗ is the saddle point F 0 (p∗ ) = 0. Then the result of each p integral is
Z ∞
dp iF (p) 1
e =p eiF (p∗ ) .
−∞ 2π −2iπF 00 (p∗ )
We will be interested in the special case where the coefficient f2 = F 00 (p∗ )/2 is
independent of q just like in the case of non-relativistic QM where f2 = 1/2m.
In this case the multiplicative factor is just a constant and it will contribute
to the definition of the q path integral. We are therefore only interested in the
exponential, which is given by p∗ q̇ − H(p∗ , q) in the δt → 0 limit, where p∗ is
determined from the equation
∂ ∂
(pq̇ − H(p, q)) =0 ⇒ q̇ = H(p∗ , q) .
∂p p=p∗ ∂p
This is precisely the Legendre transformation of the Hamiltonian, which means
that the exponential is precisely the Lagrangian of the theory L(qj , q̇j ). We
therefore arrive at our second expression for the path integral representation
which involves a single path integral over q:
Z q(tf )=qf R tf
hqf , tf |qi , ti i = Dq(t) ei ti dt L(q,q̇) . (3.7)
q(ti )=qi

This tells us that the probability amplitude of a particle at point qi at time ti to

arrive to point qf at time tf is given my summing over all paths between these
two points weighed exponential of the action.
28 CHAPTER 3. PATH INTEGRAL QUANTIZATION

Figure 3.1: Definition of the path integral for a quantum mechanical particle
between initial position qi and final position qf and initial and final times ti and
tf .

3.1.1 Path integrals and time ordering

Path integral formulation is very useful to calculate scattering amplitudes

because it is automatically time ordered. Consider for example the expectation
value of the position operator between two position eigenstates at different times
ti and tf , hqf , tf | Q(t1 ) |qi , ti i where ti < t1 < tf . Here Q is the position operator
in the Heisenberg formalism

Q(t1 ) = eiHt1 Qe−iHt1 .

The state |q, ti is defined as the instantaneous position eigenstate defined as

follows. Given an arbitrary state |ψ, ti, its position basis value is given by

ψ(q, t) = hq, t|ψi = hq|ψ, ti = hq| e−iHt |ψi ⇒ hq, t| = hq| e−iHt

hence
|q, ti = e+iHt |qi . (3.8)
Then we have

hqf , tf | Q(t1 ) |qi , tt i = hqf | e−iH(tf −t1 ) Qe−iH(t1 −ti ) |qi i .

This is now precisely the same form from which we derived the path integral
description. We can use the same trick to divide the intervals t1 − ti and tf − t1
into infinitesimal steps, inserting position and momentum eigenstates etc. This
leads to
Z q(tf )=qf R tf
hqf , tf | Q(t1 ) |qi , ti i = Dq(t) q(t1 ) ei ti dt L(q,q̇) . (3.9)
q(ti )=qi
3.1. PATH INTEGRALS IN QUANTUM MECHANICS 29

Note that the operator is replaced by its eigenvalue in the path integral. Obvi-
ously this can be immediately d to an arbitrary operator
Z q(tf )=qf R tf
hqf , tf | F [Q(t1 ), P (t1 )] |qi , ti i = Dq(t) F [q(t1 ), p(t1 )] ei ti dt L(q,q̇)
,
q(ti )=qi
(3.10)
following the same steps that lead to expression (3.6).

Now consider the situation where two operators inserted at different times
hqf , tf | Q(t1 )Q(t2 ) |qi , tt i. Since the path integral formalism is automatically
ordered in time, this expression can be represented in the path integral formalism
only for t1 > t2 . Introducing the time ordering symbol

A(t1 )B(t2 ) t1 > t2
T A(t1 )B(t2 ) = , (3.11)
B(t2 )A(t1 ) t2 > t1

we find that the path integral automatically yields the time-ordered two-point
or higher point functions:
Z q(tf )=qf R tf
hqf , tf | T Q(t1 )Q(t2 ) |qi , ti i = Dq(t) q(t1 )q(t2 ) ei ti dt L(q,q̇)
. (3.12)
q(ti )=qi

This has the obvious generalization to the product of n arbitrary operators:

Z qf
hqf , tf | T A1 [Q, P, t1 ] · · · An [Q, P, t2 ] |qi , ti i = Dq(t) A1 [q, p, t1 ] · · · An [q, p, tn ] eiS ,
qi
(3.13)
where we use the notation A[Q(t), P (t)] = A[P, Q, t].

3.1.2 Sources and n-point functions

Time ordered correlation functions measure how two or more points in time
are correlated with each other. For example how a measurement at time t1 affects
another measurement at t2 etc. They constitute the basic observables both in
quantum mechanics and in quantum field theory. As we have discussed above the
2-point correlators describe the probability amplitude of propagation from one
point in time to another. This idea can easily be d to n-point correlators. As we
discuss later in the course, the basic observables in the high energy experiments,
the scattering amplitudes are determined by these correlation functions.

The n-point functions can be determined using the generating function. To

do this, we first define the functional derivative:
∂
j(t2 ) = δ(t1 − t2 ) . (3.14)
∂j(t1 )

This is just a generalization of the ordinary derivation δji /δjj = δij to contin-
uous variables. Now we modify the Hamiltonian by adding the source terms:

H(p, q) → H(p, q) − j(t)q(t) − h(t)p(t) ,

30 CHAPTER 3. PATH INTEGRAL QUANTIZATION

and we the propagator in the presence of sources as,

Z qf Z R tf
hqf , tf |qi , ti ij,h = Dq(t) Dp(t)ei ti dt{p(t)q̇(t)−H((p(t),q(t))−j(t)q(t)−h(t)p(t)} .
qi
(3.15)
Now a general n-point correlator of a series of Q and P operators can be obtained
as
δ δ
hqf , tf | T Q(t1 ) · · · P (tn ) |qi , ti i = (−i)n
··· hqf , tf |qi , ti if,h=0 .
δj(t1 ) δh(tn )
(3.16)
So far we introduced the sources j(t) and h(t) as auxiliary functions to effec-
tuate calculation of n-point functions. In fact, a source, has a clear physical
meaning, as an external disturbance on the system. It is an external function
that measures how much the energy of the system is altered by adding a linear
deformation of position and momentum.

3.1.3 Transition amplitude between arbitrary states

In the previous sections we studied the time-ordered transition amplitudes

and n-point correlators between two position eigenstates. How about transition
amplitudes and n-point correlators between arbitrary states? These can be
obtained from inserting complete set of instantaneous position eigenstates in
hψf , tf |ψi , ti i as
Z
hψf , tf |ψi , ti ij,h = dqi dqf ψf∗ (qf )ψi (qi ) hqf , tf |qi , ti ij,h . (3.17)

In particular we will be interested in the vacuum-to-vacuum transition amplitude

which will be precisely the quantity required to describe a scattering process. In
a scattering experiment two beams of particles come from spatial infinity, collide
and produce new particles which then propagate toward spatial infinity again.
In the lab frame the state at infinite past is the vacuum, then stuff happens at
intermediate times and the state at infinite future is again the vacuum. The
corresponding generating function is given by
Z
h0|0ij,h = lim dqi dqf ψ0∗ (qf )ψ0 (qi ) hqf , tf |qi , ti ij,h (3.18)
ti →−∞,tf →+∞

This quantity can be obtained directly from the path integral by employing the
following trick. Suppose that the state at initial time is some generic state |ψi i:
Z
|ψi , ti i = dqi ψi (qi ) |qi , ti i .

The latter ket can be expanded in the energy eigenstates as

X
|qi , ti i = eiHti |qi i = eiEn ti |ni hn|qi i
n=0

where En denote the energy eigenvalues, we used (3.8) and assumed a discrete
energy spectrum for simplicity. Now, instead of taking the initial time to −∞
3.1. PATH INTEGRALS IN QUANTUM MECHANICS 31

as in (3.18) consider taking it to −∞ with a small imaginary component

lim |qi , ti i = |0i h0|qi i ,

ti →−∞(1−i)

where  > 0 and infinitesimal. We managed to single out and project onto the
vacuum state because the energy of the vacuum state is defined to be 0 and all
the excited states have positive energies. We then find,
Z
lim |ψi , ti i = dqi ψi (qi )ψ0∗ (qi ) |0i = ci |0i ,
ti →−∞(1−i)

where ci = h0|ψi i. By the same token we have

lim hψf , tf | = c∗f h0| ,

ti →+∞(1−i)

where cf = h0|ψf i. Now the limits ti → −∞(1 − i) and tf → +∞(1 − i) in
the path integral can be effectuated by replacing the time integral on the real
axis in the action in (3.6) or (3.7) by a contour C that is slightly tilted toward
the lower half plane, see figure 3.2. All in all we arrived at the result1
Z R
h0|0ij = (ci c∗f )−1 Dq(t) ei C dt (L+jq) , (3.19)

where the contour C is given in figure 3.2.

Normalization: There are two ambiguities in normalization of this expres-

sion, first from the unknown coefficients ci and cf , second from the ambiguity
of including an overall constant in defining the path integral Dq itself. We can
fix both by the following physical requirement: the probability amplitude for
the vacuum state to remain vacuum in the absence of any source perturbing it
should be 1, that is h0|0i0 = 1:
Z R
∗ −1
(ci cf ) Dq(t) ei C dt L = 1 .

You may be confused that the LHS seems to depend on the choice of the initial
and the final states in the path integral but there is no inconcistency. There is
also a hidden multiplicative dependence coming from the boundary conditions
on the path integral. In particular for the choice2 which ψi and ψf are the
vacuum state, all the multiplicative factors drop out. Substituting this relation
in equation (3.19) we finally obtain
Z
Z0 [J] R
h0|0ij = , Z0 [J] ≡ Dq(t) ei C dt (L+jq) . (3.20)
Z0 [0]
Here the subscript 0 denotes a non-interacting theory i.e. with an Hamiltonian
that involves only terms up to and including quadratic order in the canonical
variables. Our final result (3.20) is a great simplification because the entire
1 From now on I present the result with only the position source j turned on. This is

immediately generalised to include the momentum source h using (3.6) instead of (3.7).
2 This is what we will assume from now on.
32 CHAPTER 3. PATH INTEGRAL QUANTIZATION

Im t

ϵ
Re t
ϵ
C

Figure 3.2: Definition of the contour C that defines the vacuum-to-vacuum

amplitude.

dependence on the initial and final states in the path integral disappeared in
this expression. They are all absorbed into the normalization of the vacuum
state. All of this happens thanks to the “i trick” we used above by rotating
the time integral infinitesimally into the lower half plane. This projects the
initial and final states in the path integral to the vacuum state.

3.2 Path integrals in QFT

All the notions we introduced in the previous section, path integrals, gener-
ating function, n-point correlators, vacuum-to-vacuum amplitude etc carry over
to quantum field theory. Here, however we define the path integral over field
configurations. This means that the propagation amplitude of a field from a
given field configuration φi (~x, ti ) at initial time ti to another field configuration
φf (~x, tf ) at final time tf is given by summing over all possible field paths in
the intermediate times which start and end on these configurations weighed by
exponential of the action.

Consider for example the free real scalar field. The Hamiltonian density for
this Klein-Gordon field is given by
1 2 1 1
H= Π + (∇φ)2 + m2 φ2 , (3.21)
2 2 2
where Π is the canonical momentum that is given by the variation Π = ∂L/∂ φ̇
of the Lagrangian density
1 1
L = − (∂φ)2 − m2 φ2 . (3.22)
2 2
To the path integral of the previous section to this field we replace q(t) → φ(~x, t),
p(t) → Π(~x, t) and j(t) → J(~x, t)
3.2. PATH INTEGRALS IN QFT 33

With these replacements the propagation amplitudes (3.6) and (3.7) become
Z Z φf R tf
dtd3 x(Πφ̇−H)
hφ(~x, tf )|φ(~x, ti )i = DΠ(x) Dφ(x)ei ti
(3.23)
φi

for tf > ti , recalling that path integral automatically orders in time, and,
Z φf R tf
dtd3 x L
hφ(~x, tf )|φ(~x, ti )i = Dφ(x)ei ti
. (3.24)
φi

Similarly, the generating function in the presence of a source J becomes

Z
4
R
Z0 [J] = Dφei C d x(L+Jφ) (3.25)

where the contour is the same as in figure 3.2. Recall that in this final expression
we do not have to specify the boundary conditions on the φ path integral, as
the infinite past and future time limit with the slight tilt of contour C (see
fig. 4) automatically projects onto the vacuum state regardless of the boundary
conditions in time. Now the the vacuum-to-vacuum amplitude in the presence
of source can be obtained from the generating function by normalizing

Z0 [J]
h0|0iJ = , (3.26)
Z0 [0]

such that the transition between two vacuum states in the absence of any source
perturbing the Hamiltonian is unity.

3.2.1 Definition by discretization

Recall that we defined the quantum mechanics path integral over the par-
ticle trajectories q(t) by dividing the propagation time into infinitesimal time
intervals with length  that we labelled by an index i, and performing the ordi-
nary integrals over the positions qi in these time intervals
Z N Z
Y
Dq(t) = N () dqi , (3.27)
i=1

where N () is a normalization factor which you determined in a problem set:

 N/2
−im
N () = . (3.28)
2π~

This normalization is part of the definition of the path integral measure, makes
sure that the path integral is finite in the limit N → ∞,  → 0 and it is
completely invisible in the observables such as correlation functions as it can be
absorbed in the normalization of the vacuum state.

This procedure is straightforward to generalize to QFT. Here we are inte-

grating over field configurations in space φi (~x) = φ(~x, t(i)) at each time step
34 CHAPTER 3. PATH INTEGRAL QUANTIZATION

t(i) = ti + i with ti being the initial time. The final time is tf = t(N + 1)
as before. We can also define this path integral by discretization, not only in
time, as in the QM example above, but also in space. Imagine that we confine
the space in which the field φ lives into a 3D box with volume V3 = L3 . We
then divide this cube into infinitesimal boxes of volume V /N 3 . Then the path
integral over field configurations is defined similar to above as,
Z YZ
Dφ(x) = dφ(xi ) (3.29)
i

where i labels the infinitesimal boxes of space-time with the total space-time
“volume” V4 = L3 T . We suppress a proportionality constant, analogous to
N () in eq. (3.28) in this definition, which we will not need as it will cancel out
in all the observables, see below.

Let us now work out the vacuum-to-vacuum transition amplitude in the

absence of a source J, Z
Z0 [0] = Dφ(x)eiS (3.30)

in the simple case of a real scalar field with the Klein-Gordon action
Z
1

S=− d4 x ∂µ φ∂ µ φ + m2 φ2 . (3.31)
2
We will eventually set to 1 by normalization of the vacuum state but this will
serve as a nice and simple illustration of the computation. To avoid unnecessary
cluttering let’s assume T = L with the total space-time volume V4 = L4 . Then,
the fields in each space-time box φ(xi ) can be expanded in a discrete Fourier
series as,
1 X −ikn ·xi
φ(xi ) = e φ(kn ) , (3.32)
V4 n
with knµ = 2πnµ /L, nµ denoting a set of four integers. I will closely follow the
discussion in Peskin & Schroder, section 9.2. The Fourier coefficients φ(kn ) are
complex but they satisfy φ∗ (kn ) = φ(−kn ) as φ(xi ) are constraint to be real.
We can treat the real and imaginary parts of φ(kn ) for kn0 > 0 independent
(they become related to one another for negative and positive kn0 .), then we
change variables φ(xi ) → φ(kn ) which is a unitary transformation, see (3.32),
hence with unit Jacobian. Thus we can write
Z Y Z
Dφ(x) = dRe[φ(kn )]Im[φ(kn )] . (3.33)
0 >0
kn

This defines for us the path integral measure. Now, to determine the integrand,
substitute (3.32) in the action (3.31) and obtain
1 X

S=− (m2 + kn2 ) Re[φ(kn )]2 + Im[φ(kn )]2 . (3.34)
V4 0
kn >0

Now we will evaluate the path integral (3.30) using (3.33) and (3.34):
Y Z − Vi (m2 +kn
2
)Re[φ(kn )]2
Z
i 2 2 2
Z0 [0] = dRe[φ(kn )]e 4 dIm[φ(kn )]e− V4 (m +kn )Im[φ(kn )]
0 >0
kn
(3.35)
3.2. PATH INTEGRALS IN QFT 35

These are all Gaussian integrals and we know how to evaluate them when the
real part of the coefficient in the exponent is negative definite:
Z ∞
r
−ax2 π
dx e = , Re[a] > 0 . (3.36)
−∞ a

The exponents in (3.35) are not necessarily satisfy this. However, recall that,
to obtain the vacuum-to-vacuum transition amplitude we needed to integrate
over the contour C̃ in the momentum space (or over the contour C in time)
which is obtained by replacing kn0 → kn0 ei which means m2 + kn2 → m2 + kn2 − i
for infinitesimal  which indeed makes sure that the integrals are convergent for
 > 0. The result is
s s s
Y −iπV4 −iπV4 Y −iπV4
Z0 [0] = 2 2 2 2
= , (3.37)
0
m + kn m + kn n
m2 + kn2
kn >0

where in the last equation we removed the constraint kn0 > 0 noting that the
product is identical for kn0 > 0 and kn0 < 0. This result looks terribly divergent
and indeed it is. To see where the divergence comes from, you can take a
logarithm of (3.37), obtaining a sum over n. Clearly the divergence comes from
the high frequency-high momentum modes n  1. This is our first encounter of
what is called the “UV divergence” in QFT. Later in the class we will see how
to make sense of such divergences by renormalization. For the moment, we will
interpret this expression in the following formal manner.

To understand what (3.37) really is, let us consider a Gaussian integral over
an n-component vector ~v with indices vi :
YZ
I= dvi e−vi Bij vj , (3.38)
i

where B is a n × n symmetric matrix with positive definite eigenvalues. We can

diagonalize the exponent by an orthogonal transformation

vi = Oik ṽk OT BO = Bd = diag(b1 , b2 , · · · bn )

where bi are the eigenvalues of B. This is an orthgonal change of variables with

no Jacobian. We can now perform the Gaussian integrations and obtain
Yrπ − 21
I= ∝ (detB) . (3.39)
i
bi

We have a completely analogous expression in (3.30) and (3.31). To see this

explicitly, do an integration by parts in the action3
Z Z
1
1

S=− d4 xφ(x) −∂ 2 + m2 φ(x) = − d4 yd4 xφ(y)δ 4 (x−y) −∂x2 + m2 φ(x) .
2 2
(3.40)
3 The surface term drops with our usual boundary conditions.
36 CHAPTER 3. PATH INTEGRAL QUANTIZATION

except the vector becomes “continuous” vi → φ(x) now labelled by a continuous

4
index x and the
2 2

matrix B is replaced by a continuous matrix Bij → δ (x −
y) −∂x + m . Therefore we can formally write eq. (3.37) as

− 1
Z0 [0] ∝ det(−∂ 2 + m2 ) 2 (3.41)
This formal, infinite dimensional, continuous determinant of an operator is an
example of what is called the functional determinant detO where O is an opera-
tor mapping functions to functions on space-time. It can be properly defined by
discretization as in (3.37) or by introducing a complete set of eigenfunctions of
the operator O with appropriate boundary conditions and then taking the prod-
uct over all the corresponding eigenvalues. The latter method is exemplified in
Exercise 5.

3.2.2 Generating function: free case

We can evaluate the generating function Z0 [J] in (3.25) directly. This is

most easily done by Fourier transforming to momentum space. Introduce the
Fourier transforms:
Z
R d4 k ikx
φ̃(k) = d4 xe−ikx φ(x) , φ(x) = e φ̃(k) , (3.42)
(2π)4
Z
R d4 k ikx ˜
˜
J(k) = d4 xe−ikx J(x) , J(x) = e J(k) . (3.43)
(2π)4
Substituting these in the action (3.22) we find
Z Z
1 d4 k n ˜ φ̃(−k) + J(−k)
˜
o
S= d4 xL = − φ̃(k)(k 2
+ m 2
)φ̃(−k) + J(k) φ̃(k) .
C 2 C̃ (2π)4
(3.44)
Note that we changed the contour C to C̃ in the integral when we pass from
spatial to momentum
R∞ integration.

In order to preserve the form of the Fourier
transforms −∞ dk 0 exp −ik 0 t as one shifts from t → t exp(−i), one needs
to shift k 0 → k 0 exp{+i}. Therefore the contour in the momentum space is
slightly tilted toward the upper half plane as shown in figure 3.3. We now
Fourier transform the integration variable in the path integral (3.25) from φ(x)
to φ̃(k). This is a unitary transformation hence it has unit Jacobian. We can
now evaluate the path integral
Z R d4 k
i ˜ φ̃(−k)+J(−k)
{−φ̃(k)(k2 +m2 )φ̃(−k)+J(k) ˜ φ̃(k)}
Z0 [J] = Dφ̃(k)e 2 (2π)4 (3.45)

as a Gaussian integral, by making another change of variable4

˜
J(k)
φ̃(k) → φ̃(k) + .
k 2 + m2
The result of the Gaussian integration is
 − 21 ( Z )
i 2 i ˜ J(−k)
d4 k J(k) ˜
2
Z0 [J] = N det[− (k + m )] exp , (3.46)
2 2 C̃ (2π)4 k 2 + m2
4 Again the Jacobian is unity as the RHS is linear in φ̃ with unit coefficient.
3.2. PATH INTEGRALS IN QFT 37

Im k0

k0

ϵ
Re k0
ϵ ˜
C

Figure 3.3: Definition of the contour C̃ in the complex frequency space that
defines the vacuum-to-vacuum amplitude.

where N is a yet unknown multiplicative factor arising from multiplicative am-

biguities in defining the path integral. The determinant in front is independent
of J˜ hence it does not play an important role in calculation of the n-point cor-
relators. It is divergent hence should be defined carefully. We will do this later
in this course. Here we just note that all these ambiguities can be absorbed into
the normalisation of the generating function by writing:
( Z )
i ˜ J(−k)
d4 k J(k) ˜
Z0 [J] = Z0 [0] exp , (3.47)
2 C̃ (2π)4 k 2 + m2

where the contour of the k 0 integral is shown in fig. 5. We can now impose the
same physical requirement we imposed below equation (3.19), that the probabil-
ity amplitude for the system to remain in vacuum in the absence of any source
perturbing it is unity. Then, from equation (3.26) we obtain
( Z )
i ˜ J(−k)
d4 k J(k) ˜
h0|0iJ = exp . (3.48)
2 C̃ (2π)4 k 2 + m2

By a change of variables k 0 → k 0 ei we can put back the contour on the real
axis. Keeping track of the  we find the following expression for the exponent
Z ∞ ˜ J(−k)
˜
J(k)
dk 0 d3 k ,
−∞ k 2 + m2 − i

up to O(). Then, using (3.43) the generating function is rewritten as

 Z 
i 4 4
h0|0iJ = exp d xd yJ(x)∆(x − y)J(y) (3.49)
2

where Z
d4 k eik(x−y)
∆(x − y) = . (3.50)
(2π)4 k 2 + m2 − i
38 CHAPTER 3. PATH INTEGRAL QUANTIZATION

Figure 3.4: Pole structure in the Feynman propagator.

This is called the Feynman propagator. Poles structure in the Feynman propa-
gator is shown in figure 3.4. The time ordered propagator for the Klein-Gordon
field obtained from the generating function is precisely the Feynman propagator:

1 1 ∂ 1 ∂
h0| T φ(x)φ(y) |0i = Z0 [J] = −i∆(x − y) . (3.51)
Z0 [0] i ∂J(x) i ∂J(y) J=0

We can further perform the k 0 integral in (3.50) by using the Cauchy’s integral
theorem
Z
1 f (z)
dz = f (z0 ) ,
2πi C0 z − z0

where the contour encloses the pole z0 . Using the pole structure shown in fig.
6 we find
Z Z
˜ ik(x−y) + iΘ(y 0 − x0 ) dke
∆(x − y) = iΘ(x0 − y 0 ) dke ˜ −ik(x−y) , (3.52)

where Θ is the Heaviside theta function. To arrive at this expression we closed

the contour C0 from above for y 0 > x0 and below for x0 > y 0 .

3.2.3 Wick’s theorem

Working out the same calculation that led to (3.51), we find the following
result for the higher point correlators:

• If there are odd number of φs then the result vanishes when we set J = 0
at the end of the calculation:

h0| T φ(x1 )φ(x2 ) · · · φ(xn ) |0i = 0, for n odd . (3.53)

3.2. PATH INTEGRALS IN QFT 39

• If there are even number of φs then the result is

1 X
h0| T φ(x1 )φ(x2 ) · · · φ(x2n ) |0i = ∆(xi1 − xi2 ) · · · ∆(xi2n−1 − xi2n ) ,
in pairs
(3.54)

where the sum is over all different pairs. The result (3.53) and (3.54) is called
the Wick’s theorem.

3.2.4 Effective potential

The generating function in general can be written as exponential of an object

that is called the effective potential W [J]:

Z[J] = exp(iW [J]) . (3.55)

The meaning of the effective potential becomes clear recalling that the generat-
ing function is, at the same time, the vacuum-to-vacuum transition amplitude,

h0|0iJ = lim h0| e−iH[J]T |0i = lim exp(−i∆E0 [J]T ) = exp(iW [J] − iW [0]) ,
T →∞ T →∞
(3.56)
where
R 4 H[J] is the hamiltonian corrected with the source term: H[J] = H−
d xJ(x)φ(x) and ∆E0 [J] is the change in the energy of the vacuum state
caused by addition of the source term. Therefore the effective potential measures
how much the vacuum energy is perturbed by adding an external source to the
theory.

The source term itself has a clear physical meaning. The generating function
can be written as the expectation value of the source term,
Z R 4 R 4
Z[J] = DφeiS ei d xJ(x)φ(x) = h0| ei d xJ(x)φ̂(x) |0i . (3.57)

The operator in the last expression is reminiscent of coherent states in quantum

mechanics. Recall that a coherent state is a collective excitation in the quantum
system with the minimum uncertainty, hence a coherent addition to the system
that behaves classical as much as the quantum uncertainty permits. The oper-
ator above is precisely the quantum field theory analog of this. Recalling the
expansion of φ in terms of creation and annihilation operators
Z  
φ̂(x) = dk ˜ â(~k)eik·x + ↠(~k)e−ik·x , (3.58)
R

we see that exp i Jφ adds a collective excitation of particle-anti-particle pairs
to the vacuum. This is how we probe the quantum field theory, by adding a
classical like perturbation and see how the system reacts to it by computing the
n-point functions.

The function J(x) then determines the amplitude of this collective excita-
tion. For example we can choose this collective excitation to be localised around
40 CHAPTER 3. PATH INTEGRAL QUANTIZATION

two points in space by taking

J(x) = J1 (x) + J2 (x) ,
where the functions J1 and J2 has support around points x1 and x2 .

Let us determine how this perturbation modifies the energy of the vacuum
state by computing the effective potential W [J]. For a general source, by trans-
forming to Fourier space we have,
Z
1 1
∆W [J] = d4 k J˜∗ (k) 2 ˜ ,
J(k) (3.59)
2 k + m2 − i
where ∆W [J] ≡ W [J] − W [0], J˜ is the Fourier transform of J(x) and I used
J˜∗ (k) = J(−k) that comes from reality of J(x). For a source of the form above
˜
we have J(k) = J˜1 (k) + J˜2 (k) and there are four terms in (3.59) that arise from
˜
expanding J. These are W11 and W22 which correspond to self interaction of
J1 with J1 , self interaction of J2 with J2 and the cross interaction terms W12
and W21 between J1 and J2 . In general we are interested in how the excitations
interact with each other, hence the term we are after is the latter. This is
Z
1 1
∆W12 [J] = d4 k J˜1∗ (k) 2 J˜2 (k) , (3.60)
2 k + m2 − i
This expression makes a significant contribution when k 2 +m2 = 0, that is when
the dispersion relation is satisfied. Thus, we learn that the interactions in QFT
are carried by particles that correspond to the excitations of the field! In addition
to this “on-shell” contribution there are contributions from k 2 6= m2 , which are
smaller. These are the “off-shell” contributions or “virtual excitations”.

To calculate the shift in the energy in the vacuum due to this interaction
let us choose the sources as
J(x) = δ 3 (~x − ~x1 ) + δ 3 (~x − ~x2 ) .
This corresponds to test particles, that are infinitely massive as they are an-
chored to the space points ~x1 and ~x2 . Inserting the fourier transform
Z Z
˜ −ik·x 3 0 0 ~
4
Ji (k) = d xe δ (~x − ~xi ) = dx0 e+ik x e−ik·~xi

in (3.60) and carrying out the x0 integral which sets k 0 = 0, one easily finds
Z ~
1 d3 k e−ik·(~x1 −~x2 )
∆W12 [J] = T , (3.61)
2 (2π)3 |~k|2 + m2
where I ignored the i in the denominator as the real part of the denominator is
positive definite. The multiplicative
R factor T is the total time of the interaction
and it comes from the dx01 integral which factored out exactly.

Comparison to (3.56) shows that the shift in the energy of the vacuum due
to the interaction between the two test particles is (W21 contributes equally)
Z ~
d3 k e−ik·(~x1 −~x2 )
∆E12 [J] = − . (3.62)
(2π)3 |~k|2 + m2
3.2. PATH INTEGRALS IN QFT 41

The ~k integral can be done exactly and one finds

1
∆E12 [J] = − e−m|~x1 −~x2 | . (3.63)
|~x1 − ~x2 |

This is the potential energy of two test particles that are interacting by exchang-
ing φ field excitations.

Two remarks:

1. In the massless limit this is precisely the Coulomb interaction in electro-

statics. In fact, as we will see later in the course, in the quantum theory
of electromagnetism, we find precisely the coulomb interaction which arise
due to exchange of a massless photon between the test particles. This type
of potentials, including the mass term, is called Yukawa potential.
2. We observe that the interaction potential is negative definite! This shows
that, in this particular case of the real scalar field the presence of delta-
function sources lowers the total energy of the system rather than increas-
ing, hence the interaction, which arise from coupling the test particles to
the quantum field φ turns out to be attractive! This last point i.e. the
sign of the potential, depends on the theory under consideration.
Chapter 4

Interactions in QFT

Interaction terms in quantum field theory are given by the terms higher
than quadratic order in the Lagrangian (or Hamiltonian). This can be un-
derstood as follows: As discussed in the previous section the quadratic terms
in the Lagrangian determine the propagator which describe propagation of a
point-like particle from one space time point to another. Particle interactions
on the other hand are described by exchange of another particle between two
particles. In order to exchange a particle the initial particle should emit one.
This emission process necessitates presence of at least 3 particles at the same
space-time point:incoming particle, emitted particle, outgoing particle. This
spacetime point which includes at least three particles is called an interaction
vertex. As the single-particle states are created and annihilated by the quantum
field φ(x), see e.g. (2.21), presence of an interaction vertex requires a cubic or
higher order term in the Lagrangian. These terms are called the interaction
terms.

4.1 Generating function: interacting case

We will now calculate the generating function of an interacting QFT. We

will develop the necessary techniques in the path integral formalism using a
simple example: the real scalar field with a cubic interaction term with the
lagrangian density

1 1 g
L = − (∂φ(x))2 − m2 φ(x)2 + φ(x)3 . (4.1)
2 2 3!

Here g is a constant called the coupling constant and we inserted 3! for simplicity,
as will become clear below. To calculate the generating function of this theory
we will employ the following trick.

One can obtain a formal expression for the generating function with inter-

42
4.2. PERTURBATIVE EXPANSION AND FEYNMAN DIAGRAMS 43

actions using the following trick,

∞ ∞  l !n
X (gq l )n jq X ∂ l
gq l +jq
ejq = eg( ∂j ) ejq .
∂
e = e = g
n=0
n! n=0
∂j

This trick can easily be generalised to functional derivatives (3.14) in place of

ordinary derivatives. Replacing q with φ(x) and j with J(x) we find
l
d4 x(gφ(x)l +J(x)φ(x)) d4 xg ( δJ(x)
δ
) d4 xJ(x)φ(x)
R R R
e =e e .

Now we apply this trick to an interacting QFT with a generic interaction given
by a polynomial in φ(x). Denoting this polynomial Lint and the free, quadratic
part of the Lagrangian L0 we have
Z R 4
R 4 δ
Z[J] = Dφ(x)ei d x(L0 +Lint +Jφ+) = ei d xLint [−i δJ(x) ] Z0 [J] . (4.2)

As we have already calculated Z0 [J] in (3.49) (4.2) is the final expression for the
interacting generating function, at least formally. Unfortunately this expression
cannot be evaluated analytically in most of the cases and we will apply a pertur-
bative expansion to calculate. In particular we will expand the first exponential
in a Taylor expansion when the coupling constants are small. This perturba-
tive evaluation of the generating function is best described in the language of
Feynman diagrams that we turn to in the next section.

4.2 Perturbative expansion and Feynman dia-

grams

4.2.1 Feynman rules

We will evaluate the generating function in (4.2) by expanding the expo-

nentials in (4.2) and (3.49) in a Taylor series. To be definite let us consider a
cubic interaction term Lint = g/3!φ3 . We obtain

∞
X Z 3 !V X∞  Z P
1 ig δ 1 i
Z[J] ∝ d4 x −i 4 4
d yd zJ(y)∆(y − z)J(z) .
V! 3! δJ(x) P! 2
V =0 P =0
(4.3)
where the proportionality constant will be determined below. The letters P
and V stand for the number of propagators ∆ and vertices in this expression.
Now consider a specific term (P, V ) in the double sum. When the functional
derivatives hit the sources the leftover number of Js is given by

E = 2P − 3V . (4.4)

The letter E stands for the number of “external” sources. There are 3V func-
tional derivatives in this expression acting on 2P sources in 2P (2P −1) · · · (2P −
3V + 1) = 2P !/(2P − 3V )! combinations.
44 CHAPTER 4. INTERACTIONS IN QFT

vertex

δJ δJ δJ ⟺

JΔJ JΔJ external

source

Figure 4.1: There is a single type of contractions of δJ s with Js in the case

E = 1, V = 1, P = 2 shown on the left. Multiplicity of these contractions
is 24 as explained in the text. Following these contractions one can draw a
corresponding Feynman graph shown on the right.

An important observation is that most of these combinations will be iden-

tical. Furthermore each of these classes can be represented by a simple graph,
called the Feynman graph. Let us derive this notion by working out an example.

4.2.2 Example: One point function

Consider the O(g) contribution to the one-point function. The one-point

function is obtained by taking a single derivative w.r.t the source of the gener-
ating function(4.3) and setting J = 0 in the end. Therefore, only the diagrams
with E = 1 give a non-trivial contribution. We can organise all these diagrams
with E = 1 according to their number of vertices, that is, we can expand the
one-point function in powers of g V . As obvious from (4.4) there is no contri-
bution for E = 1 and V = 0. The first contribution arises from E = 1, V = 1
which requires P = 2 according to (4.4). Let us calculate the corresponding
term in the generating function
Z  3  Z 2
ig δ 1 i
Z[J] ∝ d4 x −i d4 yd4 zJ(y)∆(y − z)J(z) .
3! δJ(x) 2! 2

Clearly all the contributions in this expression are of the same form, one δ/δJ
hitting one of the J∆J terms and the two other δ/δJ hitting the other J∆J,
see Fig. 4.1.

In Fig. 4.1 we show one possibility of contractions between δJ s and J. In

this example there is a single class of contractions that contribute to the final
result and the example in the figure is a representative of this class. Now,
suppose you go through these contractions starting from the external source
taken in circle in the figure. You first encounter a propagator, ∆, from the
location of the source to the location of the vertex. Then from the second leg of
the vertex, there is another propagator to the third leg of the same vertex. This
reading of the contractions immediately suggests a graphical representation of
the contraction operators which we described, that is shown on the right of Fig.
4.1. To obtain this representation, we denote an external source by a black dot,
4.2. PERTURBATIVE EXPANSION AND FEYNMAN DIAGRAMS 45

1
: i ∆(x − y)

R
: i d4 x J(x)

R
: ig d4 x

Figure 4.2: Feynman rules for the real scalar field with cubic interaction.

a propagator by a single line and a vertex by three legs meeting at a point,

a tri-star. Adding these pieces together in the order you encounter them on
the left diagram you obtain the right diagram in Fig. 4.1. These rules which
we summarised in Fig. 4.2 are the Feynman rules for constructing a Feynman
graph. Obviously there are separate rules per theory. We will see later how to
read off these rules directly from the action.

Let us consider a more complicated example, O(g 2 ) contribution to the two

point function. For this we need E = 2, V = 2. All possible contractions
of δJ s with Js in the corresponding contribution to the generating function is
shown in Fig. 4.3. I also drew the corresponding Feynman graphs in the figure.
You get the idea. Now, all of this suggest a much quicker way to arrive at the
final result. We notice that every distinct class of contracting the functional
derivatives with the sources correspond to a topologically distinct Feynman
graph. Similarly every topologically distinct Feynman graph corresponds to a
distinct contribution to contractions, simply because we introduced these graphs
as a way to represent the contractions, they are in one-to-one correspondence.
We should also know the multiplicities of distinct contractions and it would be
great to have a way to read off these multiplicities from the Feynman graphs
as well. In fact there is a way as I discuss below. Now the prescription is very
clear. Given E external sources, V vertices , and P = (E + 3V )/2 propagators,
one draws all possible ways to connect these “lego pieces” in a graph. If we also
know the multiplicity of these graphs then we are done!

Before going further we make a note that there are two general classes of
graphs, the connected and disconnected Feynman graphs. The only contribution
to the E = V = 1 case was a connected one shown in Fig. 4.1. On the other
hand there were two connected and three disconnected contributions in the
E = V = 2 case, shown in Fig. 4.3. We will see below that the disconnected
graphs will not contribute to the n-point functions, hence we will be able to
ignore them. I show one more example of distinct connected Feynman graphs
46 CHAPTER 4. INTERACTIONS IN QFT

δJ δJ δJ δJ δJ δJ
δJ δJ δJ δJ δJ δJ

JΔJ JΔJ JΔJ JΔJ

JΔJ JΔJ JΔJ JΔJ

⟺
⟺

δJ δJ δJ δJ δJ δJ δJ δJ δJ δJ δJ δJ

JΔJ JΔJ JΔJ JΔJ JΔJ JΔJ JΔJ JΔJ

⟺

δJ δJ δJ δJ δJ δJ

JΔJ JΔJ JΔJ JΔJ

⟺

Figure 4.3: Five distinct contractions and the corresponding Feynman graph
contributions to the two point function at order O(g 2 ) that arise from the E =
2, V = 2 term in the generating function. The symmetry factors of these
contributions read from left to right, top to bottom are 4, 4, 8, 16 and 24.
4.2. PERTURBATIVE EXPANSION AND FEYNMAN DIAGRAMS 47

that contribute to the one-point function at O(g 3 ) in Fig. 4.4.

Now let’s discuss the multiplicities of each separate class of contractions.

Consider the one-point function example above, E = V = 1. The number of
ways we can contract δJ ’s with Js in (4.1) is: 2 (for choosing the J∆J that is
hit by a single δJ ) times 3 (for choosing the δ/δJ that hits that J∆J) times 2
(for choosing which of the Js in that J∆J it hits) times 2 (for the order of the
leftover Js hitting the other J∆J, which gives 24. We note that this almost
cancels the overall coefficient of the relevant term in the partition function, that
is, (2!)P (3!)V V !P ! = 48 but not quite. If you think of the different ways we can
contract δJ s with Js it becomes clear that these two numbers should almost
match. The mismatch, a factor of 2 in this example, arises as follows. Note
that in the final contribution we have 2 instead of 4 because exchanging two
δJ s simultaneously with the two Js in J∆J gives the same permutation. This
mismatch between the actual number of contractions and the overall coefficient
is called the symmetry factor of the diagram and we will study this in more
detail below.

All in all, the result for the one point function which follows from the Feyn-
man rules on the table above, is,
Z
δ −i
h0| O(x) |0i = −iZ[0]−1 Z[J] = g∆(0) d4 z ∆(x − z) . (4.5)
δJ(x) J=0 2

In fact it is better to organise the expansion using the pair (E, V ), that is the
number of external sources and vertices, instead of the pair (P, V ) as in (4.3).
This is because we are eventually interested in computing the n-point correlators
that is obtained from the generating function by taking n functional derivatives
with respect to the sources J and setting J = 0 in the end. Clearly, the only
non-vanishing contributions to these n-point correlators would have n = E. If
we organize the expansion of Z[J] by E and V then we know that we only have
to consider the E = nth order terms (which are further distinguished by the
label V ). Furthermore, we define the perturbative expansion for a small value
of the coupling constant g, which makes possible to expand in the number of
vertices. In order to calculate an n-point correlator then one needs to calculate
all the Feynman graphs with E = n to the desired order in the number of
vertices.

4.2.3 Symmetry factor of diagrams

We will now determine the multiplicity factor in each Feynman diagram. In

a diagram with E external sources, V vertices and P = (E + 3V )/2 propagators
the naive multiplicity factor is V !×P !×(2!)P ×(3!)V . The first factor arise from
swapping the different vertices, the second arises from swapping different prop-
agators and the last two terms arise from swapping the end points in individual
propagator and vertex terms. This nicely cancels out the multiplicative factor
in the double Taylor expansion in (4.3). However this naive result overlooks the
fact that a some rearrangement of the derivatives give the same match-up to
sources as some rearrangement of sources. We have to divide the contribution
48 CHAPTER 4. INTERACTIONS IN QFT

With E = 1, V = 1 ⇒ P = 2

With E = 1, V = 3 ⇒ P = 5

With E = 2, V = 2 ⇒ P = 4

Figure 4.4: Some examples of connected Feynman diagrams with different num-
ber of external sources E and vertices V constructed using the Feynman rules.
Number of propagators follow using the formula (4.4). The top and bottom ex-
amples are worked out in more detail in figures 4.1 and 4.3 respectively, where
in the latter we also showed the disconnected contributions.
4.2. PERTURBATIVE EXPANSION AND FEYNMAN DIAGRAMS 49

Figure 4.5: Some examples of Feynman diagrams with their symmetry factors.

of each diagram by this factor which is called the symmetry factor. Therefore
each Feynman diagram has a symmetry factor associated to it. A useful obser-
vation is that this overcounting i.e. the match-up of derivatives with sources,
is represented by the symmetry of a diagram under the simultaneous exchange
of the legs in a vertex with the propagators that connect to these legs, and a
simultaneous exchange of two vertices and the endpoints of the propagators that
connect these two vertices. Some examples are presented in figure 4.4. As clear
from the explanation above and the examples in figure 4.5, the symmetry factor
is literally the number of permutations of vertices, propagators and external
sources that leave the diagram invariant. In more detail there are the following
types of permutations that contribute to the symmetry factor:

1. Swapping the endpoints of a propagator that starts and ends at the same
vertex. This gives a factor of 2.

2. Permuting the propagators that connect the same two vertices. This gives
a factor of p! if there are p propagators.

3. Permuting the vertices themselves (together with all the propagators and
sources that are connected to them). This gives v! if there are v vertices
that are swapped.

4. Permuting the external sources that are connected to the same internal
vertex (or to each other as in the case of 2 external sources connected to
each other by a propagator). This gives s! when there are s number of
identical source terms connected to the same vertex.

5. Permuting identical disconnected Feynman graphs which gives a factor of

nI ! if there are NI number of “I-type” graphs.

4.2.4 Connected vs. disconnected diagrams

We noted above that they were generally two classes of graphs, connected
and disconnected. The connected diagrams are the ones we can trace a path
50 CHAPTER 4. INTERACTIONS IN QFT

Figure 4.6: Example of a disconnected Feynman diagram with V = 4, E = 2

and P = 7.

through the diagram between any two points on it. A generic diagram, on the
other hand, consists of several disconnected pieces as we have seen above. These
arise in the sum (4.3) as a product of disconnected integrals. Another example
is given in figure 4.6. Therefore the most general diagram can be represented
by the following expression:
1 Y n
D{n1 ,n2 ,··· } = (CI ) I . (4.6)
SD
I

Here we label all possible connected Feynman diagrams with the index I, CI
represent their values including the symmetry factors, nI is the number of the
same diagram I appearing in the product, and SD is the symmetry factor as-
sociated with the product itself. The latter gives nI ! if there are nI identical
diagrams present in the sum. Generally it is given by
Y
SD = nI ! . (4.7)
I

Then the generating function (4.3) can be written in terms of the values corre-
sponding to the Feynman diagrams generically as follows:
!
X X Y (CI )nI X
Z[J] ∝ D{n1 ,n2 ,··· } = = exp CI , (4.8)
nI !
{n1 ,n2 ,··· } {n1 ,n2 ,··· } I I

where we exchanged the sum and the product in the last step and summed up
each of the sums over nI into exponentials. This is a remarkable result: the
generating function in general can be written as the exponential of sum over all
possible connected diagrams.

We are eventually interested in the n-point functions that are derived from
the vacuum-to-vacuum amplitude h0|0iJ = Z[J]/Z[0]. Now, there is a simple
way to express this normalisation condition in the expression above1 . In the
absence of any sources, Feynman diagrams only include the so called vacuum
diagrams, diagrams with no sources, that only involve closed loops, see figure 4.7.
Denoting the set of these vacuum diagrams by {0} the normalisation condition
above means  
X 
Z[0] = exp CI , (4.9)
 
I∈{0}

1 Note
that this is in general a different condition than the previous one h0|0i0 = 1 in (3.48)
because the vacuum state in the presence of interactions is generically different than the free
case. We denote both vacua by |0i with a slight abuse of notation.
4.3. FEYNMAN RULES IN THE MOMENTUM SPACE 51

Figure 4.7: Examples of vacuum diagrams. The symmetry factors from top left
to bottom right are S = 24 , 23 , 24 , 3! × 23 and 4! respectively.

thus we arrive at our final result

 
X
h0|0iJ = exp  CI  ≡ ei∆W [J] , (4.10)
I ∈{0}
/

where we denoted the log of the generating function as W [J], see section 3.2.4,
and ∆W [J] = W [J] − W [0]. We learn that the effective potential is the generat-
ing function for the connected n-point functions. This will be extremely useful in
determining the physical observables such as the scattering amplitudes in QFT.

4.3 Feynman rules in the momentum space

The basic observables in a QFT, the n-point functions can be calculated

much easily in the momentum instead of position space. To derive the Feynman
rules directly in the momentum space let us first derive graphical representations
of the n-point functions from (4.10).

We begin by the simplest, the 2-point function. This is what we already

computed in the free theory and named the Feynman propagator:
1 ∂ 1 ∂
h0| T φ(x1 )φ(x2 ) |0i = h0|0iJ = −i∆(x1 − x2 ) . (4.11)
i ∂J(x1 ) i ∂J(x2 ) J=0

The expression for ∆(x − y) is given in equation (3.50). Written in terms of the
effective potential we have
1 ∂ 1 ∂ ∂ ∂
h0| T φ(x1 )φ(x2 ) |0i = i∆W [J]+ ∆W [J] ∆W [J] ,
i ∂J(x1 ) i ∂J(x2 ) ∂J(x1 ) ∂J(x2 ) J=0
(4.12)
where we used the the normalization h0|0iJ=0 = 1 hence ∆W [0] = 0. Now,
the second term in (4.12) is proportional to the one-point function (6.1) which
52 CHAPTER 4. INTERACTIONS IN QFT

vanishes in the free case as one can directly see from the canonical expression of
the quantum field (2.21). We will require this one point function vanish also in
the fully interacting case as we will discuss in chapter 6. The motivation behind
this is that we want the creation operators to create a one-particle state, and
not a superposition of a one-particle state and the vacuum state. We want this
physical requirement also in the presence of interactions. So we will always set

h0| φ̂(x) |0i = 0 . (4.13)

Clearly the only non-vanishing contributions will come from Feynman diagrams
with two external sources. These contributions can be organised in an expansion
in g as usual and the lowest order contribution contains no vertices. Using the
Feynman rules in fig. 4.1 we find
Z
1
i∆W [J] = + · · · = −i d4 x1 d4 x2 J(x1 )∆(x1 − x2 )J(x2 ) + O(g 2 ) ,
2
(4.14)
where 1/2 comes from the symmetry factor. Now we take the variational deriva-
tives and we note that each variational derivative pins down an external vertex,
that is represented by a blob in the Feynman diagram, to a fixed point in space
time. Graphically we have

1 ∂ 1 ∂ 1 1
= (x1 x2 + x2 x1 ) = x1 x2 .
i ∂J(x1 ) i ∂J(x2 ) 2 2
(4.15)
This leads to a graphical representation of the two-point function
 
h0| T φ(x1 )φ(x2 ) |0i = x1 x2 + x1 x2 + x1 x2 + · · · (4.16)

where we also showed the g 2 contributions which arise from higher order terms
in (4.14) and the ellipsis represent O(g 4 ) contributions. All contributions to
the two point function beyond g 0 necessarily involve loops. Now let us take the
Fourier transform of this two point function. Consider the tree-level term first.
Using (3.50), we find
Z Z
0 0
h0| T φ(k)φ(k 0 ) |0i = −i d4 xe−ikx d4 x0 e−ik x ∆(x − x0 )
−i
= (2π)4 δ 4 (k + k 0 ) . (4.17)
k 2 + m2 − i

The overall delta function is nothing else but the statement that the momentum
carried by the particles at the two ends of the two-point function should be the
same, that is momentum is conserved. It is immediate to see that this overall
factor will be the same in all connected Feynman diagrams, as it just arises from
combination of the incoming and outgoing space-time integrals combined with
the exp{ik(x − y)} factors in the propagators.

Generally Feynman diagrams in the momentum space much simpler. One

only has to impose momentum conservation at every vertex in the diagram and
integrate over the unfixed momenta (which arise from the loops) and divide by
 x2 x’2

4.3. FEYNMAN RULES IN THE MOMENTUM SPACE 53

k
i
k 2 + m2 i✏

ig

Figure 4.8: Momentum space Feynman rules for the real scalar field with cubic
interaction.

the symmetry factor of the diagram. For example the value of the first diagram
in the one-loop contribution in (4.16) is
Z
−g 2 1 1
(2π)4 δ 4 (k + k 0 ) 2 d4 k1 ,
(k + m2 − i)2 (k − k1 )2 + m2 − i k12 + m2 − i
and the second one-loop contribution (the “tadpole” diagram) is2
Z
4 4 0 −g 2 1 d4 k1
(2π) δ (k + k ) 2 .
(k + m2 − i)2 m2 − i k12 + m2 − i
The same prescription (label all propagators by a momentum, impose momen-
tum conservation at the vertices and include an overall factor of momentum
conservation) applies to higher point functions. In the general case the over-
all momentum conservation factor obviously becomes δ 4 (kin − kout ) where kin
and kout are the sum of all ingoing and outgoing momenta respectively. The
momentum space Feynman rules are summarized in figure 4.8.

Consider the 4-point function as another example. One finds two separate
type of contributions
∂ ∂ ∂ ∂
h0| T φ(x1 )φ(x2 )φ(x01 )φ(x02 ) |0i = i∆W [0]
∂J(x1 ) ∂J(x2 ) ∂J(x1 ) ∂J(x02 )
0

∂ ∂ ∂ ∂
+ i∆W [0] i∆W [0] + (x1 ↔ x01 ) + (x1 ↔ x02 )
∂J(x1 ) ∂J(x2 ) ∂J(x1 ) ∂J(x02 )
0

given in the first and the second lines, where we used the condition that the
one-point function, hence δW [0]/δJ(x) = 0. Let us first work out the second
line. The only non vanishing contributions in W in the second line involve two
external sources, the tree level contribution being the same as (4.14). In fact
the second line is nothing else but pairs of two-point functions. At the zeroth
order in g they look like this:
x1 x2
+ (x1 ↔ x01 ) + (x1 ↔ x02 ) . (4.18)
x01 x02
2 We will see later that this tadpole contribution is cancelled in the renormalized theory.
54 CHAPTER 4. INTERACTIONS IN QFT

There will of course be higher order g corrections on these two separate propa-
gators but they will always correspond to the disconnected contributions to the
4-point function. As we will see in the next section they play no role in the
scattering processes and we typically ignore such disconnected contributions in
calculation of experimental observables. They are there however, are physical
and represent processes with no interactions. Their contributions are completely
determined by the knowledge of the two-point functions we studied above.

The novel contribution is the fully connected piece that arises from the first
line above, (4.18). Only the terms with 4 external sources in W can contribute
to this. Graphically,

1 4
i∆W [J] = + O(g
x1 ), x’1 (4.19)
8
where the numerical factor is the symmetry factor x2
and the higher x’
order contri-
2
butions involve loops. As in the two-point function above, when the 4 functional
derivatives in (4.18) act onx1this they pin thex’external
1
sources to the indicated
x1 x’1
specific values. There are 4! = 24 different ways these derivatives can act and,
clearly, they produce only three distinct classes of Feynman diagrams shown
below x2 x’2
x2 x’2
x1 x’1
x1 x’1
x1 x’1

x2 x’2 x2 x’2 x2 x’2

+ + (4.20)
x1
each of which come x1
x’1 with a multiplicity factorx’of
1
8 which neatly cancels the 1/8
factor in (4.19). The value of the tree-level contribution to the 4-point function,
using figure 4.8 is x2 x’2
x2 x’2
h0| T φ(k1 )φ(k2 )φ(k10 )φ(k20 ) |0i = A(k)(ig 2 )(2π)4 δ 4 (k1 + k2 − k10 − k20 )
x1  x’1 
1 1 1
× + + , (4.21)
(k1 + k2 )2 + m2 (k1 − k10 )2 + m2 (k1 − k20 )2 + m2
x2 x’2
where we ignored the i’s in the denominators which only become relevant when
appear inside integrals, hence can be ignored in the tree level diagrams and we
stripped out the overall factor that arises from the external propagators that is
common to all three contributions
−i −i −i −i
A(k) ≡ 02 02 , (4.22)
k12 2 2 2
+ m − i k2 + m − i k1 + m − i k2 + m2 − i
2

This product arise from the 4 external propagators that enter in the diagram
above.
Chapter 5

From theory to experiment:

scattering in high energy
physics

Consider a scattering experiment where incoming particles scatter and pro-

duce outgoing particles that are captured by a detector. How do we prepare
the initial states in this scattering experiment? In E the free theory a one-particle
states are created by the creation operator as ~k = a† (k) |0i with energy of the
q
particle given by ω = |~k|2 + m2 and the vacuum is normalized as h0|0i = 1.
As we discussed earlier, it is crucial that this one-particle state obeys the nor-
malization condition D E
~k ~k 0 = (2π)3 2ωδ 3 (~k − ~k 0 ) . (5.1)

The crucial assumption we will make in the following is that the space-time re-
gion where we prepare the states and region where they interact with one another
are well separated. The region where the initial states are prepared and the final
states are observed are called the asymptotic region. In an experiment this is
where the particle beams are created and where the final states are captured
by the detectors. In theory we idealize the situation1 by taking the asymptotic
region at t → ±∞ and |~x| → ∞. Thus our initial and final states are created
by the creation operators of the free theory. This set-up makes sense because of
the following observation: A basic assumption of QFT is locality, in particular
interaction terms e.g. gφ3 (x) are local functions of fields. This means that, if
we assume that particles interact with each other somewhere at the origin of the
lab (this can always be chosen as the origin of the coordinate system using the
translation symmetry), then these particles should be very close to each other
there. But this means that they should be well separated in the asymptotic
1 This setup is perfectly suitable for the high energy experiments. To use the same descrip-

tion in a condensed matter system to describe scattering of electrons, holes and phonons etc
one has to assume that the mean free path of the particle-like excitations is sufficiently large.
This assumption typically fails in strongly correlated systems.

55
56CHAPTER 5. FROM THEORY TO EXPERIMENT: SCATTERING IN HIGH ENERGY PHYSICS

region t = ±∞, |~x| → ∞ since particles have finite momentum. That is, the
inter particle distance goes to infinity |~x1 − ~x2 | → ∞ in the asymptotic region.
Then, indeed it makes sense that they are free hence created and destroyed by
creation/annihilation operators of the free theory.

Summarizing the discussion above the scattering amplitude which deter-

mines the overlap of two incoming particles with momenta k1 and k2 with two
outcoming ones with momenta k10 and k20 is given by hf |ii with

|ii = lim a† (k1 )a† (k2 ) |0i , |f i = lim a† (k10 )a† (k20 ) |0i . (5.2)
t→−∞ t→+∞

5.1 LSZ reduction formula

To relate this physical setup to the quantum field φ we need a relation

between the creation operator and the field. In the free theory this is given by2
Z
←→
a† (k, t) = −i d3 xeikx ∂0 φ(x) , (5.4)

←→
where we define a derivative with double arrow as A ∂ B ≡ A∂B − (∂A)B.
We also indicate the time-dependence of the creation operator explicitly. This
equation is easy to obtain from the oscillator expansion (2.21).

Now, the crucial step is to define the one-particle creation operator as in

(5.4) also in the interacting theory. The motivation for this is that the cre-
ation/annihilation operators should reduce to that of the free theory in the
asymptotic regions. There is no obvious justification for this, as we derived
(5.4) from the oscillator expansion of the free field φ in creation and annihila-
tion operators and this does not necessarily hold in the interaction theory. For
the moment, this should be viewed as a definition of the one-particle creation
operator. We will justify this definition in the next section by showing that
the state created by (5.4) indeed satisfies all the properties associated to a one-
particle state also in the interacting theory i.e. it is orthogonal to the vacuum
and multi-particle states, and it has the correct normalization as in (5.1). Then
the derivation that leads to (5.5) also holds in the interacting case.

Let us see how the on-particle states evolve in time by computing the time

2 More precisely, a one-particle state will be given by a wave-packet, which is created by a

superposition of different momentum eigenstates weighed by a Gaussian of the form

|~
k−~k1 |2
Z
−
a†1 ≡ d3 kf1 (~k)a† (k) , f1 ∝ e 4σ 2 , (5.3)

where σ determines the spread of the wave-packet. Here I will simply assume that the particle
is well localized in momentum by taking the function f1 (~k) = δ 3 (~k − ~k1 ).. This corresponds
to the σ → 0 limit of the Gaussian. Use of a Gaussian wave-packet means that the particle
is also localised in space. You can see this by taking the Fourier transform of f1 (k) to obtain
f˜1 (x) ∝ exp −|x|2 σ .


5.1. LSZ REDUCTION FORMULA 57

derivative of (5.4). We find,

Z
†


∂0 a (k, t) = −i d3 x eikx +∂t2 + ω 2 φ(x) ,
Z  
= −i d3 x eikx +∂t2 + |~k|2 + m2 φ(x) ,
Z


= −i d3 x eikx −∂ 2 + m2 φ(x) , (5.5)

where in the second step we used ω 2 = |~k|2 + m2 , which is part of the defini-
of a† (k, t) in (5.4): the exponential term reads ikx = −iωt + i~k · ~x with
tion q
ω = |~k|2 + m2 . This may be confusing as we obtained this dispersion relation
directly followed from the Klein-Gordon equation in the free theory. Does it
imply that the dispersion relation (a.k.a the mass-shell condition) in the free
and interacting theories the same? In fact, they are not. The dispersion rela-
tion can be read off from the poles of the two-point function in the interacting
theory, and indeed one finds a different pole structure as we briefly discuss in
the next section. The point here is that the exponent in (5.4) is chosen, again,
to make sure that it reduces to that of the free theory in the asymptotic regions.
Other than that it is a part of the definition. In the last step in (5.5) we did
an integration by parts where we wrote |~k|2 as −∇ ~ 2 exp(ikx) and moved the
derivatives onto φ. To see that this is justified, one should use the definition of
the one-particle creation operator with the Gaussian wave-packet (see Appendix
A). I will assume that this has indeed been done and the limit σ → 0 is taken
at the end of the computation and ignore such boundary terms.

As a consistency check we see that the RHS of this equation vanishes in the
free theory upon use of Klein-Gordon equation. Hence the one-particle states
do not change in time in the free theory, as one expects. However, the initial
and final states of the same momentum are different in the presence of some
scattering region

in space-time because the equation of motion now differs from
−∂ 2 + m2 φ(x) = 0 hence the RHS of (5.5) does
not vanish. For example in
the real scalar theory above we have −∂ 2 + m2 φ(x) = gφ2 /2.

Using (5.5) we can derive a very useful formula that relates a one-particle
state created at t = −∞ to the same state captured at t = +∞:
Z +∞
† ~ † ~
a (k, +∞) − a (k, −∞) = dt ∂0 a† (~k, t)
−∞
Z


= −i d4 x eikx −∂ 2 + m2 φ(x) . (5.6)

This formula is valid in general when φ is an interacting field. We will now make
the assumption that the creation operator in the interacting theory is also given
in terms of the field as in equation (5.1). This is motivated with our discussion
above, that the interacting field approaches to the free field in the asymptotic
regions in space-time where the initial states are prepared and final states are
measured.

The two-to-two scattering amplitude is given by

h0| a(k10 , ∞)a(k20 , ∞)a† (k1 , −∞)a† (k2 , −∞) |0i . (5.7)
58CHAPTER 5. FROM THEORY TO EXPERIMENT: SCATTERING IN HIGH ENERGY PHYSICS

As this is already time-ordered we may rather write

h0| T a(k10 , ∞)a(k20 , ∞)a† (k1 , −∞)a† (k2 , −∞) |0i , (5.8)

where T is the time-ordering symbol introduced in (3.11). Now use the equa-
tion (5.6) to rewrite a† (k1 , −∞) and a† (k2 , −∞) in terms of a† (k1 , +∞) and
a† (k2 , +∞) plus the integrals and use the hermitean conjugate of (5.6) to rewrite
a(k1 , ∞) and a(k2 , ∞) in terms of a(k1 , −∞) and a(k2 , −∞) plus the integrals,
then reorder these creation annihilation operators according to time-ordering.
One finds that all the annihilators are to the right of the creators inside the
bracket hence they destroy the vacuum states leaving only the contributions
from the integrals. This is a huge simplification and leads to our final result,
Z Z
0


hf |ii = in+n d4 x1 eik1 x1 −∂12 + m2 · · · d4 xn eikn xn −∂n2 + m2
Z  0  Z  0 
4 0 −ik10 x01 0 0
d x1 e −∂1 + m · · · d4 x0n0 e−ikn0 xn0 −∂n20 + m2
2 2

h0| T φ(x1 ) · · · φ(xn )φ(x01 ) · · · φ(x0n0 ) |0i , (5.9)

where we presented the general result for the amplitude of n incoming initial one-
particle states scattering into n0 outgoing final particle states. This is called the
LSZ reduction formula named after Lehmann, Symanzik and Zimmermann, and
it is one of the fundamental equations of QFT. Its importance is clear: it relates
scattering matrices, hence experimental observables to the basic observables of
QFT namely the n-point functions.

5.2 Feynman rules for scattering amplitudes

The LSZ reduction formula when combined with the momentum space Feyn-
man rules that we derived in section 4.3 result in a very simple prescription for
the scattering amplitudes. Let us exemplify this in the 2 → 2 scattering ampli-
tude hf |ii = hk1 , k2 |k10 , k20 i. The relevant n-point function is the 4-point function
that we studied in section 4.3. First consider the disconnected contributions in
(4.18). Substituting it in the LSZ formula we find, for the leftmost diagram in
(4.18),
Z
0 0 0 0
hk1 , k2 |k10 , k20 idisconnected ∝ d4 x1 d4 x2 d4 x01 d4 x02 ei(k1 x1 +k2 x2 −k1 x1 −k2 x2 )

×(−∂12 + m2 )(−∂22 + m2 )δ 4 (x1 − x01 )δ 4 (x2 − x02 )

where we used (−∂x2 + m2 )∆(x − y) = δ 4 (x − y). Integration by parts in x1 and

x2 and use of the delta functions finally yield

hk1 , k2 |k10 , k20 idisconnected ∝ (k12 + m2 )(k22 + m2 )δ 4 (k1 − k10 )δ 4 (k2 − k20 ) . (5.10)

Hence, as expected, the disconnected contributions only describe propagation

without scattering. The non-vanishing contribution to the scattering amplitude
instead arises from the connected 4-point function which we computed in the
momentum space in (4.21). Fourier transforming this back to position space and
5.2. FEYNMAN RULES FOR SCATTERING AMPLITUDES 59

substituting in the LSZ formula we discover a nice simplification: all the external
propagators are cancelled by the differential operators in (5.9). It is immediate
to see that this is a general result and hold for a general n → n0 scattering
amplitude in QFT. In fact the LSZ formula is called a reduction formula precisely
for this reason. The external legs always drop out of the scattering calculations
and the corresponding diagrams with external propagators removed are called
the amputated diagrams. The 2 → 2 scattering amplitude in the real scalar
theory with cubic interaction is then:
 
1 1 1
hk1 , k2 |k10 , k20 i = ig 2 (2π)4 δ 4 (k1 +k2 −k10 −k20 ) + + +· · · ,
m2 − s m2 − t m2 − u
(5.11)
where the ellipsis represent the higher order loop corrections and we defined the
so-called Mandelstam variables

s ≡ −(k1 + k2 )2 , t ≡ −(k1 − k10 )2 , u ≡ −(k1 − k20 )2 . (5.12)

These Lorentz invariant combinations provide a very useful parametrisation of

the scattering process. In particular the variable s is the total incoming 4-
momentum squared in a scattering experiment and the contribution given in
terms of s is called the “s-channel” depicted in the first diagram in (4.20).
Similarly the other contributions are called the t and the u-channels.

By generalising this computation we arrive at the following Feynman rules

in the momentum space for scattering amplitudes. The scattering matrix T is
defined as
hf |ii = (2π)4 δ 4 (kin − kout )iT , (5.13)
and the Feynman rules to compute T read as follows:

1. Draw external lines for each incoming and outgoing particle with 4-momentum
labels ki . Put an arrow towards the vertex for the incoming and away from
the vertex for the outgoing.

2. Connect these lines in all topologically inequivalent ways using internal

lines and vertices.

3. Assign a 4-momentum for each internal line and use momentum conser-
vation at the vertices to label them.

4. The value of the diagram is obtained by using the momentum space Feyn-
man rules in figure 4.8.

5. Integrate over the momenta that are not fixed by momentum conserva-
tion (these arise in loop diagrams) with the measure d4 `i /(2π)4 for loop
momenta `i .

6. Divide by the symmetry factor for the loop diagrams (the other symmetry
factors which arise from the external legs are all removed already).

7. The scattering matrix iT is given by the sum over the values of all these
diagrams.
60CHAPTER 5. FROM THEORY TO EXPERIMENT: SCATTERING IN HIGH ENERGY PHYSICS

k10
k1
✓

k20 k2

Figure 5.1: Setup of the 2 → 2 scattering process.

5.3 Cross sections and decay rates

Having derived the prescription to calculate the scattering matrix T , we

are now ready to study scattering processes. In general, we divide the high
energy processes into two classes: i) An incoming particle decays into two or
more particles. In this case the relevant quantity to calculate is the decay rate.
ii) Two incoming particles collide in which case the relevant quantity is the
scattering cross section.

We first consider the scattering process and pick 2 → 2 scattering as a

definite example. The setup is shown in figure 5.1. The on-shell states (incoming
and outgoing particles) satisfy ki2 + m2i = 0 where we allow for different masses3
The momenta and energies are conserved, as imposed by the delta function
inside (5.13 as a result of the underlying Poincare symmetry: ~k1 + ~k2 = ~k10 + ~k20
and E1 + E2 = E10 + E20 .

Since the scattering amplitude is Lorentz invariant we can choose any frame
to evaluate it. A particularly useful frame for the scattering process is the
“center of mass” (CM) frame where the total incoming q and outgoing qmomentum
vanishes k1 + k2 = k1 + k2 = 0 and Ein = E1 +E2 = |k1 | + m1 + |~k1 |2 + m22 .
~ ~ ~ 0 ~0 ~ 2 2

A parameter, that is the Lorentz-invariant analog of the total incoming energy

in a scattering, therefore more useful in relativistic processes, is the Mandelstam
s−variable s = −(k1 + k2 )2 . In the CM frame this becomes s = Ein 2
. Solving
for the individual momenta we find in the CM frame
q
1
|~k1 | = √ s2 − 2(m21 + m22 )s + (m21 − m22 )2 (5.14)
2 s
q
1
|~k10 | =
0 0 0 0
√ s2 − 2(m12 + m22 )s + (m12 − m22 )2 . (5.15)
2 s

Another parameter that enters in scattering is the angle θ between the “beam
3 All particles have the same mass m2 in the real scalar theory. However most of the

discussion will be valid for a generic scattering process.

5.3. CROSS SECTIONS AND DECAY RATES 61

axis” that is the axis of the incoming momentum and the axis of the outgoing
momentum (see figure 5.1. The Lorentz-invariant analog of this parameter is
the Mandelstam t-variable:
0
t = −(k1 − k10 )2 = m21 + m12 − 2E1 E10 + 2|~k1 ||~k10 | cos θ . (5.16)

Another frame that is commonly used in scattering processes is the lab frame
(or ”fixed-target frame”) with has ~k2 = 0. In this frame the incoming particle
has q
1
|~k1 | = s2 − 2(m21 + m22 )s + (m21 − m22 )2 . (5.17)
2m2
Note also that the momentum of the incoming√ particle in the CM and lab frames
are related to each other by m2 |~k1 |lab = s|~k1 |CM .

We are now ready to calculate the scattering cross section that is the relevant
observable for the high-energy collisions. The probability for the scattering to
take place is
| hf |ii |2
P = , hf |ii = (2π)4 δ 4 (kin − kout )T . (5.18)
hf |f i hi|ii
The numerator is | hf |ii |2 = (2π)4 δ 4 (kin −kout )×(2π)
R 4
4 4
δ (0)|T |2 . To understand
4 4 4
the meaning of δ (0) write it as (2π) δ (0) = d x exp{i0} = vol × time where
vol. × time is the total volume of space-time which we take to infinity at the
end of the calculation. Using this and the norm of a single particle hk|ki =
(2π)3 2k 0 δ 3 (0) = 2k 0 vol we arrive at the following formula for the probability
of the 2 → 2 scattering to take place per unit time:
P (2π)4 δ 4 (kin − kout ) vol|T |2
Ṗ = = Q2 . (5.19)
4E1 E2 vol2 j=1 (2kj0 vol)
0
time

This is the probability of scattering into definite momenta kj0 per unit time.
To obtain the total probability of scattering we need to integrate this over the
final momenta. This integration can be performed by regularising space time
as a 3D box of side-length L and vol = L3 . The momenta become discrete
~k 0 = (2π/L)~nj by imposing periodic boundary conditions in this 3D box4 where
j
~nj = (nxj , nyj , nxj ) is a collection of integers. Then one should think of (5.19)
as the probability for a particular collection of outgoing momenta {~nj }. To
turn this discrete output into continuous and finally integrate over the outgoing
momenta we need the replacement
X  L 3 X vol
Z
= −→ d3 k 0 .
2π (2π)3
~
n ~
k0
Q
This gives the multiplicative factor j vol/(2π)3 d3 kj0 leads to (5.19) and the
final result is given by
2
(2π)4 δ 4 (kin − kout ) 2 Y ˜ 0
Ṗ = |T | dk j , (5.20)
4E1 E2 vol j=1
4 Periodic boundary conditions are chosen for the sake of the argument, the final result

which we obtain by taking the continuum limit is independent of how we discretize the mo-
menta.
62CHAPTER 5. FROM THEORY TO EXPERIMENT: SCATTERING IN HIGH ENERGY PHYSICS

l
l’
}A
ρ
ρ’
v v’

Figure 5.2: Schema of two beams of incoming particles with densities ρ and ρ0 ,
lengths l and l0 and velocities v and v 0 in a scattering experiment. A denotes
the overlaping area of the beams.

where we used dk ˜ = d3 k/(2π)3 2k 0 . This is the differential probability of scat-

tering which is the probability of two incoming states to scatter into two final
states with momenta kj0 . Integrating (5.25) over the momenta of the final states
yields the total probability for the scattering event.

Now, let us define the concept of a scattering cross section. Consider the
set-up of a scattering experiment in Fig. 5.2 where two beams are approaching
each other. Denote the total number of scattering events in this experiment by
Nscat . As clear from the setting Nscat is proportional to the density of particles
ρ and ρ0 in the beams (which we assume to be constant for simplicity), the
length of the beams l and l0 and the area A shared by the beams, see the figure.
The total cross section is defined as Nscat divided by all these parameters:

Nscat
σ= (5.21)
ρ ρ0 l l0 A

This is called a “section” because it has the dimension of an area. The number
of scattering events per unit time is proportional to Ṗ . Based on this scattering
probability we can define the differential cross section as the differential rate
Ṗ , given in (5.25) divided by the incident flux of particles where the total
flux of particles, in the lab frame where the target is at rest, is given by the
number of incident particles per unit volume rho times their speed v times the
total number of target particles that participate in the scattering N 0 = ρ0 l0 A.
Writing Nscat = Ṗ × ttot in (5.21) where ttot is the total time of the scattering
process, bringing ttot in the denominator and using v = l/ttot , one sees that
the differential cross section defined above is the same as the differential cross
section defined from (5.21).

Using the expression (5.25) and the aforementioned definition of the differ-
ential cross section it is straightforward to obtain the following expression (see
exercise) for the differential cross section for a 2→ 2 scattering per differential
solid angle dCM = sin θdθdφ in the CM frame:

dσ 1 |~k10 | 2
= |T | . (5.22)
dCM 64π 2 s |~k1 |
5.3. CROSS SECTIONS AND DECAY RATES 63

Using (5.16) we have

dΩCM
dt = 2|~k1 ||~k10 |d cos θ = 2|~k1 ||~k10 |CM ,
2π
where ΩCM is the solid angle in the CM frame. We can now express (5.22) in
a Lorentz invariant form
dσ 1
= |T |2 , (5.23)
dt 64π 2 s|~k1 |2

where T is the scattering amplitude and |~k1 | is expressed in terms of Lorentz

invariants as
q
1
|~k1 |CM = √ s2 − 2(m21 + m22 )s + (m21 − m22 )2 . (5.24)
2 s
This differential cross section is easily generalised to 2 → n scattering as
n
(2π)4 δ 4 (kin − kout ) 2 Y ˜ 0
Ṗ = |T | dk j , (5.25)
4E1 E2 vol j=1

and,
Yn
1 ˜0.
dσ = √ |T |2 (2π)4 δ 4 (kin − kout ) dk j (5.26)
~
4|k1 |CM s j=1

The total cross section is obtained by integrating dσ taking into account the
0
Q if 0 there are ni identical
symmetry factor for the identical outgoing particles:
particles of type i then the symmetry factor is S ≡ i ni and one has,
Z
1
σ= dσ . (5.27)
S

5.3.1 Example: Real scalar with cubic interaction

Let us work out the 2 → 2 scattering in the real scalar theory with the
cubic interaction term. We will only consider the contribution of the tree level
diagrams to the scattering process, which provides a good approximation when
the coupling constant is sufficiently small g  1. We obtained the scattering
amplitude in (5.11) above which we reproduce here:
 
1 1 1
T = g2 + + + ··· . (5.28)
m2 − s m2 − t m2 − u
Using the formulae above t and u can be expressed in terms of the total center-
of-mass energy s and the scattering angle θ, see fig. (5.1). Then the scattering
amplitude is only a function of s and θ: T = T (s, θ).

To analyse the result it is useful to study it in certain limit. For example in

the non-relativistic limit ~k1 |  m that is s − 4m2  m2 one has
 
5g 2 8 s − 4m2
T ≈ 1− + ··· . (5.29)
3m2 15 m2
64CHAPTER 5. FROM THEORY TO EXPERIMENT: SCATTERING IN HIGH ENERGY PHYSICS

where the angle dependence is hidden in the subleading terms suppressed in this
low momentum limit. What we learn from this is that the 2 → 2 scattering in
this theory is very isotropic in the non-relativistic limit. On the other hand, in
the extreme relativistic limit ~k1 |  m that is s  m2 one has

g2  
T ≈ 2 3 + cos2 θ + O(m2 /s) , (5.30)
s sin θ
from which we learn that the extreme relativistic limit is sharply peaked in the
forward and backward directions θ = 0 and θ = π.

Using (5.27) we compute the total cross sections in the non-relativistic and
extreme relativistic limits
 
25g 4 79 s − 4m2
σ = 1− + ··· , (5.31)
1152πm6 60 m2
 
g4 7 m2
σ = 1 + + ··· , (5.32)
16πm2 s2 2 s
(5.33)

respectively.
Chapter 6

An introduction to
renormalization in QFT

In the previous chapter we made the connection between the generating

function of QFT and experimental observables. If one is able to determine the
generating function (4.10) then one can obtain the n-point functions and calcu-
late observables such as the scattering amplitudes etc using the LSZ reduction
formula. However, the final expression we obtained for the generating function
in (4.10) is not completely well-defined. For one thing, it involves infinitely
many divergent terms! In this chapter we will see where the divergences arise
from and introduce the procedure to cure them called renormalization .

6.1 Loop divergences

To see the divergences in a simple example let us consider the one-point

function. Using the (4.4) formula for E = 1 we see that the lowest order
contribution in g to the one-point function comes from the term with V = 1,
P = 2. The corresponding Feynman diagram is the top figure in figure 4.2. We
can easily evaluate this diagram using the Feynman rules in figure 4.1. The
result of this calculation is
Z
∆(0)
g d4 xd4 y J(x) ∆(x − y) ,
2

where the factor of 1/2 comes from the symmetry factor of the diagram and the
propagator ∆(y − y) = ∆(0) comes from the propagator in the loop. The one-
point function is obtained from this by taking the functional derivative w.r.t.
the source:
Z
δ ∆(0)
h0| φ(x) |0i = −i Z[J] = −i g d4 y ∆(x − y) + O(g 2 ) , (6.1)
δJ(x) 2

65
66 CHAPTER 6. AN INTRODUCTION TO RENORMALIZATION IN QFT

Figure 6.1: (Left): Time-ordered (Feynman) Green’s function has no poles inside
the contour C which consists of the real and the imaginary axes and the infinite
semi circles (dashed). (Right): Vanishing of the integral on C means one can
rotate the integral on the real axis toward the one on the imaginary axis. This
operation is called the Wick rotation.

where the higher order contributions arise from Feynman diagrams with a single
source and higher number of vertices. This expression is divergent because the
factor in front ∆(0) is infinite. From (3.50) we have
Z
d4 k 1
∆(0) = .
(2π)4 k 2 + m2 − i

Recalling the pole structure of the propagator, shown in figure 3.4, we see that
we can rotate the integral of k 0 from −∞ to +∞ on the real axis to an integral
−i∞ to +i∞ on the imaginary axis. This is because, the contour C shown
in figure 6.1 does not contain any poles. Make then the change of variables
k 0 = ikE with k 2 = −(k 0 )2 + ~k · ~k = kE2
+ kx2 + ky2 + kz2 ≡ |k|2 defined a 4-
sphere with radius |k|. This rotation is called the Wick rotation. After the Wick
rotation, the integral can be written as
Z ∞
d|k| |k|3
∆(0) = iΩ3 ,
0 (2π) |k| + m2 − i
4 2

where Ω3 = 2π 2 is the sphere of a unit 3 dimensional sphere.

This integral is clearly divergent. These type of divergences are ubiquitous

in QFT: they appear whenever there is a closed loop in a Feynman diagram
which instructs us to integrate over the loop 4-momentum from −∞ to +∞.
One way to deal with this problem is to place a high 4-momentum cut-off, that
is to replace the upper limit of the integral with a large but finite number
Z ΛE
d|k| |k|3
∆(0) → iΩ3 ,
0 (2π)4 |k|2 + m2 − i

and then take ΛE → ∞ at the end of the calculation hoping that this diver-
gence does not show up in the observables. This is called the regularization
procedure. However, we need to be very careful with the way we regularize
the loop divergences: we should make sure that we preserve the symmetries of
the action in the regularization procedure. For instance the UV cut-off we just
6.2. COUNTERTERMS 67

R
= iY d4 x
Figure 6.2: The new Feynman rule associated to the linear counter-term.

describe destroys the Lorentz invariance of the theory, that maps to rotational
invariance in 4D in the Euclidean momentum space above. There are different
ways to regularize which respects the Lorentz symmetry: dimensional regular-
ization, Pauli-Villars regularization, differential regularization etc. Pauli-Villars
instructs us to add another hypothetical highly massive particle to the theory
and replace the propagator of the scalar field as
 2
1 1 Λ2
→ 2 .
k 2 + m2 − i k + m2 − i k + Λ2 − i
2

The k-integral with this replacement is now finite. The result of the integral is
i
∆(0) → Λ2 . (6.2)
16π 2
We still need to take Λ → ∞ at the end of the calculation. The only thing
we achieved by this regularization is that we made the one-point function, and
the generating function in general well-defined at the expense of introducing
another parameter in the theory. The final results should not depend on this
unphysical cut-off Λ and the way to make that such an unphysical dependence
cancels out is called the renormalization procedure.

6.2 Counterterms

In these lectures I will not give a complete account of renormalization in

QFT which is a very important and rich subject. I will instead take a practical
approach and explain the minimum amount of modifications we need to make
in the original theory to make sure that the generating function is well-defined.
I will again do this using the example of the one-point function. Now, imagine
that the action we started with (4.1) is problematic for some reason – we will see
what the problem is shortly – and we need to modify it to make it well-defined.
The practical approach is to find the minimum number of modifications needed.
In the case of the one-point function, consider adding a linear term to the
Lagragian L → L + Y φ(x) where Y is a new coupling constant. Repeating
the calculation in section (4.2.1) with this new interaction term we find that
we need to add a new Feynman rule, that is a new vertex with a single line
(because the interaction is linear in φ), that we show in figure 6.2. This new
rule should be included in the Feynman graphs. Possible contributions to the
one-point function are shown in figure 6.3 with the various number of Y and
g-vertices. Now, let’s see that the divergences we encountered in the last section
can completely be cancelled by adjusting this new vertex. Anticipating that Y
68 CHAPTER 6. AN INTRODUCTION TO RENORMALIZATION IN QFT

with E = 1, VY = 1, V = 0

with E = 1, VY = 1, V = 2

with E = 1, VY = 2, V = 1

Figure 6.3: Examples of Y-vertex contributions to Feynman graphs. Here we

show possible contributions to the one-point function h0| φ̂(x) |0i with the indi-
cated number of Y and g-vertices.

should be chosen O(g) – to at least have a chance of cancelling the divergence

(6.2) in (6.1) – we see that there are two contributions to the one-point function
now: the one from the g-vertex that we calculated in (6.1) and the new, lowest
order contribution from the new vertex, that is the topmost diagram shown in
fig. 6.3. Adding these we find,
 Z
1
h0| φ̂(x) |0i = −i iY + g∆(0) d4 y∆(x − y) + O(g 3 ). (6.3)
2

We learn that we can cancel the divergence completely by setting

1
Y = i g∆(0) . (6.4)
2
Therefore including a new vertex in the theory allows us to cancel the unwanted
divergences. The divergences that arise from loops in the Feynman graphs that
are higher order in g can similarly be cancelled by correcting Y to the desired
order.

6.3 Renormalization procedure

All of this is called the renormalization procedure. To recap the procedure

is: first regularise the theory by including a cut-off Λ, then add new terms to the
Lagrangian a.k.a the counterterms, and finally choose them wisely to cancel the
loop divergences. The result will be well-defined, finite n-point functions. As,
you should have noticed, there is a lot of ambiguity in this procedure: i) how do
we know which counterterms to add? ii) How do we fix these terms precisely?
e.g. why not to choose Y = i 12 g∆(0) + constant which would still cancel the
divergence but would leave an arbitrary finite contribution to the one-point
function. iii) Do we need to introduce an infinite number of counterterms to
cancel all possible divergences in the Feynman graphs? If so, the theory would
have absolutely no predictive power!?

In this introductory course I will only provide answers to these questions

without proofs:
6.4. RENORMALIZATION CONDITIONS 69

• The answer to i) is add all possible polynomials in φ up to the maximum

order of φ in the original Lagrangian, that is φ3 in the present case. Then
the full counterterm action for the real scalar theory is

1 1 1
Lct = − (Zφ − 1)∂ µ φ∂µ φ − (Zm − 1)m2 φ2 + (Zg − 1)gφ3 + Y φ , (6.5)
2 2 3!
where the coefficients Zf , Zm , Zg and Y are to be fixed below. We
parametrised these coefficients so that when added to the original La-
grangian, the total Lagrangian reads

1 1 1
Ltot = − Zφ ∂ µ φ∂µ φ − Zm m2 φ2 + Zg gφ3 + Y φ . (6.6)
2 2 3!

• The answer to ii) is: renormalization conditions. These are certain phys-
icality requirements that determine the counterterms solely by physical
requirements. We discuss them in the next section.

• The answer to iii) is: no we do not need an infinite number of counterterms.

We only need the four terms in (6.5) in this theory. This is sufficient
to fix all the divergences! In general, when we need a finite number of
counterterms the theory is called renormalisable. A general rule of thumb
to determine whether a theory is renormalisable is to require that the
mass dimension of the coupling constants in the original Lagrangian are
non-negative. To calculate the mass dimension of the field [φ] just require
that the action should be dimensionless. Then from the kinetic term and
the rest we learn that

[φ] = 1, [m] = 1, [g] = 1 . (6.7)

Hence the theory is renormalisable according to the rule. A quartic inter-

action g4 φ4 would still be renormalisable, but a quintic one g5 φ5 would
not, as [g5 ] = −1.

6.4 Renormalization conditions

Now we will answer the question how to fix the counterterms unambiguously.
We will present the argument for the real scalar theory but the basic idea is the
same and applies to all interacting field theories. We need to determine the four
unknown coefficients Zφ , Zm , Zg and Y in (6.6). To do so, it is important to
consider a physical situation, e.g. a typical experiment to which we would apply
our quantum field theory and predict the outcome. We can consider for example
the 2 → 2 scattering experiment of the precious section. A crucial ingredient in
this experimental setup was that the asymptotic one-particle states as t → ±∞
were taken to be the eigenstates of the free Hamiltonian. This is discussed in the
beginning of the previous section. To make sure that these asymptotic states
are well defined we need the following three conditions on the one-particle state
operator a† (k, t) defined in (5.4)
70 CHAPTER 6. AN INTRODUCTION TO RENORMALIZATION IN QFT

1. The asymptotic one-particle state is not a superposition of a single particle

state and the ground state:

lim h0| a† (k, t) |0i = 0 .

t→±∞

2. This one-particle state is properly normalised with Lorentz invariant nor-

malisation:
lim hp| a† (k, t) |0i = (2π)3 2ωδ 3 (~k − p~) .
t→±∞

3. There is no overlap between this state and the many-particle states

lim hΨ| a† (k, t) |0i = 0 ,

t→±∞
P R
where |Ψi = n d3 pψn (p) |p, ni represents a multi-particle state with to-
tal momentum p~ and n collectively denotes all the other degrees of freedom
such as the relative momenta of the particles and other possible quantum
numbers such as charge etc.

In essence these conditions ascertain orthogonality in the asymptotic Hilbert

space. Using the definition (5.4) we find
Z
~ ← →
h0| a† (k, t) |0i = −i d3 x e−iωt+ik·~x ∂0 h0| φ(x) |0i

Now using the space-time translation operator and the Poincare invariance of
the vacuum we have h0| φ(x) |0i = h0| e−iP̂ x φ(0)eiP̂ x |0i = h0| φ(0) |0i where we
used invariance of the vacuum state under translations i.e. P̂ |0i = 0. Thus one
has Z
~
h0| a (k, t) |0i = ω d3 x e−iωt+ik·~x h0| φ(0) |0i .
†

Note that h0| φ(0) |0i is a constant with no dependence on space-time. This ex-
pression vanishes in the limit t → ±∞ only if and only if this constant vanishes.
Thus we arrive at the condition:

h0| φ(x) |0i ≡ v = 0 . (6.8)

This one-point function is just a Lorentz invariant number and called the vacuum
expectation value sometimes abbreviated as VeV. This can easily be set to zero
by making a field redefinition φ(x) → φ(x) − v. This shift is achieved by adding
the counterterm Y in section 6.2. Note that the condition (6.8) is the same as
the choice (6.4) which cancels the unwanted UV divergences that arise in the
one-point function at O(g). From the discussion above we learn that this choice
is not only required to cancel the divergences but also to ascertain well defined
asymptotic one-particle states in the theory. In particular the renormalization
condition (6.8) forbids a possible additional constant in (6.4). We learn that
the condition (6.8) fixes the value of the counterterm Y in (6.6) unambiguously.

Now, consider the second condition above, that hp| a† (k, t) |0i is properly
normalised. Using again (5.4) we find that this will have proper normalisa-
tion only if hp| φ(x) |0i has the proper normalisation asymptotically. We have
6.4. RENORMALIZATION CONDITIONS 71

hp| φ(x) |0i = hp| e−iP̂ x φ(0)eiP̂ x |0i = e−ipx hp| φ(0) |0i. Now, we know one case
where the proper normalization condition of one-particle states was indeed sat-
isfied: the free field. Then using the expression for the canonically quantized
free field, (2.21) in the LHS of the equation above, we find that, in that case,
the correct normalisation of the one-particle states is equivalent to

hp| φ(0) |0i = 1 . (6.9)

What can happen in the interacting case at most is that the RHS will not be
unity but another (infinite) constant. We conclude that this infinite constant
can be removed and the condition (6.9) can be satisfied also in the interacting
case by a rescaling the field φ. This is indeed what the counterterm Zφ in (6.6)
achieves. Therefore we learn that the condition (6.9) fixes the value of Zφ in
(6.6) unambiguously. The third condition above is in fact automatically satisfied
and it does not help fixing any counterterm. This is derived in Appendix A.

All in all, we have discovered that the requirement of orthogonality of the

asymptotic Hilbert space fixes two of the counterterms in (6.6) leaving out the
other two Zm and Zg . What are the physical requirements to fix these? They
are entirely fixed by the experiment. Consider a 2 → 2 scattering process we
studied in the previous section with the scattering amplitude given by equation
(5.28).

First of all, we see that this amplitude is peaked1 around s = m2 . This

mass is generically different than the bare mass we wrote in the Lagrangian
because it should include all the corrections that arise from the interactions.
We adjust the counterterm Zm such that that the theory reproduces the correct
value of the observed mass of the φ particle. For example a similar counterterm
in quantum electrodynamics is fixed by calculating the amplitude of a similar
2 → 2 scattering and fixing the counterterm by the observed value of the electron
mass around 0.5 MeV.

Second we have the overall magnitude of this peak which is clearly affected
by the “loop corrected” coupling constant Zg g in (6.6). Comparison of the
scattering cross obtained from the theory (6.6) with the actual experimental
value then fixes Zg . All these four conditions (6.8), (6.9) and the final two
above are called the renormalization conditions. They are sufficient to fix all the
counterterms and fix all the unknowns in the theory (6.6). One can now consider
a different scattering or decay process and (6.6) will make a precise prediction for
the outcome of this process. QFT — in particular QED — has been enormously
successful in this regard. I have to warn you that I oversimplified the discussion
in this section.

In fact the story is more complicated. In particular, as we have seen in the

example of the scattering cross section above, the renormalization conditions
depend on the energy scale s we probe the theory with. As a result of this all
the parameters in the theory, for example the mass and the coupling constant,
1 The pole in (5.28) at s = m2 is in fact smoothened by the higher loop corrections. For

instance the intermediate propagator that gave rise to this pole becomes 1/(m2 −s+iσ) where
σ = O(g 2 ) arises from the loop corrections to this intermediate propagator. Therefore one
obtains a peak around s = m2 with the width determined by σ.
72 CHAPTER 6. AN INTRODUCTION TO RENORMALIZATION IN QFT

acquire dependence on the energy scale. This signifies the fact that the higher
order quantum corrections (loops) become more or less relevant depending on
the energy scale you probe the theory with. This is called the renormalization
group flow. In particular the dependence of the coupling constant on the so
called “renormalization scale” µ(you can think of the same as s) is given by the
so-called beta-function:
dg(µ)
= β(g) .
d log µ
For β(g) > 0 the coupling constant increases with the energy scale and the
theory becomes non-perturbative i.e. the perturbation theory in g breaks down
at some high energy scale. QED is an example of this type. For β(g) < 0
the theory becomes non-perturbative at low energies. QCD, the QFT of the
strong force which glues the quarks inside the atomic nuclei is an example of
this situation. Finally β(g) = 0 is a special case, called a fixed point of the
renormalization group flow, and the theory around such fixed points are typically
described by a conformal field theory. An example is the QFT corresponding to
a 2d spin system (Ising model) at criticality.
Chapter 7

Representations of the
Lorentz group and particles

7.1 Group representations

We only discussed a specific type of quantum field up to here: the scalar

field φ. Under a Lorentz transformation

x̄µ = Λµ ν xν ,

this field transforms as a scalar:

φ̄(x̄) = φ(x) ⇒ φ(x) → φ̄(x) = φ(Λ−1 x) , (7.1)

hence we say that φ transforms trivially, that is the field in the transformed
frame is the same as in the old frame. Here is an example of field that transforms
non-trivially: ∂µ φ(x). By taking the derivative of the expression above and
making a change of variables we find,

Λνµ ∂¯ν φ̄(x̄) = ∂µ φ(x) ⇒ ∂µ φ(x) → ∂µ φ(x) = (Λ−1 )νµ (∂ν φ)(Λ−1 x) .
(7.2)
This is a non-trivial transformation because not only the index of the field but
also the field itself transforms ∂µ φ → (Λ−1 )νµ ∂ν φ by a factor of Λ−1 . This
is called the covariant vector representation. In general a covariant vector Aµ
transform as
Aµ → (Λ−1 )νµ Aν .
Since contraction of this with a vector with an upper index B µ Aµ is a scalar,
we learn that the vectors with upper indices transform as

B µ → Λµν B ν .

This is called a contravariant vector representation. By repeating this calcula-

tion we learn that a general object with n number of lower and m number of

73
74CHAPTER 7. REPRESENTATIONS OF THE LORENTZ GROUP AND PARTICLES

upper indices transform as

···βm
Cαβ11 ···αn
→ (Λ−1 )µα11 · · · (Λ−1 )µαnn Λβν11 · · · Λβνm
m
Cµν11 ···ν
···µn .
m
(7.3)

This general object is called a tensor representation. Classification of all possible

transformation rules under a group action is called the representation theory.
In order to find all possible representations under the Lorentz group we need to
work out the representation theory of the Lorentz group.

Let us first establish that the Lorentz transformations indeed form a group:

(i) If Λ and Λ0 are Lorentz transformations, then Λ00 = ΛΛ0 is also a Lorentz
transformation.
(ii) There exists an identity Λµν = δνµ
(iii) For ∀Λ, there exists an inverse Λ−1
(iv) Associativity: (Λ · Λ0 ) · Λ00 = Λ · (Λ0 · Λ00 ).

It is easy to show these properties using only the definition of a Lorentz trans-
formation:
ηµν Λµα Λνβ = ηαβ (7.4)
Let us denote a generic transformation rule as

DA → D̄A = LB
A (Λ)DB ,

where A, B denote a generic collection of indices. For example

−1 µ1
LB
A = (Λ )α1 · · · (Λ−1 )µαnn Λβν11 · · · Λβνm
m

for a tensor. Now, a representation is defined by the following property:

0 0
LB C C
A (Λ)LB (Λ ) = LA (Λ · Λ ) . (7.5)

The representation should become unity for a trivial Lorentz transformation:

LB B
A (1) = δA . From this it follows that

−1
LB
A (Λ ) = (L−1 )B
A (Λ) . (7.6)

Representations can be worked out by expanding the group action near unity:

Λµν = δνµ + δwνµ + O(δw2 ) . (7.7)

Inserting this in the definition (7.4) we find that w should be an antisymmetric

4 by 4 matrix. Therefore it has 6 independent components which correspond to
3 rotations and 3 boosts. Denoting a unit vector by n̂ we can parametrize them
as follows:
Rotations: δωij = −ijk n̂k δθ ≡ −ijk δθk generates a rotation around the k-
axis with angle δθ,
Boosts: δωi0 = n̂i δη ≡ δηi generates a boost in the i direction with rapidity
δη.
7.2. LORENTZ GENERATORS AND THEIR ALGEBRA 75

Now consider a generic representation L(Λ). Using LB B

A (1) = δA and demanding
continuity we find that, for a transformation Λ close to unity, L itself should
have an expansion near identity:

i ˆ µν )B ,
LB B
A (1 + δω) = δA + δωµν (M A (7.8)
2
where M µν are called the generators. They are a collection of N by N dimen-
sional matrices where N is called the dimension of the representation. They
should be anti-symmetric under the exchange of Lorentz indices.

7.2 Lorentz generators and their algebra

Starting from the generic properties of a representation (7.5) and (7.6) we

have
L−1 (Λ)L(Λ0 )L(Λ) = L(Λ−1 Λ0 Λ) .

Assuming that Λ0 is close to unity Λ0 = 1 + δw0 and expanding in δw0 we find

0
δwµν L−1 (Λ)M µν L(Λ) = δwµν
0
Λµρ Λνσ M ρσ

which needs to hold for all δw0 , thus we have:

L−1 (Λ)M µν L(Λ) = Λµρ Λνσ M ρσ . (7.9)

This is how the generators transform under a Lorentz transformation. Note

that this transformation rule is consistent with the general rule (7.3) that we
established above. Assuming that Λ is close to unity Λ = 1 + δw and expanding
(7.13) we discover that the Lorentz generators M should satisfy the Lorentz
algebra:

[M µν , M ρσ ] = iη µρ M νσ − iη νρ M µσ + iη νσ M µρ − iη µσ M νρ (7.10)

Lorentz generators generate the transformations associated with the parameters

δwµν that they couple to. In particular the boosts and rotations are generated
respectively by
M i0 ≡ Ki , M ij ≡ ijk Jk . (7.11)

Thus a general transformation can be written as

1 1
δwµν M µν = δwi0 M 0i + δwij M ij = δηi K i − δθk J k . (7.12)
2 2

Decomposing the Lorentz algebra (7.10) into rotations and boosts we find

[Ji , Jj ] = iijk Jk (7.13)

[Ji , Kj ] = iijk Kk (7.14)
[Ki , Kj ] = −iijk Jk . (7.15)
76CHAPTER 7. REPRESENTATIONS OF THE LORENTZ GROUP AND PARTICLES

7.3 Finite dimensional representations of the Lorentz

group

A very important result from the representation theory is that the repre-
sentations of a group algebra is in one-to-one correspondence with the repre-
sentations of the group elements that are continuously connected to the unit
transformation, that is, of the form (7.7). Generally the Lorentz group also
contain elements that are not connected to the identity transformation. They
are given by the parity and time-reversal transformations. For the moment we
ignore these elements–which we discuss parity and time-reversal in the appendix
B – and focus on the connected subgroup. For this subgroup it suffices to work
out the representations of the Lorentz algebra (7.10) or (7.13-7.15). Define
1 1
JiL ≡ (Ji − iKi ), JiR ≡ (Ji + iKi ) . (7.16)
2 2
From (7.13-7.15) we find that these combinations satisfy
[JiL , JjL ] = iijk JkL , [JiR , JjR ] = iijk JkR , [JiL , JjR ] = 0 , (7.17)
that is the algebra of two sets of SU(2)s. We already know the representations
of SU(2) algebra from a basic quantum mechanics course. A generic finite
dimensional representation of SU(2) is labelled by a non-negative (half)-integer
1 3
J = 0, , 1, , · · ·
2 2
and given by (2J + 1) × (2J + 1) dimensional matrices. These matrices have
2J + 1 eigenvalues that correspond to the value of J3 :
J3 : −J, −J + 1, · · · + J
In fact, these are the finite dimensional irreducible representations, usually ab-
breviated as an “irrep”. A more generic reducible representation is obtained by
tensor products of these. For example a 2(J + J 0 ) + 2 dimensional matrix that
can be put in a block diagonal form with 2J + 1 and 2J 0 + 1 dimensional blocks
is a reducible representation.

We conclude that the finite dimensional irreducible representations of the

Lorentz algebra (hence the connected Lorentz group) is labelled by two non-
negative (half)-integers (JL , JR ) that correspond to the subalgebras generated
by (7.16) and given by the tensor product of a 2JL + 1 and 2JR + 1 dimensional
matrices. Therefore the dimension of the representation is (2JL + 1) × (2JR + 1).
As the angular momentum generator is given by, see (7.16),
Ji = JiL + JiR
the states of different angular momentum contained in the representation (JL , JR )
are
|JL − JR |, |JL − JR | + 1, · · · , JL + JR .
This means that, even though the representation (JL , JR ) is irreducible under
the full Lorentz algebra, it is generically reducible under the rotation subgroup.

Here are some examples of finite dimensional Lorentz irreps:

7.3. FINITE DIMENSIONAL REPRESENTATIONS OF THE LORENTZ GROUP77

• Scalar, (0, 0): This is the trivial representation. According to the rule
above, it only contains a spin-zero state. The quantum field that is based
on this representation is nothing else but the scalar field φ(x) that we
studied above.

• Left and right handed Weyl spinors, ( 12 , 0) and (0, 12 ): These are spinor rep-
resentations that transforms as a 2 dimensional spinor under the left(right)
SU(2) and trivially under the right(left) SU(2) respectively. They only
contain a spin-1/2 state under rotations. The quantum field that is based
on these representation are called left(right)-handed Weyl spinor fields
and will be denoted by ψaL (x) (ψaR (x)). The index a = 1, 2 denotes the
components of the spinor with spins up and down respectively. These
representations play a key role in QFT and represent the massless spin
1/2 fields. The Dirac spinor ψa is obtained as a tensor sum of the left and
right handed Weyl spinors. Therefore these representations form the basis
of the largest class of elementary particles, electrons, muons, quarks etc.

• Vector representation, ( 12 , 12 ): This is a 4 dimensional representation that

transforms as a (covariant or contravariant) vector. It contains two types
of spin states, a spin-zero and a spin-one. The quantum field that is based
on this representation is called a vector field and will be denoted by Aµ (x).
The index µ is the space-time index that runs from 0 to 3 and labels the
4 different components. The spin-zero and spin-one components can be
thought of the A0 and A ~ respectively1 . This representation plays a key role
in QFT and corresponds to the gauge field. When this field is massless, it
is nothing else but the photon!

• Self-dual and anti-self dual antisymmetric tensors, (1, 0) and (0, 1): These
are 3-dimensional representations that only contains a spin-one state under
rotations. The associated quantum field can be written as an antisymmet-
±
ric matrix and denoted by Bµν . Note that the property of antisymmetry
under exchange of space-time indices is preserved under Lorentz transfor-
mations. That is, the transformed matrix B̄µν = (Λ−1 )α µ (Λ
−1 β
)ν Bαβ is also
anti-symmetric as the Lorentz matrices are symmetric. Therefore the anti-
symmetry property is respected by the irreps, hence can be used to reduce
general reducible representations. An antisymmetric 4 by 4 matrix has 6
components whereas this representation is supposed to have 3. There is a
+ −
further decomposition of an antisymmetric matrix Bµν = Bµν +Bµν where
the individual components satisfy the duality and anti-duality conditions
Bµν+
= 21 µναβ (B + )αβ and Bµν
−
= − 12 µναβ (B − )αβ . Note that these dual-
ity properties are also invariant under Lorentz transformations, hence are
respected by irreducible representations. The field strength of the gauge
field Fµν = ∂µ Aν − ∂ν Aµ transforms as a sum of the B + and B − repre-
sentsations. They further play an important role in characterising certain
topological field configurations called instantons.
1 Thisdecomposition is not Lorentz invariant and holds only in a Lorentz center of mass
Lorentz frame. A Lorentz invariant way to decompose the spin states canpbe obtained by
introducing the 4-velocity of the Lorentz frame uµ = γ(1, ~v ) where γ = 1/ 1 − |~v |2 . Note
that u2 = −1 in our metric convention. Then the decomposition reads Aµ = −uµ u · A +
(Aµ + uµ u · A). The first component is spin-zero and the second one is spin-one.
78CHAPTER 7. REPRESENTATIONS OF THE LORENTZ GROUP AND PARTICLES

• Left and right-handed Rarita-Schwinger spinors, (1/2, 1) and (1, 1/2): These
are 6 dimensional representations that contain spin 1/2 and spin 3/2 states
each. The associated quantum fields possess one vector one and spinor in-
L R
dex and denoted as ψµ,a and ψµ,a . Here µ is a space-time vector index and
a = 1, 2 is a spinor index. This contains 4 x 2 = 8 components which are
further reduced to 6 by the gamma-tracelessness condition γ µ ψµ,a L,R
= 0
µ
that are two independent conditions, where γ is the Dirac-gamma ma-
trix. Note that this condition is preserved by the Lorentz transformations.
They play an important role in supergravity theories and represent the su-
persymmetric partner of the graviton, called the gravitino.

• Symmetric traceless tensor, (1, 1): This is a 9 dimensional representation

that can be represented by a symmetric matrix Gµν , which is 10 dimen-
sional that is further reduced by demanding tracelessness η µν Gµν = 0.
Notice that both symmetry and tracelessness is preserved by the Lorentz
transformations. This representation contains one spin-zero, one spin-one
and one spin-two states. The latter one is nothing else but the graviton!

The list goes on.

7.4 Representing the Poincare group in quan-

tum mechanics

Reading the examples I listed above you should have noticed that the finite
dimensional irreps of the Lorentz group are in one-to-one correspondence with
the different type of elementary (also non-elementary) particles. Indeed, in gen-
eral we label particle excitations by their spin, their mass and their charge under
the various global symmetry groups such as the electric charge, hypercharge, lep-
ton number, baryon number etc. The most fundamental among these, in the
sense that it applies to all elementary particles is the mass and the spin. This
is indeed dictated by the Lorentz symmetry. There are other quantum numbers
that enter the labelling of a particle state, such as the 4-momentum pµ , spin
projected on some axis Ji etc. but these numbers are frame-dependent, they
change from one Lorentz frame to another. Only the Lorentz invariant labels
enter the classification, i.e. we do not say an electron is a particle with momen-
tum 2.5 MeV in some direction, we say it is a particle with mass 0.5 MeV and
spin 1/2. These invariant labels are indeed determined by the different finite
dimensional irreps of the Lorentz algebra as we discussed above.

The story is in fact more subtle. The reason for this subtlety is that in
a quantum theory we need to represent all symmetry operations by unitary
operators acting on the Hilbert space, because only unitary operators leave
the norm of a state invariant, hence preserve probabilities. That is, Lorentz
transformation acting on a quantum field in a specific representation is defined
by
φA (x) → φ̄A (x) = U (Λ)−1 φA U (Λ) = LB
A φB (Λ
−1
x) , (7.18)
7.4. REPRESENTING THE POINCARE GROUP IN QUANTUM MECHANICS79

with U (Λ)−1 = U † (Λ). What we need to define an elementary particle is then

an unitary irrep. A general result from representation theory, on the other
hand, is that non-compact groups (groups with a non-compact transformation
parameter, such as the boosts in the Lorentz group) do not possess any unitary
finite dimensional representations. This means that we need infinite dimensional
representations of the Lorentz group – as opposed to the examples discussed
above – in order to represent the Lorentz group in a quantum mechanical theory.

In fact we need infinite dimensional representations for a different reason

too: our full symmetry group is Poincare which consists of space-time transla-
tion in addition to the Lorentz transformations. Space-time translations act on
the Hilbert space by the unitary operator U (a) = exp(iaµ Pµ ) where Pµ is the
translation generator. Acting on a one-particle state with definite 4-momentum
it gives
U −1 (a) |pi = e−ia·p |pi , (7.19)
that is to say 4-momentum of the particle is its “charge” under translations.
However, as we have seen above, Lorentz transformations change p and the irreps
of Lorentz contain states with all pµ but fixed mass and spin. But then how
can this Lorentz irrep, which is a collection of states with different p represent
translations? The way out is, as we have seen before, is to consider fields. A
field φ(x) indeed provides a unitary representation under translations:

U −1 (a)φA (x)U (a) = φ̄A (x) = φA (x − a) . (7.20)

This is a unitary operation because U (a) in a field representation is given by

U (a) = exp{ia · P } = exp{aµ ∂µ } which is a unitary operator. The reason that
representing translations as fields does not mess up the representation structure
under the Lorentz group is because a field φA (x) is comprised of one-particle
(and anti-particle) states with all different pµ . Let us consider the simplest
example of the free scalar field:
Z  
φ̂(x) = dk ˜ â(~k)eik·x + ↠(~k)e−ik·x . (7.21)

First, let’s see how the translations (7.20) work out. When (7.20) applies to
(7.21) it acts on the creation and annihilation operators inside φ. From (7.19)
and the fact that |pi = a(p)† |0i we learn that

U −1 (a)a(p)† U (a) = e−ia·p a(p)† , U −1 (a)a(p)U (a) = e+ia·p a(p) , (7.22)

which immediately yields (7.20). At the same time it does not mess up the
transformation property under Lorentz because, similar to.(7.22) the Lorentz
generators act on the creation/annihilation operators as

U −1 (Λ)a(p)† U (Λ) = a(Λ−1 p)† , U −1 (Λ)a(p)U (Λ) = a(Λ−1 p) , (7.23)

therefore a change of integration variables k → Λk yields (7.18) for the trivial

representation, that is LB A = I. Note that, it is crucial to have an integral over
all k for this to work, so indeed representation of the Poincare symmetry in
terms of a field is sufficient and does the job. Representing the Poincare group
by fields turns out to also automatically solve the issue with unitary vs finite
80CHAPTER 7. REPRESENTATIONS OF THE LORENTZ GROUP AND PARTICLES

dimensional representations of the Lorentz group. We will show this below. For
now, we can make the following precise definition of an elementary particle:

An elementary particle is a unitary irreducible representation of the Poincare

group.

As we discussed above this unitary representation is infinite dimensional and

given in terms of quantum fields. Now, it seems like our entire discussion of finite
dimensional representations of the Lorentz group in the previous section become
obsolete. Luckily, this is not the case. As Wigner showed in 1939, in fact, one
can construct these infinite dimensional unitary representations starting from
the finite irreps following a procedure called “induction from the Little group”.

7.5 Wigner’s construction of unitary represen-

tations

First we identify the “Casimir operators” of the Poincare group. These are
operators that are quadratic in the group generators with the essential property
that they commute with all of the generators of the group. The Poincare group
is generated by Lorentz generators M µν and space-time translations P µ . Its
algebra is given by

[M µν , M ρσ ] = iη µρ M νσ − iη νρ M µσ + iη νσ M µρ − iη µσ M νρ
[P µ , M ρσ ] = iη µσ P ρ − iη µρ P σ (7.24)
µ ρ
[P , P ] = 0.

It is easy to show, by using the Poincare algebra, that there are only two Casimir
operators: the mass and the spin,

1
M 2 = P µ Pµ , J 2 = W µ Wµ , Wµ = µνρσ M νρ P σ , (7.25)
2
where W µ is called the Pauil-Lubansky vector. These are invariant under both
translations and Lorentz transformations thus can be used to label represen-
tations. Clearly the first operator corresponds to the invariant mass of the
representation.

To figure out the value of the second operator it is easiest to work in a specific
Lorentz frame. Clearly the value should be independent of this frame choice.
One should distinguish two cases M 2 > 0 and M 2 = 0. For M 2 > 0 one can
always choose the rest frame P µ = (m, 0, 0, 0). Then one finds J 2 ∝ Ji Ji . This
is the total spin operator (up to a proportionality constant m2 ) whose eigenvalue
is j(j + 1) in a representation with spin j. Therefore we characterise a massive
Poincare representation by its mass M 2 and its total spin j. Both are conserved
under time evolution because both of the operators M 2 and J 2 commute with
the Hamiltonian P 0 . They are also clearly invariant under changes of Lorentz
frame.
7.5. WIGNER’S CONSTRUCTION OF UNITARY REPRESENTATIONS 81

When M 2 = 0 there is no center-of-mass frame and the best one can do is

to fix the spatial momentum in a given direction, say z. Then P µ = (ω, 0, 0, ω).
In this frame one can show that J 2 = 0. Heuristically this is because a massless
physical excitation cannot have non-trivial spin on the plane perpendicular to
its motion. This means that W µ ∝ P µ with a single independent component
W 0 = P~ · J.~ The representation in a generic frame is characterized by the
eigenvalue of W 0 /|P~ | which is indeed frame independent. This operator is
called the helicity operator:
~ ~
~h ≡ P · J . (7.26)
|P~ |

This is the value of spin in the direction of motion. It is quantized for the finite
dimensional representations discussed above. For a massless scalar h = 0, for
a massless Weyl spinor it is h = ±1/2. For a massless vector, h = ±1, for the
graviton h = ±2 etc. It is conserved under time evolution because both P~ and
J~ commute with P 0 . It is invariant under changes of Lorentz frame only when
M 2 = 0. Physically this is because one can only change the spin in the direction
of motion by a boost that reverses the direction of motion, which is impossible
for massless particles.

Now that we have identified the Casimir’s to characterize the representa-

tions, we move on to construct the infinite dimensional unitary representations
based on the finite dimensional representations of the Lorentz group above. The
problem of non-unitarity is caused by the boost generators because this is the
part of the transformations with a non-compact range parameters i.e. rapidi-
ties. We will see this non-unitarity explicitly when we discuss the spinors below.
Wigner’s idea is simply to get rid of these unwanted boosts by fixing a frame,
and then constructing the representations in that given frame. In other words,
first fix a particular representative with a fixed momentum, then find the unitary
representation of the Lorentz group that commutes with this representative, and
then boost to an arbitrary frame to get the more general representation.

Let’s see how this works in practice. Again, we have to distinguish the
massive and the massless cases. In the massive case we can go to the rest frame
P µ = (m, 0, 0, 0). This means that all the boosts are fixed, because a boost
would change this fixed momentum. But the rotation subgroup O(3) leaves it
invariant, and we need to identify the irrep under rotations. This O(3) is called
the “little group”. It turns out that in quantum mechanics the correct group to
use is SU (2) as above. Finite dimensional representations are just the usual spin
states labelled by the (half)-integers J = 0, 1/2, 1, · · · . In a given representation
J there are 2J + 1 states −J, −J + 1 · · · , +J. This representation is both finite
dimensional and unitary.

For example, the field representation Aµ (x) for a massive vector particle
works as follows. Determine the wave-function of a given state with momen-
tum p~ for J = 1. This wave function µi (p) should be labelled by an index
i = 1, 2, 3 that corresponds to −1, 0, +1. In the special frame, they should sat-
isfy pµ µi (p) = 0 by Wigner’s construction. One can also normalize them by
requiring i,µ µi = 1. They are called the polarization vectors. Then the generic
field representation can be obtained by coupling these wave-functions with the
82CHAPTER 7. REPRESENTATIONS OF THE LORENTZ GROUP AND PARTICLES

creation and annihilation operators as usual and summing over all i and p~:
Z
µ
A (x) = dp ˜ eip·x ai (p)µ (p) + e−ip·x a† (p)µ∗ (p) . (7.27)
i i i

The trick works: even though the individual finite dimensional representations
µi (p), i.e. the wave-functions, transform non-unitarily under the Lorentz trans-
formations,
µi (p) → µi (p) = Λµν Λji νj (Λ−1 p) (7.28)
— because the Λµν is a non-unitary matrix for the boosts — the infinite di-
mensional representation Aµ (x) transforms unitarily because the action of the
Poincare group on A is given in terms of its action on the basis of creation/annihilation
operators

U (Λ)−1 ai (p)U (Λ) = Λji aj (Λ−1 p), U (Λ)−1 a†i (p)U (Λ) = Λji a†j (Λ−1 p) ,
(7.29)
where Λji is a rotation generator and it is unitary. This means that the one-
particle states transform unitarily under the Lorentz transformations. Yet the
action (7.30) generates the desired transformation property of Aµ :

U (Λ)−1 Aµ (x)U (Λ) = Λµν Aν (Λ−1 x) , (7.30)

as you can easily show by using (7.30), (7.28) and a change of integration variable
p → Λ · p in (7.27).

In the massless case instead one fixes the boosts by choosing P µ = (ω, 0, 0, ω).
Then the little group is O(2) that are rotations on the xy plane. Finite dimen-
sional representations are given by (half)-integers and they are labelled by the
helicities h = 0, 1/2, 1, · · · . In a given representation there are only two states
s = +h, −h. This is a unitary representation. The infinite dimensional field
representation is determined by first obtaining the wave-functions associated to
these states for a given momentum, then coupling them to the corresponding
creation/annihilation operators and finally summing over the helicity states s
and momenta p~. For example the field representation for a massless left or
right Weyl spinor works as follows. The wave functions for each helicity state
and fixed momentum are the spin 1/2 particles denoted as usa (p) and the spin
1/2 anti-particles denoted as vas (p). Here s is the helicity index running over
s = −1/2, +1/2 and a = 1, 2 is the spin index. The field representation is given
by Z
ψa (x) = dpe ˜ ip·x as (p)us (p) + e−ip·x b† (p)v s (p) . (7.31)
a s a

Wigner’s trick works as before: even though the individual spinors usa (p) and
vas (p) transform non-unitarily, the creation/annihilation operators transform as
0 0
U (Λ)−1 as (p)U (Λ) = Λss as0 (Λ−1 p), U (Λ)−1 b†s (p)U (Λ) = Λss b†s0 (Λ−1 p) ,

where Λji is a rotation generator and it is unitary. Therefore, the spinor one-
particle states transform unitarily under the Lorentz group action.
Chapter 8

Spinors

8.1 Weyl and Dirac Lagrangians

In this chapter we analyse in detail the first non-trivial representation of

the Poincare group, namely the spinor representations, and construct the cor-
responding quantum field theory. As we motivated in section 7.5 this repre-
sentation, that is necessarily in terms of fields will be induced by the finite
dimensional spinor representation of the Lorentz algebra that we discussed in
section 7.3. In that section we have seen that there are two such spinors, namely
the left and right Weyl spinor representations (1/2, 0) and (0, 1/2). These are
both 2 dimensional objects which we denote by ψaL and ψaR , a = 1, 2 respectively.
Thus the Lorentz transformation acting on them should be represented by a two
dimensional matrix. Using (7.12) and (7.16) a generic Lorentz transformation
is
 
L(Λ) = exp [i(ηi Ki − θi Ji )] = exp −(iθi + ηi )JiL + (−iθi + ηi )JiR . (8.1)

The left and right Weyl representation correspond to the two dimensional rep-
resentations of the SU(2) x SU(2) algebra (7.17) in terms of Pauli’s σ matrices
on one of the SU(2)s and trivially in the other:

1
(JiL )ba ψbL = (σi )ba ψbL , (JiR )ba ψbL = 0 , (8.2)
2
1
(JiR )ba ψbR = (σi )ba ψbR , (JiL )ba ψbR = 0 . (8.3)
2
Expanding (8.1) for infinitesimal transformations we find the following infinites-
imal transformations on the fields
1 1
δψ L (x) = − (iθi +ηi )σi ψ L (Λ−1 x), δψ R (x) = (−iθi +ηi )σi ψ R (Λ−1 x) (8.4)
2 2
What is the simplest Lagrangian for these fields? We need to require that this
Lagrangian

83
84 CHAPTER 8. SPINORS

• is a Lorentz scalar ,
• is real ,
• contains a single derivative in the kinetic term.

The last requirement is based on our discussion in section 1.3.2. The second
requirement is non-trivial unlike in the case of the real scalar field, because —
as clear from the transformation properties (8.4) — components of ψ L,R are in
general complex numbers. The first requirement, Lorentz invariance, however
is the most powerful one in constructing Lagrangians.

First, consider the kinetic term for ψ L . To construct a Lorentz scalar that
contains a single derivative, one needs to contract ∂µ with another object that
contains a space-time derivative. One almost have such an object for the spinor
fields, that is the collection of the sigma matrices σ i . In fact, the following
combinations
σ µ = (1, ~σ ), σ µ = (1, −~σ ) , (8.5)
does the job, and the simplest kinetic terms for the Weyl spinors which satisfy
all the requirements above can be written as,
LR = iψ R† σ µ ∂µ ψ R , LL = iψ L† σ µ ∂µ ψ L . (8.6)
You should check explicitly, using (8.4) above, that they transform as a scalar
under Lorentz transformations.

Now let us consider a mass term. The simplest possibility mψ L† ψ L is not

Lorentz invariant. In fact the only Lorentz invariant quadratic combination of
the fields ψ L,R with no derivatives, is the one that couples the left and the right
spinors
Lm = m(ψ L† ψ R + ψ R† ψ L ) . (8.7)
This is called the Dirac mass term1 . We learn that a generic massive spinor
needs to transform in the combined left-right Weyl representation (1/2, 0) ⊕
(0, 1/2) which is of course reducible to left and right spinors. This representation
is called the Dirac spinor. It is conventional to represent it by a 4-component
spinor by writing  L 
ψ
ψ= . (8.8)
ψR
One can also combine the matrices (8.5) to form the 4x4 gamma matrices
 
µ 0 σµ
γ = (8.9)
σ̄ µ 0
The Lagrangian for the Dirac spinor is obtained by combining the left and
right components of the Weyl kinetic terms (8.6) and the Dirac mass (8.7) and
compactly written as
L = ψ̄(iγ µ ∂µ − m)ψ . (8.10)
where we also defined
ψ̄ ≡ ψ † γ 0 . (8.11)
1 It
is in fact possible to construct a mass term only for the left or the right spinors with help
of an additional “reality” condition. This is called the Majarona mass term and is discussed
in the relevant Problem set.
8.2. DIRAC ALGEBRA AND LORENTZ TRANSFORMATIONS 85

8.2 Dirac algebra and Lorentz transformations

Before we discuss the solutions to the Dirac lagrangian and discuss the
quantization of spinors, it is important to note some properties of the Dirac
gamma-matrices and the spinors. As we already noted in section 1.3.2, the γ
matrices satisfy the Dirac algebra
{γ µ , γ ν } = −2η µν . (8.12)
The Lorentz generators in the Dirac representation are given by
i µ ν
M µν ≡ S µν = [γ , γ ] . (8.13)
Dirac 4
You can easily check that they reduce to (8.4) for the left and the right compo-
nents. It is also straightforward to verify that Sµν satisfy the Lorentz algebra
using (8.12). Generators of rotations are given by the spatial components
 k 
1 σ 0
S ij = ijk (8.14)
2 0 σk
as expected. For example the upper(lower) components of both ψ L and ψ R
have spin +1/2 (-1/2) in the z-direction. The boost generators are given by
 
i σi 0
S i0 = (8.15)
2 0 −σ i
Using the hermiticity properties of the γ matrices (8.9)
γ i† = −γ i , γ 0† = +γ i , (8.16)
or directly from (8.14) and (8.15) that one finds that the spin generators are
hermitean whereas the boost generators are anti-hermitean:
S ij† = S ij , S i0† = −S i0 . (8.17)
This means that, as we alluded to in the previous sections, the boost trans-
formations on the Weyl and Dirac spinors are non-unitary. Because of this
non-unitarity, the simplest expression you would guess for a mass term i.e. ψ † ψ
is in fact not a Lorentz scalar:
   
ψ † ψ → ψ † exp iδωi0 S i0 − iδωij S ij exp iδωi0 S i0 + iδωij S ij ψ 6= ψ † ψ .
The boosts screw up the Lorentz invariance. However, noting that γ 0 satisfies
S i0 γ 0 = −γ 0 S i0 , S ii γ 0 = +γ 0 S ij ,
the desired minus sign can be generated in the combination ψ̄ψ where ψ̄ is
defined in (8.11). Indeed
   
ψ̄ψ = ψ † γ 0 ψ → ψ † γ 0 exp −iδωi0 S i0 − iδωij S ij exp iδωi0 S i0 + iδωij S ij ψ = ψ̄ψ .
Therefore the combination ψ̄ is essential in obtaining expressions that have well-
defined Lorentz transformation properties. For instance, the combination
ψ̄γ µ1 · · · γ µn ψ
transforms as a contravariant n-tensor. In particular ψ̄γ µ ∂µ ψ is a Lorentz scalar.
Thus the Dirac lagrangian in (8.10) is indeed the simplest Lorentz scalar that
can be constructed from ψ with both a kinetic and a mass term.
86 CHAPTER 8. SPINORS

8.3 Chirality

The fact that the Dirac representation of the Lorentz algebra (8.13) is re-
ducible is directly seen by the presence of an operator that commutes with all
generators. This generator is called the chirality and given by

γ 5 ≡ iγ 0 γ 1 γ 2 γ 3 . (8.18)

Using the definitions (8.9) one can write it as,

 
5 −1 0
γ = (8.19)
0 1

Therefore it distinguishes the left and the right handed components in

 L 
ψ
ψ= ,
ψR

by assigning the eigenvalue -1(+1) to ψ L (ψ R ):

γ 5 ψ L = −ψ L , γ 5 ψ R = +ψ R . (8.20)

It is immediate to check that

{γ 5 , γ µ } = 0 , ⇒ [Sµν , γ 5 ] = 0 . (8.21)

Thus every state ψ in the Dirac representation should further be characterised by

its chirality eigenvalue, which means that the Dirac representation is reducible
to its left and right handed Weyl components. Even though chirality does not
change under Lorentz transformations, it is not necessarily conserved under time
evolution. For example a Dirac mass term in the Hamiltonian (8.7) transforms
a left-handed Weyl spinor into a right-handed one vice versa, thus chirality is
not conserved in the presence of a mass term. In the case m = 0it is conserved
and, as we demonstrate below, it coincides with the helicity defined in (7.26).

8.4 Solutions to the Dirac Lagrangian

Let us now derive the solutions to the Dirac Lagrangian

L = ψ̄(iγ µ ∂µ − m)ψ . (8.22)

The Dirac equations are derived from (8.10) by taking variations with respect
to ψ and ψ̄. We find
(i∂/ − m) ψ = 0 , (8.23)
and the hermitean conjugate of this equation. Here we introduced the shorthand
notation ∂/ ≡ γ µ ∂µ . Multiplying both sides of (8.23) by (i∂/ + m) we learn that
ψ automatically satisfies the Klein-Gordon equation


∂ 2 − m2 ψ = 0 .
8.4. SOLUTIONS TO THE DIRAC LAGRANGIAN 87

Therefore the most general solutions to (8.23) should

p be superpositions of plane-
waves exp(ipx) with p2 + m2 = 0 hence p0 = ± |~ p|2 + m2 = ±ω. When we
quantize the field we will interpret the solution with negative p0 as the anti-
particle. Then the general solution with p0 > 0 is
Z
ψ+ (x) = dp ˜ us (~
p)eipx , (8.24)

˜ = d3 p/2ω(2π)3 as before. Here us (~

where pµ = (ω, p~) and dp p) are spinors that
satisfy
(p/ + m) us (~
p) = 0 , (8.25)
for any fixed pµ , p0 > 0 with s the labelling the linearly independent solutions.
Physically s = 1, 2 can be chosen to correspond to the spin-up and spin-down
states respectively. One can easily solve (8.25) in the rest frame pµ = (m, 0, 0, 0)
where it becomes  
1 −1
m us (0) = 0 ,
−1 1
with the solution  
ξs
us (0) = (8.26)
ξs
with ξs some arbitrary constant spinor. One can for example choose a basis
where the two independent solutions read
   
1 0
ξ1 = , ξ2 = . (8.27)
0 1

In this basis s = 1, 2 correspond to spin-up and down respectively. One obtains

the solution in an arbitrary frame with momentum p~ by boosting this solution
using the boost transformation (8.15) (see the relevant Problem set ) and finds
√ 
−p · σ ξs
p) = √
us (~ . (8.28)
−p · σ̄ ξs

The solution with p0 < 0 is obtained following the same steps. For the
generic solution we have
Z Z
ψ− (x) = dp˜ ṽs (~ ˜ vs (~
p)e+iωt+i~p·~x = dp p)e−ipx , (8.29)

where pµ = (ω, p~) as before and vs (~

p) = ṽs (−~
p) are generic spinors. The latter
satisfies
(−p
/ + m) vs (~p) = 0 . (8.30)
Solution in the rest frame is  
ηs
vs = (8.31)
−ηs
where we can choose the constant spinors ηs as
   
0 1
η1 = , η2 = . (8.32)
1 0
88 CHAPTER 8. SPINORS

The solution in an arbitrary frame with momentum p~ by boosting this solution

using the boost transformation (8.15 and finds
√ 
−p · σ ηs
vs (~
p) = √ . (8.33)
− −p · σ̄ ηs
Let us count the number of degrees of freedom in these solutions. Dirac spinors
in general have 4 complex components that is 8 real degrees of freedom. Sub-
ject to Dirac equation the number of independent components become 4 real
numbers. As demonstrated above they correspond to spin up and down states
of a particle and an antiparticle.

Having found all the solutions we can now write down the most general
solution to (8.23) as follows:
2 Z
X

ψ= ˜ eipx us (~
dp p) + e−ipx vs (~
p)as (~ p)b∗s (~
p) , (8.34)
s=1
p
with p0 = ω = |~ p2 | + m2 and a(~p) and b(~
p) are arbitrary complex coeffi-
cients. These coefficients will be promoted to creation/annihilation operators
of one-particle states when we quantize the theory. This solution is an infinite
dimensional field representation of the Poincare algebra.

To get an idea about the spinor solutions let us consider the example of
a spinor moving in the z direction with momentum pz . The corresponding
solutions are
√   
ω − pz 0
   √ 
0  , u2 (pz ) =  ω + pz  ,
u1 (pz ) =  √
 ω + pz   
0
√
0 ω − pz
  √ 
0 ω − pz
√
 ω + pz   
v1 (pz ) =   , v2 (pz ) =  √ 0 
 0  − ω + pz  ,
√
− ω − pz 0
where the components s = 1 and 2 correspond to spin up and down states
respectively. In the massless limit m → 0 the energy of the particle becomes
ω → |pz | thus pz = ±ω. These solutions become eigenstates of the helicity
operator (??) in both cases. Using (8.14) one can write the helicity operator as
~σ · p~
h= . (8.35)
p|
2|~
For example for pz = +ω we have
      
0 √0 √0 0
 0   2ω   2ω   0 
u1 →  √  
 2ω  , u2 →  0  ,
 v1 →  
 0 , v2 →  √ 
− 2ω  ,
0 0 0 0
with helicity eigenvalues h = +1/2, −1/2, −1/2 and +1/2 respectively. We
observe that they coincide perfectly with the chirality eigenvalues which is −1
8.5. SPIN AND STATISTICS 89

for the left-handed components (upper two) and +1 for the right-handed com-
ponents (lower two). The same is true for the other case, pz = −ω. We learn
that the helicity and chirality eigenstates coincide in the massless limit. These
eigenvalues are also conserved under the time evolution in the massless limit.

8.5 Spin and Statistics

Before we move on to quantization of the spinors we need to discuss their

statistics. You are already familiar with the fact that the spinors are fermions
hence they satisfy anti-commutation relations rather than commutators in the
quantized theory. One can easily see that this should be the case by considering
their behavior under rotations.

One needs to first understand what spin-1/2 means. Consider rotating a

spinor around an axis, say z-axis, by some angle. We find that the operator
that rotates the spinors is
 θ 
ei 2 0 0 0

 0 e−i 2
θ
0 0 

L(θ) = exp iθS 12 =  θ  (8.36)
 0 0 ei 2 0 
θ
0 0 0 e−i 2

using (8.14). We find that under a full 2π rotation the spinor picks up a minus
sign. Only a double rotation by 4π would get us back the same spinor.

Now consider the following setting: two spin-up states located o the x-axis
at ±x and rotate the system by π. Applying (8.36) for θ = π to both spin states
one generates an overall minus sign in the total state

|ψ↑ (x)ψ↑ (−x)i → − |ψ↑ (−x)ψ↑ (x)i .

But, since the two particles are identical we cold obtain the same final configura-
tion by exchanging the positions of the two states. That is to say, |ψ↑ (−x)ψ↑ (x)i =
− |ψ↑ (x)ψ↑ (−x)i , hence spinors should anti-commute:

{ψ(x), ψ(y)} = 0 . (8.37)

This is opposite of fields with integer spin, in which case the particles have boson
statistics i.e. they commute.

8.6 Canonical Quantization of Spinors

We can now quantize the spinors using the canonical quantization method.
The canonical “position” field is ψ(x). We obtain the canonical “momentum”
field by varying the Dirac lagrangian Start with the Langrangian:

L = iψ̄∂/ψ − mψ̄ψ , (8.38)

90 CHAPTER 8. SPINORS

and obtain
δL
πψ = = iψ̄γ 0 . (8.39)
δ∂0 ψ
As demonstrated in the previous section we should impose the canonical equal
time anti-commutation relations:
{ψα (~x, t), πψ,β (~y , t)} = iδ 3 (~x − ~y )δαβ (8.40)
{ψα (~x, t), ψβ (~y , t)} = 0
{πψ,α (~x, t), πψ,β (~y , t)} = 0.
Then the question is how to write the general quantum spinor that solves the
Dirac equation (8.23). We determined the solutions in the previous section. The
most general solution is of the form (8.34) which I reproduce here:
2 Z
X

ψ= ˜ eipx us (~
dp p) + e−ipx vs (~
p)as (~ p)b∗s (~
p) . (8.41)
s=1

We quantize the theory simply by promoting the coefficients to operators in the

Hilbert space just as in the case of the scalar field. The only difference is that
the two coefficients a and b are not hermitean conjugates of each other as in the
real scalar but they are different operators as in the complex scalar (which you
studied in the exercises). This is because Dirac spinors are complex fields. The
quantized field then becomes:
X2 Z


ψ= ˜ eipx us (~
dp p) + e−ipx vs (~
p)as (~ p)b†s (~
p) , (8.42)
s=1

where now, just as in the complex scalar field, âs (~

p) annihilates a one-particle
†
state and b̂s (~
p) creates
p a one-anti-particle state, both with momentum p~, spin
s and energy ω = p|2 + m2 . The other linearly independent solution ψ̄
|~
determines the canonical momentum as in (8.39) and it becomes
2 Z
X

ψ= ˜ e−ipx ūs (~
dp p)a†s (~
p) + eipx v̄s (~
p)bs (~
p) , (8.43)
s=1

The annihilation operators are defined to annihilate the vacuum state, the state
with neither particles nor anti-particles:
p) |0i = 0 ,
as (~ p) |0i = 0
bs (~ ∀s , ∀~
p, (8.44)
We normalise the vacuum state as usual h0|0i = 1. The creation operators
create one-particle and one-anti-particle states out of the vacuum instead:
a†s (~
p) |0i = |p, s, +i , b†s (~
p) |0i = |p, s, −i , (8.45)
where I denote the particles by + and anti-particles by -.

Then, imposing the anticommutation relations (8.40) on this solution we

can determine the anticommutation relations between a and b’s. As shown in
appendix C we find
p), a†s0 (~
{as (~ p0 )} p − p~0 )δss0 2ω,
= (2π)3 δ 3 (~ (8.46)
0 † 3 3 0
{bs (~ p) }
p), bs0 (~ p − p~ )δss0 2ω .
= (2π) δ (~ (8.47)
8.7. SPINOR PROPAGATOR 91

and all the rest vanish. We can immediately see that the one-(anti)-particle
states are properly normalised, that is, the probability amplitudes are invariant
under Lorentz transformations
p), a†s0 (~
hp, s, +|p0 , s0 , +i = h0| {as (~ p0 )} |0i = (2π)3 δ 3 (~
p − p~0 )δss0 2ω , (8.48)
which is invariant under the Lorentz transformations. This already indicates
that our representation (8.42) is a unitary representation. As discussed at the
end of section 6, a general boost will act on the creation/annihilation operators
as (this directly follows from (??)):
0 0
U (η)−1 as (p)U (η) = Rss (η)as0 (Λ(η)−1 p), U (η)−1 b†s (p)U (η) = Rss b†s0 (Λ−1 p) ,
(8.49)
where R is a rotation matrix that represents how the spin of the particle ro-
tates, which depends on the orientation of the boost ~η . For example when
0 0
we boost along the spin direction then Rss = δss . It is straightforward to see
that this action, which is unitary on the basis states, when applied on the field
representations (8.42), (8.43) generates the desired spinor transformation law:
U (Λ)−1 ψa (x)U (Λ) = Lba (Λ)ψb (Λ−1 x) , (8.50)
which indeed demonstrates that going from finite dimensional spinor represen-
tations to the the infinite dimensional spinor field representation we managed to
represent the Lorentz group with a unitary action on the Hilbert space states.

8.7 Spinor propagator

The commutation relations (8.46), (8.47) and the oscillator expansions (8.42)
and (8.43) are sufficient to obtain the correlator of the spinor fields h0| ψ(x)ψ̄(y) |0i
etc. Using the definition of the vacuum state (8.44) we find
Z
h0| ψa (x)ψ̄b (y) |0i = ˜ ip(x−y) (−p/ + m)ab ,
dpe (8.51)
Z
h0| ψ̄b (x)ψa (y) |0i = ˜ −ip(x−y) (−p/ + m)ab ,
dpe (8.52)

(8.53)
where we use the completeness relations (exercise)
X
us (~
p)ūs (~
p) = −p/ + m . (8.54)
s

The other correlators just vanish h0| ψb (x)ψa (y) |0i = h0| ψ̄b (x)ψ̄a (y) |0i = 0. As
we discussed for the scalar fields the Feynman propagator is obtained by time
ordering of this correlator. However, there is one important difference in the
prescription for time ordering. Since spinors are fermions we need to take into
account the extra minus sign which arises when we reorder the two operators
inside the correlator:

ψa (t1 )ψb (t2 ) t1 > t2
T ψa (t1 )ψb (t2 ) = , (8.55)
−ψb (t2 )ψa (t1 ) t2 > t1
92 CHAPTER 8. SPINORS

Now let us prove the following identity:

Z Z
d4 p eip(x−y) f (p) 0 0 ˜ (p)eip(x−y)
= iθ(x − y ) dpf (8.56)
(2π)4 p2 + m2 − i
Z
˜ (−p)e−ip(x−y) .
+iθ(y 0 − x0 ) dpf

This follows from doing the p0 integral as a contour integral and closing the
contour from above for x0 < y 0 and from below for y 0 < x0 . Using this identity
we obtain our final expression for the time-ordered spinor propagator:
Z
d4 p ip(x−y) (−p/ + m)ab
h0| T ψa (x)ψ̄b (y) |0i = −i e . (8.57)
(2π)4 p2 + m2 − i
A similar calculation shows that the others vanish:
h0| T ψa (x)ψb (y) |0i = h0| T ψ̄a (x)ψb (y) |0i = 0 . (8.58)

8.8 Path integral Quantization of Spinors

The generating function for the spinors is written as a path integral com-
pletely analogous to the scalar path integral we discussed above: integrating
over all canonical spinor field configurations from t = −∞ to ∞ (with a slight
tilt in the time-contour) yields the vacuum-to-vacuum amplitude with sources:
Z
Z(η̄, η) = DψDψ̄ ei d x{−ψ̄(−i∂/+m)ψ+η̄ψ+ψ̄η}
R 4
(8.59)

where η and η̄ are the sources for the ψ̄ and ψ fields respectively. The main
difference from the scalar fields is that these path integrals should now be de-
fined as Grassmann variables. Properties of Grassmann integrals are detailed
in section 43 of Srednicki (also in Statistical Field Theory). We will only need
the following modification of the functional derivatives here:
Z
δ  
d4 y η̄(y)ψ(y) + ψ̄(y)η(y) = −ψ̄(x) (8.60)
δη(x)
Z
δ  
d4 y η̄(y)ψ(y) + ψ̄(y)η(y) = +ψ(x) . (8.61)
δ η̄(x)
Now we can obtain the time-ordered spinor propagator directly from the gen-
erating function. Let us recall that the propagator, as it follows from the path
integral is nothing but inverse of the kinetic term operator in the lagrangian.
For example for the complex scalars we have:
Z
Z(J , J) = DφDφ† ei {−φ Oφ+J φ+φ J } ,
† † †
R
†
(8.62)

where the kinetic term operator is O = −∂ 2 + m2 . Then the propagator follows

by completing the square in the exponential and taking functional derivatives
wrt Js in the end:
−iδ −iδ
h0| T φ(x)φ† (y) |0i = Z(J † , J) = −i O−1 (x, y) . (8.63)
δJ † (x) δJ(y) J=J † =0
8.8. PATH INTEGRAL QUANTIZATION OF SPINORS 93

Here the space-time components of the operator O follows by writing

Z Z Z
φ Oφ = d x d4 yφ† (y)δ 4 (x − y)(−∂x2 + m2 )φ(x) ,
† 4

hence O(x, y) = δ 4 (x − y)(−∂x2 + m2 ) and the inverse is defined as

Z
d4 yO−1 (x, y)O(y, z) = δ 4 (x − z) .

This expression in the momentum space becomes the literal inverse: For
Z
d4 p ip(x−y)
O(x, y) = e Õ(p)
(2π)4

Õ−1 (p) = 1/Õ(p) hence

Z
−1 d4 p eip(x−y)
O (x, y) = . (8.64)
(2π)4 p2 + m2 − i

The bottomline of this computation is that you can read off the time-ordered
propagator directly from the lagrangian without any computation, as inverse of
the kinetic term operator.

The same is true for the spinors. In this case the generating function is
given by (8.59) and the kinetic term operator is O = −i∂/ + m with the inverse
in the momentum space O−1 (p) = (−p/ + m)/(p2 + m2 − i) hence we obtain the
same expression as (8.57).
Chapter 9

Vector fields and gauge

invariance

9.1 Lagrangian for the vector field

The next irrep on our list in section 7.3 is (1/2,1/2). This representation
has 4 degrees of freedom, therefore it contains 4 fields when we represent it
using fields . Under rotations these 4 degrees of freedom will be decomposed
as 1+3, the first being a scalar (spin-0), the second a vector (spin-1). We
denote this collection of four fields by a space-time vector Aµ . We need to
find a lagrangian that represents the kinetic term for the 3-vector degrees of
freedom inside this Aµ , and this should be done in a Lorentz covariant way.
~ and just ignoring A0 would not work because this
Hence writing Aµ = (A0 , A)
separation is not invariant under Lorentz transformations which naturally mix
time and space components.

Let us begin with the most natural candidate, a direct generalization of the
scalar field:
1 m2
LA = − ∂µ Aν ∂ µ Aν − Aν Aν .
2 2
There are two issues with this lagrangian. Issue i): it leads to the equation of
motion (∂ 2 − m2 )Aµ (x) = 0. This EoM implies however that the components
of Aµ do not mix at all under time evolution, hence they decompose into four
scalar fields, A0 , · · · A3 . This is not a spin-1 representation but 4 times a spin-
0 representation. Issue ii): Upon quantization, we would write the oscillator
expansion, in analogy with the real scalar case as,
Z


Aµ (x) = dp ˜ aµ (p)eipx + a† (p)e−ipx ,
µ

where aµ (p) (a†µ (p)) with p2 + m2 = 0 destroys (creates) a particle with index
µ. Just as in the real scalar case these creation/annihilation operators should

94
9.1. LAGRANGIAN FOR THE VECTOR FIELD 95

satisfy the commutation relations of the form

[aµ (p), a†ν (p0 )] = (2π)3 2ωδ 3 (~

p − p~0 )ηµν ,
p
with ω = p|2 + m2 . But then, consider the norm of the state with µ = 0:
|~

|a†0 (p) |0i |2 = h0| a0 (p)a†0 (p) |0i = −(2π)3 2ωδ 3 (0) .

The norm is negative definite! This leads to negative probabilities which does
not make sense. This is an indication that the quantum theory that follows
from quantizing the lagrangian above is non-unitary.

Both problems can be solved at once, by realizing that in fact there is

another type of Lorentz invariant kinetic term we can write for the vector fields:
∂µ Aν ∂ ν Aµ . This term solves issue i) we mentioned above, namely that the
mixing between ∂µ and Aµ in the new kinetic term forces Aµ to transform
as a 4-vector instead of a collection of four scalars. Now consider the general
lagrangian obtained by combining all these terms with arbitrary coefficients a
and b
1
LA = a∂µ Aν ∂ µ Aν + b∂µ Aν ∂ν Aµ − m2 Aν Aν . (9.1)
2
This lagrangian also allows us decompose the 4-vector into a spin-0 and a spin-1
representation and construct the field theory for the spin-1 part. This can be
achieved by a particular choice of the coefficients a and b in (9.1) as follows.
Equations of motion now read

−aAv − b∂ν ∂µ Aµ − m2 Aν = 0 . (9.2)

Now multiply both sides with ∂ ν and choose a = −b and assume that m2 6= 0:
Then we obtain the constraint

m 2 ∂ ν Aν = 0 . (9.3)

In the momentum space this constraint reads k · A = 0. This is great, because

when we decompose a generic 4-vector into components parallel and perpendic-
ular to k µ as
 
kν (k · A) kν (k · A)
Aν (k) = Aν − + ≡ A1ν + A0ν , (9.4)
k2 k2

and recognize the first term as the spin-1 part and the second as the spin-0 part,
we see that the constraint (9.3) allows us to set the second one to zero, making
sure that the field Aµ only contains the component with spin-1.

It also solves issue ii) we mentioned above. The problem with unitarity that
arose from the temporal component of the gauge field upon quantization is no
longer there because the constraint (9.3) relates the temporal component to the
spatial ones:
A0 (k) = −ki Ai /k 0 ,
which means that we do not need to quantize the A0 component independently.
Hence the negative norm states would not arise (at least at the tree level).
96 CHAPTER 9. VECTOR FIELDS AND GAUGE INVARIANCE

9.2 Classical massive vector

We saw that the choice of a = −b in (9.1) solves our problems. With the
standard normalization choice a = −1 we can write the lagrangian neatly as
1 1
LA = − F µν Fµν − m2 Aν Aν , (9.5)
4 2
where we introduced the antisymmetric tensor

Fµν = ∂µ Aν − ∂ν Aµ . (9.6)

For those of you who have studied electrodynamics in the tensor formalism will
recognize this as the tensor whose components are the electric and the magnetic
fields. This is not the theory of electromagnetism yet as we have a finite mass
term in (9.5). We will discuss QED in the next chapter. Using the constraint
(9.3) the equation of motion that follows from varying (9.5) with respect to Aν
(or just setting a = −b = −1 in (9.2) reads


∂ 2 − m 2 Aν = 0 , (9.7)

which still looks like the equation of motion for a collection of 4 scalars, but it
is not because of the constraint we should impose on the solutions. The general
solutions to (9.7) are of the same form as the scalar except that we now have
the index ν. We also have to take into account the fact that only the spin-1 of
the vector Aν survives due to the constraint. The situation resembles the case
of spinors which also satisfied the Klein-Gordon equation but then imposing the
Dirac equation on top of the solutions to the Klein-Gordon equation required
p) and vas (~
us to include the spinor solutions usa (~ p) which maps the spinor index
a to the spin-1/2 polarization index s. Here, for the massive spinor, we need
a solution which maps the space-time vector index ν to the spin-1 polarization
index λ. These solutions are called the polarization vectors and denoted by
λν (~
p). In analogy with spinors, we can now write the most general solution as,
Z 3
X

Aν (x) = ˜
dp p) + e−ipx aλ∗ (p)λ∗
eipx aλ (p)λν (~ ν (~
p) , (9.8)
λ=1

with p2 + m2 = 0 and we also required that Aν is real. Now we have to impose

the constraint (9.3) which requires

pµ λµ (~
p) = 0 , ∀λ . (9.9)

As a result of this constraint there are only three independent basis vectors
which we denote by λ = 1, 2, 3. This label corresponds to the 3 spin states in
the spin-1 representation of the rotation group. In this sense λ are analogous
to us and v s of the spinors. In this field representation there are 3 basis vectors
per given momentum p~. Therefore in total there are infinitely many basis states.
Upon quantization these states λ (~ p) will correspond to the basis vectors in the
Hilbert space.

The polarization vectors can be fixed as follows. In addition to (9.9) they

can be normalized as ∗µ µ = +1. For the field propagating in the z-direction we
9.3. LAGRANGIAN FOR THE MASSLESS VECTOR 97
p
have pµ = (ω, 0, 0, pz ) with ω = p2z + m2 in the rest frame. Then the three
linearly independent solutions are
p ω
z
1µ = (0, 1, 0, 0) , 2µ = (0, 0, 1, 0) , 3µ = , 0, 0, . (9.10)
m m
1 and 2 are in the direction transverse to the direction of motion, hence they
are called the transverse polarizations whereas 3 is called the longitudinal po-
larization.

It is now straightforward to quantize the theory by promoting the co-

efficients aλ (p) and aλ∗ (p) to quantum annihilation and creation operators
respectively and working out their commutation relations starting from the
equal time commutators of Aν (x) and the corresponding canonical momentum
Πν (x) = δLA /δ∂0 Aν . However, what we are really interested in is the quanta
of the electromagnetic field that is the photon which corresponds to the limit
m → 0 of the Proca field. This is what we consider next.

9.3 Lagrangian for the massless vector

We can obtain the lagrangian of the massless vector field directly from (9.5)
by taking the m2 → 0 limit:
1
LEM = − F µν Fµν , (9.11)
4
with the field strength Fµν defined in (9.6). You recognize this as the lagrangian
for the electromagnetic field. Indeed we interpret Aµ as the scalar and the
vector potentials of electrodynamics as A0 = −A0 = V the scalar potential
and A~ the vector potential. Then the electric and magnetic fields are given as
E~ = −∇V − ∂t A ~ and B~ =∇ ~ ×A ~ which can be written as

1
E i = ∂ 0 Ai − ∂ i A0 = F 0i Bi = ijk ∂j Ak = ijk F jk . (9.12)
2
Then we indeed obtain the lagrangian for the electromagnetic field from (9.11)
1
1 2 
LEM = − 2F0i F0j η 00 δ ij + Fij Fkl δ ik δ jl = ~ − |B|
|E| ~ 2 . (9.13)
4 2
The equations of motion follow from setting the variation of (9.11) with respect
to Aν to zero and they read,
∂µ F µν = 0 . (9.14)
The individual components ν = 0 and ν = i give
~ ·E
∇ ~ =0 ~ ×B
∇ ~ = −∂t E
~

respectively, which we recognize as two of the Maxwell’s equations in the absence

of matter. The other two of the Maxwell’s equations arise from the Bianchi
identity
µναβ ∂ν Fαβ = 0 , (9.15)
98 CHAPTER 9. VECTOR FIELDS AND GAUGE INVARIANCE

which just follows from the definition Fµν = ∂µ Aν −∂ν Aµ . Physically this means
that the total number of 6 components in the electric and magnetic fields are
not independent but they arise from an underlying gauge potential Aµ . Written
in components for µ = 0 and µ = i equation (9.15) indeed produce the two
other Maxwell’s equations
~ ·B
∇ ~ =0 ~ ×E
∇ ~ = ∂t B
~.

Therefore we can be sure that this is the correct theory to quantize to obtain
the quantum mechanical theory of light.

However, the problems we mentioned in section 9.1 are back to haunt us

in the massless limit. Decomposition into spin-1 and spin-0 parts as in (9.4)
and extraction of the spin-1 part from the space-time vector Aµ was based on
the constraint (9.3) which no longer holds in the massless limit. Then, the
problem that the temporal component A0 giving rise to negative norm states
upon quantization is also back. Finally the longitudinal polarization in (9.10)
seem to diverge in the massless limit. In fact the longitudinal polarization
becomes proportional to the unwanted spin-0 part of the vector field in (9.4):
3µ → m−1 (pz , 0, 0, pz ) and A0ν ∝ kν → m−1 (pz , 0, 0, pz ). Related to this issue,
there is a contradiction with Wigner’s representation theory based on the little
group, see section 7.5, which says that whereas the massive vector should have
3 states per momentum, the massless vector should have only 2 that correspond
to the helicity eigenstates. All of these issues point toward the fact that we have
one degree of freedom in Aν that should somehow be removed.

9.4 Gauge invariance

All aforementioned issues are solved at once by realizing that the lagrangian
(9.11) has the following continuous symmetry called the gauge symmetry:

Aν (x) → Aν (x) + ∂ν α(x) , (9.16)

where α(x) is non-singular and real, otherwise arbitrary. This means that two
vector fields Aν (x) and A0ν (x) which differ by gradient of a function are physi-
cally equivalent, and the system is said to be gauge invariant. From here on I
will call the massless vector field A the gauge field.

Gauge symmetry is extremely important for solving the problems mentioned

above. Since α can be chosen arbitrarily we can use this freedom to kill one
whole real function at every space-time point. In particular we can use it to
kill both the spin-0 part of the vector field, A0ν in (9.4) and the longitudinal
polarization 3ν at the same time because they become the same in the massless
limit. This procedure of setting the gauge freedom by some physical constraint
is called the gauge fixing procedure. There are many different choices of gauge
fixing conditions

1. Lorentz gauge: ∂µ Aµ = 0 ,
9.5. GAUGE FIXING IN THE COULOMB GAUGE 99

2. Coulomb gauge: ∂i Ai = 0

3. Temporal gauge: A0 = 0 ,

etc. Of course, all the physical results such as scattering amplitudes, decay prob-
abilities etc. should be independent of the choice of gauge but certain gauges
have the advantage of making certain property of the theory manifest. For in-
stance the Lorentz gauge makes the Lorentz invariance manifest, the temporal
gauge makes the unitarity of the theory manifest, the Coulomb gauge simplifies
the canonical quantization etc.

9.5 Gauge fixing in the Coulomb gauge

Let us now work out the gauge-fixing procedure in the Coulomb gauge:
~ ·A
∇ ~ = 0. This can easily be reached by appropriately choosing α such that
~ ~
∇·A → ∇ ~ ·A~ + ∇2 α = 0, that is because, the Poisson equation ∇2 α = −∇~ ·A~
1
can always be solved The equations of motion (9.14) in the Coulomb gauge
when written in components read
 
−∇2 A0 + ∂t ∇ ~ ·A
~ = −∇2 A0 = 0 , (9.17)

and  
∂ 2 Ai − ∂ i ∇~ ·A
~ − ∂t A0 = ∂ 2 Ai + ∂t ∂i A0 = 0 (9.18)

Now, note that a generic gauge transformation Aµ → Aµ + ∂µ α which preserves

the Coulomb gauge should satisfy ∇2 αCoulomb = 0. This means that one can
choose this remaining gauge freedom in αCoulomb to completely get rid of A0
because A0 and αCoulomb satisfy the same condition (see (9.17)) in the Coulomb
gauge. Thus we can set
A0 = 0 ,
in this gauge which now fixes the entire gauge freedom of the theory. Then the
equation of motion for the remaining spatial components of the gauge field are
obtained from (9.18) as
∂ 2 Ai = 0 .
The solution is of the same form as for the massive vector field in (9.8) except
that the number of independent polarizations is now only two:
Z 2
X

Aν (x) = ˜
dp p) + e−ipx aλ∗ (p)λ∗
eipx aλ (p)λν (~ ν (~
p) , (9.19)
λ=1

with p2 = 0. This is because, as we found above, in the Coulomb gauge the

polarization vectors should satisfy

λ0 = 0 , pj λj = 0 ∀λ .
1 This is possible at least locally, up to possible global obstructions which we assume absent.
100 CHAPTER 9. VECTOR FIELDS AND GAUGE INVARIANCE

Now recall that for p2 = 0 one can go to a frame where the four-momentum reads
pµ = (ω, 0, 0, ω). Then it is immediate to see that there are only two linearly
independent solutions to these equations. Imposing normalization ∗µ µ = +1
one can choose them as

1µ = (0, 1, 0, 0) , 2µ = (0, 0, 1, 0) . (9.20)

These two polarizations are called transverse. One can also solve the equations
above by a different choice, for example,
1 1
+
µ = √ (0, 1, i, 0) , −
µ = √ (0, 1, −i, 0) . (9.21)
2 2
With this choice the vector field is circularly polarized and they correspond to
the right-handed and left-handed helicity eigenstates respectively.

9.6 Quantum Gauge Field: Canonical Quanti-

zation

Canonical quantization of the theory proceeds by first fixing the gauge and
then promoting the Poisson brackets between the canonical position and canon-
ical momenta to commutators in this gauge. Therefore the quantized field will
look different for different gauge choices. Of course the physical quantities
should in the end be independent of this choice. Let us work out canonical
quantization of the gauge field in the Coulomb gauge.

We have seen that the Coulomb gauge constrains both the divergence of the
spatial components and the temporal component of the gauge field as

∂i Ai = 0 , A0 = 0 .

Then the lagrangian in this gauge reads,

1 1 1
LEM = − F µν Fµν = ∂t Ai ∂t Ai − ∂j Ai ∂j Ai , (9.22)
4 2 2
where we used integration by parts. The canonical positions in this gauge are
Ai , hence the corresponding canonical moment are obtained as
δLEM
Πi = = ∂t Ai . (9.23)
δAi
The hamiltonian density then reads,
1 1
HEM = Πi ∂t Ai − LEM = Πi Πi + ∂j Ai ∂j Ai . (9.24)
2 2
The non-standard aspect of canonical quantization of a system with constraints
is the form of the canonical commutation relations. Normally we would require

[Πi (~x, t), Aj (~y , t)] = iδij δ 3 (~x − ~y ) . (9.25)

9.6. QUANTUM GAUGE FIELD: CANONICAL QUANTIZATION 101

This is indeed the correct form when the gauge field Aj does not have to satisfy
any constraint. It would be inconsistent with the constraint ∂i Ai = 0 because
when we take ∂i of both sides we see that the LHS vanishes but the RHS
does not. In fact, the Coulomb constraint projects the gauge field to a certain
combination of its components with vanishing divergence. If we denote the
gauge field that satisfies the Coulomb constraint as AC
i then this projection can
be written as
 
C ∂i ∂j
Ai (x) = Pij (x)Aj (x), Pij (x) = δij − 2
∇

in the position space or

!
ki kj
AC
i (k) = Pij (k)Aj (k), Pij (k) = δij −
|~k|2

in the momentum space. Therefore the correct form of the commutator for the
gauge field AC
i which does satisfy the constraint is given by taking the projection
of both sides of (9.25) and reads,
 
  ∂i ∂j
C
Πi (~x, t), Aj (~y , t) = i δij − 2 δ 3 (~x − ~y ) . (9.26)
∇

The quantum field is obtained from the classical solution (9.19) as usual by
promoting the coefficients ai and ai∗ to annihilation and creation operators
respectively:
Z 2
X

Aν (x) = ˜
dp p) + e−ipx aλ† (p)λ∗
eipx aλ (p)λν (~ ν (~
p) , (9.27)
λ=1

with p2 = 0. The polarization vectors satisfy

λ0 = 0 pj λj = 0 ∀λ

as in the classical solution. The vacuum is defined in the standard way by

requiring
aλ (p) |0i = 0 , ∀~
p, λ ,
and the one-particle states are created out of the vacuum by acting with the
creation operators
p, λi = aλ† (p) |0i .
|~
This quanta of the massless vector field is called the photon. It has momentum
p~, energy ω = |~
p| and polarization λ.

It is now straightforward (but tedious) to obtain the commutation relations

for the annihilation/creation operators starting from (9.26)
0 0
[aλ (k), aλ † (k 0 )] = (2π)3 δ 3 (~k − ~k 0 )2ωδ λλ , (9.28)

with other commutators vanishing.

102 CHAPTER 9. VECTOR FIELDS AND GAUGE INVARIANCE

9.7 Quantum Gauge Field: Path Integral Quan-

tization

Quantum vector fields can be studied using path integrals by formally writ-
ing down the path integral over all possible gauge field configurations just like
for scalars and spinors. The vacuum-to-vacuum transition amplitude for the
massless vector is
Z R 4 µ
Z[Jext ] = DAµ ei d x(LEM −Aµ Jext ) , (9.29)

where we introduced and “external current” that is a source for the gauge field
Aµ . To make sure that this defines vacuum-to-vacuum amplitude we have to
choose the tilted time contour in the complex plane as before.

Before we start using this to calculate probability amplitudes, however, we

need to discuss a couple of issues with this path integral. First, as we have
seen above, it is crucial to maintain the gauge-invariance of the theory to make
sure that the theory is free of negative norm states. The source term seems to
Rviolate the gauge symmetry Aµ → Aµ + ∂µ α which generates a term of the form
µ
d4 x∂µ αJext . An integration by parts shows that this term is proportional to
the divergence of the external current. Then this issue is solved by requiring
that any external current we couple to the gauge field should satisfy
µ
∂µ Jext = 0. (9.30)

This also makes sense from the point of view of counting the degrees of freedom.
In general we should introduce one source per degree of freedom. As we have
seen above, the gauge field has only 3 real degrees of freedom: one per each
component of Aµ minus the degree of freedom removed by gauge symmetry.
Therefore we need to couple it to 3 degrees of freedom in total which is precisely
µ
the case with Jext with the constraint (9.30).

However, there is another issue with (9.29). There is a new type of diver-
gence associated with this path integral which arises from the gauge symmetry:
As the action is invariant under a gauge transformation the value of the inte-
grand of the path integral is the same for any gauge field configurations Aµ and
A0µ that differ by ∂µ α for all choices of α. Since the path integral sums over
all possible configurations a particular value of the integrand that arises from
a particular configuration Aµ is counted infinitely many times, causing a diver-
gence. There is an elegant way to fix this issue by the so called Faddeev-Popov
procedure which I describe in Appendix C. For the sake of our discussion here it
suffices to proceed by introducing a lagrange multiplier to fix the gauge symme-
try inside the path integral. This can be done by introducing a term of the form
λG[A] where λ is a lagrange multiplier and G[A] = 0 some functional of Aµ .
That is, minimizing the lagrangian with respect to λ imposes the gauge-fixing
condition G[A] = 0 . We chose the Coulomb gauge to quantize the field in the
canonical formalism in the previous section. This gauge manifestly breaks the
Lorentz symmetry which is not a big issue in the canonical formalism as it treats
time and space differently by defining the time evolution with the Hamiltonian,
9.8. PHOTON PROPAGATOR 103

by defining the commutation relations at a fixed time slice etc. On the other
hand keeping the space-time symmetries e.g. Lorentz invariance manifest in
the path integral formalism is crucial as we derive all the associated conserved
charges, including the Hamiltonian from the invariance of the action. Thus it
makes more sense to use the Lorentz gauge ∂µ Aµ = 0 instead, so G[A] should
be proportional to ∂µ Aµ = 0. We also have to make sure that the kinetic term
in the full lagrangian which involves the time derivatives is positive definite for
stability. The simplest term to add then is a quadratic of the form −λ(∂ · A)2
with λ > 0. For historical reasons, one denotes this term as −(∂ · A)2 /2ξ and
the theory that is defined with this gauge fixing term is said to be in the “Rξ
gauge”. Thus, our generating function is
Z R 4 µν µ 2 µ
Z[Jext ] = DAµ ei d x(− 4 Fµν F − 2ξ (∂µ A ) −Aµ Jext ) .
1 1
(9.31)

One obtains the n-point functions of the gauge fields from this generating func-
tion by taking variational derivatives with respect to Jext as usual.

9.8 Photon propagator

We can now derive the propagator of the photon directly from the path
integral (9.31). For this we use the method explained below equation (8.62):
Given any lagrangian density of a field φA with some generic indices A with the
quadratic part −φA OAB φB the propagator is given by
−1
h0| T φA (x)φB (y) |0i = −i OAB (x, y) . (9.32)

Here the lagrangian in (9.31) written in momentum space reads

 
1 µν µν µν 2 µ ν 1
LEM = Aµ (k)O (k)Aν (k) , O = −η k + k k 1 − . (9.33)
2 ξ
−1
We can invert this easily by making an ansatz of the form2 and Oµν = A(k 2 )ηµν +
2 µν −1 µ
B(k )kµ kν and solving the defining equation O Oνρ = δρ to determine the
functions A(k 2 ) and B(k 2 ). One finds
 
−1 1 kµ kν
Oµν (k) = − 2 ηµν − 2 (1 − ξ) .
k k

Thus photon’s Feynman propagator, a.k.a time-ordered two-point function is

Z k k
d4 k ik(x−y) ηµν − µk2 ν (1 − ξ)
h0| T Aµ (x)Aν (y) |0i = −i e (9.34)
(2π)4 k 2 − i

where the i in the denominator arises from the tilted time contour in the path
integral as usual. The fact that the photon propagator depends on the gauge
fixing parameter ξ seems disappointing at first. However, this dependence neatly
drops out of all physical quantities such as the scattering amplitudes and decay
2 This is the most general symmetric tensor consistent with Lorentz symmetry.
104 CHAPTER 9. VECTOR FIELDS AND GAUGE INVARIANCE

probabilities. One can easily see this by the fact that a scattering amplitude that
involves an external photon involves the external wave-function µ (k) contracted
with the propagator, that is, the propagator (9.34) will always appear contracted
with µ (k) in a physical observable, and the second term in (9.34) then drops out
of this quantity because of the Lorentz gauge condition k µ µ = 0, see (9.27). The
result that the gauge fixing procedure does not have any bearing on the physical
quantities in general is due to a general theorem, called the Ward identity which
shows that the second term in the numerator of the propagator (9.34) drops out
of all scattering amplitudes. I will not prove this identity in these introductory
lectures, see Peskin&Schroeder, or Schwartz for a good discussion.
Chapter 10

Quantum electrodynamics

10.1 Coupling photon to matter

In classical electrodynamics coupling of the electromagnetic fields to matter

is introduced by writing the charge density and current of charged matter on
the RHS of the Maxwell’s equations as

∂µ F µν = J ν , (10.1)

~ is the 4-vector that comprises of the charge density ρ and the

where J µ = (ρ, J)
~
current J. These, together with the Bianchi identity (9.15) gives the Maxwell’s
equations in the presence of matter. By taking the ∂ν of both sides of (10.1) we
obtain the continuity equation:

∂ν J ν = 0 . (10.2)

This is just the statement that charge cannot be created or destroyed: if the
charge density in a region of space is increasing in time then it should be that
some charge is being transported by a current from outside. This is the conser-
vation of charge.

To obtain (10.1) in from lagrangian we need to modify (9.11) as

1
LEM = − F µν Fµν − Aµ J µ , (10.3)
4

whose variation with respect to Aµ now yields (10.1). This lagrangian is in-
variant under the gauge transformation Aµ → Aµ + ∂µ α despite the second
term. This is because an integration by parts in the second term yields α∂µ j µ
which vanishes by the continuity equation (10.2). In other words, if we want
to maintain gauge invariance in the presence of matter we need to require that
the charge current that couples to.the gauge field is conserved. The coupling of
matter to gauge fields in the form (10.3) is called the minimal coupling.

105
106 CHAPTER 10. QUANTUM ELECTRODYNAMICS

Quite generally conservation of a current is associated to a continuous sym-

metry in the system. Noether’s theorem states that a continuous global sym-
metry leads to existence of a conserved current in the system. If lagrangian L
for the fields φa is invariant under an infinitesimal transformation of these fields
δφa then there exists a current
µ δL
JN oether = δφa , (10.4)
δ∂µ φa
µ
which is conserved: ∂µ JN oether = 0. It is also true that a conserved current
necessitates presence of a global continuous symmetry in the lagrangian. You
µ
can easily see this by integrating the equation ∂µ JN oether = 0 over space and
making an integration by parts on the spatial components, hence obtaining
Z
∂t Q = 0, Q = d3 xJ 0 .

As time translations in a quantum system is generated by the hamiltonian this

means that [H, Q] = 0 hence the system possesses a symmetry. That this is a
continuous symmetry rather than a discrete one is also obvious by this argument.

We learned that in order to couple a gauge theory to matter with minimal

coupling we need the matter part of the system to possess a continuous global
symmetry. The simplest kind of such symmetry is a phase rotation of the fields
φa → eiα φa . You studied this in the case of a complex scalar field in the
exercises. This symmetry is also present in the Dirac lagrangian:
Lψ = iψ̄∂/ψ − mψ̄ψ is invariant under the transformations ψ → exp(iα)ψ,
ψ̄ → exp(iα)ψ̄. The conserved Noether’s current is obtained by substituting ψ
in place of the generic field φa in (10.4) as
µ
JD = eψ̄γ µ ψ , (10.5)
where we included the unit charge of an electron e in the normalization.

In passing, let us provide another derivation of this conserved current which

is more direct and does not require Noether’s formula. Assume for the moment
that you make the parameter of the symmetry transformation space-time de-
pendent. This means in the case of the U(1) electric charge symmetry that we
consider the transformation
ψ → eiα(x) ψ , ψ̄ → e−iα(x) ψ̄ .
Then, for infinitesimal α(x) the Dirac action changes as
Z
SD → SD − e d4 x (∂µ α(x)) ψ̄γ µ ψ ,

which indeed vanishes for constant α confirming that this is a symmetry of SD .

However, one can also integrate by parts the second term on the RHS and learn
that (10.5) is conserved by consistency.

Then, we can minimally couple Dirac fermions to gauge field by taking the
µ
current in (10.3) to be JD . The resulting theory is called Quantum Electrody-
namics, QED for short:
1
LQED = − F µν Fµν + iψ̄∂/ψ − mψ̄ψ − e Aµ ψ̄γ µ ψ, . (10.6)
4
10.2. COVARIANT DERIVATIVES AND ‘GAUGING A SYMMETRY’ 107

This is the quantum theory of Dirac fermions interacting with the electromag-
netic force, and it is, as far as we know today, the most successful theory in
physics. Agreement of theory with experiment is unprecedented by any other
quantum or classical theory that we know today.

10.2 Covariant derivatives and ‘gauging a sym-

metry’

We note that the QED lagrangian (??) can be neatly written as follows

1
LQED = − F µν Fµν + iψ̄D
/ψ − mψ̄ψ , (10.7)
4
where we defined the covariant derivative

Dµ ≡ ∂µ − i e Aµ . (10.8)

The meaning of the covariant derivative becomes clear by considering the al-
ternative derivation of the Noether’s current above where we assumed a space-
time dependent symmetry transformation parameter α(x). In fact the gauge
transformation Aµ → Aµ + ∂µ α(x) is precisely what is required to make the
lagrangian (10.6) invariant under ψ → exp(iα(x))ψ , ψ̄ → exp(−iα(x))ψ̄.
This combined transformation is the gauge symmetry of the QED lagrangian:

ψ → eiα(x) ψ , ψ̄ → e−iα(x) ψ̄ , Aµ → Aµ + ∂µ α(x) . (10.9)

This is not a global symmetry anymore, it is a local symmetry of the action.

Yet, the word “symmetry” is a bit of a misnomer in some sense because all
physical observables are independent of the choice of the gauge parameter α.
Unlike a global symmetry which for example implies that the scattering ampli-
tudes should have a particular form to satisfy invariance under it, or implies
superselection rules between transition amplitudes etc., the gauge symmetry is
just a redundancy of our description of the physical system in terms of a local
field theory. It is extremely important however, as, the only known forces that
we know in the universe (including gravitation) arise from gauge symmetries.

Given any continuous global symmetry of the theory, one can try to gauge
it, i.e. one can promote the transformation parameter to a function of space-
time by including a gauge field and require that the transformation of the gauge
field cancels that of the derivative of the space-time dependent phase of the field
just as in the example above. A covariant derivative achieves this neatly as it
is precisely the combination that transforms covariantly:

Dµ ψ → eiα(x) Dµ ,

which means that once all the derivatives in the lagrangian are written in terms
of covariant derivatives it is guaranteed to be invariant. This procedure of
promoting ordinary derivatives to covariant derivatives is called gauging a global
symmetry. It works perfectly fine for all global continuous symmetries as long
108 CHAPTER 10. QUANTUM ELECTRODYNAMICS

p −i(−p/+m)
Spinors : = p2 +m2 −i

p −igµν
Photons : µ ν = p2 −i

Figure 10.1: Photon and spinor propagators in the momentum space. The
photon propagator is in the Feynman gauge ξ = 1.

as the symmetry is free of so-called anomalies. I will not dwell into anomalies
here, this will be a subject of a more advanced QFT course. It suffices here to
assume that there are no obstructions in gauging the electric U(1) symmetry of
the Dirac theory.

10.3 Feynman rules for QED

We are now ready to calculate scattering amplitudes in QED. The momen-

tum space Feynman rules are easy to state. In fact we have already driven the
Feynman rules for the spinor and gauge field propagators in sections 8.7 and
9.8. We summarize them in figure 10.1. The external states for the photon, the
spinor and the anti-spinor are shown in figure 10.2. There is a single vertex in
the theory which you can read off directly from the lagrangian (10.7), which we
show in figure 10.3.

10.4 Pair annihilation

As an application of the Feynman rules derived in the previous section let

us compute the scattering amplitude for an electron and a positron annihilating
into a pair of photons: e− e+ → γ γ. The two topologically distinct Feynman
graphs at the tree level are shown in figure 10.4 Using the Feynman rules for
QED in figure 10.2 we find the following result for the scattering amplitude:
 
p/1 − k/1 − m
T = e2 µ (k1 )ν (k2 )v̄(p2 ) iγν γ µ + (k 1 ↔ k 2 ) u(p1 ) ,
(p1 − k1 )2 + m2 − i
(10.10)
This is the result for the polarized scattering amplitude i.e. the scattering
amplitude depends on the spin polarizations of both the incoming fermions s, s0
and the outgoing photons λ, λ0 , that is the amplitude is in fact a tensor of the
form Tss0 λλ0 . The spin unpolarized result is obtained by summing over these
spins. We will not deal with spin and polarization sums in these lectures. You
can consult to either of Peskin&Schroeder or Schwartz for a readable account.
10.4. PAIR ANNIHILATION 109

p
= ✏µ* (p) incoming photon
µ
p
= ✏⇤µ (p) outgoing photon
p µ
= ✏µ (p) incoming photon
µ
pp
= ✏ū⇤µs(p)
(p) incoming electron
outgoing photon
µ
pp
= ✏uµs(p)
(p) outgoing
incoming electron
photon
µ p s
= ū (p) incoming electron
p
= ✏⇤µ (p) outgoing photon
pp µ
= uv̄ss(p) incoming
= positron
outgoing electron
pp
== ūvss(p) incoming
outgoing positron
electron
p
= v̄ s (p) incoming positron
p s
Figure 10.2: Momentum = uFeynman
space (p) outgoing
rules for the electron
external states in QED,
p
photon, electron and positron. s Dot denotes the location where the external
= v (p) outgoing positron
particle is inserted in the diagram. Thick arrows indicate the direction of mo-
mentum flow. Note that the direction of the arrow coincides with the direction
of momentum for p
electron, but it
s is the opposite for the positron.
= v̄ (p) incoming positron

p
= v s (p) outgoing positron

= i e γµ

Figure 10.3: Momentum space Feynman rule for the QED vertex.
110 CHAPTER 10. QUANTUM ELECTRODYNAMICS

µ ν µ ν
k1 k2
k1 k2

p1 p1 − k1 p2
+p p1 − k1 p2
1

Figure 10.4: Two topologically distinct diagrams that contribute to the process
e− e+ → γ γ at the tree level.

Appendix A: Orthogonality of the asymptotic Hilbert

space

In section 6.4 we outlined the criteria for orthogonality of the asymptotic

Fock space of QFT and left the proof of criterion iii, namely the proof that
the states created by the one-particle creation operator a†1 has no overlap with
the multi-particle states to this appendix. To prove this statement we need to
redefine a†1 in a slightly more general fashion as described in the first footnote
in section 5.1:
Z
|~ k 1 |2
k−~
†
a1 (t) ≡ d3 kf1 (~k)a† (k) , f1 ∝ e− 4σ2 . (10.11)

The creation operator used in section 6.4 corresponds to the σ → 0, delta-

†
function limitP R 3wave-packet. Condition iii states limt→±∞ hψ| a1 (t) |0i = 0,
of this
where |ψi = n d pψn (p) |p, ni denotes a multi-particle state. We first note

hp, n| φ(x) |0i = e−ipx hp, n| φ(0) |0i ≡ e−ipx An (p) .

Using (10.11) and (5.4) we find

XZ Z Z  ←→ 
hψ| a†1 (t) |0i = −i d3
pψn∗ (p) 3
d kf1 (k) d3 x eikx ∂0 e−ipx An (p)
n
XZ 0
−k0 )t
= d3 pψn∗ (p)f1 (p)An (p)(p0 + k 0 )ei(p ,
n

where in the last line we performed the ~x integral and used the resulting delta-
p − ~k) to eliminate thep~k integral. The multipartticle
function δ 3 (~ p and single
particle energies are given by p0 = |~ p| + M 2 and k 0 = |~
p| + m2 where M is
the total mass of the multiparticle state |ψi, Clearly RM ≤ 2m > m hence p0 >
k 0 . Recall the Riemann-Lebesque lemma limt→±∞ d3 pF (p) exp(iz(p)t) = 0
for z(p) > 0 and F (p) has a compact support in p, or it can be defined as a limit
of a function with a compact support in p. The compact support can indeed
be assured by an appropriate choice of the function f1 (p) like the Gaussian in
(10.11). Then the proof is complete. Note that the proof extends to the delta-
function peaked creation operator we used in the text because delta function
can always be defined as a limit of a function with compact support.
10.4. PAIR ANNIHILATION 111

Appendix B: Subgroups of Lorentz Transforma-

tion

We can find the determinant of a Lorentz matrix as:

det(ΛT δΛ) = det(δ) =⇒ det(Λ)2 = 1.

We divide the Lorentz transformations into two groups as:

(i) Proper: detΛ = 1

(ii) Improper: detΛ = −1

A second dividing occurs as:

(i) Orthochronous: Λ00 ≥ 1

(ii) Non-orthochronous: Λ00 ≤ −1.

where we have (Λ00 )2 = Λi0 Λi0 + 1.

Parity and Time Reversal

In QFT, we generally mean orthochronous and proper Lorentz transforma-

tions by ”Lorentz transformations”, those that are continuously connected to
δνµ .

Let us define the parity operator:

Pνµ = diag(+1, −1, −1, −1)

Then, under parity transformation we have x0 → x0 and ~x → −~x. Given

ΛT δΛ = Λ, we know that Λ is proper. Then, P Λ is improper.

Similarly, the time reversal can be written as:

Tνµ = diag(−1, 1, 1, 1)

then under time reversal we have x0 ⇐⇒ −x0 and ~x → −~x.

Appendix C: Spinor formulas and derivation of

anticommutation relations

In order to obtain the anticommutators (8.46) and (8.47) between the spinor
creation and annihilation operators, we need to express the creation/annihilation
112 CHAPTER 10. QUANTUM ELECTRODYNAMICS

operators in terms of the fields ψ(x) and ψ̄(x). To do so, let us first prove the
so called “Gordon identities”:

p0 )γ µ us (~
2mūs0 (~ p0 ) [(p + p0 )µ − 2iM µν (p0 − p)ν ] us (~
p) = ūs0 (~ p) (10.12)
p0 )γ µ vs (~
−2mv̄s0 (~ p0 ) [(p + p0 )µ − 2iM µν (p0 − p)ν ] vs (~
p) = v̄s0 (~ p) . (10.13)

We start by writing
 
1 µ ν 1 µ ν
γµp
/= γ , γ + [γ , γ ] pν = −pµ − 2iM µν pν
2 2
 
0 µ 1 ν µ 1 ν µ
p/γ = γ , γ + [γ , γ ] p0ν = −p0µ + 2iM µν p0ν .
2 2
Equations (10.12) and (10.13) follow from adding these equations and using the
spinor equations of motion:

(p
/ + m)us (~
p) = 0, (−p
/ + m)vs (~
p) = 0, ūs (~
p)(p/ + m) = 0, v̄s (~
p)(−p/ + m) = 0 .

An important special case is obtained by setting p = p0 and using the orthogo-

nality conditions

ūs0 (~
p)us (~
p) = 2mδss0 , v̄s0 (~ p) = −2mδss0 .
p)vs (~ (10.14)

We find:

p)γ µ us (~
ūs0 (~ p) = 2pµ δss0 , p)γ µ vs (~
v̄s0 (~ p) = 2pµ δss0 . (10.15)

From (8.42) we obtain

Z X 1 1 2iωt s†

3 −ipx s s s
d xe ψ(x) = a (p)u (~
p) + e b (−p)v (−~
p) .
s
2ω 2ω

p)γ 0 and using (10.15) and ūs0 (~

Multiplying this on the left by ūs0 (~ p)γ 0 vs (−~
p) =
0 (which can be obtained using the tricks above) we finally obtain
Z
as (p) = d3 xe−ipx ūs (~ p)γ 0 ψ(x) . (10.16)

Hermitean conjugate of this gives

Z
as† (p) = d3 xeipx ψ̄(x)γ 0 us (~
p) . (10.17)
R
Repeating the same method starting from d3 xe+ipx ψ(x) instead we find
Z
b (p) = d3 xeipx v̄s (~
s†
p)γ 0 ψ(x) , (10.18)

and Z
bs (p) = d3 xe−ipx ψ̄(x)γ 0 vs (~
p) . (10.19)

The anticommutators (8.46) and (8.47) then immediately follows from (10.16),
(10.17), (10.19) and (10.18).