Quantum Field Theory I

Lecture Notes

ETH Zurich, 2019 HS

Prof. N. Beisert

c 2012–2019 Niklas Beisert
This document as well as its parts is protected by copyright.
This work is licensed under the Creative Commons
“Attribution-NonCommercial-ShareAlike 4.0 International”
License (CC BY-NC-SA 4.0).

To view a copy of this license, visit:

https://creativecommons.org/licenses/by-nc-sa/4.0
The current version of this work can be found at:
http://people.phys.ethz.ch/~nbeisert/lectures/
Contents

Contents 3

Overview 5
0.1 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
0.2 Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
0.3 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1 Classical and Quantum Mechanics 1.1

1.1 Classical Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1
1.2 Hamiltonian Formulation . . . . . . . . . . . . . . . . . . . . . . . . 1.2
1.3 Quantum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3
1.4 Quantum Mechanics and Relativity . . . . . . . . . . . . . . . . . . 1.6
1.5 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.9

2 Classical Free Scalar Field 2.1

2.1 Spring Lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1
2.2 Continuum Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4
2.3 Relativistic Covariance . . . . . . . . . . . . . . . . . . . . . . . . . 2.6
2.4 Hamiltonian Field Theory . . . . . . . . . . . . . . . . . . . . . . . 2.7

3 Scalar Field Quantisation 3.1

3.1 Quantisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1
3.2 Fock Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4
3.3 Complex Scalar Field . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7
3.4 Correlators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.9
3.5 Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.14

4 Symmetries 4.1
4.1 Internal Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1
4.2 Spacetime Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . 4.5
4.3 Poincaré Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.9
4.4 Poincaré Representations . . . . . . . . . . . . . . . . . . . . . . . . 4.11
4.5 Discrete Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . 4.15

5 Free Spinor Field 5.1

5.1 Dirac Equation and Clifford Algebra . . . . . . . . . . . . . . . . . 5.1
5.2 Poincaré Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4
5.3 Discrete Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7
5.4 Spin Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.9
5.5 Graßmann Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 5.12
5.6 Quantisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.14

3
 5.7 Complex and Real Fields . . . . . . . . . . . . . . . . . . . . . . . . 5.18
5.8 Massless Field and Chiral Symmetry . . . . . . . . . . . . . . . . . 5.20

6 Free Vector Field 6.1

6.1 Classical Electrodynamics . . . . . . . . . . . . . . . . . . . . . . . 6.1
6.2 Gauge Fixing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3
6.3 Particle States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.6
6.4 Casimir Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12
6.5 Massive Vector Field . . . . . . . . . . . . . . . . . . . . . . . . . . 6.16

7 Interactions 7.1
7.1 Interacting Lagrangians . . . . . . . . . . . . . . . . . . . . . . . . 7.1
7.2 Interacting Field Operators . . . . . . . . . . . . . . . . . . . . . . 7.6
7.3 Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 7.9

8 Correlation Functions 8.1

8.1 Interacting Time-Ordered Correlators . . . . . . . . . . . . . . . . . 8.1
8.2 Time-Ordered Products . . . . . . . . . . . . . . . . . . . . . . . . 8.2
8.3 Some Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6
8.4 Feynman Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.13
8.5 Feynman Rules for QED . . . . . . . . . . . . . . . . . . . . . . . . 8.17

9 Particle Scattering 9.1

9.1 Scattering Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1
9.2 Cross Sections and Matrix Elements . . . . . . . . . . . . . . . . . . 9.3
9.3 Electron Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4
9.4 Pair Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.12
9.5 Loop Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.15

10 Scattering Matrix 10.1

10.1 Asymptotic States . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.1
10.2 S-Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4
10.3 Time-Ordered Correlators . . . . . . . . . . . . . . . . . . . . . . . 10.7
10.4 S-Matrix Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . 10.10
10.5 Unitarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.15

11 Loop Corrections 11.1

11.1 Self Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1
11.2 Loop Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.4
11.3 Regularisation and Renormalisation . . . . . . . . . . . . . . . . . . 11.7
11.4 Counterterms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.11
11.5 Vertex Renormalisation . . . . . . . . . . . . . . . . . . . . . . . . . 11.14

Schedule of Lectures 9

4
Quantum Field Theory I Chapter 0
ETH Zurich, 2019 HS Prof. N. Beisert
18. 12. 2019

0 Overview
Quantum field theory is the quantum theory of fields just like quantum mechanics
describes individual quantum particles. Here, a the term “field” refers to one of
the following:
• A field of a classical field theory, such as electromagnetism.
• A wave function of a particle in quantum mechanics. This is why QFT is
sometimes called “second quantisation”.
• A smooth approximation to some property in a solid, e.g. the displacement of
atoms in a lattice.
• Some function of space and time describing some physics.
Usually, excitations of the quantum field will be described by “particles”. In QFT
the number of these particles is not conserved, they are created and annihilated
when they interact. It is precisely what we observe in elementary particle physics,
hence QFT has become the mathematical framework for this discipline.
This lecture series gives an introduction to the basics of quantum field theory. It
describes how to quantise the basic types of fields, how to handle their quantum
operators and how to treat (sufficiently weak) interactions. We will focus on
relativistic models although most methods can in principle be applied to
non-relativistic condensed matter systems as well. Furthermore, we discuss
symmetries, infinities and running couplings. The goal of the course is a derivation
of particle scattering processes in basic QFT models.
This course focuses on canonical quantisation along the lines of ordinary quantum
mechanics. The continuation of this lecture course, QFT II, introduces an
alternative quantisation framework: the path integral.1 It is applied towards
formulating the standard model of particle physics by means of non-abelian gauge
theory and spontaneous symmetry breaking.

What Else is QFT? There are many points of view.

After attending this course, you may claim QFT is all about another 1000 ways to
treat free particles and harmonic oscillators. True, these are some of the few
systems we can solve exactly in theoretical physics; almost everything else requires
approximation. After all, this is a physics course, not mathematics!
If you look more carefully you will find that QFT is a very rich subject, you can
learn about many aspects of physics, some of which have attained a mythological
status:
1
The path integral is much more convenient to use than canonical quantisation discussed here.
However, some important basic concepts are not as obvious as in canonical quantisation, e.g. the
notion of particles, scattering and, importantly, unitarity.

5
 • anti-particles, anti-matter,
• vacuum energy,
• tachyons,
• ghosts,
• infinities,
• mathematical (in)consistency.

Infinities. How to deal with infinities?

There is a famous quote due to Dirac about QED (1975): “I must say that I am
very dissatisfied with the situation, because this so-called ‘good theory’ does
involve neglecting infinities which appear in its equations, neglecting them in an
arbitrary way. This is just not sensible mathematics. Sensible mathematics
involves neglecting a quantity when it is small – not neglecting it just because it is
infinitely great and you do not want it!”
This is almost true, but QFT is neither neglecting infinities nor in an arbitrary
way.
Infinities are one reason why QFT is claimed to be mathematically ill-defined or
even inconsistent. Yet QFT is a well-defined and consistent calculational
framework to predict certain particle observables to extremely high precision.
Many points of view; one is that it is our own fault: QFT is somewhat idealised; it
assumes infinitely extended fields (IR) with infinite spatial resolution (UV);2 there
is no wonder that the theory produces infinities. Still, it is better to stick to
idealised assumptions and live with infinities than use some enormous discrete
systems (actual solid state systems).
There is also a physics reason why these infinities can be absorbed somehow: Our
observable everyday physics should neither depend on how large the universe
actually is (IR) nor on how smooth or coarse-grained space is (UV).
We can in fact use infinities to learn about allowable particle interactions. This
leads to curious effects: running coupling and quantum anomalies.
More later, towards the end of the semester.

Uniqueness. A related issue is uniqueness of the formulation. Alike QM, QFT

does not have a unique or universal formulation.
For instance, many meaningful things in QM/QFT are actually equivalence classes
of objects. It is often more convenient or tempting to work with specific
representatives of these classes. However, one has to bear in mind that only the
equivalence class is meaningful, hence there are many ways to describe the same
physical object.
The usage of equivalence classes goes further, it is not just classes of objects.
Often we have to consider classes of models rather than specific models. This is
something we have to accept, something that QFT forces upon us.
2
The UV and the IR are the two main sources for infinities.

6
We will notice that QFT does what it wants, not necessarily what we want. For
example, we cannot expect to get what we want using bare input parameters.
Different formulations of the same model naively may give different results. We
must learn to adjust the input to the desired output, then we shall find agreement.
We just have to make sure that there is more output than input; otherwise QFT
would be a nice but meaningless exercise because of the absence of predictions.
Another nice feature is that we can hide infinities in these ambiguities in a
self-consistent way.

Enough of Talk. Just some words of warning: We must give up some views on
physics you have become used to, only then you can understand something new.
For example, a classical view of the world makes understanding quantum
mechanics harder. Nevertheless, one can derive classical physics as an
approximation of quantum physics, once one understands the latter sufficiently
well.
Let us start with something concrete, we will discuss the tricky issues when they
arise.

Important Concepts. Some important concepts of QFT that will guide our
way:
• unitarity – probabilistic framework.
• locality – interactions are strictly local.
• causality – special relativity.
• symmetries – exciting algebra and geometry.
• analyticity – complex analysis.

0.1 Prerequisites
Prerequisites for this course are the core courses in theoretical physics of the
bachelor syllabus:
• Classical mechanics (brief review in first lecture)
• Quantum mechanics (brief review in first lecture)
• Electrodynamics (as a simple classical field theory)
• Mathematical methods in physics (HO, Fourier transforms, . . . )

0.2 Contents
1. Classical and Quantum Mechanics (105 min)
2. Classical Free Scalar Field (110 min)
3. Scalar Field Quantisation (210 min)
4. Symmetries (240 min)
5. Free Spinor Field (270 min)

7
 6. Free Vector Field (250 min)
7. Interactions (170 min)
8. Correlation Functions (240 min)
9. Particle Scattering (185 min)
10. Scattering Matrix (200 min)
11. Loop Corrections (225 min)
Indicated are the approximate number of lecture minutes for each chapter.
Altogether, the course consists of 51 lectures of 45 minutes including one overview
lecture.

0.3 References
There are many text books and lecture notes on quantum field theory. Here is a
selection of well-known ones:
• M. E. Peskin, D. V. Schroeder, “An Introduction to Quantum Field Theory”,
Westview Press (1995)
• C. Itzykson, J.-B. Zuber, “Quantum Field Theory”, McGraw-Hill (1980)
• P. Ramond, “Field Theory: A Modern Primer”, Westview Press (1990)
• M. Srendnicki, “Quantum Field Theory”, Cambridge University Press (2007)
• M. Kaku, “Quantum Field Theory”, Oxford University Press (1993)
• online: D. Tong, “Quantum Field Theory”, lecture notes,
http://www.damtp.cam.ac.uk/user/tong/qft.html
• online: M. Gaberdiel, “Quantenfeldtheorie”, lecture notes (in German),
http://people.phys.ethz.ch/~mrg/Lectures/QFT.pdf
• ...
Peskin & Schroeder may be closest to this lecture course, but we will not follow it
literally.

8
Quantum Field Theory I Chapter 1
ETH Zurich, 2019 HS Prof. N. Beisert
23. 09. 2019

1 Classical and Quantum Mechanics

To familiarise ourselves with the basics of quantum field theory, let us review some
elements of classical and quantum mechanics. Then we shall discuss some issues of
combining quantum mechanics with special relativity.

1.1 Classical Mechanics

Consider a classical non-relativistic particle in a potential. In Lagrangian
mechanics it is described by the position variables q i (t) and the action functional
S[q] 1 2 Z t2
dt L q i (t), q̇ i (t); t .


S[q] = (1.1)
t1

A typical Lagrangian function is

L(~q, ~q˙) = 21 m~q˙ 2 − V (~q). (1.2)

with mass m and V (q) as the external potential.

A classical path extremises (minimises) the action S. One therefore determines the
saddle point δS = 0 by variation of the action3
Z t2  
i ∂L i ∂L
δS = dt δq (t) i + δ q̇ (t) i
t1 ∂q ∂ q̇
Z t2   Z t=t2  
i ∂L d ∂L i ∂L
= dt δq (t) − + d δq (t) i . (1.3)
t1 ∂q i dt ∂ q̇ i t=t1 ∂ q̇

The first term is the equation of motion (Euler–Lagrange)

δS ∂L d ∂L
i
= i− = 0. (1.4)
δq (t) ∂q dt ∂ q̇ i

The second term due to partial integration is the boundary equation of motion;
usually we ignore it.4
1
In many cases, L is time-independent: L(q i , q̇ i ; t) = L(q i , q̇ i ).
2
A single time derivative q̇ i usually suffices.
3
Einstein summation convention: there is an implicit sum over all index values for pairs of
matching upper/lower indices.
4
More precisely, we usually fix the position qi (tk ) = const (Dirichlet) or the momentum
∂L/∂ q̇ i (tk ) = 0 (Neumann) at the boundary.

1.1
Example. Throughout this chapter we will use the harmonic oscillator and the
free particle as an example to illustrate the abstract formalism. The harmonic
oscillator is described by the following Lagrangian function and corresponding
equation of motion

L(~q, ~q˙) = 21 m~q˙ 2 − 21 mω 2~q 2 , −m(~q¨ + ω 2 ~q) = 0. (1.5)

For ω = 0 this system becomes a free particle.

1.2 Hamiltonian Formulation

The Hamiltonian framework is the next step towards canonical quantum
mechanics.
First, define the momentum pi conjugate to q i as5
∂L
pi = (1.6)
∂ q̇ i

and solve for q̇ i = q̇ i (q, p; t).6 Phase space is defined as the space of the position
and momentum variables (q i , pi ).
The Lagrangian function L(q, q̇; t) is replaced by the Hamiltonian function
H(q, p; t) on phase space. We define H(q i , pi ; t) as the Legendre transformation of
L
H(q, p; t) = pi q̇ i (q, p; t) − L q, q̇(q, p; t); t .


(1.7)
Let us express the equations of motion through H: A variation of the Hamiltonian
function w.r.t. all q i and pi reads
∂L
δH = δpi q̇ i − δq i , (1.8)
∂q i
where we substituted the definition of the momenta pi twice to cancel four further
terms that arise. We use the Euler–Lagrange equation and momenta to simplify
the variation further
δH = δpi q̇ i − δq i ṗi . (1.9)
Comparing this expression to the general variation
δH = δq i (∂H/∂q i ) + δpi (∂H/∂pi ) we obtain the Hamiltonian equations of motion
∂H ∂H
q̇ i = and ṗi = − . (1.10)
∂pi ∂q i

Next, we introduce the Poisson bracket for two functions f, g on phase space
∂f ∂g ∂f ∂g
{f, g} := − . (1.11)
∂q i ∂pi ∂pi ∂q i
5
This is a choice, one might also use different factors or notations.
6
We suppose the equation can be solved for q̇.

1.2
The Poisson bracket allows to express the time evolution for phase space functions
f (q, p; t) 7
df ∂f
= − {H, f }. (1.12)
dt ∂t
In particular, this works well for the functions f (q, p; t) = q i and f (q, p; t) = pi , and
yields the canonical equations of motion.

Example. For the harmonic oscillator we find p~ = m~q˙ and

1 2 1
H = p~·~q˙ − 12 m~q˙ 2 + 12 mω 2~q 2 = ~p + 2 mω 2~q 2 . (1.13)
2m
The Hamiltonian equations of motion read
∂H 1
~q˙ = −{H, ~q} = = p~,
∂~p m
∂H
p~˙ = −{H, p~} = − = −mω 2 ~q. (1.14)
∂~q
A convenient change of variables reads
1 1
~a = √ (mω~q +ı̊~p), ~a∗ = √ (mω~q −ı̊~p). (1.15)
2mω 2mω
Using these coordinates of phase space, the Poisson brackets read
∂f ∂g ∂f ∂g
{f, g} = −ı̊ ∗
+ı̊ . (1.16)
∂ai ∂ai ∂a∗i ∂ai
We obtain a separated first-order time evolution for ~a, ~a∗

H = ω~a∗ ·~a, ~a˙ = −ı̊ω~a, ~a˙ ∗ = +ı̊ω~a∗ . (1.17)

Note that the Poisson bracket with this Hamiltonian simply measures the degree of
a function in a vs. a∗
∂f ∂f
{H, f } =ı̊ωai −ı̊ωa∗i . (1.18)
∂a i ∂a∗i
E.g. for f = (a)m (a∗ )n one finds {H, f } =ı̊ω(m − n)f .

1.3 Quantum Mechanics

Let us revisit the canonical quantisation procedure highlighting the harmonic
oscillator. Quantum field theory is built on the same methods, and we shall
encounter the same kinds of problems, yet in some more elaborate fashion.
In canonical quantisation, classical objects are replaced by elements of linear
algebra:
7
The Hamiltonian H is itself a phase space function.

1.3
 • the state (q i , pi ) becomes a vector |ψi in a Hilbert space V ;
• a phase space function f becomes a linear operator F on V ;
• Poisson brackets {f, g} become commutators −ı̊~−1 [F, G].8
The equation of motion for a state (Schrödinger) is a wave equation
d
ı̊~ |ψ(t)i = H|ψ(t)i. (1.19)
dt
The (normalised) wave function of a state has a probabilistic interpretation:
|hφ|ψi|2 is the probability of finding the system described by |ψi in the state |φi.
This requires the following essential features of wave functions:
• hψ|ψi is positive;
• hψ|ψi can be normalised to 1 by scaling |ψi;
• hψ|ψi is conserved.
Conservation requires
d  1
hψ|ψi = hψ|(H − H † )|ψi = 0.

(1.20)
dt ı̊~
Therefore, the Hamiltonian must be hermitian (self-adjoint). We then have time
evolution with a unitary operator U (t2 , t1 )
|ψ(t2 )i = U (t2 , t1 )|ψ(t1 )i. (1.21)
The matrix element hψ|F |ψi describes the expectation value of the operator F in
state |ψi. Curiously, it obeys the quantum analog of the classical time evolution
 
d ∂F 1
hψ|F |ψi = hψ| − [H, F ] |ψi. (1.22)
dt ∂t ı̊~

Example. For the harmonic oscillator and free particle we need to represent the
canonical commutation relations
[q̂ i , p̂j ] =ı̊~{q i , pj } =ı̊~δji (1.23)
on phase space. We introduce a basis of position eigenstates |~qi. The position and
momentum operators q̂ i and p̂i then act as9
∂
q̂ i |~qi = q i |~qi, p̂i |~qi =ı̊~ |~qi. (1.24)
∂q i
We introduce a wave function ψ(t, ~q) = h~q|ψ(t)i to represent a general state |ψ(t)i
on phase space Z
|ψ(t)i = d~q d ψ(t, ~q)|~qi. (1.25)

8
Poisson brackets cannot always be translated literally to commutators; the idea of
quantisation is to represent them up to “simpler” terms, i.e. up to polynomials of lower degree in
the operators and of higher orders in ~.
9
As usual in quantum mechanics, the action of the operators on the states effectively inverts
the order of terms in operator products. Plain insertion of q̂ i = q i and p̂i =ı̊ ∂/∂q i into the
commutation relations leads to the wrong sign.

1.4
By construction, the position operator q̂ i acts by multiplying the wave function by
q i . The momentum operator p̂i effectively acts by the derivative −ı̊∂/∂q i on the
wave function.10 For acting directly on the wave function ψ(t, ~q) we can thus write
∂
q̂ i ψ(t, ~q) ' q i ψ(t, ~q), p̂i ψ(t, ~q) ' −ı̊~ ψ(t, ~q). (1.26)
∂q i
The Hamiltonian acting on the wave function reads
 2
~2 ∂ mω 2 2
H'− + ~q . (1.27)
2m ∂~q 2

The free particle is solved exactly by momentum eigenstates (Fourier

transformation)

d~p d −ı̊~−1 p~·~q

Z Z
d ı̊~−1 p
~·~
q
|~pi = d~q e |~qi, |~qi = e |~pi. (1.28)
(2π~)d

The momentum eigenstate |~pi is an energy eigenstate with E = ~p 2 /2m.

For the harmonic oscillator we use the operators ai and a†i which act on a wave
function as
 
1 1 ∂
~a = √ (mω~q +ı̊~p) ' √ mω~q + ~ ,
2mω 2mω ∂~q
 
† 1 1 ∂
~a = √ (mω~q −ı̊~p) ' √ mω~q − ~ . (1.29)
2mω 2mω ∂~q
They obey the commutation relations

[ai , a†j ] = ~δji . (1.30)

The quantum Hamiltonian has an apparent extra vacuum energy E0 = 12 d~ω

compared to its classical counterpart H = ω~a† ·~a

H = 21 ωai a†i + 21 ωa†i ai = ω~a† ·~a + 21 d~ω = ω~a† ·~a + E0 . (1.31)

• One can add any numerical energy E0 to the Hamiltonian. This has no effect
on any commutation relations, it merely shifts the frequencies by a common
amount. E0 is largely irrelevant for physics.11
• This has the same effect as adding the term ı̊α(~q·~p − p~·~q) to the Hamiltonian
H.12 Such a term is classically invisible, but it leads to a quantum energy shift
∆E0 = −dα~. There is a quantum ordering ambiguity. Here it is very
harmless, it merely affects the trivial energy E0 .
10
The opposite sign arises from a partial integration. In this presentation direct substitution
into commutators yields the desired result [q̂ i , p̂j ] 'ı̊~δji .
11
This statement holds unless the energy couples to something else, e.g. in gravity theories. A
similar case is the relativistic rest energy E0 = mc2 which is irrelevant unless particles are created
or annihilated.
12
There appears to be no ordering ambiguity for the terms ~p 2 and ~q 2 , but it is instructive to
consider all ordered degree-two polynomials in p~ and ~q which include the terms ~q·~ p and p~·~q.

1.5
 • Quantum theory does as it pleases, e.g. it can introduce a non-trivial E0 or
shift its value. It is generally preferable to consider all allowable terms in the
first place, i.e. also introduce an energy shift E0 into the classical Hamiltonian
(which has almost no consequences).
• The conventional vacuum energy 12 ~ω of the harmonic oscillator is by no means
uniquely determined by quantum mechanics. Further input such as symmetry
is needed to arrive at this result.
• The reason for being very picky about this minor issue is that some of the
infinities of quantum field theory will be absorbed into a redefinition of this
constant.
We can now construct the spectrum of the harmonic oscillator: Start from the
vacuum state |0i which is defined to be annihilated by the lowering operator ~a
ai |0i = 0. (1.32)
This state has energy E = E0 (although this information is irrelevant as argued
above). Then add ni ≥ 0 excitations of the raising operator a†i and normalise the
state to h~n|~ni = 1 !
d
Y (a†i )ni
|~ni = √ |0i. (1.33)
i=1
ni ! ~ni /2
This is an energy eigenstate with E = E0 + ~ωN where N = di=1 ni is the total
P
excitation number. The crucial property to arrive at this result is
[H, ~a†i ] = ~ω~a†i , (1.34)
which follows directly from the commutator algebra.

1.4 Quantum Mechanics and Relativity

From now on, let us set ~ = 1, c = 1 for convenience.13
Attempts to set up a relativistic version of quantum mechanics have failed. Let us
see why. In investigating the reasons we shall encounter the key characteristics of a
quantum (field) theory such as unitarity, locality and causality.
First, compare the typical non-relativistic and relativistic energy relations for a
free particle
~p 2
q
e= , vs. e = ~p + m or e = ~p 2 + m2 .
2 2 2
(1.35)
2m
A natural guess for a relativistic wave equation is (Klein–Gordon)
 2  2 !
∂ ∂ 2
− + − m ψ(t, ~q) = 0. (1.36)
∂t ∂~q
This equation has several conceptual problems:
13
The constants can always be recovered from considerations of physical units.

1.6
Probabilistic Properties. There is no notion of the norm of the state |ψi
which satisfies all the required properties of non-relativistic quantum mechanics:
• The norm hψ|ψi of non-relativistic quantum mechanics is conserved only for a
first-order wave equation.
• There is a real conserved quantity
ı̊  
Q= hψ|ψ̇i − hψ̇|ψi . (1.37)
2m
This charge has a major problem: it is not positive definite. Consequently, it is
not suitable for a probabilistic interpretation!14
• Alternatively, one can define a positive definite measure. Unfortunately, this
measure is not local.
Altogether, one might ask why to consider probabilities defined in a time slice: A
time slice is not a Lorentz invariant concept in a relativistic model.

Causality. Suppose, the issues of the probabilistic interpretation are not severe
by themselves, and let us discuss causality instead. Consider the overlap

h~q2 |U (t2 , t1 )|~q1 i (1.38)

for a pair of spacetime points (t1 , q1 ) and (t2 , q2 ). It describes the probability
amplitude for a particle moving from 1 to 2.
For space-like separated points this leads to a conceptual problem in a relativistic
model:
• the overlap is non-zero even if points are space-like separated;
• this is a forbidden region, non-zero probabilities appear to violate causality;
• at least there is exponential suppression in this region; this may be interpreted
as a quantum tunnelling effect.

Negative-Energy Solutions. The relativistic wave equation is a second-order

partial differential equation. For every positive-energy solution
Z
|~p, +, ti = d~q d eı̊~p·~q−ı̊e(~p)t |~qi (1.39)

there is a negative-energy solution

Z
|~p, −, ti = d~q d eı̊~p·~q+ı̊e(~p)t |~qi. (1.40)

This leads to the following problems:

• negative-energy particles have not been observed (they are not tachyons,
though);
• one might extract energy from making this particle faster;
14
As we shall see, Q is rather similar to an electric charge.

1.7
 • a positive-energy particle could fall to a negative-energy state; such a process
would release a lot of energy to produce further particles.
One could insist on positive energies by postulating the relativistic Schrödinger
wave equation s  
2
∂ ∂
ı̊ ψ(t, ~q) = − + m2 ψ(t, ~q). (1.41)
∂t ∂~q
Likewise this leads to several problems:
• The square root of a (differential) operator is hard to define;
• any suitable definition would certainly imply a non-local wave-equation.

Particle Creation. Special relativity allows energy to be converted to rest mass

of particles:
• relativistic quantum mechanics should allow such processes;
• ordinary quantum mechanics usually assumes a fixed particle number.

Dirac Equation. The Dirac equation was an attempt to overcome some of the
above problems
∂ ∂
ψ = αi i ψ + mβ ψ. (1.42)
∂t ∂q
For a suitable choice of matrices αi , β it is a relativistic wave equation which
implies the Klein–Gordon equation. Let us consider the implications on the above
problems.
Probabilistic interpretation:
• it is a first-order wave equation;
• the combination hψ|ψi is conserved and positive definite;
• positivity requires Bose statistics for the wave function.
Spin:
• ordinarily, operators αk imply spin-1/2 particles;
• half-integer spin requires Fermi statistics.
Negative-energy solutions:
• they exist (with different spin d.o.f.);
• they are separated from positive-energy solution in a non-local fashion.
The Dirac equation essentially has the same problems as the Klein–Gordon
equation.

Conclusion. Klein–Gordon and Dirac equations:

• they are perfectly acceptable relativistic wave equations;
• they do not offer a probabilistic interpretation;
• they represent a model without particle production.

1.8
1.5 Conventions
Units. We shall work with natural units ~ = c = 1.
• c = 299 792 458 m s−1 therefore s := 299 792 458 m.
• ~ = 1.055 . . . ·10−34 kg m2 s−1 therefore kg := 2.843·1042 m−1 .
• one can always reinstall appropriate units by inserting factors of 1 = c or 1 = ~
until the units come out as desired;
• in particle physics one often uses electron Volt (eV) instead of length:
m = 5.068·106 eV−1 , s = 1.519·1015 eV−1 , kg = 5.610·1035 eV;
• one can convert back to SI units:
eV = 5.068·106 m−1 = 1.519·1015 s−1 = 1.783·10−36 kg.

Euclidean Space. We write a three-vector x alternatively as

• xj = xj with Latin indices k, l, . . . = 1, 2, 3;
• ~x = (x1 , x2 , x3 ) = (x, y, z).
Scalar product between two vectors:
3
X
~a·~b := ak b k = a1 b 1 + a2 b 2 + a3 b 3 . (1.43)
k=1

Vector square:
~a 2 := ~a·~a = a21 + a22 + a23 . (1.44)
Totally anti-symmetric tensor εijk with normalisation:

ε123 = +1. (1.45)

Cross product (in terms of ε):

(a × b)k = εijk ai bj . (1.46)

Minkowski Space. Four-vectors, Greek indices µ, ν, . . . = 0, 1, 2, 3:

• position vector xµ := (x0 , x1 , x2 , x3 ) = (+t, ~x).
• momentum covector pµ := (p0 , p1 , p2 , p3 ) = (−e, p~).
Summation convention: repeated index µ means implicit sum over µ = 0, 1, 2, 3
3
X
µ
x pµ := xµ pµ = −et + ~x·~p. (1.47)
µ=0

Minkowski metric: signature (−+++)

ηµν = η µν = diag(−1, +1, +1, +1). (1.48)

Raise and lower indices (wherever needed):

xµ := ηµν xµ = (−t, ~x), pµ := η µν pµ = (+e, p~). (1.49)

1.9
Scalar products of two vectors or two covectors, e.g.

p·p := −e2 + ~p 2 . (1.50)

Totally anti-symmetric tensor εµνρσ with normalisation:

ε0123 = +1, ε0123 = −1. (1.51)

Our conventions:
• Mass shell p2 = −m2 : p2 < 0 massive, p2 = 0 massless, p2 > 0 tachyonic.
• Light cone: (x − y)2 < 0 time-like, (x − y)2 = 0 light-like, (x − y)2 > 0
space-like.
p2 = −m2 < 0
massive (x − y)2 < 0
2
p =0 time-like, future
massless (x − y)2 = 0
light-like

p2 > 0 (x − y)2 > 0

(1.52)
tachyonic space-like

(x − y)2 < 0
2 2
p = −m < 0 time-like, past
massive

Why (−+++) signature?

• notation follows space (not time): xµ = (t, ~x), pµ = (e,~p);
• xi = xi but x0 = −x0 = t;
• pi = pi but p0 = −p0 = e;
• Wick rotations more natural: just rotate time t →ı̊t and obtain Euclidean
metric.
How to convert from (+−−−) to (−+++) signature?
• flip sign of every η µν and ηµν ;
• find out which (co)vectors match: xµ , pµ and ∂µ agree literally, xµ , pµ and ∂ µ
flip the sign;
• flip sign for every scalar product of vectors of same type: e.g.
p2 + m2 ↔ −p2 + m2 ;
• preserve the scalar product between vectors of opposite type, e.g. xµ ∂µ .

Clifford Algebra. We use the standard form of the Clifford algebra

{γ µ , γ ν } = 2η µν . (1.53)

In comparison to gamma-matrices γ̃ in (+−−−) signature the gamma-matrices in

(−+++) signature carry a factor of ±ı̊ to compensate for the overall factor of −1

1.10
in the metric η: γ̃ = −ı̊γ. For the Weyl representation it makes sense to apply a
further change of basis
 
µ −1 µ 1 0
γ̃ = −ı̊U γ U, U= . (1.54)
0 ı̊

Then the various conjugation matrices are related by

γ̃P = U −1 γP U,
γ̃C = U −1 γC U ∗ ,
γ̃ 5 = U −1 γ 5 U,
γ̃† = U −1 γ† U −† ,
γ̃T̄ = U −1 γT̄ U ∗ . (1.55)

Finally, our definition of the extended Pauli matrices σ, σ̄ in signature (−+++)

˜ in signature (+−−−)
relates to the conventional one σ̃, σ̄

σ̃ = σ, ˜ = −σ̄.
σ̄ (1.56)

This not only yields the conventional expression for the Weyl representation of
gamma-matrices, but also has the benefit that σ k = σk = σ̄ k = σ̄k irrespectively of
the index position and the conjugation (whereas σ 0 = −σ0 = −σ̄ 0 = σ̄0 = 1).

Name Spaces. We have only 26 Latin letters at our disposal and some are more
attractive than others. We have to recycle:
• e may be 2.71 . . ., but also energy e (no distinction made on blackboard);
• π may be √3.14 . . ., but also momentum conjugate to field φ;
• ı̊ may be −1 , but also useful for counting i (no distinction made on
blackboard);
• κ may look like k or K on the blackboard;
• H may be Hamilton function or operator;
• ...
We will typically not say explicitly which letter means what:
• we may even use same letter for different meanings in one formula;
one can guess meaning from the context, e.g. i in exp(πı̊ . . .) vs. ni=1 ;
P
•
• indices typically do not mix with other symbols;
• one may try to avoid clashes, but this may end up cluttering notation;
• it’s a fact of life (and the literature) and one needs to get used to it.

1.11
Quantum Field Theory I Chapter 2
ETH Zurich, 2019 HS Prof. N. Beisert
21. 11. 2019

2 Classical Free Scalar Field

In the following chapter we will discuss the transition from classical mechanics to
the mechanics of fields, and introduce one of the simplest field theory models, the
classical non-interacting relativistic scalar field.1

2.1 Spring Lattice

Before considering a field, we start with an approximation we can certainly handle:
a lattice.
Consider an atomic lattice:
• 1D or 2D cubic lattice,
• atoms are coupled to neighbours by springs,2
• atoms are coupled to rest position by springs,
• atoms can move only orthogonally to lattice (transverse),
• boundaries: periodic identification.

(2.1)
qj,k µ
κ r λ
The model has the following parameters and variables:
• lattice separation r,
• number of atoms N (in each direction),
• mass of each atom µ,
• lattice spring constant κ,
• return spring constant λ,
• shift orthogonal to lattice qj,k

Lagrangian Formulation. The Lagrange function reads

L = Lkin − Vlat − Vrest . (2.2)

1
This chapter will be somewhat repetitive and present similar relationships from different
points of view. These are the basic relationships of relativistic quantum field theory, and it helps
to be able to recognise and interpret them.
2
Springs are useful approximations because they model first deviation from rest position;
always applies to small excitations.

2.1
We make the standard non-relativistic ansatz for the kinetic terms, where µ is the
mass of an atom
XN
1 2
Lkin = 2 µ q̇i,j . (2.3)
i,j=1

The potential for the springs with spring constants κ between the atoms reads3
N
X N
X
Vlat = 12 κ (qi−1,j − qi,j )2 + 21 κ (qi,j−1 − qi,j )2 . (2.4)
i,j=1 i,j=1

Finally, there is some spring potential with constant λ to drive the atoms back to
their rest position
N
X
2
Vrest = 21 λ qi,j . (2.5)
i,j=1

Altogether the interactions are quadratic in q’s. In other words, we have a bunch
of coupled harmonic oscillators. The equations of motion read

− κ(qi−1,j − 2qi,j + qi+1,j )

µq̈i,j + λqi,j = 0. (2.6)
− κ(qi,j−1 − 2qi,j + qi,j+1 )

Note that these are spatially homogeneous equations. We can use a discrete
Fourier transform (respecting periodicity) to solve them
N  
1 X γk,l 2πı̊
qi,j (t) = 2 p exp (ki + lj) −ı̊ωk,l t
N k,l=1 2µωk,l N
N ∗
γk,l
 
1 X 2πı̊
+ 2 p exp − (ki + lj) +ı̊ωk,l t . (2.7)
N k,l=1 2µωk,l N

We have used the freedomp to (re)define the Fourier coefficients γk,l in order to
introduce prefactors 1/ 2µωk,l N 2 which will later lead to a canonical
∗
normalisation of all harmonic oscillators.4 5 The complex conjugate coefficients γk,l
3
In fact, we can ignore the contribution of the spring extension along the x and y directions of
the lattice. This is because a three-dimensional distance is given by d2 = d2x + d2y + d2z and
therefore the extensions in the three spatial directions decouple in the spring potential.
Moreover, dx and dy describe the distance of two lattice sites. In our model we assume transverse
excitations only, hence dx and dy are constant leading to an irrelevant constant shift of the
potential. Even if dx and dy are considered dynamical, the results for the transverse excitations
would not change because longitudinal and transverse excitations decouple exactly.
4
We would lose nothing by using different prefactors, only the resulting expressions will have a
(slightly) more complicated form.
5
Admittedly, this is violating causality. A priori, there is little motivation to put precisely
these factors here right now. In practice, one would start with trivial factors, go through the
calculation and then redefine the coefficients. Out of lazyness, nostalgia or a different
interpretation of being pedagogical, one might stick to the old coefficients for part of the
calculation, state the translation to new coefficients, and use the latter for the rest. Tampering
with causality may, however, improve the readability even though it leaves the reader somewhat
puzzled about the starting point. Thank you for your attention and back to the subject.

2.2
 ∗
in the second term ensure reality of qi,j . Note that γk,l represents a Fourier mode
of opposite momentum and energy as compared to γk,l !
The equation of motion translates to a dispersion relation:
2 πk πl
µωk,l = λ + 4κ sin2+ 4κ sin2 . (2.8)
N N
Along one axis it takes the following shape:
ωk

(2.9)

k
− 21 N 0 + 12 N

Hamiltonian Formulation. First, define canonical momentum variables

∂L
pi,j := = µq̇i,j . (2.10)
∂ q̇i,j
Then derive the Hamiltonian function as the Legendre transform of L
N
1 X 2
H = Hkin + Vlat + Vrest with Hkin = p . (2.11)
2µ i,j=1 i,j

Finally, define canonical Poisson brackets

N  
X ∂f ∂g ∂f ∂g
{f, g} := − . (2.12)
i,j=1
∂qi,j ∂pi,j ∂pi,j ∂qi,j

In other words {qi,j , pk,l } = δi,k δj,l and {q, q} = {p, p} = 0.

Fourier Modes. Next we introduce new complex variables ck,l as the discrete
Fourier transformation of the position and momentum variables qi,j and pi,j
N  
1 X 2πı̊

ck,l = p exp − (ki + lj) µωk,l qi,j +ı̊pi,j . (2.13)
2µωk,l i,j=1 N

The transformed Hamiltonian becomes very simple

N
1 X
H= 2 ωk,l c∗k,l ck,l . (2.14)
N k,l=1

The Poisson brackets for new variables are simple, too

{ci,j , c∗k,l } = −ı̊ N 2 δi,k δj,l ,
{ci,j , ck,l } = {c∗i,j , c∗k,l } = 0. (2.15)

2.3
One can convince oneself that the canonical equations of motion apply

ċk,l = −{H, ck,l } = −ı̊ωk,l ck,l ,

ċ∗k,l = −{H, c∗k,l } = +ı̊ωk,l c∗k,l . (2.16)

They are solved by the above solution in the Lagrangian framework

ck,l (t) = γk,l exp(−ı̊ωk,l t), ∗

c∗k,l (t) = γk,l exp(+ı̊ωk,l t). (2.17)

2.2 Continuum Limit

Now turn this spring lattice into a smooth field ϕ:
• send the number of sites N → ∞;
• consider a box of size L in all directions; lattice separation r = L/N → 0;
• positions x = ir = iL/N ;
• field qi,... = ϕ(~x);
• generalise to d spatial dimensions, e.g. d = 1, 2, 3.
Two useful rules in applying this limit are as follows:
N Z
X 1
→ dx , qi − qi−1 → r(∂ϕ) (2.18)
i=1
r

Lagrangian Formulation. Substitute the above rules in the Lagrangian

function Z  
d µ 2 κ ~ 2 λ 2
L → d~x ϕ̇ − d−2 (∂ϕ) − dϕ . (2.19)
2rd 2r 2r
This expression diverges as r → 0, but we can rescale the parameters to
compensate the divergences. Suitable rescalings are given by6 7

µ = rd µ̄, κ = rd−2 κ̄, λ = rd λ̄. (2.20)

The parameters µ, κ, λ become densities µ̄, κ̄, λ̄, and the Lagrangian functional
takes the form
Z  
d 1 2 1 ~ 2
L[ϕ, ϕ̇](t) = d~x 2 µ̄ϕ̇ − 2 κ̄(∂ϕ) − 2 λ̄ϕ .1 2
(2.21)

6
A crucial step in formulating a continuum limit is to choose the limiting behaviour of
constants: Typically there are two dominant domains. If the scaling is too weak (e.g. λ ∼ ra with
a < d), the resulting limit is inconsistent (e.g. infinities remain). If the scaling is too strong (e.g.
λ ∼ ra with a > d), some terms are discarded from the resulting model (e.g. m = 0). However,
the most interesting physical models (e.g. with 0 < m < ∞) can be found at the interface of the
two domains where the scaling is fine-tuned (e.g. λ ∼ ra with a = d). Note that the location of
this interface in parameter space is defined by the discrete model itself, but often some freely
tunable parameters remain on the interface.
7
Such rescalings are closely linked to what we will later understand as renormalisation. In
fact, the lattice serves as a UV regulator for field theory, and this is a very basic example of
renormalisation.

2.4
We can furthermore rescale the field ϕ = κ̄−1/2 φ to remove the coefficient in front
of the momentum term and make the field canonically normalised
Z  
~ 2 − 1 λ̄κ̄−1 φ2 .
L[φ, φ̇](t) = d~x d 12 µ̄κ̄−1 φ̇2 − 12 (∂φ) (2.22)
2

From here, we can derive the equations of motion for the field: We start with the
action functional S[φ]
Z Z
S[φ] = dt L[φ](t) = dt d~x d L φ(~x, t), ∂i φ(~x, t), φ̇(~x, t) ;


(2.23)

it is useful to express the (homogeneous) Lagrangian functional L[φ](t) through

the Lagrangian density L(φ, ∂φ,~ φ̇) (the Lagrangian).
We vary the action functional (and discard boundary terms)
Z  
d ∂L ∂ ∂L d ∂L !
δS[φ] = dt d~x δφ − i − + . . . = 0. (2.24)
∂φ ∂x ∂(∂i φ) dt ∂ φ̇
We thus write the general Euler–Lagrange equation for fields
∂L ∂ ∂L d ∂L
(~x, t) − i (~x, t) − (~x, t) = 0. (2.25)
∂φ ∂x ∂(∂i φ) dt ∂ φ̇
In our case the result agrees with the continuum limit of the discrete equations of
motion 2
−µ̄κ̄−1 φ̈ + ~∂ φ − λ̄κ̄−1 φ = 0. (2.26)
Now denote the speed of sound by c and the mass by m
µ̄κ̄−1 = c−2 = 1, λ̄κ̄−1 = m2 ; (2.27)
we discover the Klein–Gordon equation (set c = 1)
2
−c−2 φ̈ + ~∂ φ − m2 φ = 0. (2.28)

Plane Wave Solutions. Consider solutions on infinitely extended space and

time. The homogeneous equation is solved by a Fourier transformation
d~p d
Z


φ(~x, t) = d
α(~p) exp ı̊~p·~x −ı̊e(~p)t
(2π) 2e(~p)
d~p d
Z
∗


+ α (~
p) exp −ı̊~
p ·~
x +ı̊e(~
p)t (2.29)
(2π)d 2e(~p)
with (positive) energy e(~p) on the mass shell (the energy e(~p) is often denoted as
the angular velocity ω(~p)) q
e(~p) = ~p 2 + m2 . (2.30)
This agrees with the discrete solution when identifying momenta as
N Z
X L α(~p)
p = 2πk/L, → dp , γk,... = p . (2.31)
k=1
2π 2e(~p)rd
Some remarks on factors and conventions:

2.5
 • Fourier transforms on R produce factors of 2π; they must be put somewhere.
The dominant convention in physics is to associate (2π)−1 to every dp:
d̄p := dp /2π. There are no factors of 2π associated to dx. There are no factors
of 2π in the exponent.
• The combination d~p d /2e(~p) is useful for relativistic covariance. This is also the
reason for the conversion factor for γk,... vs. α(~p).

2.3 Relativistic Covariance

The Klein–Gordon equation is invariant under Lorentz and Poincaré symmetry. If
we interpret c as the speed of light rather than the speed of sound, we obtain a
relativistic theory.8
The Klein–Gordon equation can be written manifestly relativistically9 10

−∂ µ ∂µ φ + m2 φ = −∂ 2 φ + m2 φ = 0. (2.32)
Also the Lagrangian density and the action have a manifestly relativistic form with
D =d+1 Z
2 1 2 2
1
L = − 2 (∂φ) − 2 m φ , S = dxD L. (2.33)
To understand the relativistic behaviour of the general solution derived above, we
take a look at the momentum space representation of the field
dpD
Z
φ(x) = exp(ı̊p·x) φ(p) (2.34)
(2π)D
and its corresponding equation of motion
(p2 + m2 )φ(p) = 0. (2.35)
The solution of this equation of motion is thus a delta-function supported on the
solution of the algebraic equation p2 + m2 = 0
φ(p) = 2πδ(p2 + m2 ) θ(p0 )α(~p) + θ(−p0 )α∗ (−~p) .


(2.36)
The fields α(~p) and α∗ (~p) define amplitudes on the forward/backward mass shells
α(~
p)

e = ±e(~p),
p (2.37)
e(~p) := ~p 2 + m2 .

α∗ (−~
p)
8
There appears to be no deeper reason why the limiting model should be relativistic.
However, as long as the lattice spacing is finite, there is a periodic band structure with Brillouin
zones which is clearly violates Poincaré symmetry.
9 2
∂ is often written as the D’Alembertian operator .
10
Our signature of spacetime is −+++!

2.6
Note that the reality condition in position space translates to a non-local reality
condition in momentum space

φ(x)∗ = φ(x) ⇐⇒ φ(p)∗ = φ(−p). (2.38)

To investigate the relativistic behaviour of the solution, we perform the integration

of a function f (p) over the mass shell p2 + m2 = 0 with positive energy

dpD
Z
2πδ(p2 + m2 )θ(p0 ) f (p)
(2π)D
d~p d de
Z
2 2 2


= 2πδ −e + ~p + m θ(e) f (e, p~)
(2π)d 2π
d~p d de
Z
2πδ −e2 + e(~p)2 θ(e) f (e, p~)


= d
(2π) 2π
d~p d de
Z


= 2πδ (e(~
p ) + e)(e(~p) − e) θ(e) f (e, p~)
(2π)d 2π
d~p d
Z
de

= d
δ e − e(~p) f (e, p~)
(2π) e(~p) + e
d~p d
Z


= f e(~
p), p
~ . (2.39)
(2π)d 2e(~p)

Since we started from a manifestly covariant integral, the above measure on the
positive mass shell is Poincaré covariant. This is precisely the measure used in the
formulation of the general solution.

2.4 Hamiltonian Field Theory

Now that we have a nice relativistic formulation for the Klein–Gordon field φ(x),
let us separate space from time.11

Position Space. Define the momentum variables π(x) (field) conjugate to the
field φ(x)
δL ∂L
π(~x, t) = (t) = (~x, t) = φ̇(~x, t). (2.40)
δ φ̇(~x) ∂ φ̇
Next, determine the Hamiltonian function
Z
H[φ, π] = d~x d π φ̇ − L[φ, φ̇]
Z  
~ 2 + 1 m2 φ2 .
= d~x d 12 π 2 + 12 (∂φ) (2.41)
2

11
The Hamiltonian formalism breaks manifest relativistic invariance, but they physics remains
(secretly) relativistic.

2.7
This expression is not relativistically covariant, but it was not designed to be.12
Finally, define Poisson brackets for two phase space functionals f, g with
f = f [φ, π] Z  
d δf δg δf δg
{f, g} = d~z − . (2.42)
δφ(~z) δπ(~z) δπ(~z) δφ(~z)
The Poisson brackets of the fundamental fields yield delta-functions13
Z
φ(~x), π(~y ) = d~z d δ d (~x − ~z)δ d (~y − ~z) = δ d (~x − ~z).

(2.43)

Momentum Space. Now introduce momentum modes by a suitable Fourier

transformation14
Z
a(~p) = d~x d exp(−ı̊~p·~x) e(~p)φ(~x) +ı̊π(~x) ,

Z
∗
a (~p) = d~x d exp(+ı̊~p·~x) e(~p)φ(~x) −ı̊π(~x) ,


(2.44)

and the inverse Fourier transformation

d~p d
Z
φ(~x) = a(~p) exp(+ı̊~p·~x)
(2π)d 2e(~p)
d~p d
Z
+ d
a∗ (~p) exp(−ı̊~p·~x),
(2π) 2e(~p)
d~p d
Z
ı̊
π(~x) = − a(~p) exp(+ı̊~p·~x)
2 (2π)d
d~p d ∗
Z
ı̊
+ a (~p) exp(−ı̊~p·~x). (2.45)
2 (2π)d
We compute the Poisson brackets for the Fourier modes15 16
a(~p), a∗ (~q) = −ı̊ 2e(~p) (2π)d δ d (~p − ~q).

(2.46)
In other words,
Z  
d d δf δg δf δg
{f, g} = −ı̊(2π) d~p 2e(~p) − . (2.47)
δa(~p) δa∗ (~p) δa∗ (~p) δa(~p)
The Hamiltonian translates to17
d~p d ∗ d~p d
Z Z
1
H= d
a (~p)a(~p) = d
e(~p) a∗ (~p)a(~p). (2.48)
2 (2π) (2π) 2e(~p)
12
The Hamiltonian as a phase space function governs translation in time. The notion of a
specific time is not relativistically covariant.
13
Use the formula for variation δφ(~z)/δφ(~x) = δ d (~z − ~x).
14
We have additional factors compared to some literature.
15
There is a conventional factor of 2π for the delta-function in momentum space
complementary to the factor of 1/2π for the momentum space measure.
16
The factor 2e(~p) is the appropriate relativistic measure for the mass shell complementary to
the factor of 1/2e for the relativistic measure on the mass shell.
17
The second form highlights the relativistic covariance of the mass shell measure as well as the
fact that the Hamiltonian measures the integral of energy e(~ p).

2.8
Hence the equations of motion for momentum space fields take the simple form of
a harmonic oscillator

ȧ(~p) = − H, a(~p) = −ı̊e(~p)a(~p),
ȧ∗ (~p) = − H, a∗ (~p) = +ı̊e(~p)a∗ (~p).

(2.49)

There is one harmonic oscillator for each momentum p~. The solution reads


a(~p, t) = α(~p) exp −ı̊e(~p)t ,
a∗ (~p, t) = α∗ (~p) exp +ı̊e(~p)t .


(2.50)

2.9
Quantum Field Theory I Chapter 3
ETH Zurich, 2019 HS Prof. N. Beisert
21. 11. 2019

3 Scalar Field Quantisation

We can now go ahead and try to quantise the classical scalar field using the
canonical procedure described before. We will encounter some infinities, and
discuss how to deal with them. Then we shall investigate a few basic objects in
quantum field theory.

3.1 Quantisation
We start with the Hamiltonian formulation of the scalar field discussed earlier.

Equal-Time Commutators. Phase space consists of the field φ(~x) and the
conjugate momentum π(~x) with Poisson bracket1
φ(~x), π(~y ) = δ d (~x − ~y ).

(3.1)

Hence the canonical quantisation implies operators φ̂(~x) and π̂(~x)

φ̂(~x), π̂(~y ) =ı̊δ d (~x − ~y ).
 
(3.2)
Note that φ and π are now operator-valued fields (rather: distributions).2

Field States. Next we have to define some states. Straight application of

quantum mechanics would lead to a state |φi for every field configuration φ(~x)
such that
δ
φ̂(~x)|φi = φ(~x)|φi, π̂(~x)|φi =ı̊ |φi. (3.3)
δφ(~x)
This can be done formally, but it is not very convenient. For example, the ground
state wave functional takes the following form with a suitable function Ω
Z  Z 
1 d d
|0i = Dφ exp − d~x d~y Ω(~x, ~y ) φ(~x) φ(~y ) |φi. (3.4)
2
This is to be compared to the wave function of a harmonic oscillator
Z !
1 X
|0i = d~x d exp − ωij xi xj |~xi. (3.5)
2 i,j
1
At the moment, the fields are defined on a common time slice t, e.g. t = 0. Later we discuss
unequal times.
2
The delta-function is a distribution, also the fields should be considered distributions.
Distributions are linear maps from test functions to numbers (or operators in this case). In
physics, we write them as integrals with a distributional kernel (e.g. the delta-function).
Sometimes we also perform illegal operations (e.g. evaluate delta-function δ(x) at x = 0).

3.1
Momentum Space. The classical field is a bunch of coupled harmonic
oscillators, let us diagonalise them and use creation and annihilation operators.
We go to momentum space, and pick a and a† appropriately, using the same
transformation as above3
Z
a(~p) = d~x d exp(−ı̊~p·~x) e(~p)φ(~x) +ı̊π(~x) ,

d~p d
Z
φ(~x) = a(~p) exp(+ı̊~p·~x)
(2π)d 2e(~p)
d~p d
Z
+ a† (~p) exp(−ı̊~p·~x),
(2π)d 2e(~p)
d~p d
Z
ı̊
π(~x) = − a(~p) exp(+ı̊~p·~x)
2 (2π)d
d~p d †
Z
ı̊
+ a (~p) exp(−ı̊~p·~x). (3.6)
2 (2π)d

The canonical commutation relations in momentum space read

a(~p), a† (~q) = 2e(~p) (2π)d δ d (~p − ~q).

 
(3.7)

We substitute these fields into the Hamiltonian of the scalar field paying attention
to ordering
Z
H = d~x d 12 π 2 + 21 (∂φ) ~ 2 + 1 m2 φ2


2

d~p d
Z
1
a† (~p)a(~p) + a(~p)a† (~p)


= d
4 (2π)
d~p d
Z
1 †
 †


= 2a (~
p )a(~
p) + a(~
p), a (~
p)
4 (2π)d
d~p d †
Z Z
1 1
= a (~p)a(~p) + d~p d e(~p) δ d (~p − p~). (3.8)
2 (2π)d 2

Vacuum Energy. We introduce a vacuum state |0i annihilated by all a(~p).

There are two problems with the vacuum energy H|0i = E0 |0i where
Z
1
d
E0 = ε0 δ (~p − p~), ε0 = d~p d e(~p). (3.9)
2
Namely:
• evaluation of the delta-function at the origin, δ d (~p − p~) = δ d (0), is ill-defined;
• the integral ε0 = 12 d~p d e(~p) diverges.
R

These are self-made problems:

3
We will henceforth drop the indication of quantum operators for the fields φ(~x), π(~x).

3.2
 • We consider an infinite volume. It is not reasonable to expect a finite overall
energy. This is an IR problem! The delta-function has units of volume. If one
traces its appearance carefully, it turns out to measure the volume of the
system δ d (~p − p~) ∼ V → ∞. We should consider the energy density ε0 instead!
RP
• The integral ε0 = 21 0 d~p d e(~p) represents the vacuum energy density. There
are infinitely many oscillators per volume element, and it is not reasonable to
expect a finite energy density. This is a UV problem!R We can introduce a
P
momentum cutoff |~p| < P to obtain a finite result 21 0 d~p d e(~p) ∼ P d+1 → ∞
which diverges as the cutoff is sent to infinity.
The finite lattice has a similar effect to regularise the IR and the UV. There are
many other ways to avoid the infinities that arise in a QFT.
To avoid IR infinities in QFT: One may put the system in a finite box. This is
possible but inconvenient, e.g. one would have to deal with non-trivial boundary
conditions. One would rather work with an infinite system and cope with the
arising IR divergences in a different manner. E.g. one might consider the energy
density instead of the overall energy.
To avoid UV infinities in QFT: By definition we want a field theory, not a discrete
model. The intuition from physics is that one should be able to approximate a
field by a discrete model (or vice versa). The procedure involves several steps:
1. regularisation; there are several options: impose a momentum cut-off, consider
a lattice, use other tricks without physical motivation, . . .
2. renormalisation: absorb the terms that would otherwise diverge.
3. remove regularisation: obtain finite results.
In our case we simply drop E0 . We can do it because:
• There is no meaning to an absolute vacuum energy. It is only a philosophical or
religious issue.
• We may add any constant to the Hamiltonian to compensate E0
(renormalisation).
• This makes no observable difference in any physical process.
• Here we keep in mind that, later on, infinities may also lead to interesting
effects.
Renormalised Hamiltonian Hren

d~p d †
Z
1
Hren := H − E0 = a (~p)a(~p). (3.10)
2 (2π)d

This result is nice, the vacuum has zero energy. Note that this result or definition
is essential for Poincaré invariance.

Normal Ordering. The ordering of variables in the classical Hamiltonian H

plays no role. In quantum theory it does! It is responsible for the vacuum energy
E0 .
How to map some classical observable O to a quantum operator?

3.3
A possible map is normal ordering N(O) which is defined as follows:
• express O in terms of a and a∗ → a† ;
• write all a† ’s to the left of all a’s.
Here, the renormalised Hamiltonian is the normal ordering of H
Hren = N(H). (3.11)

There are other ordering prescriptions for operators:

• Normal ordering depends on the choice of the vacuum state; there may be
other normal-ordering prescriptions associated to different “vacuum” states.
• There are other useful ordering prescriptions which we will encounter, e.g. time
ordering, symmetric ordering.
• Note that one cannot consistently lift classical to quantum observables while
respecting their Poisson brackets lifted to commutators in all cases.

3.2 Fock Space

We already have:
• a collection of HO’s operators a(~p), a† (~p) labelled by the momentum p~,
• a vacuum state |0i with energy E = 0.
We can now discuss the other related states.

Single-Particle States. We can excite the vacuum

|~pi := a† (~p)|0i. (3.12)
This is an energy eigenstate with energy
E = +e(~p). (3.13)
It is precisely the relativistic energy of a particle with momentum p~ and mass m.

Negative-Energy Solutions. Please note:

• The energy of the above state is positive definite E > 0.
• The state a(~p)|0i would have negative energy. Gladly, it does not exist by
construction.
Our earlier problem of negative-energy particles is solved!
• a† (~p) creates a particle of momentum p~ and positive energy +e(~p) from the
vacuum. a† (~p) is the particle creation operator.
• a(~p) removes a particle of momentum p~ and thus removes the positive energy
+e(~p) from the state.
a(~p)|~qi = a(~p)a† (~q)|0i = [a(~p), a† (~q)]|0i
= 2e(~p) (2π)d δ d (~p − ~q) |0i. (3.14)

3.4
 a(~p) is the particle annihilation operator.4
a† (~
p)

(3.15)

a(−~
p)

The interpretation of the position space operators is not as straight-forward:

φ(~x) = φ† (~x) as well as π(~x) = π † (~x) either create or annihilate a particle at
position ~x. The result of application of this operator to a state is typically a
superposition of states with different particle numbers.

Normalisation. Let us declare a proper normalisation for the vacuum state

h0|0i = 1. (3.16)

The normalisation of a single-particle state follows:

h~p|~pi = 2e(~p) (2π)d δ d (~p − p~) = ∞. (3.17)

This is a known problem from quantum mechanics: plane-wave states have an

infinite extent; they are smeared over all space. This is an unphysical assumption.
Recall that δ d (~p − p~) typically represents some volume of space.
Consider instead a peaked wave packet state |ψi defined by the test function ψ(~p)

d~p d ψ(~p)
Z
|ψi := |~pi. (3.18)
(2π)d 2e(~p)

For a suitable test function ψ(~p) this state has a finite normalisation

d~p d |ψ(~p)|2
Z
hψ|ψi = . (3.19)
(2π)d 2e(~p)

Importantly, the normalisation is positive definite, which overcomes one of the

difficulties of relativistic quantum mechanics.
4
a(~
p) is not the generator of an anti-particle (as sometimes claimed). It has negative energy
while anti-particles (like particles) have positive energy. For the real scalar, the particle is its own
anti-particle.

3.5
Multi-Particle States and Fock Space. Now we excite more than one
harmonic oscillator
|~p1 , p~2 , . . . , p~n i := a† (~p1 ) a† (~p2 ) · · · a† (~pn ) |0i. (3.20)
The energy is given by the sum of particle energies
n
X
E= e(~pk ). (3.21)
k=1

To arrive at this result, use the elementary commutator

H, a† (~p) = e(~p)a† (~p)
 
(3.22)
to show
H|ni = Ha†1 . . . a†n |0i
= [H, a†1 ]a†2 . . . a†n |0i + . . . + a†1 . . . a†n−1 [H, a†n ]|0i
= (e1 + . . . + en )a†1 a†2 . . . a†n |0i
= E|ni. (3.23)

Moreover all particles are freely interchangeable

|. . . , p~, ~q, . . .i = |. . . , ~q, p~, . . .i (3.24)
because creation operators commute
 †
a (~p), a† (~q) = 0.

(3.25)
This corresponds to Bose statistics for indistinguishable particles: In QFT the
wave function is automatically totally symmetric under particle exchange.
A generic QFT state (based on the vacuum |0i) is a linear combination of
k-particle states with k not fixed. This vector space is called Fock space. It is the
direct sum
VFock = V0 ⊕ V1 ⊕ V2 ⊕ V3 ⊕ . . . , Vn = (V1 )⊗s n (3.26)
of n-particle spaces Vn where
• V0 = C merely contains the vacuum state |0i;
• V1 = Vparticle is the space of single particle states |~pi with positive energy;
• Vn is the symmetric tensor product of n copies of V1 .
To understand Fock space better, consider non-relativistic physics:
• In practice, the amount of available energy is bounded from above. It is much
smaller than the particle rest mass m = e(0).
• The relevant part of Fock space has n bounded from above. For example: V1 ,
V2 or V1 ⊕ V2 .
• Multiple-particle quantum mechanics is a low-energy limit of quantum field
theory. It becomes quantum field theory when the number of particles is
unbounded.

3.6
Conservation Laws. The Hamiltonian H measures the total energy E.
There is also a set of operators to measure total momentum P~

d~p d
Z
P~ := p~ a† (~p)a(~p). (3.27)
(2π)d 2e(~p)

with eigenvalue P~ = nk=1 p~k on state |~p1 , . . . , p~n i. The vacuum state carries no
P

momentum. We can combine this into a relativistic vector P µ = (H, P~ ) of

operators

d~p d
Z
µ
P := pµ (~p) a† (~p)a(~p), pµ (~p) := (e(~p), p~). (3.28)
(2π)d 2e(~p)

Another useful operator is the particle number operator

d~p d
Z
N := a† (~p)a(~p). (3.29)
(2π)d 2e(~p)

It measures the number of particles n in a state |~p1 , . . . , p~n i

N Vn = nVn . (3.30)

The relativistic momentum vector and the number operator are conserved

[H, Pµ ] = [H, N ] = 0. (3.31)

Moreover, they carry no momentum

[P µ , P ν ] = [P µ , N ] = 0. (3.32)

In fact, there are many conservation laws. Any operators composed from number
density operators commute

n(~p) := a† (~p)a(~p),
 
n(~p), n(~q) = 0. (3.33)

Hence such operators are conserved, they carry no momentum and no particle
number:      
H, n(~p) = Pµ , n(~p) = N, n(~p) = 0. (3.34)
In a free theory there are infinitely many conservation laws.5

3.3 Complex Scalar Field

Let us discuss a slightly more elaborate case of the scalar field, the complex scalar,
where we first encounter anti-particles.
5
I should not say this: the free theory is trivial.

3.7
The complex scalar field φ(x) has the Lagrangian6 7

L = −∂ µ φ∗ ∂µ φ − m2 φ∗ φ = −|∂φ|2 − m2 |φ|2 . (3.35)

The equation of motion associated to the above Lagrangian is the very same
Klein–Gordon equation. However, now complex solutions φ(x) are allowed. The
general solution is given by

d~p d
Z
φ(x) = b(~p) exp(+ı̊p·x)
(2π)d 2e(~p)
d~p d
Z
+ a∗ (~p) exp(−ı̊p·x),
(2π)d 2e(~p)
d~p d
Z
∗
φ (x) = a(~p) exp(+ı̊p·x)
(2π)d 2e(~p)
d~p d
Z
+ b∗ (~p) exp(−ı̊p·x). (3.36)
(2π)d 2e(~p)

Note the skewed appearance of the coefficients a and b. For a 6= b we have φ 6= φ∗

while a = b implies a real field φ = φ∗ .

b(~
p) a(~
p)

(3.37)

a∗ (−~
p) b∗ (−~
p)

For the conjugate momentum we obtain8

∂L ∂L
π∗ = = φ̇∗ , π= = φ̇. (3.38)
∂ φ̇ ∂ φ̇∗
In the quantum theory we then impose the canonical commutators

φ(~x), π † (~y ) = φ† (~x), π(~y ) =ı̊δ d (~x − ~y ).

   
(3.39)
6
The prefactor 1/2 of the real scalar field (φ2 ) is now absent. It is a convenient symmetry
factor. Here the appropriate symmetry factor is 1 because φ∗ 6= φ. More on symmetry factors
later. √ √
7
One can also set φ(x) = (φ1 +ı̊φ2 )/ 2 and φ∗ (x) = (φ1 −ı̊φ2 )/ 2 and obtain two
independent scalar fields with equal mass.
8
This implies that π is canonically conjugate to φ∗ . One may just as well define
π = ∂L/∂ φ̇ = φ̇∗ in which case π is conjugate to φ. This is a matter of convention, and it makes
no difference if applied consistently.

3.8
Equivalently, we quantise the coefficients a(~p), b(~p) to operators. The non-trivial
commutation relations of these operators read:
a(~p), a† (~q) = b(~p), b† (~q) = 2e(~p) (2π)d δ d (~p − ~q).
   
(3.40)
The quantum Hamiltonian takes the form
d~p d
Z
1 † †


Hren := a (~
p)a(~
p) + b (~
p)b(~
p) . (3.41)
2 (2π)d

The complex scalar carries a charge, +1 for φ and −1 for φ† .

• The operator a† creates a particle with charge +1.
• The operator b has negative energy, it should remove a particle. The operator
carries the same charge as a† , hence the corresponding particle should have
charge −1.
There are two types of particles: the particle and the anti-particle. They have
opposite charges, but equal masses and all positive energies. In conclusion:
• a† creates a particle,
• b annihilates an anti-particle,
• b† creates an anti-particle,
• a annihilates a particle,
• the vacuum is annihilated by a’s and b’s.

3.4 Correlators
We have now quantised the scalar field. The states have an adjustable number of
indistinguishable particles with definite momenta. Now what? Let us consider
particle propagation in space and time.

Schrödinger Picture. For particle propagation we consider the following steps:

• create a particle at xµ = (t, ~x);
• let the state evolve for some time s − t;
• measure the particle at y µ = (s, ~y ).
We could use φ(~x) or π(~x) to create the particle from the vacuum.9 Let us use φ
because π = φ̇ can be obtained from its time derivative.
In the Schrödinger picture, the states evolve in time
d    
ı̊ Ψ (t) = H Ψ (t) . (3.42)
dt
We can solve this equation as10
 
 
Ψ (s) = exp −ı̊H(s − t) Ψ (t) . (3.43)
9
These operators also annihilate particles, but evidently, there are none present in the vacuum.
10
The Hamiltonian H is time-independent, hence time translation is governed simply by
exp(−ı̊H∆t). Later we shall encounter a more difficult situation.

3.9
We substitute the field φ in terms of creation and annihilation operators, and
altogether we find for the correlator


∆+ (y, x) :=ı̊h0|φ(~y ) exp −ı̊(s − t)H φ(~x)|0i
d~p d
Z


=ı̊ exp ı̊~
p ·(~
y − ~
x ) −ı̊(s − t)e(~
p)
(2π)d 2e(~p)
d~p d
Z


=ı̊ exp ı̊p·(y − x) , (3.44)
(2π)d 2e(~p)

where p0 = e(~p) is the relativistic energy of the particle. In this derivation, space
and time take different roles. Nevertheless, the final answer is manifestly Poincaré
covariant.

Heisenberg Picture. Before we continue to investigate this function, let us go

to a different formulation of quantum mechanics, which makes the relativistic
properties more manifest: the Heisenberg picture. We translate the time
dependence of the state to time dependence of operators11



FH (t) := exp +ı̊H(t − t0 ) FS (t) exp −ı̊H(t − t0 ) , (3.45)

where t0 is the reference time slice on which quantum states are defined, commonly
t0 = 0. States are therefore time-independent in the Heisenberg picture. Let us
compare the application of an operator to a state



FH (t)|Ψ (t0 )i = exp ı̊H(t − t0 ) FS (t) exp −ı̊H(t − t0 ) |Ψ (t0 )i


= exp ı̊H(t − t0 ) FS (t)|Ψ (t)i. (3.46)


The difference between the applications is the factor exp ı̊H(t − t0 ) which is
required to translate between the two time slices but which eventually drops out in
correlators.
The field operator φ(x) := φH (t, ~x) in the Heisenberg picture recovers a
dependence on time compared to the field operator φ(~x) := φS (~x) = φH (0, ~x)



φ(x) = exp +ı̊Ht φ(~x) exp −ı̊Ht
d~p d
Z
+ı̊p·x −ı̊p·x †


= e a(~
p) + e a (~
p) (3.47)
(2π)d 2e(~p)

again with p0 = e(~p) implied. Some features:

• it has complete spacetime dependence;
• it is manifestly relativistic;
• there is no need to consider π = φ̇ separately;
• it obeys the Klein–Gordon equation (∂ 2 − m2 )φ = 0;
11
In most relevant cases, the operator FS (t) = FS has no explicit time dependence. The
implicit time dependence from applying the operator FS to a time-dependent state is made
explicit in FH (t).

3.10
 • it has the same form as the solution of the Euler–Lagrange equations.
• in momentum space it takes the form12
a(~
p)

dp4 ı̊p·x
Z
φ(x) = e φ(p),
(2π)4
(3.48)
φ(p) = δ(p2 + m2 )
· θ(p0 )a(~p) + θ(−p0 )a† (−~p) .

a† (−~
p)

In the Heisenberg picture the correlator takes the following form13

∆+ (y, x) =ı̊h0|φ(~y ) exp ı̊H(t − s) φ(~x)|0i
=ı̊h0|φ(y)φ(x)|0i. (3.49)

This yields the same result, but by means of a more immediate and relativistic
derivation.

Correlator. Let us discuss the correlation function

d~p d
Z


∆+ (y, x) =ı̊ exp ı̊p·(y − x) . (3.50)
(2π)d 2e(~p)
We know the function in momentum space from the derivation of the integration
measure over the mass shell
dpd+1
Z


∆+ (y, x) = d+1
∆+ (p) exp ı̊p·(y − x) ,
(2π)
∆+ (p) = 2πı̊δ(p2 + m2 )θ(p0 ). (3.51)

How about position space? Translational symmetry implies that only differences of
positions can matter in agreement with the above momentum space representation

∆+ (y, x) = ∆+ (y − x). (3.52)

Furthermore, Lorentz invariance implies that the result can only depend on
Lorentz invariant quantities. The only Lorentz invariant quantity that can be
12
It is curious to notice that the particle creation
R operator a† is formally the contribution with
−4 ı̊p·x
4
negative energy in the Fourier transform φ(x) = dp (2π) e φ(p). This is because a† in
0
φ(x) creates a particle at time x , whereas a quantum mechanical state |Ψ (t)i describes the
evolution of a particle (which has been created earlier). The latter is determined by the
Schrödinger equation which defines the notion and sign of energy according to |Ψ (t)i ∼ e−ı̊et .
This time evolution is appropriately measured by the particle annihilation operator a which
therefore is attributed a positive energy e in the Fourier transform; correspondingly, the particle
creation operator resides at negative energy.
13
The vacuum state is the same in both pictures; it is not time dependent because the vacuum
energy E0 is (was defined to be) zero. Therefore exp(ı̊Ht)|0i = |0i.

3.11
constructed from the vector xµ is x2 , hence the function can only depend on the
single variable x2
∆+ (x) = ∆+ (x2 ). (3.53)
The mass is the only dimensionful constant at our disposal, and therefore
dimensional analysis implies

∆+ (x2 ) = md−1 F (m2 x2 ), (3.54)

where F is an numerical function. The above considerations hold locally, we

should discuss three distinct regions of spacetime: future, past, elsewhere.14
The Klein–Gordon equation for F becomes

4rF 00 (r) + 2(d + 1)F 0 (r) − F (r) = 0. (3.55)

This is a differential equation for Bessel functions Jα (z).15 There are two solutions:
√
F± (r) = r−(d−1)/4 J±(d−1)/2 (ı̊ r ). (3.56)
√
First, consider the future with x = (t, 0) where t = ±kxk with kxk = −x2 .16
Substitute this into the above momentum space expression

d~p d
Z


∆+ (x) =ı̊ d
exp −ı̊te(~p) .
(2π) 2e(~p)
ı̊ Vol(S d−1 ) ∞ dp pd−1
Z p

= d
p exp −ı̊t p2 + m2 .
2(2π) 0 p2 + m2
ı̊ Vol(S d−1 ) ∞
Z
= de (e2 − m2 )(d−2)/2 exp(−ı̊te)
2(2π)d m
∼ e−ı̊mt for t → ±∞. (3.57)

One can see that the function oscillates with a positive frequency m. The same
conclusion holds for the past. This fixes the relevant linear combination of F± ,
which is in fact the Hankel function Hα 17

m(d−1)/2 p

2 .
∆+ (y − x) ∼ H(d−1)/2 m −(y − x) (3.58)
(y − x)(d−1)/2

For space-like separation, however, one finds the asymptotic behaviour

√
∆+ (x) ∼ e−mr for r = kxk := x2 → ∞. (3.59)
14
These regions are disconnected in the sense that they are not related by orthochronous
Lorentz transformations.
15
Bessel functions are well-known solutions for spherical waves.
16
For time-like separation one can go to such a frame with x = (t, 0).
17
The asymptotic behaviour e−ı̊mt determines which of the two Hankel functions H (1) and
(2)
H applies to the past and the future.

3.12
The result is non-zero, but it decays exponentially with range 1/m. This behaviour
fixes the linear combination of F± which is a modified Bessel function Kα
m(d−1)/2 p

2 .
∆+ (y − x) ∼ K (d−1)/2 m (y − x) (3.60)
(y − x)(d−1)/2
This non-vanishing result is the same as in relativistic quantum mechanics.
Before we discuss causality, let us summarise in this figure:
oscillations ∼ e−ı̊mkxk

exponential
∼ e−mkxk (3.61)
decay

oscillations ∼ e+ı̊mkxk
Note that there are additional delta-function contributions for light-like separation
in ∆+ (y, x) which we will not discuss here.

Unequal-Time Commutator. It is in fact acceptable to violate causality as

long as this effect is never measured.
The correct question to ask is: Can one measurement at spacetime point x
influence the other at spacetime point y? To that end, consider the commutator
 
∆(y − x) :=ı̊ φ(y), φ(x) . (3.62)

We can relate it to the above correlators by inserting the relationship in between

two vacua

∆(y − x) =ı̊h0|[φ(y), φ(x)]|0i = ∆+ (y − x) − ∆+ (x − y). (3.63)

We obtain the following observations for space-like separations:

• ∆+ is a symmetric function;
• the commutator ∆ vanishes;
• φ(x) and φ(y) commute;
• this insight follows also from invariance and the equal-time commutator which
is a delta-function at coincident points;
• causality is preserved!
Note: In the calculation, we observe a cancellation. The particle created at x and
annihilated at y cancels against the particle created at y and annihilated at x; one
of these evidently travels backwards in time.
However, the commutator is non-trivial for time-like separations; one finds the
Bessel functions of the first kind Jα
m(d−1)/2 p

2 .
∆(y − x) ∼ J (d−1)/2 m −(y − x) (3.64)
(y − x)(d−1)/2

3.13
This means that time-like separated measurements can indeed influence each other.
Finally, we can recover the equal-time commutators. For two fields φ the
commutator follows from asymmetry of the integrand
 
φ(~y ), φ(~x) = −ı̊∆(0, ~y − ~x)
d~p d
Z
e+ı̊~p·(~y−~x) − e−ı̊~p·(~y−~x)


= d
(2π) 2e(~p)
= 0. (3.65)
The commutator between a field φ and its conjugate momentum π yields

∂ 
[φ(~y ), π(~x)] = −ı̊ ∆(s − t, ~y − ~x)
∂t s=t
d~p d
Z
ı̊ p·(~
+ı̊~ y −~
x) −ı̊~
p·(~
y −~
x)


= e + e
2 (2π)d
=ı̊δ d (~y − ~x) (3.66)
in agreement with the fundamental commutation relations.

3.5 Sources
We have quantised a free field, we have discussed correlators of two fields, but
there is not much else we can do besides adding interactions (later).
As a first step towards interactions, let us discuss driving the field by an external
source ρ(x)
−∂ 2 φ(x) + m2 φ(x) = ρ(x). (3.67)
We make the assumption that the source field is non-zero only for some finite
interval of time.
Question: Given an initial field φi (x) obeying the homogeneous Klein–Gordon
equation, how to determine the solution φ(x) of the inhomogeneous Klein–Gordon
equation such that φ(x) = φi (x) for all t < t0 before the activation of the source at
t0 ? In particular, what is the final field φf (x) after deactivation of the source at t1 ?
Note that the latter must also obey the homogeneous Klein–Gordon equation.
φf
φ
(3.68)
φi
t0 t1 t
ρ
Due to linearity and translation invariance we make the general ansatz18 19
Z
φ(x) = φi (x) + ∆φ(x), ∆φ(x) = dy d+1 GR (x − y) ρ(y), (3.69)

18
Admittedly, this is a perfectly classical problem already encountered in electrodynamics.
19
Due to linearity each element of the source field ρ will influence the field φ independently.
Furthermore the influence will be invariant under translations.

3.14
where GR is the retarded propagator (Green function)
−∂ 2 GR (x) + m2 GR (x) = δ d+1 (x), GR (x) = 0 for x0 < 0. (3.70)

The first equation is conveniently solved in momentum space

dpd+1 ı̊p·x
Z
G(x) = e G(p) (3.71)
(2π)d+1
with the inhomogeneous Klein–Gordon equation and its solution20
1
p2 G(p) + m2 G(p) = 1, G(p) = . (3.72)
p2 + m2

Let us see how to incorporate the second relation. Write the function as
d~p d ı̊~p·~x −1
Z Z
de −ı̊et
G(x) = e e
(2π)d 2π e2 − e(~p)2
d~p d ı̊~p·~x
 −ı̊et
e−ı̊et

de −1
Z Z
e
= e − . (3.73)
(2π)d 2π 2e(~p) e − e(~p) e + e(~p)
Then solve the Fourier integrals by the residue theorem in the complex plane. The
integral over e runs from −∞ to +∞; close the contour! This is done by a
semi-circle in the complex plane with very large radius. Its contribution must
vanish in order not to alter the overall integral; therefore consider the exponent:
exp(−ı̊et) = exp(−ı̊t Re e) exp(t Im e). (3.74)
Only the second term is able to suppress the contribution:
• For t > 0 we need Im e < 0: close contour in lower half.
• For t < 0 we need Im e > 0: close contour in upper half.
Im e Im e Im e

t<0
Re e (3.75)
Re e Re e
t>0

We have two poles at e = ±e(~p) on the real axis i.e. on the integration contour.
We need to decide how they contribute to residues.
For the retarded propagator, we want GR (x) = 0 for t < 0. This is achieved by
shifting the poles slightly into the lower half plane
Im e
1
GR (p) = 2 . (3.76)
p + m2 −ı̊p0 
−e(~ p) Re e
p) +e(~

20
Due to the poles on the mass shell this solution is in fact ill-defined; in particular, it is left
unspecified how to treat them in the Fourier transformation. This corresponds to the freedom to
add a solution of the homogeneous equation, δ(p2 + m2 )f (p), which is fixed by imposing the
second condition as we shall see.

3.15
Alternatively, we can deform the contour slightly to close above the poles.21
Im e Im e

t<0 −e(~
p) +e(~
p)
Re e
(3.77)
−e(~
p) +e(~
p) Re e
t>0

There are no poles in the upper half plane, hence GR (x) = 0 for t < 0. For t > 0,
however, both poles contribute a residue,22 and we obtain
d~p d
Z
e+ı̊p·x − e−ı̊p·x


GR (x) =ı̊θ(t) d
(2π) 2e(~p)


= θ(t) ∆+ (x) − ∆+ (−x)
= θ(t)∆(x). (3.78)
This is nice: we have found a relationship between correlation functions and
propagators. It yields the position space form for the propagator.
We can confirm that it satisfies the defining relation:23
(−∂ 2 + m2 )GR (x)
∂ ˙
= θ(t) (−∂ 2 + m2 )∆(x) +


δ(t) ∆(x) + δ(t) ∆(x)
∂t
= δ d+1 (x). (3.79)
• the first term vanishes because ∆ satisfies the equations of motion;
• the second term vanishes because [φ(~x), φ(~y )] = 0;
• the third term uses [φ(~x), π(~y )] =ı̊δ d (~x − ~y ).
We can now determine the contribution to φ from the source. Let us focus on the
future after the source is switched off24
Z
∆φ(x) = φf (x) − φi (x) = dy d+1 ∆(x − y) ρ(y). (3.80)

Transform this expression to momentum space with the Fourier transformation

dpd+1 ı̊p·x
Z
ρ(x) = e ρ(p), ρ(p)∗ = ρ(−p). (3.81)
(2π)d+1
It yields after substitution and evaluation of two integrals
d~p d
Z
ı̊p·x ∗ −ı̊p·x


∆φ(x) = ı̊ρ(p) e −ı̊ρ (p) e . (3.82)
(2π)d 2e(~p)
21
Another option is to write the function as the principal value of the pole 1/(p2 + m2 ) and a
distributional contribution ı̊πδ(p2 + m2 ) sign(p0 ) representing a homogeneous solution.
22
Note that the contour is clockwise.
23
One has to distribute the derivatives in a suitable way between θ and ∆ to arrive at the
intermediate expression.
24
We can therefore set θ(y 0 − x0 ) = 1.

3.16
As we know, the homogeneous Klein–Gordon equation is solved by

d~p d
Z
ı̊p·x ∗ −ı̊p·x


φ(x) = α(~
p) e + α (~
p) e . (3.83)
(2π)d 2e(~p)

Let this represent the solution in the distant past. Then the solution φf in the
distant future is obtained by replacing


αf (~p) = αi (~p) +ı̊ρ e(~p), p~ . (3.84)

We notice that only the Fourier modes of the source ρ on the mass shell can
actually drive the field φ.
We can now ask how much energy, momentum or particle number the source ρ
transfers to the field. We compare the corresponding charges for φf and φi to find

d~p d
Z
2
µ µ   + 2 Im αi (~p)ρ∗ (p) ,


∆P = d
p ρ(p)
(2π) 2e(~p)
d~p d
Z
ρ(p)2 + 2 Im αi (~p)ρ∗ (p) .
 

∆N = (3.85)
(2π)d 2e(~p)

The contribution proportional to |ρ|2 originates from the source alone, and it is
manifestly positive for the energy ∆E = ∆P 0 and particle number ∆N .25 The
other contribution is an interference term with the initial field configuration which
can raise or lower these charges.

25
∆N is not manifestly integer for an classical external field. However, if ρ is itself a quantum
operator related to another field, the combination ρ† (~
p)ρ(~
p) should again lead to an integer N .

3.17
Quantum Field Theory I Chapter 4
ETH Zurich, 2019 HS Prof. N. Beisert
14. 10. 2019

4 Symmetries
So far we have not discussed symmetries. QFT does not actually need symmetries,
but they help very much in restricting classes of models, providing stability and
simplifying calculations as well as results.1
For example, in most cases QFT’s have some symmetry of space and time.
Particularly in fundamental particle physics all models have relativistic invariance
or Poincaré symmetry.
Symmetries are some transformations of the fields φ → φ0 that map solutions of
the equations of motion to other solutions. Hence they can be used to generate a
whole class of solutions from a single one.
We shall discuss the action of various types of symmetries, their groups and
representations, and the resulting conserved charges via Noether’s theorem. Most
of the discussion applies to classical and quantum field theories.

4.1 Internal Symmetries

Let us first discuss internal symmetries. In a QFT with several fields, these
typically transform the fields into each other in some way without making
reference to their dependence on space or time.

Internal Transformations. The simplest example is a complex scalar field φ(x)

with Lagrangian and corresponding equation of motion

µ ∗ 2 ∗ ∂ 2 φ − m2 φ = 0,
L = −∂ φ ∂µ φ − m φ φ, (4.1)
∂ 2 φ∗ − m2 φ∗ = 0.
Consider a global transformation of the fields

φ0 (x) = e+ı̊α φ(x), φ∗ 0 (x) = e−ı̊α φ∗ (x). (4.2)

It maps a solution of the equations of motion to another solution2

∂ 2 φ0 − m2 φ0 = eı̊α ∂ 2 φ − m2 φ = 0.


(4.3)

Moreover the symmetry leaves the Lagrangian and the action invariant

L(φ0 , ∂µ φ0 ) = L(φ, ∂µ φ), S[φ0 ] = S[φ]. (4.4)

1
For free particles symmetries are not that helpful, the true power of symmetries arises in
interacting situations.
2
The transformed field φ0 satisfies the equation of motion because φ does.

4.1
Any such transformation must be a symmetry because it maps extrema of the
action to extrema and hence solutions to solutions. Symmetries of the action are
more powerful than mere symmetries of the equations of motion. In the following
we will only consider symmetries of the action.3

Noether’s Theorem. Every continuous global symmetry of the action leads to

a conserved current and thus a conserved charge for solutions of the equations of
motion.
Let us derive the theorem: Consider a solution φ of the equations of motion. By
construction, any variation of the Lagrangian evaluated on this solution is a total
derivative4
δL δL
δL = δφ + ∂µ δφ
δφ δ(∂µ φ)
 
δL δL δL
= ∂µ δφ + ∂µ δφ = ∂µ δφ . (4.5)
δ(∂µ φ) δ(∂µ φ) δ(∂µ φ)

Suppose now δφ is the infinitesimal field variation generated by a continuous

symmetry. We know that δS = 0, hence the Lagrangian can only change by some
total derivative
δL = δα ∂µ K µ . (4.6)
Equating the two expressions for δL we find a current5
δL δφ
Jµ = − K µ, (4.7)
δ(∂µ φ) δα

which is conserved for every solution φ

∂µ J µ = 0. (4.8)

Furthermore, a conserved current implies a conserved charge

Z
Q(t) = d~x d J 0 (t, ~x) (4.9)

if we assume that the field vanishes sufficiently fast at spatial infinity

Z Z
Q̇ = d~x ∂0 J = − d~x d ∂k J k = 0.
d 0
(4.10)

3
For example the scaling transformation φ(x) → eβ φ(x) also maps solutions to solutions, but
it rescales the Lagrangian L0 = e2β L and likewise the action. If one considers QFT’s to be
specified by their Lagrangians, then this symmetry of the equations of motion relates two
different models L and L0 . We typically use the freedom to redefine the fields to bring the
Lagrangian to some canonical form.
4
Usually we can ignore the total derivative term, here it is relevant.
5
Any term of the form ∂ν B µν with anti-symmetric indices on B µν can be added to J µ without
modifying any of the following relations.

4.2
The conserved charge actually generates an infinitesimal symmetry transformation
via the Poisson brackets
δF
{Q, F } = − (4.11)
δα
as can be shown using its defining relations.6

Example. Let us consider the complex scalar field. The field variation is defined
by
δφ =ı̊φ δα, δφ∗ = −ı̊φ∗ δα. (4.12)
The Lagrangian is invariant under the transformation δL = 0, hence K µ = 0. The
other term reads
δL δφ δL δφ∗
Jµ = +
δ(∂µ φ) δα δ(∂µ φ∗ ) δα
= (−∂ µ φ∗ ) (ı̊φ) + (−∂ µ φ) (−ı̊φ∗ )
= −ı̊(∂ µ φ∗ φ − φ∗ ∂ µ φ). (4.13)
The naive divergence of the current reads
∂µ J µ = −ı̊(∂ 2 φ∗ φ − φ∗ ∂ 2 φ). (4.14)
This indeed vanishes for a solution of the equations of motion.
The conserved charge reads
Z Z
Q =ı̊ d~x (φ̇ φ − φ φ̇) =ı̊ d~x d (π ∗ φ − φ∗ π).
d ∗ ∗
(4.15)

Transformed to momentum space we get

d~p d
Z
∗ ∗


Q= a (~
p)a(~
p) − b (~
p)b(~
p) . (4.16)
(2π)d 2e(~p)
This charge is indeed time-independent and (Poisson) commutes with the
Hamiltonian. As expected, it obeys
δφ δφ∗
{Q, φ} = −ı̊φ = − , {Q, φ∗ } = +ı̊φ∗ = − . (4.17)
δα δα
We furthermore observe a relation to the number operators
Q = Na − Nb . (4.18)
In the quantum theory, Q therefore measures the number of particles created by a†
minus the number of particles created by b† .
Despite the similarities, there is a crucial difference to the number operator: the
charge Q is associated to a symmetry, whereas the individual number operators
Na , Nb are not.7 In a symmetric theory with interactions, Q is conserved while N
is in general not.
6
The difficulty is to deal with the term K which is only implicitly defined. However, it suffices
to show the relationship for the canonical fields F = φ, π.
7
One might construct a non-local symmetry transformation corresponding the number
operator in a free field theory. However, this symmetry would not generalise to interactions.

4.3
Quantum Action. Let us briefly state how to represent this symmetry in the
quantum theory where Q = Na − Nb becomes a quantum operator. It is obviously
hermitian
Q† = Q. (4.19)
It obeys the following commutation relations with creation and annihilation
operators

[Q, a(~p)] = −a(~p), [Q, b(~p)] = +b(~p),

† †
[Q, a (~p)] = +a (~p), [Q, b† (~p)] = −b† (~p). (4.20)

This tells us that particles of type a carry positive unit charge while the
anti-particles of type b carry negative unit charge.
The commutators of spacetime fields φ ∼ a† + b read

Q, φ† (x) = −φ† (x),

   
Q, φ(x) = +φ(x), (4.21)

which tell us that φ and φ† carry charges +1 and −1, respectively. The
commutators are also in agreement with the classical result that charges generate
infinitesimal transformations, i.e.

δφ δφ†
[Q, φ] = +φ = −ı̊ , [Q, φ† ] = −φ† = −ı̊ . (4.22)
δα δα

For finite transformations we introduce the operator

U (α) = exp(ı̊αQ). (4.23)

We can convince ourselves that it obeys the following algebra with the fields8

U (α) φ(x) U (α)−1 = e+ı̊α φ(x) = φ0 (x),

U (α) φ† (x) U (α)−1 = e−ı̊α φ† (x) = φ† 0 (x). (4.24)

So U (α) generates a finite symmetry transformation by means of conjugation while

Q generates the corresponding infinitesimal transformation by means of
commutators.
Note that the operator U (α) is unitary because Q is hermitian

U (α)† = exp(−ı̊αQ† ) = exp(−ı̊αQ) = U (−α) = U (α)−1 . (4.25)

A crucial property of symmetries in QFT is that they are represented by unitary

operators. This is required to make expectation values invariant under symmetry.
The symmetry group for the complex scalar is simply U(1).
The above discussions only apply to operators; let us finally discuss
transformations for states. States transform under finite transformations as

|Ψ 0 i = U (α)|Ψ i. (4.26)
8
Note that Qφ = φ(Q + 1) implies exp(ı̊αQ)φ = φ exp(ı̊α(Q + 1)) = eı̊α φ exp(ı̊αQ).

4.4
Typically, the vacuum is uncharged under symmetries9

Q|0i = 0. (4.27)

The transformation for all other states in the Fock space then follows from the
transformation of creation operators.

4.2 Spacetime Symmetries

Next we shall consider symmetries related to space and time. In relativistic
theories these are the spatial rotations and Lorentz boosts (altogether called
Lorentz symmetries) as well as spatial and temporal translations. In total they
form the Poincaré group. We will see different ways in which the symmetry is
implemented in QFT.

Translations. Let us start with simple translations in space and time

(x0 )µ = xµ + aµ . (4.28)

We demand that the fields merely change by shifting the position argument

φ0 (x0 ) = φ(x). (4.29)

In other words the new field evaluated at the new position equals the old field at
the old position.10 Explicitly,11

φ0 (x) = φ(x − a) or δφ(x) = −δaµ ∂µ φ(x). (4.30)

In order for translations to be a symmetry, we have to require that the Lagrangian

does not explicitly depend on the position
∂L


= 0, L φ(x), ∂µ φ(x), x = L φ(x), ∂µ φ(x) . (4.31)
∂xµ

Energy and Momentum. The Noether theorem equally applies to this

situation, let us derive the associated currents and charges. The symmetry
variation of the Lagrangian is (by definition) through the field variables only
 
µ δL δL
δL = −δa ∂µ φ + ∂µ ∂ν φ . (4.32)
δφ δ∂ν φ
9
An uncharged vacuum is not a requirement in QFT. In fact, a charged vacuum is related to
spontaneous symmetry breaking and Goldstone particles, see QFT II. Note that ordering
ambiguities can arise in the determination of the charges, and are resolved by specifying the
intended charge of the vacuum.
10
Alternatively, one might define φ0 (x) = φ(x0 ) resulting in the opposite transformation rules.
11
The variation is defined via φ0 (x) = φ(x) + δφ(x).

4.5
This expression can be written as a total derivative because there is no other
(explicit) dependence on spacetime
δL = −δaµ ∂µ L(φ, ∂φ) .


(4.33)
We therefore obtain a contribution K µ ν = −δνµ L. Together with the ordinary
contribution (δL/δ(∂µ φ))δφ to the Noether current, we obtain a vector of
conserved currents J µ ν =: T µ ν (with ∂µ T µ ν = 0) where the index ν labels the d + 1
dimensions for shifting
δL
T µν = − ∂ ν φ + η µν L. (4.34)
δ(∂µ φ)
This object is called stress-energy (or energy-momentum) tensor. For a real scalar
field it reads
T µν = ∂ µ φ ∂ ν φ − 21 η µν (∂φ)2 + m2 φ2 .


(4.35)
The corresponding conserved charge is the momentum vector
Z Z
P = d~x T = d~x d −π ∂µ φ − 12 η 0µ (∂φ)2 + m2 φ2 .
µ d 0µ



(4.36)

We recover the Hamiltonian as its time component

Z
~ 2 + 1 m2 φ2 ,
H = P = d~x d 12 π 2 + 21 (∂φ)
0


2
(4.37)

while the total spatial momentum simply reads

Z
~ ~
P = − d~x d π ∂φ. (4.38)

Quantum Action. We have already encountered the quantum operators for

energy and momentum. Recall that in momentum space they read
d~p d
Z
µ
P = pµ (~p) a† (~p)a(~p). (4.39)
(2π)d 2e(~p)
Performing a quantum commutator with the field yields the expected shift
 µ
P , φ(x) =ı̊∂ µ φ(x).

(4.40)
As before, we can introduce an operator U (a) for finite shift transformations as the
exponential
U (a) = exp(ı̊aµ Pµ ). (4.41)
Conjugating a field with it yields the shifted field12
U (a)φ(x)U (a)−1 = exp(−aµ ∂µ )φ(x) = φ(x − a) = φ0 (x). (4.42)
Note that the operator U (a) is unitary because P µ is hermitian.
12
The exponentiated derivative exp(−aµ ∂µ )φ(x) generates all the terms in the Taylor
expansion of φ(x − a) for small a.

4.6
Lorentz Transformations. Next, consider Lorentz transformations
(x0 )µ = (Λ−1 )µ ν xν . (4.43)
All upper (contravariant) indices transform according to the same rule as xµ under
Lorentz transformations, whereas lower (covariant) indices transform with the
matrix Λ, just as ∂µ does, e.g.
(∂ 0 )µ = Λν µ ∂ν . (4.44)
A product between a covariant and contravariant index is Lorentz invariant
x0 ·∂ 0 = (x0 )µ (∂ 0 )µ = Λρ µ (Λ−1 )µ ν xν ∂ρ = xν ∂ν = x·∂. (4.45)

The matrix Λ has the defining property that it leaves the metric ηµν invariant
0
ηµν = ηρσ Λρ µ Λσ ν = ηµν . (4.46)
We can write this relation also as
(Λ−1 )µ ν = ηνσ Λσ ρ η ρµ =: Λν µ . (4.47)
It implies that it makes no difference whether indices are raised or lowered before
or after a Lorentz transformation. Correspondingly, scalar products between equal
types of vectors are invariant.
Lorentz transformations combine spatial rotations (the matrix acts on two of the
spatial dimensions)
   
cos ϕ − sin ϕ 0 −ϕ
= exp (4.48)
sin ϕ cos ϕ ϕ 0
and Lorentz boosts (the matrix acts on time and one of the spatial dimensions)
   
cosh ϑ sinh ϑ 0 ϑ
= exp . (4.49)
sinh ϑ cosh ϑ ϑ 0
There are also some discrete transformations which we shall discuss below. Here
we restrict to proper orthochronous Lorentz transformations which form the Lie
group SO+ (d, 1).
We note that spatial rotations are generated by anti-symmetric matrices while
Lorentz boosts are generated by symmetric matrices. Composing various such
transformations in two-dimensional subspaces of spacetime we conclude that
Lorentz rotations are generated as
Λµ ν = exp(ω)µ ν (4.50)
where ω µ ν is a matrix satisfying
ω k l = −ω l k , ω0k = ωk 0, ω 0 0 = ω k k = 0. (4.51)
Lowering the first index ωµν = ηµρ ω ρ ν , this is equivalent to an anti-symmetric
matrix
ωµν = −ωνµ . (4.52)

4.7
Angular Momentum. For a scalar field the transformation reads

φ0 (x) = φ(Λx), δφ = δω µ ν xν ∂µ φ. (4.53)

Lorentz invariance of the action requires the Lagrangian to transform in the same
way13
δL = δω µ ν xν ∂µ L = δωµν ∂ µ (xν L). (4.54)
Note that the measure dxd+1 is Lorentz invariant. Comparing this to an explicit
variation of L(φ, ∂φ) in terms of the field variables implies the relation
δL δL
∂ν φ = ∂ µ φ. (4.55)
δ(∂µ φ) δ(∂ν φ)
This relation holds whenever ∂µ φ appears only in the Lorentz-invariant
combination (∂φ)2 = η µν (∂µ φ)(∂ν φ). For the stress-energy tensor it implies
symmetry in both indices
T µν = T νµ . (4.56)

The currents J µ,ρσ = −J µ,σρ corresponding to the anti-symmetric matrix δωρσ can
be expressed in terms of the stress-energy tensor T

J µ,ρσ = −T µρ xσ + T µσ xρ . (4.57)

Conservation of J µ,ρσ is then guaranteed by conservation and symmetry of T

∂µ J µ,ρσ = −T σρ + T ρσ = 0. (4.58)

The integral of J is the relativistic angular momentum tensor

Z Z
d 0,µν
µν
= d~x d −T 0µ xν + T 0ν xµ .


M = d~x J (4.59)

For a scalar field in d = 3 dimensional space we obtain the well-known spatial

angular momentum
Z
J = − 2 ε M = d~x 3 π ~x × (∂φ)
m 1 mkl kl ~ m.


(4.60)

Furthermore, the momentum for Lorentz boosts reads14

Z
K = M = d~x 3 T 00 xm − P m t.
m m0


(4.61)

We can also write the Lorentz generators in momentum space15

d~p d
Z
µν µ ν ∗ ν µ ∗


M =ı̊ p ∂ a (~
p)a(~
p) − p ∂ a (~
p)a(~
p) . (4.62)
(2π)d 2e(~p)
13
The anti-symmetry of ωµν allows to pull xν past the derivative.
14
Conservation implies that the motion of the centre of gravity (first term) is governed by the
momentum (second term).
15
Its form is reminiscent of the position space form because Lorentz rotations in of xµ and pµ
are practically the same.

4.8
In the quantum theory, all components of the tensor M µν are hermitian operators.
Consequently, the operators for finite transformations are unitary

U (ω) = exp( 2ı̊ ωµν M µν ), U (ω)† = U (ω)−1 . (4.63)

The interesting conclusion is that we have found a unitary representation of the

Poincaré group. As the latter is non-compact this representation is necessarily
infinite-dimensional. Indeed, the field φ(x) and Fock space carry infinitely many
degrees of freedom.

4.3 Poincaré Symmetry

Above we have derived Lorentz (M µν ) and momentum (P µ ) generators for
relativistic transformations of a scalar field. Let us now discuss the algebraic
foundations and generalisations.

Some Basic Definitions. Here are some sketches of basic definitions in group
and representation theory.16
Group. A set G with an associative composition law G × G → G (usually called
multiplication), a unit element and the inverse map G → G.
Algebra. A vector space A with a bi-linear composition law A ⊗ A → A (usually
called multiplication).
Lie group. A group G that is also a manifold. Continuous symmetries in physics
are realised as Lie groups.
Lie algebra. A non-associative algebra g with an anti-symmetric product [·, ·]
(called Lie bracket) that satisfies the Jacobi identity
     
[a, b], c + [b, c], a + [c, a], b = 0. (4.64)

The tangent space of a Lie group G at the unit element is the corresponding Lie
algebra g.
Quantum group, quantum algebra. The algebra of operators in quantum mechanics
is called a quantum group or a quantum algebra. In QFT it is spanned by the field
operators and their products modulo their commutation relations. In addition to
being an algebra, it has a unit element and an inverse for most elements.
Interestingly, it can act as any of the above structures: It is an algebra by
definition. It may contain Lie groups, e.g. the symmetries of a quantum theory. A
Lie algebra can be realised by the map [a, b] = ab − ba which automatically
satisfies the Jacobi identity.
Representation. A map R : X → End(V ) from a group or an algebra X to linear
operators (matrices, endomorphism) on some vector space V . The representation
16
See a textbook for proper definitions.

4.9
must reflect X’s composition law by operator composition (matrix multiplication).
If ab = c then R(a)R(b) = R(c).
Representation of a Lie algebra. The Lie bracket must be represented by a
commutator: If [a, b] = c then R(a)R(b) − R(b)R(a) = R(c).
Representation space. The vector space V on which a representation
R : X → End(V ) can act. Also known as a module.
Physics. The notation in physics often does not distinguish between abstract Lie
algebra generators a and their representations R(a), both may be denoted simply
by a. Likewise the distinction between Lie brackets and commutators may be
dropped (this is perfectly reasonable in a quantum algebra). Moreover the term
representation is used not only for an operatorial version of algebra generators, but
also for the space on which these operators act.
Unitary groups and algebras. The simplest class of Lie groups are the unitary
groups U(N ) consisting of unitary N × N matrices. In quantum field theory, they
frequently appear as flavour and gauge symmetries. The unitary matrices
themselves also form the fundamental or defining representation. The special
unitary subgroup SU(N ) is given by matrices with unit determinant. The
corresponding Lie algebras u(N ) are spanned by hermitian matrices.17 For the
subalgebra su(N ) the matrices must furthermore be traceless.
Orthogonal groups and algebras. Another large class of Lie groups are the special
orthogonal groups SO(N ) which describe rotations in N -dimensional real space.
The group SO(N, M ) is a generalisation which describes rotations in an
(N + M )-dimensional space of signature (N, M ). In particular, SO(d, 1) is the
Lorentz group for a spacetime of dimension d + 1. The corresponding Lie algebras
so(N ) and so(N, M ) are spanned by anti-symmetric matrices (with respect to a
metric of the given signature).
Double cover of the orthogonal groups. A peculiar feature of the orthogonal groups
SO(N, M ) is that they have a double cover called Spin(N, M ). In the latter group,
a rotation by an angle of 2π is a non-trivial central element. It can act as either
+1 for vectorial representations (integer spin) or −1 for spinorial representations
(half-integer spin). In SO(N, M ) this element is projected out reducing the set of
allowable representations to the vectorial ones. The double cover is a topological
issue which does not affect the local Lie algebra; it is so(N, M ) for both groups.

Algebra and Group. It is straight-forward to derive the algebra of infinitesimal

transformations from the spacetime symmetry operators derived earlier

[M µν , M ρσ ] =ı̊η νρ M µσ −ı̊η µρ M νσ −ı̊η νσ M µρ +ı̊η µσ M νρ ,

[M µν , P ρ ] =ı̊η νρ P µ −ı̊η µρ P ν ,
[P µ , P ν ] = 0. (4.65)
17
Depending on conventions for the basis of Lie algebras one uses either hermition matrices
(physics) or anti-hermitian matrices (mathematics).

4.10
These define the so-called Poincaré algebra. The operators M µν generate the
algebra so(d, 1) of Lorentz (orthogonal) transformations in d + 1 spacetime
dimensions. The spatial components M jk generate the algebra so(d) of rotations in
d spatial dimensions.
The Poincaré group is obtained by exponentiating the algebra18

U (ω, a) = exp 2ı̊ ωµν M µν +ı̊aµ P µ .

(4.66)

More precisely it is the component of the Poincaré group connected to the identity
element. It includes Spin+ (d, 1), the double cover of the proper orthochronous
Lorentz group, along with translations.

Fock Space Representations. The above symbols (M, P ) and their

relationships can be interpreted in two ways:
They can serve as abstract elements of a Lie algebra without immediate
connection to a physics problem. In that case the above relations should be viewed
as the defining Lie brackets and group multiplication laws.
This is not the approach we had taken. Instead, we have derived them as quantum
operators from Noether’s theorem of a physical system. Their algebraic relations
are expressed as commutators and products of operators. They form an explicit
representation (M, P ) of the Poincaré algebra and the corresponding
representation U (ω, a) of the Poincaré group. The representation space is the Fock
space of a scalar particle.
Since (M, P ) commute with the number operator N , the representation is
reducible.19
• The most relevant representation is the one acting on single particle states.
• The multi-particle representations are symmetric tensor powers of it; they are
generally reducible.
• The vacuum transforms in the trivial representation.
• The single-particle representation is complex, unitary, infinite-dimensional and
irreducible.

4.4 Poincaré Representations

We have seen how the Poincaré algebra acts on the free Klein–Gordon field and its
Fock space. In the following we will reverse the logic, and use the abstract
Poincaré algebra to derive the concept of relativistic particles. We will see that the
18
The exponentiation can be specified in different forms such as exp( 2ı̊ ωµν M µν ) exp(ı̊aµ P µ ) or
exp(ı̊aµ P µ ) exp( 2ı̊ ωµν M µν ). These forms are all equivalent but a would take different values in
the various cases.
19
In other words, the action of (M, P ) neither creates nor annihilates particles and therefore
maps Vn → Vn . The representation on Fock space thus splits into representations on the
individual n-particle subspaces Vn .

4.11
particles are characterised by their mass and spin, which leads to generalisations of
the Klein–Gordon field in the following chapters.
Let us therefore investigate the unitary irreducible representations (UIR’s) of the
Poincaré group (Wigner’s classification). These will be the elementary building
blocks for physical theories with relativistic invariance. The derivation will parallel
the derivation of unitary irreducible representations of the rotation group
SO(3) ' SU(2) (known from elementary quantum mechanics) which leads an
understanding of spin. Here the result will characterise the types of admissible
particles in a relativistic QFT.
The following construction will start with a set of states. The action of the
symmetry operators requires the introduction of further states. As soon as the
symmetry generators close on the space of states, we have found a complete
representation and its corresponding representation space.

Mass. First, we should look for commuting combinations C of elements of the

algebra (respectively commuting elements of the quantum algebra). Their
eigenvalues classify irreducible representations because if measured on one state
|Xi, any other state |Y i = J|Xi related to it by some symmetry operator J must
have the same eigenvalue c:
C|Y i = CJ|Xi = JC|Xi = cJ|Xi = c|Y i. (4.67)
The principal example in so(3) is the total spin operator C = J 2 . There, a
representation of spin j is uniquely characterised by the eigenvalue j(j + 1) of J 2 .
We notice that the Poincaré algebra possesses a quadratic invariant
P 2 = P µ Pµ . (4.68)
This combination obviously commutes with all the momenta P µ . It also commutes
with the Lorentz generators M µν because it is constructed as a scalar product.
The combination P 2 must be represented by a unique number on an irreducible
representation, see above. Clearly, P 2 measures the mass of a particle
P 2 = −m2 . (4.69)
For unitary representations P 2 must be real. There are three cases to be
distinguished:
• P 2 < 0, i.e. massive particles;
• P 2 = 0, i.e. massless particles;
• P 2 > 0, i.e. tachyons.
We shall discuss the massive case in detail and comment only briefly on the others.

Momentum. The next observation is that the momentum generators P µ span

an abelian ideal20 This property is useful because we can treat the representation
20
An ideal is a subalgebra such that brackets between its elements and elements of the algebra
always end up in the subalgebra, here [M, P ] ∼ P .

4.12
of the ideal independently of the other generators of the Poincaré algebra.21
Moreover the ideal is abelian, and therefore we can choose simultaneous eigenstates
|pi of all the generators P µ as basis vectors for the representation space,
P µ |pi = pµ |pi. (4.70)

As the representation of a finite translation by aµ is given by exp(ı̊a·P ), the

representation is necessarily complex.
We have already fixed P 2 = −m2 within our irreducible representation, and hence
we must restrict the representation space to a mass shell p2 = −m2 . For a given
spatial momentum the constraint has two solutions for the energy
pµ = ±e(~p), p~ .


(4.71)
We can thus label the basis states for our representation space by means of their
spatial momentum p~ and the sign of energy.
|~p im,+

|~0im,+
(4.72)

|~p im,−

This completes the discussion of the representation of momentum generators P µ .

What about the Lorentz generators M µν ?
Orthochronous Lorentz boosts can map between any two momentum vectors on
the same mass shell. However, they cannot map between the forward and
backward mass shells; this is achieved only by discrete time reversal
transformations which we will consider later. Consequently, for an irreducible
representation of the orthochronous Poincaré group, all energies must have the
same sign. The irreducible representations are thus labelled by the sign of the
energy in addition to the mass m. The positive-energy representation space is now
spanned by the vectors
|~p im,+ = |+em (~p), ~p i. (4.73)
For negative energies the representation space is spanned by
|~p im,− = |−em (~p), ~p i. (4.74)
Note that, in physics, the negative-energy representation is typically obtained as
the hermitian conjugate of the positive-energy representation
|~p im,− ∼ h+em (~p), ~p |† . (4.75)
21
More precisely, individual states of the representation of the ideal will become subspaces of
the representation of the complete algebra.

4.13
Let us consider only positive energies from now on; negative energy representations
are analogous.

Spin. Among the Lorentz generators, there are some which change a given vector
pµ = (e(~p), p~) on the mass shell. For the particle at rest, (m, ~0), these are the
Lorentz boosts. The Lorentz boosts can be used to map the momentum vector
(m, ~0) to any other admissible momentum vector (e(~p), p~). Therefore, they ensure
that the following discussion for (m, ~0) equivalently applies to any other pµ on the
same mass shell.
The transformations which do not change pµ = (m, ~0) are the spatial rotations
which form the orthogonal group SO(d) or its double cover Spin(d).22 23 This
group is called the little group (physics) or stabiliser (mathematics) of pµ . The
representation subspace with fixed p~ must therefore transform under a
representation of Spin(d).
For the most relevant case of d = 3 spatial dimensions, the unitary irreducible
representations of Spin(3) = SU(2) are labelled by a non-negative half-integer j.
Their representation space is spanned by 2j + 1 vectors
|−jij , |−j − 1ij , . . . , |+j − 1ij , |+jij with definite z-component of spin. An
equivalent representation of Spin(d) must apply to all momenta p~ because it can
be shifted to the point p~ = ~0.24
We have now considered all algebra generators and hence the representation is
complete. The representation space is thus spanned by the states

|~p, j3 i(m,±,j) = |~p im,± ⊗ |j3 ij . (4.76)

Unitary Irreducible Representations. Altogether we find that the massive

UIR’s of the Poincaré algebra are labelled by their mass m > 0, the sign of energy
and a unitary irreducible representation of Spin(d). In the case of d = 3, the latter
UIR are labelled by a non-negative half-integer j. The representation space for
(m, ±, j) is spanned by the vectors

|~p, j3 i(m,±,j) (4.77)

with continuous p~ and discrete j3 = −j, −j + 1, . . . , j − 1, j.

For spin j = 0 the representation space is simply spanned by momentum
eigenstates |~pi with arbitrary three-momentum p~. These are just the single-particle
states of a scalar field. The conjugate states h~p| also transform in a UIR, but one
with negative momentum.
22
Reflections extend SO(d) to O(d) or Spin(d) to Pin(d), but they are not included in the
identity component of the Poincaré algebra.
23
One obtains the same group SO(d) for any time-like p, but they form different subgroups
within SO+ (d, 1). In fact, the mass shell is a coset space SO+ (d, 1)/SO(d).
24
Alternatively, one can introduce the Pauli–Lubanski vector Wµ = 12 εµνρσ P ν M ρσ . It
generates the appropriate so(3) stabiliser subalgebra of so(3, 1) for arbitrary momentum p~. Its
invariant square W 2 = m2 j(j + 1) yields the total spin j.

4.14
The next interesting case is j = 1/2 which we shall discuss in the following chapter
of the lecture.
In addition, there are massless representations of positive or negative energy. They
are classified by a representation of Spin(d − 1).25 For d = 3 the massless
representations of Spin(2) = U(1) are labelled by a positive or negative half-integer
h known as helicity. There is only one state in the representation (0, ±, h) with
given helicity
|~p i(0,±,h) . (4.78)
This includes the case h = ±1 which is used for the photon excitations of the
electromagnetic field.
Last but not least, there is the trivial representation with P = 0. The Fock space
vacuum transforms under it. Finally, there are tachyonic representations with
P 2 > 0, but they are clearly unphysical. In particular, the stabiliser of a tachyonic
momentum vector is the non-compact group SO(d − 1, 1) whose unitary
representations are necessarily trivial or infinite-dimensional.

4.5 Discrete Symmetries

In addition to the continuous symmetries discussed above, there are also relevant
discrete symmetries. The most prominent ones are parity, time reversal and charge
conjugation. Let us discuss them for the example of a quantised complex scalar
field φ(t, ~x).

Parity. Spatial rotations in d dimensions form the special orthogonal group

SO(d). However, also spatial reflections preserve all distances, and it is natural to
consider them among the symmetries, too. Reflections were long believed to be a
symmetry of nature, until the electroweak interactions were shown to violate
parity symmetry. On the mathematical side, reflections flip the orientation and
together with the rotations they form the general orthogonal group O(d).26
For an odd number of spatial dimensions d, it is convenient to introduce parity P
as the unique transformation which inverts all spatial components of the position
vector27
P : (t, ~x) 7→ (t, −~x). (4.79)
It is an element of the group O(d), but not of SO(d), and it is convenient to choose
this element because it does not introduce any preferred directions. There are
many more orientation-inverting elements in O(d); these can be obtained as
25
In fact, the stabiliser is the Euclidean group in d − 1 dimensions which also allows for
so-called continuous spin representations.
26
The double cover of O(d) is called Pin(d) in analogy to Spin(d) which is the double cover of
SO(d).
27
For even d this transformation would be a rotation rather than a reflection, and there is no
reflection that commutes with all rotations.

4.15
products P R of parity P with elements R of SO(d). Hence it is sufficient to
consider only P . This parity reflection acts as an involution on spacetime,

(t, ~x) 7→ (t, −~x) 7→ (t, ~x), (4.80)

i.e. it makes sense to demand P 2 = 1 at the level of the operator. In spacetime,

introducing parity enlarges the proper orthochronous Lorentz group SO+ (d, 1) to
the orthochronous Lorentz group O+ (d, 1).
In the quantum theory, P is represented by a linear operator. A scalar field φ(t, ~x)
is also an operator, and as such it transforms under parity by conjugation with P .
By definition, the result of this operation should flip the position vector ~x and we
expect the following relationships

P φ(t, ~x)P −1 = ηP φ(t, −~x),

P φ† (t, ~x)P −1 = ηP∗ φ† (t, −~x), (4.81)

where the constant ηP is the intrinsic parity of the field φ. We want that two
consecutive parity transformations equal the identity P 2 = 1, therefore the parity
of scalar fields can be either positive or negative, ηP = ±1. Typically, the vacuum
state has positive parity, P |0i = |0i.
For the creation and annihilation operators the above relationship implies a
transformation which reverses the momentum

P a(~p)P −1 = ηP a(−~p), P a† (~p)P −1 = ηP a† (−~p),

P b(~p)P −1 = ηP b(−~p), P b† (~p)P −1 = ηP b† (−~p). (4.82)

This is a unitary operation because P † = P = P −1 .

Time Reversal. The other discrete transformation of the Lorentz group is time
reversal
T : (t, ~x) 7→ (−t, ~x). (4.83)
It enlarges the orthochronous Lorentz group O+ (d, 1) to the complete Lorentz
group O(d, 1).
Time reversal is a rather special transformation due to the distinguished role of
time in quantum mechanics and special relativity.
For a complex scalar field φ we expect

T φ(t, ~x)T −1 = ηT φ(−t, ~x),

T φ† (t, ~x)T −1 = ηT∗ φ† (−t, ~x). (4.84)

Comparing to the mode expansion of fields, this could be implemented by a linear

transformation of the type a† (~p) 7→ b(−~p). In fact, it makes perfect sense that time
reversal interchanges particle creation and annihilation as much as it flips the sign
of energy E. However, such a transformation could not act consistently within a

4.16
Fock space because all states of the latter have positive definite energy, and
naively, T would seem to annihilate all states but the vacuum.28
Instead, time reversal (sometimes called motion reversal) is defined by an
anti-linear operator T̄ ,29 which also conjugates plain complex numbers

T̄ a|Xi = a∗ T̄ |Xi. (4.85)

This inverts the plane wave factors e±ı̊p·x of the mode expansion of φ(t, ~x), and
allows to map creation to creation operators a† 7→ a† such that the structure of
Fock space is preserved.30 More explicitly

T̄ a(~p)T̄ −1 = ηT̄∗ a(−~p), T̄ a† (~p)T̄ −1 = ηT̄ a† (−~p),

T̄ b(~p)T̄ −1 = ηT̄ b(−~p), T̄ b† (~p)T̄ −1 = ηT̄∗ b† (−~p). (4.86)

The difference with respect to parity is merely the anti-linear feature of T̄ . Time
reversal actually allows for a complex ηT̄ only restricted by |ηT̄ |2 = 1. For the
vacuum we may define T̄ |0i = |0i.

Charge Conjugation. Also the internal symmetry groups can come along with
several connected components. For example the complex scalar field has a global
U(1) = SO(2) symmetry. This can be extended to O(2) by adding a charge
conjugation symmetry.
We already know that the complex conjugate scalar field φ∗ or φ† satisfies the
same equations of motion as the original field φ. Charge conjugation symmetry
thus maps between the fields φ and φ†

Cφ(x)C −1 = ηC φ† (x),
Cφ† (x)C −1 = ηC∗ φ(x). (4.87)

Requiring that two charge conjugations square to unity, the parity ηC must be on
the complex unit circle |ηC |2 = 1.

Ca(~p)C −1 = ηC∗ b(~p), Ca† (~p)C −1 = ηC b† (~p),

Cb(~p)C −1 = ηC a(~p), Cb† (~p)C −1 = ηC∗ a† (~p). (4.88)

Although C maps φ 7→ φ† , it is a perfectly linear map. Charge conjugation is not

complex conjugation. One might as well make an anti-linear ansatz for C, but it
28
More accurately, time reversal would map the ordinary vacuum |0i to the vacuum |00 i of a
conjugate Fock space with negative definite energy. Such an alternative vacuum is annihilated by
all creation operators a† , b† and therefore it is analogous to the hermitian conjugate vacuum h0|.
In fact, this identification naturally makes the linear operation equivalent to the conventional
anti-linear operation upon hermitian conjugation.
29
We denote this operation by T̄ in order to emphasise its anti-linearity; more commonly it is
denoted by T in the literature.
30
The idea underlying the anti-linear operation T̄ is to view the reverse time evolution of the
original system as forward time evolution of some mirror system. The spaces of states for these
two systems are formally not the same, but they should be the equivalent, hence we can identify
their Fock spaces and let T̄ act within it.

4.17
would lead to a transformation of the kind a† 7→ a which could not act within Fock
space.
There are several conceptual difficulties with charge conjugation parity:
• In the presence of a corresponding internal symmetry the parity ηC actually
does not have deeper meaning. In this case one can define a new charge
conjugation operation C 0 by conjugating C with the internal symmetry. This
would lead to a different ηC , and it makes sense to choose C 0 such that ηC = 1.
• Even if there is no continuous internal symmetry, there can be discrete internal
symmetries. For example, a possible transformation for a real scalar field is
φ 7→ −φ.
• In a model with multiple fields, several independent internal parities can
coexist, and there may not be a distinguished charge conjugation symmetry. In
general, one would expect C to invert all internal charges.
• In the presence of some internal parity C, the spacetime parities P and T
become somewhat ambiguous, as one could define P 0 = P C.31
Hence the choice of discrete symmetries C, P , T can be ambiguous, and it is
sometimes tricky to identify the most suitable (or the established) one.32

Implications. A discrete transformation is a symmetry of a Hamiltonian system

if it commutes with the Hamiltonian. It is natural to assume parity and time
reversal as symmetries of relativistic QFT models and of nature. For a single
scalar field, it appears impossible to violate parity or time reversal. However, as we
shall see, this need not be so for other types of fields. In nature, indeed, some of
these symmetries are violated.
Discrete symmetries also lead to conservation laws in the quantum theory. States
can be classified by their eigenvalue (parity) under the discrete symmetry.
Typically these parities take only finitely many values (e.g. +1 or −1).

31
This may appear
√ strange at first sight as P 0 would conjugate the field φ. But by writing
φ = (φ1 +ı̊φ2 )/ 2 we merely get two fields with opposite parities ηP .
32
For example, the statement that a system has a parity symmetry does not imply that the
parity acts in the conventional fashion; it merely meant that there is some definition of parity
that is a symmetry, e.g. by transforming the fields in a non-diagonal fashion.

4.18
Quantum Field Theory I Chapter 5
ETH Zurich, 2019 HS Prof. N. Beisert
12. 11. 2019

5 Free Spinor Field

We have seen that next to the scalar field there exist massive representations of
the Poincaré algebra with spin. The next higher case is spin j = 1/2 . It is described
by the Dirac equation, and as a field with half-integer spin it should obey Fermi
statistics.

5.1 Dirac Equation and Clifford Algebra

Dirac Equation. Dirac attempted to overcome some of the problems of
relativistic quantum mechanics by introducing a first-order wave equation.1 2

γ µ ∂µ ψ − mψ = 0. (5.1)

Here, the γ µ are some suitably chosen operators acting locally on the wave
function ψ. This wave equation can be viewed as a factorisation of the
second-order Klein–Gordon equation as follows:

(γ ν ∂ν + m)(γ µ ∂µ − m)ψ = (γ ν γ µ ∂ν ∂µ − m2 )ψ = 0. (5.2)

The latter form becomes the Klein–Gordon equation provided that the γ’s satisfy
the Clifford algebra3 4

{γ µ , γ ν } = γ µ γ ν + γ ν γ µ = 2η µν . (5.3)

This means that every solution of the Dirac equation also satisfies the
Klein–Gordon equation and thus describes a particle of mass m.

Clifford Algebra. The Clifford algebra obviously cannot be realised in terms of

plain numbers, but finite-dimensional matrices suffice. The realisation of the
Clifford algebra strongly depends on the dimension and signature of spacetime.
1
The combination of a gamma-matrix and an ordinary vector γ µ Bµ is often denoted by a
slashed vector B;
/ however, it is questionable whether this notation improves readability. As a less
common alternative one could simply write B (without vector index) as a short form of γ µ Bµ .
2
The sign in front of the mass term has no physical significance as long as chosen consistently.
Analogously, the mass m could be a negative number, but for simplicity will always assume
m ≥ 0.
3
The indices of the two derivatives are automatically symmetric, hence only the
symmetrisation of γ µ γ ν must equal η µν .
4
We use gamma-matrices for signature −+++ which implies that some factors of −1 and ı̊ are
needed to match with other conventions used in the literature.

5.1
The simplest non-trivial case is three-dimensional space (without time). A
representation of the corresponding Clifford algebra is given by the 2 × 2 Pauli
matrices      
1 0 1 2 0 −ı̊ 3 +1 0
σ = , σ = , σ = . (5.4)
1 0 +ı̊ 0 0 −1
One can convince oneself that these matrices obey the algebra5

σ j σ k = δ jk +ı̊εjkl σ l . (5.5)

This also implies the three-dimensional Clifford algebra

{σ i , σ j } = 2δ ij . (5.6)

In this course we will be predominantly interested in the case of d = 3 spatial

dimensions plus time, i.e. spacetime with D = d + 1 = 4 dimensions. There, the
smallest non-trivial representation of the Clifford algebra is four-dimensional
(coincidence!). The elements of this four-dimensional vector space are called
spinors, more precisely, Dirac spinors or 4-spinors.
There are many equivalent ways to write this representation as 4 × 4 matrices.
The best-known ones are the Dirac, Weyl and Majorana representations. These
are often presented in a block form of 2 × 2 matrices whose elements are again
2 × 2 matrices. The latter are written using the Pauli matrices σ i or the 2 × 2 unit
matrix 1. We shall mainly use the Weyl representation
   
0 0 1 k 0 σk
γ = , γ = . (5.7)
−1 0 σk 0
One can easily confirm that these matrices obey the Clifford algebra
{γ µ , γ ν } = 2η µν by means of the three-dimensional Clifford algebra. A useful
property of the Weyl representation is that all four gamma-matrices are block
off-diagonal. The Dirac and Majorana representations have different useful
properties. In most situations, it is however convenient not to use any of the
explicit representations, but work directly with the abstract Clifford algebra.

Solutions. The Dirac equation is homogeneous, therefore it is conveniently

solved by Fourier transformation
Z
ψ(x) = dp4 eı̊p·x ψ(p), (ı̊pµ γ µ − m)ψ(p) = 0. (5.8)

To construct the solutions, let us introduce the matrices

1
Π± = (m ∓ı̊p·γ) (5.9)
2m
such that the Dirac equation becomes Π + ψ = 0. We are interested in the kernel of
Π +.
5
We will not introduce a distinguished symbol for unit matrices. 1 is the unit element. Here
the term δ ij has an implicit 2 × 2 unit matrix.

5.2
As noted above, we have the identity
1
Π −Π +ψ =
(m2 + p2 )ψ. (5.10)
4m2
This operator acts identically on all components of ψ. Any solution therefore
requires the mass shell condition p2 = −m2 .
On the mass shell p2 = −m2 , the operators Π ± act as a complete set of orthogonal
projectors:
Π ±Π ± = Π ±, Π ± Π ∓ = 0, Π + + Π − = 1. (5.11)
Both of them have half-maximal rank, tr Π ± = 2. The Dirac equation thus has
two solutions for each on-shell momentum p.
A basis of two positive-energy solutions is denoted by
uα (~p), α = ±, (ı̊p·γ − m)uα (~p) = 0. (5.12)
Instead of introducing negative-energy solutions, we prefer to consider equivalent
positive-energy solutions of the opposite Dirac equation Π − ψ = 0 6
vα (~p), α = ±, (ı̊p·γ + m)vα (~p) = 0. (5.13)

To write such solutions explicitly, we can recycle the projectors Π ± and set
u = Π − λ, v = Π + λ, (5.14)
where λ is some spinor. The properties of the projectors immediately show that u
and v are solutions to their respective equations. Note, however, that some
components of λ are projected out in u and in v.
Let us consider explicitly solutions in the Weyl representation. E.g. setting
λ = (κ, 0) with κ some 2-spinor, we find
  
1 m −ı̊e(~p) +ı̊~p·~σ κ
u(~p) =
2m ı̊e(~
p) +ı̊~
p ·~
σ m 0
 
1 mκ
= ,
2m ı̊e(~p)κ +ı̊~p·~σ κ
 
1 mκ
v(~p) = . (5.15)
2m −ı̊e(~p)κ −ı̊~p·~σ κ
There are two independent choices for the 2-spinor κ, hence there are two solutions
for u and v, respectively. One typically considers uγ (~p), vγ (~p), γ = ±, as two pairs
of fixed basis vectors for each momentum p~.
Altogether the general solution can now be expanded as
d~p 3
Z X
ı̊p·x −ı̊p·x †


ψ(x) = e u γ (~
p)b γ (~
p) + e vγ (~
p)a γ (~
p) . (5.16)
(2π)3 2e(~p) γ=±
Here the negative-energy coefficient bγ is chosen differently from aγ because
gamma-matrices are generally complex and therefore also the Dirac spinor ψ is
complex. This means that the anti-particles (e.g. positrons) are different from the
particles (e.g. electrons).
6
At the level of the Dirac equation the replacement m → −m is equivalent to p → −p.

5.3
5.2 Poincaré Symmetry
The Dirac equation is a relativistic wave equation. Translational invariance is
evident, but we have not yet shown its Lorentz covariance (although the resulting
Klein–Gordon equation certainly is covariant).

Lorentz Symmetry. Let us therefore consider a Lorentz transformation

x0 = Λ−1 x with Λ(ω) = exp(ω). Suppose ψ is a solution of the Dirac equation. It is
not sufficient to use the transformation rule for scalar fields ψ 0 (x0 ) = ψ(x). In
analogy to vectors we should also transform spinors. We make the ansatz
ψ 0 (x0 ) = S(ω)ψ(x), (5.17)
where S(ω) is a matrix that acts on Dirac spinors. We then substitute
ψ 0 (x) = Sψ(Λx) into the Dirac equation
0 = γ µ ∂µ − m ψ 0 (x) = γ µ ∂µ − m Sψ(Λx)

= γ ν SΛµ ν ∂µ ψ − Smψ (Λx)

= S S −1 γ ν SΛµ ν ∂µ ψ − γ µ ∂µ ψ (Λx)

= S Λµ ν S −1 γ ν S − γ µ (∂µ ψ)(Λx).


(5.18)
So the term in the bracket must vanish for invariance of the Dirac equation.
Indeed, the canonical Lorentz transformation of gamma-matrices
γ 0µ = (Λ−1 )µ ν Sγ ν S −1 , (5.19)
where not only the vector index is transformed by Λ−1 , but also the spinor matrix
is conjugated by the corresponding spinor transformation S.7 In analogy to the
invariance of the Minkowski metric, η 0 = η, the Dirac equation is invariant if the
gamma-matrices are invariant
γ 0µ = γ µ . (5.20)
This condition relates S to the Lorentz transformation Λ.
The infinitesimal form of the invariance condition reads
[δS, γ µ ] − δω µ ν γ ν = 0. (5.21)
This implies that δS must be proportional to δωµν . The latter carries two vector
indices, while δS carries none. The only possibility is to contract the vector indices
to gamma-matrices, and we make the ansatz δS = 21 α δωµν γ µ γ ν . Substituting this
into the invariance condition and using
[γ ρ γ σ , γ µ ] = γ ρ {γ σ , γ µ } − {γ ρ , γ µ }γ σ , (5.22)
we arrive at −(2α + 1)δω µ ν γ ν = 0. We conclude that a Lorentz transformation for
spinors is given by the matrix
δS = − 14 δωµν γ µ γ ν or S(ω) = exp − 41 ωµν γ µ γ ν .


(5.23)
7
In general, spinors are transformed by S from the left, co-spinors by the S −1 from the right.

5.4
Comparing this result to the abstract form of finite Lorentz transformations as
U (ω) = exp( 2ı̊ ωµν M µν ) we have derived a new representation on spinors8
M µν = 4ı̊ [γ µ , γ ν ]. (5.24)
This representation obeys the Lorentz algebra derived above, i.e.
[M µν , M ρσ ] =ı̊M + . . ..

Double Cover. Spinor representations exist only for the double cover Spin(N )
of an orthogonal group SO(N ). Let us observe this fact in a simple example.
Consider a rotation in the x,y-plane with angle ω12 = −ω21 = ϕ. The associated
finite Lorentz transformation matrix in the x,y-plane reads
 
cos ϕ sin ϕ
Λ(ϕ) = exp(ω) = . (5.25)
− sin ϕ cos ϕ
The associated spinor transformation reads
S(ϕ) = diag(e−ı̊ϕ/2 , e+ı̊ϕ/2 , e−ı̊ϕ/2 , e+ı̊ϕ/2 ). (5.26)
The vector rotation Λ(ϕ) is 2π-periodic in ϕ whereas the spinor rotation is merely
4π-periodic. The rotation by ϕ = 2π is represented by the unit matrix for vectors,
but for spinors it is the negative unit matrix
Λ(2π) = 1 = (−1)F , S(2π) = −1 = (−1)F . (5.27)
The spin group thus has an element which represents a rotation by 2π
(irrespectively of the direction). On vector representations (integer spin) it acts as
the identity, on spinor representations (half-integer spin) it acts as −1. Due to the
relation between spin and statistics, the extra element is equivalent to (−1)F
where F measures the number of fermions (odd for spinors, even for vectors).

Chiral Representation. There is an important feature of the spin

representation M µν which is best observed in the Weyl representation of
gamma-matrices  
µ 0 σµ
γ = . (5.28)
σ̄ µ 0
Here we have introduced the sigma-matrices σ µ , σ̄ µ as an extension of the Pauli
matrices σ k to four spacetime dimensions as follows
 
0 0 1 0
σ = −σ̄ := , σ̄ k := σ k . (5.29)
0 1
The Lorentz representation now reads
 
µν ı̊ σ µ σ̄ ν − σ ν σ̄ µ 0
M = . (5.30)
4 0 σ̄ µ σ ν − σ̄ ν σ µ
8
We have to make sure that M µν is anti-symmetric in its vector indices, hence the explicit
anti-symmetrisation.

5.5
This representation has block-diagonal form and therefore reduces to two
independent representations M µν = diag(MLµν , MRµν ) with
MLµν = 4ı̊ (σ µ σ̄ ν − σ ν σ̄ µ ), MRµν = 4ı̊ (σ̄ µ σ ν − σ̄ ν σ µ ). (5.31)

In other words, the Dirac spinor ψ = (ψL , ψR ) transforms in the direct sum of two
(irreducible) representations of the Lorentz group. The 2-spinors ψL and ψR are
called left-chiral and right-chiral spinors. The massive Dirac equation, however,
mixes these two representations
σ µ ∂µ ψR − mψL = 0,
σ̄ µ ∂µ ψL − mψR = 0. (5.32)
It is therefore convenient to use Dirac spinors for massive spinor particles whereas
massless spinor particles can also be formulated using 2-spinors; we shall discuss
the massless case later on.
The decomposition into chiral parts is not just valid in the Weyl representation of
the Clifford algebra. More abstractly, it is due to the existence of the matrix
γ5 = ı̊
ε
24 µνρσ
γ µγ ν γ ργ σ =ı̊γ 0 γ 1 γ 2 γ 3 . (5.33)
In the Weyl representation it reads γ 5 = diag(−1, +1), it therefore measures the
chirality of spinors. In general, it anti-commutes with all the other
gamma-matrices,
{γ 5 , γ µ } = 0. (5.34)
This property implies that a single gamma-matrix maps between opposite
chiralities, i.e. it inverts chirality. The property is also sufficient to prove
commutation with M µν . Alternatively, it follows by construction of γ 5 as a
(pseudo)-scalar combination of gamma-matrices.
A further useful property is
γ 5 γ 5 = 1. (5.35)
1 5
It can be used to show that the combinations 2
(1 ± γ ) are two orthogonal
projectors to the chiral subspaces.

Sigma-Matrices. Let us briefly discuss the sigma-matrices which are chiral

analogs of the gamma-matrices. The sigma-matrices obey an algebra reminiscent
of the Clifford algebra9 10
σ µ σ̄ ν + σ ν σ̄ µ = 2η µν = σ̄ µ σ ν + σ̄ ν σ µ . (5.36)
The two variants of the sigma-matrices are related by transposition (modulo
conjugation by the anti-symmetric matrix ε)
 
µ T −1 µ −1 2 0 +1
ε(σ ) ε = −σ̄ , ε = −ε :=ı̊σ = . (5.37)
−1 0
9
The assignment of bars enables a 2-dimensional representation for this algebra unlike the
Clifford algebra which requires a larger 4-dimensional representation.
10
Note that for any reasonable product of sigma-matrices the sequence of factors will alternate
between σ and σ̄. This agrees with the fact that a single γ maps between the two chiralities.

5.6
Inspection shows that all sigma-matrices are hermitian

(σ µ )† = σ µ , (σ̄ µ )† = σ̄ µ . (5.38)

Since there are 4 independent 2 × 2 hermitian matrices, the four sigma-matrices σ µ

(or equivalently σ̄ µ ) form a real basis for such matrices. Likewise, the 6 matrices
MLµν (or equivalently MRµν ) form a real basis of 2 × 2 complex traceless matrices.11
Furthermore,
(MLµν )† = MRµν . (5.39)
These are just the defining relations for the fundamental representation of sl(2, C)
along with its conjugate representation.12
The Lorentz algebra so(3, 1) is indeed equivalent to the algebra sl(2, C). At the
level of groups, Spin+ (3, 1) = SL(2, C) is the double cover of SO+ (3, 1).
• A chiral 2-spinor of Spin+ (3, 1) transforms in the fundamental representation of
SL(2, C).
• Similarly, a 2-spinor of opposite chirality transforms in the conjugate
fundamental representation of SL(2, C).
• Spinor representations exist only for the double-cover group Spin+ (3, 1), but
not for the original Lorentz group SO+ (3, 1).

5.3 Discrete Symmetries

In addition to the continuous Poincaré symmetry and an obvious U(1) internal
symmetry, there are several discrete symmetries and transformations which we
shall now discuss. These are also needed to formulate a Lagrangian.

Parity. Spatial parity ~x 0 = −~x is the simplest discrete symmetry. We make the
usual ansatz
ψ 0 (t, −~x) = γP ψ(t, ~x), (5.40)
where γP is a matrix that induces the reflection on spinors.
The new field obeys the same old Dirac equation provided that the
gamma-matrices are invariant
!
γ 0µ := Λµ ν γP γ ν γP−1 = γ µ . (5.41)

We need to find a matrix γP that

• commutes with γ 0 (because Λ0 0 = 1),
• anti-commutes with γ k (to compensate Λk k = −1),
11
A complex traceless 2 × 2 matrix has 2 · 2 − 1 = 3 complex degrees of freedom which are
equivalent to 6 real ones.
12
The group SL(N ) of matrices with unit determinant is generated by the algebra sl(N ) of
traceless matrices.

5.7
 • squares to unity (because P 2 = 1).13
This matrix is easily identified (up to a sign) as

γP = −ı̊γ 0 . (5.42)

Note that γP interchanges the two chiralities. Hence the Dirac spinor is
• reducible to chiral spinors under proper orthochronous Lorentz rotations,
• but irreducible under orthochronous Lorentz rotations which include the spatial
reflections.

Time Reversal. Anti-linear time (motion) reversal also has a representation on

spinors
ψ 0 (−t, ~x) = γT̄ ψ(t, ~x). (5.43)
The anti-linear nature of T̄ implies that a solution of the Dirac equation should be
mapped to a solution of the complex conjugated Dirac equation. In the Weyl
representation this is achieved by the matrix

γT̄ =ı̊γ 0 γ 2 . (5.44)

The gamma-matrices satisfy the following identity with the time reversal matrix

Λµ ν γT̄ (γ ν )∗ γT̄−1 = γ µ . (5.45)

Charge Conjugation. The Dirac field is charged, it therefore makes sense to

define charge conjugation. We will use it later to investigate the statistics
associated to spinor fields.
Linear charge conjugation maps a field to its conjugate field14

ψ 0 (x) = γC ψ †T (5.46)

such that ψ 0 solves the same wave equation as ψ. Let us substitute

†T
(γ µ ∂µ − m)ψ 0 = (γ µ ∂µ − m)γC ψ †T = (γ µ )∗ ∂µ − m γC∗ ψ .


(5.47)

This vanishes if
γC (γ µ )∗ γC−1 = γ µ . (5.48)
In the Weyl representation only γ 2 is imaginary, and the condition is solved by the
matrix
γC =ı̊γ 0 γ 1 γ 3 . (5.49)
13
In fact, it may also square to a rotation of 2π, i.e. the element (−1)F , which is achieved by
the replacement by γP →ı̊γP .
14
The composition of adjoint and transpose operations is almost the same as complex
conjugation. There is however a slight difference which becomes relevant only later.

5.8
CPT-Transformation In QFT a discrete transformation of fundamental
importance is the combination of charge conjugation, parity and time reversal,
called CPT. Effectively, it flips the sign of all coordinates15 and performs a
complex conjugation.
A spinor transforms according to
ψ 0 (x) = γT̄ γP γC ψ †T (−x). (5.50)
We find that the combination of matrices is just the additional gamma-matrix γ 5
γT̄ γP γC =ı̊γ 0 γ 2 γ 0 γ 0 γ 1 γ 3 =ı̊γ 0 γ 1 γ 2 γ 3 = γ 5 . (5.51)
This anti-commutes with all gamma-matrices
γ 5 γ µ γ 5 = −γ µ . (5.52)
The sign is compensated by flipping the sign of all vectors.
The CPT-theorem states that all reasonable relativistic QFT’s must be invariant
under the CPT-transformation. They need not be invariant under any of the
individual transformations.

Hermitian Conjugation. The Dirac spinor ψ is complex. To construct real

quantities for use in the Lagrangian or the Hamiltonian one typically uses
hermitian conjugation. However, the various gamma-matrices transform differently
under this operation.
The transformation can be uniformised by conjugation with some matrix γ† such
that16
γ† (γ µ )† γ†−1 = −γ µ . (5.53)
In most relevant representations, in particular in the chiral one, one finds
γ† = −ı̊γ 0 . (5.54)

Therefore, one should modify hermitian conjugation for a spinor ψ and likewise for
a spinor matrix X as
ψ̄ = ψ † γ†−1 , X̄ = γ† X † γ†−1 . (5.55)
The gamma-matrices are anti-hermitian under hermitian conjugation, γ̄ µ = −γ µ ,
with respect to the scalar product defined by γ† .

5.4 Spin Statistics

So far we have only considered the Dirac equation. For quantisation, conserved
charges and later for adding interactions we should construct a Lagrangian.
15
This is an orientation-preserving transformation which belongs to Spin(3, 1), but not to
Spin+ (3, 1).
16
Hermitian conjugation can be defined with a plus or minus on the r.h.s. leading to two
distinct operations. Consistent physics requires the minus sign.

5.9
Lagrangian. It is straight-forward to guess
L = ψ̄(γ µ ∂µ − m)ψ. (5.56)
The variation with respect to ψ † obviously yields the Dirac equation. Variation
with respect to ψ gives the hermitian conjugate equation
−∂µ ψ̄γ µ − mψ̄ = (γ µ ∂µ ψ − mψ)† γ†−1 = 0. (5.57)
In fact, the Lagrangian is almost real
L† = ψ † ((γ µ )† ∂µ† − m)γ†−1 ψ = ψ̄(−γ µ ∂µ† − m)ψ
= L − ∂µ (ψ̄γ µ ψ). (5.58)
There is merely an imaginary topological term left,17 the action is manifestly real.

Hamiltonian Formulation. To go to the Hamiltonian framework we compute

the conjugate momenta π = ∂L/∂ ψ̇ † = 0 and π † = ∂L/∂ ψ̇ =ı̊ψ † . It turns out that
the conjugate momenta are proportional to the fields without (time) derivatives:18
• Dirac equation is a first-order differential equation.
• The dynamical data on a time slice consists of the fields ψ and ψ̄ alone.
• There is no need for independent momenta keeping track of time derivatives;
the latter are determined by the equations of motion.
• Phase space equals position space.
• ψ and ψ † are canonically conjugate fields.
We shall not follow the canonical framework towards quantisation because it has
some complications that we cannot yet understand. Instead, we can compute the
stress-energy tensor of which the Hamiltonian is a component
T µν = −ψ̄γ µ ∂ ν ψ + η µν L. (5.59)
Unfortunately this tensor is not symmetric as it should be due to Lorentz
invariance. Gladly, the anti-symmetric part can be written as (making use of the
equations of motion)19
T [µν] = ∂ρ (ψ̄γ [ρ γ µ γ ν] ψ) = ∂ρ K ρµν . (5.60)
The contribution from K is a boundary term for the integral defining the total
momentum integral P µ . We can thus subtract ∂ρ K ρµν from T µν to make the latter
symmetric.
The Hamiltonian for the Dirac equation now reads
Z Z
H = d~x T = d~x 3 ψ̄ −~γ ·∂~ + m ψ.
3 00


(5.61)
17
The topological term can be removed from the Lagrangian to obtain a manifestly real
0
L = L − 21 ∂µ (ψ̄γ µ ψ).
18
Here, π † is not the complex conjugate of π because L is not real. For the real L0 we get
instead π = 2ı̊ ψ † and π † = − 2ı̊ ψ in which case the definition of the canonical structure is even less
evident. The point is that the Lagrangian is already in the first-order form of the Hamiltonian
with ψ † and ψ being canonically conjugate.
19
Here, the square brackets within indices imply total anti-symmetrisation of all the enclosed
indices.

5.10
Charge Conjugation. A naive treatment and quantisation of the above
framework of the Dirac equation leads to several undesirable features.
For example, there is an issue for charge conjugation. For every solution ψ of the
Dirac equation, there is a charge conjugate solution ψC = γC ψ †T . Let us compute
its energy20
Z
H[ψC ] = d~x 3 ψ T γC† γ†−1 −~γ ·∂~ + m γC ψ †T

Z
= d~x 3 ψ T γC† γ†−1 γC −~γ ∗ ·∂~ + m ψ †T

Z
= d~x 3 ψ T γ†−1 −~γ ∗ ·∂~ + m ψ †T

Z
= d~x 3 ψ T +~γ T ·∂~ + m γ†−1 ψ †T

Z
∗ T
= − d~x 3 ψ † γ†−1 +~γ ·~∂ + m ψ

Z
= − d~x 3 ψ̄ −~γ ·∂~ + m ψ = −H[ψ].


(5.62)

We notice that the charge conjugate solution has the opposite energy

H[ψC ] = −H[ψ]. (5.63)

This implies that there must exist solutions with positive and negative energy.
We compare this result to the complex scalar field where charge conjugation is
defined by φC = φ∗ . There we obtain a similar relationship for the energy,

H[φC ] = +H[φ], (5.64)

but the sign is the same for both solutions. This is consistent with the Hamiltonian
being a positive definite functional and with the physical concept of positivity of
the energy. Conversely, the naive field ψ does not have positive definite energy.
Related issues arise for the naive evaluation of propagators and for causality.
Almost all of the above steps are elementary and have to be accepted. Only the
step marked by ∗ can be altered: transposition. We have used

ψ T Xψ †T = ψ a Xa b ψb† = ψb† Xa b ψ a = ψ † X T ψ. (5.65)

Instead of ψ a ψb† = ψb† ψ a we could use a different rule21

ψ a ψb† = −ψb† ψ a . (5.66)

20
We denote by ∂ T a derivative operator which acts towards the left. The ordinary derivative
operator is obtained by partial integration and in the absence of boundary contributions we have
∂ T ' −∂.
21
In view of quantum mechanics, changing the commutation relationship is not as severe as it
may seem: eventually, quantisation will make ψ’s become operators which do not commute either.

5.11
This change inserts a minus sign at ∗ and the energy of a solution and its charge
conjugate become equal
H[ψC ] = +H[ψ]. (5.67)
This modification of the algebra of the fields ψ actually solves all the other issues
of the spinor field as well.

Spin-Statistics Theorem. The spin-statistics theorem states that consistent

quantisation of fields with half-integer spin requires the use of anti-commutation
relations22
{ψ, ψ † } ∼ ~. (5.68)
Such fields are called fermionic, they obey the Fermi–Dirac statistics.
Multi-particle wave functions will be totally anti-symmetric.
Conversely, fields with integer spin require commutation relations
[φ, φ† ] ∼ ~. (5.69)
These fields are called bosonic, they obey the Bose–Einstein statistics.
Multi-particle wave functions will be totally symmetric.

5.5 Graßmann Numbers

Quantisation can be viewed as a deformation of classical physics. Therefore, the
anti-commutation relations of the quantum theory {ψ, ψ † } ∼ ~ should be reflected
by anti-commuting fields {ψ, ψ † } = 0 in the classical theory. More generally,
ψ a ψ b = −ψ b ψ a , ψ a ψb† = −ψb† ψ a , ψa† ψb† = −ψb† ψa† . (5.70)
Besides these additional signs, the fields ψ will commute with numbers and scalar
fields.
We therefore cannot use ordinary commuting numbers to represent the field ψ in
the classical Lagrangian, we need something else.23

Description. The required extension of the concept of numbers is called

Graßmann numbers:
• Graßmann numbers form a non-commutative ring.
• Graßmann numbers are Z2 -graded, they can be even or odd: |a| = 0, 1.24
22
{A, B} := AB + BA denotes the anti-commutator of two operators A and B.
23
Put differently, classical physics arises as a limit of quantum physics. The precise definition
of the limit is important. For the Dirac equation there is a choice. Formulation of the fields in
terms of ordinary commuting numbers leads to an ill-defined classical limit. Formulation of the
fields in terms of anti-commuting numbers, as described below, leads to a well-defined classical
limit, albeit with an extended notion of numbers. The new class of numbers allow to work with
the Lagrangian and canonical framework with minor alterations.
24
Linear combinations of even and odd numbers could be defined, but usually they do not
appear.

5.12
 • Sums and products respect the even/odd grading

|a + b| = |a| = |b|, |ab| = |a| + |b|. (5.71)

• The product is commutative unless both factors are odd in which case it is
anti-commutative:
ab = (−1)|a||b| ba. (5.72)
• Ordinary numbers are among the even Graßmann numbers.
• The field ψ takes values in odd Graßmann numbers.
• Real and complex Graßmann numbers can be defined. Graßmann numbers
then form an algebra over the respective field.
A basis an of odd Graßmann numbers can be constructed out of a Clifford algebra
{γ j , γ k } = 2δ jk
1
an = √ γ 2n +ı̊γ 2n+1 .


(5.73)
2
In other words, Graßmann numbers can be represented in terms of (exponentially
large) matrices. One should view the basis an to be sufficiently large or infinite.25

Calculus. One can do calculus with Graßmann numbers much like ordinary
numbers, but note:
• odd numbers square to zero: (an )2 = 12 {an , an } = 0.
• the square root of zero is ill-defined.
• odd numbers have no inverse.
• some even numbers (e.g. products of two odd numbers) have no inverse.
A derivative for odd numbers can be defined as usual
∂ n n
a = δm . (5.74)
∂am
Note that derivatives are also odd objects

{∂/∂am , ∂/∂an } = 0. (5.75)

The above derivative relation can be written as an anti-commutator

{∂/∂am , an } = δm
n
. (5.76)

We can also define the derivatives as elements of the same Clifford algebra
∂ 1 2n 2n−1


= √ γ −ı̊γ . (5.77)
∂an 2
This leads to the same anti-commutation relations as above.
25
There is no distinguished element such as i which extends the real numbers to complex
numbers. We therefore do not have universal means to assign a value to a Graßmann variable.
We will mainly use Graßmann variables to describe classical (fermionic) fields without assigning
values.

5.13
Complex Conjugation. A complex Graßmann number a can be written as a
combination of the real Graßmann numbers ar , ai as

a = ar +ı̊ai . (5.78)

Spinor fields are typically complex and we often need to complex conjugate them.
Confusingly, there are two equivalent definitions of complex conjugation for
Graßmann numbers.
One is reminiscent of complex conjugation

a∗ = ar −ı̊ai . (5.79)

It obviously satisfies
(ab)∗ = a∗ b∗ . (5.80)

The other conjugation is reminiscent of hermitian conjugation. It satisfies the rule

that the order of factors within a product is flipped

(ab)† = b† a† . (5.81)

For ordinary numbers it would be the same as complex conjugation, but this
cannot hold for odd Graßmann numbers which do not commute. The two
definitions are in fact related as follows
(
a∗ if a is even,
a† = (5.82)
−ı̊a∗ if a is odd.

Quantum mechanics and QFT frequently use the adjoint operation, hence it is
convenient to stick to hermitian conjugation for Graßmann numbers as well.
One should pay close attention in defining real and odd Graßmann numbers: An
odd number which satisfies a† = a is not real. In particular, the even product of
two such numbers is imaginary

(ab)† = b† a† = ba = −ab. (5.83)

Instead, real odd numbers are defined by the relationship

a† = −ı̊a or a∗ = a. (5.84)

5.6 Quantisation
Poisson Brackets. We have also seen that ψ and ψ † are canonically conjugate
fields, there is no need to introduce additional conjugate momenta. The Poisson
bracket for the spinor field should read26
Z  
3 δF δG δF δG
{F, G} =ı̊ d~x + . (5.85)
δψ a (~x) δψa† (~x) δψa† (~x) δψ a (~x)
26
The correct normalisation can be derived from Ḟ = −{H, F }.

5.14
This expression can also be written as27

ψ (~x), ψb† (~y ) = ψb† (~y ), ψ a (~x) =ı̊δba δ 3 (~x − ~y ).

 a 
(5.86)

Anti-Commutators. For quantisation, these Poisson brackets are replaced by

an anti-commutator28

ψ (~x), ψb† (~y ) = δba δ 3 (~x − ~y ).

 a
(5.87)

By Fourier transformation to momentum space29 30

d~p 3
Z
ı̊p·x −ı̊p·x †


ψ(x) = e uα (~
p)b α (~
p) + e vα (~
p)a α (~
p) (5.88)
(2π)3 2e(~p)

we obtain anti-commutation relations for the Fourier modes31

uγ (~p)bγ (~p), ūδ,b (~q)b†δ (~q) = (ı̊p·γ + m)a b 2e(~p)(2π)3 δ 3 (~p − ~q),
 a

v̄γ,b (~p)aγ (~p), vδa (~q)a†δ (~q) = (ı̊p·γ − m)a b 2e(~p)(2π)3 δ 3 (~p − ~q).

(5.89)

It is convenient to split these relations into contributions from quantum operators

and contributions from spinor solutions. The spinors u and v form a complete set
of basis vectors for the space of spinors. Therefore, they obey certain completeness
relations. In a typical normalisation, these read

uα (~p)ūα (~p) =ı̊p·γ + m,

vα (~p)v̄α (~p) =ı̊p·γ − m. (5.90)

In the Weyl representation these relations are reproduced by the following choice
for the spinors u, v 32
   
(−p·σ)1/2 ξα (−p·σ)1/2 ξα
uα (~p) = , vα (~p) = , (5.91)
ı̊(p·σ̄)1/2 ξα −ı̊(p·σ̄)1/2 ξα

where the ξα form an orthonormal basis of 2-spinors.

With the above normalisation and completeness relations, we obtain simple
anti-commutation relations for the creation and annihilation operators that mimic
their scalar field counterparts

aα (~p), a†β (~q) = bα (~p), b†β (~q) = δαβ 2e(~p)(2π)3 δ 3 (~p − ~q).
 
(5.92)
27
Although the Poisson brackets are anti-symmetric in most cases, they are symmetric for two
Graßmann odd elements.
28
The sign can be determined from the relation [H, ψ a (~x)] = −ı̊ψ̇ a (~x).
29
The field ψ(x) is now extended to all times by means of the equations of motion. In other
words, ψ(x) satisfies the equations of motion by construction.
30
A sum over repeated indices γ, δ, . . . = ± is implicit.
31
This form is determined by the fact that the field ψ obeys the Dirac equation.
32
The square roots of matrices are defined such that (−p·σ)1/2 (−p·σ)1/2 = −p·σ and
(−p·σ)1/2 (p·σ̄)1/2 = m, etc.. Note that this definition has no branch cuts on the mass shell
because the eigenvalues e ± k~ pk of −p·σ and p·σ̄ are positive definite.

5.15
Dirac Sea. The Pauli exclusion principle for fermions states that each state can
be occupied only once. It follows from the (implicit) anti-commutators

aα (~p), a†β (~q) = b†α (~p), b†β (~q) = 0

 † 
(5.93)

that
2
2
a†α (~p) = b†α (~p) = 0. (5.94)

Dirac used the exclusion principle to make useful proposals concerning

negative-energy states in relativistic quantum mechanics and the prediction of
anti-particles.
The Dirac equation has positive and negative solutions. Furthermore, the solutions
carry an (electrical) charge. The positive-energy states are regular, and we do not
need to discuss them. Dirac proposed that all negative-energy states are already
occupied in the vacuum and cannot be excited further. This picture is called the
Dirac sea, and it explained how to avoid negative-energy solutions.

a†

a-particles empty
Dirac sea (5.95)
c-particles filled

b = c†

Continuing this line of thought, there is now the option to remove an excitation
from one of the occupied states. This hole state would not only have positive
energy, but also carry charges exactly opposite to the ones of the regular
positive-energy solutions. In this way he predicted the existence of positrons as the
anti-particles of electrons. The prediction was soon thereafter confirmed in
experiment.
Our view of QFT today is analogous, but also different, so let us compare:
• Positive-energy solutions of ψ are associated to a† .
• Negative-energy solutions of ψ are associated to b.
• We may define an operator c such that c† := b.
• The vacuum is annihilated by c† . All c-states are occupied.
• A hole in the Dirac sea c = b† creates an anti-particle.
Applying the QFT framework to the Dirac equations works as predicted, but:
• There is no need for a Dirac sea.
• Negative-energy solutions are commonly defined as annihilation operators, not
as creation operators with a sea of occupied states.
• Dirac’s argument relies on the exclusion principle, it works for fermions only.
QFT deals with bosons in the very same fashion.

5.16
 • The Dirac equation has real solutions (see later) just as well as the
Klein–Gordon equation has complex solutions. The existence of anti-particles is
unrelated to spin and the Dirac equation. It is a consequence of the CPT
theorem.

Correlators and Propagators. We now have all we need to compute

correlators of the free quantum fields. There are two non-vanishing correlators of
two fields

∆D a a
+ b (x − y) = +ı̊h0|ψ (x)ψ̄b (y)|0i,
∆D a a
− b (x − y) = −ı̊h0|ψ̄b (y)ψ (x)|0i. (5.96)

Evaluation of the correlators yields

d~p 3 eı̊p·x
Z
Da
∆+ b (x) =ı̊ (ı̊p·γ + m)a b ,
(2π)3 2e(~p)
d~p 3 eı̊p·x
Z
Da
∆− b (−x) =ı̊ 3
(−ı̊p·γ + m)a b . (5.97)
(2π) 2e(~p)

They can be expressed in terms of the correlator ∆+ of two scalar fields with an
additional operator acting on spinors

∆D a a
+ b (x) = (γ·∂ + m) b ∆+ (x),
∆D a a
− b (−x) = (−γ·∂ + m) b ∆+ (x). (5.98)

When acting with the Dirac equation on the correlator, it combines with the
operator to give the Klein–Gordon equation acting on ∆+ , e.g.

a
γ µ ∂µ − m b ∆D b
+ c (x)
= (γ·∂ − m)a b (γ·∂ + m)b c ∆+ (x)
= δca ∂ 2 − m2 ∆+ (x) = 0.


(5.99)

Likewise, the unequal-time anti-commutator

ı̊ ψ a (x), ψ̄b (y) = ∆Da b (y − x)


(5.100)

can be written in terms of the one for the scalar field

∆D (x) = (γ·∂ + m)∆(x). (5.101)

As such it satisfies the Dirac equation and vanishes for space-like separations33
 a
ψ (x), ψ̄b (y) = 0 for (x − y)2 > 0.

(5.102)
33
The fact that the anti-commutator rather than the commutator vanishes is not in
contradiction with causality. Typically we can observe only fermion bilinears which are bosonic
and which do commute.

5.17
For the Dirac equation with a source, the same methods we introduced earlier for
the scalar field apply. The propagator is a spinor matrix and it is defined via the
equations

(−γ·∂ + m)a b GDb c (x) = δca δ 4 (x),

(−γ·∂ + m)b c GDa b (x) = δca δ 4 (x), (5.103)

supplemented by suitable boundary conditions. By the same reasons as above, we

can express the Dirac propagator through the scalar propagator

GDa b (x) = (γ·∂ + m)a b G(x). (5.104)

Obviously, one has the same relations as before, e.g. for the retarded propagator

GD D
R (x) = θ(t)∆ (x). (5.105)

There are some other useful relationships between correlators and propagators in
momentum space which are worth emphasising because they hold generally.
First of all, by construction the propagator is the inverse of the kinetic term in the
action34
ı̊γ·p + m
GD (p) = (−ı̊γ·p + m)−1 = 2 . (5.106)
p + m2
The corresponding correlators and unequal-time commutators take the form

∆D 2 2 0
± (p) = 2πı̊δ(p + m )θ(±p )(ı̊p·γ + m),
∆D (p) = 2πı̊δ(p2 + m2 ) sign(p0 )(ı̊p·γ + m). (5.107)

These reflect precisely the residues times a delta-function localised at the position
of the pole along with some restriction to positive or negative energies.
The construction of the propagator and its relationship to correlators and
commutators can be used as a shortcut in deriving the latter. Large parts of the
canonical quantisation procedure can thus be avoided in practice.

5.7 Complex and Real Fields

Complex Field. The Dirac spinor is complex and the Lagrangian has an
obvious U(1) global symmetry

ψ 0 (x) = eı̊α ψ, ψ̄ 0 (x) = e−ı̊α ψ̄. (5.108)

The symmetry has a corresponding conserved Noether current35

δψ a δL
Jµ = =ı̊ψ̄γ µ ψ. (5.109)
δα δ∂µ ψ a
34
The poles should be shifted away from the real axis to accommodate for the desired
boundary conditions.
35
Note the order of terms.

5.18
The time component of the current was used earlier to define a positive definite
probability density, −J 0 = ψ † ψ. However, if one follows the spin-statistics theorem
and let ψ be Graßmann odd, the density is not positive. In particular, it changes
sign for the charge conjugate solution

JCµ =ı̊ψ̄C γ µ ψC = −ı̊ψ̄γ µ ψ = −J µ . (5.110)

Nevertheless, the current is conserved and it defines a conserved Noether charge

Z Z
Q = d~x J = − d~x 3 ψ † ψ.
3 0
(5.111)

It leads to the usual charge assignments for a complex field

   
Q, ψ(x) = +ψ(x), Q, ψ̄(x) = −ψ̄(x). (5.112)

Reality Condition. The Dirac field has four independent particle modes, a†α (~p)
and b†α (~p), for each three-momentum. From the classification of Poincaré UIR’s we
know that the irreducible representation for spin j = 1/2 has only two spin
orientations for each three-momentum.
This discrepancy is associated to the existence of charge conjugate solutions. We
can remove the additional solutions by imposing a reality condition on ψ, namely36

ψC = ψ. (5.113)

A spinor which satisfies this condition is called a Majorana spinor. There exist
representations of the Clifford algebra where all γ µ are purely real. In this basis
the Dirac equation is real, and it makes sense to restrict ψ to real (Graßmann odd)
numbers.
For the momentum modes we can use the identity (which also involves a change of
basis between α and α0 )
uα (~p) = γC vα∗ 0 (~p), (5.114)
to show that the identification ψC = ψ implies

aα (~p) = bα0 (~p). (5.115)

It reduces the modes of the Dirac field by a factor of two.

2-Spinors. Let us consider a real spinor ψ = (ψL , ψR ) in the Weyl

representation. The reality condition implies
1
ψL = −ı̊εψR†T = √ χ. (5.116)
2
36
One could also use any other complex phase eı̊α between ψC and ψ.

5.19
This allows to write the Lagrangian in terms of the 2-spinor field χ as37
L = −χ† ı̊σ̄·∂χ + mχT εχ − mχ† εχ†T . (5.117)
The Lagrangian for the Dirac field can be written as two identical copies of this.38

Parity. Note that parity interchanges ψL and ψR . The reality condition relates
the two, hence
χ0 (t, −~x) = εχ†T (t, ~x). (5.118)
As this transformation also sends the field χ to its complex conjugate χ†T , it is
usually viewed as the combination CP of charge conjugation C and parity P
CP χ(t, ~x)(CP )−1 = εχ†T (t, −~x). (5.119)
In that sense there cannot be individual C and P transformations and only CP
can be a symmetry.
An alternative point of view is that C was used to define the reality condition.
Hence C is preserved by construction, and the parity operation P is well-defined
on its own.
Technically, both points of view have the same content: They merely use the same
words to refer to different operators. They are related by identifying C 0 = 1 and
P 0 = CP , where the primed operations refer to the latter approach.

5.8 Massless Field and Chiral Symmetry

So far we have assumed a non-zero mass m. Let us now consider the massless case
which has some special features. We will assume a real (Majorana) field.
First, let us compare to the irreps of the Poincaré group: The massless real spinor
field gives rise to two particle states a†α (~p) for each momentum. Conversely, a
massless irrep with fixed helicity has merely one state for each momentum. The
two particles correspond to irreps with helicity h = ± 1/2 . In fact, helicity states
must always come in pairs in QFT due to the CPT theorem. One cannot construct
a real Lagrangian which describes just one helicity.
Interestingly, the splitting of representations leads to an enhancement of
symmetry. The massive real spinor field has no symmetry beyond the Poincaré
transformations. For m = 0, the Lagrangian in terms of 2-spinors simplifies to
L = −χ† ı̊σ̄·∂χ. (5.120)
Quite obviously, this Lagrangian has a global U(1) symmetry39
χ0 = eı̊α χ. (5.121)
37
The two mass terms are anti-symmetric in χ, which requires the classical field χ to be an odd
Graßmann number.
38
The U(1) global symmetry of the Dirac equation is recovered as a SO(2) rotation symmetry
of the two fields χ.
39
The mass terms χT εχ and χ† εχ†T evidently break this symmetry.

5.20
It is called chiral symmetry. The associated Noether current reads

J µ = χ† σ̄ µ χ. (5.122)

At the level of 4-spinors, chiral symmetry is represented by the transformation

ψ 0 = exp(−ı̊αγ 5 )ψ, ψ̄ 0 = ψ̄ exp(−ı̊αγ 5 ). (5.123)

These equal transformation factors cancel when they are separated by a single
gamma-matrix as in the kinetic term

exp(−ı̊αγ 5 )γ µ = γ µ exp(ı̊αγ 5 ). (5.124)

The massless Lagrangian is therefore invariant under chiral transformations.40

Here, the conserved current is the so-called axial vector current

J µ =ı̊ψ̄γ 5 γ µ ψ. (5.125)

40
In the presence of masses the chiral transformation converts between the two allowable mass
terms ψ̄ψ and ı̊ψ̄γ 5 ψ.

5.21
Quantum Field Theory I Chapter 6
ETH Zurich, 2019 HS Prof. N. Beisert
11. 11. 2019

6 Free Vector Field

Next we want to find a formulation for vector fields. This includes the important
case of the electromagnetic field with its photon excitations as massless relativistic
particles of helicity ±1. This field will be the foundation for a QFT treatment of
electrodynamics called quantum electrodynamics (QED). Here we will encounter a
new type of symmetry which will turn out to be extremely powerful but at the
price of new complications.

6.1 Classical Electrodynamics

We start by recalling electrodynamics which is the first classical field theory most
of us have encountered in theoretical physics.

Maxwell Equations. The electromagnetic field consists of the electric field

~ ~x) and the magnetic field B(t,
E(t, ~ ~x). These fields satisfy the four Maxwell
equations (with ε0 = µ0 = c = 1)

0 = div B ~B
~ := ∂· ~ = ∂k Bk ,
~ +B
0 = rot E ~˙ := ∂×
~ E~ +B ~˙ = εijk ∂j Ek + Ḃi ,
ρ = div E ~E
~ = ∂· ~ = ∂k E k ,
~˙ = ∂×
~ −E
~ = rot B ~ B ~˙ = εijk ∂j Bk − Ėi .
~ −E (6.1)

The fields ρ and ~ are the electrical charge and current densities.
The solutions to the Maxwell equations without sources are waves propagating
with the speed of light. The Maxwell equations were the first relativistic wave
equations that were found. Eventually their consideration led to the discovery of
special relativity.

Relativistic Formulation. Lorentz invariance of the Maxwell equations is not

evident in their usual form. Let us transform them to a relativistic form.
The first step consists in converting Bi to an anti-symmetric tensor of rank 2
 
0 −Bz +By
Bi = − 21 εijk Fjk , F = +Bz 0 −Bx  . (6.2)
−By +Bx 0

6.1
Then the Maxwell equations read

0 = −εijk ∂k Fij , ρ = ∂k Ek ,
0 = εijk (2∂j Ek − Ḟjk ), ji = ∂j Fji − Ėi . (6.3)

These equations are the 1 + 3 components of two 4-vectors which can be seen by
setting
Ek = F0k = −Fk0 , J µ = (ρ, ~). (6.4)
Now the Maxwell equations simply read

εµνρσ ∂ν Fρσ = 0, ∂ν F νµ = J µ . (6.5)

Electromagnetic Potential. For QFT purposes we need to write a Lagrangian

from which the Maxwell equations follow. This is however not possible using Fµν
as the fundamental degrees of freedom. A Lagrangian can be constructed by the
help of the electromagnetic vector potential Aµ . To some extent the potential is a
convenient technical tool which is also needed to couple the field to charged
matter.1
The first (homogeneous) equation is an integrability condition for the field Fµν . It
implies that it can be integrated consistently to an electromagnetic potential Aµ

Fµν = ∂µ Aν − ∂ν Aµ . (6.6)

With this parametrisation of F the homogeneous equation is automatically

satisfied.
The electromagnetic potential is not uniquely defined by the electromagnetic fields
F . For any solution A, we can add the derivative of an arbitrary scalar field α(x)

A0µ (x) = Aµ (x) + ∂µ α(x). (6.7)

The extra term cancels out when anti-symmetrising the two indices of ∂µ Aν and
hence
0
Fµν (x) = Fµν (x). (6.8)
This freedom in defining Aµ is called a gauge symmetry or gauge redundancy. It is
called a local symmetry because the transformation can be chosen independently
for every point of spacetime as opposed to globally for all of spacetime. Gauge
symmetry will turn out to be very important in quantising the vector field.
1
This fact can be observed in the Aharonov–Bohm effect, where the phase of a charged
quantum particle feels the presence of a non-trivial electromagnetic potential A, although it is
confined to a region of spacetime where the field strength F vanishes. Nevertheless, the field
strength F is the primary physical quantity whereas A is secondary concept because the physical
content of A can be derived from F unambiguously. Note that this requires integration and
therefore the vector potential contains non-local information on the physics. Even though
non-local interactions are undesired in classical physics, they do influence phases in quantum
mechanics.

6.2
Lagrangian. A Lagrangian for the electromagnetic fields can now be formulated
in terms of the potential Aµ
~ 2 − 1 B[A]
L = − 14 F µν [A]Fµν [A] = 12 E[A] ~ 2. (6.9)
2

Here and in the following, Fµν [A] is not considered a fundamental field, but merely
represents the combination

Fµν [A] = ∂µ Aν − ∂ν Aµ . (6.10)

The equation of motion yields the second (inhomogeneous) Maxwell equation

(here, with a trivial source term)

∂ν F νµ = 0. (6.11)

The first (homogeneous) Maxwell equation is already implied by the definition of

F in terms of A.
Due to Poincaré symmetry we can also derive a conserved stress-energy tensor
T µν . It takes the form2

T µν = F µρ F ν ρ − 41 η µν F ρσ Fρσ . (6.12)

6.2 Gauge Fixing

Hamiltonian Framework. Towards quantisation we should proceed to the
Hamiltonian framework. The canonical momentum Π conjugate to the field A
reads
∂L
Πµ = = F0µ . (6.13)
∂ Ȧµ
Here, a complication arises because the component Π0 is strictly zero and the field
A0 has no conjugate momentum. The non-zero spatial components form the
electrical field F0k = Ek . Moreover, the equations of motion imply ∂k Πk = 0 which
is an equation without time derivative. This so-called constraint has to be
implemented on the dynamical data provided in each time slice.3 It is related to
Gauß’ law which determines the electrical charge density (zero) from a field
configuration.
Here, the missing of the momentum Π0 and the constraint for Πk are related to
gauge redundancy of A. Although Aµ has four components, two of them can be
chosen arbitrarily using gauge symmetry. Effectively Aµ has only two physically
relevant components, which is matched by only two independent conjugate
momenta. Under-determined and constrained variables require extra work in the
Hamiltonian framework. Here we can avoid the complication by fixing a gauge.
2
A naive derivation from the Lagrangian yields T µν = F µρ ∂ ν Aρ + η µν L which is neither
symmetric nor gauge-invariant. Equivalently, we can use a gauge-invariant form of the
translation δAρ = −δaν Fνρ which differs from the naive translation δAρ = −δaν ∂ν Aρ by a gauge
transformation with the field δα = δaν Aν .
3
Typically it suffices to satisfy constraints on the initial time slice.

6.3
Coulomb Gauge. A simple ansatz to resolve the problem of Π0 = 0 is to
demand that
A0 = 0. (6.14)
This can always be achieved by a suitable gauge transformation. It eliminates the
need for the problematic canonical Poisson bracket between A0 and Π0 .
Fixing A0 in the Lagrangian leads to a completely determined dynamical system.
However, since A0 is now missing from the set of dynamical variables, the
corresponding equation of motion is absent. It amounts to the constraint

~A
∂· ~˙ = ∂·
~Π ~ = 0. (6.15)

Thus, the gauge fixed system is more general than electrodynamics. Only if the
initial data satisfies the Gauß law constraint, along with its time derivative
~A
∂~ 2 ∂· ~ = 0, our dynamical system agrees with electrodynamics.
The choice A0 = 0 does not completely eliminate all gauge freedom for Ak , a
time-independent gauge redundancy α(~x) remains. It can be eliminated by
demanding
~A
∂· ~ = 0. (6.16)
This is called the Coulomb gauge-fixing condition.
Now Πk = F0k = Ek = Ȧk and for the Hamiltonian we obtain
Z
H = d~x 3 21 E~2 + B~2 ,


(6.17)

which indeed represents the energy of the electromagnetic field.

With the Coulomb gauge, we can now quantise the electromagnetic field. The
gauge is however not always convenient, since it specialises the time direction and
therefore breaks relativistic invariance. For instance, it leads to instantaneous
contributions to field correlators, which appears odd in a relativistic model.
However, these contributions are gauge artifacts of the gauge potentials. In
physical gauge-invariant observables, eventually such instantaneous or
causality-violating contributions will always cancel.

Lorenz Gauges. A more general class of gauge-fixing conditions are the Lorenz
gauges
∂ µ Aµ = 0. (6.18)
They are particularly convenient in a relativistic theory because they respect
Poincaré symmetry. They do not completely fix the gauge freedom since any gauge
transformation with ∂ 2 α = 0 will preserve the Lorenz gauge condition.4 For
example, one may furthermore demand A0 = 0 to recover the Coulomb gauge.
4
This degree of freedom will later serve as an additional scalar field without physical
relevance. Gladly, the field decouples from all physical processes.

6.4
Gauge-Fixing Term. Instead of choosing a particular gauge, we can alter the
Lagrangian such that it imposes a gauge on its own. To that end, we can add a
gauge-fixing term Lgf = − 12 ξ −1 (∂·A)2 to the Lagrangian which is known as the Rξ
gauge5
L = LED + Lgf ' − 12 ∂ µ Aν ∂µ Aν + 12 (1 − ξ −1 )∂ µ Aµ ∂ ν Aν . (6.19)
The equations of motion now read

∂ 2 Aµ − (1 − ξ −1 )∂µ ∂ ν Aν = 0. (6.20)

The gauge-fixing term spoils gauge invariance and makes time evolution for all
four gauge potentials well-defined. As before, the new system describes more than
electrodynamics. However, the additional gauge-fixing term vanishes quadratically
in the Lorenz gauge, so that in this gauge we recover electrodynamics. In
particular, we have to implement the Lorenz gauge on the initial data by hand
~A
∂·A = −Ȧ0 + ∂· ~ = 0,
2
∂·Ȧ ∼ −~∂ A0 + ∂·
~A ~˙ = 0. (6.21)

The equations of motion propagate the gauge condition to all other time slices.
Therefore, the extended model with constrained phase space yields
electrodynamics.

Generators of Residual Symmetries. The gauge fixing also has a relevant

effect on the canonical structure. Let us determine the latter to see this effect. The
canonical momenta now read67
~ A,
Π0 = ξ −1 Ȧ0 + (1 − ξ −1 )∂· ~ ~˙
~ = A,
Π (6.22)

and we define canonical Poisson brackets

Aµ (t, ~x), Πν (t, ~y ) = ηµν δ 3 (~x − ~y ).


(6.23)

The constraint function ∂·A has non-trivial Poisson brackets with some phase
space function F

{∂·A, F } = ξ{−Π0 + ∂· ~ A,
~ F },
{∂·Ȧ, F } = ξ{−∂~ 2 A0 + ∂·
~ Π,
~ F }. (6.24)
5
Note that the equations of motion imply ∂ 2 ∂ µ Aµ = 0. This condition holds in the Lorenz
gauge, but it does actually not fix it; instead, it also holds for general massless gauge potentials
not satisfying the gauge. To impose the Lorenz gauge one can take a Lagrange multiplier field B
and add the term B ∂·A to the Lagrangian (or alternatively take the limit ξ → 0 with the above
term).
6
All time derivatives can be solved for canonical momenta unless ξ = ∞.
7
The definition of the canonical momenta is ambiguous due to the equivalence of the terms
∂ µ Aν ∂ ν Aµ and ∂ µ Aµ ∂ ν Aν in the Lagrangian by partial integration. This ambiguity, however,
does not make any difference for the Poisson brackets.

6.5
It is desirable to be able to set the left-hand sides to zero, but for a general
function F the right-hand side does not vanish. However, one can convince oneself
that for all components of the field strength tensor, i.e. F = Ek and F = Bk , the
right-hand side is zero. In fact, the functions ∂·A and ∂·Ȧ generate the residual
gauge transformations with ∂ 2 α = 0 given by their initial data α and α̇. For
gauge-invariant observables F , the canonical structure thus becomes consistent
with the gauge-fixing condition, but we will also have to deal with
gauge-dependent quantities at intermediate stages of some calculations, and thus
we cannot impose the Lorenz gauge in a strict sense.

Feynman Gauge. To simplify the subsequent analysis, we shall set ξ = 1. This

so-called Feynman gauge has a simple Lagrangian
L = − 21 ∂ µ Aν ∂µ Aν (6.25)
with simple equations of motion
∂ 2 Aµ = 0. (6.26)
Effectively, it describes four massless scalar fields Aµ with the peculiarity that the
sign of the kinetic term for A0 is wrong. With the canonical momenta Πµ = Ȧµ
the Poisson brackets read
Aµ (~x), Πν (~y ) = ηµν δ 3 (~x − ~y ),

(6.27)
where again the relation for A0 has the opposite sign. Likewise all correlation
functions and propagators equal their scalar counterparts times ηµν .
As such, the model described by the above simple Lagrangian is not
electrodynamics. Only when taking into account the constraint ∂·A = 0 it becomes
electrodynamics. Moreover, the constraint will be crucial in making the QFT
model physically meaningful. Nevertheless, we have to be careful in implementing
the constraint since it is inconsistent with the Poisson brackets.

Light Cone Gauge. The above Rξ gauge does not eliminate all unphysical
degrees of freedom, which introduce some complications later. There are other
useful gauges which avoid these problems, but they trade them in for different
problems. A prominent example is the light cone gauge which eliminates a
light-like component A− = A0 − A3 = 0 of the gauge potential Aµ . The equations
of motion then allow to solve for a non-collinear like-like component of
A+ = A0 + A3 . The remaining two degrees of freedom of Aµ then represent the two
helicity modes of the electromagnetic field. Let us nevertheless continue in the
Feynman gauge.

6.3 Particle States

Next we quantise the model and discuss its particle states. The construction of
Fock space is the same as for a set of four massless scalar fields, but we need to
implement the gauge-fixing constraint.

6.6
Quantisation. We quantise the vector field Aµ (x) in Feynman gauge
analogously to four independent scalar fields where merely one of the kinetic term
has the opposite sign. This leads to the equal-time commutation relations

Aµ (t, ~x), Ȧν (t, ~y ) =ı̊ηµν δ 3 (~x − ~y ).

 
(6.28)

We then solve the equation of motion ∂ 2 Aµ = 0 in momentum space

d~p 3
Z
ı̊p·x −ı̊p·x †


Aµ (x) = e a µ (~
p) + e a µ (~
p) (6.29)
(2π)3 2e(~p)

and translate the above field commutators to commutators for creation and
annihilation operators

aµ (~p), a†ν (~q) = ηµν 2e(~p)(2π)3 δ 3 (~p − ~q).

 
(6.30)

Fock Space. We define the vacuum state |0i to be annihilated by all aµ (~p)

aµ (~p)|0i = 0. (6.31)

As before, multi-particle states are constructed by acting with the creation

operators a†µ (~p) on the vacuum |0i.
There are two problems with this naive Fock space. The first is that there ought to
be only two states (with helicity h = ±1) for each momentum. Here we have
introduced four states. The other problem is that one of these states has a
negative norm: To see this we prepare a wave packet for a†0

d~p 3
Z
|f i = 3
f (~p)a†0 (~p)|0i. (6.32)
(2π) 2e(~p)

The norm of this state is negative definite

d~p 3
Z
f (~p)2 < 0.
 
hf |f i = − (6.33)
(2π)3 2e(~p)

A negative-norm state violates the probabilistic interpretation of QFT, hence it

must be avoided at all means.8

Physical States. The above problems are eventually resolved by implementing

the gauge-fixing constraint ∂·A = 0 which we have not yet considered. This is not
straight-forward:
• The non-trivial commutation relations of ∂·A prevent us from implementing
the constraint ∂·A = 0 at an operatorial level.
8
Alternatively, one might define the vacuum state to be annihilated by a†0 instead of a0 . In
that case, all states have a positive norm. However, the particles created by a0 now have a
negative energy, leading to an equally undesirable situation.

6.7
 • We cannot implement it directly on states: E.g. requiring the vacuum |0i to be
physical means setting p·a† (~p)|0i = 0 which is inconsistent with the
commutation relations.
• The weakest implementation is to demand that the expectation value of ∂·A
vanishes for all physical states. This is the Gupta–Bleuler formalism.
For two physical states |Ψ i, |Φi we thus demand

hΦ|∂·A|Ψ i = 0. (6.34)

This is achieved by demanding

p·a(~p)|Ψ i = 0 (6.35)

for any physical state. An adjoint physical state then obeys hΦ|p·a† (~p) = 0.
• Both conditions together ensure that hΦ|∂·A|Ψ i = 0.
• Moreover, the vacuum is physical by construction.
We conclude that Fock space is too large in agreement with the discussion at the
classical level. The space of physical states |Ψ i is a subspace of Fock space such
that for all p~
p·a(~p)|Ψ i = 0. (6.36)
Nevertheless, we cannot completely abandon the larger Fock space in favour of the
smaller space of physical states. For instance, the action of Aµ (x) cannot be
confined to the physical subspace since it does not commute with the operator
p·a(~p).
Evidently, the negative-norm state |f i discussed above is not physical since

p·a(~p)|f i = −f (~p)e(~p)|0i. (6.37)

The right-hand side vanishes for all p~ only if the function f is identically zero, in
which case the state |f i = 0 is trivial. Therefore, the state is an element of Fock
space, but not of its physical subspace.

Basis of Polarisation Vectors. To investigate the space of physical states

further, we introduce a convenient basis for polarisation vectors µ(α) (~p) of the
vector field aµ (~p) and a†µ (~p) on the light cone p2 = 0.9
We denote the four polarisations α by G for gauge, L for longitudinal and 1, 2 for
the two transverse directions.10
• We first define (G) as a light-like vector in the direction of p. For definiteness,
we set (G) = p.
• We construct another light-like vector (L) which has unit scalar product with
(G) , i.e. (L) ·(G) = 1.
9
The polarisation vectors are similar to the spinors u(~ p) and v(~
p) for the Dirac equation.
10
There is a lot of arbitrariness in defining the polarisation vectors (L) , (1) , (2) for each
momentum p. Moreover, there is no universal Lorentz-invariant choice of basis. Gladly, the
arbitrariness does not affect physics.

6.8
 • We then construct two orthonormal space-like vectors (1,2) which are also
orthogonal to (G) and (L) .
For example, suppose the light-like momentum is given by
pµ = (e, 0, 0, e). (6.38)
Then we can define the following four vectors11
µ(G) (~p) = (e, 0, 0, e),
µ(L) (~p) = (−1/2e, 0, 0, 1/2e),
µ(1) (~p) = (0, 1, 0, 0),
µ(2) (~p) = (0, 0, 1, 0). (6.39)

These four polarisations define a complete basis for the vector space. We can thus
decompose the creation and annihilation operators as follows
a(α) = µ(α) aµ , a†(α) = µ(α) a†µ . (6.40)

Likewise, we can write the non-trivial commutation relations12

a(1) (~p), a†(1) (~q) = a(2) (~p), a†(2) (~q)
   

= a(L) (~p), a†(G) (~q) = a(G) (~p), a†(L) (~q)

   

= 2e(~p) (2π)3 δ 3 (~p − ~q). (6.41)

By construction we know that

p·a(~p) = (G) ·a(~p) = a(G) (~p). (6.42)
hence the physical state condition in this basis reads
a(G) (~p)|Ψ i = 0. (6.43)

The physical state condition together with the commutation relations implies that
a physical state cannot have any longitudinal excitations a†(L) (~p). It must be of the
form13
|Ψ i = a†(G) · · · a†(G) a†(1,2) · · · a†(1,2) |0i. (6.44)
Since negative norm states can originate exclusively from the commutators
[a(L) , a†(G) ] and [a(G) , a†(L) ], and since the a†(L) ’s are absent, the norm of any such
state is positive semi-definite
hΨ |Ψ i ≥ 0. (6.45)
The modes a†(1,2) have a positive norm while a†(G) is null.14
11
The vectors (1) and (2) define a basis with linear polarisation. The complex combinations
(±) ∼ (1) ±ı̊(2) define circular polarisation with definite helicity.
12
The non-trivial overlap between L and G is due to the construction of the basis using two
light-like directions.
13
Note that a(G) commutes with a†(G) and a†(1,2) but not with a†(L) .
14
Full Minkowski space has the indefinite signature (− + ++) while the subspace spanned by
(G,1,2) has positive semi-definite signature (0 + +).

6.9
Null States. Let us discuss a physical state |Ψ i which contains an excitation of
type a†(G) = p·a† , i.e. a state which can be written as15

|Ψ i = p·a† (~p)|Ωi (6.46)

with some other physical state |Ωi. This state has zero norm by the physical state
condition

hΨ |Ψ i = hΩ|p·a(~p) p·a† (~p)|Ωi = hΩ|p·a† (~p) p·a(~p)|Ωi = 0. (6.47)

Null states are not normalisable and therefore have to be interpreted appropriately.
Typically null states are considered irrelevant because QM is a probabilistic
framework. Something that takes place with probability zero does not happen in
practice. Nevertheless, some consistency requirements have to be fulfilled:
By the same argument as above, we can show that a null state

|Ψ i = p·a† (~p)|Ωi (6.48)

actually has vanishing scalar products with any physical state |Φi due to the
physicality condition of the latter

hΦ|Ψ i = hΦ|p·a† (~p)|Ωi = 0. (6.49)

In particular, this implies that the sum |Ψ 0 i of a physical state |Ψ i and some null
state
|Ψ 0 i = |Ψ i + p·a† (~p)|Ωi (6.50)
behaves just like the original physical state |Ψ i in scalar products

hΦ|Ψ 0 i = hΦ|Ψ i + hΦ|p·a† (~p)|Ωi = hΦ|Ψ i. (6.51)

We should thus impose an equivalence relation on the physical Fock space

|Ψ i ' |Ψ 0 i = |Ψ i + p·a† (~p)|Ωi. (6.52)

Any two states which differ by a state which is in the image of some p·a† are
physically equivalent. In other words, physical states of the gauge field are not
described by particular states but by equivalence classes of states.
s
las

physical (6.53)
.c
eq

Fock space
15
For simplicity, we assume the gauge excitation to be a plane wave. As such the norm will
contain a factor of ∞ leading to some potential ambiguities. Nevertheless, all the following
statements will apply if the plane wave is replaced by a normalisable wave packet.

6.10
We may use states which have no contribution of p·a† as reference states of the
equivalence classes16
|Ψ i = a†(1,2) · · · a†(1,2) |0i. (6.54)
These representatives show that we have two states for each momentum p~. It
matches nicely with the massless UIR’s of the Poincaré group with positive and
negative helicity h = ±1. The particle excitations of the electromagnetic field are
the photons.

Gauge Transformations. However, inserting some gauge potentials Aµ (x) into

the scalar product may actually lead to dependence on null states. Let us therefore
compute

hΦ|Aµ (x)|Ψ 0 i = hΦ|Aµ (x)|Ψ i + hΦ|Aµ (x)p·a† (~p)|Ωi

= hΦ|Aµ (x)|Ψ i + hΦ| Aµ (x), p·a† (~p) |Ωi.
 
(6.55)

The commutator evaluates to

Aµ (x), p·a† (~p) = pµ eı̊p·x = −ı̊∂µ eı̊p·x .

 
(6.56)

The expectation value of Aµ (x) thus changes effectively by a derivative term

hΦ|Ωi
Aµ (x) 7→ Aµ (x) − ı̊∂µ eı̊p·x . (6.57)
hΦ|Ψ i

This is just a gauge transformation of the potential Aµ (x). We observe that the
states |Ψ i and |Ψ 0 i lead to two expectation values which differ by a gauge
transformation of the fields within the expectation value. Note that the gauge
transformation does not leave the Lorenz gauges

∂·A(x), p·a† (~p) = −ı̊∂ 2 eı̊p·x = 0.

 
(6.58)

Hence null states induce residual gauge transformation within the Lorenz gauges
as discussed in the classical context.
Now it appears that the choice of representative in an equivalence class has
undesirable impact on certain expectation values. Gladly, this does not apply to
gauge-invariant observables. For instance, the electromagnetic field strength is
unaffected
Fµν (x), p·a† (~p) = ∂µ (pν eı̊p·x ) − ∂ν (pµ eı̊p·x ) = 0.
 
(6.59)
Moreover, the coupling of the gauge potential to a conserved current J µ
Z
J[A] = dx4 J µ (x)Aµ (x) (6.60)

16
Although this appears to be a useful choice at first sight, it is not at all unique. By a change
of basis for the polarisation vectors at any given p~ we can add any amount of a†(G) to a†(1,2) . The
new states are certainly in the same equivalence class, but they are different representatives.

6.11
commutes with p·a†
Z
†
J[A], p·a (~p) = −ı̊ dx4 J µ (x)∂µ eı̊p·x
 

Z
=ı̊ dx4 eı̊p·x ∂µ J µ (x) = 0. (6.61)

The expectation value of any gauge-invariant operator composed from Fµν , J[A] or
similar combinations thus does not depend on the choice of representatives, and it
is consistent to define physical states as equivalence classes.

6.4 Casimir Energy

At this point we can already compute a quantum effect of the electromagnetic
field, the Casimir effect. The Casimir effect is a tiny force between nearby
conductors which exists even in the absence of charges, currents or medium. In
classical electrodynamics no forces are expected in this setup. There are several
alternative explanations for the quantum origin of the force. One is the exchange
of virtual photons between the conductors. An equivalent explanation attributes
the force to a change of vacuum energy of the electromagnetic field induced by the
presence of the plates. The latter one has a quite efficient derivation, and we shall
present it here.

Setup. We place two large planar metal plates at a small distance into the
vacuum (much smaller than their size, but much larger than atomic distances). In
our idealised setup, the plates extend infinitely along the x,y-directions. They are
separated by the distance a in the z-direction. We will not be interested in the
microscopic or quantum details of the metal objects. We simply assume that they
are classical conductors and that they shield the electromagnetic field efficiently.

A (6.62)

a
At the surface of the plates, the electric fields must be orthogonal Ex = Ey = 0
while the magnetic field must be parallel Bz = 0. In order to match these
conditions simultaneously at both plates, the z-component of the wave vector
(momentum) must be quantised

Bz
π
pz ∈ Z. (6.63)
a

6.12
Careful analysis shows that for pz = 0 only one of the two polarisation vectors is
permissible. Conversely, for pz 6= 0 both polarisations are good. To achieve
cancellations in this case, each wave must be synchronised to its reflected wave
where pz → −pz . Hence we should only count the contributions with pz > 0.

Vacuum Energy. Just like the scalar field, the electromagnetic field carries some
vacuum energy.17 The discretisation modifies the vacuum energy E0 , which results
in a force between the plates if the new vacuum energy depends on the distance a.
The sum and integral of all permissible modes between the plates yields the energy
E per area A 18
∞
Z !
A dpx dpy 1 X
1
E= 2
e(px , py , 0) + 2 2
e(px , py , πn/a) . (6.64)
(2π)2 n=1

For convenience we shall exploit the rotation symmetry in the x,y-plane to simplify
the expression to
∞
Z ∞ !
E p dp 1 X p
= 2
p+ p2 + π 2 n2 /a2 . (6.65)
A 0 2π n=1

As discussed earlier, this expression diverges due to UV contributions at large

momenta.

Regularisation. We also emphasised earlier that infinities are largely our own
fault. The idealised setup was somewhat too ideal.
For macroscopic electromagnetic waves, we certainly made the right assumption of
total reflection. But it is also clear that the conducting plates will behave
differently for hard gamma radiation. This is precisely where the problem arises,
so we seem to be on the right track. Electromagnetic waves with wave length much
smaller than atomic distances or energies much larger than atomic energy levels
will pass the conducting plates relatively unperturbed. These modes therefore
should be discarded from the above sum.19
Therefore, we must introduce a UV cutoff for the modes. Define a function f (e)
which is constantly 1 for sufficiently small energies, constantly 0 for sufficiently
17
Earlier, we had argued that vacuum energies are infinite, ambiguous and unobservable. Here,
we will learn how to deal with the infinity. The ambiguity could in principle be resolved by
considering how the electromagnetic field couples to the conductors. The standard expression for
the vacuum energy 12 ~ω of harmonic oscillators from the symmetric ordering prescription turns
out to do the job.
18
A sum over the modes in some box of volume V in d dimensions turns into an integral over
momenta when the volume is very large. In a box, the positive and negative modes are coupled,
so the integral is over positive p only with integration measure V d~p d /π d . In the absence of
boundary contributions, the integration domain extends to positive and negative p which is
compensated by the measure V d~p d /(2π)d .
19
The modes do contribute to the vacuum energy between the plates. Importantly, the
distance between the plates will hardly enter their contribution, and consequently they cannot
contribute to forces.

6.13
large energy and which somehow interpolates between the 1 and 0 for intermediate
energy.
f (e)
1 cutoff
IR (6.66)
UV
0 e
1 1
The cutoff replaces each contribution 2
e by 2
f (e)e
∞
!
Z ∞
EIR p dp 1
X p √
= 2
pf (p) + p2 + π 2 n2 /a2 f ( . . . ) . (6.67)
A 0 2π n=1

Let us keep in mind the remaining contribution in the ultraviolet

Z ∞
p dp ∞
Z
EUV p √

= dn p2 + π 2 n2 /a2 1 − f ( . . . ) . (6.68)
A 0 2π 0
Here, we have converted the sum to an integral due to the absence of quantisation
in the z-direction.20

Summation. The regularised expression EIR is now finite, but it certainly

depends on the cutoff in f (e). We continue its evaluation and write it as a sum of
integrals
∞
EIR 1
X
= 2 F (0) + F (n), (6.69)
A n=1

with Z ∞
p dp p 2 p

F (n) = p + π 2 n2 /a2 f p2 + π 2 n2 /a2 . (6.70)
0 2π
It is convenient to use energy as the integration variable
p
e = p2 + π 2 n2 /a2 , p dp = e de , (6.71)

and write the integral as

Z ∞
1
F (n) = de e2 f (e). (6.72)
2π πn/a

The Euler–MacLaurin summation formula writes the above sum for EIR /A as an
integral Eint /A plus discrete correction terms EC /A
Z ∞
Eint
= dn F (n),
A 0
∞
EC X B2k (2k−1)
=− (−1)k F (0), (6.73)
A k=1
(2k)!
20
The high-energy electromagnetic waves can effectively pass through the conducting plates.

6.14
where we have used that the function F (n) is constantly zero at infinity due to the
cutoff. Here Bn is the n-th Bernoulli number.
Let us analyse the two terms: The first term we can rewrite as
Z ∞ Z ∞ Z ∞ Z ∞
Eint 1 2 2a 0
= dn de e f (e) = 2 de de e2 f (e). (6.74)
A 2π 0 πn/a 4π 0 e0

It depends on the cutoff, but it is manifestly linear in a. In fact, it combines nicely

with the contribution from UV modes above the cutoff that we dropped earlier
Z ∞ Z ∞
E0 EUV + Eint 2 0
= = 2 de de e2 . (6.75)
V Aa 4π 0 e0

As such it represents the vacuum energy of the enclosed volume V = Aa in the

absence of plates. The same vacuum energy density is present outside the plates.21
This term therefore does not contribute to the force because any shift of the plate
would merely transfer some vacuum energy from the inside to the outside leaving
the overall energy invariant. The fact that E0 is formally infinite does not play a
role. We therefore consider only the change in energy EC = E − E0 arising from
the second term of the Euler–MacLaurin summation.
The second term can be evaluated near n = 0 by extending the domain of
integration down to e = 0 and subtracting the extra part
Z ∞ Z πn/a
1 2 1
F (n) = de e f (e) − de e2 f (e)
2π 0 2π 0
2 3
π n
= F (0) − , (6.76)
6a3
where we used that f (e) = 1 for the second term near e = 0. Quite surprisingly,
F (n) is a polynomial with only two terms. All cutoff dependence is in the constant
term F (0) which does not appear in the summation formula.22 The single
correction term contributes the following vacuum energy (B4 = − 1/30 )

EC B4 (3) π2
=− F (0) = − . (6.77)
A 4! 720a3
The presence of the conducting plates decreases the vacuum energy by some
amount proportional to 1/a3 .

Casimir Force. The Casimir force can be expressed as the pressure

F E 0 (a) π2 ~cπ 2
P = = C = = . (6.78)
A A 240a4 240a4
Some properties:
21
The factor of 2/2π is interpreted as follows: 1/2π is the correct measure for integration over
pz and the factor of 2 compensates for the restricted integration region pz ≥ 0.
22
The reason is apparently that the cutoff is in a region of energies where the difference
between a sum and an integral does not matter.

6.15
 • Bringing the plates closer decreases the energy, hence the Casimir force is
attractive.
• It increases with the fourth power of the inverse distance as the plates come
closer.
• It is a quantum effect, and there are hidden factors of ~ and c. Due to the
fourth-power behaviour in 1/a it can nevertheless be detected at reasonable
separations. It becomes relevant at micrometer distance.
• It does not depend on the coupling strength of the electromagnetic field or on
the elementary charge.

6.5 Massive Vector Field

So far we have discussed the massless vector field. Among the UIR’s of the
Poincaré group there is also the massive representation with spin 1. Massive vector
particles exist in nature as the W± and Z0 bosons transmitting the weak nuclear
interactions.23

Lagrangian. We can add a mass term to the vector Lagrangian to obtain the
corresponding quantum field
L = − 21 ∂µ Vν ∂ µ V ν + 21 ∂µ Vν ∂ ν V µ − 21 m2 V µ Vµ . (6.79)
The corresponding equation of motion reads
∂ 2 Vµ − ∂µ ∂ ν Vν − m2 Vµ = 0. (6.80)
By taking the total derivative of this equation, we see that it implies the simpler
equation −m2 ∂ µ Vµ = 0. Substituting this result in the original equation of motion
then yields a system of two equations
∂ 2 Vµ − m2 Vµ = 0, ∂ µ Vµ = 0. (6.81)
The first equation is the Klein–Gordon equation for each component of Vµ , the
second equation removes one of the four potential orientations. The degrees of
freedom agree with the classification of UIR’s.

Correlators. We now want to quantise this system. In the canonical approach

we first derive the conjugate momenta
Πµ = V̇µ − ∂µ V0 . (6.82)
As before, there is no conjugate momentum for the field V0 hinting at the fact that
the equation of motion for V0 is a constraint24
∂k Πk + m2 V0 = 0. (6.83)
23
The implementation of interacting massive vector fields actually needs much more care.
Interacting vectors fields can acquire mass only through the Higgs mechanism.
24
Note that unlike in the massless case it is possible to solve the constraint without integration
by V0 = −m−2 ∂k Πk .

6.16
Constrained systems are somewhat tedious to handle in the Hamiltonian
framework and therefore in canonical quantisation.
Instead, let us take a shortcut. We consider the fields to be operators and cook up
unequal-time commutation relations

∆V
 
µν (x − y) =ı̊ Vµ (x), Vν (y) . (6.84)

Our previous experience has shown that correlators can be composed from
derivatives acting on the correlator of the scalar field. This automatically
implements the Klein–Gordon equation. Here we propose25 26
−2
∆V


µν (x) = ηµν − m ∂ µ ∂ ν ∆(x). (6.85)

The combination of derivatives was constructed such that ∆V satisfies the

polarisation equations
∂ µ ∆V ν V
µν (x) = ∂ ∆µν (x) = 0. (6.86)

Equal-Time Commutators. Next let us see what this proposal implies for the
equal-time commutators. The non-vanishing ones read as follows

V0 (~x), Vk (~y ) =ı̊m−2 ∂k δ 3 (~x − ~y ),

 

V0 (~x), V̇0 (~y ) = −ı̊m−2 ∂k ∂k δ 3 (~x − ~y ),

 

Vk (~x), V̇l (~y ) =ı̊δkl δ 3 (~x − ~y ) −ı̊m−2 ∂k ∂l δ 3 (~x − ~y ),

 

V̇0 (~x), V̇k (~y ) =ı̊∂k δ 3 (~x − ~y ) −ı̊m−2 ∂k ∂l ∂l δ 3 (~x − ~y ).

 
(6.87)

These relations appear somewhat unusual since they mix time and space
components of Vµ .
Let us replace the time derivatives V̇k by the associated conjugate momenta Πk .
For the spatial components we recover the canonical commutator

Vk (~x), Πl (~y ) =ı̊δkl δ 3 (~x − ~y ).

 
(6.88)

The commutators involving V0 and V̇0 can be recovered using the equations of
motion. The latter actually give an explicit solution for the field V0 and its time
derivative V̇0
V0 = −m−2 ∂k Πk , V̇0 = ∂k Vk . (6.89)
In other words, V0 is not an elementary field and its commutation relations follow
from the canonical one above

V̇0 (~x), Πk (~y ) =ı̊∂k δ 3 (~x − ~y ),

 

Vk (~x), V0 (~y ) =ı̊m−2 ∂k δ 3 (~x − ~y ),

 

V0 (~x), V̇0 (~y ) = −ı̊m−2 ∂k ∂k δ 3 (~x − ~y ).

 
(6.90)
25
The correlator ∆+ and the propagator G take an equivalent form in terms of their scalar field
counterparts.
26
The factor of 1/m2 in not as innocent as it may appear. When adding interactions, this term
involving an inverse mass scale actually makes the theory behave badly for large momenta.

6.17
Hamiltonian Framework. We have obtained a reasonable QFT framework for
our massive scalar field. Now we can revisit the Hamiltonian framework. First we
perform a Legendre transformation of the Lagrangian for spatial components of
the fields Vk 27
Z
H = d~x 3 Πk V̇k − L

Z 
= d~x 3 21 Πk Πk + 21 m−2 ∂k Πk ∂l Πl

+ 12 ∂k Vl ∂k Vl − 12 ∂l Vk ∂k Vl + 12 m2 Vk Vk . (6.91)

Here, we have also substituted the solution for the field V0 and its time derivative.
We note that the Hamiltonian is slightly unusual in that it contains derivatives of
the momenta along with inverse powers of the mass. The inverse powers of the
mass in fact prevent us from taking the massless limit.28
Gladly, this Hamiltonian implies the desired equations of motion

V̇k =ı̊[H, Vk ] = Πk − m−2 ∂k ∂l Πl ,

Π̇k =ı̊[H, Πk ] = ∂l ∂l Vk − ∂k ∂l Vl − m2 Vk . (6.92)

It is not at all obvious that these equations imply the Klein–Gordon equation.
However, their twisted form is required to be able to solve for the field V0 easily
and thereby obtain the correct energy.

27
The Hamiltonian is manifestly positive since 21 ∂k Vl ∂k Vl − 12 ∂l Vk ∂k Vl = 14 (∂k Vl − ∂l Vk )2 .
28
We may impose a gauge by demanding ∂k Πk = −m−2 V0 = 0. This eliminates the inverse
mass from the Hamiltonian and validates the massless limit. Using ∂k Πk = −m−2 V0 the gauge
also implies V̇0 = ∂k Vk = 0, i.e. the gauge is the Coulomb gauge.

6.18
Quantum Field Theory I Chapter 7
ETH Zurich, 2019 HS Prof. N. Beisert
18. 11. 2019

7 Interactions
We have learned a lot about the three basic constituents of QFT in four
dimensions:
• scalar fields (spin j = 0 or helicity h = 0),
• spinor fields (spin j = 1/2 or helicity h = ± 1/2 ),
• vector fields (helicity h = ±1 or spin j = 1).
So far we considered only free fields. The particle number was conserved by all
processes and most operators we encountered so far.
Now, we would like to introduce interactions between such fields. Unfortunately,
interactions cannot be treated exactly, neither classically nor quantum
mechanically.
We have to assume the strength of interactions to be sufficiently small. The
well-understood free fields will dominate, and we insert interactions as small
perturbations. This eventually leads us to Feynman diagrams to describe particle
interactions order by order.

7.1 Interacting Lagrangians

One of the main reasons to consider QFT is its ability to deal with processes that
do not conserve the number of particles.
Quantum fields are particle creation and annihilation operators: φ → a, a† . So far,
we used them for two purposes:
• to build the multi-particle Fock space from a vacuum state, i.e. a† . . . a† |0i;
• to write conserved charges as quadratic combinations of the fields which
conspired to yield one creation and one annihilation operator a† a and thus
conserve the particle number.
Combining more than two fields typically yields a quantum operator which
changes the particle number.
Time evolution of a quantum system is governed by its Hamiltonian, therefore it is
natural to include such higher-order terms in it and consequently in the
Lagrangian.

Scalar Interactions. For a real scalar field we could consider an interacting

Lagrangian of the form

L = − 21 ∂ µ φ ∂µ φ − 21 m2 φ2 − 61 µφ3 − 1
24
λφ4 . (7.1)

7.1
This model called the φ4 theory.1 It is perhaps the conceptually simplest
interacting QFT model, but it leads to very non-trivial physics. The φ4 term is also
an interaction of the scalar Higgs field which is essential for the Higgs mechanism.
We might also add higher-order terms or terms involving derivatives such as

φ5 , φ(∂φ)2 , φ2 (∂φ)2 , (∂φ)4 , ... . (7.2)

Such terms are in principle allowable in QFT, but they have some undesirable
features regarding renormalisation which will be discussed later. We will, however,
never add non-local terms of the type to the action
Z Z
4
dx φ(x)φ(x + a), dx4 dy 4 f (x, y)φ(x)φ(y). (7.3)

These terms represent some unphysical action at a distance; we consider only local
interactions which can be written using a local Lagrangian2
Z
S = dx4 L(x), L(x) = L[φ(x), ∂φ(x), . . .]. (7.4)

The classical equation of motion for the above Lagrangian reads

∂ 2 φ − m2 φ − 21 µφ2 − 61 λφ3 = 0. (7.5)

It is a non-linear differential equation. Our usual strategy to deal with the

differential equation of motion was to go to momentum space
dq 4
Z
2 2 1
0 = −p φ(p) − m φ(p) − 2 µ φ(q)φ(p − q)
(2π)4
dq14 dq24
Z
1
− 6λ φ(q1 )φ(q2 )φ(p − q1 − q2 ). (7.6)
(2π)8
Unfortunately, we obtain an integral equation instead of an algebraic equation. We
cannot solve it in general, but for small µ  m and small λ  1 we can try to find
useful approximations.

Quantum Electrodynamics. Electrons and positrons carry an electrical charge

±q, respectively,3 and their conserved current J µ =ı̊q ψ̄γ µ ψ couples to the Maxwell
equations. Putting together the Dirac and Maxwell Lagrangians we can simply
add a source term for the electromagnetic potential J µ Aµ

LQED = ψ̄(γ µ ∂µ − m)ψ − 41 F µν Fµν −ı̊q ψ̄γ µ ψ Aµ . (7.7)

1
One often drops the term φ3 and gains a discrete symmetry φ 7→ −φ. A term φ3 without a
term φ4 would lead to a potential unbounded from below.
2
Non-local terms could be recovered as Taylor series involving derivatives of arbitrary order,
e.g. φ(x)φ(x + a) = φ2 + aφ∂φ + 12 a2 φ∂ 2 φ + . . .. It is therefore desirable to restrict the number of
derivatives that can appear in L.
3
We shall assume that q = −e < 0 is the charge of the electron where e > 0 is the unit
electrical charge. For practical purposes, the overall sign of the charges does not matter.

7.2
This yields the desired inhomogeneous Maxwell equations, but also a modification
of the Dirac equation

∂µ F µν =ı̊q ψ̄γ ν ψ,
(γµ ∂ µ − m)ψ =ı̊qγ µ Aµ ψ. (7.8)

The above model is called quantum electrodynamics (QED). It is a model that has
been tested at a remarkable accuracy within its domain of validity, i.e. at low
energies where the other elementary particles play no essential role. For instance,
the electron anomalous magnetic dipole moment, also known as “g − 2”, was
predicted to more than 10 digits, and many of the leading digits are due to plain
QED alone.
In the standard model, the above type of interaction between vectors and spinors
is arguably the most important one because it couples matter in the form of spinor
fields (leptons and quarks) to forces in the form of vector fields (photons, gluons
and others).

Gauge Invariance. A crucial property of the electromagnetic potential is its

gauge symmetry. In the quantisation procedure it eliminates an unphysical degree
of freedom of the electromagnetic potential. We therefore want to preserve this
symmetry in the presence of interactions.
The interaction term breaks the original gauge symmetry, but the latter can be
restored by extending the symmetry to the Dirac field

A0µ (x) = Aµ (x) + ∂µ α,

ψ 0 (x) = exp ı̊qα(x) ψ(x).


(7.9)

Note that the latter transformation rule is just the global U(1) symmetry of the
Dirac field which is responsible for conservation of the current J µ . This global
symmetry is enhanced to a local transformation parameter α(x). The derivative
terms of α(x) are now compensated by the inhomogeneous gauge transformation of
the potential Aµ .
There is a construction which makes the gauge invariance more manifest.
Introduce the gauge covariant derivative

Dµ = ∂µ −ı̊qAµ . (7.10)

Under gauge transformations this operator transforms homogeneously4

Dµ0 = ∂µ −ı̊qA0µ = ∂µ −ı̊qAµ −ı̊q∂µ α

= Dµ + [Dµ , −ı̊qα]
= exp(+ı̊qα)Dµ exp(−ı̊qα). (7.11)
4
This is an equation for operators. As such O1 = O2 is equivalent to the statement
O1 f = O2 f for all functions f .

7.3
In the QED Lagrangian written with a covariant derivative

LQED = ψ̄(γ µ Dµ − m)ψ − 14 F µν Fµν , (7.12)

the factors of exp(±ı̊qα) trivially cancel between ψ̄, Dµ and ψ. Moreover the
electromagnetic field strength can be written as
ı̊
Fµν = [Dµ , Dν ], (7.13)
q
which makes manifest its invariance under gauge transformations.
Consequently, we can also couple the complex scalar field to the electromagnetic
field via its kinetic term

LSQED = −(Dµ φ)∗ Dµ φ − m2 |φ|2 − 41 λ|φ|4 − 14 F µν Fµν . (7.14)

This model is called scalar QED.

Further Interactions. Let us list some other simple interactions. We want to

consider only those interactions which respect Lorentz symmetry. Curiously, all of
them appear in the standard model.
Two Dirac spinors can be multiplied to form a scalar combination. This can be
multiplied by a scalar field
ψ̄ψ φ. (7.15)
This term was originally proposed by Yukawa for the interaction between nucleons
of spin 1/2 and scalar pions. In the standard model such terms couple the Higgs
field to the leptons and quarks.
A similar term coupling a scalar and two Dirac fermions but with opposite parity
properties is
ı̊ψ̄γ 5 ψ φ. (7.16)
There is also an analog of the spinor-vector coupling with opposite parity
properties
−ı̊ψ̄γ 5 γ µ ψ Aµ . (7.17)
This so-called axial vector coupling term is relevant to the weak nuclear
interactions. Here, gauge invariance needs to extend to local chiral transformations
of the spinors.
The above interactions for Dirac 4-spinors can be written in terms of more
elementary chiral 2-spinor fields,

χT εχ φ, −χ† σ̄ µ χ Aµ . (7.18)

Note that the first interaction is complex, and therefore only some real projection
can appear in the Lagrangian. This leads to two couplings, one for the real part
and one for the imaginary part. The second term is perfectly real and requires a
single real coupling constant.

7.4
Power Counting. We have encountered several types of interaction terms.
These have a rather simple form with very few factors. Moreover, most of the
simple terms have been observed directly or indirectly in nature. However, there
are many more local terms one could imagine, but which have not been observed.
What distinguishes the above interactions?
To answer this question, consider the mass dimension. The action S must Rbe a
dimensionless quantity.5 The action is the integral of the Lagrangian S = dx4 L
and length counts as inverse mass, dx ∼ m−1 , therefore the Lagrangian must have
mass dimension 4,
L ∼ m4 . (7.19)

The kinetic terms (∂φ)2 , ψ̄∂ψ and F 2 where the derivative counts as a mass,
∂ ∼ m, determine the mass dimensions of the scalar, spinor and vector fields
φ ∼ Aµ ∼ m, ψ ∼ m3/2 . (7.20)

The mass dimension of the remaining terms is now fixed, e.g. for the mass terms
φ2 ∼ m2 , ψ̄ψ ∼ m3 , (7.21)
and for the simple interaction terms
φ3 ∼ m3 , φ4 ∼ ψ̄γ µ ψAµ ∼ ψ̄ψφ ∼ m4 . (7.22)
All of these terms have mass dimension at most 4. When they appear in the
Lagrangian L ∼ m4 , their coupling constant must compensate for the missing mass
dimension. The scalar and fermion mass terms therefore read m2 φ2 , mψ̄ψ. Among
the interaction terms, only φ3 requires a dimensionful coupling µ ∼ m. All the
other terms have mass dimension 4 and their coupling constants are plain numbers.
We can take the bound of mass dimension 4 6 as an experimentally observed
principle. There are good reasons to consider only terms of this type:
• Such interactions are reasonably simple.
• There are only finitely many such terms, hence finitely many parameters for
the model.
• All higher-dimensional terms require a coupling constant with negative mass
dimension.
• Coupling constants with negative mass dimension lead to undesirable effects in
the ultraviolet or short-distance regimes.
• Such theories are called non-renormalisable. Renormalisability will be
considered later in QFT II.
• In the infrared or long-distance regime,7 only the interactions of mass
dimension up to 4 are relevant. The higher-dimensional terms have small
effects and are mostly irrelevant.
5
Quantities that appear in an exponent must be dimensionless numbers. The action carries
the same units as Planck constant ~ which in natural units is a number ~ = 1.
6
More generally, the number of spacetime dimensions.
7
The meaning of long-distance depends on the point of view. It can be astronomical units,
everyday length scales, atomic scales or even less when interested in fundamental description of
nature.

7.5
 • The mass for a vector field reappears as an inverse power in the massive vector
propagator. This also leads to a non-renormalisable model. To explain the
mass of the W and Z vector bosons we have to rely on the Higgs mechanism.8

Symmetries. Unfortunately, it is generally hard to extract information from

interacting QFT models. Usually we can only do certain approximations.
Symmetries are powerful concepts in QFT because they can apply to interacting
models as well:
• Free theories are somewhat trivial because they have infinitely many
conservation laws. For example, the particle number is conserved, but it is
related only to a non-local transformation.
• Only few of the conservation laws typically survive when interactions are
added. Those are related to global symmetries preserved by the interactions.
• The conservation laws allow to make certain statements on the result of QFT
observables even when actual computations are not feasible.
• Symmetries of the classical theory are not necessarily respected by the
quantum theory. Such symmetries are called anomalous.
• Anomalies of local symmetries are typically bad because they spoil gauge
redundancies which are required for consistency.
For example, consider φ4 theory with a complex field given by the Lagrangian

L = −∂ µ φ∗ ∂µ φ − m2 |φ|2 − 41 λ|φ|4 . (7.23)

It is invariant under global multiplication by a complex phase φ → eı̊α φ. This

leads to the same conserved current as for the complex scalar

J µ = −ı̊(∂ µ φ∗ φ − φ∗ ∂ µ φ). (7.24)

The associated current Q = Na − Nb is exactly conserved even in the presence of

interactions. Conservation of the individual number operators Na and Nb ,
however, is broken by interactions.

7.2 Interacting Field Operators

Consider an interacting field theory whose fields (and conjugate momenta) we will
collectively denote by φ(x). More concretely, we can consider φ4 theory.9
We want to compute some correlation function, for example a correlator of two
fields at different times t1 , t2

F (x2 , x1 ) = h0|φ(t2 , ~x2 )φ(t1 , ~x1 )|0i. (7.25)

8
Also known as the Englert–Brout–Higgs–Guralnik–Hagen–Kibble–Anderson–and–perhaps–
also–’t–Hooft mechanism.
9
The spatial dependence of the fields will hardly be relevant, and the following discussion
applies just as well to any weakly interacting quantum mechanical system.

7.6
Interacting Field. As before, we can quantise the field φ on one time slice at
some time t0 . This step is identical to a free field φ0 because the canonical Poisson
brackets are the same for φ and φ0 .10 The full time dependence of φ is recovered
by conjugating with the Hamiltonian



φ(t, ~x) = exp ı̊(t − t0 )H φ(~x) exp ı̊(t0 − t)H . (7.26)

Supposing that the vacuum is time-invariant,11 we can write the correlator as

F (x2 , x1 ) = h0|φ(~x2 ) exp −ı̊(t2 − t1 )H φ(~x1 )|0i. (7.27)

Now everything is explicitly known except how to exponentiate H in practice. The

latter is a hard problem.

Interaction Picture. We can do slightly better whenever the interactions are

weak. In this case, the dominant contribution should come from the free
Hamiltonian H0 .12 The quantisation of fields at a given time slice is the same. We
can thus identify the fields at time t0
!
φ0 (~x) = φ(~x). (7.28)

Time evolution of the free field φ0 is governed by the free Hamiltonian H0

φ0 (t, ~x) = exp ı̊(t − t0 )H0 φ(~x) exp ı̊(t0 − t)H0 . (7.29)

We know almost everything about this field. For weak interactions and small times
t ≈ t0 , we expect the free field φ0 (t, ~x) to be a suitable approximation for the full
field φ(t, ~x).
Comparing φ to φ0 we can write

φ(t, ~x) = U (t)−1 φ0 (t, ~x)U (t) (7.30)

with the transformation operator

U (t) = exp ı̊(t − t0 )H0 exp ı̊(t0 − t)H[φ0 ] . (7.31)

Note that we expressed the full Hamiltonian H in terms of the free field
φ0 = φ0 (t0 ) = φ(t0 ). For small interactions and small times t ≈ t0 , this operator is
approximately the identity.
The correlator in question becomes

h0|U (t2 )−1 φ0 (t2 , ~x2 )U (t2 )U (t1 )−1 φ0 (t1 , ~x1 )U (t1 )|0i. (7.32)
10
Interactions may alter the precise relationship between the conjugate momentum π and φ̇,
but the Poisson brackets between φ and π retain their canonical form by construction.
11
A constant energy of the vacuum can always be eliminated by subtracting it from H.
12
Any free Hamiltonian H0 for the field φ0 is permissible for the following construction of the
interaction picture. In order for perturbation theory to be applicable, the interaction terms in
H − H0 must be small (but they may for instance include corrections to the mass of the free field,
see the following chapters for applications).

7.7
Note that products of transformation operators can be combined in the pertinent
time evolution operator

U (t2 , t1 ) := U (t2 )U (t1 )−1

= exp ı̊(t2 − t0 )H0 exp ı̊(t1 − t2 )H exp ı̊(t0 − t1 )H0



= exp ı̊(t2 − t1 )H0 exp ı̊(t1 − t2 )H[φ0 (t1 )] . (7.33)

In the last transformation we have used that H = H[φ] = H[φ(t0 )] = H[φ0 (t0 )] and
the relationship



X[φ0 (t1 )] exp ı̊(t1 − t0 )H0 = exp ı̊(t1 − t0 )H0 X[φ0 (t0 )], (7.34)

which holds for any quantum operator X composed from φ0 (t) without further
dependence on time. This follows from (repeated) use of the time evolution of the
free field φ0 (t).
Finally, we write the resulting correlator as

F (x2 , x1 ) = h0|U (t0 , t2 )φ0 (x2 )U (t2 , t1 )φ0 (x1 )U (t1 , t0 )|0i. (7.35)

This is called the interaction picture, it is a mixture between the Schrödinger and
the Heisenberg pictures
• In the Schrödinger picture, the field is defined on a constant time slice
φ = φ(t0 ) and the operator exp(ı̊H(t2 − t1 )) evolves states in time.
• In the Heisenberg picture, the field φ(t) carries the full time dependence, there
is no need for a time evolution operator.
• In the interaction picture, the field φ0 (t) carries the time dependence of a free
particle and the operator U (t2 , t1 ) evolves states in time.

Interacting Ground State. All the operators are expressed using the free field
φ0 , but the state |0i is a state of the interacting theory and we do not know how to
act on it.
Luckily we can express the interacting ground state |0i in terms of the vacuum |00 i
of the free theory with a trick: The free vacuum |00 i should be some linear
combination of the interacting ground state |0i and excited eigenstates |ni with
definite energy En > E0 = 0 13
X
|00 i = c0 |0i + cn |ni. (7.36)
n

Letting this state evolve for some time T with the interacting Hamiltonian H we
obtain X
exp(−ı̊HT )|00 i = c0 |0i + cn exp(−ı̊En T )|ni. (7.37)
n

All eigenstates oscillate with their respective frequencies. Suppose we give the time
T ∈ (1 −ı̊)R some small negative imaginary part with En−1  | Im T |  |T |.
13
It is reasonable to assume c0 = h0|00 i =
6 0 when interactions are sufficiently small.

7.8
Then almost all eigenstates will get exponentially suppressed compared to the
interacting ground state. The latter remains as the dominant contribution

exp(−ı̊HT )|00 i ≈ c0 |0i. (7.38)

Primarily, this identification is a formal trick. In terms of physics, we let a system

in some exited state |00 i evolve for some long time and find it in its ground state
|0i. This is reasonable if we assume the system to be open or damped in some way.
All real world systems, at least those we can expect to observe, are finite and open;
the effect of ı̊ is to implement this assumption into our calculations.
We can thus express the interacting vacuum at time t0 as the evolution of the free
vacuum at time −T

|0i ≈ c−1 −1


0 exp ı̊H(−t0 − T ) |00 i = c0 U (t0 , −T )|00 i (7.39)

Analogously,
h0| ≈ (c∗0 )−1 h00 |U (+T, t0 ). (7.40)

Our final result for the correlation function F (x2 , x1 ):

h00 |U (+T, t2 )φ0 (x2 )U (t2 , t1 )φ0 (x1 )U (t1 , −T )|00 i
lim . (7.41)
T →∞(1−ı̊) h00 |U (+T, −T )|00 i
The denominator implements the desired normalisation h0|0i = 1.

Interacting Correlators. In conclusion, the recipe for determining some

correlation function in the interacting theory is the following
h00 |U (T, t0 )XU (t0 , −T )|00 i
h0|X|0i = lim , (7.42)
T →∞(1−ı̊) h00 |U (T, −T )|00 i
where all the interacting quantum operators in X are replaced by free fields
evolved from time t0 to the desired time slice

φ(t, ~x) → U (t0 , t) φ0 (t, ~x) U (t, t0 ). (7.43)

Effectively two consecutive time evolution operators can always be combined into
one
U (t2 , t0 ) U (t0 , t1 ) = U (t2 , t1 ). (7.44)

7.3 Perturbation Theory

We still cannot evaluate the time evolution operator U (t, t0 ), but at least we know
that it is close to the identity when interactions are sufficiently small

U (t, t0 ) ≈ 1. (7.45)

This approximation is too crude, it is equivalent to computing the correlator in the

free theory, and we gain nothing.

7.9
Schrödinger Equation. To improve the approximation, consider the time
derivative of U (t, t0 )



ı̊∂t U (t, t0 ) = exp ı̊(t − t0 )H0 H[φ0 (t0 )] − H0


· exp −ı̊(t − t0 )H0 U (t, t0 )


= H[φ0 (t)] − H0 U (t, t0 ). (7.46)

We see that the time evolution operator is determined by a differential equation

and a trivial initial value condition

ı̊∂t U (t, t0 ) = Hint (t)U (t, t0 ), U (t0 , t0 ) = 1. (7.47)

This is a Schrödinger equation, and its Hamiltonian is the so-called interaction

Hamiltonian
Hint (t) := H[φ0 (t)] − H0 [φ0 ]. (7.48)
This Hamiltonian is time-dependent, therefore the solution cannot be as simple as
exp(−ı̊(t − t0 )Hint ).14 For weak interactions, one can use the Dyson series to solve
the equation perturbatively.

Dyson Series. The interaction Hamiltonian is the quantity which we should

assume to be small. It appears in the Schrödinger equation, so at first order we
can use the above approximation for U

i∂t U (t, t0 ) = Hint (t)U (t, t0 ) ≈ Hint (t). (7.49)

Integrating with proper initial value this yields

Z t
U (t, t0 ) ≈ 1 −ı̊ dt1 Hint (t1 ). (7.50)
t0

This is certainly better than before, it involves interactions at first order.

Nevertheless we can do better.
To go further systematically, write the differential equation in integral form
Z t
U (t, t0 ) = 1 −ı̊ dt1 Hint (t1 )U (t1 , t0 ). (7.51)
t0

Substitute the above solution yields a better solution

Z t
U (t, t0 ) ≈ 1 −ı̊ dt1 Hint (t1 )
t0
Z t Z t1
− dt1 dt2 Hint (t1 ) Hint (t2 ). (7.52)
t0 t0

14
Hint is time-dependent because its time evolution is governed by H0 with which it does not
commute in general.

7.10
Now use the new solution instead
Z t
U (t, t0 ) ≈ 1 −ı̊ dt1 Hint (t1 )
t0
Z t Z t1
− dt1 dt2 Hint (t1 )Hint (t2 )
t0 t0
Z t Z t1 Z t2
+ı̊ dt1 dt2 dt3 Hint (t1 ) Hint (t2 ) Hint (t3 ). (7.53)
t0 t0 t0

And so on.
The picture should be clear, we could go to arbitrarily high orders. More
importantly, everything is expressed in terms of free fields φ0 and the interaction
Hamiltonian Hint [φ0 ].

Time-Ordered Exponential. The multiple integral with a nested sequence of

boundaries is hard to manage. We can formally improve the situation, and let us
assume t > t0 for definiteness. Consider the quadratic term:
Z t Z t1
− dt1 dt2 Hint (t1 ) Hint (t2 ). (7.54)
t0 t0

We can also write it as

Z t Z t
− dt1 dt2 Hint (t2 ) Hint (t1 ). (7.55)
t0 t1

The integration region assumes t1 ≥ t2 in the first integral and t2 ≥ t1 in the

second integral. Importantly, in both integrands the operator Hint (tk ) with larger
tk is to the left of the operator Hint (tj ) with smaller tj .
t2
t
Hint (t2 )Hint (t1 )

t2 > t1
(7.56)
t1 > t2

Hint (t1 )Hint (t2 )

t0

t0 t t1

We introduce a time ordering symbol T which puts the affected operators in an

order with time decreasing from left to right, e.g.15
(
  X(t1 )Y (t2 ) for t1 > t2 ,
T X(t1 )Y (t2 ) := (7.57)
Y (t2 )X(t1 ) for t1 < t2 ,
15
For fermionic operators X, Y one would insert suitable signs for flipping the order.

7.11
and similarly for multiple operators. This allows to write the integrand of both
above integrals as the time-ordered product T[Hint (t1 )Hint (t2 )]. We can thus write
the integral as the average of the two equivalent representations where the
integration regions combine to a square

1 t
Z Z t
 
− dt1 dt2 T Hint (t1 ) Hint (t2 ) . (7.58)
2 t0 t0

Even better, we can write this as the time-ordered square of a single integral
Z t 2
1 0 0
− T dt Hint (t ) . (7.59)
2 t0

As all terms of the perturbative expansion of U (t, t0 ) are naturally in time ordering,
the above construction generalises straight-forwardly to the n-th order term
 Z t n
1 0 0
T −ı̊ dt Hint (t ) . (7.60)
n! t0

Here the integration region is a hypercube in n dimensions. It contains n!

simplices,16 which form the integration regions for the terms in the Dyson series.
Summing up all terms yields the time-ordered exponential17


U (t2 , t1 ) = T exp ı̊Sint (t2 , t1 ) , (7.61)
18
where we introduced the interaction action Sint between times t1 and t2
Z t2
Sint (t2 , t1 ) := − dt0 Hint (t0 ). (7.62)
t1

The time-ordered exponential represents both the formal solution to the above
Schrödinger equation for U (t2 , t1 ) as well as a concrete perturbative prescription to
evaluate it.19

16
A hypercube is the generalisation of a cube to n dimensions, a simplex is the generalisation
of a triangle.
17
Here, we have assumed t2 > t1 . For the opposite order of external times t1 > t2 , the reverse
time ordering is required.
18
It is clear that Lint = −Hint unless the canonical momenta are non-linear functions of the
fields. However, in the case of non-linear canonical momenta (such as in scalar QED), the
definition of the canonical commutators involves additional terms which make up for the
differences.
19
In a mathematical sense, it represents a formal power series. It may or may not converge as
an actual series; in a QFT setting it is typically merely an asymptotic series.

7.12
Quantum Field Theory I Chapter 8
ETH Zurich, 2019 HS Prof. N. Beisert
21. 11. 2019

8 Correlation Functions
We have seen how to formally write the time evolution operator


U (t1 , t0 ) = T exp ı̊Sint (t1 , t0 ) (8.1)

in an interacting QFT model based on the interaction picture and time-ordered

products.
A particularly convenient correlator is one where the operators are already in
proper time order
  
X[φ] := h0| T X[φ] |0i. (8.2)
Such time-ordered correlation functions have multiple applications in QFT, for
example, they can be used for particle scattering processes. In this chapter we will
develop methods to compute them in more practical terms. The outcome will be a
set of graphical rules, the Feynman rules.
For simplicity we will drop all free field indices φ0 → φ from now on and instead
mark interacting correlators and fields by an index “int”, φ → φint .

8.1 Interacting Time-Ordered Correlators

Consider the correlator of two time-ordered fields with t1 > t2


F = φ(t1 , ~x1 )φ(t2 , ~x2 ) int = h0int |φint (t1 , ~x1 )φint (t2 , ~x2 )|0int i. (8.3)

In the expression in terms of free fields

X = U (+T, t1 ) φ(x1 ) U (t1 , t2 ) φ(x2 ) U (t2 , −T ), (8.4)

we notice that all operators are in proper time order and we can extend the time
ordering over all the operators

 

X = T exp ı̊Sint (+T, t1 ) φ(x1 ) T exp ı̊Sint (t1 , t2 )


· φ(x2 ) T exp ı̊Sint (t2 , −T )



= T exp ı̊Sint (+T, t1 ) φ(x1 ) exp ı̊Sint (t1 , t2 )


· φ(x2 ) exp ı̊Sint (t2 , −T ) . (8.5)

Inside the time-ordering symbol the order of operators does not matter. The
exponents can now be combined nicely:


X = T φ(x1 )φ(x2 ) exp ı̊Sint (+T, −T ) . (8.6)

8.1
We thus find the correlation function hφ(x1 )φ(x2 )iint


h0| T φ(x1 )φ(x2 ) exp ı̊Sint (+T, −T ) |0i
F = lim 
 . (8.7)
T →∞(1−ı̊) h0| T exp ı̊Sint (+T, −T ) |0i

This formula generalises to vacuum expectation values of arbitrary time-ordered

combinations X[φ] of quantum operators



  h0| T X[φ] exp ı̊Sint (+T, −T ) |0i
T X[φ] int = lim 

T →∞(1−ı̊) h0| T exp ı̊Sint (+T, −T ) |0i
 
h0| T X[φ] exp(ı̊Sint ) |0i
'   . (8.8)
h0| T exp(ı̊Sint ) |0i

Here the complete interaction action Sint implies a small imaginary part for the
time coordinate in the distant past and future. We can thus express time-ordered
correlators in the interacting theory in terms of similar quantities in the free theory.
This expression has several benefits and applications:
• Typically, there are no ordering issues within X because time ordering puts all
constituent operators into some well-defined order. This is useful when
interested in the quantum expectation value of some product of classical
operators.
• It directly uses the interaction terms Sint in the action.
• Time-ordered products and expectation values can be evaluated conveniently.
• This expression appears in many useful observables, for example in particle
scattering amplitudes.

8.2 Time-Ordered Products

We now look for a method to evaluate a time-ordered correlator of a combination
of free-field operators X[φ]

  
X[φ] := h0| T X[φ] |0i. (8.9)

Feynman Propagator. We start with two fields

 
GF (x1 , x2 ) =ı̊h0| T φ(x1 )φ(x2 ) |0i. (8.10)

By construction and earlier results it reads

(
ı̊h0|φ(x1 )φ(x2 )|0i for t1 > t2 ,
GF (x1 , x2 ) =
ı̊h0|φ(x2 )φ(x1 )|0i for t2 > t1
= θ(t1 − t2 )∆+ (x1 − x2 ) + θ(t2 − t1 )∆+ (x2 − x1 ). (8.11)

8.2
Comparing this to the retarded propagator GR (x)
GF (x) = θ(t)∆+ (x) + θ(−t)∆+ (−x),
GR (x) = θ(t)∆+ (x) − θ(t)∆+ (−x), (8.12)
we can write
GR (x) = GF (x) − ∆+ (−x). (8.13)
As such it obeys the equation of a propagator,
−∂ 2 GF (x) + m2 GF (x) = δ d+1 (x), (8.14)
but with different boundary conditions than for the retarded propagator. It is
called the Feynman propagator.
The momentum space representation of the Feynman propagator for the scalar
field reads
1
GF (p) = 2 . (8.15)
p + m2 −ı̊
Here the two poles at e = ±e(~p) are shifted up and down into the complex plane
by a positive infinitesimal amount  1
 
1 1 1
GF (p) = − . (8.16)
2e(~p) e − (−e(~p) +ı̊) e − (+e(~p) −ı̊)
Im e Im e
−e(~
p) +ı̊
Re e −e(~
p) +ı̊ Re e (8.17)
p) −ı̊
+e(~
p) −ı̊
+e(~

Concerning the relation to the position space representation:

• The positive energy pole e = e(~p) −ı̊ is below the real axis and thus relevant
to positive times.
• The negative energy pole e = −e(~p) +ı̊ is above the real axis and thus relevant
to negative times.
Alternatively, to obtain the correct contour around the two poles, we could
integrate on a slightly tilted energy axis in the complex plane
e ∈ (1 +ı̊)R. (8.18)
Equivalently, we can assume times to be slightly imaginary, but with the time axis
tilted in the opposite direction
t ∈ (1 −ı̊)R. (8.19)
The ı̊ prescription of the Feynman propagators is thus directly related and
equivalent to the i prescription for converting the free vacuum to the interacting
one.
Im e Im t
−e(~
p) +ı̊ Re t (8.20)
p) −ı̊ Re e
+e(~

1
The ’s in the above and below expressions are not literally the same, but importantly they
are both positive infinitesimals.

8.3
Wick’s Theorem. To evaluate more complex time-ordered vacuum expectation
values one typically employs Wick’s theorem. It relates a time-ordered product of
operators T[X[φ]] to a normal-ordered product of operators N[X[φ]]. The
normal-ordered product is useful when evaluating vacuum expectation values
because the vacuum expectation value picks out field-independent contributions
only.
Let us recall the definition of normal ordering: Split up the free fields φ into pure
creation operators φ+ and pure annihilation operators φ−

φ = φ+ + φ− , φ+ ∼ a† , φ− ∼ a. (8.21)

Normal ordering of a product is defined such that all factors of φ+ are to the left of
all factors φ− . For example,

N φ(x1 )φ(x2 ) = φ+ (x1 )φ+ (x2 ) + φ− (x1 )φ− (x2 )

 

+ φ+ (x1 )φ− (x2 ) + φ+ (x2 )φ− (x1 ), (8.22)

where the latter two terms are in normal order and the ordering of the former two
terms is irrelevant.
In comparison, time-ordering of the same product is defined as

T φ(x1 )φ(x2 ) = φ+ (x1 )φ+ (x2 ) + φ− (x1 )φ− (x2 )

 

+ θ(t1 − t2 ) φ+ (x1 )φ− (x2 ) + φ− (x1 )φ+ (x2 )

+ θ(t2 − t1 ) φ+ (x2 )φ− (x1 ) + φ− (x2 )φ+ (x1 ) .

(8.23)

The difference between the two expressions reads2

 
(T − N) φ(x1 )φ(x2 )
= θ(t1 − t2 ) φ− (x1 ), φ+ (x2 ) + θ(t2 − t1 ) φ− (x2 ), φ+ (x1 )
   

= −ı̊θ(t1 − t2 )∆+ (x1 − x2 ) −ı̊θ(t2 − t1 )∆+ (x2 − x1 )

= −ı̊GF (x1 − x2 ). (8.24)

Wick’s theorem is a generalisation of this result to an arbitrary number of fields: It

states that the time-ordered product of a set of fields equals the partially
contracted normal-ordered products summed over multiple contractions between
pairs of fields. A Wick contraction between two, not necessarily adjacent, fields
φ(xk ) and φ(xl ) replaces the relevant two field operators by their Feynman
propagator −ı̊GF (xk , xl ), in short:

[. . . φk−1 φk φk+1 . . . φl−1 φl φl+1 . . .]

:= −ı̊GF (xk − xl ) [. . . φk−1 φk+1 . . . φl−1 φl+1 . . .]. (8.25)
2
Note that the commutators are between creation and annihilation parts φ± rather than the
full fields φ. Therefore they yield the function ∆+ rather than ∆ as can be seen by taking the
vacuum expectation value.

8.4
For example:
   
T φ1 φ2 = N φ1 φ2 + φ1 φ2 ,
   
T φ1 φ2 φ3 = N φ1 φ2 φ3 + φ1 φ2 φ3 + φ1 φ2 φ3 + φ1 φ2 φ3 ,
  h
T φ1 φ2 φ3 φ4 = N φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4
+ φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4
i
+ φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4
+ φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4 . (8.26)

To prove the statement by induction is straight-forward:

• Assume the statement holds for n − 1 fields.
• Arrange n fields in proper time order φn . . . φ1 with tn > . . . > t1 .
−
• Consider T[φn . . . φ1 ] = (φ+ n + φn ) T[φn−1 . . . φ1 ] and replace T[φn−1 . . . φ1 ] by
contracted normal-ordered products.
• φ+n is already in normal order, it can be pulled into N[. . .].
• Commute φ− n past all the remaining fields in the normal ordering.
• For every uncontracted field φk in N[. . .], pick up a term
[φ−n , φk ] = −ı̊∆+ (xn − xk ) = −ı̊GF (xn − xk ) because tn > tk .
• Convince yourself that all contractions of n fields are realised with unit weight.
• Convince yourself that for different original time-orderings of φn . . . φ1 , the step
functions in GF do their proper job.

Time-Ordered Correlators. To compute time-ordered correlators we can use

the result of Wick’s theorem. All the normal-ordered terms with remaining fields
drop out of vacuum expectation values. The only terms to survive are those where
all the fields are complete contracted in pairs
X
hφ1 . . . φn i := φ1 φ2 . . . . . . . . . . . . φn−1 φn . (8.27)
complete
contractions

In particular, it implies that correlators of an odd number of fields must be zero.

This formula applies directly to a single species of real scalar fields, but for all the
other fields and mixed products there are straight-forward equivalents:
• For fields with spin, use the appropriate propagator, e.g. (GD )a b for contracting
the Dirac fields ψ a and ψ̄b .
• For any crossing of lines attached to fermionic fields, multiply by a factor of
(−1).

8.5
8.3 Some Examples
We have learned how to reduce time-ordered correlators in a weakly interacting
QFT to free time-ordered correlators

 hX[φ] exp(ı̊Sint [φ])i
X[φ] int = . (8.28)
hexp(ı̊Sint [φ])i

We have also learned how to evaluate the latter

X
hφ1 . . . φn i := φ1 φ2 . . . . . . . . . . . . φn−1 φn . (8.29)
complete
contractions

We will now apply these formulas to some basic types of time-ordered correlators
in order to develop an understanding for them.

Setup. We will consider φ4 theory, i.e. a single real scalar field with a φ4
interaction
L = − 21 ∂ µ φ ∂µ φ − 12 m2 φ2 − 24
1
λφ4 . (8.30)
We define the interaction picture using the quadratic terms in the action

L0 = − 21 ∂ µ φ ∂µ φ − 12 m2 φ2 . (8.31)

What remains is the interaction term

1
Lint = − 24 λφ4 , (8.32)

whose coefficient, the coupling constant λ, is assumed to be small. The interaction

part of the action is thus Z
Sint := dx4 Lint (x). (8.33)

We would like to evaluate some correlators of the form

X = hφ1 . . . φm iint , (8.34)

where φk denotes the field φ(xk ) evaluated at position xk . These are functions of
the coupling constant λ, and we formally expand them for small λ as
∞
X
X(λ) = X (n) , X (n) ∼ λn , (8.35)
n=0

and consider the leading two terms X (0) and X (1) .

8.6
Two-Point Function. First, we would like to evaluate the correlator of two
fields
T12 (λ) = hφ1 φ2 iint . (8.36)
At leading order we simply set λ = 0 and obtain the correlator in the free theory

T (0) = hφ1 φ2 i. (8.37)

Using Wick’s theorem this evaluates to

T (0) = φ1 φ2 = (−ı̊)G12 , (8.38)

where Gkl denotes the Feynman propagator GF (xk − xl ).

The contributions to the interacting two-point function at the next perturbative
order read3

T (1) = hφ1 φ2 ı̊Sint [φ]i − hφ1 φ2 ihı̊Sint [φ]i

Z
ı̊λ
=− dy 4 hφ1 φ2 φy φy φy φy i
24
Z
ı̊λ
+ dy 4 hφ1 φ2 ihφy φy φy φy i. (8.39)
24
Using Wick’s theorem the two terms expand to 15 and 3 contributions. Consider
the first term only: The 15 contributions can be grouped into two types. The first
type receives 12 identical contributions from contracting the 4 identical φy ’s in
superficially different ways. The remaining 3 terms in the second group are
identical for the same reason. We summarise the groups as follows
Z
(1)
T1a = − 2 λ dy 4 φ1 φ2 φy φy φy φy ,
ı̊

Z
(1)
T1b = − 8 λ dy 4 φ1 φ2 φy φy φy φy .
ı̊
(8.40)

The term originating from the denominator of the interacting correlator evaluates
to Z
(1) (1)
T2 = − 8 (−ı̊λ) dy 4 φ1 φ2 φy φy φy φy = −T1b .
1
(8.41)

It precisely cancels the second contribution to the first term. Altogether we find
the following expression for the first-order correction to the two-point function
Z
(1)
T = T1a = 2 (−ı̊) λ dy 4 G1y G2y Gyy .
(1) 1 4
(8.42)

In fact, the term Gyy is problematic as it stands; we shall comment on this issue
later.
3
The first term originates from the expansion of the numerator of the interaction correlation
function, the second term from the denominator.

8.7
Diagrams. Before we continue, it will be useful to introduce a graphical
notation to represent our results which will later be generalised to the concept of
Feynman diagrams. A term consists of collection of nodes or vertices which are
connected by lines or edges. Each vertex represents a spacetime point xk and each
line connecting two vertices k and l represents a propagator −ı̊GF (xk − xl ).
In this notation, the free two-point function is written as
T (0) = x1 x2 . (8.43)
The terms arising at the first perturbative order can be written as
Z
(1)
T = T1a = − 2 λ dy 4 x1
(1) ı̊
x2 ,
y
Z y
(1) (1)
T1b = −T2 = − 8ı̊ λ dy 4 . (8.44)
x1 x2
We notice that the second, disconnected term drops out from the overall result;
this is a general feature of interactions to be discussed later.

Four-Point Function. As a more elaborate example for an interacting

correlation function, we now consider the four-point correlator
F1234 (λ) = hφ1 φ2 φ3 φ4 iint . (8.45)
We shall first evaluate it at lowest order in the coupling strength, where we set
λ = 0 and obtain the correlator in the free theory
F (0) = hφ1 φ2 φ3 φ4 i. (8.46)
Using Wick’s theorem this evaluates to
F (0) = φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4 + φ1 φ2 φ3 φ4

= (−ı̊)2 G12 G34 + (−ı̊)2 G13 G24 + (−ı̊)2 G14 G23 . (8.47)
In a graphical notation we could write this as
x1 x4 x1 x4 x1 x4
F (0) = + + . (8.48)
x2 x3 x2 x3 x2 x3
Here, it is natural to add diagrams, each of which represents a certain functional
term.
The first-order contributions to the interacting four-point function take a similar
form as for the two-point function
F (1) = hφ1 φ2 φ3 φ4 ı̊Sint [φ]i − hφ1 φ2 φ3 φ4 ihı̊Sint [φ]i
Z
ı̊λ
=− dy 4 hφ1 φ2 φ3 φ4 φy φy φy φy i
24
Z
ı̊λ
+ dy 4 hφ1 φ2 φ3 φ4 ihφy φy φy φy i. (8.49)
24

8.8
These expressions are not as innocent as they may look: Using Wick’s theorem the
two terms expand to 7 · 5 · 3 · 1 = 105 and 3 · 3 = 9 terms. Gladly, most of these
terms are identical and can be summarised, we group them into 24, 6 · 12 and 3 · 3
terms from the first contribution and 3 · 3 terms from the second one
Z
(1)
F1a = (−ı̊λ) dy 4 φ1 φ2 φ3 φ4 φy φy φy φy ,
Z
(1)
F1b = 1
2
(−ı̊λ) dy 4 φ1 φ2 φ3 φ4 φy φy φy φy + 5 perm.,
Z
(1)
F1c = 1
8
dy 4 φ1 φ2 φ3 φ4 φy φy φy φy + 2 perm.,
(−ı̊λ)
Z
(1)
F2 = − 8 (−ı̊λ) dy 4 φ1 φ2 φ3 φ4 φy φy φy φy + 2 perm..
1
(8.50)

The graphical representation of these terms is

x1 x4
(1)
F1a ' ,
y
x2 x3

x1 x4 x1 x4 x1 x4
(1) y
F1b ' y + +
y
x2 x3 x2 x3 x2 x3
x1 x4 x1 x4 x1 x4
y
y
+ y + + ,
x2 x3 x2 x3 x2 x3

x1 x4 x1 x4 x1 x4
(1)
y
F1c ' y + + . (8.51)
y
x2 x3 x2 x3 x2 x3

Let us now discuss the roles and features of the three terms.

Vacuum Bubbles. In the above result we notice that again the contribution
(1) (1)
F2 from the denominator of the interacting correlator cancels the term F1c from
the numerator. This effect is general:

x1 x4
(8.52)
y
x2 x3

Some graphs have components which are coupled neither to the rest of the graph
nor to the external points. Such parts of the graph are called vacuum bubbles.
• Vacuum bubbles represent some virtual particles which pop out of the quantum
mechanical vacuum and annihilate among themselves. They do not interact
with any of the physically observed particles, hence one should be able to
ignore such contributions.

8.9
 • VacuumR bubbles are usually infinite. Here we obtain as coefficient
GF (0)2 dy 4 . ThisRcontains two divergent factors of GF (0) and an infinite
spacetime volume dy 4 .
• Formally, we could remove such terms by adding a suitable field-independent
term to the interaction Lagrangian. Alternatively we could normal order it.
• In any case, vacuum bubbles are generally removed by the denominator of the
interacting correlation function. This cancellation ensures that the interacting
vacuum is properly normalised, h0|0iint = 1. Any diagram containing at least
one vacuum bubble can be discarded right away.

(1)
Disconnected Graphs. The contribution from F1b is reminiscent of the
correction T (1) to the two-point function. In fact it can be written as a sum of
products of two-point functions
(1) (0) (1) (0) (1) (0) (1)
F1b = T12 T34 + T13 T24 + T14 T23
(1) (0) (1) (0) (1) (0)
+ T12 T34 + T13 T24 + T14 T23 . (8.53)

This combination is precisely the first-order contribution to a product of two

T (λ)’s
x1 x2 = x1 x2 + λ x1 x2 + ... . (8.54)
y
This is also a general feature of correlation functions:
• Correlation functions contain disconnected products of lower-point functions.
The corresponding graphs contain disconnected components (each of which is
connected to at least one external field).
• Such contributions are typically put aside because their form can be composed
from correlators of fewer fields.4 Nevertheless, they are essential and
non-negligible contributions to the correlation functions.
• Such disconnected contributions represent processes that take place
simultaneously without interfering with each other.
Quite generally one can split the contributions into connected and disconnected
terms. Here we know, to all orders in λ

F (λ) = T12 (λ)T34 (λ) + T13 (λ)T24 (λ) + T14 (λ)T23 (λ)
+ Fconn (λ),
x1 x4 x1 x4 x1 x4 x1 x4
= + +
x2 x3 x2 x3 x2 x3 x2 x3
x1 x4
+ , (8.55)
x2 x3
4
When computing an n-point function one will typically already have computed all the
k-point functions with k < n anyway.

8.10
where Fconn (λ) summarises all connected contributions. In our case
Z
(1)
Fconn = F1a = −ı̊λ dy 4 G1y G2y G3y G4y ,
(1)

x1 x4 x1 x4
=λ + ... . (8.56)
y
x2 x3 x2 x3

Tadpoles. We were careful enough not to write the leading correction to the
two-point function too explicitly
Z
T = 2 λ GF (0) dy 4 GF (x1 − y)GF (x2 − y).
(1) 1
(8.57)

We notice that one of the propagators decouples from the function. Moreover its
argument is precisely zero because the propagator connects a point to itself.
• The result is divergent, it is very similar to the vacuum energy we encountered
much earlier in QFT.
• In our derivation of time ordering we were sloppy in that we did not discuss the
case of equal times. In a local Lagrangian, however, all terms are defined at
equal time, moreover at equal spatial position. It would make sense to employ
normal ordering in this case, which eliminates the term right at the start.
• Whatever the numerical value of GF (0), even if infinite, it does not yield any
interesting functional dependence to T (1) . In fact it could be eliminated by
formally adding a term − 4ı̊ λGF (0)φ2 to the interaction Lagrangian. This has
the same effect as normal ordering the Lagrangian.
This term is called a tadpole term because the corresponding diagram looks like a
tadpole sitting on the propagator line.

(8.58)

More generally, tadpoles are internal parts of a diagram which are attached to the
rest of the diagram only via a single vertex. In most cases, they can be
compensated by adding suitable local terms to the interaction Lagrangian. Even
though this correction term is somewhat dangerous and somewhat trivial, let us
pretend it is a regular contribution and carry it along.

Symmetry Factors. In our computation, we have encountered many equivalent

contributions which summed up into a single term. We observe that these sums
have conspired to cancel most of the prefactors of 1/24 . The purpose of having a
prefactor of 1/24 for φ4 in the action is precisely to be cancelled against
multiplicities in correlators, where λ typically appears without or with small
denominators.

8.11
We can avoid constructing a large number of copies of the same term by
considering the symmetry of terms or the corresponding graphs. The symmetry
factor is the inverse size of the discrete group that permutes the elements of a term
or a graph while leaving its structure unchanged.
To make use of symmetry factors for the calculation of correlation functions, one
should set up the Lagrangian such that every product of terms comes with the
appropriate symmetry factor. For example, the term φ4 allows arbitrary
permutations of the 4 φ’s. There are 4! = 24 such permutations, hence the
appropriate symmetry factor is 1/24 .5
The crucial insight is the following: when the symmetry factors for the Lagrangian
are set up properly, the summed contributions to correlation function also have
their appropriate symmetry factors.
Occasionally, determining the symmetry factors correctly can be difficult, as one
has to identify all permissible permutations. It is tricky, for example, when the
graphical representation has a complicated topology or when it hides some relevant
information.
Let us consider the symmetry factors of the terms we have computed so far. The
(1)
contributions T (0) , F (0) and Fconn have trivial symmetry factors.

1 4
(8.59)
2 3
Permutations of any of the elements would change the labelling of the external
fields. The symmetry factor for the tadpole diagram is 1/2 .

(8.60)

The relevant Z2 symmetry flips the direction of the tadpole line.6 Finally, the
vacuum bubble diagram has a symmetry factor of 1/8 .

(8.61)

There are two factors of 1/2 for flipping the direction of the tadpole lines. Then
there is another factor of 1/2 for permuting the two tadpole lines.
5
After all, we are free to call the term that multiplies φ4 either λ/24 or λ0 . It is not even
difficult to translate between them.
6
In fact, the symmetry acts on the connections of lines to vertices. Here, exchanging the two
endpoints of the tadpole line is the only symmetry.

8.12
8.4 Feynman Rules
We have seen how to evaluate some perturbative contributions to interacting
correlators. Following the formal prescription leads to a lot of combinatorial
overhead as the results tend to be reasonably simple compared to the necessary
intermediate steps. Feynman turned the logic around and proposed a simple
graphical construction of correlators:
The interacting correlator F of several fields can be expressed as a sum of so-called
Feynman graphs X
F = Fgraph . (8.62)
graphs

Each Feynman graph represents a certain mathematical expression which can be

evaluated from the graph by the Feynman rules. Moreover a Feynman graph nicely
displays the physical process that leads to the corresponding term of the correlator.
For every weakly coupled QFT there is a set of Feynman rules to compute its
correlators.7 Here we list the Feynman rules for the scalar φ4 model.

Feynman Rules in Position Space. Consider a generic correlator of several

fields

F (x1 , . . . , xn ) = φ(x1 ) . . . φ(xn ) int . (8.63)
A permissible graph for this correlator:
• has undirected and unlabelled edges;

zk zl (8.64)

• has n 1-valent (external) vertices labelled by xj ;

xj (8.65)

• has an arbitrary number m of 4-valent (internal) vertices labelled by yj ;

yj (8.66)

• can have lines connecting a vertex to itself (tadpole);

(8.67)

• can have multiple lines connecting two vertices;

(8.68)
7
Similar graphs and rules can actually be set up and applied to a wide range of algebraic
problems not at all limited to relativistic QFT’s.

8.13
 • can have several connection components;

(8.69)

• must not have components disconnected from all of the external vertices xj
(vacuum bubble).

(8.70)

For each topologically distinct graph we can compute a contribution according to

the following rules:
• For each edge connecting two vertices zk and zl write a factor
zk zl −→ −ı̊GF (zk − zl ). (8.71)
• For each 4-valent vertex yj , write a factor of −ı̊λ and integrate over dyj4
Z
yj −→ −ı̊λ dyj4 . (8.72)

• Multiply by the appropriate symmetry factor, i.e. divide by the number of

discrete symmetries of the graph.

Feynman Rules in Momentum Space. One of the problems we have not yet
mentioned is that the Feynman propagator GF is a complicated function in
spacetime. Moreover, we need to compute multiple convolution integrals of these
(1)
functions over spacetime, e.g. the integral defining Fconn . This soon enough exceeds
our capabilities.
These computations can be simplified to some extent by going to momentum
space. Such a momentum space representation will be particularly useful later
when we compute the interaction between particles with definite momenta in
particle scattering experiments.
The momentum space version is defined as follows8
Z
F (p1 , . . . , pn ) = dx41 . . . dx4n e−ı̊x1 ·p1 −...−ı̊xn ·pn hφ1 . . . φn i. (8.73)

A Feynman graph in momentum space:

8
Note that we are evaluating a time-ordered correlator. This is well-defined in position space
and we have to perform the Fourier integrals after computing the correlator. It implies that the
momenta pj can and should be taken off-shell p2j + m2 6= 0. This is different from computing a
p1 ) . . . a† (~
correlator such as h0|a(~ pn )|0i where all the momenta are defined only on shell
2 2
pj + m = 0.

8.14
 • has edges labelled by a directed flow of 4-momentum `j from one end to the
other;
`j (8.74)
• has n 1-valent (external) vertices with an outflow of 4-momentum pj ;
pj (8.75)
• has an arbitrary number m of 4-valent (internal) vertices which conserve the
flow of momentum;

(8.76)

• shares the remaining attributes with the position space version.

The Feynman rules for evaluating a graph read:
• Work out the flow of momentum from the external vertices across the internal
vertices. Label all edges with the appropriate momenta `j .

`i + `j
(8.77)
`i `j

• There is a momentum-conservation condition pj1 + . . . + pjm = 0 for each

connected component of the graph. It includes all contributing external
momenta pj . Write a factor enforcing momentum conservation

−→ (2π)4 δ 4 (pj1 + . . . + pjm ). (8.78)

• For each internal loop of the graph, there is one undetermined 4-momentum `j .
Integrate the final expression over all such momenta

d`4j
Z
`j −→
(2π)4
. (8.79)

• For each edge write a factor

`j −ı̊
−→ . (8.80)
`2j + m2 −ı̊
• For each 4-valent vertex, write a factor

−→ −ı̊λ. (8.81)

• Multiply by the appropriate symmetry factor, i.e. divide by the number of

discrete symmetries of the graph.

8.15
General Models. We observe that the Feynman graphs and rules for a QFT
model reflect quite directly the content of its action:
• In particular, the free part of the action S0 determines the types and features
of the fields and particles. These are reflected by the Feynman propagator GF
which is associated to the edges.
• The interaction part of the action Sint contains all the information about the
set of interaction vertices.

Examples. Let us apply the Feynman rules to compute the mathematical

expressions for a few Feynman graphs.
(1)
Consider first the graph for the leading connected contribution Fconn to the
four-point function.
x1 x4 p1 p4
(8.82)
x2
y x3 p2 p3
Applying the rules for position space, we obtain right away
Z 4
Y
(1) 4
Fconn = (−ı̊λ)(−ı̊) dy 4 GF (xj − y). (8.83)
j=1

In momentum space, the corresponding result is

4
(1) 4 4 4
Y 1
Fconn = (−ı̊λ)(−ı̊) (2π) δ (p1 + p2 + p3 + p4 ) . (8.84)
p2
j=1 j
+ m2 −ı̊

This expression is merely a rational function and does not contain any integrals. It
is therefore conceivably simpler than its position space analog. Unfortunately, it is
generally not easy to perform the Fourier transformation back to position space.9
Next, consider a slightly more complicated example involving an internal loop.

x1 x4 p1 ` p4
y1 y2 (8.85)
x2 x3 p2 p3
`0

Evaluation of the Feynman graph in position space is straight-forward

Z
2 6
1
F = 2 (−ı̊λ) (−ı̊) dy14 dy24 GF (y1 − y2 )2

· GF (x1 − y1 )GF (x2 − y1 )GF (x3 − y2 )GF (x4 − y2 ). (8.86)

The symmetry factor is 1/2 because the two lines of the internal loop can be
interchanged.
9
A notable exception is the massless case where the correlation functions in position space has
a manageable form.

8.16
For momentum space, we first have to label the remaining lines along the internal
loop: The total flow of momentum out of the left vertex to the external lines is
p1 + p2 , whereas the momenta on both internal lines are yet undetermined. The
sum of internal momenta flowing out of the vertex must therefore equal −p1 − p2 by
momentum conservation.10 One internal momentum remains undetermined, let us
call it ` and eventually integrate over it. The other one must equal `0 = p1 + p2 − `.
4
Y 1
F = 1
2
(−ı̊λ)2 (−ı̊)6 (2π)4 δ 4 (p1 + p2 + p3 + p4 )
p + m2 −ı̊
j=1 j

d`4
Z
1 1
. (8.87)
(2π)4 `2 + m2 −ı̊ (p1 + p2 − `)2 + m2 −ı̊

Now we are left with a multiple integral over a rational function. There exist
techniques to deal with this sort of problem, we will briefly discuss some of the
basic ones at the end of this course. Some integrals like this one can be performed,
but most of them remain difficult and it is an art to evaluate them. Unfortunately,
numerical methods are not applicable straight-forwardly either.11 This is a generic
difficulty of QFT with no hope for a universal solution. The Feynman rules are a
somewhat formal method and it is hard to extract concrete numbers or functions
from them.

8.5 Feynman Rules for QED

Finally, we would like to list the Feynman rules for the simplest physically relevant
QFT model, namely quantum electrodynamics (QED). We shall use the
Lagrangian in Feynman gauge

L0 = ψ̄(γ µ ∂µ − m)ψ − 21 ∂ µ Aν ∂µ Aν ,
Lint = −ı̊q ψ̄γ µ ψ Aµ . (8.88)

Feynman Graphs in Momentum Space. A non-trivial interacting correlation

function in this model must contain as many fermionic fields ψ as conjugates ψ̄
due to global U(1) symmetry. Consider such a correlation function

Aµ1 (k1 ) . . . Aµm (km ) ψ̄b1 (q1 )ψ a1 (p1 ) . . . ψ̄bn (qn )ψ an (pn ) .


(8.89)

Admissible Feynman graphs have the following properties in addition to or instead

of the ones of the φ4 model:
10
By considering the right vertex, it must also equal p3 + p4 . This requirement is consistent by
means of overall momentum conservation p1 + p2 + p3 + p4 = 0.
11
The integrals are over an unbounded space with a large dimension. Moreover the poles and
asymptotic behaviour of the integrand lead to numerical instabilities and divergences. The letter
are inherent to QFT and need to be dealt with.

8.17
 • There are two types of edges: undirected wavy lines (photons) or directed
straight lines (electrons and/or positrons).
`j `j (8.90)
ck dl νk νl
• The edges are labelled by a directed flow of 4-momentum `j .
• The ends of wavy lines are labelled by indices νk and νl ; the ends of straight
lines are labelled by indices ck and dl in the direction of the arrow of the
straight line.
• There is a 1-valent (external) vertex for each field in the correlator. The
momentum outflow and the label at the end of the edge are determined by the
corresponding field.

ψ: pj ψ̄ :
qj A: kj (8.91)
aj bj µj
• There is one type of (internal) vertex: It is 3-valent and connects an ingoing
and an outgoing straight line (fermion) with a wavy (photon) line.

ck νj
(8.92)
dl
The QED-specific Feynman rules read as follows:
• The graph can have only fermion loops, which contribute an extra factor of
(−1) due to their statistics

d`4j
Z
`j −→ (−1)
(2π)4
. (8.93)

• For each wavy edge write a factor

`j −ı̊ηνk νl
−→ ; (8.94)
νk νl `2j −ı̊
for each straight edge write a factor12
`j −ı̊(−ı̊`j ·γ + m)ck dl
ck −→ . (8.95)
dl `2j + m2 −ı̊
• For each 3-valent vertex, write a factor

ck νj
−→ −q(γ νj )ck dl . (8.96)
dl
Note that the arrows on the fermion lines conveniently describe the ordering of
gamma-matrix factors from left to right and from ψ to ψ̄.
12
The signs in the numerator of the Feynman propagator match with the direction of
momentum flow required for the Dirac equation to hold.

8.18
Example. Let us consider the simplest interacting correlator in QED

F = Aµ (k) ψ̄b (q)ψ a (p) .


(8.97)

We shall restrict to the leading order with one interaction vertex.

p
a
c ν µ (8.98)
d k
bq
Applying the Feynman rules leads to the following expression

F (1) = (2π)4 δ 4 (k + p + q)(−q)(γ ν )c d

−ı̊ηµν −ı̊(ı̊p·γ + m)a c −ı̊(−ı̊q·γ + m)d b
· . (8.99)
k 2 −ı̊ p2 + m2 −ı̊ q 2 + m2 −ı̊
Summing over the internal indices c, d, ν and combining the three spinor matrices
we obtain a matrix product as the answer
 a
−ı̊q (ı̊p·γ + m)γµ (−ı̊q·γ + m) b
F (1) = (2π)4 δ 4 (k + p + q) 2 2 2 2 2
. (8.100)
(k −ı̊)(p + m −ı̊)(q + m −ı̊)

Note that the ordering of spinor matrix factors from left to right follows the flow of
the spinor lines in the Feynman diagram.

QED and Gauge Invariance. Note that QED is a gauge theory which requires
some gauge fixing. Feynman gauge is very convenient, but any other consistent
gauge is acceptable, too. Different gauges imply different propagators which lead
to non-unique results for correlation functions. Unique results are only to be
expected when the field data within the correlation function is gauge invariant.
More precisely, the correlator should contain
R 4 the gauge potential Aµ (x) only in the
µ
combination Fµν (x) or as the coupling dx J Aµ to some conserved current
Jµ (x). Moreover, charged spinor fields should be combined into uncharged
products, e.g. ψ̄(x) . . . ψ(x) potentially dressed with covariant derivatives.

8.19
Quantum Field Theory I Chapter 9
ETH Zurich, 2019 HS Prof. N. Beisert
28. 11. 2019

9 Particle Scattering
A goal of this course is to understand how to compute scattering processes in
particle physics.

9.1 Scattering Basics

Setup. The usual setup for scattering experiments at particle colliders is the
following:

q1

p1
(9.1)

p2 qn

• Two bunches of particles are accelerated to high or relativistic velocities and

made to collide.
• Whenever two particles from the bunches come very close, they produce some
complicated interacting quantum state.
• After a while this state evolves into several particles moving away in various
directions.
• The outgoing particles of each scattering event are measured and recorded.
Some additional comments:
• Quantum mechanics is probabilistic, so a large number of particle collisions
must be measured.
• To measure collisions of three or more incoming particles is technically
challenging because they would all have to be focused within a tiny region of
space simultaneously. The likelihood for two particles to scatter is much higher.
• By Lorentz symmetry, the directions of the two ingoing momenta p~1,2 can be
adjusted arbitrarily. In some reference frame, the momenta will be parallel and
along the z-axis. The relevant quantity√is the centre-of-mass energy squared
s = −(p1 + p2 )2 . The highest energies s are obtained where the collisions are
head-on with equal but opposite momenta. For practical purposes, the particles
can have momenta of different magnitude or one of the two bunches could be a
fixed target at rest.
• The particle momenta in the beam are not perfectly aligned. By the

9.1
 uncertainty principle this is actually impossible if the beam is also focused on a
finite area.
• The particle detectors are not perfect: They have a certain spatial and
temporal resolution. They measure the energy and momenta at a certain
resolution. They may not be able to detect and distinguish all kinds of
particles; they may miss some particles; they may misidentify some. Scattered
particles along the beam direction are hardest to detect.

Cross Sections. How to quantify scattering? Consider a simple classical

scattering experiment:

A
d
r2
r1
(9.2)

r1 + r2
σ

• Take two hard balls of radii r1 , r2 .

• Throw them towards each other along the z-axis in opposing directions.
• Depending on the transverse offset d, the balls will either hit (d < r1 + r2 ) or
miss (d > r1 + r2 ).1
• When the balls hit, they bounce off in different directions.
Quantum mechanics is probabilistic, there cannot be such deterministic output.
One has to repeat the experiment many times or perform an experiment with
many identical particles and count:
• Accelerate two bunches of n1 and n2 particles.
• Focus each bunch on a cross-sectional area A.
• Repeat the experiment nex times.
• Count the number of individual scattering events N .
The expectation value for N is
nex n1 n2 σ
N= , (9.3)
A
where the characteristic quantity is the scattering cross section σ. For two classical
hard balls one obtains σ = π(r1 + r2 )2 : Given the transverse position of the first
ball, the centre of the second ball must be within an area of σ to make the two
collide.
In collider experiments one measures scattering cross sections:
1
More accurately, at a near miss, the flow of air will also deform the balls’ trajectories slightly.

9.2
 • Total or inclusive cross sections σ simply count the number of overall collision
events.
• Differential cross sections dσ measure the number of events where the outgoing
particles have predetermined momenta.2 The definition of dσ depends on the
number of outgoing particles. The so-called phase space is furthermore
constrained by Poincaré symmetry.
• One may even resolve the polarisation of the outgoing particles and measure
polarised cross sections.

9.2 Cross Sections and Matrix Elements

The computation of the scattering cross section is not straight-forward. Naively,
we prepare initial and final states with definite momenta p1 , p2 and q1 , . . . , qn at
some times tin and tout in the distant past and distant future

hf| ∼ hq1 , . . . , qn |, |ii ∼ |p1 , p2 i. (9.4)

The probability is given by the square of the correlator hf| exp(−ı̊H(tout − tin ))|ii

2
σ ∼ hf| exp −ı̊H(tout − tin ) |ii . (9.5)

For initial and final states with definite momenta, the correlator contains a
delta-function δ 4 (Pin − Pout ) to conserve momentum. It cannot be squared because
this would result in a factor of δ 4 (0) = ∞. We know that such factors represent
some volume of spacetime relevant to the problem. A proper treatment requires
the use of wave packets. They actually account for the finite extent of the ingoing
bunches, namely the cross-sectional area A, and for the finite resolution of the
detector. The factor δ 4 (0) represents this area A among others.
A somewhat tedious calculation in terms of wave packets yields a meaningful
result for the differential cross section of 2 → n scattering. At the end of the day,
the wave packets can be focused to definite momenta3
n
3n (2π)4 δ 4 (Pin − Pout ) Y d~qk 3
d σ=

3 0
|M |2 . (9.6)
4 e2 (~p2 )~p1 − e1 (~p1 )~p2 k=1 (2π) 2ek (~qk )

Here, M is the appropriate element of the scattering matrix with the

momentum-conserving delta-function stripped off

lim hf| exp −ı̊H(tout − tin ) |ii = (2π)4 δ 4 (Pin − Pout )ı̊M.


(9.7)
tin,out →∓∞

2
The direction of the scattered classical balls is determined by the impact parameter d, and
hence certain regions of the scattering cross section correspond to specific angles. In quantum
mechanics this is mostly a matter of probability.
3
This quantity is not invariant under Lorentz transformations due to the denominator
ke(~ p2 − e(~
p1 )~ p1 k. Nevertheless, it is covariant and transforms like an area as it should.
p2 )~

9.3
The normalisation is such that in the free theory the correlator for n = 2 final
state particles equals4

2e1 (~p1 ) 2e2 (~p2 ) (2π)6 δ 3 (~p1 − ~q1 )δ 3 (~p2 − ~q2 ). (9.8)

For 2 → 2 scattering it makes sense to fix the outgoing overall momentum Pout ,
and express the result in terms of the spherical angle element d2 Ω of the direction
of outgoing particle 1

d2 σ 1 k~q1 k3 |M |2
= e2 (~p2 )~p1 − e1 (~p1 )~p2 e02 (~q2 )~q1 − e01 (~q1 )~q2 ·~q1 64π 2 .

(9.9)
d2 Ω

The formula simplifies in the centre-of-mass frame to

d2 σ k~q k |M |2
= , (9.10)
d2 Ω k~p k 64π 2 s

where s = −Pin2 = −(p1 + p2 )2 is the centre-of-mass energy squared. Its form

becomes even simpler in case all four particles are identical

d2 σ |M |2
= . (9.11)
d2 Ω 64π 2 s

9.3 Electron Scattering

We can now compute some realistic scattering event in quantum electrodynamics.
We shall consider scattering of two electrons into two electrons (Møller
scattering).5
e−
q1
p1 θ
e− p2 e− (9.12)

q2
e−
Here, we will not distinguish the two polarisation modes of the electron spin. One
might as well consider the polarised cross section, but the experimental setup as
well as the theoretical calculation is more challenging.

Initial and Final States. To prepare the initial and final states we use the
interaction picture. The free reference field provides the creation and annihilation
operators for the in- and outgoing particles which are assumed not to interact
4
This contribution representing no scattering is actually removed from M for 2 → 2 particle
scattering.
5
Depending on conventions, our calculation may also represent positron-positron scattering.
Obviously, the cross section is exactly the same by charge conjugation symmetry.

9.4
when sufficiently far away. Moreover the initial and final states will be practically
independent of tin and tout as long as the latter are sufficiently large.
The initial state is composed from two ingoing electrons

|ii = a†α (~p1 )a†β (~p2 )|0i. (9.13)

The electrons have definite momenta p1 , p2 . Let us assume they are in their
centre-of-mass frame with momenta aligned along the z-axis

p1,2 = (e, 0, 0, ±p), (9.14)

where e2 = p2 + m2 . The polarisations α, β are required to set up the state

properly. We will not care about them, so we should eventually average over all
ingoing polarisation configurations.
We want to probe the final state for two outgoing electrons

hf| = h0|aδ (~q2 )aγ (~q1 ). (9.15)

In the centre-of-mass frame they will escape in two opposite directions with the
same magnitude p of momentum. Due to rotational symmetry around the z-axis,6
we only need to probe for particles in the x,z-plane

q1,2 = (e, ±p sin θ, 0, ±p cos θ). (9.16)

Again we shall not care about the polarisations. We therefore have to sum over all
outgoing polarisation configurations.
For a fixed particle momentum p or energy e, we will be interested in the angular
distribution of outgoing particles. Due to rotational symmetry the differential
cross section d2 σ/d2 Ω must be an even function of the scattering angle θ alone.
This function also has the symmetry θ → π + θ because the outgoing particles are
indistinguishable

d2 σ d2 σ d2 σ d2 σ
= (θ) = (−θ) = (π − θ). (9.17)
d2 Ω d2 Ω d2 Ω d2 Ω

Time Evolution. We now insert the time evolution operator Uint of the
interaction picture between the initial and final states to determine the probability
amplitude7

F = lim hf|Uint (tout , tin )|ii = (2π)4 δ 4 (Pin − Pout )ı̊M. (9.18)
tin,out →∓∞

The matrix element is a function of the momenta and the polarisations

Mαβγδ (p1 , p2 , q1 , q2 ).8
6
We will not measure polarisations which would otherwise break the symmetry.
7
The conventional factor of ı̊ typically makes the leading contributions to M (mostly) real.
8
We can write it as a function of all the external momenta noting that we shall only evaluate
it for p1 + p2 = q1 + q2 .

9.5
The expansion of the amplitude at leading order reads simply

F (0) = hf|ii
= 2e(~p1 ) 2e(~p2 ) (2π)6 δ 3 (~p1 − ~q1 )δ 3 (~p2 − ~q2 )
− 2e(~p1 ) 2e(~p2 ) (2π)6 δ 3 (~p1 − ~q2 )δ 3 (~p2 − ~q1 ). (9.19)

The contribution from the free theory represents the situation when the two
particles pass by each other without scattering at all. Note that there are two
terms corresponding to the fact that the particles are indistinguishable.
At first perturbative order the matrix element vanishes
Z
F =ı̊hf|Sint |ii =ı̊ dx4 hf|Lint (x)|ii
(1)

Z
= q dx4 hf|Aµ (x)ψ̄(x)γ µ ψ(x)|ii = 0 (9.20)

because there is a single electromagnetic field which cannot contract to anything

else and thus directly annihilates either of the vacua.

Second Order. For the next order we insert two interaction Lagrangians
Z
(2) 1 2
dx4 dy 4 hf| T Lint (x)Lint (y) |ii.
 
F = 2 ı̊ (9.21)

Each of the interaction Lagrangians contains an electromagnetic field. As they

would otherwise annihilate the vacua, they have to be contracted via Wick’s
theorem (for the field A) by a Feynman propagator
Z
1 2
dx4 dy 4 hf| T Aµ (x)ψ̄(x)γ µ ψ(x) Aν (y)ψ̄(y)γ ν ψ(y) |ii.
 
2
q (9.22)

Next, Wick’s theorem should be applied to the time-ordered spinor fields yielding
several contributions:
• There are two vacuum bubble contributions.

(9.23)

These vacuum processes take place everywhere and all the time, and they do
not interact with the scattering process. As discussed earlier, they must be
discarded.
• There are two correction terms with two remaining external fields.

(9.24)

9.6
 They contribute to two point functions of spinor fields, but cannot be
non-trivial functions of all the four scattering particle momenta. Here they
contribute only to forward scattering, and we can safely ignore their
contribution. We will discuss their relevance later.
• Finally, there is one connected diagram.

(9.25)

This is the leading non-trivial contribution to the scattering process.

Connected Contribution. In our case, the connected diagram is obtained by

replacing time ordering by normal ordering
Z
1 2
dx4 dy 4 GFµν (x − y)hf| N ψ̄(x)γ µ ψ(x) ψ̄(y)γ ν ψ(y) |ii.
 
− 2 ı̊q (9.26)

Now we need to contract the fields with the external particles. This is achieved by
the following two anti-commutators which follow from the mode expansion of the
free Dirac field
ψ̄(x), a†α (~p) = eı̊p·x v̄α (~p),


aα (~q), ψ(x) = e−ı̊q·x vα (~q).


(9.27)
Putting everything together we obtain the matrix element
q1 , γ q2 , δ q1 , γ q2 , δ

(2) (2)
Fconn = Ft + Fu(2) = − ,

p1 , α p2 , β p1 , α p2 , β
Z
(2)
Ft = −ı̊q 2 dx4 dy 4 GFµν (x − y) eı̊p1 ·x+ı̊p2 ·y−ı̊q1 ·x−ı̊q2 ·y

· v̄α (~p1 )γ µ vγ (~q1 ) v̄β (~p2 )γ ν vδ (~q2 ),

Fu(2) = − . . . . (9.28)
The omitted term takes the same form, but with the two outgoing particles
exchanged (q1 ↔ q2 , γ ↔ δ). The reason for the doubling of terms is that two
identical types of particles are scattered. The two-particle wave function is
anti-symmetric because the particles are fermionic.
The two remaining integrals are Fourier transforms. One of them transforms the
Feynman propagator to momentum space. The other one generates the momentum
conserving delta-function. Altogether the integral yields
Z
dx4 dy 4 GFµν (x − y) eı̊p1 ·x+ı̊p2 ·y−ı̊q1 ·x−ı̊q2 ·y

= (2π)4 δ 4 (Pout − Pin ) GFµν (q1 − p1 ). (9.29)

9.7
We now separate off the momentum conserving delta-function and write the
matrix element M 9
q 2 ηµν
Mt = − v̄α (~p1 )γ µ vγ (~q1 ) v̄β (~p2 )γ ν vδ (~q2 ),
(p1 − q1 )2
Mu = − . . . . (9.30)

We could try to evaluate the various spinor products. It turns out to be much
simpler to square the matrix element first
1 X ∗
|M |2 = Mαβγδ Mαβγδ . (9.31)
4 α,β,γ,δ

The factor of (1/2 )2 originates from averaging over the polarisations of the ingoing
particles.

Sum over Polarisations. A pleasant feature of the polarisation sums is that

they can be performed by the completeness relations for spinor solutions
X
vα (~p)v̄α (~p) =ı̊p·γ − m. (9.32)
α

We obtain the following three terms

q 4 Ttt q 4 Tuu q 4 Ttu

|M |2 = + − , (9.33)
4(p1 − q1 )4 4(p1 − q2 )4 2(p1 − q1 )2 (p1 − q2 )2

corresponding to the diagrams

∗ ∗ ∗
p1M p2 p1M p2 p1 M p2
q2
q1 q2
. (9.34)
q1 q2 q1
p1 M p2 p1 M p2 p1 M p2

The spinor products have turned into the traces

Ttt = tr[(ı̊p1 ·γ − m)γµ (ı̊q1 ·γ − m)γν ]

· tr[(ı̊p2 ·γ − m)γ µ (ı̊q2 ·γ − m)γ ν ],
Tuu = tr[(ı̊p1 ·γ − m)γµ (ı̊q2 ·γ − m)γν ]
· tr[(ı̊p2 ·γ − m)γ µ (ı̊q1 ·γ − m)γ ν ],
Ttu = tr[(ı̊p1 ·γ − m)γµ (ı̊q1 ·γ − m)γ ν
· (ı̊p2 ·γ − m)γ µ (ı̊q2 ·γ − m)γν ]. (9.35)
9
The i prescription for the Feynman propagator will not be relevant here.

9.8
The double-trace terms are most conveniently evaluated using the spinor trace
formulas10

tr(1) = 4,
tr(γ µ ) = 0,
tr(γ µ γ ν ) = 4η µν ,
tr(γ µ γ ν γ ρ ) = 0,
tr(γ µ γ ν γ ρ γ σ ) = 4(η µν η ρσ − η µρ η νσ + η µσ η νρ ). (9.36)

This brings Ttt into the following form

Ttt = 16(p1µ q1ν + q1µ p1ν − (p1 ·q1 + m2 )ηµν )

· (pµ2 q2ν + q2µ pν2 − (p2 ·q2 + m2 )η µν )
= 32(p1 ·p2 )(q1 ·q2 ) + 32(p1 ·q2 )(q1 ·p2 )
+ 32m2 (p1 ·q1 + p2 ·q2 ) + 64m4 . (9.37)

The other double-trace term takes a similar form with q1 and q2 interchanged. The
crossed single-trace term can be simplified by means of the enveloping identities

γµ γ µ = 4,
γµ γ ν γ µ = −2γ ν ,
γµ γ ν γ ρ γ µ = 4η νρ ,
γµ γ ν γ ρ γ σ γ µ = −2γ σ γ ρ γ ν . (9.38)

After some algebra we obtain

Ttu = −32(q1 ·q2 )(p1 ·p2 ) − 32m4

− 16m2 (p1 ·p2 + p1 ·q1 + p1 ·q2 + p2 ·q1 + p2 ·q2 + q1 ·q2 ). (9.39)

Mandelstam Invariants. In order to simplify the expressions we introduce the

Mandelstam invariants11

s = −(p1 + p2 )2 = −(q1 + q2 )2 ,
t = −(p1 − q1 )2 = −(p2 − q2 )2 ,
u = −(p1 − q2 )2 = −(p2 − q1 )2 . (9.40)

Inverting the relations we can write all scalar products of momenta using the s, t, u

p1 ·p2 = q1 ·q2 = m2 − 12 s,
p1 ·q1 = p2 ·q2 = 12 t − m2 ,
p1 ·q2 = p2 ·q1 = 21 u − m2 . (9.41)
10
The latter of these formulas follow from anti-commuting one gamma-matrix past all the
others.
11
For external momenta pi in cyclic order (here p3 := −q2 , p4 := −q1 ), the Mandelstam
invariants are given by s = −(p1 + p2 )2 , t = −(p1 + p4 )2 , u = −(p1 + p3 )2 .

9.9
Note furthermore that momentum conservation implies the relation12
s + t + u = 4m2 . (9.42)
Using Mandelstam invariants, the traces can be expressed very compactly as13
Ttt = 8(t2 − 2su + 8m4 ),
Tuu = 8(u2 − 2st + 8m4 ),
Ttu = −8(s2 − 8m2 s + 12m4 ). (9.43)
The squared matrix element now reads14
2
16q 4 m2 (5m2 − 2s)

2 4 u−s t−s
|M | = q + + . (9.44)
t u tu
This expression is symmetric under exchange of t and u as it should because the
outgoing particles are of the same kind.

Angular Distribution. In order to understand the angular distribution of

scattered particles we express the invariants in terms of the scattering angle θ and
the three-momentum magnitude p = k~pk of the particles
s = 4p2 + 4m2 , t = −2p2 (1 − cos θ), u = −2p2 (1 + cos θ). (9.45)
and insert everything into the differential cross section
d2 σ q4
 2
(4p + 2m2 )2 8p4 + 12m2 p2 + 3m4

= − +1 . (9.46)
d2 Ω 64π 2 e2 p4 sin4 θ p4 sin2 θ
This expression is the leading non-trivial contribution to the angular distribution
of scattered electrons.
We can however notice a problem by inspecting the expression. It diverges when
• the electron momenta p are small or
• the scattering angle θ is close to 0 or π.
In those regions of parameter space, the formula cannot be trusted. The deeper
reason for the divergences is that the photons which transmit the electromagnetic
force are massless. Massless particles cause some conceptual problems in scattering
processes.
The divergences are also relevant to the total cross section15
Z 1
d2 σ
σ= 2π d cos θ 2 . (9.47)
0 dΩ
12
This constraint implies that functions of s, t, u can be written in several alternative ways
much alike functions of p1 , p2 , q1 , q2 which are constrained by p1 + p2 − q1 − q2 = 0.
13
It is not straight-forward to derive these particular expressions, but it is easy to confirm that
they match with some other expression upon substituting, e.g. s = 4m2 − t − u.
14
We can identify the first term as the corresponding result in scalar QED.
15
The outgoing particles are indistinguishable, hence the integration extends only over one half
of the spherical angles. This is sufficient since at leading order hf|ii has two terms one of which
covers the opposite angles π − θ. Alternative the integral over all spherical angles must be
multiplied by a factor of 1/2 .

9.10
This integral diverges at cos θ = 1.
In order to properly address the above problematic regimes, one would have to
take higher perturbative corrections and competing processes into account.
However, only the full non-perturbative expression can provide exact results in
those regimes.
Nevertheless one should not expect a meaningful result for the total cross section
because the electromagnetic force is long-ranged: The photon propagator is not
exponentially suppressed at long distances. Effectively all particles scatter at least
by tiny amount and therefore the overall probability for scattering is 1. The
scattering cross section σ is the complete area A of the bunches which is infinite
due to our assumption of exactly defined momenta.

Crossing Symmetry. A closely related process is the scattering of electrons and

positrons (Bhabha scattering).

e−
q1
p1 θ
e− p2 e+ (9.48)

q2
+
e
It can be computed in much the same way.
The relevant connected diagrams for electron-positron scattering are
q1 , γ q2 , δ q1 , γ q2 , δ

. (9.49)

p1 , α p2 , β p1 , α p2 , β
The cross section turns out to be exactly the same as for electron scattering but
with s and u interchanged
s ↔ u. (9.50)
The resulting leading contribution to the squared matrix element is16
2
16q 4 m2 (5m2 − 2u)

2 4 s−u t−u
|M | = q + + . (9.51)
t s st

Indeed the computation is exactly the same when replacing

X X
p2 ↔ −q2 , vα (~p2 )v̄α (~p2 ) ↔ − uα (~q2 )ūα (~q2 ). (9.52)
α α
16
Apart from effects due to identical particles, the electron-positron scattering cross section
does not differ substantially from the case of electron-electron scattering. The difference between
attraction and repulsion manifests in the phase of matrix elements rather than in their absolute
value.

9.11
This relationship is called crossing symmetry. In terms of Feynman diagrams, the
positron line is equivalent to the electron line in the reverse direction
q1 q2 q1 −p2

←→ . (9.53)

p1 p2 p1 −q2

9.4 Pair Production

The above electron-positron scattering involves a process where the two particles
combine into a photon and subsequently split up into a pair. This process is mixed
with photon exchange in the t-channel.
A class of similar processes is pair production where the oppositely charged
particles annihilate and create a pair of charged particles of a different kind.

Y−
q1
p1 θ
X− p2 X+ (9.54)

q2
Y+
Let us compute scattering cross sections for such processes.
We assume that the outgoing particles have a mass mf and charge ±qf which is
different from the ingoing ones labelled by mi and ±qi .

Spinor Processes. First we consider the case where all external particles are
spinors.
q1 q2

(9.55)

p1 p2
There is now only one spinor trace to be evaluated

T = tr[(ı̊p1 ·γ − mi )γµ (ı̊p2 ·γ + mi )γν ]

· tr[(ı̊q1 ·γ + mf )γ µ (q2 ·γ − mf )γ ν ],
= 16 −p1µ p2ν − p2µ p1ν − (−p1 ·p2 + m2i )ηµν

−q1µ q2ν − q2µ q1ν − (−q1 ·q2 + m2f )η µν

= 4(t − u)2 + 4s2 + 16(m2i + m2f )s. (9.56)

9.12
The Mandelstam invariants are defined as above, but due to the different masses
their relationships have to be adjusted
p1 ·p2 = m2i − 21 s,
q1 ·q2 = m2f − 12 s,
p1 ·q1 = p2 ·q2 = 12 t − 21 m2i − 12 m2f ,
p1 ·q2 = p2 ·q1 = 12 u − 12 m2i − 12 m2f ,
s + t + u = 2m2i + 2m2f . (9.57)

Next we express the invariants as functions of the scattering angle

s = 4e2 , t = −2pi pf (1 − cos θ), u = −2pi pf (1 + cos θ) (9.58)
q
with the three-momentum magnitudes pi,f = e2 − m2i,f of the in- and outgoing
particles. The unpolarised matrix element squared now reads
qi2 qf2 T
|M |2 =
4s2
e2 − m2i e2 − m2f m2i + m2f

= qi2 qf2 2
cos θ + +1 . (9.59)
e2 e2 e2
The formula for the differential cross section for our configuration of masses and
momenta in the centre-of-mass frame reads
s
d2 σ e2 − m2f |M |2
= . (9.60)
d2 Ω e2 − m2i 256π 2 e2

Total Cross Section. This expression is free from singularities and can be
integrated to a total cross section
Z 1 s Z 1
2
d cos θ d σ 1 e2 − m2f d cos θ
σ = 4π 2
= 2 2 2
|M |2 . (9.61)
−1 2 dΩ 64πe e − mi −1 2
Upon integration we obtain the final result
s
2 2
qi qf e2 − m2f e2 + 12 m2i e2 + 12 m2f
σ= . (9.62)
48πe2 e2 − m2i e2 e2

We can plot the energy-dependence of this function.

σ

(9.63)

mf e
Quite clearly the total energy 2e of the scattered particles must be at least as large
as the sum of masses 2mf of produced particles. There is a sharp increase above
production threshold, a maximum slightly above threshold (for mi < mf ), and a
slow 1/e2 asymptotic descent.

9.13
Processes Involving Scalars. It is interesting to compare this process to the
corresponding one of charged scalars. The matrix element reads
(t − u)2 2 2 2
2 2 e − mi e − mf
2
|M |2 = qi2 qf2 = q q
i f cos2 θ, (9.64)
s2 e2 e2
and after integration we obtain the total cross section
s
2 2
qi qf e2 − m2f e2 − m2i e2 − m2f
σ= . (9.65)
192πe2 e2 − m2i e2 e2

Let us finally consider a mixed process of spinors scattering into scalars for which
the matrix element reads
−(t − u)2 + s2 − 4m2f s
|M |2 = qi2 qf2
s2
e2 − (e2 − m2i ) cos2 θ e2 − m2f
= qi2 qf2 . (9.66)
e2 e2
For the total cross section we obtain
s
qi2 qf2 e2 − m2f e2 + 21 m2i e2 − m2f
σ= . (9.67)
192πe2 e2 − m2i e2 e2

The opposite process of scalars scattering into spinors merely has a different
overall factor s
qi2 qf2 e2 − m2f e2 − m2i e2 + 21 m2f
σ= . (9.68)
48πe2 e2 − m2i e2 e2

We observe that, although the matrix elements differ substantially, the final cross
sections take a very predictable form: There are particular factors for scalars and
spinors in the initial and final states, namely
• e2 − m2i for ingoing scalars,
• e2 + 21 m2i for ingoing spinors,
• e2 − m2f for outgoing scalars,
• 4(e2 + 12 m2f ) for outgoing spinors.
The square root on the other hand is a kinematical factor corresponding to the
number/volume of initial and final states (phase space).
Let us discuss the physics of these factors in view of the total spin of the particle
pairs: Two scalars form a spin-0 state, whereas the 4 polarisations of two spinors
make up 1 spin-0 and 3 spin-1 states. We can deduce that a spin-0 state couples to
the photon by a factor e2 − m2 whereas a spin-1 state must couple by the factor
e2 + m2 such that the combination of two outgoing spinors yields the observed
overall factor of
1(e2 − m2 ) + 3(e2 + m2 ) = 4(e2 + 21 m2 ). (9.69)
For ingoing spinors the factor of 4 is compensated by taking the average rather
than a sum. It is also worth taking a look at the factors close to threshold

9.14
e2 = m2 , where we find complete suppression for the spin-0 state with e2 − m2 = 0,
whereas a spin-1 state couples with a non-zero strength e2 + m2 = 2m2 . This
makes sense because the photon is a vector particle, and by conservation of
angular momentum it can only couple to a state of overall angular momentum 1.
At threshold, the outgoing particles have no orbital angular momentum, so their
angular momentum equals the overall spin. A photon can therefore only transform
into to the spin-1 combination of two spinors, but not to pair of scalars. Above
threshold this restriction is lifted by the availability of orbital angular momentum.

9.5 Loop Contributions

We have obtained the leading-order contributions to some particle scattering
processes. Let us finally take a peek at contributions at higher orders in the
perturbation series.
For the electron scattering process at the next order q 4 there are several classes of
diagrams contributing.

(9.70)

Here we listed only the connected diagrams up to obvious symmetric copies.

It is easy to see that these diagrams lead to two types of problems:
• The diagram with a bubble on the external leg is ill-defined.
`

p p
(9.71)
p−`

Due to momentum conservation the momentum on both sides of the bubble is

the same. All external momenta originate from particle creation and
annihilation operators a† (~p) and a(~p). These momenta are exactly on the mass
shell p2 = −m2 . Conversely, the internal line represents a Feynman propagator
1/(p2 + m2 −ı̊) which is to be evaluated right on the pole
1
at p2 = −m2 . (9.72)
p2 + m2 −ı̊
This diagram therefore makes no sense as a contribution to the scattering
process.
• Most of the integrals are actually divergent in the UV, i.e. where the loop
momentum ` is very large. For example, the bubble on the photon line yields
`
`2 + . . . d`4
Z Z
4
∼ d` 4 ∼ → ∞. (9.73)
k k ` + ... `2
`+k

9.15
We have to understand how to deal with these two problems. This is going to be
the subject of the final two chapters.

9.16
Quantum Field Theory I Chapter 10
ETH Zurich, 2019 HS Prof. N. Beisert
16. 12. 2019

10 Scattering Matrix
When we computed some simple scattering processes we did not really know what
we were doing. At leading order this did not matter, but at higher orders
complications arise. Let us therefore discuss the notion of asymptotic particle
states and their scattering matrix in more detail. We shall start from scratch using
the results of the previous few chapters only as a guideline.

10.1 Asymptotic States

First we need to understand asymptotic particle states in the interacting theory

|p1 , p2 , . . .i. (10.1)

In particular, we need to understand how to include them in calculations by

expressing them in terms of the interacting field φ(x).
Asymptotic particles behave like free particles at least in the absence of other
nearby asymptotic particles.1 For free fields we have seen how to encode the
particle modes into two-point correlators, commutators and propagators. Let us
therefore investigate these characteristic functions in the interacting model.

Two-Point Correlator. Consider first the correlator of two interacting fields

φ(x)
∆+ (x − y) :=ı̊h0|φ(x)φ(y)|0i. (10.2)
Due to Poincaré symmetry, it must take the form

dp4 ı̊p·x
Z
∆+ (x) =ı̊ e θ(p0 )ρ(−p2 ). (10.3)
(2π)4

The factor θ(p0 ) ensures that all excitations of the ground state |0i have positive
energy. The function ρ(s) parametrises our ignorance. We do not want tachyonic
excitations, hence the function should be supported on positive values of s = m2 .
We now insert a delta-function to express the correlator in terms of the free
1
For any actual measurement this will be an approximation because particles will continue to
interact at arbitrary distances, yet with extremely low strength or probability. In that sense, the
precise mathematical formulation of scattering is an idealisation which cannot be achieved in
experiment (somewhat similar to Fourier transforms). In an experiment one has to make a choice
at what point a particle is considered asymptotical. This introduces at least one extra length
scale into the problem.

10.1
 √
two-point correlator ∆+ (s; x, y) with mass s (Källén, Lehmann)
Z ∞
dp4 ı̊p·x
Z
ds
∆+ (x) =ı̊ ρ(s) 4
e θ(p0 )2πδ(p2 + s)
0 2π (2π)
Z ∞
ds
= ρ(s)∆+ (s; x). (10.4)
0 2π

This identifies ρ(s) as the spectral

√ function for the field φ(x): It tells2 us by what
amount particle modes of mass s will be excited by the field φ(x).

Spectral Function. For a free field of mass m0 we clearly have

ρ0 (s) = 2πδ(s − m20 ). (10.5)

For weakly interacting fields, we should obtain a similar expression. In typical

situations we expect the spectral function to have the following shape.
ρ
multi-
original particle
particle continuum
(10.6)
bound
mass states
gap
√
m 2m s

The sharp isolated peak represents a single particle excitation with mass m. Now
the interacting field φ(x) may also excite multi-particle modes (with the same
quantum numbers). Multi-particle modes in the free theory would have energy
e ≥ 2m: For example, a two-particle state |~p1 , p~2 i carries the overall momentum
P = p1 + p2 . Under Lorentz transformations √ it is related
√ to similar states on the
2
same mass shell given by the invariant mass s = −P ≥ 2m (equivalently,
energy E = P 0 at rest P~ = 0).3 In the spectral function they form a continuum
since the momenta of the individual particles can sum up to arbitrary energies in
the frame at rest. In the presence of interactions, bound states may form whose
rest energies are somewhat below e = 2m. Whenever these bound states are stable
they will also be represented by sharp peaks.
We observe that our spectral function has at least two mass gaps: One separates
the vacuum from the lowest excitation; the other separates the latter from bound
states and the multi-particle continuum. The isolated modes are called asymptotic
2
The spectral function describes the spectrum of all quantum states only to some extent. For
example, not all states may be excited by the action of a single field φ(x). In particular, in a
model with several kinds of fields, each field can excite only a subset of particles or states (e.g.
the appropriate charges have to match).
3
In the Poincaré representation theory the invariant mass serves the same purpose as the mass
of a particle. However, it makes sense to distinguish the two notions: In a multi-particle state, one
can vary the momenta of the individual particles and thus change the invariant mass. Conversely,
the mass of a particle is a fixed quantity and cannot be varied by changing the momentum.

10.2
particles. This is the type of particle which we would like to collide. The
assumption of a mass gap is crucial in this definition.
For weak interactions, we expect that the free particle mode approximates the
asymptotic particle well.4 The interactions may shift the mass m0 → m slightly;
they may also change the strength with which this mode is excited by the field
φ(x). Therefore, the weakly interacting spectral function takes the form
ρ(s) = 2πZδ(s − m2 ) + bound states + continuum. (10.7)
The factor Z is called field strength or wave function renormalisation.

Asymptotic Particles. Based on the above discussion we can expand the

interacting field φ(x) as5
√
φ(x) = Z φas (x) + |bound states{z+ continuum} + operators . (10.8)
| {z } | {z }
∼ a† + a ∼ a† n + an ∼ a† m an
Here φas (x) is a canonically normalised free field of the physical mass m expressed
by means of creation and annihilation operators a† , a
d~p 3
Z
eı̊p·x a(~p) + e−ı̊p·x a† (~p) .


φas (x) = 3
(10.9)
(2π) 2e(~p)
The other terms in the field φ(x) are multiple creation and/or annihilation
operators which we shall not consider in detail.
Single-particle asymptotic states are created simply by a† (~p) from the vacuum.
The Hamiltonian Has for the free asymptotic field reads
d~p 3
Z
Has = e(~p) a† (~p)a(~p). (10.10)
(2π)3 2e(~p)
The characteristic property of Has is that it reproduces exactly the time evolution
of the vacuum and single-particle states
Has |0i = 0 = H|0i,
Has a† (~p)|0i = e(~p)a† (~p)|0i = Ha† (~p)|0i. (10.11)

We shall use the free creation and annihilation operators as a convenient,

distinguished basis to expand our interacting fields using the interaction picture.6
4
For reasonably strong interactions, bound states may approach the single particle states and
even acquire lower energies. This case shows that the notion of fundamental particles is not
evident in general QFT, but it belongs to weakly interacting models. In fact, some models may
have alternative formulations where the fundamental degrees of freedom are some bound states of
the original formulation.
5
Note that the following discussion does not follow from a concrete Lagrangian. It rather
makes use of the basic ingredients of QFT to formulate asymptotic particles based on their
physical properties.
6
The interaction picture is based on a reference free model which can be chosen freely.
Typically one would choose it so that interactions represent small corrections, however that
leaves the residual freedom of shifting small free contributions between the reference model and
the interaction terms. The asymptotic model is fixed such that it is exact in the one-particle
sector, i.e. all terms due to interactions cancel there.

10.3
The extra operatorial terms in the field φ(x) are some higher-order polynomials in
the operators a† , a which create and annihilate bound state particles and states
from the multi-particle continuum. They cannot be observed in the spectral
function.

Commutator and Normalisation. The other characteristic functions now

follow from our expression for the correlator. As before these can be expressed as
convolutions of the same spectral function ρ(s) with their free counterparts.
The expectation value of the unequal-time commutator
 
∆(x − y) :=ı̊h0| φ(x), φ(y) |0i (10.12)

therefore reads Z ∞
ds
∆(x) = ρ(s)∆(s; x). (10.13)
0 2π
We know that for a normalised free field the equal-time commutation relations
˙ 0, ~x) = δ 3 (~x). Hence
imply ∆(s;
Z ∞
˙ 3 ds
h0|[φ(~x), φ̇(~y )]|0i =ı̊∆(0, ~x − ~y ) =ı̊δ (~x − ~y ) ρ(s). (10.14)
0 2π

Assuming that the field φ(x) is canonically normalised,7 we have the constraint
Z ∞
ds
ρ(s) = 1. (10.15)
0 2π

When using the above expansion of the real field φ(x) in terms of creation and
annihilation operators, it also follows that the function ρ(s) must be positive.
Hence the coefficient Z for the asymptotic modes should be between 0 and 1.

10.2 S-Matrix
For the scattering setup we define two asymptotic regions of spacetime, one in the
distant past ti → −∞ and one in the distant future tf → +∞.

Asymptotic Regions. On the initial time slice we create wave packets which
are well separated in position space and narrowly peaked in momentum space. We
let these quantum mechanical wave packets evolve in time. At some instance the
wave packets collide. Then the state is evolved further until all outgoing wave
packets are sufficiently well separated


|fi = exp −ı̊H(tf − ti ) |ii. (10.16)
7
This is evident at least if the interaction terms do not contain derivatives.

10.4
Now the initial and final states are in the Schrödinger picture and they evolve even
at asymptotic times. It is hard to compare them to see what the effect of
scattering is.8

(10.17)
ti |ii t0 tf |fi

At asymptotic times the wave packets are assumed to be sufficiently well separated
such that they effectively do not interact. Therefore, we can use the asymptotic
Hamiltonian of the asymptotic field φas 9

d~p 3
Z
Has = e(~p) a† (~p)a(~p). (10.18)
(2π)3 2e(~p)
to shift the two time slices onto a common one conventionally positioned at t = 0



|outi = exp ı̊Has tf |fi, |ii = exp −ı̊Has ti |ini. (10.19)

The relationship between the in and out states is the following

|outi = exp ı̊Has tf exp −ı̊H(tf − ti ) exp −ı̊Has ti |ini
=: Uas (tf , ti )|ini. (10.20)

The in and out states |ini and |outi are both defined at time t = 0. Consequently,
they are elements of the same Hilbert space and can be compared directly. The
operator Uas is the time evolution operator for the interaction picture based on the
asymptotic Hamiltonian Has and the reference time slice at t = 0.

(10.21)
ti → −∞ |ini t0 tf → +∞ |outi

S-Matrix Definition. As interactions have become negligible at asymptotic

times, the in and out states are almost independent of ti and tf . It therefore makes
sense to take the limit ti,f → ∓∞. The limit of the time evolution operator for
infinite times is called the S-matrix




S = lim exp ı̊Has tf exp ı̊H(ti − tf ) exp −ı̊Has ti
ti,f →∓∞

= lim Uas (tf , ti ) = Uas (+∞, −∞). (10.22)

ti,f →∓∞

8
The latter figure is somewhat misleading in a quantum mechanical setting. It shows only one
out of many potential final states.
9
This asymptotic Hamiltonian is a specialisation of the free Hamiltonian H0 used previously
in the interaction picture. The free Hamiltonian was merely required to agree with the full
Hamiltonian at leading order. The asymptotic Hamiltonian furthermore has to agree with the
full Hamiltonian exactly when acting on the vacuum or one-particle states.

10.5
It transforms in states to out states

|outi = S|ini. (10.23)

Note that the in and out Hilbert spaces are isomorphic.10 This allows us to
compare states between the two. To compute matrix elements of the S-matrix,
prepare definite in and out states11 using creation and annihilation operators a† , a
for the reference time slice at t = 0

|ini = |p1 , . . . , pm i := a† (~p1 ) . . . a† (~pm )|0i,

hout| = hq1 , . . . , qn | := h0|a(~q1 ) . . . a(~qn ). (10.24)

Conventionally, scattering amplitudes M are defined as the matrix elements of

S − 1 with the overall momentum-conserving delta-function stripped off

hout|(S − 1)|ini = (2π)4 δ 4 (P − Q)ı̊M (p1 , . . . , pm ; q1 , . . . , qn ). (10.25)

The combination S − 1 is particularly useful for 2 → n scattering processes: It

removes all direct contractions between the in and out states as well as all other
disconnected contributions.12

Properties of the S-Matrix. The S-matrix has a number of useful properties,

let us list a few relevant ones.
First of all, the S-matrix is trivial for the ground state and for single-particle states

S|0i = |0i, S|~pi = |~pi. (10.26)

This follows from the definition of the asymptotic Hamiltonian to strictly emulate
the action of the interacting Hamiltonian on these states.
The S-matrix is a unitary operator

S † = S −1 . (10.27)

This property follows from the definition. It reflects the fact that probabilities are
conserved across scattering processes.
The S-matrix is also Poincaré invariant

U (ω, a)SU (ω, a)−1 = S. (10.28)

10
It is natural to assume that outgoing particles of some scattering process can be used as
ingoing particles of another scattering process. Therefore the in and out spaces must be
isomorphic.
11
These in and out states are not to be related by |outi = S|ini. The out and in states hout|,
|ini rather define the basis for the matrix elements M .
12
When one of the ingoing particles does not participate in the scattering, the S-matrix must
act trivially on the other. For general m → n scattering, the matrix elements indeed contain
direct contractions and disconnected contributions.

10.6
10.3 Time-Ordered Correlators
When we expressed the time-evolution operator in the interaction picture, we
realised that time-ordered correlation functions hφ(x1 ) . . . φ(xn )i are very natural
objects. The S-matrix is defined as the time evolution operator for the interaction
picture in terms of asymptotic states. Lehmann, Symanzik and Zimmermann
derived a relationship between the S-matrix elements and time-ordered expectation
values.

Asymptotic States. First we need to understand how to represent particle

creation and annihilation operators a† , a in terms of the field φ(x). Above, we have
expanded the field φ(x) as
√ Z
d~p 3 ı̊p·x −ı̊p·x †


φ(x) = Z e a(~
p) + e a (~
p) + ... . (10.29)
(2π)3 2e(~p)

The omitted terms represent the contributions from multi-particle states and
operators which annihilate the vacuum.
Previously we were able to isolate a†0 (~p) from a time slice of the free field φ0 (x) as
Z
†
a0 (~p) = d~x 3 eı̊p·x e(~p)φ0 (x) −ı̊φ̇0 (x) .


(10.30)

This was easy because there are only two modes with energy e = ±e(~p) in the free
field φ0 . The linear combination of φ0 and φ̇0 selects the correct one.
The interacting field, however, carries many other modes whose precise nature we
do not understand a priori. To select the modes corresponding to a† and a we need
to drive the field φ(x) for a sufficiently long time with a frequency that is in
resonance with the relevant modes. Let us sketch the construction for a single
oscillator f (t) = c eı̊ωt with resonance frequency ω
Z t2
ı̊c
dt e−ı̊et f (t) = e−ı̊(e−ω)t2 − e−ı̊(e−ω)t1 .


F (e) = (10.31)
t1 e−ω

The longer the time, the stronger will be the amplitude at e = ω. At infinite time
the function F (e) develops a pole at e, so we set t1 = −∞
t2
ı̊c e−ı̊(e−ω)t2
Z
ı̊c
F (e) = dt e−ı̊et f (t) = = + finite. (10.32)
−∞ e−ω e−ω

Here we have discarded the term that keeps oscillating at asymptotic times.13
What remains is an isolated pole at e = ω whose residue is proportional to the
amplitude c. The residue is in fact independent of the time t2 where the driving
stops.
13
As usual, one could formally dampen this term by introducing some small imaginary part.
This may be an approximation, but even in practice, one can never isolate a resonance perfectly.

10.7
Applied to the field φ(x) we find
Z t2 Z
dt d~x 3 eı̊p·x φ(x)
−∞
√
ı̊ Z
θ(−p0 )ain (−~p) − θ(p0 )a†in (~p) + . . . .


= 2 2
(10.33)
p +m

What remains are isolated poles at p0 = ±e(~p) whose residues are creation and
annihilation operators for ingoing asymptotic particles. The remaining terms are
either finite or irrelevant when creating well-separated wave packets.
An important subtlety is that the operators ain (~p) belong to an asymptotic Fock
space in the distant past. The S-matrix, however, is defined via operators a(~p)
belonging to an asymptotic Fock space at the reference time t = 0. These Fock
spaces are isomorphic but not identical, they are related by the appropriate
evolution operator14

ain (~p) = Uas (0, −∞)a(~p)Uas (−∞, 0). (10.34)

We note:
• The residues of the pole 1/(p2 + m2 ) isolate the creation and annihilation
operators.
• The residues at positive and negative energies correspond to creation and
annihilation operators, respectively.
• The residues do not depend on the final time t2 .
• Bound state particles correspond to similar poles at different energies.
A similar expression with opposite sign is obtained for driving the field into the
distant future
Z ∞ Z
dt d~x 3 eı̊p·x φ(x)
t1
√
ı̊ Z 0 0 †


=− 2 θ(−p )aout (−~
p) − θ(p )a out (~
p) + ... . (10.35)
p + m2
Here we have to identify

aout (~p) = Uas (0, +∞)a(~p)Uas (+∞, 0). (10.36)

LSZ Reduction. We want to express the elements of the S-matrix in terms of

time-ordered correlation functions in momentum space. Let us start with the
14
The subtlety is that the higher operatorial contributions to the field φ(x), which we never
made explicit, will have a different form in both pictures. These operators can share the energy
and momentum with the particle excitations, and thus they contribute to the residues. In fact,
they constitute the difference between ain and a, and at the end of the day, they yield the
S-matrix itself.

10.8
time-ordered expectation value15
m
Z Y n
4 ı̊pk ·xk

Y
dyk4 e−ı̊qk ·yk


Fm,n (p, q) = dxk e
k=1 k=1
 
· h0| T φ(x1 ) . . . φ(xm )φ(y1 ) . . . φ(yn ) |0i. (10.37)
Consider just the integral of the quantum operator over one of the xk
Z
X := dx4 eı̊p·x T φ(x) Y .
 
(10.38)

Now split up the time integral into three regions at the times tmin and tmax
representing the minimal and maximal times within the operator Y
Z tmin Z
X= dt d~x 3 eı̊p·x T[Y ] φ(x)
−∞
Z tmax Z
dt d~x 3 eı̊p·x T φ(x) Y
 
+
t
Z min
+∞ Z
+ dt d~x 3 eı̊p·x φ(x) T[Y ]. (10.39)
tmax

According to the results of the above consideration of resonances, the two integrals
extending to t = ±∞ produce a pole when the momentum is on shell, p2 = −m2 .
Conversely, the middle integral is finite and therefore does not produce a pole. We
can express the residue of the pole using creation and annihilation operators of in
and out particles. By assumption, the energy p0 is positive, and hence only
creation generators contribute
√
−ı̊ Z  † †

X' 2 T[Y ] a in (~
p) − a out (~
p) T[Y ] , (10.40)
p + m2
where we discard finite contributions at p2 = −m2 .
Performing this step for all ingoing particles yields
m √ !Z n
Y −ı̊ Z Y
dyk4 e−ı̊qk ·yk


Fm,n ' 2 2
k=1
pk + m k=1

· h0| T φ(y1 ) . . . φ(yn ) a†in (~p1 ) . . . a†in (~pm )|0i

 

m √ !Z n
Y −ı̊ Z Y
dyk4 e−ı̊qk ·yk


= 2 2
k=1
pk + m k=1
· h0| T φ(y1 ) . . . φ(yn ) Uas (0, −∞)a† (~p1 ) . . . a† (~pm )|0i.
 
(10.41)

Note that all outgoing creation operators a†out directly annihilate the vacuum h0|,
and the asymptotic evolution operator Uas (0, −∞) acts trivially on the vacuum by
construction.
15
The two types of momenta pk and qk are equivalent up to their overall sign. In the following,
the two will correspond to ingoing and outgoing momenta, respectively, and thus we assume that
the energy components in all pk and qk are positive.

10.9
Now we perform equivalent steps for the outgoing particles. We first insert a
trivial evolution operator Uas (+∞, 0) next to the vacuum h0|. We then use a
similar relation as above dressed by factors of Uas (+∞, 0) and Uas (0, −∞)
Z
X := dy 4 e−ı̊q·y Uas (+∞, 0) T φ(y)Y Uas (0, −∞)
 

√
ı̊ Z 
' 2 Uas (+∞, 0) T[Y ] ain (~q)Uas (0, −∞)
q + m2

− Uas (+∞, 0)aout (~q) T[Y ] Uas (0, −∞)
√
−ı̊ Z  
' 2 2
a(~q), Uas (+∞, 0) T[Y ] Uas (0, −∞) . (10.42)
q +m
Here, we obtained annihilation operators as opposed to creation operators due to
the assumption of positivity of the energy q 0 .16 For each particle this yields one
commutator of the remaining fields Uas (+∞, 0) T[Y ]Uas (0, −∞) with an
annihilation operator. After performing this step for all the outgoing particles, we
are left with the S-matrix
Uas (+∞, 0) T[1] Uas (0, −∞) = S. (10.43)
Altogether we find that the residue of Fm,n is given by an element of the S-matrix
m √ ! n √ !
Y −ı̊ Z Y −ı̊ Z
Fm,n '
k=1
p2k + m2 k=1 qk2 + m2
· h0| a(~q1 ), . . . [a(~qn ), S] . . . a† (~p1 ) . . . a† (~pm )|0i.
 
(10.44)
Here, the commutators make sure that all the operators a(~qk ) and a† (~pm ) connect
0 0
only to the S-matrix: Suppose that S is expanded in terms of the kind (a† )n am .
Then, the commutators with the a(~qk ) eliminate n factors a† from S. This only
yields a non-trivial contribution if there are sufficiently many factors and if none
remain to annihilate the vacuum h0|, so we may assume n0 = n. Finally, each of
the m0 factors a need to connect to one of the m factors a† (~pk ), hence also m0 = m
for a non-trivial contribution. Thus, none of the operators a(~qk ) will connect
directly to the a† (~pl ).17 18

10.4 S-Matrix Reconstruction

We have seen that time-ordered correlation functions have poles when the external
fields are on the mass shell of asymptotic particles. The residue of these poles is
given by the corresponding element of the scattering matrix.
16
Note that the Fourier exponent for q is opposite to the one for p.
17
S-matrix elements are often denoted by h~q1 , . . . ~qn |S|~
p1 , . . . p~m i, in the literature, but this
expression (for m > 2) also allows direct connections between the external particles which are
explicitly prohibited in the above formula.
18
In fact, the correlation function does contain pairwise contractions between external fields,
but these have a single pole (p2k + m2 )−1 rather than a double pole (p2k + m2 )−1 (ql2 + m2 )−1 .
Therefore they do not appear in the above formula.

10.10
We can therefore fully reconstruct the S-matrix from time-ordered correlation
functions.

Two-Point Correlator. In the construction of the S-matrix, the two-point

correlation function takes a special role. First, consider the above residue formula
for two legs
√ √
−ı̊ Z −ı̊ Z
F1,1 ' 2 h0|a(~q)(S − 1)a† (~p)|0i. (10.45)
p + m2 q 2 + m2
Momentum conservation implies p = q, hence the residue of a double pole at
p2 = −m2 is given by h0|a(~q)(S − 1)a† (~p)|0i. However, the S-matrix should act as
the identity on single-particle states. We conclude that there is no double pole in
F1,1 at p2 = −m2 . There is no reason to expect a double pole in the first place,
therefore the above residue statement is empty for m = n = 1.
There is nevertheless a single pole at p2 = −m2 as can be shown using the spectral
representation of the time-ordered two-point function

  
K(x − y) :=ı̊ φ(x)φ(y) =ı̊h0| T φ(x)φ(y) |0i. (10.46)
Using the spectral function ρ(s) of the interacting
√ field φ(x), it can be written in
terms of the free Feynman propagator of mass s
Z ∞
ds
K(x − y) = ρ(s) GF (s; x − y). (10.47)
0 2π
Most importantly, its momentum space representation
Z ∞
ds ρ(s) Z
K(p) = 2
= 2 + ... (10.48)
0 2π p + s −ı̊ p + m2 −ı̊
contains the parameters of the asymptotic particle: the function K(p) has an
isolated pole at the physical mass m, and its residue is the wave function
renormalisation factor Z.
Now we can nicely expand K in terms of Feynman diagrams with two external legs
and thus determine m and Z.

Amputation. The residue formula for the time-ordered correlation functions can
be inverted to a complete expression for the S-matrix
∞ Z Y n m
X d~qk 3 a† (~qk ) Y d~pk 3 a(~pk )
S =1+
m,n=2 k=1
(2π)3 2e(~q) k=1 (2π)3 2e(~p)
m n
!
Fm,n (p, q) Y p2k + m2 Y qk2 + m2
· √ √ . (10.49)
m! n! k=1 −ı̊ Z k=1 −ı̊ Z

Importantly, the poles and zeros of the latter term must be combined before the
momenta are set on shell p2k = qk2 = −m2 . The construction of this expression with
m, n ≥ 2 ensures that

10.11
 • the vacuum does not scatter, S|0i = |0i,
• single-particle states do not scatter, S|pi = |pi,
• for two or more particles, the residue of Fm,n is reproduced according to the
above formula.
It is now convenient to replace each factor (p2k + m2 ) by the inverse of the
corresponding two-point function in the construction of the S-matrix
∞ Z Y n m
X d~qk 3 a† (~qk ) Y d~pk 3 a(~pk )
S =1+
m,n=2 k=1
(2π)3 2e(~q) k=1 (2π)3 2e(~p)
m √ n √ !
Fm,n (p, q) Y Z Y Z
· . (10.50)
m! n! k=1 −ı̊K(pk ) k=1 −ı̊K(qk )

This formula has a useful interpretation in terms of Feynman graphs for Fm,n

K
K

Fn = F̃n . (10.51)
K K
K K

In the second representation we have cut the graph into a smaller (m + n)-function
F̃m,n and m + n 2-point functions K according to the rules:
• Each 2-point function connects an external leg to the (m + n)-function at the
core.
• Each 2-point function is maximal.
• The Feynman propagator that connects the 2-point function to the core is
attributed to the 2-point function.
Essentially one chops each leg of the graph as much as possible. Such a graph is
called amputated.
Now it is clear that each 2-point fragment of the graph is a Feynman graph for the
two-point function K. Moreover all these graphs have natural relative weights.
The sum of all Feynman graphs contributing to Fm,n therefore contains the sum of
all graphs contributing to K separately for each leg. What remains is a sum over
all amputated Feynman graphs at the core. This expression separates cleanly into
factors because all the weights are naturally defined
m
Y n
Y
m+n
Fm,n (p, q) = (−ı̊) F̃m,n (p, q) K(pk ) K(qk ). (10.52)
k=1 k=1

The function F̃m,n therefore is precisely what is needed for reconstruction of the
S-matrix
∞ Z Y n m √
X d~qk 3 a† (~qk ) Y d~pk 3 a(~pk ) Z m+n
S =1+ 3 2e(~ 3 2e(~
F̃m,n . (10.53)
m,n=2 k=1
(2π) q ) k=1
(2π) p ) m! n!

10.12
In other words, the elements of the S-matrix are
√ determined precisely by the sum
of amputated Feynman graphs multiplied by Z for each external leg.
Note that this also solves one problem we encountered in the earlier naive
computation of scattering matrix elements. At higher orders, one contribution was
evaluated right on a pole of a Feynman propagator. However, this contribution is
not amputated, and therefore does not actually contribute to the S-matrix.

General Picture. The general picture is as follows: Poles in the time-ordered

two-point function K(p) indicate stable asymptotic particle states.19 20
• These may be deformations of the poles in the free theory.
• They may as well be poles corresponding to bound states.
• Also poles for correlators of composite fields are permissible.
• Unbound multi-particle states forming the continuum of the spectral function
ρ(s) do not (independently) contribute to the S-matrix: they are already taken
into account in S by individual scattering of their constituent particles and
moreover the continuum leads to branch cuts rather than pole singularities.
The location p2 = −m2 of the pole defines the mass m of the particle.
Time-ordered multi-point correlation functions have poles at these locations. Their
overall residue yields the corresponding element of the S-matrix. Some comments:
• It is clear that all the external legs of the S-matrix must be exactly on shell.
• The S-matrix is completely determined in terms of time-ordered correlation
functions. No reference is made to the original formulation of the QFT, e.g. the
Lagrangian. This fact will be crucial when we go to higher perturbative orders
where Feynman diagrams have internal loops.

Feynman Rules. Let us summarise the Feynman rules for elements of the
S-matrix in φ4 theory
hq1 , . . . , qn |S|p1 , . . . , pm i. (10.54)
The matrix element is given by the sum of all graphs with certain properties. The
properties are similar to the properties of Feynman graphs for correlation functions
in momentum space, but mainly the external legs are handled differently. Let us
state the modified and additional rules:
• The graph has m ingoing and n outgoing external lines labelled by momenta pk
19
When a field has several components, the notion of pole is more subtle in the sense that the
residue of a pole is typically a matrix of non-maximal rank, e.g. ı̊p·γ + m for spinor fields. In this
case only the vectors which are not projected out correspond to asymptotic particles.
20
In practice one may not be able to distinguish an exact pole from a very narrow resonance.
One might consider such resonances at the same level as stable external particles and allow them
as legs of the S-matrix. Such an S-matrix would not rest on rigorous assumptions and therefore
not all theorems apply in a strict sense. In this regard, one should remember that in quantum
physics one has to make some separation of scales into the microscopic quantum regime and the
regime of macroscopic classical objects. Alternatively, resonances can be viewed as asymptotic
particles with a complex mass parameter.

10.13
 and qk , respectively.
pj qk (10.55)
• The external momenta must be on the mass shell, p2k = qk2 = −m2 , and must
have positive energy.
• The graph can have several connection components.
• Cutting the graph at any internal line must not split off a graph with only two
external lines (amputated graph).

Fn (10.56)

In particular, a connection component with two external lines must be a direct

connection.
The Feynman rules for evaluating a graph are the same as for correlation functions
in momentum space except:
√
• For each external line write a factor of Z instead of a Feynman propagator
−ı̊/(p2j + m2 −ı̊)
√
pj , qk −→ Z . (10.57)
• Any external line which directly connects an ingoing to an outgoing particle
contributes a factor of hqk |pl i = 2e(~pl )(2π)3 δ 3 (~pl − ~qk ). This line simply
bypasses the S-matrix21

pj qk −→ 2e(~pj ) (2π)3 δ3 (~pj − ~qk ). (10.58)

• Each connection component contributes a factor of (2π)4 δ 4 (Pk ) where Pk is the

total momentum outflow of the component. One typically defines the matrix
element M of a single connected component with the delta-function stripped
off as
hq1 , . . . , qn |S|p1 , . . . , pm i = (2π)4 δ 4 (P − Q)ı̊M. (10.59)
The other Feynman rules then yield ı̊M without the delta-function.
For quantum electrodynamics the Feynman rules for scattering matrix elements
also have to be adjusted w.r.t. the Feynman rules in momentum space, namely:
p
• For each external spinor line, write a factor of Zψ along with uα (~q), ūα (~p),
vα (~p) or v̄α (~q) depending on whether the particle is ingoing (u, v̄) or outgoing
(v, ū) and whether it is an electron (v, v̄) or a positron (u, ū)
pj p qk −→ pZ v (~q ),
αj −→ Zψ v̄αj (~pj ),
αk ψ α j j

pj p qk −→ pZ ū (~q ).
αj −→ Zψ uαj (~pj ), (10.60)
αk ψ α j j

21
Such contributions do not directly correspond to the identity within S, i.e. they are present
in S − 1, but only for at least 3 ingoing particles.

10.14
 √
• For external photon lines, write a factor of ZA along with a normalised
transverse polarisation vector εµ (~p)
pj p qk −→ pZ ε∗ (~q ).
αj −→ ZA εαj (~pj ), (10.61)
αk A α j j

10.5 Unitarity
The S-matrix is a unitary operator

S † = S −1 . (10.62)

This is an essential feature of any physical QFT. However, when deriving the
S-matrix from time-ordered correlators by means of the LSZ reduction, unitarity is
not evident at all. Therefore, we can use the property to derive some non-trivial
relations between elements of the S-matrix.

Optical Theorem. Commonly, an identity operator is removed from the

S-matrix as
S = 1 +ı̊T. (10.63)
This split is useful because for small coupling T is small. Moreover, the identity in
S is never seen in the LSZ reduction.
Unitarity SS † = 1 for the operator T is then written as the optical theorem

2 Im T = −ı̊T +ı̊T † = T T † = T † T. (10.64)

It relates the imaginary part of T to its absolute square. The latter is a quantity
we have already encountered: In the form of matrix elements it appears in the
scattering cross section. It allows to determine the total cross section of some
process in terms of the imaginary part of a matrix element.22 Alternatively, the
imaginary part of T can be obtained as a total cross section.23 The remaining real
part of T can be reconstructed from arguments of complex analyticity.
A graphical representation of the optical theorem is as follows

... q1 . . . qn
q1 qn

∞ Y T∗
X l Z
d~kj3 X k ...
1 kl
2 Im T = . (10.65)
l=2 j=1 (2π)3 2e(~kj ) pol
T
p1 ... pm
p1 . . . pm
22
In this matrix element one would choose the ingoing and outgoing momenta to be the same.
Evidently, this requires to split off the momentum-conserving delta-function first.
23
In fact one needs a generalisation of the total cross section where the ingoing particles of T
are chosen independently of the outgoing particles of T † .

10.15
The optical theorem implies that one has to integrate and sum over all allowed
degrees of freedom for the lines which connect T to T † . This is similar as for
internal lines within T and T † with one important distinction: The cut lines
originate from contracting two operators a and a† inside T and T † , respectively,

a(~p), a† (~q) = 2e(~p) (2π)3 δ 3 (~p − ~q).

 
(10.66)

Therefore the momenta associated to these lines must be on shell, p2 = −m2 , with
directed flow of energy p0 from T towards T † . Conversely, the internal lines are
integrated over all off-shell momenta.

One-Loop Unitarity. A common application of the optical theorem is to

express the imaginary part of a one-loop amplitude in terms of the total cross
section at tree level. For example, the imaginary part of the one-loop scattering
amplitude for two electrons is determined by the cross section we computed
earlier24 25

2 Im = . (10.67)

This insight is helpful because the imaginary part of a complex analytic function
determines to some extent the real part as well. In other words, unitarity can be
used to construct higher-order perturbative terms from the lower orders.
Note that there are further contributions to the imaginary part from the
annihilation of an electron with a positron

2 Im = . (10.68)

These actually happen in a different channel, i.e. in a different kinematical regime

distinguished by the signs of the particle energies.26
Another point to note is that the above unitarity relationship refers only to
continuous contributions to the imaginary part. There are also discrete
contributions which are harder to spot. These are present already at tree level
where they can be identified easily, as we shall see below.

Tree Level. It is instructive to discuss the optical theorem at tree level. At first
sight one might think that tree-level contributions to T are manifestly real because
24
Recall that the total cross section for this process is divergent implying that the imaginary
part of the loop amplitude is divergent, too. We can ignore this issue for the time being. We shall
deal with divergent loop integrals in the next section (albeit only for UV divergences; here the
divergence comes from the IR).
25
These equations are merely meant to sketch the unitarity relationship. Commonly, there are
several similar diagrams contributing to either side of the equation.
26
In our case, the channels separate the various imaginary contributions. In more general
situations at higher loops, typically many contributions overlap.

10.16
they are rational functions of the momenta and masses with real coefficients.27
Although the ı̊ prescription for Feynman propagators appears negligible, it does
have a considerable impact on the imaginary part
1 1
= 2 +ı̊πδ(p2 + m2 ). (10.69)
p2 2
+ m −ı̊ p + m2

Now in the conjugate S-matrix T † all Feynman propagators are conjugated

1
G∗F (p) = 6= GF (p). (10.70)
p2 + m2 +ı̊
When computing the imaginary part of T one therefore frequently encounters the
difference
1 1
2 2
− 2 2
= 2πı̊δ(p2 + m2 ). (10.71)
p + m −ı̊ p + m +ı̊
This identity replaces the Feynman propagator for an internal line by an on-shell
correlator for a cut line connecting T and T † . The restriction to positive energies
on the cut is a more subtle issue. It is resolved by the fact that in the sum over all
possible cuts each line appears twice, once for every direction of energy flow.
With these remarks one can show that the optical theorem holds at tree level.28
Here, we showed that at tree level T has an imaginary part concentrated at isolated
momentum configurations. However, the optical theorem is most frequently
applied at loop level where Im T has continuous (non-distributional) contributions.

27
The various prefactors of ı̊ for propagators and interaction vertices conspire to cancel out in
the appropriate definition of the imaginary part.
28
It is crucial to also take the disconnected contributions to T into account.

10.17
Quantum Field Theory I Chapter 11
ETH Zurich, 2019 HS Prof. N. Beisert
16. 12. 2019

11 Loop Corrections
Finally, we will discuss some basic loop effects and how to deal with the divergent
loop integrals we shall encounter.
We use a scalar theory with a quartic potential

L = − 21 (∂φ)2 − 12 µ2 φ2 − 61 κµφ3 − 1 2
24
κ λφ4 . (11.1)

The mass is denoted by µ and there are two dimensionless constants κ and λ. We
assume κ to be small and λ to be a freely tunable parameter.

11.1 Self Energy

First, we consider the time-ordered two-point correlator in order to isolate the pole
associated to asymptotic particles. In momentum space we know by means of
momentum conservation

ı̊ φ(p)φ(q) = (2π)4 δ 4 (p + q)K(p).


(11.2)

Leading Orders. Let us evaluate the first few orders of the function K from
Feynman diagrams.

+ + + ... (11.3)

Obviously the leading contribution is the isolated Feynman propagator

1
K (0) (p) = GF (p) = . (11.4)
p2 + µ2 −ı̊

The first correction involves one loop in the diagram. The tadpole diagram can be
safely ignored.1 The bubble diagram amounts to the following loop integral2
ı̊(−ı̊µκ)2 (−ı̊)4
K (2) (p) = ı̊I(−p2 ),
(p2 + µ2 −ı̊)2
−ı̊ d`4
Z
2 1 1 1
I(−p ) = . (11.5)
2 (2π) ` + µ −ı̊ (p − `) + µ2 −ı̊
4 2 2 2

1
We shall see later on how this can be achieved in practice.
2
The resulting integral is Lorentz invariant and can therefore be written as a function of p2
(or equivalently of −p2 ).

11.1
The resulting quadruple integral is difficult to perform, in particular due to two
different denominators. We will postpone the evaluation, and discuss the
implications for a generic function I(−p2 ).

Mass Shift. Altogether we obtain for the two-point function

1 µ2 κ2 I(−p2 )
K(p) = + + ... . (11.6)
p2 + µ2 −ı̊ (p2 + µ2 −ı̊)2

Now this expression appears to have a double pole at p2 = −µ2 . From our earlier
discussion we know that this should not happen; there should only be single poles
in the two-point function.
Taking a peek at higher loop orders, we find, among others, a sequence of iterated
loop integrals terms.

+ + + + ... (11.7)

All of these we can express easily in the form of a geometric series

∞  2 2 k
1 X µ κ I(−p2 )
K(p) = 2 + ... . (11.8)
p + µ2 −ı̊ k=0 p2 + µ2 −ı̊

Pushing convergence questions aside, we sum the series

1
K(p) = + ... . (11.9)
p2 + µ2 − µ2 κ2 I(−p2 ) −ı̊
In other words we have moved the correction term to the denominator of the
propagator.
Why should we include precisely these higher-order terms into our first order
correction? There are several reasons:
• There is nothing wrong with it.3 We just have to make sure to eventually count
every higher-order diagram with the correct weight.
• We avoid unwanted higher poles in the two-point function.4 The diagrams
contributing to the denominator are one-particle irreducible. They cannot be
cut into two parts by cutting a single internal line. Such diagrams are not
expected to produce additional poles at the mass shell.
• The inverse of the two-point function at leading order is directly related to the
action. It therefore appears somewhat more natural to expand the inverse
1/K(p) rather than the original function K(p). Our correction terms are
simply the first correction to 1/K(p).
• A mass counterterm (to be discussed below) yields the same result.
• In QFT2 we will introduce a useful functional that includes 1/K(p).
3
Note that truncation of the perturbative series to a polynomial is just as good as any other
choice that assumes some non-trivial contributions beyond the perturbative order of interest.
4
The earlier general discussion of singularities in correlation functions and S-matrix elements
revealed that the two-point function can only have simple poles.

11.2
In this form we generically expect only single poles. Particularly, the pole which
was originally at p2 + µ2 = 0 may now have shifted to a new location p2 + m2 = 0.
Assuming that this is the case, we can determine the new mass and also the
residue at the pole. For the mass, we should solve the equation

−m2 + µ2 − µ2 κ2 I(m2 ) + . . . = 0. (11.10)

The assumption of the perturbative treatment is that κ is small and that m is

approximated well by µ at this order with corrections contributing to yet higher
orders. Hence we can replace I(m2 ) by I(µ2 ) and therefore the new mass to
leading order reads
m2 = µ2 − µ2 κ2 I(µ2 ) + . . . . (11.11)
How about the residue? The expansion of the denominator at p2 + m2 = 0 reads

(p2 + m2 ) 1 + µ2 κ2 I 0 (m2 ) + . . . .


(11.12)

Altogether the pole of the two-point function corresponding to the asymptotic

particle takes the form
Z
K(p) = 2 + ... (11.13)
p + m2 −ı̊
with the field strength renormalisation

Z = 1 − µ2 κ2 I 0 (µ2 ) + . . . . (11.14)

Spectral Function. We can now extract the spectral function ρ(s) from K(p).
Recall their relationship
Z ∞
ds ρ(s)
K(p) = 2
. (11.15)
0 2π p + s −ı̊
To that end, it is most convenient to consider the imaginary part originating from
the ı̊ prescription
1 1
= +ı̊πδ(x). (11.16)
x −ı̊ x
We thus find that the imaginary part of K is directly related to the spectral
function up to a factor of 2

ρ(−p2 ) = 2 Im K(p). (11.17)

We can evaluate the imaginary part of our K starting from the naive perturbative
expansion without making use of the above geometric series. Let us illustrate this
option and check the consistency of both approaches. To extract the imaginary
part, we use the above identity for the ı̊ prescription and its derivative
1 1
2
= 2
−ı̊πδ 0 (x). (11.18)
(x −ı̊) x

11.3
We then find the spectral function5 6

ρ(s) = 2πδ(−s + µ2 ) − 2πµ2 κ2 Re I(s) δ 0 (−s + µ2 )

2µ2 κ2 Im I(s)
+ + ...
(−s + µ2 )2
= 2πδ s − µ2 + µ2 κ2 Re I(µ2 ) 1 − µ2 κ2 Re I 0 (µ2 )

2µ2 κ2 Im I(s)
+ + ...
(s − µ2 )2
2 2µ2 κ2 Im I(s)
= 2πZδ(s − m ) + + ... , (11.19)
(s − µ2 )2
which matches precisely with our expectations of a shifted delta-peak and a
continuum from Im I(s).
ρ

(11.20)

√
m µ 2m s

11.2 Loop Integral

We will now turn back to the loop integral I(−p2 ) and evaluate it.

Combining Denominators. A standard trick to proceed due to Feynman is to

combine the denominators by virtue of a new integral. For two propagators we use
the integral7 Z 1
1 dz
= 2
, z̄ := 1 − z. (11.21)
AB 0 (zA + z̄B)
An alternative is the trick used by Schwinger to convert each numerator to an
exponent where they automatically sum up
Z ∞
1
= dz e−Az . (11.22)
A 0

The loop integral then takes a form with a single squared denominator8
Z 1 Z
2 −ı̊ d`4 1
I(−p ) = dz 4
2 . (11.23)
0 32π z` + z̄(p − `)2 + µ2 −ı̊
2

5
Note that f (s)δ 0 (s − s0 ) = f (s0 )δ 0 (s − s0 ) − f 0 (s0 )δ(s − s0 ).
6
The combination of the delta-function and its derivative into a delta-function with shifted
arguments is driven by the desire to avoid derivatives of delta-functions in the spectral function
(in much the same way as the geometric series removes multiple poles). It is a non-trivial feature
of Feynman diagrams that this absorption is perfect.
7
There are similar formulas for more than two denominators and for higher powers.
8
The advantage of this expression is that the new integrand is spherically symmetric for fixed
value of z which simplifies integration drastically. The centre of the sphere, however, varies with
z.

11.4
Now the denominator has quadratic (`2 ), linear (`·p) and constant terms. We can
remove the linear term by a shift ` → ` + z̄p of the integration variable9 and thus
simplify the integral
Z 1 Z
2 −ı̊ d`4 1
I(−p ) = dz
. (11.24)
0 32π 4 `2 + z z̄p2 + µ2 −ı̊ 2

Momentum Integrals. Performing the momentum integrals is not very

difficult. Let us start with the integral over the energy component.10 The
integrand has double and single poles at
q
`0 = ± ~` 2 + z z̄p2 + µ2 −ı̊ . (11.25)

The integration contour is along the real axis and passes right between the two
poles. It decays sufficiently fast at |l0 | → ∞ so that we can close the contour by a
large semicircle in the upper or lower half of the complex plane. Either of the
single poles contributes the same residue
1
d~` 3
Z Z
2 1
I(−p ) = dz
. (11.26)
0 64π 3 ~`2 + z z̄p2 + µ2 −ı̊ 3/2

We notice that the integral merely reduces the exponent of the denominator by 1/2
and multiplies by a suitable overall factor. The remaining three spatial momentum
integrals yield a very similar result, each of them reducing the exponent by 1/2 .
The last integral in fact is logarithmically divergent for large momenta. We have
to cut it off at some bound k~lk ' Λcut , and we obtain
Z 1
2 1 z z̄p2 + µ2 −ı̊
I(−p ) = − dz log . (11.27)
32π 2 0 Λ2cut

For a large UV cutoff Λcut the integral diverges logarithmically.

Wick Rotation. Another trick which is commonly used is to rotate the

integration contour for the energy `0 from the real axis to the imaginary axis

`0 =ı̊`4E . (11.28)

This Wick rotation is permissible since physical integrands are typically perfectly
analytic in the first and third quadrants of the complex plane. For instance, poles
9
For an infinite integration domain shifting the integration variable does not change the
integral as long as it is convergent. For divergent integrals this point is subtle.
10
The integrand is constant over surfaces of constant `2 whose area is infinite. Hence there is
definitely a convergence issue in any loop integral. It is avoided by always starting with the
integral over energy.

11.5
of Feynman propagators are located slightly below the positive real axis in
quadrant four or slightly above the negative real axis in quadrant two.

Im e
II I

(11.29)
Re e
III IV

By means of the residue theorem, the value of the integral does not change by the
Wick rotation Z Z
d` F (` , `) = ı̊ d`4E F (ı̊`4E , ~`).
4 0 ~
(11.30)

Applying this rotation to our loop integral we obtain an integral over

4-dimensional Euclidean space
Z 1 Z
2 (d`E )4 1
I(−p ) = dz . (11.31)
0 32π (`E + z z̄p + µ2 −ı̊)2
4 2 2

Now the integrand depends only on k`E k, and we can use rotational symmetry to
replace the integral over three spherical angles at fixed k`E k by the volume of a
three-sphere 2π 2 k`E k3
Z Z ∞
4 2
`3E d`E F (`E ).


d`E F k`E k = 2π (11.32)
0

For our integral this implies

1 ∞
`3E
Z Z
2 d`E
I(−p ) = dz . (11.33)
0 0 16π 2 (`2E + z z̄p2 + µ2 −ı̊)2

The integral over `E is divergent again and needs to be cut off at k`E k ' Λcut . The
result is compatible with the above expression up to some minor adjustment of the
cutoff parameter Λcut .

Final Integral. Gladly the remaining integral over the Feynman parameter z
can be performed for our simple integral yielding
s s
1 p 2 + 4µ2 −ı̊ −p2
I(−p2 ) = − arctan
16π 2 −p2 p2 + 4µ2 −ı̊
1 µ2
− log . (11.34)
32π 2 Λ2cut e2

11.6
The integral is manifestly real for 0 < −p2 < 4µ2 . It is also real for −p2 < 0.
However, for −p2 > 4µ2 it develops an imaginary part
s
1 −p2 − 4µ2
Im I(−p2 ) = . (11.35)
32π −p2

It signals the opening of the two-particle creation channel at −p2 > 4µ2 .

Spectral Function. We can now write the spectral function for our model at
next-to-leading order
r
2

µ2 κ2 θ(s − 4µ2 ) s − 4µ2
ρ(s) = 2πZδ s − m + + ... . (11.36)
16π(s − µ2 )2 s

The two terms correspond to the asymptotic particle and the two-particle
continuum. Let us consider its normalisation, we find
√
2 2 3π−9
Z
ds
ρ(s) = Z + κ + ... . (11.37)
2π 288π 2
√
Since 2 3 π > 9 we see that the correction term due to the two-particle continuum
is indeed small and positive. Reassuringly the field strength renormalisation
precisely compensates for the correction term
√
2 2 0 2 0 2 2 3π−9
Z = 1 − µ κ I (µ ) + . . . , I (µ ) = . (11.38)
288π 2 µ2

11.3 Regularisation and Renormalisation

Above, we have encountered a divergent integral I(−p2 ) and in order to evaluate it
anyway, we somewhat arbitrarily introduced a momentum cutoff Λcut . Gladly the
cutoff has only a mild impact on the function I(−p2 )

d 1
I(−p2 ) = 2
. (11.39)
dΛcut 16π Λcut
In particular, the dependence on Λcut does not mix at all with the dependence on
the momenta and masses! This allowed us to extract some information from
I(−p2 ) without caring too much about the cutoff.

Regularisation Schemes. In order to extract precise information from a QFT

model at higher orders, one has to introduce a consistent regularisation scheme.
Such a scheme should make all relevant quantities finite.
There are several schemes, e.g.:

11.7
 • Cutoff. Our choice was to cut off momentum integrals at very large momenta
(UV). Similarly one could cut off very small momenta (IR). Unfortunately, a
cutoff is not easy to formulate consistently for all quantities. It is often used to
quickly derive individual leading order results.
• Pauli–Villars. Replace Feynman propagators by a difference of two propagators
1 1 1
→ 2 − 2 . (11.40)
p2 2
+ m −ı̊ p + m −ı̊ p + M 2 −ı̊
2

For large M 2 and small p2 the second propagator is suppressed. For large p2 ,
however, the two propagators almost cancel. Therefore this scheme suppresses
UV divergences. It is similar to a UV cutoff, but it can be applied universally
to Feynman diagrams.
• Point Splitting. In position space, the problem of UV divergences is related to
putting several fields at the same point in spacetime. By separating the field
insertion points in the action by a tiny amount, UV divergences can be avoided.
• Lattice. For the lattice regulator one approximates infinite spacetime by a finite
lattice. For finitely many degrees of freedom there cannot be divergences,
neither from the UV (finite spacing), nor from the IR (finite extent).
• Dimensional Regularisation. The types and degrees of divergences depend
crucially on the number of spacetime dimensions D. In the dimensional
regularisation scheme, one works in a spacetime of dimension D, where D is
taken to be an unconstrained real (or even complex) number. Observables
become functions of D, and divergences appear as poles in the D-dependence,
e.g. 1/(D − 4). Although the definition of this scheme is somewhat abstract, it
is one of the favourite ones because it works well in almost all circumstances.
• Finite Observables. Sometimes divergences can be avoided by considering
physical observables only. In our example, one could try to argue that all
observables can be deduced from I 0 (s) which
R is 0perfectly finite. The constant
term of I(s) is an integration constant of ds I (s). The divergence happens to
be located precisely in this undetermined coefficient.
In a regularised QFT, all observables are perfectly finite. However, they are not
quite what we are interested in, we are interested in observables of the original
QFT.

Renormalisation. The next step called renormalisation is to somehow absorb

the divergences consistently. To that end the most important insights are the
following:
• All physically relevant information and all observables for a QFT model are
encoded into its quantum correlation functions.
Example. The spectrum of asymptotic particles is encoded into time-ordered
two-point functions. Moreover, the scattering matrix can be derived from the
poles of higher-point functions.
• The Lagrangian and the action are devices to derive suitable correlation
functions. They are not fundamental objects, in particular their parameters

11.8
 such as masses and coupling constants are not directly observable.11
Example. The mass terms (µ) in the Lagrangian do not exactly reflect the
masses (m) of asymptotic particles.
• Correlation functions depend on the so-called bare parameters of the
Lagrangian. One should tune the parameter values such that the correlation
functions behave as expected. At the end of the day, the numerical values of
the parameters are not important.
Example. One would adjust µ and κ such that the physical mass m has the
desired physical value.
• In terms of differential geometry: The parameters of a QFT form a manifold,
where each point represents a particular physical setup. A Lagrangian
description (with a particular regularisation, renormalisation and, where
applicable, a particular gauge fixing scheme) represents a chart of the manifold.
The parameter values correspond to coordinates on this particular chart. There
is, however, no universal meaning to coordinates without reference to the
specific chart.
In this picture, renormalisation is the step to adjust the Lagrangian parameters to
the physical parameters. In the regularised and finite QFT this step is
well-defined. We express the bare Lagrangian parameters in terms of the physical
parameters, e.g.

µ = µ(m, Λcut ) = m + 12 mκ2 I(m2 , Λcut ) + . . . . (11.41)

with √
2 − π/ 3 1 m2
I(m2 , Λcut ) = − log . (11.42)
32π 2 32π 2 Λ2cut

Running Coupling. We can now remove the regulator by sending the

regularisation parameters to some appropriately chosen limit. We shall keep the
physical parameters fixed in the limit, but the resulting bare parameters may well
be divergent
µ(m, Λcut ) → ∞ as Λcut → ∞. (11.43)
This by itself is not a problem, since we attribute no meaning to µ. We just need
to keep in mind that the definition of bare parameters such as µ depends on a
scale such as Λcut . Changing the scale must be compensated by a change in the
bare parameter. In our case this dependency is governed by the equation

dµ mκ2
= + ... . (11.44)
dΛcut 16π 2 Λcut
Such an equation is often written in logarithmic form

d log µ κ2
= + ... , (11.45)
d log Λcut 16π 2
11
Some traces of the complete Lagrangian or action may remain valid in the QFT, such as the
exact equations of motion and normalisation of the interacting field, as well as Noether’s theorem.

11.9
which suggests the scaling behaviour
2 /16π 2 +...
µ ∼ m(Λcut /m)κ . (11.46)

Note that the dependence of the bare mass µ on the physical mass m has become
non-linear by quantum effects. In the quantum field theory the mass term has
acquired a so-called anomalous dimension.
A similar effect can be observed for coupling constants governing the interactions
of several particles. In quantum field theory one may find an anomalous
2
dependence on the particle momenta, e.g. λ ∼ (p/Λ)∗κ +... . It means that one will
measure a different effective coupling strength depending on which length or
energy scale the interaction is probed (e.g. the energy of the probe photon).
Naturally, one would like to define a universally valid coupling strength to appear
in the Lagrangian, e.g. by considering the limit of very high or very low energies.
However, in this limit, the coupling strength often diverges. Therefore one needs to
define the coupling constant at a particular energy scale Λ, and the value of this
coupling constant depends on Λ. This effect is called a running coupling constant.
An alternative interpretation is charge screening: The electromagnetic interactions
depend on the product of the coupling constant and the charges. Hence a running
of the coupling has the same effect as a scale-dependent effective charge of the
particles. In electrodynamics one can view the bare charges of particles as partially
screened by the surrounding quantum effects of the electromagnetic and matter
fields. The effectively perceived charge then decreases slightly with the distance.

Renormalisability. The question is whether all physical quantities remain finite

in the limit Λcut → ∞, and whether they are independent of the chosen
regularisation scheme. We can only adjust one bare parameter per physical
parameter, are there sufficiently many bare parameters to absorb all the
divergences?
• In the case of our model the answer is yes.
• QFT models where all divergences can be absorbed are called renormalisable.
• In principle, one can introduce further terms and couplings in the Lagrangian
to compensate for more and more divergences. As long as only finitely many
terms are needed to absorb the divergences at all perturbative orders, the
model is called renormalisable.
• Some models, such as general relativity, appear to require infinitely many
coupling constants to absorb all divergences. These models are called
non-renormalisable. Here one would need infinitely many measurements to
adjust infinitely many parameters, and effectively the model loses its predictive
power.
Before discussing which models are renormalisable, let us consider some technical
aspects of absorbing divergences into the coupling constants.

11.10
11.4 Counterterms
In principle, we know how to absorb divergences by writing the bare parameters of
the Lagrangian as functions of the physical parameters. Let us discuss the origin of
the divergence in detail towards absorbing the divergences into redefinitions of the
Lagrangian parameters.

Localised Divergences. To that end we investigate the integral I(−p2 ) more

closely. The overall dependence on p2 is some inverse trigonometric function.12
However, the divergent or cutoff-dependent contribution to I(−p2 ) is much simpler
d 1
I(−p2 ) = 2
. (11.47)
dΛcut 16π Λcut
It is actually independent of p2 !
This behaviour is in fact general; divergences typically couple to polynomials of
the momenta only. This statement becomes more meaningful when translated to
position space: Any polynomial of the momenta translates to a localised
distribution such as δ 4 (x) under a Fourier transformation.
In terms of the loop integral in position space we can localise the origin of the
divergence. The loop integrand diverges as r−4 when the two vertices are nearby at
a distance of r, otherwise it is perfectly finite.
(11.48)
This divergence cannot be compensated by the measure dx4 ∼ r3 dr, and the
region r ≈ 0 contributes an infinite amount to the integral. The divergence is
therefore localised in spacetime, and it can be absorbed by a suitable local term in
the Lagrangian.
(11.49)

Asymptotic Lagrangian. Indeed we see that we can absorb the divergence by

a suitable definition of the bare mass term µ = µ(m, Λcut ) in the Lagrangian. The
prescription may be somewhat confusing because the integral I(−p2 ) depends on
the mass µ, so the definition of µ apparently is implicit. For our calculations we
use the interaction picture where we decided to split the Lagrangian
L = L0 + L0,int into a free and an interaction contribution
L0 = − 21 (∂φ)2 − 12 µ2 φ2 , L0,int = − 16 κµφ3 − 1 2
24
κ λφ4 . (11.50)
However, we are not forced to do this naive split, we are free to choose any free
field as a reference. In this sense, it makes perfect sense to choose the asymptotic
field with the physical mass m as a reference
Las = − 12 (∂φ)2 − 12 m2 φ2 . (11.51)
12
We were extremely lucky to find a simple function. Typically one finds much more
complicated special functions such as polylogarithms or hyper-geometric functions. And this only
if one is lucky. Often the encountered integrals lead to functions which do not even have a name.

11.11
This will automatically position all the poles due to Feynman propagators at the
desired physical location. Now we have to add a compensating mass term to the
interaction terms13

Las,int = − 12 µ0 m2 φ2 − 61 κmφ3 − 1
24
λκ2 φ4 . (11.52)

The new term is called a counterterm. Its role is to compensate all potential mass
shifts due to loop effects. The mass of the reference field will thus conveniently be
the physical mass to all orders.14 In this picture, the loop integral I(−p2 ) is
defined directly in terms of the physical mass m instead of µ.
When including the counterterm in our example, we obtain the following one-loop
contributions15
+ (11.53)

and the following corrected two-point function

1
K(p) = + ... . (11.54)
p2 + m2 + µ0 m2 − m2 κ2 I(−p2 ) −ı̊
We impose the consistency equation that the physical mass equals the
asymptotical mass

m2 = m2 + µ0 m2 − m2 κ2 I(m2 ) + . . . , (11.55)

which is solved by µ0 = κ2 I(m2 ). Concerning the counterterm, the rule is that

whenever a bubble with two legs appears in a Feynman graph, there is a
compensating counterterm.16

:=

+ + (11.56)

It therefore makes sense to introduce a subtracted bubble integral

Isub (−p2 ) := I(−p2 ) − I(m2 ). (11.57)

13
The definition of µ0 , κ and λ has changed, but this does not matter since the values of the
bare parameters are not directly measurable.
14
Only the mass is stabilised by the counterterm. The two Feynman diagrams do not cancel
exactly because the loop integral has a complicated dependence on p2 whereas the counterterm is
a constant function.
15
The loop order does not necessarily refer to the literal number of loops in a Feynman graph.
It makes sense to also count counterterms to loop divergences as loops. The loop order commonly
refers to the order in a small coupling constant, in our case κ2 .
16
We have reintroduced the contribution from the tadpole graph which is equivalent to a mass
term with infinite mass and can therefore be absorbed entirely into a suitable redefinition of µ0 .

11.12
Obviously, this function is finite and satisfies
Isub (m2 ) = 0. (11.58)
The corrected two-point function then takes the simplified form
1
K(p) = + ... . (11.59)
p2 + m2 − m2 κ2 Isub (−p2 ) −ı̊

We can even go one step further and decide to add a counterterm for the kinetic
term
Las,int = − 21 ζ(∂φ)2 − 21 µ0 m2 φ2 − 16 κmφ3 − 24
1
λκ2 φ4 . (11.60)
This term changes the overall normalisation of the field φ(x) away from its
canonical choice, but allows us to normalise the residue of the asymptotic particle
pole in the two-point function to 1 at all orders. In this case, we can drop the field
strength renormalisation factor Z by setting it to 1.
In conclusion, the asymptotic Hamiltonian describes the canonically normalised
fields with appropriate physical masses. The interaction Hamiltonian contains all
types of allowable interactions terms. Their parameters are tuned to stabilise the
masses and normalisations and to match with physical interaction processes.

Power Counting. Can we understand under which circumstances a QFT model

is renormalisable? We can use a crude argument in terms of the mass dimensions
of interaction terms.
A Feynman diagram evaluates to an integral over a rational function of the
momenta and masses of the particles
d`D Q(`)
Z
I∼ , (11.61)
P (`)
where the denominator P (`) originates from Feynman propagators and the
numerator Q(`) from vertices and spinor traces.
• Earlier, we have discussed that UV divergences of the integral are polynomials
of the momenta.
• Furthermore, we can argue that particle masses in the denominator P can be
safely ignored for the purpose of UV divergences because they are always
dominated by the momenta.17 Contributions from the masses to divergences
therefore originate from Q alone (or equivalent parts of the integral), and will
only appear as polynomials in the final answer.
• Evidently, the coupling constants will appear as overall factors.
Altogether this implies that the structure of UV divergences of a loop integral is
given by polynomials in the momenta and the masses
dI
∈ Poly(αk , pk , mk ). (11.62)
d log Λ
17
A notable exception is the propagator of a massive vector field which has unaccompanied
factors of mass in the denominator. Such factors upset the power counting scheme, and in order
to preserve it, masses of vector field should be generated by the Higgs mechanism.

11.13
The polynomials in the momenta pk determine the appropriate local counterterm.
This has an important consequence for a Lagrangian L whose terms Lk have a
mass dimension bounded from above by the dimension D of spacetime
X
L= αk L k . (11.63)
k

The class of potential divergences is restricted by the following consideration:

• The mass dimension of the integral is essentially determined by the external
lines of the Feynman diagram. It is a fixed number usually bounded from
above by the number of spacetime dimensions D, in our case D = 4.
• Since the mass dimension of the Lagrangian L must equal D, all coupling
constants αk have a non-negative mass dimension.
• Then the overall polynomial divergence terms must have non-negative mass
dimension. The remaining mass dimension must be carried by the momenta
and masses.
• This implies that there is only a very restricted set of momentum polynomials
which can carry divergences. In other words, only few counterterms are needed
to compensate the divergences.
• These counterterms have mass a dimension bounded from above by D. The
counterterm couplings in turn have non-negative mass dimension.
• Usually, there are finitely many such terms, and therefore such models are
renormalisable.
Note that there is a crucial difference between coupling constants with positive
mass dimension and dimensionless coupling constants. As the dimension of
coupling constants accumulates, dimensionful coupling constants contribute
divergences only at low perturbative orders. When there are also dimensionless
coupling constants present, arbitrary loop orders will be divergent.
In our example, we have included all interaction terms of mass dimension bounded
from above by 4. For each divergence which can possibly arise, there is a
corresponding counterterm.
When considering Feynman diagrams which contain a divergent loop along with
some other structures, it can be shown that the loop can be replaced by a
universal counterterm corresponding to the loop to cancel the divergence. A subtle
issue at higher loops are overlapping loops, where it may not be evident which
counterterms to use. Gladly, it can be shown that this situation does not leave
behind divergences which cannot be accounted for.

11.5 Vertex Renormalisation

Let us briefly discuss how to renormalise the remaining divergences at the one-loop
level in our example.

11.14
Divergent Interactions. By power counting arguments we can derive that
divergences can appear in our model only for two-sided loops.18 Loops with three
or more sides are perfectly finite. There are only three potentially divergent terms
in our model.

(11.64)

One of them we have already discussed, the other two integrals can be made finite
by adding appropriate counterterms of the form φ3 and φ4 to the Lagrangian
Lint = − 61 κmφ3 − 241
λκ2 φ4 + Lct ,
Lct = − 21 µ0 m2 φ2 − 16 κ0 κmφ3 − 24
1 0 2 4
λκ φ

= + + . (11.65)

Suitable counterterm coefficients to make all observables finite at one loop read
µ0 = κ2 I(m2 ), κ0 = 3κ2 λI(m2 ), λ0 = 3κ2 λ2 I(m2 ). (11.66)
As discussed above, µ0 is determined by a stable physical mass at m. There is no
similar universal condition for the coefficients κ0 and λ0 ; any finite shift with
respect to the above values is permissible, it merely leads to a reparametrisation of
our model.
We have already seen that the φ2 counterterm effectively replaces the loop integral
by a finite subtracted loop integral
I(−p2 ) → Isub (−p2 ). (11.67)
Exactly the same replacement is achieved by our above choice of counterterms for
φ3 and φ4 .

(11.68)

Three-Point Function. Let us briefly consider the resulting one-loop

contributions to the three-point function.
(1) (3)
F3 = F3 + F3 + . . . ,

(1)
F3 = ,

(3)
F3 = + +

18
One-sided tadpoles can be removed entirely by suitable counterterms.

11.15
 + +

= + + . (11.69)

The latter two terms involve bubbles only for which we know the integral already.
The
R counterterms make both integrals finite.19 The first triangle integral
4 6
∼ d` /` is UV finite by itself. In fact it is the most complicated contribution
(3)
F3 = −ı̊κ3 m3 (2π)4 δ 4 (p1 + p2 + p3 )
−ı̊ d`4
Z
1
·
(2π) ` + m2 −ı̊
4 2

1 1
· (11.70)
(` + p1 )2 + m2 −ı̊ (` − p3 )2 + m2 −ı̊
and the only one which allows all three momenta to interact non-trivially.

Four-Point Function. There are many diagrams contributing to a four-point

process. Here we can merely plot all the graphs (up to permutations of the
external legs).
(2) (4)
F4 = F4 + F4 + . . . ,

(2)
F4 = + ,

(4)
F4 = + +

+ + +

+ + + . (11.71)

19
Note that because Isub (m2 ) = 0 the internal Feynman propagator which connects the bubble
to the 3-vertex is precisely cancelled, and only one Feynman propagator with associated
asymptotic particle pole remains for each leg.

11.16
Schedule of Lectures
The following table lists the locations in the text at the beginning of each lecture:

19.09. 1 [40]: 0. Overview

19.09. 2 [45]: 1. Classical and Quantum Mechanics
23.09. 1 [45]: 1.3. . . . § Example.
23.09. 2 [40]: 1.4. . . . § Negative-Energy Solutions.
26.09. 1 [40]: 2.1. . . . § Fourier Modes.
26.09. 2 [45]: 2.3. Relativistic Covariance
30.09. 1 [45]: 3. Scalar Field Quantisation
30.09. 2 [45]: 3.1. . . . § Normal Ordering.
03.10. 1 [50]: 3.4. Correlators
03.10. 2 [45]: 3.4. . . . § Correlator.
07.10. 1 [50]: 3.5. Sources
07.10. 2 [50]: 4.1. . . . § Noether’s Theorem.
10.10. 1 [45]: 4.2. . . . § Energy and Momentum.
10.10. 2 [35]: 4.3. Poincaré Symmetry
14.10. 1 [45]: 4.4. Poincaré Representations
14.10. 2 [40]: 4.5. Discrete Symmetries
17.10. 1 [45]: 5. Free Spinor Field
17.10. 2 [45]: 5.2. Poincaré Symmetry
21.10. 1 [50]: 5.3. Discrete Symmetries
21.10. 2 [40]: 5.4. . . . § Charge Conjugation.
24.10. 1 [55]: 5.6. Quantisation
24.10. 2 [35]: 5.7. Complex and Real Fields
28.10. 1 [45]: 6. Free Vector Field
28.10. 2 [40]: 6.2. . . . § Coulomb Gauge.
31.10. 1 [45]: 6.3. Particle States
31.10. 2 [35]: 6.3. . . . § Null States.
04.11. 1 [50]: 6.4. Casimir Energy
04.11. 2 [35]: 6.5. Massive Vector Field
07.11. 1 [50]: 7. Interactions
07.11. 2 [40]: 7.1. . . . § Further Interactions.
11.11. 1 [50]: 7.2. Interacting Field Operators
11.11. 2 [30]: 7.3. Perturbation Theory
14.11. 1 [40]: 8. Correlation Functions
14.11. 2 [55]: 8.2. . . . § Wick’s Theorem.
18.11. 1 [55]: 8.3. . . . § Four-Point Function.
18.11. 2 [40]: 8.4. Feynman Rules
21.11. 1 [50]: 8.4. . . . § Examples.
21.11. 2 [40]: 9. Particle Scattering

9
25.11. 1 [45]: 9.3. Electron Scattering
25.11. 2 [45]: 9.3. . . . § Sum over Polarisations.
28.11. 1 [55]: 9.4. Pair Production
28.11. 2 [40]: 10. Scattering Matrix
02.12. 1 [45]: 10.2. S-Matrix
02.12. 2 [40]: 10.3. . . . § Asymptotic States.
05.12. 1 [50]: 10.4. S-Matrix Reconstruction
05.12. 2 [25]: 10.5. Unitarity
09.12. 1 [40]: 11. Loop Corrections
09.12. 2 [40]: 11.2. Loop Integral
12.12. 1 [45]: 11.3. Regularisation and Renormalisation
12.12. 2 [40]: 11.3. . . . § Running Coupling.
16.12. 1 [60]: 11.4. . . . § Asymptotic Lagrangian.

10