Physics 215A: Particles and Fields

Fall 2017

Lecturer: McGreevy
These lecture notes live here. Please email corrections to mcgreevy at physics dot
ucsd dot edu.

Last updated: 2017/12/14, 13:16:23

1
Contents
0.1 Introductory remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
0.2 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1 From particles to fields to particles again 7

1.1 Quantum sound: Phonons . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Scalar field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Quantum light: Photons . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4 Lagrangian field theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5 Casimir effect: vacuum energy is real . . . . . . . . . . . . . . . . . . . 30

2 The path integral makes some things easier 34

2.1 Fields mediate forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Euclidean path integral and Wick rotation . . . . . . . . . . . . . . . . 38
2.3 Feynman diagrams from the path integral . . . . . . . . . . . . . . . . 42

3 Lorentz invariance and causality 46

3.1 Causality and antiparticles . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 Propagators, Green’s functions, contour integrals . . . . . . . . . . . . 52
3.3 Interlude: where is 1-particle quantum mechanics? . . . . . . . . . . . . 58

4 Interactions, not too strong 61

4.1 Where does the time dependence go? . . . . . . . . . . . . . . . . . . . 61
4.2 S-matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3 Time-ordered equals normal-ordered plus contractions . . . . . . . . . . 67
4.4 Time-ordered correlation functions by diagrams . . . . . . . . . . . . . 70
4.5 Interlude: old-fashioned perturbation theory . . . . . . . . . . . . . . . 80
4.6 From correlation functions to the S matrix . . . . . . . . . . . . . . . . 83
4.7 From the S-matrix to observable physics . . . . . . . . . . . . . . . . . 91

5 Spinor fields and fermions 97

5.1 More on symmetries in QFT . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Representations of the Lorentz group on fields . . . . . . . . . . . . . . 103
5.3 Spinor lagrangians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4 Free particle solutions of spinor wave equations . . . . . . . . . . . . . 116
5.5 Quantum spinor fields . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

6 Quantum electrodynamics 126

6.1 Vector fields, quickly . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.2 Feynman rules for QED . . . . . . . . . . . . . . . . . . . . . . . . . . 131

2
 6.3 QED processes at leading order . . . . . . . . . . . . . . . . . . . . . . 136

A Why care about GR ? 145

B More about the non-perturbative proof of the Ward identity 147

3
0.1 Introductory remarks

Quantum field theory (QFT) is the quantum mechanics of extensive degrees of freedom.
What I mean by this is that at each point of space, there’s some stuff that can wiggle.
It’s not surprising that QFT is so useful, since this situation happens all over the
place. Some examples of ‘stuff’ are: the atoms in a solid, or the electrons in those
atoms, or the spins of those electrons. A less obvious, but more visible, example is the
electromagnetic field, even in vacuum. More examples are provided by other excitations
of the vacuum, and it will be our job here to understand those very electrons and atoms
that make up a solid in these terms. The vacuum has other less-long-lasting excitations
which are described by the Standard Model of particle physics.
Some examples of QFT are Lorentz invariant (‘relativistic’). That’s a nice simplifi-
cation when it happens. Indeed this seems to happen in particle physics. We’re going
to focus on this case for most of this quarter. Still I would like to emphasize: though
some of the most successful applications of QFT are in the domain of high energy
particle physics, this is not a class on that subject, and I will look for opportunities to
emphasize the universality of QFT.
A consequence of relativity is that the number of particles isn’t fixed. That is: there
are processes where the number of particles changes in time. This is a crucial point
of departure for QFT, worth emphasizing, so let me stop and emphasize it. (Later on
we’ll understand in what sense it’s a necessary consequence of Lorentz symmetry. The
converse is false: particle production can happen without relativity.)
Single-particle QM. In classes with the title ‘Quantum Mechanics’, we generally
study quantum systems where the Hilbert space H1 holds states of a single particle (or
sometimes a fixed small number of them), which are rays in H1 .
The observables of such a system are represented by hermitian operators acting on
H1 . For example, the particle has a position ~x and a momentum ~p each of which is a
d-vector of operators (for a particle in d space dimensions). The particle could be, for
example, an electron (in which case it also has an inherent two-valuedness called spin)
or a photon (in which case it also has an inherent two-valuedness called polarization).
Time evolution is generated by a Hamiltonian H which is made from the position
and momentum (and whatever internal degrees of freedom it has), in the sense that
i~∂t |ψi = H |ψi. Finally, the fourth (most ersatz) axiom regards measurement: when
measuring an observable A in a state |ψi ∈ H, we should decompose the state in the
P
eigenbasis of A: A |ai = a |ai, |ψi = a |aiha|ψi; the probability to get the answer a
is | ha|ψi |2 .

4
 By the way: The components of the state vector in the position basis h~x|ψi = ψ(~x)
is a function of space, the wavefunction. This looks like a field. It is not what we mean
by a field in QFT. Meaningless phrases like ‘second quantization’ may conspire to try
to confuse you about this.

Now suppose you want to describe quantumly the emission of a photon from an
excited electron in an atom. Surely this is something for which we need QM. But
it’s not something that can happen within H1 , since the number of particles changes
during the process. How do you do it?
In the first section of this course we’ll follow an organic route to discovering an
answer to this question. This will have the advantage of making it manifest that the
four axioms of QM just reviewed are still true in QFT. It will de-emphasize the role of
Lorentz symmetry; in fact it will explicitly break it. It will emerge on its own!
‘Divergences’. Another intrinsic and famous feature of QFT discernible from the
definition I gave above is its flirtation with infinity. I said that there is ‘stuff at each
point of space’; how much stuff is that? Well, there are two senses in which ‘the number
of points of space’ is infinite: (1) space can go on forever (the infrared (IR)), and (2)
in the continuum, in between any two points of space are more points (the ultraviolet
(UV)). The former may be familiar from statistical mechanics, where it is associated
with the thermodynamic limit, which is where interesting things happen. For our own
safety, we’ll begin our discussion in a padded room, protected on both sides from the
terrors of the infinite.

Sources and acknowledgement. The material in these notes is collected from

many places, among which I should mention in particular the following:
Peskin and Schroeder, An introduction to quantum field theory (Wiley)
Zee, Quantum Field Theory in a Nutshell (Princeton, 2d Edition)
Le Bellac, Quantum Physics (Cambridge)
Schwartz, Quantum field theory and the standard model (Cambridge)
David Tong’s lecture notes
Many other bits of wisdom come from the Berkeley QFT course of Prof. Lawrence
Hall.

5
0.2 Conventions

Following most QFT books, I am going to use the + − −− signature convention for
the Minkowski metric. I am used to the other convention, where time is the weird one,
so I’ll need your help checking my signs. More explicitly, denoting a small spacetime
displacement as dxµ ≡ (dt, d~x)µ , the Lorentz-invariant distance is:
 
+1 0 0 0
 0 −1 0 0 
 
ds2 = +dt2 − d~x · d~x = ηµν dxµ dxν with η µν = ηµν =  .
 0 0 −1 0 

0 0 0 −1 µν
 µ
(spacelike is negative). We will also write ∂µ ≡ ∂ ~x
= ∂t , ∇ , and ∂ µ ≡ η µν ∂ν . I’ll
∂xµ
use µ, ν... for Lorentz indices, and i, k, ... for spatial indices.
The convention that repeated indices are summed is always in effect unless otherwise
indicated.
A consequence of the fact that english and math are written from left to right is
that time goes to the left.
h
A useful generalization of the shorthand ~ ≡ 2π
is

dk
d̄k ≡ .
2π
d
I will also write /δ (q) ≡ (2π)d δ (d) (q). I will try to be consistent about writing Fourier
transforms as
dd k ikx ˜
Z Z
e f (k) ≡ d̄d k eikx f˜(k) ≡ f (x).
(2π)d
IFF ≡ if and only if.
RHS ≡ right-hand side. LHS ≡ left-hand side. BHS ≡ both-hand side.
IBP ≡ integration by parts. WLOG ≡ without loss of generality.
+O(xn ) ≡ plus terms which go like xn (and higher powers) when x is small.
+h.c. ≡ plus hermitian conjugate.
We work in units where ~ and the speed of light, c, are equal to one unless otherwise
noted. When I say ‘Peskin’ I mean ‘Peskin & Schroeder’.
Please tell me if you find typos or errors or violations of the rules above.

6
1 From particles to fields to particles again
In modern
physics, when
Here is a way to discover QFT starting with some prosaic ingredients. Besides the I hear the
word ‘particle’
advantages mentioned above, it will allows us to check that we are on the right track I reach for
my Fourier
with simple experiments. analyzer...
– J. Ziman,
Electrons and Phonons.

1.1 Quantum sound: Phonons

Let’s think about a crystalline solid. The specific heat of solids (how much do you
have to heat it up to change its internal energy by a given amount) was a mystery
before QM. The first decent (QM) model was due to Einstein, where he supposed that
each atom is a (independent) quantum harmonic oscillator with frequency ω. This
correctly predicts that the specific heat decreases as the temperature is lowered, but
is very crude. Obviously the atoms interact: that’s why they arrange themselves in a
nice crystal pattern, and that’s why there are sound waves, as we will see. By treating
the elasticity of the solid quantum mechanically, we are going to discover quantum
field theory. One immediate benefit of this will be a framework for quantum mechanics
where particles can be created and annihilated.
As a more accurate toy model of a one-dimensional crystalline solid, let’s consider
a linear chain of particles of mass m, each connected to its neighbors by springs with
spring constant κ. When in equilibrium, the masses form a regular one-dimensional
crystal lattice (equally spaced mass points). Now let qn denote the displacement of the
nth mass from its equilibrium position xn and let pn be the corresponding momentum.
Assume there are N masses and (for simplicity) impose periodic boundary conditions:
qn+N = qn . The equilibrium positions themselves are

xn = na, n = 1, 2...N

where a is the lattice spacing. The Hamiltonian for the collection of particles is:
N  2 
X p n1 2
H= + κ (qn − qn−1 ) + λq4 . (1.1)
n=1
2m 2

Notice that this system is an ordinary QM system, made of particles. In particular,

the whole story below will take place within the fixed Hilbert space of the positions of
the N particles.
I’ve included a token anharmonic term λq4 to remind us that we are leaving stuff
out; for example we might worry whether we could use this model to describe melting.
Now set λ = 0. (It will be a little while before we turn back on the interactions

7
resulting from nonzero λ; bear with me.) This hamiltonian above describes a collection
of coupled harmonic oscillators, with a matrix of spring constants V = kab qa qb . If we
diagonalize the matrix of spring constants, we will have a description in terms of
decoupled oscillators, called normal modes.
Since our system has (discrete) translation invariance, these modes are labelled by
a wavenumber k 1 :
N N
1 X ikxn 1 X ikxn
qk = √ e qn , pk = √ e pn ,
N n=1 N n=1

(Notice that in the previous expressions I didn’t use boldface; that’s because this step is
really just classical physics. Note the awkward (but in field theory, inevitable) fact that
we’ll have (field) momentum operators pk labelled by a wavenumber aka momentum.)
The nice thing about the fourier kernel is that it diagonalizes the translation operator
T defined by Tf (x) = f (x + a):

Teikx ≡ eik(x+a) = eika eikx .

Regulators: Because N is finite, k takes discrete values (1 = eikN a ); this is a

long-wavelength “IR” property. Because of the lattice structure, k is periodic (only
eikan , n ∈ Z appears): k ≡ k + 2π/a; this is a short-distance “UV” property. The range
of k can be taken to be
2π(N − 1)
0≤k≤ .
Na
Because of the periodicity in k, we can equivalently label the set of wavenumbers by2 :
2π π π
0<k≤ or − <k≤ . (1.3)
a a a

Summary: Because the system is in a box (periodic), k-space is discrete. Because

the system is on a lattice, k-space is periodic. There are N oscillator modes altogether.

1
The inverse transformation is:
2π/a 2π/a
1 X −ikxn 1 X −ikxn
qn = √ e qk , pn = √ e pk . (1.2)
N k>0 N k>0

2
This range of independent values of the wavenumber in a lattice model is called the Brillouin
zone. There is some convention for choosing a fundamental domain which prefers the last one but I
haven’t found a reason to care about this.

8
 The whole hamiltonian is a bunch of decoupled oscillators, labelled by these funny
wave numbers:
X  pk p−k 1 
2
H= + mωk qk q−k (1.4)
k
2m 2
where the frequency of the mode labelled k is
r
κ |k|a
ωk ≡ 2 sin . (1.5)
m 2

Why might we care about this frequency? For one thing, consider the Heisenberg
equation of motion for the deviation of one spring:
pn
i∂t qn = [qn , H] = , i∂t pn = [pn , H]
m
Combining these gives:

mq̈n = −κ ((qn − qn−1 ) − (qn − qn+1 )) = −κ (2qn − qn−1 − qn+1 ) .

In terms of the fourier-mode operators:

mq̈k = −κ (2 − 2 cos ka) qk .

P −iωt
Plugging in a fourier ansatz in time qk (t) = ωe qk,ω
turns this into an algebraic equation which says ω = ωk2 =
2
2κ

2 |k|a
m
sin 2 for the allowed modes. We see that (the classical
version of) this system describes waves:
k1/a
0 = ω 2 − ωk2 qk,ω ω 2 − vs2 k 2 qk,ω .



'

The result for small k is the fourier transform of the wave equation:

∂t2 − vs2 ∂x2 q(x, t) = 0 .

(1.6)

vs is the speed of propagation of the waves, in this case the speed of sound. Comparing
to the dispersion relation (1.5), we have found
r
∂ωk κ
vs = |k→0 = a .
∂k m

The wave looks something like this:

9
 So the story I am telling is a quantization of sound waves. Soon we will quantize
electromagnetic (EM) waves, too.
QM. So far the fact that quantumly [qn , pn0 ] = i~δnn0 1 hasn’t really mattered in
our analysis (go back and check – we could have derived the wave equation classically)3 .
3
In case you are rusty, or forget the numerical factors like I do, here is a concise summary of the
operator solution of the quantum harmonic oscillator:

p2
 
1 2 2 ~ω 2 2

† 1
H= + mω q = P + Q = ~ω a a +
2m 2 2 2

with
1 1
a ≡ √ (Q + iP) , a† ≡ √ (Q − iP) .
2 2
Here I’ve defined these new operators to hide the annoying factors:
 mω 1/2  1/2
1
Q≡ q, P ≡ p.
~ m~ω

[q, p] = i~1 =⇒ [a, a† ] = 1.

The number operator N ≡ a† a satisfies

[N, a] = −a, [N, a† ] = +a† .

So a and a† are lowering and raising operators for the number operator. The eigenvalues of the number
operator have to be positive, since

0 ≤ || a |ni ||2 = hn| a† a |ni = hn| N |ni = n hn|ni

which means that for n = 0 we have a |n = 0i = 0. If it isn’t zero (i.e. if n ≥ 1), a |ni is also an
eigenvector of N with eigenvalue n − 1. It has to stop somewhere! So the eigenstates of N (and hence
of H = ~ω N + 21 are



n
|0i , |1i ≡ a† |0i , ..., |ni = cn a† |0i ...
where we must choose cn to normalize these states. The answer which gives hn|ni = 1 is cn = √1 .
n!

10
For the Fourier modes, this implies the commutator
X X
[qk , pk0 ] = Ukn Uk0 n0 [qn , pn0 ] = i~1 Ukn Uk0 n = i~δk,−k0 1.
n,n0 n

(In the previous expression I called Ukn = √1 eikxn the unitary matrix realizing the
N
discrete Fourier kernel.)
To make the final step to decouple the modes with k and −k, introduce the anni-
hilation and creation operators4
r r
~   1 ~mωk  
For k 6= 0: qk = ak + a†−k , pk = ak − a†−k .
2mωk i 2
They satisfy
[ak , a†k0 ] = δkk0 1.
In terms of these, the hamiltonian is
p20
 
X 1
H= ~ωk a†k ak + +
k
2 2m

– it is a sum of decoupled oscillators, and a free particle describing the center-of-mass.

The discovery of Fock space. The ground state satisfies ak |0i = 0 for all k (and
has eigenvalue zero for the center-of-mass momentum, p0 = 0). The first excitation
above the ground state

a†k |0i ∝ |one phonon with momentum ~ki (1.7)

has energy ~ωk . (This is an eigenstate because [Nk , a†k ] = a†k ], where Nk = a†k ak .) It is
called a phonon with momentum ~k.5 This is what in undergrad QM we would have
called “|ki”; we can make a state with one phonon in a position eigenstate by taking
superpositions:
X X
|one phonon at position xi = eikx |one phonon with momentum ~ki ∼ eikx a†k |0i .
k k

This is the state which in single-particle QM we would have called “|xi”.6

4
You might notice that when k = 0, ωk = 0. This last step applies to the modes with ωk 6= 0,
hence k 6= 0. The ‘zero-mode’ must be treated specially. It is neglected in many treatments of this
topic but actually as a lot of physics in it. If you are curious see this discussion, page 11.
5
I put ‘proportional to’ rather than ‘equal’ in (1.7) because there can be a k-dependent normaliza-
tion factor. We’ll see soon that Lorentz symmetry prefers a particular normalization here which we
will adopt.
6
For now, the fact that |ki and |xi are related by a Fourier transform will have to serve for evidence
that ~k is the momentum of the particle we’ve just discovered. Later we will show that indeed the
state a†k |0i has momentum ~k above the groundstate.

11
 The number operator (of the SHO with label k) Nk ≡ a†k ak counts the number of
phonons with momentum k. The ground state is the state with no phonons – for this
reason we could also call it the ‘vacuum’. We can also make a state with two phonons:
|k, k 0 i = a†k a†k0 |0i
whose energy is E = ωk + ωk0 . Note that all these states have non-negative energy.
So this construction allows us to describe situations where the number of particles
P
N = k Nk can vary! That is, we can now describe dynamical processes in which the
number of particles changes. Let me emphasize: In QM, we would describe the hilbert
space of two (distinguishable) particles as a tensor product of the hilbert space of each.
How can we act with an operator which enlarges the hilbert space?? We just figured
out how to do it.
We can specify basis states for this Hilbert space
 nk1  nk2
a†k1 a†k2 · · · |0i = |{nk1 , nk2 , ...}i
by a collection of occupation numbers nk , eigenvalues of the number operator for each
normal mode (and the center-of-mass momentum p0 ).
Notice that in this description it is manifest that a phonon has no identity. We
only keep track of how many of them there are and what is their wavenumber. They
cannot be distinguished. Also notice that we can have as many we want in the same
mode – nk can be any non-negative integer. These are an example of bosons. (Further
evidence for this is that a†k a†k0 = a†k0 a†k .)
[End of Lecture 1]
Lessons from the hard work of the first lecture:

• Starting from a collection of particles, we chained them together, and made a

field; treating this system quantumly, we found a new set of particles. The new
particles (the normal modes) are collective excitations: their properties can be
very different from those of the constituent particles. For example, the constituent
particles are distinguishable by their locations, but phonons are indistinguishable
from each other. The state of such indistinguishable particles is determined
merely by specifying a collection of positions (or momenta) and of spin states –
we don’t need to say which is which (and in fact, we cannot).

• Notice that there are some energies where there aren’t any phonon states. In
particular, the function (1.5) has a maximum. More generally, in a system with
discrete translation invariance, there are bands of allowed energies. In the con-
tinuum limit (roughly a → 0 with the correct quantities held fixed), to which we
devolve soon, this maximum (which occurs at k = πa ) goes off to the sky.

12
 • Lorentz invariance can emerge. The dispersion relation for the long-wavelength
(ka  1) sound mode was ω 2 = v 2~k 2 . This is the fourier transform of

~ 2 )φ(x) = ∂µ ∂ µ φ(x) = 0,
(∂t2 − v 2 ∇

s a wave equation which has Lorentz symmetry (if v is the speed appearing in the
Minkowski metric). We had to ignore the O(a4 k 4 ) terms in the long-wavelength
expansion of the dispersion relation, 2(1 − cos(ka)). The lattice breaks Lorentz
symmetry, but its effects go away for ka  1. This point might make you think
that the Lorentz symmetry which is so precious to us in particle physics could
emerge in a similar way, but with a much smaller a than the lattice spacing in
solids. There are strong constraints on how small this can be (e.g. this well-
appreciated paper) so it is very useful to treat it as a fundamental principle.

Phonons are real: heat capacity of (insulating) solids. The simplest demon-
stration that phonons are real is the dramatic decrease at low temperatures of the heat
capacity of insulating solids. At high temperatures, the equipartition theorem of clas-
sical thermodynamics correctly predicts that the energy of the solid from the lattice
vibrations should be T times the number of atoms, so the pacity, CV = ∂T E should be
independent of T .
At low temperatures T < ΘD , this is wrong. ΘD is the tem-
perature scale associated with the frequencies of the lattice
vibrations (say the maximum of the curve ωk above). The
resolution lies in the thermal energy of a quantum harmonic
oscillator for T < ω, the energy goes to a constant 21 ~ω:

So the heat capacity (the slope of this curve) goes to zero as T → 0.

Phonons are real: the Mössbauer effect. Here is another dramatic conse-
quence of the quantization of the lattice vibrations of solids, known as the Mössbauer
effect, first described in words. The nuclei of the atoms in a solid have various en-
ergy levels; when hit with a γ-ray photon, these nuclei can experience transitions from
the groundstate to some excited energy level. If an excited nucleus somewhere in
the lattice gets hit by a very energetic photon (a γ-ray) of some very specific energy
Eγ = ∆E ≡ Eexcited − E0 , the nucleus can absorb and re-emit that photon. The
resulting sharp resonant absorption lines at Eγ = ∆E are indeed observed.
This sounds simple, but here is a mystery about this: Consider a nucleus alone
in space in the excited state, after it gets hit by a photon. The photon carried a
momentum pγ = Eγ /c. Momentum is conserved, and it must be made up by some
recoil of the absorbing nucleus. When it emits a photon again, it needn’t do so in

13
the same direction. This means that the nucleus remains in motion with momentum
∆~p = p~1 − p~2 . But if some of its energy ∆E = Eexcited − E0 goes to kinetic energy of
recoil, not all of that energy can go to the final photon, and the emitted photon energy
2 p)2 2
will be less than Eγ by Erecoil = ∆p2M
max
. This can be as big as Erecoil = (2~
2M
= (2Eγ/c)
2M
(in
the case of scattering by angle π). So instead of a sharp absorption line, it seems that
2
we should see a broad bump of width (Eγ/c) M
. But we do see a sharp line!
The solution of the puzzle is phonons: for a nucleus in a lattice, its recoil means
that the springs are stretched – it must excite a lattice vibration, it must create some
phonons. But there is a nonzero probability for it to create zero phonons. In this case,
the momentum conservation is made up by an acceleration of the whole solid, which is
very massive (its mass is N M where N is the number of nuclei!), and therefore does
p2
not recoil very much at all (it loses only energy 2NγM ).

This allows for very sharp resonance lines. In turn, this effect
has allowed for some very high-precision measurements. The
different widths in these cartoon absorption spectra don’t
do justice to the relative factor of N . An essentially similar
effect makes it possible to get precise peaks from scattering
of X-rays off of a solid (Bragg scattering) – there is a finite
amplitude for the scattering to occur without exciting any
phonons.

This is actually a remarkable thing: although solids seem ordinary to us because

we encounter them frequently, the rigidity of solids is a quantum mechanical emergent
phenomenon. You can elastically scatter photons off of a solid only because the atoms
making up the solid participate in this collective behavior wherein the whole solid acts
like a single quantum object!

14
 Towards scalar field theory. It is worthwhile to put together the final relation
between the ‘position operator’ and the phonon annihilation and creation operators:
s
~ X 1  ikxn  1
qn = √ e ak + e−ikxn a†k + √ q0 (1.8)
2µ k ωk N

and the corresponding relation for its canonical conjugate momentum

s
µ ~ X √  ikxn  1
pn = ωk e ak − e−ikxn a†k + √ p0 .
i 2µ k N

The items in red are the ways in which p and q differ; they can all be understood from
the relation p = µq̇ as you will see on the homework. These expressions are formally
identical to the formulae we’ll find below expressing a scalar field in terms of creation
and annihilation operators (such as Peskin eqns. (2.25) and (2.26) ). The stray factors
of µ arise because we didn’t ‘canonically normalize’ our fields and absorb the ms into
√
the field, e.g. defining φ ≡ mq would get rid of them. The other difference is because
we still have IR and UV regulators in place, so we have sums instead of integrals.

Path integral reminder in a box.

If we use the path integral description, some of these things (in particular the
continuum, sound-wave limit) are more obvious-seeming.
Let’s remind ourselves how the path integral formulation of QM works for a particle
p2
in one dimension with H = 2m +V (q). The basic statement is the following formula for
the propagator – the amplitude to propagate from position eigenstate |q0 i to position
eigenstate |qi during a time interval t is
Z q(t)=q Rt 1 2
hq| e −iHt
|q0 i = [dq]ei 0 dt ( 2 q̇ −V (q)) .
q(0)=q0
Q Mt
Here [dq] ≡ N l=1 dq(tl ) – the path integral measure is defined by a limiting procedure
t
(Mt ≡ ∆t → ∞, ∆t → 0, t fixed), and N is a normalization factor that always drops
out of physical quantities so I don’t need to tell you what it is.
Recall that the key step in the derivation of this statement is the evaluation of the
propagator for an infinitesimal time step:
p2
hq2 | e−iH∆t |q1 i = hq2 | e−i∆t 2m e−i∆tV (q) |q2 i + O(∆t2 ) .

An integral expression for this can be obtained by inserting resolutions of the identity
Z  Z 
2
1=1 = dp |pi hp| dq |qi hq|

15
in between the two exponentials. For a more extensive reminder, please see §2.4 of this
document.

1.2 Scalar field theory

Scalar field theory in one dimension. [Zee §1.3] The path integral for our collection
of oscillators is Z
Z = [dq1 · · · dqN ]eiS[q]
R P 1
R
with S[q] = dt n 2
mn q̇n2 − V ({q}) ≡ dtL(q, q̇). The potential is V ({q}) =
P 1 2
n 2 κ (qn+1 − qn ) . Now suppose we have poor eyesight and can’t resolve the indi-
vidual atoms in the chain; rather we’re only interested in the long-wavelength (small-
wavenumber) physics. So let’s try to take the continuum limit a → 0, N → ∞.
Basically the only thing we need is to think of qn = q(x = na) as defining a smooth

function: [Note that the continuum field is often

called φ(x) instead of q(x) for some reason. At least the letters q(x) and φ(x) look
similar.]
We’ll use
X Z
2 2 2
(qn − qn−1 ) ' a (∂x q) |x=na , a f (qn ) ' dxf (q(x)).
n

The path integral becomes: Z

Z= [dq]eiS[q]

with [dq] now representing an integral over all configurations q(t, x) (defined by this
limit) and
Z Z Z Z
1 2 2 2 2 4


S[q] = dt dx µ (∂t q) − µvs (∂x q) − rq − uq − ... ≡ dt dxL
2
where I’ve introduced some parameters µ, vs , r, u determined from m, κ, a... in some
ways that we needn’t worry about, except to say that they are finite in the continuum
limit. The · · · includes terms like a4 (∂x q)4 which are small when k  a1 , so we
ignore them. L is the Lagrangian density whose integral over space is the Lagrangian
R
L = dxL.
The equation of motion is the stationary phase condition,
δS
0= = −µq̈ + µvs2 ∂x2 q − rq − 2uq 3 − ...
δq(x, t)

16
In this expression I have written a functional derivative; with our lattice regulator, it is
simply a(n extremely useful) shorthand notation for the collection of partial derivatives
∂
∂qn
. 7
From the phonon problem, we automatically found r = u = 0, and the equation
of motion is just the wave equation (1.6). This happened because of the symmetry
qn → qn + . This is the operation that translates the whole crystal, It guarantees
low-energy phonons near k = 0 because it means q(x) can only appear in S via its
derivatives. (This is a general property of Goldstone modes; more on this later.)
The following will be quite useful for our subsequent discussion of quantum light.
We can construct a hamiltonian from this action by defining a canonical field-momentum
π(x) = ∂∂L
tq
= µ∂t q and doing the Legendre transformation:

π(x)2
X Z Z  
2 2 2 4
H= (pn q̇n − Ln ) = dx (π(x)q̇(x) − L) = dx + µvs (∂x q(x)) + rq + uq + ... .
n
2µ
(1.10)
Note that I suppress the dependence of all the fields on t just so it doesn’t get ugly,
not because it isn’t there. Also, I emphasize that the position along the chain x here
is just a label on the fields, not a degree of freedom or a quantum operator.
The field q is called a scalar field because it doesn’t have any indices decorating it.
This is to be distinguished from the Maxwell field, which is a vector field, and which
we’ll discuss next. (Note that vibrations of a crystal in three dimensions actually do
involve vector indices! We omit this complication.)
The lattice spacing a and the size of the box N a in the discussion above are playing
very specific roles in regularizing our 1-dimensional scalar field theory. The lattice
spacing a implies a maximum wavenumber or shortest wavelength and so is called an
“ultraviolet (UV) cutoff”, because the UV is the short-wavelength end of the visible
7
Functional derivatives will be very useful to us. The definition is

δφ(x)
= δ(x − y) (1.9)
δφ(y)

plus the Liebniz properties (linearity, product rule). More prosaically, they are just partial derivatives,
if we define a collection of values of the independent variable {xn } to regard as grid points, and let

φn ≡ φ(xn )

so that (1.9) is just

∂φn
= δnm .
∂φm
If you are not yet comfortable with the machinery of functional derivatives, please work through pages
2-28 through 2-30 of this document now.

17
light spectrum. The size of the box N a implies a maximum wavelength mode which
fits in the box and so is called an “infrared (IR) cutoff”.
If we also take the infinite volume limit, then the sums over k become integrals. In
this limit we can make the replacement
Z
1 X
d̄d k, Ld δkk0 (2π)d δ (d) (k − k 0 ).
Ld k

A check of the normalization factors comes from combining these two rules
X Z
1= δk,k0 = d̄d k(2π)d δ (d) (k − k 0 ).
k

Continuum (free) scalar field theory in d + 1 dimensions. These continuum

expressions are easy to generalize to scalar field theory in any number of dimensions.
Let’s do this directly in infinite volume and set µ = 1 by rescaling fields. The action is
Z  
d 1 2 1 2~ ~
S[φ] = d xdt φ̇ − vs ∇φ · ∇φ − V (φ) . (1.11)
2 2

This is almost what we would have found for the long-wavelength (ka  1) description
of a d-dimensional lattice of masses on springs, like a mattress (except that there would
have been one φ for each direction in which the atoms can wiggle). The equation of
motion is
δS[φ]
0= = −∂t2 φ + vs2 ∇2 φ − V 0 (φ). (1.12)
δφ(x)
For the harmonic case V (φ) = 21 m2 φ2 we know what we’re doing, and (1.12) is called
the Klein-Gordon equation,
0 = ∂µ ∂ µ + m2 φ.


(1.13)
(Notice that I’ve set vs = c = 1 here, and this is where we have committed to a choice
of signature convention; take a look at the conventions page §0.2.). In relativistic
notation, the Lagrangian density is just L = 21 (∂µ φ∂ µ φ − m2 φ2 ). This describes free
continuum real massive relativistic scalar quantum field theory. (Match the adjectives
to the associated features of the lagrangian; collect them all!)
The canonical momentum is π = ∂L ∂ φ̇
= φ̇ and the Hamiltonian (which we can
instantly promote to a quantum operator by using boldface symbols) is then

π(x)2 1 2  ~
Z   1 
H= d x d
+ vs ∇φ · ∇φ ~ + mφ .2 2
2 2 2
Note that all these terms are positive.

18
 A translation invariant linear problem is solved by Fourier transforms: φ(x) =
~ ~
d̄ k e−ik·~x φk , and π(x) = d̄d k e−ik·~x πk , this is
d
R R
Z  
d 1 1 2 2 2


H = d̄ k πk π−k + v k + m φk φ−k
2 2 s

where k 2 = (−i~k) · (i~k) = ~k · ~k. Just as in (1.4), this is merely a sum of decoupled
oscillators, except for the coupling between wavenumbers k and −k. Comparing with
(1.4), we can read off the normal mode frequencies, aka the dispersion relation:

ωk2 = vs2 k 2 + m2 .
~
Notice that this is also the condition for a Fourier mode eik·~x−iωt to solve the Klein-
Gordon equation (1.13).
We can decouple the modes with wavenumber k and −k as above by introducing
the ladder operators8
r r
~  †
 1 ~ωk  
φk ≡ ak + a−k , πk ≡ ak − a†−k , [ak , a†k0 ] = (2π)d δ (d) (k − k 0 ).
2ωk i 2
Their commutator follows from [φ(x), π(y)] = iδ (d) (x − y). In terms of the ladder
operators, Z  
d † 1
H = d̄ k ~ωk ak ak + .
2
The field operators
Z r
~  i~k·~x ~

φ(~x) = d
d̄ k e ak + e−ik·~x a†k ,
2ωk
Z r
1 d ~ωk  i~k·~x −i~k·~
x †

π(~x) = d̄ k e ak − e ak , (1.14)
i 2
satisfy the canonical commutation relation

[φ(~x), π(~x0 )] = i~1δ d (~x − ~x0 ).

I emphasize that this is really the same equation as our starting point for each ball on
springs:
[qn , pn0 ] = i~1δnn0 .

The mode expansions (1.14) contain a great deal of information. First notice that
φ is manifestly hermitian. Next, notice that from φ(~x) ≡ φ(~x, 0) by itself we can-
~
not disentangle ak and a†k , since only the combination ak + a†−k multiplies eik·~x . The
8
Beware that the mode operators ak defined here differ by powers of 2π/L from the finite-volume
objects in the previous discussion. These agree with Peskin’s conventions.

19
momentum π contains the other linear combination. However, if we evolve the field
operator in time using the Heisenberg equation (as you did on the HW), we find
Z r
~  i~k·~x−iω~ t −i~k·~
x+iω~k t †

φ(~x, t) ≡ eiHt φ(~x)e−iHt = d̄d k e k a + e
k ak . (1.15)
2ωk

Indeed we can check that the relation π = φ̇ holds.

Notice that the dependence on spacetime is via a sum of terms of the form:
~ µ
eik·~x−iω~k t = eikµ x |k0 =ω~k

and their complex conjugates. These are precisely all the solutions to the wave equation
(1.13). For each ~k, there are two solutions, one with positive frequency and one with
negative frequency. You might have worried that solutions with both signs of the
frequency mean that the world might explode or something (like it would if we tried to
replace the Schrödinger equation for the wavefunction with a Klein-Gordon equation).
This danger is evaded in a beautiful way: the coefficient of the positive frequency
solution with wavenumber ~k is the destruction operator for the mode; the associated
negative frequency term comes with the creation operator for the same mode, as a
consequence of reality of the field. (Some words about antimatter might be appropriate
here, but it will be clearer later when we talk about examples of particles that are not
their own antiparticles.) [End of Lecture 2]
Finally, in a relativistic system, anything we can say about time should also be
true of space, up to some signs. So the fact that we were able to generate the time
dependence by conjugation with the unitary operator eiHt (as in (1.15)) says that we
should be able to generate the space dependence by conjugating by a unitary operator
~ ~ is the last in a long list of objects in this section with
of the form e−iP·~x . Here P
a claim to the name ‘momentum’. It is the conserved charge associated with spatial
translation symmetry, the generator of spatial translations. Its form in terms of the
fields will be revealed below when we speak about Noether’s theorem. For now, let me
emphasize that it is distinct from the objects pn , π(x) (which were ‘momenta’ in the
sense of canonical momenta of various excitations) and also from the wavenumbers ~k
of various modes, which (when multiplied by ~) are indeed spatial momenta of single
particles. (This statement gives us an expectation for what is the total momentum of
a state of a collection of particles which we will check below in §1.4.) In terms of the
momentum operator, then, we can write
µ µ
φ(xµ ) = eiPµ x φ(0)e−iPµ x

~ µ.
with Pµ ≡ (H, P)

20
1.3 Quantum light: Photons

The quantization of the Maxwell field is logically very similar to the preceding dis-
cussion. There are just a few complications from its several polarizations, and from
the fact that quantum mechanics means that the vector potential is real and necessary
(whereas classically it is just a convenience). This is a quick-and-dirty version of the
story. I mention it here to emphasize that the machinery we are developing applies to
a system you have already thought a lot about!
Maxwell’s equations (with c = 1) are:

µνρσ ∂ν Fρσ = 0 ~ ·B
∇ ~ = 0, ~ ×E
∇ ~ = −∂t B,
~ (1.16)
µ
∂ Fµν = 4πjν ~ ·E
∇ ~ = 4πρ, ∇×B~ = ∂t E
~ + 4π~j (1.17)

(where the familiar electric and magnetic fields are E i = −F 0i and ijk B k = −F ij ).
The first two equations (1.16) are constraints on E ~ and B ~ which mean that their
components are not independent. This is annoying for trying to treat them quantumly.
~ µ which determine the fields by
To get around this we introduce potentials Aµ = (Φ, A)
taking derivatives and which automatically solve the constraints (1.16):

Fµν = ∂µ Aν − ∂ν Aµ , aka ~ = −∇Φ

E ~ − ∂t A,
~ ~ =∇
B ~ × A.
~

Potentials related by a gauge transformation

~→A
A ~λ = A
~ − ∇λ,
~ Φ → Φλ = Φ + ∂t λ

~ B.
for any function λ(~r, t), give the same E, ~ The Bohm-Aharonov effect is proof that
(some of the information in) the potential is real and useful, despite this redundancy.
We can partially remove this redundancy be choosing our potentials to satisfy Coulomb
gauge
∇~ ·A~=0.

In the absence of sources ρ = 0 = ~j, we can also set Φ = 0. In this gauge, Ampere’s
law becomes
   
2~ ~ ~ 2~ ~ ~ ~ = −∂ 2 A ~ − c2 ∇2 A
~ i.e. ∂ 2 A ~=0.
c ∇ × ∇ × A = c ∇ · ∇ · A − c2 ∇ 2 A t t

This wave equation is different from our scalar wave equation (1.6) in three ways:

• we’re in three spatial dimensions,

• the speed of sound vs has been replaced by the speed of light c,

21
 ~ is a vector field obeying the constraint ∇
• the field A ~ ·A
~ = 0. In fourier space
~ ~ ~
= k eik·~x A(k)
P
A(x) this condition is

0 = ~k · A(k)
~

– the vector field is transverse.

An action which gives rise to Maxwell’s equations is

Z   Z
4 1 µν 1 1
= d4 xLMaxwell . LMaxwell = − Fµν F µν = E2 − B2 .


S[A] = d x − Fµν F
4 4 2

Note that we must regard A as the dynamical variable to obtain (1.17) by 0 = δAδS
µ (x)
.
∂LMaxwell i 9
The canonical momentum of A is then ΠAi = ∂ Ȧ = E . So the Hamiltonian is :
i

Z
1 
~ 2 + c2 B

~2 .
H= d3 x E (1.18)
2

Here E~ = −∂t A~ plays the role of field momentum π(x) in (1.10), and B
~ =∇
~ ×A~ plays
the role of the spatial derivative ∂x q. We immediately see that we can quantize this
system just like for the scalar case, with the canonical commutator

[φ(x), π(x0 )] = i~δ(x − x0 ) [Ai (~r), Ej (~r0 )] = −i~δ 3 (~r − ~r0 )δij

where i, j = 1..3 are spatial indices10 . So we can immediately write down an expres-
sion for the quantum Maxwell field in terms of independent creation and annihilation
operators: Z
~ r) = d̄3 k √ 1
X ~ ~

A(~ a~k,s~es (k̂)eik·~r + a~†k,s~e?s (k̂)e−ik·~r
2ωk s=1,2

The field momentum is E~ = −∂t A

~:
r
~ ωk X  i~k·~
r † ? −i~k·~
r

E(~r) = i a~ ~es (k̂)e − a~k,s~es (k̂)e
2 s=1,2 k,s
9
 You may  also recall that the energy density of a configuration of Maxwell fields is u =
1
E~2 + B
~ 2 . This result can be obtained either by Legendre transformation of LMaxwell , or from
2
T00 , the energy momentum tensor as you’ll see on the HW.
10
As a check, note that using this Hamiltonian and the canonical commutator, we can reproduce
Maxwell’s equations using Ehrenfest’s theorem:

~ = − i [H, E]
D E D E D E D E
~ = ∂t E
∂t2 A ~ = c2 ∇
~ 2A
~ .
~

22
Also, the magnetic field operator is
r  
~ =∇~ ×A~ =
XX ~ ~k × a~ ~es (k̂)ei~k·~r − a† ~e? (k̂)e−i~k·~r ;
B i k,s ~k,s s
s
20 ωk L3
~k

~ in the scalar field theory11 . Plugging these

the magnetic field is analogous to ∇φ
expressions into the Hamiltonian (1.18), we can write it in terms of these oscillator
modes (which create and annihilate photons). As for the scalar field, the definitions of
these modes were designed to make this simple: It is:
 
X † 1
H= ~ωk a~k,s a~k,s + .
2
~k,s

Notice that in this case we began our story in the continuum, rather than with
microscopic particles connected by springs. (However, if you read Maxwell’s papers
you’ll see that he had in mind a particular UV completion involving gears and cogs. I
actually don’t understand it; if you do please explain it to me.)
The vacuum energy is

L3
Z
1X
E0 = ~ωk = d3 k~ck.
2 (2π)3
~k,s

P
The fact that k is no longer a finite sum might be something to worry about. We
will see below in §1.5 that this vacuum energy has physical consequences.

11
I should say a little more about the polarization vectors, ~es . They conspire to make it so that
there are only two independent states for each ~k and they are transverse ~k · ~es (k̂) = 0, so s = 1, 2.
The polarization vectors of a given ~k can be chosen to satisfy the following completeness relation:
X
esi (k̂)e?sj (k̂) = δij − k̂i k̂j . (1.19)
s

This says that they span the plane perpendicular to k̂.

23
 Consolidation of understanding
So far in this chapter, we have studied systems of increasing complexity: the simple
harmonic oscillator, a non-interacting scalar field, and the EM field in vacuum (i.e. in
the absence of charge). All these free field theories have the same structure, in the
following sense.
In the following, Here ReA ≡ 12 A + A† as usual. The normalization constant is


q
N = 12 2mω ~
.
 
1 2 1 2 2 † 1
HSHO = p + mω q = ~ω a a +
2m 2 2
[q, p] = i~ =⇒ [a, a† ] = 1.
q = ReN a, p = mImωN a.

Z    
1 2 1 2 X † 1
Hscalar = dx π + µc (∂x φ)2 = ~ωk ak ak +
2µ 2 k
2

[φ(x), π(x0 )] = i~δ(x − x0 ) =⇒ [ak , a†k0 ] = i~δkk0 .

! !
X X
φ(x) = Re Nk eikx ak , π(x) = µIm ωk Nk eikx ak .
k k

0 ~ 2 0 c2 ~ 2
Z    
3
X † 1
HEM = dx E + B = ~ωk aks aks +
2 2 k,s=1,2
2

[Ai (x), Ej (x0 )] = i~δ 3 (x − x0 )δij =⇒ [aks , a†k0 s0 ] = ~δkk0 δss0 .

! !
~ ~
X X
~
A(x) = Re ~
Nk eik·~x aks~es (k̂) , E(x) = µIm ωk Nk eik·~x aks~es (k̂) .
k k

~ is the canonical momentum of A

Note that E ~ since (in Coulomb gauge) E
~ = −∂t A.
~

24
1.4 Lagrangian field theory

[Here we fill in the bits of Peskin §2.2 that we missed above.] Let’s consider a classical
field theory in the Lagrangian description. This means that the degrees of freedom
are a set of fields φr (x), where r is a discrete index (for maybe spin or polarization
or flavor), and we specify the dynamics by the classical action. If the world is kind
to us (in this class we assume this), the action is an integral over space and time of a
Lagrangian density Z
S[φ] ≡ dd+1 xL(φ, ∂ µ φ).

This important assumption is an implementation of locality.

This central object encodes the field equations, the canonical structure on the phase
space, the Hamiltonian, the symmetries of the theory.
I’ve sneakily implied that we are going to assume Lorentz invariance, so that L
depends on the 4-vector ∂ µ φ, and not its components separately.
I am also going to assume that the action S is real.
We’ve seen basically two examples so far
1 1
LKG = ∂µ φ∂ µ φ − m2 φ2
2 2
and (you got to do the Legendre transformation from the Hamiltonian of §1.3 on the
homework)
1 1
LEM = − 2 Fµν F µν = 2 E 2 − B 2


4e 4e
with Aµ regarded as the independent degrees of freedom.
A word about units: in units with ~ = c = 1, everything has units of
mass to some power, called its mass dimension. Energy and momen-
tum pµ = ~kµ have mass dimension +1. The space and time coordi- object mass dim.
µ
nates x have mass dimension −1. The action goes in the exponential pµ 1
R iS
µ
of the path integral measure [Dφ]e ~ and so must be dimensionless. x -1
So the Lagrangian density has mass dimension d+1. This means that S 0
d−1
the KG field has mass dimension 2 (and the mass m has mass di- L d+1
~
mension 1 (yay!)). In d + 1 = 3 + 1 dimensions, E ∼ Ȧ, B ∼ ∇A have φ d−1
2
mass dimension 2 and A has mass dimension one (and e is dimension- Aµ 1
less). This is nice because then the covariant derivative ∂µ + Aµ has E,~ B,
~ Fµν 2
2 2
mass dimension one. Notice that E + B has dimension 4 which is
good for an energy per unit volume.
The equation of motion is
δS
0= .
δφr (x)

25
Note the functional derivative. You can check that in the case when L depends only
on φ and ∂µ φ, this is the same as the Lagrange EOM

∂L ∂L
0= − ∂µ
∂φr ∂(∂ µ φr )

(for each r) which I can’t remember. Note that since we are interested here in the bulk
equations of motion, we ignore boundary terms unless we are interested in field theory
on a space with boundary. That is a worthy subject but an unnecessary complication
for now.
1
By redefining the field by e.g. φ ≡ D
(χ − B/C), we can make the KG theory uglier

1 1
L = A + Bχ + Cχ2 + D∂ µ χ∂µ χ + ....
2 2
From the path integral point of view, the field is just an integration variable. Some-
times, its normalization is meaningful, like in the phonon example where it began its
life as the displacement of the atoms from their equilibrium. So you see that we are
not losing generality except in our neglect of interactions, and in our neglect of terms
with more derivatives. The former neglect we will repair little by little in this course,
by doing perturbation theory. The latter is justified well by the renormalization group
philosophy, which is a subject for next quarter.
Canonical field momentum and Hamiltonian. The Hamiltonian viewpoint in
field theory has the great virtue of bringing out the physical degrees of freedom. It
has the great shortcoming that it picks out the time coordinate as special and obscures
Lorentz symmetry.
The canonical field momentum is defined to be
∂L
π(x) = .
∂(∂t φ(x))
Notice that this expression assumes a local Lagrangian density. π is actually a ‘field
∂
momentum density’ in the sense that the literal canonical momentum is ∂ φ̇(x) L =
dd xπ(x) (as opposed to L). I will often forget to say ‘density’ here.
The hamiltonian is then
X Z   Z
H= pn qn − L = d x π(x)φ̇(x) − L ≡ dd x h.
d

Noether’s theorem and the Noether method. Yay, symmetries. Why do

physicists love symmetries so much? One reason is that they offer possible resting
places along our never-ending chains of ‘why?’ questions. For example, one answer

26
(certainly the one given in Weinberg’s text, but just as certainly not the only one) to
the question “Why QFT?” is: quantum mechanics plus Poincaré symmetry.
They are also helpful for solving physical systems: Continuous symmetries are as-
sociated with conserved currents. Suppose the action is invariant under a continuous
transformation of the fields φ, φ(x) 7→ φ0 (x). (The invariance of the action is what
makes the transformation a symmetry.) ‘continuous’ here means we can do the trans-
formation just a little bit, so that φ(x) 7→ φ(x) + ∆φ(x) where  is an infinitesimal
parameter. If the transformation with constant  (independent of space and time) is a
symmetry, then the variation of the action with  = (x, t) must be proportional to ∂µ 
(at least assuming some smoothness properties of the action), and so that it vanishes
∀φ when  is constant:
Z Z
µ IBP
S[φ + (x)] − S[φ] = d xdt∂µ (x)j = − dd xdt(x)∂µ j µ .
d

But if the equations of motion are obeyed, then the action is invariant under any
variation, including this one, for arbitrary (x). But this means that ∂µ j µ = 0, the
current is conserved. These words are an accurate description of the equation because
they mean that the charge Z
QR ≡ dd x j 0
R
in some region of space R can only change by leaving the region (assume the definition
of R is independent of time):
Z Z Z
∂t QR = d 0 d ~
d x ∂t j = − d x ∇ · j =~ dd−1 xn̂ · ~j
R R ∂R

where in the last step we used Stokes’ theorem.

This trick with pretending the parameter depends on space is called the Noether
method. More prosaically, the condition that the action is invariant means that the
Lagrangian density changes by a total derivative (we assume boundary terms in the
action can be ignored):
symmetry
L(φ0 , ∂µ φ0 ) = L(φ, ∂µ φ) + ∂µ J µ

but on the other hand, by Taylor expansion,

 
0 0 calculus ∂L ∂L
L(φ , ∂µ φ ) = L(φ, ∂µ φ) +  ∆φ + ∂µ ∆φ
∂φ ∂ (∂µ φ)
 
IBP  ∂L ∂L  ∂L
= L(φ, ∂µ φ) +  
 ∂φ − ∂ µ
 ∆φ + ∂µ ∆φ .
∂ (∂µ φ)  ∂ (∂µ φ)
| {z }
eom

27
By combining the previous two equations for L(φ0 ), we see that on configurations which
satisfy the EOM, 0 = ∂µ j µ with
∂L
jµ = ∆φr − J µ . (1.20)
∂ (∂µ φr )
Notice that I stuck the index back in at the last step.
[End of Lecture 3]
There is a converse to the Noether theorem, which is easier to discuss directly in
quantum mechanics. Given a conserved charge Q, that is, a hermitian operator with
[H, Q] = 0, we can make a symmetry transformation of the fields φ by

δφ ≡ i[Q, φ]. (1.21)

We’ll say that Q generates the symmetry, for the following reason. (1.21) is the in-
finitesimal version of the finite transformation

φ → φ0 ≡ eiQ φe−iQ .

The object U ≡ eiQ is a unitary operator (since Q is hermitian) which represents the
action of the symmetry on the Hilbert space of the QFT. It is a symmetry in the sense
that it commutes with the time evolution operator eiHt .
Some examples will be useful:

• For example, suppose S[φ] only depends on φ through its derivatives, for example,
S[φ] = 12 ∂µ φ∂ µ φ. Then there is a shift symmetry φ → φ0 ≡ φ + . Letting 
R
R
depend on spacetime, the variation of the action is S[φ + (x)] = ∂µ ∂ µ φ, so the
R
current is jµ = ∂µ φ. Let’s check the converse: Indeed, the charge Q = space j0
generates the symmetry in the sense that for small , the variation in the field is

δφ ≡ φ0 − φ =  = i[Q, φ]

(if we were doing classical mechanics, we should replace i[Q, φ] with the Poisson
bracket). Using our expression for the current this is
Z
dd y φ̇(y), φ(x) = 
 
δφ = i
|{z}
=π(y)

which is indeed true. In this case the finite transformation is again φ → φ + .

• On the homework you’ve seen a complex scalar Φ, with S[Φ, Φ? ] is invariant under
Φ → ei Φ = Φ + iΦ + O(2 ). This U(1) phase transformation can be rewritten

28
 in terms of the real and imaginary parts as an SO(2) rotation. The charge can
be written as Z Z
Q = d xj = d̄d p a†p ap − b†p bp
d 0

where the two sets of creation and annihilation operators are associated with
excitations of Φ and Φ† respectively. (That is, quantize φ1,2 as we did for a
single real scalar field, in terms of mode operators a1,2 respectively. Then let
a ≡ a1 + ia2 , b ≡ a1 − ia2 , up to numerical prefactors which I tried to get right
in the solutions.) So the particles created by a and b have opposite charge (this
is essential given the mode expansion Φk ∼ ak + b†−k ) and can be interpreted as
each others’ antiparticles: there can be symmetry-respecting processes where an
a particle and b particle take each other out.

• Consider spacetime translations, xµ → xµ − aµ . We can think of this as a trans-

formation of the fields by

φ(x) 7→ φ(x + a) = φ(x) + aν ∂ν φ +O(a2 ).

|{z}
≡∆ν φ

Our transformation parameter is now itself a four-vector, so we’ll get a four-

vector of currents Tνµ . This will be a symmetry as long as the lagrangian doesn’t
depend explicitly on space and time ( so ∂ν L = 0) but rather depends on space
chain rule
and time only via the fields (so 0 6= dxdν L = ∂ν φ ∂L ∂φ
+ ∂µ ∂ν φ ∂(∂∂Lµ φ) ). Let’s use
the prosaic method for this one: the shift in the Lagrangian density also can be
found by Taylor expansion
d
L 7→ L + aµ L = L + aν ∂µ (δνµ L) .
dxµ
So the formula (1.20) gives
∂L
Tνµ = ∂ν φ −Lδνµ .
∂ (∂µ φ) |{z}
∆ν φ

For the time translation, the conserved charge T00 gives back the hamiltonian
density h = π φ̇−L obtained by Legendre transformation. The conserved quantity
from spatial translations is the momentum carried by the field, which for the KG
field is Z Z
Pi = d x Ti = − dd x π∂i φ.
d 0

For the Maxwell field, this gives the Poynting vector.

There is some ambiguity in the definition of the stress tensor which you’ll study
on the homework.

29
 Let’s check that expression above for the conserved momentum agrees with our
expectations.
 E In particular, in free field theory the total momentum of Pthe state
~
k1 , · · · ~kn should be just the sum of the momenta of the particles, P~ = `=1 ~~k`
n

(with interactions the story can be more complicated). Indeed

Z Z
Pi = − d x π∂i φ = d̄d kki a~†k a~k
d

agrees with this. (Notice that I used rotation invariance of the vacuum to not
worry about a possible constant term.)

• We’ll say more about the rest of the Poincaré group, i.e. rotations and boosts,
later on.

1.5 Casimir effect: vacuum energy is real

[A. Zee, Quantum Field Theory in a Nutshell, §I.9] This subsection has two purposes.
One is to show that the 21 ~ω energy of the vacuum of the quantum harmonic oscillator
is real. Sometimes we can get rid of it by choosing the zero of energy (which doesn’t
matter unless we are studying dynamical gravity). But it is meaningful if we can vary
ω (or the collection of ωs if we have many oscillators as for the radiation field) and
compare the difference.
The other purpose is to give an object lesson in asking the right questions. In
physics, the right question is often a question which can be answered by an experiment,
at least in principle. The answers to such questions are less sensitive to our silly
theoretical prejudices, e.g. about what happens to physics at very short distances.
In the context of the bunch of oscillators making up the radiation field, we can
change the spectrum of frequencies of these oscillators {ωk } by putting it in a box and
varying the size of the box. In particular, two parallel conducting plates separated by
some distance d experience an attractive force from the change in the vacuum energy
of the EM field resulting from their presence. The plates put boundary conditions on
the field, and therefore on which normal modes are present.
To avoid some complications of E&M which are not essential for our point here,
we’re going to make two simplifications:

• we’re going to solve the problem in 1+1 dimensions

• and we’re going to solve it for a scalar field.

30
 To avoid the problem of changing the boundary conditions outside the plates we
use the following device with three plates:

| ← d → | ←− L−d −→ |

(We will consider L  d, so we don’t really care about the far right plate.) The
‘perfectly conducting’ plates impose the boundary condition that our scalar field q(x)
vanishes there. The normal modes of the scalar field q(x) in the left cavity are then

qj = sin (jπx/d) , j = 1, 2, ...

with frequencies ωj = π|j|

d
c. There is a similar expression for the modes in the right
cavity which we won’t need. We’re going to add up all the 12 ~ωs for all the modes in
both cavities to get the vacuum energy E0 (d); the force on the middle plate is then
−∂d E0 .
The vacuum energy in the whole region of interest between the outer plates is the
sum of the vacuum energies of the two cavities

E0 (d) = f (d) + f (L − d)

where ∞ ∞
1 X π X ?!?!!
f (d) = ~ ωj = ~c j = ∞.
2 j=1 2d j=1
We have done something wrong. What?
Our crime is hubris: we assumed that we knew what the modes of arbitrarily large
mode number k (arbitrarily short wavelength, arbitrarily high frequency) are doing,
and in particular we assumed that they cared about our silly plates. In fact, no metal
in existence can put boundary conditions on the modes of large enough frequency –
those modes don’t care about d. The reason a conductor puts boundary conditions
on the EM field is that the electrons move around to compensate for an applied field,
but there is a limit on how fast the electrons can move (e.g. the speed of light). The
resulting cutoff frequency is called the plasma frequency but we don’t actually need to
know about all these details. To parametrize our ignorance of what the high-frequency
modes do, we must cut off (or regularize) the contribution of the high-frequency modes.
Let’s call modes with ωj  π/a high frequency, where a is some short time12 . Replace
∞
π X −aωj /π
f (d) f (a, d) = ~ e j
2d j=1
12
You can think of a as the time it takes the waves to move by one lattice spacing. If we work
in units where the velocity is c = 1, this is just the lattice spacing. I will do so for the rest of this
discussion.

31
 ∞
!
π~ X
−aj/d
= − ∂a e
2 j=1
| {z }
1
= −1
1−e−a/d

π~ ea/d
= +
2d (ea/d − 1)2
 
ad πd π πa2
' ~ − + + ... (1.22)
 
2a2
|{z} 24d 480d 3

→∞ as a→0

Answers which don’t depend on a have a chance of being meaningful. The thing we
can measure is the force:

F = −∂d E0 = − (f 0 (d) − f 0 (L − d))

   π 
π π 2 π 2
= −~ + + O(a ) − + + O(a )
2a2 24d2 2a2 24 (L − d)2
 
a→0 π~ 1 1
= − −
24 d2 (L − d)2
Ld π~c
= − (1 + O(d/L)) . (1.23)
24d2
This is an attractive force between the plates. (I put the c back in the last line.)
The analogous force between real conducting plates, caused by the change of bound-
ary conditions on the electromagnetic field, has been measured.
The string theorists will tell you that ∞ 1
P
j=1 j = − 12 , and our calculation above
agrees with them in some sense. But what this foolishness means is that if we compute
something which is not dependent on the cutoff we have to get the same answer no
matter what cutoff we use. Notice that it is crucial to ask the right questions.
An important question is to what extent could we have picked a different cutoff
function (instead of e−πω/a ) and gotten the same answer for the physics. This interest-
ing question is answered affirmatively in Zee’s wonderful book, 2d edition, section I.9
(available electronically here!).
A comment about possible physical applications of the calculation we actually did:
you could ask me whether there is such a thing as a Casimir force due to the vacuum
fluctuations of phonons. Certainly it’s true that the boundary of a chunk of solid
puts boundary conditions on the phonon modes, which change when we change the
size of the solid. The problem with the idea that this might produce a measurable

32
force (which would lead the solid to want to shrink) is that it is hard to distinguish
the ‘phonon vacuum energy’ from the rest of the energy of formation of the solid,
that is, the energy difference between the crystalline configuration of the atoms and
the configuration when they are all infinitely separated. Certainly the latter is not
well-described in the harmonic approximation (λ = 0 in (1.1)).

A few comments about the 3+1 dimensional case of E&M. Assume the size
of the plates is much larger than their separation L. Dimensional analysis shows that
the force per unit area from vacuum fluctuations must be of the form
~c
P =A
L4
where A is a numerical number. A is not zero!
Use periodic boundary conditions in the xy planes (along the plates). The allowed
wave vectors are then  
~k = 2πn x 2πn y
,
Lx Ly
with nx , ny integers.
We have to do a bit of E&M here. Assume the plates are perfect conductors
(this where the hubris about the high-frequency modes enters). This means that the
transverse component of the electric field must vanish at the surface. Instead of plane
waves in z, we get standing waves: φ(z) ∝ sin (nπz/L) .
The frequencies of the associated standing waves are then
r
π 2 n2 ~ 2
ωn (~k) = c + k , n = 0, 1, 2
L2
Also, there is only one polarization state for n = 0.
So the zero-point energy is
 
0
~ X
E0 (L) = 2 ωn (~k)
2
n,~k

where it’s useful to define

0
X 1 X X
≡ +
2
n,~k n=0,~k n≥1,~k

Now you can imagine introducing a regulator like the one we used above, and replacing
0 0
~
X X
· e−aωn (k)/π ·
n,~k n,~k

33
and doing the sums and integrals and extracting the small-a behavior.

2 The path integral makes some things easier

2.1 Fields mediate forces

[Zee §1.3] Consider again our chain of balls on springs. Suppose a giant hand reaches
in and pushes the atom at position xn a little bit. This can be described by adding to
the hamiltonian a term
δV (q) = −Jn (t)qn (t)
which applies a force Jn (t) to the nth atom. We can ask, in the presence of such a
force, what is the amplitude to go from state I to state F in time T :
Z
−i 0T dtH(t) 1 2
|Ii = [Dφ]ei dtd x( 2 (∂φ) −V (φ)+J(x)φ(x)) .
d
R R
hF | e

As you see, this is a quantity for which we have a path integral representation. Here’s
a reason to care about this quantity: take the initial and final states to be the ground-
state: RT RT
h0| e−i 0 dtH(t) |0i ' e−i 0 dtEgs (J) .
If the time-dependence is slow enough, the answer is obtained by the adiabatic approx-
imation: just add up the instantaneous groundstate energy at each time step.
To do this integral, we retreat to the case where the action is quadratic in φ, so
that
1
IBP 1
∂µ φ∂ µ φ − m2 φ2 = − φ ∂ 2 + m2 φ + total derivative.


L(φ) = (2.1)
2 2
Going back to the lattice to make the integrals slightly less scary, we have

∞ M t ,N
r
(2πi)N Mt − i Jx A−1
Z Z
i
R Y
eiW [J] ≡ [Dφ]ei (L+Jφ)
= dqn,t e q A q +iJx qx
2 x xy y = e 2 xy Jy .
−∞ n,t det A
R
Here repeated indices are summed as usual: qx Axy qy = dxdyφ(x)Axy φ(y), etc... So
you can see that the matrix A multiplying the quadratic term in this gaussian integral
is Axy = −δ d+1 (x − y) (∂x2 + m2 ). It is an N Mt × N Mt matrix. Its inverse A−1 satisfies
by definition Axz A−1
zy = δxy , which is the differential equation

− (∂ 2 + m2 )D(x − y) = δ(x − y). (2.2)

34
This equation says that D is a Green’s function for the operator −(∂ 2 + m2 ). The fact
that there is no special point in spacetime says A−1
xy = D(x − y) only depends on the
difference of its arguments.
Does this integral actually converge? The integral you saw on the homework was
1
of the form R dqe− 2 qAq , which surely converges if A is a positive matrix. Actually, this
R

is overkill – it is enough to replace m2 → m2 − i to make all the integrals converge.

Here  is an infinitesimal, whichRmeans 2 = 0 and c =  for any positive c. Then each
2
dqnt will have a factor of e− qnt which suppresses the integrand in the dangerous
R

large-field region13 .
[End of Lecture 4]
The equation (2.2) is translation-invariant and linear so you should not be surprised
that it is solved by going to Fourier space (in space and time):
Z Z
d+1 ikµ xµ µ
D(x) = d̄ k e Dk , δ (x) = d̄d+1 k eikµ x .
d+1

in terms of which (2.2) becomes the algebraic equation 1 = (k 2 − m2 + i)Dk . Hence

eikx
Z
D(x) = d̄d+1 k 2 .
k − m2 + i
Notice that the shift by  saves the day here: it keeps the inte-
gration contour from running right over the pole at k 2 = m2 , by
moving slightly in the imaginary direction. More explicitly,

k 2 − m2 + i = ω 2 − ~k 2 − m2 + i

is zero when
q q
ω = ± k + m − i = ± (ωk − i) , ωk ≡ ~k 2 + m2 .
~ 2 2

p p
In the second step I Taylor expanded ωk2 − i = ωk2 − ωik +
O()2 and used the facts that ωk > 0, and that anything positive
times an infinitesmal is an infinitesimal.
We can then do the ω integral by contours14 : if t > 0 (t < 0), we can close the
contour in the UHP (LHP) since the integrand goes like e−Im ωt , and the integral equals
13
You might notice that I am swindling you here: there could be other ways to make the integral
well-defined (there are). Another thing you might be bothered by is the boundary conditions on the
fields and their relation to the initial and final states. In the next subsection, we’ll say more.
14
H P
We are using the Cauchy residue theorem C dz = f (z) = 2πi zj Resz=zj f where zj are the
poles of f . To remember the sign, consider a small circle C0 counterclockwise around the origin and
R 2π
f (z) = 1/z, so C0 dz
H
z = i 0 dθ = 2πi.

35
the residue of the pole at ω = ωk ∓ i (times 2πi):
~ ~
!
e−i(ωk t−k·~x) ei(ωk t−k·~x)
Z
D(x) = −i d̄d k θ(t) + θ(−t) . (2.3)
2ωk 2ωk

We’ll learn to call this time-ordered in a moment.

The propagator. Who is D(x), besides some quantity in terms of which we did a
Gaussian integral? Recall from the homework that the inverse matrix can be extracted
via a two-point correlation function:
Z
1
A = dqq 2 e− 2 qAq /Z.
−1

Putting back all the labels, the same manipulations show that
?
D(x − y) = h0|φ(x)φ(y)|0i (2.4)

– the amplitude to propagate an excitation created from the vacuum by φ(x) to be

annihilated by φ(y). The propagator, for short.
(Notice that if the system is Lorentz invariant (which starting from (2.1) it is) then
since D(x) is a scalar quantity, it can only depend on x through Lorentz invariants
made from xµ , namely the proper distance x2 = t2 − ~x2 , and the sign of t.)
Why the ‘?’ in (2.4)? For one thing, φ(x) and φ(y) are operators – the order matters.
How do I know which order in which to write them? To reproduce (2.3) the thing to
do is to time-order them:

h0|T φ(x)φ(y)|0i ≡ θ(x0 − y 0 ) h0|φ(x)φ(y)|0i + θ(y 0 − x0 ) h0|φ(y)φ(x)|0i .

To see this, plug in the mode expansion (1.14) to see e.g.

Z d d Z d
d̄ k d̄ q −ikx+iqy † d̄ k −ik(x−y)
h0|φ(x)φ(y)|0i = √ e h0| ak aq |0i = e
2 ωk ωq 2ωk

(where k 0 = ωk , q 0 = ωq to satisfy the KG equation), which reproduces the first term

in (2.3)15 .
Now why should we care about the propagator? Look again at W [J]. We’ve learned
that (up to terms independent of J),
Z Z Z
1 1 1
W [J] = − d+1 d+1
d xd yJ(x)D(x − y)J(y) = − d̄d+1 kJk? 2 Jk
2 2 k − m2 + i
15
The other ways of making the path integral well-defined correspond to other ways of ordering the
φs, and other initial and final states.

36
Here J(x) = d̄d+1 keikx Jk , Jk? = J−k . We get to pick J(x). Choose J = J1 + J1 to
R
3
describe (in Zee’s words) two lumps sitting still
 on the mattress: Ja (x) = δ (x−xa ), a =
R 0 −ik0 x0  i~k·~x ~
1, 2. Then Jk = dx e e 1 + eik·~x2 . The interaction between the two lumps
mediated by the mattress field φ will then be described by the J1 J2 cross-terms in W [J]:
~
eik·(~x1 −~x2 )
Z Z Z Z
2 0 ik0 (x0 −y 0 )
W [J] = − dx 0
dy 0
d̄k e d̄3 k 2 + .. (2.5)
2 k − m2 + i
~
eik·(~x1 −~x2 )
Z Z Z
0 0 0 3
= − dx d̄k 2πδ(k ) d̄ k 2 + .. (2.6)
k − m2 + i
| {z }
=1
Z Z i~k·(~
x1 −~
x2 )
e
=+ dx0 d̄3 k + .. (2.7)
~k 2 + m2 − i

For this choice of J, the Hamiltonian is time-independent, and eiW = h0| e−iHT |0i =
e−iEgs (J)T , so W = −Egs (J)T . We learn that
~
eik·~r
Z
Egs (J) = − d̄d k .
~k 2 + m2

Notice that we can drop the i now, because this integrand is nonsingular for real ~k.
In d = 1, there are poles at k = ±im, and we can close the contour in the UHP for
free to get16
2πi e−mx e−mx
Egs (J) = − =− .
2π 2im 2m
Since x is the separation between the lumps, this means that our field has produced
an attractive force between the lumps
1
F = −∂x Egs (J) = + e−mx
2
which falls off exponentially with the separation between the lumps. The range of the
potential goes inversely with the mass of the ‘force carrier’ φ. The 3d version of this
−mr
potential e r (see footnote 16) is called the Yukawa potential.
16
For convenience, here’s the integral in 3d:
~ Z ∞ Z 1 Z ∞
eik·~x y≡cos θ 1 k 2 dk
Z
1 dkk sin kr
d̄3 k = dyeikyr
=
~k 2 + M 2 2 2
(2π) 0 k + M −1 2 (2π) r −∞ k 2 + M 2
2
| {z }
=2 sin
kr
kr
 Z ∞ ikr

1 1 ke
= dk + h.c. close contour in UHP for free
(2π)2 r 2i −∞ k 2 + M 2
1 1 iM ei(iM )r e−M r
= 2πi · 2 = .
(2π)2 r 2i 2iM 4πr

37
2.2 Euclidean path integral and Wick rotation

Here is a route to defining the path integral (actually the same as the replacement
m2 → m2 − i) which makes clearer what is going on with the initial and final states.
For simplicity, let us focus on a single mode of the field – a single harmonic oscillator
Z Z
1 2 2 2


S[q] = dt (∂t q) − Ω q − Jq
2

(where Ω2 = ~k 2 + m2 if you like). Consider the replacement τ = it in the action:

Z Z Z  
1 2 2 2

1 2 2 2


S[q] = i dτ −(∂τ q) − Ω q − i dτ Jq = −i dτ (∂τ q) + Ω q + Jq .
2 2
Then the path integral becomes
Z
1
[Dq]e− dτ ( 2 ((∂τ q) +Ω q )+Jq) .
2 2 2
R

This integrand suppresses configurations with large q, and large ∂τ q, and the integral
is therefore totally well-defined. The euclidean action is17
Z   Z  
1 2 2 2

1 2 2


SE [q] = dτ (∂τ q) + Ω q + Jq = dτ q −∂τ + Ω q + Jq
2 2

where (−∂τ2 + Ω2 ) is a positive operator (meaning all of its eigenvalues are positive).
Call its inverse GE , which then, by definition, satisfies

−∂σ2 + Ω2 G(s, t) = δ(σ − τ )

The fact that our system is time-translation invariant means G(σ, τ ) = G(σ − τ ). We
R
can solve this equation in fourier space: G(s) = d̄ωeiωσ Gω makes it algebraic:
1
Gω =
ω 2 + Ω2
and we have
eiωσ
Z
1
G(σ) = d̄ω 2 2
= e−Ω|σ| . (2.8)
ω +Ω 2Ω
(Do it by residues: the integrand has poles at ω = ±iΩ (see the figure 1 below). The
absolute value of |σ| is crucial, and comes from the fact that the contour at infinity
converges in the upper (lower) half plane for σ < 0 (σ > 0).)
17
It is called euclidean because the (∂τ q)2 has the same sign as the spatial derivatives (∂x q)2 , so
this is the action we get in euclidean spacetime with metric δµν , rather than ηµν . Exercise: put back
the spatial derivative terms and check that this is the case.

38
 Figure 1: Poles of the integrand of the ω integral in (2.8).

I claim that the real-time calculation which keeps the oscillator in its groundstate
is the analytic continuation of the one we did above, where we replace
ωMink = e−i(π/2−) ωabove (2.9)
where  is (a familiar) infinitesimal. In the picture of the euclidean frequency plane
in Fig. 1, this is a rotation by nearly 90 degrees. We don’t want to go all the way to
90 degrees, because then we would hit the poles at ±iΩ. The replacement (2.9) just
means that if we integrate over real ωMink , we rotate the contour in the integral over ω
as follows:

as a result we pick up the same poles at ωabove = ±iΩ as in the euclidean calculation.
Notice that we had better also rotate the argument of the function, σ, at the same
time to maintain convergence, that is:
ωeucl = −iωMink , ωeucl teucl = ωMink tMink , teucl = +itMink . (2.10)

So this is giving us a contour prescription for the real-frequency integral. The

result is the Feynman contour, and it is the same as what we got from m2 → m2 − i:

39
 Figure 2: The Feynman contour in the ωMink complex plane.

depending on the sign of the (real) time separation of the two operators (recall that t
is the difference), we close the contour around one pole or the other, giving the time-
ordered propagator. For the case of a free scalar field, the replacement m2 → m2 − i
had the same effect of rotating the real-frequency contour away from the poles. It is
also the same as shifting the frequency by Ω → Ω − i, as indicated in the right part
of Fig. 2. This prescription works in a case where there is no m2 term.

Notice for future reference that the euclidean action and real-time action are related
by
Z  2 ! Z  2 !
1 ∂Q 1 ∂Q
Seucl [Q] = dteucl + Ω2 Q2 = −iSMink [Q] = −i dtMink − Ω2 Q2 .
2 ∂teucl 2 ∂tMink

because of (2.10). This means the path integrand is e−Seucl = eiSMink .

Now, why does the contour coming from the euclidean path integral put the oscil-
lator into its groundstate? The point in life of the euclidean path integral is to prepare
the groundstate from an arbitrary state:
Z
[dq]e−S[q] = hq0 | e−βH |...i = ψgs (q0 ) (2.11)
q(β)=q0

– the euclidean-time propagator e−βH beats down the amplitude of any excited state
relative to the groundstate, for large enough β.
More slowly: the path integral representation for the real-time propagation ampli-
tude is Z Rt
−iHt
hf | e |ii = [dq]ei 0 dtL .
f ←i

On the RHS here, we sum over all paths between i and f in time t (i.e. q(0) = qi , q(t) =

40
 R
qf ), weighted by a phase ei dtL
. But that means you also know a representation for
X
hf | e−βH |f i ≡ tre−βH
f

– namely, you sum over all periodic paths qi = qf in imaginary time t = −iβ. So:
I I
− 0β dτ L
R
−βH
Z(β) = tre = [dq]e = [dq]e−Seucl [q]

The LHS is the partition function in quantum statistical mechanics. The RHS is the
euclidean functional integral we’ve been using. [For more on this, see Zee §V.2]
The period of imaginary time, β ≡ 1/T , is the inverse temperature. We’ve been
studying the limit as β → ∞. Taking β → ∞ means T → 0, and you’ll agree that at
T = 0 we project onto the groundstate (if there’s more than one groundstate we have
to think more).
Time-ordering. To summarize the previous discussion: in real time, we must
choose a state, and this means that there are many Green’s functions, not just one:
hψ| q(t)q(s) |ψi depends on |ψi, unsurprisingly.
But we found a special one which arises by analytic continuation from the euclidean
Green’s function, which is unique18 . It is

G(s, t) = hT (q(s)q(t))i ,

the time-ordered, or Feynman, Green’s function, and I write the time-ordering symbol
T to emphasize this. I emphasize that from our starting point above, the time ordering
arose because we have to close the contour in the UHP (LHP) for t < 0 (t > 0).
Let’s pursue this one more step. The same argument tells us that the generating
functional for real-time correlation functions of Q is
D R E R
i Jq
Z[J] = T e = h0| T ei Jq |0i .
18
Another important perspective on the uniqueness of the euclidean Green’s function and the non-
uniqueness in real time: in euclidean time, we are inverting an operator of the form −∂τ2 + Ω2 which is
positive (≡ all its eigenvalues are positive) – recall that −∂τ2 = p̂2 is the square of a hermitian operator.
If all the eigenvalues are positive, the operator has no kernel, so it is completely and unambiguously
invertible. This is why there are no poles on the axis of the (euclidean) ω integral in (2.8). In real
time, in contrast, we are inverting something like +∂t2 + Ω2 which annihilates modes with ∂t = iΩ
(if we were doing QFT in d > 0 + 1 this equation would be the familiar p2 − m2 = 0). These are
called ‘on-shell states’, they are actual states in the spectrum of the Hamiltonian of the system. So
the operator we are trying to invert has a kernel and this is the source of the ambiguity. In frequency
space, this is reflected in the presence of poles of the integrand on the contour of integration; the
choice of how to negotiate them encodes the choice of Green’s function.

41
In the second step I just emphasized that the real time expectation value here is really
a vacuum expectation value. This quantity has the picturesque interpretation as the
vacuum persistence amplitude, in the presence of the source J.
So we see that in general, the correlation functions that are computed by this “i
prescription” of Wick rotating from Euclidean spacetime are time-ordered:
Z
1
[Dφ]eiSm2 →m2 −i f (φ) = h0|T f (φ)|0i .
Z

Causality. In other treatments of this subject, you will see the Feynman contour
motivated by ideas about causality. We’ll discuss this in §3. This was not the logic of
our discussion but it is reassuring that we end up in the same place. Note that even
in 0+1 dimensions there is a useful notion of causality: effects should come after their
causes. [End of Lecture 5]

2.3 Feynman diagrams from the path integral

The previous subsection was a sophisticated discussion of QFT in 0 + 1 dimensions,

since we focussed on a single mode. To attempt to demystify some more of the structure
we’ll discover in QFT, let’s regress even further, and consider the case of QFT in 0 + 0
dimensions. By the path-integral representation, this means ordinary integrals. If
everything is positive, this is probability theory.
Suppose we want to do the integral
Z ∞ Z
g 4
− 12 m2 q 2 − 4!
Z(J) = dq e q +Jq
≡ dq e−S(q) . (2.12)
−∞

It is the path integral for φ4 theory with fewer labels. For g = 0, this is a gaussian
integral which we know how to do. For g 6= 0 it’s not an elementary function of its
arguments. We can develop a (non-convergent!) series expansion in g by writing it as
Z ∞  
− 21 m2 q 2 +Jq g 4 1  g 4 2
Z(J) = dq e 1− q + − q + ···
−∞ 4! 2! 4!
and integrating term by term. And the term with q 4n (that is, the coefficient of
1 −g n


n! 4!
) is
Z ∞  4n Z ∞  4n r
− 21 m2 q 2 +Jq 4n ∂ − 21 m2 q 2 +Jq ∂ 1
J 1
J 2π
dq e q = dq e = e 2 m2 .
−∞ ∂J −∞ ∂J m2
So: r
2π − 4!g ( ∂J
∂ 4 1
) e 2 J m12 J .
Z(J) = 2
e
m

42
 This is a double expansion in powers of J and powers of g. The process of computing
the coefficient of J n g m can be described usefully in terms of diagrams. There is a factor
of 1/m2 for each line (the propagator), and a factor of (−g) for each 4-point vertex
(the coupling), and a factor of J for each external line (the source). For example, the
coefficient of gJ 4 comes from:  4
1
∼ gJ 4 .
m2

There is a symmetry factor which comes from expanding the exponential: if the
diagram has some symmetry preserving the external labels, the multiplicity of diagrams
does not completely cancel the 1/n!.
As another example, consider the analog of the two-point function:

dq q 2 e−S(q)
R

2 ∂
G ≡ q |J=0 = R −S(q)
= −2 log Z(J = 0).
dq e ∂m2

In perturbation theory this is:

G' + O(g 3 )
 
−2 1 2 2 −8
=m 1 − gm−4 + g m + O(g )3
(2.13)
2 3

The Feynman diagrams we’re going to draw all the time are the same but with more
labels. Notice that each of the qs in our integral could come with a label, q → qa . Then
each line in our diagram would be associated with a matrix (m−2 )ab which is the inverse
of the quadratic term qa m2ab qb in the action. If our diagrams have loops we get free
sums over the label. If that label is conserved by the interactions, the vertices will have
some delta functions. In the case of translation-invariant field theories we can label
lines by the conserved momentum k. Each comes with a factor of the free propagator
i D
k) (2π)D , and we
P
2 2
k +m +i
, each vertex conserves momentum, so comes with igδ (
R D
must integrate over momenta on internal lines d̄ k.

43
Brief comments about large orders of perturbation theory.

• The perturbation series about g = 0 does not converge. How do I know? One
way to see this is to notice that if I made g even infinitesimally negative, the
integral itself would not converge (the potential would be unbounded below),
and Zg=−|| is not defined. Therefore Zg as a function of g cannot be analytic in
a neighborhood of g = 0. This argument is due to Dyson, and applies also in
most QFTs. This means there is more to QFT than perturbation theory: the
perturbation series does not define the field theory amplitudes.

• The expansion of the exponential in the integrand is clearly convergent for each
q. The place where we went wrong is exchanging the order of integration over q
and summation over n.

• The integral actually does have a name – it’s a Bessel function:

2 √ ρ 3m4
Z(J = 0) = √ ρe K 1 (ρ), ρ≡
m2 4 4g
√
(for Re ρ > 0), as Mathematica will tell you. Because we know about Bessel
functions, in this case we can actually figure out what happens at strong coupling,
when g  m4 , using the asymptotics of the Bessel function.

• In this case, the perturbation expansion too can be given a closed form expression:
1
r
2π X (−1)n 22n+ 2
 
1  g n
Z(0) ' Γ 2n + . (2.14)
m2 n n! (4!)n 2 m4

• The expansion for G is of the form

∞
X  g n
−2
G'm cn .
n=0
m4
n1
When n is large, the coefficients satisfy cn+1 ' − 23 ncn (you can see this by
looking at the coefficients in (2.14)) so that |cn | ∼ n!. This factorial growth of
the number of diagrams is general in QFT and is another way to see that the
series does not converge.

• The fact that the coefficients cn grow means that there is a best number of orders
to keep. The errors start getting bigger when cn+1 mg4 ∼ cn , that is, at order


4
n ∼ 3m
2g
. So if you want to evaluate G at this value of the coupling, you should
stop at that order of n.

44
• I said above that the fact that the perturbation series doesn’t converge means
that it doesn’t define the field theory amplitudes. What does it miss? To answer
this, consider trying to do the integral (2.12) by saddle point (at J = 0 for
simplicity):
g
0 = S 0 (q? ) = m2 q? + q?3 .
3!
(Note the resemblance to the equations of motion.) This has three solutions:
s
3!m2
q? = 0, q? = ±i .
g

The expansion about the ‘trivial’ saddle at q? (where the action is S(q? = 0) = 0)
reproduces the perturbation series. At the other saddles,
s !
3!m2 3m4
S q? = ±i =− , (2.15)
g 2g

3m4
which means their contribution would go like e+ 2g , which actually would blow
up at weak coupling, g → 0. These saddles are not on the contour and don’t
contribute for small positive g, but more generally (as for example when m2 < 0),
a
there will be effects that go like e− |g| . This is a function whose series expansion
in g at g = 0 is identically zero. You can never find it by doing perturbation
theory in g about g = 0.

• A technique called Borel resummation can sometimes produce a well-defined func-

tion of g from an asymptotic series whose coefficients diverge like n!. The idea is
to make a new series X cm
B(z) ≡ zm
m=0
n!
whose coefficients are ensmallened by n!. Then to get back Z(g) we use the
identity
1 ∞
Z
1= dze−z z n
n! 0
and do the Laplace transform of B(z):
R∞ ∞
dze−z/g z m
Z ∞ X X
−z/g
dzB(z)e = cm 0 =g cm g m = gZ(g).
0 m=0
m! m=0

This procedure requires both that the series in B(z) converges and that the
Laplace transform can be done. In fact this procedure works in this case.

45
 The fact that the number of diagrams at large order grows like n! is correlated
with the existence of saddle-point contributions to Z(g) which go like e−a/g .
This is because they are associated with singularities of B(z) at z = a; such a
singularity means the sum of cn!n z n must diverge at z = a. (More generally, non-
1/p
perturbative effects which go like e−a/g (larger if p > 1) are associated with
(faster) growth like (pn)!. See this classic work.)

• The function G(g) can be analytically continued in g away from the real axis,
and can in fact be defined on the whole complex g plane. It has a branch cut on
the negative real axis, across which its discontinuity is related to its imaginary
a
part. The imaginary part goes like e− |g| near the origin and can be computed by
a tunneling calculation like (2.15).
How did we know Z has a branch cut? One way is from the asymptotics of the
Bessel function. But, better, why does Z satisfy the Bessel differential equation
as a function of the couplings? The answer, as you’ll check on the homework, is
that the Bessel equation is a Schwinger-Dyson equation,
Z ∞
∂
something e−S(q)


0=
−∞ ∂q

which results from demanding that we can change integration variables in the
path integral.

For a bit more about this, you might look at sections 3 and 4 of this recent paper from
which I got some of the details here. See also the giant book by Zinn-Justin. There is a
deep connection between the large-order behavior of the perturbation series about the
trivial saddle point and the contributions of non-trivial saddle points. The keywords
for this connection are resurgence and trans-series and a starting references is here.

3 Lorentz invariance and causality

[Peskin §2.2, 2.3, 2.4] Now we take Lorentz invariance to be an exact symmetry and
see what its consequences are for QFT.
Relativistic normalization of 1-particle states. Fock space is spanned by the
states |~p1 , · · · p~n i ∝ a†p1 · · · a†pn |0i where ap |0i = 0. Now it is time to turn that ∝
into an =. Fock space is a direct sum of sectors labelled by the number of particles:
P
k Nk = 0, 1, 2.... (Without interactions, the hamiltonian is block diagonal in this

46
decomposition.) In the no-particle sector, it is clear what we should do: h0|0i = 1.
A one-particle state is |pi ≡ cp a†p |0i. How best to choose cp ?

(This discussion is shaded because it contains equations which will not be true in
the normalization we’ll use below. In this regard, beware the section of Peskin called
?
“how not to quantize the Dirac equation”.) Suppose we choose cp = 1. Then
D  E ak |0i=0
?
~k p~ = h0| ak a†p |0i = h0| [ak , a†p ] |0i = (2π)d δ (d) (~k − p~) ≡ /δ(~k − p~).

Suppose the previous equation is true in my rest frame F . Since 1 = d̄d p/δ(p − k), we
R

see that d̄d p/δ(p − k) is Lorentz invariant. More precisely,

X δ(x − x0 )
δ(f (x)) = .
zeros x0 of f
|f 0 (x0 )|

If another F 0 is obtained by a boost in the x direction, p0µ = Λµν pν ,

dp0x
 

0
   dE
E γ −βγ 0 0 E = γ 1−β
 0  dpx dpx
 px  −βγ γ 0 0 px  px 
  
 0  =⇒ = γ 1−β
 py   0 0 1 0 py 
 
E
0
pz 0 0 0 1 pz γ E0
= (E − βpx ) =
E E
dE dE px
where we used E 2 = p~2 + m2 = p2x + p2⊥ + m2 and 2E dp x
= 2px and dpx
= E
.
So
d̄d p0 dp0 E0
δ (d) (~p − ~k) = d δ (d) (~p0 − ~k 0 ) = x δ (d) (~p0 − ~k 0 ) = δ (d) (~p0 − ~k 0 ).
d̄ p dpx E

Which means that in F 0 we would have

D  E E0
?
~k 0 p~0 = /δ(d) (~p0 − ~k 0 ).
E
There is a special frame, it’s no good.

There is an easy fix:

2ωp~ ap†~ |0i .
p
|~pi ≡
In that case the calculation in the shaded text is replaced by
D E p
~k|~p = 4ωk ωp /δ(d) (k − p) = 2ωp /δ(d) (k − p)

47
while
~k 0 |~p0 = 2ωp ωp0 /δ(d) (k 0 − p0 ).
D E
ωp
So the expression is the same in any frame, yay.
√
Now you can ask why the factor of 2. We’d like to use these states to resolve the
P
identity in the 1-particle sector, 1 1 ≡ i |iihi|. I claim that the following expression
does this and makes Lorentz symmetry manifest:
Z
?
1 1 = d̄d+1 k θ(k 0 ) 2πδ(k 2 − m2 ) |ki hk|
| {z }
E>0
  p 
Z Z 0
δ k − k −m ~ 2 2 Z d 
d 0  d̄ k  ~ E D ~ 
= d̄ k dk0 θ(k )  |ki hk| = ω~ , k ω~k , k  .
2k 0 2ω~k k

1
P
We used the general fact δ(f (x)) = x0 |f (x0 )=0 |f 0 (x0 )| δ(x − x0 ).
So in retrospect, a quick way to check the normalization is to notice that the
following combination is Lorentz invariant:

d̄d k d̄d k 0
Z
d 2 2
= dk0 θ(k0 ) d̄ kδ(k − m ) = .
2ωk 2ωk0

Actually, this statement has a hidden assumption, that m2 > 0. In that case, the 4-
vector k µ satisfying k 2 = m2 > 0 is timelike, and no Lorentz transformation connected
to the identity can change the sign of k0 , it can only move it around within the lightcone.
So the θ(k0 ) is Lorentz invariant.
Notice that our convenient choice of normalization doesn’t show that our Hamilto-
nian description of scalar field theory is actually Lorentz invariant. For example, we
have
[φ(~x), π(~y )]ET CR = iδ (d) (~x − ~y )
at equal times, in one frame. What about other frames?
A second reason to study commutators is ...

3.1 Causality and antiparticles

Causality: This is the very reasonable condition on our description of physics that
events should not precede their causes.
It will be worthwhile to think about how to implement this condition in a QFT.
(The following discussion is based on appendix D of this paper.) Suppose B wants

48
to send a message to A. How does he do this? He applies an operator19 , call it B,
localized near B, to the shared state of their quantum many body system |ψiABE .
(Here E is for ‘environment’, the rest of the world besides A and B.) Then A measures
P
some observable A; let’s assume 1 = trA = (eigenvalues of A). To send a different
message, he should apply a different operator, say B 0 .
Under the circumstances just stated, the expectation for A’s measurement of A is

hAiB = hψ| B† e|iHt Ae−iHt

† 
{z } B |ψi = hA(t)i − B [B, A(t)] .
=A(t)

Therefore, if [B, A(t)] = 0, the expectation doesn’t depend on what B did. In fact,
replacing A with Aη for any η and using ( [B, A(t)] = 0 =⇒ [B, A(t)η ] = 0 ) shows
that all the moments of the distribution for A’s measurement will also be independent
of what B did, so no message gets through20 . [End of Lecture 6]
So nonvanishing commutators are essential for sending information. Notice that
entangled states of well-separated local degrees of freedom (such as a Bell pair of
distant spins |↑0 ↓r i − |↓0 ↑r i ∈ H0 ⊗ Hr ) are insufficient for sending signals: whatever
we do to the spin at 0 has no effect on the outcomes of measurements of the spin at r.
Here’s a proof: the spin operators at 0, S ~0 and the spin operators at r, S ~r commute:
~0 acts the identity on Hr and vice versa. The groundstate of a
[S0i , Srj ] = 0, since S
quantum field is somewhat similar: there is quite a bit of entanglement between nearby
points in space, but the equal-time commutators [φ(~x), φ(~y )] = 0 vanish, so it can’t be
used to send instantaneous signals.

Causality in relativistic QFT. [Peskin §2.4] In a Lorentz invariant system, ‘pre-

cede’ is sometimes a frame-dependent notion. If A is in the future lightcone of B,
19
‘Applying an operator’ is more complicated than it seems. Actually it means ‘changing the
Hamiltonian so that the time evolution operator is B’.
20
The loophole-seeking reader (ideally, this is everyone) will worry that a distribution is not in
general determined by its moments. (For example, there are perfectly normalizable
√ 3
distributions with
2a /π
finite averages but where the higher moments are all infinite, such as p(x) = x4 +a4 on the real line:

2

x = a2 , but x2n = ∞ for n > 1.) What we would really like to show is that the conditional
probability p(a|B) is independent of B, in which case for sure A couldn’t learn anything about what
B did. That is

p(a|B) = hψ| B† eiHt |ai ha| e−iHt B |ψi = hPa (t)i − B† [B, Pa (t)] .


| {z }
=Pa (t)

Does [A, B] = 0 imply that the projector onto a particular eigenvalue of a commutes with B? In a
finite dimensional Hilbert space, it does for sure, since 0 = [Aη , B] = a aη [Pa , B] is true for all η,
P

which gives infinitely many equations for [Pa , B]. In the case of infinite dimensional H I think there is
some room for functional analysis horror. On the other hand, any measurement has finite resolution.
Thanks to Sami Ortoleva and Chuncheong Lam for help with this point.

49
i.e. 0 < (xA − xB )2 = (tA − tB )2 − (~xA − ~xB )2 and tA > tB , then everyone agrees
that A is after B. This is the easy case. But if A and B are spacelike separated,
0 > (xA − xB )2 = (tA − tB )2 − (~xA − ~xB )2 , then there is a frame in which they occur
at the same time, and frames where they occur in either order. This is the dangerous
case.
So: causality will follow if, for all local operators A, B, [A(xA ), B(xB )] = 0 when-
ever xA and xB are spacelike separated, 0 > (xA −xB )2 . Recall that spacelike separated
means that there is a Lorentz frame where A and B are at the same time.
A general operator in a scalar QFT can be made from φs and ∂µ φs, so the general
statement will follow from considering commutators of
d̄d p  −ipµ xµ
Z 
µ
φ(x) = p ap~ e + ap†~ e+ipµ x |p0 =ωp~ ≡ φ(+) + φ(−) .
2ωp~
Here we have decomposed the field into positive- and negative-frequency parts. Notice
that since φ(+) (φ(−) ) only involves annihilation (creation) operators, [φ(±) (x), φ(±) (y)] =
0 for any x, y. Using the ladder algebra [a, a† ] = 1,
Z d
d̄ p −ipµ (x−y)µ µ

[φ(x), φ(y)] = e − e+ipµ (x−y)
Z 2ωp~
µ µ

= d̄d+1 p 2πδ(p2 − m2 )θ(p0 ) e−ipµ (x−y) − e+ipµ (x−y) (3.1)
| {z }
Lorentz inv’t

Here comes the slippery stuff: Suppose x − y is spacelike. Let’s choose a frame
where they (the points labelled by x, y) are at the same time, and let Λ be the Lorentz
matrix that gets us there: Λµν (x−y)ν = (0, ∆x) ~ µ ≡ x̃µ . Then we can change integration
variable to p̃µ = Λµν pν , so that pµ (x − y)µ = p̃µ x̃µ . Then
Z
(x − y)2 < 0
 
d+1 2 2 0 ~ ∆x
−ip̃· ~ ~ ∆x
+ip̃· ~
[φ(x), φ(y)] = d̄ p̃ 2πδ(p̃ − m )θ(p̃ ) e −e = 0. (3.2)
| {z }
~ → −p̃
odd under p̃ ~

We conclude that [φ(x), φ(y)] = 0 if (x − y)2 < 0, i.e. for spacelike separation.
The same argument works for [φ, π] and [π, π]. For [φ, π], the strict inequality
(x − y)2 < 0 is important.
So: vanishing equal-time commutators (for nonzero separation) plus Lorentz sym-
metry implies causality.
Notice that the argument fails if (x − y)2 > 0, since then we can’t get rid of the
time component of the exponents by a Lorentz transformation, and they don’t cancel.
It is possible to send signals inside the light cone.

50
 Now let’s think more about the bit which is nonzero:

[φ(x), φ(y)] = [φ(+) (x), φ(−) (y)] + [φ(−) (x), φ(+) (y)] .
| {z } | {z }
ˆ + (x−y)
≡∆ ˆ − (x−y)
≡∆

ˆ ± is a c-number, independent of what state it acts on. So, for

Because [a, a† ] ∝ 1, ∆
any normalized state |ψi,

ˆ + (x − y) |ψi = h0| ∆
∆+ (x − y) = hψ| ∆ ˆ + (x − y) |0i
= h0| [φ(+) (x), φ(−) (y)] |0i
= h0| φ(+) (x)φ(−) (y) |0i − h0| φ(−) (y) φ(+) (x) |0i
| {z }
=0:φ+ 3 a, a |0i = 0
= h0| φ(x)φ(y) |0i ≡ D(x − y)

where in the last step we again used φ+ |0i = 0. 21

This is the vacuum-to-vacuum
amplitude, or propagator22 , in the sense that

φ(y) |0i = |‘particle created at y’i

h0| φ(x) = h‘particle destroyed at x’|

That is, ∆+ is the amplitude to propagate the excitation created by the field from x
to y. Z d
+
X d̄ p −ip·(x−y)
∆ (x − y) = (x · − − − ← − − − · y) = e |p0 =ωp~ .
p
2ωp~

This integral can be done in terms of functions with names23 , but the most useful
information is about its asymptotics in the very timelike ( t ≡ |x0 − y 0 |  |~x − ~y | )
and very spacelike ( |x0 − y 0 |  |~x − ~y | ≡ r ) limits.
(
e−imt , |x0 − y 0 |  |~x − ~y |, (x − y)2 ≡ t2
∆+ (x − y) '
e−mr , |x0 − y 0 |  |~x − ~y |, (x − y)2 ≡ −r2

You can read more about how to arrive at these expressions in Peskin (page 27); the
spacelike case related by a Lorentz boost (to the rest frame) to the calculation of the
Yukawa potential that we did in the last section.
Notice that this quantity hφ(x)φ(y)i is not zero outside the lightcone. There are
nonzero spatial correlations (and indeed entanglement) in the scalar field vacuum.
21
Note from the definition that ∆± (x − y) = [φ(±) (x), φ(∓) (y)] = −∆∓ (y − x).
22
We’ll understand its connection to the time-ordered propagator just below.
23
Specifically, in four spacetime dimensions and spacelike separation, (x − y)2 ≡ −r2 , ∆+ (x − y) =
m
2π 2 r K1 (r).

51
What gives? Nothing. Causality only requires the vanishing of commutators outside
the lightcone, which we already showed in (3.2).
The cancellation in (3.2) can be interpreted as destructive interference between
particles and antiparticles. It’s clearer for the complex scalar field, where

d̄d p −ipx d̄d p +ipx †

Z Z
(+) (−)
Φ = p e ap , Φ = p e bp
2ωp 2ωp

(with the expressions for the + and − frequency components for Φ? following by taking
hermitian conjugates).
So consider the analogous24

D(x − y) ≡ h0| Φ(x)Φ? (y) |0i = h0| [Φ+ (x), Φ?− (y)] |0i = ∆+ (x − y)
| a {z }
from [a, a† ]
? ? ?+ −
D (y − x) ≡ h0| Φ (y)Φ(x) |0i = h0| [Φ (y), Φ (x)] |0i = − ∆−b (y − x)
| {z }
from [b, b† ]

So if we consider the commutator,

h0| [Φ(x), Φ? (y)] |0i = D(x− y) − D? (y − x) 

X
= (x · − − − ← p − − − ·y) − (x · − − −p → − − − · y)
| {z } | {z }
p
particle antiparticle!

then in the spacelike case, the antiparticle bit from the first term of the commutator
cancels the particle bit of the second, as above in (3.2). Antimatter makes QFT causal.

3.2 Propagators, Green’s functions, contour integrals

[This subsection is mostly in a different color because I am mostly skipping it in lecture:

we already learned all these things from the point of view of the path integral in the
last section. I’m leaving it in because it has nice pictures, and because it has a possibly-
useful summary of the discussion of choosing the contour for the frequency integrals in
the propagator.]
24
Notice the extremely annoying fact that the positive frequency component of Φ? , Φ?(+) , is actually
the dagger of the negative frequency component of Φ:
 ?
(+)
(Φ? ) = Φ(−) .

52
 I claim that the propagator for a real free scalar field can be represented as :
Z
µ i
∆(x) = d̄d+1 p e−ipµ x 2 .
C p − m2
|R {z R }
≡ C d̄p0 d̄d p

To see that this is related to the object we discussed above, first note that the denom-
inator is p
p2 − m2 = (p0 − ωp~ ) (p0 + ωp~ ) , ωp~ ≡ p~ · p~ + m2 ,
so there are two poles, which seem to be on the real axis; this means that our integral
is ambiguous and we need more information, indeed some physical input.

We can specify the contour C by taking linear combinations of C±

which are small circles going clockwise around the poles at ±ωp~ .

These basic integrals are (for a reminder about Cauchy’s theorem see footnote 14):
Z Z
d+1 −ipµ xµ i 1 −ipx
d̄ p e 2 2
= d̄d p e |p0 =ωp~ = ∆+ (x). (3.3)
C+ p − m 2ω p
~
Z Z
−ipµ xµ i 1 −ipx let q~ ≡ −~
p
d+1
d̄ p e 2 − m2
= d̄ d
p e |p0 =−ωp~ = −∆+ (−x) = ∆− (x).
C− p −2ω p
~

If we add these up, we get the full propagator:

Z
− i
+
∆(x) = ∆ (x) + ∆ (x) = d̄d+1 p e−ipx .
C=C+ +C− p2 − m2

This is one particular choice of contour, and others are also interesting. Consider
the retarded propagator,

DR (x − y) ≡ θ(x0 − y 0 ) h0|[φ(x), φ(y)]|0i .

This is the Green’s function which governs linear response:

δ hφ(x)i
= DR (x − y).
δJ(y)

(I explain how to see this in §A; for now, notice that the θ(x0 − y 0 ) guarantees that the
response happens after the perturbation that causes it.) We can reproduce DR by rout-
ing our contour to go above the poles in the complex p0 plane: if x0 − y 0 > 0, then the
0 0 0
factor e−ip (x −y ) decays when Imp0 < 0, so we can close the contour for free in the LHP,
and we pick up both poles; if x0 − y 0 > 0, we must close in the UHP and we pick up no

53
poles and get zero. Notice that we could get the same result by replacing p0 → p0 +i in
the denominator, where  is an infinitesimal (this means that 2 = 0 and c =  for any

positive quantity c).

Another interesting way to navigate the poles is by replacing p2 −m2 with p2 −m2 +i.
This shifts the poles to
q
0
p = ±ωp 1 − i/ωp = ±ωp (1 − i) .

This is called

the Feynman contour, CF , and we saw in the last section that it arises by continuation
from Euclidean spacetime. Consider again the euclidean propagator, where we get rid
of the relative sign in the metric :
−i
Z PD
IE (~x) ≡ d̄D p PD 2 e−i i=1 pi xi .
2
i=1 pi + m
p
Its poles are at pD = ±i p~ · p~ + m2 , far from the integration contour, so there is no
problem defining it. Now consider smoothly rotating the contour by varying θ from
0 to π/2 −  in p0 ≡ eiθ pD . The Feynman contour is the analytic continuation of the
euclidean contour, and the  is the memory of this.

54
To be more precise about this ‘smooth rotation’ of the contour,
we can use Cauchy’s theorem again, with the contour at right
(the figure is from Matthew Schwarz’s book). The semicircular
arcs are at |k 0 | = Λ, and give a contribution that decays like
1
Λ
, and we must take Λ → ∞. The integrand is analytic inside
e−ikx
R
the contour, no poles, so the integral k2 −m2
= 0. Therefore

the real-frequency integral (the real axis) and euclidean integral

(minus the imaginary axis contribution, with xD = it) are equal.

In position space, the Feynman propagator is

Z
µ i
∆F (x) ≡ d̄d+1 p e−ipµ x 2 2
= θ(x0 )∆+ (x)+θ(−x0 )∆+ (−x) = θ(x0 )∆+ (x)−θ(−x0 )∆− (x).
CF p −m

If x0 > 0 (< 0), we must close the contour in the LHP (UHP) and get ∆+ (∆− ).
Recalling that ∆+ (x − y) = h0|φ(x)φ(y)|0i,

∆F (x − y) = h0| θ(x0 − y 0 )φ(x)φ(y) + θ(y 0 − x0 )φ(y)φ(x) |0i ≡ h0| T (φ(x)φ(y)) |0i .

The T is the time-ordering symbol: the operators at the earliest times go on the right,
so we can regard time as increasing to the left.
The propagator is a Green’s function. So we’ve learned that
i ˜
≡ ∆(p)
p2 − m2
is the Fourier transform of ∆(x), the momentum-space propagator (either retarded or
Feynman). From this we can see that ∆(x) is a Green’s function for the differential
operator ∂µ ∂ m + m2 in the sense that

∂µ ∂ µ + m2 ∆(x) = −iδ(x)

(by plugging in the Fourier expansion of ∆ and of the delta function, δ (d+1) (x) =
R d+1 −ipx
d̄ p e , and differentiating under the integral). Notice that this did not depend
on the choice of contour, so this equation in fact has several solutions, differing by
the routes around the poles (hence by ∆± , which are solutions to the homogeneous
equation, without the delta function). On the homework, you will show this directly
from the position space definition.
Physics preview. Here is a preview of the physics of the Feynman propagator.
Imagine we introduce some interactions, such as a cubic term in the Lagrangian, e.g.

L 3 φp (x)φn (x)φπ (x) + h.c. (3.4)

55
where the fields appearing here destroy or create particles with the names in the sub-
scripts. Here are two stories we might tell about a collection of such particles.

In both pictures, time goes to the left. In the first picture, a ∆− emits a π − ,
becoming a ∆0 at spacetime point P . This π − propagates to Q where it is absorbed
by a p, which turns into an n. In the second picture, a p emits a π + at Q, and becomes
n; that π + is absorbed by a ∆− which becomes a ∆0 .
But these two stories are related by a Lorentz boost which exchanges the relative
times of the interaction events – they must be the same story. The Feynman propagator
includes both automatically.

Antiparticles going backwards in time. This story is clearer if we discuss the

complex scalar, where particles (created by a† ) and antiparticles (created by b† ) are
distinct:
d̄d p d̄d p † +ipx
Z Z
−ipx
Φ(x) = p ap e + p bp e .
2ωp 2ωp
| {z } | {z }
≡Φ+ (x) ≡Φ− (x)

The commutator is
[Φ(x), Φ? (y)] = [Φ+ (x), Φ?− (y)] + [Φ− (x), Φ?+ (y)]
| {z } | {z }
from [a, a† ], particles from [b, b† ], antiparticles
= ∆+a (x − y) + ∆−b (x − y)
| {z }
=−∆+
b (y−x)
Z Z
i i
= d̄d+1 e−ip(x−y) + d̄d+1 e−ip(x−y) . (3.5)
C+ p2 − m2 C− p2 − m2
The propagator that we’ll really need to compute S-matrix elements is
∆F (x − y) ≡ h0| T (Φ(x)Φ? (y)) |0i

56
= θ(x0 − y 0 ) h0| Φ(x)Φ? (y) |0i
| {z }
=h0|Φ+ (x)Φ?− (y)|0i=h0|[Φ+ (x),Φ?− (y)]|0i=∆+
a (x−y), particles

+ θ(y 0 − x )
0 ?
h0| Φ (y)Φ(x) |0i
| {z }
=−h0|[Φ?+ (y),Φ− (x)]|0i=−∆+
b (x−y), antiparticles

57
3.3 Interlude: where is 1-particle quantum mechanics?

[Tong, §2.8] Consider a relativistic complex free massive scalar field Φ, with mass m.
The minimum energy of a single-particle state is ωp=0 = m (above the vacuum); in its
rest frame, the wavefunction is e−imt . Consider the change of variables:
1
Φ(~x, t) =: e−imt Ψ(~x, t) √ . (3.6)
2m

The Klein-Gordon equation is

√  
0 = ∂t2 Φ − ∇2 Φ + m2 Φ = 2me−imt Ψ̈ − 2imΨ̇ − ∇2 Ψ ,

where the terms with m2 cancel; so far we’ve just p changed variables. The non-
2 2 2 p2
relativistic limit is |~p|  m which says ωp = p~ + m ' m + 2m + · · · . The
−imt
point of factoring out the phase e is that the time evolution of Ψ looks like
~2
p
e−i(E−m)t ∼ e−i 2m t ; so in the NR limit, |Ψ̈|  m|Ψ̇| meaning that we can ignore
that term in the KG equation. The remaining equation of motion is
1 2
i∂t Ψ = − ∇ Ψ. (3.7)
2m
This looks like the Schrödinger equation for a particle in no potential, in position
space, but that is a coincidence: Ψ is not a wavefunction. This equation (3.7) is the
eom associated with the lagrange density
1 ~ ? ~
L = iΨ? Ψ̇ − ∇Ψ · ∇Ψ
2m
from which πΨ = iΨ? , πΨ? = 0 (which you can also get by plugging (3.6) into ∂µ Φ? ∂ µ Φ).
The ETCRs are then

[Ψ(~x), Ψ? (~y )] = δ d (~x − ~y ), [Ψ, Ψ] = 0 = [Ψ? , Ψ? ]

and the Hamiltonian is Z

1 ~ ? ~
H= dd x
∇Ψ · ∇Ψ.
2m
The solution in terms of creation operators is then
Z Z
Ψ(x) = d̄ pe ap , Ψ (x) = d̄d pe−i~p·~x a†p
d i~p·~
x ?

with [ap , a†q ] = (2π)d δ d (p − q) as before. The hamiltonian is then

p~2 †
Z
H = d̄d p a ap (3.8)
2m p

58
(with no normal-ordering constant – the vacuum of this non-relativistic theory is
extremely boring and has no dance of birth and death of the fluctuating particle-
antiparticle pairs).
The crucial point here is that the antiparticles are gone, despite the fact that the
field is complex. They disappeared when we dropped the second-order time derivative
(recall that a first-order equation has half as many integration constants as a second-
order equation). In the limit we’ve taken, we don’t have enough energy to make any.
(More precisely, we’re promising to only study states where there is not enough energy
to make any antiparticles. ) The states are

ap†~1 · · · ap†~n |0i ≡ |{p}i , ap |0i = 0

P p~2i
and these are energy eigenstates with energy H |{p}i = i 2m |{p}i, a nice non-
relativistic (NR) dispersion for each particle. The particle-number symmetry is still
present Ψ → e−iα Ψ, but now the Noether current is
 µ  
i ?~ µ
µ
j = Ψ Ψ, ?
Ψ ∇Ψ + h.c. = ρ, ~j .
2m

Now we can find the QM of a single particle which cannot go away (since we
got rid of the antiparticles), with some position and momentum operators. In fact
the momentum operator is just the charge associated with translation invariance, and
takes the form (just like on the homework)
Z
~
P = d̄d p~pa†p ap

and P~ |{p}i = P
p~a |{p}i. What’s the position operator? A state with a particle at
a
position ~x is Z
?
|~xi = Ψ (x) |0i = d̄d pe−i~p·~x a†p |0i .

If we let Z
~ ≡
X dd xΨ? (x)~xΨ(x)

then indeed X~ |~xi = ~x |~xi. To see that the Heisenberg algebra [X, P] = i works out,
consider the general 1-particle state
Z
|ψi = dd xψ(x) |xi .

The function ψ(x) here is the usual position-basis Schrödinger wavefunction. You can
check on the homework that
Z Z  
i d i i d ∂
X |ψi = d xx ψ(x) |xi , P |ψi = d x −i i ψ(x) |xi
∂x

59
which implies the Heisenberg commuator. Finally, the hamiltonian (3.8) gives the time
evolution equation
∇2
i∂t ψ = − ψ
2m
which really is the Schrödinger equation.
Many particles which one studies in NR QM are actually fermions (e, p, n...) and
therefore not described by a scalar field. But in the 1-particle sector, who can tell? No
one. Later we’ll see the NR limit of the Dirac equation, which is basically the same,
but with some extra juicy information about spin.
Next we will speak about ‘interactions’. This term is used in two ways. In NR QM,
it is sometimes used to describe an external potential V (x) appearing as an extra term
in the Schrödinger equation

∇2
i∂t ψ = − ψ + V (x)ψ(x).
2m
Such a term explicitly violates translation symmetry. It can be accomplished by adding
to the action the quadratic term
Z Z
∆SV = − d xΨ (x)Ψ(x)V (x) = − dd xρ(x)V (x).
d ?

This says that nonzero density of particles at x costs energy V (x). A second sense of
‘interaction’ which is how it will be used forever below is interaction between particles.
With only one particle this cannot happen. NR QM theory does accommodate more
than one particle, and we can consider an interaction between them like
Z Z
∆S = − d x dd yΨ? (x)Ψ(x)V (x − y)Ψ? (y)Ψ(y).
d

If V (x − y) = δ d (x − y), this interaction is local. [End of Lecture 7]

60
4 Interactions, not too strong

4.1 Where does the time dependence go?

[Peskin chapter 4.2] Now we must get straight where to put the time dependence. Dif-
ferent ways of doing the book-keeping are called different ‘pictures’. At some reference
time, say t = 0, they all agree: label states by |ψ, 0i and operators by O(0). At a later
time, in picture P , these evolve to |ψ, tiP , OP (t). Physics, such as any amplitude like

P hφ, t| OP (t) |ψ, tiP (4.1)

is independent of P . Let’s assume the hamiltonian H has no explicit time dependence.

In Heisenberg picture (P = H), |ψ, tiH ≡ |ψ, 0i for all time, and the burden of the
time dependence is all on the operators

OH (t) = UH† (t)O(0)UH (t).

The Heisenberg equations of motion are

iȮH = [OH , HH ] (4.2)

so in particular ḢH = 0 so HH (t) = HH (0) = H. Then (4.2) is solved by UH (t) = e−itH .

Another example of an operator is the free field:

d̄d p
Z
ap e−ipx + b†p eipx


ΦH (~x, t) = p
2ωp
which you time-evolved this way on the homework. In fact, this equation is basically
the whole story of free field theory. The field makes particles which don’t care about
each other.
In Schrödinger picture (P = S), dtd OS = ∂t OS time dependence of operators comes
only from explicit, external dependence in the definition of the operator (which will
not happen here), so OS (t) = O(0), and (4.1) then requires

|ψ, tiS = UH (t) |ψ, 0i .

And the unitary evolution operator is

UH (t) = e−iH(0)t = e−iHS t = US = U

so does not require a picture label.

Interactions. So, in an interacting field theory, all we need to do is to find U to
figure out what it does. For better or worse, this isn’t a realistic goal in general. In

61
this class we are going to focus on the special case where the interactions are weak, so
that the hamiltonian takes the form

H = H0 + V

where H0 is quadratic in fields (linear terms are allowed but annoying) and we assume
that the interaction term V is proportional to a small parameter. This by no means
exhausts all interesting questions in field theory; on the other hand, a surprisingly
enormous amount of physics can be done using this assumption.
Interaction picture. (P = I) In this case, it is very convenient to use a hybrid
picture where the time-dependence of the operators is as in the Heisenberg picture for
the hamiltonian with V → 0. This free field evolution is solvable:

OI (t) ≡ U0† O(0)U0 , U0 (t) ≡ e−iH0 t . (4.3)

Note that in this picture, H0 (t) = H0 (0) = H0 . Equivalently, iȮI = [OI , H0 ], where in
this expression, crucially, H0 = H0 (ΦI ) is made from interaction picture fields, whose
evolution we know from above; for example, for a complex scalar,

d̄d p
Z
ap e−ipx + b†p eipx .


ΦI (~x, t) = p
2ωp

The catch is that the interaction-picture states are still time-dependent:

(4.1)!
H hϕ, t| OH (t) |ψ, tiH = I hϕ, t| OI (t) |ψ, tiI
| {z } | {z }
†
=hψ,0|UH (t)O(0)UH (t)|ψ,0i =U0† (t)O(0)U0 (t)

∀ϕ, ψ which says that

|ψ, tiI = U0† (t)UH (t) |ψ, 0i ≡ UI (t) |ψ, 0i , that is, UI (t) = U0† (t)UH (t). (4.4)

In the interaction picture, the interaction hamiltonian itself evolves according to

d
i VI = [VI , H0 ] =⇒ VI (t) = U0† V (0)U0 .
dt
So for example, if V (0) = dd xgφ3 (x, 0), then using 1 = U0 U0† repeatedly,
R

Z Z Z
† † 3 † † †
U0 V (0)U0 = g d xU0 φ (x, 0)U0 = g d xU0 φ(x, 0)U0 U0 φ(x, 0)U0 U0 φ(x, 0)U0 = g dd x (φI (x, t))3 .
d d

This trick wasn’t special to φ3 and works for any local interaction:

(V (t))I = V |t=0 (φI (t))

62
– just stick the interaction-picture-evolved fields into the form of the interaction at
t = 0, easy.
How do the states evolve? Notice that [U0† , UH ] 6= 0, if the interactions are inter-
esting. So
 

∂t |ψ, tiI = ∂t (UI (t) |ψ, 0i) = ∂t U0† (t) U (t) |ψ, 0i
 
| H{z }
  =e−iH(0)t =e−i(H0 +V )t

= U0† iH0 − iH  UH |ψ, 0i

| {z } | {z }
=−iV (0) †
=UH (t)U0 (t)|ψ,tiI

= −iU0† V (0) UH UH U0 |ψ, tiI = −i U0† V

†
(0)U |ψ, tiI . (4.5)
| {z } | {z }0
=1 =V (t)

That is
i∂t |ψ, tiI = V (t) |ψ, tiI .
Alternatively, the interaction-picture evolution operator satisfies

i∂t UI (t) = V (t)UI (t).

Notice how this differs from the Heisenberg evolution equation (4.2): although the full
H is time-independent, VI (t) actually does depend on t, so [V (t), V (t0 )] 6= 0, and so the
solution is not just a simple exponential. We’ll find a nice packaging for the solution
next in the form of Dyson’s expansion.
Peskin’s notation for this object is UI (t) = U (t, t0 )|t0 =0 . We can figure out how to
change the reference time from zero as follows:

|ψ, tiI = UI (t) |ψ, 0i , |ψ, t0 iI = UI (t0 ) |ψ, 0i =⇒ |ψ, 0i = UI† (t) |ψ, t0 i

|ψ, tiI = UI (t)UI† (t0 ) |ψ, t0 iI

| {z }
=U (t,t0 )

From which we infer that

0 0
U (t, t0 ) = U0† (t)UH (t)UH† (t0 )U0 (t0 ) = eiH0 t e−iH(t−t ) e−iH0 t .

From now on we drop the P = I subscripts: everything is I.

Definition of S-matrix. What are we going to do with the evolution operator?
Here is a basic (only slightly naive) three-step framework for doing particle physics
(not yet for making predictions).

63
 • At time ti , specify (e.g. measure) all the particle types, spins, momenta in the
form of an initial state |ii in the QFT Hilbert space.

• Wait. At time t, the state is

U (t, ti ) |ii = |ψ, ti .

• At time tf , measure all the particle data and get some state |f i.

Quantum mechanics says that the probability for this outcome is

| hf | U (tf , ti ) |ii |2 .
(One way in which this is significantly naive is that the space of outcomes is continuous,
so we must instead make probability distributions. More soon.) Because particle
interactions are a fast-moving business, a useful idealization is to take ti → −∞ and
tf → ∞, and let
Sf i ≡ hf | Ŝ |ii , Ŝ ≡ U (∞, −∞)
the S-matrix (‘S’ is for ‘scattering’).
This has only three ingredients: initial state, time evolution operator, and final
state. Let’s focus on the middle of the sandwich some more:
Dyson expansion. We need to solve the equation
∂t |ψ, ti = −iV (t) |ψ, ti , with initial condition |ψ, ti i = |ii .
Here’s a “solution”: Z t
|ψ, ti = |ii + (−i) dt1 V (t1 ) |ψ, t1 i
ti
as you can check by the fundamental theorem of calculus. The only small problem is
that we don’t know |ψ, t1 i. But we can use this expression for that too:
Z t  Z t1 
|ψ, ti = |ii + (−i) dt1 V (t1 ) |ii + (−i) dt2 V (t2 )
ti
Z t t
Z ti Z t1
2
= |ii + (−i) dt1 V (t1 ) |ii + (−i) dt1 dt2 V (t1 )V (t2 ) |ψ, t2 i
ti ti ti

Why stop there? Two comments: (1) This is a good idea when V ∝ λ  1. (2) Notice
the time-ordering: the range of integration restricts t1 ≥ t2 , and the earlier operator
V (t2 ) is to the right. The ad absurdam limit is
X∞ Z t Z t1 Z tn−1
n
|ψ, ti = (−i) dt1 dt2 · · · dtn V (t1 )V (t2 ) · · · V (tn ) |ii = U (t, ti ) |ii
n=0 ti ti ti
(4.6)
which since this is true for any |ii tells us a formula for U .

64
 To review, the equation we are trying to solve is:

i∂t |ψ, ti = i∂t U |ψ, ti i = i∂t U U † |ψ, ti .

| {z }
=V (t)|ψ,ti

This is true for all |ψ, ti, so it means i∂t U U † = V . Multiplying the BHS on the right
by U gives
=⇒ ∂t U = −iV U.

We might expect that an equation like this has a solution which is something like
?
U ∼ e−iV t .

Now we must deal with what Lawrence Hall calls “the wretched n!”. Starting from
our series solution (4.6),
∞
X Z t Z t1 Z tn−1
n
U (t, ti ) = (−i) dt1 dt2 · · · dtn V (t1 )V (t2 ) · · · V (tn )
n=0 ti ti ti
∞
X Z t Z t1 Z tn−1
= (−i)n dt1 dt2 · · · dtn T (V (t1 )V (t2 ) · · · V (tn ))
n=0 ti ti ti
∞ Z t Z t Z t
X 1 n
= (−i) dt1 dt2 · · · dtn T (V (t1 )V (t2 ) · · · V (tn )) (4.7)
n=0
n! ti ti ti

In the first step I used the fact that the operators are already time ordered (this followed
from the differential equation we are solving, since the V always acts from the left). In
the second step we used the fact that the time-ordered integrand doesn’t change if we
permute the labels on the times. So we can just average over the n! possible orderings
of n times. If we pull out the time-ordering symbol, this is an exponential series:
 Rt 0 0 
−i dt V (t )
U (t, ti ) = T e ti .

The time-ordered exponential is defined by its Taylor expansion.

4.2 S-matrix

Taking the times to ±∞ in the previous equation gives an expression for the S-matrix:
 R∞ 
−i −∞ dtV (t)
Ŝ = U (−∞, ∞) = T e . (4.8)

The practical value of these expressions is that they give a (compact) recipe for evalu-
ating the time evolution operator as a series in powers of the small parameter in front
of V (0): we know V (t) in terms of things like a, a† , can pull them down term-by-term.

65
 I should have called the previous expression the ‘S-operator’, since the thing we
are after is really the S-matrix elements, hf | Ŝ |ii, for which we still need hi| and |f i.
Here we encounter some small trouble. Can we just use the states like 2ωp a†p |0i
p

(the eigenstates of the free hamiltonian) which we’ve grown to love? In fact, even the
vacuum |0i is not an eigenstate of the actual H0 + V (since [H0 , V ] 6= 0), so it will not
stay where we put it. The vacuum of the interacting theory |Ωi is itself an object of
mystery (a boiling sea of virtual particles and antiparticles), and the stationary excited
states are too (a particle carries with it its disturbance of the vacuum). We’ll learn
to deal with this in perturbation theory, but here’s an expedient: pick a function f (t)
which is zero at one end, one in the middle, and then zero again at the far end. Now
replace the interaction hamiltonian V with f (t)V (t). Then, if we take ti < the time
before which the interaction turns on, and tf > the time after which we turn it off,
then we can use the free hamiltonian eigenstates. This is in fact wrong in detail, but
it will get us started.
Example. Let’s return to the ‘scalar Yukawa theory’ that we briefly encountered
earlier in (3.4). Simplifying the notation a bit, the whole Lagrangian density is
1 1 1 1
L = ∂µ Φ? ∂ µ Φ − m2 Φ? Φ + ∂µ φ∂ µ φ − M 2 φ2 + LI (4.9)
2 2 2 2
with LI = −gΦ? Φφ.
The mode expansions are

dd p
Z
ap e−ipx + a†p eipx |p0 =ωp


φ= p
2ωp

dd p
Z
bp e−ipx + c†p eipx |p0 =Ep


Φ= p
2Ep
p p
where I’ve written ωp ≡ M 2 + p2 , Eq ≡ m2 + q 2 . Notice that the Φ → e−iα Φ
symmetry is conserved; the charge is

q = Nc − Nb .

But the φ particles are not conserved.25

Artisanal meson decay. [Tong §3.2.1] The interaction LI can change the number
of φ particles. What is the amplitude for a φ particle to decay to a Φ-anti-Φ pair? So
consider
|ii = 2ωp a†p |0i , |f i = 2Eq1 2Eq2 b†q1 c†q2 |0i .
p p

25
You might notice a possible problem with this theory: what happens to the quadratic term for Φ
when φ is very negative? Let’s not take it too seriously.

66
The S-matrix element between these states is
 Z 
d+1 ? 2
hf | Ŝ |ii = hf | T 1 + (−i) d xgφx Φx Φx + O(g ) |ii

where all operators are in interaction picture. The first term dies because hf |ii = 0. In
the O(g) term, the time-ordering doesn’t matter because all the operators are at the
same time. The φ ∼ a + a† takes a one-particle state into a superposition of states with
zero and two φ-particles. We need to end up with zero φ-particles. The leading-order
nonzero term is
d̄d k −ikx √
Z Z
d+1 ?
= −ig d x hf | Φx Φx √ e ak a†p |0i 2ωk
2ωk | {z }
d
=/δ (k−p)|0i
d̄d k d̄d k
Z Z Z
dd+1 xe−ipx h0| bq1 cq2 4Eq1 Eq2 p 1 eik1 x b†k1 p 2 eik2 x c†k2 |0i
p
= −ig
Z 2Ek1 2Ek2
= −ig dd+1 xei(q1 +q2 −p)x = −ig(2π)d+1 δ d+1 (q1 + q2 − p)

This is a small victory. The delta function imposes conservation of energy and mo-
mentum on the transition amplitude. In the φ rest frame, pµp= (M, 0) which says the
amplitude is only nonzero when ~q1 = −~q2 and when M = 2 |q1 |2 + m2 . Notice that
this can only happen if M ≥ 2m.
How do we get from this amplitude to a probability? We have to square it:

2
Pf i ∼ |Sf i |2 = g 2 δ d+1 (pf − pi ) .

The square of a delta function is infinity. What did we do wrong?

Not so much, we just asked a dumb question. Here is where it helps to be a physicist.
Consider:
Z

2
δ (p) = δ (p)δ (0) = δ (p) dd+1 xei0x = δ d+1 (p)V T
d+1 d+1 d+1 d+1

where V T is the volume of spacetime – the size of the box times how long we’re willing
to wait. There is a nonzero probability per unit time per unit volume that a φ particle
in a plane wave state will decay. We’ll get its lifetime out momentarily.
For more complicated examples, it will help to streamline this process, which is the
job of §4.3. [End of Lecture 8]

4.3 Time-ordered equals normal-ordered plus contractions

We have an expression for Ŝ in (4.8) involving only time-ordered products of operators.

If we stick this between states with just a few particles, the annihilation operators in

67
there very much want to move to the right so they can get at the vacuum and annihilate
it, as is their wont. Wick’s theorem tells us how to do this, along the following lines:

Wick: T (φ...φ) = : φ...φ : + ? (4.10)

| {z } | {z }
have want

In the previous schematic non-equation, I introduced a notation for a normal-ordered

product : φ · · · φ : which means each term has all the annihilation operators to the right
of all the creation operators, for example

: φ(x)φ(y) :≡ φ− (x)φ+ (y) + φ− (y)φ+ (x) + φ+ (x)φ+ (y) + φ− (x)φ− (y). (4.11)

Normal-ordering difficulty. There is a sticky point in the definition of normal

ordering26 . Notice that we are defining normal ordering as a lexicographic operation of
moving symbols around, not as an operator. If we tried to define the normal-ordering
operation to be linear, then we would need to have
 
?
a† a  :=: aa† : − : 1 : .
:  |{z}
=aa† −1

?
But 0| : a† a : |0 = 0| : aa† : |0 = 0 so we would need to have h0| : 1 : |0i = 0, which




means we would require the shocking-looking equation:

?
: 1 := 0

that is: the normal-ordered product of a c-number would be be zero. This definition
(which, beware, differs from Peskin’s) would have the advantage that the vacuum
expectation value (VEV) of any normal ordered product is zero (with no exceptions
for c-numbers). The price is that we is that we wouldn’t be able to put the normal-
ordering symbol around the c-number bits, as Peskin does.

Instead we’ll define : c := c for c a c-number. More generally, let A, B, C be the

positive- and negative-frequency bits of some fields. Then
 
0 0 0
: ABC · · · :≡ A
| B{zC · }· |{z}
···· 
only a† s only as

Peskin writes N (· · · ) ≡: · · · :. Notice that h0| : anything except a c-number : |0i = 0.

68
 A comment about fermions. Later we will use anticommuting operators, which
have
ck c†p + c†p , ck ≡ {ck , c†p } = /δ(k − p), {c†k , c†p } = 0.

2
In particular, the equation c†p = 0 is an algebraic realization of the Pauli principle.
The cost is that even the φ− bits generate signs when they move through each other.
In that case, we define the normal ordered product as
 
0 0 0
: ABC · · · :≡ A · · · ·  (−1)P
| B{zC · }· |{z}
only a† s only as

where P is the number of fermion interchanges required to get from ABC · · · to

A0 B 0 C 0 · · · . Keeping track of these signs, and replacing commutators with anticommu-
tators, everything below goes through for fermion fields.

Let’s go back to (4.11) and compare with (4.10). Because [φ± , φ± ] = 0, the order
in the last two terms doesn’t matter. This can differ from the time-ordered product
only in the first or second term. If y 0 > x0 , it differs by [φ− (x), φ+ (y)] = −∆+ (x − y),
and if x0 > y 0 , it differs by [φ− (y), φ+ (x)] = +∆+ (x − y). Altogether:

: φ(x)φ(y) := T (φ(x)φ(y)) −∆F (x − y) ≡ T (φ(x)φ(y)) −φ(x)φ(y). (4.12)

More generally, writing φa ≡ φa (xa ), Wick’s theorem says

: φ1 · · · φn := T (φ1 · · · φn ) − (all contractions) (4.13)
where a contraction is defined as the price for moving a pair of operators through each
other to repair the time ordering, as in (4.12), and denoted by the symbol in (4.12).
For example, for four fields, the theorem says
 
T (φ1 · · · φ4 ) =: φ1 · · · φ4 + (φ1 φ2 φ3 φ4 + 5 more) : +(φ1 φ2 φ3 φ4 + 2 more)

Notes: The fully-contracted bits are numbers, so (with Peskin’s convention) it doesn’t
matter
 if they
  are inside
 thenormal-ordering
   symbol. For a product of n fields, there
n n 4 n 6 4 n
are + + ··· + (= many) contractions. But if
2 4 2 6 4 2 bn/2c
we take the vacuum expectation value (VEV) of the BHS, most terms go away.
Here’s the idea of the proof of (4.13) [Peskin page 90], which is by induction on the
number of fields m in the product. We showed m = 2 above. Assume WLOG that
x01 ≥ · · · ≥ x0m , or else relabel so that this is the case. Wick for φ2 · · · φm says
T (φ1 · · · φm ) = φ1 : φ2 · · · φm : + (all contractions w/o φ1 )
|{z}
−
=φ+
1 +φ1

69
The φ−1 term is already in the right place and can slip for free inside the normal-ordering
−
sign. The φ+1 needs to move past all the uncontracted φj≥2 s; this process will add a
term for every possible contraction involving φ1 . 

4.4 Time-ordered correlation functions by diagrams

Time-ordered correlation (or Green’s) functions of local operators will be useful:

G(n) (x1 · · · xn ) ≡ hΩ| T φH H

1 (x1 ) · · · φn (xn ) |Ωi .

Here, the operators are in Heisenberg picture for the full hamiltonian, and Ω is its
actual lowest-energy eigenstate, H |Ωi = E0 |Ωi. The fourier transform is also useful:
Z Z Pn
G̃ (p1 · · · pn ) ≡ d x1 · · · dd+1 xn e−i i pi xi G(n) (x1 · · · xn ) .
(n) d+1

In the free theory of a real scalar, we know something about these:

(2)
Gfree (x1 , x2 ) = ∆F (x1 − x2 ) =

d+1 i d+1
G̃(2) (p1 , p2 ) = /δ (p1 + p2 ) = /δ (p1 + p2 ) · (4.14)
p21 2
− m + i
The higher correlations are Gaussian, in the sense that they are sums of products of
the two point functions:
(4)
Gfree (x1 · · · x4 ) = ∆F (12)∆F (34) + ∆F (13)∆F (24) + ∆F (14)∆F (23)

= + + . (4.15)

Expectations. Our next goal is to construct a perturbative expansion in the case

of V = dd z 4!λ φ4 (z). We expect a correction of order λ of the form:

R
In

momentum space, we have blobs (unspecified sums of diagrams) with external lines

labelled by pµi : Notice that there is no need to restrict their values to the

mass shell p2i = m2 , that is, G̃(p) is nonzero even when pi are off-shell: these Green’s
functions contain “off-shell” information, more information than is available in just the
scattering matrix. However, something special will happen when the external legs are

70
on-shell. As you can see from the free two-point function, (4.14), they blow up on the
mass-shell. The existence of a singularity of G̃ on the mass-shell is a general fact, and
their residues give the S-matrix elements:
p2 →m2i Y i
G̃(p1 · · · pn ) i → S(p1 · · · pn ).
i
p2i − m2i + i

27
This is the content of the LSZ theorem, about which more later.
Perturbative expansion of time-ordered correlators. We’ll do this in three
steps: (1) Relate |Ωi to |0i. (2) Relate φH to φI . (3) Wick expand and organize the
diagrams.
Step (1): [Peskin page 86-87] Some preparations:

• Fix the additive normalization of the hamiltonian by H0 |0i = 0.

P
• Label the spectrum of H by |ni, so 1 = n |ni hn|. This is a very scary sum over
the whole QFT Hilbert space, really an integral.

• Assume that hΩ|0i = 6 0. A necessary condition for this is that the actual Hamil-
tonian is in the same phase as the H0 . Also, let’s keep the volume of space finite
for awhile.

Now consider
X X
h0| e−iHT = h0|ni hn| e−iHT = h0|ni hn| e−iEn T + h0|Ωi hΩ| e−iE0 T .
n n6=Ω

Since E0 < En for all other n, by given T a large negative imaginary part, T → ∞(1−i)
we can make the contribution of Ω arbitrarily larger than the others. Multiplying by
eiE0 T / h0|Ωi gives

h0| e−iHT eiE0 T h0|H0 =0 h0| eiH0 T e−iHT eiE0 T

   
hΩ| = lim = lim
T →∞(1−i) h0|Ωi T →∞(1−i) h0|Ωi

Since T is infinite anyway, we can shift it to T → T − t0 without change:

h0| eiH0 (T −t0 ) e−iH(T −t0 ) eiE0 (T −t0 ) (4.4) h0| UI (T, t0 )eiE0 (T −t0 )
   
hΩ| = lim = lim .
T →∞(1−i) h0|Ωi T →∞(1−i) h0|Ωi
(4.16)
27
Big picture comment: This is a long chapter. We are working our way towards a useful and
correct perturbative expansion of the S-matrix, from which we can extract observable physics.

71
Do the same for |Ωi (rather than just taking the dagger of (4.16)):

UI (t0 , −T ) |0i
|Ωi = lim .
T →∞(1−i) e−iE0 (T +t0 ) hΩ|0i

Now step (2), relate φH and φI : in terms of the field φ at some reference time t0
(could be zero), we have φI = U0† φU0 , φH = UH† φUH and therefore φH = UH† U0 φI U0† UH =
UI† φI UI , using (4.4). Now put together the Green’s function. First assume x0 > y 0 :

G(2) (x, y) = hΩ| T (φH (x)φH (y)) |Ωi

(4.16)  −iE0 (T −t0 )
−1
= lim e h0|Ωi h0| UI (T, t0 ) hΩ|
T →∞(1−i)
· UI† (x0 , t0 )φI (x)UI (x0 , t0 ) · UI† (y 0 , t0 )φI (y)UI (y 0 , t0 ) φH φH

−1 
· UI (t0 , −T ) |0i e−iE0 (T +t0 ) hΩ|0i |Ωi

−1
= lim e−i2E0 T | h0|Ωi |2 h0| U (T, x0 )φ(x)U (x0 , y 0 )φ(y)U (y 0 , −T ) |0i
T →∞(1−i) | {z }
=T (φ(x)φ(y)U (T,−T ))

In the last expression I’ve gone back to implying the I subscripts on the interac-
tion picture fields. In observing that the big underbraced product is time ordered we
are appealing to the Dyson formula  for the interaction-picture evolution operators,
R t 00 00
e.g. UI (t, t0 ) = T e−i t0 dt V (t ) – so it is a sum of time-ordered products, evaluated
between the times in the argument. Notice that I did something slippery in the first
step by combining the factors into one big limit; this is OK if each factor converges
separately. If y 0 > x0 , the relation between the first and last expressions is unchanged.
What’s the denominator? The norm of the vacuum is one, but we can assemble it
from these ingredients (4.16):

−1
1 = hΩ|Ωi = lim e−i2E0 T | h0|Ωi |2 h0| U (T, t0 )U (t0 , −T ) |0i
T →∞(1−i) | {z }
=U (T,−T )

Therefore  RT 
0 0
h0| T φ(x)φ(y)e−i −T dt V (t ) |0i
G(2) (x, y) = lim  RT 0 0 
T →∞(1−i)
h0| T e−i −T dt V (t ) |0i

The same methods give the analogous formula for G(n) (x1 · · · xn ) for any number
of any local operators. Now we can immediately perturbate to our hearts’ content
by expanding the exponentials. Let’s do some examples, then I will comment on the
familiarity of the prescription for T , and we will see that the denominator is our friend
because it cancels annoying (disconnected) contributions in the numerator.

72
 Examples. For V = 4!λ φ4 , let’s study the numerator of G(2) (x, y) in the first few
orders of λ:
 R T d+1 λ 4 
(2) −i −T d z 4! φ (z)
Gnum (x, y) = h0| T φ(x)φ(y)e |0i
−iλ
Z
= h0| T φ(x)φ(y) |0i + dd+1 z h0| T (φ(x)φ(y)φ(z)φ(z)φ(z)φ(z)) |0i + O(λ2 )
4!  
−iλ
Z
d+1
= ∆F (x − y) + d z 3φ(x)φ(y)φ(z)φ(z)φ(z)φ(z) + 4 · 3φ(x)φ(z)φ(y)φ(z)φ(z)φ(z) +
4!

The O(λ2 ) contribution is

 2 Z
1 −iλ
dd+1 z1 dd+1 z2 h0| T φ(x)φ(y)φ(z1 )4 φ(z2 )4 |0i


2! 4!
With ten fields, there will be five propagators in each diagram. The ingredients which

we must connect together are: The answer is

For example,

 2
φ(x)φ(z1 )φ(y)φ(z1 ) φ(z1 )φ(z2 ) φ(z2 )φ(z2 ) ∝

up to the symmetry factor. [End of Lecture 9]

Feynman rules for φ4 theory in position space. The set of diagrams is made
by drawing one external vertex for each xi , and m internal vertices, and connecting
them in all possible ways with propagators,

{diagrams} ≡ {A} = {A0 } ∪ {A1 } ∪ · · ·

where Am gives contributions proportional to λm . Let’s call MA the amplitude asso-

ciated to diagram A, and the Green’s function is G(n) (x1 · · · xn ) = A MA . To get
P

MA ,

73
 1
R
• Put a −iλ dd+1 za for each vertex (notice no 4!
).

• Put a ∆F (yi − yj ) for each propagator where y may be an

internal point xi or an internal point za .

• Multiply by the symmetry factor s(A). The symmetry factor is defined to be

s(A) = |Aut(A)|−1 , the inverse of the order of the automorphism group of the
diagram, that is the number of ways of permuting the ingredients of the diagram
which map the diagram to itself. Symmetries of the diagram mean that the sum
over contractions fails to completely cancel Dyson’s wretched n!1 and the 4!1 in the
interaction term. For example:
1 1   1 1
s = ·3= , s = ·4·3= .
4! 8 4! 2
Do not get hung up on this right now. Much more important to understand is
the structure of the set of diagrams, which we get started on next.

Let’s do the numerator of G(4) through order λ2 :

O(λ0 ) : + +

O(λ1 ) :

Notice that only the last term here is “fully connected” in the sense that you can’t
divide the diagram up into disjoint pieces without cutting propagators. The other
diagrams follow a simple pattern: the first three are obtained from the O(λ0 ) diagrams
by multiplying by a figure-eight bubble. The second set is obtained by mutiplying
(2) (2) (n)
G0 · G1 , where Gm denotes the order-λm fully-connected contribution to G(n) .

O(λ2 ) :

(4.17)

74
 The exponentiation of the disconnected diagrams. [Peskin page 96] There
are some patterns in these sums of diagrams to which it behooves us to attend. (The
following discussion transcends the φ4 example.) The general diagram has the form:

Only some of the components are attached to the external legs; for a given diagram
A, call the factor associated with these components Ac (note that Ac need not be
fully connected). The rest of the diagram is made of a pile of ‘bubbles’ of various
types Vi (each one internally connected, but disconnected from the external lines) and
multiplicities ni (e.g. V1 could be a figure eight, and there could be n1 = 2 of them, as
in the second term indicated in (4.17)). These bubbles (or ‘vacuum bubbles’) would be
there even if we didn’t have any external lines, and they would have the same value;
they are describing the fluctuations intrinsic to the vacuum. The amplitude associated
with the general diagram is then
V1n1 V2n2 Vαnα
MA = MA c · · ···
n1 ! n2 ! nα !
where the ni ! factors are the most important appearance of symmetry factors: they
count the number of ways to permute the identical copies of Vi amongst themselves.
The numerator of G(n) is then
(n)
 R  X X X V n1 V n2 V nα
Gnumerator = h0| T φ1 · · · φn e−i V |0i = MA = MA c 1
· 2 ··· α
A Ac
n1 ! n2 ! nα !
{ni =0}
X
= MAc · eV1 · eV2 · · · eVα
Ac
X P
Vi
= MA c e i (4.18)
Ac
P
– the bubbles always exponentiate to give the same factor of e i Vi , independent of
the external data in G. In particular, consider the case of n = 0, where there are no
external lines and hence no Ac :
R
(0)
P
Gnumerator = h0| T e−i V
|0i = 1 · e i Vi

But we care about this because it is the denominator of the actual Green’s function:
R

(n)
(n) h0| T φ1 · · · φn e−i V |0i Gnumerator X
G = R = (0) = MAc . (4.19)
h0| T e−i V |0i Gnumerator A c

75
And with that we can forget all about the bubbles. So for example,

G(2) =

G(4) =

Notice that in this manipulation (4.19) we are adding terms of many orders in
perturbation theory in the coupling λ. If we want an answer to a fixed order in λ, we
can regard anything of higher order as zero, so for example, it makes perfect sense to
write
· (1 + 8 + 88 + · · ·) eV
G(2) = + O(λ) = · + O(λ) = + O(λ).
(1 + 8 + 88 + · · ·) eV

(I only drew one kind of bubble in the previous expression since that one was easy to
type.)
Momentum space Green’s functions from Feynman diagrams. In translation-
invariant problems, things are usually a little nicer in momentum space. Let’s think
about n Z
Y
(n)
G̃ (p1 · · · pn ) ≡ dd+1 xi e−ipi xi G(n) (x1 · · · xn ).
i=1

Again, this an off-shell Green’s function, a function of general p, not necessarily

p2 = m2 . It will, however, vanish unless i pµi = 0 by translation invariance. Con-
P

sider a fully-connected contribution to it, at order λN . (We’ll get the others by

multiplying these bits.) In φ4 theory, we need to make a diagram by connecting
n external position vertices xi to N 4-valent vertices za using Feynman propagators
∆F (yA − yB ) = d̄d+1 qr e−i(yA −yB )qr q2 −mi 2 +i , where {yA } = {za , xi }. Since each propa-
R
r
gator has two ends, the number of internal lines (by the fully-connected assumption)
is
# of ends of lines n + 4N n
NI = = = + 2N.
2 2 2
The associated amplitude is then

MNFC =
Z Z Z NI
Y
P
d+1 d+1 −i i pi xi N d+1 d+1
d x1 · · · d xn e (−iλ) · s(F C) d z1 · · · d zN ∆F (yA − yB ) =
r=1
Z Z Z NI Z
P Y i
dd+1 x1 · · · dd+1 xn e−i i pi xi
(−iλ)N · s(F C) dd+1 z1 · · · dd+1 zN d̄d+1 qr e−i(yA −yB )qr .
r=1
qr2 − m2 + i
(4.20)

76
For example, consider a particular contribution with n = 4 external legs and N = 2
interaction vertices:

4+2·4
has NI = =6
2

Notice that we are doing a silly thing here of labelling the momenta of the external
lines (the first n momenta qi=1...n ). Here’s why it’s silly: Look at the integral over
x1 . Where is the dependence on x1 ? There is the external factor of e−ip1 x1 that
we put to do the Fourier transform, and there is the propagator taking x1 to z1 ,
∆F (x1 − z1 ) = d̄d+1 q1 e−i(x1 −z1 )q1 q2 −mi 2 +i . So the integral over x1
R
1

Z
d+1
dd+1 x1 e−ix1 (p1 −q1 ) = /δ (p1 − q1 )

just sets p1 = q1 , and eats the d̄d+1 q1 . The same thing happens for each external line,
R

and reduces the number of momentum integrals to NI − n.

Where is the dependence on z2 ?
Z
d+1
dd+1 z2 e−iz2 (q3 +q4 +q5 +q6 ) = /δ (q3 + q4 + q5 + q6 ).

Similarly, the z1 dependence is all in the exponentials:

Z
d+1
dd+1 z1 e−iz1 (−q5 −q6 +q1 +q2 ) = /δ (q5 + q6 − q1 − q2 ).

These two factors combine to set q1 + q2 = q3 + q4 = −q5 − q6 : momentum is conserved

at the vertices. Notice that in the example q5 − q6 is not determined.
Each internal vertex reduces the number of undetermined momenta by one. One
combination of the momenta is fixed by overall momentum conservation so we have
left
n
NI − n − (N − 1) = N − + 1 (4.21)
2
momentum integrals. This number is ≥ 0 for fully connected diagrams, and it is the
number of loops in the diagram. (This counting is the same as in a Kirchoff’s law
resistor network problem.) In the example, NL = 2 − 2 + 1 = 1 which agrees with one
undetermined momentum integral.
Here’s a proof that (4.21) is the number of loops in the diagram: place the N + n
internal and external vertices on the page. Add the propagators one at a time. You

77
must add N + n − 1 just to make the diagram fully connected. After that, each line
you add makes a new loop.
In practice now, we need not introduce all those extra qs. Label the external lines
by p1 · · · pn , and the loop momenta by kα , α = 1..NL . In the example, we might do it

like this: for which the amplitude is

X Z NL
(d+1)
Y Y i
MF C (p1 · · · pn ) = N
(−iλ) · s(F C)/δ ( pi ) d̄d+1 kα
loops,α=1
q2
lines,r r
− m2 + i
4 n=4
(−iλ)2 d+1
Z
X Y i i i
= /δ ( pi ) 2
d̄d+1 k
2! i=1
p − m2 − i
i=1 i
k2 − m + i (p1 + p2 + k)2 − m2 + i
2

(You might notice that the integral over k is in fact formally infinite, since at large k
R Λ d4 k
it goes like k2
∼ log(Λ). Try to postpone that worry.) The propagators for the
external lines just factor out, and can be brought outside the momentum integrals.
Let’s celebrate my successful prediction, for this particular graph, that there would be
poles when the external particles are on-shell, p2i = m2 . (It would be more correct to
call it Lehmann, Symanzik and Zimmerman’s successful prediction.)
So here are the momentum space Feynman rules for Green’s function in φ4 theory:

• Every line gives a factor of ˜ F (p). Notice that since

= p2 −mi 2 +i = ∆
∆F (x − y) = ∆F (y − x), the choice of how we orient the lines is not so fateful.

P d+1
dd+1 ze−i
R pi z
= (−iλ)/δ (
P
• An internal vertex gives (−iλ) i
i pi ),
P
momentum conservation at each vertex. So, set i pi = 0 at each vertex (I’ve as-
sumed the arrows are all pointing toward the vertex). After imposing momentum
conservation, the remaining consequence of the vertex is

= −iλ.

• Integrate over the loop momenta N

Q L d+1
α=1 d̄ qα for each undetermined momentum
variable. There is one for each loop in the diagram. You should think of these
integrals as just like the Feynman path integral: if there is more than one way
to get from here to there, we should sum over the amplitudes.

78
 • Multiply by the wretched symmetry factor s(A).
d+1 P
• For G̃(p), multiply by an overall /δ ( p) in each diagram.

• An external vertex at fixed position, = e−ipx . (Such vertices would arise

if we wanted to compute G(x) using momomentum-space feynman rules. ) More
generally, external vertices are associated with the wavefunctions of the states we
are inserting; here they are plane waves.

Comment on T → ∞(1 − i). What happened to the limit on T ? It’s hidden in

the integrals over the vertices:
Z Z
dz 0 dd z e−i(z ( i qi − i ~z·~qi )) · · ·
0
P P 0 P
d+1 −iz( i qi )
d ze ··· = lim
T →∞(1−i)

One end of the integral z 0 = ±∞ is going to be infinite unless i qi0 z 0 ∈ iR, in which
P

case it just oscillates. This seems scary. We can make ourselves feel better about it if
we just replace every q 0 with q 0 (1 + i) for some infinitesimal . This means that the

integrals will look like: That is:

if we use the Feynman contour for every propagator

Z
i
∆F (x) = d̄d+1 qe−iqx 2
CF q − m2 + iF

with F =  then this problem goes away.

The factors of T give another perspective on the exponentiation of the vacuum
bubbles. Consider the diagram:

4 Z
d+1 d+1
Y
= (−iλ) 2
d̄d+1 pi /δ (p1 + p2 )/δ (p1 + p2 ) · · ·
i=1

The two delta functions come from the integrals over z1,2 , and we can restore sense by
remembering this:
 2 Z
/δ (p1 + p2 ) = /δ (p1 + p2 ) dd+1 z2 = /δd+1 (p1 + p2 )2T V
d+1 d+1

79
where V is the volume of space. This factor arises because this process can happen any-
where, anytime. There is one such factor for each connected component
 of a collection
of vacuum bubbles, so for example the diagram is proportional to
(V T )2 . But the free energy ∝ log Z = log G(0) should be extensive, ∝ V T . Therefore,
the vacuum bubbles must exponentiate. [End of Lecture 10]
The whole two point function in momentum space is then (through order λ2 ) :

G̃(2) = O(λ3 )

(4.22)
I draw the blue dots to emphasize the external propagators.
Here’s an easy one: G(n) = 0 (in either position or momentum space) when n is
odd. Technically, we can see this from the fact that there is always a φ left over after
all contractions, and h0|φ|0i = 0. Slightly more deeply, this is because of the φ → −φ
symmetry.

4.5 Interlude: old-fashioned perturbation theory

[Schwartz, chapter 4] I want to take a brief break from the inexorable building of
theoretical machines to demonstrate some virtues of those machines. It will explain
what I was really mumbling about when I said that the Feynman propagator involves
antiparticles going backwards in time.
Consider a system which is a small perturbation of a solvable system H = H0 +
V . Suppose that the initial system H0 has a continuous spectrum, so that there are
eigenstates at every nearby energy. Then given an eigenstate of H0 , H0 |ϕi = E |ϕi, we
expect an eigenstate of H with the same energy H |ψi = E |ψi. Palpating the previous
equation appropriately gives
1
|ψi = |ϕi + V |ψi
E − H0 + i
| {z }
≡Π

(the Lippmann-Schwinger equation). This represents the perturbed eigenstate as the

free one plus a scattering term, in terms of the ‘propagator’ Π. The i is a safety factor
which helps us negotiate the fact that E − H0 is not actually invertible. To write this
entirely in terms of the known free state |ϕi, iterate. Let V |ψi ≡ T |ϕi where T is the
transfer matrix:
|ψi = |ϕi + ΠT |ϕi .

80
Now act on both sides with V to get V |ψi = T |ϕi = V |ϕi + V ΠT |ϕi, which will hold
if
 
1
T = V +V ΠT = V +V Π(V +V ΠT ) = V +V ΠV +V ΠV +V ΠV ΠV +· · · = V
1−VΠ
P
Given a complete set of eigenstates of H0 , with i |ϕi i hϕi | = 1,
Tf i ≡ hϕf | T |ϕi i = Vf i + Vf j Π(j)Vji + Vf j Π(j)Vjk Π(k)Vki + · · ·
1
where Vf i ≡ hϕf | V |ϕi i gives the first Born approximation, and Π(j) ≡ E−Ej
, and
E = Ei = Ef , energy is conserved.
For a vivid example, consider the mediation of a force by a boson field. Let
Z
1
V = e dd xΨe (x)φ(x)Ψe (x)
2
where ‘φ’ is for ‘photon’ and ‘e’ is for ‘electron’ but we’ve omitted spin and polarization
information, and got the statistics and charge of the electron wrong, for simplicity.
Consider the free eigenstates |ii = |~p1 , p~2 i , hf | = h~p3 , p~4 |. Then,
X 1
Tf i = Vf i + Vf n Vni + · · · .
|{z}
n
Ei − En
p1,2 6=p3,4
= 0

What are the possible intermediate states |ni? It has to be two e and one φ, as in
the following visualization (not a Feynman diagram in the same sense we’ve been dis-
cussing):

Time goes to the left, as always. The wiggly line represents the quantum of φ. You
see that there are two classes of possibilities: |nR i = |p3 , pγ ; p2 i , |nA i = |p1 ; pγ , p4 i.
Consider them from the point of view of particle 2. In the first (R) case, e2 absorbs a
photon emitted by particle 1, after the emission happens:
 
VniR = p3 , pγ , p2  V |p1 , p2 i = p3 , pγ  V |p1 i hp2 |p2 i

| {z }
=1

81
 Z Z
e    e
  
d x p , pγ  Ψe (x)φ(x)Ψe (x) p1 =
d 3
dd x hpγ | φ(x) |0i p3  Ψ(x)2 p1

=
2 2 | {z }| {z }
=e−i~pγ ·~x =2e−i(~p3 −~p1 )·x
d
= e/δ (~p1 − p~3 − p~γ ) (4.23)

– momentum is conserved. Note that energy is not, Ei 6= Em (or else the denominator
is zero).
The other possibility is |nA i = |p1 ; pγ , p4 i, which means e2 feels the effects of a
photon it emitted, which is later absorbed by e1 (!!).

   d
VniA = p4 pγ  V p2 = e/δ (~p2 − p~4 − p~γ ).

Altogether, to leading nonzero order,

X 1 X Z d d e2
Tf i = Vf n Vni = dd pγ /δ (p1 − p3 − pγ )/δ (p2 − p4 + pγ ) .
n
Ei − E n
n=R,A
E i − En

q
A bit of kinematics: let the φ have mass mγ , so for a given p~γ , Eγ = |~pγ |2 + m2γ .
Notice that these are real particles, they satisfy the equations of motion. For the R
case, the intermediate energy is
q
EnR = E3 + EγR + E2 = E3 + |~p1 − p~3 |2 + m2γ + E2

so
Ei − EnR = E1 + E2 − (E3 + EγR + E2 ) = E1 − E3 − EγR = −∆E − Eγ
where ∆E ≡ E3 − E1 = E2 − E4 (by overall energy conservation). Momentum conser-
vation means p~γ = p~2 − p~4 = p~3 − p~1 so
q q
EγA = |~p1 − p~3 |2 + m2γ = |~p2 − p~4 |2 + m2γ = EγR ≡ Eγ .

Therefore
Ei − EnA = E1 + E2 − (E1 + E4 + Eγ ) = +∆E − Eγ .
The sum of these factors is
X e2 e2 e2 2Eγ e2 e2
= + = 2 − E2
= 2Eγ 2 2
.
Ei − En −∆E − Eγ ∆E − E γ ∆E γ k − mγ
n=R,A

Here we defined k µ = pµ1 − pµ3 = (∆E, p~γ )µ , and k 2 = kµ k µ is a Lorentz-invariant

inner product. Ignoring the normalization factor 2Eγ , this is the Lorentz-invariant
momentum-space propagator for the φ particle with four-momentum k µ . Notice that
there is no actual particle with that four-momentum! It is a superposition of a real
particle going forward in time and its (also real) antiparticle going backward in time.
If we followed the i that would work out, too, to give the Feynman propagator.

82
4.6 From correlation functions to the S matrix

Now we resume our inexorable progress towards observable physics (such as cross sec-
tions and lifetimes). We would like to unpack the physics contained in the correlation
functions which we’ve learned to compute in perturbation theory. The first interesting
one is the two-point function.
Recall our expression for the momentum-space two-point function (4.22) in terms
of a sum of connected diagrams, ordered by the number of powers of λ. Let’s factor
out the overall delta function by writing:
d+1
G̃(2) (p1 , p2 ) ≡ /δ (p1 + p2 )G̃(2) (p1 ).

It will be useful to re-organize this sum, in the following way:

Here’s the pattern: we define a diagram to be one-particle irreducible (1PI) if it cannot

be disconnected by cutting through a single internal propagator. So for example,
is 1PI, but is not; rather, the latter contributes to the bit with two
1PI insertions. Then

G̃(2) (p) = +· · ·

So that we may write equations without pictures, let

−iΣ(p) ≡

denote the 1PI two-point function. Σ being 1PI means that the external lines sticking
out of it are ‘nubbins,’ placeholders where propagators may be attached. That’s why
there are no blue dots at the ends.

83
 Now suppose we know Σ. It is known as the self-energy, for reasons we will see
next. Then we can write
i i i i i i
G̃(2) (p) = 2
+ 2 2
(−iΣ(p)) 2 2
+ 2 2
(−iΣ(p)) 2 2
(−iΣ(p)) 2 + ···
p2
− m0 p − m0 p − m0 p −!m0 p − m0 p − m20
 2
i Σ Σ
= 2 1+ 2 + + ···
p − m20 p − m20 p2 − m20
i 1 i
= 2 2 Σ
= 2 2
. (4.24)
p − m0 1 − p2 −m2 p − m0 − Σ(p)
0

We see that the self-energy shifts the m2 of the particle, it moves the location of the
pole in the propagator. In the interacting theory, m20 + Σ(p)|pole is the physical mass,
while m0 (what we’ve been calling m until just now) is deprecatingly called the ‘bare
mass’. For p2 ∼ m2 , we will write
 
(2) iZ
G̃ (p) ≡ + regular bits (4.25)
p2 − m2
This equation defines the residue Z which is called the ‘wavefunction renormalization
factor’. It is 1 in the free theory, and represents the amplitude for the field to create
a particle, and the other terms, which are not singular at p2 = m2 , represent the
amplitude for the field to do something else (such as create multiparticle states), and
are absent in the free theory. Later (in 215B) we will see that unitarity requires Z < 1.
Notice that if we know Σ only to some order in perturbation theory, then (4.24) is still
true, up to corrections at higher order.
The notion of 1PI extends to diagrams for G̃(n>2) (p1 · · · pn ). Let

(n)
G̃1P I (p1 · · · pn ) ≡

where the blob indicates the sum over all 1PI diagrams with n external nubbins (notice
that these do not have the blue circles that were present before). This means G1P I
does not include diagrams like:

or .

Notice that 1PI diagrams are amputated – their external limbs have been cut off.

84
 This is almost what we need to make S-matrix elements. If we multiply the n-
Q p2 −m2
point function by ni=1 i√Z we cancel out the propagators from the external legs.
This object is naturally called the amputated n-point function. (It differs from the

1PI n-point Green’s function because of diagrams like this one

which is amputated but not 1PI.) If we then take p2i → m2 , we keep only the part of
G̃ which is singular on the mass-shell. And here’s why we care about that:

Claim (the LSZ reduction formula):

n+m
!
Y Pa2 − m2
Sf i ≡ h~p1 · · · p~n |S|~k1 · · · ~km i = lim √ G̃(n+m) (k1 · · · km , −p1 · · · − pn )
a=1
Pa0 →EP~
a i Z
(4.26)

where Pa ∈ {pi , ki }. In words: the S-matrix elements are obtained from Green’s
functions by amputating the external legs, and putting the momenta on-shell. Notice
that choosing all the final momenta pi different from all the initial momenta ki goes a
long way towards eliminating diagrams which are not fully connected.
This formula provides the bridge from time-ordered Green’s functions (which we
know how to compute in perturbation theory now) and the S-matrix, which collects
probability amplitudes for things to happen to particles, in terms of which we may
compute cross sections and lifetimes. [End of Lecture 11]
Let us spend just another moment inspecting the construction of this fine con-
veyance.

Why is LSZ true? Here’s the argument I’ve found which best combines concision
and truthiness. [It is mainly from the nice book by Maggiore §5.2; I also like Schwartz’
chapter 6; Peskin’s argument is in section 4.6.] The argument has several steps. The
field operators in this discussion are all in Heisenberg picture.

1. First, for a free field, the mode expansion implies that we can extract the ladder
operators by:
√
Z
2ωk ak = i dd x eikx (−iωk + ∂0 ) φfree (x)

85
 √
Z
2ωk a†k = −i dd x e−ikx (+iωk + ∂0 ) φfree (x) (4.27)

Notice that the LHS is independent of time, but the integrand of the RHS is not.

2. Now, recall our brontosaurus expedient (introduced previously after (4.8)): turn
the interactions off at t = ±∞.28 This allows us to write the field in terms of
some pretend free fields of mass m (not m0 !)
(t→−∞ 1
Z 2 φin (x)
φ(x) t→+∞ 1 .
Z 2 φout (x)
√
The factors of Z are required to get the correct two point functions (4.25) near
the mass shell. The mode operators for φin are called a(in) etc. φin, out are free
fields: their full hamiltonian is H0 . They are in Heisenberg picture, and the
reference time for φin, out is ±∞ respectively. Since they are free fields, we can
use (4.27) to write
√
Z Z
−ikx −1/2
2ωk a (in)† d
= −i d x e (+iωk + ∂0 ) φin (x) = −iZ dd x e−ikx (+iωk + ∂0 ) φ(x)|t→−∞

where in the second step we used the independence on time in (4.27), even though
φ(x) is not a free field. An expression for a(out)† obtains if we take t → +∞
instead.

3. Now make this expression covariant using the fundamental theorem of calculus:
Z ∞
√
Z 
(in)† (out)† −1/2 d −ikx


2ωk a −a = iZ dt∂t d xe (iωk + ∂0 ) φ(x)
−∞  
Z
−1/2
 −ikx 2 µ
= iZ dd+1 x 
e ∂0 φ − φ · ∂02 e−ikµ x 
| {z } 
~ 2 −m2 )e−ikx
(∇
Z
IBP
= iZ −1/2 dd+1 xe−ikx 2 + m2 φ(x)


(4.28)

In the last step we made a promise to only use wavepackets for external states,
so that we can do IBP in space.

4. Now, here’s where the S-matrix enters. Assume none of the incoming momenta
ki is the same as any outgoing momentum pj .

hp1 · · · pn | S |k1 · · · km i
28
Here’s why this is really bad: nearly everything we might scatter is a boundstate. For example:
atoms, nuclei, nucleons etc... But if there are no interactions there are no boundstates.

86
 Y√ Y Y
= 2ω hΩ| aout
p S ain†
k |Ωi
p,k
Y√ Y Y 
= 2ω hΩ| T aout
p S ain†
k |Ωi a
out
lives at t = +∞

p,k
m
!
Y√ Y 
in† out†
Y
= 2ω hΩ| T aout
p S ak1 − ak1 ain†
k |Ωi since pi 6= kj , use h0| a
out†
=0

p,k 2
Z m
!
(4.28) Y Y
−1/2 d+1 −ik1 x1
aout 2
ain†


= iZ d x1 e hΩ| T p S 2+m φ(x1 ) k |Ωi
2 !
Z Y m
Y
= iZ −1/2 dd+1 x1 e−ik1 x1 2 + m2 hΩ| T aout ain†


p Sφ(x1 ) k |Ωi + X
2

In the last step, X comes from where the 2x1 hits the time ordering symbol. This
gives terms which will not matter when we take k 2 → m2 , I promise.

5. Now do this for every particle to get

hp1 · · · pn | S |k1 · · · km i = m d yj e+ipj yj iZ −1/2 (2j + m2 )

Q R d+1
Qj=1
n
dd+1 xi e−iki xi iZ −1/2 (2i + m2 ) hΩ| T φ(xi ) · · · φ(yj )S |Ωi + X
R
i=1

The x and y integrals are just Fourier transforms, and this says that near the
mass shell,
n+m √
Y i Z
(n+m)
G̃ (k1 · · · km , −p1 · · · − pn ) = 2 − m2
hp1 · · · pn | S |k1 · · · km i + regular
a
P a

(where Pa ∈ {pj , ki }) which is the same as (4.26).

Comment: In our discussion of QFT, a special role has been played by fields called
φ. Suppose we have some other (say hermitian) local operator O such that

hp| O(x) |Ωi = ZO eipx

where hp| is a one-particle state made by our friend φ (we could put some labels, e.g. for
spin or polarization or flavor, on both the operator and the state, but let’s not). Such
an O is called an ‘interpolating field’ or ‘interpolating operator’. And suppose we have
information about the correlation functions of O:
(n)
GO (1 · · · n) ≡ hΩ| T (O1 (x1 ) · · · On (xn )) |Ωi .

In this case, there is a more general statement of LSZ:

Q  −1/2 R d+1 −ipa xa 2

a∈i Za i d xa e (2a + ma )

87
  
−1/2 R d+1 +ipb xb (n)
m2b )
Q
b∈f Zb i d xb e GO (1 · · · n)
(2b +
= h{pf }| S |{pa }i (4.29)
t→−∞ √
This more general statement follows as above if we can write Oa Za φin . This
more general formula allows us to scatter particles that are not ‘elementary’ in the
sense that they are made by the fields in terms of which we write our Lagrangian.

Here is a summary of the long logical route connecting Feynman diagrams to mea-
surable quantities:

One step is left.

88
 S-matrix from Feynman diagrams. The end result of the previous discussion is
a prescription to compute S-matrix elements from Feynman diagrams. In a translation-
invariant system, the S matrix always has a delta function outside of it. Also we are
not so interested in the diagonal elements of the S matrix where nothing happens. So
more useful than the S matrix itself are the scattering amplitudes M defined by
!
X X
hf | (S − 1) |ii ≡ (2π)d+1 δ (d+1) pf − pi iMf i . (4.30)
f i

d+1
(The object iM/δ
P
( p) is sometimes called the transfer matrix. The i is a conven-
tion.)
The rules for the Feynman diagram calculation of M (for φ4 theory, as a represen-
tative example) are:

1. Draw all amputated diagrams with appropriate external nubbins for the initial
and final states. For a diagram with NL loops think of NL letters that are like k
or q or p to call the undetermined loop momenta.

2. For each vertex, impose momentum conservation and multiply by the coupling
(−iλ).

3. For each internal line, put a propagator.

4. For each loop, integrate over the associated momentum d̄d+1 k.

R

A comment about rule 1: For tree-level diagrams (diagrams with no loops), ‘am-
putate’ just means leave off the propagators for the external lines. More generally, it

means leave off the resummed propagator (4.24). In particular, a diagram like

is already included by using the correct Z and the correct m.

Example: snucleon scattering. [Here we follow Tong §3.5 very closely] Let’s
return to the example with a complex scalar field Φ and a real scalar field φ with
Lagrangian (4.9). Relative to φ4 theory, the differences are: we have two kinds of
propagators, one of which is oriented, and instead of a 4-point vertex which costs −iλ,
we have a 3-point vertex for φΦ? Φ which costs −ig.
Let’s consider 2 → 2 scattering of Φ particles [recall HW 5 or see Tong §3.3.3 for
the artisanal version of this calculation], so
q
|ii = |~p1 , p~2 i , |f i = |~p3 , p~4 i with |~pi , p~j i ≡ 2Ep~i 2Ep~j bp†~i bp†~j |0i .
p

89
The Feynman rules above give, to leading nonzero order,

iM =

 
2 i i
= (−ig) 2 + 2 . (4.31)
(p1 − p3 ) − M 2 + i (p1 − p4 ) − M 2 + i

The diagrams depict two ‘snucleons’ Φ (solid lines with arrows indicating snucleons
versus antisnucleons) exchanging a meson φ (double gray line, with no arrow) with
momentum k ≡ p1 − p3 = p4 − p2 (first term) or k ≡ p4 − p1 = p2 − p3 (second term).
Time goes to the left as always. Notice that here I am being careful about using arrows
on the lines to indicate flow of particle number through the diagram, while the extra
(light blue) arrows indicate momentum flow.
The meson in these diagrams is virtual, or off-shell, in the sense that it does not
satisfy its equation of motion k 2 6= M 2 . As we saw in §4.5, each of these diagrams is
actually the sum of retarded and advanced exchange of real on-shell particles. The two
diagrams included in (4.31) make the amplitude symmetric under interchanging the
two particles in the initial or final state, as it must be because they are indistinguishable
bosons.
Two more examples with the same ingredients are useful for comparison. If we
instead scatter a snucleon and an anti-snucleon, so |ii = 2Ep~1 2Ep~2 bp†~1 cp†~2 |0i, then
p p

the leading diagrams are

iM =

 
2 i i
= (−ig) 2 + 2 . (4.32)
(p1 + p2 ) − M 2 + i (p1 − p3 ) − M 2 + i

This one has a new ingredient: in the first diagram, the meson momentum is k = p1 +p2 ,
which can be on-shell, and the i matters. This will produce a big bump,pa resonance,
√
in the answer as a function of the incoming center-of-mass energy s ≡ (p1 + p2 )2 .

90
 Finally, we can scatter a meson and a snucleon:

iM =

 
2 i i
= (−ig) 2 + 2 . (4.33)
(p + k) − m2 + i (p − k 0 ) − m2 + i

Now the intermediate state is a snucleon.

There is a common notation for the Lorentz-invariant combinations of the momenta
appearing in these various processes, called Mandelstam variables, of which s is one.
A concise summary appears in §3.5.1 of Tong’s notes.

4.7 From the S-matrix to observable physics

Now, finally, we extract some physics that can be measured from all the machinery
we’ve built.
Mediation of forces. Consider the non-relativistic (NR) limit of the snucleon-
snucleon scattering amplitude (4.31). In the center-of-mass frame p~ ≡ p~1 = −~p2 and
 2
p~ ≡ p~3 = −~p4 . In the NR limit, |~p|  m, and so p1 = m(1 + 2 mp| + · · · ).
0 0 1 |~

Energy-momentum conservation says p1 + p2 = p3 + p4 , so |~p0 | = |~p|  m as well.

In this limit, the meson propagator (in the first diagram) depends on (p1 − p3 )2 =
(p01 − p03 )2 − (~p − p~0 )2 ' −(~p − p~0 )2 , so the amplitude reduces to
 
2 1 1
iM = +ig + .
(~p − p~0 )2 − M 2 (~p + p~0 )2 − M 2

Now compare to NR QM. The scattering amplitude in the COM frame for two particles
with relative position ~r and potential U (~r) is, in the first Born approximation,
Z
0
iABorn (~p → p~ ) = −i NR h~p | U (~r) |~piNR = −i dd rU (~r)e−i(~p−~p )·~r
0 0

where the two-particle state with NR normalization is

1 1
|~piN R = √ √ |p1 , p2 i = |p1 , p2 i .
2E1 2E2 2m
The two diagrams in the relativistic answer come from Bose statistics, which means we
can’t distinguish p~ → ±~p0 from each other; to infer the potential we can just compare

91
the first diagram, (2m)2 iABorn (~p → p~0 ) = +ig 2 (~p−~p0 )12 +M 2 to find:

g 2
Z

d −i(~ p0 )·~
p−~ r 2m
d r U (~r)e = −MNR limit = −
(~p − p~0 )2 + M2

which means, in d = 3,29

g 2


2m −M r
e . U (~r) = −
4πr
This is the Yukawa potential again. It has a range, M −1 , determined by the mass of
the exchanged particle. If we take M → 0, it becomes the Coulomb potential. The
sign means that it is attractive, even though this is the potential between particle
and particle; this is a general consequence of scalar exchange. Notice that in d = 3,
! R
the Yukawa coupling between scalars has 0 = [ d4 xgφΦ2 ] = −4 + 3 + [g] so g/m is
dimensionless. [End of Lecture 12]
A brief warning: while it is satisfying to make contact with something familiar
here, the way we actually measure any such potential is by scattering the particles and
measuring cross-sections.
Lifetimes. [Schwartz, chapter 4; Peskin §4.5] How do we compute the lifetime of
an unstable particle in QFT? Consider such a particle in its rest frame, pµ = (M, ~0)µ .
Let dP be the probability that the particle decays (into some set of final states f )
during a time T . The decay rate is then dΓ ≡ T1 dP , the probability per unit time. I
put a dΓ to indicate a differential decay rate into some particular set of final states.
If we sum over all possible final states, we can make a practical, frequentist definition
of the decay rate, with the idea that we have a big pile of particles and we just count
how many go away in some time window:
# of decays per unit time 1
Γ≡ ≡ (4.34)
# of particles τ
where τ is the lifetime.
29
For convenience, here’s the integral again:
~ ∞ 1 ∞
eik·~x k 2 dk
Z Z Z Z
y≡cos θ 1 1 dkk sin kr
d̄3 k = dyeikyr =
~k 2 + M 2 (2π)2 0 k2 + M 2 −1 (2π)2 r −∞ k2 + M 2
| {z }
=2 sin
kr
kr
 Z ∞
keikr

1 1
= dk + h.c. close contour in UHP for free
(2π)2 r 2i −∞ k 2 + M 2
1 1 iM ei(iM )r e−M r
= 2πi · 2 = .
(2π)2 r 2i 2iM 4πr

92
 Fortunately for us, particles which are stable in the free theory can decay because
of weak interactions; in such a case, we can relate dP  to an S matrixelement for a
process which takes one particle to n particles, Sn←1 {pj }nj=1 ← (M, ~0) . (To get the
full decay rate Γ we would have to sum over n.) So:
1 1 |hf | S |ii|2
dΓ ≡ dP = dΠf (4.35)
T T hf |f i hi|ii
Here are two new ingredients:

(1) dΠf is the volume of the region of final-state phase space, dΠf ∝ nj=1 d̄d pj . We
Q

are allowing, as we must, for imperfect measurements. We will normalize the

R
density of final states so that dΠ = 1. Putting back the IR and UV walls of
our padded room as in (1.3), we take the continuum limit (N → ∞) of
i 2π i
xi = L, pi = , i = 1···N
N L N
which requires, for each spatial dimension,
Z Z
X L X N →∞,Lfixed 1 X 1 2π X N →∞,Lfixed
∆x = dx = L and ∆p = d̄p = L−1 .
i
N i
2π i
2π LN i

This gives
n
Y
dΠ = V d̄d pj ,
j=1

a factor of the volume of space V = Ld for each final-state particle.

(2) The normalization factors hf |f i hi|ii are not so innocent as they look, partly
because of our relativistic state normalization. Recall that |~pi = 2ωp~ ap†~ |0i, the
p

price for the relativistic invariance of which is

D E p
~k|~p = 2ωp~ 2ω~ · h0| a a† |0i = 2ωp~ /δd (~p − ~k)
k
| {zk p }
d
=h0|[ak ,a†p ]|0i=/
δ (~k−~
p)

Therefore,
Z d
d i(p=0)x
h~p|~pi = 2ωp~ /δ (0) = 2ωp~ dx e = 2ωp~ V.

Therefore,
√
|ii = 2M a~†0 |0i =⇒ hi|ii = 2M V
Y
|f i = |{~pj }i =⇒ hf |f i = (2ωj V ) (4.36)
j

where I’ve abbreviated ωj ≡ ωp~j .

93
 Now it is time to square the quantum amplitude
d+1
hf |S − 1|ii = i/δ (pT ) hf | M |ii
P P
(pT = pi − pf is the total momentum change) to get the probability (4.35). Again
we encounter a δ 2 , and again we use (2π)d+1 δ d+1 (0) = T V , so as long as f 6= i, we have
d+1 d+1 d+1
| hf | (S − 1) |ii |2 = /δ (0)/δ (pT )| hf | M |ii |2 = V T /δ (pT )|M|2

so that
n
d+1 1 Y
dP = T V /δ (pT ) Qn 2
|M| V d̄d pj
2M V j (2ωj V ) j
T
= |M|2 dΠLI (4.37)
2M
where all the factors of V went away (!), and
Y d̄d pj d+1
dΠLI ≡ /δ (pT )
final state,j
2ωj

is a Lorentz-invariant measure on the allowed final-state phase space. You can see that
this is the case by the same calculation that led us to stick those 2ωj s in the states.
One more step to physics:
1 1 T 1
dΓ = dP = |M|2 dΠLI = |M|2 dΠ
T T 2M | {z } |2M{z LI}
dynamics
kinematics

1
dΓ = |M|2 dΠLI .
2M
On the RHS is all stuff we know how to calculate (recall the Feynman rules for M that
we listed after (4.30)), and on the LHS is a particle decay rate.
The boxed formula gives the decay rate in the rest frame of the unstable particle.
In other frames, the lifetime gets time-dilated. This must be true on general grounds
of special relativity, but we can see this directly: in a general frame, the normalization
of the initial state is not hi|iirest frame = 2m but hi|ii = 2E. Therefore

Γrest frame E 1
= = ≤1
Γ m γ

and τ = τ rest frame γ ≥ τ rest frame .

94
 Cross sections. If we are not in the convenient situation of having in our hands a
big pile of particles which are stable in the free theory and decay because of not-too-
strong interactions, we need to be more proactive to get physics to come out: we have to
smash the particles together. When doing this, we send beams of particles at each other
and see what comes out. We will treat these beams as perfectly collimated momentum
eigenstates; if something goes wrong, we’ll make a more accurate representation and put
them in better-localized wavepackets. A quantity which is good because it is intrinsic
to the particles composing the beams is the scattering cross section, σ, defined by
NA NB
Number of events of interest ≡ dσ
A
where A is the common area of overlap of the beams A and B, and NA,B are the
number of particles in each beam. (Peskin does a bit more
worrying at this point, for example, about whether the beams have constant density
of particles.) By ‘events of interest’ I mean for example those particles which end up
going in a particular direction, for example in a solid angle dΩ(θ, ϕ). Restricting to
dσ
events of interest in particular direction gives the differential cross section, dΩ . The
dσ
R
notation is motivated by the idea that σ = dΩ dΩ .
The cross-section is the effective cross-sectional area of the
beam taken out of the beam and put into the particular state
of interest. Here is a picture (adapated from Schwartz’ book)
which I think makes vivid the idea behind the definition of
a cross section:
dP
Now we relate σ to the S-matrix. The scattering rate dwf i ≡ Tf i is the scattering
probability per unit time, for some fixed initial and final particle states. In a beam,
this is related to the cross section by
dw
dσ = (4.38)
j
where j is the particle current density (aka the particle flux), which for the case of
scattering from an initial state with two particles A + B → ... is
relative velocity of A and B |~vA − ~vB |
j= = .
volume V
The number of particles in each beam does not appear in (4.38) because the BHS is
intensive, and this is the point of introducing σ: it’s independent how many opportuni-
ties the particles have to interact because details of the experiment. Putting together
these statements, we can relate the cross section to the scattering probability:
dwf i dPf i V 1
dσ = = = dPf i . (4.39)
j Tj T |~vA − ~vB |

95
In (4.39) we used a practical frequentist expression for (4.38), analogous to (4.34) for
decay rates. And just as in the discussion of lifetimes above,
|hf | S |ii|2
dP = dΠf .
hf |f i hi|ii
Everything is as before except for the different initial state:

|ii = |~pA , p~B i =⇒ hi|ii = (2ωA V )(2ωB V ).

Squaring the amplitude gives

T 1
dP = |M|2 dΠLI ;
V 2ωA 2ωB
1
the only difference is that we replace 2M
with the factors for the 2-particle initial state.
1
Finally, dσ = VT |~vA −~vB |
dP gives

1 1
dσ = |M|2 dΠLI .
2ωA 2ωB |~vA − ~vB |

Again all the IR-divergent factors of V and T went away in the intrinsic physical
quantity, as they must.

4.7.1 Two-body phase space

[Schwartz §5.1] To make the formulae of the previous section more concrete, let’s sim-
plify them for the case of n = 2: two particles in the final state, whose momenta we’ll
call p1 , p2 . Note that overall momentum conservation implies p1 + p2 = pCM ; we can
use this to eliminate p2 . In that case
d̄d p1 d̄d p2
q
d+1
dΠLI = /δ (pT ) Ei ≡ |pi |2 + m2i
2E1 2E2
1 1
= 2d−(d+1)
δ(E1 + E2 − ECM ) dd p 1 p1 ≡ |p1 | > 0
4(2π) E1 E2 |{z}
=dd−1 Ωpd−1
1 dp1
1 d d−1
Ωpd−1
1 dp1
= θ(p1 )δ(x) x(p1 ) ≡ E1 (p1 ) + E2 (p2 = pCM − p1 ) − ECM
4(2π)d−1 E1 E2
CoM 1 dd−1 Ωpd−2
1 E1 E2 dp1 E1 E2 dx
= d−1
dxδ(x) θ(ECM − m1 − m2 ) dp1 = dx =
4(2π) E1 E2 ECM | {z } dx E1 + E2 p1
=1

In the last step, we went30 to the center-of-mass frame where pCM = 0, and we used
the fact that p1 ≥ 0 means E1 (p1 ) ≥ m1 , E2 (p2 = pCM − p1 ) ≥ m2 .
30
Thanks to Akhil Premkumar for pointing this out.

96
 2 → 2 scattering in d = 3. In the special case where the initial state also consists
of two particles, we can further simplify the formula for the cross section. Let the
initial momenta be kA , kB . In particular, the relative velocity factor is
 
CoM: ~kA =−~kB  |kA | |k |
B  = |kA | ECM

|~vA − ~vB | = 
 Ek +
A
EkB  EkA EkB
Therefore  
dσ 1 |~p1 |
= 2
|M|2 θ(ECM − m1 − m2 ). (4.40)
dΩ COM 64π 2 ECM |~kA |
Warning: for identical particles in the final state, one must be careful about over-
counting in the integral over angles, since a rotation by π exchanges them. In this case
dσ
σ = 21 4π dΩ dΩ
R
.

5 Spinor fields and fermions

[Peskin chapter 3] Now we need to confront the possibility of fields which transform in
more interesting ways under Lorentz transformations. To do so let’s back up a bit.

5.1 More on symmetries in QFT

Lightning summary of group theory. A group G = {gi } is a set of abstract

elements,

1. two of which can be multiplied to give a third g1 g2 ∈ G.

2. The product is associative (g1 g2 )g3 = g1 (g2 g3 )

3. and has an identity element g0 gi = gi for all gi

4. and every element has an inverse, ∀i, ∃gi−1 such that gi gi−1 = g0 .

The order of G, denoted |G|, is the number of elements of the group. G is abelian if
the product is commutative.
A Lie group is a group whose elements depend smoothly on continuous parameters,
g(θ). These then provide local coordinates on the group manifold. The dimension
of a Lie group is the number of coordinates (to be distinguished from |G|, which is
continuously infinite for a Lie group).
A (linear) representation R of a group assigns to each abstract element g of the
group a linear operator D̂R (g) on some vector space H, R : g 7→ D̂R (g) in a way which

97
respects the group law (it is a group homomorphism): meaning that D̂R (g0 ) = 1 and
D̂R (g1 )D̂R (g2 ) = D̂R (g1 g2 ).31 If we choose a basis of the vector space, then D̂R (g) is
a matrix. Two representations R and R0 are regarded as the same R ' R0 if they are
related by a change of basis on H, DR (g) = S −1 DR0 (g)S (with S independent of g!). A
reducible representation is one for which the matrices can be made block diagonal by
a basis change. A reducible representation   to R1 ⊕ R2 ⊕ ... a direct sum
is equivalent
DR1 0
 0 D
of irreducible representations, DR '  .

R2
...

The dimension of R is the dimension of H as a vector space. Notice that different

representations of the same group G can have different dimensions!
What properties of G are inherent in all of its representations? For the case of
Lie groups, one answer is the Lie algebra relations. Consider a (say n-dimensional)
representation of a group element near the identity (which let’s label the identity
element g0 ≡ e ≡ g(θ = 0) by the coordinate value θ = 0):

DR (g(θ ∼ 0)) = 1 + iθa TRa + O(θ2 ), i.e. TRa ≡ −i∂θa DR (g(θ))|θ=0

where TRa are the generators of G in the representation R. In a basis for the vector
space, they are n × n matrices.
The generators T a determine a basis of the tangent space of
G at the identity, Te G (or equivalently, by the group action,
at any other point). A finite transformation (in the compo-
nent of the Lie group which is continuously connected to the
identity element) can be written as
a
DR (g(θ)) = e−iθa TR

which is unitary if T = T † .
1 a 2 a
Given two such elements DR (g(θ1 )) = e−iθa TR and DR (g(θ2 )) = e−iθa TR , their prod-
uct must give a third:
3 a
DR (g1 )DR (g2 ) = DR (g1 g2 ) = e−iθa TR (5.1)

for some θ3 . Expanding the log of the BHS of (5.1) to second order in the θs (see
Maggiore chapter 2.1 for more detail), we learn that we must have
1
θa3 = θa1 + θa2 − θb1 θc2 f bc a + O(θ3 )
2
31
Since the overall phase of a vector in H is unphysical, quantum mechanics allows for projective
representations where the group law is only satisfied up to phases. We’ll see an example below.

98
which implies that
[T a , T b ] = ifcab T c
which relation is called the Lie algebra g of G, and the f s are called structure constants
of g or G. f does not depend on the representation. For those of you comfortable
with differential geometry, an easy way to see this is that the commutator is the Lie
bracket between two tangent vectors (which gives another tangent vector). Note that
the normalization of the T a is ambiguous, and rescaling T rescales f . A common
convention is to choose an orthonormal basis
1
trT a T b = δ ab . (5.2)
2
Notice that we often use lowercase letters to denote the algebra and uppercase letters
to denote the group, which is natural since the algebra generates small group transfor-
mations. The Lie algebra is defined in the neighborhood of the identity element, but
by conjugating by finite transformations, the tangent space to any point on the group
has the same structure, so it determines the local structure. It doesn’t know about
global, discrete issues, like disconnected components, so different groups can have the
same Lie algebra.
A casimir of the algebra is an operator made from the generators which commutes
with all of them. Acting on an irreducible representation (≡ one which is not reducible
≡ irrep), where all the states can be made from each other by the action of products
of generators, it is proportional to the identity.
Example: representations of the rotation group. This will be a fancy pack-
aging of familiar stuff which will make the step to Lorentz transformations painless (I
hope). Recall from QM that generators of rotations about the axes x, y, z = 1, 2, 3,
Ji=1,2,3 , satisfy the algebra so(3) = su(2):

[Ji , Jj ] = iijk Jk . (5.3)

2
So the structure constants are fkij = ijl δlk . A Casimir of this algebra is J2 = i (Ji ) ,
P

which acts of j(j + 1) on the spin-j representation, whose dimension is 2j + 1, any

non-negative integer. A finite rotation on H is
~
D(n̂, θ) = e−iθn̂·J

where n̂ is a unit vector and θ is an angle, so three real parameters. Familiar matrix
solutions of (5.3) are its action on vectors, where the generators are 3 × 3 matrices:
i
= iijk


J(j=1) jk

99
and its 2d representation on the Hilbert space of a spin- 21 object:

1
J(i j= 1 ) = σ i .
2 2
i
Also, its one-dimensional representation, on a scalar, has J(j=0) = 0, so eiθJ(j=0) = 1.
a a
More generally, the 2j + 1 dimensional representation is D(j) (θ) = e−iθ J with
p
J 3 mm0 = δmm0 m, J ± mm0 ≡ J 1 ± iJ 2 mm0 = δm0 ,m∓1 (j ∓ m)(j ± m + 1),

with the basis labels taking the 2j + 1 values m, m0 ∈ {−j, −j + 1 · · · j − 1, j}.

Notice that the rotation algebra (5.3) is the statement that Ji itself transforms as
a vector (j = 1) under infinitesimal rotations. What I mean by this is: the action of G
†
on H by |ψi → DR |ψi implies an action on linear operators on H by O 7→ DR ODR .
i
Relabelling the reference axes x, y, z that we used to label J by a rotation g produces
a rotation by the same angle in the 3d representation. The reference axes themselves
transform in the spin-1 representation:

D(j=1) (g)kj Jj = DR (g)Jk DR (g)† ,

the infinitesimal version of which is (5.3) (or maybe its complex conjugate). More
generally, the equation
[J i , K j ] = iijk K k
is the statement that K transforms as a vector. [End of Lecture 13]
General d. Some of what I have said so far about rotations is special to rotations
in d = 3. In particular, the notion of “axis of rotation” is (d = 3)-centric. More
generally, a rotation is specified by a (2d) plane of rotation; in d = 3 we can specify a
plane by its normal direction, the one that’s left out, J i ≡ ijk J jk , in terms of which
the so(3) lie algebra is (using  identities)

[J ij , J kl ] = i δ jk J il + δ il J jk − δ ik J jl − δ jl J ik .


(5.4)

The vector representation is

 k
ij
= i δ ik δlj − δ jk δli


J(1) l (5.5)

(that is, there is a −i in the ij entry and an i in the ji entry and zeros everywhere
else). In d = 3, the spinor representation is

1 i
J ij1 = ijk σ k = [σ i , σ j ]. (5.6)
(2) 2 4

100
For general d, we can make a spinor representation of dimension k if we find d k × k
matrices γ i which satisfy the Clifford algebra {γ i , γ j } = 2δ ij (as the Paulis do). More
on this soon.
Define the group O(d) by its action on the d-dimensional vector representation:
!
d × d real matrices O preserving lengths of vectors under ni 7→ Oi j nj : |On|2 = |n|2 =
ni nj δij , ∀ni :
Ot O = 1 or, in a basis, Ot i j δjk Ok l = δil .


(5.7)
In words: O(d) transformations preserve the bilinear form δij . Looking in the connected
ij ij
component with the identity32 , O = e−iθ J , (5.7) implies that the generators J ij ,
i, j = 1..d satisfy (5.4). Reality of O says J ij are antisymmetric and pure imaginary
matrices. There are d(d−1) 2
of them, and a good basis is given by (5.5) for general d.
This agrees with the d = 3 case above where there are 3·2 2
= 3 such generators.
 
0 −i
A special case is SO(2) where the one generator is T = = σ 2 , and the finite
i 0
transformation is
 
iβT 2 cos β sin β
e = 1 cos β + iσ sin β = .
− sin β cos β

U(N ). Another important example is the Lie group U(N ) defined by its N -
dimensional representation as N × N complex unitary matrices 1 = M † M = M M † .
This one doesn’t arise as a spacetime symmetry, but is crucial in the study of gauge
theory, and already arose as an example of a global symmetry on the homework. We
a a
can generate these M = e−iβ T by any hermitian N × N matrices T a . A basis is given
by the following set of generators satisfying (5.2):
 
  1
  1
1  1
 
 1  
1  −1 1
 
  1    1 
−2
1 2 3

T =  , T =√  , T =√  ,···
  
2 0  
12 

24  −3 
0
 
...

0
  
...  
..
.

k 1 ik j
k i ik j 2 1
Txij δ δl + δ jk δli , Tyij δ δl − δ jk δli , T N = √ 1 N ×N



for i 6= j : l = l =
2 2 2N
32 2
For real matrices Ot O = 1 says 1 = det Ot O = (det O) , so det O = ±1 gives two disconnected
components. The component with det O = 1 (containing the identity) is called SO(d). The other
component is obtained by multiplying by a diagonal matrix with an odd number of minus signs on
the diagonal.

101
Altogether there are N (N2−1) · 2 + N = N 2 of these. Only the last one has a nonzero
trace. The ones called Txij and Tyij only have nonzero entries in the ij and ji place,
and are like σ x and σ y respectively. A finite transformation is
aT a
PN 2 −1 N2 N2
βaT a
e−iβ = e|−i a=1
{z
−iβ T
} |e {z }
√
=M N 2 / 2N
=e−iβ

P 2 −1 a
where the first factor has det M = 1 (since log det M = tr log M = −i N a
a=1 β tr (T ) =
0) and the second is just a phase. The subgroup with det M = 1 is called SU(N ). This
shows that U(N ) = SU(N ) × U(1).
A special case is SU(2) which has N 2 − 1 = 22 − 1 = 3 generators, which are 2 × 2
and are T a = 12 σ a Pauli matrices. So SU(2) and SO(3) have the same lie algebra. They
are not the same group, though, since SU(2) is twice as big: a 2π rotation is not the
identity (but squares to it). Half-integer spin representations of SU(2) are projective
representations of SO(3) – the SO(3) group law is only satisfied up to a phase (in fact,
a sign).
Lorentz group. The Lorentz group can be defined, like O(d) above, as the linear
transformations preserving (proper) lengths of vectors. This implies

η = Λt ηΛ i.e. ηµν = Λt ρ
ηρσ Λσ ν .


µ
(5.8)

There are four disconnected components of solutions to this condition. As for O(d),
taking the det of both sides of (5.8) implies that such a matrix has det Λ = ±1; the
two components are called proper and improper Lorentz transformations, respectively.
The µν = 00 component of (5.8) says

2 X i
2
2
1 = Λ00 − Λ0 =⇒ Λ00 ≥ 1
i

which has two components of solutions, Λ00 ≥ 1 (orthochronous) and Λ00 ≤ −1 (not
orthochronous).

102
Below we will focus on the proper, or-
thochronous component. The other three
components are obtained by multiplying
one of these by one or both of the follow-
ing extra discrete symmetries (whose action
 
1
on (real) vectors is P = and
−1 3×3
 
−1
T = . (A warning about time
1 3×3
reversal: In order to preserve the time evo-
lution operator e−iHt while reversing t and
preserving the Hamiltonian, the time rever-
sal transformation T must also be accom-
panied by complex conjugation K : i → −i,
and the combined operation T = T ⊗ K is
therefore antilinear.)

The identity-containing component is called SO(1, d). More generally O(m, n) is

the group of linear operations preserving the matrix with m −1s and n +1s on the
diagonal, which I will also call ηµν . All the steps leading to the associated algebra
(5.4) (and generators in the m + n dimensional representation (5.5)) are the same as
for SO(d) with the replacement δij 7→ ηµν (and δij 7→ δµν ). We will nevertheless resort
to some special features of the case d = 3 to build representations.

5.2 Representations of the Lorentz group on fields

Consider a Lorentz-invariant field theory of a collection of fields φr = (φ1 ...φn , ψα , Aµ · · ·)r .

Together they form a (in general reducible) representation of the Lorentz group

φr (x) 7→ Drs (Λ)φs (Λx)

where Drs (Λ) is some matrix representation. So far we know two possibilities: the scalar
(one-dimensional) representation, where D(Λ) = 1, and the vector (d + 1-dimensional)
representation, where D(Λ) = Λ is a (d + 1) × (d + 1) matrix. (We can also take direct
sums of these to make reducible representations.)
But there are other irreps. To find more, let’s think about the algebra in more
detail by extracting it from the representation on 4-vectors
 

V µ → V 0µ = Λµ ν V ν , Λ(θa , β a ) = exp −iθa Trot

a
−iβ a Tboost
a .
|{z} | {z }
≡J a ≡K a

103
Let’s find the fcab by building the Js and Ks: with the time component first, the matrix
representations are  
i 0
J = (5.9)
Ji
where the 3 × 3 matrix is (Ji )jk = iijk and

j
0
K i 0 = iδji = K i j (5.10)

and other components zero. To check this, consider a boost in the x direction:
1 −β γ −βγ
   
1
e−iβK = 1 − iβK 1 + O(β)2 = −β 1  + O(β 2 ) = −βγ γ  + O(β 2 ).
1 2×2 1 2×2
(5.11)
0 1 1 0 2,3
That is, δV = βV , δV = βV , δV = 0. The others are related to this one
by a rotation. In (5.11), we only checked the infinitesimal transformation; but this
is enough, by the uniqueness of solutions of linear first-order differential equations:
∂β Λ(β) = −iKΛ with initial condition Λ(β = 0) = 1 has a unique solution, and so our
solution must be the correct one. We’ll use this strategy several times below. 33
Notice with slight horror that the boost generators are not hermitian, and hence
the finite boost operator is not unitary. This is a symptom of the fact that the Lorentz
group is non-compact (in the sense that its group manifold is not compact: think of
the orbits of rotations on a 4-vector (a sphere, compact), and the orbits of a boost on
a 4-vector (a hyperbola, non-compact)). For (faithful) representations of non-compact
groups, ‘unitary’ and ‘finite-dimensional’ are mutually exclusive.
The commutators of these objects are34

[J i , J j ] = iijk J k , [J i , K j ] = iijk K k (5.12)

cosh β − sinh β
 
1
33
In case you are wondering, the finite transformation is e−iβK = − sinh β cosh β . Note
1 2×2
that the parameter β here is the rapidity; it is additive under successive finite boosts, unlike the velocity
(though they agree when infinitesimal, sinh β = vc + O(v/c)2 .)
34
Like we did for O(d), we can slick this up, and generalize to other SO(1, d) by collecting the
generators into an antisymmetric matrix J µν with components J ij = ijk J k , J 0i = K i = −J i0
~ B
(exactly as E, ~ are collected into F µν ). This object satisfies the direct analog of (5.4)

[J µν , J ρσ ] = i (η νρ J µσ + η µσ J νρ − (µ ↔ ν))

and the fundamental (d + 1-dimensional vector) representation matrices solving this equation are
ρ
(J µν ) σ = i (η νρ δσµ − (µ ↔ ν)) .

104
(which are respectively the statements that the rotation and boost generators each
form a vector) and
[K i , K j ] = −iijk J k (5.13)
which says two boosts commute to a rotation. Notice that these equations are not
changed by the (parity-like) operation K → −K, J → J.
 
Now consider J~± ≡ 12 J~ ± iK~ . The observation that K → −K changes nothing
implies that they satisfy

[J+i , J−j ] = 0, [J±i J±j ] = iijk J±k ,

two independent su(2) algebras, which will be called left and right. Formally, we’ve
shown that as algebras over the complex numbers, so(1, 3) ' su(2)L × su(2)R . But
we know what the representations of su(2)L × su(2)R are! We just have to specify a
representation of each. So we can label states in an irrep by (j+ , m+ , j− , m− ) with
m± ∈ {−j± · · · + j± }; this has dimension (2j+ + 1)(2j− + 1).

(j+ , j− ) dim Preview of physics

(0, 0) 1 scalar
1
( 2 , 0) 2 left-handed Weyl spinor
1
(0, 2 ) 2 right-handed Weyl spinor
1 1 1 1
( 2 , 0) ⊗ (0, 2 ) = ( 2 , 2 ) 2 × 2 = 4 4-vector
( 12 , 0) ⊕ (0, 21 ) 2+2=4 Dirac spinor (reducible)
(1, 0) ⊕ (0, 1) 3+3=6 V µν = ± ρσ V ρσ , V µν = −V νµ , antisymmetric tensor
µν

Table 1: Lorentz representations on fields.

Let me emphasize here that we are identifying the possible ways that the Lorentz
group can act on fields, not on the particle excitations of such fields. The resulting
unitaries on the Fock space will come later. [End of Lecture 14]
Weyl spinors. Let’s focus
 on the first nontrivial entry in the table. This is a
ψ1
2-component field ψ = on which 2 × 2 Lorentz matrices act as
ψ2

D( 1 ,0) (θ, β) = e−i(θ J ).

i i +β i K i

But the fact that it’s a singlet of SU(2)R means that

1
0 = J−i ψ = (J − iK)i ψ
2

105
that is J = iK when acting on ψ, which says that the nontrivial generators act as
1 1
J+i ψ = (J + iK)i ψ = (J + J)i ψ = J i ψ.
2 2
~ = −i 1 ~σ .
But we know a 2 × 2 representation of this object: J~( 1 ) = 21 ~σ and hence K 2
2
You can check that these satisfy the three relations (5.12), (5.13). Therefore
 1 1
  1 
ψα 7→ e−i 2 θ·σ− 2 β·σ β ψβ = e− 2 σ·(β+iθ) β ψβ ≡ Mα β ψβ .
α α

Notice that this matrix M is an ordinary rotation with a complexified angle; it is

actually an SL(2, C) matrix, a general 2 × 2 complex matrix, with unit determinant.
It is common to call the ( 12 , 0) representation a left-handed (L) Weyl spinor.
1
 the (0, 2 ) or right-handed representation, χ, the same story obtains but now
 For
J~+ β̇ χβ̇ = 0 and hence J = −iK. Note the dotted indices to distinguish reps of the
α̇
two SU(2)s. Therefore
 1 1
  1 
χα̇ 7→ e−i 2 θ·σ+ 2 β·σ β̇
χβ̇ = e+ 2 σ·(β−iθ) β̇
χβ̇ = σ 2 M ? σ 2 β̇


α̇
χβ̇ . (5.14)
α̇ α̇

In the last step of (5.14) we used the identity

σ 2~σ ? σ 2 = −~σ . (5.15)

Please don’t get too hung up on dotting the indices, since as we’ll see, there are ways to
turn an L spinor into an R spinor. For example, the parity operation K → −K, J → J
interchanges the two (consistent naming these things right- and left-handed). Also,
notice that (5.14) shows that if a left-handed Weyl spinor transforms as ψ → M ψ then

σ2ψ? → σ2M ?ψ? = σ2M ?σ2 σ2ψ?

transforms like a right-handed Weyl spinor.

Invariants. In order to write Lorentz-invariant local lagrangians, we need to know
how to make Lorentz-invariant quantities out of products of fields and their derivatives.
For example, given Lorentz vectors V µ , U µ , the object V µ Uµ = V µ U ν ηµν is a Lorentz
scalar (by the defining property of the Lorentz matrices). Can we make a singlet from
two Weyl spinors, ( 21 , 0) ⊗ ( 12 , 0)? Yes: we know (e.g. from basic QM) that SU(2)
representations combine as 12 ⊗ 21 = 0 ⊕ 1 where the triplet (spin one) part is symmetric
and the singlet (spin 0) is the antisymmetric combination, ↑↓ − ↓↑. More explicitly,
ψα ξβ αβ ≡ ψα ξ α is a singlet. To see this explicitly:
1
iσ 2 ψ →7 iσ 2 e− 2 (β+iθ)·σ ψ Insert 1 = σ 2 σ 2 before ψ

106
  
 1 
~ + iθ~ · σ 2~σ σ 2  iσ 2 ψ



= exp  −
 2 β 
| {z }
(5.15)
σt
= −~

which means that if ψα 7→ Mα β ψβ , then

t  1 
ψ α ≡ iσ 2 ψ α 7→ ψ β e+ 2 (β+iθ)·~σ α
≡ ψ β M −1 α


β β

so ψ α ψα is invariant. Notice that on Weyl spinors, we raise and lower indices with
 αβ
αβ 0 1
αβ ≡ (iσ 2 ) = which is a good idea because it is an invariant tensor, as
−1 0
we just showed.
The same story holds for the (0, 21 ) “right-handed” Weyl representation, that is, we
can also make a singlet out of two of them using an epsilon tensor. Writing out the
matrices again:
t α̇
 1   1 
− 2 (β−iθ)·σ
χα̇ ≡ α̇β̇ χβ̇ 7→ e 2 (β−iθ)·σ β̇ χ β̇
= χβ̇
e α̇
β̇

which is indeed the inverse transformation of (5.14).

( 21 , 12 ) = vector. Next, we will show that the ( 21 , 12 ) = ( 21 , 0) ⊗ (0, 12 ) representation is
indeed a 4-vector. Introduce the following ‘intertwiners’:

σαµα̇ ≡ (1 αα̇ , ~σαα̇ )µ , µ

σ̄α̇α ≡ (1 α̇α , −~σα̇α )µ .

Our next job is to show that these objects eat a L and an R Weyl spinor and spit
out a vector, or vice versa. So for example, I claim that if Vµ is a vector then Vµ σαµα̇
transforms as ( 21 , 0)⊗(0, 21 ). And given any two L and R Weyl spinors ψ, χ, ψ α σαµα̇ χα̇ Vµ
is a singlet.
To summarize:
1
( 12 , 0) 3 ψL 7→ ML ψL = e 2 (−iθ−β)·σ ψL
1
(0, 12 ) 3 ψR 7→ MR ψR = e 2 (−iθ+β)·σ ψR
?
You can see from this expression and σ 2 σ i σ 2 = − (σ i ) that σ 2 ML? σ 2 = MR and
therefore σ 2 ψL? ∈ (0, 21 ).
Claim: for any two right-handed spinors ξR , ψR , the object ξR† σ µ ψR is a (complex)
1
4-vector. To see this, first notice (using σ † = σ) that ξR† 7→ ξR† e+ 2 (iθ+β)·σ , so
1 1
ξR† σ µ ψR 7→ ξR† e| 2 (iθ+β)·σ σ{z
µ 2 (−iθ+β)·σ
e } ψR
?
=Λ(θ,β)µ ν σ ν

107
where

µ
Λ(θ, β)µ ν = e−i(θ·J+β·K) ν

is the vector representation of the Lorentz transformation with rotation θ~ and boost
~ To check this, it suffices to check the infinitesimal version (by the uniqueness of
β.
solutions to linear first-order ODEs):
 
δ ξR† σ µ ψR = δξR† σ µ ψR + ξR† σ µ δψR
 
† 1 j j µ µ1 j j
= ξR (iθ + β) σ σ + σ (−iθ + β) σ ψR
2 2
 †

 ξR 21 2βj σ j ψR ...if µ = 0

  
= † 1 j i i j

j i i j 




 ξR 2  βj σ σ + σ σ +iθ j σ σ − σ σ  ψR ...if µ = i
 | {z } | {z }
=2δ ij =−2iijk σ k

On the other hand, using the form of the vector Lorentz generators (5.9), (5.10), the
transformation of a vector is
(
µ µ βj V j ...if µ = 0
δV µ = − iβj K j ν + iθj J j ν V ν =




βi V 0 − θj jim V m ...if µ = i

which is just the form we’ve found.

A few more claims and consequences:

• Since the vector representation matrices K, J in (5.9), (5.10) are pure imaginary,
the matrices are real for any θ, β and we can impose the reality condition V µ =
(V µ )? consistent with Lorentz orbits. This is not true of the Weyl spinors by
themselves.

• Similarly, for any ξL , ψL ξL† σ̄ µ ψL is a (complex) 4-vector. Notice that we need to

use the σ̄ µ , so that
   µ
† µ † † †
δ ξL σ̄ ψL = − −βi ξL σi ψL , βi ξL ψL + ijk θj ξL σk ψL

• Our explicit calculation was about ξR† σ µ ψR . But we showed that ξR† transforms
like a χL . So χαL σαµα̇ ψRα̇ is a vector, too. And χαL σαµα̇ ψRα̇ Vµ is Lorentz invariant.

• Notice that ξR† ψR is not Lorentz invariant (even if ξ = ψ), it is rather the 0
component of a four-vector. The dotted index can be a bit helpful in reminding
 α̇
us that ψL† ψL is not a singlet, since if the index on (ψL )α is undotted then ψL†
has a dotted index.

108
 On the other hand, given an R and an L spinor,

δξL† = ξL† (+iθ − β) · σ/2

the combination ξL† ψR is Lorentz invariant, since

 

†

† 1 1
δ ξL ψR = ξL (iθ − β) · σ + (−iθ + β) · σ ψR = 0. (5.16)
2 2

• Two more occasionally-useful facts:

 0
V + V 3 V 1 − iV 2

µ
V σµ = 7→ M V µ σµ M †
V 1 + iV 2 V 0 − V 3

Also, det V µ σµ = V µ Vµ is Lorentz invariant.

5.3 Spinor lagrangians

Given a Weyl 
spinor field ψR , we’d like to make a local lorentz-invariant lagrangian
of the form L ψR , ψR† , ∂µ ψR , ∂µ ψR† . The sort of obvious generalization of the KG
lagrangian is ψR† (2 + m2 ) ψR , which transforms like ψR† ψR , which is not boost invariant.
On the other hand, the object ψR† σ µ ψR is a vector, and we can find another index with
which to contract by taking a derivative:
 
LWeyl ≡ ψR† σ µ i∂µ ψR = ψR† i∂t ψR + ψR† ~σ · i∇
~ ψR

is a nice Lorentz invariant kinetic term. The factor of i is to make i∂µ hermitian, so
 
IBP
L†Weyl = −i ∂µ ψR† (σ µ )† ψR = ψR† σ µ i∂µ ψR = LWeyl .
 
Notice that we neglected the total derivative ∂µ ψR† σ µ ψR which does not change the
equations of motion.
For a left-handed field, the Lorentz-invariant Weyl lagrangian involves σ̄ µ = (1, −~σ )µ :

LWeyl (ψL ) = ψL† σ̄ µ i∂µ ψL .

What about mass terms? For a single R Weyl field, we could use the  tensor:

ψR iσ 2 ψR + h.c.

is Lorentz invariant. It is not invariant under ψR → eiθ ψR , it violates the particle

number. Neutrinos may have such a term, but electrons don’t.

109
 Dirac spinors. To make a particle-number-conserving Lorentz-invariant mass
term, we need one of each L and R, and the Dirac mass pairs them up via the in-
variant ψL† ψR + h.c.. We can slick this up, by combining the two 2-component spinors
into one 4-component spinor Ψ:
     
1 1 ψL
, 0 ⊕ 0, 3Ψ≡ .
2 2 ψR

Now let  
  0 1 2×2
Ψ̄ ≡ ψR† , ψL† =ψ †
≡ Ψ† γ 0 .
1 2×2 0
Then we can package the whole thing beautifully as
 
† † † †
LDirac = ψR iσ ∂µ ψR + ψL iσ̄ ∂µ ψL − m ψL ψR + ψR ψL = Ψ̄ (iγ µ ∂µ − m) Ψ
µ µ

with
0 σµ
 
µ
γ ≡ . (5.17)
σ̄ µ 0
The equations of motion are
∂S
0= = (iγ µ ∂µ − m) Ψ ≡ (i/
∂ − m) Ψ.
∂ Ψ̄
µ
Being explicit about indices, the Dirac equation is 0 = (iγab ∂µ − mδab ) Ψb with a, b =
1..4.
Notice that by dimensional analysis of the kinetic terms, [Ψ] = 3/2, so [m] = 1, so
m is indeed a mass. Its sign has not been fixed (and I will probably mix up m and −m
at various points).

• The Dirac (or gamma) matrices satisfy the Clifford algebra

{γ µ , γ ν } = 2η µν .

Any set of four matrices satisfying this equation can be combined to form a 4d
representation of so(1, 3) in the form

µν i
JDirac ≡ [γ µ , γ ν ].
4
If you want to see the algebra involved in this statement, see David Tong’s notes.

110
• The particular basis of gamma matrices we’ve chosen (5.17) is called the Weyl
basis. It makes the reducibility of the Dirac rep manifest, since the resulting J µν
are block diagonal:

0 σµ 0 σν
   
µν Weyl basis i
JDirac = , ν
4 σ̄ µ 0 σ̄ 0
 µ ν ν µ

i σ σ̄ − σ σ̄ 0
=
4 0 σ̄ µ σ ν − σ̄ ν σ µ
 !
i
 i −2σ 0

 = − 2i σ i ⊗ σ 3 if µ = 0, ν = i
4

0 2σ i
= ! (5.18)
i j

i
−[σ , σ ] 0 1 ijk k
= + 2  σ ⊗ 1 2×2 if µ = i, ν = j


4

0 −[σ i , σ j ]

So you see that in the Weyl basis, we already know that these satisfy the so(1, d)
algebra since it is just the Lorentz generators for the WeylL and WeylR repre-
sentations in blocks. This proves that it works in any basis, since the algebra is
basis-independent.

• This 4-dimensional Dirac representation is not the 4d vector representation. We

can see this in several ways: It is complex (the generators are not pure imaginary),
though more on this below. It is reducible (we built it by adding together two
irreps!). And it is definitely different because, using (5.18), we have, e.g. J 12 =
1 3
2
σ ⊗ 1, so

= eiπσ ⊗1 = cos π1 + sin πσ 3 ⊗ 1 = −1.

12 3
ΛDirac (θ = 2πẑ) = e−i2πJ

A 2π rotation gives minus one. This is just as for spin- 12 representation of SO(3).

• The Weyl spinors ψL , ψR are irreps. What’s the big deal about the Dirac rep?
Only that the electron is a Dirac spinor (and some other folks are too). Before we
learned about neutrino masses, they could have been Weyl spinors. Now we have
two possibilities: either there is a secret (heavy, non-interactive) partner with
whom the neutrinos pair up by a Dirac mass, and/or lepton number is violated
by a Majorana mass term (see the homework).

• Other bases of the gamma matrices are possible and sometimes useful. If we
replace γ µ with
γ µ 7→ γ̃ µ = U γ µ U † , Ψ 7→ Ψ̃ = U Ψ
for some 4 × 4 unitary U then this gives an equivalent representation, since
{γ̃ µ , γ̃ ν } = 2η µν still and hence S̃ µν = U S µν U † will still solve so(1, d).

111
 A particular useful other basis is the Majorana basis

0 σ2
 1
0 −σ 2
      3 
0 1 iσ 0 2 3 iσ 0
γm = , γm = , γm = , γm = .
σ2 0 0 iσ 1 σ2 0 0 −iσ 3

They have the property that they are all imaginary, which means so are the
µν
resulting Lorentz generators Jm , which means that all the matrix elements of
µν
iθµν Jm
e are real. So it is consistent to impose a reality condition on the spinors
?
in this basis, ηm = ηm . (The reality condition can be imposed in any basis, but
in another γm = U γ̃ U , the condition looks like (U ? )−1 U ψ = ψ ? .) This 4d real
µ µ †

representation is still different from the vector; a proof is that a 2π rotation is still
−1. A good analogy is (real scalar):(complex scalar)::(majorana spinor):(Dirac
spinor). For example, a Majorana spinor particle will be its own antiparticle, just
like for a real scalar.

• The Dirac equation (i/

∂ − m) Ψ = 0 implies the wave equation. Act on the BHS
by

(i/
∂ + m) (BHS) =⇒ 0 = (i/
∂ + m) (i/∂ − m) Ψ
= −γ γ∂µ ∂ν − m2 Ψ
µ ν


 
1
= −  [γ µ , γ ν ] + {γ µ , γ ν } ∂µ ∂ν −m2  Ψ
2 | {z } | {z } |{z}
antisymmetric =2η µν symmetric
= − ∂ 2 + m2 Ψ.

• The equation of motion for Ψ̄ can be obtained by taking the dagger35 , or by IBP
in LDirac :  ← 
IBP
LDirac = Ψ̄ −i ∂ µ γ µ − m Ψ + total deriv
 ← 
so 0 = ∂S∂Ψ Dirac
= Ψ̄ −i ∂ µ γ µ − m .

• The Dirac lagrangian is real if m = m? , since we already checked the kinetic

terms.
µν
−iθµν JDirac
• Lorentz transformations
  Dirac spinors. We have Ψ 7→ e
of Ψ ≡
M 0
Λ 1 Ψ with Λ 1 = . The Dirac conjugate spinor transforms as
2 2 0 σ M ?σ2
2

µν †
Ψ̄ 7→ Ψ† e+iθµν (JDirac ) γ 0 = Ψ† γ 0 Λ−1
1 = Ψ̄Λ−1
1 .
2 2

35

† †
In doing so, note that γ 0 = γ 0 , but the spatial ones (~γ ) = −~γ are anti-hermitian. This
compensates the fact that that the ~γ s acquire a minus sign in moving through the γ 0 in Ψ̄.

112
 Here we used that the ij components of JDirac are hermitian and commute with
γ 0 (same for L, R), while the 0i components are antihermitean and anticommute
with γ 0 (opposite sign for L, R). This makes it clear that the mass term Ψ̄Ψ is
Lorentz invariant.
The gamma matrices also provide nice packaging of the relation we showed above
between vectors and bispinors36 :

Λ−1 µ µ ν
1 (θ)γ Λ 1 (θ) = Λ ν (θ)γ . (5.19)
2 2

[End of Lecture 15]

This means that any product of gamma matrices between two spinors V µ1 ···µn ≡
Ψ̄γ µ1 · · · γ µn Ψ is a tensor, in the sense that V µ1 ···µn 7→ Λµ1 ν1 · · · Λµn νn V ν1 ···νn . To
see this, just use V µ1 ···µn 7→ Ψ̄Λ−1 1 γ
µ1
· · · γ µn Λ 1 Ψ, insert 1 = Λ 1 Λ−1
1 in between
2 2 2 2
each pair of gammas, and use (5.19).
Notice that any combination of Aµ1 ···µn Ψ̄γ µ1 · · · γ µn Ψ which is symmetric under
interchange of indices can be related using the Clifford algebra to a tensor with
fewer indices. Let γ µν ≡ 12 [γ µ , γ ν ] be just the antisymmetric bit, and similarly for
more indices. In fact, any bispinor Γab Ψ̄a Ψb can be decomposed as a sum of these
tensors: n Aµ1 ···µn Ψ̄γ µ1 ···µn Ψ. This follows from counting: 4×4 = 1+4+6+4+1.
P

• Consider the object γ 5 ≡ iγ 0 γ 1 γ 2 γ 3 = − 4!i µνρσ γ µ γ ν γ ρ γ σ . The factor of i is

† 2
chosen so that (γ 5 ) = γ 5 . Notice that (γ 5 ) = 1 so its eigenvalues are ±1. Since
it contains one of each of the other four gamma matrices, it anticommutes with
µν
each of them: {γ 5 , γ µ } = 0, ∀µ. Since the Lorentz generators JDirac are quadratic
5 µν 5
in γs, this implies [γ , JDirac ] = 0, i.e. γ is a Casimir, proportional to the identity
on irreps, indeed ±1. By direct calculation, in the Weyl basis,
 
5 Weyl basis −1
γ =
1

is ±1 on right-handed and left-handed spinors, respectively. This means that the

5
chirality projectors PR/L ≡ 1±γ
2
2
project onto R/L spinors, respectively, PR/L =
µ µ
PR/L . Notice that PL γ = γ PR .
36
This is really the same equation as we showed above. In terms of Dirac fermions, its infinitesimal
version is,
µ
(1 + iθ · JDirac ) γ µ (1 − iθ · JDirac ) = (1 − iθ · Jvector ) ν γ ν ,
which follows from
µν ρσ µ
[γ µ , JDirac ] = (Jvector ) νγ
ν
.

113
 • Our basis of bispinors can be rewritten using γ 5 . Using γ µνρσ = −iµνρσ γ 5 and
γ µνρ = +iµνρσ γσ γ 5 , we can make a basis of (hermitean) bispinors as in the
table below. The modifier ‘pseudo’ here refers to the properties under parity; for

Bispinor multiplicity representation

Ψ̄1Ψ 1 scalar
µ
Ψ̄γ Ψ 4 vector
µν
iΨ̄γ Ψ 6 antisymmetric tensor
µ 5
iΨ̄γ γ Ψ 4 pseudovector
5
iΨ̄γ Ψ 1 pseudovector

 
† †

5
example, in terms of Ψ = ψL ψR , consider iΨ̄γ Ψ = i ψL ψR − ψR ψL . It is
Lorentz invariant (since it is of the form (5.16)), but under parity P : ψL ↔ ψR
(that is, parity acts by γ 0 on Dirac spinors) it goes to minus itself.
Why care about these bispinors? One reason is that we can make 4-fermion
interactions out of them. For example, Ψ̄γ µν Ψ · Ψ̄γµν Ψ is a Lorentz-invariant
local interaction term which we might add to our Lagrangian.
Another reason is that the vector combinations play an important role:

j µ ≡ Ψ̄γ µ Ψ = ψR† σ µ ψR + ψL† σ̄ µ ψL

is the Noether current associated with the symmetry Ψ → e−iα Ψ of the Dirac
Lagrangian. You can directly check that ∂µ j µ = 0 using the Dirac equation.
Similarly, the axial current

j5µ ≡ Ψ̄γ µ γ 5 Ψ = ψR† σ µ ψR − ψL† σ̄ µ ψL

5
would be the Noether current associated with the transformation Ψ → eiαγ Ψ.
This transformation rotates the L and R bits oppositely, and is only a symmetry
if m = 0. Indeed, the Dirac equation implies that ∂µ j5µ = 2imΨ̄Ψ. For E  m,
the breaking of this symmetry by m can be ignored and it is still useful37 . The
combinations
1 ± γ5
 
µ µ
jR/L ≡ Ψ̄γ Ψ
2
involve only the Weyl components and are separately conserved if both j µ and
j5µ are conserved.
37
Beware that the coupling to gauge field also breaks this symmetry, quantum mechanically. This
is called the chiral anomaly; more later.

114
 Coupling to the electromagnetic field. Here’s another purpose for the current.
Suppose our spinor field is propagating in a background electromagnetic field with
vector potential Aµ . The whole thing should be Lorentz invariant, so we should be
able to couple them via a Lorentz-invariant Lagrangian. How can we resist adding
LEM = −ej µ Aµ for some constant e (blame Ben Franklin for the sign). The full
Lagrangian is then
Ψ̄[i (∂µ + ieAµ ) γ µ − m]Ψ
and the Dirac equation is modified to

0 = (iγ µ Dµ − m) Ψ, where Dµ Ψ ≡ (∂µ + ieAµ ) Ψ

is the gauge covariant derivative in the following sense: Dµ Ψ 7→ e−iα(x) Dµ Ψ under

Ψ → e−iα(x) Ψ(x), Aµ → Aµ + 1e ∂µ α. We could have used the demand that the action
respects this transformation to determine the coupling to Aµ to replace ∂µ → Dµ .
The solutions of the Dirac equation in a background EM field are no longer solutions
of the KG equation:

/ − m Ψ = iDµ iDν γ µ γ ν − m2 Ψ .




0 = iD / + m iD

Whereas mixed partials commute, [∂µ , ∂ν ] = 0, the covariant derivatives need not:

[Dµ , Dν ] = ei (∂µ Aν − ∂ν Aµ ) = eiFµν

so the antisymmetric term matters:

 
2 i µ ν 2
0 = (∂µ + ieAµ ) + e [γ , γ ]Fµν + m Ψ
 4   
~ + iE
B ~ · ~σ
Weyl basis
= (∂µ + ieAµ )2 − e     + m2  Ψ . (5.20)
~ ~
B − iE · ~σ
µν
In the last step we used the form of the Lorentz generators JDirac = 4i [γ µ , γ ν ] in the
Weyl basis. This extra term (relative to the gauge covariant scalar wave equation) is
an intrinsic magnetic dipole moment of a Dirac particle. This is a consequence of the
Dirac equation with implications for the non-relativistic limit.
Notice that we could add extra terms coupling the spin to the EM field strength

LDM = Fµν gm Ψ̄iγ µν Ψ + ge Ψ̄iγ µν γ 5 Ψ

but now [ge,m ] = −1 so these coefficients (which would change the magnetic and electric
dipole moments respectively) are suppressed by the inverse of some new mass scale (a
priori independent of m), which is presumably large or we would have noticed it, and
hence we will ignore such terms for a while.

115
5.4 Free particle solutions of spinor wave equations

[Peskin §3.3] To understand a quantum scalar field, we had to know that the solutions
of the KG equation were plane waves e−ipx with p2 = m2 (then we associated a mode
operator ap~ with each solution and added them up). To do the analog for spinors, we’ll
need to know the free particle solutions.
Let’s focus on the Dirac equation. This implies the wave equation, so solutions can
be made from superpositions of plane waves with p2 = m2

Ψ(x) = e−ipx u(p)

but the Dirac equation places a further restriction on the constant spinor u(p):

0 = (γµ pµ − m) u(p).

Let’s assume m 6= 0 and solve this in the rest frame, p0 = (m, ~0). Then we can find
the answer for general pµ (with p0 > 0) by a boost: u(p) = Λ 1 u(p0 ).
2
 
−1 1
0 = (mγ 0 − m)u(p0 ) = m u(p0 )
1 −1
 
ξ
which is solved by u(p0 ) ∝ for any 2-component spinor ξ. The fact that there are
ξ
two solutions for each p is the “intrinsic two-valuedness” associated with spin 21 . It will
be convenient to normalize the solutions by
√
 
ξ
u(p0 ) = 2m , ξ † ξ = 1.
ξ
   
3 1 1 2 0
We can choose a basis for such ξs which diagonalize σ , e.g. ξ = ,ξ = in
0 1
the standard basis for the Paulis. ξ is an ordinary non-relativistic spinor.
Now, under a boost in the z direction (suppressing the x, y components which
remain zero),
        
E 01 m m cosh η
p0 7→ = exp η = ,
p3 10 0 m sinh η
and the positive-energy solution of the Dirac equation becomes
√
  3   
1 σ ξ
u(p0 ) 7→ Λ 1 (η)u(p0 ) = exp − η 3 m
2 2 −σ ξ
| {z }
σ3
 

cosh(η/2)1 −sinh(η/2) 
−σ 3

116
  p p  
3 3
E + p P + + E − p P− ξ 1
1 ± σ3


=  p   P± ≡
2
p
E − p 3 P + + E + p 3 P− ξ
 p  
p √
E + p 3 σ 3 P+ + E + p 3 σ 3 P − ξ

 = √ p · σξ
=  p .
p · σ̄ξ
p
E − p 3 σ 3 P+ + E − p 3 σ 3 P − ξ

In the final expression, we define the square root of a matrix by its action on eigenstates.
The last expression also works for any boost direction since it’s rotation invariant.
Using the identity
(p · σ) (p · σ̄) = p2 (5.21)
(check it on the homework) we can check directly that this expression actually solves
the Dirac equation for general p:
√    √  √ p √
2 p·σ
! 
p · σξ −m σ · p p · σξ −m p · σ + p ξ p2 =m2
(pµ γ µ − m) √ = √ = √ p √ = 0.
p · σ̄ξ σ̄ · p −m p · σ̄ξ σ̄ · p p2 − σ̄ · pm ξ

Negative-energy solutions. Just as for the KG equation, there are also negative-
energy solutions with the same p~ (which are not related to the previous by any or-
thochronous Lorentz transformation):

Ψ(x) = v(p)e+ip·x , p2 = m2 , p0 > 0

where the Dirac equation further imposes

√
p · ση s

s
v (p) = √ , s = 1, 2.
− p · σ̄η s

Normalization. A Lorentz-invariant normalization condition is

ūs ur = 2mξs† ξr = 2mδsr v̄ r v s = −2mδ rs .

This is equivalent to the following statements about the Lorentz-variant quantities:

u†r (p)us (p) = 2Ep ξr† ξs = 2Ep δrs vr† vs = +2Ep ηr† ηs = 2Ep δ rs .

Notice that for each p, 0 = ūr (p)v s (p) = v̄ r (p)us (p) (but (ur )† (p)v s (p) 6= 0).
Completeness relations. Suppose we choose a basis
X
1 2×2 = ξ s (ξ s )† . (5.22)
s=1,2

117
Then
X √p · σξ s    (5.22) √ √ √
2 !
X
s s † √ † √ p · σ p · σ̄ p · σ
u (p)ū (p) = √ s (ξ s ) p · σ̄, (ξ s ) p · σ = √ 2 √ √
s=1,2 s
p · σ̄ξ ( p · σ̄) p · σ̄ p · σ
 
(5.21) m p·σ
= = γ · p + m ≡ p/ + m. (5.23)
p · σ̄ m
P
Similarly, s vv̄ = p/ − m. Note that these completeness relations for spinor polariza-
tions are analogous to the relation I quoted for photon polarizations (1.19).
Helicity. Define the helicity operator, acting on one-particle states
 i 
~ ≡ p̂ · JDirac = p̂
i i i 1 σ
ĥ ≡ p̂ · S .
2 σi

p̂ = |~pp~| is a unit vector, so ĥ2 = 1/4 and the eigenvalues are ±1/2, which are called
right-handed and left-handed. (Sometimes people normalize the helicity so that h =
±1.) Naturally, positive-energy (massless) Weyl R/L spinors are helicity eigenstates,
with h = ±1/2 respectively, since the R/L Weyl equation is 0 = p0 1 ∓ pi σ i , and
p0 = p|~p|. More generally, consider the ultra-relativistic limit of a Dirac spinor, where
E = p~2 + m2 → |~p|, with (WLOG) p~ = ẑp3 ,
 !  
 p 1 0
E − p3



p| √
   
0  0

  Ep →|~
! → if σ 3 = +1


   2E  

,
1 1

 p

E + p3

    

 0 0
u(Ep , p3 ) =  !  
 p 0 0
E + p3



p| √
   
1  1

Ep →|~
 
if σ 3 = −1



  ! →
 2E  

,
0  0

 p
E − p3

 

1 0


These are helicity eigenstates. More generally, any ultrarelativistic spinor wavefunction
can be decomposed into a linear combination of such helicity eigenstates. For massive
particles (away from this limit), you can switch the helicity by outrunning the particle;
it is not Lorentz invariant.

118
5.5 Quantum spinor fields

What did we need to build the Fock space of a relativistic scalar field? A quick
recapitulation: For each mode, we had ladder operators, a 6= a† , and a number operator
N = a† a which was hermitian and hence observable. The ladder operators are so called
because
[N, a] = −a, [N, a† ] = a† . (5.24)
This says that given a number eigenstate N |ni = n |ni, we can make others by
(5.24)
(5.24)
N (a |ni) = (n − 1) (a |ni) and N a† |ni = (n + 1) a† |ni . And we know n ≥ 0

since 0 ≤ || a |ni ||2 = hn| N |ni = n hn|ni, so there must exist a lowest n0 that we can’t
lower any further, a |n0 i = 0, but then n0 |n0 i = N |n0 i = a† a |n0 i = 0, so n0 = 0.
| {z }
Hence, a ladder of eigenstates of N with eigenvalues 0, 1, 2, 3, ....

OK, but here’s why I just did that: the necessary equation (5.24) did not require that
[a, a† ] = 1. In fact, (5.24) would also follow from the anticommutation relation

aa† + a† a ≡ {a, a† } = 1, {a, a} = 0 = {a† , a† }.

To see this, note the identity [AB, C] = A{B, C} − {A, C}B, so

[a† a, a] = a† {a, a} − {a† , a}a = −a.

2
But now 0 = {a, a} = 2a2 means a2 = 0 and a† = 0: the ladder only has only one

2
rung. |0i with a |0i = 0 and |1i = a† |0i, and there is no |2i ∝ a† |0i = 0. This is
Pauli exclusion. [End of Lecture 16]
For multiple modes, we can take

{ai , a†j } = δij , {ai , aj } = 0 = {a†i , a†j }.

And states look like

a†i a†j |0i = ± |00001i 00...001j 0000i = −a†j a†i |0i (5.25)

we have to remember the order in which the quanta are created to get the sign right (the
overall choice is a convention, but relative signs are physics). This is Fermi statistics.
Given a field, how do we know whether to use commutators or anticommutators? A
practical answer: try both and one of them will be bad somehow. The general answer,
with Lorentz invariance, is the spin-statistics theorem: fields with half-integer spin are
fermionic, and those with integer spin are bosonic. Schwartz, §12 has an illuminating

119
discussion of many points of view on the connection between spin and statistics. I’ll
say a little more below.
R d
Anticommuting scalar fields? Consider a real scalar φ(x) = √d̄ p ap e−ipx +
2ωp

h.c., with π = φ̇ and

Z   Z
1  2
~ ωp †
π + ∇φ + m φ = d̄d p
2 2 2
ap ap + ap a†p .


H=
2 2
d
You see that if {ap , a†p0 } = /δ (p − p0 ) then we get a hamiltonian which is just an infinite
constant, independent of the state. Not such a useful energy functional. We also get
{φ(x), π(y)} = 0.
R d
The situation is worse for a complex scalar, with Φ(x) = √d̄ p ap e−ipx + b†p eipx .


2ωp
Then with anticommutators we get
Z Z
d 1
† ? 1
no assumptions † † {b,b }=δ
d̄d p ωp a†p ap − b†p bp + cst (5.26)



H = d̄ p ωp ap ap + bp bp =
2 2
and the energy is unbounded below. We could alternatively allow negative norm states
or non-local anticommutators. Is that any better?
Dirac Hamiltonian. From the Dirac lagrangian, we have the canonical momen-
tum density Π = ∂∂L
Ψ̇
= iΨ̄γ 0 = iΨ† and the hamiltonian density
 
~ eom
†
h = ΠΨ̇ − L = iΨ ∂t Ψ −L = Ψ̄ i~γ · ∇ + m Ψ = Ψ† i∂t Ψ. (5.27)
| {z }
=iΨ̄γ 0 ∂ tΨ

Following our nose and writing the operator-valued field as a sum over all solutions
of the eom weighted by ladder operators, we have

d̄3 p X s
Z
u (p)e−ipx asp + v s (p)eipx bs†


Ψ(x) = p p
2ωp s=1,2
Z 3
d̄ p X s −ipx s
ū (p)eipx as† s


Ψ̄(x) = p p + v̄ (p)e bp (5.28)
2ωp s=1,2

where, as for the scalar, we implicitly set p0 = ωp~ .

The hamiltonian is then (using the last expression in (5.27))
Z
H = d3 x h
d̄3 p d̄3 q X s†
Z Z Z
3 −ipx s −iqx s
u (p)eipx as† s† s s iqx s†



= dx p p p + v (p)e b p ω q u (q)e aq − ω q v (q)e b q
2ωp 2ωq ss0

120
This is of the form u† u + v † v + u† v + v † u. In the first two terms, the x integral is
R 3 i(p−q)x
d xe = δ (3) (p − q) while in the mixed uv terms, we have ~q = −~p. We use the
spinor identities

u†s (p)us0 (p) = vs† (p)vs0 (p) = +2ωp δss0 , u†s (p)vs0 (−p) = vs† (p)us0 (−p) = 0

and get (no assumptions yet)

Z X
d̄3 p ωp as† s s s†


H= p ap − bp bp . (5.29)
s

Now, if [b, b† ] = 1, this is p ωp (Npa − Npb ) + constant and the world explodes in a
P

spontaneous shower of antiparticles lowering the energy by coming from nowhere. If

instead we have anticommutation relations,
d
{bs (p), bs0 (q)† } = /δ (p − q) ,

then this is Z X
H= d̄3 p ωp (Nsa (p) + Nsb (p)) + const
s

and all is well, H − E0 ≥ 0.38

This gives canonical equal time local anticommutators

{Ψ(~x)a , Π(~y )b }ET = iδ d (~x − ~y )δab

or
0 d
{Ψ(~x)a , Ψ̄(~y )b }ET = iγab δ (~x − ~y ).

Comments on the spin-statistics connection. [Schwartz’s very nice §12] Two

more comments about spin-statistics. First, the general pattern seen above continues
for all spins. The crucial ingredient is that integer-spin fields have Lagrangians that
involve two time derivatives:

LZ 3 φρ... ∂µ ∂ µ φρ... = φρ... φ̈ρ + ...,

while half-integer-spin fields have Lagrangians with a single time-derivative:

µ
LZ+ 1 3 ψ̄αρ... γαβ ∂µ ψβρ... .
2

38
Here we only used {b, b† } = δ, without any assumption about the commutators or anticommuta-
tors of a. You could ask why we can’t have [a, a† ] = 1 still. That would just be gross. Slightly more
scientifically, we can show (see below) that the Lorentz invariance of the propagator would fail.

121
This is a consequence of rotation invariance. In the former case, this means the Hamil-
†
tonian density is hZ = φρ... φ̈ρ... , while in the other case, hZ+ 1 = ψαρ... ∂t ψ αρ... , leading to
2
the same relative signs we found in (5.26) and (5.29).
The statistics are defined by the swap operation as in (5.25). More generally,
consider a two-particle state |φ1 (x1 )φ2 (x2 )i where we temporarily label the particles 1
and 2, and the wavefunctions φ1,2 are localized at positions x1,2 . The swap operator
acts to interchange the positions of the particles

S |φ1 (x1 )φ2 (x2 )i = eiφκ |φ2 (x1 )φ1 (x2 )i .

As you can see, we’ve allowed for a phase to appear; κ is a label depending on the
species of particles.

In d = 3, the phase eiφκ can only be a sign, ±1. To see this, ask: how do we
implement the swap operator? More generally, if the particles are identical, what can
happen in between two states with a particle at x1 and x2 ? We must move the particles
on some trajectory which ends up with 1 → 1, 2 → 2 or 1 ↔ 2 at the end. Let φ ≡ the
angle by which particle 1 went around 2. It is a topological property of the trajectory
of the two particles. But in d = 3, it is defined only mod 2π, because there is no
non-trivial linking of curves in 3+1 dimensions. You can see this by visualizing the

following figure in three plus one dimensions:

So the phase φκ must be φκ = φκ with κ ∈ Z, a property of the particles (in fact,

twice the spin), so that in d = 3,

S |φ1 (x1 )φ2 (x2 )i = (−1)κ |φ2 (x1 )φ1 (x2 )i .

In d = 2, there is a lot more to the story of particle statistics (because two curves in
2+1 dimensions can have a topologically-nontrivial linking).

Now, this swap of two particles can be accomplished using Poincaré transformations
(Lorentz times translations), as follows. First translate by x2 − x1 . Then rotate around
x1 by π.

~x2 T~x2 −~x1 Rπ

~x1

122
Therefore, whatever phase φκ is going to happen under interchange is already built
into the rotation matrices (translations don’t produce any phase).
On a Dirac spinor particle at rest, the action of a Lorentz transformation is (recall
µν
our expression for the Lorentz current) Λ(θ) = e−iθ·JDirac . For the particular case of a
rotation in the xy-plane, this is a diagonal matrix
 z   
eiθ /2 i
z
e−iθ /2  θ →π  −i
  z  
~ iθz σ 3 ⊗1 L/R
Λ(θ = θz ž, β = 0) = e =  → 
 
iθz /2
e i 

  
z
e−iθ /2 −i
These factors of i are a direct consequence of the fact that the 2π rotation gives −1,
so a π rotation gives ±i (where the sign depends on the spin σ z ).
Now suppose the initial spin of the particle is in the ž direction (if not, the state will
not be an eigenstate of Λ), so that (e.g. for the σz = +1 eigenstate) ψ1 = (1, 0, 1, 0) 7→
Λψ1 = (i, 0, i, 0) = iψ1 . Then the Poincaré transformation which implements the swap
of two such particles is
T~x2 −~x1 Λ(θ~ = πž) |ψ1 (~x1 )ψ2 (~x2 )i = (±i)2 |ψ1 (~x2 )ψ2 (~x1 )i .
For either spin (as long as the spins of the two particles are the same), the two π
rotations combine to give a (−1) – Fermi statistics follows from the Dirac Lagrangian.

Dirac propagator. [Peskin §4.7] Now we will efficiently redo the story of interac-
tion picture perturbation theory and Wick’s theorem for Dirac spinor fields. The story
differs by just a few very important signs.
Time ordering for fermions is defined with an extra minus sign:
T (A1 (x1 ) · · · An (xn )) ≡ (−1)P A10 (x10 ) · · · An0 (xn0 ), x010 > x020 > · · · > x0n0
where P ≡ the number of fermion interchanges required to get from the ordering 1...n
to the ordering 10 ...n0 (mod two). Similarly, normal ordering is
: ABC · · · :≡ (−1)P A0 B 0 C 0 · · · ,
where on the RHS all annihilation operators are to the right of all creation operators,
and P is defined the same way. 39 Wick’s theorem is then still
T (ABC · · · ) =: ABC · · · : +all contractions.
39
Why the extra signs? One way to see that they must be there is if they weren’t everything would
be zero. With the sign, the following two choices for a normal ordered product are equivalent:
: ap aq a†r := (−1)2 a†r ap aq = (−1)3 a†r aq ap ,
but without it, we would conclude that the LHS would be zero.

123
The possible contractions of Dirac fields are:

Ψ(x)Ψ(y) = 0, Ψ̄(x)Ψ̄(y) = 0, Ψ(x)Ψ̄(y) = SF (x − y),

where SF is the Feynman propagator for the Dirac field, i.e. the time-ordered two-point
function. It is

SFab (x − y) = h0| T Ψa (x)Ψ̄b (y) |0i

= θ(x0 − y 0 ) h0| Ψa (x)Ψ̄b (y) |0i −θ(y 0 − x0 ) h0| Ψ̄b (y)Ψa (x) |0i
(−) (+)
= θ(x0 − y 0 ) h0| Ψ(+) 0 0
a (x)Ψ̄b (y) |0i −θ(y − x ) h0| Ψ̄b (y)Ψa (x) |0i
(−)
Ψ≡Ψ (+)
Ψ(−)
|{z} + |{z}
| {z } | {z }
(+) (−) (+) (−) a b†
=h0|{Ψa (x),Ψ̄b (y)}|0i≡S + h0|{Ψ̄b (y),Ψa (x)}|0i≡S −

The S + bit, made only from as, is a c-number

d̄3 p −ipx X d̄3 q +iqy X s

Z Z
(−) 0 0
+ (+)
Sab (x − y) = {Ψa (x), Ψ̄b (y)} = p e p e ua (p)ūsb (q) {asp , asq † }
2ωp~ s=1,2
2ωq~ s0 =1,2
| {z }
δ (p−q)δ ss0
d
=/
Z 3
d̄ p −ip(x−y) X s
= e ua (p)ūsb (p)
2ωp~
|s {z }
(5.23)
= (p /+m)ab
Z 3
d̄ p
= ∂x + m)ab e−ip(x−y)
(i/
2ωp~
Z 3 Z
d̄ p −ip(x−y) (3.3) i
= (i/
∂x + m)ab e +
∆ (x − y) = d̄4 pe−ip(x−y)
2ωp~ C+ p2 − m2
| {z }
=∆+ (x−y)


i p/ + m ab
Z
4 −ip(x−y)
= d̄ pe (5.30)
C+ p2 − m 2

The same calculation for S − , the bit involving {b, b† }, gives the same integrand, the
only difference, as for the complex KG field, is the contour C + → C − . Getting the
same integrand with the same sign required the relative minus (the red one above)
which for bosons came from the sign in the commutator [b, b† ] = bb† − b† b. Adding
the two terms together, we learn that the momentum space Dirac propagator is



i p/ + m i p/ + m i
S̃(p) = 2 =

= . (5.31)
p − m2 p/ + m p/ − m p/ − m

(These matrices all commute with each other, so my cavalier manipulation of them
can be done in the eigenbasis of p/ without trouble.) It is not a coincidence that the
numerator of the propagator is the polarization sum: s us (p)ūs (p) = p/ + m
P

124
 The position-space Feynman propagator comes from integrating (5.31) over the
Feynman contour, as for scalars:
Z
i
SF (x − y) = d̄4 p e−ip(x−y) .
CF p
/ − m

Fermions and causality. Earlier I made a big deal that we need commutators
to vanish outside the lightcone to prevent acausal communication. But the Dirac field
Ψ(x) doesn’t commute with Ψ̄(y) for spacelike x − y (rather, they anticommute). Why
is this OK? What saves the day is the fact that we can’t measure a single fermion
operator. The operators we can measure (such as the number density of fermions Ψ† Ψ,
~
or their momentum density Ψ† ∇Ψ) are all made of even powers of Ψ and Ψ̄. And these
do commute outside the lightcone.
A principle which would make this restriction on what we can measure precisely
true and inevitable is if fermion parity is gauged. By ‘fermion parity’ I mean the
transformation which takes Ψ → −Ψ for every fermionic operator in the world. By ‘is
gauged’ I mean that this transformation should be regarded as an equivalence relation,
rather than a transformation which relates distinct physical configurations. In that
case, a local operator with an odd number of fermions would not be gauge invariant.
40

40
This point of view, that locality should be built into the Hilbert space, and that therefore that
fermion parity should be gauged, is advocated forcefully by Wen, Quantum Field Theory of Many-
Body Systems (Oxford, 2004) in Chapter 10. It is not clear how to include gravitational degrees of
freedom in this principle.

125
6 Quantum electrodynamics
We must fill a small hole in our discussion, to help our fermions interact. In §1.3, we
figured out a bit of the quantum theory of the radiation field. A few things we did not
do: study the propagator, figure out the data on external states, and the relation of
between the masslessness of the photon and gauge invariance. After that we will couple
electrons and photons and study leading-order (tree-level) processed in the resulting
theory of quantum electrodynamics (QED).

6.1 Vector fields, quickly

[We’ll follow Ken Intriligator’s efficient strategy for this discussion.] Consider the
following Lagrangian for a vector field Aµ (which I claim is the most general quadratic
Poincaré-invariant Lagrangian with at most two derivatives):
 
1
L = − ∂µ Aν ∂ µ Aν + a ∂µ Aµ ∂ν Aν +bAµ Aµ + cµνρσ ∂µ Aν ∂ρ Aσ  .

2 | {z }
=(∂A)2

The sign is chosen so that spatial derivatives are suppressed, and the normalization of
the first term is fixed by rescaling A. (Another possible-seeming term, ∂µ Aν ∂ν Aµ ,
is related to the second term by two IBPs.) The last term is a total derivative,
µνρσ ∂µ Aν ∂ρ Aσ ∝ ∂µ (µνρσ Aν ∂ρ Aσ ), and will not affect the EoM or anything at all
in perturbation theory; it is called a θ term. [End of Lecture 17]
The EoM are
Z
δ
0= ν
L = −∂ 2 Aν − a∂ν (∂ · A) + bAν (6.1)
δA (x)
which (like any translation-invariant linear equation) is solved by Fourier transforms
Aµ (x) = µ e−ikx , if
k 2 µ + akµ (k · ) + bµ = 0.
There are two kinds of solutions: longitudinal ones with µ ∝ kµ (for which the disper-
b
sion relation is k 2 = − 1+a ), and transverse solutions  · k = 0 with dispersion k 2 = −b.
The longitudinal mode may be removed by taking b 6= 0 and a → −1, which we will
do from now on. This gives the Proca Lagrangian:
1 1
La=−1,b≡−µ2 = − Fµν F µν + µ2 Aµ Aµ ,
4 2
where as usual Fµν ≡ ∂µ Aν − ∂ν Aµ . Note that the EOM (Proca equation) 0 = ∂ · F·ν +
µ2 Aν implies 0 = ∂ ν Aν by 0 = ∂ µ ∂ ν Fµν . So each component of Aµ satisfies (by (6.1))

126
the KG equation, k 2 = µ2 , and the transverse condition  · k = 0. In the rest frame,
k µ = (k 0 , ~0)µ , we can choose a basis of plane-wave transverse solutions which are
eigenstates of the vector rotation generator
     
0 0
z

+1 

(±) 1  1  (0) 0
  
J = i  , namely,  = √   ,  =   .
 
 −1  2 ∓i 0
0 1

(r)? (r)
They are normalized so that (r) · (s) = +δ rs and r=±1,0 µ ν = −ηµν + kµµk2 ν so
P

that they project out  ∝ k. Notice that in the massless case, only two of these three
polarization states will be transverse to k µ . If ~k ∝ ẑ (for example in the massless case
with k µ = (E, 0, 0, E)µ ) then these  are also all helicity eigenstates: h = J~ · k̂ = J z .
Canonical stuff: The canonical momenta are π i = ∂∂L Ȧi
= −F 0i = E i (as for
electrodynamics in §1.3) and π 0 = ∂∂LȦ0
= 0. This last bit is a little awkward, but it
just means we can solve the equations of motion for A0 algebraically in terms of the
other (real) dofs:

~˙
 
Z − ~ ·A
∇
δS ~ E−µ
~ 2 ~ A~˙ =⇒ A0 (~x) = d3 ye−µ|~x−~y|
0= = ∇· A0 = (−∇2 +µ2 )A0 +∇· .
δA0 4π|~x − ~y |
(6.2)
So at each moment A0 is determined by Ai . (Notice that this is still true for µ →
0.) The hamiltonian density is (after using π i = F 0i , integration by parts, and the
equations of motion for A0 )
 
1 2 1 2 2 2 2 2 1 ~2 ~2 
~ 2 + µ2 A2 ≥ 0,
h=+ F0i + Fij + µ Ai + µ A0 = E + B + µ2 A 0
2 2 2

where positivity follows from the fact that it is a sum of squares of real things.
The canonical equal time commutators are then

[Ai (t, ~x), F j0 (t, ~y )] = iδij δ (3) (~x − ~y )

which if we add up the plane wave solutions as

X Z d̄3 k  
+ikx r† (r)?
Aµ (x) = √ e−ikx ark (r)
µ + e a 
k µ
r=1,2,3
2ωk

give the bosonic ladder algebra for each mode

[ark , as† /(3) ~ ~)δ rs .

p ] = δ (k − p

127
The normal-ordered hamiltonian is
XZ
: H := d̄3 k ωk ar† r
k ak .
r

The propagator for the Aµ (x) field is

2
 
−ik(x−y) −i(ηµν − kµ kν /µ )
Z
4
h0|T Aµ (x)Aν (y)|0i = d̄ ke . (6.3)
k 2 − µ2 + i

Notice that like in the spinor case the polarization sum r r? r
P
µ ν appears in the numer-
ator of the propagator. The quantity in square brackets is then the momentum-space
propagator. Since h0| Aµ (x) |k, ri = rµ (k)e−ikx , a vector in the initial state produces a
factor of rµ (k), and in the final state gives ? .
Massless case. In the limit µ → 0 some weird stuff happens. If we couple Aµ
to some object j µ made of other matter, by adding ∆L = j µ Aµ , then we learn that
∂µ Aµ = µ−2 ∂µ j µ . This means that in order to take µ → 0, it will be best if the current
is conserved ∂µ j µ .
One example is the QED coupling, j µ = Ψ̄γ µ Ψ. We saw that this coupling Aµ j µ
arose from the ‘minimal coupling’ prescription of replacing ∂µ → Dµ = ∂µ + ieqAµ
in the Dirac Lagrangian. In that case, the model had a local invariance under Aµ →
Aµ + ∂µ λ(x)/e, Ψ(x) → eiqλ(x) Ψ(x). For λ non-constant (and going to zero far away),
this is a redundancy of our description rather than a symmetry (for example, they
~ B,
~ A). That is, configurations related by this
H
have the same configuration of E,
gauge transformation should be regarded as equivalent.
Another example can be obtained by taking a complex scalar and doing the same
replacement: L = Dµ Φ? Dµ Φ + ... Notice that in this case the vertex involves a deriva-
tive, so it comes with a factor of = −ieq(pΦ +pΦ? )µ . Also, there is a Aµ Aν Φ? Φ

coupling, which gives a vertex = −ie2 q 2 ηµν .

How do I know that configurations related by a gauge transformation should be

regarded as equivalent? If not, the kinetic operator for the massless vector field
(ηµν (∂ ρ ∂ρ ) − ∂µ ∂ν ) Aν = 0 is not invertible, since it annihilates Aν = ∂ν λ.
What’s the propagator, then? One strategy is to simply ignore the gauge equiva-
lence and use the same propagator (6.3) that we found in the massive case with µ → 0.
Since the dynamics are gauge invariant, it will never make gauge-variant stuff, and the
longitudinal bits ∝ kµ kν in (6.3) (which depend on µ) will just drop out, and we can
take µ → 0 in the denominator at the end. This actually works. The guarantee that

128
it works is the QED Ward identity: any amplitude with an external vector (k)µ is of
the form

= iM = iMµ (k)µ (k)

and if all external fermion lines are on-shell then

Mµ (k)kµ = 0.

There is a complicated diagrammatic proof of this statement in Peskin; Schwartz §8.4

argues that it is a necessary condition for Lorentz invariance of M = µ · Mµ ; and
we will see some illustrations of it below (I also recommend Zee §II.7). But it is
basically a statement of current conservation: such an amplitude is made (by LSZ
and the photon Schwinger-Dyson equation) from a correlation function involving an
insertion of the electromagnetic current j µ (k) = d4 x e−ikx j µ (x), in the form, Mµ ∼
R

... hΩ|....j µ (k)...|Ωi, and kµ j µ (k) = 0 is current conservation41 .

This property guarantees that we will not emit any longitudinal photons, since the
amplitude to do so is the µ → 0 limit of
 
!
emit Lλ = µ1 (k, 0, 0, −ω)λ L µ 1 0 3

1 p
kM0 − k 2 + µ2 M3 

A ∝ µ M = kM − ωM =
with k λ = (ω, 0, 0, k)λ µ µ | {z } 
2
=k+ µ
2k
+...
1 µ µ→0
= kµ Mµ − M3 + O(µ3 ) → 0.
µ | {z } 2k
| {z }
=0,by Ward
→0 as µ→0

Gauge fixing. You might not be happy with the accounting procedure I’ve advo-
cated above, where unphysical degrees of freedom are floating around in intermediate
states and only drop out at the end by some formal trick. In that case, a whole zoo of

41
Current conservation ∂µ j µ is a statement which requires the equations of motion (recall the proof
of Noether’s theorem). Recall that we proved that equations of motion are true in correlation functions,
up to contact terms, using the independence of the path integral on choice of integration variables.
By contact terms, I mean terms which are only nonzero when two operators are at the same point.
So you can worry about the contact terms in argument for the Ward identity. The reason they do
not contribute is that all the operators in the correlation function (using the LSZ formula) correspond
to external states. A collision between the operators creating the external particles would lead to a
disconnected amplitude, which could only contribute for degenerate kinematical configurations, and
we can ignore them. If you would like to read more words about this, look at Schwartz §14.8, or
Appendix B.

129
formal tricks called gauge fixing has been prepared for you. Here’s a brief summary to
hold you over until 215B.
At the price of Lorentz invariance, we can make manifest the physical dofs, by
choosing Coulomb gauge. That means we restrict ∂µ Aµ = 0 (so far, so Lorentz in-
~ ·A
variant) and also ∇ ~ = 0. Looking at (6.2), we see that this kills off the bit of A0
that depended on A.~ We also lose the helicity-zero polarization ∇
~ ·A
~ ∝ (0) . But the
Coulomb interaction is instantaneous action at a distance.
To keep Lorentz invariance, we can instead merely discourage configurations with
∂ · A 6= 0 by adding a term to the action
1 1
L = − Fµν F µν − (∂ · A)2
4 2ξ
for some arbitrary number ξ. Physics should not depend on ξ, and this is a check on
calculations. The propagator is
2
 
−ik(x−y) −i(ηµν − (1 − ξ)kµ kν /µ )
Z
4
hT Aµ (x)Aν (y)iα = d̄ k e
k 2 − µ2 + i

and again the bit with kµ kν must drop out. ξ = 1 is called Feynman gauge and makes
this explicit. ξ = 0 is called Landau gauge and makes the propagator into a projector
onto k⊥ .
It becomes much more important to be careful about this business in non-Abelian
gauge theory.

42

42
By the way, you might be bothered that we didn’t go back to our table 1 of possible Lorentz
representations on fields to think about spin one particles. Indeed, we could start with the (1, 0)⊕(0, 1)
antisymmetric tensor Fµν as our basic object. (See, for example, the book by Haag, page 47.) Indeed,
in this way we could construct a theory of a free EM field. But don’t we need a vector potential to
couple to charged matter? The answer turns out to be ‘no,’ as explained by Mandelstam here. The
price is that the charged fields depend on not just a point, but a choice of path; if we did introduce
the vector potential, they would be related to our fields by
R
Φ(x, P ) = ei P
A
Φ(x),
R
where P is a path which ends at x and infinity. This Wilson line ei P A carries away the gauge
transformation, so that Φ(x, P ) is actually invariant under gauge transformations that fall off at
infinity.
Thanks to Wei-Ting Kuo for asking about this.

130
6.2 Feynman rules for QED

First, Feynman rules for Dirac fermion fields, more generally43 . As always in these
notes, time goes to the left, so I draw the initial state on the right (like the ket) and
the final state on the left (like the bra).

1. An internal fermion line gives

i
=
k/ − mΨ

which is a matrix on the spinor indices.

There are four possibilities for an external fermion line of definite momentum.

2. = Ψ|k, ri = ur (k)

3. = hk, r|Ψ̄ = ūr (k)

4. = Ψ̄|k, ri = v̄ r (k)

5. = hk, r|Ψ = v r (k)

6. Some advice: When evaluating a Feynman diagram with spinor particles, always
begin at the head of the particle-number arrows on the fermion lines, and keep
43
Another good example of a QFT with interacting fermions is the Yukawa theory theory of a Dirac
fermion field plus a scalar φ and an interaction

0
V = gφΨ̄Ψ =⇒ = − igδ rr . (6.4)

Notice that in 3 + 1 dimensions, [g] = +4 − [φ] − 2[Ψ] = 4 − 1 − 2 32 = 0, the coupling is dimensionless.

This describes more realistically the interactions between nucleons (which are fermions, as opposed to
snucleons) and scalar pions, which hold together nuclei. It also is a crude sketch of the Higgs coupling
to matter; notice that if φ is some nonzero constant hφi, then there is a contribution to the mass of
the fermions, g hφi.

131
 going along the fermion line until you can’t anymore. This will keep the spinor
indices in the form of matrix multiplication. Why: every Lagrangian you’ll ever
encounter has fermion parity symmetry, under which every fermionic field gets
a minus sign; this means fermion lines cannot end, except on external legs. The
result is always of the form of a scalar function (not a matrix or a spinor) made
by sandwiching gamma matrices between external spinors:
0
X
r 0 p0 rp = ūr (p0 )a (pile of gamma matrices)ab ur (p)b
a,b...=1..4

Furthermore, in S-matrix elements the external spinors u(p), v(p) satisfy the equa-
tions of motion (p/ − m)u(p) = 0, a fact which can be used to our advantage to
shrink the pile of gammas.
There can also be fermion lines which form internal loops (though not at tree
level, by definition). In this case, the spinor indices form a trace,
X
(pile of gamma matrices)aa ≡ tr (pile of gamma matrices) .
a

We’ll learn to compute such traces below (around (6.6)); in fact, traces appear
even in the case with external fermions if we do not measure the spins.
[End of Lecture 18]

7. Diagrams related by exchanging external fermions have a relative minus sign.

8. Diagrams with an odd number of fermion loops have an extra minus sign.

The last two rules are best understood by looking at an example in detail.
To understand rule 8 consider the following amplitude in the Yukawa theory with

interaction (6.4): It is a contribution to the meson propagator.

It is proportional to
X
Ψ̄a (x)Ψc (x)Ψ̄c (y)Ψd (y) = (−1)trΨ(x)Ψ̄(y)Ψ(x)Ψ̄(y) = (−1)trSF (x − y)SF (x − y)
abcd

[Peskin page 119] To understand rule 7 consider ΨΨ → ΨΨ (nucleon) scattering

in the Yukawa theory: The blob represents the matrix

132
element
V d4 z
R
−i
0 hp3 r3 ; p4 r4 | T e |p1 r1 ; p2 r2 i0
where the initial state is
|p1 r1 ; p2 r2 i0 ∝ arp11† arp22† |0i
and the final state is

0 hp3 r3 ; p4 r4 | = (|p3 r3 ; p4 r4 i0 )† ∝ h0| arp44 arp33 = − h0| arp33 arp44

where note that the dagger reverses the order.

The leading contribution comes at second order in V :
 Z Z 
1 2 4 4



0 hp3 r3 ; p4 r4 | T (ig) d z1 d z2 Ψ̄Ψφ 1 Ψ̄Ψφ 2 |p1 r1 ; p2 r2 i0
2!

To get something nonzero we must contract the φs with

each other. The diagrams at right indicate best the pos-
sible ways to contract the fermions. Exchanging the roles
of z1 and z2 interchanges two pairs of fermions so costs
no signs and cancels the 2!1 .
The overall sign is annoying but can be fixed by demand-
ing that the diagonal bit of the S-matrix give

hp3 p4 | (1 + ...) |p1 p2 i = +δ(p1 − p3 )δ(p2 − p4 ) + · · ·

The relative sign is what we’re after, and it comes by comparing the locations of fermion
operators in the contractions in the two diagrams at right. In terms of the contractions,
these t− and u− channel diagrams are related by leaving the annihilation operators
alone and switching the contractions between the creation operators and the final state.
Denoting by a†1,2 the fermion creation operators coming from the vertex at z1,2 ,

h0| ap4 ap3 a†1 a†2 ... + h0| ap4 ap3 a†1 a†2 ...
| {z } | {z }
= h0| ap4 ap3 a1 a2 ... − h0| ap4 a†1 ap3 a†2 ...
† †
| {z } | {z } | {z }

In the last expression the fermion operators to be contracted are all right next to each
other and we see the relative minus sign.

While we’re at it, let’s evaluate this whole amplitude to check the Feynman rules
I’ve claimed and get some physics out. It is

e−iq(z1 −z2 ) i
Z Z
2
dz1 dz2 d̄4 q 2 e−ız2 (p1 −p3 ) ūr3 (p3 )ur1 (p1 ) · e−ız1 (p2 −p4 ) ūr4 (p4 )ur2 (p2 ) − (3 ↔ 4) .


Sf i = −g 2
q − m + i

133
In the first (t-channel) term, the integrals over z1,2 gives /δ(p1 − p3 − q)/δ(p2 − p4 − q),
and the q integral then gives δ(p1 + p2 − p3 − p4 ), overall momentum conservation. In
the second (u-channel) term, q = p1 − p4 = p3 − p2 . Altogether,
4
Sf i = 1 + /δ (pT )iM
with, to leading order,
 
2 1 1
iM = −ig (ū3 u1 ) (ū4 u2 ) − (ū4 u1 ) (ū3 u2 ) (6.5)
t − m2 u − m2
with t ≡ (p1 − p3 )2 , u ≡ (p1 − p4 )2 . This minus sign implements Fermi statistics.
Yukawa force revisited. In the non-relativistic limit, we can again relate this
amplitude to the force between particles, this time with the actual spin and statistics of
nucleons. In the COM frame, p1 = (m, p~), p2 = (m, −~p) and p3 = (m, p~0 ), p4 = (m, −~p0 ).
√
σ · pξ r
 r
√

r ξ
In the non-relativistic limit, the spinors become up = √ r → m r so that
σ̄ · pξ ξ
r3 r1 †
ū3 u1 ≡ ū(p3 ) u(p1 ) = 2mξr3 ξr1 = 2mδr3 r1 . Let’s simplify our lives and take two
distinguishable fermions (poetically, they could be proton and neutron, but let’s just
add a label to our fermion fields; they could have different masses, for example, or
different couplings to φ, call them g1 , g2 ). Then we only get the t-channel diagram.
The intermediate scalar momentum is q = p1 − p3 = (0, p~ − p~0 ) so t = (p1 − p3 )2 =
−~q2 = − (~p − p~0 )2 and
1
iMN R,COM = ig1 g2 4m2 δ r1 r3 δ r2 r4 .
~q2 + m2φ
Compare this to the NR Born approximation matrix element
 
2πδ(Ep − Ep0 ) −iṼ (~q) = N R h~p0 | S |~piN R
XZ 4
Y 1
= d̄3 p4 V √ S(34 ← 12)
r4
2Ei
|i=1 {z }
=√ 1 4
2m
ig1 g2
= 2πδ(Ep − Ep0 )δ r1 r3
~q2 + m2φ
where in the second line we summed over possible final states of the second (target)
3
particle, and corrected the relativistic normalization, so that N R h~p0 |~piN R = /δ (p − p0 ).
This is completely independent of the properties of the second particle. We infer that
−mφ r
the scalar mediates a force with potential U (x) = − g1 g24πr
e
. It is attractive if g1 g2 > 0.

Back to QED. The new ingredients in QED are the propagating vectors, and the
interaction hamiltonian V = eΨ̄γ µ ΨAµ . The rest of the Feynman rules are

134
 9. The interaction vertex gets a

= − ieγ µ

10. An external photon in the initial state gets a µ (p), and in the final state gets a
µ? (p).

11. An internal photon line gets a

i µν µ ν 2


= −η + (1 − ξ)k k /k
k 2 − m2γ

where mγ = 0 (it’s sometimes useful to keep it in there for a while as an IR

regulator) and the value of ξ is up to you (meaning that your answers for physical
quantities should be independent of ξ).

Spinor trace ninjutsu.

The trace is cyclic: tr (AB · · · C) = tr (CAB · · ·) . (6.6)

Our gamma matrices are 4 × 4, so tr1 = 4.

(6.6) 5 µ
{γ ,γ }=0
2
trγ µ = tr γ 5 γ µ = trγ 5 γ µ γ 5 −trγ µ = 0.


= (6.7)
The same trick shows that the trace of any odd number of gammas vanishes. The idea
is that an odd number of gammas is a map between the L and R subspaces, so it has
only off-diagonal terms in the Weyl basis.
clifford (6.6)
trγ µ γ ν = −trγ ν γ µ + 2η µν tr1 = −trγ µ γ ν + 8η µν =⇒ trγ µ γ ν = 4η µν . (6.8)

trγ µ γ ν γ ρ γ σ = 4 (η µν η ρσ + η σµ η νρ − η µρ η νσ ) . (6.9)
Why is this? The completely antisymmetric bit vanishes because it is proportional to
γ 5 which is traceless (by the same argument as (6.7)). If any pair of indices is the same
then the other two must be too by (6.8). If adjacent pairs are the same they can just
square to one and we get +1; if alternating pairs are the same (and different from each
other) then we must move them through each other with the anticommutator. If they
are all the same we get 4.

trγ µ γ ν γ ρ γ σ γ 5 = −4iµνρσ .

135
6.3 QED processes at leading order

Now we are ready to do lots of examples, nearly all of which (when pushed to the
end) predict cross sections which are verified by experiments to about one part in
1 e2
137.44 Here 137 ≈ α ≡ 4π is the small number by which the next order corrections are
45
suppressed.
Did I mention that the antiparticle of the electron, predicted by the quantum Dirac
theory (i.e. by Dirac), is the positron? It has the same mass as the electron and the
opposite electromagnetic charge, since the charge density is the 0 component of the
electromagnetic current, j µ = Ψ̄γ µ Ψ, so the charge is
Z Z Z Z X
†
d xj (x) = Ψ̄γ Ψ = Ψ Ψ = d̄3 p
3 0 0
a†p,s aps − b†p,s bps .


s

So b† creates a positron.
[Schwarz §13.3, Peskin §5.1] Perhaps the simplest to start with is scattering of
electrons and positrons. We can make things even simpler (one diagram instead of
two) by including also the muon, which is a heavy version of the electron46 , and asking
about the process µ+ µ− ← e+ e− . At leading order in e, this comes from

iMµ+ µ− ←e+ e− =

 
(1−ξ)kµ kν
−i ηµν − k2
= (−ieūs3 (p3 )γ µ v s4 (p4 ))muons (−iev̄ s2 (p2 )γ ν us1 (p1 ))electrons (6.10)
k2
with k ≡ p1 + p2 = p3 + p4 by momentum conservation at each vertex. I’ve labelled
the spinors according to the particle types, since they depend on the mass.
Ward identity in action. What about the kµ kν term in the photon propagator?
The spinors satisfy their equations of motion, p/1 u1 = me u1 (where u1 ≡ usp11 for short)
44
I guess it is this overabundance of scientific victory in this area that leads to the intrusion of so
many names of physicists in the following discussion.
45
This statement is true naively (in the sense that the next diagrams which are nonzero come with
two more powers of e), and also true in fact, but in between naiveté and the truth is a long road of
renormalization, which begins next quarter.
46
Who ordered that? (I. Rabi’s reaction to learning about the muon.) I hope you don’t find it too
jarring that the number of ‘elementary’ particles in our discussion increased by three in the last two
paragraphs. People used to get really disgruntled about this kind of thing. But here we have, at last,
uncovered the true purpose of the muon, which is to halve the number of Feynman diagrams in this
calculation (compare (6.16)).

136
and v̄2 p/2 = −me v̄2 . The kν appears in
 
ν
kν v̄2 γ u1 = v̄2 p/1 + p/2 u1 = v̄2 p/1 u1 + v̄2 p/2 u1 = (m − m)v̄u = 0.

(The other factor is also zero, but one factor of zero is enough.) Therefore
e2
M= ū3 γµ v4 · v̄2 γ µ u1
s
where s ≡ k 2 = (p1 + p2 )2 = ECoM
2
is the Mandelstam variable. And I am relying on
you to remember which spinors refer to muons (3,4) and which to electrons (1,2).
Squaring the amplitude. We need to find M† (the dagger here really just means
complex conjugate, but let’s put dagger to remind ourselves to transpose and reverse
the order of all the matrices). Recall the special role of γ 0 here:

㵆 γ0 = γ0 γµ , γ0† = γ0 .

This means that for any two Dirac spinors,

†
Ψ̄1 γ µ Ψ2 = Ψ̄2 γ µ Ψ1 .

(This is the same manipulation that showed that the Dirac Lagrangian was hermitian.)
So
e2
M† = (v̄4 γ µ u3 ) (ū1 γµ v2 ) .
s
and therefore
e4
|Mµ+ µ− ←e+ e− |2 = (v̄4 γ µ u3 ) (ū3 γ ν v4 ) · (ū1 γµ v2 ) (v̄2 γν u1 ) . (6.11)
s2 | {z } | {z }
out in

These objects in parentheses are just c-numbers, so we can move them around, no
problem. I’ve grouped them into a bit depending only on the initial state (the electron
stuff 1, 2) and a bit depending only on the final state (the muon stuff 3,4).
Average over initial, sum over final. In the amplitude above, we have fixed the
spin states of all the particles. Only very sophisticated experiments are able to discern
this information. So suppose we wish to predict the outcome of an experiment which
does not measure the spins of the fermions involved. We must sum over the final-state
spins using X   X
vas4 (p4 )v̄bs4 (p4 ) = p/4 − mµ = v̄bs4 (p4 )vas4 (p4 )
ab
s4 s4

(where I wrote the last expression to emphasize that these are just c-numbers) and
X  
s3 s3
ua (p3 )ūb (p3 ) = p/3 + mµ .
ab
s3

137
Looking at just the ‘out’ factor of |M|2 in (6.11), we see that putting these together
produces a spinor trace, as promised:
X µ
ū(p3 )sa3 γab v(p4 )sb4 v̄(p4 )sc4 γcd
ν s3



u (p3 )d
s ,s
| {z }
3 4
/4 −mµ )bc
(p
µ ν
= γab (p − mµ )bc γcd (p/3 + mµ )da
 /4    
µ ν
= tr γ p/4 − mµ γ p/3 + mµ [End of Lecture 19]
= p4ρ p3σ trγ µ γ ρ γ ν γ σ − m2µ trγ µ γ ν
 
(6.8),(6.9)
= 4 pµ4 pν3 + pν3 pµ4 − p3 · p4 η µν − m2µ η µν  (6.12)
| {z }
≡p34

If also we don’t know the initial (electron) spins, then the outcome of our experiment
is the average over the initial spins, of which there are four possibilities. Therefore, the
relevant probability for unpolarized scattering is
1 X e4  µ         
|M|2 = tr γ p
/4 − m µ γ ν
p
/3 + m µ tr γν /2p − m µ γ p
µ /1 + m µ
4s 4s2
1,2,3,4
4
(6.12) twice 8e 2 2 2 2


= p 13 p 24 + p 14 p 23 + m µ p 12 + m e p 34 + 2m e m µ
s2
algebra 2e4 2
t + u2 + 4s(m2e + m2µ ) − 2(m2e + m2µ )2


= 2
(6.13)
s
In the second step of (6.13) the p12 p34 terms cancel. In the
last step of (6.13) we used all the Mandelstam variables:

s ≡ (p1 + p2 )2 = (p3 + p4 )2 = 2
ECoM = 4E 2
t ≡ (p1 − p3 )2 = (p2 − p4 )2 = m2e + m2µ − 2E 2 + 2~k · p~
u ≡ (p1 − p4 )2 = (p2 − p3 )2 = m2e + m2µ − 2E 2 − 2~k · p~

where the particular kinematic variables (in the rightmost

equalities) are special to this problem, in the center of
mass frame (CoM), and are defined in the figure at right.
Really there are only two independent Lorentz-invariant
kinematical variables, since s + t + u = i m2i .
P

Now we can use the formula (4.40) that we found for a differential cross section
with a two-body final state, in the CoM frame:
  !
dσ 1 |~p| 1 X
= 2
|M|2
dΩ CoM 2 ~
64π ECoM |k| 4 spins
α2 |~p|  4 ~ 2 2 2 2 2 2

= E + |k| |~
p| cos θ + E (m e + mµ ) (6.14)
16E 6 |~k|

138
 2
e
where α ≡ 4π ispthe fine structure constant. This can be boiled a bit with kinematical
~
p
relations |k| = E 2 − m2e , |~p| = E 2 − m2µ to make manifest that it depends only on
two independent kinematical variables, which we can take to be the CoM energy E
and the scattering angle θ in ~k · p~ = |~k||~p| cos θ (best understood from the figure). It
simplifies a bit if we take E  me , and more if we take E  mµ ∼ 200me , to

dσ α2
1 + cos2 θ .


= 2
(6.15)
dΩ 4ECoM

In fact, the two terms here come respectively from spins transverse to the scattering
plane and in the scattering plane; see Schwartz §5.3 for an explanation.
There is a lot more to say about what happens when we scatter an electron and a
positron! Another thing that can happen is that the final state could be an electron
and positron again (Bhabha scattering47 ).
They are not necessarily the same e− and e+ , though
(except in the sense that they are all the same), because
another way to get there at tree level is the second, t-
channel, diagram, at right. The intermediate photon in
that diagram has kt = (p1 −p3 ), so that the denominator
of the propagator is t = kt2 = (p1 − p3 )2 instead of s.

Squaring this amplitude gives

|Ms + Mt |2 = |Ms |2 + |Mt |2 + 2Re(Ms M?t ), (6.16)

interference terms. Interference terms mean that you have to be careful about the
overall sign or phase of the amplitudes.
You may be surprised that the cross section (6.15) decreases with energy. Mechan-
ically this comes mainly from the 1/s2 from the photon propagator: as s grows, the
intermediate photon is more and more off-shell. But more deeply, it’s because above
we’ve studied an exclusive cross-section, in the sense that we fixed the final state to be
exactly a muon and an antimuon. At higher energies, nothing new happens, because
the final state is fixed.
It has also been very valuable to think about inclusive cross-sections for e+ e− scat-
tering, because in this way you can make anything that the s-channel photon couples
to, if you put enough energy into it. The inclusive cross section for (e+ e− goes to
anything) does grow with energy, and jumps at energies which are thresholds for new
particles in the final state. In this way, for example, we can also make quarks (more
47
See figure 3 here. Now remember that a person doesn’t have much control over their name. By
the way, I totally believe the bit about non-perturbative strings = lint.

139
specifically quark-antiquark pairs) since they also carry electric charge. See Peskin pp
139-140 for a bit more about that, and in particular how this observable gives evidence
that there are three colors of quarks.

e− e− ← e− e− . What happens if instead we scatter two

electrons (Möller scattering)? In that case, the leading
order diagrams are the ones at right. Now the interme-
diate photons have kt = (p1 − p3 ) and ku = (p1 − p4 )
respectively, so that the denominator of the propaga-
tor is t and u in the two diagrams. The evaluation of
these diagrams has a lot in common with the ones for
e+ e− → e+ e− , namely you just switch some of the legs
between initial and final state.

The relation between such amplitudes is called crossing

symmetry. Let’s illustrate it instead for e− µ− ← e− µ− ,
where again there is only one diagram, related by cross-
ing to (6.18). The diagram is the one at right. (The
muon is the thicker fermion line.)

 
(1−ξ)kµ kν
−i ηµν − k2
iM = = (−ieū3 γ µ u1 ))electrons (−ieū2 γ ν u4 )muons (6.17)
k2

with k ≡ p1 − p3 = p2 − p4 . It differs from (6.18) by replacing the relevant vs with

us for the initial/final antiparticles that were moved into final/initial particles, and
relabelling the momenta. After the spin sum,

1 X 2 e4  µ   
ν
     
|M| = 2 tr γ p/4 + mµ γ p/2 + mµ tr γν p/3 + me γµ p/1 + me
4s 4t
1,2,3,4

this amounts to the replacement (p1 , p2 , p3 , p4 ) → (p1 , −p3 , p4 , −p2 ); on the Mandelstam
variables, this is just the permutation (s, t, u) → (t, u, s).
Crossing symmetry more generally. If you look at a Feynman diagram on its
side (for example because someone else fails to use the convention that time goes to the
left) it is still a valid amplitude for some process. Similarly, dragging particles between
the initial and final state also produces a valid amplitude. Making this relation precise
can save us some work. The precise relation for dragging an incoming particle into the

140
final state, so that it is an outgoing antiparticle, is:

iMf ←iA (pf ; pi , pA ) = = iMĀf ←i (pf , k = −pA ; pi ) = .

(If you must, note that this is another sense in which an antiparticle is a particle
going backwards in time.) If A is a spinor particle, the sum relations for particles and
antiparticles are different:
X X
ur (p)ūr (p) = p/ + m, v r (k)v̄ r (k) = k/ − m = −(p/ + m)
r r

– after accounting for k = −pA , they differ by an overall sign. Hence we must also ap-
pend a fermion sign factor (−1)number of fermions shuffled between in and out in the unpolarized
scattering probability. We’ll study a well-motivated example in more detail next.
Mott formula. By studying scattering of an electron from a heavy charged fermion
(a muon is convenient) we can reconstruct the cross section for scattering off a Coulomb
potential (named after Mott). This example will be important next quarter, where we’ll
figure out how it is corrected by other QED processes.
µ+ µ− ← e+ e− . Consider again the process µ+ µ− ← e+ e− . To try to keep things
straight, I’ll call the electron momenta p, p0 and the muon momenta k, k 0 , since that
won’t change under crossing. We found the amplitude

iMµ+ µ− ←e+ e− =

 
(1−ξ)qµ qν
  −i ηµν − q2
 
s µ s0 0 r0 0 ν r
= −ieū (k)γ v (k ) −iev̄ (p )γ u (p) (6.18)
muons q2 electrons

(with q ≡ p + p0 = k + k 0 )48 and the (unpolarized) scattering probability density

1X spinor traces 1 e4 µν
|M|2 = E Mµν ,
4 spins 4 s2

where the tensor objects E µν , M µν come respectively from the electron and muon lines,
1
Eµν = pµ p0ν + p0µ pν − ηµν (p · p0 + m2e )
4
48
Relative to the notation I used earlier, p1 = p, p2 = p0 , p3 = k, p4 = k 0 .

141
 1
Mµν = kµ kν0 + kµ0 kν − ηµν (k · k 0 + m2µ ).
4
and they are contracted by the photon line, with s = q 2 = (p + p0 )2 .
e− µ− ← e− µ− . To get from this the amplitude (tree level, so far) for the process
e− µ− ← e− µ− , we must move the incoming positron line to an outgoing electron line,
and move the outgoing antimuon line to an incoming muon line (hence the sign in σ will
be (−1)number of fermions shuffled between in and out = (−1)2 = 1). Relative to the amplitude
for µ+ µ− ← e+ e− (6.18), we must replace the relevant vs with us for the initial/final
antiparticles that were moved into final/initial particles, and we must replace p0 →
−p0 , k 0 → −k 0 :
t qt
 
(1−ξ)qµ ν
−i ηµν − qt2
iM = = (−ieū(p0 )γ µ u(p)))electrons 2
(−ieū(k)γ ν u(k 0 ))muons(6.19)
qt

with qt ≡ p − p0 = k − k 0 . After the spin sum,

1 X e4
|M|2 = 4 2 −pµ p0ν − p0µ pν − ηµν (−p · p0 + m2e )


4 s,s0 ,r,r0 t
· −kµ kν0 − kµ0 kν − ηµν (−k · k 0 + m2µ )


(6.20)

On the Mandelstam variables, this is just the permutation (s, t, u) → (t, u, s).
Payoff: the Mott formula. Recall other ways of figuring out the scattering cross

section from a Coulomb potential from a point charge of charge ze.

We think about scattering from a fixed electrostatic potential A0 = ze

r
and do classical
mechanics. I can never remember how this goes. Instead, let’s just scatter an electron
off a heavy charge, such as a muon. If the charge of the heavy object were z times that
of the electron, we would multiply the amplitude by z and the cross section by z 2 .
‘Heavy’ here means that we can approximate the
CoM frame by its rest frame, andq its initial and fi-
nal energy as k 0 = mµ , k0 = m2 + ~k 2 = mµ +
0 µ
1~ 2
2
k /mµ + · · · ' mµ . Also, this means the collision
is approximately elastic, E 0 ' E. In the diagram of
the kinematics at right, annoyingly, s ≡ cos θ, c ≡
sin θ. (Sorry.)

142
 This means that the muon-line tensor factor Mµν in (6.20) simplifies dramatically:
 
1
− Mµν ' kµ kν0 + kµ0 kν − ηµν k · k 0 − m2µ  ' δµ0 δν0 2m2µ .
 
4 | {z }
=m2µ −m2µ =0

In the electron line, we’ll need the ingredient.

− p · p0 + m2e = −E 2 + p~2 cos θ + m2e = −~p2 (1 − cos θ). (6.21)

So

E µν Mµν = 32m2µ E 00 = 32m2µ (2E 2 + η 00 (−p · p0 + m2e ))

(6.21)
= 32m2µ (2E 2 − p~2 (1 − cos θ))
trig β 2 ≡~
p2 /E 2
= 32m2µ 2(E 2 − p~2 sin2 θ/2) = 64m2µ E 2 (1 − β 2 sin2 θ/2)
(6.22)
.

From the two-body phase space, the cross section is

1 1 1 z 2 e4 dΩ p
dσ = 2
64m2µ E 2 (1 − β 2 sin2 θ/2)
vrel 2E 2mµ t 16π 2 Etotal
|{z}
=β
Etotal ∼mµ 4E z 2 e4 (1 − β 2 sin2 θ/2)
= dΩ
β t2
Noting that t = (p − p0 )2 = −2~p2 (1 − cos θ), we get
2 2 2
dσ 2 α (1 − β sin θ/2)
=z .
dΩ Mott 4β 2 p~2 sin4 θ/2
If we take β  1 in this formula we get the Rutherford formula. Notice that it blows
up at θ → 0. This is a symptom of the long-range nature of the Coulomb potential,
i.e. the masslessness of the photon.
Electron-proton scattering. The answer is basically the same if we think of the
heavy particle in (6.19) as a proton (we have to flip the sign of the charge but this gets
squared away since there is no interference in this case). ep → ep is called Rutherford
scattering, for good reason49 .
Electron-photon scattering. In the case of the process e− γ ← e− γ, 50
we meet
49
If you don’t know why, you should go read Inward Bound, by Abraham Pais, as soon as possible.
50
which at high energy is called Compton scattering and at low energies is called Thomson scattering.
Despite my previous curmudgeonly footnote chastising the innocent reader for an imagined incomplete
knowledge of the history of science, I do have a hard time remembering which name goes where.
Moreover, as much as I revere the contributions of many of these folks, I find that using their names
makes me think about the people instead of the physics. No one owns the physics! It’s the same
physics for lots of space aliens, too.

143
a new ingredient, namely external photons:

iM = ≡ iMs + iMt
 
ik/s + m ik/t + m
= (−ie)2 µ1 ?ν
4 ū3 γν γµ + γµ γν u2 . (6.23)
s − m2 t − m2
The two amplitudes have a relative plus since we only mucked with the photon contrac-
tions, they just differ by how the gamma matrices are attached. If you don’t believe
me, draw the contractions on this:
/ 1 (Ψ̄AΨ)
hγe| (Ψ̄AΨ) / 2 |γei

(I’m not going to TeX it, thank you).

Now, if we don’t measure the photon polarizations, we need
1 X
P = |M|2 .
4 polarizations, spins

The key ingredient is the completeness relation

X
i? i µ ν
µ (k)ν (k) = −ηµν + something proportional to k k .
i=1,2

We can do various incantations to find a definite coefficient, but it will not matter
because of the Ward identity: anytime there is an external photon (k)µ , the amplitude
is M = Mµ µ (k) and satisfies k µ Mµ = 0. Therefore, we can ignore the term about
which I was vague and we have
X X
|M|2 = i? µ? ν i µ? ν
µ M M ν = −ηµν M M + (terms with Mµ k )
µ

polarizations i
= −M?µ Mµ .

Don’t be scared of the minus sign, it’s because of the mostly minus signature, and
makes the thing positive. But notice the opportunity to get negative probabilities if
the gauge bosons don’t behave!
You get to boil the Compton scattering probability some more on the homework.
Good luck!
A dramatic process related by crossing to Compton scattering is pair annihilation,
Mγγ←e+ e− . See the end of Peskin §5, where he has a nice plot comparing to experi-
dσ
mental data the result for dΩ as a function of scattering angle.
[End of Lecture 20]
To be continued ... here.

144
A Why care about GR ?
From discussions of QFT in particle physics, you might be hard-pressed to understand
why in the world anyone would care about real-time Green’s functions other than the
Feynman Green’s function. Here I would like to try to ameliorate that situation.
The retarded Green’s function for two observables OA and OB is
Z
GOA OB (ω, k) = −i dD−1 xdt eiωt−ik·x θ(t)h[OA (t, x), OB (0, 0)]i
R

θ(t) = 1 for t > 0, else zero.

We care about this because it determines what hOA i does if we kick the system via
OB .
The source is a time dependent perturbation to the Hamiltonian:
Z
δH(t) = dD−1 xφB(0) (t, x)OB (x) .

hOA i(t, x) ≡ Tr ρ(t) OA (x)

= Tr ρ0 U −1 (t)OA (t, x)U (t)
Rt
δH(t0 )dt0
in interaction picture: U (t) = T e−i (e.g. ρ0 = e−βH0 )
linearize in small perturbation:

Z t
δhOA i(t, x) = −iTr ρ0 dt0 [OA (t, x), δH(t0 )]
Z t
= −i dD−1 x0 dt0 h[OA (t, x), OB (t0 , x0 )]iφB(0) (t0 , x0 )
Z
= dD x0 GR (x, x0 )φB (x0 )

fourier transform:

δhOA i(ω, k) = GR
OA OB (ω, k)δφB(0) (ω, k)

Linear response, an example

perturbation: an external electric field, Ex = iωAx
couples via δH = Ax J x where J is the electric current (OB = Jx )
response: the electric current (OA = Jx )

145
 δhOA i(ω, k) = GR
OA OB (ω, k)δφB(0) (ω, k)

it’s safe to assume hJiE=0 = 0:

Ex
hOJ i(ω, k) = GR R
JJ (ω, k)Ax = GJJ (ω, k)
iω

Ohm’s law: J = σE defines the conductivity σ. (Really it is a tensor since J and

E are both vectors.)

GR (ω, k)
=⇒ Kubo formula: σ(ω, k) = JJ
iω

146
B More about the non-perturbative proof of the
Ward identity
[Schwartz §14.8] First, consider a Green’s function from which we might make an S-
matrix element by LSZ,
Z
G ≡ hΩ|T O1 (x1 ) · · · On (xn )|Ωi = DΨeiS O1 (x1 ) · · · On (xn )

where the operators O1 (x1 ) 7→ e−iQi α O1 (x1 ) have charge Qi under a global U(1) sym-
metry. For example the O(x) could be just the elementary field Ψ(x) 51 .
Now change variables in the path integral so that Oi (xi ) 7→ e−iQi α(xi ) Oi (xi ); the
R
action will shift by S 7→ S − α∂µ j µ where j µ is the Noether current. The path integral
doesn’t change at all, so its infinitesimal variation is
Z Z !
X
0 = δG = DΨ − iα∂ µ jµ eiS O1 · · · On − i Qi α(xi )eiS O1 · · · On (2.1)
i
Z " #
X
d µ
= d xα(x) i∂µ hj (x)O1 · · · On i − Qi δ(x − xi )G . (2.2)
i

Since this is true for any α(x), we learn that the thing in square brackets is zero:
∂µ j µ = 0 up to contact terms. This is called the Ward-Takahashi identity.
Now suppose we do this same manipulation in a gauge theory, like QED. The addi-
tional terms in S are Fµν F µν +iAµ Ψ̄γ µ Ψ, which are invariant under the transformation,
so don’t change these statements. Notice that the transformation we’re doing here is
not the gauge transformation, since Aµ doesn’t transform – we’re only doing the gauge
transformation on the matter fields here, so their kinetic terms actually shift and pro-
duce the α∂ µ jµ term above. Photon field insertions in G don’t contribute, since they
have charge zero here.
Next, think about the LSZ formula for an S-matrix element with (say) two external
photons:
Z Z Z
LSZ µ ρ n 4 ipx 4 ipk xk k
M = h, ...k ...|S|...i =  k i d xe 2µν d x1 e 2ρσ ... hAν (x)...Aσ (xk )...i
(2.3)

where 2µν is shorthand for the photon kinetic operator 2µν = 2ηµν − ∂µ ∂ν /µ2 . The
51
You’ll have to trust me for now that the path integral for fermionic fields exists. That’s the only
information about it we’ll need here.

147
Schwinger-Dyson equation for Aµ then implies that

2kρσ 2µν hAν (x)...Aσ (xk )...i = 2kρσ hj µ (x)...Aσ (xk )...i − iδ 4 (x − xk )ηµσ h...i


(2.4)
= hjµ (x)...jσ (xk )i − i2δ(x − xk )ηµσ h...i (2.5)

First of all, this is why I said we could get the S-matrix elements with photons from
correlators with currents. But notice that this is only true up to the contact terms.
But those are disconnected amplitudes which we can ignore.
Finally, set the polarization of one of the photons equal to its momentum  = p.
Then
Z Z Z
µ ρ n 4 ipx 4 ipk xk
dyeiq1 y i∂/y + m1 ... h−i∂µ j µ ...jρ (x1 )...Ψ(y)..i


p Mµ = k i d xe d x1 e

(2.6)
   X
= /q1 − m1 /q2 − m2 · · · Qj G̃(..., qj ± p, ...) (2.7)
j

where the ± depends on whether particle j is incoming or outgoing. At the last step
we used the Fourier transform of (2.2).
Now here’s the punchline: The G̃ on the RHS of (2.7) has poles at (qj ± p)2 = m2j ,
q 2 −m2
and not at qj2 = m2j . So when it’s multiplied by /qj − mj = j/q+mj it will vanish. End of
story. Notice that no use of perturbation theory was made here.

148