Quantum Field Theory II

Babis Anastasiou
Institute for Theoretical Physics, ETH Zurich,
8093 Zurich, Switzerland
E-mail: babis@phys.ethz.ch
July 24, 2008
Abstract
The subject of the course is modern applications of quantum eld
theory with emphasis on the quantization of non-Abelian gauge theo-
ries. The following topics are discussed:
Classical gauge transformations.
Quantization for fermionic and bosonic elds and perturbation
theory with path-integrals is developed.
Quantization of non-Abelian gauge-theories.
The Fadeev-Popov method for gauge xing.
BRST symmetry.
The quantum eective action and the eective potential.
Classical symmetries of the eective action.
Slavnov-Taylor identities.
The Zinn-Justin equation.
Power-counting and ultraviolet innities in eld theories.
Renormalizable Lagrangians.
Renormalization and symmetries of non-Abelian gauge theories.
1
Contents
1 Non-abelian gauge theories 4
1.1 Gauge invariance . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Non-abelian (global) SU(N) transformations . . . . . . . . . . 6
1.3 Local non-abelian gauge symmetries . . . . . . . . . . . . . . . 9
2 Path integral quantization 15
2.1 Path integral formulation of quantum mechanics . . . . . . . . 15
2.2 An adventurous transition . . . . . . . . . . . . . . . . . . . 22
2.3 Functional dierentiation . . . . . . . . . . . . . . . . . . . . . 26
2.4 Vacuum to vacuum transitions . . . . . . . . . . . . . . . . . . 27
2.5 The simple harmonic oscillator . . . . . . . . . . . . . . . . . . 29
3 Path integrals and scalar elds 34
3.1 Functional Integration . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Path integrals and interacting elds . . . . . . . . . . . . . . . 40
3.3 Perturbation theory for
4
/4! interactions . . . . . . . . . . 42
3.4 Fermionic path integrals . . . . . . . . . . . . . . . . . . . . . 45
4 Quantization of non-abelian gauge theories 51
4.1 Feynman rules for QCD . . . . . . . . . . . . . . . . . . . . . 59
4.2 BRST symmetry . . . . . . . . . . . . . . . . . . . . . . . . . 64
5 Quantum eective action and the eective potential 71
5.1 The eective potential . . . . . . . . . . . . . . . . . . . . . . 81
6 Symmetries of the path integral and the eective action 84
6.1 Slavnov-Taylor identities . . . . . . . . . . . . . . . . . . . . . 84
6.2 Symmetry constraints on the eective action . . . . . . . . . . 85
6.3 Contraints on the eective action from BRST symmetry trans-
formations of the classical action . . . . . . . . . . . . . . . . 87
6.4 Slavnov-Taylor identities in QED . . . . . . . . . . . . . . . . 89
7 Renormalization: counting the degree of ultraviolet diver-
gences 93
7.1 Cancelation of supercial divergences with counterterms . . . 98
7.2 Nested and overlapping divergences . . . . . . . . . . . . . . . 101
2
8 Proof of renormalizability for non-abelian gauge theories 102
3
1 Non-abelian gauge theories
1.1 Gauge invariance
We will introduce a new principle which is found in realistic quantum eld
theories, i.e. theories such as QED, QCD and the Standard Model. These
theories describe nature to the experimentally accessible accuracy. Their
Lagrangian is invariant under local gauge transformations. We start with
the familiar case of Quantum Electrodynamics. You have already seen the
Lagrangian in the course of QFT I. Here we will construct it by imposing
invariance under local U(1) transformations. We consider a free fermion eld:
L =

(x) (i m) (x) (1)
and dene a transformation:
(x)

(x) = exp(ig)(x). (2)

We rst consider a global gauge transformation
_

x
= 0
_
. The free La-
grangian is clearly invariant under the gauge transformation of Eq. 2. How-
ever, under a local gauge transformation,
U(x) = exp (ig(x)) , (3)
the Lagrangian is no longer invariant.

+

e
ig
_
e
ig
_
. (4)
We cannot make any local gauge transformation which leaves invariant
the Lagrangian of a single eld. The problem was that
e
ig
+ something else .
We will modify the derivative so that it transforms more conveniently. We
will look for a new derivative which, under a local gauge transformation
transforms as:
D U(x)D. (5)
We add to the normal derivative a new function (eld):
D

igA

(x). (6)
4
We need:
D

(x) D

= U(x)D

;
_

igA

_
(U(x)) = U(x) (

igA

)
; U(x)

+ [

U(x)] igA

U(x) = U(x)

igA

U(x)
; A

= A

i
g
U
1
(x)

U(x) (7)
The covariant derivative transforms as:
D

igA

ig
_
A

i
g
U
1
U
_
=

igA

U
1
(

U)
=

igA

+ U
_

U
1
_
= U(x) (

igA

) U
1
(x) (8)
Therefore:
D

= U(x)D

U
1
(x) (9)
We can now replace the free Lagrangian of the spin-1/2 eld with a new
Lagrangian which is also gauge invariant.
L =

[iD m]

U
1
U [iD m] U
1
U

[iD m] .
If A

is a physical eld, we need to introduce a kinetic term in the La-

grangian for it. We will insist on constructing a fully gauge invariant La-
grangian. To this purpose, we can use the covariant derivative as a building
block. Consider the gauge transformation of the product of two covariant
derivatives:
D

= UD

U
1
UD

U
1
= UD

U
1
.
This is not a gauge invariant object. Now look at the commutator:
[D

, D

]
_
D

, D

_
= U [D

, D

] U
1
(10)
5
This is gauge invariant. To convince ourselves we write the commutator
explicitly:
[D

, D

] = (

igA

) (

igA

) [ ]
=

ig (

) igA

igA

+ (ig)
2
A

[ ]
= ig [

] . (11)
Inserting Eq. 11 into Eq. 10, we nd that the commutator of covariant deriva-
tives (in the abelian case) is gauge invariant. We have also found that it is
proportional to the eld strength tensor of the gauge (photon) eld:
F

=
i
g
[D

, D

] =

(12)
We now have invariant terms for a Lagrangian with an electron and a
photon eld. The Lagrangian for QED reads
L =

(D m)
1
4
F

. (13)
Exercise: Find the Noether current and conserved charge due to the invari-
ance under the U(1) gauge transformation of the QED Lagrangian.
1.2 Non-abelian (global) SU(N) transformations
We can think of more complicated transformations than a simple phase on
a complex eld. Lets consider as an example a collection of N scalar elds.
A simple Lagrangian for them could be:
L = (

i
) (

i
) m
2
(
i

i
)

4
(
i

i
)
2
, i = 1 . . . N. (14)
The Lagrangian is required to be invariant if we perform a transformation

i
= V
ij

j
, (15)
where a summation convention is assumed (this will always be the case unless
explicitly stated otherwise).

i
= V

ij

j
V
ik

k
= V

ji
V
ik

k
=

i
, (16)
6
if
V

ji
V
ik
=
jk
. (17)
This is a unitary U(N) transformation under which our example Lagrangian
is invariant. The Lagrangian is also invariant under a stricter SU(N) trans-
formation. We can write
V = e
i
U,
such that
U

U = 1 and detU = 1. (18)

Exercise: Think of a concrete example for a unitary 2 2 matrix V where
this can be done. The N N matrices U represent the group of special
unitary transformations SU(N).
It is sucient to study small SU(N) transformations. Due to them form-
ing a group, large transformation can be obtained by repeating (innitely)
many small ones. We write:
U
ij
=
ij
i
a
T
a
ij
+O(
2
), (19)
where we choose
a
to be real parameters. The N N matrices T
a
are
generators of SU(N) matrices. They are N
2
1: An arbitrary N N
complex matrix has 2N
2
real elements. For a unitary matrix U

= U
1
,
only N
2
elements are independent. The specialty condition detU = 1 adds
one more constraint, leaving N
2
1 independent elements. Remember the
dimensionality of the indices concerning the generators.
T
a
i,j
: a = 1 . . . (N
2
1) and i, j = 1 . . . N.
They will be needed in various situations.
Exercise SO(N) group: N N matrices with
R
ij
R
kl

jl
=
ik
.
Find the number of generators.
Exercise The symplectic group Sp(2N) can be dened as 2N 2N matrices
S with
S
ij
S
kl

jl
=
ik
,
where,

ij
=
ji
and
2
= 1.
7
Find the number of generators.
The SU(N) generators are hermitian:
U

U = 1
;
_
1 + ig(T
a
)

a
_ _
1 igT
b

b
_
= 1 +O(
2
)
; T
a
= T
a
(20)
and traceless:
detU = 1 ;log (detU) = 0
; Tr (log U) = 0
; Tr log (1 i
a
T
a
) = 0
; Tr (i
a
T
a
) = 0
; Tr (T
a
) = 0 (21)
We can choose a normalization condition for the SU(N) generators. By
convention, we choose:
Tr
_
T
a
T
b
_
T
a
ij
T
b
ji
=

ab
2
. (22)
A very basic property of the generators is that they satisfy a Lie algebra:
_
T
a
, T
b
_
= if
abc
T
c
, (23)
where f
abc
are the structure constants of the algebra.
Exercise: Prove Eq. 23 by considering a transformation U
1
U
1
U

U, with
U, U

independent SU(N) transformations.

From Eq. 23, we can derive
_
T
a
, T
b
_
= if
abd
T
d
;
_
T
a
, T
b
_
T
c
= if
abd
T
d
T
c
; f
abc
= 2iTr
__
T
a
, T
b
_
T
c
_
. (24)
Exercise: Prove that the structure constants are fully antisymmetric and real
8
1.3 Local non-abelian gauge symmetries
We now consider N elds
i
(scalar or spinor). We are interested in La-
grangians which are invariant under a local SU(N) transformation:

i
(x)

i
(x) = U
ij
(x)
j
(x). (25)
As in QED, the building block for the construction of a gauge invariant
Lagrangian will be a covariant derivative:
D

igA

, (26)
such that
D

= UD

, with U

= U
1
. (27)
For a scalar eld we have
(D

)(D

) (D

)(D

)
Similarly, the kinetic term with the same covariant derivative for a fermion
eld is invariant. There are many similarities with QED, however there are
also many important dierences. Lets start by pointing out that the gauge
eld A

is an N N matrix.
We can easily nd the transformation for the gauge eld:
D

= U(x)D

(x)
;
_

igA

_
= U (

igA

) U

igA

+ U
_

_
UigA

; A

= U(x)A

(x) +
i
g
U(x)
_

(x)
_
(28)
This formula is analogous to the gauge transformation of the photon in QED.
However, here the gauge eld A

and the gauge transformation U are complex

N N matrices rather than complex numbers.
We now compute the corresponding commutator:
[D

, D

] = (

igA

) (

igA

) [ ]
=

ig (

) igA

igA

+ (ig)
2
A

[ ]
= ig {

ig [A

, A

]} . (29)
9
The commutator term was absent in the case of QED.
The gauge eld strength:
G

i
g
[D

, D

] =

ig [A

, A

] , (30)
is no longer gauge invariant:
G

= U(x)G

(x).
However, the trace
Tr(G

)
Tr(UG

UG

)
Tr(U

UG

UG

)
Tr(G

)
is gauge invariant.
We can expand the gauge eld in the basis of generators:
A

= A
a

T
a
, (31)
Equivalently,
A
a

= 2Tr(A

T
a
) (32)
We also have
G

= G
a

T
a
, (33)
with
G
a

= 2Tr(G

T
a
). (34)
It is
G
a

= 2Tr(G

T
a
)
= 2Tr (

T
a

T
a
ig [A

, A

] T
c
)
=

A
a

A
a

+ gf
abc
A
b

A
c

. (35)
A concrete example is the Lagrangian for QCD which is invariant under
SU(3) gauge transformations.
L =

(iD m1)
1
4
G
c
G
c

. (36)
10
Exercise:Expand all terms in the QCD Lagrangian using the explicit expres-
sions in terms of the gauge eld for the covariant derivative and the gauge
eld strength. Sketch the interactions (the precise Feynman rules will be de-
rived in forthcoming lectures)
Exercise:Write a gauge invariant Lagrangian under SU(N) transformations
for a scalar eld. This case appears in supersymmetric theories for the scalar
partners of quarks. Sketch the interactions.
Adjoint representation:
By expanding the commutators we can prove that
__
T
a
, T
b
_
, T
c
_
+
__
T
b
, T
c
_
, T
a
_
+
_
[T
c
, T
a
] , T
b
_
= 0. (37)
From this we derive a relation for the structure constants:
;
_
f
abd
T
d
, T
c
_
+
_
f
bcd
T
d
, T
a
_
+
_
f
cad
T
d
, T
b
_
= 0
; f
abd
f
dce
+ f
bcd
f
dae
+ f
cad
f
dbe
= 0 (38)
We dene the matrices

T
b
ac
= if
abc
. (39)
Then the above relation can be written as
_

T
b
,

T
c
_
= if
bcd

T
d
. (40)
Therefore, the matrices

T furnish a representation of the same Lie algebra.
This is called the adjoint representation.
We can now consider elds
a
which transform in the adjoint represen-
tation. An example is the gluino, the supersymmetric partner of the gluon,
which transforms in the adjoint.

a
= U
A
ab

b
with U
ab
A
= e
i
c
T
c
ab
and a, b, c = 1 . . . (N
2
1).
(41)
Or, for small transformations,

a
=
a
i
b

T
b
ac

c
;

a
=
a
i
b
_
if
abc

c
_
;

a
=
a
+ f
abc

c
(42)
11
We can nd a covariant derivative for the transformations of the adjoint
representation in a complete analogy as for the fundamental representation:
D

a
=

a
igA
b

T
b
ac

c
=

a
+ gf
abc
A
b

c
. (43)
Now consider a general member of the Lie algebra =
a
T
a
. The covariant
derivative acts on it like:
D

=
_

a
+ gf
abc
A
b

c
_
T
a
=

ig
_
T
b
, T
c
_

c
A
b

, (44)
or equivalently
D

ig [A

, ] ,
a
T
a
, A

A
a

T
a
. (45)
Exercise: Take an element =
a
T
a
of the Lie algebra, transforming in
the adjoint representation, and a eld transforming in the fundamental
representation. Prove the Leibniz rule for the covariant derivative:
D

() = (D

) + (D

) . (46)
Euler-Lagrange equations/ conserved currents:
We consider the variation of the gauge eld
A

+ A

and

+ (

) .
The corresponding variation of the gauge eld strength is
G

= (

ig [A

, A

])
=

ig [A

, A

] ig [A

, A

]
; G

= D

(A

) D

(A

) . (47)
We can now look at the variation of this term in the QCD action

_
d
4
xTr (G

)
12
= 2
_
d
4
xTr (G

)
= 2
_
d
4
xTr (G

[D

(A

) D

(A

)])
= 4
_
d
4
xTr (G

(A

)) antisymmetry
= 4
_
d
4
x{D

[Tr (G

(A

))] Tr [D

(G

) A

]} (48)
The rst trace is gauge invariant. We can then nd a gauge transformation
for the gluon eld so that D

UD

, and drop the surface term.

So, we have:

_
d
4
xTr
_
1
2
G

_
= 2
_
d
4
xTr [D

(G

) A

] . (49)
The fermionic term in the Lagrangian varies as:

_
d
4
x

(iD m) =
_
d
4
x

(i + gA)
= g
_
d
4
x

A = g
_
d
4
x

T
a
A
a

=
= 2g
_
d
4
x
_

T
a
ij

j
_
Tr [A

T
a
]
= 2gTr [J

] , (50)
with
J

= J
a

T
a
, (51)
and
J
a

= g

T
a
. (52)
Combining the variation of both fermionic and gauge boson terms in the
Lagrangian of Eq. 36 we derive the Euler-Lagrange equation:
D

= J

. (53)
We can act with a second covariant derivative on the above equation:
D

= D

(54)
The lhs is:
D

= D

ig [A

, G

])
13
=

ig [

, G

]
ig [A

] ig [A

]
g
2
[A

[A

, G

]] (55)
Using that G

= G

and
[A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0
you can prove that
D

=
ig
2
[G

, G

] = 0 (56)
Therefore:
D

= 0. (57)
The fermionic current is thus no longer (as in QED) a conserved current. It
is rather covariantly conserved!
Exercise: In an abelian gauge theory, consider the dual tensor

=
1
2

.
Show that
F

, (58)
with
K

Exercise: In a non-abelian gauge theory, consider the dual tensor

=
1
2

.
Show that
G

, (59)
with
K

Tr
_
G

+
2
3
A

_
.
14
2 Path integral quantization
In the previous lessons we constructed Lagrangians and studied them at the
classical level. We will use eld theories in order to describe particles which
can be created from the vacuum or annihilated into it and their interactions.
This can be done by quantizing the classical eld theories. You have carried
this quantization program for simple eld theories in the course of QFT I,
by means of canonical quantization imposing commutation and anticom-
mutation relations. Here we will quantize gauge invariant eld theories with
an alternative method, using a formalism based on path integrals.
2.1 Path integral formulation of quantum mechanics
As a warm up we will quantize systems from classical mechanics, and we
will see how quantum mechanics can be postulated dierently by using path
integrals.
We consider a quantum mechanical state | which satises the Schrodinger
equation:
i h
t
| =

H| . (60)
The solution of this equation
|(t
2
) = e

i
h

H(t
2
t
1
)
|(t
1
) , (61)
determines the evolution of this state from an initial moment t
1
to a later
moment t
2
. The wave function (x, t) x| (t) is then
(x, t
2
) = x| e

i
h

H(t
2
t
1
)
|(t
1
) (62)
The sum of the Hamiltonian eigenstates

H| = E

| (63)
is unitary,
1 =

| | . (64)
Inserting it in the rhs of the evolution equation we nd
(x, t
2
) = x| e

i
h

H(t
2
t
1
)
_
_

| |
_
_
|(t
1
)
=

i
h
E

(t
2
t
1
)
x| | (t
1
) . (65)
15
We now insert another unit operator
1 =
_
d
3
x

|x

| , (66)
obtaining
(x, t
2
) =

i
h
E

(t
2
t
1
)
x| |
__
d
3
x

|x

|
_
|(t
1
)
=
_
d
3
x

_
_
_

i
h
E

(t
2
t
1
)
x| | x

_
_
_
(x

, t
1
) . (67)
In other words, if we know the wave-function at one time, we can determine
it fully at a later time by integrating it,
(x, t
2
) =
_
d
3
x

K(x, x

; t
2
t
1
) (x

, t
1
) , (68)
with a kernel
K(x, x

; t
2
t
1
) =

i
h
E

(t
2
t
1
)
x| | x

, (69)
which depends on the Hamiltonian of the system and the elapsed time t
2
t
1
during the evolution of the quantum state. This integration kernel is called
the propagator. For t
2
= t
1
the propagator is a delta function
K(x, x

; t
1
t
1
) =

e
0
x| | x

=
3
(x x

). (70)
Exercise: Prove that the propagator K(x, x

; t t

) satises the Schrodinger

equation in the variables x, t for times t > t

.
Exercise: Prove that
_

dxe
ax
2
=
_

a
.
Exercise: Compute the propagator K(x, x

; t t

) for
a free particle,
the simple harmonic oscillator.
16
The propagator is the amplitude for a particle measured at position |x

at time t
1
to propagate to a new position |x at time t
2
> t
1
. We can verify
this easily. Consider a particle measured at a position x

. After time t
2
t
1
it will be evolved to a new state
e

i
h

H(t
2
t
1
)
|x

. (71)
The probability amplitude to be measured at a position x is
x| e

i
h

H(t
2
t
1
)
|x

= x| e

i
h

H(t
2
t
1
)
_
_

| |
_
_
|x

i
h
E

(t
2
t
1
)
x| | x

= K(x, x

; t
2
t
1
). (72)
We can attempt to compute the propagator for a transition which takes
a very small time t 0.
K(x, x

; t) = x| e

i
h

Ht
|x

= x| 1
i
h

Ht |x

+O
_
t
2
_
= x|
_
1
i
h

Ht
___
d
3
p |p p|
_
|x

+O
_
t
2
_
=
_
d
3
p
_
x| p p| x

i
h
t x|

H|p p| x

_
+O
_
t
2
_
(73)
We now specialize to Hamiltonian operators of the form

H = f
1
( p) + f
2
( x).
Then
x|

H|p = Hx| p , (74)
where H is not an operator anymore and it is simply a number. Recall that
the position and momentum states are related via a Fourier transform,
x| p =
e
i
h
px

2 h
. (75)
For simplicity let us consider one only dimension; the three-dimensional case
is a faithful repetition of the same steps. Then we nd for the propagator at
17
small time intervals:
K(x, x

; t) =
_
dp
2 h
e
i
h
p(xx

)
_
1
i
h
Ht
_
+O
_
t
2
_
=
_
dp
2 h
exp
_
i
h
{p (x x

) Ht}
_
+O
_
t
2
_
. (76)
An interesting form for the propagator for small time intervals arises when
the Hamiltonian is of the form
H =
p
2
2m
+ V (x). (77)
Then,
K(x, x

; t) =
_
dp
2 h
exp
_
i
h
_
p (x x

)
p
2
2m
t V (x)t
__
+ O
_
t
2
_
=
_
dp
2 h
exp
_
_
i
h
_
_
_

t
2m
_
p m
x x

t
_
2
+ t
1
2
m
_
x x

t
_
2
V (x)t
_
_
_
_
_
+O
_
t
2
_
=
_
d p
2 h
exp
_
_
it
h
_
_
_

p
2
2m
+
1
2
m
_
x x

t
_
2
V (x)
_
_
_
_
_
+O
_
t
2
_
. (78)
Finally,
K(x, x

; t)
1
N(t)
exp
_
_
it
h
_
_
_
1
2
m
_
x x

t
_
2
V (x)
_
_
_
_
_
, (79)
with
1
N(t)
=
_
dp
2 h
exp
_
itp
2
2m h
_
=
_
m
2i ht
_
1/2
(80)
The path integral:
We consider now the transition from an initial position (x
i
, t
i
) to a nal
position (x
f
, t
f
) with propagator:
K(x
f
, x
i
; t
f
t
i
).
We can take a snapshot at an intermediate time t
1
during this transition:
t
i
< t
1
< t
f
.
18
If the particle is measured at a position x
1
at the moment t
1
, then the
amplitude for the full transition will be:
K(x
f
, x
i
; t
f
t
i
) = K(x
f
, x
1
; t
f
t
1
)K(x
1
, x
i
; t
1
t
i
). (81)
If we dont perform a measurement of the particle at t
1
, we should add up
together all the amplitudes for the particle to have performed this transition
via any point: We then have:
K(x
f
, x
i
; t
f
t
i
) =
_

dx
1
K(x
f
, x
1
; t
f
t
1
)K(x
1
, x
i
; t
1
t
i
). (82)
We can decide to take more than one snapshots during the transition from
x
i
to x
f
, in times t
i
< t
1
< t
2
< . . . < t
n
< t
f
. Using again the superposition
principle we must write:
K(x
f
, x
i
; t
f
t
i
) =
_

dx
1
. . . dx
n
K(x
f
, x
n
; t
f
t
n
)K(x
n
, x
n1
; t
n
t
n1
)
. . . K(x
2
, x
1
; t
2
t
1
)K(x
1
, x
i
; t
1
t
i
). (83)
We can now consider innitesimal equally fast t
k+1
t
k
= t =
t
f
t
i
n+1
inter-
mediate transitions. Then we obtain
K(x
f
, x
i
; t
f
t
i
) = lim
n
_

dx
1
. . . dx
n
K(x
f
, x
n
; t)K(x
n
, x
n1
; t)
. . . K(x
2
, x
1
; t)K(x
1
, x
i
; t). (84)
We have discretized and taken the innite limit, which is the dening proce-
dure of an integration. However, this new integral is rather unusuall: for each
time interval t, we consider the transition from a point x
j
to a new point
x
j+1
; then we integrate over all possibilities for the successive points that we
considered. Let us now x the initial and nal points x
i
and x
f
and think of
all possible paths that a particle may follow in going from x
i
x
f
. All the
points which form all the paths are accounted for by the limit of innitesimal
n transitions of Eq. 84. We interpret the rhs of this equation as an
integral over all paths that a particle may take in going from x
i
x
f
.
We can insert the expressions for the propagator at small time intervals
of Eq. 76 or Eq. 79 into Eq. 84. Notice that in the case Eq. 79 I can use that
x
n
x
n1
t
x
n
19
in the limit t 0. Then Eq. 84 becomes:
K(x
f
, x
i
; t
f
t
i
) = lim
n
1
N(t)
n
_

dx
1
. . . dx
n
exp
_
it
h
_
m
2
x
1
2
V (x
1
)
_
_
exp
_
it
h
_
m
2
x
2
2
V (x
2
)
_
_
. . .
exp
_
it
h
_
m
2
x
f
2
V (x
f
)
_
_
. (85)
Equivalently,
K(x
f
, x
i
; t
f
t
i
) = lim
n
1
N(t)
n
_

dx
1
. . . dx
n
exp
_
i
h
{L(x
1
(t
1
))t + L(x
2
(t
2
))t + . . . L(x
f
(t
n
))t}
_
, (86)
where
L(x) =
m
2
x
2
V (x),
is the Lagrangian of the system. We now can have a more concrete picture
about the above integral. We integrate over all paths connecting the xed
points x
i
and x
f
.We write, symbolically,
K(x
f
, x
i
; t
f
t
i
) =
1
N
_
Dxexp
_
i
h
_
t
f
t
i
dtL[x(t)]
_
. (87)
Even shorter, we can write:
K(x
f
, x
i
; t
f
t
i
) =
1
N
_
Dxexp
_
i
h
S[x]
_
. (88)
The sum (integral) in the exponential was the action integral evaluated on
each of the paths S[x].
Exercise: Consider a Lagrangian of the form
L =
1
2
f(x) x
2
+ g(x) x V (x).
This is more similar to a quantum mechanics analog of the tr [G

] terms
in a non-abelian gauge theory if we identify A

x and

x.
20
a) Compute the Hamiltonian
b) Compute the propagator for a small transition c) Write the path-integral
expression for the propagator at large time integrals. Notice that the measure
of the path integration is modied
Dx Dxf(x)
1
2
21
2.2 An adventurous transition
We look now at a more eventful transition amplitude. We rst prepare a
particle on an initial position x
i
at a time t
i
and let it evolve for some time
t t
i
according to a Hamiltonian

H:
e

i
h

H(tt
i
)
|x
i

At the time t, something abrupt occurs (e.g. an interaction with another

particle which was originally far away) and modies the particle state. We
will see later how we can describe interactions of particles using path inte-
grals; now, let us consider an easy modication of the state where the state
is mixed up in a simple way, acting on it with the position operator:
xe

i
h

H(tt
i
)
|x
i

Then we allow the particle to evolve undistracted for a time t

f
t,
e

i
h

H(t
f
t)
xe

i
h

H(tt
i
)
|x
i
,
and then we place a detector at x
f
:
x
f
| e

i
h

H(t
f
t)
xe

i
h

H(tt
i
)
|x
i
.
We can compute this matrix-element as a path integral. Before we proceed,
we should use some language which is more convenient to describe eventful
transitions. We can write the same transition amplitude as:
_
x
f
| e

i
h

Ht
f
_ _
e
i
h

Ht
xe

i
h

Ht
_ _
e
i
h

Ht
f
|x
f

_
= x
f
, t
f
| x(t) |x
i
, t
i
. (89)
We have dened states
|, t e
i
h

Ht
| , (90)
which refer to a xed moment t only and do not evolve
i h

t
|, t = i h

t
_
e
i
h

Ht
|
_
= e
i
h

Ht
_
i h

|
_
+ i h
_

t
e
i
h

Ht
_
|
= e
i
h

Ht
_

H

H
_
| = 0. (91)
22
We have dened operators which do change with time

O(t) = e
i
h

Ht

Oe

i
h

Ht
. (92)
As you recognize, this is the Heisenberg picture of evolution. For us, it
is convenient to assign a time date on a state which denotes a particle or a
collection of particles at the beginning of an experiment or at the end of it.
However, one could have equally well chosen to work in the probably more
familiar Schrodinger picture.
Lets us now compute this eventful transition:
x
f
, t
f
| x(t
j
) |x
i
, t
i
with t
i
< t < t
f
,
following the method we used for the simple transition x
f
, t
f
| x
i
, t
i
=
x
f
| e
i
h

H(t
f
t
i)
|x
i
, and subdividing the transition in small time intervals.
We will nd a similar/related path integral for the new case as well.
x
f
, t
f
| x(t
j
) |x
i
, t
i
=
_
dx
1
. . . dx
j1
dx
j
. . . dx
n
x
f
, t
f
| x
n
, t
n
. . . x
j+1
, t
j+1
| x
j
, t
j

x
j
, t
j
| x(t
j
) |x
j1
, t
j1

x
j1
, t
j1
| x
j2
, t
j2
. . . x
1
, t
1
| x
i
t
i
. (93)
The subdivision of time is carefully chosen. We have
x(t
j
) |x
j
, t
j
=
_
e
i
h

Ht
j
xe

i
h

Ht
j
_ _
e
i
h

Ht
j
|x
j

_
= e
i
h

Ht
j
( x|x
j
)
= x
j
e
i
h

Ht
j
|x
j

= x
j
|x
j
, t
j

;x
j
, t
j
| x(t
j
) |x
j1
, t
j1
= x
j
x
j
, t
j
| x
j1
, t
j1
. (94)
We then obtain for the transition amplitude the same succession of propaga-
tors as for the simple transition multiplied with an additional factor x
j
:
x
f
, t
f
| x(t
j
) |x
i
, t
i
=
_
dx
1
. . . dx
j
x
j
. . . dx
n
x
f
, t
f
| x
n
, t
n
. . . x
j+1
, t
j+1
| x
j
, t
j
x
j
, t
j
| x
j1
, t
j1
. . . x
1
, t
1
| x
i
, t
i
. (95)
23
Introducing, as before, the exlpicit form for the propagator during a small
time transition
x
b
, t
n+1
| x
a
, t
n

_
m
2i h (t
n+1
t
n
)
_1
2
exp
_
_
i
h
_
_
_
m
2
_
x
b
x
a
t
n+1
t
n
_
2
V (x
a
)
_
_
_
(t
m+1
t
m
)
_
_
and taking equal time intervals, we obtain the path integral
x
f
, t
f
| x(t
j
) |x
i
, t
i
= lim
n
1
N(t)
n
_

dx
1
. . . dx
j
x
j
. . . dx
n
exp
_
n

r=1
L(x
r
, x
r
)t
_
with t =
t
f
t
i
n + 1
. (96)
In compact notation we can write
x
f
, t
f
| x() |x
i
, t
i
=
_
Dxx()e
i
h
S[x]
. (97)
Exercise:
a) Prove that
x
f
, t
f
| x() |x
i
, t
i
=
_

dxx x
f
, t
f
| x, x, | x
i
, t
i
.
b) Evaluate the above integral explicitly for a free particle
c) Observe the dependence of the result on the intermediate time
d) You (could) have considered the simpler matrix element x
f
, t
f
| x
i
, t
i

and written down an analogous integral. Observe how the intermediate

time drops out from the nal expression, when nothing special occurs
then!
We can work further on the form of the new path integral that we have
found. Remember the procedure to compute exponential integrals of the
form
I
n
=
_

dxx
n
e
ax
2
with n = 1, 2, . . . (98)
after we have worked out the result
I
0

_

dxe
ax
2
=
_

a
. (99)
24
We can compute I
n
with a rather very simple dierentiations, after we add
a source term on the exponent of the integrand
I
J
=
_

dxe
ax
2
+Jx
=
_

dxe
a(x
J
2a
)
2
+
J
2
4a
=
_

d xe
a x
2
+
J
2
4a
=
_

a
e
J
2
4a
(100)
To compute the I
1
it is sucient to dierentiate the last expression with
respect to the source and substitute J = 0.
dI
J
dJ

J=0
=
_

dx
de
ax
2
+Jx
dJ

J=0
=
_

dxxe
ax
2
.
Similarly,
d
n
I
J
d
n
J

J=0
=
_

dxx
n
e
ax
2
.
Adding a source term to the exponent does not increase the diculty of the
computation and it allows us to calculate all integrals where the integrand
is multiplied with a polynomial in the integration variable. This is very
suggestive, and we will do the same trick for path integrals such as the
one that we found in Eq. 96. We then add a linear term (source) in the
Lagrangian; this is only a computational trick and eventually we will compute
all interesting physical quantities as in the above example with the source
term set to zero. The simple transition from a state |x
i
, t
i
to a state |x
f
, t
f

in the presence of the source has a probability amplitude:

x
f
, t
f
| x
i
, t
i

J
= lim
n
1
N(t)
n
_
dx
1
. . . dx
n
e
i
h

k
dt
k
(L(x
k
, x
k
)+J
k
x
k
)
(101)
We can then compute
x
f
, t
f
| x(t
l
) |x
i
, t
i
=
h
i

J
l
x
f
, t
f
| x
i
, t
i

J
l
=0
. (102)
Let us now dierentiate two times with respect to the source.
_
h
i
_
2

2
J
l
J
q
x
f
, t
f
| x
i
, t
i

J
l,q
=0
= lim
n
1
N(t)
n
_
dx
1
. . . dx
l
x
l
. . . dx
q
x
q
. . . dx
n
e
i
h

k
dt
k
(L(x
k
, x
k
)+J
k
x
k
)
. (103)
25
We can recognize the rhs as the expectation value for the product of two
position operators x(t
l
) x(t
q
) if t
q
is earlier than t
l
or x(t
q
) x(t
l
) otherwise.
We then write
x
f
, t
f
| T ( x(t
l
) x(t
q
)) |x
i
, t
i
=
_
h
i
_
2

2
J
l
J
q
x
f
, t
f
| x
i
, t
i

J
l,q
=0
, (104)
where we have introduced the notation T(

O(t
1
)

O(t
3
)

O(t
2
) . . .

O(t
n
)) to re-
mind us that we should put the operators in the correct time order once the
sequence of the moments t
i
is known. For example, if t
1
< t
2
< t
3
< . . . < t
n
we should write
T(

O(t
1
)

O(t
3
)

O(t
2
) . . .

O(t
n
)) =

O(t
n
) . . .

O(t
3
)

O(t
2
)

O(t
1
).
Exercise: Compute x
f
, t
f
| T ( x(t
l
) x(t
q
)) |x
i
, t
i
for a free particle.
2.3 Functional dierentiation
It is cumbersome to work with path integrals by writing explicitly the inite
limit of discretized paths. We introduced earlier a more compact notation,
x
f
, t
f
| x
i
, t
i

J
=
1
N
_
Dxe
i
h
_
t
f
t
i
dt[L(x(t), x(t))+J(t)x(t)]
. (105)
We can write neatly expressions for the expectation values of operators,
x
f
, t
f
| x(t
1
) |x
i
, t
i
=
1
N
_
Dx x(t
1
) e
i
h
_
t
f
t
i
dtL(x(t), x(t))
or
x
f
, t
f
| T ( x(t
1
) x(t
2
)) |x
i
, t
i
=
1
N
_
Dx x(t
1
) x(t
2
) e
i
h
_
t
f
t
i
dtL(x(t), x(t))
by dening a functional derivative. We consider an integral F[f] over a func-
tion f(t) (for example a path-line). We dene a functional derivative by
changing slightly the function f(y):
F[f(y)]
f(t)
= lim
0
F [f(y) + (y t)] F [f(y)]

. (106)
26
For example, consider the derivative of the action integral in the presence of
a source with respect to the source:

_
dy {L(x(y), x(y)) + J(y)x(y)}
x(t)
=
_
dy(t y)x(y)
= x(t). (107)
Practically, we need the chain rule and to remember that
f(x)
f(y)
= (x y).
The expectation values of time-ordered operators can then be written as
x
f
, t
f
| T ( x(t
l
) . . . x(t
n
)) |x
i
, t
i
=
_
h
i
_
n

n
J(t
1
) . . . J(t
n
)
x
f
, t
f
| x
i
, t
i

J=0
.
(108)
2.4 Vacuum to vacuum transitions
Field theory allows us to compute transitions between states with dierent
particle content. Interestingly, we can build particle states acting with cre-
ation (or eld) operators on the ground state which contains no particles
(vacuum). All transition amplitudes can be described as expectation values
of operators in the vacuum:
0, t
f
| T(. . .) |0, t
i
.
We can then restrict our discussion to expectation values of operators in the
vacuum. We describe this transition using path integrals. First, we try a
direct approach
0, t
f
| T(. . .) |0, t
i
=
_
dxdx

0, t
f
| x, t x, t| T(. . .) |x

, t

, t

| 0, t
i
.(109)
This formula can make use of the path integral formalism but it is very
complicated. It requires that we know the wave function of the vacuum and
that we are able to convolute it with the result fpr a path integral. We
will be able to describe vacuum expectation values of time-ordered operators
without convoluting the path integral, using a nice trick which is easy to
implement.
27
Consider a Hamiltonian

H with eigenstates |n,

H|n = E
n
|n (in the Scrhodinger picture)
and a general state
| =

n
c
n
|n .
Taking a Heisenberg photograph of the state in the very past, we nd:
|, t = e
i

H(t)

n
c
n
|n . (110)
Now, we play a mathematical game; we give a very small imaginary part to
the Hamiltonian

H

H(1 i), ( 0),
which has the same energy eigenstates as the physical Hamiltonian. The
state will evolve as
|, t = e
i

H(t)(1i)

n
c
n
|n
=

n
c
n
e
(+i)Ent
|n . (111)
For a very long time t in the past the exponential e
Ent
vanishes; it vanishes
faster for larger energy eigenvalues. For the vacuum, (E
0
= 0), there is no
such suppression.
|, t = e
t(i+)E
0
_
c
0
|0 +

n=1
e
(i+)(EnE
0
)t
c
n
|n
_
e
t(i+)E
0
c
0
|0 = c
0
|0, t . (112)
This is a very convenient property. For vacuum to vauum transitions, with
any kind of events occuring during them, we do not have to distinguish
the vacuum state from the rest. Instead, we can consider an arbitrary state
of the sytem and include in the evolution a very small imaginary part for the
Hamiltonian. Excited states will not contribute to the transition amplitude
since the friction will lter them out after some time.
You might wonder what happens with the , t| states in the future.
, +t| =

n
c

n
n| e
iHt(1i)
c

0
0| e
iE
0
t(1i)
c

0
0, t| , t +. (113)
28
Therefore
0, t| T(. . .) |0, t =
1
c

0
c

0
, t| T(. . .) |

, t
=
, t| T(. . .) |

, t
| 0 0|

t , (114)
as long as we set H H(1 i) in the rhs. We can choose as states two
position states; one in the very past and one in the very future. Then we
immediately write the expectation values of operators in the vacuum state
as a path integral
0, t| T(. . .) |0, t =
x
2
, t| T(. . .) |x
1
, t
x
2
| 0 0| x
1

. (115)
The wave function of the vacuum still appears in this expression, but only as
a prefactor. We will see that the overall normalization of the path integral
is not important for the physics questions that we are interested in.
2.5 The simple harmonic oscillator
We have studied properties of path integrals at a theoretical level. We will
now put them at test and see whether we can compute something useful with
them. A very instructive example is the simple harmonic oscillator, with a
Hamiltonian
H =
p
2
2m
+
1
2
m
2
x
2
. (116)
The Lagrangian of the system is
L = p x H
p x H(1 i)
m x
2

m x
2
2
(1 i) +
1
2
m
2
x
2
(1 i)

1
2
m(1 + i) x
2

1
2
m
2
(1 i)x
2
, (117)
where we see that we can account for the correct prescription to select the
vacuum as asymptotic states by giving a small positive imaginary part to m
and small negative imaginary part to m
2
m m(1 + i),
m
2
m
2
(1 i).
29
From our earlier discussions we conclude that expectation values of time
ordered operators in the ground state can be given directly from the path
integral adding source terms and having an appropriate i prescription. We
then need to compute a generating functional integral
W[J] =
_
Dxe
i
_
dt(L+J(t)x(t))
,
=
_
Dxe
i
_
dt[
1
2
m(1+i) x
2

1
2
m
2
(1i)x
2
+J(t)x(t)]
. (118)
Things become easier if we perform a Fourier transformation of all quantities
which depend on time:
x(t) =
_

dE
2
e
iEt
x(E), (119)
J(t) =
_

dE
2
e
iEt

J(E). (120)
(121)
The action integral in the exponent of the generating functional becomes
S[x] =
_
dt
_
1
2
m(1 + i) x
2

1
2
m(1 i)
2
x
2
+ J(t)x(t)
_
=
1
4
2
_
dEdE

dt
1
2
e
it(E+E

__
mEE

(1 + i) m
2
(1 i)
_
x(E) x(E

)
+

J(E) x(E

) + x(E)

J(E

)
_
=
_
dE
4
_
m
_
E
2

2
+ i
_
E
2
+
2
__
x(E) x(E)
+

J(E) x(E) + x(E)

J(E)
_
=
_
dE
4
_
m
_
E
2

2
+ i
_
x(E) x(E) +

J(E) x(E) + x(E)

J(E)
_
=
_
dE
4
__
x(E) +

J(E)
m[E
2

2
+ i]
_
m
_
E
2

2
+ i
_
_
x(E) +

J(E)
m[E
2

2
+ i]
_

J(E)

J(E)
m[E
2

2
+ i]
_
(122)
In the above we redened (E
2
+
2
) . We can also dene
y(E) = x(E) +

J(E)
m[E
2

2
+ i]
(123)
30
The second term of the rhs corresponds to the Fourier transform of the
solution of the Euler-Lagrange equations; i.e. it corresponds to the classical
path. We can verify it easily. The classical equation of motion for the
harmonic oscillator with a source term and a small negative imaginary part
for the frequency is
m
_
d
2
x
cl
dt
2
+
_

2
i
_
x
cl
_
= J(t), (124)
and taking the Fourier transform
x
cl
=
_

dE
2
e
iEt
x
cl
,
we obtain
m
_
E
2

2
+ i
_
x
cl
(E) =

J(E)
; x
cl
(E) =

J(E)
m[E
2

2
+ i]
. (125)
Then the path integral over paths x(t), under a shift
x(t) = x
cl
(t) + y(t), (126)
becomes
W[J] = e

1
2
_
dE
2

J(E)i

J(E)
m[E
2

2
+i]
_
Dye
i
2
_
dE
2
y(E)m[E
2

2
+i] y(E)
(127)
The dependece of the path integral on the source terms becomes a simple
prefactor. It is easy to see that the new path integral corresponds to the
harmonic oscillator Hamiltonian with no sources. We can therefore write the
above expression as,
x
f
, t
f
| x
i
, t
i

J
= e
i
2
_
dE
2

J(E)(1)

J(E)
m[E
2

2
+i]
x
f
, t
f
| x
i
, t
i
. (128)
This is a remarkable result since vacuum expectation values of operators are
functional derivatives of the above expression, and the only path integral
which we seem to require is the one for the source-free transition and enters
as a normalization factor. Since we need the derivatives with respect to the
31
time-dependent sources we introduce the inverse Fourier transformation on
the exponential. We write
_

dE
2

J(E)

J(E)
E
2

2
+ i
=
_

dE
2

_
dte
iEt
J(t)
_ _
dt

e
iEt

J(t

)
_
E
2

2
+ i
=
_
dtdt

J(t)G(t, t

)J(t

), (129)
with
G(t, t

) =
_
+

dE
2
e
i(tt

)E
E
2

2
+ i
(130)
We will evaluate this integral later. The expression for W[J] now becomes
x
f
, t
f
| x
i
, t
i

J
= e
i
2
_
dtdt

J(t)G(t,t

)J(t

)
x
f
, t
f
| x
i
, t
i
. (131)
We can now dierentiate this expression with respect to the sources as many
times as we need. We then obtain:
x
f
, t
f
| T x(t
n
) . . . x(t
1
) |x
i
, t
i
=
_
1
i
n

n
J(t
1
) . . . J(t
n
)
e
i
2
_
dtdt

J(t)G(t,t

)J(t

)
_
x
f
, t
f
| x
i
, t
i
. (132)
The i prescription will select the ground state on both sides of the equation
as asymptotic states (the normalization factors are also the same on both
sides of the equation). We therefore have for the vacuum expectation value
of time ordered position operators:
0, t| T x(t
n
) . . . x(t
1
) |0, t =
_
1
i
n

n
J(t
1
) . . . J(t
n
)
e
i
2
_
dtdt

J(t)G(t,t

)J(t

)
_
0, t| 0, t . (133)
The vacuum state does not evolve with time,
0, t| 0, t = 0| e
i2tE
0
|0 = 0| e
0
|0 = 0| 0 = 1.
We can therefore write
0, t| T x(t
n
) . . . x(t
1
) |0, t =
_
1
i
n

n
J(t
1
) . . . J(t
n
)
e
i
2
_
dtdt

J(t)G(t,t

)J(t

)
_
(134)
Exercise:
32
a) Show that
0, t| T x(t
n
) . . . x(t
1
) |0, t = 0,
for n odd.
b) Compute 0, t| T x(t
2
) x(t
1
) |0, t and 0, t| T x(t
4
) . . . x(t
1
) |0, t in terms
of G(t, t

).
c) Find a general expression for 0, t| T x(t
2n
) . . . x(t
1
) |0, t in terms of
G(t, t

).
We now compute the propagator G(t
1
, t
2
) explicitly.
G(t
1
, t
2
) = =
_
dE
2
e
i(t
2
t
1
)E
E
2

2
+ i
=
_
dE
2
e
i(t
2
t
1
)E
(E + i) (E + i)
=
_
dE
2
e
i(t
2
t
1
)E
2
_
1
E + i

1
E + i
_
. (135)
We can compute this integral using the residue theorem. If t
2
> t
1
we can
close the contour of integration on the lower complex half plane where the
exponential vanishes. If t
1
< t
2
we can only close the contour on the upper
complex half-plane. The result is:
G(t
1
, t
2
) =
i
2
e
i|t
2
t
1
|
. (136)
33
3 Path integrals and scalar elds
We will apply the quantization formalism of path integrals in eld theory.
The canonical quantization of the quantum mechanical system was an inte-
gral over all paths.
[ x, p] = i ;
_
Dx(t)e
i
_
dtL
. (137)
In quantum eld theory, the canonical quantization procedure postulates
commutation relations to scalar elds
[(x, t), (y, t)] = i
3
(x y). (138)
We therefore anticipate that applying a similar program as in quantum me-
chanics will lead to an integral over all eld functions (x). We discretize
space-time in small cells V
n
, and the eld may take a value
n
in each cell.
We can also dene discretized derivatives of the eld comparing its value in
neighbouring cells. As in quantum mechanics, when passing from commu-
tation relations to an innite number of integrals we will need to sum over
all possible eld values in each cell. Taking the innite limit yields a new
integral over eld functions; in other words, a path integral over elds.
We will not attempt to derive a path integral formulation starting with
canonical quantization. Instead we will dene quantum eld theory in terms
of path integrals from the beginning. We will test the equivalence of the
two approaches in some cases by verifying that vacuum expectation values
or equivalently Feynman rules are the same in the path integral formalism
and canonical quantization (which you have already seen in QFT I).
In analogy to quantum mechanics we postulate a generating functional
for elds
Z [J] = N
_
De
i
_
d
4
x(L+J+i
2
)
(139)
where
L =
_
d
3
xL,
and the i
2
is in order to dissipate the contributions from expectation val-
ues in between states other than the vacuum. (Recall that in the harmonic
oscillator the H H(1 i) substitution yielded equivalent Greens function
with adding a quadratic imaginary part in the potential) We start with the
Lagrangian for a free-scalar eld. We will see that the results are as simple as
34
in the case of the harmonic oscillator in quantum mechanics. The Lagrangian
density is
L =
1
2
_

m
2

2
_
. (140)
The action integral in the exponent is:
S =
_
d
4
x
_
1
2
_

m
2

2
_
+ i
2
+ J
_
=
_
d
4
x
_
1
2
_

)
2
m
2

2
_
+ i
2
+ J
_
=
_
d
4
x
_
1
2

2
+ m
2
i
_
J
_
(141)
The path integral for this action
Z [J] = N
_
De
iS
,
is a straightforward generalization in four dimensions of the path-integral of
the simple harmonic oscillator. We can then compute it repeating faithfully
the steps of the previous lecture. An important step in this computation is
to shift the eld by the solution of the classical equations of motion.
+
classical
We then nd a path integral free of source terms:
Z [J] = e
i
2
_
d
4
xd
4
yJ(x)
F
(x,y)J(y)

N
_
De

i
2
_
d
4
x(
2
+m
2
i)
. (142)
with the propagator

F
(x, y) =
_
d
4
k
(2)
4
e
ik(xy)
k
2
m
2
+ i
. (143)
Exercise:
a) Write the Euler-Lagrange equations for the free real scalar eld.
b) Evaluate the generating path integral for the free real scalar eld working
in Fourier space and following the analogous derivation of the simple
harmonic oscillator.
35
c) Find the Fourier representation of the propagator.
d) Integrate the Fourier representation of the propagator over the energy
using the Cauchy theorem. Pay attention to the conditions on the time
variable in order to be able to use a closed contour of integration.
The path integral for the free scalar eld is no more dicult than for the
harmonic oscillator. We can act with functional derivatives with respect to
the sources on the above expressions. We then derive expressions which, in
analogy to Quantum Mechanics, are, likely, vacuum expectation values of
eld operators. From here, we can easily develop a method for performing
perturbative calculations and derive Feynman rules. Before doing so, we will
develop some additional mathematical tools.
3.1 Functional Integration
We start with the familiar integral
_

dxe
ax
2
=
_

a
, a > 0. (144)
We then obtain
_

dx
1
. . . dx
n
e

n
i=1
a
i
x
2
i
=

n
2
(a
1
. . . a
n
)
1
2
, a
i
> 0. (145)
We can dene
A = diag (a
1
, a
2
, . . . , a
n
) , (146)
x
T
= (x
1
, . . . , x
n
). (147)
We then rewrite the above integral as
_

i
dx
i

_
e
x
T
Ax
=
1
(detA)
1
2
. (148)
Let us now peform a transformation
x
i
= R
ij
y
j
, or x = Ry. (149)
36
The integral can be written as
_

_
n

i=1
dy
i

_
(detR) e
y
T
(R
T
AR)y
=
1
(detA)
1
2
. (150)
We dene the matrix
B = R
T
AR. (151)
We can easily verify that B is symmetric.
B
T
= (R
T
AR)
T
= R
T
A
T
_
R
T
_
T
= R
T
AR = B.
The determinant of B is then
detB = det(R
T
AR) = (detA)(detR)
2
.
We then nd
_

_
n

i=1
dy
i

2
_
e

1
2
y
T
By
=
1
(detB)
1
2
, (152)
where B is any real, positive denite (positive eigen-values), symmetric ma-
trix.
We can add a linear term (source) in the exponent.
_

_
n

i=1
dy
i

2
_
e

1
2
y
T
By+y
T
J
, (153)
The minimum (classical eld) of the exponent (action) is at
y
0
= B
1
J. (154)
Substituting
y = y
0
+ x,
we nd
_

_
n

i=1
dy
i

2
_
e

1
2
y
T
By+y
T
J
=
_

_
n

i=1
dx
i

2
_
e

1
2
x
T
Bx+
1
2
J
T
B
1
J
=
e
1
2
J
T
B
1
J
(detB)
1
2
(155)
37
Let us naly take the limit n . The exponent in the rhs is originally a
double sum over the two indices of the n n matrix B. In the innite limit,
sums turn into integrals and the inverse matrix turns into an integration
kernel which is the inverse of an operator. We can then write:
_
De

_
d
4
x[
1
2
(x)

A(x)(x)J(x)(x)]
=
e
1
2
_
d
4
xd
4
yJ(x)(x,y)J(y)
_
det

A
. (156)

A may be even a dierential operator which has an inverse, as long as the

path integral is convergent (the i prescription seems to achieve this). For
example, for the free real scalar eld we nd:
Z[J] = N
_
De
_
d
4
x{
1
2
i[
2
+m
2
i]+iJ}
= N
e

i
2
_
d
4
xd
4
yJ(x)
F
(x,y)J(y)
_
det [i (
2
+ m
2
i)]
(157)
It is true that we have not given a precise meaning to the determinant of this
operator. Al of the above is a language to describe quickly the separation of
the sources from the path-integral when an action is quadratic in the elds.
However, we do not need to give a precise denition of the determinant. The
generating fuctional should have a normalization such that for zero sources
it describes a vacuum to vacuum transition; the corresponding amplitude is
1. This xes the normalization to be
N =
_
det [i (
2
+ m
2
i)].
We can thus consider the identity
_
De

1
2
_
(x)

A(x)(x)
=
1
_
det

A
, (158)
as the denition of the determinant.
To conclude, we must identify the

F
(x) =
_

2
x
+ m
2
i
_
1
, (159)
or, more precisely,

2
+ m
2
i
_

F
(x) =
4
(x) (160)
38
Writing the Fourier transform of it

F
(x) =
_
d
4
k
(2)
4
e
ikx

F
(k) (161)
and substituting in Eq. 160 we nd

F
(x) =
_
d
4
k
(2)
4
e
ikx
k
2
m
2
+ i
. (162)
We can now dierentiate the generating functional Z[J] with respect to
the sources twice. We nd:

F
(x, y) =

2
Z[J]
J(x)J(y)

J=0
. (163)
We can now obtain a rst evidence that the path integral formulation yields
the same Greens functions as the canonical quantization formalism. You
can verify that indeed we also obtain
0, T| T(x)(y) |0, T =
F
(x, y). (164)
Exercise: Prove the above using canonical quantization. You will need the
Fourier integral of the propagator after you inegrate out the energy.
It is useful to know one more technical denition. We can dene the
determinant of a Hermitian operator via path integration over complex elds.
Consider

a
=
_

dxdye
a(x
2
+y
2
)
=
_

dxdye
a(xiy)(x+iy)
. (165)
We dene
x =
z + z

2
, y =
z z

2i
We obtain
_
dzdz

2i
e
azz

=
1
a
. (166)
Following the same steps as before, we can write the determinant of a Her-
mitian operator as:
_
DD

_
d
4
x

(x)

A(x)(x)
=
1
detA
. (167)
39
3.2 Path integrals and interacting elds
The only computation needed for a free scalar eld in the path integral
formalism is the evaluation of the propagator
F
. The generating functional
requires this propagator as a kernel and it yields Greens functions by taking
functional derivatives with respect to the sources. It is however an extremely
dicult or even impossible problem to compute exactly the path integral for
a Lagrangian with interacting elds. However the formalism can be easily
used to perform calculations of Greens functions for interacting elds using
perturbation theory.
Let us consider a real scalar eld which can interact with itself,
L = L
0
+L
I
, (168)
with
L
0
=
1
2
(x)
_

2
+ m
2
i
_
(x), (169)
and
L
I
=

4!
(x)
4
. (170)
Vacuum expectation values of operators will then be given via functional
derivatives of
Z[J] = N
_
De
i
_
d
4
{L
0
+(x)J(x)+L
I
}
. (171)
We require that the vacuum to vacuum transition has a unit amplitude, so
we can x
1
the normalization N:
Z[0] = 1.
We then write:
Z[J] =
_
De
i
_
d
4
{L
0
+(x)J(x)+L
I
}
_
De
i
_
d
4
{L
0
++L
I
}
. (172)
We consider the case with 1. We separate the action into a part cor-
responding to the free scalar particle with source terms and the interaction
part
S = S
0
+ S
I
, (173)
1
We have assumed that if necessary a constant is subtracted from the Hamiltonian,
so that the ground energy is zero; in canonical quantization this can be achieved with a
normal ordering of the operators.
40
with
S
0
=
_
d
4
L
0
+ (x)J(x), (174)
and
S
I
=
_
d
4
xL
I
. (175)
We can perform a perturbative expansion in lambda,
_
De
iS
0
+iS
I
=
_
D

n=0
(iS
I
)
n
n!
e
iS
0
_
D

n=0
_
i
_
d
4
y
_

4!

4
__
n
n!
e
iS
0

n=0
_
i
_
d
4
y
_

4!
_

iJ(y)
_
4
__
n
n!
_
De
iS
0
. (176)
In the last step we used the fact that we can produce integrals with poly-
nomials on multiplying the exponential of the action in a path integral
by acting with functional derivatives. The result that we have derived is an
all-orders result. We can undo the expansion and cast the sum back into
an exponential. We obtain for the generating functional
Z[J] =
e
i
_
d
4
xL
I(i

J(x)
)
_
De
iS
0
e
i
_
d
4
xL
I(i

J(x)
)
_
De
iS
0

J=0
. (177)
We can use the result for the generating path integral for the free els, which
we computed earlier.
Z
0
[J] = N
_
De
i
_
d
4
x(L+J(x)(x))
= e
i
2
_
d
4
xd
4
yJ(x)
F
(x,y)J(y)
. (178)
The generating fucntional for the interacting theory is related to the gener-
ating functional for the free theory via,
Z[J] =
e
i
_
d
4
xL
I(i

J(x)
)
Z
0
[J]
e
i
_
d
4
xL
I(i

J(x)
)
Z
0
[J]

J=0
. (179)
41
3.3 Perturbation theory for
4
/4! interactions
Let us consider the numerator of the generating functional Z[J],
Num[J] = e
i
_
d
4
xL
I(i

J(x)
)
Z
0
[J]
=
_
1

4!
_
d
4
z

4
J(z)
4
_
e
i
2
_
d
4
xd
y
J(x)(xy)J(y)
+O(
2
).(180)
We introduce Feynman diagram notation dening:
y x
(x y),
y

_
d
4
xJ(x)(x y),
and

_
d
4
xd
4
yJ(x)(x y)J(y).
With this notation, we write:
Num[J] =
_
1

4!
_
d
4
z

4
J(z)
4
_
e
i
2
+O(
2
). (181)
We need to dierentiate the exponential four times. It is easy to see that

z
=
z
+
z
= 2
z
,
and

w
z
=
z w
.
We nd,
e
i
2
z
= i
z
e
i
2
, (182)
42

2
e
i
2
J(z)
2
=
_

_
i
z

z
_

_
e
i
2
, (183)
and so on. We can then compute the generating function, and you will nd
that:
Z[J] =
Num[J]
Num[0]
=
_
1

4!
_
6 +i
_
_
e
i
2
+O(
2
) .
(184)
Notice that in the above expression, all diagrams are connected to a source.
This is due to the normalization of Z[0] = 1. The denominator, after ex-
panding in , removes all vacuum graphs. We can now compute Greens
functions using this expression. It is useful to cast the generating functional
as an exponential. We can write,
Z[J] = e
iW[J]
, (185)
with
W[j] =
1
2
+
i
2


4!
+O(
2
). (186)
Interestingly, only W[J] generates Greens functions which contribute to the
S-matrix. We nd:
1
i
Z[J]
J(x1)

J=0
=
W[J]
J(x1)

J=0
= 0| T(x
1
) |0 ,
1
i
2

2
Z[J]
J(x1)J(x2)

J=0
=
1
i

2
W[J]
J(x1)J(x2)

J=0
+0| T(x
1
) |0 0| T(x
2
) |0 ,
1
i
3

3
Z[J]
J(x1)J(x2)J(x3)

J=0
=
1
i
2

3
W[J]
J(x1)J(x2)J(x3)

J=0
+0| T(x
1
)(x
2
) |0 0| T(x
3
) |0
+0| T(x
1
)(x
3
) |0 0| T(x
3
) |0
+0| T(x
3
)(x
2
) |0 0| T(x
1
) |0
2 0| T(x
1
) |0 0| T(x
2
) |0 0| T(x
3
) |0 ,
43
and so on. The diagrams which contribute to a scattering are the ones where
all derivatives act on a single W. The remaining terms produce disconnected
diagrams where subsets of particles scatter independently.
44
3.4 Fermionic path integrals
We have discussed the path integral for scalar elds. Bosons, in canonical
quantization, have a non-zero commutator. In the path integral language this
gave rise to integrals over exponentials with commuting functions, and not
operators, at the exponent. For fermions, the spin-statistics cannot be ac-
counted for by using our known commuting functions in the exponent of the
path integral. We can formulate path integration over anti-commuting func-
tions (complex-numbers) which can describe fermions. We will use Grass-
mann numbers, which are dened to obey:
{c
i
, c
j
} = c
i
c
j
+ c
j
c
i
= 0. (187)
This dention means that any function constructed out of Grassmann vari-
ables is at most linear in any of them, since:
c
2
i
= 0.
Let us consider a concrete example of a function of two Grassmann variables
c
1
, c
2
. Its general form is:
f (c
1
, c
2
) = a
0
+ a
1
c
1
+ a
2
c
2
+ a
12
c
1
c
2
, (188)
where the coecients a
0
, a
1
, a
2
, a
12
are normal commuting numbers. We will
now develop our calculus for functions of Grassmann variables. We start by
dening the derivative:
c
i
c
j
=
ij
. (189)
However, rearrangements of the Grassmann variables give rise to minus signs.
For example, we can decide to write the c
2
variable at the left of any other
Grassmann variable. The derivative with respect to c
2
is then

L
f
c
2
=

L
c
2
(a
0
+ a
1
c
1
+ a
2
c
2
+ a
12
c
1
c
2
)
=

L
c
2
(a
0
+ a
1
c
1
+ a
2
c
2
a
12
c
2
c
1
)
= a
2
a
12
c
1
. (190)
This is dened to be a left derivative (as the superscript denotes). We
could have dened a right derivative, where we always anti-commute a
45
Grassmann variable to the right of any other variable before we dierentiate
with it. In our example, we have:

R
f
c
2
= a
2
+ a
12
c
1
. (191)
Exercise: Prove that
_
C
i

C
j
_
=
ij
, (192)
and
_

C
j
,

C
j
_
= 0. (193)
We will always use left derivatives, unless it is explicilty stated otherwise.
Our next step is to dene an integration over Grassmann variables c
i
. We
only require to know two integrals
_
dc 1 =?
and _
dc c =?
since quadratic and higher terms in c vanish. We have
__
dc
_
2
=
__
dc
1
___
dc
2
_
=
__
dc
2
___
dc
1
_
=
__
dc
_
2
.
Therefore,
_
dc 1 = 0. (194)
The only remaining integral cannot be specied from the algebra,
_
dc c = x.
and it is an arbitrary (but xed) commuting number x. We can always divide
the integration measure by x. It is then convenient to dene:
_
dc c = 1. (195)
We should be careful in integrating multiple variables, since
_
dc
1
c
2
c
1
=
_
dc
1
c
1
c
2
= c
2
.
46
We can make a striking observation: dierentiation and integration are
identical operations.
_
dc
2
f =
_
dc
2
(a
0
+ a
1
c
1
+ a
2
c
2
+ a
12
c
1
c
2
)
=
_
dc
2
(a
0
+ a
1
c
1
+ a
2
c
2
a
12
c
2
c
1
)
= a
2
a
12
c
1
=
f
c
2
. (196)
We must also investigate how we can perform a change of variables in inte-
grals over Grassmann variables. We consider with the integral
_
dc
1
dc
2
. . . dc
n
f(c
1
, c
2
, . . . , c
n
),
and perform a transformation (which is always linear)
c
i
= M
ij
b
j
. (197)
We shall have
_
dc
1
dc
2
. . . dc
n
f(c
i
) = (Jacobian)
_
db
1
db
2
. . . db
n
f (M
ij
b
j
) . (198)
What is the Jacobian? In the case of a double integral, we have:
_
dc
1
dc
2
c
1
c
2
= (Jacobian)
_
db
1
db
2
(M
11
b
1
+ M
12
b
2
) (M
21
b
1
+ M
22
b
2
)
= (Jacobian)
_
db
1
db
2
(M
11
M
22
b
1
b
2
+ M
12
M
21
b
2
b
1
)
= (Jacobian)
_
db
1
db
2
b
1
b
2
(M
11
M
22
M
12
M
21
)
= (Jacobian)det(M)
_
db
1
db
2
b
1
b
2
;Jacobian =
1
det(M)
. (199)
Exercise: Prove that the same results holds for multiple integrals of arbitrary
dimensions.
If we recall that Grassmann integration is in reality dierentiation, it is not
surprising that the Jacobian of the transformation on Grassmann variables is
the inverse of what emerges in integrations of normal commuting variables.
47
After we have dened integration over Grassmann variables we can study
multiple exponential integrals which, in analogy with the bosonic case, could
be used to dene a fermionic path-integral. Let us consider two independent
vectors of Grassmann variables x
T
= (x
1
, x
2
) and y
T
= (y
1
, y
2
). We have
x
T
y = x
1
y
1
+ x
2
y
2
, (200)
and
_
x
T
y
_
2
= (x
1
y
1
+ x
2
y
2
)(x
1
y
1
+ x
2
y
2
)
= x
1
y
1
x
2
y
2
+ x
2
y
2
x
1
y
1
= 2x
1
y
1
x
2
y
2
. (201)
We can easily see that
_
x
T
y
_
n
= 0, n > 2. (202)
Therefore, using a Taylor expansion, we have
_
dx
1
dy
1
dx
2
dy
2
e
x
T
y
=
_
dx
1
dy
1
dx
2
dy
2
[1 (x
1
y
1
+ x
2
y
2
) + x
1
y
1
x
2
y
2
]
= 1. (203)
We can now perform linear transformations on both x and y.
x Mx

,
y My

.
We nd,
1 =
_
dxdye
x
T
y
= det(M
T
)
1
det(N)
1
_
dxdye
x
T
M
T
Ny
= det(M
T
N)
1
_
dxdye
x
T
M
T
Ny
. (204)
Dening A = M
T
N, we obtain that
_
dxdye
x
T
Ay
= det(A). (205)
Recall that for normal commuting variables we have
_
dxdx

e
x

Ax

1
detA
. (206)
48
Let us now dene a path integral over Grassmann variables by considering an
innite number of them and taking the continuous limit. This path integral
will quantize the eld psi as in the scalar eld case. It will dier however from
the Greens functions of the bosonic eld. Let us consider the Lagrangian of
a free Dirac fermion.
L =

(i m) . (207)
We cab write a generating functional
Z
0
[a, a] = N
_
DD

e
i
_
d
4
x[L+ a+

a]
. (208)
In this path integral the integration variables ,

and the sources a, a are
all independent Grassmann functions. The constant N is xed as usual by
requiring that
Z
0
[0, 0] = 0| 0 = 1.
Everything works in an analogous way as in the case of the path integral
for the scalar eld. For a theory without interactions, we can complete the
square and compute the generating functional explicitly if we shift he elds
,

by a constant corresponding to their classical value which minimizes the
action. We need the inverse of the Dirac wave operator,
(i m1)S(x y) =
4
(x y) 1. (209)
As for the scalar eld and the harmonic oscillator we can write a Fourier
representation,
S(x y) =
_
d
4
k
(2)
4
k + m
k
2
m
2
e
ik(xy)
(210)
The values of ,

which minimize the action are now given by

cl
=
_
d
4
yS(x y)a(y),

cl
=
_
d
4
y a(y)S(x y),
and the change of variables
=
cl
+ ,

=

cl
+

eta,
49
yields
Z
0
[a, a] = e
i
_
d
x
d
y
a(x)S(xy)a(y)
. (211)
We can also develop a perturbative expansion repeating faithfully the steps
we performed in the
4
scalar eld theory. It is not hard to convince ourselves
that a completely analogous formula should be valid here for the generating
functional when interactions are present in the Lagrangian:
Z [a, a] = e
i
_
d
z
L(
1
i

a
,
1
i

a
,)Z
0
[a, a].
(212)
Besides the main similarities in the appearance of the formulae there are also
very important dierences which are encoded in the Grassmann algebra of
the functions which we integrate upon. What is very important to remember,
is the fact that functional derivatives anticommute, generalizing the result
that we found for the derivatives of discrete Grassmann variables:
_

a(x)
,

a(x)
_
=
_

a(x)
,

a(x)
_
=
_

a(x)
,

a(x)
_
= 0. (213)
We should also remember that these derivatives are left (by convention)
derivatives, and the order in which sources appear in the integrand matters
indeed.
50
4 Quantization of non-abelian gauge theories
We now have a sucient formalism to quantize a non-abelian gauge theory.
The classical Yang-Mills Lagrangian is,
L
Y M
=
1
4
G
a
G
a

. (214)
with
G
a

A
a

A
a

+ gf
abc
A
b

A
c

. (215)
Proceeding in analogy with the quantization of the scalar eld theory, we
write the following path integral:
Z[J

a
] = N
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)+J

a
Aa]
(216)
We would like to develop a perturbation theory program. Remember what
we required in the

4
4!
case. We could try to repeat the same steps here:
- nd the propagator of the free-eld by inverting the dierential oper-
ator in the quadratic part of the Lagrangian. This would give us an
expression for the path integral when all interactions are switched o
(g = 0)
Z
0
[J

a
] = N

e
i
_
d
4
xd
4
yJ

a
(x)
ab
(xy)J

b
(x)
. (217)
- derive the perturbative expansion from
Z[J] e
i
_
d
4
zL
int
_
1
i

a
(z)
_
Z
0
[J

a
]. (218)
The free-eld action (g = 0) is
S
free
=
1
4
_
d
4
x
_

A
a

A
a

_
(

A
a

A
a
)
=
1
2
_
d
4
x
_
(

A
a

) (

A
a
)
_

A
a

_
(

A
a
)
_
=
1
2
_
d
4
x
_

(A
a

A
a
) A
a

A
a

_
A
a

A
a
_
+ A
a

A
a
_
=
1
2
_
d
4
xA
a

ab
_

2
g

_
A
b
. (219)
51
We now need to nd the inverse of the operator
_

2
g

ab
.
However, it turns out that there is none! This operator has zero eigenvalues,
and its determinant is zero. In particular we obtain zero when it acts on any
function that can be written as a total derivative:
_

2
g

(x) =
2

(x) +
2

(x) = 0. (220)
Our naive attempt to establish a perturbation expansion in g using a path
integral formalism has failed at the rst step. However, there is a property of
the theory, gauge invariance, which we have not yet used and we can exploit
it to remove the zero-modes of the operator in the free part of the Lagrangian.
Let us start by dening a -functional, in analogy to a -function, which
we will need in a while. The integral over a -function is
_
df(f) = 1.
We can change variables, f = f(w), and we obtain
2
:
_
dw
f
w
(f(w)) = 1.
The multidimensional generalization of this equation is:
1 =
_
dw
1
. . . dw
1
det
_
f
i
w
j
_
(f
1
(w
1
, . . . , w
n
)) . . . (f
n
(w
1
, . . . , w
n
)).
We can take the limit n which yields a functional integral over w(x),
where x is the continuous variable corresponding to the index i = 1 . . . n. We
dene the innte product of delta functions as a delta functional. We write:
_
Dwdet
_
f(x)
w(y)
_
[f(w)]
lim
n
_
dw
1
. . . dw
1
det
_
f
i
w
j
_
(f
1
(w
1
, . . . , w
n
)) . . . (f
n
(w
1
, . . . , w
n
))
= 1. (221)
2
rigor is on holidays concerning absolute values, signs, and the support of the delta-
functions...
52
Notice the emergence of a functional determinant, due to changing variables
in the measure of a functional integral.
We now return to the gauge-theory; the action is of course invariant under
gauge transformations
A

= U(x)A

(x) +
i
g
U(x)
_

(x)
_
, (222)
with A

A
a

T
a
. The transformation matrices U are determined by as many
independent parameters as the generators of the Lie group,
U(x) = e
i
a
(x)T
a
= 1 i
a
(x)T
a
+O(
2
). (223)
The integration
_
DA
a

in the path integral formalism does not discriminate

among the elds which are connected via a gauge transformation.
Consider all the elds A
a(independent)

which cannot be connected via a

gauge transformation. We never know them explicitly, but we can impose
that they satisfy gauge-xing conditions of the form
F
1
(A
a

) = 0,
which remove the superuous degrees of freedom. For example, a very com-
mon choice is a Lorentz gauge xing condition,

A
a
= 0.
Of course, we are allowed to choose other conditions to x the gauge,
F
2
(A
a

) = 0,
The gauge elds A

a
(2) and A

a
(1) satisfying the two dierent gauge condi-
tions are related via gauge transformations; for each of the solutions of the
second gauge-xing condition there is a unique set of gauge transofrmation
parametetes
a
which maps it to a solution of the rst gauge-xing condition.
Therefore, if we consider all innite possibilities for gauge-xing conditions,
we enumerate all innite members of the Lie algebra parameters
a
.
Let us now go back to the path integral for the gauge theory without
sources:
Z = N
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
(224)
53
This integrates over all elds including the ones related by gauge transfor-
mations. In other words, had we thought of all possible gauge xings, it
integrates over all these possibilities. We can write:
Z = N
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
1
= N
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]

_
DF
a
[F
a
(A
a
)], (225)
where
F
a
(A
a
) = 0 ;A
a
= A
a
(
b
).
Dierent gauge xings F
a
correspond to dierent group parameters, so we
may change variables F
a

a
in the second functional integration.
Z = N
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
_
D
b
det
_
F
a
(A
a
)

b
_
[F
a
(A
a
(
b
))]
= N
_
D
b
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
F
a
(A
a

b
_

F
a
(A
a
(
b
))=0
[F
a
(A
a
(
b
))]
(226)
Now there is a crucial observation to be made. No term in the inner functional
integral depends on
b
. Let us justify this statement. We have,
1 =
_
DF
a
[F
a
(A
a
)]
;
1
det
_
F
a
(A
a
)

b
_

F
a
(A
a
(
b
))=0
=
_
D
b
(F
a
(A
a
(
b
))) (227)
The rhs is gauge invariant. We have
U = e
i
a
T
a
;
a
T
a
= lnU ;
a
= 2Tr(T
a
ln U).
Under a gauge transformation U

= e
iTb
b
we have

c
=
a
+
b
, (228)
and
D
c
= D
a
det
_

a
_
= D
a
.
54
A gauge transformation only reshues the elements of the Lie algebra; we
can compute the measure over all possible Lie algebra elements in any gauge.
Therefore:
1
det
_
F
a
(A
a
)

b
_

F
a
(A
a
(
b
))=0
=
_
D
b
(F
a
(A
a
(
b
)))
=
_
D
c
(F
a
(A
a
(
c
)))
det
_
F
a

c
_

F
a
(A
a
(
c
))=0
. (229)
This determinant is therefore just a gauge-invariant number. The exponential
of the Yang-Mills action is of course gauge invariant:
e
i
_
d
4
xL
Y M
(A)
= e
i
_
d
4
xL
Y M
(A
a

(
b
))
.
Finally, the gauge-boson elds are also members of the Lie algebra and only
get reshued by gauge transformations. The measure DA
a

can be evaluated
at any gauge.
DA
a

= DA
a

(
b
). (230)
Exercise: Consider a gauge transformation A
a

A
a

. Prove that DA
a

=
DA
a

. You only need to consider an innitesimal gauge transformation.

We then compute all terms in the inner path inegral at the specic-gauge
chosen by the delta-functional.
Z = N
_
D
b
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
F
a
(A
a

b
_

F
a
(A
a
(
b
))=0
[F
a
(A
a
(
b
))]
= N
_
D
b
_
DA
a

(
b
)e
i
_
d
4
x[L
Y M
(A

a
(
b
))]
det
_
F
a
(A
a

(
b
))

b
_
[F
a
(A
a
(
b
))]
= N
__
D
b
__
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
F
a
(A
a

b
_
[F
a
(A
a
)] (231)
In the last line we noticed that the gauge-xed eld variable A
a

(
b
) is a
dummy integration variable. The path integration over all possible gauge
transformations (corresponding to all possible gauge xings) is an overall
normalization factor. This is an innite integration over the measure of
innite Lie algebra parameters. However, this is not a problem if we want to
55
compute physical quantities, since the overall normalization will cancel. We
therefore end up with the gauge-xed path integral
Z = N

__
D
b
__
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
F
a
(A
a

b
_
[F
a
(A
a
)]. (232)
It is very important to notice that (up to an irrelevant innite normalization)
this path integral does not depend on the gauge-xing condition F(A
a
) that
we may choose!
This new-path integral in Eq. 232 integrates over elds which cannot
be related via gauge transformations; it should therefore be ne to derive
Greens functions for elds which are physically distinct. However, the new
expression has two features, the gauge-xing delta-functional and the deter-
minant, which were absent in our formulation of perturbation theory. It is
therefore unclear at rst sight how to establish a calculable perturbative ex-
pansion with the mathematics that we now. Two very clever tricks will come
to our rescue.
Without loss of generality, and given the non-dependence of the path-
integral on the gauge condition, we write
F(A
a
) = G(A
a
) w(x). (233)
We are allowed to multiply the path-integral with an overall constant without
any physical consequences. We then multiply with the factor,
C =
_
Dwe
i
_
d
4
x
w(x)
2
2
. (234)
We can do this because Z in Eq. 232 does not depend on w(x). We have
Z
__
Dwe
i
_
d
4
x
w(x)
a
w(x)
a
2
__
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
G
a
(A
a

b
_
[G
a
(A
a
) w(x)]
=
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)]
det
_
G
a
(A
a

b
_
_
Dwe
i
_
d
4
x
w(x)
2
2a
[G
a
(A
a
) w(x)]
=
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)
1
2
(G
a
(A
a
))
2
]
det
_
G
a
(A
a

b
_
(235)
With this trick, we remove the delta-function from the integrand add modify
the exponent of the path-integral. This is helpful in order to nd a well-
dened propagator for the gauge boson eld. If we choose for example the
gauge-xing condition
G
a
(A
a
) =

A
a
,
56
the free-part (g = 0) of the action in the exponent becomes:
S
free
=
1
2
_
d
4
xA
a

ab
_

2
g

+
_
1
1

_
A
b
. (236)
Now, the new dierential operator has an inverse:

ab

=
ab
_
d
4
k
2
4
e
ikx
k
2
+ i
_
g

(1 )
k

k
2
+ i
_
. (237)
Exercise: Find the inverse of the operator:
_
_

2
+ M
2
_
g

+
_
1
1

.
_
This will be the case of a massive gauge-boson such as W, Z.
Exercise: Find the gauge boson propagator in an axial gauge
G(A) = n

A
a
,
where n is a light-like vector n
2
= 0.
We now need to deal with the determinant in the integrand of the path
integral. Here we will use a result that we found from innite integration
over Grassmann variables. We proved earlier that:
_
dx
1
. . . dx
n
dy
1
. . . dy
n
e
x
T
Ay
= det(A). (238)
where A is an n n matrix and x, y are Grassmann variables. We can take
the limit of n . We then obtain express a functional determinant as a
fermionic path integrals over two independent Grassmann functions x and y:
ig det(A) =
_
DyDxe
i
_
d
4
x
1
d
4
x
2
y(x
1
)(gA(x
1
x
2
))x(x
2
)
. (239)
The Fadeev-Popov idea was to introduce two new elds with odd spin-
statistics (Grassmann variables in the path-integral), a ghost and an anti-
ghost, and write the determinant as functional integral over an exponential.
We write,
Z
_
DA
a

e
i
_
d
4
x[L
Y M
(A

a
)
1
2a
(G
a
(A
a
))
2
]
det
_
ig
G
a
(A
a

b
_

_
DA
a

D
a
D
a
e
i
_
d
4
x[L
Y M
(A

a
)
1
2
(G
a
(A
a
))
2
]

e
i
_
d
4
x
1
d
4
x
2

a
(x
1
)
_
g
G
a
(A
a

(
a
))

b
_

b
(x
2
)
. (240)
57
We need not to worry about computing precisely the overall normalization
of the path-integral. This will drop out when we require that vacuum to
vacuum transitions have a unit amplitude. The factor ig is convenient, in
order to later combine easierly all terms under a common exponential.
Let us consider a local gauge transformation
U(x) = e
i
a
T
a
,
under which a gauge-eld transforms as
A

() = U(x)A

(x) +
i
g
U(x)
_

(x)
_
, (241)
where, as usual, A

= A
a

T
a
. For a small transformation, we obtain
A
,a
() = A
,a

1
g
_

ab
gf
abc
A
,c
_

b
. (242)
We recognize in the above expression the covariant derivative in the adjoint
representation:
D
ab

ab
gf
abc
A
c

, (243)
so we can write:
A
,a
() = A
,a

1
g
D
,ab

b
. (244)
We then have,
A
,a
((x))

b
(y)
=
1
g
(D
,ac
(x)
c
(x))

b
(y)
=
1
g
D
,ab
(y)(x y). (245)
We can now compute the functional derivative
g
G(A
,a
((x)))

b
(y)
= g
_
d
4
z
G(A
,a
((x)))
(A
,c
((z)))
(A
,c
((z)))

b
(y)
=
_
d
4
z
G(A
,a
(x))
(A
,c
(z))
D
,cb
(z)(z y)
=
G(A
,a
(x))
(A
,c
(y))
D
,cb
(y) (246)
We can then write the following expression for the path integral:
Z
_
DA
a

D
a
D
a
e
i
_
d
4
x[L
Y M
(A

a
)
1
2
(G
a
(A
a
))
2
]

e
i
_
d
4
x
1
d
4
x
2

a
(x
1
)
_

G
a
(A
a

(x
1
))
A
c

(x
2
)
D
cb

(x
2
)
_

b
(x
2
)
. (247)
58
This is valid for any arbitrary gauge xing condition G
a
(A
a

).
It will be instructive and practical (this is what we need to do when we
compute elements of the S-matrix) to choose a gauge. A common choice is
the Lorentz gauge:
G
a
(A
a

) =

A
,a
, (248)
where

A
,a
(x
1
)
A
,c
(x
2
)
=

ac
(x
1
x
2
) =

ac
(x
1
x
2
). (249)
Therefore, in the Lorentz gauge, the path integral is (up to a normalization):
Z =
_
DA
a

D
a
D
a
e
i
_
d
4
x[
1
4
G
a

G
a,

1
2
(A
a
)
2

a
(x)D
;ab

b
(x)]
. (250)
After a partial integration, we obtain:
Z =
_
DA
a

D
a
D
a
e
i
_
d
4
x[
1
4
G
a

G
a,

1
2
(A
a
)
2
+(
a
)D
;ab

b
]
. (251)
In this last expression, we have exponentiated all terms which arose from
gauge-xing. The argument of the exponential is a new action, modied by
a gauge-xing term and contributions from the ghost elds.
Notice that the form of the quadratic term in the ghost elds is the same
as for a complex scalar eld. However, the variables , in the path-integral
are anti-commuting Grassmann variables. Therefore, the ghost eld is a
scalar eld with the wrong spin-statistics.
We also observe that for an abelian gauge theory f
abc
= 0 we have
D
ab

=
ab

and there is no interaction term for the ghost eld and the
gauge-boson. In this case, the ghost eld is a free eld and we can integrate
out its contribution to the path integral (changing the irrelevant overall nor-
malization). This is why in QED you never needed to introduce it.
4.1 Feynman rules for QCD
After gauge-xing and the Fadeev-Popov method, we can formulate a path
integral for QCD, where the path-integral action is:
S =
_
d
4
xL,
with
L = L
Y ANGMILLS
+L
FERMION
+L
GAUGEFIXING
+L
FADEEV POPOV
.
59
The classical Yang-Mills Lagrangian is:
L
Y ANGMILLS
=
1
4
G
a

G
a,
,
and
G
a

A
a

A
a

+ gf
abc
A
b

A
c

.
The gauge xing and the Fadeev-Popov terms, in the Lorentz gauge, are:
L
GAUGEFIXING
=
1
2
_

A
a

_
2
.
L
FADEEV POPOV
= (

a
) D
ab

b
,
and the fermion term:
L
FERMION
=

i
_
i

D
ij

m
ij
_

j
.
The covariant derivtives in the adjoint and fundamental representation are
D
ab

ab
gf
abc
A
c

and
D
ij

ij
igT
c
ij
A
c

accordingly.
We would like to compute Greens function using perturbation theory.
If we switch o the coupling, g = 0, then we are left with terms which are
quadratic in the elds, and we can compute the corresponding path integral
for the free action. We dene,
L = L
free
+L
interacting
, (252)
with
L
free
= L|
g=0
.
Explicitly,
L
free
=
1
4
_

A
a

A
a

_
(

A
a,

A
a,
)
1
2
_

A
a

_
(

A
a

)
+(

a
) (

a
)
+

i
(i

m)
i
(253)
60
It is convenient to use integration by parts and cast the free Lagrangian in a
standard form:

1
2
(Field
A
)

O(Field
A
) +

(. . .)
or, for terms with independent elds (appearing separately in the measure
of the path-integral),
(Field
A
)

O(Field
B
) +

(. . .) .
We have
L
free
=
1
2
A
a

K
ab,
A
b

a
K
ab

ij

j
+

(. . .) , (254)
with (
2

)
K
ab,
=
ab
_
g

2
+
_
1
1

_
, (255)
K
ab
=
ab

2
, (256)

ij
=
ij
(mi) . (257)
An essential step in order to compute the generating functional for the free
path-integral is to nd the inverse of the above operators, i.e. objects which
diagonilize them in all indices. We dene:
K
ac

(x)D
cb,
(x y) =
ab
g

4
(x y),
K
ac
(x)D
cb
(x y) =
ab

4
(x y),

ik
(x)S
kj
(x y) =
ij

4
(x y). (258)
Exercise: Solve these equations. The solutions are given next in the text
and it is easy to verify whether it is correct or not. However, it will be
instructive to think how to nd them if nobody told you the answer!
61
We nd that
D
ab

(x) =
ab
_
d
4
k
(2)
4
e
ikx
k
2
+ i
_
g

(1 )
k

k
2
+ i
_
(259)
D
ab
(x) =
ab
_
d
4
k
(2)
4
e
ikx
k
2
+ i
(1) (260)
S
ij
(x) =
ij
_
d
4
k
(2)
4
e
ikx
k
2
m
2
+ i
(1) (k + m) . (261)
The generating functional for the free-Lagrangian is
Z
0
_
J

, J

, J

, J

J
A
_
=
_
DD

DD DA

exp
_
i
_
d
4
x
_

1
2
A
a

K
ab,
A
b

a
K
ab

ij

j
+ J
a,
A
A
a

+ J
i

i
+

i
J
i

+ J
a

a
+
a
J
a

_
_
(262)
where we have included independent sources for all fermion, anti-fermion,
ghost, anti-ghost, and gauge elds. We should keep in mind that only the
source J
a,
A
for the gauge eld is a bosonic (commuting) variable; all other
sources are Grassmann variables. We can now complete squares by shifting
the elds as,
A
a,
(x) A
a,
(x) +
_
d
4
yD
ab

(x y)J
b,
A
(y) (263)

a
(x)
a
(x) +
_
d
4
yD
ab
(x y)J
b

(y) (264)

a
(x)
a
(x) +
_
d
4
yJ
b

(y)D
ba
(x y) (265)

i
(x)
i
(x) +
_
d
4
yS
ij
(x y)J
j

(y) (266)

i
(x)

i
(x) +
_
d
4
yJ
j

(y)S
ji
(x y). (267)
We then obtain (with an undetermined overall constant factor),
Z
0
_
J

, J

, J

, J

J
A
_
= N exp
_
i
_
d
4
xd
4
y
_
1
2
J
a,
A
(x)D
ab

(x y)J
b,
A
(y) + J
a

(x)D
ab
(x y)J
b

(y) + J
i

(x)S
ij
(x y)J
j

(y)
_
_
(268)
62
We will now deal with the interaction Lagrangian; this is dened as
L
interaction
= L L
free
. (269)
We nd
L
interaction
= g

i
T
a
ij
A
a

j
g (

a
) f
abc
A
a,

b
2gf
abc
A
b,
A
c,
_

A
a

A
a

g
2
4
f
abc
f
ade
A
b

A
c

A
d,
A
e,
. (270)
The generating functional for the full theory can be obtained perturbatively,
expanding in g,
Z
_
J

, J

, J

, J

J
A
_
= exp
_
i
_
d
4
zL
interaction
_
i
J(z)
__
Z
0
_
J

, J

, J

, J

J
A
_
, (271)
where we must replace in the interaction Lagrangian the eld variables with
functional derivatives with respect to their corresponding sources.
63
4.2 BRST symmetry
We recall here the path-integral for a non-abelian gauge theory with a fermion
eld,
Z
_
DA
a

i
D
i
D
a
D
a
e
i
_
d
4
x[L
Y M
+L
fermion

1
2
(G
a
(A
a
))
2
+
a

a
]
, (272)
where we have dened,

a
(x) =
_
d
4
y
_
G
a
(A
a

(x))
A
c

(y)
D
cb

(y)
_

b
(y). (273)
It is convenient (not imperative) to linearize the action in the gauge-xing
term. If we introduce (yet) another (bosonic) eld, w
a
, we can re-write,
_
Dw
a
e
i
_
d
4
x(

2
w
a
w
a
+w
a
G
a
)
=
_
Dw
a
e
i

2
_
d
4
x
_
(w
a
+
G
a
2
)
2

G
2

2
_
= N
x
e
i
_
d
4
x
G
a
G
a
2
(274)
We will denote,
L
cl
= L
Y M
+L
fermion
, (275)
the classical Lagrangian which satises local gauge-invariance. The path-
integral, up to a normalization, is equal to
Z =
_
DA
a

i
D
i
D
a
D
a
Dw
a
e
i
_
d
4
x[L
cl
+
a

a
+w
a
G
a
+

2
G
a
G
a
]
. (276)
If we want to go back to the original form of Z we simply need to integrate out
the eld w
a
. The exponent is not gauge invariant, except the term with the
classical Lagrangian L
cl
. We will nd, however, a closely related symmetry
which is called Becchi-Rouet-Stora; Tyutin or, shorter, BRST symmetry.
In this new symmetry, the gauge boson A
a

and fermion elds transform

as before with local gauge transformations albeit with a gauge parameter
which involves the ghost eld. We will counter the eect of these gauge
transformations on the ghost and anti-ghost so that the variation due to the
gauge transformation is exactly canceled in the sum of the Fadeev-Popov
and gauge-xing terms.
We would like the BRST transformation to be a generalization ( bigger
symmetry) of gauge invariance for the classical Lagrangian. The easiest
64
way to achieve this is by demanding that BRST transformations are gauge
transformations of a special form for all the elds which enter the classical
Lagrangian L
cl
.
Let us require a BRST transformation of the fermion eld which is also
a gauge transformation:

= iT
a
(
a
) . (277)
The gauge parameter in the above transformation is
a
. The ghost eld

a
is a Grassmann variable, however gauge transformation parameters are
bosonic variables. We have
(gauge parameter) =
a
.
Therefore, the parameter of the BRST transformation is also a Grassmann
variable.
Exercise: Prove that the product of two Grassmann variables is a commuting
variable.
We take the parameter to be global

= 0.
This transformation mixes bosonic and fermionic variables.
We now require one more property for the BRST transformations. Two
successive BRST transformations on an arbitrary function of elds
should leave the function invariant (nilpotent).

1
F(A, ,

, , ) = 0. (278)
If we insist on this property, we obtain:
0 =

2
(

1
)
=

2
(iT
a

a
)
= iT
a

1
[(

a
) +
a
(

2
)]
= iT
a

1
_
(

a
) +
a
_
iT
b

__
= i
1

2
[(

a
)T
a
]
(279)
65
Equivalently,
T
c

c
= iT
b
t
c

c
; tr(T
a
T
c
)

c
= i
2
tr(T
a
T
b
T
c
)
b

c
beware of Grassmann!
;

ac
2

c
= i
2
tr
_
T
a
T
b
T
c
_

b

b
2
;

c
= i
2
tr
_
T
a
_
T
b
, T
c
__

c
;

c
= f
bcd
tr(T
a
T
d
)
b

c
, (280)
Therefore we require the ghost eld to transform as:

a
=

2
f
abc

c
. (281)
For the gauge boson, we require that the BRST transformations are the
gauge transformation with the same parameter
a
as for the fermion .
Specically,

A
a

g
D
ab

b
. (282)
Two successive transformations produce,

1
A
a

1
g

2
_

a
gf
abc

b
A
c

_
=

1
g
_

a
) gf
abc
_

b
_
A
c

gf
abc

b
_

2
A
c

__
=

1
g
_
D
ab

a
) +
1
2
f
abc

2
D
cd

d
_
= . . . = 0. (283)
Exercise:Fill the dots... Prove the above statement using the anti-commutation
of Grassmann variables and the Jacobi identity for the structure constants
The BRST transformation of the anti-fermion is also xed by the require-
ment of gauge invariance for the classical Lagrangian:

=

iT
a
(
a
) .
Notice that the Grassmann variable entering here is , so that we perform
the same gauge-transformation on all classical elds ,

, A
a

.
For the remaining two independent elds in the action of the path in-
tegral, we have no unabiguous guidance in order to construct their BRST
66
transformations. We will make two very simple choices. We perform no
transformation on the auxiliary bosonic eld,

w
a
= 0.
For the anti-ghost we require that under a a BRST transformation it gets
shifted by a a constant.

a
=
1
g
w
a
.
This choice as we will see guarantees BRST invariance of the quantum action.
Notice that

2

a
=

2
w
a
= 0.
Let us now compactify the notation. We consider any eld F from
{A

, ,

, ,

eta, w}. We will introduce the short-hand notation:

F (sF)
We have pulled an explicit prefactor of the BRST transformation parame-
ter, and the notations sF denotes the remainder of the expression after we
transformed the eld F. If we perform two consecutive transformations, we
have:

2
F =
1

2
s
2
F.
Nilpotency means that:
s
2
F = 0. (284)
Nilpotency is a property valid for a product of two variables as well.

1
(F
1
F
2
) = (

1
F
1
)F
2
+ F
1
(

1
F
2
)

1
(sF
1
)F
2
+ F
1

1
(sF
2
)

1
[(sF
1
)F
2
F
1
(sF
2
)] , (285)
where the minus sign arises if F
1
is a Grassmann variable.
We observe that the eld F and sF have always opposite statistics. If we
perform a double BRST transformation on the product F
1
F
2
we then nd,

1
(F
1
F
2
) =
1

2
[(sF
1
)F
2
F
1
(sF
2
)]
=
1
[(sF
1
)
2
(sF
2
)
2
(sF
1
)(sF
2
)]
=
1

2
[(sF
1
)(sF
2
) (sF
1
)(sF
2
)]
= 0. (286)
67
We can continue in the same spirit. We ns that:

1
(F
1
F
2
. . . F
n
) = 0. (287)
In fact every functional of these elds satises,

1
G[(F
1
, F
2
, . . . , F
n
)] = 0.
We will return to this shortly.
We should investigate the eect of a BRST transforamtion on the gauge
xing term G
a
in the Lagrangian. G
a
is a function of the gauge eld and we
should use the chain-rule.

G
a
(A

(x)) =
_
d
4
y
G
a
(x)
A

(y)

(y)
=
1
g
_
d
4
y
G
a
(x)
A

(y)
D
ab

b
(y)
=
1
g

a
(288)
Let us now consider the variation:

_

a
q
_
G
a
+

2
w
a
__
= (

a
)
_
G
a
+

2
w
a
_
+
a
(

G
a
)
=

g
_

a

a
+ w
a
G
a
+

2
w
a
w
a
_
. (289)
In other words, the non-classical part of the Lagrangian is already a total
variation under a BRST-transformation. Due to the property of nilpotence,
such terms remain invariant under a BRST-transformation.

_
L
cl
+
a

a
+ w
a
G
a
+

2
w
a
w
a
_
= (

L
cl
) + g

_
s
_

a
_
G
a
+

2
w
a
___
= 0
(290)
We remind here that the classical-part of the Lagrangian is BRST-invariant
due to its gauge-invariance.
Exercise:Prove that the Jacobian of a BRST transformation is unit. This
completes a proof that the full path-integral is BRST-invariant.
The BRST symmetry provides us with the asymptotic states of the S-
matrix. Let us compute the S-matrix element
| ,
68
in a two dierent gauges, G
1
and G
2
, where the two gauge xing conditions
dier by little:
G
2
= G
1
+

G.
We require that the initial | and nal | states are physical nad thus the
same in both gauges:
|
G
1
= |
G
2
, |
G
1
= |
G
2
.
The matrix-element is computed in terms of the LSZ reduction formula. A
dierence due to the dierent gauge-xing conditions will be introduced due
to the change of the action in the exponent of the path-integral. We will
have:

Z = Z
G
2
Z
G
1
=
_
DA. . . e
i
_
d
4
Lcl+gs K|
G
1

_
DA. . . e
i
_
d
4
Lcl+gs K|
G
2
Since we have considered innitesimally dierent gauge-xing conditions, we
have that:

Z = Z
G
1
igs

K.
The induced dierence in the S-matrix element should be proportional to:

| | s

K| .
We can construct an operator Q which is the generator of the BRST
transformations for canonical elds.

(Field) = i [Q, Field]

where a commutator for a eld with even spin and an anti-commutator for
a eld with odd spin are understood with the subscript. We can write:
0 = s(s Field) =
_
Q, [Q, Field]

=
_
Q
2
, Field
_

.
For the above to be satised for every eld we need:
Q
2
= 0.
Using the BRST generator we write:
| s

K| = |
_
Q,

K
_

| = | QK| | KQ| .
69
The matrix-element is the same in all gauges if the above vanishes; this
provides us with a condition for physical elds. Computations of matrix-
elements in dierent gauges yield the same result if the physical external
states annihilate the generator of BRST transformations:
| Q = Q| = 0. (291)
Exercise:
3
Construct the BRST generator Q in QED, i.e. nd a set of com-
mutation or anticommutation relations which it should satisfy with respect to
the creation and annihilation operators of all elds in the QED Lagrangian
3
This is only for the very curious!
70
5 Quantum eective action and the eective
potential
We have started to collect essential tools for the renormalization of gauge
eld theories by proving the existence of the BRST symmetry. In the re-
maining of this lecture series we will convince ourselves that gauge theories,
such as QCD, are renormalizable. It turns out that for renormalization we
need to worry only about one-particle-irreducible (1PI) Feynman diagrams.
If these are rendered nite, then the full S-matrix elements will possess no
other ultraviolet singularities. In this Section, we wll introduce a new func-
tional, the quantum eective action, which generates only 1PI graphs. The
quantum eective action is also a very important tool in order to dene the
ground state of the quantum eld theory, and to study symmetry breaking
via quantum eects.
We have worked with the generating functional
Z[J] =
_
De
i(S[]+
_
d
4
xJ(x)(x))
;
Greens functions are obtained via:
0| T(x
1
) . . . (x
n
) |0 =
1
i
n
1
Z[J]

n
Z[J]
J(x
1
) . . . J(x
n
)

J=0
.
We found that if we require only connected graphs to be generated, which
are relevant for computing S-matrix elements, we should use the generating
functional W[Z], with
Z[J] = e
iW[J]
;W[J] = ilnZ[J].
Then,
0| T(x
1
) . . . (x
n
) |0
CONNECTED
=
1
i
n1

n
W[J]
J(x
1
) . . . J(x
n
)

J=0
.
An important Greens function is the one-point function
0| |0 (x) =
W[J]
J(x)

J=0
,
which we have required to vanish if the eld is in the nal or the initial
state. This is necessary in order to guarantee that the asymptotic states
71
in the LSZ formula are pure one-particle states and they do not have an
overlap with the vacuum state. There are situations however, that a eld
(external eld) lls the vacuum 0| |0 (x) = 0. Such a eld cannot be in
the nal or initial state of a scattering process, but it can be a background
where scattering of other elds takes place. For example, the gravitation
or the Higgs boson eld may have such a role. It is useful to study such
cases; a quick way to do so is to consider the sources J when we compute
Greens functions as part of the action, and not set them to zero after taking
derivatives.
In the presence of sources in the action, we have

J
0| |0
J
(x) =
W[J]
J(x)
.
The above equation denes a relation between the source J(x) and the vac-
uum expectation of the corresponding eld. We can then trade source func-
tions in the path-integral and in functional derivatives with the corresponding
vevs (vacuum expectation values).

J
= function(J), J = function
1
(
J
).
There is a systematic method to perform this change of variables. We start
with the integral
_
d
4
x
j
(x)J(x), (292)
where we assume that
j
is computed from the rst derivative of W[J] for
an arbitrary source J(x). We then dierentiate with a vacuum expectation
value which corresponds to a source J

(y)
_
d
4
x
j
(x)J(x) =
_
d
4
x
_

J
(x)

J

(y)
J(x) +
J
(x)
J(x)

(y)
_
=
_
d
4
x
_
(x y)J

(x) +
W[J]
J(x)
J(x)

(y)
_
= J

(y) +
W[J]

(y)
(293)
We dene the quantum eective action as:
[
J
] = W[J]
_
d
4
xJ(x)
J
(x). (294)
72
From the above we nd the simple equation,
[ ]

J
(x)
= J(x). (295)
Recall the role of the classical action S[]. The equations of motion for
the classical eld are found by requiring that the action takes a minimal value
S[]

=
classical
= 0.
At the quantum level, elds are promoted into operators. It is meaningful
to ask what is the average value of the eld operators in the state with the
minimum energy (vacuum). In the absence of sources, the quantum eective
action yields the equations of motion for the average values of quantum elds:
[]

0
(x)
= 0. (296)
We may take a next step and promote the quantum eective action to
generate new Greens functions. We use it in the exponent of a path integral:
e
iW

[g,J]
=
_
D e
i
g
{[]+
_
d
4
xJ(x)(x)}
. (297)
We can establish perturbation theory using this path integral. It is possible
to derive propagators by identifying the quadratic terms in the action and
inverting the corresponding operator. We will not do this explicitly here; we
are rather interested in counting powers of the arbitrary constant g.
Every propagator in a graph (since it is produced by inversion), will con-
tribute a single power of g. Vertices are derived from the non-quadratic terms
in the Lagrangian without any inversion. Thus, each vertex will contribute
a power 1/g to a Feynman graph. For a graph with N
I
propagators and N
V
vertices the overall power of the coupling is:
g
N
I
N
V
.
All connected graphs generated by W

4
with N
I
propagators and N
V
vertices
have L = 1 + N
I
N
V
loops. It only takes trying a few examples out in
4
they are connected because W

is the logarithm of a path integral

73
order to convince ourselves about the above statement. Otherwise, assign N
I
unconstrained momenta for each propagator. Each vertex will provide one
constraint, of which one combination is an overall delta function stating that
the sum of momenta of incoming an outgoing particles is zero. N
I
N
V
+1
momenta are left unconstrained and they are thus loop momenta. The power
of g for a graph is therefore determined exclusively from the number of loops
that it posseses:
g
L1
.
We can then perform an expansion:
W

[g, J] =

L=0
g
L1
W
(L)

[J]. (298)
Of course, we can still be interested in the case with g = 1. What the above
expression tells us,
W

[1, J] =

L=0
W
(L)

[J], (299)
is that the generating functional W

[1, J] can be decomposed as a sum of

independent generating functionals W
(L)

[J] corresponding to dierent loop

orders. The functionals W
(

L) are independent in the sense that shifts in the

measure do not mix them; symmetries of the full action should therefore be
symmetries of each one of the loop contributions sepearately.
Let us explore the possibility of a very small parameter g. We can expand
the exponent in the path integral around the value:
=
J
+ , with
[]

J
= J.
We have
[] +
_
d
4
xJ = [
J
] +
_
d
4
xJ
J
+
+
_
d
4
x(x)
_
[]

J
+ J
_
+
_
d
4
xd
4
y(x)

2
[]

J
(x)
J
(y)
(y) + . . . (300)
The linear term in vanishes. We therefore have,
[] +
_
d
4
xJ =
_
[
J
] +
_
d
4
xJ
J
_
+O(
2
) = W[J] +O(
2
).
(301)
74
We then have,
e
i
g

L=0
g
L
W
(L)

[J]
= e
i
g
W[J]
_
De
i
g
_
O(
2
)
. (302)
The path integral over can be computed perturbatively. The factor
1
g
can
be eliminated by redening =

g
1/2
. Then we are left with a canonically
normalized quadratic part. The important observation to make is that this
perturbative expansion will start at order O(g
0
) the earliest. After taking
the logarithm of both sides of the above equation, and by comparing the
1
g
coecients, we nd that:
W [J] = W
(0)

[J]. (303)
In other words, the full generating functional of connected graphs W[J] can
be obtained by a generating function where we have replaced the classical
action with the quantum eective action:
W[J] =
_
TREE
De
i[[]+
_
d
4
xJ(x)(x)]
, (304)
and keeping only the tree-contributions (denoted with the subscript in
the integral symbol).
This is a remarkable result; it states that it is possible to re-organize
the perturbative expansion, which gives rise to both tree and loop graphs,
into a new expansion where only tree-graphs appear. Of course, W[J] and
W
(0)

[J] are equal. The corresponding perturbative expansions are therefore

equivalent; the apparent lack of loops in the the expansion from the path
integral wth the eective action W
(0)

[J] should be explained by a re-writing

of the usual expansion from W[J] with modied vertices and propagators.
These new exotic vertices and propagators should be dressed to account for
all loop eects that we have encountered in the path integral of the classical
action.
The above result is of great importance for renormalization. Trees do
not have any ultraviolet divergences. Therefore, we only need to establish
a renormalization procedure which renders nite the dressed propagators
and vertices of the quantum eective action.
Let us take the tree-only statement seriously, and write down all pos-
sible graphs that we might have for two-, three-, and four-point functions.
This will be sucient to establish a pattern for the Greens functions derived
from the eective action. Actually, we can only have a very small number of
tree-graphs for small number of external legs.
75
Figure 1: Possible tree-graphs for two-, three-, and four-point functions. This
very small number of connected graphs, which arises from the perturbative
expansion of the path integral W

[J] with the eective action [], should

contain in the propagators and vertices all loop eects found in the usual
path integral W[J].
We can gure out the propagators and vertices of the tree diagrams of
Fig 1 by comparing with the usual Feynman diagrams which we obtain by
expanding W[J]. The two-point function in Fig 1 must be equivalent to the
full propagator, computed at all orders in perturbation theory from W[J]:
= ( this is the full propagator )
= +
1PI
+
1PI 1PI
+
1PI 1PI 1PI
+ . . . (305)
where we sum all possible Feynman diagrams with two external legs. We can
write the sum of all graphs contributing to the full propagator as a geometric
76
series of one-particle-irreducible two-point loop Feynman diagrams.
1PI
= + + + . . . (306)
One-particle irreducible diagrams are these which cannot be separated into
two diagrams after we cut one of the propagators. Knowledge of the 1PI
propagator graphs is sucient to determine the full propagator. Let us work,
as an example, with a scalar eld theory. We denote, in Fourier-space,
1PI
=
i
p
2
m
2
_
i(p
2
)
_
i
p
2
m
2
.
From Eq. 305 we nd,
= =
i
p
2
m
2

n=0
_
(p
2
)
p
2
m
2
_
n
=
i
p
2
m
2
(p
2
)
.
Indeed, the full propagator is then computed using only 1PI graphs.
We proceed to compare the three-point Greens functions of Fig. 1 with
the result of the full perturbative expansion from W[J]. We now that the
propagators connected to the triple-vertex are full propagators.
=
The triple vertex must then be only the sum of all 1PI three-point functions.
=
+
+
+
77
Three-point graphs which are one-particle reducible, always contribute a two-
point subgraph to the full propagators of the external legs and an 1PI 3-point
subgraph to the vertex.
We can now convince ourselves easily that the four point vertex in Fig. 1
contains all one-particle irreducible four point functions.
The same of course holds for higher multiplicities. Our conclusion is that we
can always rearrange the sum of graphs in the perturbative exansion, derived
via W[J] and containing both loops and trees, to an equivalent tree-level
graphs only expansion where the propagator is the full two-point function
and the vertices are all one-particle irreducible graphs with the same number
of external legs as in the vertex.
The important statement is that W[J] = W
(0)

[J] and such a tree-expansion

can be obtained automatically from a generating functional if the eective ac-
tion. We then conclude that the full propagator and one-particle irreducible
Greens functions should be derivable from the eective action [J]. Let us
verify this statement for the two, three, and four point functions.
We rst introduce the short notation

J
(x)
x
, J(x) J
x
.
We start from the equation,

x
= J
x
.
78
Dierentiating with a source, we obtain:
(x y) =

J
y

x
=
_
d
4
z

x
=
_
d
4
z
_

2

i
x

z
_ _

2
W
iJ
z
J
y
_
. (307)
From the above we see that
1
i

y
is the inverse of the full two-point function
(x
1
x
2
) 0| T(x
1
)(x
2
) |0 =
1
i

2
W
J
x
J
y
.
Before we compute the three-point function we need two tricks.
- Chain rule:

J
x
=
_
d
4
z

z
J
x

z
=
_
d
4
z

2
W[J]
J
x
J
z

z
= i
_
d
4
z(x z)

z
. (308)
- Dierentiation of an inverse matrix
1 = MM
1
; 0 =
(MM
1
)

=
M

M
1
+ M
M
1

;
M
1

= M
1
M

M
1
. (309)
We have:
=
1
i
2

3
W[J]
J
x
1
J
x
2
J
x
3
79
=
_
d
4
y
1
(x
1
y
1
)

y
1
_

2
W
iJ
x
2
J
x
3
_
=
_
d
4
y
1
(x
1
y
1
)

y
1
_

2

i
x
2

x
3
_
1
=
_
d
4
y
1
d
4
y
2
d
4
y
3
(x
1
y
1
)
_

2

i
x
2

y
2
_
1

i
y
1

y
2

y
3
_

2

i
y
3

x
3
_
1
=
_
d
4
y
1
d
4
y
2
d
4
y
3
(x
1
y
1
)(x
2
y
2
)(x
3
y
3
)

3

i
y
1

y
2

y
3
(310)
Now we may compare the graph on the lhs and the rhs of this equation.
We have explicitly found that the full three-point function is the convolution
of propagators, one for each external leg, and the third derivative of the
eective action. From our earlier discussion we now that after we factor out
full propagators for the external legs, the remainder is the sum of one-particle
irreducible three-point diagrams.
Exercise: Prove that

i
y
1

y
2

y
3

y
4
is the sum of 1PI 4-point functions.
In summary, we can deduce from the Quantum Eective Action all phys-
ical predictions in a quantum eld theory.
The second derivative of [] is the inverse propagator. The zeros of
the inverse propagator yield the mass values of the physical particles
in the theory.
Higher derivatives of the eective action are 1PI Greens function. Con-
necting them with full propagators to from trees we can derive akk
connected amplitudes which are required for S-matrix element compu-
tations.
Additionally, solving the equation

= 0
yields the values of vevs where the eective action is minimal. This will
serve to dene the true ground-state of the theory. The location of the
minimum will also reveal whether any symmetries of the Lagrangian are
broken spontaneously.
80
5.1 The eective potential
We have just observed that by dierentiating the eective action functional
with respect to the eld vevs, we generate one-particle-irreducible Feynman
diagrams. All functional derivatives of [] are therefore represnted in
terms of Feynman diagrams; if we could compute all these diagrams we could
compute the full eective action by summing up all the terms of a Taylor
series expansion.
Specically, we can expand
[] =

n=1
1
n!
[]
(x
1
) . . . (x
n
)
(x
1
) . . . (x
n
)
=

n=1
1
n!

(n)
(x
1
, . . . , x
n
) (x
1
) . . . (x
n
) , (311)
where

(n)
(x
1
, . . . , x
n
)
[]
(x
1
) . . . (x
n
)
. (312)
are one-particle-irreducible Greens functions (in coordinate space).
We consider the case where the ground state (vacuum) is translation
invariant; it does not distinguish among dierent points in space-time. There
are situations where this is not true (instantons), however the space-time
blind vacuum case is interesting and common. We then have:
(x) = constant . (313)
In this case, the Greens functions simplify enormously if we use a Fourier
transformation ( momentum space):
_
d
4
x
1
. . . d
4
x
n

(n)
(x
1
, . . . , x
n
) (x
1
) . . . (x
n
)
=
n
_
d
4
x
1
. . . d
4
x
n

(n)
(x
1
, . . . , x
n
)
=
n
_
d
4
x
1
. . . d
4
x
n
_
d
4
k
1
(2)
4
. . .
d
4
k
n
(2)
4
e
ik
1
x
1
. . . e
iknxn
(2)
4

(4)
(k
1
+ k
2
+ . . . + k
n
)

(n)
(k
1
, . . . , k
n
)
=
n
_
d
4
k
1
. . . d
4
k
n

(4)
(k
1
) . . .
(4)
(k
n
)(2)
4

(4)
(k
1
+ k
2
+ . . . + k
n
)
81

(n)
(k
1
, . . . , k
n
)
=
n
(2)
4

(n)
(0, 0, . . . , 0)
(4)
(k
1
+ k
2
+ . . . + k
n
). (314)
Notice that we have explicitly shown the delta-function which imposes mo-
mentum conversation. The multiple integrations over space-time x
i
are sim-
ple because of the assumption of x-independent vev . The factor
(2)
4
(0) =
_
d
4
xe
i0 x
=
__
d
4
x
_
.
We then have for the eective action,
[] =
__
d
4
x
_

n=1

n
n!

(n)
(0, 0, . . . , 0). (315)
The eective potential is dened from the eective action, factoring out the
space-time volume:
V
eff
()]
[]
(
_
d
4
x)
(316)
We obtain:
V
eff
() =

n=1

n
n!

(n)
(0, 0, . . . , 0). (317)
Therefore, the recipe to compute the eective potential is:
Compute all 1PI Greens fucntions with increasing number of external
legs in momentum-space and setting all external momenta to zero.
For each external leg include a power of the classical vev of the corre-
sponding eld.
Sum the series up without forgetting to include the 1/n! from the Taylor
series expansion.
Let us consider the Lagrangian of a real scalar eld with a quartic inter-
action,
L =
1
2
(
m
u)
2

m
2
2

2

4!

4
. (318)
A computation of the eective potential including all orders in perturbation
theory is impossible. We can compute the eective potential easily in the
tree and one-loop approximation.
82
We observe that the only two 1PI Greens functions that we can write in
the tree approximation are the 2-point (amputated propagator) and 4-point
(vertex). From Eq. 317 we nd,
V
tree
eff
= +
m
2
2

2
+

4!

4
. (319)
It turns out that the eective potential at tree-level is the same as the po-
tential of the classical Lagrangian.
The 1-loop computation of the eective potential can be found in the
solutions of the exercises provided in this lecture.
83
6 Symmetries of the path integral and the
eective action
Our guiding principle in constructing realistic theories for particle inter-
actions is invariance of the classical action under certain symmetries (e.g.
BRST symmetry for Yang-Mills theories). Symmetries of the classical action
S may not be automatically symmetries of the eective action . However,
the eective action satises very general equations (Slavnov-Taylor identi-
ties) due to these classical symmetry constraints.
6.1 Slavnov-Taylor identities
Consider a theory of
i
interacting elds with arbitrary (bosonic or fermionic)
spin-statistics. We assume that this theory is symmetric under an innitesi-
mal symmetry transformation:

i

i

=
i
+ F
i
(x,
i
),
where is a small parameter and F
i
is usually an ordinary function of the
elds
i
and their derivatives. Then, we require that both the action and the
path-integral measure of the elds are invariant under this transformation:
S[
i
+ F
i
(x,
i
)] = S[
i
]
D(
i
+ F
i
(x,
i
)) = D
i
.
After transforming the elds, the generating path-integral is
Z[J
i
] =
_
D

i
e
iS[
i
]+i
_
d
4
x

i
J
i
=
_
D(
i
+ F
i
(x,
i
)) e
iS[
i
+F
i
(x,
i
)]+i
_
d
4
x(
i
+F
i
(x,
i
))J
i
=
_
D
i
e
iS[
i
]+i
_
d
4
x(
i
+F
i
(x,
i
))J
i
.
We can now expand in the small parameter ,
Z[J
i
] =
_
D
i
e
iS[
i
]+i
_
d
4
x
i
J
i
_
1 + F
i
(x,
i
) +O(
2
)
_
= Z[J
i
]
_
D
i
__
d
4
yF
i
(y,
i
)J
i
(y)
_
e
iS[
i
]+i
_
d
4
x
i
J
i
;
_
D
i
__
d
4
yF
i
(y,
i
)J
i
(y)
_
e
iS[
i
]+i
_
d
4
x
i
J
i
= 0, (320)
84
or, dividing by the path inegral,
_
d
4
y
_
_
_
D
i
F
i
(y,
i
)e
iS[
i
]+i
_
d
4
x
i
J
i
Z[J
i
]
_
_
J
i
(y) = 0 (321)
In the square brackets we recognize the average of the transformation over
all eld congurations,
F
i
(y,
i
)
J

_
D
i
F
i
(y,
i
)e
iS[
i
]+i
_
d
4
x
i
J
i
Z[J
i
]
. (322)
We then nd the identity,
_
d
4
y F
i
(y,
i
)
J
J
i
(y) = 0, (323)
concluding that if there exists an infnitesimal symmetry transformation of
the classical action, then there is a constraint on the average value of the
transformation. Eq. 323 depends on abritrary sources, and by dierentiating
multiple times with the sources, we can obtain an innite number of identi-
ties. These are called Slavnov-Taylor identities; we shall consider an example
soon.
6.2 Symmetry constraints on the eective action
The generating Slavnov-Taylor identity of Eq. 323 identity tells us that there
exists a symmetry for the eective action. Substituting
J
i
(y) =


i

J
(y)
,
we obtain:
_
d
4
y F
i
(y,
i
)
J

J
(y)
= 0. (324)
Equivalently,
[
i

J
]

+
_
d
4
y F
i
(y,
i
)
J

J
(y)
=
[
i

J
]

; [
i

J
] +
_
d
4
y F
i
(y,
i
)
J

J
(y)
= [
i

J
]
; [
i

J
+ F
i
(y,
i
)
J
] = [
i

J
] +O
_

2
_
. (325)
85
Therefore, the eective action is symmetric under the transformation

i

i

=
i
+ F
i
(
i
) . (326)
We recall that the classical action is symmetric under the transformation

i
=
i
+ F
i
(
i
) . (327)
Are these two transformations the same? Otherwise, is it F
i
= F
i
?
In general they are not! The symmetries of the classical action are usually
no symmetries of the quantum eective action. Consider an example of a
classical action symmetric under a eld transformation
(x) (x) +
2
(x)
The quantum action should be symmetric under the transformation
(x) (x) +
_

2
(x)
_
.
Given that
(x)
2
=
_

2
(x)
_
,
the two transformations are dierent.
Nevertheless, we can identify many symmetries in classical actions for
realistic eld theories which are linear:
F
i
[
i
, x] = c
i
(x) +
_
d
4
yT
ij
(x, y)
j
(y). (328)
The equivalent symmetry transformation for the eective action is
F
i
[
i
, x] =
_
c
i
(x) +
_
d
4
yT
ij
(x, y)
j
(y)
_
= c
i
(x) +
_
d
4
yT
ij
(x, y)
j
(y) , (329)
and it is identical. It is usefull to remember that linear symmetry trans-
formations of the classical action, are automatically symmetry trans-
formations of the eective action.
86
6.3 Contraints on the eective action from BRST sym-
metry transformations of the classical action
The BRST transformations are not linear; therefore they are only a sym-
metry of the classical action
5
and not of the eective action. Nevertheless,
the eective action is constrained by the BRST symmetry of the classical
Lagrangian (Eq. 323). For these nillpotent transformations Eq. 323 takes a
very special form, the so called Zinn-Justin equation.
We start with a classical action S[
i
] of elds
i
which is symmetric under
the BRST transformation

i
= B
i
. (330)
Since B
i
is nilpotent we also have

i
= 0
;

B
i
= 0 (331)
We realize that, because of Eq. 331, there is a more general action which has
the same symmetry as the original S[
i
]. It is easy to verify that the action,
S[
i
, K
i
] = S[
i
] +
_
d
4
xB
i
(x)K
i
(x), (332)
is indeed symmetric under the same transformation.
The functions K
i
are arbitrary (sources). We recall, however, that the
functions B
i
have the opposite spin-statistics of the corresponding eld
i
.
Since the product B
i
K
i
must have even spin-statistics (the same as the action
S), we deduce that the source K
i
and the eld
i
have also opposite spin-
statistics.
We can write the generating functional W for connected graphs,
e
iW[J
i
,K
i
]
=
_
D
i
e
iS[
i
]+i
_
d
4
xB
i
K
i
+i
_
d
4
x
i
J
i
. (333)
The elds
i
may be bosonic or fermionic, therefore the order that we have
chosed in writing the integrands in the exponential is important. Conven-
tionally, we have placed source terms J
i
, K
i
to the right.
5
From now on, by classical action for a gauge theory we mean the action obtained
after gauge-xing using the Fadeev-Popov method.
87
The vacuum expecation value
i
is given by

i
(y) =
_
D
i
(y)e
iS[
i
]+i
_
d
4
xB
i
K
i
+i
_
d
4
x
i
J
i
_
Dx
i
e
iS[
i
]+i
_
d
4
xB
i
K
i
+i
_
d
4
x
i
J
i
=

R
W[J
i
, K
i
]
J
i
(y)
. (334)
This is an implicit relationship among J
i
, K
i
,
i
, and we will consider
K
i
,
i
as independent variables, and the sources J
i
as being expressed in
terms of these two variables:
J
i
= J
i
(
i
, K
i
).
The exact form of J
i
(
i
, K
i
) can be found if we evaluate explicitly the
eective action,
[
i
, K
i
] = W [J
i
(
i
, K
i
), K
i
]
_
d
4
x
i
J
i
(
i
, K
i
), (335)
and take a left derivative,

L
[
i
, K
i
]

i
(x)
= J
i
(
i
, K
i
)(x). (336)
We now compute the derivative of the eective action with respect to the
sources K
i
.

R
[
i
, K
i
]
K
i
(x)
=

L
K
i
(x)
_
W [J
i
(
i
, K
i
), K
i
]
_
d
4
x
i
J
i
(
i
, K
i
)
_
=

R
W [J
i
, K
i
]
K
i
(x)

J
i
=J
i
(
i
,K
i
)
+
_
d
4
y
_
_

R
W [J
i
, K
i
]
J
m
(y)

Jm=Jm(
i
,K
i
)
_
_
_

R
J
m
(
i
, K
i
)(y)
K
i
(x)
_

_
d
4
y
m
(y)

R
J
m
(
i
, K
i
)(y)
K
i
(x)
=

R
W [J
i
, K
i
]
K
i
(x)

J
i
=J
i
(
i
,K
i
)
+
_
d
4
y
m
(y)

R
J
m
(
i
, K
i
)(y)
K
i
(x)
88

_
d
4
y
m
(y)

R
J
m
(
i
, K
i
)(y)
K
i
(x)
=

R
W [J
i
, K
i
]
K
i
(x)

J
i
=J
i
(
i
,K
i
)
= i

R
K
i
(x)
ln
__
D
i
e
iS[J
i
]+i
_
d
4
x
i
J
i
+
_
d
4
xB
i
K
i
_

J
i
=FIXED
= B
i
. (337)
From the general Slavnov-Taylor identity we have,
_
d
4
xB
i
J
i
= 0
;
_
d
4
xB
i

= 0
;
_
d
4
x

K
i

= 0. (338)
This is a constraint which depends only on the eective action [
i
, K
i
]
(Zinn-Justin equation). It is a very useful form in order to study the con-
sequences of symmetry for the eective action, especially in connection with
renormalization proofs and studying anomalies.
For later use, we dene the product
(F, G)
_
d
4
x
_

R
F
K
i

L
G

i

R
F

i

L
G
K
i
_
. (339)
for functionals F[
i
, K
i
], F[
i
, K
i
] of the functions
i
, K
i
. Recall that

i
and K
i
have opposite spin-statistics. Then, the Zinn-Justin equation
can be written as
(, ) = 0. (340)
6.4 Slavnov-Taylor identities in QED
We now consider an example of Slavnov-Taylor identities in QED. The clas-
sical Lagrangian is, in the Lorentz gauge,
L =
1
4
F

+

(iD m)
1
2
(

)
2
. (341)
89
The corresponding path integral is,
Z [J

, , ] =
_
DA

De
i
_
d
4
[L+A

J+

+ ]
. (342)
An inniterimal local gauge transformation is:
A

1
q

(x),
(1 iq(x)),

(1 + iq(x))

.
The path integral measure is invariant under the gauge transformation. In
the integrand of the path-integral exponent, we can identify a part which is
invariant under these transformations, while the remaining, which includes
tha gauge-xing term and the source term, is not invariant.
L
noninvariant
=
1
2
(

)
2
+ A

+

+ . (343)
By performing a gauge-transformation on the path-integral we can derive, as
before, the Slavnov-Taylor identity,
_
d
4
xL
noninvariant
= 0. (344)
We can work out what is the change in the non-invariant part of the La-
grangian. The gauge-xing transforms as:

1
2
(

)
2

1
2
_

_
A

1
q

__
2
=
1
2
(

)
2
+
1
q

2
(x) +O
_

2
_
;
_

1
2
(

)
2
_
=
1
q

2
(x). (345)
Adding the variation of the source terms, we obtain:
L
noninvariant
=
1
q

2
(x) iq(x) + iq(x)

1
q
J

(x).
(346)
90
The Slavov-Taylor identity is:
_
d
4
x
_
1
q

2
(x) iq(x) + iq(x)

1
q
J

(x)
_
= 0
;
_
d
4
x
_
1
q

2
(x) iq(x) + iq(x)
_

_

1
q
J

(x)
_
= 0
;
_
d
4
x(x)
_
1

2
A

iq
2
+ iq
2
_

_
+

_
= 0, (347)
where we have used integration by parts. The above should be valid for
arbitrary (x), therefore the kernel of the integration in the square brackets
should be identically zero.
1

2
A

iq
2
+ iq
2
_

_
+

= 0. (348)
Substituting vacuum expectation values with functional derivatives of the
path-integral for connected graphs, W = ilnZ, we write:
1

2
W
J

iq
2

W

iq
2
W

= 0, (349)
where the functional derivatives are left derivatives for the fermionic sources.
Eq. 349, provides constraints for Greens functions in QED at all orders
in perturbation theory. We nd the simplest example, by dierentiating this
equation with a photon source and then set all sources to zero,

(y)
_
1

2
W
J

iq
2

W

iq
2
W

J,, =0
= 0,
;
1

2

2
W
J

(x)J

(y)

J,, =0
+

(x y) = 0.
;
1

2
0| TA

(x)A

(y) |0 =

(x y). (350)
We now write the Fourier representations,
(x y) =
_
d
k
(2)
4
e
ik(xy)
,
and
0| TA

(x)A

(y) |0 =
_
d
k
(2)
4
e
ik(xy)
D

(k).
91
Substituting into Eq. 350 we nd that the photon propagator in momentum
space should satisfy,
k

(k) =
k

k
2
(351)
We can write, in complete generality,
D

= A(k
2
)g

+ B(k
2
)
k

k
2
. (352)
Substituting in Eq. 351, we nd
A(k
2
) + B(k
2
) =

k
2
. (353)
Thus, the photon propagator in momentum space has the form,
D

(k) = A(k
2
)
_
g

k
2
_
+

k
2
k

k
2
(354)
This is a result valid at all orders in perturbation theory.
Notice that the term which depends on the gauge-xing parameter is fully
known. We can compare this with the result at leading order in perturbation
theory,
D

(k) =
1
k
2
_
g

k
2
_
+

k
2
k

k
2
+O(g
2
). (355)
We can see that the gauge-xing contribution is accounted fully in the leading
order result, and therefore it is not modied at higher orders in perturbation
theory. Higher order corrections modify only the function A(k
2
). For this
reason, the gauge-xing parameter does not receive any renormalization.
A second important observation to make is that the part of the propagator
which does not depend on ,
D

T
= A(k
2
)
_
g

k
2
_
,
is transverse to the photon-momentum. Indeed, we easily nd that
D

T
(k)k

= 0.
92
7 Renormalization: counting the degree of
ultraviolet divergences
Consider a Lagrangian with elds f and generic interaction operators O
i
.
L = kinetic terms +g
1
O
1
+ g
2
O
2
+ . . . + g
N
O
N
. (356)
Each operator O
i
is a product of elds and/or their derivatives. In QED for
example, we have one such interaction term:

A;
In QCD more operators emerge, e.g.
f
abc

A
a

A
,b
A
,c
, f
abe
f
cde
A
a

A
b

A
,c
A
,d
, . . .
We would like to keep this discussion as general as possible; we shall nally
be able to make statements on whether we can remove via renormalization
ultraviolet innities from arbitrary Lagrangians. Most of our arguments will
be derived using simple dimensional analysis.
We consider a generic one-particle irreducible Feynman diagram in per-
turbation theory. We will rst nd a simple formula to test whether it has the
most obvious of all possible divergences, the so called supercial ultra-violet
divergence. If the diagram has L loops, a supercial divergence corresponds
to an innity of the diagram in the limit
|k
1
| = |k
2
| = . . . = |k
L
| = ,
where k
i
are the loop-momenta. A Feynman diagram might have divergences
in other limits, where only some momenta or linear combinations of them are
taken to innity while the remaining independent momenta remain xed. A
Feynman diagram in the supercial ultraviolet limit behaves as
_

d
D1
, (357)
where D is an integer, called the supercial degree of divergence.
If D > 0 the Feynman diagram has a powerlike divergence,
if D = 0 it diverges logarithmically,
93
if D < 0 it is convergent (only in the supercial limit, since it might
have other divergences).
We can compute D (or an upper bound of it) for any Feynman diagram
on general grounds. We assume that our 1PI Feynman diagram has
I
f
internal propagators for each of the elds f,
E
f
external legs for each of the elds f, and
N
i
vertices corresponding to the term g
i
O
i
in the Lagrangian.
Recall, as examples, the Feynman rules for propagators in gauge theories,
and how they behave at innity.
a photon propagator,

+
kk
k
2
k
2

2
;
a fermion propagator,

k + m
k
2
m
2

1
;
for a scalar,

1
k
2
m
2

2
.
For each of the internal propagators of the eld f in the Feynman diagram
there is a contribution to the supercial divergence,

f
k
2+2s
f
,
where, s
f
= 0 for a boson and s
f
=
1
2
for a fermion. The total contribution
to the asymptotic limit from propagators is then

f
2I
f
(s
f
1)
. (358)
The contribution from vertices is easy to nd, if we know how many loop
momenta appear in the corresponding Feynman rules. For a vertex with an
operator O
i
this is equal to the number of space-time derivatives d
i
which can
94
be found in the expression for O
i
. Recall that a Feynman rule for a vertex is
esentially the Fourier transform of the corresponding operator, and momenta
may only arise from derivatives. The total contribution from vertices to the
asympotic limit is then

i
N
i
d
i
. (359)
Finally, due to the integration measure d
4
k
i
for each loop, the total contri-
bution from the loop-momenta to the asymptotic limit is

4L
, (360)
where L is the number of loops in the graph. L is known if we are given the
number of internal propagators I
f
and the number of vertices N
i
in the graph.
The number of loop-momenta carried from internal propagators is

f
I
f
. The
vertices provide

i
N
i
constraints of which one is not for loop-momenta but
for the external momenta. Therefore, the number of loop-momenta is
L =

f
I
f

i
N
i
+ 1.
Putting together the contributions from the loop integration measures, ver-
tices, and internal propagators, we nd that the asymptotic behavior at
innity has a supercial degree of divergence
D =

f
2I
f
(1 + s
f
)

i
N
i
(4 d
i
) + 4. (361)
We can express the number of internal propagators in terms of the number
of external legs. Let us assume that we have N
if
particles f in the vertex
corresponding to the operator O
i
in the Lagrangian. The total number of
(internal) legs of the particle f connected in all the vertices of the graph are

i
N
i
N
if
.
From these E
f
are external and the remaining are internal. Every propagator
of f has two edges, so the number of internal legs is
2I
f
.
We then have the identity
2I
f
+ E
f
=

i
N
i
N
if
. (362)
95
We can therefore write the degree of divergece as
D = 4

f
E
f
(1 + s
f
)

i
N
i
_
_
4 d
i

f
N
if
(1 + s
f
).
_
_
(363)
Notice that the square bracket in the last expression depends only on
the functional form of the operator O
i
. If this operator is multiplied with a
coupling constant g
i
in the Lagrangian, i.e. L = g
i
O
i
+. . . we can prove that
this square bracket is exactly the mass dimensionality of the coupling g
i
:
[g
i
] = 4 d
i

f
N
if
(1 + s
f
). (364)
We can prove this rather easily. The mass dimensionality of each term in
the action should be zero. We then have that
_
d
4
x
_
+ [g
i
] + [O
i
] = 0,
where the operator O
i
has d
i
derivatives and N
if
elds f. Thus,
4 + [g
i
] + d
i

f
N
if
[f] = 0,
and [f] = 1 +s
f
is the mass dimensionality of the eld. Indeed,
0| Tf(x
1
)f(x
2
) |0
_
k
d
4
kk
2+2s
f
e
ikx
; 2[f] = 4 2 + 2s
f
;[f] = 1 +s
f
. (365)
In conclusion, we can write a very suggestive expression for the supercial
degree of divergence:
D = 4

f
E
f
(1 + s
f
)

i
N
i
[g
i
]. (366)
If the Lagrangian does not contain any couplings with negative mass dimen-
sions, [g
i
] 0, we nd a supercial ultraviolet divergence, D 0, in Feynman
diagrams with a small number of external legs.
D 0 ;

f
E
f
(1 + s
f
) 4.
96
Figure 2: Both two-loop graphs have the same supercial degree of divergence
D = 1. However the Feynman diagram on the right has a self-energy one-
particle-irreducible subgraph which has a supercial degree of divergence
D = 1. A necessary condition for a graph to be UV nite is that the graph
and all its subgraphs have D < 0
In particular, supercial divergences do not appear in (one-particle-irreducible)
Feynman diagrams with ve external legs or more.
Examples of theories where supercial divergences may appear in only
a limited number of Greens fucntions are QED and QCD. All interaction
operators have dimension four and their coecients are dimensionless. Su-
percial ultraviolet divergences are limited in 1PI Greens functions, such as
0| T

(x
1
)(x
2
) |0 , 0| TA
a

(x
1
)A
b

(x
2
) |0 , 0| T

(x
1
)A
b

(x
2
)(x
3
) |0 , . . .. On
the contrary 0| T

(x
1
)(x
2
)

(x
3
)(x
4
) |0
1PI
is (supercially) nite.
Theories with [g
i
] 0 are called renormalizable. As we shall see,
these supercial divergences in a nite number of Greens functions can be
removed by adding a nite number of extra terms in the original Lagrangian
(counterterms).
If the Lagrangian contains a coupling with negative mass dimension [g
j
] <
0, then from Eq. 7.1 we see that all Greens functions, at some loop-order,
will develop a supercial divergence. It is therefore impossible to cancel the
innities by adding a nite number of counterterms. Such theories are called
non-renormalizable.
We should stress once again that the counting of the supercial degree of
divergence is not sucient to prove that a Feynman is nite or it is divergent.
Consider the two example graphs of Fig. 2. Both two-loop graphs have the
same supercial degree of divergence D = 1. One could then conclude that
both Feynman two-loop graphs are likely to be nite. We know, though, that
this is not the case. Let us compute the supercial degree of divergence for
97
all one-loop subgraphs that we can spot in the two diagrams. For the left
diagram, we nd that all subgraphs have a supercial degree of divergence
D = 1. Indeed, if one computes this integral explicitly it is found to be UV
nite. However the two-loop Feynman diagram on the right has an one-loop
self-energy subgraph; this has a supercial degree of divergence D = 1. The
self-energy is 1/ divergent. Such a divergence remains even after we embed
it inside a two-loop graph (there is no mechanism to cancel it), and against
our naive counting for the global supercial degree of divergence the two-loop
diagram on the right is divergent. The lesson from the above examples is
that for a diagram to be UV nite it is necessary that the superecial degrees
of divergences for the full graphs and all of its sub-graphs must be D < 0.
7.1 Cancelation of supercial divergences with coun-
terterms
We derived a criterion to decide whether a Greens function will develop
the most obvious type of divergence (supercial) in the limit where the
magnitudes of all loop momenta approach concurrently innity. We have
found that such innities appear, in renormalizable theories, in only a nite
number of Greens functions with a small number of external legs.
Innities are not acceptable for physical theories. A way out of this
problem is to recognize that the Lagrangian that we started with has a certain
degree of arbitrariness, which was constructed with the guiding principle of
respecting a given symmetry (e.g. BRST symmetry). However, this is not a
tight enough constraint to x, for example, the actual values of independent
mass and coupling parameters. In fact, we are allowed to add more terms in
the Lagrangian as long as they do not break the original symmetry.
Let us not worry for now about how restrictive are symmetry contraints.
We have seen that for renormalizable theories the disease of innities
is in only spread to a few Greens functions. We can then attempt a mod-
ication of the Lagrangian, adding a few new terms (counterterms) with
coecients engineered to cancel exactly these innities order by order in
perturbation theory. We shall worry whether this is a modication compati-
ble with symmetry in the next chapter. As a rst priority, we will investigate
if mathematically innities from loop diagrams can be cancelled from coun-
terterms. For this method to work, it is essential that a counterterm and the
loop innities yield the same type of kinematic functions. If, for example,
98
the 1/ terms of a Greens function at the one-loop order are logarithms of
external momenta,
ln(p
2
)

such a contribution cannot be cancelled by the tree-level contribution of a

countertem. Recalling the Feynman rules for vertices in QCD for example, we
nd no such logarithms in their expressions; we rather encounter constants
or simple polynomials of momenta. This suggests that a minimal necessary
condition for the counterterm program to be succesfull is that one-loop in-
nities are local, i.e. they appear as simple polynomials in the external
momenta as the usual Feynman rules do.
Let us look at the fuctional form of the supercial innities in perturba-
tion theory, and convince ourselves that indeed this is exactly what happens
in practice. We nd such an innity when computing, for example, the one-
loop correction to the four-point function in the

4!

4
theory. The supercial
degree of divergence is D = 4 4 1 = 0 and the graph is indeed diver-
gent. In dimensional regularization, the corresponding Feynman parameter
integral yields,
0| T(x
1
)(x
2
)(x
3
)(x
4
) |0|
1loop

2
_
d
d
k
1
(k
2
m
2
) [(k + p)
2
m
2
]

2
()
_
1
0
dx
1
dx
2
(1 x
1
x
2
)
(m
2
x
1
x
2
s)

2

2
_
1
0
dxln
_
m
2
x(1 x)s
_
(367)
with s = (p
1
+ p
2
)
2
= (p
3
+ p
4
)
2
.
Loop integrals, in general, contain logarithms or integrals of logarithms
(polylogaritms) with arguments kinematic invariants formed from external
momenta. Our example result is not an exception and we indeed nd loga-
rithmic contributions in the nite part. We cannot escape to observe how-
ever, that the divergent part is very simple; it is just a constant. We can
then modify the interaction terms in the Lagrangian,

4!

4

4!

4
+

2

4
and adjust the coecient so that it cancels exactly the divergent part of
this one-loop integral.
99
Supercial divergences essentially are singular parts of one-loop ampli-
tudes (as we explained, we set all loop-momenta equal before taking the
innite limit in order to nd the supercial degree of divergence). Crucially,
the divergent parts of one-loop integrals are simple polynomials of the exter-
nal momenta, as in the above example. If we use Feynman parameters, any
one-loop integral may be written as,
I
1loop
(N
d
2
)
_
1
0
dx
1
. . . dx
n
(1 x
1
. . . x
n
)
(m
2
1
x
1
+ . . . + m
2
n
x
n

s
i...j
x
i
x
j
i)
N
d
2
(368)
where N is an integer (equal to the number of propagators) and d = 4 2
the dimension. The denominator contains a sum over masses and kinematic
invariants of the external momenta. Divergences may arise from two terms;
the Gamma function (N 2 + ) and the denominator of the integrand.
The denominator does not have any ultraviolet divergences. It could become
divergent when masses or invariants become zero, but it is nite when all
the propagators are massive. If this is not the case, and there are massless
partices propagating in a loop, singularities from the denominator are of in-
frared nature connected to the small or zero values of |k| rather than the UV
|k| limit. Infrared singularities can also be regulated by attributing
a small mass to massless particles and/or considering them to be slighlty
o-shell. We shall not worry here about infrared one-loop singularities (they
cancel when computing physical cross-sections, but this is an interesting topic
which unfortunatelly cannot be discussed in this course). In short, the ul-
traviolet divergences of one-loop amplitudes are only due to the the Gamma
function term:
(N d/2) = (1 + ), ()
for N = 1, 2. Using the identity,
(x) =
(1 +x)
x
,
and
(1 +) = 1 +O(
2
),
and the fact that for N = 1, 2 (responsible for the UV divergences) the de-
nominator of Eq. 368 turns into a numerator Nd/2 < 0 in four dimensions,
we can see that the coecient of the 1/ pole can only be a polynomial in the
external momenta. In fact, the rank of the polynomial is necessarily smaller
100
than or equal to the supercial degree of divergence. A loop diagram with
supercial degree of divergence D
_
|k|
|k|
D1
, (369)
has a mass dimensionality D. Therefore, this polynomial can only be of
rank D in the external momenta. Each term in this polynomial multiplying
1/epsilon should be cancelled by a dierent counterterm, and it is very good
news that the rank of the polynomial is bounded by the supercial degree of
divergence. Each term in the polynomial must be cancelled with a separate
counterterm with a dierent number of derivatives. We can only introduce
counterterms with couplings of zero of negative mass dimension,
[g
i
] = 4 d
i

f
N
if
(1 + s
f
) 0 (370)
so the number of derivatives that we are allowed to introduce is:
d
i
4 d
i

f
N
if
(1 + s
f
).
This is the same bound as of the rank of the divergent polynomial,
D 4

f
E
f
(1 + s
f
).
7.2 Nested and overlapping divergences
This is a very nice topic of historic proportions which you should not miss
the opportunity to interview Klauss Hepp about!
It can be proven that we only to worry about removing supercial diver-
gences from loop integrals. Nested and overlapping singularities are auto-
matically removed with this procedure as well. More details will be added
to this section at a future version.
101
8 Proof of renormalizability for non-abelian
gauge theories
This (last) lecture followed Section 17.2 from The Quantum Theory of Fields,
Vol. II by Steven Weinberg. This is NOT AN EXAM MATERIALunless
you choose otherwise. You have the possibility to present within 25
minutes a COMPREHENSIVE REVIEW of Section 17.2 during
your examination session. In such a case, any questions will be connected
directly to your presentation.
102
References
[1] The Quantum Theory of Fields, Volume I Foundations, Steven Wein-
berg, Cambridge University Press.
[2] The Quantum Theory of Fields, Volume II Modern Applications, Steven
Weinberg, Cambridge University Press.
[3] Gauge Field Theories, Stefan Pokorski, Cambridge monographs on
mathematical physics.
[4] An introduction to Quantum Field Theory, M. Peskin and D. Schroeder,
Addison-Wesley
[5] An introduction to Quantum Field Theory, George Sterman, Cambridge
University Press.
[6] Quantum Field Theory, Mark Srednicki, Cambridge University Press.
[7] Modern Quantum Mechanics, J.J. Sakurai, Addison-Wesley.
103