Advanced Topics in Quantum Field Theory

Advanced Quantum Field Theory :
Renormalization, Non-Abelian Gauge Theories

and Anomalies
Lecture notes
Brussels, September 2011
Adel Bilal
Laboratoire de Physique Theorique, Ecole

Normale Superieure - CNRS
24 rue Lhomond, 75231 Paris Cedex 05, France
Abstract
This is part of an advanced quantum field theory course intended for graduate students
in theoretical high energy physics who are already familiar with the basics of QFT. The
first part quickly reviews what should be more or less known: functional integral methods
and one-loop computations in QED and 4 . The second part deals in some detail with
the renormalization program and the renormalization group. The third part treats the
quantization of non-abelian gauge theories and their renormalization with special emphasis on the BRST symmetry. The fourth part discusses gauge and gravitational anomalies,
how to characterise them in various dimensions, as well as anomaly cancellations.
Unite mixte du CNRS et de lEcole Normale Superieure associee `a lUniversite Paris 6 Pierre et Marie Curie
Contents
1 Functional integral methods
1.1 Path integral in quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Functional integral in quantum field theory . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Derivation of the Hamiltonian functional integral . . . . . . . . . . . . . . .
1.2.2 Derivation of the Lagrangian version of the functional integral . . . . . . .
1.2.3 Propagators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Green functions, S-matrix and Feynman rules . . . . . . . . . . . . . . . . . . . . .
1.3.1 Vacuum bubbles and normalization of the Green functions . . . . . . . . . .
1.3.2 Generating functional of Green functions and Feynman rules . . . . . . . .
1.3.3 Generating functional of connected Green functions . . . . . . . . . . . . .
1.3.4 Relation between Green functions and S-matrix . . . . . . . . . . . . . . .
1.4 Quantum effective action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Legendre transform and definition of [] . . . . . . . . . . . . . . . . . . .
1.4.2 [] as quantum effective action and generating functional of 1PI-diagrams
1.4.3 Symmetries and Slavnov-Taylor identities . . . . . . . . . . . . . . . . . . .
1.5 Functional integral formulation of QED . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1 Coulomb gauge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.2 Lorentz invariant functional integral formulation and -gauges . . . . . . .
1.5.3 Feynman rules of spinor QED . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
2
3
5
7
7
10
12
14
16
16
17
20
21
21
23
26
2 A few results independent of perturbation theory

2.1 Structure and poles of Green functions . . . . . . . . . . . . . . . .
2.2 Complete propagators, the need for field and mass renormalization
2.2.1 Example of a scalar field . . . . . . . . . . . . . . . . . .
2.2.2 Example of a Dirac field . . . . . . . . . . . . . . . . . . . .
2.3 Charge renormalization and Ward identities . . . . . . . . . . . . .
2.4 Photon propagator and gauge invariance . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
28
28
31
31
34
35
37
3 One-loop radiative corrections in 4 and QED

3.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Evaluation of one-loop integrals and dimensional regularization
3.3 Wick rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Vacuum polarization . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Electron self energy . . . . . . . . . . . . . . . . . . . . . . . . .
3.6 Vertex function . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Cancellation of the divergent piece . . . . . . . . . . . .
3.6.2 The magnetic moment of the electron: g 2 . . . . . .
3.7 One-loop radiative corrections in scalar 4 . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
41
42
46
48
51
53
54
55
57
. . . . . . . .
. . . . . . . .
and theorem
. . . . . . . .
. . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
61
61
63
66
68
68
5 Renormalization group and Callan-Szymanzik equations

5.1 Running coupling constant and -function: examples . . . . . . . . . . . .
5.1.1 Scalar 4 -theory . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.2 QED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Running coupling constant and -functions: general discussion . . . . . .
5.2.1 Several mass scales . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 Relation between the one-loop -function and the counterterms . .
5.2.3 Scheme independence of the first two coefficients of the -function
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
70
70
70
73
75
75
76
78
4 General renomalization theory

4.1 Degree of divergence . . . . . . . . . . . . . . . . . .
4.2 Structure of the divergences . . . . . . . . . . . . . .
4.3 Bogoliubov-Parasiuk-Hepp-Zimmermann prescription
4.4 Summary of the renormalization program and proof
4.5 The criterion of renormalizability . . . . . . . . . . .
Adel Bilal : Advanced Quantum Field Theory
.
.
.
.
.
.
.
.
.
Lecture notes - September 27, 2011
5.3
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
in asymptotic
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
.
.
.
.
.
.
.
.
.
.
79
79
79
80
81
82
83
83
85
87
.
.
.
.
.
88
89
89
90
91
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
92
92
96
97
98
100
102
7 Renormalization of non-abelian gauge theories

7.1 Slavnov-Taylor identities and Zinn-Justin equation . . . . . . . . . . . . . . .
7.1.1 Slavnov-Taylor identities . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.2 Zinn-Justin equation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.3 Antibracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.4 Invariance of the measure under the BRST transformation . . . . . . .
7.2 Renormalization of gauge theories theories . . . . . . . . . . . . . . . . . . . .
7.2.1 The general structure and strategy . . . . . . . . . . . . . . . . . . . .
7.2.2 Constraining the divergent part of . . . . . . . . . . . . . . . . . . .
7.2.3 Conclusion and remarks . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Background field gauge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4 One-loop -functions for Yang-Mills and supersymmetric Yang-Mills theories
7.4.1 -function for Yang-Mills theory . . . . . . . . . . . . . . . . . . . . .
7.4.2 -functions in supersymmetric gauge theories . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
106
106
106
107
108
108
110
110
112
118
119
121
122
125
8 Anomalies : basics I
8.1 Transformation of the fermion measure: abelian anomaly
8.2 Anomalies and non-invariance of the effective action . . .
8.3 Anomalous Slavnov-Taylor-Ward identities . . . . . . . .
8.4 Anomaly from the triangle Feynman diagram: AVV . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
127
127
127
127
127
5.4
5.5
-functions and asymptotic behaviors of the coupling . . . . . . . . . . . . . .

5.3.1 case a : the coupling diverges at a finite scale M . . . . . . . . . . . .
5.3.2 case b : the coupling continues to grow with the scale . . . . . . . . .
5.3.3 case c : existence of a UV fixed point . . . . . . . . . . . . . . . . . .
5.3.4 case d : asymptotic freedom . . . . . . . . . . . . . . . . . . . . . . . .
5.3.5 case e : IR fixed point . . . . . . . . . . . . . . . . . . . . . . . . . . .
Callan-Symanzik equation for a massless theory . . . . . . . . . . . . . . . . .
5.4.1 Renormalization conditions at scale . . . . . . . . . . . . . . . . . .
5.4.2 Callan-Symanzik equations . . . . . . . . . . . . . . . . . . . . . . . .
5.4.3 Solving the Callan-Symanzik equations . . . . . . . . . . . . . . . . .
5.4.4 Infrared fixed point and critical exponents / large momentum behavior
free theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Callan-Symanzik equations for a massive theory . . . . . . . . . . . . . . . . .
5.5.1 Operator insertions and renormalization of local operators . . . . . . .
5.5.2 Callan-Symanzik equations in the presence of operator insertions . . .
5.5.3 Massive Callan-Symanzik equations . . . . . . . . . . . . . . . . . . .
6 Non-abelian gauge theories: formulation and quantization

6.1 Non-abelian gauge transformations and gauge invariant actions
6.2 Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Faddeev-Popov method . . . . . . . . . . . . . . . . . .
6.2.2 Gauge-fixed action, ghosts and Feynman rules . . . . . .
6.2.3 BRST symmetry . . . . . . . . . . . . . . . . . . . . . .
6.3 BRST cohomology . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9 Anomalies : basics II
127
9.1 Triangle diagram with chiral fermions only . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
9.2 Locality and finiteness of the anomaly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
9.3 Cancellation of anomalies, example of the standard model . . . . . . . . . . . . . . . . . . . . 127
10 Anomalies : formal developments
127
10.1 Differential forms and characteristic classes in arbitrary even dimensions . . . . . . . . . . . . 127
10.2 Wess-Zumino consistency conditions and descent equation . . . . . . . . . . . . . . . . . . . . 127
11 Anomalies in arbitrary dimensions

11.1 Relation between anomalies and index theorems . . . . . . . . . .
11.2 Gravitational and mixed gauge-gravitational anomalies . . . . . .
11.3 Anomaly cancellation in ten-dimensional type IIB supergravity
E8 E8 heterotic superstring . . . . . . . . . . . . . . . . . . . .
127
. . . . . . . . . . . . . . . . 127
. . . . . . . . . . . . . . . . 127
and in type I SO(32) or
. . . . . . . . . . . . . . . . 127
PART I :
A QUICK REVIEW OF WHAT SHOULD BE KNOWN
1
Functional integral methods
1.1
Path integral in quantum mechanics
The usual description of quantum mechanics is in the Schrodinger picture where

[Qa , Pb ] = i ab
Qa |qi = qa |qi
Pa |pi = pa |pi
hq |pi =
Y eiqa pa
.
2
a
(1.1)
Go to the Heisenberg picture by Qa (t) = eiHt Qa eiHt and Pa (t) = eiHt Pa eiHt . The eigenstates of
these Heisenberg picture operators are
|q, ti = eiHt |qi
Qa (t) |q, ti = qa |q, ti ,
|p, ti = eiHt |pi
Pa (t) |p, ti = qa |p, ti .
(1.2)
Note that these are not the Schrodinger states |qi or |pi evolved in time (which would be eiHt |qi,
resp. eiHt |pi). It follows that |q, t + ti = eiHt |q, ti and hq 0 , t + t| = hq 0 , t| eiHt . Hence

hq 0 , t + t |q, ti = hq 0 , t| eiHt |q, ti = hq 0 , t| 1 iHt + O(t2 ) |q, ti
(1.3)
Now H = H(P, Q) = eiHt H(Q, P )eiHt = H(Q(t), P (t)) and we assume that H is written with all
P s to the right of all qs (by using P Q = QP i if necessary). Then one has
hq, t| H(Q(t), P (t)) |p, ti = H(q(t), p(t)hq, t |p, ti ,
(1.4)
so that
hq 0 , t + t |q, ti =
!
Y

hq 0 , t| 1 iH(Q(t), P (t))t + O(t2 ) |p, ti hp, t |q, ti
d pa
Z
=
!
Y

hq 0 , t| 1 iH(q 0 (t), p(t))t + O(t2 ) |p, ti hp, t |q, ti
d pa
Z
=
!
Y
d pa
iH(q 0 (t),p(t))t+O(t2 )
Y eipa (qa0 qa )
b
(1.5)
Now one can take a finite interval t0 t and let t = t Nt . We write tk = t + kt with k = 0, . . . N
and t0 = t, tN = t0 as well as qN = q 0 , q0 = q. Then
Z Y
0 0
a
hq , t |q, ti =
dq1a . . . dqN
1 hqN , tN |qN 1 , tN 1 i hqN 1 , tN 1 | . . . |q1 , t1 i hq1 , t1 |q0 , t0 i
(
)
Z Y
N
N
a
Y dpa
X
X
dp
1
a
=
dq1a . . . dqN
. . . N exp i
H(qk , pk )t + i
pk (qk qk1 ) . (1.6)
1
2
2
k=1
k=1
Then, for any configuration {q0 , q1 , . . . qN } define an interpolating q( ), so that qk+1 qk '
Q
Q
Q dpa Q
q(
) . Also k,a dqka ' a Dqa and k,a 2k ' a Dpa , so that finally
0
hq , t |q, ti =
qa (t)=qa , qa (t0 )=qa0
Dqa
Dpb exp i
t0
t0
d H(q( ), p( )) + i
t
d p( )q(
)
(1.7)
This can be easily generalized to yield not only transition amplitudes but also matrix elements of
products
of
operators.
Going
through
the
same
steps
again
for
hq 0 , t0 | OA (Q(tA ), P (tA ))OB (Q(tB ), P (tB )) . . . |q, ti with tA tB . . ., one easily sees that the path
integral just gets OA (q(tA ), p(tA ))OB (q(tB ), p(tB )) . . . inserted. Thus
hq 0 , t0 | T {OA (Q(tA ), P (tA ))OB (Q(tB ), P (tB )) . . .} |q, ti
Z
Y
Y
=
Dqa
Dpb OA (q(tA ), p(tA )) OB (q(tB ), p(tB )) . . .
qa (t)=qa , qa (t0 )=qa0
b
t0
exp i
Z
d H(q( ), p( )) + i
1.2
t0
)
d p( )q(
)
(1.8)
Functional integral in quantum field theory
An advantage of the canonical formalism is that unitarity is manifest, but Lorentz invariance is
somewhat obscured (although guaranteed by general theorems). In the functional integral formalism
with covariant Lagrangians to be discussed next, Lorentz invariance is manifest, but unitarity is not
guaranteed, unless the formalism can be derived from the canonical one (and then extra terms might
be present).
1.2.1
Derivation of the Hamiltonian functional integral
The path integral formula for matrix elements in quantum mechanics immediately generalizes at
least formally to quantum field theory by the obvious generalizations of the labels a to include the
position in space:
X
XZ
d3 x ,
etc.
(1.9)
a (n, ~x) ,
a
However, in field theory we do not want to compute transition amplitudes between eigenstates |(~x)i
of the field operator (~x)(the analogue of Q) but between in and out states having definite numbers
of particles, or often simply between the in and out vacuum states. In order to obtain these one
has to multiply the transition amplitudes obtained from generalizing (1.7) to field theory by the
appropriate vacuum wave functions which for a real scalar e.g. are

Z
Z
q
1
d3 p i~p(~x~y)
2
3
3
2
h(~x), |vac, i = N exp
p~ + m (~x)(~y ) .
(1.10)
d xd y
e
2
(2)3
Note that, contrary to the exponentials appearing in the transition amplitudes or matrix elements,
the exponential in (1.10) is real. Note also that it only contains 3-dimensional space integrals (if it
p
R
were not for the p~ 2 + m2 the whole expression would collapse to a single d3 x integral), and in this
sense it is infinitesimal as compared to the 4-dimensional space-time integrals in the exponents of the
transition amplitudes or matrix elements. Hence we are let to expect that the effect of multiplying
with (1.10) is only to add terms of the form i (infinitesimally small) to the exponent. It can
indeed be shown that they precisely provide the correct i terms that result in the correct Feynman
propagator. Again, this was to be expected since this must be the role of the initial and final
conditions imposed by h, |vac, i. Hence, one arrives at the functional integral representation for
the time-ordered product of Heisenberg picture operators between the in and out vacuum states:
n

o
hvac, out| T OA (tA , ~xA ), (tA , ~xA ) OB (tB , ~xB ), (tB , ~xB ) . . . |vac, ini
Z Y
Y

2
Dn OA (tA , ~xA ), (tA , ~xA ) OB (tB , ~xB ), (tB , ~xB ) . . .
Dl
= |N |
(l Z
exp i
"Z
#)
d3 x
0 l (, ~x)l (, ~x) H((, ~x)(, ~x)) + iterms
, (1.11)
where we have denoted the fields and their conjugate momenta as l and l while the corresponding
Heisenberg picture operators are l and l . The functional measures can be thought of as being
Y
Y
Dl =
d (l (, ~x)) , Dl =
d (l (, ~x)) .
(1.12)
,~
x
1.2.2
,~
x
Derivation of the Lagrangian version of the functional integral
In many theories the Hamiltonian is a quadratic functional of the momenta l :

Z
XZ
1X
3
3
d3 x Bn~x ()n (, ~y ) + C() ,
d x d y An,~x,m~y ()n (, ~x)m (, ~y ) +
H((, ~x)(, ~x)) =
2 n,m
n
(1.13)
with a real, symmetric, positive and non-singular kernel An,~x,m~y (). Then the functional integral
over n (, ~x) in the vacuum to vacuum amplitude is gaussian and can be performed explicitly. More
generally, if the OA only depend on the fields l and not on the l , one can also perform the Dn
-integration in (1.11). Before giving the result it is useful to recall the following remark on gaussian
integrations.
Let f (x) be a quadratic form in xi , i = 1, . . . N , i.e. f (x) = 21 xi aij xj + bi xi + c, with a real, symmetric,
positive and non-singular matrix a. Then by straightforward computation (completing the square)
Z Y
1
1 ij
dxi ef (x) = (2)N/2 (det a)1/2 e 2 bi (a ) bj c .
(1.14)
i
Now the exponent 12 bi (a1 )ij bj c is just f (x0 ) where xi0 is the value which minimizes f . Indeed, f /xi =
aij xj + bi and hence xi0 = (a1 )ij bj and f (x0 ) = c 12 bi (a1 )ij bj . This is just the statement that for a
gaussian integration the saddle-point approximation is exact. Indeed, expanding f (x) around its minimum
we have f (x) = f (x0 ) + 12 (x x0 )i aij (x x0 )j from which follows immediately
Z Y
dxi ef (x) = (2)N/2 (det a)1/2 ef (x0 ) .
(1.15)
i
We now apply this remark to the quadratic form given by

"Z
#
Z
X
d
d3 x
0 l (, ~x)l (, ~x) H((, ~x)(, ~x)) ,
(1.16)
with H given by (1.13). Note that for the second term there is a double integral d3 x d3 y but only
a single d integral. We rewrite everything as full 4-dimensional integrals by adding a ( 0 ).
Hence the corresponding kernel is An ~x,m 0 ~y () = ( 0 )An,~x,m~y (). The saddle-point value of l
R 3 P
H
.
But
evaluating
d x l 0 l l H(l , l ) at
extremizing (1.16) is the solution
l of 0 l =
l
l =
l is exactly doing the (inverse) Legendre transformation that gives back the Lagrange function:
Z
Z
X
3
dx
0 l
l (, 0 ) H(l ,
l (, 0 )) = L(l , 0 l ) d3 x L(l , l ) .
(1.17)
l
Putting everything together we find for Hamiltonians that are quadratic in the l :
n

o
hvac, out| T OA (tA , ~xA ) OB (tB , ~xB ) . . . |vac, ini
Z Y
1/2

2
= |N |
Dl Det [2iA()]
OA (tA , ~xA ) OB (tB , ~xB ) . . .
l
Z

4
exp i d x L(l (x), l (x)) + iterms .
(1.18)
A few remarks are in order:

The overall constant |N |2 drops out when computing amplitudes that do not involve vacuum bubbles, which is achieved by dividing by hvac, out |vac, ini. This is the case in particular for the
connected n-point amplitudes. Most of the times, this is implicitly understood, and we drop this
factor, as well as other overall constants. Similarly, if A is field independent, Det [2iA()] is a constant and can be dropped. Moreover, even if it is field-dependent, it can be replaced by Det [2iA()]
(Det [2iA(0)])1 , which may be easier to handle.
If A is field-dependent, e.g. An x,m y () = nm ((x)) (4) (x y) it gives a contribution to an effective
Lagrangian. To see this note that
DetA = exp [Tr log A] .
(1.19)
A is the quantum-mechanical operator whose matrix elements are

hx, n| A |y, mi = An x,m y () = nm ((x)) (4) (x y) = nm ((x)) hx |yi ,
with ((x)) an ordinary matrix-valued function. It follows that

log ((x)) nm hx |yi
hx, n| log A |y, mi =
Z
Z

4
Tr log A =
d x hx, n| log A |x, ni = d4 x tr log ((x)) hx |xi ,
(1.20)
(1.21)
where tr is an ordinary matrix trace over the indices n = m, and hx |xi = (4) (0) is to be interpreted,
R d4 p
as usual, as (2)
4 (which is divergent, of course, and has to be regularized and renormalized). Thus
Z
Z

d4 p
4
DetA = exp
d x tr log ((x)) ,
(1.22)
(2)4
which can indeed be interpreted as an additional contribution to the Lagrangian.
As just mentioned, one encounters diverging expressions and there is the need to regularize and
renormalize as will be extensively discussed later-on. Actually, the need to renormalize occurs in any
interacting theory, whether there are divergences or not. In particular, the fields that appear in the
Lagrangian in the first place are so-called
bare fields l,B . They are related to the renormalized fields
l,R by a multiplicative factor, l,B = Zl l,R . For the time being, it is understood that the fields
l are bare fields, although we do not indicate it explicitly.
In the presence of constraints, e.g. if some of the fields have vanishing canonical momentum, the
corresponding l are absent in the Hamiltonian. Integrating over these l when deriving (1.18) formally
still gives the r.h.s. of (1.18) but with the Lagrangian missing certain auxiliary fields. This can be
cured by adding in the Hamiltonian formulation a constant factor which is an integral over the auxiliary
fields. In the end one recovers (1.18) with the full Lagrangian.
Functional integrals for anticommuting fields (fermions) can be defined similarly. The relevant formula
for fermionic gaussian integrals is
Z

(1.23)
DD exp M + + = N Det M exp M 1 .
The power of the determinant is positive rather than negative because the integration variables are
anticommuting. Furthermore, it is +1 = 2 21 because the fields and are to be considered as
independent fields (just as bosonic and are considered independent). Another difference with the
bosonic case is that the Hamiltonian is not quadratic in the momenta (they are anticommuting, too),
e.g. for the Dirac field the free Hamiltonian density is H = 0 ( j j + m), where = 0 . As
a result, to pass from the Hamiltonian formalism to the Lagrangian one, one should not integrate the
but only rename = 0 . The analogue of our bosonic formula (1.18) for Dirac fields is
n
o

hvac, out| T OA (tA , ~xA ), (tA , ~xA ) OB (tB , ~xB ), (tB , ~xB ) . . . |vac, ini
Z Y

2
= |N |
Dl D l OA (tA , ~xA ), (tA , ~xA ) OB (tB , ~xB ), (tB , ~xB ) . . .
l

Z
4
.
exp i
d x L(, , ) + iterms
1.2.3
(1.24)
Propagators
The free propagators or simply propagators are defined as

ilk (x, y) = hvac, out| T (l (x)k (y)) |vac, ini |no
interactions
(1.25)
They are not to be confused with the complete propagators (denoted 0 )

i0lk (x, y) = hvac, out| T (l (x)k (y)) |vac, ini ,
(1.26)
to be discussed later on. Recall that in a free theory we do not need to distinguish between the
Heisenberg picture and interaction picture field operators l (x) and l (x). Evidently, the free propagators are determined by the free part of the action, i.e. the part of the Lagrangian density that is
quadratic in the fields. Hence, the computation of the propagators reduces to computing a Gaussian
integral. As before, we consider the bosonic case where the free action is of the form1
Z
Z
X
1
4
d x L0 =
d4 x d4 y
l (x)Dl,l0 (x, y)l0 (y) .
2
0
l,l
(1.27)
For a hermitean scalar field e.g. D(x, y) = ( + m2 i) (4) (x y). The general formula (1.18)
b = 1) gives
(with A
Z

Z Y
4
e
ilk (x, y) = N
Dl0 l (x) k (y) exp i d x L0 .
(1.28)
l0
In a free theory one has

1 = hvac, out |vac, ini |no
interactions
e
= N
Z Y

Z
4
Dl0 exp i d x L0 ,
(1.29)
l0
which allows us to rewrite

RQ
ilk (x, y) =
l0
R

Dl0 l (x) k (y) exp i d4 x L0
R

RQ
.
d4 x L0
l0 Dl0 exp i
(1.30)
Actually, in a free theory, it is not much more difficult to compute the n-point functions:
hvac, out| T (l1 (x1 ) . . . ln (xn )) |vac, ini |no interactions
R 4

RQ
d x L0
l0 Dl0 l1 (x1 ) . . . ln (xn ) exp i
R

RQ
=
4x L
0 exp i
D
d
0
l
0
l

1

...
Z0 [J]
,
(i)n
= Z0 [0]

Jl1 (x1 )
Jln (xn )
(1.31)
J=0
where
Z0 [J] =
Z Y
Z

4
l
Dl0 exp i d x L0 (x) + Jl (x) (x)
.
(1.32)
l0
(One should not confuse the generating functional Z0 [J] with the field renormalization factors Zl .)
With the quadratic L0 given by (1.27), the integral is Gaussian and one gets

Z

b 1/2
iD
i
1
4
4
d xd y Jl (x)Dlk (x, y)Jk (y)
Z0 [J] = Det
exp
2
2
Z

i
1
4
4
= Z0 [0] exp
d xd y Jl (x)Dlk (x, y)Jk (y) .
(1.33)
2
We then get for the free propagator
ilk (x, y) = (i)2
1
Z0 [0]

Z0 [J]

Jl (x) Jk (y)
1
= iDlk
(x, y) ,
(1.34)
J=0
1
As already mentioned, for the time being, our fields are bare fields. Indeed, the fact that the quadratic part of the
action equals the free action is true for the bare fields with a bare mass parameter, while for the renormalized fields
the quadratic part of the action contains the free part determining the free propagator, as well as a counterterm
part which is at least of first order in the coupling constant. This will be discussed in detail in section 2.
or
lk (x, y) = (D1 )lk (x, y) .
(1.35)
R d4 p ip(xy)
From translation invariance one has Dl,k (x, y) Dl,k (xy) = (2)
Dl,k (p) so that the inverse
4e
1
1
operator (D )lk (x, y) is given by the Fourier transform of (D )lk (p), which is the inverse matrix of
Dlk (p):
Z
d4 p ip(xy) 1
e
lk (x, y) lk (x y) =
(D )lk (p) .
(1.36)
(2)4
For
R
=
the
scalar
field
with
D(x, y)
=
( + m2
R
d4 p ip(xy)
1
eip(xy) (p2 + m2 i) this leads to (x y) = (2)
.
4e
(2)4
p2 +m2 i
i) (4) (x
y)
d4 p
1.3
Green functions, S-matrix and Feynman rules
1.3.1
Vacuum bubbles and normalization of the Green functions
1 ...ln
most important result that the n-point Green functions Gl(n)
(x1 , . . . , xn )
h
i
= hvac, out| T l1 (x1 ) . . . ln (xn ) |vac, ini (where the l are Heisenberg picture operators of the
interacting theory) are given by the sum of all Feynman diagrams with n external lines terminating
at x1 , . . . xn . We will now derive this result and at the same time obtain the Feynman rules from the
functional integral formalism.
It will be useful to consider normalized n-point Green functions (or simply n-point functions)
obtained by dividing by the 0-point function:
It
is
bl1 ...ln (x1 , . . . , xn ) =

G
(n)
h
i
hvac, out| T l1 (x1 ) . . . ln (xn ) |vac, ini
hvac, out |vac, ini
(1.37)
bB (n) , while if
Obviously, if the fields are bare fields, this is the so-called bare n-point function G
bR (n) . Since
the fields are renormalized fields, this is the so-called renormalized n-point function G
l,B = Zl l,R one simply has

" n
#
Yp
bl1 ...ln (x1 , . . . , xn ) .
bl1 ...ln (x1 , . . . , xn ) =
Z lr G
(1.38)
G
B (n)
R (n)
r=1
For the time being, we will concentrate on the bare n-point functions, although we will not indicate
it explicitly.
We use the functional integral representation of the numerator and the denominator2 in the
Lagrangian formalism and obtain
R
R
i d4 x L(x)
D
(x
)
.
.
.
(x
)
e
l
1
l
n
n
1
bl1 ...ln (x1 , . . . , xn ) =
R
R
G
.
(1.39)
4
(n)
D ei d x L(x)
2
In the following we simply write D instead of
Dl
Note that the normalization constant |N |2 has been eliminated when dividing by
Z
R 4
2
Svac,vac hvac, out |vac, ini = |N |
D ei d x L(x) .
(1.40)
In the absence of time-varying external fields Svac,vac is just a number. Contrary to a free field theory,
however, in general this number is not just 1. Recall the definition of the in and out states: |vac, ini
is the state that resembles the vacuum |0i of particles without interactions if an observation is made
at t . Recall also that the separation of H into H0 and V must be such that H and H0 have
the same spectrum. In particular, H |vac, ini = 0 and H0 |0i = 0. Hence |vac, ini cannot contain any
particles that would necessarily contribute a positive energy. We will suppose that the vacuum is
unique3 and stable, so that there are no transitions h, out |vac, ini for any 6= vac. (For a unique
vacuum, this follows from energy conservation.) Hence,
S,vac = Svac,vac ,vac .
Unitarity of the S-matrix implies
X
1=
|S,vac |2 = |Svac,vac |2
(1.41)
Svac,vac hvac, out |vac, ini = eivac .
(1.42)
It is instructive to compute Svac,vac in perturbation theory and verify that it is a pure phase. Indeed,

R
Svac,vac = h0| T exp i d4 x Hint (x) |0i, which equals 1 plus all Feynman diagrams without external
lines, cf Fig. 1. One can convince oneself that the sum of all such diagrams equals the exponential
of the connected diagrams only:
h
i
Svac,vac = exp sum of all connected vacuum-vacuum diagrams
(1.43)
In such a diagram, every propagator contributes a i, and each vertex also gives a factor i (since
Figure 1: Svac,vac is given by the sum of all vacuum bubbles which equals the exponential of the sum
of all connected vacuum bubbles.
3
In many theories with symmetries, the vacuum is degenerate. In this case the present discussion is slightly more
complicated but can be adapted accordingly.
Hint is real, but the vertex equals i times the numerical factor). Finally, each loop contributes an i
due to the Wick rotation (to be discussed below). If we let I be the number of internal lines, V the
number of vertices and L the number of loops, this yields a total factor
(i)I (i)V iL = ()V iV I+L = ()V i ,
(1.44)
where we used the diagrammatic identity

I V =L1 ,
(1.45)
valid for each connected component of a diagram. Thus, every connected vacuum-to-vacuum diagram
is purely imaginary and Svac,vac is indeed pure phase.
What is the effect of normalizing the Green functions as in (1.37), i.e. of dividing by
hvac, out |vac, ini ? Suppose the numerator in (1.37) is given by the sum of all Feynman diagrams
with n external lines (including propagators) terminating at x1 , . . . xn . This sum then corresponds to
connected and disconnected diagrams. The disconnected diagrams, in particular, contain diagrams
with vacuum-bubbles. There may be 0, 1, 2, . . . vacuum bubbles. It is easy to convince oneself that
the sum of all diagrams is the product of a) the sum of diagrams without vacuum-bubbles and of
b(n) as given by
b) 1 plus the sum of all vacuum bubbles, i.e. of Svac,vac = hvac, out |vac, ini. Thus G
(1.37) should exactly be the sum of all diagrams (connected and disconnected) with n external lines
(with their propagators) not containing any vacuum bubbles:
bl1 ...ln (x1 , . . . xn ) is given by the sum of all Feynman diagrams with n external lines
G
(n)
(with propagators) terminating at x1 , . . . xn and not containing any vacuum bubbles.
(1.46)
This is the result we will show starting from the identity (1.39). Actually, this result applies both
to the bare and the renormalized Green functions, provided one uses the Feynman rules with bare
propagators and interactions in the first case, and renormalized propagators and interactions (and
counterterms) in the second case. This will become clearer in section 2.
b(n) in a simpler-looking way. Indeed, still assuming a non-degenerate
One can also rewrite G
vacuum, |vac, ini and |vac, outi only differ by the phase factor eivac as is easily seen from (1.41) and
(1.42):
X
X
|, outi S,vac = eivac |vac, outi .
(1.47)
|vac, ini =
|, outi h, out |vac, ini =
It follows that for any operator or product of operators M one has

D E
hvac, out| M |vac, ini
e M
e
= h| M |i =
hvac, out |vac, ini
|i |vac, ini ,
E
e
|vac, outi , (1.48)
and hence
h
i
h
i
bl1 ...ln (x1 , . . . , xn ) = h| T l1 (x1 ) . . . ln (xn ) |i hT l1 (x1 ) . . . ln (xn ) ivac .
G
(n)
(1.49)
1.3.2
Generating functional of Green functions and Feynman rules
Just as we defined Z0 [J] for a free theory, eq. (1.32), the generating functional for the interacting
theory is defined by
Z
Z[J] =
Z

4
l
D exp i d x L(x) + Jl (x) (x)
.
(1.50)
Equation (1.39) can then be written as

1

n
bl1 ...ln (x1 , . . . , xn ) =
(i)
.
.
.
Z[J]
G

(n)

Z[0]
Jl1 (x1 )
Jln (xn )
(1.51)
J=0
b(n) by successive
We see that indeed Z[J], or rather Z[J]/Z[0], generates the n-point Green functions G
b(n) appear as the coefficients in the development of Z[J] in
functional derivatives. Conversely, the G
powers of the J:
Z
X
1
bl1 ...ln ((x1 , . . . , xn ) iJl1 (x1 ) . . . iJln (xn ) .
Z[J] = Z[0]
d4 x1 . . . d4 xn G
(1.52)
(n)
n!
n=0
To make the relation with the Feynman diagrams, recall that the sum of Feynman diagrams corresponds to a perturbative expansion in the coupling constant(s). So let us compute Z[J] in perturbation theory. To do so, separate

L (x), (x) = L0 (x), (x) + Lint (x), (x) ,
(1.53)
R
with the free Lagrangian L0 given by the quadratic part, cf. (1.27), and develop ei Lint in a power
series.4 Hence
Z
Z
n Z
X
iN

o
iN h
4
d x Lint (x), (x)
exp i d4 x L0 (x) + Jl (x) l (x)
Z[J] =
D
N!
N =0
Z
Z
n Z
X
o

iN h
iN
4
=
d x Lint i
, i
D exp i d4 x L0 (x) + Jl (x) l (x)
N!
J(x)
J(x)
N =0
Z
X
iN h
iN
4
=
, i
(1.54)
d x Lint i
Z0 [J] .
N!
J(x)
J(x)
N =0
Z0 [J] is the generating functional of the free theory computed before, cf. eq. (1.33) with D1 equal
to (0) :
Z

i
4
4
Z0 [J] = Z0 [0] exp
d xd y Jl (x)lk (x, y)Jk (y)
2
Z

1
4
4
= Z0 [0] exp
d xd y iJl (x) ilk (x, y) iJk (y)
.
(1.55)
2
4
As it stands, this applies to the computation of bare Green functions. To compute the renormalized Green
functions, one simply takes the corresponding L0 while including all counterterms into Lint , even the quadratic ones.
1/2
The bare and renormalized generating functionals then are the same provided one also defined JB,l = Zl
JR,l so
that JB,l B,l = JR,l R,l .
10
We see that i/J(x) acting on Z0 [J] yields a propagator i(x, y) attached to a vertex at x
(times iJ(y) and integrated over d4 y). There are as many propagators attached to a vertex at x as
there are fields in Lint (x). All propagators are attached to some vertex or to an external iJ(zi ).
Obviously, a term of given order N in (1.54) corresponds to a diagram with N vertices. It is also
not difficult to work out that the combinatorial factors accompanying a diagram are the usual ones.
Hence, Z[J] is the product of Z0 [0] and the sum of all Feynman diagrams with an arbitrary number
of external lines at the end of which are attached the factors iJ(zi ) (integrated d4 zi ).
g 4
, and
Lets look at an example. Take a hermitean scalar field with an interaction Lint = 24
compute Z[J] up to first order in g, meaning we only keep the terms of order N = 0 and N = 1 in
(1.54):
)
(
Z

R
g
4
d4 x i
exp 12 d4 xd4 y (iJl (x))(ilk (x, y))(iJk (y))
Z (g) [J] = Z0 [0] 1 i
24
J(x)
"
(
Z
Z
Z
2
4
g
4
4
d4 z (i(x, z)iJ(z)
dx
d z (i(x, z)iJ(z) + 6(i(x, x)
= Z0 [0] 1 i
24
#)

R
+3(i(x, x))2
exp 12 d4 xd4 y (iJl (x))(ilk (x, y))(iJk (y)) .
(1.56)
First, take J = 0. At order g there is only one term and:
Z
n
o
g
(g)
Z [0] = Z0 [0] 1 i
d4 x (i(x, x))2
8
(1.57)
The term of order g corresponds to a single vertex with 4 lines, joined two by two (two loops). This
is a vacuum-bubble diagram. The factor ig8 is in agreement with the usual combinatoric factor: the
vertex gives a factor ig and the symmetry factor is 12 21 12 = 81 . More generally, Z[0] is the sum
of 1 and all vacuum-bubbles.
If one first takes the derivatives (iJ(x

. . . (iJ(x
of Z[J] and only then sets J = 0, one generates
n ))
1 ))
a sum of products of propagators (i) attached either to the external xi or to internal x
ei of vertices
which are integrated. One sees that each vertex contributes i times the numerical factors in Lint ,
and the symmetry factors again are automatically generated. As explained above, this sum of all
diagrams factorizes into a sum of diagrams without vacuum bubbles and the sum of 1 plus all vacuum
bubbles. Thus, dividing by Z[0] exactly eliminates these vacuum bubbles and we have shown (1.46)
b(n) as defined by the functional integral (1.51).
for the n-point Green functions G
g 4
Let us come back to the example of the scalar theory with Lint = 24
. Here, we get for the
11
4-point function up to order g:

Z (g) [J]
...

(iJ)(x1 )
(iJ)(x4 ) Z (g) [0]
J=0
"
Z
g
d4 x 24(i(x, x1 ))(i(x, x2 ))(i(x, x3 ))(i(x, x4 ))
= i
24
#
+6(i(x, x)) 2(i(x, x1 ))(i(x, x2 ))(i(x3 , x4 )) + 5 permutations . (1.58)
The two terms correspond to the two diagrams shown in Figure 2.
Figure 2: Diagrams corresponding to (1.58).

Loop-counting : It is sometimes convenient to introduce a loop-counting parameter by replacing
the action S 1 S and J 1 J. This multiplies all vertices by 1 and all propagators by . Each
external line also gets a factor 1 from the J . Thus external lines get a net factor 0 , and the overall
factor of a diagram is IV = LC , where I is the number of internal lines, V the number of
vertices, L the number of loops and C the number of connected components of the diagram and we
used (1.45). Thus for fixed C, is a loop-counting parameter. In particular, a connected diagram is
R
accompanied by a factor L1 . Note that the exponent in the functional integral is ~i (S + J) if one
does not use units where ~ = 1. One sees that ~ is a loop-counting parameter, and the limit ~ 0
isolates the diagrams with L = 0, i.e. tree diagrams. In this sense, tree amplitudes are referred to as
classical, while loop corrections are quantum corrections.
1.3.3
Generating functional of connected Green functions
b(n) (x1 , . . . , xn ) without vacuum-bubbles contain the imporThe n-point (n > 0) Green functions G
tant subclass of connected n-point Green functions GC
(n) (x1 , . . . , xn ). They can be defined by an
C
b(1) (x) and then GC (x1 , x2 ) = G
b(2) (x1 , x2 )
algebraic recursion relation: by definition G (x) = G
(1)
(2)
C
G(n) (x1 , . . . , xn )
C
GC
(1) (x1 )G(1) (x2 ),
etc. One can show that this is equivalent to

being the sum of the
corresponding connected Feynman diagrams. The algebraic recursion relation is best summarized
as a relation between generating functionals. Let iW [J] be the generating functional of connected
12
Green functions, (cf. Fig. 3)

Z
X
1
l1 ...ln
iW [J] = iW [0] +
d4 x1 . . . d4 xn GC,
(x1 , . . . , xn ) iJl1 (x1 ) . . . iJln (xn ) .
(n)
n!
n=1
(1.59)
Figure 3: W [J] is the generating functional of connected Green functions.

We separated the part iW [0] which corresponds to connected 0 point Green function, i.e. to connected
vacuum-bubbles. Note that for n 1, the GC
n cannot contain vacuum-bubbles. As one sees from
Fig. 3 or the definition (1.59), the connected full propagator is given by

0
C
iC (x, y) G(2) (x, y) = i
W [J]
i W (2) (x, y) .
(1.60)
J(x) J(y)
J=0

2
Consider now exp iW [0] = 1 + iW [0] + 21 W [0] + . . .. Here, iW [0] contains all vacuum-bubbles
2
with a single connected component, while 12 W [0] contains all vacuum-bubble diagrams with two
connected components (the factor 12 is the appropriate symmetry factor for those diagrams having
two identical components, while it is compensated by a factor 2 for the product of two different

components), etc. Hence, exp iW [0] is the sum of 1 and all possible vacuum-bubble diagrams,

connected or not, i.e. it equals Z[0]. In the same way one sees that exp iW [J] equals 1 plus the
sum of all diagrams, connected or not, i.e. Z[J] :

Z[J] = exp iW [J] .
(1.61)
Lets look at the example of connected 1- and 2-point functions. As already noted, the 1-point
function without vacuum-bubbles is necessarily connected:
b(1) (x) = GC
G
(1) (x) .
(1.62)
Next, the relation (1.61) indeed leads to the correct relation between the 2-point functions (without
b(2) and the connected 1- and 2-point functions GC and GC :
vacuum-bubbles) G
(1)
(2)

GC
iW [J]
=
log Z[J]
(2) (x, y) =
(iJ)(x) (iJ)(y)
(iJ)(x) (iJ)(y)
J=0
J=0

1
= Z[J] (iJ)(x) (iJ)(y) Z[J]

Z[J] (iJ)(x) Z[J]
Z[J]
Z[J] (iJ)(y)
J=0
J=0
b(2) (x, y) G
b(1) (x) G
b(1) (y)
= G
C
b(2) (x, y) GC
= G
(1) (x) G(1) (y) .
J=0
(1.63)
13
Loop-counting : If one introduces the loop-counting parameter as before, one also has W [J] =
P L1
WL [J], where WL [J] is the L-loop contribution to W [J]. In the limit 0 one isoL=0
lates the contributions of the tree-diagrams. On the other hand, in this limit, one can evaluate
the functional integral in a saddle-point approximation (stationary phase) and then the integral is
S
dominated by those J that solve
l + Jl = 0. It follows that
Z
W0 [J] = S[J ] +
d4 xJl (x)Jl (x) ,
(1.64)
i.e the tree contribution W0 [J] is the (inverse) Legendre transform of the classical action.
1.3.4
Relation between Green functions and S-matrix
The basic quantity in particle physics is the S-matrix from which measurable transition rates like
cross-sections and life-times can be extracted. The S-matrix elements are defined as
S = h, out |, ini ,
(1.65)
and give the transition amplitudes between the in-states |, ini and the out-states |, outi. Here,
and are short-hand for a complete collection of momenta pi , helicities i and (anti)particle
types ni describing the state. Recall that the in-state |, ini |p1 , 1 , n1 ; p2 , 2 , n2 ; . . . ini is a
(time-independent Heisenberg-picture) state that looks, if an observation is made at t , as
a collection of non-interacting particles with momenta pi , helicities i and of type ni . A similar
definition holds for the out-states with t +.
To relate the S-matrix elements to the Green-functions, we first define the Fourier transform of
the latter as
Z
Pn
l1 ...ln
1 ...ln
e
G(n) (p1 , . . . pn ) = d4 x1 . . . d4 xn ei i=1 pi xi Gl(n)
(x1 , . . . xn ) ,
(1.66)
with all momenta pi considered as entering the diagram. These momenta are off-shell and are those
of the propagators associated with the external lines. S-matrix elements are computed between onshell external states, i.e. precisely at those values of the momenta where the external propagators
of the Green functions have poles. We will see in the next sections, that loop-corrections to the
free propagators shift the pole from p2 = m2B (mB is the bare mass entering the Lagrangian) to
p2 = m2 , where m must be interpreted as the physical mass. Thus the full propagators have poles
at p2 = m2 . To get a finite result for on-shell external states, one obviously has to remove the full
external propagators. This can be done by multiplying with the inverse full propagators i(0 )1 .
The result is called the amputated n-point Green function, cf. Fig. 4.
" n
#
Y
l
...l
0
1
e 1 n (p1 , . . . pn ) =
el1 ...ln (p1 , . . . pn ) .
G
i( ) (pj ) G
(1.67)
(n,amp)
(n)
j=1
Again, this definition holds with all Green functions and full propagators being the bare or renormalized ones.
14
p3
p3
p2
pr
pr
p2
p1
p1
Figure 4: n-point Green function (left) and corresponding amputated n-point Green function (right)
It can be shown that the S-matrix elements are obtained from the on-shell amputated renormalized Green functions simply by multiplication with the appropriate wave-functions of the initial
and final (anti)particles. More precicely, to obtain the S-matrix element with r (anti)particules in
the initial state and n r in the final state : (i) take the corresponding amputated renormalized npoint Green function (with for any initial particle or final antiparticle and for any final particle
or initial antiparticle), (ii) take the pi on-shell for the initial (anti)particles, and similarly the pj
i ,i )
on-shell for the final (anti)particles, (iii) multiply with the appropriate wave-function factors u(p
(2)3/2
etc., that enter in the expansions of the corresponding free fields. Thus
Sp01 ,10 ,n01 ,...;p1 ,1 ,n1 ,...
"nr 0 0
#" r
#
Y ulj0 (pj , j )/vlj0 (p0j , j0 ) Y
uli (pi , i )/vli (pi , i )
=
(2)3/2
(2)3/2
j=1
i=1
(1.68)
l1 ...lr l0 ...l0
nr
1
e
G
(p1 , . . . pr , p01 , . . . p0nr ) .
R (n,amp)
It follows from (1.38) that 0B = Z 0R and, combining with the definition of the amputated Green
function (1.67) one immediately sees that
#
" n
Y 1/2
el1 ...ln (p1 , . . . pn ) .
el1 ...ln (p1 , . . . pn ) =
G
(1.69)
G
Z lj
R (n,amp)
B (n,amp)
j=1
Thus we can rewrite the relation between the S-matrix elements and the amputated Green functions
in terms of the bare amputated Green functions as
Sp01 ,10 ,n01 ,...;p1 ,1 ,n1 ,...
r
h nr
ih Y
i
Y ulj0 (p0j , j0 )/vlj0 (p0j , j0 ) q
uli (pi , i )/vli (pi , i ) p
=
Zlj0
Z li
(2)3/2
(2)3/2
j=1
i=1
0
(1.70)
el1 ...lr l1 ...lnr (p1 , . . . pr , p01 , . . . p0nr ) .

G
B (n,amp)
It is in this second form that the relation, first derived by Lehman, Symanzik and Zimmermann, is
usually referred to as LSZ reduction formula. However, (1.68) has the advantage of expressing the
finite S-matrix elements solely in terms of renormalized quantities that have a finite limit as the
regularization is removed.
15
1.4
1.4.1
Quantum effective action

Legendre transform and definition of []
We already defined the generating functional Z[J] of Green functions and the generating functional
W [J] of connected Green functions. They correspond to the sum of all Feynman diagrams and of
connected diagrams only. Connected diagrams are more basic since all diagrams can be constructed
from them. The algebraic relation was simply Z[J] = eiW [J] . Here we will define yet another
generating functional [] that generates an even smaller subclass of connected diagrams, namely
the one-particle-irreducible diagrams, or 1PI for short. A 1PI diagram is a connected diagram that
does not become disconnected by cutting a single line. (There is a slight subtlety with this definition
for the 1PI 2-point diagram to be discussed below.) Since a tree diagram becomes disconnected by
cutting a single line, tree diagrams are not 1PI. A one-loop diagram with the external propagators
removed always is 1PI. Higher-loop diagrams may or may not be 1PI. For n 3, a 1PI n-point
diagram is also called an n-point proper vertex.
The functional [] is defined as the Legendre transform of W [J]. First, let

1
W [J] = i
log Z[J] =
i
Z[J] .
(1.71)
rJ (x)
Jr (x)
Jr (x)
Z[J]
Jr (x)
The expression on the r.h.s. is similar to the one-point Green function without vacuum bubbles
b r (x) GC (x) except that we have not set J = 0.
(which is the connected one-point function) G
(1)
(1), r
Not setting J = 0 amounts to keeping the additional interaction terms r Jr in the Lagrangian. Thus
rJ (x) is the connected one-point function in the presence of the additional interactions generated by
the sources. This is also called the vaccum expectation value of the corresponding Heisenberg field
r in the presence of the sources J:
rJ (x) = hr (x)ivac, J h| r (x) |iJ
hvac, out| r (x) |vac, iniJ

.
hvac, out |vac, iniJ
(1.72)
One can invert the relation rJ (x) = Jr(x) W [J] to get Jr (x) as a function of r (x). More precisely, for
every (c-number) function r (x), we let jr (x) be the (c-number) function such that rJ (x) = r (r) if
Jr (x) = jr (x), i.e. jr (x) is the current such that the vacuum expectation value of r equals r (x).
We can now use as variable5 and define the Legendre transform of W as
Z
[] = W [j ] d4 x r (x)jr (x) .
(1.73)
is called the quantum effective action. Let us show why: one has
Z
Z
jr (x)
4 jr (x) W [j ]
[] = d x
js (y) d4 x r (x) s
s
s
(y)
(y) jr (x)
(y)
Now

W [j ]
W [J]
=

jr (x)
Jr (x)
1/2
Since JB,s = Zs
J r = j r
JR,s one obviously has (sJ )B =

= rJ (x)
J r = j r
= r (x) ,
Zs (sJ )R and thus also sB =
16
(1.74)
(1.75)
Zs sR .
so that the first and third terms in (1.74) exactly cancel. Hence,
s (y)
[] = js (y) .
(1.76)
[]
Suppose that for a given function one has
s (y) = 0, i.e. the corresponding jr vanishes. This
means that the vacuum expectation values of the r (x), in the absence of any current, equal r (x).
[]
Conversely, the vacuum expectation values of r , for vanishing current, must be solutions of
s (y) =
0, i.e. be stationary points of []. This shows that can indeed be interpreted as some quantum
action.
Note that the preceding careful discussion usually is simply summarized as
W
= r
Jr
= Jr
r
Z
,
d4 x r (x)Jr (x)
[] = W [J]
(1.77)
,
, etc remain valid for
Note also that all these manipulations involving functional derivatives J
fermionic fields and sources, provided one correctly uses left or right derivatives, paying attention to
RW
L
the order of the fields. Thus one should define e.g J
= r and
r = Jr
r
1.4.2
[] as quantum effective action and generating functional of 1PI-diagrams
The interpretation of [] as quantum effective action is confirmed further if we recall that in the
classical limit, i.e. at tree-level, W [J] is just the inverse Legendre transform of the classical action, cf.
(1.64). Since [] is the Legendre transform of W [J], it follows that, in the classical limit, [] just
is the classical action. Thus [] equals the classical action S[] plus quantum-, i.e. loop-corrections.
Actually, in a sense, [] captures all loop effects since one has the following property:
One may compute iW [J] as a sum of connected tree diagrams with vertices and propagators
determined as if the action were [] rather than S[].
To prove this, let us proceed as for the loop-counting above: we compute the generating functional
of connected Green functions W [J, ] using as action [] and having divided and J by :
Z
o
n
o Z
ni
exp iW [J, ] = D exp
[] + d4 x r (x)Jr (x)
.
(1.78)
If one does a perturbative (Feynman diagram) expansion of W [J, ], the propagators are given by
the inverse of the quadratic piece in and hence contribute a factor , while every vertex gets a
factor 1 as does an external line. This yields an overall factor IV = L1 where L is the number
of loops. Thus the loop-expansion of W [J, ] reads
W [J, ] =
(L)
L1 W [J, = 1] .
(1.79)
L=0
17

(0)
One isolates the tree graphs (L = 0) by taking the limit 0 : lim0 W [J, ] = W [J, = 1].
(0)
(0)
But iW [J, = 1] = iW [J] is the sum of connected tree diagrams computed as if the action were
[]. On the other hand, in the limit 0, one can use the stationary phase (saddle point) to
evaluate (1.78) and get

Z
o
ni
o
ni

(0)
4
r
= J . (1.80)
W [J] exp
[J ] + d x J (x)Jr (x)
where
exp

= J
There is some constant of proportionality which has some finite limit as 0 and which contributes
(0)
an order 0 piece to the exponent, but nothing at order 1 . We see that W is the (inverse) Legendre
transform of . On the other hand, the (inverse) Legendre transform of is the ordinary W [J]. We
conclude that
(0)
W [J] = W ,
(1.81)
and the full generating functional of connected Green functions is indeed given as a sum of connected
tree diagrams computed with propagators and vertices taken from the effective action .
If we let
Z
X
1
r1
rn
[] =
d4 x1 . . . d4 xn (n)
(1.82)
r1 ...rn (x1 , . . . , xn ) (x1 ) . . . (xn ) ,
n!
n=0
the (n) for n 3 are the so-called proper vertices, and the complete (connected) propagators

(2) 1
(x, y). This can also be seen more formally
GC
(2) (x, y) are given (cf. (1.27) and (1.35)) by i
as
W [J] = i
s (y)
Jr (x) Js (y)
Jr (x) J
[] = r js (y) .
(2)
r,s (x, y) =
r
s
(x) (y)
(x)
r,s
GC,
(2) (x, y) = i
(1.83)
It follows that
0
(2)
GC
(2) i = i
1
(1.84)
Since an arbitrary connected diagram is obtained once and only once as a tree diagram using these
complete propagators and proper vertices, the proper vertices must be one-particle irreducible (1PI)
amputated n-point functions:
[] is the generating functional of one-particle irreducible (1PI) diagrams.
As an example, consider a hermitean scalar field. The full propagator is of the form i0 (p) =
1
so that (2) (p) = p2 m2 + (p). Clearly, p2 m2 is the contribution
i p2 + m2 (p)
from the quadratic part of the classical action and contains the loop-contributions.
A few remarks:
18
In later sections, we will be much concerned with possible divergences occurring in loopdiagrams and their cancellation by counterterms. Since a tree diagram is never divergent
if the vertices and propagators are finite, it is clear that any diagram will be finite if the (n)
are. Hence the issue of renormalisation can be entirely discussed at the level of the (n) . More
precisely, one can expand in powers of the bare sB or of the renormalized sR related by the
same relation as the fields B and R , namely
sB =
p
Zs sR ,
(1.85)
implying
"
(n)
r1 ...rn (x1 , . . . , xn )
n
Y
#
Zr1/2
j
(n)
r1 ...rn (x1 , . . . , xn )
(1.86)
j=1
(n)
(n)
The B and R are called the bare and renormalized n-point vertex functions. The vertex
(n)
functions that should be finite after removing the regularization are the R .
Quite often one encounters a somewhat different notion of effective action: in a theory with
two sorts of fields, say and , one might only be interested in Green functions of one sort
of fields, say the . This happens in particular if the other sort corresponds to very heavy
particles that do not appear as asymptotic states in a scattering experiment, though they still
do contribute to intermediate loops. Let S[, ] = S1 [] + S2 [] + S12 [, ]. We only introduce
sources J for the and define
Z
Z
Z
n
o Z
n
o
r
r
f
Z[J] = DD exp i S[, ]+ Jr
= D exp i S1 []+W []+ Jr
, (1.87)
where
n
o Z
n
o
f
exp iW [] = D exp i S2 [] + S12 [, ]
.
(1.88)
f [] is referred to as the effective

Then, for reasons that are obvious from (1.87), S1 [] + W
f [] still
action for the field obtained after integrating out the field . Note that often W
allows to obtain certain Green function of the -field. Suppose e.g. that the coupling between
f []
the two sorts of fields is S12 [, ] F(). Then, by taking functional derivatives of W
with respect to one generates vacuum expectation values of time-ordered products of the
F(). A standard example is spinor quantum electrodynamics with playing the role of the
fermions and of the gauge field. It is relatively easy to integrate out the fermions since they
only appear quadratically in the action. This yields a determinant which can be exponentiated
f and is interpreted as a single fermion loop with arbitrarily many gauge fields attached.
into W
There is a different, sometimes more direct way to compute the quantum effective action []:
Z

exp i[] =
D exp iS[ + ] ,
(1.89)
1PI only
19
where the subscript 1PI instructs us to keep only 1PI diagrams in a perturbative evaluation
of the functional integral. To see why this equation is correct, it is best to look at an example.
R
2
g 4
Consider a scalar 4 -theory with S[] = ( 12 ( )2 m2 2 24
). Then
Z
Z
Z
g
1
g
g
g
S[+] = S[]+ ( 2 m2 3 )
(()2 +m2 2 ) ( 2 2 + 3 + 4 ) . (1.90)
6
2
4
6
24
If one computes the functional integral (1.89) in perturbation theory one sees that (i) S[] can
be taken in front of the integral, (ii) the (free) -propagator is the same as before, (iii) one now
has vertices with one, two, three and four -lines attached. However, the vertices with only one
line attached cannot lead to 1PI-diagrams and we can drop the term linear in . Thus only the
interactions quadratic, cubic and quartic in remain and they exactly generate all diagrams
where at every vertex one has either two external and two internal -lines, or one external
and three internal -lines or only four internal -lines. With the restriction to 1PI diagrams
only, the perturbation theory will exactly yield the generating functional of all 1PI diagrams,

connected or not, i.e. exp i[] . It should also be clear that the cubic and higher terms
in only contribute to two- and higher-loop 1PI diagrams. Thus if we are only interested in
the one-loop approximation to [] it is enough to keep only the part of the interactions that
is quadratic in . On the other hand, this quadratic part cannot generate any contributions
beyond one loop and the latter are necessarily 1PI. We have in general:
Z

Z
2 S[]
i
4
4
i1loop []
iS[]
d xd y [x]
(y)
e
= e
D exp
2
(x)(y)

1/2
2 S[]
iS[]
= e
,
Det
(1.91)
(x)(y)
with the power of the determinant depending on whether is bosonic or fermionic.
1.4.3
Symmetries and Slavnov-Taylor identities
Symmetries of the classical action lead, via Noethers theorem, to conserved currents, at least classically, and in many cases also at the quantum level. Since the quantum effective action equals the
classical action plus quantum corrections, one might expect that the former share the symmetries of
the latter. We will show when this is indeed the case.
Suppose that under the infinitesimal transformation
0
r (x) r r (x) + F r (x, )
(1.92)
the action and the functional integral measure are invariant:

Y
Y
0
S[0 ] = S[] , D0
D r = D
Dr .
(1.93)
One then has (suppressing the indices r)

Z
Z
Z
R
R
R
iS[]+i J
0 iS[0 ]+i 0 J
Z[J] =
D e
= D e
= D eiS[]+i (+F )J
Z
Z
Z
Z
R
R
iS[]+i J
=
D e
(1 + i F J) = Z[J] + i D
F J eiS[]+i J ,
20
(1.94)
where we first renamed the integration variable from to 0 and then identified 0 with the transformed field (1.92). Hence,
Z
Z
Z
R
4
r
iS[]+i J
d x F (x, )Jr (x) e
= Z[J] d4 x hF r (x)iJ Jr (x) ,
(1.95)
0 = D
r
r
for every J. Recall that
r = J,r with J,r such that h iJ = . Choosing Jr in (1.95) to equal
this J,r , we can rewrite (1.95) as
d4 x hF r (x)iJ
=0.
r (x)
(1.96)
This identity is called Slavnov-Taylor identity. It states that [] is invariant under r r +

hF r (x)iJ . In general, for a non-linear transformation, hF r (x)iJ is different from F r (x, ), and the
symmetry of the quantum effective action is different (quantum-corrected) from the symmetry of
the classical action. For a linear classical symmetry, one can go further. Suppose now that F r (x, ) =
R
R
R
f r (x) + d4 y trs s (y). Then hF r (x)iJ = f r + d4 y trs hs (y)iJ = f r + d4 y trs s (y) F r (x, ).
0
In this case, (1.96) states that [] is invariant under r (x) r r (x) + F r (x, ) :
If the action and measure are invariant under a linear field transformation,
then so is the quantum effective action [].
1.5
Functional integral formulation of QED
We will now apply the functional integral formalism to the particularly important example of quantum electrodynamics. We will consider Lagrangian densities of the form
1
L = F F + J A + Lmatter (l , l ) ,
4
(1.97)
where J is a conserved matter current (i.e. J = 0 by the classical Euler-Lagrange equations).

Lagrangians of the form (1.97) include in particular those of spinor electrodynamics, which describe
the coupling of a charged spin 12 Dirac field to the electromagnetic fields.6
1.5.1
Coulomb gauge
Due to the gauge symmetry A A + and l eiql l , the naive canonical formalism
does not apply. In particular, the canonical momenta are = F 0 and obviously then 0 = 0 :
A0 has a vanishing canonical momentum. A vanishing momentum is a constraint on the canonical
variables. One has to distinguish so-called first class and second class constraints. The first class
constraints always correspond to a local (gauge) symmetry and can be eliminated by a gauge choice.
6
They do not include scalar electrodynamics though, which has couplings A A or, equivalently, in which
case the current J depends on A .
21
Possibly remaining second class constraints can be dealt with either by Dirac quantization or by the
functional integral formalism in the way we will see now.
~ A
~ = 0. This fixes A0 in terms of J 0 and thus eliminates A0 and
We adopt the Coulomb gauge
0 as canonical pair, hence eliminates the first class constraints. It leaves as second class constraint
~
~ = 0,
the Coulomb gauge condition itself and a corresponding condition on the momenta:
where ,j = j j A0 = A j . Upon working out the Hamiltonian one finds that it is given by

Z
1
1
l
3
2
2
~
~ , , Pl ) =
~ + (
~ A)
~ J~ A
~ + VCoulomb + Hmatter (l , Pl ) ,
H(A,
dx
2 2
Z
Z
0
x)J 0 (t, ~y )
1
3
3 J (t, ~
3 1 0 0
d xd y
VCoulomb =
dx J A =
,
(1.98)
2
2
4|~x ~y |
where Hmatter is the part of the Hamiltonian that does not depend on the gauge field or the .
Our starting point for the functional integral formulation is the Hamiltonian formalism. The two
~ A
~ = 0 and
~
~ = 0 will be enforced by inserting the factors Q (
~ ~a) and
constraints
~
x
Q
~ ) inside the integral. (We write a instead of A and ~ instead of
~ for the integration
~
x ( ~
variables.) To simplify the discussion suppose the matter Hamiltonian Hmatter is quadratic in the Pl
(with a constant matrix Alk ) and that the operators OAj do not depend on the Pl , so that they can
be straightforwardly integrated out. Hence7

hT OA OB . . . ivac
Z
Y
Y
Y
~ ~a)
~ ~ ) OA OB . . .
Dl
(
(
= D~a D~
~
x
lZ

~x
Z
1 2 1 ~
2
4
~
exp i d x ~ 0~a ~ ( ~a) + j ~a + Lmatter i dtVCoulomb .
2
2
To appreciate the role of the (. . .), recall the formula (f (x)) =
1
a |f 0 (xa )| (x
(1.99)
xa ) with the xa being
1
(N ) (xi xi ) with
the solutions of f (x) = 0. For N variables
this reads
=
a
a | det J(xa )|

Q
Q
1
1
i
i
j
~
Jj = f /x . Thus, ~x ( ~a) = |Det3 | ~x a3 + 3 (1 a1 + 2 a2 ) , and we see that imposing the
xi ,
(N ) (f i (x))
Coulomb gauge amounts to eliminating the functional integration over one out of the 3 fields aj (t, ~x), as
Q
~ ) eliminates the integration over the corresponding canonically
expected. Similarly, the insertion of ~x (~
conjugate momentum.
It is often useful to rewrite a functional as a functional integral over an auxiliary field, e.g.
Z

Z
Y
4
~ ~ ) = Df exp i d x f (x)
~ ~ (x) .
(
(1.100)
~
x
We will also suppose that the operators O do not depend on the ~ . Then the only part in (1.99)
ivac
is just a (constant)
Above we denoted h(. . .)ivac = h| (. . .) |i = hvac,out|(...)|vac,ini
hvac,out|vac,ini . Since hvac, out |vac, ini = e
phase, we will drop it together with other constants and simply write h(. . .)ivac instead of hvac, out| (. . .) |vac, ini, with
the understanding that overall constants are either unimportant or should be fixed in the end by dividing by the same
expression without the operators OA OB . . ..
7
22
that does depend on the ~ is

Z
Z
Z

Z
Y
1 2
4
4
~
~ ~ + ~ 0~a
D~
( ~ ) exp i d x ~ 0~a ~
= D~ Df exp i d x f
2
~
x
Z
Z

2
4 1
~
0~a f
=
Df exp i d x
2
Z
Z
Z

1 ~2
4 1
2
4
~
= exp i d x (0~a)
,
Df exp i d x f f + f 0~a
2
2
1 2
~
2

(1.101)
up to an irrelevant overall multiplicative constant which we do not write explicitly. This expression
~ ~a = 0 which
(1.101) is to be inserted into the remaining integral. But then ~a is constrained by
~ 0~a = 0 and the term f
~ 0~a in the exponent in the third line of (1.101) does not
implies
contribute. The integral over Df then only gives another irrelevant constant. We arrive at
Z
Y
Y

~ ~a) OA OB . . .
(
hT OA OB . . . ivac = D~a
Dl
l
~
x
Z

Z
1
1 ~
4
2
2
exp i d x (0~a) ( ~a) + ~j ~a + Lmatter i dtVCoulomb .
2
2
1.5.2
(1.102)
Lorentz invariant functional integral formulation and -gauges
Let us rewrite this functional integral in a manifestly Lorentz invariant form. First note that

Z

Z
Z
1 ~ 0 2
4
0 0
4 1 0 ~ 2 1 0
Da0 exp i d x a j + (a )
= exp i d x j ( ) j
2
2

Z

Z
Z
0
0
x)j (t, ~y )
1
3
3 j (t, ~
d xd y
= exp i dt VCoulomb . (1.103)
= exp i dt
2
4|~x ~y |
Furthermore,
1
1
1
f f = a a + a a
4
2
2
1
1
1
1
1
1
=
0 ai 0 ai i aj i aj + i a0 i a0 + i aj j ai 0 ai i a0 i a0 0 ai
2
2
2
2
2
2
1
1
1
~ (0~a) .
~ 0~aa0 ) + a0
~ 0 )2 (
~ ~a)2 (
=
(0~a)2 + (a
(1.104)
2
2
2
The last term of the last line vanishes due to the constraint, and the next to last term is a total
R

derivative. Inserting the expression (1.103) of exp i dt VCoulomb into (1.102) one gets
Z Y
Y
Y

~ ~a) OA OB . . .
hT OA OB . . . ivac =
Da
Dl
(
~
x
Z

1 ~
1
1
4
2
2
0
0
0
2
~ ) + Lmatter
exp i d x (0~a) ( ~a) + ~j ~a a j + (a
2
2
2
Z Y
n
Y
Y
o
~ ~a) OA OB . . . exp i S[a , l ] ,
=
Da
Dl
(
(1.105)
~
x
23
with

1
S[a , l ] = d x L[a , l ] = d x f f + j a + Lmatter .

(1.106)
4
Q
~ ~a) which fixes the
Now everything is manifestly Lorentz invariant except the insertion ~x (
gauge. Let us now suppose that not only the action S[a , l ] is gauge invariant but also the operators

H
dx A (x) . Moreover, we will assume that
OA OB . . ., e.g.8 O1 (x) = F (x)F (x) or O2 = exp
Q
Q
the product of the mesures Da et l Dl is gauge invariant. One can show rather easily that
Q
Da is gauge invariant, but the invariance of Dl is not always warranted. As we will see later
on, in the presence of chiral fermions, this measure generally is not invariant and one has an anomaly.
Different chiral fermions contribute additively to the anomaly and, in a consistent theory, the sum
Q
of all anomalous contributions must vanish so that l Dl indeed is gauge invariant. With these
Q
~ a). Recall
assumptions, the only gauge non-invariant term in (1.105) is the gauge-fixing term ~x (~
that the gauge transformations act as
Z
a a = a + ,
l l, = eiql l ,
(1.107)
with = (x) completely arbitrary. It could even depend on the a themselves.9

One can rewrite the functional integral (1.105) by first changing the names of the integration
variables from a and l to a and l, , then identifying the latter with the gauge transformed
fields (1.107). The gauge invariance of the action and the operators O gives
Z Y
Y
Y

~ ~a ) OA OB . . . exp {iS[a , l ]} .
(
(1.108)
Da
Dl,
~
x
Since the -dependence came about by a simple change of integration variables, we know that the
expression on the r.h.s. actually does not depend on , whatever this function may be. Let us choose
Z
0
e ~x) d3 y 0 a (t, ~y ) ,
(t, ~x) = (t,
(1.109)
4|~y ~x|
with an a independent .
Let us check what happens to the measure
has
Da
under this field-dependent gauge transformation. One
0 a0 (t, ~y )
e
a (t, ~x) = a (t, ~x) + (t, ~x) d3 y
x
4|~y ~x|

0 0
Z
a (t , ~y )
3
0
0
e
(t t )
= a (t, ~x) + (t, ~x) + d y dt
,
0
x
t
4|~y ~x|
(1.110)
The definition of composite operators like F (x)F (x) requires

some

normal order type prescription preserving
the gauge invariance. In practice,
one
most
often
computes
hT
O
O
.
.
.
ivac with OAi that are not gauge invariant,
A B

as e.g. the propagator hT A (x)A (y) ivac . Nevertheless, such gauge non-invariant quantities should only appear at
an intermediate stage, and the final result should be gauge invariant.
9
A familiar example of depending on a is the transformation that allows oneself to go to a given gauge, e.g.
R 3 ~
~ (t,~
y)
1
(t, ~x) = 4
d y |~xa~
y | to go to Coulomb gauge.
24
so that
a0 (t, ~x)
a0 (t0 , ~y )
ai (t, ~x)
a0 (t0 , ~y )
1
t t0 (t t0 )
4|~y ~x|
a (t, ~x)
1
,
t0 (t t0 )
= i (4) (x y) ,
i
x 4|~y ~x|
ai (t0 , ~y )
= (4) (x y) +
=
(1.111)
resulting in a non-trivial Jacobian.

Y
Da =

Da Det
(4)

1
00
0
(x y)
(t t ) .
4|~y ~x|
(1.112)
e
Although non-trivial, this Jacobian only contributes an irrelevant field- and -independent
constant to the
Q
Q
functional integral (which we drop as usual). Similarly, in the absence of anomalies, l Dl, = l Dl .
Thus the only effect of this gauge transformation with is

~ ~a ) = (
~ ~a +
~ 2
e + 0 a0 ) = ( a +
~ 2 )
e ,
(
(1.113)
which allows to write (1.108) as

Z Y
Y
Y

~ 2 )
e OA OB . . . exp {iS[a , l ] + iterms} .
( a +
Da
Dl
~
x
(1.114)
e
By construction,
both isides of this equation are independent
can multiply both sides by
of . We
h
1
R
4
2
2
2
~ )
e
e = Det
~
~ 2 ).
e On the l.h.s. this
(with > 0) and integrate D
D(
exp i 2 d x (
results in yet another irrelevant constant factor. Interchanging the order of integrations on the r.h.s.,
we finally arrive at
Z Y
Da
Dl OA OB . . . exp {iSeff [a , l ]} ,
(1.115)
with
Seff [a , l ] = S[a , l ]
2
d4 x ( a )2 ,
(1.116)
where the parameter is often called the gauge parameter. Starting from the manifestly unitary canonical formalism in Coulomb gauge, we have obtained a manifestly Lorentz invariant functional integral representation of the vacuum expectation values of time-ordered products of gauge
invariant Heisenberg operators. As already noted, we will use this equation (1.115) to compute

hT OA OB . . . ivac even if the O are not gauge invariant. In this case, one has to remember that the
result is unphysical and depends on the gauge-parameter . Nevertheless, any final physical result
(like S-matrix elements) must be gauge invariant and independent of .
25

Let us now determine the propagator of the gauge field hT A (x)A (y) ilibre
vac = i (x, y).
According to (1.36), the propagator is given by the inverse of the quadratic part of the action Seff :

Z
1
2
Seff |quadratic =
d x f f ( a ) + (i terms)
4
2
Z
1
=
d4 x a [ (1 ) ] a + (i terms)
2
Z
1
d4 x d4 y a (x)D (x, y)a (y) ,

(1.117)
2
with

+ (1 ) i (4) (x y)
D (x, y) =
x x
x x
Z
4
iq(xy)
dq
2
=
(1
)q
q
i
e
.
(2)4
(1.118)
The propagator is i (x, y) where = D1 , i.e.

Z
(x, y) (x y) =
d4 q
(q)eiq(xy) ,
(2)4
(1.119)
with
(q) =
1 q q
.
+
i
(q 2 i)2
q2
(1.120)
As expected for a gauge-dependent quantity, the propagator depends explicitly on . Note that
the limit 0 is singular since it would remove the gauge-fixing. The choice = 1 is called
Feynman gauge and yields (q) = q2
which is particularly simple, while gives (q) =
i
q q
(q2 i)2 and is called the Landau or Lorenz gauge (since strictly enforces the Lorenz
q 2 i
gauge condition a = 0).
1.5.3
Feynman rules of spinor QED
Let us now specify the matter part of the action to be that of an electron Dirac field (of charge
q = e with e > 0) interacting with the electromagnetic field:
1
L = F F (/ + ieA
/ + m) ,
4
and
Leff = L
( a )2 .
2
(1.121)
(1.122)
The Feynman rules for S-matrix elements then are:

i
1 q q
,
photon propagator :
+
(2)4 q 2 i
(q 2 i)2
26
electron/positron propagator :
i
i (ik/ + m)
1
4
(2) ik/ + m i
(2)4 k 2 + m2 i
vertex : (2)4 e (4) (k k 0 + q) ,

e
p
initial photon :
(2)3/2
2p0
final photon :
e
p
(2)3/2
initial electron :
u
,
(2)3/2
final electron :
u
,
(2)3/2
initial positron :
v
,
(2)3/2
final positron :
v
,
(2)3/2
2p0
integrate over all internal four-momenta.

The Feynman rules for Green-functions are the same, except that one associates propagators to the
external lines instead of the initial/final particle wave-function factors u, v or .
Most of the integrations over internal momenta are fixed by the (4) s from the vertices. Of course,
one overall (4) only enforces conservation of the external four-momenta and thus cannot serve to
fix any internal momentum. Thus the number of unconstrained internal momenta is I V + 1 if
the number of vertices is V and the number of internal lines I. We have already seen that there is
the general topological relation (1.45) between I, V and the number of independent loops L in a
diagram, I V = L 1. It follows that in any Feynman diagram there are exactly L unconstrained
four-momenta to be integrated, one for every loop.
Note that in spinor QED all vertices are tri-valent (3 lines attached). This gives another relation
between V , I and the number E of external lines: 3V = 2I + E. Thus in spinor QED
3V = 2I + E
I V =L1
V = 2L + E 2 ,
(1.123)
and for a given S-matrix element or given Green function (fixed number of external lines) one gets an
additional factor of e2 for every additional loop: one sees very clearly that the perturbative expansion
is an expansion in the number of loops and the expansion parameter is the fine structure constant
(not to be confused with the gauge parameter) 10
=
e2
1
'
.
4
137
(1.125)
One can argue that the expansion parameter for a given S-matrix element is 4
rather than : every vertex
4
4
contributes a factor (2) e and every internal line a (2) . Every integration over a loop momentum d4 k can be
expected to give a factor 2 (the angular integration is estimated to give the volume 2 2 and k 3 dk = 12 k 2 dk 2 gives
another 21 ). Altogether, one has a factor
10
(2)4V eV (2)4I 2L = (2)4 eE2

so that every loop can be expected to yield a factor
e2
16 2
e2
16 2
L
= (2)4 eE2
L
,
4
(1.124)
' 6 104 1.
27
2
2.1
A few results independent of perturbation theory

Structure and poles of Green functions
There are a few statements that can be made about the structure of the various Green functions
independently of any explicit (perturbative) computation, just based on arguments of symmetry, in
particular Poincare invariance. Consider the Fourier transform of a general n-point (Green) function11
Z
P

b
G(n) (q1 , . . . qn ) d4 x1 . . . d4 xn ei r qn xn hT O1 (x1 ) . . . On (xn ) ivac .
(2.1)
Recall that hT (. . .)ivac = h| T (. . .) |i where h| and |i are both the in-vacuum. (In perturbation
theory, this would be given by the sum of all the corresponding Feynman diagrams with n external
lines but excluding all diagrams with vacuum bubbles.) From translational invariance, this Green
P
e(n) (q1 , . . . qn ). The latter may contain pieces
function must be a product of (4) ( r qr ) times some G
which are again proportional to some (4) (corresponding to a disconnected part of the Green function)
and pieces without such further (4) -singularities, but with various poles and branch cuts in various
combinations of the momenta. We will concentrate on the poles and their residues. As an example,
e(2) is just the propagator with a pole at q12 = q22 = m2 and
consider a free scalar theory where G
residue i.
Here we will establish the general structure of the 2-point Green functions close to their poles and then
just state the corresponding result for the n-point functions. To begin with, we write explicitly
Z
h
b (2) (q1 , q2 ) = d4 x1 d4 x2 eiq1 x1 iq2 x2 (x01 x02 ) h| O1 (x1 )O2 (x2 ) |i
G
i
(2.2)
+ (x02 x01 ) h| O2 (x2 )O1 (x1 ) |i ,
We now insert a complete set of states
inEthe in-basis of the Hilbert space. This basis contains, besides the

in
in-vacuum, the one-particle states p~,,n , as well as all the multi-particle states. These one-particle states
2
2
correspond

E to the physical
E particles with masses mn that one can measure as mn = p p p and where

in
P in
p
~,,n = p p
~,,n . Thus
1 = |i h| +
XZ

in
d3 p in
p~,,n + . . . ,
p
~,,n
(2.3)
n,
where + . . . indicates all the contributions from multi-particle states. These are defined as states depending
on the total momentum p~tot , as well as at least one more continuous variable. Thus
h| O1 (x1 )O2 (x2 ) |i = h| O1 (x1 ) |i h| O2 (x2 ) |i
XZ

in
+
p~,,n O2 (x2 ) |i + . . . .
d3 p h| O1 (x1 ) in
p
~,,n
(2.4)
n,
By translational invariance one has

ip x
iP x
1 O (0)eiP x1 in
1 h| O (0) in
h| O1 (x1 ) in
1
1
p
~,,n = h| e
p
~,,n = e
p
~,,n ,
(2.5)
11
b (n) for G
b (n) (x1 , . . . xn ) and its Fourier transform G
b (n) (q1 , . . . qn ). Also, we consider
Here we use the same notation G
general Heisenberg operators Oj rather than just the elementary fields lj , since most of the argument does not
depend on the form of the operators.
28
as well as h| O1 (x1 ) |i = h| O1 (0) |i. Most often, the Oj transform non-trivially under the Lorentz
group or some internal symmetry group (that leaves the vacuum invariant) in which case h| Oj (0) |i = 0.
In general, one has
h| O1 (x1 )O2 (x2 ) |i = h| O1 (0) |i h| O2 (0) |i
XZ

in
+
d3 p eip(x1 x2 ) h| O1 (0) in
p~,,n O2 (0) |i + . . . .
p
~,,n
|
{z
}
{z
}
|
n,
(2.6)
p,,n)
M (~
MO1 (~
p,,n)
O2
p
p). When inserted into (2.2), the first line
Let us insist that the p0 are on-shell, i.e. p0 = p~R2 + m2n n (~
of (2.6), if non-vanishing, yields a contribution d4 x1 d4 x2 eiq1 x1 iq2 x2 (4) (q1 ) (4) (q2 ) corresponding
to a disconnected piece. Concentrate now on the contributions of the one-particle states. Writing
(x01
x02 )
Z
=
they are

b (2) (q1 , q2 )
G
one particle
d ei(x1 x2 )
,
2i + i
(2.7)
Z
X i Z d
3
=
d p
d4 x1 d4 x2 eiq1 x1 iq2 x2
2
+
i
n,
h
i
0
0
0
0
eip(x1 x2 ) ei(x1 x2 ) MO1 (~

p, , n)MO
p, , n) + eip(x1 x2 ) e+i(x1 x2 ) MO2 (~
p, , n)MO
p, , n)
(~
(~
2
1
X Z d
= i(2)7 (4) (q1 + q2 )
d3 p
+
i
n,
h
i
(~
p ~q1 )( q10 + p0 )MO1 (~q1 , , )MO
q1 , , n) + (~
p ~q2 )( q20 + p0 )MO2 (~q2 , , n)MO
q2 , , n)
(~
(~
2
1
#
"
X MO1 (~q1 , , n)MO (~q1 , , n) MO2 (~q2 , , )MO (~q2 , , n)

1
2
(2.8)
+
= i(2)7 (4) (q1 + q2 )
0 (~
0 (~
q
q
q
)
+
i
q
)
+ i
n
1
n
2
1
2
n,
This expression clearly exhibits
the poles due to the one-particle intermediate states. The poles are at
p
q10 = q20 = n (~q1 ) = m2n + ~q12 , i.e. on the mass shell of the intermediate physical particle. One can
show that the multi-particle intermediate states do not lead to poles but to branch cuts.
We will be mostly interested in the case where the Heisenberg operators Oj correspond to the elementary
fields l appearing in the Lagrangian, specifically O1 = l and O2 = k so that the above result reads
M (~q2 , , n)M (~q2 , , n)

(~

X
M
(~
q
,
,
n)M
q
,
,
n)
1
1
l
k
k
l
b (2) (q1 , q2 )
. (2.9)
G
= i(2)7 (4) (q1 + q2 )
+
0 (~
0 (~
poles
q
q
)
+
i
q
q
)
n 1
n 2 + i
1
2
n,
Let us compare with the result that would have been obtained in a free theory of a field of species n and

P R d3 p
mass m where l (x) = l (x) = (2)
p, , n )a(~
p, , n )eipx +vl (~
p, , n )ac (~
p, , n )eipx . In this
3/2 ul (~
case, the only intermediate states that contribute are the one-particle states of species n created by a and
p, , n ) and M (~q1 , , n ) = (2)13/2 vk (~
p, , n ), so that
ac . Furthermore, Ml (~q1 , , n ) = (2)13/2 ul (~
k
b free (q1 , q2 ) = i(2)4 (4) (q1 + q2 )

G
(2)
X
4 (4)
i(2)
ul (~q1 , , n )uk (~q1 , , n ) vl (~q2 , , n )vk (~q2 , , n )

+
q10 n (~q1 ) + i
q20 n (~q2 ) + i
(q1 + q2 ) m
lk (q1 )
(2.10)
where im
lk (q) is the usual free propagator with mass m . The similarity between (2.9) and (2.10) is
no coincidence. Indeed, by Lorentz invariance, the matrix element Ml (~q1 , , n) is constraint to equal
the corresponding ul (~q1 , , n), up to a normalization, and similarly for the M (~q1 , , n) and vl (~q1 , , n).
k
29
(Recall that for every irreducible representation of the Lorentz group one can determine the corresponding
coefficients ul and vl solely from the transformation properties up to a normalization). Hence:
Ml (~q1 , , n) = Nn
ul (~q1 , , n)
(2)3/2
M (~q2 , , n) = Nn
l
vl (~q2 , , n)
,
(2)3/2
(2.11)
where the normalization constants Nn and Nn may differ at most by a phase. In (2.9), the contributions
to the residue of a given pole at some12 q12 = m2 come from those one-particle states n that have a mass
mn equal to m .
Combining the results (2.9), (2.11) and (2.10), we finally get for the behaviour of the 2-point
function:

h X
i

4 (4)
n 2
blk
(2)
(q
+
q
)
|N
|
(i) m
G
(q
,
q
)
1
2
(2) 1 2
lk (q1 )
2
2
pole at q1 =m
n | mn =m
|Nm |2
blk
G
(2)free,m (q1 , q2 )
(2.12)
The lesson to remember is the following: in general, the 2-point function of the interacting theory
is very complicated, with branch cuts and poles. Equation (2.12) states that, as q12 m2 , where

m is the mass of a physical one-article state such that h| l (0) in
p
~,,n 6= 0, the 2-point function
behaves as the 2-point function of a free field of mass m, up to a normalization constant.
These results can be generalized to an arbitrary n-point function depending on momenta q1 , . . . qn :
Such an n-point function has a pole whenever, for any subset I of {1, . . . n}, the combination qI =
in
P
2
2

q
is
such
that
q
=
m
with
m
being
equal
to
the
mass
of
any
one-particle
state
j
I
p
~,,n
jI
Q
Q
O
|i.
More
precisely,
if
that has non-vanishing matrix elements with jI Oj |i and with j I
j
/
p
0
we suppose I = {1, . . . r}, q qI = q1 + . . . + qr = qr+1 . . . qn then, as q ~q2 + m2
p
X
2i ~q 2 + m2
7 (4)
G 2
(2)
(q
+
.
.
.
+
q
)
M0|q (q2 , . . . qr )Mq,|0 (qr+2 , . . . qn ) ,
(2.13)
1
n
q + m2 i
with
Z
r
X
Pr

(2) (
qs p) M0|p (q2 , . . . qr ) = d4 x1 . . . d4 xr ei s=1 qs xs h| T O1 (x1 ) . . . Or (xr ) |p, i
4 (4)
4 (4)
(2) (
s=1
n
X
qs + p) Mp|0 (qr+2 , . . . qn ) =
d4 xr+1 . . . d4 xn ei
Pn
r+1 qs xs
s=r+1

hp, | T Or+1 (xr+1 ) . . . On (xn ) |i .
(2.14)
Again, the proof uses only translation invariance, the causal structure implied by the time-ordering
and the fact that multiparticle intermediate states produce branch cuts rather than poles. Note that
the above pole structure is exactly what one expects from a Feynman diagram with a single internal
line for a particle of mass m connecting a part of the diagram, with the first r operators Oi attached,
to another part, with the last n r operators Oi attached, as shown in the figure. However, the
above property is much more general in that the particle of mass m need not be one corresponding
to an elementary field in the Lagrangian but could correspond to a complicated bound state.
12
This is an abuse of language: when we say a pole at q 2 = m2 , since

0
1
q 2 +m2
1
2m (~
q)
1
q 0 +m (~
q)
1
q 0 m (~
q)
, we
really mean a pole at q = m (~q) and a pole at q = m (~q).
30
r+1
r+2
3
2
1
2.2
Complete propagators, the need for field and mass renormalization
In the above formula (2.12) the 2-point function on the left-hand-side is the Fourier transform of

hT l (x1 )k (x2 ) ivac where the l (x) are the Heisenberg operators that evolve with the full Hamiltonian. This is also referred to as the full or complete propagator, while on the right-hand-side appears
the free propagator as entering the Feynman rules. More precicely, the Heisenberg operators l correspond to the fields as they appear in the (interacting) Lagrangian and are accordingly normalized.
Such fields will be called bare fields and we write l,B (x). Actually, in most theories (at least in
in

perturbation theory) the only one-particle states in
6 0 all
p
~,,n that are such that p
~,,n l,B (0) |i =
2
2
have the same mass, and then there is only a pole at q1 = m , with m being in general different
from the mass parameter appearing in the Lagrangian of the bare field l,B and which we call the
bare mass mB . Hence, we can rewrite (2.12) as
Z

2
4 (4)
im
d4 x1 d4 x2 eiq1 x1 iq2 x2 hT l,B (x1 )k,B (x2 ) ivac 2
lk (q1 ) |N | (2) (q1 + q2 ) .
q1 m2
(2.15)
2
We can get rid of the factor |N | on the right-hand-side by deviding by it and defining
l,R =
1
l,B .
N
(2.16)
Then, close to its pole at q12 = m2 , the two-point function of l,R behaves as the two-point function
of a free field with mass m. A field with this behaviour is called a renormalized field and m the
renormalized mass. In the sequel, to simplify the notation, we will not write the subscript R for the
renormalized fields and masses. We will study in some detail how this goes for the different types of
fields.
2.2.1
Example of a scalar field
Call the interacting real scalar B with Lagrangian

1
1
(2.17)
L = ( B )2 m2B 2B VB (B ) .
2
2
There is no reason to expect that B has a correctly normalized two-point function or that this
function has a pole at m2B . Let
B =
= Z 1/2 B
(2.18)
m2 = m2B + m2 ,
31

and require that is correctly normalized, i.e. that h| (0) in
=
q~
and that its two-
q~ 2 +m2
of m2B and
(2)3/2 2
point function have a pole at m2 . This will fix Z and m2 as functions

the couplings.
Then is called the renormalized field and m the renormalized mass. A straightforward rewriting
of the Lagrangian yields
1
L = Z( )2
2
1
=
( )2
2
|
{z
L0
1 2
(m m2 )Z2 VB ( Z)
2

1
1 2 2
1
2
2 2
2 2
m + (Z 1) ( ) + m + Zm VB ( Z) .
2
2
2
} |
{z
}
L1
(2.19)
The strategy is to treat L0 as the free part of the Lagrangian and L1 as the interaction. All the
terms in L1 involving factors of (Z 1) or m2 will be called counterterms.
The full propagator of the renormalized field is called 0 (q). It is conveniently expressed in terms
of the one-particle irreducible propagator. In general, a one-particle irreducible (1PI) diagram is a
connected diagram that will not become disconnected by just cutting a single line. More precisely,
let i(2)4 (q 2 ) be the complete one-particle irreducible propagator of with two external free
i
1
4
propagators (2)
4 q 2 +m2 i removed. Graphically, with a interaction is given by
2
(q ) =
The first term equals (Z 1)(q 2 + m2 ) + Zm2 and is entirely due to the counterterms, while
the other terms involve loops (possibly including counterterms inside the loops) and contribute to
loops (q 2 ). The full propagator then is related to the one-particle irreducible propagator13 by (see
Figure below)
i
0 (q)
(2)4

i
i
i
1
1
1
4 2
+
i(2) (q )
+ ...
=
(2)4 q 2 + m2 i
(2)4 q 2 + m2 i
(2)4 q 2 + m2 i

1
i
(q 2 )
i
1
1
=
1 2
=
.
(2.20)
4
2
2
2
4
2
2
(2) q + m i
q + m i
(2) q + m (q 2 ) i
In summary:
0 (q) = q 2 + m2 (q 2 ) i
1
(2.21)
13
To simplify the discussion, we exclude the possibility of tri-linear self-interactions 3 which would lead to
tadpole diagrams.
32
with
(q 2 ) = (Z 1)(q 2 + m2 ) + Zm2 + loop (q 2 ) .
(2.22)
The so far arbitrary constants Z and m2 must be fixed by the normalization requirements: we
required that 0 has a pole at q 2 = m2 . This implies
(m2 ) = 0 .
(2.23)
Correct normalization of the two-point function translates into the residue of the pole of 0 be one:
1

1
2
2
2
2 d
0
2
2 2
= q (m + q ) i
= q q
(q)
(m ) + O((q ) ) i
dq 2
q 2 =m2 +q 2
1

1
d
2
,
(2.24)
1 2 (m )
2
dq
q i
which leads to
d 2
(q )
=0.
dq 2
q 2 =m2
(2.25)
Inserting (2.22) into (2.23) and (2.25) yields

Zm2 = loop (m2 )

d

Z = 1 + 2 loop (q 2 )
,
dq
q 2 =m2
(2.26)
(2.27)
which determines Z and m2 in terms of loop .

Clearly, in any generic interacting theory, loop will be non-vanishing (and in particular nonvanishing at q 2 = m2 with a non-vanishing first derivative). Hence, in such a generic interacting
theory there will always be renormalization of the wave-function (Z 6= 1) and of the mass (m2 6= 0).
The necessity of renormalization has nothing to do with diverging, infinite loop-integrals, but is a
generic feature of interacting theories in order to have correctly normalized two-point functions with
poles at physical values of q 2 .
It is interesting to substitute the values (2.26) and (2.27) of Z and Zm2 into eq. (2.22) for (q 2 )
to get
(q 2 ) = loop (q 2 ) loop (m2 )

d
2
(q
)
(q 2 + m2 ) .
2
dq 2 loop
q =m2
(2.28)
We see that the (renormalized) one-particle irreducible complete propagator is given by its loop
contribution loop with its two first terms in a Taylor series expansion around q 2 = m2 subtracted!
To be completely clear, let us insist that i0 is the full propagator of the renormalized field ,
related to the full propagator i0B of the bare field B by
i0 (x y) = hT (x)(y)i =
1
1
hT B (x)B (y)i = (i)0B (x y) .
Z
Z
33
(2.29)
From our previous relations one finds

1
1
0 (q) = q 2 + m2 (q 2 ) i
= q 2 + m2 + (Z 1)(q 2 + m2 ) Zm2 loop (q 2 ) i
1
1
1 2
1
1 2
=
q + m2 m2 loop (q 2 ) i
=
q + m2B B, loop (q 2 ) i
,
(2.30)
Z
Z
Z
where we defined
1
B, loop = loop .
(2.31)
Z
Comparing (2.29) and (2.30) we see that
0B (q) =
0 (q)
q 2 + m2B B, loop (q 2 ) i
q 2 + m2 (q 2 ) i
1
1
(2.32)
0B (q) = Z 0 (q) .
In particular, in the bare propagator 0B one has B, loop , missing the contributions from the counterterms. Nevertheless the relation between the full bare propagator and the full renormalized propagator
is very simple: they only differ by the factor Z. Let us insist that the renormalization conditions
(2.23) and (2.25) are such that the renormalized propagator satisfies
h
i
h
i
1
0
2
2
2
2
(q) = 2
1 + O(q + m ) = (q) 1 + O(q + m ) ,
(2.33)
q + m2 i
i.e. up to corrections that vanish on shell, the full renormalized propagator equals the free propagator.
2.2.2
Example of a Dirac field
The Lagrangian is
L = B (/ + mB )B VB (B ) ,
(2.34)
with
1/2
= Z2
B =
p
Z2
m = mB + m .
(2.35)
As above, we rewrite L as
L = L0 + L1
/ + m)
L0 = (
p
/ + m) + Z2 m
VB ( Z2 ) .
L1 = (Z2 1)(
(2.36)
Denote by i(2)4 (k/) the one-particle irreducible complete fermion propagator.14 Let the complete
i
0
propagator be (2)
4 S (k). Then
S 0 (k) =
1
1
1
1
+
(k/)
+ ... =
,
ik/ + m i ik/ + m i
ik/ + m i
ik/ + m (k/) i
(2.37)
When we write (k
/) we mean the following: we will see that is of the form = f (k 2 ) ik
/ + g(k 2 ) 1 =
2
f (k
/ ) ik
/ + g(k
/ ) 1, which is indeed a function of /k.
14
34
and
(k/) = (Z2 1)(ik/ + m) + Z2 m + loop (k/) .
(2.38)
The correct normalization for the complete propagator (pole at k 2 = m2 and the residue condition)
yields
(im) = 0 ,

=0.
(k/)
k/
/=im
k
(2.39)
This fixes Z2 and m as

Z2 m = loop (im) ,

.
Z2 = 1 i loop (k/)
k/
/=im
k
(2.40)
(2.41)
Again, if we insert these values for Z2 and Z2 m into (2.38) we get
(k/) = loop (k/) loop (im)

1

loop (k/)
(ik/ + m) ,
i k/
/=im
k
(2.42)
showing again that the renormalized (k/) is given by its loop-contributon with its two first terms
in a Taylor expansion around k/ = im subtracted.
2.3
Charge renormalization and Ward identities
The Lagrangian for charged particles is invariant under phase rotations of the associated fields. This
R
leads, as usual to a conserved current J and associated conserved charge Q = d3 x J 0 which, upon
quantization, become operators acting on the states. Since Q commutes with the Hamiltonian we
can take all (one-particle) states |p,,n i to be eigenstates of Q with
Q |p,,n i = qn |p,,n i ,
(2.43)
as well as Q |i = 0. The eigenvalue qn is called the charge of the particle. On the other hand, in
the Lagrangian L appear parameters ql via the covariant derivatives D l = ( i
ql A )l of l .
How are they related?
l ,D l ]
From the definition of J 0 = L[A
= L
(i
ql )l = i
ql Pl l and the canonical commutation
0
0 l
relations we get [Q, l ] =
ql l . Hence
0 = h| Q l |p,,n i = h| (l Q + [Q, l ]) |p,,n i = (qn ql ) h| l |p,,n i .
(2.44)
Thus, whenever h| l |p,,n i =

6 0 we must have ql = qn : the charge qn as measured by the eigenvalue
of Q equals the parameter qn appearing in the covariant derivatives of n in the Lagrangian. Suppose
now we rescale A A = A0 . Then J (J 0 ) = 1 J and hence qn qn0 = 1 qn . According
to the previous argument, then also qn qn0 = 1 qn . As a result, qn0 A0 = qn Am and the covariant
35
derivative remains unchanged. It follows that if A is renormalized by some multiplicative factor, all
charges are renormalized by the inverse of this factor:
1/2
A = Z 3
AB
ql =
p
Z3 qB,l
l .
(2.45)
The charge renormalization is the same for all fields! Of course, this is related to the gauge invariance
which forces all charged particles to couple in the same way via the covariant derivative to the
electromagnetic field A . In particular (2.45) shows that if qB,l = qB,l0 then also ql = ql0 , even if
l and l0 have very different non-electromagnetic couplings like e.g. a proton and a positron, as
p
shown in Fig. 5 for a quark and a positron/electron. One sometimes writes qB = Zq q so that (2.45)
implies
Zq Z3 = 1 .
(2.46)
Figure 5: Contributions to the electromagnetic vertex function for a quark (left) and an electron/positron (right). The solid lines are fermion propagators, the wavy lines photon propagators
and the dashed lines gluon propagators.
One defines the vertex function mn (p, p0 ) by
Z

m (z) |i
d4 x d4 y d4 z eipxiky+ilz h| T J (x)n (y)
0
0
4 (4)
iq Snn
(p + k l) .
0 (k)n0 m0 (k, l)Sm0 m (l) (2)
(2.47)
It follows from this definition that is the sum of all vertex graphs (with the two complete Dirac
propagators removed, and also no photon propagator): it is the one-particle irreducible 3-point
1
1
function. To lowest order (free fields) the l.h.s. of (2.47) is ik/+mi
(iq ) il/+mi
(4) (p + k l), so
that

tree = .
(2.48)
Above we have derived identities due to the universal coupling of the electromagnetic field through the
covariant derivative. One of the tools was the commutation relation of the charge operator with the
Heisenberg picture quantum fields. Similarly, we now derive a relation between the vertex function
and the full fermion propagators, known as Ward identity. Using J = 0 and [J 0 (t, ~x), n (t, ~y )] =
36
q (3) (~x ~y ) n (t, ~y ) we get

T J (x)n (y)m (z) = T J (x)n (y)m (z) +

x

m (z) + (x0 z 0 )T n (y)[J 0 (x),
m (z)]
+ (x0 y 0 )T [J 0 (x), n (y)]

m (z) + q (4) (x z)T n (y)
m (z) .
(2.49)
= 0 q (4) (x y)T n (y)
This, together with
R
0
= iSnm
(y z) = i
the definition (2.47),

d4 q iq(yz) 0
e
Snm (q), yields
(2)4
and
recalling
that

m (z) |i
h| T n (y)
(l k) S 0 (k) (k, l)S 0 (l) = iS 0 (l) iS 0 (k) ,
(2.50)
or
1
(l k) (k, l) = iS 0 (k) iS 0 (l) ,
(2.51)
which is known as the generalized Ward identity. The original form of the Ward identity is obtained
by letting l k so that
(k, k) = i

0 1
S (k) = + i
(k/) .
k
k
(2.52)
Due to (2.39), the last term vanishes on-shell, i.e. for k/ = im, and hence when evaluated between
on-shell spinors one simply has
u(k) (k, k)u(k) = u(k) u(k) ,
(2.53)
so that radiative corrections to the vertex function for the interaction of an on-shell fermion with a
zero-momentum photon vanish. But this is exactly the way the electric charge of particles is defined,
ren . Similarly, (2.51) leads to
and we find again that q is the coefficient of A ()
(l k) u(k) (k, l)u(l) = 0 .
2.4
(2.54)
Photon propagator and gauge invariance
Gauge invariance implied that the current J is conserved, J = 0. Then, much as above, one has

0
0
0
T
J
(x)J
(y)J
(z)
.
.
.
=
(x
y
)T
[J
(x),
J
(y)]J
(z)
x

+ (x0 z 0 )T J (x)[J 0 (y), J (z)] + . . . .
(2.55)
Now J is a neutral operator and hence [J 0 (x), J (y)] = 0. This coud be violated by so-called
Schwinger terms from defining J properly in terms of the elementary fields at coinciding points. For
Dirac fermions and using dimensional regularization, however, no such Schwinger terms arise and
one indeed has
[J 0 (x), J (y)] = 0
(2.56)
37
as an operator identity. As a consequence, the r.h.s. of (2.55) vanishes. Let

Z

...
T J (x1 )J (x2 ) . . . +
.
M (q1 , q2 , . . .) = d4 x1 . . . eiq1 x1 ...
(2.57)
Then (2.55) implies

...
(qi ) M
(q1 , q2 , . . .) = 0 .
(2.58)
Now any S-matrix element in electrodynamics is of the form

Z
S
d4 q1 d4 q2 . . . 1 1 (q1 )2 2 (q2 ) . . . 1 (k1 )2 (k2 ) . . . 1 (l1 )2 (l2 ) . . .
c1 1 2 2 ...1 2 ...1 2 ... (q1 , q2 , . . . , k1 , k2 . . . , l1 , l2 . . .) ,
M
(2.59)
c being the matrix element of all the matter currents as defined above but with all electrowith M
c and we see that the S-matrix is
magnetic interactions turned off. Clearly, (2.58) still holds for M
unchanged if we replace any of the photon propagators (q) or any of the polarization vectors
according to
(q) (q) + a q + q b ,
(k) (k) + c k ,
(2.60)
with arbitrary four-vectors a , b or scalar c.

The complete photon propagator necessarily is given by
+
0 (q)
= (q) + (q)M (q) (q)
(2.61)
P c1 ...2n
where M
1 2 . . . 2n1 2n . Indeed, the first term in (2.61) just is the free
n Mvac,vac
propagator, and all higher-order corrections are summarized in the second term. Of course, M is
not one-particle irreducible. Obviously,

it follows
from (2.58) that q M = M q = 0. In a general

gauge we had (q) =
1
q 2 i
(q 2 ) qq2q
and hence,
0 (q) = (q) +
(q 2
1
M .
i)2
(2.62)
)
Then q 0 (q) = q (q) = 1(q
q .
q 2 i
In terms of the complete one-particle irreducible photon propagator = one has (with
obvious index contractions)
0 = + + . . . = (1 )1 ,
(2.63)
but also
0 = + + + . . . = + ( + + . . .) = + 0 .
38
(2.64)
Contracting the last equation with q on the left yields15 0 =

follows that
1(q 2 )
q 0 ,
q 2 i
hence q = 0. It
(q) = (q 2 q q )(q 2 ) .
1
1
so that (1 ) = q2 i
(q 2 (1
q 2 i

1
(q 2 ) qq2q . Using this in (2.63) gives
1(q 2 )
Then ( ) =
(1 )1
=
0
(2.65)
) + q q ) whose inverse is
e 2 ) q2q
(q
[(q 2 )(1 (q 2 )) + (q 2 )] qq2q

q
1
= (1 ) =
2
2
2
2
(1 (q ))(q i)
(1 (q ))(q i)
(2.66)
As before, we require that the complete photon propagator should have a pole at the physical
mass and close to this pole be normalized as the free propagator. Of course, we expect the physical
mass to be still zero, although this needs to be verified. Here we do not prove this statement but
give evidence for it. Indeed, since is one-particle irreducible it is not expected to have any poles
at q 2 = 0. It certainly has branch cut singularities and it could, at least in principle, also have poles
at q 2 = M 2 due to intermediate bound states of mass M . However, we do not expect the latter
to have zero mass. If has no pole at q 2 = 0, then (q 2 ) does not have such a pole either, and
then 0 keeps its pole at q 2 = 0. Note that if (q 2 ) had a pole at q 2 = 0, say (q 2 ) qa2 b then
1
1
a
1
1
0 q2 (1a/q
would have its pole shifted to q 2 = 1+b
. On the other hand,
2 +b) = q 2 a+bq 2 = 1+b q 2 a
1+b
a
0
2
if (q 2 ) has a pole at q 2 = M 2 with M 2 6= 0, say (q 2 ) q2 +M
2 b then keeps its pole at q = 0,
a
but there is an additional pole that appears at q 2 = M 2 + 1+b
. Clearly, this is undesirable, too.
2
Henceforth we assume that indeed (q ) does not have any pole. Then, for the correctly normalized
A the residue of 0 at the pole should be (q 2 ) qq2q which requires
(0) = 0
and
(q 2 ) should not have a pole .
(2.67)
As for the scalar or Dirac fields we can rewrite the (bare) Lagrangian, written in terms of the bare
fields AB , in terms of the renormalized fields A and the constant Z3 , and then separate a free and
an interaction part. In order to do so, we also need to start with a bare parameter B for the
gauge-fixing term:
B Z3
1 B B
1
L = F
FB
( AB )2 + . . . = Z3 F F
( A )2 + . . .
4
2
4
2
1
1
= F F ( A )2 (Z3 1)F F + . . . ,
4
2
4
(2.68)
with = Z3 B . The terms + . . . involve the couplings to the matter fields and enter only in the
loop-corrections to the photon propagator. We see that the free propagator for the renormalized
field A now is the same with the same () as before, while the quadratic counterterm
15
Except for one has 6= 1.
39
14 (Z3 1)F F gives a purely transverse contribution to the one-particle irreducible propagator
. Hence it contributes a piece 1 Z3 to (q 2 ) and we conclude
(q 2 ) = 1 Z3 + loop (q 2 ) .
(2.69)
The residue condition (2.67) then determines Z3 in terms of loop as

Z3 = 1 + loop (0) .
(2.70)
We have shown how all the renormalization constants Z, Z2 and now Z3 are determined by the
renormalization conditions in terms of the loop contributions to the various one-particle irreducible
functions. In practice though, if one wants to compute these one-particle irreducible functions one
does not need to explicitly determine the Zs, as we have seen above for (for the scalar) or for
for the fermions. Similarly, for the photon, inserting the value (2.70) of Z3 into (2.69) simply gives
(q 2 ) = loop (q 2 ) loop (0) ,
(2.71)
which clearly satisfies (2.67), and we see (again) that the renormalized (q 2 ) is given by its loopcontribution, with its first Taylor coefficient subtracted.
40
One-loop radiative corrections in 4 and QED
We will first work out the one-loop radiative corrections in some detail for QED and at the end of
this section quickly consider those of scalar 4 theory.
3.1
Setup
Recall that the renormalized fields are those that have correctly normalized residues of their propagators close to their poles. They are related to the bare fields which appear in the original Lagrangian
by field renormalization factors. The renormalized masses are defined as the positions of the poles of
the complete propagators (poles at q 2 = m2 ) and are related to the bare masses appearing in the
Lagrangian. Finally, coupling constants are also renormalized. For electrodynamics with charged
Dirac fermions
1/2
= Z2
A
1/2
Z3 AB
m = mB + m
e = Z3 eB ,
1/2
(3.1)
while for scalars

= Z 1/2 B
m2 = m2B + m2 .
(3.2)
The original Lagrangian is always written in terms of the bare fields with the same coefficients as for
the free fields, plus the usual interactions. For spinor QED we have
B
1 B
L = F
( AB )2 .
FB B (/ + mB )B ieB AB B B
4
2
(3.3)
We have included the gauge-fixing term with a bare parameter as discussed above (2.68). Using (3.1)
together with = Z3 B , L is rewritten as
L = L0 + L1 + L2 ,
1
/ + m) ( A )2 ,
L0 = F F (
4
2
,
L1 = ieA
1
/ + m) + Z2 m
ie(Z2 1)A
.
L2 = (Z3 1)F F (Z2 1)(
4
(3.4)
(3.5)
(3.6)
(3.7)
Clearly, L0 is exactly like the free Lagrangian but now with renormalized fields and masses. Similarly,
L1 is exactly like the original interaction term but now with renormalized fields and couplings, resp.
charges. The third term, L2 is due to the difference between bare and renormalized quantities. Its
terms are called the counterterms. If we would take all couplings to zero so that all fields become
free, there no longer would be a distinction between bare and renormalized quantities and in this
limit L2 would vanish.
41
The complete propagators for the renormalized fields are iS 0 (k/), i0 (q) and i0 (q) with
for
S 0 (k/) =
for A
ik/ + m (k/) i
0 = (1 )1

or 0 (q) =
1
(3.8)
with (q) =
[ + (1 )(q
qq2q
q 2 i
(q) = (q 2 q q )(q 2 )
)] qq2q
(3.10)
(1 (q 2 ))(q 2 i)
(3.9)
(3.11)
for
0 (q) =
q 2 + m2 (q 2 ) i
1
(3.12)
The (k/), (q 2 ) and (q 2 ) get contributions from the loops (including also counterterms in the
loops) and from the counterterms at tree-level:
fermions : (k/) = (Z2 1)(ik/ + m) + Z2 m + loop (k/) ,
(3.13)
photons : (q 2 ) = (Z3 1) + loop (q 2 ) ,
(3.14)
scalars : (q 2 ) = (Z 1)(q 2 + m2 ) + Z m2 + loop (q 2 ) .
(3.15)
One imposes the renormalization conditions (correct poles and residues)

fermions : (im) = 0 ,
photons :
(0) = 0
scalars : (m2 ) = 0 ,
3.2

(k/)/=im
k
/
k
= 0,
and (q 2 ) should have no pole at any q 2 ,

d
(q 2 )q2 =m2 = 0 .
dq 2
(3.16)
(3.17)
(3.18)
Evaluation of one-loop integrals and dimensional regularization
When evaluating one-loop diagrams one typically encounters integrals of the type
Z
Z
1
k 1 . . . k r
1 ...r
4
IN = d k
and
IN
= d4 k
,
D1 D2 . . . DN
D1 D2 . . . DN
(3.19)
where
Di = [(k pi )2 + m2i i] .
(3.20)
The pi are combinations of the external momenta and the mi are the masses appearing in the
r
propagators. Since each Di contains a k 2 the integrand of IN1 ...r behaves for large k as d4 k kk2N
k 3+r2N dk. (This is easy to see once the integral has been continued to Euclidean signature, but the
following discussion is equally valid in Minkowski signature.) Hence:
The integral diverges for 3 + r 2N > 1, i.e. 2N r < 4.
The integral diverges logarithmically for 3 + r 2N = 1, i.e. 2N r = 4.
The integral converges for 3 + r 2N < 1, i.e. 2N r > 4.
42
The 4 comes from the space-time dimension and would have to be replaced by d in d space-time
dimensions. Actually, one can formally continue to non-integer and more-over to complex dimensions
d where the integral can be evaluated and yields a meromorphic function of d with poles when d is
an even integer. This is the basis for dimensional regularization.
In dimensional regularization one replaces
Z
Z
4
d k (. . .)
dd k (. . .) .
(3.21)
As we will see, this makes convergent all our integrals (which we continue to call IN and IN1 ...r ) for
d 6= 2, 4, . . . . The indices , , , . . . formally become d-dimensional indices. One still has k k = k 2
but this now is a sum of d terms. In particular, = = d or, when dealing with -matrices,
= = d. Also, the usual rules for replacements in tensor integrals have to be accordingly
modified, e.g.
Z
Z
1
d

2
d k k k f (k ) = dd k k 2 f (k 2 ) ,
(3.22)
d
where the factor d1 can be checked by contracting with . Dimensional regularization can be
consistently implemented except when the antisymmetric 4-index tensor plays an important
role, as in the definition of 5 and of chiral fermions.16
Feynmans trick : To facilitate the integration, one rewrites the denominators appearing in the
integrals (3.19), using the formula
1
1
(a1 + . . . aN )
aN =
a1
a2
D1 D2 . . . DN
(a1 ) . . . (aN )
Z
dx1 . . .
dxN (1
0
aN 1
xa11 1 . . . xN
P
xj )
. (3.23)
[x1 D1 + . . . + xN DN ] aj
(Recall (x + 1) = x(x), (n) = (n 1)!, in particular (1) = (2) = 1, (3) = 2.) Note that the
P
P
i terms in
xj Dj add up to
xj i = i. If all aj = 1 eq. (3.23) simplifies:
P
Z 1
Z 1
1
(1 xj )
,
(3.24)
= (N 1)!
dx1 . . .
dxN
D1 D2 . . . DN
[x1 D1 + . . . + xN DN ]N
0
0
and in particular
1
D1 D2
Z
=
dx
0
1
,
[xD1 + (1 x)D2 ]2
(3.25)
1
= 2
D1 D2 D3
Z
dx
1x
1
dy
.
[xD1 + yD2 + (1 x y)D3 ]3
Since Dj = (k pj )2 + m2j i = k 2 2kpj + p2j + m2j i, one has

x1 D1 + . . . + xN DN = k 2 2kP (xj ) + M 2 (xj ) i = (k P (xj ))2 + M 2 (xj ) P 2 (xj ) i , (3.26)
16
This is one way to see why chiral fermions can lead to anomalies: one cannot simply use the gauge invariant
dimensional regularization in the presence of chiral fermions.
43
with
P (xj ) = x1 p1 + x2 p2 + . . . + (1 x1 . . . xN 1 )pN
M 2 (xj ) = x1 (p21 + m21 ) + x2 (p22 + m22 ) + . . . + (1 x1 . . . xN 1 )(p2N + m2N ) .
(3.27)
Then
Z
IN =
d k (N 1)!
Z
dx1 . . .
dxN
P
(1 xj )
[(k P (xj ))2 + M 2 (xj ) P 2 (xj ) i]N
One interchanges the now convergent k-integration with the xj -integrations, so that
Z 1
Z 1
X

IN = (N 1)!
dx1 . . .
dxN (1
xj ) IeN P (xj ), M 2 (xj ) ,
0
with
Z
IeN (P, M ) =
(3.28)
(3.29)
dd k
=
[(k P )2 + M 2 P 2 i]N
dd k 0
,
[k 0 2 + R2 i]N
(3.30)
where we set R2 = M 2 P 2 and shifted the integration variables form k to k 0 = k P . Note that
such shifts are justified only because we have a convergent integral. Of course, one has an analogous
formula for the IN1 ...r with
Z
Z
1
r
0
1
0
r
k
.
.
.
k
...
d
d 0 (k + P ) . . . (k + P )
r
1
IeN
(P, M ) = d k
=
d
k
, (3.31)
[(k P )2 + M 2 P 2 i]N
[k 0 2 + R2 i]N
At this point one needs to make a Wick rotation, to be discussed in the next subsection. This
results in a factor of i and the four-momentum k then is a Euclidean four-momentum, which we
denote by kE . (Strictly speaking, the P appearing as argument in IeN (P, M ) should then also be
continued to a Euclidean PE .) The last expression in (3.30) for the integral is easily evaluated due
to its spherical symmetry. Using dd kE = dd1 kEd1 dkE and
Z
2 d/2
dd1 = vol(S d1 ) =
,
(3.32)
( d2 )
S d1
we get
Z
d
kEd1 dkE
d/2 x 2 1 dx
=
i
( d2 ) 0 [x + R2 ]N
[kE 2 + R2 ]N
0
Z
d
d/2 2 d N
= i d (R ) 2
dy y 2 1 (1 + y)N .
( 2 )
0
2 d/2
IeN (P, M ) = i d
( 2 )
(3.33)
The last integral is

Z
dy y
d
1
2
(1 + y)
d ( d2 ) (N d2 )
,
= B ,N
2
2
(N )
d
(3.34)
so that finally (R2 = M 2 P 2 )

d N
(N d2 )
IeN (P, M ) = i d/2
M2 P2 2
.
(N 1)!
44
(3.35)
Inserting this into (3.29) and performing the integrations over the Feynman parameters xj then
yields IN . Performing explicitly these xj integrations is often the less obvious part of the story. Let
us summarize: so far, we have shown that, using Feynmans trick, the dimensionally regularized
momentum integration reduces to
Z
d
d
dd k
d/2 (N 2 )
=
i
(R2 ) 2 N .
2
2
N
[k + R i]
(N )
(3.36)
By differentiating IeN (P, M ) with respect to the P one can generate the integrals with factors of
k in the numerator. Indeed, rewriting the integrand in (3.30) as [k 2 2kP + M 2 ]N it is obvious
that
e
(N r 1)!
...
IN r (P, M ) .
(3.37)
IeN1 ...r (P, M ) = r
2 (N 1)! P1
Pr
Plugging in the explicit expression (3.35) for IeN r (P, M ) one gets
d
d
1 ...r
d/2 (N r 2 )
2
2 2 N +r
e
(3.38)
.
.
.
M
IN
(P, M ) = i
P
2r (N 1)! P1
Pr
R1
R1
P
Integrating this expression, with M 2 = M 2 (xj ) and P 2 = P 2 (xj ), over 0 dx1 . . . 0 dxN (1 xj )
yields the IN1 ...r .
The most commonly encountered cases are r = 1 with
d
d N
d/2 (N 2 )
e
IN (P, M ) = i
P M2 P2 2
,
(N 1)!
(3.39)
and r = 2 where17
d
n
d
d o
(N d2 )
d/2 (N 2 1)
2
2 2 N +1
2
2 2 N
e
+
. (3.40)
IN (P, M ) = i
M P
P P M P
2(N 1)!
(N 1)!
There is an alternative, often simpler way, to compute the integrals IN , IN , etc. First note that a
slight generalization of equations (3.33) to (3.36) yields the useful formula
Z
d
d
d
dd k (k 2 )s
d/2 ( 2 + s)(N 2 s)
=
i
(R2 ) 2 +sN .
d
2
2
N
(k + R i)
( 2 )(N )
(3.41)
Next, it follows from Lorentz (resp. Euclidean rotational) invariance (cf. (3.22)) that
Z
17
k
=0 ,
d k 2
(k + R2 i)N
d
kk
1
d k 2
=
2
N
(k + R i)
d
d
dd k k 2
.
(k 2 + R2 i)N
(3.42)
In the Euclidean, one should replace by .
45
One then has from (3.31), e.g. for r = 2

Z
Z
Z
d
2
d
k
k
dd k
(k
+
P
)(k
+
P
)
d
=
+
P
P
IeN (P, M ) =
d k
(k 2 + R2 )N
d
(k 2 + R2 )N
(k 2 + R2 )N
d
n (N d 1)
o
d
d/2
2 d2 +1N
(N 2 )
2
= i
(3.43)
(R )
+P P
(R2 ) 2 N ,
2(N 1)!
(N 1)!
which, of course, coincides with (3.40).
Finally one needs the expansions of the various -factors as d 4. We always let
d=4 .
(3.44)
Then

d
2
2
=
+ O() ,
2

d
d
2
2
2
1
=
= + 1 + O() ,
2
2d
2

where ' 0.5772 . . . is Eulers constant.
3.3
Wick rotation
In the previous subsection we had to evaluate integrals like

Z
dd k
e
,
IN (P, M ) =
[k 2 + R2 i]N
(3.45)
where R2 = M 2 P 2 . Furthermore, dd k = dk0 dd1~k and k 2 = ~k 2 k02 . Thus the integral explicitly
is
N
Z
Z
1
p

p
. (3.46)
IeN (P, M ) = dd1~k
dk0
2
2
0
2
2
0
~
~
k0
k + R i
k0 +
k + R i
p
p
The integrand has poles at ~k 2 + R2 i0 and at ~k 2 + R2 + i0 , as shown on the left of Fig. 6.
As it is also clear from this figure, one can deform the k0 -integration contour away from the real axis
without crossing any of these poles until on gets the integration contour depicted on the right part
of the figure and denoted 1 2 3 :
Z
Z
dk0 (. . .) =
dk0 (. . .) +
1
dk0 (. . .) +
2
dk0 (. . .) .
(3.47)
Now with 1 and 3 being quarter-circles of radius going to infinity, the corresponding integrals
vanish, so that only the integral over 2 remains. But 2 is the imaginary axis in the complex k0
plane, and if one sets
k0 = ik0E ,
(3.48)
46
Figure 6: The integration contours for the k0 -integration

then, as k0E runs from to +, k0 runs along the imaginary axis, i.e. along 2 . Hence
Z
Z
dk0 f (k0 ) = i
dk0E f (ik0E ) .
(3.49)
Applying this to (3.46) or directly to (3.45) finally gives

N
Z
Z
1
p
p

IeN (P, M ) = i
dd1~k
dk0E
E
E
2
2
2
2
0
~
~
ik0
k +R
ik0 +
k + R i
Z
dd kE
,
(3.50)
= i
[kE2 + R2 ]N
where, of course, kE2 = (k0E )2 + ~k 2 .
What happens if the integrand contains some expression involving k p = ~k p~ k0 p0 ? Obviously, this
E E
E
becomes ~k p~ ik0E p0 = ~k p~ + k0E pE
0 k p if we also let p0 = ip0 . Finally, consider
IeN
(P, M ) =
Doing the Wick rotation yields

Z d
j j
d kE kE
kE
jk
e
,
IN (P, M ) = i
2
2
[kE + R ]N
00
IeN
(P, M ) = i
dd k k k
.
[k 2 + R2 i]N
0 k0
dd kE ()kE
E
[kE 2 + R2 ]N
(3.51)
0j
IeN
(P, M ) = i
Due to the spherical symmetry of the Euclidean integral one has

Z d
Z

2
d kE kE
kE
dd kE kE
1
=
,
2
2
d
[kE + R2 ]N
[kE + R2 ]N
0 kj
dd kE (i)kE
E
.
[kE 2 + R2 ]N
(3.52)
(3.53)
which together with (3.52) gives

1
IeN,d
(P, M ) = i
d
2
dd kE kE
.
2
[kE + R2 ]N
(3.54)
This is consistent with what would have been obtained by first replacing k k by d1 k 2 , and then doing
the Wick rotation. One can summarize all this as the following
47
Recipe for the Wick rotation: do the continuation k0 = i k0E in the integrand, replace similarly all
external momenta that may appear in scalar products by their Euclidean counterparts, replace dd k
by dd kE (with all components of kE real), and multiply the whole integral by a factor of i. Integrands
involving k k f (k 2 ) can be replaced by d1 k 2 f (k 2 ), independently of the Wick rotation.
3.4
Vacuum polarization
The computation of the one-loop vacuum-polarization diagram shown in Fig. 7 is straighforward. To

4
order e2 this is the only contribution to
loop (q), while at order e one would get 2-loop diagrams,
as well as one-loop diagrams with counterterms inserted, as shown in Fig. 8.
k
q
kq
Figure 7: The one-loop vacuum polarization diagram. Of course, although we have drawn the
external photon propagators, they are not to be included in .
Figure 8: Loop contributions to the vacuum polarization at order e4 . The two upper diagrams are
two-loop diagrams, while the two lower diagrams are one-loop diagrams with O(e2 ) counterterms
inserted
Now, applying the Feynman rules to the order e2 diagram of Fig. 7 we get

Z
(i) ik/ + m
i(k/ /q ) + m
4
4
(i)
4
4
(2) e
(2) e
,
i(2) loop,e2 (q) = d k tr
(2)4 k 2 + m2 i
(2)4 (k q)2 + m2 i
(3.55)
where the overall minus sign is due to the fermion loop. Simplifying a bit gives
Z
i e2
/ + m) (i(k/ /q ) + m) ]
4 tr [(ik
loop,e2 (q) =
d
k
.
(3.56)
(2)4
[k 2 + m2 i][(k q)2 + m2 i]
48
This integral clearly diverges for large |k| where the integrand behaves like |k|2 while the
integration measure goes like |k|3 d|k|. If we simply cut off the large |k| values et |k|max = we
R 3
expect the resulting integral to be dominated by a term
|k| d|k| |k|2 2 . If this is the
case one says the integral is quadratically divergent. We have seen that gauge invariance requires
2
(q) = (q 2 q q )(q 2 ) and hence also
q q )loop,e2 (q 2 ). As long as our
loop,e2 (q) = (q
regularization procedure does not destroy gauge invariance this must still be true for the regularized
integrals. In order to manifestly preserve gauge invariance we will use dimensional regularization.
Before doing the computation, let us argue a bit more what we should expect. It is not difficult to
3
see that taking a derivative of
loop,e2 (q) with respect to q results in an integrand behaving as |k|
for large |k| and hence an integral that is less divergent, i.e. behaves as rather than 2 . Taking
one more derivative with respect to q results in an integral that is only logarithmically divergent,
and taking a third derivative with respect to a q gives a convergent integral. This means that the
diverging part of the integral is annihilated by 3 derivatives with respect to the external momentum.
Thus the diverging part of
loop,e2 (q) must be a polynomial of second degree in the q, and by gauge
invariance:
q q )loop,e2 ,div ,
loop,e2 ,div = 0 .
(3.57)
loop,e2 ,div (q) = (q
q
Hence, although loop,e2 will be a non-trivial function of q 2 , its diverging part (the coefficient of the
poles in ) will be constant. From (3.57) we see that we can e.g. extract this divergent part of loop,e2
by taking two derivatives of (3.56) and setting q = 0. The resulting integral clearly is logarithmically
divergent. One often says that gauge invariance has reduced the degree of divergence of the vacuum
polarization from 2 (quadratic) to 0 (logarithmic).
Let us now do the computation. As already mentioned, we choose dimensional regularization as explained
above. Applying also the Feynman trick yields
Z
Z
tr [. . .]
i e2 1
dx dd k
.
(3.58)
loop,e2 (q) =
2
(2)4 0
[(k xq) + m2 + x(1 x)q 2 i]2
The Dirac trace can still be evaluated as in 4 dimensions18 so that

tr [. . .] = 4 k (k q) k (k q) + k(k q) + m2 .
Shifting the integration variable k xq k this becomes
Z
Z
h
4i e2 1
d
loop,e2 (q) =
dx
d
k
(k + xq) (k (1 x)q) (k + xq) (k (1 x)q)
(2)4 0
i
1
+(k + xq)(k (1 x)q) + m2
.
[k 2 + m2 + x(1 x)q 2 ]2
(3.59)
(3.60)
Now one can do the Wick rotation. As explained in the previous subsection, in order for this to make
sense one must also continue the external q to a Euclidean qE

. Furthermore, due to rotational symmetry,
All one needs is that { , } = 2 1 and the only issue concerns tr 1 which can be chosen to be 2d/2 = 4 2/2
or just 4. The ambiguity consisting in an overall 2/2 is of the same type as the ambiguity in choosing to continue
1
1
also the factors (2)
4 to (2)d . In the end this only changes the renormalization constants Zi by finite amounts and
does not affect the renormalized or .
18
49
terms linear in k dont contribute to the integral, while terms k k can be replaced by d1 k 2 which
2 after the continuation. Hence
become d1 kE
2

Z
Z
2 + (1 2 )k 2 + 2x(1 x)q q
m x(1 x)qE
4i e2 1
E E
d
d E
loop,e2 (qE ) =
dx i d kE
.
(3.61)
2
2
(2)4 0
k + m2 + x(1 x)q 2
E
2 (and taking into account that here

Using the formula (3.41) for s = 0 and s = 1 with R2 = m2 + x(1 x)qE
19
we already have done the Wick rotation ) yields
(
Z 1
h
i

d 2 d 2
4e2

2
2
d/2
m
x(1
x)q
+
2x(1
x)q
q
loop,e2 (qE ) =
dx
(R ) 2
E
E
E
(2)4 0
2
)

2 d/2 ( d2 + 1)(1 d2 ) 2 d 1

+ 1
(R ) 2
(3.62)
.
d
( d2 )
Observing that (1 d2 )
(q )
loop,e2 E
( d2 +1)(1 d2 )
( d2 )
= (1 d2 ) d2 (1 d2 ) = ( d2 1)(1 d2 ) = (2 d2 ) we get
Z
h
i

d
4e2
d d/2 1

2
2
2
+
2x(1
x)q
q
dx
m
(R2 ) 2 2
x(1
x)q
R
E
E
E
4
(2)
2
{z
}
|
0
2
2x(1 x)qE
Z
1
2
d
d d/2
8e2
2
2 2 2
q
q
dx
x(1
x)
m
+
x(1
x)q
.
(3.63)
=
E
E
E E
(2)4
2
0

Recall that (q) = q 2 q q (q 2 ), so that we read from (3.63), upon changing back the argument
from qE to q,
Z

d 2
d d/2 1
8e2
2
.
(3.64)
2
dx x(1 x) m2 + x(1 x)q 2 2
loop,e2 (q ) =
4
(2)
2
0
Note that, when doing the dimensional regularization, one has various options. In addition to continuing
d
1
1
d4 k dd k one can also continue (2)
4 (2)d and, as already mentioned, tr 1 = 4 2 2 . Furthermore,
in d dimensions the coupling e no longer would be dimensionless. In order to avoid this, one replaces e by
d
e is some mass scale and then e remains dimensionless.
e2 2 e where
All this results in an additional factor
(C
e)4d where C possibly includes the additional factors 2 or 1/ 2, so that
d 2
2
Z
e2
d d/22 4d 1
m + x(1 x)q 2 2
loop,e2 (q ) = 2 2
C
.
dx x(1 x)
2
2
e2
0
2
(3.65)
Next one sets d = 4 and expands the result in . Recall from (3.45) that (2 d2 ) = ( 2 ) = 2 +O(),
and also a = e log a = 1 + log a + O(2 ), so that

2

d
d d/22 4d

2
C
[. . .] 2 2 =
+ O() 1 log log[. . .] + log C + O(2 )
2

2
2
2
=
log + 2 log C log[. . .] + O() .
(3.66)

R1
Since 0 dx x(1 x) = 61 , we finally get
e2
loop,e2 (q ) = 2
6
2
1
C
+ log 3

2
Z
0

m2 + x(1 x)q 2
dx x(1 x) log
+ O() . (3.67)
e2

19
Indeed in eq. (3.41) the l.h.s. is still Minkowskian, so in using (3.41) for our present Euclidean integral one has to
omit the i on the r.h.s. of (3.41).
50
Recall from (2.69) that also contains the contribution from the counterterm 1 Z3 , and hence
e2 includes 1 Z3 only up to order e2 :
e2 (q 2 ) = (1 Z3 )e2 + loop,e2 (q 2 ) .
The renormalization condition (2.70) then gives

2

e2 1
C
m
1
(Z3 1)e2 = loop,e2 (0) = 2
+ log log
+ O() .
6

2
e2
2
(3.68)
(3.69)
For later reference we note that the divergent part simply is

(Z3 1)e2
=
div
e2 1
.
6 2
(3.70)
Finally (cf (2.71)),

e2 (q 2 ) =
loop,e2 (q 2 ) loop,e2 (0)
e2
=
2 2
Z
0

q2
dx x(1 x) log 1 + x(1 x) 2 + O() .
m
(3.71)
This now has a finite limit as 0, so that one can remove the regularization and simply set = 0.
Note that this renormalized e2 (q 2 ) does not depend on
e nor on the arbitrariness of the continuation
which showed up through the constant C. We also note that e2 (q 2 ) is a monotonuous function of
q2.
Note that e2 (q 2 ) is positive for q 2 > 0 and negative for 4m2 < q 2 < 0, while it develops an
imaginary part for q 2 < 4m2 . This imaginary part translates the possibility that a photon with
such q can yield an on-shell electron-positron pair.
3.5
Electron self energy

k
p
pk
Figure 9: The electron self-energy diagram. As before, the external propagators are not to be
included in .
The electron self-energy diagram shown in Fig. 9 gives after dimensional regularization

Z
i(p/ k/) + m
(i)
4
d
4
(i)
i(2) loop,e2 (p/) = d k
(2) e
(2)4 e ,
4
2
4
2
2
(2) k i
(2) (p k) + m i
51
(3.72)
or after simplifying:
loop,e2 (p/)

dd k
1 i(p/ k/) + m
= ie
(2)4 k 2 i (p k)2 + m2 i
Z
1 (d 2)i(p/ k/) + d m
dd k
= ie2
(2)4 k 2 i (p k)2 + m2 i
2
(3.73)
Introducing the Feynman parameter and shifting the integration variables gives
Z 1
Z
(d 2)i(p/ k/) + d m
dd k
2
loop,e2 (p/) = ie
dx
4
(2) [(k xp)2 + x(1 x)p2 + xm2 i]2
0
Z 1
Z
0
dd k 0 (d 2)i (1 x)p/ k/ + d m
2
= ie
dx
(2)4 [k 0 2 + x(1 x)p2 + xm2 i]2
0
Z 1
Z
dd k
1
2
= ie
dx [(d 2) (1 x)ip/ + d m]
4
2
(2) [k + x(1 x)p2 + xm2 i]2
0
Z

2
d
e2
d 1
d/2
2 2 2
=
dx
[(d
2)
(1
x)ip
/
+
d
m]
xm
+
x(1
x)p
.
(2)4
2 0
(3.74)
As for the vacuum polarization, we introduce (C

e)4d to keep e dimensionless and to allow for the other
options in doing the dimensional continuation. Expanding in = 4 d as above in eq. (3.66) gives

Z 1
e2
C2
xm2 + x(1 x)p2
2
[(d 2) (1 x)ip/ + d m]
loop,e2 (p/) =
dx
+
log
log
16 2 0

e2
(

e2
C2
d2
2
=
+ log
ip/ + d m
16 2

2
)
Z 1
xm2 + x(1 x)p2
.
(3.75)
dx [(d 2) (1 x)ip/ + d m] log
e2
0
Now we can drop the O() terms to get
e2
loop,e2 (p/) =
16 2

2
C2
(ip/ + 4m) ip/ 2m + log
(ip/ + 4m)

)
Z 1
xm2 + x(1 x)p2
dx [2(1 x)ip/ + 4m] log

.
e2
0
(3.76)
As discussed in general above, the self-energy (p/) also receives contributions from the counterterms,
cf. eq. (2.38):
e2 (p/) = (Z2 1)e2 (ip/ + m) + (Z2 m)e2 + loop,e2 (p/) .
(3.77)
The renormalization conditions then fix Z2 and m according to (2.40) and (2.41) which now read
(Z2 1)e2 = i
so that
2
me2 =
e
m
8 2

loop,e2 (p/)
p/
/=im
p
3 1 + 3
C
+ 3 log

2
52
me2 = loop,e2 (im) ,
dx(1 + x) log
0
xm
e2
(3.78)
)
,
(3.79)
and
e2
(Z2 1)e2 = 2
8
1 1+
C
+ log

2
)

x2 m2 2(1 x2 )
dx (1 x) log 2 +
.
e
x
(3.80)
Note that the dx-integral for Z2 diverges as x 0. This is actually an infrared divergence which
occurs for on-shell electrons and is related to the possibility of emitting soft (very low energy) photons.
The proper treatment of such infrared divergences would be a chapter by itself. Let us only say that
when summing appropriate diagrams corresponding to physically measurable and distinguishable
situations such infrared divergences cancel. Note that the UV-divergent part simply is

(Z2 1)e2
=
div
e2 1
.
8 2
(3.81)
We finally get
e2 (p/) =
e
8 2
dx
(
h
(1 x) ip/ + 2m log
h m2 + (1 x)p2 i
xm2
2(1 x )
(ip/ + m)
x
(3.82)
Although there are infrared divergences as x 0 as just discussed, all ultraviolet divergences ( 1 poles)
have cancelled.
3.6
Vertex function
p
pk
k
pp
pk
Figure 10: The one-loop vertex function diagram. The external photon and electron propagators are
not to be included in (p0 , p).
The vertex function (p0 , p) was defined in (2.47). Since we are computing with renormalized fields
(and not bare fields) we will automatically compute the renormalized vertex function, and not the
bare one. Note that (p0 , p) is normalized such that its lowest order contribution is just , cf.
eq. (2.48), while the tree vertex Feynman diagram would give a (2)4 e . Taking this into account
53
and applying the Feynman rules to the one-loop diagram shown in Fig. 10 gives
Z
(i) i(p/0 k/) + m
(i) i(p/ k/) + m
0
loop,e2 (p , p) =
d4 k (2)4 e
4
0
2
2
(2) (p k) + m i
(2)4 (p k)2 + m2 i
1
(i)
(2)4 e
4
2
(2) k i
Z
0
4
d k i(p/ k/) + m
i(p/ k/) + m
1
= i e2
2
. (3.83)
4
0
2
2
2
2
(2)
(p k) + m i
(p k) + m i k i
R 4
For large |k| the integral behaves like dk4k which diverges logarithmically. We have to introduce
again some regularization, e.g. dimensional regularization as before. Actually, this diagram could
2
also be regularized by simply including an additional factor k2M
into the photon propagator with
+M 2
M being taken to in the end. It can be shown that this does not affect the gauge invariance. Of
in L2 :
course, there is also a contribution from the counterterm ie (Z2 1)A
e2 (p0 , p) = loop,e2 (p0 , p) + (Z2 1) .
(3.84)
We will not do the complete computation of loop,e2 (p0 , p) here. However, we will extract its divergent
piece and show that it is precisely cancelled by the counterterm with the same Z2 as already determined from the self-energy. We will also extract a certain finite part which gives the first higher-order
correction to the magnetic moment of the electron, the famous g 2.
3.6.1
Cancellation of the divergent piece
The divergent part of loop,e2 (p0 , p) arises from the large |k| limit of the integrand. In this limit one
can neglect the external momenta and the mass so that with dimensional regularization
Z

dd k ik/
ik/
1

2
loop,e2 (0, 0)
= ie
2
2
2
4
(2)
k i
k i k i
m=0
Z
2
d
d k k 2k k
= i e2 (d 2)
(2)4
(k 2 i)3
Z
(d 2)2
dd k
1
= i e2
.
(3.85)
d
(2)4 (k 2 i)2
The trouble with this integral is that for d < 4 it is UV convergent but IR divergent, and vice versa
for d > 4 (and UV and IR divergent for d = 4). To avoid the IR divergence, we should have kept
the electron mass m. If we are only interested in the UV divergent behavior it is enough to keep m
in the denominators so that

2 Z
dd k
1

2 (d 2)
loop,e2 (0, 0)
i e
4
2
d
(2) (k + m2 i)2
div
(d 2)2 d/2
d 2 d 2
e2
(m ) 2
d
(2)4
2
e2 1
.
(3.86)
8 2

54
This divergence should be cancelled by the diverging part of (Z2 1) . However, Z2 is already
determined from the above computation of the electron self-energy and, hence, it is by no means
obvious that this cancellation does indeed take place. Nevertheless, we know from the Ward identity
(2.52) that if is finite, then (p, p), and in particular (0, 0) must also be finite. It follows
that the divergences must cancel between (3.86)
and (Z2 1) . Of course, this is confirmed by the

e2 1
explicit expression (3.81) of Z2 : (Z2 1)e2 = 8
2 . We conclude that the renormalized vertex
div
function is finite (at least to the order we computed).

3.6.2
The magnetic moment of the electron: g 2
Historically, the computation of the vertex function has played an important role since it gives the
first correction to the magnetic moment of the electron. This magnetic moment is usually expressed
e
s, where s j = 12 is the spin of the electron. The tree-level result
through the g-factor as = g 2m
which can also be obtained from studying the Dirac equation in a magnetic field is g = 2. Let us
sketch how to obtain the corrections to this result.
First note that one is interested in evaluating (p0 , p) between on-shell wave-functions u(p0 , 0 )
and u(p, ) as appropriate when computing e.g. the S-matrix elements between an incoming electron
of momentum p and polarization and an outgoing one with p0 and 0 . The interaction is with an
electromagnetic field A (q) where q = p0 p. Now (p0 , p) can involve various products of , p/
0
and p/0 , together with p and p . Moving all p/ to the right till one can use p/u(p, ) = imu(p, ) and
moving all p/0 to the left until u(p0 , 0 )p/0 = u(p0 , 0 )im, one is left with the general structure
n
o
i
1
u(p0 , 0 ) (p0 , p)u(p, ) = u(p0 , 0 ) F (q 2 )
G(q 2 ) (p+p0 ) +
H(q 2 ) (pp0 ) u(p, ) , (3.87)
2m
2m
where the coefficient functions can depend on the only scalar available, i.e. q 2 = 2m2 2p p0 and,
of course, on m. If one contracts this equation with (p p0 ) , the left-hand-side vanishes by the Ward
q2
identity (2.54), while the right-hand-side equals 2m
H(q 2 )
uu and thus
H(q 2 ) = 0 .
(3.88)
The following identity is valid between the on-shell wave-functions u(p0 , 0 ) and u(p, ) :
i[ , ](p0 p) = i (p/0 p/) i(p/0 p/) = 2i(p + p0 ) + 4m . We can use it to rewrite
(3.87) as
n
o

i
0
0
0
0
0
2
2
2
0
u(p , ) (p , p)u(p, ) = u(p , ) F (q ) + G(q )
G(q ) [ , ](p p) u(p, )
4m
n

i
F (q 2 ) + G(q 2 ) (p + p0 )
= u(p0 , 0 )
2m
o
i
2
0
F (q ) [ , ](p p) u(p, ) ,
(3.89)
+
4m
This form is particularly useful when studying the limit where q = p0 p 0. The Ward identity
(2.53) states u(p, 0 ) (p, p)u(p, ) = u(p, 0 ) u(p, ), so that
F (0) + G(0) = 1 .
55
(3.90)
Indeed, the vertex function captures the quantum corrections to the coupling of the electron to the
electromagnetic field. It contributes ie u(p0 , 0 ) (p0 , p)u(p, )A (q) to an effective Lint = Hint .
For p = p0 and in the rest frame of the electron this is just +eA0 = (e)A0 , stating that e is indeed
the charge of the electron one can measure. To determine the magnetic moment, consider the second
rewriting in (3.89). The term (p + p0 ) is blind to the spin and cannot contribute to the magnetic
moment. The second term yields a contribution to the effective interaction Hamiltonian
ie
i
F (q 2 ) u(p0 , 0 )[ , ]u(p, )(p0 p) ieA (q) =
F (q 2 ) u(p0 , 0 )[ , ]F (q)u(p, ) , (3.91)
4m
8m
where F (q) = iq A (q) iq A (q). In an almost static situation (i.e. to first order in q), one has
~ S
~0 so that the spin-dependent terms in Hint are e F (0)B
~ S
~
u(0, 0 )[ , ]F (0)u(0, ) = 8iB
m
~ ~, with the magnetic moment of the electron being
B
e
~ g e S
~ ,
~ =
F (0) S
(3.92)
m
2m
so that the celebrated g-factor equals
g = 2 F (0) = 2 2 G(0) ,
(3.93)
where we used (3.90).

It remains to explicitly compute G(0) at one loop. To do so, we only need to keep the part
(p + p0 ) in (p0 , p), while we can drop the part . Due to the explicit factor (p + p0 ) we
expect this part to be given by a finite (converging) integral. We start with (3.83) and introduce
dimensional regularization and Feynman parameters as usual:
loop,e2 (p0 , p)

Z 1 Z 1x Z
(p/0 k/) + im (p/ k/) + im
2i e2
d
dx
dy d k n
=
o3 . (3.94)

(2)4 0
0
x (p0 k)2 + m2 + y (p k)2 + m2 + (1 x y)k 2 i
The actual computation is a bit lengthy: First one does the -matrix algebra in the numerator. The

3
denominator is (k xp0 yp)2 + (x + y)2 m2 + xyq 2 i and one shifts k k + xp0 + yp and drops
the terms linear in k. The result still contains many terms, but taking between u(p0 ) and u(p)
this can be further simplified as above. Dropping then the terms one gets
Z 1 Z 1x h

i
e2

0
0
u(p )loop,e2 (p , p)u(p) = 4 m dx
dy u (x+y) (1y)p +(1x)p0 yp xp0 u Id (x, y, q 2 )
2
p+p0
0
o
(3.95)
2
where Id (x, y, q ) is a convergent integral for d < 6. For d = 4 it equals
Z

3 i 2
1
2
I4 (x, y, q ) = d4 k k 2 + (x + y)2 m2 + xyq 2 i
=
(x + y)2 m2 + xyq 2
.
(3.96)
2
R 1 R 1x
R1 R1
R1 R1
Writing 0 dx 0 dy f (x, y) = 0 dx 0 dy (1 x y)f (x, y) = 0 dx 0 dy (1 x y)f (y, x), we
see that we can symmetrize the integrand of (3.95) in x and y, and thus in p and p0 :
Z 1 Z 1x

ie2
(x + y)(2 x y) (x + y)

0
0
0
u(p )loop,e2 (p , p)u(p) = 2 m u(p + p ) u dx
, (3.97)
dy
8
(x + y)2 m2 + xyq 2
p+p0
0
0
56
from which we identify

e2
G(q ) = 2 m2
4
2
1x
dx
0
dy
0
(x + y)(2 x y) (x + y)
.
(x + y)2 m2 + xyq 2
(3.98)
We only need
so that
e2
G(0) = 2
4
e2
g =2 1+ 2
8
dx
0
1x
dy
0
1xy
e2
= 2 ,
x+y
8

=2 1+
2

g2=
(3.99)
(3.100)
This is the classical result of Schwinger. Since then, the art of measuring and computing g 2 has
been pushed to an extreme refinement (four loops in QED!) with an excellent agreement.
3.7
One-loop radiative corrections in scalar 4
The Ward identity of QED linked the charge or coupling constant renormalization to the photon
wave-function renormalization, see eq. (3.1). In general though, such a relation is not expected, and
here we will briefly discuss scalar 4 theory where the coupling constant gets renormalized separately.
We start with (cf. eqs. (2.17), (2.18) and (2.19))
gB
1
1
L = ( B )2 m2B 2B 4B ,
2
2
4!
and let
B =
m2B = m2 m2 ,
Z ,
so that
gB =
(3.101)
Zg
g,
Z2
(3.102)
L = L0 + L1 + L2 ,
(3.103)
1
1
L0 = ( )2 m2 2 ,
(3.104)
2
2
g
L1 = 4 ,
(3.105)
4!
1
1
1
g
L2 = (Z 1)( )2 (Z 1)m2 2 + Zm2 2 (Zg 1)4 .
(3.106)
2
2
2
4!

1
0
2
2
2
Propagator: Recall that the complete propagator is (q) = q + m (q ) i
with
(q 2 ) = (Z1)(q 2 +m2 )+Zm2 +loop (q 2 ) = loop (q 2 )loop (m2 )(q 2 +m2 ) dqd2 loop (q 2 )|q2 =m2 .
Then the one-loop contribution of order g to the one-particle irreducible complete propagator is
given by20
k
i(2)
20
loop,g (q 2 )
1
=
2

(i)
1
i
d k
ig(2)4 = g
4
2
2
(2) k + m i
2
d
dd kE
,
kE2 + m2 i
(3.107)
There are also one-loop contributions with additional insertions of the counterterms. These are higher order in g.
57
where the factor

we get21
1
2
loop,g (q 2 )
comes from the symmetry factor of the diagram. In dimensional regularization

d d2
g 4d d/2
e
1
m
=
2
(2)4
2
2

g
2

m2
= m2
1
log
1
+
O()
1
log
2
(2)4
2

2
e2

1 + log 1 1
m2
g
2
m
log
.
=
(3.108)
16 2

2
2
e2
Since this is q-independent, Z = 1 at this order (cf. (2.27)):

Z = 1 + O(g 2 ) .
(3.109)
For the mass renormalization we then get from (2.26)

g
1 + log 1 1
m2
2
2
2
m
mg = loop,g (m ) =
log 2
16 2

2
2
e
so that
g (q 2 ) = 0 .
(3.110)
(3.111)
4-point function: We define F (q1 , q2 q10 , q20 ) as the amputated four-point function, i.e. with
the external (full) propagators removed and normalized such that to lowest order F = g. Then the
connected two-particle to two-particle scattering S-matrix element is given by
1
p
F (q1 , q2 q10 , q20 ) , (3.112)
S c (q1 , q2 q10 , q20 ) = i(2)4 (4) (q1 + q2 q10 q20 )
6
(2) 16E1 E2 E10 E20
with all momenta being on-shell. In the present 4 theory, the amputated 4-point function is just
the renormalized 1PI 4-point vertex function (4) and, taking into account the normalization of F
we have22
(4) (q1 , q2 , q10 , q20 ) = (2)4 (4) (q1 + q2 q10 q20 ) F (q1 , q2 q10 , q20 ) .
(3.113)
It follows that, up to this order g 2 , the four-point function i(2)4 F is given by the tree-level vertex
i(2)4 g, the one-loop diagrams shown in Fig. 11, as well as the contribution from the counterterms
i(2)4 g(Zg 1). Hence
F = (q1 , q2 q10 , q20 ) = g + Floop (q1 , q2 q10 , q20 ) + g(Zg 1) ,
(3.114)
where in dimensional regularization

i(2)4 Floop,g2 (q1 , q2 q10 , q20 )
2 ( Z

1
(i)
1
2
d
i(2)4 g
d
k
=
2
(2)4
[(k + q1 ) + m2 i][(q2 k)2 + m2 i]
)
+(q2 q10 ) + (q1 q20 )
(3.115)
21
Just as in QED, if we want to keep a dimensionless coupling constant for d 6= 4 we must replace the coupling in
the d-dimensional Lagrangian by ge
4d , where
e is some arbitrary mass scale.
22
Recall that we usually take the arguments of Green functions or vertex functions as incoming four-momenta, i.e.
an outgoing qi appears as qi .
58
q1
q2
q1
q
2
q1
q2
q2
q1
q1
q1
q2
q2
Figure 11: The three one-loop contributions at order g 2 to the 4-point function. Other one-loop
contributions also involve counterterm insertions and are of higher order in g.
Introducing the Feynman parameter, as well as the standard notation
s = (q1 + q2 )2 ,
t = (q1 q10 )2 ,
u = (q1 q20 )2 ,
(3.116)
the denominator written explicitly in (3.115) becomes

2
2
k + 2xq1 k + xq12 2(1 x)q2 k + (1 x)q22 + m2 i

2
2
2
=
k + xq1 (1 x)q2 xq1 (1 x)q2 + xq12 + (1 x)q22 + m2 i

2
2
=
k + xq1 (1 x)q2 x(1 x)s + m2 i .
[. . .][. . .] =
(3.117)
We then get
i g2
Floop,g2 (q1 , q2 q10 , q20 ) =
2(2)4
(Z
dx
0
)
dd k
+ (s t) + (s u) , (3.118)
[k 2 + m2 sx(1 x) i]2
or after the by now familiar Wick rotation and evaluation of the Euclidean integral:
)
(
Z 1

2
d
g
d
Floop,g2 (q1 , q2 q10 , q20 ) =
d/2 2
dx [m2 sx(1x)] 2 2 +(s t)+(s u) . (3.119)
2(2)4
2 0
As already mentioned, if we want to keep a dimensionless coupling constant for d 6= 4 we must replace the
coupling in the d-dimensional Lagrangian by g
4d , where
is some arbitrary mass scale. This results in an
4d
extra factor
accompanying every factor of g. However, we want to normalize F such that its tree-level
d
value is just g, not g
4d , so that we just need to include a single factor of
4d = (
2 )2 2 in the r.h.s.
of (3.119). This then nicely combines with the terms in the braces to make them dimensionless. Finally,
expanding in = 4 d we get
(
)

Z 1
2
2 sx(1 x)
g
m
2
Floop,g2 (q1 , q2 q10 , q20 ) =
log
dx log
+ (s t) + (s u)
32 2

2
0
)

Z 1 (
m2 sx(1 x)
3g 2 2
g2
=
log +
dx log
+ (s t) + (s u) .
(3.120)
32 2
32 2 0
The full four-point function up to order g 2 is given by adding the tree-level and the counterterm
contributions according to (3.114). Just as the wave-function renormalization factors Z, Z2 and Z3
are fixed by requiring the full propagator to satisfy certain normalization conditions, we must impose
some condition on the full four-point function to fix Zg . A rather standard condition is to require
59
that it equals g at the symmetric (off-shell) point s = t = u = 34 2 , where is some mass scale
(which may or maynot equal the
already introduced):

4 2
F s=t=u= =g .
(3.121)
3
Using this condition in (3.114) fixes Zg as

g(Zg 1) = Floop
4
s = t = u = 2
3

.
Up to the order we computed this gives

Z 1
m2 + 43 2 x(1 x)
3g
2
dx log
.
(Zg 1)order g =
log
32 2
2
0
(3.122)
(3.123)
Substituting this back into (3.114) or, equivalently, subtracting (3.122) from (3.114) finally gives
(
)
2
2 Z 1
g
m
sx(1
x)
dx log 2 4 2
Fg2 (q1 , q2 q10 , q20 ) = g +
+ (s t) + (s u) .
(3.124)
32 2 0
m + 3 x(1 x)
Note that the argument of the logarithm may be negative for s 2m, so that F then has an
imaginary part. This is in agreement with the optical theorem (unitarity). Note also that the
dependence on the scale
has cancelled upon imposing the renormalization condition. However, the
latter condition has introduced a physically relevant scale : it is the scale where the measurable
coupling g is defined. In a massive theory, one could conveniently take = m or = 0, but lets
stay more general.
60
PART II :
RENORMALIZATION AND RENORMALIZATION GROUP
4
General renomalization theory
In section 2, we have seen that the notion of renormalization arises in order to keep correctly normalized full propagators, with their poles at the physical masses. This discussion was independent
of any divergences arising in loop integrals. In the previous section, we have computed various oneloop two-, three- and four-point functions. The one-loop contributions were divergent and had to be
regulated. We have observed that after performing the renormalization, according to the conditions
formulated in section 2, the renormalized quantities turned out to be finite (i.e. have finite limits
even if the regulator is removed). The purpose of the general renormalization theory is to show
that this is no accident but remains valid to all orders in perturbation theory, at least for so-called
renormalizable theories as QED or scalar 4 -theory.
4.1
Degree of divergence
The superficial degree of divergence D characterizes the behaviour of the momentum integral
R 4
d k1 . . . d4 kL [. . .] when all |ki | with a common k . More precicely: if the integral
R
behaves as k D1 dk then D is the superficial degree of divergence. An integral with D 0 is
called superficially divergent, and one with D < 0 is called superficially convergent. This does not
necessarily mean the integral really diverges or converges, but we will see that the superficial degree
of divergence D nevertheless plays an important role. Often, we will talk somewhat loosely about
the superficial degree of divergence D of a diagram, meaning the D of the associated integral. To
determine D let
If be the number of internal lines of the field f ,
Ef be the number of external lines of the field f ,
Ni the number of vertices of type i with di derivatives and nif fields f attached.
Now, propagators behave as f (k) k 2+2sf , were
for scalars: sf = 0,
for spin 1/2: sf = 1/2,
for massive spin 1: sf = 1 since (q)
1
q2
q q
m2
and
sf = 0, since when coupled to a conserved current: (q)
for photons
or gravitons:
1
q q2q q2 .
q2
61
It follows that
D = 4L +
Ni di +
If (2 + 2sf ) .
P
P
Now use the relation (cf. (1.45)) L = I V + 1 = f If i Ni + 1 to obtain
X
X
D =4+
Ni (di 4) +
If (2 + 2sf ) .
i
Use further
(4.1)
(4.2)
Ni nif = 2If + Ef to get

X
X
X
D =4+
Ni (di 4 +
nif (sf + 1)
Ef (sf + 1) ,
i
(4.3)
or
D =4
Ef (sf + 1)
Ni i ,
(4.4)
with
i = 4 di
nif (sf + 1) .
(4.5)
Note that this remains valid in an arbitrary space-time dimension d if one replaces 4 d.
There is an alternative derivation of (4.4), which works easily as long as only scalars, spin 12
particles, photons or gravitons are involved (no massive spin 1). Define the dimension of a field
R
ik(xy)
f as given by the behavior of its propagator: hT (f (x) f (y))i d4 k ek22sf has (mass) dimension
4 (2 2sf ) = 2(1 + sf ) and hence the field f has (mass) dimension Df = 1 + sf . (This does not
work for massive spin 1 fields that have sf = 1 but (mass) dimension 1, just as photons - it is the
explicit appearence of the mass m which messes up the argument.) Then any interaction of type i
P
involving nif fields f and di derivatives has dimension f nif Df + di . Since the Lagrangian must
P
have dimension 4, the coupling constant gi must have dimension [gi ] = 4 f nif Df di or
[gi ] = i .
(4.6)
Now in a diagram with Ni interactions of type i and Ef external lines for fields of type f the
P
corresponding hT (f (x1 ) . . .)i has dimension
f Ef Df . Its Fourier transform then has dimension
P
P
(4)
pj ) of dimension 4,
f Ef (Df 4). This dimension must arise from the products of an overall (
P
P
propagators for the external lines of total dimension f Ef (2+2sf ) = f Ef (2Df 4), the coupling
P
constants of dimension i Ni i and the momentum integrals whose dimension equals the superficial
P
P
P
degree of divergence D of the diagram. Hence f Ef (Df 4) = 4+ f Ef (2Df 4)+ i Ni i +D,
or
X
X
D =4
Ef Df
Ni i ,
(4.7)
i
in agreement with (4.4). As an example, in Table 2, we consider the various interactions in QED,
where one has D = 32 and DA = 1 and one indeed verifies that the i coincide with the dimensions
of the coupling constants.
62
/
ieA
(Z3 1)F F
/
(Z2 1)
[(Z2 1)m + Z2 m]
= 4 2 23 1 = 0
=422=0
= 4 2 23 1 = 0
= 4 2 32 = 1
Table 1: The values of for the various interactions in QED

One can similarly define assign comp
to composite operators Oa (x) inserted in a diagram. comp
is
a
a
the same as one would define for the corresponding interaction, and eq. (4.4) would be replaced by
P
P
P
D = 4 f Ef Df i Ni i a Na comp
.
a
The importance of the notion of superficial degree of divergence D resides in the following remarks:
P
If i 0 for all i then only Greens functions with 4 f Ef Df 0 can have D 0 and,
hence, there are only finitely many superficially divergent Greens functions. Note that, unless
i > 0, i, these are still infinitely many diagrams (arbitrary Ni for fixed Ef ).
Interactions with i 0 are called renormalizable interactions. Theories with only renormalizable interactions are called renormalizable theories. In such theories only finitely many Greens
functions are superficially divergent. We will see that in such theories all Greens functions
can be rendered finite by a finite number of counterterms corresponding to the redefinition of
a finite number of physical constants (couplings and masses) and the (re)normalizations of the
fields.
If i < 0 for some interaction then an infinite number of Greens functions (arbitrary numbers
Ef of external fields) are divergent and one needs an infinite number of counterterms to make
them finite. Such theories are called non-renomalizable.
The importance of the notion of superficial divergence is partly due to the following
Theorem (Weinberg): If D < 0 for the complete integration and any sub-integration (i.e. holding
some linear combination of momenta fixed) then the integral is really convergent.
In particular, at one-loop, there are no sub-integrations and, hence, by this theorem, any one-loop
integral with D < 0 is convergent. Note also that this theorem does not say anything about the
divergence of integrals with D > 0. We have seen examples in QED where the integrals are less
divergent than expected. The vacuum polarization diagram e.g. had D = 2, but gauge invariance
allows one to pull out two factors of external momentum by writing = (q 2 q q )(q 2 )
with (q 2 ) only logarithmically divergent (D = 0).
4.2
Structure of the divergences
Suppose an integral has D < 0, i.e. it is superficially convergent. This means that if all k are
simultaneously taken to be large the integral converges. Thus the only possibility for this integral not
to converge is that it diverges if some combination of momenta is held fixed and the divergence is due
to the sub-integration over the other momenta. This must necessarily correspond to a subdiagram.
63
Hence a superficially convergent diagram can only be divergent due to a diverging subdiagram (with
D0 0), as shown e.g. in Fig. 12. However, such subdiagrams are of lower order in perturbation
theory. If they have already been rendered finite by the addition of appropriate conterterms at the
lower order in perturbation theory, one no longer has to worry about such diverging subdiagrams
any more. Henceforth, we will assume that such diverging subdiagrams have been taken care of and
that superficially convergent diagrams are convergent.
Figure 12: A superficially convergent diagram containing a divergent vacuum-polarization subdiagram.
If an integral I(pi ) depending on external momenta pi has superficial degree of divergence D 0,

then by differentiating with respect to the external momenta lowers D by one unit since
1
2(k + pi )
.
=
2
2
pi (k + pi ) + m i
[(k + pi )2 + m2 i]2
(4.8)
(There is a caveat to this argument to be discussed soon.) Differentiating D + 1 times results in

an integral having degree of divergence 1, i.e. which is superficially convergent. According to the
above remarks it is then convergent. Hence
pi11
1+1 I(pi ) = J(pi ) ,

. . . piD+1
(4.9)
where J(pi ) is convergent. Upon integrating with respect to the external momenta pi we get
b i) ,
I(pi ) = PD (pi ) + I(p
(4.10)
with PD (pi ) a polynomial of order D in the external momenta with (a priori) divergent coefficients
b i ) a convergent integral. Obviously, the diverging part PD (pi ) just are the first D + 1 terms
and I(p
in a Taylor expansion of I(pi ). We may then rephrase this by saying that I(pi ) minus the first D + 1
terms in a Taylor expansion in the external momenta is finite. We have seen examples of this in our
one-loop computations, see e.g. (3.77) and (3.78) for the electron self-energy which one can rewrite
as
loop,e2 (p/) = (Z2 1)e2 (ip/ + m) (Z2 m)e2 + e2 (p/)

(ip/ + m) + loop,e2 (im) + e2 (p/) .
= i loop,e2 (p/)
p/
/=im
p
64
(4.11)
The first two terms indeed constitute a first order polynomial in the external p/ which are the first
two terms in a Taylor expansion of loop,e2 in p/ around p/ = im, while the last term e2 (p/) is the
renormalized finite electron self-energy.
It is very important that the divergent part of the integral I(pi ) is a polynomial in the external
momenta since it is precisely such divergences that can be cancelled by local counterterms. Indeed,
a local counterterm is of the form
Ef
YY
(i,f )di,f f (x) ,
f
di,f = d ,
(4.12)
i=1
being the product of Ef fields of type f with a total number of d derivatives. Such a counterterms
gives a (tree-level) contribution to the diagram with Ef external lines of type f which is a polynomial
of order d in the momenta. Thus, to cancel the diverging PD (pi ) one can always find a sum of such
a local counterterms involving up to D derivatives. Now, such a counterterm will also appear as
a new vertex in loop diagrams and it is important to check that it does not render the theory less
P
renormalizable. Indeed, with at most D derivatives, we have c.t. 4 f Ef Df D. But in
P
a renormalizable theory the possible values of D are constrained by D 4 f Ef Df so that
P
4 f Ef Df D 0 and c.t. 0, so the counterterm is part of the finitely many renormalizable
interactions. Moreover, if one includes all (finitely many) renormalizable interactions in the bare
Lagrangian, then all possible counterterms are necessarily of the same form as the terms already
present, and they just renormalize the couplings, masses and wave-functions
p+q
p+q
q
pp
Figure 13: A two-loop diagram with overlapping divergences: two divergent subdiagrams share a
common propagator.
There is a caveat in the argument that taking a derivative with respect to the external momenta
lowers the superficial degree of divergence by one unit and which has to do with so-called overlapping
divergences. An overlapping divergence occurs if two divergent subdiagrams share a common line
(propagator), see e.g. the two-loop cotribution to the vacuum polarization shown in Fig. 13. Then
taking a derivative with respect to an external momentum typically results in a sum of terms with
the large momentum behaviour in one of the sub-loops improved in one term but not in the other,
and vice versa for the other sub-loop. To be specific, for the example of Fig. 13 one has (neglecting
the electron mass for simplicity)

Z 4 4 0
d pd p
1
1
1
1
overlap
(q)
tr
.
(4.13)
(p p0 )2
p/0 p/0 + /q p/ + /q p/
65
Taking a derivative with respect to the external momenum q gives a sum of two integrals:

Z 4 4 0
overlap
d pd p
1
1 1
1
1
tr
.
(q)
+ 0 0
q
(p p0 )2
p/0 (p0 + q)2 p/ + /q p/
p/ p/ + /q (p + q)2 p/
(4.14)
0
In the first term the large p behavior is improved but not the large p behaviour, while in the second
term things are reversed. Of course, the superficial degree of divergence is lowered by one unit, and
taking two more derivatives would result in a sum of terms all with a D0 = 1, but the criteria of
the theorem cited above are not satisfied and one cannot conclude that up to a divergent subgraph
the integral is convergent. Although this is true for each individual term in the sum, the trouble is
that for each term the diverging subgraph is a different one. For example, in a given term in the
sum, the divergence may be traced to a divergent p sub-integration corresponding to the divergent
subgraph 1 in Fig. 14, while in another term the divergence would be due to the p0 sub-integration
corresponding to the subdiagram 2 in Fig. 14.
2
3
Figure 14: The nested sequences of subdiagrams used in the BPHZ construction: 1 3 and
2 3 .
Although such overlapping divergences are more complicated, the previous discussion shows that
one still has a sum of divergences associated with the various subdiagrams. To deal with such overlapping divergences in a systematic way and show that the usual counterterms are exactly sufficient to
cancel these divergences, too, was the achievement of Bogoliubov, Parasiuk, Hepp and Zimmermann
(BPHZ).
4.3
Bogoliubov-Parasiuk-Hepp-Zimmermann prescription and theorem
One defines a forest U as a family of nested (sub)graphs i : if 1 U and 2 U then either

1 2 or 2 1 or 1 2 = /0. For the diagram of Fig. 14 e.g., 1 and 2 cannot be in the same
forest. More generally, overlapping loops cannot be in the same forest. On the other hand, in Fig. 12
the vacuum polarization subdiagram and the whole diagram can be in the same forest. Hence the
notion of forest is exactly what one needs to distinguish overlapping subdiagrams (divergences) from
no-overlapping ones. Again, in Fig. 14 the list of all forests is:
U0 = /0 ,
U1 = {1 } ,
U2 = {2 } ,
U3 = {3 } ,
66
U4 = {1 , 3 } ,
U5 = {2 , 3 } .
(4.15)
For each diagram G one then considers such a family F(G) of all forests of G. Then we have the
BPHZ prescription: For each forest Ui F(G) consider the nested sequence a b . . . of the
(sub)graphs in Ui . Starting with the innermost a , one defines a subtraction term S(Ui ) by replacing
the integrand Ib of each subdiagram b (in the nested sequence) with superficial degree of divergence
Db by the first Db + 1 terms in its Taylor expansion (e.g. around 0) in the momenta flowing into
(or out of) this subdiagram. Since at each step one only keeps a polynomial in the corresponding
momenta, the final subtraction term is still a polynomial in the external momenta.
To see how this works, consider the diagram of Fig. 13, resp. Fig. 14:
U0 = /0 : no subtraction, S(U0 ) = 0.
U1 = {1 } : This subdiagram is just the vertex function e1loop (p0 +q, p0 ) and we know from our
one-loop computations that e1loop (p0 + q, p0 ) = eL + efinite (p0 + q, p0 ), with finite (p0 + q, p0 )
being the finite (renormalized) vertex function and L = (Z2 1)e2 . This subdiagram has
D = 0 and hence the corresponding subtraction term is just the first term in the Taylor
expansion: S(U1 ) = e[(Z2 1)e2 + c] with c = finite (0, 0) and (Z2 1)e2 given in (3.80).
.
Obviously, this corresponds to a counterterm A
U2 = {2 } : This similarly gives the same subtraction term S(U2 ) = e[(Z2 1)e2 + c] .
U3 = {3 } : This is just the full diagram we want to study. It has D = 2 and thus the
subtraction term must be a second-order polynomial in the q. By gauge invariance this must
be of the form ( q 2 q q )C with C = overlap (q 2 = 0). We are not claiming that this is the
only divergence of this diagram, but BPHZ tell us that this is the subtraction term for this
forest: S(U3 ) = ( q 2 q q ) overlap (q 2 = 0). This corresponds to an order e4 contribution
to Z3 .
U4 = {1 , 3 } : Here we must first replace the integrand of the 1 -part by S(U1 ) as defined
above, so that the integrand for the 3 part is the same as for
1loop but with e in the
left vertex replaced by e[(Z2 1)e2 c] . Then the subtraction term is S(U4 ) = (q 2
q q )[(Z2 1)e2 c](Z3 1)e2 . This also corresponds to an order e4 contribution to Z3 .
U5 = {2 , 3 } : same subtraction term as for U4 .
BPHZ theorem: The original (regularized) diagram minus the sum of all subtraction terms is finite.
The latter is exactly generated by the (usual) counterterms.
We will not give the proof of this theorem, but just mention that it proceeds by recursion. One
assumes that all diagrams at order N are made finite by the counterterms. Then one finds a recursion
relation for the large momentum behavior at the next order N + 1 (including on subspaces p0 ||p ,
etc, as relevant for overlapping divergences). The solution to this recurrence relation uses the sum
over forests. It can then be shown that the original diagram minus the sum of all subtraction terms
is indeed finite.
67
4.4
Summary of the renormalization program and proof
i is the (mass) dimension of the coupling gi (provided one only deals with scalars, spin- 12 ,
photons or gravitons). Renormalizable interactions have i 0, a renormalizable theory only
has interactions with i 0. The superficial degree of divergence (divergence in the region
P
P
where all ki ) is D = 4 f Ef Df i Ni i . In a renormalizable theory only a finite
set of Ef gives D 0. Diagrams with D < 0 are superficially convergent.
A superficially convergent diagram can be divergent only due to divergent subdiagrams. Then
the same counterterms that render finite these subdiagrams also make the whole diagram finite.
Thus we only have to deal with superficially divergent diagrams, i.e. only with finitely many
Ef if the theory is renormalizable.
Ignoring overlapping divergences for the moment, a superficially divergent diagram with D 0
is made superficially convergent by D + 1 derivatives with respect to the external momenta.
This implies that its diverging part is a polynomial in the external momenta and this can
be subtracted by a local counterterm. In a renormalizable theory, even though at each order
in perturbation theory there are new diverging diagrams, for fixed Ef , the degree D cannot
increase and the structure of the polynomial, i.e. of the counterterms remains the same, only
the coefficients (the Zs) get higher and higher order contributions. Since D 0 only for
finitely many Ef , only finitely many counterterms are needed. Adding these counterterms
renders all these diagrams finite to all orders in perturbation theory. On the other hand, in
a non-renormalizable theory, at every order in perturbation theory, diagrams with more and
more external lines become divergent and one needs an infinite number of counterterms to
render them all finite.
To deal with overlapping divergences, BPHZ define the forests with nested boxes (subdiagrams)
a of degree Da = D(a ), each forest giving a subtraction term obtained by replacing each
subdiagram in the nested sequence by its Da + 1 first terms in a Taylor series in the external
momenta. This corresponds to the same counterterms as defined above. BPHZ show that
this procedure renders also finite these diagrams with overlapping divergences to all orders in
perturbation theory. Again, in a renormalizable theory, finitely many counterterms make all
diagrams finite.
4.5
The criterion of renormalizability
P
Renormalizability restricts interactions to those with i = 4 di f nif (1 + sf ) = 4 di
P
f nif Df 0 only. For a finite number of fields these are only finitely many interactions. Of
course, there are also other restrictions on the possible interactions like Lorentz invariance or other
symmetries one might want to preserve.
If the regulator preserves a given symmetry then the finite and the diverging parts of any Greens
function (not necessarily of individual diagrams) at any order in perturbation theory must also
preserve this symmetry. Hence the counterterms also respect this symmetry.
68
F
What if one allowed to include non-renormalizable interactions, like e.g. adding a term
to the QED Lagrangian? This interaction indeed has dimension 5 so that = 1 and it is nonrenormalizable. For dimensional reasons its coupling constant must be of the form M with dimensionless and some mass scale M . Such a term would change the magnetic moment of electrons by
e
an additional amount m
which would thus become an adjustable parameter. From experiment
M
we know that this quantity must be extremely small. Said differently, if we take of the same order
as e, the mass scale M must be very large compared to me .
More generally, a non-renormalizable interaction has i < 0 and coupling constants of dimension
i = |i |. Write gi = Mge|i i | . Then non-renormalized diagrams have divergences that behave as

|i |N
(e
gi )N M
at order N in perturbation theory, being some UV cutoff. Also, infinitely many
Greens functions are divergent and we need infinitely many counterterms, possibly with an arbitrary
number of derivatives. However, once these divergences are cancelled, we have renormalized Greens
p | |N
functions G(pi ), and at order N in perturbation theory (in gi ) they will behave as (e
gi )N Mj i .
Thus, as long as |pj | M one can neglect the effect of these non-renormalizable interactions.
Theories like QED are presently thought to be only effective theories, in the sense that they
provide the effective description of electromagnetic interactions at energies that are low compared
to some scale at which new physics could be expected, like e.g. the grand unification scale of 1015
GeV or even the Planck scale of 1019 GeV. Such an effective theory then has an effective Lagrangian
obtained by integrating out the very heavy additional fields that are present in such theories. (We
will discuss such integrating out a bit in the next section). This necessarily results in the generation
of (infinitely) many non-renormalizable interactions in this effective Lagrangian with couplings Mge|i i | ,
M being e.g. the grand unification scale. From the previous argument it is then clear that at energies
well below this scale these additional non-renormalizable interactions are completely irrelevant, and
this is why we only see the renormalizable interactions. Our low-energy world is described by
renormalizable theories like QED not because such theories are somehow better behaved, but because
these are the only relevant ones at low energies:
Renormalizable interactions are those that are relevant at low energies, while non-renormalizable
interactions are irrelevant at low energies.
This is also the teminology in statistical mechanics where one studies infrared physics at scales
|p| a1 , where a is the lattice spacing.
69
5
5.1
5.1.1
Renormalization group and Callan-Szymanzik equations

Running coupling constant and -function: examples
Scalar 4 -theory
Consider the 4-point vertex function F in the scalar 4 -theory we computed in section 3. At tree-level
it equals g, independent of the momenta, while up to order g 2 it was given by (cf. (3.124))
(
)
2 Z 1
2
g
m
sx(1
x)
Fg2 (q1 , q2 q10 , q20 ) = g +
dx log 2 4 2
+ (s t) + (s u) ,
(5.1)
32 2 0
m + 3 x(1 x)
which depends on the momenta or, equivalently, on s, t and u. It also depends on the scale used
to fix the renormalization condition (3.121). Below, we will consider different scales, so let us call
the present scale . To emphasize the dependence of F on we will write F (s, t, u; ) and call
the corresponding coupling g . Thus

4 2
F s = t = u = ; = g ,
(5.2)
3
and eq. (5.1) is rewritten as
g2
F (s, t, u; ) = g + 2
32
Z
0
)
m2 sx(1 x)
dx log 2 4 2
+ (s t) + (s u) .
m + 3 x(1 x)
(5.3)
Recall from section 3, that this is essentially the 1PI 4-point function, since (4) (q1 , q2 , q10 , q20 ) =
(2)4 (4) (q1 + q2 q10 q20 ) F (q1 , q2 q10 , q20 ).
Let us remark that the condition (5.2) can be viewed as defining what we mean by the coupling
constant of our theory. It is one convenient definition and obviously one might have chosen a
different one. For example, one might have defined g in terms of the function F and a scale but
at cs s = ct t = cu u = 2 with some unequal coefficients cs , ct , cu . One might also have defined the
coupling not in terms of F , resp. (4) , but in terms of the 4-point Green function at some convenient
off-shell point. We will come back to this later-on, for the time being we keep the definition (5.2).
To make the discussion of the following paragraph more intuitive, suppose m 6= 0 and that
one has chosen = m. Then, as long as the momenta remain of the same order of magnitude
R1
as m, the argument of the logarithm is of order 1, and 0 dx{. . .} = O(1), so that the one-loop
g
correction is small with respect to the tree contribution as long as 32
2 1. If, on the other hand,
one is interested sin extreme high-energy scattering with s, |t|, |u| m2 , then the logarithms will
1
x(1x)
x(1x)
s
be large: log 1+m42x(1x) ' log m
2 log 1+ 4 x(1x) . To fix the ideas, suppose m is of the order of 1
3
3
MeV and s of the order of 100 MeV. Then log ms2 ' 10 which is large but is easily compensated
1
3
by the 32
, and the one-loop contribution will certainly be small with respect to the
2 ' 3 10
tree-level result if g is small. However, as one considers higher and higher energy scales, the log
p p
will increase further. For s, |t|, |u| ' 10 TeV e.g. the sum of the 3 logarithms is about 100.
Eventually, perturbation theory will break down as one goes to very high energies. Actually, like
70
for most quantum field theories, the perturbation series in g has a vanishing radius of convergence.
Nevertheless, the series is asymptotic, meaning that one gets a good approximation, if g is small
enough, by computing the first few orders in g. The approximation will be better for smaller g
or, at fixed g, if large logarithms like the ones discussed do not appear. In this sense, to improve
perturbation theory one has to avoid such large logarithms. This can be done as follows.
If one is only interested in the high-energy scattering, one might define the coupling constant g
p p
right away at a scale comparable to s, |t|, |u|. Then no large logarithms will ever appear23 .
In general, however, one wants to be able to compute at different scales and most often the original
definition of the coupling g is at low or zero momentum, like the definition of the elementary charge e
in QED. The solution is to define a different coupling g() for every scale as the value of F (s, t, u; )
at s = t = u = 43 2 :

Z 1
m2 + 34 2 x(1 x)
3g2
4 2
dx
log
g() = F s = t = u = ; = g +
.
(5.4)
3
32 2 0
m2 + 43 2 x(1 x)
Again, if the logarithm will be large, but this can be easily avoided by first defining g(1 )
at some 1 just slightly larger than , and then use this g1 to compute a g2 at a scale 2 slightly
larger than 1 , etc. More precisely, introduce intermediate 0 = , 1 , . . . N 1 , N = and define
Z
m2 + 34 2n+1 x(1 x)
3g(n )2 1
dx
log
.
(5.5)
g(n+1 ) = g(n ) +
32 2 0
m2 + 43 2n x(1 x)
Iterating this relation allows us to go from g( ) = g to g(N ) = g() while keeping the logarithms
small at every step.24 Thus eq. (5.5) provides the desired relation between the g() at different
scales. Let us insist that no large logarithms will appear when computing scattering amplitudes at
typical s, t, u of order 2 if one uses this coupling g() and, in this sense, g() is the natural coupling
constant at this scale. Obviously, as increases, g() also increases.
It is much more convenient to turn this relation into a differential equation by considering n =
and n+1 = + with infinitesimal . We get
Z 1
4 2
x(1 x)
d
3
2
3
g()
dx
g() =
(, m) .
(5.6)
2
d
16
m2 + 34 2 x(1 x)
0
Let us be slightly more precise. What one does is to express g(0 ) in terms of 0 , g() and and
take the derivative with respect to 0 at fixed (and fixed g()) and set 0 = in the end. With
d
g(). In general one defines the functions (, m) and
this being understood, one simply writes d
() as
(, m) =
d
g() ,
d
() = (, 0) .
(5.7)
23
except in certain kinematical regions like forward scattering

Note that g(n+1 ) differs from g(n ) only by a term of order g 2 . Thus, to the order we work, we could just as
well replace the g(n ) in front of the integral by g(0 ) = g . If one does so, iterating eq. (5.5) N times exactly yields
eq. (5.4).
24
71
Obviously, if m one can neglect the effect of the mass m and approximate (, m) by ().
Equation (5.6) yields for the 4 -theory:
() =
3
g()2 0 g()2 .
16 2
(5.8)
For the present 4 -theory, the -function is positive and the differential equation shows that for
positive (negative) -function the coupling increases (decreases) when the scale increases. It is
easy to solve the differential equation (5.7) with (, m) replaced by (5.8):
d
g() = 0 g()2
d log
1
1
= log
g(1 ) g(2 )
2
1
0
.
(5.9)
One sees again that g() increases when increases. A useful rewriting of this solution is as follows:
1
exp
M ,
0 g()

independent of .
(5.10)
The quantity M is called the renormalization group invariant scale, since the change in g() is exactly
such that this quantity does not depend on . Being independent of , we expect the quantity M
1
to have a physical meaning. Indeed, if = M we have 0 g()
= 0 and we see that M is the scale
where g() . Of course, we have only done a one-loop computation and as g() becomes large
we no longer can trust a one-loop result. What one can say is the following: even if one starts at
some scale with a small g( ), as the scale is increased, the coupling grows and for scales of the
order of M the theory enters a strong-coupling regime. We may rewrite eq. (5.9) as
h
i1
.
(5.11)
g() = g(0 ) 1 0 g(0 ) log
0
This is valid for any 0 m (since we neglected the mass) and as long as g() is small, whether
g(0 ) log 0 is small or not.
It is useful to note that, in practice, one can compute (, m) directly form (5.4) by taking
d
and replacing g2 by g()2 . Note though, that one should not take dd of (5.3) as this would
d
result in the opposite sign. If one traces back our computation one sees that the -function is the
coefficient of the log -term which is minus the coefficient of the log
-term in Floop which, in turn is
1
the coefficient of the -pole of the Zg -factor times g, cf. (3.123). :
()1loop = coefficient of
1
in Zg g.

(5.12)
We have also seen that, at one loop in the 4 theory, the wave function renormalization factor is
Z = 1. Thus, we can just as well say that the one-loop -function is the coefficient of 1 of ZZg2 g = gB :
()1loop = coefficient of
1
in gB when expressed in terms of g.

(5.13)
If we go to two loops, however, Z 6= 1 and the two expressions for the -function become different.
Also in other theories, like 3 in 6 dimensions, Z 6= 1 already at one loop. Does this mean that
72
in general (5.12) is right and (5.13) is wrong? The answer depends on how exactly one defines the
coupling constant g. Obviously, if it is defined in terms of the 1PI vertex function, eq. (5.12) should
be correct. More generally, one might define couplings gn in terms of the renormalized amputated
n-point Green functions as is relevant for computing (on-shell) S-matrix elements at typical energy
scales . On the other hand, 1PI vertex functions and full propagators at off-shell momenta appear
as building blocks of larger Feynman diagrams, as we have extensively discussed in the previous
section. Thus one not only has to avoid large logarithms in the 1PI vertex functions but also in
the full propagators. It then makes sense to include corrections from the full propagators into the
definition of the coupling g(). Since each propagator is linked to two 1PI vertex function, its
contributions should be split between the two vertices, assigning half its contribution to each. Again,
if one traces the appearance of the log and 1 terms, one sees that taking into account half the
contribution of a full propagator amounts to picking the coefficient25 of the 1 -pole in 1Z . In a k
1
and with this new definition of the coupling g the
theory this results in an extra factor of Z k/2
-function would indeed be given by (5.13).
Since the coefficients of the 1 poles are determined by the divergent parts of the one-loop integrals
only, this actually gives a very easy way to get the leading coefficient of the -function in a large
class of theories. We will discuss this in somewhat more detail below.
5.1.2
QED
To see how the effective coupling of QED evolves with the scale of energy, consider how the tree-level
coupling of a photon to an electron gets corrected by one-loop effects. Following the above discussion,
we think of a complicated Feynman diagram built from 1PI-vertices , complete electron propagators
and complete photon propagators, all with off-shell momenta, as shown in Fig. 15. (Note that we
use only the electron-electron-photon 1PI vertex explicitly, not the higher vertex functions. Hence
the resulting diagram is not necessarily a tree diagram.)
Then we need to include half of the corrections coming from the propagators and the entire
correction coming from the 1PI-vertex function to define what we mean by the running coupling
e(). At the one-loop level, there are the following diagrams to take into account. First, the photon
propagator gets corrected by the one-loop vacuum-polarization diagram. We have seen that this
e2
1
2
2
2
modifies e2 1(q
2 ) ' e (1 + (q )) or e e(1 + 2 (q )). Next, the electron propagators get
corrected by the self-energy diagrams on the fermion lines (there are two of them but each counts
half). Finally, there is the one-loop corrected vertex . The latter two are related by the Ward
identity. Recall that their divergencies were cancelled by the same counterterm Z2 1 and,
similarly, one can see that their contributions to the -function cancel. (Schematically, the vertex
diagram contributes the coefficient of 1 in Z2 , while each of the two electron self-energy diagrams
contribute the coefficient of 1 in 1Z2 , giving a vanishing total contribution.) Thus, the effective
25
Recall that Z = 1 + z with z of order g at least. Thus
1
Z
' 1 z2 .
73
Figure 15: A complicated Feynman diagram in QED is built from the 1PI electron-electron-photon
vertex functions, the full electron propagators and the full photon propagators.
coupling is entirely determined by the vacuum polarization diagram:

1
2
2
e e(q ) = e 1 + (q )
2Z
1

q2
e3
dx x(1 x) log 1 + x(1 x) 2 .
= e+ 2
4 0
m
(5.14)
Repeating the argument done for the coupling in the 4 -theory, we get
Z 1

2
e3
dx x(1 x) log 1 + x(1 x) 2 ,
e() = e + 2
4 0
me
and
d
e3
(, m) = e() = 2
d
2
dx
0
(5.15)
x2 (1 x)2
,
m2
+
x(1
x)
2
(5.16)
which yields the -function of QED

e3
.
e () =
12 2
1/2
Note that in QED we have eB = Z3

and again
e
e with Z3 = 1 6
2
(5.17)
1

e3
+ finite so that eB = e+ 12
2
1

+ finite
1
in eB when expressed interms of e.
(5.18)

The -function is again positive and e grows with the energy scale. Sometimes, this is written in
e2
d
e
terms of the running of the fine-structure constant = 4
as = d log
= 2
e ():
e ()1loop = coefficient of
() =
2 2
,
3
74
(5.19)
2
which is of the same form as (5.8) with 0 = 3
. The -function (5.17) or (5.19) governs the
running of e() or () for values of much larger than me , say = 10 me . To get the running
for 0 , one should use the explicit expression (5.15) instead. The latter is perturbatively
2
e2
valid in this region since 4

2 log m2 . log 100 ' 0.01. As discussed in connection with the Ward
e
identity, the renormalized charge of the electron is defined at zero momentum, hence eR = e(0) and
R1
5
1
R = (0) ' 137
. Thus ( 0 dx x(1 x) log x(1 x) = 18
)
Z

22 1
2
( ) = +
dx x(1 x) log 1 + x(1 x) 2
0
me
2 Z 1

2
2
' +
dx x(1 x) log x(1 x) + log 2
0
me
h
2
2
5 i1
5

2
+ log
' 1
log
.
= +
3
6
me
3
me 6
(5.20)
d
For > , one can safely use the solution of d
() = () with initial value given by (5.20)
and given by (5.19):
h
2
i1
() = ( ) 1
( ) log
.
(5.21)
3
Combining (5.20) and (5.21) yields

h
2
5 i1
() = 1
log
,
3
me 6
(5.22)
where we neglected a term O(3 ). This equation is valid even when log me is of order 1. However,

3
just as for the scalar 4 theory, it shows that the coupling becomes strong as ' M = me exp 2
'
10280 me . Of more experimental relevance is the value of () at present day collider energies. For
1
= 100 GeV e.g. (LEP) one gets () ' 134.6
. This is the result when taking into account the
electron field only. Including the effects of the two other lepton families (as well as the quarks) one
1
gets (100 GeV) ' 128
.
5.2
Running coupling constant and -functions: general discussion
Having seen the running of the coupling and the corresponding -functions in two explicit examples,
we will now try to make some more general statements.
5.2.1
Several mass scales
We have seen in the example of QED with only electrons and positrons that the running of e() or
() was well determined by the -function for scales above a few electron masses. Below, one
had to use the exact form (5.15) or its approximation (5.20). However, in the real world, there are
many more charged particles of higher masses that also contribute to the -function. Let us consider
the case of two well-separated mass scales, 0 < m1 m2 , the generalization to more than two mass
scales will be obvious.
75
Suppose the one-loop computation yields (cf. (5.4) with = 0 or (5.15) written for () instead
of e())
Z
n

2
2 o
2
g() = g + g dx f1 (x) log 1 + h1 (x) 2 + f2 (x) log 1 + h2 (x) 2
,
(5.23)
m1
m2
where x could stand for multiple Feynman parameters xi , and thefj (x) and hj (x) are some polynomials. Let, much as before
Z
n
2h2 (x)2 o
d
2h1 (x)2
2
+
f
(x)
,
(5.24)
1,2 (, m1 , m2 ) = g() = g dx f1 (x) 2
2
d
m1 + h1 (x)2
m22 + h2 (x)2
Then, as long as m1 m2 , one sees from (5.23) that g() ' g . As becomes comparable
to m1 one still has m2 , and one can neglect the second logarithm. Actually, as long as is
considerably smaller than m2 (exactly how much smaller depends on h2 and f2 /f1 ), one can continue
to neglect the second term in the braces: in this region, the heavy particle does not contribute to
the running of g(). If, in this region, the first logarithm becomes large for m1 m2 , one must
use the differential equation with (, m1 , m2 ) to evolve g(). However, for m1 m2 , one can
neglect the second term in the braces in (5.24), while in the first term one can set m1 = 0. Hence,
in this region, the running is effectively governed by
Z
2
1 () = 1,2 (, 0, ) = g dx 2f1 (x) .
(5.25)
As becomes of the order of m2 , one has to use an expression similar to (5.23) with 6= 0 to evolve
g() from some < m2 to some > m2 beyond which it is reasonable to neglect the effect of m2
and one can use
Z

2
1+2 () = 1,2 (, 0, 0) = g dx 2f1 (x) + 2f2 (x)
(5.26)
to evolve g() further to even larger scales . It is often enough to simply evolve g() with 1 ()
from m1 to m2 and then with 1+2 () above m2 .
The general lesson to remember is that at a given scale the running of the coupling is determined
by the -function due to loops of only those particles that have masses below while particle with
masses much larger than do not influence the running.
5.2.2
Relation between the one-loop -function and the counterterms
We have already seen in the 4 -theory and in QED that, with the appropriate definition of the
running coupling constant, the one-loop -function coincided with the coefficient of the 1 pole in the
expression of the bare coupling in terms of the renormalized one. We will now establish this relation
more generally.
According to the above discussion, in general, we define the one-loop corrected n-point coupling as
given by the value of an appropriate proper n-point vertex function (n) , plus half of the propagator
corrections (as shown in Fig. 15 for the example of QED), all computed up to one loop and evaluated
at some conveniently chosen momenta pj ():
(n)
g() =
n
1X

pj () +
g j (pj ) j pj () .
2 j=1
76
(5.27)
As before, (n) and are the renormalized vertex function and 1PI propagator (self-energy).
It will be useful here to separate the contributions from the one-loop diagrams and those of the
counterterms:
(n)
(n) (pj ) = g + (Zg 1)g + 1loop (pj )

j (pj ) = (Zj 1) j (pj )1 + Zj m2j + j,1loop (pj ) .
(5.28)
(n)
We will restrict ourselves to the cases where the one-loop contribution 1loop diverges logarithmically
and the original coupling g is dimensionless in 4 dimensions. This includes most of the interesting

d
g()mj =0 , i.e. we are interested in the
theories. Moreover, since we want to compute () = d
region where |p2j | m2i , we can drop all terms involving the masses right away (except maybe to
regulate IR divergences). In particular, m2j m2j , so these terms will also be dropped. Thus
g() = g + (Zg 1)g +
(n)
1loop
n
1X

g (Zj 1) + gj (pj ) j,1loop pj () .
pj () +
2 j=1
(5.29)
The -function will be given by the coefficient of log in this expression. We will show shortly that
each log in a one-loop contribution is accompanied by a 1 with the same coefficient, and that the
finite renormalization conditions require that this is also the coefficient of + 1 in the corresponding
combination of the counterterms. Hence:
"
#
n
1X
1
in g(Zg 1)
g(Zj 1) .
(5.30)
1loop () = coefficient of

2 j=1
Note that, since Zg 1 and Zj 1 are always at least O(g), one has to the order we are interested in
i
h
Zg
1 + (Zg 1)
1X
gB g p = g p
(Zj 1) ,
(5.31)
= g 1 + (Zg 1)
2 j
j Z j
j 1 + (Zj 1)
and, hence
1loop () = coefficient of
1
in gB when expressed in terms of g.

(5.32)
Let us now give the argument: Using dimensional regularization and Feynman parameters, the one-loop
contribution to the 1PI vertex function is of the form

Z Y

R(pj , xj ) /2
2
(n)
n
1loop (pj ) = cn g
+a
dxj f (xj )
,
(5.33)

e2
j
where R(pj , xj ) is some quadratic form in the momenta and masses and
e is some scale introduced to keep
g dimensionless after dimensionally continuing to d = 4 . We let
Z Y
b0 = 2cn
dxj f (xj ) ,
(5.34)
j
so that
(n)
1loop (pj ) =
b0 n a
g b0 g n + cn g n

2
Z Y
j
77

dxj f (xj ) log
R(pj , xj )
e2

.
(5.35)
As before, we impose a renormalization condition that (n) should equal g for some fixed pj pj ( ).
Combining (5.28) and (5.35) fixes Zg as

Z Y
R(pj ( ), xj )
b0 n a
n
n
g + b0 g cn g
dxj f (xj ) log
.
(5.36)
(Zg 1)g =

2
e2
j
The (renormalized) n-point vertex function then reads

Z Y
(n) (pj ) = g + cn g n
dxj f (xj ) log
j
R(pj , xj )
R(pj ( ), xj )

.
(5.37)
We need the coefficient of log in this expression when pj = pj (), for sufficiently large so that we can
neglect all masses. Now, R(pj (), xj ) necessarily is of the form
R(pj (), xj ) = r(xj )2 +
prs (xj )mr ms
r,s
Thus the coefficient of log in (n) (pj ()) is 2cn g n
RQ

d

=2,
log R(pj (), xj )
d
mr =0
(5.38)
dxj f (xj ) = b0 g n which is also the coefficient of
1

in (Zg 1)g, cf. (5.36). The same argument can be repeated to show that the coefficient of log in

j pj () j pj () is the same as the coefficient of 1 in (Zj 1). This completes the argument.
5.2.3
Scheme independence of the first two coefficients of the -function
We have just seen that the result for the -function does not depend much on the details of the
renormalization conditions. One might ask what happens if one uses a different regularization or
makes coupling constant redefinitions. We will show the following result:
As long as one defines the coupling such that
(g) = b0 g 2 + b1 g 3 + b2 g 4 + . . . ,
(5.39)
the first two coefficients b0 and b1 are universal, i.e. are unchanged by changing the renormalization
scheme or redefining the coupling by higher-order terms.
To prove this, assume that in a different scheme with a different coupling ge one finds
e g ) = eb0 ge2 + eb1 ge3 + eb2 ge4 + . . . .
(e
(5.40)
Now, the bare coupling is scheme independent and must equal g and ge to lowest order: gb = g +
g
a1 g 2 + . . . and gb = ge + e
a1 ge2 + . . .. It follows that ge = g + ag 2 + O(g 3 ). Hence de
= 1 + 2ag + O(g 2 ).
dg
e.g. It follows
Also, ge should be a function of g only and not of m

d
de
g
de
g 2
d
3
e
(e
g ) = ge = g
= (g)
= b0 g + b1 g + . . . 1 + 2ag + . . .
d
d dg
dg

2

g 2 + . . . + b1 + 2ab0 ge3 + . . .
= b0 g 2 + b1 + 2ab0 g 3 + . . . = b0 ge ae
= b0 ge2 + b1 ge3 + . . . .
(5.41)
Hence eb0 = b0 and eb1 = b1 , as was claimed.

78
5.3
-functions and asymptotic behaviors of the coupling
Let us now study the different asymptotic behaviours of the running coupling constant, given the
different possible forms of the -function. Here we will be mainly interested in the UV asymptotics,
i.e. the behaviour as the scale becomes very large. In a massless theory however, the limit where
becomes very small (IR behaviour) is also of interest, in particular in the description of critical
phenomena in statistical physics.
d
d
The -function being defined as (g()) = d
g() = d log
g(), this gives a differential equation
for g() if (g) is known. This differential equation is integrated as

Z
g(2 )
g(1 )
5.3.1
dg
2
= log
.
(g)
1
(5.42)
case a : the coupling diverges at a finite scale M
Suppose (g) > 0. Then the coupling g() increases as the scale is increased. Suppose furthermore
R dg
that (g) increases fast enough with g so that the integral g(1 ) (g)
converges. Let the value of
M
this integral be log 1 . Comparing with (5.42) we see that at the scale M the coupling diverges:
g(M ) = . This is shown in the left part of Fig. 16. Explicitly, M is given by

Z
dg
M = exp
.
(5.43)
g() (g)
Of course, the running of g() is precisely such that the r.h.s. does not depend on . For this reason,
M is called the renormalization group invariant mass. Note that, even if we start with a massless
theory, the renormalization group equation (5.42) asserts that there is a well-defined mass scale in
this theory! Equation (5.43) allows to trade g() for M and vice versa. This is sometimes referred
to as dimensional transmutation.
We have seen that this would be the behaviour in scalar 4 -theory or QED if we could trust
R dg
1
the one-loop result (g) = b0 g 2 with b0 > 0 (g = for QED) since then g() (g)
= b0 g()
and

1
. We have also remarked that for QED this M is extremely large and certainly
M = exp b0 g()
well beyond the energy scales at which QED has to be embedded in a larger and/or more fundamental
theory. Of course, as already emphasized, unless one knows (g) exactly, perturbation theory must
break down as gets closer to M and g() becomes large. In any case, as becomes of the order of
M , the theory enters a strong-coupling regime.
5.3.2
case b : the coupling continues to grow with the scale
Rg
dg 0
Suppose (g) > 0 but (g) increases slowly enough with g so that the integral g(1 ) (g
0 ) diverges as
g . In this case, (5.42) shows that g() as , but for any finite the coupling g()
remains finite. This situation is depicted in the right part of Fig. 16.
79
Figure 16: Shown are the -functions and corresponding scale dependence of the couplings for case
a (left) and case b (right).
5.3.3
case c : existence of a UV fixed point
Suppose that (g) starts out positive for small g and has a zero at some finite g with 0 (g ) < 0
(see the left part of Fig. 17):
(g ) = 0 , 0 (g ) a < 0 .
(5.44)
Then, if one starts with some initial g(1 ) < g , the -function is positive and the coupling will
increase as is increased. For some large enough 2 , g(2 ) will become close to g and we can
approximate (g) ' a(g g ). Equation (5.42) then gives
log
=
2
g()
g(2 )
dg
g() g
' a log
(g)
g(2 ) g
g() g ' (g(2 ) g )
a
.
(5.45)
We see that, as , the coupling g() is driven to g . For this reason, g is called a UV fixed
point:
g > 0 is a UV fixed point if (g ) = 0 and 0 (g ) < 0.
(5.46)
So far, we have considered positive (g) for small enough g. This resulted in a coupling that
increased as was increased.
80
Figure 17: Shown are the -functions and corresponding scale dependence of the couplings for case
c (left) and case d (right).
5.3.4
case d : asymptotic freedom
Now suppose that (g) < 0 at least for 0 < g < g . Then, as long as we start with a g(1 ) < g , the
coupling g() will decrease as is increased (see the right part of Fig. 17).
R g( ) dg0
To be more specific, suppose g 1 (g
This will be always realized in
0 ) diverges as g 0.
n
n+1
perturbation theory since (g) g + O(g ) with n 0. Then (5.42) tells us that as one
has g() 0: the theory becomes a free theory in the UV limit. This is called asymptotic freedom.
To see this in more detail, suppose (with b0 < 0)
(g) = |b0 |g n + O(g n+1 ) ,
n2.
(5.47)
Neglecting the higher order contributions to (since g becomes small this will be more and more
justified as becomes large!), we have
1
1
=
log
|b0 |
g(1 )
g()
dg
1
=
gn
(n 1)|b0 |
1
1
g(1 )n1 g()n1

.
(5.48)
One sees that as one must have g() 0. Indeed

1
n1
n1
g() = g(1 ) 1 + (n 1)|b0 |g(1 )
log
.
1
81
(5.49)
On the other hand, as is decreased, the coupling g() increases. It is easy to see at which scale the
coupling must become strong. Again, we see from (5.48) that the combination

1
1
M = exp
= exp
(5.50)
(n 1)|b0 |g()n1
(n 1)b0 g()n1
does not depend on (at least, as long as one can trust the lowest order result for (g)). It is again
a renormalization group invariant scale. As M the argument of the exponent must vanish and
hence g(M ) . Of course, perturbation theory breaks down before reaches M , but this shows
that M is the scale where the coupling becomes large. The difference with case a considered before
is that now this strong coupling region is approached from large values of as is decreased.
Figure 18: Shown are the -function and corresponding scale dependence of the coupling for case e
(IR fixed point).
5.3.5
case e : IR fixed point
Finally, consider a situation where the -function starts out negative for small g and then has a
zero for some g , necessarily with 0 (g ) = a > 0. In this case, if at some initial scale 1 one has
g(1 ) < g , the -function is negative. This means that g() decreases as is increased, and g()
increases as is decreased, just as for case d, above. However, as gets smaller, and g() gets larger,
the -function becomes less negative and as g() approaches g it can be well approximated by its
linearized form (just as for case c): (g) ' a(g g ). Suppose that g(2 ) is close enough to g so
that the linearized approximation is good enough. Then
log
2
g(2 ) g
' a log
,
g() g
(5.51)
and g() g as 0, i.e. g is an IR fixed point:

g > 0 is an IR fixed point if (g ) = 0 and 0 (g ) > 0.
(5.52)
Such IR fixed points are of particular interest in the study of critical phenomena in statistical physics.
82
5.4
Callan-Symanzik equation for a massless theory
I most textbooks, when presenting the Callan-Symanzik equation, the discussion is somewhat simplified by switching from dimensional regularization to an explicit UV cut-off . This has the advatage
that one does not have to deal with the extra mass scale
e introduced in dimensional regularization
(to keep the coupling g dimensionless) and which is different from the renormalization group scale
. Thus with an explicit UV cut-off one only has to deal with and , while in dimensional regularization one has to deal with , and
e. We will nevertheless derive the Callan-Symanzik equations
entirely within the framework of dimensional regularization. Indeed, the presence of
e presents only
a slight complication of the discussion.
5.4.1
Renormalization conditions at scale
We have seen that one can define a running coupling constant g() as the value of (a certain combination of 1PI propagators and) appropriate vertex function evaluated at the scale , cf eq. (5.27).
Of course, this equation (5.27) can be viewed as imposing a renormalization condition on the vertex
function at the scale . It is then most convenient to also impose the renormalization condition on
the full propagator at scale and accordingly require for the corresponding 1PI propagator that
2
(p )
=0,
(5.53)
p2
p2 =2
while still keeping the condition that the pole be at the physical mass:
(m2 ) = 0 .
(5.54)
(Note that p2 = 2 corresponds to a space-like momentum.) These two equations are for scalars,
but their generalization to Dirac fields or the electromagnetic field is obvious as can be seen on the
examples to follow.
We will now restrict ourselves to the massless case. This will also include the case where , pj m
and m can be neglected, just as in our computation of () = (, m = 0). Lets look at the example
of massless QED. Then one has for the electron self energy (cf. (3.76) and (3.77))
h2
e2
C2
p2 i
ip/ + log
+ 1 log 2 ,
e2 (p/) = (Z2 1)e2 ip/ + (Z2 m)e2
(5.55)
16 2

e
where
e was some fixed mass scale introduced to keep the coupling e dimensionless in d = 4 . For
m = 0 the condition (im) = 0 simply yields
m = 0 .
(5.56)
26
This was
The normalization condition (5.53) now reads
to be expected for a massless theory.

(p/)
= 0 and yields
p
/
/=i
p
(Z2 1)e2 =
26
e2 h 1
C
1+
i
+
log
log
,
8 2
2
(5.57)
Although expected, it is generally not true if we use a different regularization.
83
and
h
i
e2
p2
ip
/
log
2
.
(5.58)
16 2
2
In particular, this is now free from infrared divergences!
Similarly, for the vacuum polarization (photon propagator) one gets from (3.67) and (3.68), now
with m = 0
C
5 1
q2 i
e2 h 1
2
+ log + log 2 .
(5.59)
e2 (q ) = (Z3 1)e2 2
6
e
2 6 2
e2 (p/) =
Our new condition (5.53) can be easily seen to translate into (2 ) = 0, i.e. (Z3 1) = loop . One
gets
C
i
e2 h 1
5
+ log + log
,
(5.60)
(Z3 1)e2 = 2
6
e
2 6
and
e2
q2
e2 (q 2 ) =
log
.
(5.61)
12 2
2
What is the general lesson we learn from these examples? First, we note the obvious fact that
the 1PI functions are simpler in the massless case than in the massive one. Second, we see that the
renormalized 1PI functions only depend on the coupling e, or rather e(), and explicitly on via the
2
2
dimensionless combination p 2 or q 2 . In addition, they may depend polynomially on the momenta.
They have a well-defined finite limit as 0 and they do not depend on
e. Third, the Z-factors
depend on (they have poles 1 ), on the coupling e() and on the dimensionless ratio e .
Finally, we must formalize a bit more the definition of the -function of the preceding subsection.
We will assume that the coupling is dimensionless as appropriate for a renormalizable coupling.
As explained above, the running coupling should be defined in terms of the appropriate vertex
function and half the sum of the corresponding 1PI propagators evaluated at some reference momenta
pj () j . Let us call F the corresponding combination of vertex function and 1PI propagators.
Obviously, at tree level we simply have F = g. Beyond tree level, F will depend on g as well as
on the momenta pj . Note that for dimensionless g the function F also is dimensionless and can
depend on the pj only through dimensionless ratios like pj / or pj /e
. Just as in the examples of
the 1PI propagators above, the loop-contributions will depend on g and pj /e
, but there are also the
counterterm contributions. If we write the renormalization condition for the coupling at scale 0
(more precisely for pj = pj (0 ) j 0 ) then the counterterms are fixed such that the dependence on
e will cancel out and

pj
F F g(0 ),
with F g(0 ), j = g(0 ) .
(5.62)
0
This then allows us to define (at least for close to 0 )

g() = F g(0 ), j
.
0
(5.63)
The -function is defined in terms of this running coupling g() as

d
d

(g(0 )) = g()
= F g(0 ), j
.

d
d
0 =0
=0
84
(5.64)
Note that is a function of g only with no explicit dependence. The definition (5.64) implies that
d
of g() keeping g(0 ) fixed. Now, for finite regularization,
the -function is obtained by taking d
i.e. 6= 0, one can express g(0 ) in terms of the bare coupling gB , and 0 . Thus one sees that one
can also write

(g()) = g() .
(5.65)
gB ,
5.4.2
Callan-Symanzik equations
It is now relatively easy to establish the Callan-Symanzik equations for proper vertices (n) or n-point
Green functions. They express that the choice of renormalization scale is arbitrary. Indeed, the bare
action, expressed in terms of bare fields and bare couplings (as well as bare masses if we considered
a massive theory), equals the renormalized action which is expressed in terms of renormalized fields
b(n) (cf. (1.37)),
and couplings (and masses). Thus when computing normalized Green functions G
i.e. excluding vacuum bubbles, of bare fields or of renormalized fields one uses the same functional
integral with the same action, the only difference being the explicit Z-factors. Any difference in the
measures due to the different normalizations of the fields cancels in the normalized Green-functions.
Thus (cf. (1.38))
#
" n
Yp
bl1 ...ln (p1 , . . . , pn ) ,
bl1 ...ln (p1 , . . . , pn ) =
G
Z lr G
(5.66)
(n)
B (n)
r=1
and thus for amputated Green functions

"
bl1 ...ln (p1 , . . . , pn ) =
G
B (n,amp)
n
Y
#
1/2
Z lr
bl1 ...ln (p1 , . . . , pn ) ,

G
(n,amp)
(5.67)
r=1
As for the fields, a subscript B indicates the bare quantity, while we have dropped the subscript
R for the renormalized ones. Often, the amputated Green functions coincide with the 1PI vertex
function and, obviously, the latter must satisfy the same relation as the former:
#
" n
Y
1/2
(n)
(n)
Z lr
l1 ...ln (p1 , . . . , pn ) .
(5.68)
B, l1 ...ln (p1 , . . . , pn ) =
r=1
One can also derive this relation more formally by working with the generating functionals: eiW [J]
R
R
R
R
D eiS[]+i J and eiWB [JB ] DB eiSB [B ]+i JB B . Observing that S[] = SB [B ] and B =
Z we find WB [JB ] = W [ ZJB ] up to an irrelevant additive constant. Upon Legendre transforming

W [J] and WB [JB ] one finds
(5.69)
[] = B [ Z] .
Expanding in powers of yields the relations (5.68).
To simplify the notation, we consider just one type of field and suppress all indices. However, we
will explicitly indicate the other quantities on which the vertex functions depend:
n/2

(n)
(n) pj , g(), .
(5.70)
B (pj , gB , ) = Z g(), ,
e
85
This is to be thought as the asymptotics for small where the renormalized (n) are -independent.
, holding gB (and )
On the left-hand side, nothing depends on . More precisely, if we take
fixed, we get zero. Thus

n

0=
log Z
(n) pj , g(), .
(5.71)

gB , 2
gB ,
One has

+ g()
(n) pj , g(),
g()
gB ,
g()

+ (g())
=
(n) pj , g(), ,
g()
g()

(n) pj , g(), =
gB ,
where we have used (5.65). Similarly, we have

log Z g(), , =
+ (g())
log Z g(), , .

gB ,
e
g(),
g() ,
(5.72)
(5.73)
A priori, can depend on the same arguments as Z, but looking at the above one-loop examples of
Z2 or Z3 , we see that the corresponding 2 and 3 are functions only of the renormalized coupling
g(). We will shortly see that, in general, can only be a function of g(), just as is the case for
(). Combining the last three equations, we get

(n)

n
0=
+ g()
g() pj , g(), .
g()
g() 2
(5.74)
This is the Callan-Symanzik equation for the n-point vertex function. It is now clear that cannot
depend on or
e since nothing else in this equation depends on these quantities. Since the explicit
-dependence of Z is only via the combination e it follows, since does not depend on
e it cannot
have any explicit -dependence either. Thus is a function of the dimensionless coupling g() only:

g() =
log Z .
gB ,
(5.75)
The generalization of (5.74) to the case of a vertex function involving several fields is obvious. Each
type of field comes with its own Zr factor and corresponding r -function, so that the Callan-Symanzik
equation becomes
"
#
n
(n)

1 X
+ g()
j g() l1 ,...ln pj , g(), .
(5.76)
0=

g()
g() 2 j=1
b(n) (pj , g(), ) with simply
Note that there is an analogous equation for the n-point Green function G
the sign in front of the j being + rather than .
86
As an example, let us work out the Callan-Symanzik equation for the electron-electron-photon
vertex function of QED (in the limit where one can neglect the electron mass, i.e. m
h e ), includi
e()2 e
ing the contributions to , 2 and 3 up to one loop. One has from (5.60), log Z3 = 62 C log e .

e()2

log Z3 = 62 while (g()) g() log Z3 gives a contribution e()4 comparable to
Then
g(),
,
a two-loop contribution. Hence, at leading order

e()2
4
+
O
e()
.
6 2
(5.77)

e()2
4
+
O
e()
.
2 =
8 2
(5.78)
3 =
Similarly, from (5.57)
Thus 22 + 3 =
that
5.4.3
5e()2
.
12 2
Finally, recall the one-loop -function of QED, eq. (5.17), (e()) =

e()3
5 e()2

+
p
,
e(),
.
0=

j
e() 12 2 e()
24 2
e()3
,
12 2
so
(5.79)
Solving the Callan-Symanzik equations
To solve (5.74) or (5.76) one first needs to find the running coupling constant as the solution g() of
the first order ordinary differential equation with some initial condition

d
g() = g() ,
d
g(0 ) = g0 .
(5.80)
R g() dg
We have already seen that the solution simply is given by g0 (g)
= log 0 . With this g() the
partial differential equations (5.74) or (5.76) are turned into ordinary ones:
#
"
n
(n)

1 X
d
j g() l1 ,...ln pj , g(), .
(5.81)
0=
d 2 j=1
This shows that
#
n Z

1 X d0
(n)
(n)
0
exp
g(
)
p
,
g(),
= l1 ,...ln (pj )
j
j
l
,...l
n
1
0
2 j=1 0
"
(n)
(5.82)
(n)
does not depend on . Obviously, l1 ,...ln (pj ) = l1 ,...ln (pj , g0 , 0 ). Hence, the Callan-Symanzik
equation (5.76) is solved by
(n)
l1 ,...ln
"
#
n Z

1 X d0
(n)
j g(0 ) l1 ,...ln (pj , g0 , 0 ) .
pj , g(), = exp
2 j=1 0 0
(5.83)
This is a powerful result. Suppose we have a good approximation for (n) for some g(0 ), like in
an asymptotically free theory for 0 . One can then use (5.83) to reliably compute (n) with
87
coupling g() at scale , as long as one can trust the approximations made when computing and
j .
One can make further statements if one uses dimensional analysis. Every vertex function or
Green function has some well-defined engineering dimension which is its mass dimension. We say
that some function (n) has engineering dimension n if (n) /n is a dimensionless quantity. In
4 dimensions, a scalar field has dimension 1 and a scalar n-point Green function G(n) (x1 , . . . xn )
has = n, its Fourier transform has dimension n 4n. The vertex function (n) is obtained by
multiplying with n inverse propagators which add a total of 2n to the dimension. Hence (n) has
b(n) (pj ) (4) (P pj ), as we did e.g. in QED for , we
scaling dimension n. If one writes (n) (pj ) =
b(n) is 4 n. More generally, let the engineering dimension of
see that the engineering dimension of
(n) be n . Then

p
j
, g(), 1 ,
(5.84)
(n) pj , g(), = n (n)
since the dimensionless quantity (n) /n can only depend on the dimensionless combinations
on g().
5.4.4
pj
and
Infrared fixed point and critical exponents / large momentum behavior in asymptotic free theories
Suppose one has an infrared fixed point at g = g . Recall that this occurs if (g ) = 0 and 0 (g ) =

R 0
j g(0 ) for
a > 0. Then as 0 one has g() g . Let j (g ) = j . The integral 0 d
0
0 is dominated by the small values of 0 and one can approximate j g(0 ) ' j . Hence

R d0
0
g(
)
' cj (0 ) + j log 0 . Combining (5.83) and (5.84) yields
j
0
0
#
"
n Z

1 X d0
(n)
n
0
(n) pj
g(
)
,
g(),
1
.
(5.85)
l1 ,...ln (pj , g0 , 0 ) = exp
j
2 j=1 0 0
We now let = 1 and pj = qj with fixed 1 and qj while letting 0. Then 0 and
g() g so that
"
#
n

1 X
1
(n)
(n) qj
n
cj (0 ) + j log
l1 ,...ln (qj , g0 , 0 ) ' exp
n
,
g
,
1
1
2 j=1
0
1
"
#

n

P
1 X
1
n 12 j j
(n) qj
n
=
exp
cj (0 ) + j log
,
g
,
1
. (5.86)
1
2 j=1
0
1
1
Thus as 0, the (n) scale as n 2 j j . The naive scaling exponent n has been corrected by
P
21 j j .
Recall that for Green functions one has to switch the sign in front of the j . For the two-point
function there is only a single involved. Finally recall that the momentum-space propagator is the
two-point functions without the (4) so that = 2 for scalars and = 1 for Dirac propagators.
Then e.g. for a scalar propagator

1
0
2+
c(0 )
2
0 qj
(q, g0 , 0 ) '
e
1
, g , 1
as 0 .
(5.87)
0
1
88
This infrared scaling behavior 2+ is as if the field had dimension 1 + /2. This is why /2 is
called the anomalous dimension of the field. When the critical point of the (four-dimensional) Ising
model is described by the massless 4 theory, is called a critical exponent.
The previous arguments are easily adapted to the large momentum behavior of vertex or correlation functions in asymptotic free theories. One now considers the large asymptotics. Then

(n) qj
g(1 ) 0 and it is enough to compute
, g(1 ), 1 to the lowest non-trivial order in pertur1
bation theory (i.e. tree-level if non-vanishing). Similarly j is to be replaced by the one-loop result
evaluated at the (small) coupling g(1 ).
5.5
5.5.1
Callan-Symanzik equations for a massive theory

Operator insertions and renormalization of local operators
It is often useful to consider composite operators like e.g. O(x) = 2 (x) in a scalar theory. (As
always, denotes the renormalized field.) One should think of this operator as being obtained from
(y)(x) in the limit where y x. One can then compute correlation functions (Green functions)

of O(x) with the elementary fields (xi ) by first computing hT (y)(x)(x1 ) . . . (xn ) ivac and then
letting y x. Of course, this limit is singular as one already sees for n = 0. Even in a free

R d4 p eip(yx)
R d4 p
1
theory one has limyx hT (y)(x) ivac = i limyx (2)
. This can
4 p2 +m2 i = i
(2)4 p2 +m2 i
be interpreted as a vertex with two lines attached that are joined by a propagator, i.e. a one-loop
diagram.
To get finite Green functions in the presence of such operators, one should work with the corresponding renormalized operators OR (x):
O(x) = ZO OR (x) = OR (x) + (ZO 1)OR (x) .
(5.88)
The second term is interpreted, as usual, as a counterterm. One can then show that with appropriately chosen ZO the Green functions involving OR are finite.
In the example of O = 2 we actually already know this counterterm. If we start with a massive
theory and consider all terms in the action that involve the mass terms as a small perturbation, we
see that inserting m2 2 + Zm2 2 (Z 1)m2 2 in any correlator must give a finite result. Thus
h
i
h
1
1h
m2 i1
m2 i
O2 ,R (x) = 1 2 Z m2 + (Z 1) O2 (x) = Z 1 2 O2 (x) Z2 =
1 2
.
m
m
Z
m
(5.89)
Of course, this is just such that
m2 O2 ,R (x) = m2B Z O2 (x) = m2B 2B (x) .
2
(5.90)

g
1
= 1 16
2

In the 4 -theory, eqs (3.109) and (3.110) yield, up to order g : Z2 = 1 + m

m2

2
+log 1
12 log m
. Note that this is in a massive theory. If we are in a massless theory, m2 = 0,
2
e2
as already observed above, and Z2 = 1 at this order.
As another example, consider O4 (x) = 4 (x). Again, we know from the renormalized perturbation theory of 4 -theory that insertions of gB 4B (x) in correlation functions lead to finite expressions.
89

Recall that gB = ZZg2 g and B = Z, hence gB 4B (x) = gZg 4 (x). Since g is finite, we see that the
operator which will yield finite results when inserted into correlators is
O4 ,R (x) = Zg 4 (x)
One sometimes defines m2B =
Zm 2
m
Z
Z4 = Zg1 .
h
so that Zm = Z 1
m2
m2
(5.91)
. From eq. (5.89) one sees that then
1
Zm
,
Z2 =
in complete analogy with (5.91)
Note that our definition (5.88) of ZO expresses a relation between the renormalized operator
OR (x) and the operator formed from the elementary renormalized fields. One could further express
the latter in terms of the bare fields, e.g. n (x) = Z n/2 nB (x), and define the bare composite
operators in terms of the bare fields only, e.g. OB (x) = nB (x). Thus
Yp
Yp
Zl O(x) = ZO
Zl OR (x) ZeO OR (x) .
(5.92)
OB (x) =
l
5.5.2
Callan-Symanzik equations in the presence of operator insertions
It is now straightforward to derive Callan-Symanzik equations for Green functions or vertex functions
that involve insertions of local operators as just discussed. Again one writes that the bare functions
with the bare operators OB inserted into the correlation functions of the bare fields B cannot
depend on the renormalization group scale . This is then translated into a differential equation for
the correlation functions of renormalized fields with renormalized operators inserted.
To simplify the notation, we will only consider one type of field, denoted , and one type of
operator, denoted O. We let
h
i
(n,l)
GR (x1 , . . . , xn ; y1 , . . . yl ) = hT (x1 ) . . . (xn )OR (y1 ) . . . OR (yl ) ivac ,
h
i
(n,l)
(5.93)
GB (x1 , . . . , xn ; y1 , . . . yl ) = hT B (x1 ) . . . B (xn )OB (y1 ) . . . OB (yl ) ivac .
They are related by
(n,l)
(n,l)
= Z n/2 ZeOl GR .
GB
(5.94)
If several fields and/or operators are present, one has the appropriate products of the Z factors. If
one amputates these Green functions by multiplying with n inverse propagators, one gets a similar
relation between the bare and renormalized quantities, but with Z n/2 replaced by Z n/2 . The same
relation holds for the 1PI vertex functions with operator insertions:
(n,l)
(n,l)
= Z n/2 ZeOl R .
(5.95)
As before (cf. eq. (5.69)), this relation can also be proven more properly by working through the
relations between the generating functional, now with an extra source/current for the operator OB =
ZeO OR . Doing the Legendre transform but only with respect to the elementary fields and not with
respect to this extra current, one gets a relation between the generating functional of 1PI diagrams
with additional insertions of O. Expanding this relation yields (5.95).
90
In addition to the definition (5.75) of , we also define

log ZeO .
eO g() =
gB ,
(5.96)
It is then straightforward to generalize (5.74) to

(n,l)

n
0=
+ g()
pj , g(), .
g() + l eO g()
g()
g() 2
(5.97)
The solution of this equation and study of the asymptotic behaviors of the (n,l) proceeds just as for
l = 0 with the obvious replacement n2 n2 l eO .
5.5.3
Massive Callan-Symanzik equations

R
Let us now consider a massive scalar theory and consider the mass term d4 z 21 m2B 2B (z) =
R
12 d4 z m2 O2 ,R (z) (cf. (5.90)) as a perturbation. Then any Green function of the massive theory
with l insertions of O2 can be written as
(n,l)
Gmassive (x1 , . . . xn ; y1 , . . . yl )
X
r=0
(m2 )r
(i)r r
2 r!
(n,l+r)
d4 z1 . . . d4 zr Gmassless (x1 , . . . xn ; y1 , . . . yl , z1 , . . . zr ) .
(5.98)
Each term in the sum on the right-hand side satisfies a Callan-Symanzik equation (5.97) with the
last term in the bracket being (l + r)e
2 .
But we can generate exactly this expression for each term
in the sum by acting on the sum with e2 l + m2 m

2 . We find that the left-hand side satisfies the
massive Callan-Symanzik equation:
"

n

+ g()
+ g()
0 =

g(),m
g() ,m 2
#

(n,l)
+ e2 g() l + m2
G
p
,
g(),
.
(5.99)

j
massive
m2 ,g()
The relation between the bare and renormalized 1PI functions with l insertions of an operator O was
obtained before. In particular, this also applies for the insertions of the mass operator and we get,
in analogy with the preceding Callan-Symanzik equation for the Green functions,
"

n
+ g()
g()
0 =

g(),m
g() ,m 2
#

(n,l)
+ e2 g() l + m2
(5.100)
p
,
g(),
.

j
massive
m2 ,g()
91
PART III :
NON-ABELIAN GAUGE THEORIES
6
Non-abelian gauge theories: formulation and quantization
Gauge invariance, and in particular non-abelian gauge invariance, plays a most important role in the
formulation of the quantum field theories that seem to describe (almost?) all of the particle physics as
it is known today: quantum electrodynamics to begin with and its embedding into the electro-weak
theory based on the gauge group SU (2) U (1), as well as the theory of strong interactions based
on the gauge group SU (3). Here, we will briefly show how to construct (classical) actions that are
invariant under non-abelian gauge symmetries. Then we will discuss how to quantize these theories
using the functional integral approach. This will involve the issues of gauge-fixing, Faddeev-Popov
procedure and the appearance of so-called ghost fields. The gauge-fixed action no longer is gauge
invariant but instead has a new symmetry, the BRST-symmetry which we will identify. This BRST
symmetry will play a crucial role when showing that these gauge theories are renormalizable.
6.1
Non-abelian gauge transformations and gauge invariant actions
Recall from quantum electrodynamics that the classical action is invariant under the following gauge
transformation:
i i(x)ql
e
ei(x)ql .
ql
(6.1)
i(x)ql
This is often called a U (1) gauge invariance because obviously g(x) = e
U (1). There seems
to be a different g(x) for every different charge ql , but note that ql is the eigenvalue of the charge
operator Q when acting on l . Hence we can write g(x) = ei(x)Q . Using this g(x) we can rewrite
the transformations (6.1) as
l (x) l0 (x) = ei(x)ql l (x) ,
A (x) A0 (x) = A (x) + (x) = A (x) +
(x) l0 (x) = g(x)(x) ,
QA (x) QA0 (x) = Q A (x) + ig(x) g(x)1 .
(6.2)
Of course, gauge invariant terms then are built from covariant derivatives, i.e.
D = ( iQA ) ,
(6.3)
since this transforms as

D D0 0 = iQA + (g g 1 ) g = g l + (g 1 g) iQA + (g g 1 )

= g iQA = gD .
(6.4)
Indeed, since g is just a phase, |D |2 obviously is invariant.
92
We have rewritten the simple U (1) gauge transformation of the electromagnetic (gauge) field
A and of the matter fields , as well as the definition of the covariant derivative, in a way that
makes their generalization to other gauge groups almost obvious. Let us now consider the case
where the abelian U (1) group is replaced by some, generally non-abelian Lie group G, called the
gauge group. In general this can be a product of so-called simple groups and U (1) factors, like e.g.
SU (3) SU (2) U (1) for the standard model of electro-weak and strong interactions, or even some
more exotic groups like e.g. E8 . For simplicity, one might simply think of SU (N ), the group of
unitary N N matrices with unit determinant. The structure of a Lie group is almost entirely
captured by the commutators of its generators: the group elements close to the identity are always
a
of the form ei ta with dim G small parameters a and dim G generators ta . The latter must satisfy
closed (Lie algebra) commutation relations
[t , t ] = iC t ,
(6.5)
with real structure constants C which satisfy the Jacobi identity C [ C ] = 0. It may happen
that several groups like e.g. SU (2) and SO(3) have the same Lie algebra. In this case the groups
are locally identical but not globally. We will be not so much concerned with the group elements
themselves but rather with their representations. If we use a specific representation R we write tR
R l
or (t )k for the dim R dim R matrices of this representation. For compact Lie algebras (i.e. if
tr t t is positive-definite), all finite dimensional representations are hermitian. This is the case of
R
most interest in gauge theories and, hence, (tR
) = t . Of course, the corresponding representations
of the group then are unitary. In this case one can also find a basis for the generators for which the
C are antisymmetric in all 3 indices and one may then drop the distinction between upper and
lower indices. Note that the Jacobi identity implies that the
(tadj
) = iC
(6.6)
satisfy the algebra (6.5). This special representation is called the adjoint representation.
Consider a collection of matter fields l , l = 1, . . . r transforming in some r-dimensional representation R of G:
l (x) l0 (x) = U R (g(x))l k k (x) ,
(6.7)
where U R (g(x)) is the r r matrix associated to the group element g(x) G in the representation
R. For G = SU (N ), the simplest example is the so-called fundamental (or vector)) representation
for which U (g(x))i j = g(x)i j with i and j running simply from 1 to N . In the following, we will not
write the indices explicitly so that we simply write
(x) 0 (x) = U R (g(x)) (x) ,
(6.8)
with the understanding that is a dim R-dimensional column vector of matter fields. We want to
construct a covariant derivative D involving and some gauge field A . In the abelian case the
relevant combination was A Q with Q the generator of the U (1) group. Let us try
R
A = A tR
A ,
93
(6.9)
since this can act on the dim R-dimensional column vector of matter fields , and a definition of the
covariant derivative as

DR = iAR
(6.10)
.
One often drops the superscript R on D or A , but one should remember that these quantities
take values in the representation R of the gauge group determined by the matter fields. We want
to determine the gauge transformation properties of A in such a way that the covariant derivative
simply transforms as
D D0 0 = gD .
(6.11)
First, we look at the transformation of . To further simplify the notation we will simply assume
that the representation is the fundamental representation where D(g(x)) = g(x), but one could
replace everywhere g by DR (g)). Also we will drop the arguments x, although g, and A all
depend on x. One has

0
1
= (g) = g + ( g) = g + g g
(6.12)
On the other hand,

iA iA0 0 = iA0 g = ig g 1 A0 g .
(6.13)
It is then obvious that the covariant derivative will transform covariantly, i.e. as in (6.12) provided A
transforms as ig 1 A0 g + g 1 g = iA , i.e. A0 = gA g 1 i gg 1 . Since g 1 = g 1 gg 1
one has gg 1 = g g 1 so that one gets

A0 = gA g 1 + ig g 1 = g A + i g 1 .
(6.14)
Note that for G = U (1) and g = eiQ and the replacement A QA one gets back (6.1). For a
general matter representation one simply has
0 = U R (g)

0
R 1
AR = U R (g) AR
+
i
U (g )
(6.15)
DR
iAR
0
DR
=U
(g)DR
For a unitary representation, (D ) D then is obviously invariant This will be the ingredient to
write gauge invariant matter kinetic terms.
It is often enough and simpler to consider only infinitesimal gauge transformations with g(x) =
R
i t
or U R (g) = ei (x)t . It is convenient to define R = tR
e
. Then the previous equations yield
= 0
= iR

0R
R
R
R
A = A A = R i AR
, ] .
(6.16)
The latter equation reads in components

A = + C A .
94
(6.17)
Using the generators of the adjoint representation (6.6) one may also write this as
A = (Dadj ) .
(6.18)
We still need to define the field strength F . Since D transforms exactly as , taking a further
covariant derivative D is defined in exactly the same way:
R
R
R
R
R R
DR DR = ( iAR
)( iA ) = i A iA iA A A .
(6.19)
Antisymmetrizing in and eliminates all terms with a derivative of . Hence

R
[DR , DR ] = iF
,
(6.20)
R
R
R
R
F
= AR
A i[A , A ] ,
(6.21)
F
= A A + C A , A .
(6.22)
with
R
R
or with F,
= F
t :
It follows from (6.20) that F transforms covariantly:

0
R
R
F
FR = U R (g) F
U R (g 1 )
or
R
R
F
= i[R , F
]
(6.23)
R R
Obviously then, a gauge invariant scalar density is tr F
F
. For compact groups we can choose
a basis of generators such that
R
tr tR
t = CR
R R
tr F
F
= CR F
F .
(6.24)
With the appropriate normalization, this will be the gauge kinetic term, generalizing 14 F F .
Note that, contrary to the abelian gauge theory of QED, in the non-abelian case, F
F contains
cubic and quartic terms in the gauge fields. This will lead to cubic and quartic vertices involving
only gauge boson lines.
In QED, the coupling constant is given by the electric charge, i.e. the eigenvalue of the charge
operator Q. In analogy we may include the Yang-Mills coupling constant gYM in the generators t
2
and their representations tR

. In this case CR gYM and C gYM . This is what Weinberg does,
and we will follow this convention. Alternatively one can use conventionally normalized Lie algebra
generators (where the CR do not include any factors of gYM ). Then a factor of gYM does appear
explicitly in front of the A in the covariant derivative, and similarly in front of the quadratic term
1
R
in F . This explicit factors of gYM can then be removed by rescaling AR
gYM A . The only
1
appearance of the coupling then is a factor gYM
in front of every F . If the gauge group is simple,
there is only a single coupling constant gYM . However, in the general case where the gauge group is a
product of simple and U (1) factors, every simple or U (1) factor can have its own coupling constant.
The unique Lorentz and gauge invariant Lagrangian quadratic in the field strength then is
Lgauge [F ] = 41 F
F . Of course, there is one more possibility, F
F
but this is a
total derivative. Hence the Lagrangian for matter and gauge fields is
1
L[A , ] = F
F
+ Lmatter [, D ] ,
4
95
(6.25)
where we collectively denoted all matter fields by . If temporarily denotes spin

scalars, an example of Lmatter would be
D + m) (D ) D m
Lmatter [, D , , D ] = (
2 V ( ) .
1
2
fields and
(6.26)
We only need the fact that the matter Lagrangian is gauge invariant. This will be the case if
Lmatter [, , , ] is invariant under global (rigid) transformation by elements of G.
6.2
Quantization
Just as with QED, direct canonical quantization does not work due to the presence of constraints, in
particular one again has = F0 and hence 0 = 0 which is a primary constraint. Together with a
secondary constraint which does not involve A0 these are first class constraints. As usual, they have
to be eliminated by a gauge choice. In the non-abelian case, a convenient choice is the axial gauge
A3 = 0 .
(6.27)
Then A1 and A2 become canonical variables, while A0 is given by the solution of the secondary constraint. One can then go through the canonical formulation of the functional integral in Hamiltonian
form, as we did for QED. To begin with one only integrates over the canonical fields Ai i = 1, 2 and
their conjugate fields i . Proceeding through similar steps as we did for QED27 , one ends up with
a Lagrangian version of the functional integral as (C is a normalization constant)
Z

Y
R 4
hT Oa . . . ON ivac = C DA Dl
(A3 ) Oa . . . ON ei d xL[A ,]
(6.28)
x,
R
where one intragrates over all 4 components with the manifestly gauge invariant action d4 xL[A , ].
Of course, the integration over A3 is effectively suppressed by the functional (A3 ) which imposes
our gauge choice. Note that the derivation ensures that this gives a manifestly unitary theory. On
the other hand, Lorentz invariance is not manifest.
One is usually interested in computing the vacuum expectation valued of (time-oredered) gauge
invariant operators O1 , . . . ON , although, at intermediate stagesone may also
compute and use
non-gauge invariant objects as e.g. the gauge field propagator hT A (x)A (y) ivac . Note that the
measure DA is gauge invariant: It is easy to show that the Jacobian for the gauge transformation
(6.17) equals one.28 In the absence of (massless) chiral fermions one can also show that the matter
measure is gauge invariant. However, if chiral fermions are present this is no longer guaranteed. If
the matter measure is not invariant, one has a so-called anomaly, and the entire following discussion
breaks down. It is thus important that, if present, gauge anomalies cancel between the contributions
of the different chiral matter fields.
27
28
Just as when discussing QED, in this section, we will simply write h(. . .)ivac instead of hvac, out| (. . .) |vac, ini.
0
h
i
A (x)
Jx,y = A (y) = (4) (x y) + C (x) , and due to the antisymmetry of the structure constants
C = 0 and DetJ = exp Tr log J = exp(O(2 )).
96
6.2.1
Faddeev-Popov method
We want to show that one can rewrite the expectation values (6.28) in different equivalent ways,
corresponding to different gauge choices. First it is useful to rewrite (6.28) as
Z
Z

Y
R 4
hT Oa . . . ON ivac = C DA
(A3 ) G[A ] with G[A ] = Dl O1 . . . ON ei d xL[A ,] ,
x,
(6.29)
the important point being that
is gauge invariant. Indeed, with a gauge invariant matter
R 4
R
0
0
measure and gauge invariant Oi we have G[A ] = Dl Oa . . . ON ei d xL[A ,] . Upon changing
the integration variable from l to the gauge transformed l0 and using the gauge invariance of the
R 4
R
0
0
0
measure this is G[A ] = Dl Oa . . . ON ei d xL[A , ] = G[A ], where we used the gauge invariance
of the action in the last step.
The expression of eq. (6.29) is of the general form
Z

f [A0 (x)]
IG = C DA B f [A ] DetF[A ] G[A ] ,
F (x, y) =
,
(6.30)

(y) =0
G[A ]
where A0 denotes the gauge transformed A with parameter . Indeed, if we let f [A] = A3 and
Q
0
B[f ] = x, (f (x)), we have A3 = 3 (since A3 = 0) so that F (x, y) = x 3 (4) (x y) and
DetF is just an irrelevant constant.
We want to show that IG does not depend on the choice of the gauge-fixing function f or on
B. Intuitively, what happens is the following: due to the gauge invariance of G[A ] there are many
gauge-equivalent configurations and integrating over all A would result in an infinite factor equal
to the volume of a gauge slice. The gauge-fixing condition restricts the functional integration to
exactly one configuration among the gauge equivalent ones. There are many different ways to do this
and the factor DetF ensures the independence of the specific choice of f or B. This is the functional
R
analogue of the well known fact that dx (f (x))f 0 (x)g(x) = g(x0 ) does not depend on f , as long
as f has a single root x0 .
Faddeev-Popov theorem : For gauge invariant G[A ], the functional integral IG is independent of the
gauge-fixing function f and depends on B only through an irrelevant overall factor.
To prove this theorem we first change variables from A to some A0 Ag which we identify with
A gauge transformed by some g(x) G. Using the invariance of the measure and of G[A ] we get
Z

IG = C DA B f [Ag ] DetF[Ag ] G[A ] .
(6.31)
Since we only introduced g by a change of variables, this cannot depend on g. We multiply both
R
sides by some weight function [g] such that Dg [g] = C0 is finite:
Z
Z

C0 IG = DA G[A ] H[A ] ,
H[A ] = Dg [g] B f [Ag ] DetF[Ag ] ,
(6.32)
where now (suppressing the , , . . . labels but showing the x, y, . . ., and using the fact that the gauge
transformations form a group)
Z

g
f [(Ag )0 (x)]
f [Ag(g,) (x)]
g (g, ; z)
4 f [A (x)]
g
F[A ](x, y) =
=
=
d
z
. (6.33)

(y)
(y)
g (z) g=g
(y) =0
=0
=0
97

(g,;z)
We let G[g](z, x) = g(y)
. Obviously, this only depends on g(x) and one can show that [g] =
=0
1
DetG
yields a reasonable weight function. With this choice, (6.32) yields
Z
H[A ] =

f [Ag ]
Dg [g] B f [Ag ] Det
DetG =
g

f [Ag ]
Dg B f [Ag ] Det
=
g
Z
Df B[f ] CB .
(6.34)
Obviously, this is independent of the choice of f and depends on B only through the constant CB .
Substituting this result into the first equation (6.32), the same is also true for IG , i.e IG /CB is
independent of f and B, which was to be proven.
Q
Since for the special choice f [A ] = A3 and B[f ] = x, (f ) (in this case CB = 1) the integral
IG equals (6.29), the theorem tells us that the vacuum expectation value of a time-ordered product
of gauge invariant operators can be evaluated with any choice of gauge-fixing function and function
R
B, provided Df B[f ] converges and f indeed fixes the gauge:
Z
Z

R 4

C
(6.35)
DA B f [A ] DetF[A ] Dl O1 . . . ON ei d xL[A ,] ,
hT Oa . . . ON ivac =
CB
6.2.2
Gauge-fixed action, ghosts and Feynman rules
Let us now make a convenient choice of gauge-fixing function f and function B which, in particular,
will be manifestly Lorentz invariant:

Z
i
4
d x f (x)f (x) .
(6.36)
f = A ,
B[f ] = exp
2
As one can see from (6.35), the factor B[f ] just contributes an extra term
Lg.f. =
1
f f
2
(6.37)
to the classical Lagrangian. One can then read off the gauge boson propagator i,,n (p) from
the quadratic part of L + Lg.f. and finds

p p
1
.
(6.38)
,, (p) = + ( 1) 2
p
p2 i
Except for the extra this is just like the photon propagator, which should not be surprising, given
our choice of f .
Next, we must evaluate F and compute its determinant. Since a fermionic gaussian integral
equals the determinant of the quadratic form, it is a convenient trick to rewrite DetF as

Z
Z
4
4
DetF = DD exp i d x d y (x) F, (x, y) (y) ,

(6.39)
where and are anticommuting, i.e. fermionic fields. On the other hand, given the nature of
F we will soon see that they cannot be spin 12 fields but must be scalars. Moreover, and are
independent real scalars. They do not obey the usual spin statistics relation and hence are named
98
ghosts. However, this is not a problem since we do not want to obtain any Lorentz invariant S-matrix
for scattering of these ghost fields. More precisely, is called a ghost and an anti-ghost. Note
that the ghost and anti-ghost carry an index just like the gauge field A . One sometimes says that

they are in the adjoint representation of G, just as A . One has f [A0 ] = + C A and
hence

f [A0 ](x)

=
F, (x, y) =
+ C A (x) (4) (x y) .
(6.40)
(y)
x
x
Using (6.39) we get
Z

Z
4
DetF = DD exp i d x Lghost (x)
Lghost = + C A . (6.41)
Thus the effect of the so-called Faddeev-Popov determinant DetF is to add the ghost Lagrangian
Lghost to the classical Lagrangian L and gauge-fixing Lagrangian Lg.f. :
Lmod = L + Lg.f. + Lghost .
(6.42)
Note that for an abelian theory like QED, the structure constants C vanish and the ghosts do
not couple to any of the other fields. Equivalently then, their functional integral only leads to the
determinant of which is an irrelevant constant. This is why we did not have to bother about
the ghosts in QED. In the non-abelian gauge theory, with our choice of gauge-fixing function, the
ghosts do couple to the gauge fields and do make important contributions to loop diagrams.29 Their
i
propagator (2)
4 (p) can be read off from (6.42) to be the same as for a massless spin-0 fermion:
=
,
i
(6.43)
p2
while the ghost - anti-ghost - gauge field vertex involves the structure constant C as well as a
factor of momentum. More precisely, taking all momenta as incoming (q for the ghost with label ,
p for the anti-ghost with label and k for the gauge boson with labels and ), one reads from
(6.41) that this vertex contributes
i(2)4 (4) (p + q + k) ip C .
(6.44)
We already gave the gauge boson propagator. The cubic and quartic gauge boson couplings simply
follow from the cubic and quartic terms in 41 F F . Again, with all momenta taken as incoming,
the vertex for the coupling of three gauge bosons with (p, , ), (q, , ) and (k, , ) is
h
i
i(2)4 (4) (p + q + k) (iC ) p p + q q + k k ,
(6.45)
while the vertex for the coupling of four gauge bosons (the fourth one having (l, , )) is
h
i(2)4 (4) (p + q + k + l) C C ( )
C C ( )
C C ( )
29
Of course, in axial gauge, A3 = 0, one finds F (x, y) =

3 so that the ghosts again decouple.
(4)
(x
x3
99
(6.46)
y) and the ghost Lagrangian simply is
Note that the Lagrangian preserves the ghost-number which translates into the fact that every vertex
with one incoming ghost line also has exactly one outgoing ghost line.
6.2.3
BRST symmetry
One can now start computing Feynman diagrams and generating functionals using the Lagrangian
Lmod = L + Lg.f. + Lghost of (6.42) and integrating over the matter fields, gauge and ghost fields. Note
that Lmod no longer is gauge invariant: this was the whole point about gauge-fixing. All terms in Lmod
have = 0 and hence this Lagrangian is renormalizable by power-counting as discussed in section
4: there are only finitely many divergent Greens functions and they are made finite by he addition
of finitely many counterterms with their coefficients fixed at any given order in perturbation theory.
Moreover, these counterterms themselves all have ct 0 and do not upset the renormalizability.
However, we want more than this: the counterterms should be of the form of the initial terms in
the Lagrangian. In particular, one should still be able to interpret Lmod + Lct as arising from the
gauge-fixing of some gauge invariant Lagrangian, but now with renormalized parameters. We must
find out what is the remnant of the gauge invariance of the original classical Lagrangian. This turns
out to be the BRST-symmetry.
Let us first do one more rewriting of our Lagrangian. Introducing an auxiliary field h we can
rewrite the gauge-fixing term as 2 h h +h f since doing the gaussian integration over h reproduces
21 f f . Thus, our starting point is
Lnew = L + h h + h f + ,
2
where
(x) = (x) +
C (A (x) (x))
Z
=
d4 y F (x, y) (y) .
(6.47)
(6.48)
The BRST symmetry is defined to act on the ordinary fields, i.e. the matter and gauge
fields, just like an ordinary infinitesimal gauge transformation but with the (real) parameter
replaced by the (real) ghost field . It is thus a fermionic symmetry that increases the ghost-number
by one unit. To begin with, we simply set (x) = (x) with an anticommuting parameter
( = , A = A , etc, and = 0). Then
A = + C A
= i tR
R
= i tR
= i t .
(6.49)
This is conveniently rewritten by defining a fermionic operator s such that for any functional F we
simply let
F = sF .
(6.50)
Consistency with the fermionic character requires
s(F G) = (sF )G F sG ,
100
(6.51)
with a minus sign if F is fermionic (contains an odd number of anticommuting fields). We write
R
R
furthermore R = tR
and s = i R or,
, so that we can reformulate (6.49) as s A = D
dropping the superscript R, simply
s A = i[A , ]
(6.52)
s = i
s = i .
Let us complete this by the rules how s acts on the ghosts, chosen in such a way that s is a nilpotent
operation, i.e. s2 = 0 on all fields. Choosing
s = i
s = h ,
sh = 0 ,
(6.53)
where h is the auxiliary field introduced above, we have

s2 A = s i[A , ] = (s) i(sA ) iA s + i(s)A isA
= i () i( iA + iA ) iA i A i( iA + iA ) = 0 ,
s2 = s(i) = i(s) i s = + = 0 ,
s2 = s(i) = i(s) i(s) = + = 0 ,
s2 = s(h) = 0 ,
(6.54)
which shows that the BRST operator s is nilpotent:

s2 = 0 .
(6.55)
It is now easy to show that Lnew is BRST invariant. First note that on functionals of A and
the matter fields only the BRST transformation is simply a gauge transformation with replaced
by . Thus any gauge invariant functional of the gauge and matter fields only automatically also is
BRST invariant. Hence
sL = 0 .
(6.56)
Next note that F was defined as the gauge variation of f (which only depends on A ) and hence
= sf
s = 0 .
(6.57)
It follows that
sLnew = s
h h + h f + = h sf + (s ) = h h = 0 .
2
(6.58)
Actually one has

Lnew = L + s ,
= f h ,
2
101
(6.59)
with L being the gauge invariant Lagrangian of the gauge and matter fields only. In this form, BRST
invariance is obvious. This also suggests how to obtain more general gauge-fixings: Any Lnew of this
form with an arbitrary functional of ghost-number 1 will provide a BRST invariant starting point
for quantizing the gauge theory (provided the quadratic term in the gauge fields is non-degenerate
so that one can define a propagator).
Below, we will use this BRST invariance to show that all conterterms also must respect this BRST
invariance (but with a renormalized coupling constant which we had hidden in the normalization of
the Lie algebra generators). Hence
e ,
Lnew + Lct = Le + s
(6.60)
respects the same symmetries as the original Lnew with a renormalized Yang-Mills coupling constant
and yields finite Greens functions.
6.3
BRST cohomology
We already noted that the BRST symmetry is nilpotent, i.e. s2 = 0. Obviously also, it increases the
ghost-number by one unit.
At an algebraic level, this is quite similar to the behavior of the exterior derivative d = dx acting on
differential forms30 of degree p and yielding a differential form of degree p + 1. As is well known, one has
d2 dd = 0 (which just states that = ). A p-form (p) is called closed if d (p) = 0. If there exists
a (globally well-defined) (p 1)-form (p1) such that (p) = (p1) then (p) is called exact. Since d2 = 0,
obviously every exact form is also closed. It is then interesting to find out which p-forms are closed without
being exact. As an example consider a space-time with space being just the two-dimensional sphere with
coordinates and defined in the usual way. Then d is well-defined everywhere except at = 0 or =
(north and south pole), and the 2-form = sin d d is well-defined everywhere. One has d = 0, so
it is closed. Although one has = d( cos d) or = d( sin d ), neither cos d nor sin d are
well-defined everywhere on the sphere and one finds that 6= d (1) . Hence is closed but not exact. Up
to multiplication by a constant31 this is the only closed and non-exact 2-form on the sphere. Actually,
is the volume-form on the sphere, and the volume form on any compact manifold is always closed but not
exact. For a given (compact) manifold M, the vector space of closed p-forms that are not exact is called
the pth de Rham cohomology and is denoted H (p) (M).
Similarly, one defines the classes of BRST-closed functionals and BRST-exact functionals as
follows:
The functional F [A , , , , h] is BRST closed if s F = 0,
(6.61)
it is BRST closed if there exists a functional G such that F = s G.
30
In exterior calculus, one defines an antisymmetric wedge product so that dx dx = dx dx . For every
1
rank-p antisymmetric tensor field 1 2 ...p (x) one can then define a p-form as (p) = p!
1 2 ...p dx1 dx2 . . . dxp .
In d dimensions, the maximal degree of a form is p = d.
31
For example, cos = cos sin d d = d with a = 21 sin2 d a well-defined one-form. Hence cos is
exact.
102
Of course, a BRST-closed functional is a BRST invariant functional. Moreover, the ghost-number of

a given monomial of the fields is defined as the number of ghost-fields minus the number of anti-ghost
fields. The Lagrangian Lnew e.g. has ghost-number zero. The BRST operator s always increases
the ghost-number by one unit. Since one cannot have cancellations between terms of different ghostnumbers it follows that one can define the space of BRST-closed functionals of a fixed ghost-number
and, similarly the space of BRST-exact functionals of a fixed ghost-number. Hence, the ghostnumber plays a role analogous to the degree p of the differential forms. In particular, one defines the
BRST-cohomology classes at ghost-number n :
The BRST-cohomology at ghost-number n is given by
(6.62)
the BRST-closed functionals of ghost-number n, modulo the BRST-exact functionals.
We have seen above that the gauge invariant Lagrangian L[A, ] is BRST invariant, i.e. sL[A, ] = 0,
so it is BRST closed. Moreover, it is of ghost-number zero. The gauge-fixing and ghost terms are
of the form s with of ghost-number 1, i.e. are BRST-exact terms. Since correlation functions
of gauge invariant operators do not depend on (as long as it provides some gauge-fixing), the
physics determined by Lnew = L[A, ] + s depends only on the BRST-closed L and is independent
of the BRST-exact terms s: it only depends on the BRST cohomology class of Lnew . Later-on we
will show that the BRST-cohomology at ghost-number zero is precisely given by the gauge invariant
functionals of A and (and ) only, i.e. independent of the ghost, anti-ghost and h-fields.
BRST-charge :
One can introduce a fermionic BRST charge operator (acting on a Hilbert space) by demanding
that for any (Heisenberg picture) field operator one has
(x) = i[Q, (x)] .
(6.63)
The right-hand-side equals i[Q, (x)] , i.e. a commutatot if is bosonic and an anti-commutator if
is fermionic, while the left-hand-side equals s(x). Hence the BRST-charge operator must have
the following (anti)commutation relations with the field operators:
s(x) = i[Q, (x)] .
(6.64)
One then has (using the Jacobi identity)

1
0 = s s = is[Q, ] = [Q, [Q, ] ] = [[Q, Q]+ , ] = [Q2 , ]
2
(6.65)
An operator that commutes with all fields is either the identity or has to vanish. Since Q increases
the ghost-number by one, the first option is excluded and one concludes
Q2 = 0 .
103
(6.66)
The Hilbert space on which Q acts is some big space of states which must include the
Hilbert space of physical states, but also states including an arbitrary number of ghost and antighost excitations, as well as non-physical polarization states of the gauge bosons. This is the space
which naturally arises upon the Faddeev-Popov quantization of the gauge theory. One must then
characterize the the Hilbert space of physical states. This can be conveniently done using this BRST
charge as follows. Matrix elements between physical states must be gauge invariant, i.e. independent
of the gauge-fixing functional . Hence, under an infinitesimal variation (x) (x) + (x),
the gauge-fixed Lagrangian changes by Lnew = s(x) and any matrix element changes by
h |i = h| s |i = h| [Q, ]+ |i = h| Q |i h| Q |i .
(6.67)
If h| and |i are physical states this should vanish. Since (x) is arbitrary one concludes Q |i = 0
and h| Q = 0. Hence, physical states |physi must obey
Q |physi = 0 .
(6.68)
Also, changing |i |i + Q |i does not change h |i if h| is physical since h| Q |i = 0. Moreover, a physical state should have ghost-number zero. In conclusion, physical states are determined
by Q |physi = 0 subject to the equivalence relation |physi ' |physi + Q |i and the condition of
having ghost-number zero:
Physical states are given by the ghost-number zero cohomology class of the BRST-operator Q.
(6.69)
Of course, we have only shown that this cohomology class contains the physical Hilbert space, i.e.
that the above conditions are necessary. However, one can also show that they are sufficient and thus
(6.69) exactly defines the physical Hilbert space. Note that in any covariant quantization, the modes
of the fields Aa0 , when acting on the vacuum, generate states of negative norm (since 00 = 1). One
can then also show that the space of physical state as defined by (6.69) has a positive definite norm.
Remarks :
The Faddeev-Popov procedure always leads to a Lagrangian Lnew that is bilinear in the ghost
and anti-ghost fields. (Recall that the ghosts appeared from expressing the Faddeev-Popov
R
determinant Det F as an integral over and of ei F .) Then, calling these fields the bare
fields and rewriting Lnew in terms of renormalized fields will generate various counterterms
of the same form as the original terms contained in Lnew . In particular, one will only get
counterterms that are at most bilinear in the ghost and anti-ghost fields. This turns out to be
sufficient with the gauge choice f = A , but for more general choices of f one might need
counterterms that involve two ghost and two anti-ghost fields. Although one does not have
to worry about diagrams with 4 external (anti)ghost lines, such diagrams can well appear as
divergent subgraphs, requiring a corresponding counterterm.
104
One can consider more general gauge-fixing functionals that do not rely on the Faddeev-Popov
procedure. As we have seen, all that is needed for BRST invariance is to define some BRSToperation s such that s acts on A and as a gauge transformation with the parameter replaced
by a ghost field, and with the action on the ghost and other fields (anti-ghost, h-field, and pose of ghost-number 1 and gauge invariant
sibly others) defined such that s2 = 0. Then for any
e Then just as before, there is a corresponding
L, a BRST-ivariant Lagrangian is Lenew = L + s.
BRST charge Q that defines the physical states as the zero ghost-number cohomology class
and matrix elements between physical states (in particular also vacuum expectation values of
e In
time-ordered products of gauge invariant operators) do not depend on the choice of .
particular, they are the same as with the from the Faddeed-Popov procedure and thus the
same as in axial gauge.
One can prove independently of comparison with any particular gauge, that the space of physical states, defined as the zero ghost-number cohomology of Q, has a positive definite norm,
contains no ghosts or anti-ghosts and has a unitary S-matrix.
Finally let us just mention that this BRST formalism can be rather straightforwardly extended
to other local symmetries as appear e.g. in general relativity or in string theory. If the natural
formulation of these symmetries is too large in the sense that one has introduced too many
gauge parameters and actually not all symmetries are independent, one has to introduce
ghosts of ghosts. In all these setting, the BRST operator always increases the ghost-number
by one unit.
We have already mentioned that the BRST invariant functionals of ghost-number zero are just the
gauge invariant functionals of A and the matter fields, up to adding BRST-exact terms. Let us now
prove this result.
Cohomology theorem :
The zero ghost-number cohomology consists of gauge invariant functionals of A and (and )
only, i.e. the most general BRST invariant functional of ghost-number zero is of the form I =
I0 [A , ] + s[A , , , , h].
P
The proof is relatively simple: suppose sI = 0. Write I =
N =0 where IN contains all terms that
have a total number of fields and h equal to N . (Of course, we do not allow negative powers of
the fields.) Since s = h and s h = 0, s does not change this total number N and one cannot
have any cancellations between the different sIN . Hence sIN = 0 for all N separately. Introduce
t = h . One may similarly write s = h + . . . where the unwritten terms do not involve
b IN = N IN . Thus
b where N
b = + h is such that N
or h. It follows that st + ts = N
b IN = (s t + t s)IN = s t IN ,
N IN = N
(6.70)
and we conclude that for every N 6= 0 one has IN = s N1 t IN ), i.e. IN is BRST-exact. Hence,
P
1
I = I0 + s with =
N =1 N t IN ). Now, I0 cotains no h and no and, having ghost-number
zero, it cannot contain any either. Thus I0 = I0 [A , ], as was to be proven.
105
7
7.1
7.1.1
Renormalization of non-abelian gauge theories

Slavnov-Taylor identities and Zinn-Justin equation
Slavnov-Taylor identities
Recall that the vacuum expectation values of time-ordered products of field operators can be obtained
from the generating functional
Z

Z
4
n
Z[J] = DA DDDD Dh exp i d x (Lnew +
e Jn ) ,
(7.1)
where, to simplify the notation,
en stands collectively for any of the fields A , , , , or h.
Of course, we do not really need to compute ghost correlation functions to get S-matrix elements
between physical states, but they certainly can and do appear in subdiagrams. Also, the form of Lnew
originally was derived only for time-ordered products of gauge invariant operators, but we can take
(7.1) as a definition for the gauge-dependent vacuum expectation values of time-ordered products
of the gauge non-invariant field operators. We have seen in the last subsection that the BRST
invariance implies that in the end matrix elements between gauge invariant states do not depend on
the gauge-fixing functional. We will now derive the implications of BRST invariance of the action
(and measure) for the generating functional Z[J] as defined in (7.1). These are the Slavnov-Taylor
identities.
W [J]
As usual, we define Z[J] = eiW [J] , as well as n = RJ
= he
noperator i with Jn, being the solution
n
R
of this equation for given n . Then the Legendre transform is [] = W [J ] d4 x n Jn, . Note
that, since n can be either bosonic or fermionic (i.e. anticommuting), one must specify whether
functional derivatives should act from the left (L) or from the right (R). Indeed, for fermionic n one
has L n = R n .
R
As shown in general in section 1.4.3, the invariance of Lnew and of the functional integral
measure under
n = n ,
n = s n ,
(7.2)
(dropping the tilde on the n ) implies the Slavnov-Taylor identity
Z
L
=0.
d4 x hn (x)iJ
n (x)
(7.3)
Of course, invariance of the functional integral measure under BRST transformations is not guaranteed and needs to be verified. This will be discussed at the end of this subsection. The conclusion
will be that the measure is indeed invariant under BRST transformations provided the measure for
the matter fields is gauge invariant, i.e. there are no gauge anomalies. For now, we assume that
this is the case and the Slavnov-Taylor identity (7.3) does hold. As discussed in sect. 1.4.3, for a
linear symmetry (n = cnm m ) one would have hn (x)iJ = cnm hm (x)iJ = cnm m , and then the
Slavnov-Taylor identity just states that is invariant under this symmetry. At present, however, the
BRST symmetry is non-linear, e.g. s = i, etc, and the Slavnov-Taylor identity (7.3) does not
tell us that the effective action is also BRST invariant.
106
7.1.2
Zinn-Justin equation
In order to nevertheless exploit the content of the Slavnov-Taylor identity, one uses the trick to also
introduce sources for the composite fields n . (This is somewhat similar to what we did when
discussing the renormalization of composite operators and the Callan-Symanzik equations for vertex
functions (n) with l additional insertions of some composite operator like e.g. 2 .) To begin with,
one defines
Z

Z
Z[J, K] eiW [J,K] =
Dn exp i
d4 x (Lnew + n Jn + n Kn )
(7.4)
The Kn are like additional (position-dependent) coupling constants. In particular,

R W [J, K]
= hn (x)iJ,K .
Kn (x)
(7.5)
The effective action [, K] is obtained by performing the Legendre transformation with respect to
the sources Jn only, while keeping these extra couplings Kn :
Z
n
[, K] = W [J,K , K] d4 x n (x)J,K
(x) ,
(7.6)
n
where J,K
is the solution of
R W [J, K]
= n
(7.7)
Jn
n
for given n . (For obvious notational reasons, we write J,K
rather than (Jn ),K .) Note that these
n
definitions imply on the one hand, that and Jn have the same statistics (both bosonic or both
fermionic) and n and Kn also have the same statistics. On the other hand, any field from the first
group (n or Jn ) has opposite statistics from any field of the second group (n or Kn ). Through the
usual manipulations one finds from (7.6)
Z
Z

m
m
R J,K
(y)
(y)
R J,K
R W [J, K]
R [, K]
4 R W [J, K]
=
+ dy
d4 y m (y)
. (7.8)

Kn (x)
Kn (x) J=J,K
Jm (y) J=J,K Kn (x)
Kn (x)
The last two terms cancel by (7.7) while, by (7.5), the first term is just hn (x)iJ,K ,K . Hence
R [, K]
= hn (x)iJ,K ,K .
Kn (x)
(7.9)
It is completely straightforward to generalize the Slavnov-Taylor identity to the case where the
additional couplings n Kn are present:
Z
L [, K]
d4 x hn (x)iJ,K ,K
=0.
(7.10)
n (x)
Using (7.9), we get the Zinn-Justin equation:
Z
d4 x
R [, K] L [, K]
=0.
Kn (x)
n (x)
107
(7.11)
7.1.3
Antibracket
The Zinn-Justin equation displays a nice symmetry between the roles of Kn and n . This is further
emphasized by introducing the notion of antibracket. The antibracket of two (bosonic) functionals32
F [, K] and G[, K] depending on two sets of arguments n and Kn having opposite statistics is
defined as

Z
R F [, K] L G[, K] R F [, K] L G[, K]
4
.
(7.12)
(F, G) = d x
n (x) Kn (x)
Kn (x) n (x)
This is somewhat similar to the definition of the Poisson bracket but, as we will see, it is symmetric
rather than antisymmetric under the exchange of F and G. Since the n always have the opposite
statistics from the Kn , one of the two functional derivatives in the definition of the antibracket
always yields an anticommuting expression while the other is commuting. Since exchanging a left
with a right derivative yields a minus sign for an anticommuting expression, we can flip left and right
derivatives in either of the two terms in (7.12) provided we include one extra minus sign. Thus

Z
R F [, K] L G[, K] L F [, K] R G[, K]
4
+
(F, G) =
dx
n (x) Kn (x)
Kn (x) n (x)

Z
R F [, K] L G[, K] R G[, K] L F [, K]
4
=
dx
+
n (x) Kn (x)
n (x) Kn (x)

Z
R G[, K] L F [, K] R F [, K] L G[, K]
4
= dx
(7.13)
+
.
Kn (x) n (x)
Kn (x) n (x)
This shows that the antibracket is symmetric under the interchange of F and G:
(F, G) = (G, F ) .
(7.14)
As is obvious from the last line in (7.13), the Zinn-Justin equation (7.11) can then be written as
(, ) = 0 .
7.1.4
(7.15)
Invariance of the measure under the BRST transformation
Let us now come back to the question of whether the functional integral measure is invariant under
the BRST transformation or not. Clearly, even if the measure is invariant under gauge transformations, this does not immediately imply invariance under BRST transformations. The reason is
that the latter are non-linear transformations of the fields A , , , , , h while the former are linear transformations of A , and only. A further technical complication appears since the BRST
transformation mixes commuting and anticommuting field variables. Nevertheless, we will now show
that the relevant Jacobian equals unity provided the measure for the matter fields alone is gauge
invariant.
32
A bosonic functional is a sum of terms each of which contains an even number of anticommuting fields. Hence it
is commuting. Similarly, a fermionic functional is made from terms containing an odd number of anticommuting fields
and hence is anticommuting.
108
We want to compute the Jacobian for a change of integration variables of the form n n =
n + n n + sn where n stands for the various commuting and anticommuting fields and
is a single anticommuting parameter satisfying 2 = 0. Although the n depend non-linearly on the
r , the fact that 2 = 0 will simplify the Jacobian enormously. We let

0n
n
(x)
(x)
L
L
= rn (4) (x y) +
rn (4) (x y) + Rnr (x, y) .
(7.16)
J nr (x, y) =
r (y)
r (y)
b and hence R2 = 0, so that log(1 + R) = R. We can then use
Note that R is linear in , i.e. R R,
the standard relation between the determinant and the trace of the logarithm to get

Det J = exp Tr log(1 + R) = exp Tr R = 1 + Tr R ,
(7.17)
where Tr includes a sum over the different field types of functional traces for every field. Now, for
n
n
anticommuting r (i.e. the ghost, antighost or matter fermions), one has L
= Lr . It follows
r
that one has
b,
Tr R = Str R
(7.18)
where Str is a functional supertrace which is just an ordinary functional trace but with minus signs
inserted for the anticommuting fields. With obvious notations we have
b = Tr R
bA Tr R
b Tr R
b + Tr R
bh Tr R
b Tr R
b .
StrR
(7.19)
Recall that A0 = A + sA = A + ( i[A , ]) so that A = ( i[A , ]). Similarly,
b = Tr R
bh = 0.
= i and = h, as well as = i and h = 0. Obviously then, Tr R
(4)
b ) (x, y) = C (x) (x y) so that its trace vanishes (after an appropriate gauge
Also (R
invariant regularization). Similarly,
bA ) (x, y) =
(R
(x)
= C (4) (x y) .
A (y)
(7.20)
This is again antisymmetric and its trace vanishes (after an appropriate gauge invariant regularization). Actually, this is the same operator as the one we encountered when we discussed the
invariance of the gauge field measure under gauge transformations (except that now ). ObbA = 0 is equivalent to the statement that the gauge field measure is gauge invariant.
viously, Tr R
Finally, one has to study Tr R and Tr R . Once more, the vanishing of the trace is equivalent to
the gauge invariance of the matter measure. Although formally this trace vanishes, just as for the
other fields, here we may encounter a problem if it is not possible to regularize the trace in a gauge
invariant way. This is precisely the case for chiral fermions and one has a so-called anomaly as we
will extensively discuss in the next section.
We conclude that in the absence of anomalies (no chiral matter fermions), the full functional
integral measure is indeed invariant under BRST transformations. Moreover, the BRST invariance
of the measure is equivalent to the gauge invariance of the matter measure alone.
109
7.2
7.2.1
Renormalization of gauge theories theories

The general structure and strategy
As already emphasized, non-abelian gauge theories are renormalizable by power counting. As discussed in sect. 4, it follows from power counting that there only are finitely many Green-functions
(or equivalently only finitely many 1PI n-point vertex functions (n) ) that are divergent with there
superficial degree of divergence being some fixed finite number (independent of the order of perturbation theory). Relying on the BPHZ theorem, we then know that all divergences can be removed by
finitely many local counterterms whose coefficients can be determined order by order in perturbation
theory. However, what this does not tell us is whether the divergent parts of the Green functions or
of the (n) and hence the corresponding counterterms share the same symmetries as the original
action. If one needed to add non gauge invariant (or actually non BRST invariant) counterterms at
some order N of perturbation theory to cancel some non-invariant divergence, these non-invariant
counterterms would almost certainly lead to non-invariant contributions at some higher order N 0 > N
of perturbation theory. Thus, when one asks whether a (non-abelian) gauge theory is renormalizable
the question is whether all counterterms share the same symmetries as the original (gauge-fixed)
action, namely the BRST symmetry.
The proof of the renormalizability of non-abelian gauge theories thus amounts to showing that
the divergent parts of the quantum effective action (which generates the 1PI vertex functions
(n) ) due to the loops still has the BRST symmetry, or some deformation thereof obtained after
changing the coupling constant and field normalizations. Instead of actually computing the loop
diagrams, we will exploit the algebraic structure coded in the Zinn-Justin equation to show that this
is indeed the case.
R
As usual, we write rewrite the original bare action (including now the extra couplings n Kn )
in term of renormalized fields and coupling constants and masses:
(0)
SB [B , KB ] = S[, K] = SR [, K] + Sc.t. [, K] ,
(7.21)
(0)
where SR has the same form as the original (bare) action but with all masses and couplings equal
to their renormalized values, and Sc.t. are the counterterms, c.f. the general discussion and examples
studied in sections 2 and 3. There are now two ways to rephrase the question of renormalizability
of the non-abelian gauge theories. If one insists that SB is BRST invariant, as well as the tree(0)
level renormalized action SR , then the counterterms will also be BRST invariant. The non-trivial
question then is whether these counterterms are enough to cancel all the divergent parts that arise
(0)
in any N -loop diagram. Equivalently, we may just begin with the usual BRST invariant SR and
compute the divergent parts of and adjust the counterterms Sc.t. order by order in perturbation
theory to cancel these divergent parts. The question then is whether these counterterms are BRST
invariant, i.e. whether the bare action can be BRST invariant.
Actually, there is is slight subtlety here: in the non-abelian case, the gauge and BRST transformations explicitly depend on the gauge coupling constant (which we have hidden in the generators
t , respectively in the structure constants). Thus, when one states that some quantity is gauge or
110
BRST invariant, one has to specify what is the coupling constant. Now the bare action SB is BRST
(0)
invariant using the bare coupling constant, while the renormalized action SR should be BRST invariant using the renormalized coupling constant. The precise formulation of renormalizability then
is whether we can find counterterms that cancel the divergent part and which equal the differ(0)
ence of the gB -BRST invariant bare action SB and the gR -BRST invariant action SR . Equivalently,
(0)
(0)
starting from a gR -BRST invariant SR , we must show that SR is gB -BRST invariant. In the
sequel, we will calla conterterms or simply BRST-invariant if it fulfills this requirement.
In perturbation theory, has an expansion
[, K] =
N [, K] ,
(0)
0 = SR
N = N loop +
N =0
N
X
c.t.,M
(N M )loop ,
(7.22)
M =1
where c.t.,M
(N M )loop is an (N M )-loop contribution involving vertices from lower-order counterterms
of total order M , resulting in a contribution at the same order in perturbation theory as N loop .
(In QED e.g., the order 2 contributions to the vacuum polarisation correspond to
2 and are given
by the genuine two-loop diagrams (2loop ), one-loop diagrams with one-loop (order ) counterterms
c.t.,2
c.t.,1
), as well as a new order-2 counterterm (
inserted (1loop
0loop ).
(0)
(0)
One can again introduce a loop-counting parameter and formally replace SR by 1 SR . Then an
N -loop term will have a coefficient N 1 , and N loop will be multiplied by N 1 . We can similarly
assign a factor M to every counterterm that arises from an M -loop diagram, in addition to an
P
c.t.,M
M M
overall 1 in front of Sc.t. : 1 Sc.t. = 1
M =1 Sc.t. . Then (N M )loop will be accompanied by a
P
N 1
factor N M 1 M = N 1 . As a result, we see that =
N . Inserting this into the
N =0
Zinn-Justin equation (, ) = 0 and collecting the coefficients of the N 2 -terms (recall that the
antibracket is symmetric) yields
N
X
(N 0 , N N 0 ) = 0 ,
N 0.
(7.23)
N 0 =0
We want to show that can be made finite by choosing BRST-invariant local counterterms
order by order in perturbation theory. We will show this by induction in N . First, for N = 0 one
(0)
(0)
has 0 = SR . Now, SR is expressed in terms of the renormalized (finite) parameters (couplings and
masses) and the renormalized fields. Obviously, no counterterms are needed at this order. Suppose
then that all N 0 with N 0 N 1 are finite, i.e. we assume that all divergences in N 0 loop have
P 0
c.t.,M
been cancelled by the contributions N
M =1 (N 0 M )loop , induced by appropriately chosen (BRSTM
invariant) counterterms Sc.t.
with M N 0 . This shows that in the sum (7.23) the only terms that
can involve divergences are the terms N 0 = 0 and N 0 = N . Thus, isolating the possibly divergent
parts, eq. (7.23) yields
(0)
(SR , N, ) = 0 .
(7.24)
In agreement with our above remark, we can interpret this constraint in two different ways.
First, we may consider that N contains the N -loop contributions N loop , the contributions
P 1 c.t.,M
from lower-order counterterms in loops N
M =1 (N M )loop (the corresponding counterterms are
111
BRST-invariant by the induction hypothesis) and an order-N counterterm c.t.,N

tree . The latter
is taken BRST invariant and we must then show that it can indeed cancel the divergent part
from the other contributions. This will be the case if the latter are BRST invariant. Since
the counterterm is already supposed to be BRST invariant, it is equivalent to show that the
divergent part N, of the full N is BRST invariant. (Of course, if this is the case, the
divergent parts just cancel and N, = 0.)
Second, we may consider that we have not yet included an order-N counterterm c.t.,N
in the
tree
computation of N which now only contains N loop and the contributions from lower-order
P 1 c.t.,M
counterterms in loops N
M =1 (N M )loop . Then, to be able to cancel the divergent part N,
of N by an appropriate BRST invariant counterterm, we must show that N, is BRST
invariant.
We see that, whatever interpretation we adopt, it is necessary and sufficient to show that N, is
BRST invariant. This will be done by :
exploiting the content of (7.24) which we derived from the Zinn-Justin equation,
(0)
using the various linear symmetries33 of the tree-level action SR which must be also be symmetries of the effective action at every order, i.e. of N , and in particular also of its diverging
part,
using the fact that N, is a local functional of the fields of dimension less or equal to 4 (here
we rely of course on BPHZ to exclude any trouble with overlapping divergences).
Let us recall that all computations are done in the presence of the extra couplings n Kn , i.e. with
R
(0)
SR [n , Kn ] = (Lnew,R [n ] + n Kn ). Here n = sn with the BRST-transformation s being the
(0)
one involving the same renormalized coupling as appears in SR . Let us also insists, that eq. (7.24)
will only tell us something about the diverging part of N , not its finite part.
7.2.2
Constraining the divergent part of
Let us now constrain the divergent part N, of N using dimensional arguments, the linear symmetries and equation (7.24). We begin by showing that N, can depend at most linearly on the
various Kn .
Dimensional arguments
The gauge field A has dimension 1, just as scalar matter fields. The fermionic matter fields, and
have dimension 32 . The dimensions of the ghost and antighost fields can be read from their kinetic
term (or their propagator). This depends on the choice of gauge-fixing function f . If f = A +. . .,
the ghost Lagrangian is D + . . . and dim + dim = 2. Since in any ghost-number zero
33
These are Lorentz transformations, global gauge transformations, ghost phase transformations related to ghost
number conservation and possibly antighost translations for certain choices of the gauge-fixing function like e.g.
f = A .
112
functional one always has as many ghosts as antighosts it does not matter how we distribute the
2 between the ghost and antighost and we can choose dim = dim = 1. As one sees from
s A = D , s = i, etc, s changes dimensions by dim = 1. It follows that for any field n one
has dim n = dim n + 1. Finally, n Kn must have dimension 4 so that dim Kn = 4 dim n . To
summarize:
dim A = dim = dim = 1
dim A = dim = dim = 2
dim KA = dim K = dim K = 2
3
,
2
5
dim = dim = ,
2
3
dim K = dim K = .
2
dim = dim =
(7.25)
Since the divergent part of must be a local functional of all the fields of at most dimension 4, we
see that it can be at most quadratic in the Kn .
Ghost number conservation
Since Lnew is invariant under global ghost/antighost phase rotations ei , ei it
follows that the ghost number is conserved ( has ghost number +1 and ghost number 1). This
will remain true in the presence of the extra couplings n Kn if we assign ghost numbers to the
Kn which are the opposite of the ghost numbers of n . Since n = s n and s increases the ghost
numbers by one unit, one has ngh n = ngh n + 1. Thus
ngh (A ) = ngh () = ngh () = 0
ngh () = 1 , ngh ( ) = 1 ,
ngh (A ) = ngh ( ) = ngh ( ) = 1
ngh ( ) = 2 , ngh ( ) = 0 ,
ngh (KA ) = ngh (K ) = ngh (K ) = 1
ngh (K ) = 2 , ngh (K ) = 0 , .
(7.26)
Hence, for all Kn have negative or zero ghost number.

Now, N, must have ghost number zero and we have seen above that it can be at most quadratic
in the Kn . Lets see which quadratic terms could appear. Terms involving two Ka , K or K have
dimension 4 and hence cannot involve any other field. Such terms have negative ghost numbers and
are excluded except for K K . The latter term, however, is excluded since s = h is a
R
linear transformation so that K
= h i = h tells us that [n , Kn ] must be linear in
K and cannot have a term K K . Actually, this argument tells us even a bit more: the K
R
dependence of [, K] must be precisely a term d4 x (h)K . This is just the (finite) term already
(0)
present in the tree-level action SR [, K] and, in particular, the diverging part cannot contain
any term linear in K . Similarly, h = sh = 0 and we conclude that cannot contain a term
linear in Kh either.
Next, since the ghost is a Lorentz scalar, all n transform under the same representation of the
Lorentz group as the fields n and the Kn must transform in such a way that n Kn is a Lorentz
scalar. It follows that K and K are Lorentz scalars, KA is a four-vector and K transforms as
the spinor , while K transforms as . Thus a term quadratic in the Kn and involving at least
one K or K must necessarily involve both of them, i.e. be K K . This expression has ghost
113
number 2 and dimension 3. There is no dimension 1 field of ghost number +2 available to make a
dimension 4, ghost number 0 term. Thus we conclude that N, is at most linear in any of the Kn :
Z
e nN (x)Kn (x) ,
N, [, K] = N, [] + d4 x
(7.27)
with N, [] N, [, 0] =
d4 x N, (x) . Similarly, we had for the (tree-level) action

Z
(0)
(0)
SR [, K] = SR [] + d4 x n (x)Kn (x) .
(7.28)
e n . Note also that

Just as n is a local expression in the fields n , the same must be true for the
N
n
n
e
the N must have the same ghost numbers and dimensions as the . Finally, we have seen above
that N, cannot contain any terms linear in K or Kh and we conclude that
e N =
e hN = 0 .
(7.29)
(0)
Exploiting (SR , N, ) = 0
e n (x), we insert (7.27) and (7.28)
In order to get further information on the form of N, (x) and
N
(0)
into (SR [, K], N, [, K]) = 0:
!
Z
(0)
(0)
[,
K]
S
[,
K]
[,
K]
[,
K]
S
R R
L N,
R N,
L R
0 =
d4 x
+
Kn (x)
n (x)
Kn (x)
n (x)
(
!
Z
Z
em
L N, []
4
n
4 L N (y)
=
d x (x)
+
d
y
Km (y)
n (x)
n (x)
!)
Z
(0)
m
(y)
[]
S
L
e nN (x) L R
+ d4 y
Km (y)
.
(7.30)
+
n (x)
n (x)
The terms without Km and the coefficients of Km (y) must vanish separately, giving two equations:
!
Z
(0)
L N, [] e n
L SR []
4
n
+ N (x)
=0,
(7.31)
d x (x)
n (x)
n (x)
and
Z
d4 x
m
e m (y)
L
N
e nN (x) L (y)
+
n (x)
n (x)
n (x)
!
=0.
(7.32)
The second equation (7.32) is a set of functional first-order partial linear differential equations for
e n . Note that these equations do constrain the functional form of the
e n but not their overall
the
N
N
n
e
normalization. Inserting the solutions into the first equation then should determine the functional
N
form of N, []. Note that these two equations (7.31) and (7.32) do not explicitly involve the Kn
anymore. Indeed, all we really want to determine in the end is the form of N, [] = N, [, 0]
R n
e n (x), we also know N, [, K] = N, [, 0] +
e (x)Kn (x).
although, once we know
N
N
114
To solve these equations in practice, it is useful to give them a more physical interpretation.
Define
()
(0)
()n
e nN (x) .
N [] = SR [] + N, []
and
N (x) = n (x) +
(7.33)
Consider the deformed BRST transformation
()n
s() n (x) = N (x) .
(7.34)
Obviously, for = 0 this reduces to the ordinary BRST transformation. It is then not difficult to see
()
that, up to first order in , N [] is invariant under this transformation and that this transformation
is nilpotent:
()
s() N [] = 0 + O(2 ) by (7.31)
s() s() n = 0 + O(2 ) by (7.32) .
and
(7.35)
Deformed BRST symmetry

e n couples to Kn in N, [, K], in the same way as n couples to Kn
First note (again) that, since
N
(0)
n
n
e
in SR , both N and must have the same ghost numbers, dimensions and Lorentz transformation
properties.
The strategy now is to write the most general nilpotent transformation of the fields that increases
the ghost numbers of the fields by one unit, increases their dimensions by dim and that is a
deformation of the usual BRST transformation in the sense that it depends on some small parameter
and reduces to the ordinary BRST transformation in the 0 limit. (Of course, nilpotency is
only required up to terms of order 2 .) These requirements will turn out to be stringent enough
to show that any such deformed BRST transformation must be identical to the ordinary BRST
transformation, up to changes in the normalization of the fields and a change of the gauge coupling
constant.
It is then easy to see that the deformed BRST symmetry must act as
= i T ,

A = B + D A ,
1
= E .
2
(7.36)
(Having the correct ghost number would also allow e.g. a term of the form i F T ,
e =
eh = 0
but this is excluded by the argument about the dimensions.) Recall from (7.29) that
N
N
and, hence, the deformed BRST transformation of and h just equals the undeformed one:
= h
h = 0 .
(7.37)
Let us now exploit the nilpotency of these transformations (7.36) and (7.37). First, 1 2 = 0
yields E E = 0 so that E E must vanish after antisymmetrizing in , , . i.e. the
E satisfy the Jacobi identity and must be the structure constants of some Lie algebra. As 0,
115
these are just the C and we conclude that this Lie algebra is the same as the original Lie algebra
G and the E differ from the C just by the normalization:
E = Z C .
(7.38)
(Note that if the Lie algebra is semi-simple one may have a different Zi for every simple factor.) Next,
looking at 1 2 A gives D D D D = E D = Z C D and B E = D B .
b with (D
b ) = 1 D are the generators of the
The first equation implies that the matrices D
Z
adjoint representation of the Lie algebra G, so that
D = ZC .
(7.39)
Hence, D = E and the second condition involving B just states that the matrix B commutes
b and thus is proportional to the unit matrix:
with all matrices D
B = Z N .
(7.40)
(Again, for a semi-simple G we can have a different Ni for every simple factor.) Finally, looking at
1 2 implies [T , T ] = iE T = iZC T and we identify
T = Z t .
(7.41)
We conclude that, apart from the new constants Z and N , the deformed BRST symmetry se must
act exactly as the ordinary BRST symmetry s we started with. Let us summarize:
se = iZ ,

se A = Z N i[A , ] ,
se = iZ
se = h
se h = 0 .
(7.42)
We can absorb these new constants by redefining the normalizations of the ghost fields and redefining
the coupling constant (hidden in the C ) as follows:
e = Z N
e = A
A
e = 1 C .
C
N
(7.43)
t = Z t = Z and
Then the generators are redefined accordingly by e
t = N1 t and thus
e=
e e
e = 1 A . Furthermore, one can redefine the normalization of by some factor and the one of
A
N
and h by another (common) factor. These latter factors are not yet determined by (7.42). In terms
of the redefined fields and couplings, the deformed BRST transformations (7.42) take exactly the
form of the undeformed ones.
116
Constructing a Lagrangian invariant under the deformed BRST symmetry : constraining LN,
R ()
()
()
Let us now construct N [] = LN [] with a Lagrangian LN that is invariant under the deformed
BRST transformation. Actually, since we have just seen that the latter is the same as the original
BRST transformation up to changes of normalizations our task amounts to constructing the
most general BRST invariant Lagrangian34 in the redefined fields and couplings. Of course, as
explained before, we also require that it is of dimension less or equal to four and invariant under all
linear symmetries of SR , namely Lorentz symmetry, global gauge symmetry, ghost phase rotations
(implying total ghost number zero), and (in most gauges) antighost translations, in which case
()
must appear as . Obviously, once we have constrained the form of LN = Lnew + LN, we also
have constrained the form of the diverging part LN, of LN .
Terms involving the ghosts can either involve one ghost and one antighost, or two ghosts and two
antighosts. The latter could not involve any other field or derivative for dimensional reasons. Thus,
if antighost translation is an invariance, no terms with two ghosts and two antighosts are allowed.
The terms with one ghost and one antighost must then be of the form or A .
Since dim h = 2, terms involving h but no ghost can be hh or h A or hA A (and also
h2 if scalar matter is present). We arrive at
0
h h + c h A e h A A Z d A . (7.44)
2
where global gauge invariance implies that d and e are both proportional to the structure
()
constants C . Now require invariance under the deformed BRST symmetry se, i.e. se LN = 0.
Z
Cancellation of the terms h in yield the condition c = ZN
, while cancellation of the terms
Z
h A or yields d = N C . Terms A then automatically cancel

by the Jacobi identity, and the cancellation of terms h A or hA A yields e = 0.
Finally, the action of se on L0N [A, ] is the same as the one of the undeformed BRST operator s
e . Then, cancellation of terms with
except that ZN a =
e and C N1 C = C
only one ghost field and no h or anti-ghost field implies that L0N must be gauge invariant. More
= ZN = ZN
precisely, L0N must be invariant under gauge transformations with parameter e
e = 1 C . Hence, L0 [A, ] must be of the form
and renormalized gauge coupling constants C
N
N
e ], where Fe and D
e are defined as F and D but using the C
e
14 ZA Fe Fe + Le0N [, D
instead of the C , and where Le0N [, ] is invariant under global gauge transformations. Thus we
arrive at
0

1
()
e ] + h h + Z h A Z D
e ,
(7.45)
LN = ZA Fe Fe + Le0N [, D
4
2
ZN
where
()
LN = L0N [A, ] +
e A A
= A A +C
Fe
e = iAae
D
tR
e ) = +C
e A . (7.46)
(D
This is of the same form as the initial Lnew except for the appearence of the renormalization
e , as well as further renormalization
constants ZA , Z , a changed coupling via the N1 in C
34
Strictly speaking, we only require invariance to first order in the deformation parameter , but because of the
rather rigid algebraic structure, this actually results in invariance to all orders in .
117
e ], and the replacement 0 . Let us insist that, on the redefined

constants hidden in Le0N [, D
fields, the deformed BRST symmetry is just a BRST symmetry with a redefined coupling constant
()
ge = g/N , and LN is invariant under this ge-BRST symmetry for all .
7.2.3
Conclusion and remarks

()
It is enough to recall that LN = Lnew + LN, to see that the diverging part LN, of LN is BRSTinvariant in the sense explained above, i.e. it is the difference of a ge-BRST and a g-BRST invariant
local term. Since was arbitrary, the same is true for LN, which is the required counterterm.
One can now redo the argument order by order in the loop-expansion to see that the sum of all
counterterms up to a given order N is BRST-invariant in this sense. Hence we conclude that to
any order in perturbation theory the renormalized action is g-BRST invariant precisely if the bare
action is gB -BRST invariant. Said differently, by appropriately choosing the field renormalization
constants and coupling constant renormalization we get a finite quantum effective action N at every
order N of perturbation theory.
Let us look more explicitly at the renormalization of the gauge field and of the gauge coupling
constant g. Recall that the latter was included in the Lie algebra generators as t = g b
t and
b with b
b the more conventionally
accordingly in the structure constants as C = g C
t and C
normalized, coupling-independent generators and structure constants. Thus F

= A A +
b A A , and similarly for F with AB and gB . If we let
gC
B
p
p
(7.47)
AB = ZA A , gB = Zg g ,
the relation between the bare and renormalized F 2 term is
1
1 1
FB
FB = F
F (ZA 1) ( A A )( A A )
4
4
4
1 p
3/2
b A A
g( Zg ZA 1) ( A A )C
2
1 2
b A A C
b A A .
(7.48)
g (Zg ZA2 1) C
4
To determine only ZA , at a given order, it is enough to compute the two-gauge boson function, i.e.
the vacuum-polarization. However, to determine Zg and hence the renormalization of the coupling
constant, one has to determine two of the three counterterms on the rhs of (7.48) and hence compute
e.g. the vacuum-polarization and the 3-gauge boson vertex function. This is related to the fact that
the bare and renormalized actions have their BRST invariance defined with different couplings. In
an abelian gauge theory like QED, there is no coupling dependence of the BRST symmetry and the
Ward identity implies (Zg ZA )abelian = 1 so that we could read the coupling constant renormalization
from the vacuum-polarization. Below, we will discuss a special type of gauge fixing, the so-called
background gauge, where an analogous relation holds also in the non-abelian theory and, hence, it
is enough to compute a single Greens function to obtain the coupling constant renormalization and
the -function.
We have used the anti-ghost translation invariance of the so-called -gauges f = A , and this
is why we get back this form. However, if one uses a gauge-fixing with f = A + a A A
118
with some global G-tensor a , there is no -translation invariance to enforce the absence of BRST

invariant terms with two ghosts and two anti-ghosts, like e.g. se b = b 2h +

1
ZC . Hence, such terms must be allowed in the action (and counterterms) from the
2
beginning. However, such terms cannot be obtained from the Faddeev-Popov procedure (which
always gives a ghost term linear in the ghost and linear in the anti-ghost). To get these terms
necessitates a more general Lagrangian Lnew = L[A, ] + s[A, , , , h], as already discussed
above.
We can conclude that, as long as we start with a general Lagrangian of this form, with gauge
invariant L of dimension less or equal 4, there is a counterterm available to cancel any divergence at
any loop order, and the theory is renormalizable in the usual sense.
R
Heavy use was made of the Slavnov-Taylor identity d4 x hs n i

n = 0 from which the ZinnJustin equation was obtained. The derivation of the Slavnov-Taylor identity assumed that
R
R
n
Dn ei Lnew [ ] is BRST invariant and in particular that DDA ei L[A,] is gauge invariant. Now
R
L[A, ] is gauge invariant by construction but it may happen that D is not gauge invariant. This
typically shows up at one loop and then (SR , 1 ) 6= 0. Equivalently, certain one-loop diagrams are
then seen to violate the corresponding Ward identities of covariant current conservation. In this case
the theory is said to be anomalous. Anomalies are the subject of the next chapter. Fortunately, as
we will see, anomalies can only appear for certain gauge groups and only in the presence of chiral
fermions. Moreover, in potentially anomalous theories, by carefully arranging the content of the
chiral matter fermions, the anomalies can be made to cancel. This is the case, in particular, for the
standard model based on the gauge group SU (3) SU (2) U (1) with the content of chiral matter
fermions observed in nature.
7.3
Background field gauge
Suppose we can find a gauge-fixing procedure such that the one-loop quantum effective action35
1loop [a, 0 , 0 , 0 ] must be gauge invariant under the same transformation as was the tree-level
(0)
b a and 0 = ig b
renormalized action SR , i.e. a = + g C
tR
0 , as well as some
appropriate transformations for 0 and 0 . Then necessarily we have

1 1
1
(0)
1loop
F + . . . = F
F
(ZA 1)F
F + . . . ,
SR [A, . . .] + Sc.t.
[A, . . .] + . . . = ZA F
4
4
4
(7.49)
which is the bare action. Comparing with (7.48), we see that then Zg ZA = 1 and, hence, gB =
(ZA )1/2 g. Since the one-loop -function is given by the coefficient of 1 in the expression of gB in
terms of g, we then find that36

1
1
1loop = g coeff of in ZA .
(7.50)
2

35
Although we do not want to compute scattering amplitudes or Greens functions for external ghosts, nothing
prevents us from computing also with ghost background fields 0 and 0 .
36
To lowest order in perturbation theory gB = (ZA )1/2 g = (1 + (ZA 1))1/2 g ' (1 21 (ZA 1))g.
119
This will allow us to extract the one-loop -function from the computation of the single coefficient
of the F 2 -term in 1loop .
With this motivation in mind, let us now introduce the background field gauge. Recall equations
(1.89) and (1.91) which compute [a] from S[a + A] by integrating over the quantum gauge field A
(with the restriction to 1PI diagrams - which is irrelevant at one loop). We want to find a gauge-fixing
for A such that [a] still is invariant under gauge transformations of the background field a .
We introduce two types of gauge transformations: the background field gauge transformations
B and the quantum field gauge transformations g . The background field gauge transformation
is defined to act on the background and quantum fields as follows:
B a = i[a , ] DB
B a = + C a ,
B A = i[A , ]
B A = C A .
(7.51)
Thus under this background gauge transformation a transforms as a gauge field and A as an adjoint
matter field. Obviously, on the sum of both fields B acts as an ordinary gauge transformation:
B (a + A ) = i[a + A , ] .
(7.52)
Similarly, the action on the matter quantum and background fields is defined as
B 0 = i tR
0
B = i tR
B (0 + ) = i tR
(0 + ) ,
(7.53)
B (0 + ) = C (0 + ) ,
(7.54)
and for the ghost fields

B 0 = C 0
B = C
and idem for the antighost fields. On the other hand, the quantum field gauge transformations g
should not act on the background fields and will be defined such that they do act as standard gauge
transformations on the sums a + A and 0 + . Thus
g a = g 0 = g 0 = g 0 = 0 ,
(7.55)
and
g A = i[a + A , ]
g (a + A ) = i[a + A , ]
g = i tR
(0 + )
g (0 + ) = i tR
(0 + ) .
(7.56)
In order to compute the one-loop quantum effective action as the functional integral (1.91) one
has to choose a gauge-fixing function f in the action S. The choice for the background field gauge
is
(7.57)
f = (DB A ) A + C a A .
Note that this is a generalization of the standard gauge fixing function to a non-vanishing background
field a . It is an appropriate choice for the present purpose, since DB is covariant under B and thus
DB A transforms as A under B :
B f = C f
B (f f ) = 0 ,
120
(7.58)
and we see that the gauge-fixing term is invariant under the background gauge transformations B .
On the other hand, if f is to fix the gauge in the functional integral, it better be not invariant under
the quantum field gauge transformations g . Indeed, we have
g f = g (DB A ) = (DB g A ) = (DB DB

) i(DB [A , ]) ,
from which we read the corresponding ghost Lagrangian

Lgh [a , A , , ] = DB DB
[A , ] .
(7.59)
(7.60)
Since , and A transform as adjoint matter fields under B , and since DB is covariant with
respect to B , we conclude that this ghost Lagrangian is invariant under the background field gauge
transformations, B Lgh = 0. Since also 0 and 0 transform as adjoint matter fields under B , the
same remains true if one replaces by 0 + and by 0 + :
B Lgh [a , A , 0 + , 0 + ] = 0 ,
(7.61)
so that finally
B Lmod [a , A , 0 + , 0 + , 0 + ] = 0 ,
Lmod = L
1
f f + Lgh .
2
(7.62)
Note that L[a + A, 0 + ] = 14 F [a + A]F [a + A] + Lmatter [a + A, 0 + ] with F [a + A] =

F [a] + DB A DB A i[A , A ].

R
R
Our goal was to determine [a , . . .] as 1PI DA . . . exp i Lmod (a + A, . . .) . We see that due
to the gauge fixing and ghost terms, Lnew does not only depend on the sum a + A but on a and A
separately. It is then not obvious any more that this does generate all 1PI diagrams. However, if
we restrict ourselves to the one-loop quantum effective action, things are simpler and the separate
dependence on a and A of the gauge-fixing and ghost terms does not cause any trouble. Since Lmod
is invariant under the background gauge transformations B , as is also the measure DADDD
(excluding anomalies), it is then obvious that 1loop [a, 0 , . . .] must also be invariant under these
transformations:
B 1loop [a, 0 , . . .] = 0 .
(7.63)
But the background gauge transformations act on the background fields just as ordinary gauge
transformations (in particular with the same normalisation of g a b
t as explained at the beginning
of this subsection). Then eq. (7.49) must hold, and hence Zg ZA = 1 so that we can use (7.50) to
compute the one-loop -function.
7.4
One-loop -functions for Yang-Mills and supersymmetric Yang-Mills

theories
To extract the one-loop -function we need to compute the coefficient of 1 in ZA in the background
field gauge. (ZA 1) is the coefficient of 14 F 2 in the counterterm. Hence we need the coefficient
of 1 in front of the 41 F 2 piece of the counterterm. To extract this term, we may simply consider
a constant background gauge field a and all other background fields vanishing. In particular then,
F
(a) = C a a .
121
7.4.1
-function for Yang-Mills theory
As explained above, to compute the one-loop contribution to [a] we have to identify the part in
R
Lmod [a, A, , , ] that is quadratic in the quantum fields A, , , .
In particular,

B
B
F [a + A] F [a + A]A2 part = (D A D A )(DB A DB A ) 2F [a] [A , A ]. The expansion of Lmod then yields

Z
Z
1

A
d4 xd4 y A (x)Dx,y
Lmod [a, A, , , ] quadratic part =
[a] A (y)
2

k (x)D [a] l (y) (x)D [a] (y) , (7.64)
x,y
xk,yl
where
Dx,y
[a]

C a (x)
+ C a (y) (4) (x y) ,
x
y
(7.65)
adj
or, using a matrix notation with aadj
and the generators of the adjoint representation given
= a t
in (6.6),

Dx,y
[a] = + i aadj
(x)
i
a
(y)
(4) (x y) .
(7.66)
adj
x
y
Similarly, one has
Dx,y
[a] =
i aR
(y)
+
m
(4) (x y) ,
(7.67)
and
n

A
Dx,y
i aadj (y) + i aadj

i aadj
[a] = + i aadj
(x)
(x)
(y)
x
y
x
y
o

(4)
1
adj
(7.68)
iF
+ i aadj
i aadj
(x y) .
[a] +
(x)
(y)
x
y
Then
1
i1loop [a] = Tr log DA [a] + Tr log D [a] + Tr log D [a] .
(7.69)
2
These traces are evaluated as usual. We will sketch the computation for Tr log D [a], the others
being similar.
Since we will only consider constant fields a it it most convenient to Fourier transform. One has in
general Dx,y hx| D |yi and hence
Z
Z
1
4
4
hp| D |qi = d xd y hp| xi hx| D |yi hy| qi =
d4 xd4 y eipx hx| D |yi eiqy .
(7.70)
(2)4
With a Dx,y of the form Dx,y = f ( x ) g( y ) (4) (x y) this yields
Z
Z
1
4
4
ipx
(4)
iqy
d
xd
y
e
f
(
)
g(
)
(x
y)
e
=
d4 x ei(qp)x f (ip) g(iq)
hp| D |qi =
(2)4
x
y
(2)4
= f (ip) g(ip) (4) (p q) M (p) (4) (p q) .
(7.71)
R
Hence hp| D2 |qi = d4 k hp| D |ki hk| D2 |qi = M 2 (p) (4) (p q), and similarly for any power n and any
function of D, and thus
Z
Z
4
Tr log D = d p hp| tr log D |pi = d4 p (4) (p p) tr log M (p) .
(7.72)
122
R d4 x ix(pp)
Of course, (4) (pp) arises because we work with constant fields and it has to be interpreted as (2)
=
4 e
R 4
1
d x. Our M (p) are all of the form M (p) = M0 (p) + M1 (p) + M2 (p) with iM0 (p) being the inverse
(2)4
propagator for the given field and M1 and M2 are linear and bilinear in the background fields a . One has
tr log M = tr log M0 + tr log
1+M01 (M1 +M2 )
= tr log M0 + tr
X
()n1
n=1
n
M01 (M1 +M2 ) . (7.73)
We need to pick out the quartic term in a :

2
2
4
1
1
tr log M a4 = tr M01 M2 + tr M01 M1 M01 M2 tr M01 M1 .
2
4
(7.74)
Since iM01 is just the propagator this expansion generates the relevant one-loop Feynman diagrams
with 4 external (background) gauge fields a attached. There are vertices with one and vertices with two
gauge fields, corresponding to M1 and M2 . Obviously, one could have evaluated these diagrams directly
using the appropriate Feynman rules. The present alternative computation has the advantage of giving all
combinatorial factors in a straightforward way.
Obviously, the D [a] corresponds to the ghost loop and the above expression gives
Z
Z
d4 p
4
tr log M (p) ,
Tr log D [a] =
d x
(2)4

adj
(7.75)
M (p) =
ip + i aadj
ip i aadj = p p 2p aadj
+ a aadj .
adj
We read off M0 (p) = p p , M1 (p) = 2p aadj

and M2 (p) = a aadj and (7.74) then gives

d4 p

tr
log
M
=
a4
(2)4
tr p a p a a a
tr (p a )4
d4 p 1 tr a a a a
,
+
4
4
(2)4
2 (p2 i)2
(p2 i)3
(p2 i)4
(7.76)
where all the a are in the adjoint representation although we did not write it explicitly. Since all a are
constant, corresponding to vanishing external momenta, the loop integrals are particularly simple. Using
dimensional regularization (d = 4 ) and introducing an IR regulator , we have37

Z
dd p
1
i
2 /2 2
I
=
( )
+ O()
(2)4 (p2 + 2 i)2
16 2

Z
p p
1
dd p
I
=
I
4
2
2
3
(2) (p + i)
4
Z

p p p p
1
dd p
I
=
+ + I .
(7.77)
4
2
2
4
(2) (p + i)
24
Thus
Z 4

1
1
d p
tr log M a4 = I tr a a a a + tr a a a a + + tr a a a a
4
(2)
2
6

1
1
(7.78)
= I tr a a a a tr a a a a = I tr a a a a tr a a a a .
6
6

Recall that for constant a in the
adjoint representation one has tr adj F F = tr adj (i)[a , a ](i)[a , a ] =
2 tr adj a a a a a a a a , so that finally

Z
1
Tr log D [a] = d4 x
I tr adj F F .
(7.79)
12
Note that when computing I one replaces p p by d1 p2 . The remaining scalar integral yields
= 41 I and not d1 I as one might have naively expected. Similarly for I .
37
123
d
4I
so that
The evaluation of Tr log DA [a] is very similar, except that the tensorial structure is a bit more complicated.
In the final result, the
1
12
is replaced by a 35 . The matter determinant Tr log DA [a] depends on the mass
of the fermions. However, the structure of the divergence is mass independent, i.e. the term
1

does not
depend on the mass, and if our only purpose is to compute the -function we can just as well neglect the
fermion mass.
In the limit where the fermion masses can be neglected, the three traces are:

Z
5
A
4
Tr log D |a4 piece = I
d x tr adj F F
3

Z
1
4
tr adj F F
dx
12

Z
1
4
,
d x tr R F F
3

R dd p
1
1
i
where ( is an IR regulator) I = (2)
+ finite . If we let
4 (p2 +2 i)2 = 8 2

tr adj t t = g 2 C1
(7.80)
tr R t t = g 2 C2 ,
(7.81)
we get
div
1loop [a]
g2 1
= 2
2
5
1
1
C1 + C1 C2
6
12
3
Z
1
d x F
F
4
4

,
(7.82)
and, hence

Z

1
1
g 2 1 11
4
C1 C2
d x F F
=+ 2
.
2 12
3
4
Finally we can read off the one-loop -function as
ZA |div
1loop
1loop
g3
= 2
4
1
11
C1 C2
12
3
(7.83)

,
(7.84)
which is negative as long as C2 < 11

C , i.e. as long as there are not too many matter fields.
4 1
Of course, this is a famous result. For example, if the gauge group is SU (N ) and the matter
representation R is the defining vector representation N , one has C1 = N and C2 = 12 . For nf
n
flavours of quarks in the N -representation one has C2 = 2f . Thus

g 3 11
nf
SU (N )QCD with nf flavours of quarks
1loop
= 2
N
,
(7.85)
4
12
6
Standard QCD with gauge group SU (3) could accommodate as many as nf = 16 flavours of quarks
and still remain asymptotically free. Recall from our earlier discussion that the relevant -function
at a given scale is the one that corresponds to the number of flavours of quarks having masses
(well) below . Any quarks having masses (well) above do not contribute at this scale. Let us
n g3
g 2 ()
write nf (g) = 0 f 16
. As shown before, the running of g() is such
2 and also use () =
4
that

8 2
2
nf = exp nf 2
= exp nf
(7.86)
0 g ()
0 ()
124
does not depend on . It is the RG-invariant mass scale characterizing the strength of the interaction.
More precisely, it is independent of as long as does not come close to any of the quark masses
and the effective number of flavours does not change. Since the running of the coupling below mnf +1
is well approximated by nf and above mnf +1 by nf +1 , matching the coupling at = mnf +1 leads
to the RG scale matching condition

nf
mnf +1
0nf

=
nf +1
mnf +1
0nf +1
.
(7.87)
It is useful to isolate the contributions of the various fields to the one-loop -function. There are
3 basic contributions: the gauge and ghost fields, spin- 12 matter fields and scalar matter fields. We
already have computed the first two types of contribution, while the contribution of a complex scalar
matter field can be obtained from the one of the ghost fields. Indeed, the ghosts are (anticommuting)
scalars in the adjoint representation and with the obvious changes we get (Cadj C1 and CR C2 )
gauge and
1loop
ghost
Dirac fermion
1loop
in R
complex
1loop
scalar in R
g 3 11
Cadj
4 2 12

g3
1
= 2 CR
4
3

3
1
g
= 2 CR .
4
12
(7.88)
Note that we considered standard Dirac fermions. If one considers instead chiral (Weyl) fermions or
Majorana fermions, one has to divide the corresponding matter contribution by a factor of 2.
7.4.2
-functions in supersymmetric gauge theories
In supersymmetric theories the fields are part of supersymmetry multiplets and thus the field content
is subject to important constraints. We will now show that this leads to simpler expressions for the
one-loop -functions.
N =1:
In gauge theories with the minimal amount of supersymmetry, so called unextended supersymmetry,
often referred to as N = 1 supersymmetry, the gauge fields have a corresponding partner, called
the gaugino field, which is a spin- 12 Majorana fermion (x), also in the adjoint representation
and also massless. More precisely, the boson-fermion correspondence holds for physical states, so
that the Majorana fermion field is the partner of the gauge and ghost fields. Together they form
the gauge or vector multiplet. The matter fields are organized in so-called chiral multiplets, each
multiplet consisting of a Majorana fermion (x) and a complex scalar (x), both fields in the same
representation R of the gauge group. Since a Dirac field is equivalent to two Majorana fields, the
gaugino
g3 1
-function of a Majorana fermion is half that of a Dirac fermion. Hence, 1loop
= 4
2 6 Cadj , so that
vector multiplet
gauge and
= 1loop
1loop
ghost and gaugino
125
g3 3
Cadj .
4 2 4
(7.89)
Similarly, for a chiral multiplet

chiral multiplet
1loop
1 Dirac fermion
= 1loop
2
in R
complex scalar in R
1loop
g3
= 2
4
1
CR
4

.
(7.90)
Obviously then, in an N = 1 supersymmetric gauge theory with nC,i chiral multiplets in the representation Ri :
!
3
X
g
3
1
N =1 with nC,i in Ri
1loop
= 2
Cadj
nC,i CRi .
(7.91)
4
4
4 i
Note that such a theory with 3 chiral multiplets in the adjoint representation has a vanishing one-loop
-function. Actually, with appropriate masses and couplings, this theory actually has an extended
supersymmetry.
N =2:
In N = 2 extended supersymmetric gauge theory the N = 1 multiplets are grouped into larger
N = 2 multiplets. Thus the N = 2 vector or gauge multiplet consists of the N = 1 vector multiplet
together with a massless N = 1 chiral multiplet in the adjoint representation. Hence
N =2 vector
1loop
multiplet
g3 1
Cadj .
4 2 2
(7.92)
N = 2 matter multiplets are so-called hypermultiplets which consist of two N = 1 chiral multiplets:

g3
1
N =2 hypermultiplet
1loop
= 2 CR ,
(7.93)
4
2
so that an N = 2 theory with nH,i hypermultiplets in representations Ri has its one-loop -function
given by
!
3
X
g
1
1
N =2 with nH,i in Ri
1loop
= 2
Cadj
nC,i CRi .
(7.94)
4
2
2 i
Again, we see that the special theory with a single adjoint hypermultiplet has a vanishing one-loop
-function.
N =4:
The maximally extended (global) supersymmetry is N = 4 extended supersymmetric gauge theory.
It has a single N = 4 multiplet consisting of an N = 2 vector and a massless N = 2 adjoint
hypermultiplet. As just noted, this theory has a vanishing one-loop -function:
N =4
1loop
=0.
126
(7.95)
PART IV:
ANOMALIES
For detailed notes on anomalies, see my
Lectures on Anomalies, e-Print: arXiv:0802.0634 [hep-th].
Anomalies : basics I
8.1
Transformation of the fermion measure: abelian anomaly
8.2
Anomalies and non-invariance of the effective action
8.3
Anomalous Slavnov-Taylor-Ward identities
8.4
Anomaly from the triangle Feynman diagram: AVV
Anomalies : basics II
9.1
Triangle diagram with chiral fermions only
9.2
Locality and finiteness of the anomaly
9.3
Cancellation of anomalies, example of the standard model
10
Anomalies : formal developments
10.1
Differential forms and characteristic classes in arbitrary even dimensions
10.2
Wess-Zumino consistency conditions and descent equation
11
Anomalies in arbitrary dimensions
11.1
Relation between anomalies and index theorems
11.2
Gravitational and mixed gauge-gravitational anomalies
11.3
Anomaly cancellation in ten-dimensional type IIB supergravity

and in type I SO(32) or E8 E8 heterotic superstring
127

Advanced Topics in Quantum Field Theory

Uploaded by

Copyright:

Available Formats

Advanced Topics in Quantum Field Theory

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Topics in Quantum Field Theory

Uploaded by

Copyright:

Available Formats

Advanced Quantum Field Theory :

Renormalization, Non-Abelian Gauge Theories

Laboratoire de Physique Theorique, Ecole

2 A few results independent of perturbation theory

3 One-loop radiative corrections in 4 and QED

5 Renormalization group and Callan-Szymanzik equations

4 General renomalization theory

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

7 Renormalization of non-abelian gauge theories

-functions and asymptotic behaviors of the coupling . . . . . . . . . . . . . .

6 Non-abelian gauge theories: formulation and quantization

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

11 Anomalies in arbitrary dimensions

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

Functional integral methods

Path integral in quantum mechanics

The usual description of quantum mechanics is in the Schrodinger picture where

Qa (t) |q, ti = qa |q, ti ,

|p, ti = eiHt |pi

Pa (t) |p, ti = qa |p, ti .

Lecture notes - September 27, 2011

Functional integral in quantum field theory

Derivation of the Hamiltonian functional integral

Lecture notes - September 27, 2011

0 l (, ~x)l (, ~x) H((, ~x)(, ~x)) + iterms

Derivation of the Lagrangian version of the functional integral

In many theories the Hamiltonian is a quadratic functional of the momenta l :

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

We now apply this remark to the quadratic form given by

A few remarks are in order:

A is the quantum-mechanical operator whose matrix elements are

Lecture notes - September 27, 2011

The free propagators or simply propagators are defined as

They are not to be confused with the complete propagators (denoted 0 )

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

In a free theory one has

which allows us to rewrite

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

Green functions, S-matrix and Feynman rules

Vacuum bubbles and normalization of the Green functions

bl1 ...ln (x1 , . . . , xn ) =

l,B = Zl l,R one simply has

In the following we simply write D instead of

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

Svac,vac hvac, out |vac, ini = eivac .

Adel Bilal : Advanced Quantum Field Theory

Lecture notes - September 27, 2011

where we used the diagrammatic identity

It follows that for any operator or product of operators M one has

Lecture notes - September 27, 2011

Generating functional of Green functions and Feynman rules

Equation (1.39) can then be written as

0 l (, ~x)l (, ~x) H((, ~x)(, ~x)) + iterms

r (x) r r (x) + F r (x, )