1

Quantum Chromodynamics

M.A.Braun
Department of High-Energy Physics, St. Petersburg University,
198504 St. Petersburg, Russia
 2

1 Contents

1.Lagrangian and quantization

2.Feynman rules. General properties and problems
3.BRST invariance
4.Renormalization
5.Renormalization group. The running coupling constant
6.Solution of the Callan-Symanzik equation. High momentum asymptotic
7.Calculation of the β-function in the QCD
8.e+ e− annihilation into hadrons
9.Deep inelastic lepton-hadron scattering: kinematics
10.Composite operators and their renormalization
11.Operator expansion
12.Deep inelastic lepton-hadron scattering: evolution in Q2
13.Equations DGLAP
 3

2 Literature (In Russian)

1. M.B.Voloshin and K.A.Ter-Martirosyan ”Teoriya kalibrovochnykh vzaimodeistdviy ele-

mentarnykh chastitz” , Moskva, Energoizdat, 1984
2. F. Indurain ”Kvantovaya khromodynamika”, Moskva, Mir, 1986
3. A.Slavnov and L.Faddeev ”Vvedeniye v kvantovuyu teoriyu kalibrovochnykh polei”,
Moskva. Nauka, 1978
4. K.Itzikson and Gh.-B. Zyuber, ”Kvantovaya teoriya polya”, Vol I, II, Moskva, Mir,
1984
5. S.Vainberg ”Kvantovaya teoriya polya”, Vol I, II, Moskva, Fizmatgiz, 2003
 4

3 Lagrangian and quantization

We recall that according to the modern thinking all hadrons are composed of quarks, which
are fermions with spin 1/2 existing in 6 species called flavors. They are grouped in three
generations
u c t
     
, , .
d s b
The upper quarks in each pair have their electric charge 2/3e and the lower one −1/3e. The
quarks u,d,s are light, with masses (current) 4, 7 and 150 MeV. The rest are heavy, the
masses of c-,b- and t-quarks being 1.5, 4.5 and 180 GeV. The hadrons are built of quarks of
different flavors as composites of qqq, with half integer spin (baryons) or of q q̄, with integer
spin (mesons).
Already these first experimental conclusions gave reason to suspect that quarks possess
some additional internal quantum number. Indeed the well- known resonance state ∆++
with spin 3/2 has to be composed from three u-quarks in a symmetric state, contrary to the
Fermi-statistics. To remedy this contradiction it was assumed that quarks, apart from flavor,
possess a new quantum number, color, which allows to form a fully antisymmetric state of
quarks of the same flavor. It is clear that the number of states of different color has to be
equal or larger than 3. All experimental consequences of this conjecture are consistent with
the assumption that this number is exactly three (”red, yellow and blue”). The dynamic
theory based on this assumption is naturally called Quantum Chromodynamics.
Thus the quark of any flavor exists in 3 different color states: qf,α , f = u, d, s, c, t, b,
α = 1, 2, 3. It is assumed that color itself does not show in strong interactions. In other word,
the theory is invariant under unitary transformation of quark wave function with different
colors. The group of such transformation with determinant equal to unity is SUC (3). It
does not influence flavor at all and is supposed to be an exact symmetry of the world. In the
following we shall meet no other group SU (3) and the subindex C (”color”) will be frequently
omitted.
In the following we shall use some results about the groups SU (N ) necessary for practical
calculations. Some basic properties and formulas about SU (N ) are briefly derived and can
be found in the book ”Teoriya kalibrovochnykh vzaimodeistdviy elementarnykh chastitz” by
M.B.Voloshin and K.A.Ter-Martirosyan (in Russian), Moskva, Energoizdat, 1984
The basic SU (3) spinor qα transforms under infinitesimal transformations of the group as

δq = ita ξ a q ≡ i(tξ)q, a = 1, 2, ...8. (1)

Here ξ a are 8 small (real) parameters of the transformation and ta are 8 generators of the
SU (3) in the spinor representation. They obey the standard commutation rules

[ta , tb ] = if abc tc a, b, c = 1, 2, ...8, (2)

where f abc are fully antisymmetric structure constants of the SU (3). Generators ta are 3 × 3
matrices normalized according to
1
Tr{ta tb } = δab , a = 1, 2, ...8. (3)
2
 5

We also define a scalar product in the color space as AB = Aa B a . Invariance of the theory
under the transformation of the SU (3) group means conservation of 8 total colors of the
world Qa which obey the same commutation relations

[Qa , Qb ] = if abc Qc a, b, c = 1, 2, ...8 (4)

and commute with the total Hamiltonian H

[Qa , H] = 0. (5)

All hadrons are supposed to be singlets under SU (3) having Qa = 0. They can be formed
from quarks in two different ways

M = q̄α qα , or B = ǫαβγ qα qβ qγ ,

where q are quarks of any (possibly different) flavors and ǫαβγ is a numerical completely
antisymmetric tensor. The first structure gives baryons, the second gives mesons.
To see that B is invariant introduce finite SU (3) transformations in the fundamental
representation
U (ξ) = eitξ (6)

The transformed B is then

B̄ = Uαα′ Uββ ′ Uγγ ′ ǫα,β,γ qα′ rβ ′ sγ ′ = ǭα′ β ′ γ ′ qα′ rβ ′ sγ ′ ,

where
ǭα′ β ′ γ ′ = Uαα′ Uββ ′ Uγγ ′ ǫαβγ .

One immediately observes that ǭ is totally antisymmetric:

ǭβ ′ α′ γ ′ = Uαβ ′ Uβα′ Uγγ ′ ǫαβγ .

Interchanging α ↔ β and using antisymmetry of ǫαβγ we find the desired property

ǭβ ′ α′ γ ′ = −ǭα′ β ′ γ ′ .

Now we consider the only nontrivial element

ǭ123 = Uα1 Uβ2 Uγ3 ǫαβγ = det U = 1.

So ǭ = ǫ and B is invariant.
The Quantum Chromodynamics (QCD) emerges when one imposes a condition that color
is conserved not only globally but also locally. This implies that invariance under the SU (3)
group is to fulfill also for transformations with parameters ξ a (x) depending on the space-time
point. As a result one obtains a non-Abelian gauge theory with 8 gauge bosons called gluons.
Let us follow the standard path for constructing this sort of theory. The transformation of
the basic quark spinor preserves its form under local transformations:

δq(x) = i(tξ(x))q(x), a = 1, 2, ...8. (7)

 6

However the derivative ∂q(x) acquires an additional term:

δ∂q(x) = i(tξ(x))∂q(x) + i(t∂ξ(x)) · q(x). (8)

To compensate one introduces 8 vector fields Gaµ and defines a generalized derivative

D = ∂ − ig(tG). (9)

One chooses the form of the transformation of G in accordance with its global property as a
color vector and also to obtain simple transformation properties of Dq(x)
1 1
δGa (x) = −f abc ξ b (x)Gc (x) + ∂ξ a (x) ≡ Dab ξ b (x). (10)
g g
Here, similar to (9),
Dab = δab ∂ − f abc Gc = δab ∂ − ig(T G) (11)

is the generalized derivative in the self-adjoint representation with the generators

a
Tbc = −if abc . (12)

Note that under global transformation with ∂ξ = 0 Ga transforms as a standard SU3 vector
V a:
δV a = −f abc ξ b V c = i(T b )ac ξ b V c .. (13)

It is trivial to demonstrate that Dq(x) transforms in the same way as q(x), that is does
not feel the x-dependence of ξ(x)

δDq(x) = i(tξ(x))Dq(x), a = 1, 2, ...8. (14)

Indeed  
δDq = δ ∂q − ig(tG)q

= i(tξ)∂q + i(t∂ξ) · q − ig(tG)i(tξ)q − igta f abc ξ b Gc q − i(t∂ξ)q.

The two terms with ∂ξ obviously cancel and the rest give

g(tG)(tξ)q + igf abc ta ξ b Gc q + i(tξ)∂q. (15)

We change the order of t’s in the first term

(tG)(tξ) = (tξ)(tG) + [(tG), (tξ)] = (tξ)(tG) + if cba Gc ξ b ta .

The last term here cancels the second term in (15) so we find

i(tξ)∂ + (tξ)(tG) = i(tξ)Dq,

which gives (14). As a result we find that

the product q̄Dq is invariant under the local transformations of the color group.

We have only to find an appropriate Lagrangian density for the gluons themselves. To
construct invariant quantities we first find quantities which transform in the same way under
 7

both global and local transformations, that is do not feel the dependence of the transformation
parameter ξ(x) on x.
A general SU (3) vector V a transforms according to (13). Introducing a finite transfor-
mation we have under transformation
a ξa
V → UA (ξ)V, where UA (ξ) = eiT ,

where subindex A means adjoint representation. So obviously the scalar product of two
SU (3) vectors is invariant  
δ(V1 V2 ) = δ V1a V2a = 0. (16)

Slightly more complicated is to demonstrate that the vector product of two SU (3) vectors
W a = f abc V1b V2c is also an SU (3) vector. We have

δW a = iTbc
a b d e c
(iTde ξ V1 V2 + iV1b Tde
c d e
ξ V2 ).

We change e ↔ b in the first term and e ↔ c in the second to obtain

 
δW a = −ξ d V1b V2c Tec
a e e a
Tdb + Tdc Tbc .

Now we use the symmetry

a b c c
Tbc = Tca = Tab = −Tba

to find  
a e d a
Tec Tdb = Tbe Tec = T d T a ,
bc
 
e a d a a d
Tdc Tbe = Tce Tbe = −Tbe Tec = − T a T d bc.

So we have
 
δW a = −ξ d V1b V2c T d T a − T a T d = −iξ d V1b V2c f dae Tbc
e
= i(T d )ab ξ d Wb . (17)
bc

So indeed Wa is an SU (3) vector. It is essential that these properties hold whether ξ depends
on x or not, since they do not involve its derivatives in x.
Using these properties we shall first construct fields which transform in a simple manner
under the global color group but aquire some additional terms if the transformation becomes
local.
An immediate candidate is
a
Xµν = ∂µ Gaν − ∂ν Gaµ . (18)

which is an antisymmetric tensor in the coordinate space and vector in the color space. Its
transformation under the local color transformation contains additional terms with derivatives
of ξ(x):
a 1 1
δXµν = f abc ∂µ ξ b · Gcν − f abc ∂ν ξ b · Gcµ + f abc ∂µ ∂ν ξ a − f abc ∂ν ∂µ ξ a . (19)
g g
The last two terms with the second derivative cancel, but terms with the first derivative
remain. Let us compare this with the transformation of another global color vector
a
Yµν = f abc Gbµ Gcν , (20)
 8

Figure 1: Basic QCD interactions

which is also an antisymmetric tensor in the coordinate space. Again leaving only terms with
derivatives of ξ(x) we find

a 1 abc 1
δYµν = f ∂µ ξ b Gcν + f abc Gµ ∂ν ξ c . (21)
g g
Comparing with (19) we see that in the difference

Gaµν = Xµν
a a
− gYµν = ∂µ Gaν − ∂ν Gaµ − gf abc Gbµ Gcν (22)

all terms with derivatives of ξ cancel, so that Gaµν transforms under local transformations
in exactly the same way as under global transformations, that is as a color vector. As a
result the product Gaµν Gµν
a is invariant under gauge transformations (and under Lorentz
transformations) and can serve to construct the QCD Lagrangian density.
Together with the quark part, the Lagrangian density for QCD is constructed as
1
L = − Gaµν Gµν
X
a + q̄f (iD̂ + mf )qf , (23)
4 f

where the sum goes over flavors.

From the start it is clear that flavors do not change in the strong interaction. Gluons are
massless particles. Quarks emit and absorb gluons very much in the same way as charged
particles emit and absorb photons in the Quantum Electrodynamics. However in contrast to
the latter, gluons also interact between themselves, as seen from the structure of the square of
Gaµν : there appear triple and quartic interactions of gluons, the first containing a derivative,
see Fig. 1
Expression (23) is the Lagrangian of the classical QCD. To quantize it the easiest way is
to use the generating functional of the Green functions in terms of the functional integral
Z
Z(ji ) = DGDqf Dq̄f eiS(G,qf ,q̄f )+iSE (G,qf ,q̄f ,ji ) . (24)

Here the classical action S is Z

S(G, qf , q̄f ) = d4 xL (25)

and the external action, depending on external sources ji is

Z
SE (G, qf , q̄f ) = d4 xLE . (26)

Let us recall that if

LE (x) = O(x)j(x) + ...,
 9

where O(x) is some operator made of fields G , q , q̄, then the Green functions are obtained as
1 δn Z(j)
=< Ψ0 |T {O(x1 ), ...O(xn )}|Ψ0 >
Z(0) δj(x1 )...δj(xn )
with Ψ0 the vacuum state. Note that Z(j) is determined up to an arbitrary constant inde-
pendent of the fields.
Both actions, S and SE are invariant under local transformation of the SU (3)c group.
Let us denote the fields generically as u = {G, qf , q̄f } and as uξ the fields gauge transformed
with parameters ξ a (x) (not necessarily small). Then we have

S(uξ ) = S(u), SE (uξ , ji ) = SE (u, ji ). (27)

The latter condition reflects the requirement that the physical significance have only gauge
invariant Green functions. So LE should depend only on color invariant combinations of the
field.
The invariance of the action under transformations with 8 arbitrary functions ξ a (x) means
that the integrand in (24) in fact does not depend on 8 functions in the functional space and,
as a consequence, the functional integral involves an infinite volume of integration over these
functions as a factor and so is badly defined. To define it one has to restrict integrations
prohibiting those which involve gauge transformations. Following Faddeev-Popov (FP)let us
do it by requiring the fields to obey 8 conditions fixing the gauge:

F a (u(x)) = C a (x), a = 1, ...8 (28)

with the requirement that the equation for ξ a (x)

F a (uξ (x)) = C a (x) (29)

has ξ( x) = 0 as the only solution. The functional integral over ξ a (x), as a functional of the
fields u
δF a (x)  −1
Z Y   
Dξ δ F a (uξ (x) − C a (x) = det b (x ξ=0
≡ ∆−1 (u) (30)
δξ

x
defines the so-called FP functional determinant. Obviously it is gauge invariant

∆(uξ ) = ∆(u). (31)

Using (30) we introduce this determinant into the functional integral (24):
Z Z Y  
Z(j, {C})) = DueiS(u)+iSE (u,j) ∆(u) Dξ δ F a (uξ (x) − C a (x) . (32)
x

Now we pass to integration over the fields uξ . Using gauge invariance of both the action and
the FP determinant we find
Z Y   Z
iS(u)+iSE (u,j) a a
Z(j, {C}) = Due ∆(u) δ F (u(x) − C (x) · Dξ. (33)
x
R
The gauge volume Dξ is separated as a constant factor and can be dropped. Obviously the
remaining functional integral is independent of the choice of function C a (x):

Z(j) = Z(j, {C}).

 10

So we may integrate over C a (x) with an arbitrary weight factor

Z
exp iS1 (C) = exp i d4 xL(C a (x))

This gives Z Z
Z(j) = DCZ(j, {C}) exp i d4 xL(C a (x))

and we obtain Z
Z(j) = DueiS(u)+iS1 (F (u))+iSE (u,j) ∆(u). (34)

Finally we use the well known result that integration over anticommuting scalar fields
χa (x)and χ̄a (x):
Z R
d4 xd4 y χ̄a (x)M ab (x,y)χb (y)
DχDχ̄ei = const · det M ab (x, y). (35)

This allows to substitute the FP determinant by integration over auxiliary anticommuting

scalar fields χa (x) and χ̄a (x) called ghosts:
Z
Z(j) = DuDχDχ̄eiS(u)+iS1 (F (u))+iS2 (u,χ,χ̄)+iSE (u,j) , (36)

where
δF a (x) b
Z
S2 = d4 xL2 , L2 = χ̄a (x) χ (x). (37)
δξ b (x)
This expression gives the final generating functional for the QCD.
The concrete choice of function F , the choice of gauge, may be made differently. The
standard relativistic invariant choice (’Lorentz gauge’) is

F a = ∂Ga . (38)
√
As a result δF a = (1/g)∂Dab ξ b , so that M ab = (1/g)∂Dab Standardly one rescales χ → gχ
√
χ̄ → gχ̄ to find
L2 = χ̄∂Dχ = χ̄∂ 2 χ − gf abc χ̄a ∂(Gc χb ). (39)

From this one concludes that 8 ghosts are massless as the gluons and interact with the latter
emitting and absorbing gluons with a derivative interaction (see Fig. 2). As to S1 the
standard choice leading to a renormalizable theory is
1 a a 1
Z
S1 = d4 xL1 , L1 = − F F = − ∂Ga ∂Ga , (40)
2α 2α
where α is a numerical parameter.
 11

Figure 2: Ghost interaction

µ ν

Figure 3: Propagators QCD

4 Feynman rules. General properties and problems

The Feynman rules for perturbation theory are readily read from different parts of the QCD
Lagrangian. The relevant propagators and vertices are illustrated in Figs. 3 and ??/
The propagators for quarks and ghosts are simple and not different from any quantum
field theory.
The propagator of the quark of flavor f and momentum k is
−i
Dq (k) = . (41)
mf − k̂

The propagator of the ghost of color a and momentum k is

ab iδab
Dgh = . (42)
k2
To find the propagator of the gluon Dg we have to study the part of the Lagrangian
quadratic in the gluon field
1 a 2 µν 1 a µ ν a 1  
Gµ (∂ g − ∂ µ ∂ ν )Gaν + Gµ ∂ ∂ Gν = Gaµ ∂ 2 gµν − (1 − 1/α)∂ µ ∂ ν Gaν . (43)
2 2α 2
By the definition of the propagator it is just
1 a  −1 µν a
G iDg Gν .
2 µ
Passing to the momentum space we find that the inverse propagator of the gluon with mo-
mentum k is  µν  1 µ ν
iDg−1 (k) = −k2 gµν + 1 − k k . (44)
α
 12

Figure 4: Interaction vertices QCD

To invert the matrix of the inverse propagator we introduce projectors onto the spaces parallel
and orthogonal to vector k:
kµ kν kµ kν
P⊥µν = gµν − 2
, Pkµν = 2 . (45)
k k
They are orthogonal and squared give the projectors themselves. We have found
1 2
iDg−1 (k) = −k2 P⊥ − k PL , (46)
α
which gives for the matrix of the propagator
1 α 
Dg (k) = i P +
2 ⊥
PL . (47)
k k2
Explicitly with color indices

ab
ng
µν kµ kν o
Dµν (k) = iδab − (1 − α) . (48)
k2 k4
The simplest choice is α = 1 which corresponds to the frequently used Feynman gauge:
ab gµν
Dµν (k) = −iδab . (49)
k2
From the theoretical point of view the transverse gauge α = 0 has some advantages
ab ⊥
Dµν (k) = −iδab Pµν (k). (50)

The interaction vertices are read from the corresponding non-linear terms in the La-
grangian Fig. 4
The term q̄f iD̂qf generates the vertex Vqg for the interaction of the quarks with the gluon
of color a and polarization µ
Vµa = gγµ ta . (51)
 13

Figure 5: A contribution to the gluon self-mass

The term −gf abc χ̄∂ µ (χb Gcµ ) gives rise to the vertex for the interaction of the gluon of color
b with the incoming ghost of color c and outgoing ghost of color a and momentum p

Vµabc = igf abc pµ . (52)

Note that this vertex is not symmetric in the incoming and outgoing ghosts. The triple gluon
interaction comes from the interference of linear and quadratic terms in the product GG. As-
suming all the three gluons outgoing, with momenta, color and polarizations (q, b, µ),(p, a, λ)
and (r, c, ν) we find a symmetric triple gluon vertex
abc
Vνλµ = igf abc [(p − q)ν gνλ + (q − r)λ gµν + (r − p)µ gνλ ]. (53)

There is finally the quartic gluon interaction. It does not depend on the gluon momenta. If
colors and polarizations are (a, λ), (b, µ), (c, ν) and (d, σ) then the quartic vertex is
abcd
Vλµνσ = −ig2 [f abe f cde (gλν gµσ − gλσ gµν )

+f acef bde (gλµ gνσ − gλσ gµν ) + f ade f cbe (gλν gµσ − gλµ gσν )]. (54)
Inspection of these rules allows to make first conclusions about the sort of quantum field
theory which is the QCD. By construction (gauge fixing) the theory is now not invariant
under local gauge transformations but preserves the invariance under global transformations
of the SU (3) group. However there remains a strong consequence of the initial invariance
under local gauge transformations: all couplings are fixed by it. All quarks of any flavor and
the triple gluon interaction have the same coupling constant g and the quartic gluon coupling
has the strength g2 . Unlike photons, gluons interact with each other. Also new auxiliary
scalar particles obeying the Fermi statistics, ghosts, appear in the theory and interact with
gluons.
As a result of the presence of self-interacting zero-mass particles the theory possesses
strong infrared divergences, which reveal themselves when one tries to go to the mass-shell
for gluons and ghosts. It is important that the off-mass-shell Green function are free from
infrared divergence. However when one tries to set external masses of gluons and ghosts to
zero, the corresponding quantities diverge. In fact consider the integral which appears in the
gluon self-mass (see Fig. 5)
d4 q 1 d4 q 1 d4 q
Z Z Z Z Z
= dx = dx .
q 2 (q − k)2 0 [x(q − k)2 + (1 − x)q 2 ]2 0 [(q 2 + x(1 − x)k2 ]2
 14

This integral is logarithmically divergent in the untraviolet at any value of k2 but this di-
vergence is eliminated by the mass renormalization. However at k2 = 0 there additionally
appears an infrared divergence, so that even renormalized self-mass does not exist on the
mass shell. The behavior of full Green function for gluons and ghosts around their mass shell
at k2 = 0 is in fact unknown. It is generally postulated that it is such that the massless gluon
and ghosts are excluded from the physical spectrum and so cannot be observed in nature.
This property is assumed to be also true for any objects with a non-trivial color and is called
’confinement of color’. Colored particles can only appear as intermediate virtual particles
and so are confined inside the observed hadrons. The property of confinement has not been
proved so far but is supported by the experimental facts.
From the point of view of ultraviolet divergencies, the QCD belongs to the class of renor-
malizable quantum field theories. It possesses a finite number of skeleton divergent diagrams,
of the same type as in the Quantum Electrodynamics. These are the self-masses of gluons
and ghosts with the order of divergence ω = 2, the quark self-mass with ω = 1, vertex parts
for the triple and quartic interaction of gluons and gluons with ghosts with ω = 1 and for the
interaction of quarks with gluons with ω = 0. They all are eliminated by renormalization.
However this procedure in the QCD has certain problems.
Technical problems arise, first, from the impossibility to use the gluon and ghost mass
shells as the subtraction points in the renormalization procedure, since the self-masses and
vertex parts do not exist there. This obstacle can be easily overcome by noting that in fact
any value of k2 can be used for renormalization as well. E.g. one may fix the renormalized
Green function by the condition that

k2 GR (k2 ) = 1, at k2 = −µ2 (55)

with µ2 is a suitably chosen renormalization point, and similarly for vertex parts. The second
problem is related to the necessity to preserve the consequences of gauge invariance at the
intermediate stage when the theory is regularized at large momenta. The best known way
(which also solves the first problem) is to use the so-called dimensional regularization, when
the momentum integrals are considered to be in the dimension d = 4 − ǫ. The ultraviolet
divergence then are revealed as pole terms at ǫ = 0. The renormalized quantities then can
be defined as the result of dropping all pole terms.
However there is a more profound problem related to the preservation of gauge invariance
in renormalization. Roughly speaking we have to guarantee that all the renormalized coupling
constants be related among themselves in the same manner as in the unrenormalized theory
(in particular, be equal for triple vertices). This is not trivial, since we encounter several
different vertex parts and it is not at all clear that one can eliminate ultraviolet divergence in
all of them renormalizing only one common coupling constant. As we shall see this is indeed
possible, since the divergent terms in different vertex parts are related to each other by gauge
invariance. This point will be studied in the following sections.
 15

5 BRST invariance
We recall the generating functional for the QCD:
Z
tot
Z(j) = DuDχDχ̄eiS , (56)
Z
S tot = S + S1 + S2 + SE = d4 x(L + L1 + L2 + LE ), (57)

where
1
L = − Gaµν Gµν
X
a + q̄f (iD̂ + mf )q, (58)
4 f

1 a a δF a
L1 = − F F , L2 = gχ̄a b χb . (59)
2α δξ ξ=0
Action S tot is no more gauge invariant due to the fixation of gauge. However it turns out
that it possesses an invariance under gauge transformations of a very special sort, the BRST
(Becci-Ruet-Stora-Tyutin) transformations, which also include the ghosts. This invariance is
more conveniently expressed if we somewhat transform this action to linearize it in F a . We
present
1 a a
Z  Z
iS1 4
e = exp i d x − F F = DheiS̃1 (60)
2α
with
1
Z
S̃1 = d4 xL̃1 , L̃1 = αh2a + ha Fa . (61)
2
Then the generating functional is
Z
tot
Z(j) = DuDχDχ̄DheiS̃ , (62)

where S̃ tot differs from S tot by the substitution of L1 by L̃1 .

For the physical fields G and q the BRST transformation is just the usual gauge trans-
formation with a gauge function
ξ = gχǫ, (63)

where ǫ is an anticommuting number, which anticommutes with itself and all anticommuting
fields, qf , q̄f , χ and χ̄. Naturally the product χǫ commutes with all fields as any gauge
function should do. So under the BRST transformation:

δG = Dχǫ, δq = ig(tχ)ǫq = −ig(tχ)qǫ. (64)

Additionally one postulates transformation for the ghost fields and h:

1 abc b c
δh = 0, δχ̄ = hǫ, δχa = gf χ χ ǫ. (65)
2

Before we prove that action S̃ tot is invariant under these transformations we demonstrate
that they are nilpotent. Let for any field u, by definition

δu = suǫ, (66)
 16

where s plays the role of the generator of the BRST transformations. Then we demonstrate
that
s2 = 0 (67)

by studying
δ(su) = s2 uǫ. (68)

We shall prove this successively for all fields u = {q, G, χ, χ̄, h}.
1. Quarks
We have
sq = −ig(tχ)q,

so
δ(sq) = −ig(tδχ)q − ig(tχ)δq =
1
−i g2 ta f abc χb χc ǫq − ig(tχ)ig(tχ)ǫq
2
 1 
= g2 ta tb χa χb − i ta f abc χb χc ǫq. (69)
2
However
ta tb χa χb == −ta tb χb χa = −tb ta χa χb
1 1 1
= [ta , tb ]χa χb = i f abc tc χa χb = i f abc ta χb χc .
2 2 2
So that the two terms in (69) cancel and we obtain s2 q = 0
2. Gluons
Here
sGa = ∂χa − igTabc
Gc χb .

We find
1  1 
δ(sGa ) = ∂ gf abc χb χc ǫ − igTab
c
Dcd χd ǫ · χb − igTab
c
Gc gf bde χd χe ǫ = (1) + (2) + (3).
2 2
Let us start with the second term

c e e d
(2) = −igTab (∂δcd − igTcd G )χ · ǫχb

c
= −igTab ∂χc · ǫχb − g2 Tab
c e e d b
Tcd G χ ǫχ = (21) + (22).

Consider terms which do not contain G. They are

(1) + (21) = gf abc ∂χb · χc ǫ − igTab

c
∂χc · ǫχb

= gf abc ∂χb · χc ǫ + gf abc ∂χc · χb ǫ = 0.

Now terms linear in G:

(22) = g2 Tbc
a e e d b
Tcd G χ χ ǫ = −g2 Tbc
a e e b d
Tcd G χ χ ǫ = −g2 Tdc
a e e d b
Tcb G χ χ ǫ
1 1
= −g2 Tbc
e a e d b
Tcd G χ χ ǫ = g2 [T a , T e ]bd Ge χd χb ǫ = ig2 f aec Tbd
c e d b
G χ χ ǫ.
2 2
 17

So we find
1 1
(22) + (3) = ig2 f aec Tbd
c e d b
G χ χ ǫ − ig2 Tab
c c bde d e
G f χ χ ǫ.
2 2
In the second term we change c → e,e → b and b → c to find
1 1
(22) + (3) = if aec Tbd
c e d b
G χ χ ǫ − ig2 Tac
e e cdb d b
G f χ χ ǫ.
2 2
The common factor
(1/2)g2 Ge χd χb ǫ

is multiplied by
f aec Tbd
c
− f cdb Tac
e
= −if aec f cbd + if cdb f eac = 0.

So indeed s2 Ga = 0.
3. Ghosts
By definition
sχa = (1/2)gf abc χb χc ,

so that 1  1
δ(sχa ) = gf abc
gf bde χd χe ǫ χc = g2 f abc f cde χb χd χe ǫ, (70)
2 2
(we interchanged b and c). We rewrite (70) as follows
1 1
− g2 Tbc
a e b d e
Tcd χ χ χ ǫ = − g2 Tdc
e a b d e
Tcb χ χ χ ǫ
2 2
1 1 1
= + g2 Tbc
e a b d e
Tcd χ χ χ ǫ = g2 [T e , T a ]bd χb χd χe ǫ = i g2 Tbdc cea b d e
f χχ χ ǫ
2 4 4
1 1 1
= f bdc f ceaχb χd χe ǫ = f edc f cba χe χd χb ǫ = − f abc f cde χb χd χe ǫ.
4 4 4
Comparison with (70) shows that this is in fact equal to zero. So we have proved that s2 χ = 0.
4. Antighosts
This is trivial:
sχ̄ = h, so that δ(sχ̄) = δh = 0.

5. Field h
This is still more trivial:
sh = 0, so that δsh = 0.

.
So we have proved that the BRST transformations are nilpotent.

**************************************

Now we note that

δ(χ̄F ) = δχ̄ · F + χ̄δF. (71)

However
δF a b
δF a = gχ ǫ,
δξ b
 18

so that
δF a b
δ(χ̄F ) = hǫF + gχa χǫ
δξ b
and
δF a b
s(χ̄F ) = hF + gχa χ. (72)
δξ b
We also have
δ(χ̄h) = hǫh = h2 ǫ, s(χ̄h) = h2 . (73)

Since
1 δF a
L̃1 + L2 = αh2a + ha Fa + gχ̄a b χb ,
2 δξ ξ=0
we obtain  1 
L̃1 + L2 = s χ̄F + αχ̄h , (74)
2
so that  1 
L̃tot = L + LE + s χ̄F + αχ̄h . (75)
2
Since L + LE are gauge invariant we have

s(L + LE ) = 0.

Nilpotency of s allows to prove that

 1 
sL̃tot = s2 χ̄F + αχ̄h = 0. (76)
2
Thus we have demonstrated that the theory is invariant under the BRST transformations.
Note that we have presented the generating functional in the form
Z
Z(j) = DuDχDχ̄DheiS+iSE +isΩ , (77)

where S is the classical action and

1
Ω = χ̄F + αχ̄h (78)
2
In fact sΩ is the action which is to be added to quantize the theory. Obviously the generating
functional (77) is BRST invariant with an arbitrary choice of Ω. This opens a possibility for
the so-called BRST quantization. Define for any field the BRST generator ŝ by the rule

[ŝ, u]± = isu (79)

where ”+” stands for bosonic fields and ”-” stands for fermionic fields and demand that all
physical states satisfy the BRST invariance condition

ŝΨ = 0. (80)

Obviously the theory then does not depend on the choice of Ω and is fully determined by the
classical action. Indeed the matrix elements of the additional term vanish on the physical
states:
hΨ1 |sΩ|Ψ2 i = hΨ1 |[ŝ, Ω]|Ψ1 i = 0. (81)
 19

Note finally that integration over the field h makes the substitution
1
h → − F, (82)
α
which transforms
1 2 1
h + hF → − F 2 .
2 2α
From this it follows that the theory integrated over h will be BRST invariant with the rule
of transformation of χ̄
1
δχ̄ = − F ǫ. (83)
α
 20

6 Renormalization
6.1 Formal procedure and problems
Both the fields and the parameters which appear in the generating functional are not the
observable ones but ”bare”. The physical, renormalized fields are related to them by renor-
malization. In absence of ultraviolet divergencies one can manipulate with bare quantities and
afterwards determine the physical quantities considering concrete reactions and comparing
the resulting predictions with experimental data. Divergencies do not allow this procedure,
since the bare fields and parameters turn out to be divergent. Renormalization consists in
expressing the theory directly in terms of renormalized quantities and avoid using the bare
ones altogether.
The first step in renormalization procedure is to pass to renormalized fields, which are
related to the bare ones multiplicatively. For any field u one defines the renormalized field
by
u = Z 1/2 uR , u = G, q, q̄, χ (84)

(by tradition the numerical factor ”the renormalization constant” is called Z 1/2 ). In the
future we shall have to study the Green function of our basic fields u (not gauge invariant)
for which purpose one chooses
X
LE = uju , u = G, q, q̄, χ, χ̄. (85)
u

Renormalization of the fields is to be accompanied by the inverse renormalization of the

external currents
juR = Zu1/2 ju so that LE =
X
uR jR . (86)
u

It has also be accompanied by the renormalization of the gauge parameter α to preserve the
choice of gauge
α = ZG αR . (87)

In terms of renormalized field the QCD Lagrangian takes the form

Ltot = L0 + LI + LE (88)

where L0 is a part quadratic in fields

1 1
L0 = − ZG (∂µ GRν − ∂ν GRµ )2 − (∂GR )2 + Zqf q̄f R (i∂ˆ + mf )qf R + Zχ χ̄R ∂ 2 χR , (89)
X
4 2αR f

LI is the interaction part

1 3/2 1 2 2 abc
LI = − ZG gf abc (∂µ GRµ − ∂ν GRµ )GµR GνR − ZG g [f GRµ GRν ]2
2 4
1/2 1/2
Zqf ZG gq̄f R (tĜR )qf R − Zχ ZG f abc χ̄aR ∂(GbR χcR )
X
+ (90)
f

and LE is the external action given by (86).

 21

Next step is renormalization of the quark masses. One puts for each flavor

Zq m = mR + δm. (91)

The (infinite) quantity δm is the mass renormalization constant.

Finally one has to renormalize the coupling. In contrast to the QED this can be made
in different ways, since the coupling g appears in different places. Typically one takes the
coefficient in a given interaction and defines it as the renormalized coupling gR times an
infinite vertex renormalization constant Z1 (traditional notation). E.g. using the triple gluon
interaction one can define
3/2
gZG = gR Z1G . (92)

Alternatively one can use the quark-gluon interaction defining

1/2
gZq ZG = gR Z1q (93)

or the ghost-gluon interaction

1/2
gZχ ZG = gR Z1 χ. (94)

Obviously with the fixed and unique renormalized coupling constant gR different Z1 are not
independent but related to each other as

−1
Z1G ZG = Z1q Zq−1 = Z1χ Zχ−1 . (95)

**************************************

With any definition one the splits

Zu = 1 + (Zu − 1) and Z1 = 1 + (Z1 − 1)

and assumes that quantities

δm, Zu − 1 and Z1 − 1

are of the order in gR higher than the first. Then in perturbation theory in higher orders one
finds counterterms proportional to these quantities which may be chosen to cancel ultraviolet
divergencies found in this order.
One immediately sees a problem mentioned before. The theory has divergencies coming
from each type of interaction, that is, in each vertex part in the perturbation theory. To can-
cel all these divergencies one has only one independent renormalization constant Z1 defined
by a particular vertex part. However different vertex parts will have different divergent con-
tributions which depend on the way one regularizes the theory and defines their finite parts.
It is not clear how one can eliminate all these divergencies introducing a single counterterm
and moreover guarantee relations (95). It is clear that if one introduces indiscriminately
separate counterterms for each vertex choosing them from only the condition to eliminate
the divergence then the relation (95) will be generally violated. As a result the renormal-
ized coupling constants for different vertexes will be different and gauge invariance will be
blatantly broken.
 22

We are going to demonstrate that there exists such a choice of the renormalization pro-
cedure and subtraction points which, on the one hand, guarantees fulfillment of relations
(95) and, on the other, eliminates all ultraviolet divergencies. We shall choose the vertex
renormalization constants to satisfy (95) from the start and show that we can remove all the
ultraviolet divergencies in all vertex parts. This is possible only because there exist certain
relations between the divergent terms in different vertex parts, so that in fact there is only
one independent divergence which can be eliminated by a single renormalization constant
Z1 . This is a consequence of a general relation between different Green function in the the-
ory, which generalize the well-known Ward identities in the QED, and bear the name of
Slavnov-Tailor identities.
 23

6.2 Slavnov-Tailor identities

To derive the Slavnov-Tailor identities we make use of the BRST invariance. Consider the
generating functional with the external action linear in fields:
Z
Z(j) = DuDχdχ̄eiS+iS1 +iS2 +iSE , (96)

where
LE = jG + η̄q + q̄η + χ̄ζ + ζ̄χ (97)

and
j, η̄, η, ζ̄, ζ

is a set of external sources which generate Green function for the fields

G, q, q̄, χ, χ̄

respectively. Note that this external action and with it the whole generating functional is not
gauge and BRST invariant. Now we BRST transform the fields over which the integration is
done in Z(j). Its Jacobian is unity, since all transformation except that of χ are shifts of the
variables. The matrix of the transformation of the field χ is obviously

dχ′ a
= δab − gf acb χc ǫ, (98)
dχb
which corresponds to the unitary rotation in the color space with angle χc ǫ whose determinant
is unity. After the transformation actions S, S1 and S2 do not change and the only part which
changes is SE . So we find an identity
Z  
DuDχDχ̄eiS+iS1 +iS2 +iSE eiδSE − 1 = 0. (99)

For brevity we denote the functional averaging with weight exp(iS + iS1 + iS2 + iSE ) by the
symbol < ... >: Z
< Φ(u) >≡ DuDχDχ̄Φ(u)eiS+iS1 +iS2 +iSE . (100)

We also omit the sign of integration over the coordinate space except in case when the
coordinates are explicitly indicated:
1 1
Z
δSE = d4 xδLE = jDχǫ + igη̄(tχ)ǫq − igq̄(tχ)ǫη − ∂Gǫζ + gζ̄ a f abc χb χc ǫ. (101)
α 2
From (99) we find
1 1
< jDχ − igη̄(tχ)q + igq̄(tχ)η + ∂Gζ + gζ̄ a f abc χb χc >= 0. (102)
α 2
Identity (102) is true for any values of external sources. We differentiate over ζ(x) and put
ζ = ζ̄ = 0 afterwards to obtain
  1
< iχ̄(x) jDχ − igη̄(tχ)q + igq̄(tχ)η + ∂G(x) >= 0. (103)
α
This is the generating functional (unrenormalized) for the Slavnov-Tailor identities.
 24

To renormalize it we have to take into account some important immediate consequences.

Differentiating over j(y) and putting then all sources to zero we find
1
< iG(y) ∂G(x) + iχ̄(x)Dχ(y) >= 0. (104)
α
Differentiating in y:
1
< ∂G(y)∂G(x) >= − < χ̄(x)∂Dχ(y) >= iδ4 (x − y). (105)
α
The latter equality follows from the fact that all dependence on χ̄ and χ in the Lagrangian
has the form
χ̄(x)∂Dχ(x),

so that taking as variables χ̄ and χ′ = ∂Dχ we shall have an integral

Z R
d4 xχ̄χ′
Dχ′ Dχ̄χ′ (y)χ̄(x)ei = −iδ4 (x − y).

In the momentum space we find for the full gluon Green function ∆tot
µν

1
− kµ kν ∆tot
µν = −i.
α
So that presenting the Green function as a sum of its transverse and longitudinal parts

∆tot (k) = ∆tot 2 tot 2

⊥ (k )P⊥ (k) + ∆L k )PL (k), (106)

we find
1 2 tot α
k ∆L = i, or ∆tot L = i 2. (107)
α k
Compare this with the gluon propagator, which is the gluon Green function wthout interac-
tion. 1 α 
Dg (k) = ∆(0) (k) = i 2 P⊥ + 2 PL .
k k
We observe that the the longitudinal part of the gluon Green function is not changed by the
interaction.
From this, as in QED, it follows that the gluon polarization operator Π(k) defined by the
matrix Dyson equation
∆tot (k) = ∆(0) (k) + ∆(0) (k)Π∆tot (k) (108)

is purely transverse. Splitting into the transverse and longitudinal parts we have

∆tot (k) = ∆tot 2 tot 2

⊥ (k )P⊥ (k) + ∆L (k )PL (k)

with the solution

i
∆tot 2
⊥ (k ) =
k2 − Π⊥ (k2 )
and
i
∆tot 2
L (k ) = ,
k2 /α − ΠL (k2 )
where
Πµν (k) = Π⊥ (k2 )P⊥µν + ΠL (k2 )PLµν . (109)
 25

∆ Π ∆

Figure 6: The Dyson matrix equation for the gluon Green function. Blob Π is the polarization
operator

Since the longitudinal part of ∆tot is the same as without interaction, we find

ΠL (k2 ) = 0

and
Πµν (k) = Π⊥ (k2 )P⊥µν . (110)

According to (109) the polarization operator as a whole does has no pole at k2 = 0.

However P⊥µν (k) does have a pole at k2 = 0. So the transverse polarization operator has to
have a zero at k2 = 0
Π⊥ (k2 )k2 =0 = 0. (111)

As a result, in absence of infrared divergencies the gluon mass remains equal to zero after
introduction of interaction since from the Dyson equation
i
∆tot 2
⊥ (k ) = (112)
k2 − Π⊥ (k2 )

and has a zero at k2 = 0.

We observe that the longitudinal part of the gluon field in fact does not take part in the
interaction and is in a sense trivial. Accordingly we split the external gluon current into the
transverse and longitudinal parts.
µ
j = j⊥ + jL , j⊥ = P⊥µν jν , jLµ = PLµν jν . (113)

Of course in the coordinate space both projectors become non-local. E.g.

PLµν = ∂ µ ∂ ν ∂ −2 , (114)

where ∂ −2 is an integral operator with the kernel ∆0 (x − y) which is the propagator of the
massless scalar field. So we find

jDχ = j⊥ Dχ + jL Dχ. (115)

From the second term we find a contribution inside (103)

 
< iχ̄(x) ∂j · ∂ −2 ∂Dχ > . (116)

However we have seen that < χ̄(x)∂Dχ(y) >= −iδ4 (x − y), so that (116) becomes
Z Z
4 4 −2 4
d yd z∂j(y) · ∂ (y − z)δ (x − z) = d4 yj(y)∂∆0 (y − x).
 26

This is a numerical function depending only on the longitudinal current, which may be safely
put to zero.
In the part of (115) with the transverse current we may drop the derivative in D:

a
j⊥ Dχ = −j⊥ gf abc Gc χb , (117)

so that (103) takes the form

1 
a abc c b

< ∂G(x) − igχ̄(x) j⊥ f G χ + iη̄(tχ)q − iq̄(tχ)η >= 0. (118)
α
Now we are ready to renormalize this identity. We recall that the products of the field and
the corresponding source conserve their form in the renormalized quantities:

jG = jR GR , η̄q = η̄R qR and q̄η = q̄R ηR .

The first term in (118) is rewritten as

∂GR (x)
Zg−1/2 .
αR
The pair χ̄χ in the rest part gives a factor Zχ in terms of renormalized fields. Finally we
define the renormalized charge from the ghost-gluon interaction (94) which gives a factor for
this part
Zχ−1 ZG
−1
Z1χ .
1/2
As a result, multiplying by ZG we find the renormalized generating functional as

1 
a

< ∂GR (x) − igR Z1χ χ̄R (x) jR⊥ f abc GcR χbR + iη̄R (tχR )qR − iq̄R (tχR )ηR >= 0. (119)
αR
In the following we shall use almost exclusively the renormalized fields and omit subindex R.
So we rewrite (119) with the external colour index d explicit as
1  
< ∂Gd (x) − igZ1χ χ̄d (x) j⊥
a abc c b
f G χ + iη̄(tχ)q − iq̄(tχ)η >= 0. (120)
α
Differentiating various times (119) in external sources and then putting them to zero one
obtains an (infinite) set of Slavnov-Tailor identities for the QCD. We shall only derive the
simplest of them, which are necessary to prove the renormalizability of the theory.
a (y) and j b (z), and putting the sources equal to
Differentiating (119)twice over j⊥ , in j⊥ ⊥
zero afterwards we find
1
< ∂Gd (x) · Ga⊥ (y)Gb⊥ (z) >
α
 
= gZ1χ < χ̄d (x) f aec Gc (y)χe (y)Gb⊥ (z) + f bec Gc (z)χe (z)Ga⊥ (y) >= 0. (121)

The first term in this identity is the standard triple gluon Green function with two transverse
gluons and a longitudinal one. The two rest terms correspond to a Green function in which
the longitudinal gluon is substituted by an incoming ghost, one of the transverse gluons is
retained and the other substituted by a pair gluon-ghost at the same point. The vertex for
thir coalecence looks like the ghost-gluon interaction vertex but without the derivative in the
 27

Figure 7: Graphical content of the Slavnov-Tailor identity (121)

Figure 8: Leading order of the Slavnov-Tailor identity (121)

Figure 9: Next-to-leading order of the Slavnov-Tailor identity (121)

 28

Figure 10: Graphical content of the Slavnov-Tailor identity (122)

Figure 11: Leading order of the Slavnov-Tailor identity (122)

ghost line. The identity (119) is illustrated in Fig.7. Its contents in the lowest order is shown
in Fig. 8 and in the next, third order, in Fig.9.
It is remarkable, that in contrast to the QED, this Slavnov-Tailor identity relates not
the physical Green functions among themselves but rather a physical (triple gluon) Green
function with some artificial structures which do not appear in the perturbation theory at
all.
One can obtain a similar identity for the Green function with quarks. Differentiating
(119) over η(y) from the right and over η̄(z) from the left and putting then the sources equal
to zero we get
1
< ∂Gd (x)q̄(y)q(z) >
α
   
= igZ1χ < χ̄d (x)q̄(y) tχ(y) q(z) − χ̄d (x)q̄(y) tχ(z) q(z) >= 0. (122)

The first term is the standard Green function with two quark legs and one longitudinal gluon
leg. In the rest two terms the longitudinal gluon is again substituted by the ghost, one of
the quark is retained and the other is substituted by a pair ghost-quark at the same point.
Graphically (122) is illustrated in Fig. 10. Its contents in the lowest order is shown in Fig.
11 and in the third order in Fig. 12.
Now we shall use these Slavnov-Tailor identities to demonstrate renormalizability of the
QCD.
 29

Figure 12: Next-to-leading order of the Slavnov-Tailor identity (122)

6.3 Renormalizability of the QCD

Our main goal is to show that having at our disposal only one renormalization constant
for all vertex parts we can eliminate all the vertex divergencies. We shall choose for this
single renormalization constant Z1χ corresponding to the gluon-ghost vertex part. It will
determine the common renormalized coupling constant according to (94). Then the rest of
vertex renormalization constant will be automatically found from relation (95). We shall
demonstrate that then divergencies in all vertex parts will be eliminated. We shall limit
ourselves with the lowest non-trivial order when the divergent contributions come from only
one loop.
The proof is simplest in the transverse gauge α = 0. In this gauge in the gluon-ghost vertex
the derivative acting on the outgoing ghost propagator can be transferred on the incoming
ghost propagator, since the derivative acting on the gluon propagator is zero. Therefore in
the loop diagrams with ghosts for each external pair χ and χ̄ one can separate two external
momenta, which effectively decreases the order of divergence by two units. As a result, all
diagrams with external ghosts become convergent, except the ghost self-mass diagram.
The lowest order self-mass diagrams are shown in Fig. ??.
The ghost self mass Σχ (k2 ) acquires the form

Σ(k2 ) = k2 Fχ (k2 ) (123)

where F (k2 ) diverges only logarithmically. This divergence is compensated by the renormal-
ization of the gluon wave function, which introduces interaction in the form

(Zχ − 1)χ̄∂ 2 χ

and thus gives an additional contribution to the self-mass

δΣχ (k2 ) = k2 (Zχ − 1). (124)

Requirement of the finiteness of the sum Σχ + δΣχ determines the (divergent part of) Zχ . Its
finite part is in principle arbitrary and is fixed by the concrete definition of the renormalized
 30

Figure 13: The lowest order self-masses

mass (e.g. by the choice of the renormalization point µ2 in the definition (55). Note that the
renormalized ghost mass formally remains equal to zero.
Other self-masses are renormalized in the same manner as in the QED. As we have seen,
the gluon self-mass is purely transverse:
 kµ kν 
Σµν
G (k) = g
µν
− Π⊥ (k2 ) (125)
k2
where Π⊥ (0) = 0 so that we can substitute

Π⊥ (k2 ) → Π⊥ (k2 ) − Π⊥ (0) = k2 Fg (k2 ).

This liquidates the formal quadratic divergence in Π⊥ leaving only a logarithmic one. This
is eliminated by the gluon wave function renormalization which adds a contribution to the
self mass  kµ kν 
δΣµν
G (k) = g
µν
− 2 (ZG − 1)k2 . (126)
k
One then chooses the divergent part of ZG to cancel the logarithmic divergence in Fg (k2 ).
Its finite part remains arbitrary and is fixed by the concrete renormalization conditions.
The quark self mass, as is well-known, has two different logarithmically divergent contri-
butions The divergent part of the quark self mass Σq (p) has a structure

divΣq (p) = c1 (Λ)m + c2 (Λ)p̂, (127)

where both c1 and c2 are logarithmically divergent.. The first divergency is eliminated by the
mass renormalization which introduces an additional contribution

δ1 Σq (k) = δm. (128)

The second divergency is eliminated by the quark wave function renormalization, which gives
additionally
δ2 Σq (k) = (Zq − 1)k̂. (129)
 31

Figure 14: Lowest-order nontrivial vertex parts

Divergent parts of δm and Zq − 1 are determined by the condition that the total self-mass
be finite.
So after the renormalization all self-masses become finite.
Now we pass to the vertex parts. In the lowest non-trivial order they are shown in Fig.
14.
We start from the gluon-ghost vertex part for which we have our chosen renormalization
constant z1χ . The formal order of divergence for this vertex part is unity. But in our gauge
α = 0 it decreases by two units and becomes negative, which implies that the diagram is
in fact convergent. As a result the renormalization constant Z1χ proves to be finite (and
determined by the concrete renormalization conditions, e.g. by the chosen renormalization
point).
Consider now the triple gluon vertex part V3g . We turn to the Slavnov-Tailor identity
(121). On the left-hand side we have this vertex part coupled to the three external propaga-
tors. The propagator for the longitudinal gluon αkµ kν /k4 will be multiplied by factor kµ /α
and so give kν /k2 applied to the contribution without the longitudinal external line. As we
see the result does not depend on α. Inspecting the diagrams in the third order shown in
Fig. ?? we find that the right part of the identity is finite. It involves either the (renormal-
ized) mass terms, or the finite contribution from factor (Z1χ − 1) or diagrams, which have a
structure similar to the gluon ghost vertex part but without derivatives, which is explicitly
finite. Therefore the left part of the identity is also finite. The left-hand part, apart from
finite renormalized self-masses, contains the triple gluon vertex part together with its renor-
malizing counterterm equal to the bare 3G vertex multiplied by (Z1g − 1), all multiplied by
 32

the momentum k of the longitudinal gluon:

νλρ
kν V3g

From symmetry requirements and dimensional properties (dim V3g = 1) the (possibly) di-
vergent part of the vertex V3g may include two terms with different color structures. These
structures should contain color factors which guarantee SU3 invariance once multiplied by
the three gluon fields. One of such factors is evident from the Lagrangian: it is f abc . Indeed
f abc V1a V2b V3c is evidently invariant, which follows from the vector property of wa = f abc V2b V3c
(recall (17)). However there exists a second factor with the same properties.
To see it consider a transformation of the generator ta in the fundamental representation

ta → (ta )ξ = e−i(tξ) ta e+i(tξ) .

For small ξ we find

δta = −i(tξ)ta + ita (tξ) = −iξ b [tb , ta ] = −f abc ξ b tc .

We observe that this is the transformation of an SU (3) vector:

 ab  ab
e−i(tξ) ta e+i(tξ) = ei(T ξ) tb = UA (ξ) tb . (130)

Now take the trace of three generators t. Obviously

 
Tr (ta tb tc ) = Tr e−i(tξ) ta tb tc e+i(tξ)

and using (130) we find

 aa′  bb′  cc′ ′ ′ ′
Tr (ta tb tc ) = UA (ξ) UA (ξ) UA (ξ) Tr (ta tb tc ),

which shows that this trace is an SU (3) invariant tgensor of the third rank.. One can construct
two different invariant tensors from it. A totally antisymmetric
1
Tr (ta tb tc − tc tb ta ) = if abc
2
and a totally symmetric
1 abc
Tr (ta tb tc + tc tb ta ) =
d .
2
So the group SU (3) contains two invariant tensors of the third rank: a totally antisymmetric
f abc and a totally symmetric dabc .
Correspondingly the possibly divergent part of the three-gluon vertex may contain two
different structures with f and d factors
abc
divV3G,νλρ = cf (Λ)f abc (gνλ (k1 − k2 )ρ + gλρ (k2 − k3 )ν + gνρ (k3 − k1 )λ

+cd (λ)dabc (gνλ k3ρ + gλρ k1ν + gνρ k2λ ), (131)

where k1 + k2 + k3 = 0 and cf,d are two (possibly) logarithmically divergent constants. If

we multiply (131) by kν we still are left with many different structures which all have to be
finite. From this it follows that both constants cf and cd are in fact finite.
 33

Let us finally study the quark-gluon vertex part Vq . We consider the second Slavnov-
Tailor identity (122). Again on the right-hand side all contributions turn out to be finite. A
new diagram with a pair quark-ghost at the same point is finite: its formal order of divergence
is zero, but we can separate the momentum of the external ghost to lower it to minus unity.
So the left-hand part has also to be finite. This means that lµ Vqµ a (k, k + l) is finite, where

Vq is the quark-gluon vertex part together with its renormalizing counterterm (Z1q − 1)gγµ ta
and k and k + l are the quark momenta. The divergent part of the quark-gluon vertex part
has a structure
a
divVqµ = c(Λ)γµ ta , (132)

where c is a constant, possibly logarithmically divergent. Multiplying by lµ we find that c is

in fact finite.
Thus we have demonstrated that all triple vertex parts are made finite by renormalization.
The quartic vertex parts are all finite as they are, except the quartic gluon vertex V4G To
prove that the latter vertex part is also finite one has to derive a new Slavnov-Tailor identity,
containing four gluons on the left-hand part (performing one new differentiation in the current
j⊥ ). Then similarly to the triple vertex case, on the right-hand side there appear only finite
contributions, from which it will follow that also V4g is finite.
 34

7 Renormalization group. The running coupling constant

7.1 Kallan-Symanzik equation
The multiplicative transformation of fields which enters the renormalization procedure has an
intrinsic arbitrariness involving a finite numerical factor. Changing this factor has a meaning
of changing the renormalization conditions, e.g. the renormalization point µ2 . Choosing a
new value of µ2 implies a redefinition of the physical (renormalized) coupling constant g,
which therefore is in fact a function of µ. The set of finite multiplicative transformations of
the fields forms the renormalization group under which the theory remains invariant. This
invariance allows to find the relation between the coupling constant and µ. In special cases
when physical observables involve only one variable with dimension (e.g. the momentum)
the dimensional analysis then allows to find the dependence on this single variable.
To simplify the discussion we first study the simple model of a scalar massless self-
interacting field with a Lagrangian
1 1
L= φ0 ∂ 2 φ0 − g0 φ40 . (133)
2 4!
Here φ0 is the unrenormalized field and g0 is the unrenormalized coupling constant. The
renormalized quantities are defined in the standard manner

φ = Z 1/2 φ0 , Z1 g = Z 2 g0 , (134)

where Z and Z1 are the (divergent) renormalization constants for the field and the quartic
vertex.
We recall the dimensions of our quantities Initially we have three independent units: mass
M , length L and time T . In terms of these units

cm L L2
c = 3 · 1010 = , h = 1.054 erg · sec = M T 2 .
sec T T
We use the standard choice of units

h̄ = c = 1, so that M T = M L = 1.

So we are left with a single unit mass M and length and time are measured in the inverse
unity 1/M . So we measure dimensions in powers of units M :

dim µ = 1, dim xν = −1.

Since the action s is dimensionless: dims = 0 we find for our action

dim L = 4, dim φ0 (x) = 1, dim g0 = 0.

We introduce unrenormalized Green functions

(n)
G0 (x1 , ...xn ) =< T {φ0 (x1 )...φ0 (xn )} > (135)
 35

and renormalized ones by the same formula with renormalized fields. They are related as
(n)
G(n) = Z −n/2 G0 . (136)

Here and in the following we often suppress the arguments of the functions when they are
obvious. It is convenient to define single-particle irreducible parts of the Green functions
(’amputated Green functions’) separating from G(n) n external functions G(2) and putting in
the momentum space
n
(n) (n) Y (2)
G0 (k1 , ...kn ) = V0 (k1 , ...kn ) G0 (ki ) (137)
i=1

and similarly for the renormalized Green functions. Then the relation between renormalized
and unrenormalized quantities becomes inverse:
(n)
V (n) = Z n/2 V0 . (138)

The physical coupling constant is typically defined by the requirement

 2
V (4) (ki ) = g at ki kj = −µ2 δij − i, j = 1, 2, 3, 4 (139)
5
with a single renormalization point µ. and at the symmetric point which corresponds to
3 4
ki2 = − µ2 , s = t = u = − µ2
5 5
(one takes into account that for all particles incoming s = −2k1 k2 etc).
If one introduces the vertex part Γ(4) then the relations
(4) (4)
V (4) = gΓ(4) , V0 = g0 Γ0 (140)

lead to the standard result

(4)
 2
Γ0 (ki ) = Z1−1 at ki kj = −µ2 δij − . (141)
5
Let us consider the relation between the renormalized and unrenormalized quantities in
more detail, taking into account that the latter are only determined in the theory regularized
in the ultraviolet. It is essential that such regularization requires introduction of a regularizing
parameter with dimension. A simple (but not unique) possibility is just to introduce an
ultraviolet cutoff Λ for integration Euclidean momenta. Unrenormalized quantities will then
become functions of Λ which do not exist in the limit Λ → ∞. In contrast, renormalized
quantities will admit this limit after which they become independent of Λ.
In particular the finite renormalized coupling constant is independent of Λ. But it will
obviously depend on the normalization point µ. To specify this dependence we shall first
take two different values of the renormalization point µ and µ(0) . The renormalization condi-
tion (134) together with the definition (141) will express each of the corresponding coupling
constants g and g(0) as functions depending on different values of the normalization points:

g = f (g0 , µ, Λ), g(0) = f (g0 , µ(0) , Λ). (142)

 36

Note that both are finite. From the second relation we can determine the unrenormalized
coupling constant g0 as a function of g(0) and µ(0) :

g0 = g0 (g(0) , µ(0) , Λ). (143)

Of course it is divergent. Putting this into the first relation (142) we find
 
g = f g0 (g(0) , µ(0) , Λ), µ, Λ = f (g(0) , µ, µ(0) , Λ) = f (g(0) , µ, µ(0) ). (144)

The last equality follows from the finiteness of g, which admits taking the limit Λ → ∞.
However g is dimensionless, so that it has to depend only on the ratio of the two µ’s with the
dimension of mass:  µ 
g = f g(0) , (0) ≡ g(µ). (145)
µ
This is the equation which explicitly shows the dependence of the renormalized coupling
function on the renormalization point. Considered as a function of µ, g(µ) is called the
running coupling constant. In the following we shall take the logarithmic derivative in µ and
put µ = µ(0) afterwards. Then we find

∂g µ ∂g(g(0) , z) 
µ = (0) . (146)
∂µ µ ∂z

z=µ/µ0

Putting z = 1 we obtain
∂g  ∂g(g, z) 
µ = ≡ β(g), (147)
∂µ µ=µ(0) ∂z
 
z=1

which is a function of only one variable g.

Now we differentiate over µ relation (138) keeping g(0) and µ(0) or equivalently g0 and Λ
fixed. The unrenormalized Green function then does not depend on µ at all. The renormalized
one depends on µ both explicitly and through the µ-dependence of g. So we get
 ∂ ∂  (n) ∂
µ + β(g) V = V0n µ Z n/2 . (148)
∂µ ∂g ∂µ
We have
∂ n/2 ∂ n ∂
µ Z = µ e(n/2) ln Z = µ ln Z · e(n/2) ln Z ≡ nγZ n/2 , (149)
∂µ ∂µ 2 ∂µ
where we define the so-called anomalous dimension
1 ∂
γ(g) = µ ln Z. (150)
2 ∂µ
Transferring everything to the left-hand side we obtain the Kallan-Symanzik equation
 ∂ ∂ 
µ + β(g) − nγ V (n) = 0. (151)
∂µ ∂g
One concludes from this equation that the anomalous dimension γ is a finite quantity, since
the first two terms are obviously finite. Since by construction it is dimensionless, it cannot
depend on µ explicitly but is only a function of the coupling constant g: γ = γ(g).
 37

1.5

0.5

β(g)
0

-0.5

-1

-1.5

-2
0 1 2 3 4 5 6 7
g

Figure 15: Schematic behavior of β(g) with a positive derivative at g = 0

7.2 The running coupling

Formally the running coupling is determined by the equation (147)
∂g(µ)
µ = β(g). (152)
∂µ
If the β-function is known the running coupling is trivially found. Let µ = µ0 et where t is a
dimensionless variable going from −∞ to +∞. Then the equation takes the form
∂g(t)
= β(g) (153)
∂t
with a solution
g(t) dg′
Z
t= , (154)
g(0) β(g′ )
which implicitly determines function g(t).
Let us study some general properties of this implicit solution. Since −∞ < t < +∞ the
integral on the right-hand side of (154) has to diverge at least twice. This can occur either
due to going to ±∞ of the upper limit g(t) or due to vanishing of the denominator β(g) at
some points, called fixed points g∗ : β(g∗ ) = 0. Note that one such fixed point is known: it is
g = 0, since in any theory perturbative β(g) starts from the second or third order in g.
It turns out that the behavior of g(t) crucially depends on the sign of the derivative

dβ(g)/dg at g = 0

. Consider the case when this derivative is positive, the behavior of β(g) qualitatively shown
in Fig. 15 with two fixed points g1 < g2 different from zero. One easily finds that
in region g ≤ g1 at t → −∞ g(t) → 0 and at t → +∞ g(t) → g1 .
In region g ≥ g1 at t → −∞ g(t) → g2 and at t → +∞ g(t) → g1 .
The case with
dβ(g = 0)/dg < 0

is illustrated in Fig. 16. Here the behavior of the running coupling is opposite.
In region g ≤ g1 at t → −∞ g(t) → g1 and at t → +∞ g(t) → 0.
In region g ≥ g1 at t → −∞ g(t) → g1 and at t → +∞ g(t) → g2 .
Special interest is naturally drawn to the situations when the running coupling becomes
small g(t) → 0, since in this case one may hope to apply the perturbation theory. This
 38

1.5

0.5

β(g)
0

-0.5

-1

-1.5

-2
0 1 2 3 4 5 6 7
g

Figure 16: Schematic behavior of β(g) with a negative derivative at g = 0

happens at very small values of µ if dβ(g = 0)/dg > 0 and very large µ if dβ(g = 0)/dg < 0.
The first case unfortunately has little relevance, since in realistic theories, with small values
of µ, the influence of other mass parameters becomes dominant and completely overshadows
these predictions. In contrast the case when dβ(g = 0)/dg < 0 and µ is large has direct and
most important consequences, since then the influence of other mass parameters becomes
negligible. In this case we shall show that physical processes at large momenta become
governed by the running coupling at values of µ of the order of these momenta and, since
then g(t) → 0, the interaction decreases at large momenta. This phenomenon is known as
’asymptotic freedom’.
It turns out that in the QCD precisely this situation is realized. So in the QCD phenomena
occurring at large momenta admit application of the perturbation theory. We shall see that
in the following by directly calculating the β-function at small values of g.
 39

8 Solution of the Callan-Symanzik equation. High momen-

tum asymptotic
Solution of the Callan-Symanzik equation can be easily found. Putting again µ = µ0 et we
are to solve the equation
∂ ∂ 
+ β(g) − nγ(g) V (n) = 0. (155)
∂t ∂g
We introduce a function ḡ(t) by the equation
dḡ(t)
= β(ḡ(t)) (156)
dt
and seek the solution in the form

V (n) (g, t) = f (ḡ(t), t), with ḡ(t) = g. (157)

Then we find
df ∂f ∂f ∂V (n) ∂V (n)
= + β(ḡ(t)) = + β(g) = nγ(ḡ(t))f (158)
dt ∂t ∂g ∂t ∂g
with an obvious solution Rt
dt′ γ(ḡ(t′ ))
f (t) = f (0)en 0 (159)

or in terms of µ and g:
Rt
dt′ γ(ḡ(t′ ))
V (n) (µ, g) = V (n) (µ0 , g0 )en 0 , (160)

where g and g0 are related by

g dg′
Z
t= . (161)
g0 β(g′ )

**************************************

Now we pass to the study of the dependence of amputated Green functions V (n) (pi , µ, g)
when all momenta pi are simultaneously rescaled as

pi → pi et . (162)

A change of the mass unit M by factor 1/c, M → M/c evidently leads to the rescaling of
all dimensional quantities as
A → cdim A A

In particular
pµ → cpµ , xµ → xµ /c, φ(x) → cφ(x)

The generalized vertices V (n) have dimension

dim V (n) = 4 − n ≡ Γ (163)

 40

Γ is called the canonical dimension of V (n) . Indeed the Green function G(x1 , ...xn ) has
dimension n. Transforming it to the momentum space we have to intgrate over x n − 1 times,
which lowers the dimension to be 4 − 3n. Separating n pair Green functions G(2) (k) each of
dimension (-2) we arrive to (163). So explicitly showing the dependence on pi , µ and g

V (n) (cp1 , cp2 , ...., cpn , cµ, g) = cΓ V (p1 , p2 , ...pn , µ, g)

Take here
µ → µ/c, c = et

and we get
V (n) (pi et , µ, g) = eΓt V (n) (pi , µe−t , g) (164)

Once again we introduce function ḡ(t) with a different boundary condition ḡ(−t) = g.
Then we find
V (n) (pi , µe−t , ḡ(−t)) = V (n) (pi , −t)
R −t Rt
dt′ γ(ḡ(t′ ) dt′ γ(ḡ(−t′ ))
= V (n) (pi , 0)en 0 ) = V (n) (pi , 0)e−n 0 . (165)

Here V (n) (pi , 0) = V (0) (pi , µ, g0 ) where

ḡ(−t) dg′
Z
−t = (166)
g0 β(g′ )
or
g0 dg′ g(t) dg′
Z Z
t= ≡ . (167)
ḡ(−t) β(g′ ) g β(g′ )
So we find
ḡ(−t) = g, and g0 = g(t)

that is. at fixed lower limit g the upper limit g0 becomes a function of t, which is the stanard
running coupling g(t).
Now the integral of the anomalos dimension in the exponent. We have
Z t Z t
′ ′
dt γ(ḡ(−t )) ≡ dt′ γ(g′′ ), (168)
0 0

where Z g ′′ dg1
−t′ = .
g0 β(g1 )
Using Z g0 dg1
t=
g β(g1 )
we find Z g ′′ dg1
t − t′ ≡ t′′ = ,
g β(g1 )
so that g′′ is the standard running coupling at t = t′′

g′′ = ḡ(t′′ )
 41

Passing in the integral (168) to integration over t′′ we finilly find

Z t Z t
′ ′
dt γ(ḡ(−t )) = dt′′ γ(g(t′′ )). (169)
0 0

So our final result is

Rt
dt′′ γ(g(t′′ ))
V (n) (pi et , µ, g) = eΓt V (n) (pi , µ, g(t))e−n 0 , (170)

where g(t) is the running coupling constantdefined by (167). This relation explains why γ is
called the anomalous dimension.
 42

8.1 The Callan-Symanzik equation and the running coupling in the QCD
In the QCD obviously we have the Green functions which may involve different numbers
of quark and gluon legs V (nq ,nG ) . So the relation (138) will contain different powers of
the renormalization constants Zq and ZG . Also we encounter a new dimensionless gauge
parameter α, which renormalizes according to α0 = ZG α, so that the renormalized α also
depends on µ. Accordingly, taking the derivative µd/dµ we shall find a new contribution

∂α ∂
µ .
∂µ ∂α
However
∂α ∂ ln ZG
µ = −µ α = −2γG α ≡ δ.
∂µ ∂µ
So we find the Callan-Synanzik equation in the QCD
 ∂ ∂ ∂ 
µ +β +δ − nq γq − nG γG V (nq ,nG ) (pi , µ, g, α) = 0. (171)
∂µ ∂g ∂α
Here β, δ, γq and γG are all functions of g and α. Note that in the transverse gauge α = 0
and so also δ = 0. This means that the Callan-Symanzik equation is compatible with the
purely transverse gauge.
In the perturbation theory at small values of g one finds in the QCD

β(g) = −bg3 + O(g5 ), (172)

where for the total number of flavors Nf

1
b= (33 − 2Nf ) > 0 unless Nf > 16. (173)
48π 2
Thus b is positive with the realistic Nf = 6 and the QCD belongs to a theory with the
asymptotic freedom, in which the coupling constant decreases with energy. In fact in the
lowest order we solve the equation for the running coupling as
Z g(t) dg 1 1 1
t=− 3
= 2
− 2 , (174)
g bg 2b g (t) g

which gives
g2
g2 (t) = . (175)
1 + 2g2 bt
Relating t with the change of the momentum scale

Q2
2t = ln (176)
Q20

we finally find the relation between the running coupling constant at momentum scales Q0
and Q as
g2 (Q20 )
g2 (Q2 ) = 2 . (177)
1 + g2 (Q20 )b ln Q
Q 2
0
 43

This relation can also be rewritten as

1 1
g2 (Q2 ) = Q2
= 2 , (178)
g−2 (Q20 ) + b ln Q20
b ln Λ2Q
QCD

where we defined
g−2 (Q20 ) − b ln Q20 = − ln Λ2QCD . (179)

In this way we see that at large values of Q2 the dimensionless QCD coupling constant can
be fully determined by the mass parameter ΛQCD (’dimensional transmutation’).
Finally let us see how the Green function behave when their momentum variables pi are
all equally increased from scale Q0 to scale Q. The anomalous dimensions start from terms
of the order g2 at small g:
1
γu = − ζu g2 + O(g4 ). (180)
2
Correspondingly the factor corresponding to the leg with the anomalous dimension γu is
n Z t  o n1 Z t 1 o
′ ′
exp − dt γu g(t ) = exp ζu dt′
0 2 0 2bt′
ζ
u
 1 Q2  ζ4bu
= exp ln t = ln . (181)
4b 2 Q20
This factor will appear in the Green function for each leg u = G, q, q̄.
 44

9 Calculation of the β-function in the QCD

By definition
∂g 1 ∂ ln Z
β(g) = µ , γ(g) = µ
∂µ 2 ∂µ
at fixed bare coupling constant g0 and ultraviolet cutoff Λ. In terms of bare coupling and
renormalization constants
3/2
−1 −11/2 −1 1/2
g = g0 ZG Z1G = g0 ZG Zq Z1q = g0 ZG Zχ Z1χ . (182)

This definition generally depends on the renormalization prescription and chosen gauge. How-
ever in the perturbative approach one can demonstrate that the two first terms, corresponding
to one- and two-loop approximation do not depend on gauge or on the renormalization recipe.
Indeed let g and g̃ be two renormalized coupling constant introduced in different gauges or
renormalization prescriptions. One can obviously express one through the other in the first
two orders as
g̃ = g(1 + ag2 ) or g = g̃(1 − ag̃2 ). (183)

The two β-functions are obviously related by

∂g̃ ∂g ∂g̃ ∂g̃
β̃(g̃) = µ =µ = β(g) . (184)
∂µ ∂µ ∂g ∂g
In the first two orders we have
β(g) = −bg3 + cg5 . (185)

Then we find
 
β̃(g̃) = (−bg3 + cg5 )(1 + 3ag2 ) = − bg̃3 (1 − 3ag2 ) + cg̃5 (1 + 3ag̃2 )

= −bg̃3 + cg̃5 , (186)

so that the two first term do not change. We are going to calculate only the first term
corresponding to one-loop approximation. Then we can use any gauge or renormalization
prescription.
In the perturbation theory in the single loop approximation we have for any Z
Λ Λ
Z = 1 + g2 ζ ln + f inite terms ≃ 1 + g02 ζ ln + f inite terms, (187)
µ µ
where ζ and f inite terms are some constants independent of µ, since they are dimensionless.
From this we immediately find that ∂/∂ ln µ = −∂/∂ ln Λ and so
1 ∂ ln Zu 1 1
γu = − = − g02 ζu ≃ − g2 ζu . (188)
2 ∂ ln Λ 2 2
The importance of this result is that it is independent of the renormalization prescription
and only based on the knowledge of the divergent part of Z.
In the same manner we observe that
Λ
g = g0 + bg03 ln , (189)
µ
 45

so that
∂g
β=− = −bg3 . (190)
∂ ln Λ
Now take
3/2−1 Λ Λ
g = g0 ZG Z1G , with ZG = 1 + g02 ζG ln and Z1G = 1 + g02 ζ1G ln .
µ µ
Multiplying we get in the third order
3  Λ
g = g0 + g03 ζG − ζ1G ln
2 µ
Differentiating in µ and using the rest of relations (182) we find
3 1 1
b= ζG − ζ1G = ζG + ζq − ζ1q = ζG + ζχ − ζ1χ . (191)
2 2 2
To find ζ’s and afterwards b we have to study the divergent terms in the self-masses for ζu
and vertex parts for ζ1u , u = G, q and χ.
We can use any of the definitions of g from (182) to find b. To have some contact with
analogous calculations in the QED we shall use its definition through the vertex part quark-
gluon. Then we have to know ζq , ζG and ζ1g . To simplify we shall use the Feynman gauge.

9.1 Quark self-mass

The divergent part of the quark self mass Σq (p) has a structure

divΣq (p) = c1 (Λ)m + c2 (Λ)p̂, (192)

where both c1 and c2 are logarithmically divergent.. Taking as a mass renormlization recipe
the requirement Σq (p̂ = m) = 0 we obtain as a result
 
div Σq (p) − Σq (p̂ = m) = c2 (Λ)(p̂ − m). (193)

This has to be made finite by the addition of the term coming from the quark wave function
renormalization (Zq − 1)(p̂ − m) wherefrom we find

Zq − 1 = −c2 (Λ) (194)

and to find the divergent part of Zq we have to find the divergent term proportional to p̂ in
the self-mass Σq (p). It does not depend on the quark mass, so we can put m = 0 from the
start. Then in the second order in g

d4 l a ˆl α a 1
Z
2
Σq (p) = −g 4
t γα 2 γ t (195)
(2π) i l (l − p)2

(see Fig. 11).

The invariant sum ta ta is proportional to the unit 3×3 color matrix: ta ta = c · 1. Taking
the trace with the standard normalization
1
Tr{ta tb } = δab (196)
2
 46

Figure 17: The quark self-mass in the lowest order

we obtain
4
ta ta = . (197)
3
Now the γ−matrices. We use

{γµ , γν } = 2gµν , ˆllµ = 2lµ − lµ ˆl

so that
γα ˆlγ α = (2lα − ˆlγα )γ α = 2ˆl − 4ˆl = −2ˆl. (198)

We find
8
Z
d4 l ˆl
Σq (p) = g2 . (199)
3 (2π)4 i l2 (l − p)2
The divergent contribution comes from the region l >> p so that we can expand
1 1 pl 
2
= 2 1 + 2 2 + ... (200)
(l − p) l l

and the divergent contribution is

8 d4 l ˆl  pl 
Z
divΣq (p) = g2 1 + q . (201)
3 (2π)4 i l4 l2

The first term is evidently zero. In the second term we find an integral

d4 l lα lβ 1 d4 l 1
Z Z
4 6
= gαβ . (202)
(2π) i l 4 (2π)4 i l4

The last logarithmically divergent integral is standardly calculated by the rotation to the
Euclidean momenta with a cutoff Λ at large Euclidean |l2 | to give

d4 l 1 2π 2 Λ dl 1
Z Z
I(Λ) = = = 2 ln Λ. (203)
(2π)4 i l4 (2π)4 l 8π

So we find
4 2
divΣq (p) = g p̂I(Λ) (204)
3
and consequently the divergent part of Zq is

4 8 1 Λ
Zq − 1 = − g2 I(λ) = −g2 2
ln . (205)
3 3 16π µ
 47

Figure 18: The quark-gluon vertex part in the second order

9.2 Vertex part quark-gluon

The divergent part of the quark-gluon vertex has a structure c(λ)ta γα . So to find it we can
put the quark mass and all the external momenta equal to zero. In the second order the
quark-gluon vertex part has two contributions shown in Fig. 12 a, b.
The first (Fig. 12 a) is given by

d4 l b ˆ l ˆl 1
Z
Γa1,α = g2 4
t γµ 2 ta γα 2 tb γ µ 2 (206)
(2π) i l l l
(of course this is only its divergent contribution, which we do not specify explicitly in the
following). The color factor is

tb ta tb = tb tb ta + if abc tb tc ,
4
tb tb ta = ta ,
3
1 2 abc bcd d 3
if abc tb tc = if acb tc tb = −if abc tc tb = i f f t = − ta .
2 2
where we have used that
f abc f dbc = 3δab . (207)
Thus we have found the color factor to be
4 3 a
tb ta tb = − t . (208)
3 2
Now the γ matrices.

γµ ˆlγα ˆ
lγ µ = γµ (2lα − γα ˆl)ˆlγ µ = 2lα (−2ˆl) − l2 (−2γα ) = 2(γα l2 − 2lα ˆl).

We get for the vertex

4 3 a d4 l 1 4 3
Z
Γa1,α = 2g2 − t 4 6
(γα l2 − 2lα ˆl) = g2 − ta γα I(Λ). (209)
3 2 (2π) i l 3 2
So we find the contribution from the diagram Fig. 8a
4 3
(1)
Z1q − 1 = −g2 − I(Λ). (210)
3 2
 48

Figure 19: The gluon self-mass in the second order

Now the second contribution from the diagram shown in Fig. 12 b. Putting all external
momenta equal to zero we have

d4 l abc h −ˆl 1
Z i
Γa2λ = −ig2 4
f lν gλµ − 2lλ gµν + lµ gνλ tc γµ 2 tb γ µ 4 . (211)
(2π) i l l

The color factor is

3
if abc tc tb = ta .
2
The spin factor is
h i
lν gλµ − 2lλ gµν + lµ gνλ γ ν ˆlγ µ = l2 γλ − 2lλ γν ˆlγ ν + l2 γλ = 2l2 γλ + 4lλ ˆl.

So we find
23 d4 l 1 2 9
Z
Γa2λ =g ·2·t a
(l γλ + 2lλ ˆl) = g2 ta γλ I(Λ) (212)
2 (2π)4 i l6 2
and the second part of the renormlization constant is

(2) 9
Z1q − 1 = −g2 I(Λ). (213)
2
In the sum we find
(1)
 4
Z1q − 1 = g2 3 − I(Λ). (214)
3

9.3 Gluon polarization operator

To find ZG we have to study the gluon polarization operator Πab µν (k) which, in the second
order, is represented by four Feynman diagrams shown in Fig. 13 a − d. We shall use that it
is diagonal in color and pure transverse

Πab ⊥ 2
µν = δab Pµν Π⊥ (k ) (215)
 49

and we also know that Π⊥ (k2 = 0) = 0, so that we can substitute

Π⊥ (k2 ) → Π⊥ (k2 ) − Π⊥ (0). (216)

In the following we omit subindex ⊥. Multiplying (215) by gµν δab and summing over the
indices we get
1  aaµ 
Π(k2 ) = Πµ (k) − Πaaµµ (0) . (217)
24
It is clear from the start that the diagram of Fig. 13 d will give no contribution since it does
not depend on k.
We start with the quark contribution shown in Fig. 13 a.

1
Z
d4 l n ˆ
l a µ ˆl + k̂ o  
(q)
Π = − g2 Tr t a
γµ t γ − k = 0 . (218)
24 (2π)4 i l2 (l + k)2

As before ta ta = 4/3 so that

Tr{ta ta } = 4.

lγ µ = −2ˆl, so that we find a trace

Now γµ ˆ

−2Tr{ˆl(ˆl + k̂)} = −8(l2 + kl).

Thus we find
1 d4 l  l2 + kl 1
Z
(q)
Π = − · 4 · (−8)g2 − .
24 (2π)4 l2 (l + k)2 l2
In the integrand we find a factor

l2 + kl k2 + kl k2 + kl  2kl  k2 2(kl)2
2
−1=− 2
≃− 2
1− 2 =− 2 + .
(l + k) (l + k) l l l l4

Integration will substitute (kl)2 → k2 l2 /4 so that in the sum we get −k2 /(2l2 ) This gives
1 1 2
Π(q) = − · 4 · (−8) · (− )g2 k2 I(Λ) = −g2 k2 I(Λ). (219)
24 2 3
Actually quark of each flavour gives this contribution. So the final result has to be
multiplied by Nf . Our first contribution to the gluon renormalization constant is therefore

(q) 2
ZG − 1 = −g2 Nf I(Λ). (220)
3
The second piece comes from the ghost contribution shown in Fig. 13 b:

1 d4 l bac cab l(k + l)

Z  
(χ)
Π = − g2 if if − k = 0 . (221)
24 (2π)4 i l2 (l + k)2

The color factor is if bac if cab = 24. We find

1 d4 l  l(k + l) 1 3 1
Z
Π(χ) = − · 24g2 4 2 2
− 2
= − Π(q) = k2 g2 I(Λ). (222)
24 (2π) i l (l + k) l 4 2

Thus the second part of ZG is

(χ) 1
ZG − 1 = g2 I(Λ). (223)
2
 50

Finally we study the gluon contribution shown in Fig. 13 c

1 1 2 d4 l
Z
(G)
Π = · g if abc (−if abc )
24 2 (2π)4 i
h i2 1  
(k − l)ν gµλ + (2l + k)λ gµν + (−l − 2k)µ gνλ − k = 0 . (224)
l2 (l + k)2
Note the symmetry factor 1/2. The color factor again gives if abc (−if abc ) = 24. The square
bracket squared gives

4(k − l)2 + 4(2l + k)2 + 4(l + 2k)2 + 2(k − l)(2l + k) − 2(k − l)(l + 2k) − 2(2l + k)(l + 2k)

= 3(k − l)2 + 3(2l + k)2 + 3(l + 2k)+ (k − l + 2l + k − l − 2k)2

= 18(k2 + l2 + kl).

We obtain
1 1 d4 l 1 −kl
Z
Π(G) = · 24 · 18 · g2 . (225)
24 2 (2π)4 i l2 (k + l)2
Now the integration gives

kl kl  2kl  2(kl)2 1 k2
= 1 − → − → − . (226)
(k + l)2 l2 l2 l4 2 l2

Taking together all the factors we get

9
Π(G) = k2 g2 I(Λ), (227)
2
(G)
which gives the final piece of ZG

9
Z (G) − 1 = I(Λ). (228)
2

9.4 β-function
Collecting our results and taking into account the expression for I(Λ) we find
4 1  4 1 2 1 9 1
ζq = − , ζ 1q = 3 − , ζ G = (− Nf + + . (229)
3 8π 2 3 8π 2 3 2 2 8π 2
From (191) we find
 11 1  1
b= − Nf , (230)
2 3 8π 2
which gives (173).
 51

10 e+e− annihilation into hadrons

As a first application of the perturbative QCD we shall consider the process of annihilation
of a pair of leptons onto hadrons. To be concrete we take an electron-positron pair as initial
leptons and the process in question is

e+ + e− → hadrons (231)

(see Fig. 20). Its cross-section is shown in Fig. 21. It is implied that the observer measures
the total cross-section to produce any number of hadrons in the annihilation. It is to be
noted, that apart from final hadronic states the pair may also go into leptonic states. The
typical process in this respect is production of a µ+ µ− pair:

e+ + e− → µ+ + µ− (232)

whose cross-section is well-known from the standard perturbative approach in the QED. This
process will serve as a basis for comparison. We shall assume that the c.m. energy squared
s = q 2 is large to be able to apply the asymptotic freedom in the QCD. The total cross-section
of the reaction (231) can be written as a product of the leptonic and hadronic parts times
the photon propagator. In the Feynman gauge we have
1 4 1
σ tot (s) = e Lαβ H αβ 4 . (233)
J q
Here the leptonic part is
1X
Lαβ = [v̄(−¯l)γα u(l)]ξη · [ū(l)γβ v(−¯l)]ηξ , (234)
4 pol

where l and ¯l are the momenta of the electron and positron, the sum is taken over polarizations
and ξ and η are spinor indexes. The hadronic part is
XZ
Hαβ = dτn (q)h0|jαem (0)|Xn ihXn |jβem (0)|0i, (235)
n

where j em is the hadronic electromagnetic current, |0 > and |Xn > are the vacuum and inter-
mediate hadronic state respectively. As usual dτn (q) is the phase volume element restricted
by the energy-momentum conservation:

dτn (q) = (2π)4 δ(pn − q)dτn , (236)

where dτn is the unrestricted phase volume element for n intermediate particles.
The leptonic part of (233) is known, since

u(l)ū(l) = ˆl + m, v(−¯lv̄(−¯l) = ˆ¯l − m,

X X
(237)
pol pol

so that neglecting the small leptonic mas m

1
Lαβ = Tr{ˆ¯lγα ˆlγβ } = ¯
lα lβ + lα ¯lβ − gαβ l¯l. (238)
4
 52

Figure 20: The process e+ + e− → hadrons

Figure 21: The cross-section of e+ + e− → hadrons

In the hadronic part we insert

Z
(2π)4 δ4 (pn − q) = d4 xeix(q−pn ) (239)

to lift the restriction on the energy-momentum of intermediate states. Using

e−ixpn h0|j em (0)|Xn i = h0|j em (x)|Xn i

and then summing over all states Xn we rewrite the hadronic tensor as
Z
Hαβ (q) = d4 xeiqx h0|jαem (x)jβem (0)|0i. (240)

It is instructive to compare Hαβ with the hadronic part of the photon polarization operator
(h)
Παβ which is
Z
(h)
Παβ (q) = i d4 xeiqx h0|T{j em (x)jβem (0)}|0i (241)
and is graphically illustrated in Fig. 22. It is clear that in fact the hadronic tensor H is
just the discontinuity of the polarization operator across its cut at q 2 > 0, which puts all
intermediate hadrons on their mass-shell according to Cutkosky rules:
1 (h)
Hαβ = Disc Παβ at q 2 > 0. (242)
i

Figure 22: The hadronic part of the photon polarization operator.

 53

Both Π and H are pure transverse:

 qα qβ  2 2 (h)
 qα qβ  2
Hαβ = − gαβ + q h(q ), Παβ = − gαβ + q π(q 2 ) (243)
q2 q2
with
dim h(q 2 ) = dim π(q 2 ) = 0

and obviously
1
h(q 2 ) = Disc π(q 2 ) at q 2 > 0. (244)
i
The cross-section can be expressed directly via the dimensionless function h(q 2 ).
 qα qβ  ¯
Lαβ H αβ = q 2 h(q 2 ) − gαβ + (lα lβ + lα¯lβ − gαβ l¯l).
q2

We define p = (l − ¯l)/2. Then (qp) = 0, l = q/2 + p, ¯l = q/2 − p and

¯lα lβ + lα ¯lβ − gαβ l¯l) = −2pα pβ + 1 qα qβ − gαβ (q 2 /4 − p2 ).

2
Then we find 3 
Lαβ H αβ = q 2 h(q 2 ) q 2 − p2 .
4
The c.m. energy s = q2 = 2ll, the incoming flux J = 4l¯l and also p2 = (l − ¯l)2 = −s/4. So
¯
we find
e4
σ tot (s) = h(q 2 ). (245)
4s
Naive considerations tell that the dimensionless function h(q 2 ) cannot depend on q 2 and
so is a number. Then the cross-section has the order e4 /s and rapidly falls with s. How-
ever we know that renormalization introduces a new parameter with a dimension of mass,
which transforms into the QCD-parameter ΛQCD . Therefore we may expect a non-trivial
dependence of h on q 2 .
To study this dependence we set up the Callan-Symanzik equation for h(q 2 ). This requires
knowledge of its renormalization, which is made within the QED. The one-particle irreducible
transverse photon Green function is defined as

V (2) = G−1
γ , (246)

so that we find
(2)
V (2) = Zγ V0 , (247)

where Zγ is the photon renormalization constant. Separating the transverse projector P⊥ we

have
(2)
V (2) (q 2 ) = q 2 − Π(q 2 ), V0 (q 2 ) = q 2 − Π0 (q 2 ), (248)

where Π is the photon polarization operator. Its renormalization obviously goes as q 2 − Π =

Zγ (q 2 − Π0 ) or
Π(q 2 ) = Zγ Π0 (q 2 ) − q 2 (Zγ − 1) (249)
 54

Figure 23: The lowest order contribution to the hadronic tensor

and is not purely multiplicative. However H is defined as a discontinuity in k2 and the extra
terms proportional to (Zγ − 1) are analytic in k2 so that for function h(k2 ) we find

h(q 2 ) = Zγ h0 (q 2 ). (250)

The photon renormalization constant is different from unity in the second order in the
electromagnetic coupling constant e2 /(4π) = 1/137. Neglecting this small electromagnetic
correction we find
h(q 2 ) = h0 (q 2 ), (251)
which implies that the photon anomalous dimension is zero (in fact it has order e2 ). Therefore
we find the Kallan-Symanzik equation for h:
 ∂ ∂  2
µ + β(g) h(q , µ, g) = 0 (252)
∂µ ∂g
and solving it, for q 2 = q02 e2t , t >> 1
   
h q02 e2t , µ, g = h q02 , µ, g(t) , (253)

where g(t) is the running QCD coupling. In the limit t → ∞ g(t) → 0 so that at large q 2

h(q 2 , µ, g) = h(q02 , µ, 0) + O(1/ ln(q 2 /q02 ). (254)

At g = 0 function h does not involve strong interactions and corresponds to the process in
which the lepton pair annihilates into a pair of a quark and antiquark of any flavor, which do
not interact between themselves (Fig. 23). It coincides with the analogous function for the
purely electromagnetic process e+ e− → µ+ µ− except for the difference in electromagnetic
charges and the presence of colors and flavors. Note that the analogous process of elastic
scattering e+ e− → e+ e− generates additional contributions and so does not agree with (254),
see Fig. 24.
If the electromagnetic charge of the quark of flavor f is Qf (in units e) then we find in
the high-energy limit
σ tot (e+ e− → hadrons) X
2

2 2

R= = 3 Q f 1 + O(1/ ln(q /q 0 ) . (255)
σ tot (e+ e− → µ+ µ− ) f

The number on the right-hand side in fact depends on the number of flavors which can be
√
created at a given energy s. In the limit s → ∞ it includes all 6 flavors, so that
 4 1  15
R=3 3 +3 = .
9 9 3
 55

Figure 24: Processes e+ e− → µ+ µ− and e+ e− → e+ e−

Figure 25: Contributions to h of the second order in g(t)

However up to energies of the order of several tens of GeV production of b and t quarks is
suppressed and we rather have
 4 1  10
R=3 2 +2 = .
9 9 3
This agrees quite well with the experimental data. Note that should we forget about the
color this number would be reduced to 10/9 in gross contradiction with the data. In this way
the experimental data on lepton-annihilation into hadron clearly demonstrate that the color
exists and implies exactly 3 color states for quarks.  
One can calculate corrections to R due to strong interactions developing h q02 , µ, g(t)
in powers of the small g(t). The second order term comes from the diagrams shown in Fig.
25. The contribution of these diagrams councides with similar ones in the QED, except for
the electromagnetic couplings Qf of the photon to produced quarks and the quark-gluon
couplings gta , which after summation over a give (4/3)g. So this contribution can be directly
taken from the very old results in QED (year 1953). Together with (256) they give

σ tot (e+ e− → hadrons) X

2
 g2 (q 2 ) 2 2 2

R= = 3 Q f 1 + + O(1/ ln (q /q 0 .
) (256)
σ tot (e+ e− → µ+ µ− ) f
4π 2
 56

Figure 26: The deep inelastic lepton-hadron scattering

Figure 27: The inclusive cross-section for e(l) + N (p) → e(l′ ) + X(p′ )

11 Deep inelastic lepton-hadron scattering: kinematics

We pass to the process which represents the classical application of the perturbative QCD
and has served to establish its validity as the theory of strong interaction, the deep inelastic
lepton-hadron scattering (Fig. 26):

e(l) + N (p) → e(l′ ) + X(p′ ), (257)

where N represents the target nucleon and instead of the electron also µ and ν can be used
as the projectile. The coresponding inclusive cross-section is shown in Fig. 27. Similarly
to the preceding chapter the inclusive cross-section to observe the outgoing lepton can be
factorized into leptonic and hadronic parts;
1 d3 l′ αβ e
4
dσ = L αβ H . (258)
J (2π)3 2l0′ q4
Neglecting the lepton mass we have J = 4lp. The c.m. energy squared for the reaction is
s = (l + p)2 = m2 + 2lp where m is the nucleon mass. We shall be interested in the region of
high energies when s >> m2 , so that s ≃ 2lp.
For unpolarized initial particles the leptonic part is
1X
Lαβ = [ū(l′ )γα u(l)]ξη · [ūlγβ u(l′ )]ηξ . (259)
2 pol

As for the lepton pair annihilation into hadrons, it is calculated explicitly. Neglecting the
lepton mass
1
Lαβ = Tr {ˆl′ γα ˆlγβ } = 2(lα′ lβ + lα lβ′ − gαβ ll′ ). (260)
2
 57

The hadronic part is

1 XX
Z
Hαβ = dτn (p + q − pn )hN (p)|jαem (0)|Xn ihXn |jβem (0)|N (p)i. (261)
2 pol n

Up to an unimportant coefficient, it has an obvious meaning of the total cross-section for the
virtual-photon-nucleon collision. It depends on two variables: the c.m. energy of the collision
(p + q)2 = m2 + 2pq + q 2 and the photon virtuality q 2 .
It is easy to find that q 2 = (l − l′ )2 < 0. Indeed in the antilab system
q
q 2 = 2m2e − 2me m2e + l′ 2 < 0

. Usually one introduces a positive Q2 = −q 2 . The missing mass squared

M 2 = (p + q)2 = m2 + 2pq + q 2 > m2 so that 2pq − Q2 > 0,

It follows that the so-called scaling variable

Q2
0≤x= ≤ 1. (262)
2pq

The value x = 1 obviously occurs when M 2 = m2 and corresponds to purely elastic collision.
In the general case we have

M 2 = 2pq(1 − x) ≡ 2ν(1 − x) (263)

where 2ν = 2pq is approximately the c.m. energy squared for the photon-nucleon collision.
The missing mass squared is a finite part of it and grows with ν. Deep inelastic scattering
corresponds to exactly this situation when ν → ∞ and x is a fixed finite positive number
smaller than unity. The standard variables which characterize the hadronic part of the deep
inelastic scattering are x and Q2 .
As in the preceding chapter we lift the restriction on the energy-momentum of the inter-
mediate hadronic states representing
Z
4 4
(2π) δ (q + p − pn ) = d4 zeiz(p+q−pn )

and using
eiz(p−pn ) hN (p)|jαem (0)|Xn i = hN (p)|jαem (z)|Xn i.

Summing over all intermediate states we find

1X
Z
Hαβ = d4 zeiqz hN (p)|jαem (z)jβem (0)|N (p)i. (264)
2 pol

This should be compared with the forward scattering amplitude for the virtual-photon-
nucleon scattering (see Fig. 28)

iX
Z
Aαβ = d4 zeiqz hN (p)|T {jαem (z)jβem (0)}|N (p)i. (265)
2 pol
 58

Figure 28: The forward γ ∗ − N scattering amplitude

Obviously
1
Hαβ =Disc(q+p)2 >m2 Aαβ . (266)
i
Note that both H and A are dimensionless:

dim j em (x) = 3, dim hN (p)||N (p)i = −2, dim d4 x = −4. (267)

The tensor structure of H and A is determined by the current conservation:

q α Hαβ = Hαβ q β = 0 (268)

and similarly for A. From the two vectors p and q one can construct two tensors which are
orthogonal to q:
(q) qα qβ
gαβ = gαβ − 2 (269)
q
and  pq  pq  (q) (q)
pα − q α p β − q β 2 = pα pβ . (270)
q2 q
Therefore H can be presented via two scalar dimensionless structure functions:
1 (q) 1 (q)
Hαβ = −gαβ F1 (x, Q2 ) + p(q) p F2 (x, Q2 ). (271)
2π ν α β
A similar representation is valid for A with two scalar functions Z1,2 :

1 (q) 1 (q)
Aαβ = −gαβ Z1 (x, Q2 ) + p(q) p Z2 (x, Q2 ). (272)
2π ν α β
The inclusive cross-section (258) can be directly expressed via these two structure func-
tions. Omitting the rather tedious derivation we find

1 d3 l′
dσ = I, (273)
J (2π)3 2l0′

where
πe4 
2 θ m2 2 θ

I= 2F1 sin + F2 cos . (274)
8mE 2 E ′ sin4 θ2 2 ν 2
Here E and E ′ are the initial and final energies of the lepton (electron) and θ is its scattering
angle in the lab. system (that is, where the target proton is at rest).
 59

The two structure functions F1,2 may be related to the cross-sections σ⊥ and σL for the
scattering off the nucleon of the virtual photons polarized in the plane transverse to the p − q
plane and in that plane respectively. For the transverse photon polarization vector one can
take the transverse unit vectors e⊥ orthogonal to both q and p and normalized as e2⊥ = −1.
Then one immediately finds
σ⊥ = c F1 , (275)

where the constant c depends on the choice of flux (which is unphysical for virtual photons).
The longitudinal photon polarization vector is evidently proportional to p(q) :

p(q)
eL = q .
(q) 2
−p

Then one has

2 ν2
p(q) = m2 −
q2
From the part with F1 we find
(q) (q)
pα pβ
−gαβ 2 =1
−p(q)
and from the part with F2 we find
4 2
p(q) p(q)
2 =−
−νp(q) ν

So we get
2
 p(q)  h  m2 ν i h 1 i
σL = c F1 − F2 = c F1 − − 2 F2 ≃ c F1 − F2 . (276)
2ν ν q 2x
From naive considerations we again may expect that dimensionless functions F1,2 (x, Q2 )
do not depend on variable Q2 with dimension mass squared but depend only on x. This situa-
tion is known as scaling. Again renormalization introduces parameter ΛQCD with dimension,
so one expects corrections to the scaling picture following from the QCD dynamics.
Experimental observation indeed show violation of scaling with the variation of Q2 (see
schematic illustration in Fig.29
They also testify that the longitudinal cross-section σL is very small as compared to the
transverse one, which implies that with a high precision one has

F2 (x, Q2 ) = 2xF1 (x, Q2 ). (277)

*****************************************

This relation can be understood if one assumes that the particle interacting with the
photon has spin 1/2. Indeed consider the simplest diagram for H in the lowest order of
perturbation theory, shown in Fig. 30, corresponding tto the elastic process

γ ∗ + N (q) → γ ∗ (q) + N (q) (278)

 60

Figure 29: Function F2 (x, Q2 ) at Q2 = 2 and 25 (Gev/c)2

Figure 30: The simplest diagram for the hadronic tensor H: elastic scattering
 61

Figure 31: The parton picture for e(l) + N (p) → e(l′ ) + X(p′ )

It gives
1X  
Hαβ = ū(p)γα (m + p̂ + q̂)γβ u(p) 2πδ(m2 − (p + q)2 ) .
2 pol

We neglect m2 at high enegies and get for the matrix element

     
ūνξ1 (p) γα uaξ4 (p),
X X
ū(p)γα p̂ + q̂)γβ u(p) = p̂ + q̂ γβ
ξ1 ξ2 ξ2 ξ3 ξ3 ξ4
pol a

where a = ±1/2 are the polarizations and we show the spinor indeces ξ explicitly. As we
know
uaξ4 (p)ūaξ1 (p) = p̂ξ4 ξ1 .
X

So the matrix element becomes

 
Tr {p̂γα (p̂ + q̂)γβ } = 4 pα (p + q)β + (p + q)α pβ − p(p + q)gαβ . (279)

The δ-function is
1
δ(2pq + q 2 ) = δ(2ν − Q2 ) = δ(x − 1).
2ν
It is trivial to see that with 2pq + q 2 = 0 the matrix element is orthogonal to q and can be
presented as
(q) (q)
2p(q)
α pβ − νgαβ .

From this we conclude

F1 = δ(x − 1), F2 = 2δ(x − 1) = 2xδ(x − 1), (280)

so that we indeed find F2 = 2xF1 . Obviously if the charged particle has spin 0 we shall not
find a term proportional to gαβ so that in that case F1 = 0.
This fact gave a basis for the parton model of Feynman, in which it is assumed that the
hadron is made of components, partons, the electrically charged sort of them having spin
1/2. Graphically it can be illustrated by Fig. 31. This model is best interpreted in the
antilab. system where the incident proton moves along the z-axis in the opposite direction
 62

with p+ = 0 and large p− . The viirtual photon can be taken to have only the z component
of its momentum. The partons will have a certain distribuion in their ki− = xi p− and their
”+” components will be small, since k+ = −k⊥ 2 /2k . So nothing will depend on either
−
”+” or transverse components of the partons and only the distribution in k− will determine
the interaction. As follows from the study of the elastic scattering (278), one will obtain
precisely the value x = Q2 /2ν if instead of the proton with momentum p one has the parton
with momentum xp. In other words the incident virtual photon selects the parton with
momentum xp− out of the total distribition of their momenta k− , provided it is electrically
charged. If one assumes that electrically charged partons have spins 1/2 (quarks) then relation
(277) will be valid for the resulting structure functions.

*****************************************

To finish let us study the region of z which gives the dominant contribution in the integral
representations (264) or (265). Let us choose the system in which p⊥ = q− = 0. Then q 2 = q⊥2

and ν = p− q+ . With a finite p− we have q⊥ → ∞ and q+ → ∞. In the exponent we have

qz = q+ z− + (qz)⊥ .

So the important region is

z− → 0 and z⊥ → 0. (281)

As follows z 2 → 0 although z itself does not go to zero, since z+ is finite. Therefore to find
the asymptotic of H or A in our kinematic region we have to study the matrix element in the
integrand on the light cone z 2 → 0. To this end we shall use the so-called operator expansion
on the light cone, which will be studied in the next chapter.
 63

12 Composite operators and their renormalization

To simplify we study composite operators in the theory of a single scalar real field φ(x) (with
a renormalizable interaction λφ4 ). A composite operator is any local operator constructed
from a finite number of fields φ(x) and their derivatives. Locality allows to always make a
shift and relate the composite operator to the point x = 0. Examples of composite operators
are
φn , φn1 ∂α1 ...∂αs φn2 , φn1 ∂α1 ...∂αs (∂ 2 )m φn2 and so on.

Fundamental fields φ are assumed to be renormalized. Any composite operator Oi may be

fully characterized by a set of Green functions which contain this operator and a number n
of fundamental operators Φ:

Gn (x1 , ...xn ) =< T{Oi φ(x1 )...φ(xn )} > . (282)

Amputating the external Green functions G2 for the fundamental fields we obtain the corre-
sponding generalized vertices Vn .
Each operator Oi may be characterized by the number li of operators φ which it contains.
It is also a Lorentz tensor of a certain rank, corresponding to the number of Lorentz indices
α1 , ...αs . Such a tensor is generally reducible in the tensor algebra: one can construct from
it tensors of smaller ranks contracting it with invariant tensors gαβ and ǫαβγδ . In fact we
shall mostly deal with symmetric tensors so practically one can contract them with gαβ .
Subtracting results of such contraction from the original tensor one can make it irreducible,
that is giving zero under such contractions. The rank of this irreducible tensor is called
Lorentz spin s. Each composite operator can be presented as a sum of operators with a given
spin s. E.g.
 1 1  1
O = φ∂α ∂β φ = φ ∂α ∂β φ − gαβ ∂ 2 φ + gαβ φ∂ 2 φ = O2 + gαβ O0 ,
4 4 4
where  1 
O2 = φ ∂α ∂β − gαβ ∂ 2 φ
4
is an irreducible operator of spin 2 and

O0 = φ∂ 2 φ

is an irreducible operator of spin 0. In the following we study exclusively operators with a

given spin s. One more characteristic of a composite operator is its dimension d in units of
mass. It is just a sum of dimensions of its components φ and ∂ each of them contributing one
unit. In the following an important role will be played by the twist t of the operator which
is the difference between its dimension and spin

t = d − s.

The twist does not change upon introduction of uncontracted derivatives. For all non-trivial
operators we have
d ≥ 2, s ≥ 0, t ≥ 2. (283)
 64

Figure 32: The schematic structure of the Green function Vn

(i)
Let us study generalized vertices Vn in the perturbation theory. They are represented
by standard Feynman diagrams with a single vertex of a special kind which corresponds
to the operator Oi . To this vertex exactly l lines should be attached and it also has to
contain the appropriate number of derivatives and their contractions. Let us first deternine
the dimensions of Vn . The initial Green function Gn in the coordinate space has evidently
dimension dim Gn (x1 , ..xn ) = d + n. in the momentum space its dimension is d − 3n so that
the generalized vertices will have dimension

dim Vn (pi , ..pn ) = d − n

where d is the dimension of the composite operator. In the momentum representation we

can consider the momenta pi of n external lines as independent. The total momentum
corresponding to the lines attached to the vertex Oi will then be just the sum of these
external momenta. In the lowest approximation we obviously find that n = l so that

Vn(0) = δnl Tα1 ,...αs (pi )Pq (pi pj ), (284)

where Tα1 ,...αs (pi ) is a certain irreducible tensor of rank s constructed from the external mo-
menta and Pq (pi pj ) is a homogeneous polynomial of rank q constructed from scalar products
of external momenta (corresponding to contracted derivatives). The dimnesions are

dim Tα1 ,...αs = s, dim Pq (pi pj ) = 2q

so we have
dim Vn(0) = d − l = s + 2q, d = l + 2q + s, t = l + 2q

(and t does not depend on spin s). Considered as a set of all Green functions for different n
(284) represents our operator in the lowest order of the perturbation theory;

Oi ≡ Vn(0) (i). (285)

The contribution of higher orders can be represented as a sum of all diagrams in which l
lines attached to the vertex pass to n external lines with all possible interaction in between. In
other words the vertex is attached to the Green function Mln corresponding to the transition
of l initial (off-shell) fundamental particles to n number of such particles (Fig. 32).
l
d4 ki
Z Y
Vn (pi ) = Vn(0) (pi ) + Vn(0) (ki )Mln (k1 ..kl |p1 , ..pn ), (286)
i=1
(2π)4
 65

or symbolically
(0) (0)
Oi = Oi + Oi ⊗ M. (287)

The Green function M is finite after renormalization. However integrations over the internal
lines attached to the vertex of momenta ki may introduce new divergence. So the composite
operator Oi defined by (287) is generally divergent in the ultraviolet and requires renormal-
ization. This latter is much simpler that the standard renormalization procedure since the
composite vertex does not participate in the interaction. It cannot be internal and (in our
example) is just a single vertex in all Feynman diagrams.
To study renormalization we first have to find out the character of divergent terms. As
usual it will heavily rest on the dimensional considerations. We have seen that

dim Vn (pi , ..pn ) = d − n

Possible divergent terms are to be constructed only from the external momenta and ultraviolet
cutoff Λ, since at large integration momenta all masses can be neglected. Vn has to be an
irreducible tensor of rank s and so can be expanded in all possible tensors of rank s constructed
from the external momenta Tαj 1 ,...αs (pi ). So the divergent part is

Tαj 1 ,...αs (pi )Cj (Λ, pi ).

X
divVn = (288)
j

Since dim Tj = s, we find that

dim Cj = d − n − s = t − n

. This implies that Cj have a structure

cjq Pq(j) (pi pj )Λt−n−2q , t − n − 2q ≥ 0

X
Cj =
q=0

with summation over values of q until t−n−2q ≥ 0. One observes that the leading divergence
correspond to q = 0 and has the form Λt−n . It is fully determined by the twist of the operator
and the number of external lines. For fixed t one finds a finite number of divergent diagrams
with n ≤ t and for n = t the divergence is logarithmic.
It is to be stressed that for an operator Oi with given l, d, s and t and type of tensor T i
and polynomial P i the divergent expression contains contributions with different l, d and t,
tensors T j and polynomials P j . To make this more instructive we can rewrite the divergent
part (288) as follows, introducing δnlj and an additional sunnation over lj :

cjqj Tαj 1 ,...αs (pi )Pq(j) (pi pj )δnlj Λt−lj −2qj .

X
divVn = j
(289)
j,lj ,qj

Comparing with (284) we see that in the sum we find the zero-order vertex part of different
composite operators Oj , which have the same spin but different values of number of lines lj
and twist tj = lj + 2qj . So we find
(0)
cj Λt−tj [Vj ]n ,
X
divVn = (290)
j
 66

Figure 33: The schematic structure of Vn with the separated diagrams for the complete
composite operator Oi

or, since this is true for any n simply

(0)
cij Λti −tj Oj .
X
divOi = (291)
j

Summation obviously goes over all operators Oj with tj ≤ ti . So in the divergent part
of the composite operator there appear all bare operators with the same or smaller twist.
This phenomenon is called operator mixing. The smaller the twist of the mixed operator
the stronger is the divergence. For the mixed operator of the same twist the divergence is
logarithmic.
We have to stress that the divergence we have found is the so-called proper divergence of
the diagram, which corresponds to the case when all the momenta joined at the vertex are
equally large. Apart from this divergence, as in the usual treatment, there may be partial
divergencies, when only some of these momenta are large and the rest remain finite. In this
case the dimensional argument is not valid. However it is not difficult to find the general rule
for separating these partial divergences. In fact if the divergent contribution comes from the
integration only over some of the lines attached to the vertex then this divergent part will
depend on internal momenta which are not large and on which the rest of the diagram will
depend non-trivially. This corresponds to presenting the diagram in the form shown in Fig.
33 in which all the momenta to the right of the separated full Oi are finite. This brings us
to the symbolic relation using (287)
(0) (0) (0) (0) (0)
cij Λti −tj Oj + cij Λti −tj Oj ⊗ M + f.p.Oi ,
X X
Oi + Oi ⊗M = (292)
j j

where f.p. denotes the finite part of the contribution. So we obtain

X (0) (0)
 X 
δij − cij Λti −tj Oj + δij − cij Λti −tj Oj ⊗ M =
X
f.p.Oi = Zij Oj , (293)
j j j

where
Zij = δij − cij Λti −tj . (294)
Thus taking for the renormalized operator OiR the finite part of Oi we find its relation to the
unrenormilized one in the form of matrix multiplication
X
OiR = Zij Oj , tj ≤ ti . (295)
j

This rule generalizes the standard renormalization rule for the fundamental fields.
 67

13 Operator expansion
In some applications one meets with products of operators in different space-time points, like
A(0)B(−x). It may be both a simple product or a T-product. We shall study the case of
T-product not indicating this explicitly. Operators A and B are assumed to be renormalized.
One can write the standard expansion around x = 0
X 1
B(−x) = (−1)s xα1 ...xαs ∂α1 ...∂αs B(0) (296)
s=0
s!

and subsequently an expansion of A(0)B(−x)

X 1
A(0)B(−x) = (−1)s xα1 ...xαs A(0)∂α1 ...∂αs B(0) (297)
s=0
s!

in terms of composite operators A(0)∂α1 ...∂αs B(0). These operators are however not irre-
ducible under Lorentz indices. We can present then as a sum of irreducible operators with
spins s, s − 2,...:

A(0)∂α1 ...∂αs B(0) = A(0)Tα(s)

1 ...αs
(∂)B(0) + gα1 α2 T (s−2) (∂)B(0) + ..., (298)

where T (s) is some irreducible Lorentz tensor of spin s. Then expansion (297) will be rewritten
as
A(0)Tα(s) (∂)B(0)xα1 ...xαs Pq (x2 ),
X
A(0)B(−x) = 1 ...αs
(299)
s,q

where Pq is a polynomial of rank q

(s)
In fact the composite operators A(0)Tα1 ...αs (∂)B(0) do not exist although A and B are
taken renormalized. Using the perturbative theory Wilson found a modification of this ex-
pansion valid after the composite operators in the sum are renormalized. One can easily
recognize this modification in the simple example of A = φn1 and B = Φn2 . Then already in
the lowest order of perturbation theory for A(0)B(−x) from contractions between operators
φ(0)φ(−x) one will find a sum of different operators multiplied by powers of ∆(−x) = (1/x2 ).
In higher orders of perturbation theory one finds Green functions in which these powers of
1/x2 are multiplied by functions finite at x = 0 and admitting expansion in powers of x. So
the net change is twofold. First, polynomials Pq (x2 ) are changed by functions of x2 which
(s)
may be singular at x2 → 0. Second, apart from original operators A(0)Tα1 ...αs (∂)B(0) also
other local operators appear. Renormalization will introduce still more local operators into
the sum.
This brings us to the operator expansion as proposed by Wilson:

α1 ...αs α1
x ...xαs Cis (x2 ),
X
A(0)B(−x) = Ois (300)
i,s

where summation goes over all values of Lorentz spin s and all local operators Ois with a
given spin s. Wilson coefficients C(x2 ) are some functions of x2 which may be singular at
x2 → 0. It is assumed that all operators in (300) are renormalized, so that Wilson coefficients
are finite.
 68

The benefit of the Wilson expansion consists in the possibility to extract the leading
contribution to the product A(0)B(−x) in the limit x → 0 or x2 → 0.
To start note that all operators entering the Wilson expansion may be assumed to have
definite dimensions. Then we obtain a relation: dA + dB = dis + dim Cis − s, where dA , dB
and dis are dimensions of operators A, B and Ois respectively. So we obtain

dim Cis = dA + dB − dis + s = dA + dB − tis , (301)

where tis is the twist of operator Ois . The canonical dimension dim Cis determines its crude
behavior at x2 → 0  1 dim Cis /2
Cis (x2 ) ∼ . (302)
x2
The greater the dimension the stronger is the singularity at x2 → 0. From (301) we conclude
that this dominant singularity comes from the term in the Wilson expansion with operators
Ois with a minimal twist (=2). As in the previous situations, in absence of any massive
parameters this canonical dimension would completely determine the x-dependence of C(x2 ).
However renormalization introduces a massive parameter µ which can be substituted by the
QCD parameter ΛQCD . This opens the way to corrections to the canonical behavior in
the form of logarithmic corrections containing log x2 Λ2QCD . They can be determined by the
Callan-Symanzik equation for the Wilson coefficients.
To construct it first we set up the Callan-Symanzik equation for the Green functions with
(A)
composite operators. Let Vn be the gneralized vertex with n fundamental operators and a
composite operator A. The renormalization procedure relates it with the verex as
(A)
 ′
Vn(A) = Z n/2 ZAA′ Vn(A ) )0 ,
X
(303)
A′

where on the right-hand side appear the unrenormalized Green function with all operators
A′ which mix with A in the process of renormalization together with the matrix of renormal-
(A)
ization constants ZAA′ . Applying the operator µ(∂/∂µ) to this relation we shall obtain the
Callan-Symanzik equation in the matrix form
 ∂ ∂ 
µ + β(g) − nγ(g) − γ (A) (g) Vn(A) = 0. (304)
∂µ ∂g

Here VnA is considered as column of all vertices V (A) for different A. The first three terms in
(304) are just unit matrices proportional to δAA′ . The non-trivial anomalous dimension γ (A)
is defined as a matrix
∂ −1
γ (A) = µ Z (A) · Z (A) . (305)
∂µ
Thus the Callan-Symanzik equation for composite operators has the same form as for funda-
mental operators with the only difference that due to mixing the anomalous dimension for
the composite operator has a matrix structure.
The same is true when the Green functions include several composite operators at different
space time points. In particular for the product A(0)B(−x) we get the equation
 ∂ ∂ 
µ + β(g) − nγ(g) − γ (A) (g) − γ (B (g) Vn(AB) = 0, (306)
∂µ ∂g
 69

where anomalous dimension γ (B) is defined similarly to (305) through the matrix of renor-
malization constants of operators B.
Now we take the Wilson expansion for A(0)B(−x). For brevity we include the product
of xα1 ,...αs into the expansion coefficients and rewrite it just as
X
A(0)B(−x) = Oi Ci (x), (307)
i

(AB)
where index i now also includes spin s. Correspondingly the Green function Vn will be
presented as
Vn(AB) (x) = Vn(i) Ci (x),
X
(308)
i
(i)
where Vn is the Green function with a composite operator Oi . Differentiating this relation
in µ we have to take into account that the Wilson coefficients depend on both g and µ. So
putting relation (308) into the Callan-Symanzik equation (306) we get
Xn  ∂ ∂   ∂ ∂  o
Ci (x) µ + β(g) − nγ(g)− γ (A) (g)− γ (B )(g) Vn(i) + Vn(i) µ + β(g) Ci (x) = 0.
i
∂µ ∂g ∂µ ∂g
(309)
(i)
Now we recall the Callan-Symanzik equation for Vn
 ∂ ∂ 
µ + β(g) − nγ(g) − γ (i) (g) Vn(i) = 0 (310)
∂µ ∂g

and putting this into (309) finally get the Callan-Zymanzik equation for the Wilson coeffi-
cients:  ∂ ∂ 
µ + β(g) + γ (i) (g) − γ (A) (g) − γ (B) (g) Ci (x) = 0. (311)
∂µ ∂g
In this equation it is understood that Ci (x) is in fact a matrix in three pairs of indices: (AA′ ),
(BB ′ ) and (ii′ ). Each of the three anomalous dimensions is a matrix in one pair and diagonal
in two others.
From the Callan-Symanzik equation follows the behavior of the coefficients Ci (x) when
x → 0:
n Z t  o  
−t Γt
Ci (e x, g) = e P exp − dt′ γ (A) (g′ ) + γ (B) (g′ ) − γ (i) (g′ ) Ci x, g(t) (312)
0

where the canonical dimension is Γ = dA + dB − dim Oi .

 70

14 Deep inelastic lepton-hadron scattering: evolution in Q2

14.1 Moments of the structure functions
Returning to our formulas for the hadronic tensors Hαβ or Aαβ in terms of the average in the
nucleon state of the product of two electromagnetic currents, we have to study, say, operator

T {jαem (z)jβem (0)} (313)

at z 2 → 0. We expand this product in local operators at z = 0 in the form (300). Our first
problem is how to associate the external indices αβ with the indices α1 ...αs which appear in
the expansion. Note that the product
s
z α1 ...z α Cis (z 2 )

in the expansion can be substituted by

∂ α1 ....∂ αs C̃is (z 2 ).

Indeed the difference contains terms proportional to gαi αk which give zero contracted with
α1 ,...αs
irreducible spin-s operators Ois . But then in the integral (265) each derivative ∂αi is
equivalent to qαi . Associating the structure functions with terms proportional to either gαβ
or pα pβ we find that none of the derivatives ∂αi may carry indices α or β. These indices may
appear only in the tensors g. This gives two possible structures in the expansion:
 
T {jαem (z)jβem (0)} = Oα1 ,...αs − gαβ ∂α1 ...∂αs C̃ (1) (z 2 ) + gαα1 gβα2 ∂α2 ...∂αs C̃ (2) (z 2 )
X

is
 
Oα1 ,...αs − gαβ zα1 ...zαs C (1) (z 2 ) + gαα1 gβα2 zα3 ...zαs C (2) (z 2 ) .
X
= (314)
is
We put this representation into the matrix element and take into account that
1X α1 ...αs
 
hN (p)|Ois |N (p)i = ois pα1 ...pαs − (trace terms) (315)
2 pol

where ois are some numbers; trace terms contain less numbers of pα and include several
metric tensors g necessary to make the whole construction irreducible. Inspecting (314) we
see that the first term in the average will be proportional to gαβ and the second to pα pβ . So
if
1 (q) 1 (q)
Aαβ = −gαβ Z1 (x, Q2 ) + p(q) p Z2 (x, Q2 ) (316)
2π ν α β
we find
i
Z
(1)
 
d4 zeiqz ois Cis (z 2 )z α1 ...z αs pα1 ...pαs − (trace terms) ,
X
Z1 = (317)
2π is

i
Z
(2)
 
d4 zeiqz ois Cis (z 2 )z α3 ...z αs pα3 ...pαs − (trace terms) .
X
Z2 = ν (318)
2π is
We change
∂ ∂
Z Z
(1) (1)
d4 zeiqz z α1 ...z αs Cis (z 2 ) → −i ... − i d4 zeiqz Cis (z 2 )
∂qα1 ∂qαs
 71

and since the Fourier tranform

Z
(1) (1)
d4 zeiqz Cis (z 2 ) = Cis (q 2 )

is a function of q 2 we change the derivatives in qα to the ones in q 2

∂ ∂
−i ... − i → (−2qα1 )...(−2qαs )(∂/∂q 2 )s + (trace terms)
∂qα1 ∂qαs

The arising trace terms will give nothing since they are contracted with the irreducible prod-
uct of pα . So we get
i X  s   ∂ s
(1)
Z1 = ois ν − (trace terms) (−2i)s 2
Cis (q 2 ), (319)
2π is ∂q

and similarly
i X  s−2   ∂ s−2
(2)
Z2 = ν ois ν − (trace terms) (−2i)s−2 2
Cis (q 2 ). (320)
2π is ∂q

Note that the trace terms in (319)and (320) are subleading at large ν In fact in the
simplest case of spin 2
1
pα1 pα2 − (trace terms) = pα1 pα2 − gα1 α2 m2 ,
4
so that multiplied by q α1 q α2 we get

1  m2 
ν 2 − q 2 m2 = ν 2 1 + x .
4 4ν
The second term gives a small correction of the order m2 /Q2 .
Neglecting the trace terms andf expressing ν = −q 2 /2x we find

i X  1 s 2 s  ∂ s (1) 2
Z1 = ois (iq ) Cis (q ), (321)
2π is x ∂q 2

and similarly
i  q 2  X  1 s−2 2 s−2  ∂ s−2 (2) 2
Z2 = − ois (iq ) Cis (q ). (322)
2π 2s is x ∂q 2
At this moment it is instructive to study the canonical dimensions of the Fourier trans-
(1,2)
forms Cis (q 2 ). Obviously in the coordinate space
(1) (2)
2 dim j em = dim Ois − s + dim Cis = dim Ois − s + 2 + dim Cis

so that

dim C (1) (z 2 ) = 2 dim j em (z) − tis = 6 − tis , dim C (2) (z 2 ) = 2 dim j em (x) − tis − 2 = 4 − tis
(323)
where tis is the twist of operator Ois . The Fourier transforms will have dimensions lowered
by 4
dim C (1) (q 2 ) = 2 − tis , dim C (2) (q 2 ) = −tis . (324)
 72

Figure 34: The right and left cuts of the amplitude Aαβ in variables ν and x

The leading behavior at large q 2 will come from the operators with the minimal twist equal
to 2. Then C (1) (q 2 ) ∼ const and C (2) (q 2 ) ∼ 1/q 2 .
In (321) and (322) we define new coefficients

(1)
 ∂ s (1) 2
Bis (q 2 ) = iois iq 2 C (q ), (325)
∂q 2

(2)
 q 2  2 ∂ s−2  (2) 2
Bis (q 2 ) = iois − iq C (q ) (326)
2 ∂q 2
with equal dimensions
(1) (2)
dim Bis = dim Bis = 2 − tis .

Then we finally obtain

1 X (1) 2  1 s 1 X (2) 2  1 s−1
Z1 (x, Q2 ) = Bis (Q ) , Z2 (x, Q2 ) = B (Q ) . (327)
2π is x 2π is is x

In this formula the dependence on the large q and finite p are factorized in the dependence
on q 2 and x.

*****************************************

Now we have to relate these series to the properties of the structure functions F1,2 (x, Q2 )
which are just the discontinuities of Z1,2 (x, Q2 ) considered as functions of ν at positive values
ν > (1/2)Q2 . We can consider this discontinuity as that across the cut in the variable x
at 0 < x < 1 and fixed Q2 . Obviously the forward scattering amplitude Aαβ , Fig. 28,
considered as function of its energy (p + q)2 at fixed t = 0 will also have a crossed cut at
(p−q)2 > 0, that is at −1 < x < 0. In the complex x plane the cut will extend at −1 < x < 1.
These cuts are illustrated in Fig. 34.
By crossing symmetry
Aαβ (p, q) = Aβα (p − q), (328)

which implies
Z1,2 (x, Q2 ) = Z1,2 (−x, Q2 ). (329)
 73

Accordingly we can write a dispersion relation (suppressing the fixed variable Q2 )

1 +1 dx′
Z
Z1,2 (x) = Discν Z1,2 (x′ ). (330)
2πi −1 x − x′

From (329) we find

1 +1 dx′ 1 +1 dx′
Z Z
Z1,2 (−x) = ′
Discν Z1,2 (x′ ) = Discν Z1,2 (−x′ ),
2πi −1 −x − x 2πi −1 −x + x′
which implies that the structure functions can be defined at x < 0 as odd functions in x.

Discν Z1,2 (x, Q2 ) = 2iF1,2 (x, Q2 ) = −2iF1,2 (−x, Q2 ). (331)

With this definition we have

1 +1 dx′
Z
Z1,2 (x, Q2 ) = F1,2 (x′ ). (332)
2π −1 x − x′

Expansion in values of spin s in (327) is in fact expansion of Z1,2 around infinity in the
x-plane. Taking x >> x′ in (332) we find

1 X  1 s
Z +1
2 s
Z1,2 (x, Q ) = dx′ x′ F1,2 (x′ , Q2 ). (333)
2πx s=0 x −1

Comparing this with expansions (327) we finally obtain

1
Z 1 1
Z 1
(1) (2)
dxxs−1 F1 (x, Q2 ) = Bis (Q2 ), dxxs−2 F2 (x, Q2 ) = Bis (Q2 ).
X X
(334)
2 −1 i
2 −1 i

On the left-hand side there are so-called moments of the structure functions F1 and F2
considered as functions of x. They are directly observable (if not equal to zero by oddness
of F ’s). These moments are expressed by a sum of (all) terms Bis for a given spin s related
to a set of composite operators with a given s and minimal twist equal to 2. The number of
such operators is finite (and small).
 74

14.2 Evolution in Q2 of the structure function moments

(1,2)
With the twist of the operators Ois equal 2, the coefficients Bis are dimensionless. So
(1,2) 2 2
the behavior of the coefficients Bis (Q ) at large Q is governed by the Callan-Symanzik
equation R t ′ em ′ ′
Bs(1,2) (Q20 e2t , g) = Pe− 0 dt (2γ (g(t ))−γs (g(t )) Bs(1,2) (Q20 , g(t)). (335)
(1,2) (1,2)
Here the column Bs = {Bis } is a set of all coefficients with a given spin s which
correspond to operators Ois mixed by renormalization, γ em is the anomalous dimension of
the electromagnetic current and γs is the anomalous dimension matrix for the set of operators
Ois .
The anomalous dimension of the electromagnetic current is zero. This is a consequence
of its conservation
∂em = 0 (336)

and is a parfticular case of the vanishing of the anomalous dimension for conserved currents.

*****************************************

Consider a slightly more general flavor current

(a)
jα(a) = q̄f λf f ′ γα qf′
X

(a)
where the summation is over flavours and λ(a) is any diagonal matrix in the flavours λf f ′ =
(em)
λ(a)δf f ′ . In the particular case of the electromagnetic current λf = Qf , the electric charge
of the quark of flavour f . This flavour current is conserved. Indeed the quark satisfies the
equations
(i∂ˆ − M + gtĜ)q = 0, or ∂qˆ = i(−M + gtĜ)q,

where M is the mass matrix, also diagonal. Conjugation gives

∂ˆq̄ = iq̄(−M + qtĜ)

and we find
∂ α (q̄λγα(a) q) = iq̄[−M + gtĜ, λ(a) ]q = 0

since both M and λa) are diagonal and [M, λ(a) ] = 0.

Now consider the insertion of j (a) into an arbitrary unrenormalized Green function

(G(a) (a)
α )0,n =< T {(jα )0 (x)q0 (x1 )...q0 (xn1 )q̄0 (y1 ), ...q̄0 (yn2 )G0 (z1 ), ...G0 (zn3 )} >

Taking the divergence and using (336) we are left with only contributions from the discon-
tinuities of the T -product at points when the time of x becomes equl to the times of other
operators. These discontinuities are zero when j (a) passes through the gluon operators and
reduce to commutators when it passes through q and q̄. So we get
(a)
∂(G(a) )0,n = δ(x0 − x10 < T {[(j0 )0 (x), q0 (x1 )]q0 (x2 )....q̄0 (y1 )....G0 (z1 ), ...} >
 75

(a)
+δ(x0 − x20 < T {q0 (x1 )[(j0 )0 (x), q0 (x2 )]q0 (x3 )....q̄0 (y1 )....G0 (z1 ), ...} > +...
(a)
+δ(x0 − y10 < T {q0 (x1 )q0 (x2 )....[(j0 )0 (x), q̄0 (y1 )]q̄0 (y2 )....G0 (z1 ), ... > +...} >

The equal time commutators are known:

(a)
[(j0 )0 (x), q0 (x1 )]x0 =x10 = −δ3 (x − x1 )λ(a) γ0 q0 (x1 ),
(a)
[(j0 (x)0 , q̄0 (y1 )]x0 =y10 = δ3 (x − y1 )q̄0 (y1 )λ(a) γ0

So we get
 
∂(G(a) )0,n = −δ4 (x − x1 ) < T { λ(a) γ0 q0 (x1 ) q0 (x2 )...q̄0 (y1 )...G0 (z1 )...} >
 
−δ4 (x − x2 ) < T {q0 (x1 ) λ(a) γ0 q0 (x2 ) ...q̄0 (y1 )...G0 (z1 )...} > −...
 
+δ4 (x − y1 ) < T {q0 (x1 ).... q̄0 (y1 )γ0 λ(a) q̄0 (y2 )....G0 (z1 )...} > +... ≡ X0 . (337)

Passing to renormalized quantities we get

 n  n
Z (a) Zu1/2 ∂G(a) 1/2
n = Zu X,

1/2
where there appears the product of Zu of all operators q, q̄ and G and X is renormalized
1/2
and finite. Dividing by the product of all Zu we get

Z (a) ∂G(a)
n =X

(a)
and is finite. Obviously from this follows also than Gn is finite and conseqently Z (a) is finite
and can be put to unity. Therefore the anomalous dimension is zero.

γ (a) = 0. (338)

*****************************************

Presenting
γs = ζs g2 (339)
(1,2)
we find the asymptotic of the set Bs at large Q2 as
 Q ζ/2b
Bs (Q2 , g) = ln Bs (Q20 , 0), (340)
Q0
where matrix exponentiation and multiplication is implied and we recall that in QCD
1  2 
b= 11 − Nf .
16π 2 3
To be more specific we have to find the operators of twist 2 in the QCD which may
participate in the Wilson expansion of the product of two electromagnetic currents. Our
guide will be gauge invariance of these operators implied by such invariance of the original
 76

Figure 35: Single loop diagrams for the non-singlet renormalization constant Z N S

product j em (z)j em (0). One can construct three types of gauge invariant QCD operators of
twist 2 and spin s: quark non-singlet, quark singlet and gluon.
The (Nf − 1) non-singlet operators are
h i
OsN S,a = q̄λ(a) γα1 iDα2 ...iDαs q , (341)

2
where λ(a) are (Nf − 1)× (Nf − 1) diagonal matrices with Tr λ(a) = 0 normalized as Tr λ(a) =
2. Their number depends on the number of flavours taken into account. In fact λ(a) are
(Nf − 1) diagonal matrices out of the total Nf2 − 1 Hermithean traceless matrices 2tc , where
tc are Nf2 −1 generators of the group SU (Nf ) in the fundamental representaton formed by the
quarks qf of different flavours. With only two flavours, u and d, one has only one non-singlet
matrix, for which one can take the Pauli matrix σz . With three flavours u,d and s one has
three non-singlert matrices, for wich one usually takes the standard 3 × 3 matrices 2t3 and
2t8 and so on.
The singlet operators are
h i
OsS = q̄γα1 iDα2 ...iDαs q , (342)

1h i
OsG = Gαα1 iDα2 ...iDαs−1 Gααs . (343)
2
Here D = ∂ − igtG for quarks and D = ∂ − igT G for gluons. Symbol [...] means sym-
metrization in indices α1 ...αs and addition of trace-terms necessary for making the operator
irreducible in spin.
Each non-singlet operator ON S does not mix with others and is renormalized multiplica-
tively. In the single-loop approximation the contribution to the renormalization constant
Z N S comes from 3 diagrams shown in Fig 35. Calculations (see next subsection) give
s
1 4 2 1
ζsN S = −
X
1 − + 4 . (344)
8π 2 3 s(s + 1) k=2
k

Note that ζ1N S = 0 which corresponds to conservation of the flavor current. The asymptotic
of the coefficients BsN S is
 Q ps
BsN S (Q2 ) = BsN S (Q20 ) ln , (345)
Q0
 77

where
s
|ζ N S | 4  2 X 1
ps = = 1− +4 . (346)
2b 33 − 2Nf s(s + 1) k=2
k

For Nf = 4 one finds pN 2

S = 0.427, pN S = 0.667 and slowly grows with s.
3
Note tht the non-singlet operators give the contribution which is flavour dependent and
so different for hadrons containing quarks of different flavour. As a simple example consider
the case when only the lightest u and d quarks are taken into account. In this case there is
only one non-singlet operator and λ(a) = σz . The contribution of this operator to the matrix
elements betwen quarks u and d will satisfy

¯ (a) d|
< u|j (a) |u >= − < ū|j (a) |ū >= − < d|j (a) |d >=< d|j ¯ >,

whereas the singlet and gluon operators will not discriminate between the quark flavours.
As the proton wave function is made of uud plus quark-antiquark pairs plus gluons and the
neutron wave function is made of udd plus quark-antiquark pairs plus gluons the difference
between the proton and neutron will be described by precisely j (a)

< p|j (tot) |p > − < n|j (tot) |n >= 2 < N |j (a) |N >

Here j (tot) corresponds to the sum of all three operators (341),(342) and (343) and it is
assumed that N is a column p, n.
So from the experimental point of view the behaviour of the non-singlet moments BsN S
describes the difference between the structure functions observed in the deep inelastic scat-
tering on the proton and neutron targets. This is a quantity difficult to observe due to the
lack of the single neutron target. The results on the neutron target can only be extracted
from the sacttering on the deuteron target, with the inevitable additional uncertainty coming
from the bound nature of the neutron. For this reason the more trustful comparison with
the experiment requires the study of the singlet and gluon contributions to the operator
expansion.
Singlet and gluon operators mix under renormalization. In the single loop approximation
one finds contributions to 4 components of the renormalization constant matrix illustrated in
Fig. 36. Therefore to find the asymptotic of the coefficients B S and B G one has to diagonalize
the corresponding renormalization constant matrix.
We shall not calculate the anomalous dimensions for the mixing singlet and gluon op-
ertaors, but will just give the final result folowing from the study of divergent parts of the
diagrams in Fig. 36. One gets for each spin s the matrix

g2
γ̂s = − ζ̂s , (347)
8π 2
where
s
4h 2 1i
ζsSS = ζsN S =
X
1− +4 ,
3 s(s + 1) k=1
k

s2 + s + 2
ζsSG = − ,
s(s + 1)(s + 2)
 78

Figure 36: Single loop diagrams for the matrix of singlet renormalization constants

8 s2 + s + 3
ζsGS = − ,
3 s(s2 − 1)
s
2 24 1
ζsGG
X
= 1 + Nf − + 12 . (348)
3 s(s + 2) k=2
k
Conservation of the energy-density tensor leads to relations for s = 2

2Nf ζ2SG + ζ2GG = ζ2SS + ζ2GS = 0.

 79

14.3 Calculation of anomalous dimensions for the non-singlet operator

The procedure to find the renormalization constants for composite operators in the one-loop
approximation starts from the definition
n/2
Vn (g, µ) = ZZφ V0,n (g0 , Λ), (349)

where Vn and V0,n are the renormalized and unrenormalized Green functions with n external
lines. Z is generally a matrix in the space of mixing operators. The inverse relation is
−n/2
Vn(0) g0 , Λ) = Z −1 Zφ Vn (g, µ). (350)

Note that Vn is finite and V0,n is not.

In the single loop approximation, as we have seen, the maximal divergence of the Green
function V0,n is Λt−n where t is the twist. In our case t = 2 and the divergent Green functions
we are going to study have the maximal number of external lines n = 2 abd the divergence
is logarithmic. Accordingly we present
Λ Λ
Zφ = 1 + ζφ g02 ln , Z = 1 + ζg02 ln . (351)
µ µ
Then the anomalous dimensions are
∂ ln Zφ
γφ = = −g02 ζφ ≃ −g2 ζφ (352)
∂ ln µ
and
∂Z −1
γ= Z = −g02 ζ ≃ −g2 ζ. (353)
∂ ln µ
So to find the anomalous dimension it is sufficient to search for terms proportional to lnΛ in
the unrenormalized Green functions. We have, using (351)
h Λ n  Λ n 
V0,n (g0 , Λ) = Vn (g, µ) 1 − g02 ln − ζφ − ζ ≃ Vn (g, µ) − g2 ln − ζφ − ζ Vn(0) , (354)
µ 2 µ 2
(0)
where Vn is the Green function in the lowest (zero) order of perturbation theory. So the
divergent part of V0,n is
divV0,n = g2 ζ̃ Vn(0) ln Λ, (355)

where
n
ζ̃ = ζ +
ζφ (356)
2
and is generally a matrix to be applied to the column Vn (0) of all mixing operators. Formulas
(355) and (356) give a practical recipe to find the anomalous dimension matrix coefficient
ζ. One has to find the divergent part of the Green functions with composite operators from
which coefficient ζ̃ is to be extracted. Then, knowing the anomalous dimensions for the
fundamental fields one calculates ζ from (356).
We shall limit ourselves with the calculation of the anomalous dimension for an arbitrary
non-singlet quark operator in the QCD. This operator does not mix with others and so ζ is
just a number and not a matrix. The relevant diagrams for the Green function V2 are shown
in Fig. 35. Our aim is to find their divergent parts.
 80

Figure 37: Diagram Fig. 35,a

We start with the diagram in Fig. 35,a shown separately in Fig.37. It corresponds to
taking only terms with derivatives in (341)
h i
Os(1) = q̄λγα1 i∂α2 ...i∂αs q , (357)

Note that all derivatives act on the incoming quark q. They cannot be trasferred to the out-
going quark q̄ since there is no integration over the point of the inserted composite operator.
So all derivatives act on the quark propagator with the momentum l (see Fig. 37) and each
∂α will give just lα . We get the xpression for the diagram

(1)
Z
d4 l b ˆl + q̂ h iˆl b µ 1
V2 = g2 4
t γµ 2
γα1 λ(a)
lα2 ...lαs 2 t γ , (358)
(2π) i (l + q) l (l − p)2

where, as described, [...] means symmetrization in α1 , ...αs and subtraction of trace terms.
In the following these operations will not always be explicitly shown. We are looking for a
divergent part of the form
(1)
h i
divV2 = −g2 ln Λζ̃ (a) λ(a) γα1 pα2 ...pαs . (359)

Obviously it does not depend on q so that we can put q = 0. To simplify we also put p̂ = 0.
In the following we omit flavour matrix Λ(a) common in all formulas.
As we have already known tb tb = 4/3 and

γµ ˆlγα1 ˆlγ µ = 2(γα1 l2 − 2lα1 ˆl).

So we get
8 d4 l γα1 l2 − 2lα1 ˆ
l)
Z
(1)
V2 = g2 lα2 ...lαs . (360)
3 (2π)4 i l4 (l − p)2
Now we use the Feynman parametrization

1 δ(1 − x − y) 1 δ(1 − x − y)
Z Z
= dxdy 2
, 2 =2 xdxdy (361)
ab 0 (xa + yb) a b 0 (xa + yb)3

to write
16 d4 l
Z Z
(1)
V2 = g2 xdxdyδ(1 − x − y) (γα1 l2 − 2lα1 ˆl)lα2 ...lαs D−3 , (362)
3 0 (2π)4 i
 81

where
D = xl2 + y(l − p)2 = l2 − 2ylp = (l − yp)2

In terms of k = l − yp we have l = k + yp and

16 2 1 d4 k 
Z Z 
(1) 2
V2 = g (1 − y)dy γα1 (k + 2ykp) − 2(kα1 + yp α1 )(k̂ + y p̂)
3 0 (2π)4 ik6

×(kα2 + ypα2 )...(kαs + ypαs ). (363)

Here we can drop the term with p̂ = 0. Let us classify the contributions by the terms which
come from the last product.
If we take the term which contains the product of all yp then integration over k will
effectively change
1
γα1 k2 − 2kα1 k̂ → γα1 k2 ,
2
so that in the integrand we obtain a factor
1
γα pα ...pαs k2 (1 − y)y s−1 .
2 1 2
If we retain one of kαj , j ≥ 2 then we have to retain terms odd in k in the first factor

γα1 2y(kp) − 2ypα1 k̂

to be multiplied by
kαj y s−2
Y
pαi .
i=2,i6=j

Integration over k will substitute

1 1
(kp)kαj → k2 pαj k̂kαj → k2 γαj ,
4 4
so that in the first bracket we find
1 2
yk (γα1 pαj − γαj pα1 ),
2
which gives zero after symmetrization.
If we take more terms proportional to k from the last product we obviously will not get
the structure we want, since there will not be sufficient number of p’s. So in the end we
obtain Z 1
(1) 16 2 1
divV2 = g γα1 pα2 ...pαs I(Λ) (1 − y)y s−1 dy, (364)
3 2 0
where
d4 k 1
Z
I(Λ) = 4 4
= 2 ln Λ.
(2π) k 8π
Doing the integral in y we find finally
8 1 1
ζ̃ (1) = − . (365)
3 s(s + 1) 8π 2
Now contributions from diagrams Fig. 35, b and c. They are obviously equal. We consider
Fig. 35, b separately shown in Fig. 38.
 82

Figure 38: Diagram Fig. 35,b

Its contribution is obtained by substituting in ON S one of the operators iDα by gtGα .

We shall have a sum of terms
s h i
Os(2) =
X
q̄λγα1 i∂α3 ...i∂αj gtGα2 ∂αj+1 ...i∂αs q , (366)
j=2

For convenience we denoted the Lorentz index of the term with G as α2 . In fact one has
to symmetrize over all Lorentz indices, so that their particular notation is irrelevant. The
important point is where the exceptional vertex with the Gluon stands, that is the number
of derivative on the left and on the right of gtG, since the derivatives on the left act upon the
product of Gq and on the right only on q. So in the momentum space each right derivative
gives i∂α → pα . Each left derivative gives the sum of the momenta of the quark with
momentum p and the gluon with momentum −p − q + l (the change of sign is due to the
opposite direction in the corresponding propagators). So each left derivative gives i∂α → lα .
Finally note that in the Feynman gauge tGα2 coupled to the iteraction tĜ on the outgoing
quark line generates the gluon propagator and γα2 acting on the incoming quark.
Accordingly we find the contribution
s Z ˆl
(2) d4 l b h i 1
= −g2 b
X
V2 4
t γα1 2 γα2 lα3 ...lα p α ...p αs t . (367)
j=2
(2π) i l j j+1
(l − p)2

Note that at j = 2 there appears no factor lα at all in the integrand. As before we put
q = p̂ = 0. Again tb tb = 4/3. We also have
1
γα1 ˆlγα2 = γα1 (2lα2 − γα2 ˆ
l) = 2γα1 lα2 − gα1 α2 ˆl − [γα1 , γα2 ].
2
We drop the term with gα1 α2 and the commutator will go after symmetrization. So we get
s
8 2h d4 l 1
Z
(2) X i
V2 = − g γα1 pαj+1 ...pαs 4
lα2 lα3 ...lα . (368)
3 j=2
(2π) i j
l (l − p)2
2

Using the Feynman parametrization we find

d4 l 1 1 d4 k
Z Z Z
4
lα2 lα3 ...lαj 2 = dy (kα2 + ypα2 )....(kαj + ypαj ) (369)
(2π) i l (l − p)2 0 (2π)4 ik4
 83

Terms with kα do not give the desired product of pαi , i = 2, ...j so (369) gives
Z 1 1 1
pα2 ...pαj I(Λ) dyy j−1 = pα2 ...pαj 2 ln Λ.
0 j 8π
Thus we get
s
(2) 8 h i 1 1
= − g2 γα1 pα2 ...pαs
X
divV2 2
ln Λ , (370)
3 8π j=2
j

so that s
8 1 X 1
ζ̃ (2) = 2
. (371)
3 8π j=2 j

The same contribution comes from Fig. 35, c. The total contribution from the three
diagrams of Fig. 35 gives
s
4 1  2 1
ζ̃sN S = −
X
− 4 . (372)
3 8π 2 s(s + 1) k=2
k

In the Feynman gauge the quark anomalous dimension has coefficient ζq which can be ex-
tracted from our derivation of the β-function:
4 1 1
ζq = . (373)
3 8π 2 2
From this we find the final coefficient ζsN S
s
4 1  2 1
ζsN S = ζ̃sN S + 2ζq = −
X
1 − + 4 . (374)
3 8π 2 s(s + 1) k=2
k
 84

15 Equations DGLAP
15.1 Evolution in Q2 of the non-singlet structure function
Consider structure function F1N S (x, Q2 ) which is the part of F1 corresonding to a nonsinglet
operator of twist 2 in the operator expansion. We have seen that its moments
Z 1
NS
F1s (Q2 ) ≡ B1s
NS
(Q2 ) = dxxs−1 F1 (x, Q2 ) (375)
0

defined for even s satisfy the evolution equation

Rt
dt′ (γs (g(t′ ))
FsN S) (Q20 e2t , g) = e 0 FsN S (Q20 , g(t)). (376)

At large t the running coupling constant tends to zero and in the limit we have
∂
Q2 F N S (Q2 ) = γsN S (Q2 )F1s
NS
(Q2 ), (377)
∂Q2 1s

where the Q2 dependence of the anomalous dimension enters via the running coupling g(Q2 ).
Using (375) one can construct an equation directly for the function F1N S (x, Q2 ) of two
variables which describes its evolution with Q2 . To this end we note that passing to x = e−ξ
we have Z 1 Z ∞
NS 2 dx s 2
F1s (Q ) = x F1 (x, Q ) = dξe−ξs F1 (x, Q2 ) (378)
0 x 0
N S (Q2 )
and is nothing but the Laplace transform. If one can analytically continue moments F1s
from even integer point to all complex s in the right half-plane then the inverse Laplace
transform gives Z a+∞
NS 2 1
F1 (x, Q ) = dsx−s F1s
NS
(Q2 ), (379)
2πi a−i∞
where the line (a − i∞, a + i∞) should be taken to the right of all singularities of F1s (Q2 ) as
a function of s. Obviously defined in this manner

F1 (x, Q2 ) = 0 for x > 1,

since then one may close the contour to the right.

Now we use the Callan -Symanzik equation for F1s (Q2 ):

∂ 1
Z a+∞
2
Q 2
F1N S = dsx−s γsN S (Q2 )F1s (Q2 )
∂Q 2πi a−i∞

1
Z a+∞ Z 1
s−1
= dsx−s γsN S (Q2 ) dx′ x′ F1 (x′ , Q2 ). (380)
2πi a−i∞ 0
We find an integral over s
1
Z a+∞  x′ s x
ds γsN S (Q2 ) ≡ θ(x′ − x)K N S . (381)
2πi a−i∞ x x′

Then from (380) we obtain an evolution equation for F1N S (x, Q2 ):

∂ 1 dx′ N S  x  N S ′ 2
Z
Q2 F N S (x, Q2 ) = K F (x , Q ). (382)
∂Q2 1 x x′ x′ 1
 85

This is an equation of the DGLAP type.

It is not difficult to find kernel K explicitly. Indeed we have obtained

γsN S (Q2 ) = −g2 (Q2 )ζsN S , (383)

where
s
4 1  2 1
ζsN S =
X
1 − + 4 . (384)
3 8π 2 s(s + 1) k=2
k
In terms of function
d
ψ(z) = ln Γ(z)
dz
we have s
X 1
ψ(s + 1) = 1 − C + , (385)
k=2
k
where C is the Eiler constant. So
4 1 1
Z a+∞  2 
NS 2 2
K (z) = −g (Q ) dsz −s 4C − 3 − + 4ψ(s + 1) . (386)
3 8π 2 2πi a−i∞ s(s + 1)
Let us take z < 1. Then constant terms in the brackets give zero. We have an integral
1
Z a+∞  2 2 
I(z) = dsz −s − + 4ψ(s + 1) . (387)
2πi a−i∞ s+1 s
Function ψ(s + 1) has simple poles at s = −1, −2, −3, ... with residue −1. So we get
∞
4 1 + z2
z n = −2 + 2z −
X
I(z) = −2 + 2z − 4 = −2 . (388)
n=1
1−z 1−z

Adding the coefficient we obtain

g2 (q 2 ) 8 1 + z 2
K N S (z) = , z < 1. (389)
8π 2 3 1 − z
We observe that K has a non-integrable singularity at z = 1. To eliminate it we use the
fact that γ1N S = 0, that is
Z 1
2 ∂
Q dxF1N S (x, Q2 ) = 0. (390)
∂Q2 0
On the other hand, rewriting (382) as

∂
Z 1
Q2 F N S (x, Q2 ) = dx′ dx′′ δ(x − x′ x′′ )F1N S (x′ , Q2 )K N S (x′′ , Q2 ) (391)
∂Q2 1 0

we find
∂
Z 1 Z 1 Z 1
Q2 dxF1N S (x, Q2 ) = dx′ F1N S (x′ , Q2 ) · dx′′ K N S (x′′ , Q2 ), (392)
∂Q2 0 0 0

which gives a condition Z 1

dxK N S (x, Q2 ) = 0. (393)
0
To fulfill it we regularize
1 1
→ (394)
1−z (1 − z)+
 86

by requiring that for arbitrary function f (z)

Z 1 f (z)
Z 1 f (z) − f (1)
dz = dz . (395)
0 (1 − z)+ 0 1−z

Then we find
1 1 + z2 z 1 + z2 − 2
Z Z
dz = = −3/2, (396)
0 (1 − z)+ 0 1−z
which implies that to fulfill condition (393) we have to substitute

1 + z2 1 + z2 3
→ + δ(1 − z). (397)
1−z (1 − z)+ 2

This gives the final expression

g2 (Q2 )  8 1 + z 2 
K N S (z) = + 4δ(1 − z) . (398)
8π 2 3 (1 − z)+
 87

Figure 39: Operator expansion of the singlet F1 (x, Q2 ) as a sum of quark and gluon contri-
butions

15.2 Singlet structure functions. Parton densities

Passing to the singlet structure function we find its moments as a sum of the contributions
q,f G . They are matrix
from the quarks of different flavors and gluon operators: F1s and F1s
elements between the proton state of the two singlet operators (342) and (343):
h i
OsS,f = q̄f γα1 iDα2 ...iDαs qf , (399)

1h i
OsG =Gαα1 iDα2 ...iDαs−1 Gααs . (400)
2
The operator expansion in these operators can be illustrated in Fig 39. Summation goes over
all Lorenz spins s with appropriate Wilson coefficients. So the singlet structure function is a
sum of its quark and gluon parts:

F1singlet(x, Q2 ) = F1q,f (x, Q2 ) + F1G (x, Q2 .

X

The two parts of the singlet structure function measure the quark and gluon densities in the
proton. Therefore one can also denote

F1S,f (x, Q2 ) = qf (x, Q2 ) and F1G (x, Q2 ) = G(x, Q2 ),

q
where the densities qf (x, Q2 ) and G(x, Q2 ) have their moments F1s,f G respectively for
and F1s
spin s. The latter can be denoted as

q,f
F1s (Q2 ) = qf,s (Q2 ) and F1s
G
= Gs (Q2 ).

These two densities form a column which satisfies an equation similar to (377) where, due
to mixing, γ is a 2 × 2 matrix. Correspondingly we can define quark and gluon densities
themselves by the relation similar to (379):

1
Z a+∞
q,G q,G
F1s (x, Q2 ) = dsx−s F1s (Q2 ) (401)
2πi a−i∞

and a similar derivation will bring us to the system of evolution equations for the pair q(x, Q2 )
and G(x, Q2 )

∂ 1 dx′ h SS  x  SG x
Z   i
Q2 qf (x, Q2 ) = K q f (x′
, Q 2
) + K G(x′
, Q 2
) (402)
∂Q2 x x′ x′ x′
 88

∂ 1 dx′ h GS  x  X GG x
Z   i
Q2 G(x, Q2 ) = K q f (x′
, Q 2
) + K G(x′
, Q 2
) , (403)
∂Q2 x x′ x′ f x′

where matrix K is independent of flavor and defined via the matrix of anomalous dimensions
γ (347) and (348).
x Z a+∞  ′ 
′ 1 x s
θ(x − x)K ′ = ds γs (Q2 ). (404)
x 2πi a−i∞ x
Obviously K qq coincides with our earlier K N S . Others are calculated to be

g2 (Q2 )  2 
K qG (z) = z + (1 − z)2
, (405)
8π 2
g2 (q 2 ) 8 1 + (1 − z)2
K Gq (z) = , (406)
8π 2 3 z
g2 (q 2 ) h  z 1−z   2  i
K GG (z) = 12 + + z(1 − z) + 11 − Nf δ(1 − z) . (407)
8π 2 (1 − z)+ z 3
The last formula uses conservation laws for the momentum

γ2qq + γ2Gq = 0, 2Nf γ qG + γ2GG = 0 (408)

from which one finds

∂
Z 1  
Q2 xdx F1q (x, Q2 ) + F1G (x, Q2 ) = 0 (409)
∂Q2 0

The system of equations for quark and gluon densities (402) and (403) is standardly called
the Dokshitzer-Gribov-Lipatov-Altarelly-Parisi (DGLAP) equations.
The DGLAP equations were derived originally in the perturbation theory by summing the
appropriate Feynman diagrams in the so-called leading log approximation. We shall briefly
consider this derivation in the following subsection.
 89

Figure 40: Lowest order diagrams for the deep inelstic scattering on the quark target.

15.3 DGLAP equations from the Feynman diagrams

Originally the DGLAP equations were derived not by the renormalization group technique
but by a direct summation of the Feynman diagrams taking into acccount the dominating
contributions.
Consider the structure functions in the perturbation approach, taking the quark with
momentum p and p2 = 0 as a target (the deep inelastic scattering on the quark). In absence
of the strong integraction we have the Compton effect diagram on the quark (Fig.40),a. In
the next, second order we find three diagrams shown in Fig. 40,b, c, d. Their direct calculation
shows that their contribution contains terms of the order

g2 (c ln Q2 + O(1))

The term with ln Q2 obviously gives the dominating contribution as Q2 → ∞. It is essential

that this contribution comes from the integration region where |k2 | << Q2 . In the Feynman
gauge all three diagrams in Fig. 40,b, c, d. contain this dominating term. However one can
simplify the calculation if one chooses the so-called light-cone gauge in which one requires
that the gluon field is orthogonal to the target quark momentum

(Gp) = 0 instead of (∂G) = 0.

In this particular gauge there are no ghosts. But it is not relativistically invariant, which
considerably complicates renormalization. However this is not so important in our case. The
advantage of the light-cone gauge is that in this gauge the two inteference diagrams Fig.
40,c, d. do not contain terms with ln Q2 , which come only from the single diagram Fig. 40,b.
This property is preserved when one considers diagrams of higher orders, shown in Fig. 41,
a, b, c, where the diagrams b and c illustrtate only some typical ones of these types. The
 90

Figure 41: Higher order diagrams for the deep inelstic scattering on the quark target.

Figure 42: The ladder diagrams for the structure functions expressed in terms of blobs
qf (x, Q2 ) and G(x, Q2 )

leading contribution of the order (g2 ln q 2 )n comes only from the ladder diagrams Fig. 41,a
with n gluon rungs. Also the leading terms come from the integration region in k1 , k2 , ..kn
where |kn2 | << |kn−1
2 |, , , ... << |k 2 | << Q2 , so that the virtual quark momentum grows from
1
below with each interaction. As mentioned it is not simple to perform the renormalization
procedure in the light-cone gauge, but in the end one obtains that each coupng constant g2
for the n-th gluon exchange should be changed to the runnung one g(kn2 ).
The strategy to find the structure function at large Q2 is to sum all these ladder diagrams.
Assume that g2 << 1 but g2 ln Q2 ∼ 1. Then summation of all terms of the order (g2 ln Q2 )n ,
n = 0, 1, 2... will give the correct aymptotic value of the structure function at large Q2 . This
approach is called the leading log approximation.
Including intermediate gluons one has to sum the diagrams shown in Fig. 42. One observes
that they are expressed by two blobs indicated in the pictures, which start either with the
quark lines, blob qf (x, Q2 ), or with the gluon lines,blob G(x, Q2 ) . Distributions qf (x, Q2 )
and G(x, Q2 ) actually give the probabilities to find in the target quark with momentum p and
 91

Figure 43: The pair of coupled Bethe-Salpiter equations for qf (x, Q2 ) and G(x, Q2 )

p− = 0 a quark or gluon respectively with the longitudinal momentum xp+ and all transverse
momenta squared up to Q2 : |k2 | < Q2 . The infrared cutoff |k2 |0 implies that the target
quark is slightly off-shell, so that in fact its transverse momentum is small but different from
zero: |p2⊥ | = |k2 |0 << Q2
These distributions satisfy equations which generalize the well-known Dyson equation to
the case of two particles with the ladder-type interactions (”the Bethe-Salpeter equations”).
Graphically they are illustrated in Fig.43. The target quark plays only the spectator role in
these equations. It can be substituted by any hadron, the proton included. The off-shellness
of the quarks then goes into the average transverse momenta of the quarks and gluons inside
the proton, which is related to its spatial transverse dimension.
Differentiation in ln Q2 leads to the DGLAP equations discussed earlier (402),(403).

15.4 Low-x asymptotic

At low x the dominant contribution comes from the term proportional to 1/z in K GG (z).
Correspondingly the parton densities will be dominated by the gluon one. Retaining only
the gluon density and the leading term in K GG we have the equation

∂ g2 (Q2 ) 1 1
Z
2 2
Q G(x, Q ) = 12 dx′ G(x′ , Q2 ), (410)
∂Q2 8π 2 x x

or
∂ g2 (Q2 ) 1
Z
2 2
Q xG(x, Q ) = 12 dx′ G(x′ , Q2 ). (411)
∂Q2 8π 2 x

Applying operator x∂/∂x and putting

g2 (Q2 ) λ
12 2
= , (412)
8π ln Q2
 92

where λ = 24/(11 − 2/3Nf ), we get

∂ ∂
ln Q2 xG(x, Q2 ) = −λxG(x, Q2 ), (413)
∂ ln x ∂ ln Q2

or, in variables y = − ln x and z = ln ln Q2

∂2
xG = λxG, (414)
∂y∂z
where it is supposed that both y and z are large.
The asymptotic solution of this equation is
√
xG(y, z) = e 4λyz . (415)

In fact s
∂G 1 4λz
= G
∂y 2 y
and s s
∂2G 1 4λz 1 4λy
≃ = λG.
∂zpdy 2 y 2 z
So in the normal variables the gluon density at x → 0 and Q2 → ∞ is given by
p 1
4λ ln ln ln Q2
xG(x, Q2 ) ∼ e x . (416)