Large-N Gauge Theories

LARGE-N GAUGE THEORIES
Lectures presented at the 1999 NATO-ASI on Quantum Geometry in

Akureyri, Iceland.
Preprint: ITEP-TH-80/99, hep-th/0001047
arXiv:hep-th/0001047v1 10 Jan 2000
YURI MAKEENKO
Institute of Theoretical and Experimental Physics,
B. Cheremushkinskaya 25,
117218 Moscow, Russia
email: [email protected]
Abstract. Four pedagogical Lectures at the NATO-ASI on Quantum Ge-

ometry in Akureyri, Iceland, August 1999.
Contents:
1. O(N ) Vector Models
2. Large-N QCD
3. QCD in Loop Space
4. Large-N Reduction
1. Introduction
The large-N limit of Gauge Theory was proposed in 1974 by t Hooft [1]
for quantum chromodynamics (QCD). The dimensionality of the gauge
group SU(Nc ) was used as a parameter, considering Nc as a large num-
ber and performing an expansion in 1/Nc . The motivation was an expan-
sion in the inverse number of field-components N in statistical mechanics
where it is known as the 1/N -expansion, and is a standard method for
non-perturbative investigations.
The expansion of QCD in 1/Nc rearranges diagrams of perturbation
theory in a way which is consistent with a string picture of strong inter-
action, whose phenomenological consequences agree with experiment. The
accuracy of the leading-order term, which is often called large-N QCD or
multicolor QCD, is expected to be of the order of the ratios of meson widths
to their masses, i.e. about 1015%.
2 YURI MAKEENKO
While QCD is simplified in the large-Nc limit, it is still not yet solved.
Generically, it is a problem of infinite matrices, rather than of infinite vec-
tors as in the theory of second-order phase transitions in statistical me-
chanics.
We start these Lectures by showing how the 1/N -expansion works for
the O(N )-vector models, and describing some applications to the four-Fermi
interaction and the nonlinear sigma model. Then we concentrate on the
YangMills theory at large Nc .
The methods described in these Lectures are developed mostly in the
seventies and the beginning of the eighties. They are used over and over
again when discussing a relation between Gauge Theory and String Theory.
The conformal invariance, described in Section 2 for the four-Fermi
interaction, is used in the modern approaches based on the AdS/CFT cor-
respondence [2] (discussed in the lectures by L. Thorlacius at this School).
The notion of the large-N limit, described in Section 3, is widely used, in
particular, in Matrix Theory [3] (discussed in the lectures by W. Taylor
at this School). The loop equation, described in Section 4, has been re-
cently applied [4] for studying the AdS/CFT correspondence. The large-N
reduced models, described in Section 5, are applied to a nonperturbative
formulation of superstring [5] and very recently to noncommutative gauge
theory [6].
The content of my fifth lecture at this School, which was devoted to an
application of the large-N methods to Matrix Theory at finite temperature,
is not included in the text. It is mostly contained in Ref. [7].
2. O(N ) Vector Models
The simplest models, which become solvable in the limit of a large number
of field components, deal with a field which has N components forming
an O(N ) vector in an internal symmetry space. A model of this kind was
first considered by Stanley [8] in statistical mechanics and is known as
the spherical model. The extension to quantum field theory was done by
Wilson [9] both for the four-Fermi and 4 theories.
In the framework of perturbation theory, the four-Fermi interaction
is renormalizable only in d = 2 dimensions and is non-renormalizable
for d > 2. The 1/N -expansion resums perturbation-theory diagrams after
which the four-Fermi interaction becomes renormalizable to each order in
1/N for 2 d < 4. An analogous expansion exists for the nonlinear O(N )
sigma model.
The 1/N expansion of the vector models is associated with a resumma-
tion of Feynman diagrams. A very simple class of diagrams the bubble
graphs survives to the leading order in 1/N . This is why the large-N
LARGE-N GAUGE THEORIES 3
limit of the vector models is solvable. Alternatively, the large-N solution

is nothing but a saddle-point solution in the path-integral approach. The
existence of the saddle point is due to the fact that N is large. This is to be
distinguished from a perturbation-theory saddle point which is due to the
fact that the coupling constant is small. Taking into account fluctuations
around the saddle-point results in the 1/N -expansion of the vector models.
We begin this Section with a description of the 1/N -expansion of the
N -component four-Fermi theory analyzing the bubble graphs. Then we in-
troduce functional methods and construct the 1/N -expansion of the O(N )-
symmetric nonlinear sigma model. At the end we discuss the factorization
in the O(N ) vector models at large N .
2.1. FOUR-FERMI INTERACTION
The action of the O(N )-symmetric four-Fermi interaction in a d-

dimensional Euclidean space is defined by
G
Z
2
d x + m
d

S , = . (1)
2
Here = and = (1 , . . . , N ) is a spinor field which forms an

N -component vector in an internal symmetry space so that
N
X
= i i . (2)
i=1
In d = 2 this model was studied in the large-N limit in Ref. [10] and is
often called the GrossNeveu model.
The dimension of the four-Fermi coupling constant G is dim [G] =
m 2d . For this reason, the perturbation theory for the four-Fermi inter-
action is renormalizable in d = 2 but is non-renormalizable for d > 2 (and,
in particular, in d = 4). This is why the old Fermi theory of weak interac-
tions was replaced by the modern electroweak theory, where the interaction
is mediated by the W and Z bosons.
The action (1) can be equivalently rewritten as
!
2
Z
d x + m +
d

S , , = , (3)
2G
where is an auxiliary field. The two forms of the action, (1) and (3), are
equivalent due to the equation of motion
= G (4)
4 YURI MAKEENKO
which can be derived by varying the action (3) with respect to .

In the path-integral quantization, where the partition function is defined
by Z
Z = DDD e S [,,] (5)

with S , , given by Eq. (3), the action (1) appears after performing
the Gaussian integral over . Therefore, one alternatively gets
Z
Z= DD e S [,] (6)

with S , given by Eq. (1).
The perturbative expansion of the O(N )-symmetric four-Fermi theory
can be conveniently represented using the formulation (3) via the auxiliary
field . Then the diagrams are of the type of those in Yukawa theory, and
resemble the ones for QED with and being an analog of the electron-
positron field and being an analog of the photon field. However, the
auxiliary field (x) does not propagate, since it follows from the action (3)
that
D0 (x y) h (x)(y) iGauss = G (d) (x y) (7)
or
D0 (p) h (p)(p) iGauss = G (8)
in momentum space.
It is convenient to represent the four-Fermi vertex as the sum of two
terms
@
@ @
@ @
@
I
@@ I
@@ I
@
@
=
@
@ @
@ @
@
I
@@ I
@@ I
@
@
(9)
where the empty space inside the vertex is associated with the propaga-
tor (7) (or (8) in momentum space). The relative minus sign makes the
vertex antisymmetric in both incoming and outgoing fermions as is pre-
scribed by the Fermi statistics.
The diagrams that contribute to second order in G for the four-Fermi
vertex are depicted, in these notations, in Figure 1. The O(N ) indices
propagate through the solid lines so that the closed line in the diagram in
Figure 1b corresponds to the sum over the O(N ) indices which results in a
factor of N . Analogous one-loop diagrams for the propagator of the -field
are depicted in Figure 2.
Remark on the one-loop Gell-MannLow function of four-Fermi theory

Evaluating the diagrams in Figure 1 which are logarithmically divergent
in d = 2, and noting that the diagrams in Figure 2 do not contribute to
@
@
I
@
@ @
@
- @
@
I
@ I
@@
@
6 6 @ @

@ @

I
@ - I
@
@
@ @ @
I
@@
a) b) c)
@
- @
@
@ I
@
I
@@ @
@
@
I
@
- @
@
I
@
@ @
d) e)
Figure 1. Diagrams of the second order of perturbation theory for the four-Fermi vertex.
The diagram b) involves the sum over the O(N ) indices.

6 ?
?
6
- -
- -
a) b)
Figure 2. One-loop diagrams for the propagator of the -field. The diagram b) involves
the sum over the O(N ) indices.
the wave-function renormalization of the -field, which emerges to the next

order in G, one gets for the one-loop Gell-MannLow function
(N 1) G2
B (G) = . (10)
2
The four-Fermi theory in 2 dimensions is asymptotically free as was first
noted by Anselm [11] and rediscovered in Ref. [10].
The vanishing of the one-loop Gell-MannLow function in the Gross
Neveu model for N = 1 is related to the same phenomenon in the Thirring
model. The latter model is associated with the vector-like interaction
2
of one species of fermions with being the -matrices in 2 di-
mensions. Since a bispinor has in d = 2 only two components 1 and 2 ,
6 YURI MAKEENKO
@
@

I
@@

@

@
I
@@
Figure 3. Bubble diagram which survives the large-N limit of the O(N ) vector models.
both the vector-like and the scalar-like interaction (1) for N = 1 reduce
to 1 1 2 2 since the square of a Grassmann variable vanishes. Therefore,
these two models coincide. For the Thirring model, the vanishing of the
Gell-MannLow function for any G was shown by Johnson [12] to all loops.
2.2. BUBBLE GRAPHS AS ZEROTH ORDER IN 1/N
The perturbation-theory expansion of the O(N )-symmetric four-Fermi the-

ory contains, in particular, the diagrams of the type depicted in Figure 3
which are called bubble graphs. Since each bubble has a factor of N , the
contribution of the n-bubble graph is Gn+1 N n which is of the order of
Gn+1 N n G (11)
as N since
1
G. (12)
N
Therefore, all the bubble graphs are essential to the leading order in 1/N .
Let us denote

n loops
= G + . . . + G2 + Gn+1 +...

(13)
In fact the wavy line is nothing but the propagator D of the field with
the bubble corrections included. The first term G on the RHS of Eq. (13)
is nothing but the free propagator (8).
Summing the geometric series of the fermion-loop chains on the RHS of
Eq. (13), one gets analytically1

1 dd k sp k + im k + p + im
Z
D1 (p) = N . (14)
G (2)d (k2 + m2 ) ((k + p)2 + m2 )
1
Recall that the free Euclidean fermionic propagator is given by S0 (p) = (ip + m)1
due to Eqs. (3), (5) and the additional minus sign is associated with the fermion loop.
@
@
I
@
@
@
@
I
@
@
@
@
I
@ 6 6
@

@ 6 6

@
I
@
@ @ 6 6

@
I
@
@ @
@
I
@@
Figure 4. Some diagrams of the 1/N -expansion for the O(N ) four-Fermi theory. The
wavy line represents the (infinite) sum of the bubble graphs (13).
This determines the exact propagator of the field at large N . It is O N 1

since the coupling G is included in the definition of the propagator.

The idea is now to change the order of summation of diagrams of per-
turbation theory using 1/N rather than G as the expansion parameter.
Therefore, the zeroth-order propagator of the expansion in 1/N is defined
as the sum over the bubble graphs (13) which is given by Eq. (14). Some of
the diagrams of the new expansion for the four-Fermi vertex are depicted in
Figure 4. The first diagram is proportional to G while the second and third
ones are proportional to G2 or G3 , respectively, and therefore are of order
O(N 1 ) or O(N 2 ) with respect to the first diagram. The perturbation
theory is thus rearranged as the 1/N -expansion.
The general structure of the 1/N -expansion is the same for all vector
models, say, for the N -component nonlinear sigma model which is consid-
ered in Subsection 2.4.
The main advantage of the expansion in 1/N for the four-Fermi inter-
action, over the perturbation theory, is that it is renormalizable in d < 4
while the perturbation-theory expansion in G is renormalizable only in
d = 2. Moreover, the 1/N -expansion of the four-Fermi theory in 2 < d < 4
demonstrates [9] an existence of an ultraviolet-stable fixed point, i.e. a
nontrivial zero of the Gell-MannLow function.
In order to show that the 1/N -expansion of the four-Fermi theory is
renormalizable in 2 d < 4, let us analyze indices of the diagrams of the
1/N -expansion. First of all, we shall get rid of an ultraviolet divergence of
the integral over the d-momentum k in Eq. (14). The divergent part of the
integral is proportional to d2 (logarithmically divergent in d = 2) with
being an ultraviolet cutoff. It can be canceled by choosing
g2 2d
G= , (15)
N
8 YURI MAKEENKO
where g2 is a proper dimensionless constant which is not necessarily positive

since the four-Fermi theory is stable with either sign of G. The power of in
Eq. (15) is consistent with the dimension of G. This prescription works for
2 < d < 4 where there is only one divergent term while another divergency
p2 ln emerges additionally in d = 4. This is why the consideration is
not applicable in d = 4.
The propagator D(p) is therefore finite, and behaves at large momenta
|p| m as
1
D(p) d2 . (16)
|p|
The standard power-counting arguments then show that the only divergent
diagrams appear in the propagators of the and fields, and in the --
three-vertex. These divergencies can be removed by a renormalization of
the coupling g, mass, and wave functions of and .
This completes a demonstration of renormalizability of the 1/N -expan-
sion for the four-Fermi interaction in 2 d < 4. For more detail, see
Ref. [13].
2.3. SCALE AND CONFORMAL INVARIANCE OF FOUR-FERMI THEORY
The coefficient in Eq. (15) can easily be calculated in d = 3. To evaluate

the divergent part of the integral in Eq. (14), we put p = 0 and m = 0.
Remembering that the -matrices are 2 2 matrices in d = 3, we get
d3 k sp kk d3 k 1 1
Z Z Z
=2 = 2 d|k| = . (17)
(2)3 k2 k2 (2)3 k2 2
Note that the integral is linearly divergent in d = 3 and is the cutoff
for the integration over |k|. This divergence can be canceled by choosing G
according to Eq. (15) with g equal to
g = . (18)
To calculate in d = 3 the coefficient of proportionality in Eq. (16), let us

choose G = 2 /N as is prescribed by Eqs. (15), (18) and put in Eq. (14)
m = 0 since we are interested in the asymptotics |p| m. Then the RHS
of Eq. (14) can be rearranged as
" #
d3 k k2 + kp 1 N |p|
Z
1
D (p) = 2N = . (19)
(2)3 k2 (k + p)2 k2 8
The integral is obviously convergent.

Equation (19) (or (16) in d dimensions) is remarkable since it shows
that the scale dimension of the field changes its value from l = d/2 in

-
- - @
R
@@
@
@
R
@
@
@
a) b)
Figure 5. Diagrams for the 1/N -correction to the -field propagator (a) and the
three-vertex (b).
perturbation theory to l = 1 in the zeroth order of the 1/N expansion

(remember that the momentum-space propagator of a field with the scale
dimension l is proportional to |p|2ld ). This appearance of scale invariance
in the 1/N -expansion of the four-Fermi theory at 2 < d < 4 was first
pointed out by Wilson [9] and implies that the Gell-MannLow function
B(g) has a zero at g = g which is given in d = 3 by Eq. (18).
The (logarithmic) anomalous dimensions of the fields , , and of the
-- three-vertex in d = 3 to order 1/N can be found as follows. The
1/N -correction to the propagator of the -field is given by the diagram
depicted in Figure 5a). Since we are interested in an ultraviolet behavior,
we can put again m = 0. Analytically, we have
8i d3 k k + p
Z
S 1 (p) = ip + . (20)
N (2) |k|(k + p)2
3
The (logarithmically) divergent contribution emerges from the domain of

integration |k| |p| so we can expand the integrand in p. The p-indepen-
dent term vanishes after integration over the directions of k so that we
get
" # " #
8 d3 k 1
2 2
Z
1
S (p) = ip 1 + = ip 1 + ln + finite .
N (2)3 |k|3 3 2 N p2
(21)
The diagram, which gives a non-vanishing contribution to the three-
vertex in order 1/N , is depicted in Figure 5b. It reads analytically
8 d3 k (k + p1 )(k + p2 )
Z
(p1 , p2 ) = 1 + , (22)
N (2) |k|(k + p1 )2 (k + p2 )2
3
where p1 and p2 the incoming and outgoing fermion momenta, respectively.

The logarithmic domain is |k| |p|max with |p|max being the largest of |p1 |
10 YURI MAKEENKO
and |p2 |. This gives
2 2
(p1 , p2 ) = 1 ln + finite . (23)
2 N p2max
An analogous calculation of the 1/N correction for the field is a bit
more complicated since involves three two-loop diagrams (see Ref. [14]).
The resulting expression for D1 (p) reads
" !#
|p| 1 2 2
1
N D(p) = 2+ 2+ + 2 2 |p| ln + finite .
g 8 N 3 p2
(24)
The linear divergence is canceled to order 1/N providing g is equal to
1

g = 1 + , (25)
N
which determines g to order 1/N . After this D1 (p) takes the form
!
1 N |p| 16 2
D (p) = 1 2 ln 2 . (26)
8 3 N p
To make all three expressions (21), (23), and (26) finite, we need loga-
rithmic renormalizations of the wave functions of - and -fields and of the
vertex . This can be achieved by multiplying them by the renormalization
constants
2
Zi () = 1 i ln 2 (27)

where stands for a reference mass scale and i are anomalous dimensions.
The index i stands for , , or v for the - and -propagators or the three-
vertex , respectively. We have, therefore, calculated
2
= ,
3 2 N
2
v = 2 ,
N
16
= 2 (28)
3 N
to order 1/N . Due to Eq. (4) coincides with the anomalous dimension
of the composite fields
= . (29)
Note, that
Z2 Zv2 Z = 1 . (30)
This implies that the effective charge is not renormalized and is given by
Eq. (25). Thus, the nontrivial zero of the Gell-MannLow function persists
to order 1/N (and, in fact, to all orders of the 1/N -expansion).
If g is chosen exactly at the critical point g , then the renormalization-
group equations
d ln i
= i g2 , (31)
d
where i stands generically either for vertices or for inverse propagators,
possess the scale invariant solutions
i i (g ) .
2
(32)
For the four-Fermi theory in d = 3, Eq. (32) yields
!
1 p2
S(p) = , (33)
ip 2
!
8 p2
D(p) = , (34)
N |p| 2
!v !
2 p22 p1 p2
(p1 , p2 ) = f , , (35)
p21 p21 p21
where f is an arbitrary function of the dimensionless ratios which is not
determined by scale invariance. The indices here obey the relation
1
v = + (36)
2
which guarantees that Eq. (30), implied by scale invariance, is satisfied.
The indices i are given to order 1/N by Eqs. (28). When expanded in
1/N , Eqs. (33) and (34) obviously reproduces Eqs. (21) and (26). Therefore,
one gets the exponentiation of the logarithms which emerge in the 1/N -
expansion. The calculation of the next terms of the 1/N -expansion for the
indices i is contained in Ref. [15].
Scale invariance implies, in a renormalizable quantum field theory, more
general conformal invariance as is first pointed out in Refs. [16, 17]. The
conformal group in a d-dimensional space-time has (d + 1)(d + 2)/2 param-
eters as is illustrated by Table 1. More about the conformal group can be
found in the lecture by Jackiw [18].
A heuristic proof [16] of the fact that scale invariance implies conformal
invariance is based on the explicit form of the conformal current K , which
is associated with the special conformal transformation, via the energy-
momentum tensor:
K = 2x x x2 . (37)
12 YURI MAKEENKO
Group Transformations # Parameters
d(d1) d(d1)
Lorentz 2
rotations x = x 2
d(d+1)
Poincare + d translations x = x + a 2
d2 +d+2
Weyl + 1 dilatation x = x 2
x x (d+1)(d+2)
Conformal + d special conformal (x )2
= x2
+ 2
TABLE 1. Contents and the number of parameters of groups of space-time

symmetry.
Differentiating, we get
K = 2x , (38)
which is proportional to divergence of the dilatation current. Therefore,
both the dilatation and conformal currents vanish simultaneously when
is traceless which is provided, in turn, by the vanishing of the Gell-
MannLow function.
Conformal invariance completely fixes three-vertices as was first shown
by Polyakov [19] for scalar theories. The proper formula for the four-Fermi
theory reads [20]
( d2 )( d2 v )
(p1 , p2 ) = 2v
(v )
dd k k + p1 k + p2 1
Z
d/2 1+ /2 1+ /2 d2+2
, (39)
2
[(k + p1 ) ] 2
[(k + p2 ) ] |k| /2
where the coefficient in the form of the ratio of the -functions is prescribed
by the normalization (33) and (34) and the indices are related by Eq. (36)
but can be arbitrary otherwise2 .
2
The only restriction 0 is imposed by the KallenLehmann representation of the
propagator while there is no such restriction on since it is a composite field.
Equation (39), which results from conformal invariance, unambiguously

fixes the function f in Eq. (35). In contrast to infinite-dimensional confor-
mal symmetry in d = 2, the conformal group in d > 2 is less restrictive. It
fixes only the tree-point vertex while, say, the four-point vertex remains an
unknown function of two variables.
The integral on the RHS of Eq. (39) looks in d = 3 very much like
that in Eq. (22) and can easily be calculated to the leading order in 1/N
when only the region of integration over large momenta with |k| >
|p|max
max{|p1 |, |p2 |} is essential to this accuracy.
Let us first note that the coefficient in front of the integral is v
1/N , so that one has to peak up the term 1/v in the integral for the
vertex to be of order 1. This term comes from the region of integration with
|k| >
|p|max . Recalling that |p1 p2 | <
|p|max in Euclidean space, one gets
d3 k k + p1 k + p2 1 1
Z
= ,
2 [(k + p1 )2 ]1+ /2 [(k + p2 )2 ]1+ /2 |k|1+2 /2 v (p2max )v
(40)
where Eq. (36) has been used and
!v
2
(p1 , p2 ) = . (41)
p2max
While the integral in Eq. (40) is divergent in the ultraviolet for v < 0, this
divergence disappears after the renormalization.
Equation (23) is reproduced by Eq. (40) when expanding in 1/N . This
dependence of the three-vertex solely on the largest momentum is typical
for logarithmic theories in the ultraviolet region where one can put, say,
p1 = 0 without changing the integral with logarithmic accuracy. This is
valid if the integral is fast convergent in infrared regions which is our case.
Remark on broken scale invariance

Scale (and conformal) invariance at a fixed point g = g holds only for
large momenta |p| m. For smaller values of momenta, scale invariance
is broken by masses. In fact, any dimensional parameter breaks scale
invariance. If the bare coupling g is chosen in the vicinity of g , then scale
invariance holds even in the massless case only for |p| where g(p)
approaches g while it is broken if |p| <
.
14 YURI MAKEENKO
2.4. NONLINEAR SIGMA MODEL
The nonlinear O(N ) sigma model3 in 2 Euclidean dimensions is defined by

the partition function
1
Z
n)2
1
R
d2 x( ~
Z = D~n ~n 22
e 2 (42)
g
where ~n = (n1 , . . . , nN ) is an O(N ) vector. While the action in Eq. (42) is

pure Gaussian, the model is not free due to the constraint
1
~n2 (x) = , (43)
g2
which is imposed on the ~n field via the (functional) delta-function.

The sigma model in d = 2 is sometimes considered as a toy model for
QCD since it possesses:
1) asymptotic freedom [21];
2) instantons for N = 3 [22].
The action in Eq. (42) is N as N but the entropy, i.e. a contribu-
tion from the measure of integration, is also N so that a straightforward
saddle point is not applicable.
To overcome this difficulty, we introduce an auxiliary field u(x), which
is 1 as N , and rewrite the partition function (42) as
R h i
21 n)2 u ~
d2 x ( ~ n2 1
Z Z
g2
Z Du (x) D~n (x) e , (44)

where the contour of integration over u(x) is parallel to imaginary axis.

Doing the Gaussian integration over ~n, we get
Z 1
R
N Sp ln( )
2 +u(x) + d2 x u(x)
Z Du (x) e 2 2g 2 . (45)

The first term in the exponent is nothing but the sum of one-loop diagrams
in 2 dimensions

N X1
Sp ln 2 + u (x) = (46)
2 n n
3
The name comes from elementary particle physics where a nonlinear sigma model in
4 dimensions is used as an effective Lagrangian for describing low-energy scattering of
the Goldstone -mesons.
where the auxiliary field u is denoted again by the wavy line.

Now the path integral over u(x) in Eq. (45) is a typical saddle-point
one: the action N while the entropy 1 since only one integration over
u is left. The saddle-point equation for the nonlinear sigma model
1
N G (x, x; usp ) = 0 (47)
g2
while G is defined by
* +
1
G (x, y; u) = y 2 x . (48)

+ u
The coupling g2 in Eq. (47) is 1/N as is prescribed by the con-

straint (43) which involves a sum over N terms on the LHS. This guaran-
tees that a solution to Eq. (47) exists. Next orders of the 1/N -expansion
for the 2-dimensional sigma model can be constructed analogously to the
previous Subsections.
The 1/N -expansion of the 2-dimensional nonlinear sigma model has
many advantages over perturbation theory, which is usually constructed
solving explicitly the constraint (43), say, choosing
v
u N 1
1u X
nN = t1 g 2 n2a (49)
g a=1
and expanding the square root in g2 . Only N 1 dynamical degrees of free-

dom are left so that the O(N )-symmetry is broken in perturbation theory
down to O(N 1). The particles in perturbation theory are massless (like
Goldstone bosons) and it suffers from infrared divergencies.
On the contrary, the solution to Eq. (47) has the form
4

usp = m2R 2 e Ng 2 , (50)
where is an ultraviolet cutoff. Therefore, all N particles acquires the same
mass mR in the 1/N -expansion so that the O(N ) symmetry is restored. This
appearance of mass is due to dimensional transmutation which says in this
case that the parameter mR rather than the renormalized coupling constant
2 is observable. The emergence of the mass cures the infrared problem.
gR
To show that (50) is a solution to Eq. (47), let us look for a transla-
tionally invariant solution usp (x) m2R . Then Eq. (47) in the momentum
space reads
1
Z d2 k 1 N
Z 2 dk2 N 2
=N 2 = = ln . (51)
g2 2 2
(2) k + mR 4 0 k2 + m2R 4 m2R
16 YURI MAKEENKO
The exponentiation results in Eq. (50).

Equation (51) relates the bare coupling g2 and the cutoff and allows
us to calculate the Gell-MannLow function yielding
dg2 N g4
B(g2 ) = . (52)
d 2
The analogous one-loop perturbation-theory formula for any N reads [21]
(N 2)g4
B(g2 ) = . (53)
2
Thus, the sigma-model is asymptotically free in 2-dimensions for N > 2
which is the origin of the dimensional transmutation. There is no asymp-
totic freedom for N = 2 since O(2) is Abelian.
2.5. LARGE-N FACTORIZATION IN VECTOR MODELS
The fact that a path integral has a saddle point at large N implies a very
important feature of large-N theories the factorization. It is a general
property of the large-N limit and holds not only for the O(N ) vector mod-
els. However, it is useful to illustrate it by these solvable examples.
The factorization at large N holds for averages of singlet operators, for
example
Z 1
R
1 N Sp ln[ ]
2 +u + d2 x u
h u (x1 ) . . . u (xk ) i Z Du e 2 2g 2 u (x1 ) u (xk )

(54)
in the 2-dimensional sigma model.
Since the path integral has a saddle point at some u (x) = usp (x) which
is, in fact, x-independent due to translational invariance, we get to the
leading order in 1/N :
1

h u (x1 ) . . . u (xk ) i = usp (x1 ) . . . usp (xk ) + O , (55)
N
which can be written in the factorized form
1

h u (x1 ) . . . u (xk ) i = h u (x1 ) i . . . h u (xk ) i + O . (56)
N
Therefore, u becomes classical as N in the sense of the 1/N -
expansion. This is an analog of the WKB-expansion in h = 1/N . Quan-
tum corrections are suppressed as 1/N .
We shall return to discussing the large-N factorization in the next Sec-
tion when considering the large-N limit of QCD.
3. Large-N QCD
The method of the 1/N -expansion can be applied to QCD. This was done by
t Hooft [1] using the inverse number of colors for the gauge group SU(Nc )
as an expansion parameter.
For a SU(Nc ) gauge theory without virtual quark loops, the expansion
goes in 1/Nc2 and rearranges diagrams of perturbation theory according
to their topology. The leading order in 1/Nc2 is given by planar diagrams,
which have a topology of a sphere, while the expansion in 1/Nc2 plays the
role of a topological expansion. This reminds an expansion in the string
coupling constant in string models of the strong interaction, which also has
a topological character.
Virtual quark loops can be easily incorporated in the 1/Nc -expansion.
One distinguishes between the t Hooft limit when the number of quark
flavors Nf is fixed as Nc and the Veneziano limit [23] when the
ratio Nf /Nc is fixed as Nc . Virtual quark loops are suppressed in
the t Hooft limit as 1/Nc and lead in the Veneziano limit to the same
topological expansion as dual-resonance models of strong interaction.
The simplification of QCD in the large-Nc limit is due to the fact that
the number of planar graphs grows with the number of vertices only expo-
nentially rather than factorially as do the total number of graphs. Correla-
tors of gauge invariant operators factorize in the large-Nc limit which looks
like the leading-order term of a semiclassical WKB-expansion in 1/Nc .
We begin this Section with a description of the double-line represen-
tation of diagrams of QCD perturbation theory and rearrange it as the
topological expansion in 1/Nc . Then we discuss some properties of the
1/Nc -expansion for a generic matrix-valued field.
3.1. INDEX OR RIBBON GRAPHS
In order to describe the 1/Nc -expansion of the YangMills theory, it is

convenient to represent the gauge field by a Hermitean N N matrix
Aij Aa (x)[ta ]ij .
X
(x) = g (57)
a
Here [ta ]ij are the generators of the gauge group (a = 1, . . . , Nc2 1 for
SU(Nc )) with the normalization
1
tr ta tb = ab . (58)
2
The (Euclidean) action reads
1
Z
S[A] = dd x 2
tr F , (59)
2g2
18 YURI MAKEENKO
where
F = A A i[A , A ] (60)
is the non-Abelian field strength and g is the coupling constant.
The propagator of the matrix field Aij (x) has the form
1 il kj 1 ij kl
D E
Aij kl
(x) A (y) = D (x y) , (61)
Gauss 2 Nc
where we have assumed, as usual, a gauge fixing to define the propagator
in perturbation theory. For instance, one has
g2
D (x y) = (62)
4 (x y)2
2
in the Feynman gauge.

Equation (61) can be immediately derived from the standard formula
D E
Aa (x) Ab (y) = ab D (x y) (63)
Gauss
multiplying by the generators of the SU(Nc ) gauge group according to the

definition (57) and using the completeness condition
Nc2 1
1 il kj 1 ij kl

a ij a kl
X
(t ) (t ) = for SU(Nc ) , (64)
a=1
2 Nc
where the factor of 1/2 is due to the normalization (58) of the generators.
We concentrate in this Section only on the structure of diagrams in the
index space, i.e. the space of the indices associated with the SU(Nc ) group.
We shall not consider, in most cases, space-time structures of diagrams
which are prescribed by Feynmans rules.
Omitting at large Nc the second term in parentheses on the RHS of
Eq. (61), we depict the propagator by the double line
i - l
D E
Aij kl
(x) A (y) g2 il kj = j k . (65)
Gauss
Each line represents the Kronecker delta-symbol and has orientation which
is indicated by arrows. This notation is obviously consistent with the space-
time structure of the propagator which describes a propagation from x to
y.
The arrows are due to the fact that the matrix Aij is Hermitean and
its off-diagonal components are complex conjugate. The independent fields
are, say, the complex fields Aij ii
for i > j and the diagonal real fields A .
The arrow represents the direction of the propagation of the indices of the
complex field Aij ji ij

for i > j while the complex-conjugate one, A = (A ) ,
ii
propagates in the opposite direction. For the real fields A , the arrows are
not essential.
The double-line notation appears generically in all models describing
matrix fields in contrast to vector (in internal symmetry space) fields whose
propagators are depicted by single lines as in the previous Section.
The three-gluon vertex, which is generated by the action (59), is de-
picted in the double-line notations as
j1 i1 i1 j1
. .
.... ......
.
......
. ....
g2 i1 j3 i2 j1 i3 j2 i1 j2 i3 j1 i2 j3

... ....... .. ........
.........
........ .... ...
i2 ................ ............. ............... j3 j2 .............. ............. ............... i3
.. ........ ..... ........
......... ........ ....... .......
....... .. .... ... ...
j2 i3 i2 j3 (66)
where the subscripts 1, 2 or 3 refer to each of the three gluons. The rel-
ative minus sign is due to the commutator in the cubic in A term in the
action (59). The color part of the three-vertex is antisymmetric under in-
terchanging the gluons. The space-time structure, which reads in the mo-
mentum space as
1 2 3 (p1 , p2 , p3 )
= 1 2 (p1 p2 )3 + 2 3 (p2 p3 )1 + 1 3 (p3 p1 )2 , (67)
is antisymmetric as well. We consider all three gluons as incoming so that

their momenta obey p1 + p2 + p3 = 0. The full vertex is symmetric as is
prescribed by Bose statistics.
The color structure in Eq. (66) can alternatively be obtained by multi-
plying the standard vertex
a11a22a33 (p1 , p2 , p3 ) = f a1 a2 a3 1 2 3 (p1 , p2 , p3 ) (68)
by (ta1 )i1 j1 (ta2 )i2 j2 (ta3 )i3 j3 , with f abc being the structure constants of the
SU(Nc ) group, and using the formula
1 i1 j3 i2 j1 i3 j2
f a1 a2 a3 (ta1 )i1 j1 (ta2 )i2 j2 (ta3 )i3 j3 =
i1 j2 i3 j1 i2 j3 ,
2
(69)
which is a consequence of the completeness condition (64).
The four-gluon vertex involves six terms each of them is depicted
by a cross which differ by interchanging of the color indices. We depict
the color structure of the four-gluon vertex for simplicity in the case when
i1 = j2 = i, i2 = j3 = j, i3 = j4 = k, i4 = j1 = l but i, j, k, l take on
20 YURI MAKEENKO
..........................
.......... .
...... .........................................
..... ........ ...... ....
.... ...... .... ...
. ... ...... ... ...
... ..
... ... ... ...
.... .... ... ...
... ... .. ...
... ..
... .. .. .
... ... .... ....
... ... .. ...
... ... . .
.
... ... ... ...
... ...... ... ..
.... ..... ..... ....
...... ......... ...... ....
....... ..................................
................................
Figure 6. Double-line representation of a one-loop diagram for the gluon propagator. The
sum over the Nc indices is associated with the closed index line. The relative contribution
of this diagram is g 2 Nc 1.
different values. Then only the following term is left

l i
.
.... .....
.
l i
.......
.......
.......
.......
g2
k . j
...... ......
.
k j (70)
and there are no deltas on the RHS since the color structure is fixed. In other
words, we pick up only one color structure by equaling indices pairwise.
The diagrams of perturbation theory can now be completely rewritten in
the double-line notation [1]. The simplest one which describes the one-loop
correction to the gluon propagator is depicted in Figure 6.4 This diagram
involves two three-gluon vertices and a sum over the Nc indices which is
associated with the closed index line. Therefore, the relative contribution
of this diagram is g2 Nc .
In order for the large-Nc limit to be nontrivial, the bare coupling con-
stant g2 should satisfy
1
g2 . (71)
Nc
This dependence on Nc is similar to Eqs. (12) and (47) for the vector models
and is prescribed by the asymptotic-freedom formula
24 2
g2 = (72)
11Nc ln (/QCD )
of the pure SU(Nc ) gauge theory.

Thus, the relative contribution of the diagram of Figure 6 is of order
Figure 6 g2 Nc 1 (73)
in the large-Nc limit.

4
Here and in the most figures below the arrows of the index lines are omitted for
simplicity.
The double lines of the diagram in Figure 6 can be viewed as bounding

a piece of a plane. Therefore, these lines represent a two-dimensional object
rather than a one-dimensional one as the single lines do in vector models.
These double-line graphs are often called in mathematics the ribbon graphs
or fatgraphs. We shall see below their connection with Riemann surfaces.
Remark on the U(Nc ) gauge group

As is said above, the second term in the parentheses on the RHS of Eq. (64)
can be omitted at large Nc . Such a completeness condition emerges for the
U(Nc ) group whose generators T A (A = 1, . . . , Nc2 ) are
1
T A = ta , 2N
I
, tr T A T B = AB . (74)
2
They obey the completeness condition
2
Nc ij kl 1 il kj
TA TA
X
= for U(Nc ) . (75)
A=1
2
The point is that elements of both the SU(Nc ) group and the U(Nc ) group
can be represented in the form U = exp iB, where B is a general Hermitean
matrix for U(Nc ) and a traceless Hermitean matrix for SU(Nc ).
Therefore, the double-line representation of perturbation-theory dia-
grams which is described in this Section holds, strictly speaking, only for
the U(Nc ) gauge group. However, the large-Nc limit of both the U(Nc )
group and the SU(Nc ) group is the same.
3.2. PLANAR AND NON-PLANAR GRAPHS
The double-line representation of perturbation theory diagrams in the in-

dex space is very convenient to estimate their orders in 1/Nc . Each gluon
propagator contributes a factor of g2 Each three- or four-gluon vertex con-
tributes a factor of g2 . Each closed index line contributes a factor of Nc .
The order of g in 1/Nc is given by Eq. (71).
Let us consider a typical diagram for the gluon propagator depicted
in Figure 7. It has eight three-gluon vertices and four closed index lines
which coincides with the number of loops. Therefore, the relative order of
this diagram in 1/Nc is
4
Figure 7 g2 Nc 1. (76)
The diagrams of the type in Figure 7, which can be drawn on a sheet

of a paper without crossing any lines, are called the planar diagrams. For
22 YURI MAKEENKO
........................
............
........ ......................... ................
...... ... .. ....... .....
..... ........ ..... ....
............ ... ...
... ..
.
.. ... ... ..
.... .... .. ...
... ..
... ... .. .
.. .
... ... ... .
.. ... .. ..
.. ..
.. .. .. ....
.
.
... ... .
... ...
.... .... .. ..
..... ...... ..... ....
...... ........ ...... ....
....... ........................... ................
.......... .
...............................
Figure 7. Double-line representation of a four-loop diagram for the gluon propagator.

The sum over the Nc indices is associated with each of the four closed index lines whose
number is equal to the number of loops. The contribution of this diagram is g 8 Nc4 1.
.........................
.......... .
....... .........................................
..... ....... ...... ....
.
.... ......... .... ...
... .................. .... . ...
.
. ... .... ... ... ....... ..
..
. ......... . . .... ..... .....
..
... .... .... .... ..................... .. ..
.. ... ......... ... .. .
.... .
.. .... ... ............ ..
... ... .. ............ .. ..
... ... ..
.......... ............. .. ....
.
.
... ... ..
.......... ......... .. ..
.... ... .. ...
... ... .......... ......... ....
... ....... ....
.... ..... .. ..
...... ........ ...... ....
....... ..................... .............
....
................................
.
Figure 8. Double-line representation of a three-loop non-planar diagram for the gluon

propagator. The diagram has six three-gluon vertices but only one closed index line (while
three loops!). The order of this diagram is g 6 Nc 1/Nc2 .
such diagrams, an adding of a loop inevitably results in adding of two three-

gluon (or one four-gluon) vertex. A planar diagram with n2 loops has n2
closed index lines. It is of order
n2
n2 -loop planar diagram g2 Nc 1, (77)
so that all planar diagrams survive in the large-Nc limit.

Let us now consider a non-planar diagram of the type depicted in Fig-
ure 8. This diagram is a three-loop one and has six three-gluon vertices.
The crossing of the two lines in the middle does not correspond to a four-
gluon vertex and is merely due to the fact that the diagram cannot be
drawn on a sheet of a paper without crossing the lines. The diagram has
only one closed index line. The relative order of this diagram in 1/Nc is
1
Figure 8 g6 Nc . (78)
Nc2
It is therefore suppressed at large Nc by 1/Nc2 .
The non-planar diagram in Figure 8 can be drawn without line-crossing
on a surface with one handle which is usually called in mathematics a torus
or the surface of genus one. A plane is then equivalent to a sphere and
has genus zero. Adding a handle to a surface produces a hole according
to mathematical terminology. A general Riemann surface with h holes has
genus h.
The above evaluations of the order of the diagrams in Figures 68 can

now be described by the unique formula
genus
1

genus-h diagram . (79)
Nc2
Thus, the expansion in 1/Nc rearranges perturbation-theory diagrams ac-

cording to their topology [1]. For this reason, it is referred to as the topo-
logical expansion or the genus expansion. The general proof of Eq. (79) for
an arbitrary diagram is given in Subsection 3.3.
Only planar diagrams, which are associated with genus zero, survive
in the large-Nc limit. This class of diagrams is an analog of the bubble
graphs in the vector models. However, the problem of summing the planar
graphs is much more complicated than that of summing the bubble graphs.
Nevertheless, it is simpler than the problem of summing all the graphs,
since the number of the planar graphs with n0 vertices grows at large n0
exponentially [24, 25]
#p (n0 ) # of planar graphs constn0 , (80)
while the number of all the graphs grows with n0 factorially. There is no
dependence in Eq. (80) on the number of external lines of a planar graph
which is assumed to be much less than n0 .
It is instructive to see the difference between the planar diagrams and,
for instance, the ladder diagrams which describe e+ e elastic scattering in
QED. Let the ladder has n rungs. Then there are n! ladder diagrams, but
only one of them is planar. This simple example shows why the number
of planar graphs is much smaller than the number of all graphs, most of
which are non-planar.
We shall discuss in these Lectures what is known about solving the
problem of summing the planar graphs.
Equation (79) holds, strictly speaking, only for the relative order while
the contribution of tree diagrams to a connected n-point Greens function is
(g2 )n1 which is its natural order in 1/Nc . In order to make contributions
of all planar diagrams to be of the same order 1 in the large-Nc limit,
independently of the number of external lines, it is convenient to contract
the Kronecker deltas associated with external lines.
Let us do this in a cyclic order as is depicted in Figure 9 for a generic
connected diagram with three external gluon lines. The extra deltas which
are added to contract the color indices are depicted by the single lines. They
can be viewed as a boundary of the given diagram. The actual size of the
boundary is not essential it can be shrunk to a point. Then a bounded
piece of a plane will be topologically equivalent to a sphere with a puncture.
24 YURI MAKEENKO
..........................................
.............. .........
......... ........
.......
....... .......
..
....... ......
.....
.
...... .....
....
..... ....
....
. ....
.... ...
...
...
. ...
... ...
... .
. ................................... ...
... ......... ............................... ......... ...
... ......... ...... .... ...
.... .
.. ... .
.... ... ..
... .... ............. .. . ..
... .. ... ............ ............. .... ...
. ......... .... .... ..
... .... .... .... ....
.... ... ...
....... ... ... ..
.. .. .. . .
. .
......... .......... ..
. ..
. . .. . ...... . ... . .
. . .
.. ................
... .. ...
. . . . .
. .
... ... ... ... .... ......... .
. . ... ...
... ... .... ... ....... .......... .. .
. ..
... ... ... ......... ......... .. .. . . ..
... ... .... ... .. ... . . .
.... ..... .. . ..
... .... ......... .. .. ...
..... ........ .... ... ...
... ...... .................................. .........
... ........ . ...
... ........... .................. ...
... .
... ...
... ...
... ...
....
.....
.... ..
.... ....
..... .....
...... .....
......
....... ............
........ ...
........... ........
................. ................................
Figure 9. Generic index diagram with n0 = 10 vertices, n1 = 10 gluon propagators,

n2 = 4 closed index lines, and B = 1 boundary. The color indices of the external lines
are contracted by the Kronecker deltas (represented by the single lines) in a cyclic order.
The extra factor of 1/Nc is due to the normalization (81). Its order in 1/Nc is 1/Nc2
in accord with Eq. (79).
I shall prefer to draw planar diagrams in a plane with an extended boundary

(boundaries) rather than in a sphere with a puncture (punctures).
It is clear from the graphic representation that the diagram in Figure 9
is associated with the trace over the color indices of the three-point Greens
function
1
G(3)
1 2 3 (x1 , x2 , x3 ) h tr [A1 (x1 ) A2 (x2 ) A3 (x3 )]i . (81)
Nc
We have introduced here the factor 1/Nc to make G3 of O(1) in the large-
Nc limit. Therefore, the contribution of the diagram in Figure 9 having one
boundary should be divided by Nc .
The extension of Eq. (81) to multi-point Greens functions is obvious:
(n) 1
G1 n (x1 , . . . , xn ) h tr [A1 (x1 ) An (xn )]i . (82)
Nc
The factor 1/Nc , which normalizes the trace, provides the natural normal-
ization
G(0) = 1 (83)
of the averages.
Though the two terms in the index-space representation (66) of the
three-gluon vertex look very similar, their fate in the topological expansion
is quite different. When the color indices are contracted anti-clockwise, the
first term leads to the planar contributions to G(3) , the simplest of which is
p p pp p p p p p p p p p p p p p p p pp p p p p p pp p p pp pp pp pp ppp
pp ppp pp p pp p p pp p
p p p p pp ppppppp . ..... ..... ..... ....................... ..... ..... ..... ..p.ppppp ppp p p p p p
.. ..... p p p
ppp .. ..... .... ..... ..pp p
ppppp ....... ... ... pp
...p.p
p.pp...p.p..... ....p p
.. pp
.............. .....................
........
.. p
......
.. pp
......
pp .......
.....
.....
.. p
... ....
.........
.. p
.... . . ... . . .
...................................
.. pp
.. ... .......
pp ......
........... ....... . ... ..
p p p p
.
p p p p
.
p p p p
.
p p
........ .... ......
ppp... p p p p pp p p p p ppp p ppp ....p p pp

. ... ..... .... ..
ppp ......
. .
... .....
.... .
p
... . .
..p p ppp pppp...

. . ...
.. pppp p p pp
... .. .
ppp .......
.. .. . .
. pp p p p
.. ..
p p p ...
... ..... ...
.. .
.... ....... ..
ppp ................ p p p
..
p
.
p
.
p p
... .
p p
. .
p p
. . .
p p
. . . ..
p
...................
p
...........
p p p
......pp.p.p..p..p p p p p
p p p p
... . . ... . ..
. .. .
ppp .......... ................

... ......... .. . .
..... ........... . ......... .. . .
p
............ .................
.
.... ........... . ........ .. .. ... ...
pp
.
ppp
. ...
. . ...
.......... . . .. . .
.
.... ..
p
....... ....... ..... ...
p
.. .......... ...... .........
pppp ......................... ....................................... .........................
.......
p
....... . ...
......... ..... ....
p pp
... ... .. ... ..... . ..........
pppp p
...... .... .....
pp pp
...... ...... ..
.
......
ppp pp p
.
......
p
........ .
p
.. .
p
. .
. ......
p
............. ...................... ....... .... ...... ........
............ .......... ......................... ...
p pp pp ppp ...... p
.......
pppp pp pp
...............
.. .
... ...... ..................................................
pppppppppppp......
ppp.p.pppppppppppppppppppppppp..p..p..p...pp ppppppppppppppppp
. . .....
.... ..
... ..
a) b)
Figure 10. Planar a) and non-planar b) contributions of the two color structures in
Eq. (66) for three-gluon vertex to G(3) in the lowest order of perturbation theory.
depicted in Figure 10a. The anti-clockwise contraction of the color indices

in the second term leads to a non-planar graph in Figure 10b which can be
drawn without crossing of lines only on a torus. Therefore, the two color
structures of the three-gluon vertex contribute to different orders of the
topological expansion. The same is true for the four-gluon vertex.
Remark on oriented Riemann surfaces
Each line of an index graph of the type depicted in Figure 9 is oriented.

This orientation continues along a closed index line while the pairs of index
lines of each double line have opposite orientations. The overall orientation
of the lines is prescribed by the orientation of the external boundary which
we choose to be, say, anti-clockwise like in Figure 10a. Since the lines are
oriented, the faces of the Riemann surface associated with a given graph
are oriented too all in the same way anti-clockwise. Vice versa, such
an orientation of the Riemann surfaces unambiguously fixes the orientation
of all the index lines. This is the reason why we omit the arrows associated
with the orientation of the index lines: their directions are obvious.
Remark on cyclic-ordered Greens functions
The cyclic-ordered Greens functions (82) naturally arise in the expansion of

the trace of the path-ordered non-Abelian phase factor for a closed contour.
26 YURI MAKEENKO
................................. ..............................
.............
........ ...... .......
.......
...... .... ..... ......
.... ... .... ....
....... ............ ... ...
.... ............. .......... .... .. ...
... .... ... .
... . ... .. .
.. ........ .... .... ... ...
. .
.... .... . . .
... .... ... ............. .. ..
.. ...
... ........... ....
.. .....
.. .... ..
... .. . ... ... ...
... ........................ .. ... ...
... ..
.
...... ........ ...
. ... ..
... . .
. .... ......... . ..
... .
.
. .. . ......... . .
.
... ............ ........ ..... ...
... .
...
....... .... .......... ... ...
........ .. ..... ....
......
........ . ......... .......
......... ...
.......
................................ .......................................
a) b)
Figure 11. Example of a) connected and b) disconnected planar graphs.
One gets
1
H

tr P e i dx A (x)
Nc
Zx1 xZn1
I
(n)
n
dx1 1 dx2 2 dxnn G1 n (x1 , . . . , xn ) .
X
= i ... (84)
n=0 x1 x1
The reason is because the ordering along a closed path implies the cyclic-
ordering in the index space.
Remark on generating functionals for planar graphs

By connected or disconnected planar graphs we mean, respectively, the
graphs which were connected or disconnected before the contraction of the
color indices as is illustrated by Figure 11. The graph in Figure 11a is
connected planar while the graph in Figure 11b is disconnected planar.
The usual exponential relation between the generating functionals W [J]
and Z [J] for connected graphs and all graphs, does not hold for the planar
graphs. The reason is that an exponentiation of such a connected planar
diagram for the cyclic-ordered Greens functions (82) can give disconnected
non-planar ones.
The generating functionals for all and connected planar graphs can
be constructed [26] by means of introducing non-commutative sources
j (x). Non-commutative means that there is no way to transform
j 1 (x1 )j 2 (x2 ) into j 2 (x2 )j 1 (x1 ). This non-commutativity of the sources
reflects the cyclic-ordered structure of the Greens functions (82) which
possess only cyclic symmetry.
Using the short-hand notations
XZ
jA dd x j (x)A (x) , (85)

....
=
... ... ...
Figure 12. Graphic derivation of Eq. (89): Z [j] is denoted by an empty box, W [j] is
denoted by a shadow box, j is denoted by a filled circle. By picking a leftmost external line
of a planar graph, we end up with a connected planar graph, whose remaining external
lines are somewhere to the right interspersed by disconnected planar graphs. It is evident
that jZ [j] plays the role of a new source for the connected planar graph. If we instead
pick up the rightmost external line, we get the inverse order Z [j] j, which results in
Eq. (90).
where the symbol includes the sum over the d-vector (or whatever avail-
able) indices except for the color ones, we write down the definitions of the
generating functionals for all planar and connected planar graphs, respec-
tively, as

1

tr (j A)n
X
Z [j] (86)
n=0
Nc
and
1

n
X
W [j] tr (j A) . (87)
n=0
Nc conn
The planar contribution to the Greens functions (82) and their con-
nected counterparts can be obtained, respectively, from the generating func-
tionals Z [j] and W [j] applying the non-commutative derivative which is
defined by

j (y) f (j) = (d) (x y) f (j) , (88)
j (x)
where f is an arbitrary function of js. In other words the derivative picks
up only the leftmost variable.
The relation which replaces the usual one for planar graphs is
Z [j] = W [jZ [j]] , (89)
while the cyclic symmetry says
W [jZ [j]] = W [Z [j] j] . (90)
A graphic derivation of Eqs. (89) and (90) is given in Figure 12. In other
words, given W [j], one should construct an inverse function as the solution
to the equation
j (x) = J (x)W [j] , (91)
28 YURI MAKEENKO
after which Eq. (89) says

Z [j] = W [J] . (92)
More about this approach to the generating functionals for planar graphs
can be found in Ref. [27].
An iterative solution to Eq. (91) for the Gaussian case can be easily
found. In the Gaussian case, only G(2) is nonvanishing which yields
W [j] = 1 g2 j D j , (93)
where the propagator D is given by Eq. (62). Using Eq. (89), we get ex-
plicitly
Z
2
Z [j] = 1 g dd x dd y D (x y) j (x) Z [j] j (y) Z [j] . (94)
While this equation for Z [j] is quadratic, its solution can be written only
as a continued fraction due to the non-commutative nature of the variables.
In order to find it, we rewrite Eq. (94) as
1
Z [j] = , (95)
g2 dd x dd y D (x
R
1+ y) j (x) Z [j] j (y)
whose iterative solution reads [26]
1
Z [j] = . (96)
D
1 + g2 j j
D
1 + g2 j j
D
1 + g2 j j
..
.
3.3. TOPOLOGICAL EXPANSION AND QUARK LOOPS
It is easy to incorporate quarks in the topological expansion. A quark field

belongs to the fundamental representation of the gauge group SU(Nc ) and
its propagator is represented by a single line
i -

i j ij = j . (97)
The arrow indicates, as usual, the direction of propagation of a (complex)

field . We shall omit these arrows for simplicity.
The diagram for the gluon propagator which involves one quark loop is
depicted in Figure 13a. It has two three gluon vertices and no closed index
.......................... ..........................
.......... ....... .......... .
...... ..... ...... ......... ...........................
..... .... ..... ....... ...... ....
.... ... .... ...... .... ...
.... ... . ... ...... ... ...
..
. ... ... ... ... ...
....
...
... .... .... ......... .......... ... ...
... ..
. ...
... .. ... ... .... .. .. ...
... ..
. . ..
... .. ... .. .. .
.. .. .
... ... ... .... ...
. ...
.... ....
... ... ... ....... ...... . . ...
... ... ... ... . .. .
.
... ... ... ... ... ...
... ... ... ..... ... ..
.... ... .... ...... .... ....
......
....... .....
..
.. ...... ............ .........................
...................................... ........ .. ... ........
............................
a) b)
Figure 13. Diagrams for the gluon propagator with a quark loop which is represented
by the single lines. The diagram a) involves one quark loop and has no closed index lines
so that its order is g 2 1/Nc . The diagram b) involves three loops one of which is a
quark loop. Its relative order is g 6 Nc2 1/Nc .
lines so that its order in 1/Nc is
1
Figure 13a g2 . (98)
Nc
Analogously, the relative order of a more complicated tree-loop diagram in
Figure 13b, which involves one quark loop and two closed index lines, is
1
Figure 13b g6 Nc2 . (99)
Nc
It is evident from this consideration that quark loops are not accom-
panied by closed index lines. One should add a closed index line for each
quark loop in order for a given diagram with L quark loops to have the
same double-line representation as for pure gluon diagrams. Therefore,
given Eq. (79), diagrams with L quark loops are suppressed at large Nc
by
1 L+2genus

L quark loops . (100)
Nc
The single-line representation of the quark loops is similar to the one of
the external boundary in Figure 9. Moreover, such a diagram emerges when
one calculates perturbative gluon corrections to the vacuum expectation
value of the quark operator
1
O = , (101)
Nc
where stands for one of the combinations of the gamma-matrices:
1
= I, 5 , , i 5 , = [ , ] , . . . . (102)
2i
30 YURI MAKEENKO
The factor of 1/Nc is introduced in (101) to make it O(1) in the large-

Nc limit. Therefore, the external boundary can be viewed as a single line
associated with valence quarks. The difference between virtual quark loops
and external boundaries is that each of the latter comes along with the
factor of 1/Nc due to the definitions (82) and (101).
In order to prove Eqs. (79) and its quark counterpart (100), let us
(3)
consider a generic diagram in the index space which has n0 three-point
(4)
vertices (either three-gluon or quark-gluon ones), n0 four-gluon vertices,
n1 propagators (either gluon or quark ones), n2 closed index lines, L virtual
quark loops and B external boundaries. Its order in 1/Nc is
1 (3) (4)
2 n1 n0 n0
(g ) Ncn2 Ncn2 n1 +n0 B (103)
NcB
(3) (4)
where the total number of vertices n0 = n0 + n0 is introduced. The extra
factor of 1/NcB is due to the extra normalization factor of 1/Nc in operators
associated with external boundaries.
The exponent on the RHS of Eq. (103) can be expressed via the Euler
characteristics of a given graph of genus h. Let us first mention that a
proper Riemann surface, which is associated with a given graph, is open
and has B + L boundaries (represented by single lines). This surface can be
closed by attaching a cap to each boundary. The single lines then become
double lines together with the lines of the boundary of each cap. We have
already considered this procedure when deducing Eq. (100) from Eq. (79).
The number of faces for a closed Riemann surface constructed in such
a manner is n2 + L + B, while the number of edges and vertices are n1 and
n0 , respectively. Eulers theorem says that
2 2h = n2 + L + B n1 + n0 . (104)
Therefore the RHS of Eq. (103) can be rewritten as
Ncn2 n1 +n0 B = Nc22hL2B . (105)
We have thus proven that the order in 1/Nc of a generic graph does not
depend on its order in the coupling constant and is completely expressed
via the genus h and the number of virtual quark loops L and external
boundaries B by
2h+L+2(B1)
1

generic graph . (106)
Nc
For B = 1, we recover Eqs. (79) and (100).

Remark on the order of gauge action

We see from Eq. (82) that the natural variables for the large-Nc limit are the
matrices A which include the factor of g (see Eq. (57)). In these variables,
the action (59) is O(Nc2 ) at large Nc , since g2 is 1/Nc and the trace is
Nc .
This result can be anticipated from the free theory because the kinetic
part of the action involves the sum over Nc2 1 free gluons. Therefore, the
non-Abelian field strength (60) is 1 for g2 1/Nc .
The fact that the action is O(Nc2 ) in the large-Nc limit is a generic prop-
erty of the models describing matrix fields. It will be crucial for developing
saddle-point approaches at large Nc which are considered below.
Remark on phenomenology of multicolor QCD

While Nc = 3 in the real world, there are phenomenological indications that
1/Nc may be considered as a small parameter. We have already mentioned
some of them in the text the simplest one is that the ratio of the -
meson width to its mass, which is 1/Nc , is small. Considering 1/Nc
as a small parameter immediately leads to qualitative phenomenological
consequences which are preserved by the planar diagrams associated with
multicolor QCD, but are violated by the non-planar diagrams.
The most important consequence is the relation of the 1/Nc -expansion
to the topological expansion in the dual-resonance model of hadrons. Vast
properties of hadrons are explained by the dual-resonance model. A very
clear physical picture behind this model is that hadrons are excitations of
a string with quarks at the ends.
I shall briefly list some consequences of multicolor QCD:
1) The naive quark model of hadrons emerges at Nc = . Hadrons
are built out of (valence or constituent) quark and antiquark q q, while
exotic states like qq q q do not appear.
2) The partial width of decay of the -meson, which is built out of ss
(the strange quark and antiquark), into K + K is 1/Nc , while that
into + 0 is 1/Nc2 . This explains Zweigs rule. The masses of the
- and -mesons are degenerate at Nc = .
3) The coupling constant of meson-meson interaction is small at large Nc .
4) The widths of glueballs are 1/Nc2 , i.e. they should be even narrower
than mesons built out of quarks. The glueballs do not interact or mix
with mesons at Nc = .
All these hadron properties (except the last one) approximately agree
with experiment, and were well-known even before 1974 when multicolor
QCD was introduced. Glueballs are not yet detected experimentally (maybe
because of their property listed in the item 4).
32 YURI MAKEENKO
3.4. LARGE-NC FACTORIZATION
The vacuum expectation values of several colorless or white operators,

which are singlets with respect to the gauge group, factorize in the large-
Nc limit of QCD (or other matrix models). This property is similar to that
already discussed in Subsection 2.5 for the vector models.
The simplest gauge-invariant operator in a pure SU(Nc ) gauge theory
is the square of the non-Abelian field strength:
1 2
O (x) = tr F (x) . (107)
Nc
The normalizing factor is the same as in Eqs. (81), (82), which provides the
natural normalization
* +
1 g2 a

2 a
tr F (x) = F (x) F (x) 1. (108)
Nc 2Nc
The contribution of all planar graphs to the average on the LHS of Eq. (108)
is of order 1 in accord with the general formula (106) for B = 1.
In order to verify the factorization in the large-Nc limit, let us consider
the index space diagrams for the average of the product of two colorless op-
erators O (x1 ) and O (x2 ) given by (107). It involves a factorized part when
gluons are emitted and absorbed by the same operators. The contribution
of the factorized part is or order 1 as above.
Alternatively, the connected correlator of the two operators is associated
with the general formula (106) for two boundaries B = 2. Its contribution
is suppressed by 1/Nc2 in the large-Nc limit. For this correlator, at least
one gluon line is emitted and absorbed by different operators O (x1 ) and
O (x2 ). Notice, that these graphs themselves are planar, while the suppres-
sion comes from the number of boundaries.
This example illustrates the general property that only (planar) dia-
grams with gluon lines emitted and absorbed by the same operators survive
as Nc . Since correlations between the colorless operators O (x1 ) and
O (x2 ) are of order 1/Nc2 , the factorization property holds as Nc :
1 1

tr F 2 (x1 ) tr F 2 (x2 )
Nc Nc
1 1 1

2 2
= tr F (x1 ) tr F (x2 ) + O . (109)
Nc Nc Nc2
For a general set of gauge-invariant operators O1 , . . . , On , the factor-
ization property can be represented by
1

h O1 On i = h O1 i h On i + O . (110)
Nc2
This is analogous to Eq. (56) for the vector models.

The factorization in large-Nc QCD was first discovered by A.A. Migdal
in the late seventies. An important observation that the factorization im-
plies a semiclassical nature of the large-Nc limit of QCD was done by Wit-
ten [28]. We shall discuss this in the next two Subsections.
The factorization property also holds for gauge-invariant operators con-
structed from quarks like in Eq. (101) as a consequence of Eq. (106).
Remark on factorization beyond perturbation theory

The large-Nc factorization (110) has been shown above to all orders of per-
turbation theory. It can be also verified at all orders of the strong coupling
expansion in the SU(Nc ) lattice gauge theory. A non-perturbative proof
of the factorization will be given in the next Section by using quantum
equations of motion (the loop equations).
3.5. THE MASTER FIELD
The large-Nc factorization in QCD assumes that gauge-invariant objects

behave as c-numbers, rather than as operators. Likewise the vector models,
this suggests that the path integral is dominated by a saddle point.
We already saw in Subsection 2.5 that the factorization in the vector
models does not mean that the fundamental field itself, for instance ~n in
the sigma-model, becomes classical. It is the case, instead, for a singlet
composite field.
We are now going to apply a similar idea to the YangMills theory
whose partition function reads
Z
Z= DAa e S . (111)
The action, Nc2 , is large as Nc , but the entropy is also Nc2 due
to the Nc2 1 integrations over Aa :
2
DAa e Nc . (112)
Consequently, the saddle-point equation of the large-Nc YangMills theory

is not the classical one.
The idea is to rewrite the path integral over A for the YangMills
theory as that over a colorless composite field [A], likewise it was done
in Subsection 2.4 for the sigma-model. The expected new path-integral
representation of the partition function (111) would be something like
1
Z
2
Z D [A]
e Nc S[] . (113)
Aa

34 YURI MAKEENKO
The Jacobian
[A] 2
a
e Nc J[] (114)
A
in Eq. (113) is related to the old entropy factor, so that J [] 1 in the
large-Nc limit.
The original partition function (111) can be then rewritten as
Z
2 2
Z D e Nc J[]Nc S[] , (115)
where S [] represents the YangMills action in the new variables. The new
entropy factor D is O(1) because the variable [A] is a color singlet.
The large parameter Nc enters Eq. (115) only in the exponent. Therefore,
the saddle-point equation can be immediately written:
S J
= . (116)

Remembering that is a functional of A : [A] , we rewrite the
saddle-point equation (116) as
S J
a
= ( F )a = . (117)
A Aa
It differs from the classical YangMills equation by the term on the RHS
coming from the Jacobian (114).
Given J [] which depends on the precise from of the variable [A],
Eq. (117) has a solution
A (x) = Acl
(x) . (118)
Let us first assume that there exists only one solution to Eq. (117). Then
the path integral is saturated by a single configuration (118), so that the
vacuum expectation values of gauge-invariant operators are given by their
values at this configuration:

h O i = O Acl
(x) . (119)
The factorization property (110) will obviously be satisfied.

An existence of such a classical field configuration in multicolor QCD
was conjectured by Witten [28]. It was discussed in the lectures by Cole-
man [29] who called it the master field. Equation (117) which determines
the master field is often referred to as the master-field equation.
A subtle point with the master field is that a solution to Eq. (117) is de-
termined only up to a gauge transformation. To preserve gauge invariance,
it is more reasonable to speak about the whole gauge orbit as a solution of
Eq. (117). However, this will not change Eq. (119) since the operator O is
gauge invariant.
The conjecture about an existence of the master field has surprisingly

rich consequences. Since vacuum expectation values are Poincare invari-
ant, the RHS of Eq. (119) does. This implies that Acl (x) must itself be
Poincare invariant up to a gauge transformation: a change of Acl (x) under
translations or rotations can be compensated by a gauge transformation.
Moreover, there must exist a gauge in which Acl (x) is space-time indepen-
dent: Acl
(x) = A cl (0) . In this gauge, rotations must be equivalent to a

global gauge transformation, so that Acl (0) transforms as a Lorentz vector.
In fact, the idea about such a master field in multicolor QCD may
be incorrect as was pointed out by Haan [30]. The conjecture about an
existence of only one solution to the master-field equation (117) seems to
too strong. If several solutions exists, one needs an additional averaging
over these solutions. This is a very delicate matter, since this additional
averaging must still preserve the factorization property. One might better
think about this situation as if Acl
(0) would be an operator in some Hilbert
space rather than a c-valued function. Such an operator-valued master field
is sometimes called the master field in the weak sense, while the above
conjecture about a single classical configuration of the gauge field, which
saturates the path integral, is called the master field in the strong sense.
The concept of the master field is rather vague until a precise form of
the composite field [A], and consequently the Jacobian [A] that enters
Eq. (117), is not defined. However, what is important is that the master
field (in the weak sense) is space-time independent. This looks like a simpli-
fication of the problem of solving large-Nc QCD. A Hilbert space, in which
the operator Acl (0) acts, should be specified by [A]. We shall consider in
the next Subsection a realization of these ideas for the case of [A] given
by the trace of the non-Abelian phase factor for closed contours.
Remark on non-commutative probability theory
An adequate mathematical language for describing the master field in mul-

ticolor QCD (and, generically, in matrix models at large Nc ) was found by
I. Singer in 1994. It is based on the concept of free random variables of
non-commutative probability theory, introduced by Voiculescu [31]. How
to describe the master field in this language and some other applications of
non-commutative free random variables to the problems of planar quantum
field theory are discussed in Refs. [32, 33].
36 YURI MAKEENKO
3.6. 1/NC AS SEMICLASSICAL EXPANSION
A natural candidate for the composite operator [A] from the previous
Subsection is given by the trace of the non-Abelian phase factor for closed
contours the Wilson loop. It is labeled by the loop C in the same sense
as the field A (x) is labeled by the point x, so we shall use the notation
1 H
(C) [A] = tr P e i C dx A (x) . (120)
Nc
Nobody up to now managed to reformulate QCD at finite Nc in terms
of (C) in the language of path integral. This is due to the fact that self-
intersecting loops are not independent (they are related by the so-called
Mandelstam relations [34]), and the Jacobian is huge. The reformulation
was done [35] in the language of SchwingerDyson or loop equations which
will be described in the next Section.
SchwingerDyson equations are a convenient way of performing the
semiclassical expansion, which is an alternative to the path integral. Let
us illustrate an idea how to do this by an example of the 3 theory. The
RHS of the SchwingerDyson equations is proportional to the Plancks con-
stant h. In the semiclassical limit h 0, we get
D 2 E
12 + m2 h (x1 ) . . . (xn )i + (x1 ) . . . (xn ) = 0 , (121)
2
whose solution is of the factorized form
h (x1 ) . . . (xn )i = h (x1 )i . . . h (xn )i + O (h) (122)
provided that
h (x)i cl (x) (123)
obeys
2
2 + m2 cl (x) +
(x) = 0 . (124)
2 cl
Equation (124) is nothing but the classical equation of motion for the
3
theory, which specifies extrema of the action entering the path integral.
Thus, we have reproduced, using the SchwingerDyson equations, the well-
known fact that the path integral is dominated by a classical solution as
h 0. It is also clear how to perform the semiclassical expansion in h in
the language of the SchwingerDyson equations: one should solve them by
iterations.
The reformulation of multicolor QCD in terms of the loop functionals
(C) is, in a sense, a realization of the idea of the master field in the weak
sense, when the master field acts as an operator in the space of loops.
Remark on the large-Nc limit as statistical averaging

There is yet another, pure statistical, explanation why the large-Nc limit is a
semiclassical limit for the collective variables (C). The matrix U ij [Cxx ],
that describes the parallel transport along a closed contour Cxx , can be
reduced by the gauge transformation to

U [Cxx ] = [Cxx ] diag e ig1 (C) , . . . , e igNc (C) [Cxx ] . (125)
Then (C) reads

Nc
1 X
(C) = e igj (C) . (126)
Nc j=1
The phases j (C) are gauge invariant and normalized so that j (C) 1
as Nc . For simplicity we omit below all the indices (including space
ones) except color.
The commutator of s can be estimated using the representation (126).
Since [i , j ] ij , one gets
1 1
(C) , C g2

2 (127)
Nc Nc
in the limit (71), i.e. the commutator can be neglected as Nc , and

the field (C) becomes classical.
Note that the commutator (127) is of order 1/Nc2 . One factor 1/Nc is
because of g in the definition (126) of (C), while the other has a deep
reason. Let us image the summation over j in Eq. (126) as some statistical
averaging. It is well-known in statistics that such averages weakly fluctuate
as Nc , so that the dispersion is of order 1/Nc . It is the factor which
emerges in the commutator (127).
We see that the factorization is valid only for the gauge-invariant quanti-
ties which involve the averaging over the color indices, like that in Eq. (126).
There is no reason to expect factorization for gauge invariants which do not
involve this averaging, for instance for the phases j (C). Moreover, their
commutator is 1, so that j (C)s strongly fluctuate even at Nc = .
An explicit example of such strongly fluctuating gauge-invariant quantities
was first constructed in Ref. [30].
The resume to this Remark is that the factorization is due to the ad-
ditional statistical averaging in the large-Nc limit. There is no reason to
assume an existence of the master field in the strong sense in order to
explain the factorization.
38 YURI MAKEENKO
4. QCD in Loop Space

QCD can be entirely reformulated in terms of the colorless composite field
(C) the trace of the Wilson loop for closed contours. This fact involves
two main steps:
i) All the observables are expressed via (C).
ii) Dynamics is entirely reformulated in terms of (C).
This approach is especially useful in the large-Nc limit where everything
is expressed via the vacuum expectation value of (C) the Wilson loop
average. Observables are given by summing the Wilson loop average over
paths with the same weight as in free theory. The Wilson loop average
obeys itself a close functional equation the loop equation.
We begin this Section with presenting the formulas which relate observ-
ables to the Wilson loops. Then we translate quantum equation of motion
of YangMills theory into loop space. We derive the closed equation for the
Wilson loop average as Nc and discuss its various properties, includ-
ing a non-perturbative regularization. Finally, we briefly comment on what
is known about solutions of the loop equation.
4.1. OBSERVABLES IN TERMS OF WILSON LOOPS
All observables in QCD can be expressed via the Wilson loops (C) defined
by Eq. (120). This property was first advocated by Wilson [36] on a lattice.
Calculation of QCD observables can be divided in two steps:
1) Calculation of the Wilson loop averages for arbitrary contours.
2) Summation of the Wilson loop averages over the contours with some
weight depending on a given observable.
At finite Nc , observables are expressed via the n-loop averages
Wn (C1 , . . . , Cn ) = h (C1 ) (Cn ) i , (128)
which are analogous to the n-point Green functions for 3 theory. The
appropriate formulas for the continuum theory can be found in Ref. [37].
Great simplifications occur in these formulas at Nc = , when all ob-
servables are expressed only via the one-loop average
1
H

W (C) = h (C) i tr P e C dx A
i
. (129)
Nc
This is associated with the quenched approximation.
For example, the average of the product of two colorless quark vector
currents (101) is given at large Nc by

X
(x1 ) (x2 ) = J (C) h (C) i , (130)
Cx1 ,x2
..........
....... .............. .............................
..
...
..... ....
...
...
x2 ............ .....
.....
...
...
... ... ... ..
.
. ... .. ..
..
.
.
..
.
.
..
.
...
..
. x2 .
.........
...
..
..
.
.. .. .... ..
... ..
. .
.. .
.
.... ... .... ...
....... .... ....... .....
x1 ...
... .......
.....
..... x1 ..
...
.........
.......
....... x
.....
..................... . ..
. .
.. ....
............................... 3
C C
a) b)
Figure 14. Contours in the sum over paths representing observables: a) in Eq. (130)
and b) in Eq. (131). The contour a) passes two nailed points x1 and x2 . The contour b)
passes three nailed points x1 , x2 , and x3 .
where the sum runs over contours C passing through the points x1 and x2
as is depicted in Figure 14a. An analogous formula for the (connected)
correlators of three quark scalar currents reads

X
(x1 ) (x2 ) (x3 ) conn = J (C) h (C) i , (131)
Cx1 ,x2 ,x3
where the sum runs over contours C passing through the three points x1 ,
x2 , and x3 as is depicted in Figure 14b. A general (connected) correlator
of n quark currents is given by a similar formula with C passing through n
points x1 , . . . , xn (some of them may coincide).
The weights J (C) in Eq. (130) and J (C) in Eq. (131) are completely
determined by free theory. If quarks were scalars rather than spinors, then
we would get
1 2 1
R 2 (t)
J (C) = e 2 m 2 0
dt z
= e mL(C) scalar quarks , (132)
where L(C) is the length of the (closed) contour C. For spinor quarks, an ad-
ditional disentangling of the gamma-matrices is needed (ee, e.g., Ref. [38]).
Remark on renormalization of Wilson loops

Perturbation theory for W (C) can be obtained by expanding the path-
ordered exponential in the definition (129) in g (see Eq. (84)) and averaging
over the gluon field A . Because of ultraviolet divergencies, we need a (gauge
invariant) regularization. After such a regularization introduced, the Wilson
loop average for a smooth contour C of the type in Figure 15a reads
2
2 (Nc 1) L(C)
W (C) = e g 8Nc a Wren (C) , (133)
where a is the cutoff, L(C) is the length of C, and Wren (C) is finite when
expressed via the renormalized charge gR . The exponential factor is due to
40 YURI MAKEENKO
........................ ........................
..... ..... ..... ......
... ... .... ..... ....
.. ... .. ..... ...
.. .. .. .....
... ... ..... .. .
..
..
..
..
.
..
.
...
. ......
... ..

..
.
... .. .. .....
..
.
.. .. .... .....
. . ... ...
.
... ...
. ... .
.....
. ....
... .... ... .....
..... .
..
.. ..... ..
.. .....
... ...... .... ......
............................. ............................
a) b)
Figure 15. Examples of a) smooth contour and b) contour with a cusp. The tangent
vector to the contour jumps through angle at the cusp.
the renormalization of the mass of a heavy test quark. This factor does not
emerge in the dimensional regularization where d = 4. The multiplicative
renormalization of the smooth Wilson loop was shown in Refs. [39, 40, 41].
If the contour C has a cusp (or cusps) but no self-intersections as is
illustrated by Figure 15b, then W (C) is still multiplicatively renormaliz-
able [42]:
W (C) = Z () Wren (C) , (134)
while the (divergent) factor Z () depends on the cusp angle (or angles)
(or s) and Wren (C) is finite when expressed via the renormalized charge
gR .
4.2. SCHWINGERDYSON EQUATIONS FOR WILSON LOOP
Dynamics of (quantum) YangMills theory is described by the quantum

equation of motion
w.s.
ab b
F (x) = h a
(135)
A (x)
which is understood in the weak sense, i.e. for the averages

D E
ab b
F (x) Q [A] = h Q [A] . (136)
Aa (x)
The standard set of SchwingerDyson equations of YangMills theory
emerges when the functional Q [A] is chosen in the form of the product
of As as in Eq. (82).
Strictly speaking, the last statement is incorrect, since we have not
added, in Eqs. (135) and (136), contributions coming from the variation
of gauge-fixing and ghost terms in the YangMills action. However, these
two contributions are mutually cancelled for gauge-invariant functionals
Q[A]. We shall deal below only with such gauge-invariant functionals (the
Wilson loops). This is why we have not considered the contribution of the
gauge-fixing and ghost terms.
x 6 Cxy
Cyx ?
y
Figure 16. Contours Cyx and Cxy which enter the RHSs of Eqs. (137) and (141).
It is also convenient to use the matrix notation (57), when Eq. (135) for
the Wilson loop takes on the form
* +
1 g2
H H

tr P F (x) e i C d A = tr P e i C d A ,
Nc 2Nc A (x)
(137)
where we have restored the units with h = 1.
The variational derivative on the RHS can be calculated by virtue of
the formula
Aij
(y) 1 ij kl

kl
= (d) (x y) il kj (138)
A (x) Nc
which is a consequence of
Aa (y)
= (d) (x y) ab . (139)
Ab (x)
The second term in the parentheses in Eq. (138) same as in Eq. (64)
is because A is a matrix from the adjoint representation of SU(Nc ).
By using Eq. (138), we get for the variational derivative on RHS of
Eq. (137):
H I
i d A
tr Pe C = i dy (d) (x y)
A (x)
C
1 1 1
R R
i d A i d A
R
i d A
tr P e Cyx
tr P e Cxy
tr P e C . (140)
Nc Nc Nc3
The contours Cyx and Cxy , which are depicted in Figure 16, are the parts
of the loop C: from x to y and from y to x, respectively. They are always
closed due to the presence of the delta-function. It implies that x and y
should be the same points of space but not necessarily of the contour (i.e.
they may be associated with different values of the parameter ).
42 YURI MAKEENKO
We finally rewrite Eq. (137) as

1
H

tr P F (x) e i C d A
Nc
1
I
(d)
= i dy (x y) h (Cyx ) (Cxy ) i 2 h (C) i (141)
N
C
where we have introduced

g 2 Nc
. = (142)
2
Notice that the RHS of Eq. (141) is completely represented via the (closed)
Wilson loops.
4.3. PATH AND AREA DERIVATIVES
As we already mentioned, the RHS of Eq. (141) is completely represented

via the (closed) Wilson loops. It is crucial for the loop-space formulation of
QCD that the LHS of Eq. (141) can also be represented in loop space as
some operator applied to the Wilson loop. To do this we need to develop a
differential calculus in loop space.
Loop space consists of arbitrary continuous closed loops, C. They can
be described in a parametric form by the functions x () L2 ,5 where
0 f and = 1, . . . , d, which take on values in a d-dimensional Eu-
clidean space. The functions x () can be discontinuous, generally speak-
ing, for an arbitrary choice of the parameter . The continuity of the loop
C implies a continuous dependence on parameters of the type of proper
length.
The functions x () L2 which are associated with the elements of
loop space obey the following restrictions:
i) The points = 0 and = f are identified: x (0 ) = x (f ) the
loops are closed.
ii) The functions x () and x () + , with and independent
of , represent the same element of the loop space rotational and
translational invariance.
iii) The functions x () and x ( ) with = f (), f () 0 describe
the same loop reparametrization invariance.
An example of functionals which are defined on the elements of loop space is
the Wilson loop average (129) or, more generally, the n-loop average (128).
The differential calculus in loop space is built out of the path and area
derivatives.
5
Let us remind that L2 stands for the RHilbert space of functions x () whose square

is integrable over the Lebesgue measure: f dx2 () < .
0
The area derivative of a functional F (C) is defined by the difference

F(C) 1 x
>
- x
F F
(x) | |
(143)
where an infinitesimal loop C (x) is attached to a given loop at the point
x in the -plane and | | stands for the area enclosed by the C (x).
For a rectangular loop C (x), one gets = dx dx , where the
symbol implies antisymmetrization.
Analogously, the path derivative is defined by

1 x q - x q
x F(Cxx ) F F
|x |
(144)
where x is an infinitesimal path along which the point x is shifted from
the loop and |x | stands for the length of the x .
These two differential operations are well-defined for so-called function-
als of the Stokes type which satisfy the backtracking condition they do
not change when an appendix passing back and forth is added to the loop
at some point x:

x
F = F . (145)
This condition is equivalent to the Bianchi identity of YangMills theory

and is obviously satisfied by the Wilson loop (129) due to the properties of
the non-Abelian phase factor. Such functionals are known in mathematics
as Chen integrals.
A simple example of the Stokes functional is the area of the minimal
surface, Amin (C). It obviously satisfies Eq. (145). Otherwise, the length
L(C) of the loop C is not a Stokes functional, since the lengths of contours
on the LHS and RHS of Eq. (145) are different.
For the Stokes functionals, the variation on the RHS of Eq. (143) is
proportional to the area enclosed by the infinitesimally small loop C (x)
and does not depend on its shape. Analogously, the variation on the RHS
of Eq. (144) is proportional to the length of the infinitesimal path x and
does not depend on its shape.
44 YURI MAKEENKO
If x is a regular point (like any point of the contour for the func-
tional (129)), the RHS of Eq. (144) vanishes due to the backtracking con-
dition (145). In order for the result to be nonvanishing, the point x should
be a marked (or irregular) point. A simple example of the functional with
a marked point x is
1
R
i d A ()

a [Cxx ] tr ta P e Cxx (146)
Nc
with the SU(Nc ) generator ta inserted in the path-ordered product at the

point x.
The area derivative of the Wilson loop is given by the Mandelstam
formula
1 i
H
d A i H
tr P e C = tr P F (x) e C d A .
i
(147)
(x) Nc Nc
In order to prove it, it is convenient to choose C (x) to be a rectangle in

the -plane and straightforwardly use the definition (143). The sense of
Eq. (147) is very simple: F is a curvature associated with the connection
A .
The functional on the RHS of Eq. (147) has a marked point x, and is of
the type in Eq. (146). When the path derivative acts on such a functional
according to the definition (144), the result reads
1 H 1 H
x tr P B (x) e i C d A = tr P B (x) e i C d A , (148)
Nc Nc
where
B = B i [A , B] (149)
is the covariant derivative in the adjoint representation.

Combining Eqs. (147) and (148), we finally represent the expression on
the LHS of Eq. (137) (or Eq. (141)) as
1 H i H
tr P F (x) e i C d A = x tr P e i C d A , (150)
Nc (x) Nc
i.e. via the action of the path and area derivatives on the Wilson loop. It
is therefore rewritten in loop space.
A resume of the results of this subsection is presented in Table 2 as
a vocabulary for translation of YangMills theory from the language of
ordinary space in the language of loop space.
Ordinary space Loop space
[A] phase factor (C) loop functional

F (x) field strength (x)
area derivative
x covariant derivative x path derivative
F =0 Bianchi identity Stokes functionals
F Schwinger-Dyson Loop
= /A equations equations
TABLE 2. Vocabulary for translation of YangMills theory from

ordinary space in loop space.
Remark on Bianchi identity for Stokes functionals

The backtracking relation (145) can be equivalently represented as

x (C) = 0 , (151)
(x)
by choosing the appendix in Eq. (145) to be an infinitesimal straight line
in the -direction and geometrically applying the Stokes theorem. Using
Eqs. (147) and (148), Eq. (151) can in turn be rewritten as
1 H
tr P F (x) e i C d A = 0 . (152)
Nc
Therefore, Eq. (151) represents the Bianchi identity in loop space.
Remark on relation to variational derivative

The standard variational derivative, /x (), can be expressed via the
path and area derivatives by the formula
m

xi ( i ) ,
X
= x () + (153)
x () (x()) i=1
46 YURI MAKEENKO
where the sum on the RHS is present for the case of a functional having m
marked (irregular) points xi x(i ). A simplest example of the functional
with m marked points is just a function of m variables x1 , . . . , xm .
By using Eq. (153), the path derivative can be calculated as the limiting
procedure
+0

Z
x() = d . (154)
x ( )
0
The result is obviously nonvanishing only when x is applied to a functional

with x() being a marked point.
It is nontrivial that the area derivative can also be expressed via the
variational derivative [40]:
+0

Z
= d ( )
. (155)
(x()) x ( ) x ()
0
The point is that the six-component quantity, / (x()), is expressed

via the four-component one, /x (), which is possible because the com-
ponents of / (x()) are dependent due to the loop-space Bianchi iden-
tity (151).
4.4. LOOP EQUATIONS
By virtue of Eq. (150), Eq. (141) can be represented completely in loop

space:
D E
x (C)
(x)
1
I
(d)
= dy (x y) (Cyx ) (Cxy ) 2 (C) , (156)
Nc
C
or, using the definitions (128) and (129) of the loop averages, as
1
I
x W (C) = dy (d)
(x y) W2 (Cyx , Cxy ) 2 W (C) .
(x) Nc
C
(157)
This equation is not closed. Having started from W (C), we obtain an-
other quantity, W2 (C1 , C2 ), so that Eq. (157) connects the one-loop av-
erage with a two-loop one. This is similar to the case of the (quantum)
3 -theory, whose SchwingerDyson equations connect the n-point Green
functions with different n. We shall derive this complete set of equations
for the n-loop averages in this Subsection later on.
However, the two-loop average factorizes in the large-Nc limit:

1

W2 (C1 , C2 ) = W (C1 ) W (C2 ) + O , (158)
Nc2
as was discussed in Subsection 3.4. Keeping the constant (defined by
Eq. (142)) fixed in the large-Nc limit as is prescribed by Eq. (71), we get [35]

I
x W (C) = dy (d) (x y) W (Cyx ) W (Cxy ) (159)
(x)
C
as Nc .
Equation (159) is a closed equation for the Wilson loop average in the
large-Nc limit. It is referred to as the loop equation.
To find W (C), Eq. (159) should be solved in the class of Stokes func-
tionals with the initial condition
W (0) = 1 (160)
for loops which are shrunk to points.
The factorization (158) can itself be derived from the chain of loop
equations. Proceeding as before, we get
1 x
Wn (C1 , . . . , Cn )
(x)
1
I
(d)
= dy (x y) Wn+1 (Cxy , Cyx , . . . , Cn ) 2 Wn (C1 , . . . , Cn )
Nc
C1
X 1 I h
+ dy (d) (x y) Wn1 C1 Cj , . . . , Cj , . . . , Cn
j2
Nc2
Cj
i
Wn (C1 , . . . , Cn ) . (161)
Here x belongs to C1 ; C1 Cj stands for the joining of C1 and Cj ; Cj means
that Cj is omitted.
Equation (161) looks like the SchwingerDyson equation for the 3 -
theory. Moreover, the number of colors Nc enters Eq. (161) simply as a
scalar factor Nc2 , likewise Planks constant h enters in the 3 -theory. It is
the major advantage of the use of loop space. What is said in Subsection 3.6
about the semiclassical nature of the 1/Nc -expansion of QCD is explicitly
realized in Eq. (161). Its expansion in 1/Nc is straightforward.
At Nc = , Eq. (161) is simplified to

I
x Wn (C1 , . . . , Cn ) = dy (d) (x y)Wn+1 (Cyx , Cxy , . . . , Cn ).
(x)
C1
(162)
48 YURI MAKEENKO
xn.pppppp p pppp ppp p ppp pp pppppp p p p p p p p p p pp pp pppppppppppppppp

p.p... pp ppp
pp ppp ....... pp p p
ppp ppp ......... ppp
p pp
pp p
.. .
.
ppp
..
ppp pp
..... ..
ppp
..
...... ..
p
.
.......
ppp .. .. .. .. .. ......... (n) .......... ........ ........ ........ ........ ........ .......ppppp x
.......
.. ..........
............... .....
.. .
x1 ..ppp..... ........ ........ ........ ........ ........ ......... G ......... .. .. .. .. .. ppp i

ppp pp
ppp ppp
... ...
.....
p
......... .............
ppp pp
.. ........
ppp ppp
. . .
. .
.
.
... .
p pp p p
..
.......
p p p pp ........ pp p
.
.. .
p p ppp p
.
......
ppppppp .......
p.p.p.ppppppp pppppp p p p p pp p p p p ppp ppp p p p p p p p p p
.
x2 p p p p p p pp p p
Figure 17. Graphic representation of the terms on the RHS of Eq. (164).
This equation possesses a factorized solution
1

Wn (C1 , . . . , Cn ) = W (C1 ) W (Cn ) + O (163)
Nc2
provided W (C) obeys Eq. (159) which plays the role of a classical equa-
tion in the large-Nc limit. Thus, we have given a non-perturbative proof of
the large-Nc factorization of the Wilson loops.
4.5. RELATION TO PLANAR DIAGRAMS
The perturbation-theory expansion of the Wilson loop average can be cal-

culated from Eq. (84) which we represent in the form
I I I
in dx1 1 dx2 2 . . . dxnn
X
W (C) = 1 +
C n=2 C C
(n)
c (1, 2, . . . , n) G1 2 n (x1 , x2 , . . . , xn ) , (164)
where c (1, 2, . . . , n) orders the points x1 , . . ., xn along contour in the cyclic

(n)
order and G1 n is given by Eq. (82). This -function has the meaning of
the propagator of a test heavy particle which lives in the contour C.
We assume, for definitiveness, the dimensional regularization through-
out this Subsection to make all the integrals well-defined.
Each term on the RHS of Eq. (164) can be conveniently represented
by the diagram in Figure 17, where the integration over the contour C is
associated with each point xi lying in the contour C.
These diagrams are analogous to those discussed in Subsection 3.2 with
one external boundary the Wilson loop in the given case. In the large-Nc
a) b) c)
Figure 18. Planar diagrams for W (C): a) of order with gluon propagator, and of order
2 b) with two noninteracting gluons and c) with the three-gluon vertex. Diagrams of
order 2 with one-loop insertions to gluon propagator are not drawn.
limit, only planar diagrams survive. Some of them, which are of the lowest
order in , are depicted in Figure 18.
The large-Nc loop equation (159) describes the sum of the planar dia-
grams. Its iterative solution in reproduces the set of planar diagrams for
W (C) provided the initial condition (160) and some boundary conditions
for asymptotically large contours are imposed.
Equation (164) can be viewed as an ansatz for W (C) with some un-
(n)
known functions G1 n (x1 , . . . , xn ) to be determined by the substitution
into the loop equation. To preserve symmetry properties of W (C), the
functions G(n) must be symmetric under a cyclic permutation of the points
1, . . ., n and depend only on xi xj (translational invariance). A main
advantage of this ansatz is that it automatically corresponds to a Stokes
functional, due to the properties of vector integrals, and the initial condi-
tion (160) is satisfied.
The action of the area and path derivatives on the ansatz (164) is easily
calculable. For instance, the area derivative reads

W (C)
I I
n 1
dx1 . . . dxnn c (1, 2, . . . , n)
X
= i
(z) n=1 C C
h
(n+1)
z z G1 n (z, x1 , . . . , xn )
i
(n+2)
+ ( ) G1 n (z, z, x1 , . . . , xn ) . (165)
The analogy with the Mandelstam formula (147) is obvious.

More about solving the loop equation by the ansatz (164) can be found
in Refs. [37, 43, 44].
50 YURI MAKEENKO
4.6. LOOP-SPACE LAPLACIAN AND REGULARIZATION
The loop equation (159) is not yet entirely formulated in loop space. It is a
d-vector equation whose both sides depend explicitly on the point x which
does not belong to loop space. The fact that we have a d-vector equation
for a scalar quantity means, in particular, that Eq. (159) is overspecified.
A practical difficulty in solving Eq. (159) is that the area and path
derivatives, / (x) and x , which enter the LHS are complicated, gen-
erally speaking, non-commutative operators. They are intimately related
to the YangMills perturbation theory where they correspond to the non-
Abelian field strength F and the covariant derivative . However, it is
not easy to apply these operators to a generic functional W (C) which is
defined on elements of loop space.
A much more convenient form of the loop equation can be obtained
by integrating both sides of Eq. (159) over dx along the same contour C,
which yields

I I I
dx x W (C) = dx dy (d) (x y) W (Cyx )W (Cxy ) .
(x)
C C C
(166)
Now both the operator on the LHS and the functional on the RHS are
scalars without labeled points and are well-defined in loop space. The oper-
ator on the LHS of Eq. (166) can be interpreted as an infinitesimal variation
of elements of loop space.
Equations (159) and (166) are completely equivalent. A proof of equiv-
alence of scalar Eq. (166) and original d-vector Eq. (159) is based on the
important property of Eq. (159) whose both sides are identically annihilated
by the operator x . It is a consequence of the identity
i
F = [F , F ] = 0 (167)
2
in the ordinary space. Due to this property, the vanishing of the contour
integral of some vector is equivalent to vanishing of the vector itself, so that
Eq. (159) can in turn be deduced from Eq. (166).
Equation (166) is associated with the so-called second-order Schwinger
Dyson equation

Z Z
d a w.s.
d x F (x) = h dd x dd y (d) (x y)
Aa (x) (x) Aa (y) Aa
(168)
in the same sense as Eq. (159) is associated with Eq. (135). It is called sec-
ond order since the RHS involves two variational derivatives with respect
to A .
The operator on the LHS of Eq. (166) is a well-defined object in loop

space. When applied to regular functionals which do not have marked
points, it can be represented, using Eqs. (154) and (155), in an equiva-
lent form
Zf +0

I Z
dx x = d d
. (169)
(x) x ( ) x ()
C 0 0
As was first pointed out by Gervais and Neveu [45], this operator is noth-
ing but a functional extension of the Laplace operator, which is known in
mathematics as the Levy operator.6 Equation (166) can be represented in
turn as an (inhomogeneous) functional Laplace equation
I I
W (C) = dx dy (d) (x y) W (Cyx )W (Cxy ) . (170)
C C
We shall refer to this equation as the loop-space Laplace equation.

The form (170) of the loop equation is convenient for a non-perturbative
ultraviolet regularization.
The idea is to start from the regularized version of Eq. (168), replacing
the delta-function on the RHS by the kernel of the regularizing operator:
Reg.
D E
ab (d) (x y) = y Rab x = Rab (d) (x y) (171)

with ab
2 2 /2

Rab = e a , (172)
where is the covariant derivative in the adjoint representation. The
regularized version of Eq. (168) is

Z Z D E
w.s.
dd x F
a
(x) = h dd x dd y y Rab x
.

Aa (x) (x) Ab (y) Aa
(173)
To translate Eq. (173) in loop space, we use the path-integral represen-
tation
Z
1
R a2
dt r 2 (t)
D E h i
y Rab x = Dr(t) e 2 2 tr ta U (ryx )tb U (rxy ) (174)
0
r(0)=x
r(a2 )=y
with Ry
U (ryx ) = P e i x
dr A (r)
, (175)
6
See the book by Levy [46] and a review [47].
52 YURI MAKEENKO
x 6 Cxy
Cyx ? rxy ryx
y
Figure 19. Contours Cyx rxy and Cxy ryx which enter the RHSs of Eqs. (176) and (177).
where the integration is over regulator paths r (t) from x to y whose typical
length is a.
Calculating the variational derivatives on the RHS of Eq. (173), using
Eq. (174) and the completeness condition (64), we get as N :

Z D E I I
dd x dd y y Rab x (C) = dx dy

Ab (y) Aa (x)
C C
Z R a2
12 dtr 2 (t)
Dr(t) e 0 (Cyx rxy ) (Cxy ryx ) , (176)
r(0)=x
r(a2 )=y
where the contours Cyx rxy and Cxy ryx are depicted in Figure 19. Averaging
over the gauge field and using the large-Nc factorization, we arrive at the
regularized loop-space Laplace equation [48]
I I Z
1
R a2
dt r 2 (t)
W (C) = dx dy Dr(t) e 2 0 W (Cyx rxy )W (Cxy ryx )
C C r(0)=x
r(a2 )=y
(177)
which manifestly recovers Eq. (170) when a 0.
The constructed regularization is non-perturbative while perturbatively
reproduces regularized Feynman diagrams. An advantage of this regular-
ization of the loop equation is that the contours Cyx rxy and Cxy ryx on the
RHS of Eq. (177) both are closed and do not have marked points if C does
not have. Therefore, Eq. (177) is written entirely in loop space.
Remark on functional Laplacian

It is worth noting that the representation of the functional Laplacian on
the RHS of Eq. (169), which involves the standard variational derivatives,
is defined for a wider class of functionals than Stokes functionals. It is easier
to deal with the whole operator , rather than separately with the area
and path derivatives.
The functional Laplacian is parametric invariant and possesses a num-

ber of remarkable properties. While a finite-dimensional Laplacian is an
operator of the second order, the functional Laplacian is that of the first
order and satisfies the Leibnitz rule
(U V ) = (U ) V + U (V ) . (178)
The functional Laplacian can be approximated [49] in loop space by a

(second-order) partial differential operator in such a way to preserve these
properties in the continuum limit. This loop-space Laplacian can be in-
verted to determine a Green function G (C, C ) in the form of a sum over
surfaces SC,C connecting two loops which is analogous to the sum-over-
path representation of the Green function of the ordinary Laplacian. The
standard perturbation theory can then be recovered by iterating Eq. (170)
(or its regularized version (177)) in with the Green function of the loop-
space Laplacian.
4.7. SURVEY OF NON-PERTURBATIVE SOLUTIONS
While the loop equations were proposed long ago, not much is known about
their non-perturbative solutions. We briefly list some of the results.
It was shown in Ref. [50] that area law
W (C) h (C)i e KAmin(C) (179)
satisfies the large-Nc loop equation for asymptotically large C. However, a

self-consistency equation for K, which should relate it to the bare charge
and the cutoff, was not investigated. In order to do this, one needs more
detailed information about the behavior of W (C) for intermediate loops.
The free bosonic NambuGoto string which is defined as a sum over
surfaces spanned by C
e KA(S) ,
X
W (C) = (180)
S:S=C
with the action being the area A (S) of the surface S, is not a solution
for intermediate loops. Consequently, QCD does not reduce to this kind
of string, as was originally expected in Refs. [51, 52, 53, 54, 55]. Roughly
speaking, the ansatz (180) is not consistent with the factorized structure
on the RHS of Eq. (159).
Nevertheless, it was shown that if a free string satisfies Eq. (159), then
the same interacting string satisfies the loop equations for finite Nc . Here
free string means, as usual in string theory, that only surfaces of genus
54 YURI MAKEENKO
zero are present in the sum over surfaces, while surfaces or higher gen-
era are associated with
a string interaction. The coupling constant of this
interaction is O Nc2 .
A formal solution of Eq. (159) for all loops was found by Migdal [56] in
the form of a fermionic string

Z R
d2 [k k +m4 g ]
D e
X
W (C) = , (181)
S:S=C
where the world sheet of the string is parametrized by the coordinates

1 and 2 for which the 2-dimensional metric is conformal, i.e. diagonal.
The field () describes 2-dimensional elementary fermions (elves) living
in the surface S, and m stands for their mass. Elves were introduced to
provide factorization which now holds due to some remarkable properties
of 2-dimensional fermions. For large loops, the internal fermionic structure
becomes frozen, so that the empty string behavior (179) is recovered. For
small loops, the elves are necessary for asymptotic freedom. However, it
is unclear whether or not the string solution (181) is practically useful for
a study of multicolor QCD, since the methods of dealing with the string
theory in four dimensions are not yet developed.
A very interesting solution of the large-Nc loop equation on a lattice
was found by Eguchi and Kawai [57]. They showed that the SU(Nc ) gauge
theory on an infinite lattice reduces at Nc = to the model on a hyper-
cube. The equivalence is possible only at Nc = , when the space-time
dependence is absorbed by the internal symmetry group. More about this
large-Nc reduction will be said in the next Section.
4.8. WILSON LOOPS IN QCD2
Two-dimensional QCD is popular since the paper by t Hooft [58] as a

simplified model of QCD4 .
One can always choose the axial gauge A1 = 0,so that the commutator in
the non-Abelian field strength (60) vanishes in two dimensions. Therefore,
there is no gluon self-interaction in this gauge and the theory looks, at the
first glance, like the Abelian one.
The Wilson loop average in QCD2 can be straightforwardly calculated
via the expansion (164) where only disconnected (free) parts of the corre-
lators G(n) for even n should be left, since there is no interaction. Only the
planar structure of color indices contributes at Nc = . Diagrammatically,
the diagrams of the type depicted in Figure 18a and Figure 18b are relevant
for contours without self-intersections, while that in Figure 18c should be
omitted in two dimensions.
.......................
.......... ......
...... .....
.... ...
... ...
.... ...
.
.. ..
. ...
... ..
.... .
..
x2 = y ...
2 .....
.
.
..
.
...
... ..
.... ...
..... ....
....... .....
.................................
Figure 20. Graphic representation of the contour integral on the LHS of Eq. (186) in
the axial gauge. The bold line represents the gluon propagator (184) with x2 = y2 due
to the delta-function.
The color structure of the relevant planar diagrams can be reduced by

using the completeness condition (64) at large Nc . We have
I I I I
()k dx1 1 dx21 dx2k1 dx2kk
X
k
W (C) = 1 +
k C C C C
c(1, 2, . . . , 2k) D1 1 (x1 x2 ) Dk k (x2k1 x2k ) , (182)
where the points x1 , . . ., x2k are still cyclic ordered along the contour. We
can exponentiate the RHS of Eq. (182) to get finally

H H
dx dy D (xy)
W (C) = e 2 C C . (183)
This is the same formula as in the Abelian case if stands for e2 .

The propagator D (x, y) is, strictly speaking, the one in the gauge
A1 = 0 which reads
1
D (x y) = 2 2 |x1 y1 | (1) (x2 y2 ) . (184)
2
However, the contour integral on the RHS of Eq. (183) is gauge invariant,
and we can simply choose instead
1 2
D (x y) = ln , (185)
4 (x y)2
where is an arbitrary parameter of the dimension of length. Nothing

depends on it because the contour integral of a constant vanishes.
The contour integral in the exponent on the RHS of Eq. (183) can be
graphically represented as is depicted in Figure 20, where x2 = y2 due to
the delta-function in Eq. (184) and the bold line represents |x1 y1 |. This
gives I I
dx dy D (x y) = A (C) (186)
C C
56 YURI MAKEENKO
....................
......... ...... ...................................
...... .... ........
.... ... ......
......
.... ... .....
.....
..
. ... ..
.. ....
....
.... ..
.. ... ...
... ..
..
.. A 1
..
.
.. ...
...
..
A
...
...
..
...
... .
..
. .. 1 ..
..
... ... ... ..
... .. ..
..... .... .. ..
..... ..... ... ...
....... ...... ... .
........ ............. .. .
...... ........... ...
.
. ... .....
................ ...
..
...
...
.....
...
A
...
...
....
.....
...
... A 2
..
.
.. .....
..
..
...
2 ..... ...... . .
..
...... ......... ......... ...........
.......
....... ........ .............................
.......
a) b)
Figure 21. Contours with one self-intersection: A1 and A2 stand for the areas of the
proper windows. The total area enclosed by the contour in Figure a) is A1 +A2 . The areas
enclosed by the exterior and interior loops in Figure b) are A1 + A2 and A2 , respectively,
while the total area of the surface with the folding is A1 + 2A2 .
where A (C) is the area enclosed by the contour C. We get finally

W (C) = e 2 A(C) (187)
for the contours without self-intersections.
Therefore, area law holds in two dimensions both in the non-Abelian
and Abelian cases. This is, roughly speaking, because of the form of the
two-dimensional propagator (185) which falls down with the distance only
logarithmically in the Feynman gauge.
The difference between the Abelian and non-Abelian cases shows up for
the contours with self-intersections.
We first note that the simple formula (186) does not hold for contours
with arbitrary self-intersections.
The simplest contours with one self-intersection are depicted in Fig-
ure 21. There is nothing special about the contour in Figure 21a. Equa-
tion (186) still holds in this case with A (C) being the total area A (C) =
A1 + A2 .
The Wilson loop average for the contour in Figure 21a coincides both
for the Abelian and non-Abelian cases and equals

W (C) = e 2 (A1 +A2 ) . (188)
This is nothing but the exponential of the total area.
For the contour in Figure 21b, we get
I I

dx dy D (x y) = A1 + 4A2 . (189)
C C
This is easy to understand in the axial gauge where the ends of the propa-
gator line can lie both on the exterior and interior loops, or one end at the
........................ ........................ ........................ ........................

.......... ...... .......... ...... .......... ...... .......... ......
...... ..... ...... ..... ...... ..... ...... .....
..... .... ..... .... ..... ..... ..... .....
.... ... .... ... .... ... .... ...
...
. ... .... ...
... .... ... ..
. ...
.... ... ... ...
.. .... .. .... .. .
..
. ..
... .. ... ..
.. ... .. ... ..
..
... .. ... ... .. ...
.. .. .. ..
.. . .. . .. . ... ..
.. .. .. .. .. .. ..
..
.. ... .. .. .. .. ..
...
..
... .
. ... ... ..
... .
.
. ..
... .
.
............... . .
............... . .. . .
.............. .
... ..... .. ..
. ...
... ..... .. ..
. ... ......... .........
. ... ... ..... .. ..
.
..
...
... ... .. ... ... ... .. ... ... ..
... .. ..
.. ...
... ...
.. .. ...
... .... ... ... .. ....
...
... .... . ..
....
..... ... .
.. ........ ....
..... ....
. .. ........
....
.....
... .
.. ..... ..... ... .. ........
.
.
...... .... .... ...... .
........ ........................ ...... .... ...... ........ . ...... ...... ....... ........... ...... ... . .
....... ................................
..
.... ............................. .............................. .........
a) b) c) d)
Figure 22. Three type of contribution in Eq. (189) The ends of the propagator line
lie both on a) exterior and b) interior loops, or c), d) one end on the exterior loop and
another end on the interior loop.
exterior loop and the other end on the interior loop. These cases are illus-
trated by Figure 22. The contributions of the diagrams in Figure 22a,b,c,d
are A1 + A2 , A2 , A2 , and A2 , respectively. The result given by Eq. (189) is
obtained by summing over all four diagrams.
For the contour in Figure 21b, the Wilson loop average is

W (C) = e 2 (A1 +4A2 ) (190)
in the Abelian case and

W (C) = (1 A2 ) e 2 (A1 +2A2 ) (191)
in the non-Abelian case at Nc = . They coincide only to the order

as they should. The difference to the next orders is because only the dia-
grams with one propagator line connecting the interior and exterior loops
are planar and, therefore, contribute in the non-Abelian case. Otherwise,
the diagram is non-planar and vanishes as Nc . Notice, that the expo-
nential of the total area A (C) = A1 + 2A2 of the surface with the folding,
which is enclosed by the contour C, appears in the exponent for the non-
Abelian case. The additional pre-exponential factor could be associated
with an entropy of foldings of the surface.
The Wilson loop averages (188) and (191) in QCD2 at large Nc as well as
the ones for contours with arbitrary self-intersections, which have a generic
form

W (C) = P (A1 , . . . , An ) e 2 Area (192)
where P is a polynomial of the areas of individual windows and Area is
the total area of the surface with foldings, were first calculated in Ref. [59]
by solving the two-dimensional loop equation and in Ref. [60] by applying
the non-Abelian Stokes theorem.
58 YURI MAKEENKO
Remark on the string representation

A nice property of QCD2 at large Nc is that the exponential of the area en-
closed by the contour C emerges7 for the Wilson loop average W (C). This
is as it should for the NambuGoto string (180). However, the additional
pre-exponential factors (like that in Eq. (191)) are very difficult to interpret
in the stringy language. They may become negative for large loops which
is impossible for a bosonic string. This explicitly demonstrates in d = 2 the
statement of the previous subsection that the NambuGoto string is not a
solution of the large-Nc loop equation.
5. Large-N Reduction
The large-Nc reduction was first discovered by Eguchi and Kawai [57] who
showed that the Wilson lattice gauge theory on a d-dimensional hypercu-
bic lattice is equivalent at Nc = to the one on a hypercube with peri-
odic boundary conditions. This construction is based on an extra (ZNc )d -
symmetry which the reduced theory possesses to each order of the strong
coupling expansion.
Soon after it was recognized that a phase transition occurs in the re-
duced model with decreasing the coupling constant, so that this symme-
try is broken in the weak coupling regime. To cure the construction at
weak coupling, the quenching prescription was proposed by Bhanot, Heller
and Neuberger [61] and elaborated by many authors. An elegant alterna-
tive reduction procedure based on twisting prescription was advocated by
Gonzalez-Arroyo and Okawa [62]. Each of these prescriptions results in the
reduced model which is fully equivalent to multicolor QCD, both on the
lattice and in the continuum.
While the reduced models look as a great simplification, since the space-
time is reduced to a point, they still involve an integration over d infinite
matrices which is in fact a continual path integral. It is not clear at the
moment whether or not this is a real simplification of the original theory
which can make it solvable. Nevertheless, the reduced models are useful
and elegant representations of the original theory at large Nc .
We shall start this Section by a simplest example of a pure matrix
scalar theory. The quenched reduced model for this case was proposed by
Parisi [63] on the lattice end elaborated by Gross and Kitazawa [64] in the
continuum, while the twisted reduced model was advocated by Eguchi and
Nakayama [65]. Then we concentrate on the EguchiKawai reduction of
YangMills theory.
7
This is not true, as is already discussed, in the Abelian case for contours with self-
intersections.
5.1. REDUCTION OF SCALAR FIELD
Let us begin with a simplest example of a pure matrix scalar theory on a

lattice whose partition function is defined by the path integral
Z YY P P
N tr V [x ]+
x c

x x+a
Z= dij
x e . (193)
x ij
Here x is a Nc Nc Hermitean matrix field with x running over sites of a

hypercubic lattice and V [] is some interaction potential, say
M 2 3 3 4 4
V [] = + + . (194)
2 3 4
The prescription of the large-Nc reduction is formulated as follows. We
substitute
x Sx Sx , (195)
where
[Sx ]kj = e ipk x kj = diag e ip1 x , . . . , e ipNc x (196)
is a diagonal unitary matrix which eats the coordinate dependence, so that
does not depend on x.
The averaging of a functional F [x ] which is defined with the same
weight as in Eq. (193),

1
Z Y P P
D E
x
Nc tr V [x ]+
(x)(x+a)
F [x ] dx e F [x ] , (197)
Z x
can be calculated at Nc = by
Nc
d Y
dpi D
D E Z E
a
Nc d
F [Sx Sx ]
Y
F [x ] a (198)

a =1 i=1
2 Reduced
where the average on the RHS is calculated [63] for the quenched reduced
model whose averages are defined by
D E 1
F []
Reduced ZReduced
Z Y
|ij |2 cos ((p
P P
Nc tr V []+Nc i pj )a)
dij e ij F [] . (199)
ij
The partition function of the reduced model reads

Z Y
|ij |2 cos ((p
P P
Nc tr V []+Nc i pj )a)
ZReduced = dij e ij (200)
ij
60 YURI MAKEENKO
which can be deduced, modulo the volume factor, from the partition func-
tion (193) by the substitution (195).
Notice that the integration over the momenta pi on the RHS of Eq. (198)
is taken after the calculation of averages in the reduced model. Such vari-
ables are usually called quenched in statistical mechanics which clarifies the
terminology.
Since Nc it is not necessary to integrate over the quenched mo-
menta in Eq. (198). The integral should be recovered if pi s would be uni-
formly distributed in a d-dimensional hypercube. Moreover, a similar prop-
erty holds for the matrix integral over as well, which can be substituted
by its value at the saddle point configuration s :
D E
F [x ] F [Sx s Sx ] , (201)
where the momenta pi are uniformly distributed in the hypercube. There-

fore, this saddle point configuration plays the role of a master field in the
sense of Subsection 3.5.
In order to show how Eq. (198) works, let us demonstrate how the
planar diagrams of perturbation theory for the scalar matrix theory (193)
are recovered in the quenched reduced model.
The quenched reduced model (200) is of the general type discussed in
Section 3. The propagator is given by
D E 1
ij kl = G (pi pj ) il kj (202)
Gauss Nc
with
1
G (pi pj ) = . (203)
M (pi pj )a
P
cos
It is convenient to associate the momenta pi and pj in Eq. (203) with

each of the two index lines representing the propagator and carrying, re-
spectively, indices i and j. Remember, that these lines are oriented for a
Hermitean matrix and their orientation can be naturally associated with
the direction of the flow of the momentum. The total momentum carried
by the double line is pi pj .
The simplest diagram which represents the correction of the second
order in 3 to the propagator is depicted Figure 23. The momenta pi and
pj flows along the index lines i and j while the momentum pk circulates
along the index line k. The contribution of the diagram in Figure 23 reads
23
G (pi pj )2
X
2
G (pi pk ) G (pk pj ) , (204)
Nc k
i
........
............................................
....... ......
...... .................................. ....
.. .... .
........ ...... ......
. ...
... ...... .....
.... .....
.. ......
i .
.
.. ...
... ...
k ... ...
... ...
... ..
i
...........
...... ... .... ... .. ...........
......
... ..
................. . .. .................
... ...
. ... ...
.
... ..
... ... ... ...
j ... ...
... ......
... . k .....
... ...
.... ..... j
.... ................ ..... ............... ......
..... ................... .
...... .....
........ ......
.........................................
......
Figure 23. Simplest planar diagram of the second order in 3 for the propagator in
the quenched reduced model (200). The momentum pi flows along the index line i. The
momentum pi pj is associated with the double line ij.
where the summation over the index k is just a standard one over indices
forming a closed loop.
In order to show that the quenched-model result (204) reproduces the
correction to the propagator in the original theory on an infinite lattice, we
pass to the variables of the total momenta flowing along the double lines:
pi pj = p; pk pj = q; pi pk = p q , (205)
which is obviously consistent with the momentum conservation at each

of the two vertices of the diagram in Figure 23. Since pk s are uniformly
distributed in the hypercube, the summation over k can be substituted as
Nc by the integral
Z
1 X a dd q
f (pk ) ad f (q) . (206)
Nc k a
(2)d
The prescription (198) then gives the correct expression

2 dd q
Z
a
a 3 G (p)2
d
G (q) G (p q) (207)
Nc
a
(2)d
for the second-order contribution of the perturbation theory for the prop-
agator on the lattice.
It is now clear how a generic planar diagram is recovered by the reduced
model. We first represent the diagram by the double lines and associate the
momentum pi with an index line carrying the index i. Then we write down
the expression for the diagram in the reduced model with the propaga-
tor (203). Passing to the momenta flowing along the double lines, similar
to Eq. (205), we get an expression which coincides with the integrand of the
Feynman diagram for the theory on the whole lattice. It is crucial that such
62 YURI MAKEENKO
a change of variables can always be done for a planar diagram consistently

with the momentum conservation at each vertex. The last step is that the
summation over indices of closed index lines reproduces the integration over
momenta associated with each of the loops according to Eq. (206). It is as-
sumed that the number of loops is much less than Nc which is always true
for a given diagram since Nc is infinite.
We thus have shown how planar diagrams of the lattice theory defined
by the partition function (193) are recovered by the reduced model (200).
The lattice was needed only as a regularization to make all integrals well-
defined and was not crucial in the consideration. This construction can be
formulated directly for the continuum theory [64, 66] where the propagator
turns into
1
G (pi pj ) = (208)
(pi pj )2 + m2
and a Lorentz-invariant regularization can be achieved by choosing p2 < 2 .
Remark on the twisted reduced model

An alternative reduction procedure is based on the twisting prescription
[62]. We again perform the unitary transformation (195) with the matrices
Sx being expressed via a set of d (unitary) Nc Nc matrices by
x /a x /a x /a x /a
S x = 1 1 2 2 3 3 4 4 (209)
where the coordinates of the (lattice) vector x are measured in the lattice
units. The matrices are explicitly constructed in Ref. [62] and commute
by
= Z (210)
with Z = Z being elements of Z .
Nc
For the twisting reduction prescription, Eq. (198) is valid providing the
average on the RHS is calculated for the twisted reduced model which is
defined by the partition function [65]
Z
tr
P
Nc tr V []+Nc
ZTRM = d e . (211)
We can change the order of s in Eq. (209) defining a more general

path-dependent factor Y
Sx = P . (212)
lCx
The path-ordered product in this formula runs over all links l = (z, )
forming a path Cx from infinity to the point x.
Due to Eq. (210), changing the form of the path multiplies Sx by the
Abelian factor Y
Z(C) = Z (2) (213)
2S:S=C
where (, ) is the orientation of the plaquette 2. The product runs over

any surface spanned by the closed loop C which is obtained by passing
the original path forward and the new path backward. Due to the Bianchi
identity Y
Z (2) = 1 (214)
2cube
where the product goes over six plaquettes forming a 3-dimensional cube
on the lattice, the product on the RHS of Eq. (213) does not depend on
the form of the surface S and is a functional of the loop C.
It is now easy to see that under this change of the path we get
[Sx ]ij [Sx ]kl |Z(C)|2 [Sx ]ij [Sx ]kl (215)
and the path-dependence is canceled because |Z(C)|2 = 1. This is a general

property which holds for the twisting reduction prescription of any even (i.e.
invariant under the center ZNc ) representation of SU (Nc ).
5.2. REDUCTION OF YANGMILLS FIELD
The statement of the EguchiKawai reduction of the YangMills field says

that the theory on a d-dimensional space-time is equivalent at Nc = to
the reduced model which is nothing but its reduction to a point. The action
of the reduced model is given by
1
SEK = tr [A , A ]2 , (216)
2g2 d
where A are d space-independent matrices and is a dimensionful param-

eter.
A naive statement of the EguchiKawai reduction is that the averages
coincide in both theories, for example,
1 1
H H
d A () d A
tr P e i = tr P e i (217)
Nc ddim Nc EK
where the LHS is calculated with the action (59) and the RHS is calculated
with the reduced action (216). Strictly speaking, this naive statement is
valid only in d = 2 or supersymmetric case for the reason which will be
explained in a moment.
64 YURI MAKEENKO
The precise equivalence is valid only if the average of open Wilson loops
vanish in the reduced model:
1
R
i d A
tr P e Cyx
= 0, (218)
Nc EK
as it does in the d-dimensional theory due to the local gauge invariance

under which
R R
i d A () i d A ()
Pe Cyx
(y)P e Cyx
(x) . (219)
ij ij
The point is that this gauge invariance transforms in the reduced model
into (global) rotation of the reduced field by constant matrices :
A A . (220)
which does not guarantee such vanishing in the reduced model.

There exists, however, a symmetry of the reduced action (216) under
the shift of A by a unit matrix8 :
Aij ij ij
A + a , (221)
which is often called the Rd -symmetry. Under the transformation (221), we

get R R
i d A x )a i d A
Pe Cyx
e i(y
Pe Cyx
(222)
ij ij
which guarantees, if the symmetry is not broken, the vanishing of the open
Wilson loops
1
R
i d A
WEK (Cyx ) tr P e Cyx
=0 (223)
Nc EK
in the reduced model.

The equivalence of the two theories can then be shown using the loop
equation which reads for the reduced model
1
H
i d A
x WEK (C) = tr P[A , [A , A ]] e Cxx
(x) Nc EK
1
H
d i C d A
= tr P e xx
Nc A EK
I
d
= dy WEK (Cyx ) WEK (Cxy ) . (224)
C
8
This symmetry is rigorously defined on a lattice where it is associated with a
direction-dependent ZNc transformation.
The RHS is pretty much similar to the one in Eq. (159) while (d) (x y)
is missing.
This delta function can be recovered if the Rd symmetry is not broken
since
(d) (x y)
WEK (Cyx ) WEK (Cyx ) (225)
(d) (0)
due to Eq. (223) for the open loops.
This is not a rigorous argument since a regularization is needed. What
actually happens is the following. If we smear the delta function introducing
d

(d) 2 2 /2
(x) = e x , (226)
2
then
1
(d)
2 2 2
(d)
(0) d e x (d) (x) , (227)
(0)
reproducing the delta function.
5.3. RD -SYMMETRY IN PERTURBATION THEORY
Since Nc is infinite, the Rd -symmetry can be broken spontaneously. The

point is that the large-Nc limit plays the role of a statistical averaging, as
is mentioned already in Subsection 3.6, and phase transitions are possi-
ble for infinite number of degrees of freedom. This phenomenon occurs in
perturbation theory of the reduced model for d 3.
The perturbation theory can be constructed expanding the fields around
solutions of the classical equation
[A , [A , A ]] = 0 . (228)
Any diagonal matrix
n o
Acl (1) (Nc )
p = diag p , . . . , p (229)
is a solution to Eq. (228).

The perturbation theory of the reduced model can be constructed ex-
panding around the classical solution (229):
A = Acl q
+ gA , (230)
where Aq is off-diagonal.
Substituting (230) into the action (216), we get
1 1

SEK = tr [p , Aq ]2 [p , Aq ]2 + higher orders . (231)
2 2
66 YURI MAKEENKO
To fix the gauge symmetry (220), it is convenient to add

1

Sg.f. = tr [p , Aq ]2 + [p , b][p , c] , (232)
2
where b and c are ghosts.
The sum of (231) and (232) gives
1

S2 = tr [p , Aq ]2 + [p , b][p , c] (233)
2
up to quadratic order in Aq .
Doing the Gaussian integral over Aq , we get at the one-loop level:
Z N
Z Y Yh i1d/2
dp dAq e S2 ... = dp(k)
(p(i) (j) 2
p ) ... , (234)
k=1 i<j
where the integration over p accounts for equivalent classical solutions.

For d = 1 the product on the RHS of Eq. (234) reproduces the Vander-
monde determinant. For d = 2 it vanishes and does not affect dynamics.
For d 3 the measure is singular and the eigenvalues collapse. This leads
us to a spontaneous breakdown of the Rd in perturbation theory.
The equivalence between the Nc = YangMills theory on a whole
space and the reduced model can be provided [61] introducing a quenching
prescription similar to the one described in Subsection 5.1. Then no collapse
of eigenvalues happens and d-dimensional planar graphs are reproduced by
the reduced model. More about the quenching prescription in YangMills
theory can be found in the reviews [44, 67] and cited there original papers.
Remark on supersymmetric case

In a supersymmetric gauge theory, there is an extra contribution from
fermions to the exponent on the RHS of Eq. (234). Since the integration
(i) (j)
over fermions results in the extra factor [(p p )2 ] tr I/2 , this yields fi-
nally the exponent 1d/2+ tr I/2. It vanishes in d = 4 for either Majorana
or Weyl fermions and in d = 10 for the MajoranaWeyl fermions. There-
fore, the Rd -symmetry is not broken and no quenching is needed in the
supersymmetric case [68, 5].
5.4. TWISTED REDUCED MODEL
The continuum version of the twisted reduced model can be constructed [69]
by substituting A A into the action (216), where the matrices
obey the commutation relation
[ , ] = B I , (235)
where B is an antisymmetric tensor and d is even. This is possible only

for infinite Hermitean matrices (operators). An example of such matrices is
x and p operators in quantum mechanics. Eq. (235) is a continuum version
of Eq. (210).
The Wilson loop averages in the twisted reduced model are defined by
1 1
R R
i d i d A
WTEK (Cyx ) = tr P e Cyx
tr P e Cyx
. (236)
Nc Nc TEK
They vanish for open loops which is provided by the vanishing of the trace
of the path-ordered exponential of in this definition. For closed loops
this factor does not vanish and is needed to provide the equivalence with
d-dimensional YangMills perturbation theory, since the classical extrema
of the twisted reduced model are Acl = and the perturbation theory is
constructed expanding around this classical solution.
The proof of the equivalence can be done using the loop equation quite
similarly to that of Subsection 5.2 for the EguchiKawai model with an
unbroken Rd symmetry.
Acknowledgements
I would like to thank the organizers for the warm hospitality in Iceland.
This work is supported in part by the grant RFFI 970217927.
References
1. G. t Hooft, A planar diagram theory for strong interactions, Nucl. Phys. B72
(1974) 461.
2. J. Maldacena, The large N limit of superconformal field theories and supergravity,
Adv. Theor. Math. Phys. 2 (1998) 231 hep-th/9711200.
3. T. Banks, W. Fischler, S. H. Shenker and L. Susskind, M theory as a matrix model:
a conjecture, Phys. Rev. D55 (1997) 5112, hep-th/9610043.
4. N. Drukker, D. J. Gross and H. Ooguri, Wilson loops and minimal surfaces, Phys.
Rev. D60 (1999) 125006, hep-th/9904191.
5. N. Ishibashi, H. Kawai, Y. Kitazawa and A. Tsuchiya, A large-N reduced model
as superstring, Nucl. Phys. B498 (1997) 467, hep-th/9612115.
6. H. Aoki, N. Ishibashi, S. Iso, H. Kawai, Y. Kitazawa and T. Tada, Noncommutative
Yang-Mills in IIB matrix model, hep-th/9908141.
7. Y. Makeenko, Formulation of matrix theory at finite temperature, Forschr. Phys
48 (2000) 171, hep-th/9903030.
8. H.E. Stanley, Spherical model as the limit of infinite spin dimensionality, Phys.
Rev. 176 (1968) 718.
9. K.G. Wilson, Quantum field-theory models in less than 4 dimensions, Phys. Rev.
D7 (1973) 2911.
10. D.J. Gross and A. Neveu, Dynamical symmetry breaking in asymptotically free
field theories, Phys. Rev. D10 (1974) 3235.
11. A.A. Anselm, A model of a field theory with nonvanishing renormalized charge,
Sov. Phys. JETP 9 (1959) 608.
68 YURI MAKEENKO
12. K. Johnson, Solution of the equations for the Greens functions of a two dimensional
relativistic field theory, Nuovo Cim. 20 (1961) 773.
13. G. Parisi, The theory of non-renormalizable interaction, Nucl. Phys. B100 (1975)
368.
14. W. Chen, Y. Makeenko, and G.W. Semenoff, Four-fermion theory and the confor-
mal bootstrap, Ann. Phys. 228 (1993) 341.
15. J.A. Gracey, Calculation of exponent to O(1/N 2 ) in the O(N ) Gross Neveu
model, Int. J. Mod. Phys. A6 (1991) 395, 2755(E).
16. G. Mack and A. Salam, Finite-component field representation of the conformal
group, Ann. Phys. 53 (1969) 144.
17. D. Gross and J. Wess, Scale invariance, conformal invariance, and the high-energy
behavior of scattering amplitudes, Phys. Rev. D2 (1970) 753.
18. R. Jackiw, Field theoretic investigations in current algebra, in Lectures on Current
Algebra and its Applications by S.B. Treiman, R. Jackiw, and D.J. Gross, Princeton
Univ. Press, 1972, p. 97; also in Current Algebra and Anomalies by S.B. Treiman,
R. Jackiw, B. Zumino, and E. Witten, World Sci., 1985, p. 81.
19. A.M. Polyakov, Conformal symmetry of critical fluctuations, JETP Lett. 12
(1970) 381.
20. A.A. Migdal, On hadronic interactions at small distances, Phys. Lett. 37B (1971)
98.
21. A.M. Polyakov, Interaction of Goldstone particles in two dimensions. Applications
to ferromagnets and massive YangMills fields, Phys. Lett. 59B (1975) 79.
22. A.A. Belavin and A.M. Polyakov, Metastable states of two-dimensional isotropic
ferromagnet, JETP Lett. 22 (1975) 245.
23. G. Veneziano, Some aspects of a unified approach to gauge, dual and Gribov the-
ories, Nucl. Phys. B117 (1976) 519.
24. W.T. Tuttle, A census of planar triangulations, Can. J. Math. 14 (1962) 21.
25. J. Koplik, A. Neveu and S. Nussinov, Some aspects of the planar perturbation
series, Nucl. Phys. B123 (1977) 109.
26. P. Cvitanovic, Planar perturbation expansion, Phys. Lett. 99B (1981) 49.
27. P. Cvitanovic, P.G. Lauwers, and P.N. Scharbach, The planar sector of field theo-
ries, Nucl. Phys. B203 (1982) 385.
28. E. Witten, The 1/N expansion in atomic and particle physics, in Recent develop-
ments in gauge theories, eds. G. t Hooft et al., Plenum, 1980, p.403.
29. S. Coleman, 1/N, in Proc. of Erice Int. School of Subnuclear Physics 1979,
Plenum, N.Y., 1982, p. 805 (Reprinted in S. Coleman, Aspects of symmetry, Cam-
bridge Univ. Press, 1985, pp.351402).
30. O. Haan, Large N as a thermodynamic limit, Phys. Lett. 106B (1981) 207.
31. D.V. Voiculescu, K.J. Dykema and A. Nica, Free Random Variables, AMS, Prov-
idence 1992.
32. M.R. Douglas, Stochastic master fields, Phys. Lett. B344 (1995) 117.
33. R. Gopakumar and G.J. Gross, Mastering the master field, Nucl. Phys. B451
(1995) 379.
34. S. Mandelstam, Charge-monopole duality and the phases of non-Abelian gauge
theories, Phys. Rev. D19 (1979) 2391.
35. Yu.M. Makeenko and A.A. Migdal, Exact equation for the loop average in multi-
color QCD, Phys. Lett. 88B (1979) 135.
36. K.G. Wilson, Confinement of quarks, Phys. Rev. D10 (1974) 2445.
37. Yu.M. Makeenko and A.A. Migdal, Quantum chromodynamics as dynamics of
loops, Nucl. Phys. B188 (1981) 269.
38. R.A. Brandt, F. Neri and D. Zwanziger, Lorentz invariance from classical particle
paths in quantum field theory of electric and magnetic charge, Phys. Rev. D19
(1979) 1153.
39. J.L. Gervais and A. Neveu, The slope of the leading Regge trajectory in quantum
chromodynamics, Nucl. Phys. B163 (1980) 189.
40. A.M. Polyakov, Gauge fields as rings of glue, Nucl. Phys. B164 (1980) 171.
41. V.S. Dotsenko and S.N. Vergeles, Renormalizability of phase factors in non-Abelian
gauge theory, Nucl. Phys. B169 (1980) 527.
42. R.A. Brandt, F. Neri and M. Sato, Renormalization of loop functions for all loops,
Phys. Rev. D24 (1981) 879.
43. R.A. Brandt, A. Gocksch, M. Sato, and F. Neri, Loop space, Phys. Rev. D26
(1982) 3611.
44. A.A. Migdal, Loop equations and 1/N expansion, Phys. Rep. 102 (1983) 199.
45. J.L. Gervais and A. Neveu, Local harmonicity of the Wilson loop integral in clas-
sical YangMills theory, Nucl. Phys. B153 (1979) 445.
46. P. Levy, Probleems concrets danalyse fonctionnelle, Paris, 1951.
47. M.N. Feller, Infinite-dimensional elliptic equations and operators of the type by
P. Levy, Sov. J. Usp. Mat. Nauk. 41 (1986) 97.
48. M.B. Halpern and Yu.M. Makeenko, Continuum-regularized loop-space equation,
Phys. Lett. 218B (1989) 230.
49. Yu.M. Makeenko, Polygon discretization of the loop-space equation, Phys. Lett.
212B (1988) 221.
50. Yu.M. Makeenko and A.A. Migdal, Self-consistent area law in QCD, Phys. Lett.
97B (1980) 235.
51. J.L. Gervais and A. Neveu, The quantum dual string wave functional in YangMills
theories, Phys. Lett. 80B (1979) 255.
52. Y. Nambu, QCD and the string model, Phys. Lett. 80B (1979) 372.
53. A.M. Polyakov, String representations and hidden symmetries for gauge fields,
Phys. Lett. 82B (1979) 247.
54. D. Foerster, YangMills theory a string theory in disguise, Phys. Lett. 87B
(1979) 87.
55. T. Eguchi, Strings in U(N) lattice gauge theory, Phys. Lett. 87B (1979) 91.
56. A.A. Migdal, QCD = Fermi string theory, Nucl. Phys. 189 (1981) 253.
57. T. Eguchi and H. Kawai, Reduction of dynamical degrees of freedom in the large-N
gauge theory, Phys. Rev. Lett. 48 (1982) 1063.
58. G. t Hooft, A two-dimensional model for mesons, Nucl. Phys. B75 (1974) 461.
59. V.A. Kazakov and I.K. Kostov, Non-linear strings in two-dimensional U() the-
ory, Nucl. Phys. B176 (1980) 199.
60. N. Bralic, Exact computation of loop averages in two-dimensional YangMills the-
ory, Phys. Rev. D22 (1980) 3090.
61. G. Bhanot, U. Heller and H. Neuberger, The quenched EguchiKawai model,
Phys. Lett. 113B (1982) 47.
62. A. Gonzalez-Arroyo and M. Okawa, The twisted Eguchi-Kawai model: A reduced
model for large N lattice gauge theory, Phys. Rev. D27 (1983) 2397.
63. G. Parisi, A simple expression for planar field theories, Phys. Lett. 112B (1982)
463.
64. D.J. Gross and Y. Kitazawa, A quenched momentum prescription for large-N the-
ories, Nucl. Phys. B206 (1982) 440.
65. T. Eguchi and R. Nakayama, Simplification of quenching procedure for large N
spin models, Phys. Lett. 122B (1983) 59.
66. S. Das and S. Wadia, Translation invariance and a reduced model for summing
planar diagrams in QCD, Phys. Lett. B117 (1982) 228.
67. S.R. Das, Some aspects of large-N theories, Rev. Mod. Phys. 59 (1987) 235.
68. R.L. Mkrtchyan and S.B. Khokhlachev, Reduction of the U() theory to a model
of random matrices, JETP Lett. 37 (1983) 160.
69. A. Gonzalez-Arroyo and C. P. Korthals Altes, Reduced model for large N contin-
uum field theories, Phys. Lett. B131 (1983) 396.

Large-N Gauge Theories

Uploaded by

Copyright:

Available Formats

Large-N Gauge Theories

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Large-N Gauge Theories

Uploaded by

Copyright:

Available Formats

LARGE-N GAUGE THEORIES

Lectures presented at the 1999 NATO-ASI on Quantum Geometry in

Abstract. Four pedagogical Lectures at the NATO-ASI on Quantum Ge-

2. O(N ) Vector Models

limit of the vector models is solvable. Alternatively, the large-N solution

2.1. FOUR-FERMI INTERACTION

The action of the O(N )-symmetric four-Fermi interaction in a d-

Here = and = (1 , . . . , N ) is a spinor field which forms an

which can be derived by varying the action (3) with respect to .

Remark on the one-loop Gell-MannLow function of four-Fermi theory

the wave-function renormalization of the -field, which emerges to the next

2.2. BUBBLE GRAPHS AS ZEROTH ORDER IN 1/N

The perturbation-theory expansion of the O(N )-symmetric four-Fermi the-

This determines the exact propagator of the field at large N . It is O N 1

since the coupling G is included in the definition of the propagator.

where g2 is a proper dimensionless constant which is not necessarily positive

2.3. SCALE AND CONFORMAL INVARIANCE OF FOUR-FERMI THEORY

The coefficient in Eq. (15) can easily be calculated in d = 3. To evaluate

To calculate in d = 3 the coefficient of proportionality in Eq. (16), let us

The integral is obviously convergent.

perturbation theory to l = 1 in the zeroth order of the 1/N expansion

The (logarithmically) divergent contribution emerges from the domain of

where p1 and p2 the incoming and outgoing fermion momenta, respectively.

and |p2 |. This gives

Group Transformations # Parameters

TABLE 1. Contents and the number of parameters of groups of space-time

Equation (39), which results from conformal invariance, unambiguously

Remark on broken scale invariance

2.4. NONLINEAR SIGMA MODEL

The nonlinear O(N ) sigma model3 in 2 Euclidean dimensions is defined by

where ~n = (n1 , . . . , nN ) is an O(N ) vector. While the action in Eq. (42) is

which is imposed on the ~n field via the (functional) delta-function.

where the contour of integration over u(x) is parallel to imaginary axis.

where the auxiliary field u is denoted again by the wavy line.

The coupling g2 in Eq. (47) is 1/N as is prescribed by the con-

and expanding the square root in g2 . Only N 1 dynamical degrees of free-

The exponentiation results in Eq. (50).

2.5. LARGE-N FACTORIZATION IN VECTOR MODELS

3.1. INDEX OR RIBBON GRAPHS

In order to describe the 1/Nc -expansion of the YangMills theory, it is

in the Feynman gauge.

multiplying by the generators of the SU(Nc ) gauge group according to the

complex field Aij ji ij

is antisymmetric as well. We consider all three gluons as incoming so that

a11a22a33 (p1 , p2 , p3 ) = f a1 a2 a3 1 2 3 (p1 , p2 , p3 ) (68)

different values. Then only the following term is left

of the pure SU(Nc ) gauge theory.

in the large-Nc limit.

The double lines of the diagram in Figure 6 can be viewed as bounding

Remark on the U(Nc ) gauge group

3.2. PLANAR AND NON-PLANAR GRAPHS

The double-line representation of perturbation theory diagrams in the in-

The diagrams of the type in Figure 7, which can be drawn on a sheet

Figure 7. Double-line representation of a four-loop diagram for the gluon propagator.

Figure 8. Double-line representation of a three-loop non-planar diagram for the gluon

such diagrams, an adding of a loop inevitably results in adding of two three-

so that all planar diagrams survive in the large-Nc limit.

The above evaluations of the order of the diagrams in Figures 68 can

Thus, the expansion in 1/Nc rearranges perturbation-theory diagrams ac-