Class. Quantum Grav. 17 (2000) 1225–1233.

Printed in the UK PII: S0264-9381(00)07907-7

Wilson loops in large-N theories

Hirosi Ooguri
366 Le Conte Hall, Department of Physics, University of California at Berkeley, Berkeley,
CA 94720, USA
and
Theory Group, Mail Stop 50A-5101, Physics Division, Lawrence Berkeley National Laboratory,
Berkeley, CA 94720, USA

Received 16 September 1999

Abstract. We discuss how various aspects of Wilson loops in large-N gauge theories are studied
from the point of view of the AdS–CFT correspondence.

PACS numbers: 1115T, 1125

1. Introduction

In the AdS–CFT correspondence, correlation functions of conformal fields are related to

amplitudes of a quantum theory in anti-de Sitter (AdS) theory [1–3] (for a review, see, for
example, [4]). Since the conformal group of the boundary and the isometry group of AdS are
identical, correlation functions defined in this way are conformally symmetric. However, not
all quantum theories in AdS can be related to conformal field theory (CFT) at the boundary in
this way. The correlation functions on the boundary must obey the axioms of CFT. For example,
any CFT contains the energy–momentum tensor in the operator algebra, and one must be able
to compute correlation functions including the energy–momentum tensor. There has to be a
field in AdS to which it can couple, namely the graviton. Thus a quantum theory dual to CFT
is necessarily gravitational. A quantum gravity may or may not be a string theory. In some
cases, however, there are operators in CFT which can directly couple to string states. Wilson
loop operators are examples of such operators. So we may hope to learn about stringy aspects
of the AdS–CFT correspondence by studying the Wilson loops. Also, these are fundamental
gauge-invariant operators in gauge theories, and we may hope to learn about gauge theories
from the point of view of AdS.
In this paper, I will discuss Wilson loops in the N = 4 supersymmetric Yang–Mills
theory in four dimensions, based on my work with Drukker and Gross [5]. First, we review
basic properties of the Wilson loops in N = 4 theory. This part of the paper is purely field
theoretical. The N = 4 theory contains massless scalar fields and they must be taken into
account in constructing the Wilson loop. The presence of the scalar fields improves ultraviolet
(UV) properties of the loop operators. We will then discuss how these and other properties of
Wilson loops can be seen from the point of view of string theory in AdS.
At Strings ’99, I also presented my work in progress with Cumrun Vafa about Wilson loops
in the large-N Chern–Simons gauge theory in three dimensions, which will appear elsewhere.
(This is why the title says ‘large-N theories’.)
0264-9381/00/051225+09$30.00 © 2000 IOP Publishing Ltd 1225
1226 H Ooguri

2. Wilson loops in the N = 4 super-Yang–Mills theory

Let me start with the N = 4 supersymmetric Yang–Mills theory in four dimensions. As I

mentioned, this part of the paper is purely field theoretical.

2.1. Definition
The Wilson loop is a phase factor associated with a trajectory of a quark in the fundamental
representation of the gauge group G. We will discuss the case when G = U (N ). The N = 4
theory consists of the gauge field Aµ , four Weyl fermions λa (a = 1, . . . , 4), and six scalar
fields φi (i = 1, . . . , 6) all in the adjoint representation of the gauge group U (N ). The theory
does not contain particles in the fundamental representation. Instead, we may use the W-boson
to probe the theory. We start with the U (N + 1) gauge group and break it into U (N ) × U (1)
by choosing the non-zero vacuum expectation values for the scalar fields,
 
φUi (N ) = 0 0
φUi (N +1) = . (1)
0 uθ i

Since there are six scalar fields, we parametrize their vacuum expectation values by a point θ i
on the unit 5-sphere, θ 2 = 1, corresponding to the direction of the symmetry breaking. The
absolute value of the scalar vacuum expectation value is denoted by u.
The phase factor associated with a trajectory of the W-boson gives the loop operator of
the form,
I 


W = Tr P exp ds iAµ (x(s))ẋ µ (s) + iφi (x(s))θ i (s)|ẋ(s)| , (2)

in the Minkowski space, and

I 


W = Tr P exp ds iAµ (x(s))ẋ µ (s) + φi (x(s))θ i (s)|ẋ(s)| , (3)

in the Euclidean space. We should point out that there is an important difference between the
Minkowski case and the Euclidean case; that is the absence of the imaginary unit ‘i’ in front
of the scalar field φ in the Euclidean case. In particular, the Wilson loop in the Euclidean case
is not a pure phase. This distinction is important in many of our results in the following. In
this paper, we will deal with the Euclidean case only.
Another important aspect of the Wilson loop (3) generated by the W-boson is that it couples
to the gauge field Aµ and the scalar field φi with the same strength since θ 2 = 1. Clearly this
is a consequence of the fact that the W-boson in this case is a BPS particle. This is not the
most general gauge-invariant observable one can write down. In general, one may consider a
more general loop operator whose coupling strength to the gauge field may be different from
that to the scalar field, as in
I 


W = Tr P exp µ i
ds iAµ (x(s))ẋ (s) + φi (x(s))ẏ (s) . (4)

Here we use the symbol ẏ i to denote the coupling to the scalar field. The phase factor for a
W-boson trajectory corresponds to the special case when ẋ 2 = ẏ 2 .
 Wilson loops in large-N theories 1227

2.2. UV divergence
Although the N = 4 theory is ultraviolet finite, composite operators of the theory may require
regularization. If two local operators coincide at a point, for example, there can be a divergence.
Let us discuss the ultraviolet divergence in the Wilson loop operator. It turns out that the
operator with the constraint ẋ 2 = ẏ 2 is special in this regard.
If the ’t Hooft coupling g 2 N is small, one may compute the vacuum expectation value of
the Wilson loop (4) by the perturbative expansion. The one-loop computation gives
I  
g2 N ẏ 2
hW i = 1 + ds |ẋ| 1 − 2 + · · · . (5)
(2π)2  ẋ
Here  is the UV cut-off parameter. When ẋ 2 = ẏ 2 , the divergence is cancelled due to the
cancellation between the gauge field exchange and the scalar field exchange diagrams. It seems
that this cancellation persists at higher loops. As we will see in the following, the AdS–CFT
correspondence shows that the cancellation of the UV divergence also happens at large g 2 N .
So far, we have assumed that the loop is smooth. When there is a singularity on the loop,
the divergence is not completely cancelled, and some logarithmic divergence remains even
when ẋ 2 = ẏ 2 . For example, when the loop has a cusp, the one-loop computation shows the
logarithmic divergence depending on the angle at the cusp as
 
g2 N π −  1
hWwith cusp i = 1 + (cos  + cos 2) log + ···, (6)
(2π)2 sin  
where  is the angle at the cusp (i.e. a jump in ẋ µ /|ẋ|). At the cusp, the direction θ i = ẏ i /|ẏ|
of ẏ i may also change discontinuously; the angle 2 is the amount of the discontinuity in θ i
at the cusp. Renormalizing this divergence would then give an anomalous scaling property of
the loop, which depends on  and 2. There is also a logarithmic divergence when the loop
has an intersection,
 
g2 N 1 1
hWwith intersection i = 1 + (cos  + cos 2) log + ···. (7)
2π sin  
These are somewhat similar to the logarithmic divergence in the even-dimensional observables
(such as points or surfaces) discussed by Berenstein et al [6] and by Graham and Witten [7].

2.3. Loop equation

The large-N loop equation is considered to be one of the fundamental properties of the Wilson
loop operators [8]. It is expected to hold in both small and large g 2 N . When g 2 N is small, it is
equivalent to the perturbative Feynmann rules of the gauge theory. On the other hand, there are
subtleties in the definition of the loop equation since it is derived assuming that the loops are
regularized but not renormalized. In fact, the equation depends explicitly on the UV cut-off.
The situation seems better in the N = 4 theory because of the better ultraviolet behaviour, as
we saw in the above. Since the equation is also supposed to hold at large g 2 N , it may be useful
in testing the stringy aspects of the AdS–CFT correspondence. I will comment on this later in
this paper.
In order to write down the loop equation, we need a complete set of gauge-invariant
observables which can be written in the form of loops. In particular, we also need to introduce
sources for the fermion fields. From the point of view of supersymmetry, it is natural to couple
the fermionic variable ζ to the gluino field λ in the combination,
I 


W = Tr P exp ds iAµ ẋ µ + φi ẏ i + 21 iζ̄ (iγµ ẋ µ + 0i ẏ i )λ + · · · . (8)
1228 H Ooguri

It turns out that the combination (iγµ ẋ µ + 0i ẏ i ) of 4d and 6d gamma-matrices, γµ and 0i ,

becomes nilpotent when the constraint ẋ 2 = ẏ 2 is satisfied. This simplifies our task of writing
down the loop equation considerably. We then introduce the second-order differential operator
on the loop space defined as†
I Z s+η  
0 δ2 δ2 δ2
L = lim ds ds − + . (9)
η→0 s−η δx µ (s)δx µ (s 0 ) δy i (s)δy i (s 0 ) δζ (s)δ ζ̄ (s 0 )
The two derivatives are taken at different points s and s 0 on the loop. The distance η between
the two points must be chosen to be shorter than the UV cut-off  so that we can isolate the
contact terms we need for the loop equation. This is the prescription due to Polyakov [9].
When the loop satisfies the constraint ẋ 2 = ẏ 2 , the action of the differential operator
on the Wilson loop can be written, using the Schwinger–Dyson equation and the large-N
factorization, as
I I
 
LhW i = g 2 N ds ds 0 δ (4) (x(s) − x(s 0 )) ẋ(s) · ẋ(s 0 ) − ẏ(s) · ẏ(s 0 ) hWss 0 ihWs 0 s i, (10)

where Wss 0 is given by the path-ordered exponential of the form (8) integrated over the part
of the loop, between s and s 0 . The fermionic variables ζ are set to be zero after taking the
derivative. The loop equation (10) states that LhW i is non-zero only when the loop has a
self-intersection. In the case of pure Yang–Mills theory without supersymmetry, there is an
ambiguity about whether to take into account the trivial self-intersection, namely the case when
s = s 0 , for which the delta-function constraint x µ (s) = x µ (s 0 ) in the right-hand side of (10)
is satisfied trivially. In some sense, the loop intersects with itself at each point along the loop.
In the case of the N = 4 theory, we do not have to worry about such an ambiguity since the
factor [ẋ(s) · ẋ(s 0 ) − ẏ(s) · ẏ(s 0 )] vanishes when s = s 0 .

3. Wilson loops in AdS5 × S 5

Now let us discuss how these properties of loop operators can be seen from the point of view
of a string in AdS. In the above, we started with the U (N + 1) gauge group and broke the
group into U (N ) × U (1). In string theory, this corresponds to putting N D3-branes on top
of each other, and probing it with another D3-brane. The open string stretched between the
N D3-branes and the single D3-brane probe corresponds to the W-boson of the gauge theory.
According to Maldacena’s conjecture, in the large-N limit, the N D3-branes are replaced by
the geometry of the AdS5 times a 5-sphere. The W-boson is now a string in AdS stretched from
the boundary. The large-N Wilson loop was studied from this point of view by Maldacena
[11] and by Rey and Yee [12]. In the following we will clarify some aspects of this approach
and extend it to various cases.
The metric on AdS5 times the 5-sphere is given by
p p
ds 2 = g 2 Ny −2 (dy dy + dx µ dx µ ) + g 2 N dθ 2 . (11)
It is often useful to combine the radial coordinate y of AdS with the coordinates θ of the
5-sphere into six coordinates y i = yθ i . In the coordinates x µ and y i , it is easy to see that the
total metric is conformal to the flat ten-dimensional metric,
p
ds 2 = g 2 Ny −2 (dx µ dx µ + dy i dy i ). (12)
In these coordinates, the boundary of AdS is at y = 0.
† The differential operator L defined here does not preserve the constraint ẋ 2 = ẏ 2 . Recently some improvement in
the definition of L was made, and it was found to be possible to write a loop equation which closes only among loops
preserving the constraint [10].
 Wilson loops in large-N theories 1229

3.1. Boundary conditions

To compute the Wilson loop observables in this framework, it is important to understand the
relation between the loop variables and the boundary condition on the string worldsheet. It is
well known that the conformal field theory on the string worldsheet couples to the spacetime
gauge fields Aµ and φi as
I  
∂Xµ
dσ 1 Aµ + φ i P i
, (13)
∂σ 1
where the integral is along the boundary of the worldsheet. We use σ α (α = 1, 2) for the
worldsheet coordinates with the worldsheet boundary at σ 2 = 0. The vector field Aµ couples
to the derivative of the string coordinates X µ along the boundary of the worldsheet. The scalar
fields φi , on the other hand, couple to the momentum P i conjugate to the corresponding string
coordinates Y i ,
1 ∂Y i
P i = √ g1α  αβ β , (14)
g ∂σ
where gαβ is a metric on the worldsheet. The scalar fields φi correspond to the transverse
coordinates of the D-brane, and they are T-dual of the gauge field along the D-brane. Since
the momentum P i is T-dual to the derivative of the string coordinate along the boundary, the
scalar fields couple to P i .
Since the gauge field Aµ couples to the derivative of the string coordinates ∂1 Xµ along
the boundary and the scalar field φi couples to the momentum P i conjugate to the string
coordinates at the boundary, it is clear that the Wilson loop operator of the form (4) couples to
the string worldsheet with the following boundary conditions:
X µ (σ 1 , σ 2 = 0) = x µ (σ 1 )
dy i 1 (15)
P i (σ 1 , σ 2 = 0) = (σ ).
dσ 1
In the four-dimensional directions along the D3-brane, the string worldsheet obeys the Dirichlet
condition that the string ends along the loop x µ (s). For the six transverse directions, the
string momentum is fixed following the Neumann condition. These boundary conditions are
complementary to the standard D3-brane boundary conditions (the Neumann condition for Xµ
and the Dirichlet condition for Y i ), as they should be since we are imposing extra conditions
by inserting the loop operator on the D-brane.
When g 2 N is large, the string tension becomes large and we can approximate the string
dynamics by a minimum surface in AdS [11, 12]. For a given set of the loop variables
(x µ (s), y i (s)), we expect that there is a unique minimum surface in AdS obeying these
boundary conditions. The existence and uniqueness of minimum surfaces in AdS, in the
case when ẏ i is constant, have been discussed in the mathematics literature (see, for example,
[13–15]). This, however, leads to a puzzle. In the AdS–CFT correspondence, we expect that
the boundary conditions for the bulk degrees of freedom are imposed at the boundary of AdS
at y = 0. However, the condition for the string worldsheet to terminate at y = 0 would
be an extra Dirichlet boundary condition. This may or may not be compatible with the ten
Dirichlet/Neumann boundary conditions we have already imposed.
It turns out that there is a nice resolution to this puzzle. One can show that, if the
boundary conditions are smooth, the minimum surface obeying these ten boundary conditions
can terminate at y = 0 only if the constraint ẋ 2 = ẏ 2 is satisfied [5]. This can be shown by
using the Hamilton–Jacobi equation for a minimum surface in AdS. This fits nicely with the
1230 H Ooguri

fact we saw earlier in the field theory point of view; the Wilson loop that is generated by a
trajectory the W-boson obeys the same constraint ẋ 2 = ẏ 2 . This resolves the puzzle but it
raises another question about how to define the Wilson loop operator which does not obey the
constraint. We will return to this question later.

3.2. Legendre transformation

Once we find the minimum surface obeying the boundary condition, we can compute the value
of its classical action. In the semiclassical approximation, the vacuum expectation value of
the Wilson loop is given by the exponential of the action for the minimum surface. There is
a question of which action to use. A naive guess would be the area of the surface, namely
the Nambu–Goto action. This would be appropriate if we were solving the fully Dirichlet
problem. In the fully Dirichlet problem, the boundary loop is fixed in the target space, and
there is a well defined area for each surface. The area, however, is not an appropriate action
functional for the Neumann problem. Since the Neumann problem fixes the string momentum,
rather than the location of the loop at the boundary, the area for the surface is not well defined.
The appropriate action for the Neumann problem is the Legendre transform of the area, which
we denote by Ã.
I
à = A − ds Pi Y i . (16)

à obtained by the Legendre transformation is a good functional of the string momentum.

There is a bonus in performing this Legendre transformation. Since the metric in AdS
diverges near the boundary, the area of the minimum surface is infinite if the surface terminates
at the boundary of AdS. To regularize this, we introduce a cut-off  in the y-coordinates. The
boundary of AdS is at y = 0, and the regularized area is given by an integral in the region
y > . By now, it is well known that this infrared regularization in AdS corresponds to the
ultraviolet regularization in the gauge theory [16]. So we use the same symbol  for both the
UV cut-off of the gauge theory and the IR cut-off of the string theory in AdS. If the loop is
smooth, the area is linearly divergent and the divergence is proportional to the circumference
of the loop [11].
p I
g2 N

A= ds |ẋ| + finite . (17)

It turns out that the Legendre transformation precisely cancels this linear divergence, leaving
the Legendre transformed action à to be finite. Therefore, the vacuum expectation value of
the Wilson loop, which is given by the exponential of Ã, is finite in this case. Earlier, we have
seen in the gauge theory that the Wilson loop is perturbatively finite when ẋ 2 = ẏ 2 . The fact
that à is finite fits well with this gauge theory result.

3.3. Examples
There are several types of Wilson loops for which solutions to the corresponding minimum
surface problems can be found explicitly and the areas of the surfaces can be computed
analytically.
The first example is the parallel Wilson lines. This was studied by Maldacena [11] and
by Rey and Yee [12]. By computing the area of the minimum surface connecting the Wilson
lines and by performing the Legendre transformation, one finds
√
p 4π 2 L
à = g N 1 4 ,
2 (18)
0( 4 ) R
 Wilson loops in large-N theories 1231

where L is the length of the Wilson lines and R is the distance between them. This expression is
for L  R. The parallel Wilson lines compute the potential between the quark and antiquark.
This result shows that the potentialpgoes as 1/R, as expected from the conformal invariance, and
the coefficient is proportional to g 2 N. It is interesting to compare this with the perturbative
computation. When g 2 N is small, the quark–antiquark potential is proportional to g 2 N due
2 2
p that gluon exchange. Somehow when g N becomes large, this g N behaviour turns into
to
2
g N . One may view this as a prediction of the AdS–CFT correspondence, which can in
principle be tested by a field theory computation at large N.
We can also find a minimum surface corresponding to a circular Wilson loop. The area of
the surface, after performing the Legendre transformation, turns out to be independent of the
radius of the circle, and the vacuum expectation value of the loop is given by

p 
hW i = exp g2 N . (19)

In these two cases, the Wilson loops are finite as we expect for smooth loops.
Another case for which we can find a minimum surface is a loop with a cusp. Near the
cusp singularity, the geometry is scale invariant and we can integrate the equation of motion
using the elliptic integrals. In this case, the divergence of the area is not precisely cancelled
by the Legendre transformation, and the logarithmic divergence remains. Once again, this is
similar to what we saw in the gauge theory side (6). The coefficient in front of the logarithm
is different from the perturbative result, however.

3.4. Loops with ẋ 2 6= ẏ 2

We have seen that, when the Wilson loop operator obeys the constraint ẋ 2 = ẏ 2 , we can
evaluate its vacuum expectation value at large g 2 N by computing the area of the minimum
surface in AdS. Its vacuum expectation value is the exponential of the Legendre transform of
the area, and it is finite when the loop is smooth.
It is then natural to ask how to compute the loop which does not satisfy the constraint. For
the boundary conditions which do not satisfy ẋ 2 = ẏ 2 , there is no minimum surface ending on
the boundary of AdS. So one may say that the vacuum expectation value of such a Wilson loop
should be zero. This is a reasonable answer. In fact, in other cases such as finite-temperature
theories, such an answer gave results consistent with what we expect for gauge theories [17].
For some problems, however, we need more detailed information than simply stating
hW i = 0 for ẋ 2 6= ẏ 2 at large g 2 N. Suppose, for example, we want to see whether the Wilson
loop computed in this way gives a solution to the large-N loop equation (10). When the loop
is smooth and without intersections, the equation is simply LhW i = 0 and this is satisfied by
any smooth functional of the loop. Non-trivial checks of the loop equation, therefore, have to
involve loops with cusps or intersections. For a loop with a cusp, however, a minimum surface
which can end at the boundary of AdS violates the condition ẋ 2 = ẏ 2 [5], while the loop
equation (10) is derived for loops obeying ẋ 2 = ẏ 2 . Thus, in order to test the loop equation,
we need more refined knowledge on the vacuum expectation value of such Wilson loops.
The perturbative computation suggests that loops not obeying the constraint are ultraviolet
divergent. In analogy with the distinction between chiral primary fields and non-chiral fields
in gauge theory, we expect that computation of the vacuum expectation value for loops with
ẋ 2 6= ẏ 2 requires a better understanding of stringy corrections in AdS.
1232 H Ooguri

4. Comments

At the end of my presentation at Strings ’99, I was asked whether the Wilson loop operator W
as in (4) is well defined in the Euclidean quantum field theory. Do we know that the functional
integral for hW i is convergent? Since the scalar field φi in the exponent comes with the
real coefficient ẏ i , the functional integral would be convergent only if the distribution of
the eigenvalues of φi decays sufficiently quickly for large eigenvalues. Another audience
commented that, since the Wilson loop with the constraint ẋ 2 = ẏ 2 is BPS-like (it is a phase
factor associated with a trajectory of the W-boson, which is a BPS particle in the N = 4
theory), it is likely that such an operator makes sense, and so does that with ẋ 2 > ẏ 2 since the
effect of φi in the exponent would be weaker. On the other hand, one may question whether
an operator with ẋ 2 < ẏ 2 exists.
In fact, the AdS–CFT correspondence suggests that operators with ẋ 2 = ẏ 2 , ẋ 2 > ẏ 2
and ẋ 2 < ẏ 2 behave differently. As I pointed out, the minimum surface can terminate at
the boundary of AdS at y = 0 only if the constraint ẋ 2 = ẏ 2 is satisfied. The AdS–
CFT correspondence then gives a definite prescription to compute hW i using the Legendre
transform of the area of the minimum surface. When ẋ 2 > ẏ 2 , we can still find a minimum
surface obeying the boundary conditions (15), except that the surface ends somewhere in the
interior of AdS rather than at the boundary. One may therefore hope to compute hW i using
such a minimum surface. On the other hand, in the case of ẋ 2 < ẏ 2 , there is no solution to the
minimum surface problem even if we relax the condition that the surface should terminate at
y = 0. This may be viewed as an indication that the loop operator for ẋ 2 < ẏ 2 is problematic.
It would be interesting to study properties of such loops from the point of view of the gauge
theory and to see how they fit with these behaviours of minimum surfaces.
To conclude, the Wilson loop provides us with a window for observing the stringy nature
of the correspondence between gauge theory and string theory. In the N = 4 gauge theory
in four dimensions, we have understood various aspects of loops which obey the constraint,
ẋ 2 = ẏ 2 . I think that finding a way to study loops without the constraint would teach us more
about gauge theory and string theory.

Acknowledgments

It is my pleasure to thank the organizers of Strings ’99 for giving me the opportunity to present
this work at the conference and for their hospitality. I would like to thank Nadav Drukker and
David Gross for their collaboration on this work. This research is supported in part by NSF
grant PHY-95-14797 and DOE grant DE-AC03-76SF00098.

References

[1] Maldacena J 1998 Adv. Theor. Math. Phys. 2 231

(Maldacena J 1997 Preprint hep-th/9711200)
[2] Gubser S S, Klebanov I R and Polyakov A M 1998 Phys. Lett. B 428 105
(Gubser S S, Klebanov I R and Polyakov A M 1998 Preprint hep-th/9802109)
[3] Witten E 1998 Adv. Theor. Math. Phys. 2 253
(Witten E 1998 Preprint hep-th/9802150)
[4] Aharony O, Gubser S S, Maldacena J, Ooguri H and Oz Y 2000 Phys. Rep. 323 183
(Aharony O, Gubser S S, Maldacena J, Ooguri H and Oz Y 1999 Preprint hep-th/9905111)
[5] Drukker N, Gross D J and Ooguri H 1999 Phys. Rev. D 60 125006
(Drukker N, Gross D J and Ooguri H 1999 Preprint hep-th/9904191)
[6] Berenstein D, Corrado R, Fischler W and Maldacena J 1999 Phys. Rev. D 59 105023
(Berenstein D, Corrado R, Fischler W and Maldacena J 1998 Preprint hep-th/9809188)
 Wilson loops in large-N theories 1233

[7] Graham C R and Witten E 1999 Nucl. Phys. B 546 52

(Graham C R and Witten E 1999 Preprint hep-th/9901021)
[8] Migdal A A 1983 Phys. Rep. 102 199
[9] Polyakov A M 1987 Gauge Fields and Strings (Geneva: Harwood) section 7.2
[10] Drukker N 1999 J. High Energy Phys. JHEP11(1999)006
(Drukker N 1999 Preprint hep-th/9908113)
[11] Maldacena J 1998 Phys. Rev. Lett. 80 4859
(Maldacena J 1998 Preprint hep-th/9803002)
[12] Rey S-J and Yee J 1998 Preprint hep-th/9803001
[13] Anderson M 1982 Invent. Math. 69 477
[14] Lin F-H 1989 Invent. Math. 96 593
[15] Tonegawa Y 1996 Math. Z. 221 591
[16] Susskind L and Witten E 1998 Preprint hep-th/9805114
[17] Aharony O and Witten E 1998 J. High Energy Phys. JHEP11(1998)018
(Aharony O and Witten E 1998 Preprint hep-th/9807205)