Joao Penedones PDF

TASI Lectures on AdS/CFT
arXiv:1608.04948v1 [hep-th] 17 Aug 2016
João Penedones
[email protected]
Fields and Strings Laboratory, Institute of Physics, EPFL
Rte de la Sorge, BSP 728, CH-1015 Lausanne, Switzerland
Centro de Fı́sica do Porto, Universidade do Porto, Portugal
Theoretical Physics Department, CERN, Geneva, Switzerland
We introduce the AdS/CFT correspondence as a natural extension of

QFT in a fixed AdS background. We start by reviewing some general
concepts of CFT, including the embedding space formalism. We then
consider QFT in a fixed AdS background and show that one can define
boundary operators that enjoy very similar properties as in a CFT, ex-
cept for the lack of a stress tensor. Including a dynamical metric in AdS
generates a boundary stress tensor and completes the CFT axioms. We
also discuss some applications of the bulk geometric intuition to strongly
coupled QFT. Finally, we end with a review of the main properties of
Mellin amplitudes for CFT correlation functions and their uses in the
context of AdS/CFT.
TASI Lectures on AdS/CFT 1
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Conformal Field Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1. Conformal Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2. Local Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3. Ward identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4. State-Operator Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5. Operator Product Expansion . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.6. Conformal Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.7. Embedding Space Formalism . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8. Large N Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3. Anti-de Sitter Spacetime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1. Particle dynamics in AdS . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2. Quantum Field Theory in AdS . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3. Gravity with AdS boundary conditions . . . . . . . . . . . . . . . . . . . . 28
4. The AdS/CFT Correspondence . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1. Quantum Gravity as CFT . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2. String Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3. Finite Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.4. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5. Mellin amplitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.1. Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2. OPE ) Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.3. Holographic CFTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.4. Open questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2 João Penedones
1. Introduction
The AdS/CFT correspondence [1–3] is a well established general approach

to quantum gravity. However, it is often perceived as a particular con-
struction specific to string theory. In these lectures I will argue that the
AdS/CFT correspondence is the most conservative approach to quantum
gravity. The quick argument goes as follows:
• System in a box - we work with Anti-de Sitter (AdS) boundary con-

ditions because AdS is the most symmetric box with a boundary.
This is useful to control large IR e↵ects, even without dynamical
gravity.
• QFT in the box - Quantum Field Theory (no gravity) in a fixed AdS
background leads to the construction of boundary operators that
enjoy an associative and convergent Operator Product Expansion
(OPE). The AdS isometries act on the boundary operators like the
conformal group in one lower dimension.
• Boundary stress-tensor from gravitons - perturbative metric fluctu-
ations around AdS lead to a boundary stress tensor (weakly coupled
to the other boundary operators).
Starting from these 3 facts it is entirely natural to define quantum gravity

with AdS boundary conditions as Conformal Field Theory (CFT) in one
lower dimension. Of course not all CFTs look like gravity in our universe.
That requires the size of the box to be much larger than the Planck length
and all higher spin particles to be very heavy (relative to the size of the
box). As we shall see, these physical requirements imply that the CFT is
strongly coupled and therefore hard to find or construct. The major role of
string theory is to provide explicit examples of such CFTs like maximally
supersymmetric Yang-Mills (SYM) theory.
There are many benefits that follow from accepting the AdS/CFT per-
spective. Firstly, it makes the holographic nature of gravity manifest. For
example, one can immediately match the scaling of the CFT entropy den-
sity with the Bekenstein-Hawking entropy of (large) black holes in AdS.
Notice that this is a consequence because it was not used as an argument
for AdS/CFT in the previous paragraph. More generally, the AdS/CFT
perspective let us translate questions about quantum gravity into math-
ematically well posed questions about CFT. a Another benefit of the
a It
might not be possible to formulate all quantum gravity questions in CFT language.
For example, it is unclear if the experience of an observer falling into a black hole in AdS
gauge/gravity duality is that it gives us a geometric description of QFT

phenomena, which can be very useful to gain physical intuition and to
create phenomenological models.
This introduction to AdS/CFT will not follow the historical order of
scientific developments. Section 2 reviews general concepts in CFT. This
part is not entirely self contained because this topic is discussed in detail in
the chapter Conformal Bootstrap by David Simmons-Duffin [5]. b The main
purpose of this section is to set up notation, introduce the embedding space
formalism and discuss large N factorization. Section 3 deals with Anti-
de Sitter (AdS) spacetime. The first goal here is to gain intuition about
particle dynamics in AdS and QFT in a fixed AdS background. From this
point-of-view, we will see that a gravitational theory with AdS boundary
conditions naturally defines a CFT living on its boundary. In section 4,
we discuss the AdS/CFT correspondence in more detail and emphasize its
importance for quantum gravity. We also consider what kind of CFTs have
simple AdS duals and the role of string theory. Furthermore, we discuss
several applications of the gauge/gravity duality as a tool to geometrize
QFT e↵ects. Finally, in section 5, we introduce the Mellin representation
of CFT correlation functions. We explain the analytic properties of Mellin
amplitudes and their particular simplicity in the case of holographic CFTs.
There are many reviews of AdS/CFT available in the literature. Most
of them are complementary to these lecture notes because they discuss
in greater detail concrete examples of AdS/CFT realized in string theory.
I leave here an incomplete list [8–16] that can be useful to the readers
interested in knowing more about AdS/CFT. The lecture notes [17] by
Jared Kaplan discuss in greater detail many of the ideas presented here.
2. Conformal Field Theory
This section briefly describes the basic concepts necessary to formulate a

non-perturbative definition of CFT. In the last part, we explain in more
detail the embedding space formalism for CFT and ’t Hooft’s large N ex-
pansion, which will be very important in the following sections.
is a CFT observable [4].

b See also the lecture notes [6, 7].
4 João Penedones
2.1. Conformal Transformations
For simplicity, in most formulas, we will consider Euclidean signature. We

start by discussing conformal transformations of Rd in Cartesian coordi-
nates,
ds2 = µ⌫ dx
µ
dx⌫ . (1)
A conformal transformation is a coordinate transformation that preserves

the form of the metric tensor up to a scale factor,
dx̃µ dx̃⌫
µ⌫ = ⌦2 (x) ↵ . (2)
dx↵ dx
In other words, a conformal transformation is a local dilatation.
Exercise 2.1. Show that, for d > 2, the most general infinitesimal confor-
mal transformation is given by x̃µ = xµ + ✏µ (x) with
✏µ (x) = aµ + xµ + mµ⌫ x⌫ + x2 bµ 2x↵ b↵ xµ . (3)
In spacetime dimension d > 2, conformal transformations form the

group SO(d + 1, 1). The generators Pµ and Mµ⌫ correspond to transla-
tion and rotations and they are present in any relativistic invariant QFT.
In addition, we have the generators of dilatations D and special confor-
mal transformations Kµ . It is convenient to think of the special conformal
transformations as the composition of an inversion followed by a translation
followed by another inversion. Inversion is the conformal transformationc
xµ
xµ ! . (4)
x2
Exercise 2.2. Verify that inversion is a conformal transformation.
The form of the generators of the conformal algebra acting on functions

can be obtained from

i
(xµ + ✏µ (x)) = 1 + i aµ Pµ D + mµ⌫ Mµ⌫ + i bµ Kµ (xµ ) , (5)
2
c Inversion is outside the component of the conformal group connected to the identity.
Thus, it is possible to have CFTs that are not invariant under inversion. In fact, CFTs
that break parity also break inversion.
d
which leads to
Pµ = i@µ , D= xµ @ µ , (6)
⌫ 2
Mµ⌫ = i (xµ @⌫ x⌫ @ µ ) , Kµ = 2ixµ x @⌫ i x @µ . (7)
Exercise 2.3. Show that the generators obey the following commutation
relations
[D, Pµ ] = Pµ , [D, Kµ ] = Kµ , [Kµ , P⌫ ] = 2 µ⌫ D 2i Mµ⌫ ,

[Mµ⌫ , P↵ ] = i ( µ↵ P⌫ ⌫↵ Pµ ) , [Mµ⌫ , K↵ ] = i ( µ↵ K⌫ ⌫↵ Kµ ) ,
[M↵ , Mµ⌫ ] = i ( ↵µ M ⌫ + ⌫ M↵µ µ M↵⌫ ↵⌫ M µ ) . (8)
2.2. Local Operators
Local operators are divided into two types: primary and descendant. De-
scendant operators are operators that can be written as (linear combina-
tions of) derivatives of other local operators. Primary operators can not be
written as derivatives of other local operators. Primary operators at the
origin are annihilated by the generators of special conformal transforma-
tions. Moreover, they are eigenvectors of the dilatation generator and form
irreducible representations of the rotation group SO(d),
B
[Kµ , O(0)] = 0 , [D, O(0)] = O(0) , [Mµ⌫ , OA (0)] = [Mµ⌫ ]A OB (0) .
Correlation functions of scalar primary operators obey

1 n
@ x̃ d
@ x̃ d
hO1 (x̃1 ) . . . On (x̃n )i = ... hO1 (x1 ) . . . On (xn )i (9)
@x x1 @x xn
for all conformal transformations x ! x̃. As explained above, it is sufficient

to impose Poincaré invariance and this transformation rule under inversion,
⌧ ✓ ◆ ✓ ◆
x1 xn
= x21 . . . x2n
1 n
O1 2 . . . O n hO1 (x1 ) . . . On (xn )i .
x1 x2n
This implies that vacuum one-point functions hO(x)i vanish except for
the identity operator (which is the unique operator with = 0). It also
d We define the dilatation generator D in a non-standard fashion so that it has real
eigenvalues in unitary CFTs.
6 João Penedones
fixes the form of the two and three point functions,

ij
hOi (x)Oj (y)i = 2 i
, (10)
(x y)
C123
hO1 (x1 )O2 (x2 )O3 (x3 )i = 1+ 2 3 1+ 3 2 2+ 3 1
,
|x12 | |x13 | |x23 |
where we have normalized the operators to have unit two point function.
The four-point function is not fixed by conformal symmetry because
with four points one can construct two independent conformal invariant
cross-ratios
x212 x234 x214 x223
u = z z̄ = , v = (1 z)(1 z̄) = . (11)
x213 x224 x213 x224
The general form of the four point function is
A(u, v)
hO(x1 ) . . . O(x4 )i = . (12)
(x213 x224 )
2.3. Ward identities

To define the stress-energy tensor it is convenient to consider the theory in
a general background metric gµ⌫ . Formally, we can write
1
ˆ
hO1 (x1 ) . . . On (xn )ig = [d ]e S[ ,g] O1 (x1 ) . . . On (xn ) , (13)
Z[g]
where Z[g] = [d ]e S[ ,g] is the partition function for the background
´
metric gµ⌫ . Recalling the classical definition

2 S
T µ⌫ (x) = p , (14)
g gµ⌫ (x)
it is natural to define the quantum stress-energy tensor operator via the
equation
Z[g + g] 1 p
ˆ
=1+ dx g gµ⌫ (x) hT µ⌫ (x)ig + O( g 2 ) , (15)
Z[g] 2
and
hO1 (x1 ) . . . On (xn )ig+ g hO1 (x1 ) . . . On (xn )ig
1
ˆ
p h
= dx g gµ⌫ (x) hT µ⌫ (x)O1 (x1 ) . . . On (xn )ig (16)
2
i
hT µ⌫ (x)ig hO1 (x1 ) . . . On (xn )ig + O( g 2 ) .
Under an infinitesimal coordinate transformation x̃µ = xµ + ✏µ (x), the

metric tensor changes g̃µ⌫ = gµ⌫ rµ ✏⌫ r⌫ ✏µ but the physics should
remain invariant. In particular, the partition function Z[g] = Z[g̃] and the
correlation functions e
hO1 (x̃1 ) . . . On (x̃n )ig̃ = hO1 (x1 ) . . . On (xn )ig , (17)
do not change. This leads to the conservation equation hrµ T µ⌫ (x)ig and
n
X @
✏µ (xi ) hO1 (x1 ) . . . On (xn )ig (18)
i=1
@xµi
p
ˆ
= dx g✏⌫ (x) hrµ T µ⌫ (x)O1 (x1 ) . . . On (xn )ig
for all ✏µ (x) that decays sufficiently fast at infinity. Thus rµ T µ⌫ = 0 up to

contact terms.
Correlation functions of primary operators transform homogeneously
under Weyl transformations of the metric f
hO1 (x1 ) . . . On (xn )ig

hO1 (x1 ) . . . On (xn )i⌦2 g = 1 n
. (19)
[⌦(x1 )] . . . [⌦(xn )]
Exercise 2.4. Show that this transformation rule under local rescalings of
the metric (together with coordinate invariance) implies (9) under confor-
mal transformations.
Consider now an infinitesimal Weyl transformation ⌦ = 1 + !, which

corresponds to a metric variation gµ⌫ = 2!gµ⌫ . From (16) and (19) we
conclude that
n
X
i !(xi ) hO1 (x1 ) . . . On (xn )ig
i=1
ˆ
p h
= dx g !(x)gµ⌫ hT µ⌫ (x)O1 (x1 ) . . . On (xn )ig (20)
i
hT µ⌫ (x)ig hO1 (x1 ) . . . On (xn )ig .
e If the operators are not scalars (e.g. if they are vector operators) then one also needs
to take into account the rotation of their indices.
f In general, the partition fungion is not invariant in even dimensions. This is the Weyl
anomaly Z[⌦2 g] = Z[g]e SW eyl [⌦,g] .

8 João Penedones
Consider the following codimension 1 integral over the boundary of a

region B, g
ˆ h
I= dSµ ✏⌫ (x) hT µ⌫ (x)O1 (x1 ) . . . On (xn )ig (21)
@B
i
hT µ⌫ (x)ig hO1 (x1 ) . . . On (xn )ig .
One can think of this as the total flux of the current ✏⌫ T µ⌫ , where ✏⌫ (x) is
an infinitesimal conformal transformation. Gauss law tells us that this flux
should be equal to the integral of the divergence of the current
1
rµ (✏⌫ T µ⌫ ) = ✏⌫ rµ T µ⌫ + rµ ✏⌫ T µ⌫ = ✏⌫ rµ T µ⌫ + r↵ ✏↵ gµ⌫ T µ⌫ , (22)
d
where we used the symmetry of the stress-energy tensor T µ⌫ = T ⌫µ and
the definition of an infinitesimal conformal transformation rµ ✏⌫ + r⌫ ✏µ =
2 ↵
d r↵ ✏ gµ⌫ . Using Gauss law and (18) and (20) we conclude that
X  @ i
I= ✏µ (xi ) µ + r↵ ✏↵ (xi ) hO1 (x1 ) . . . On (xn )ig . (23)
@xi d
xi 2B
The equality of (21) and (23) for any infinitesimal conformal transformation
(3) is the most useful form of the conformal Ward identities.
Exercise 2.5. Conformal symmetry fixes the three-point function of a spin

2 primary operator and two scalars up to an overall constant, h
H µ⌫ (x1 , x2 , x3 )
hO(x1 )O(x2 )T µ⌫ (x3 )i = C12T , (24)
|x12 |2 d+2 |x |d 2 |x |d
13 23
2
where
1 xµ13 xµ23
H µ⌫ = V µ V ⌫ V↵ V ↵ µ⌫
, Vµ = . (25)
d x213 x223
Write the conformal Ward identity (21)=(23) for the three point function
hT µ⌫ (x)O(0)O(y)i for the case of an infinitesimal dilation ✏µ (x) = xµ and
with the surface @B being a sphere centred at the origin and with radius
smaller than |y|. Use this form of the conformal Ward identity in the limit
of an infinitesimally small sphere @B and formula (24) for the three point
function to derive
d 1
COOT = , (26)
d 1 Sd
g Inthe notation of the Conformal Bootstrap chapter [5] this is the topological operator
Q✏ [@B] inserted in the correlation function hO1 (x1 ) . . . On (xn )ig .
h You can try to derive this formula using the embedding space formalism of section 2.7.
2⇡ d/2
where Sd = (d/2) is the volume of a (d 1)-dimensional unit sphere.
2.4. State-Operator Map

Consider Rd in spherical coordinates. Writing the radial coordinate as
r = e⌧ we find
ds2 = dr2 + r2 d⌦2d 1 = e2⌧ d⌧ 2 + d⌦2d 1 . (27)
Thus, the cylinder R ⇥ S d 1
can be obtained as a Weyl transformation of
euclidean space Rd .
Exercise 2.6. Compute the two-point function of a scalar primary operator
on the cylinder using the Weyl transformation property (19).
A local operator inserted at the origin of Rd prepares a state at ⌧ = 1
on the cylinder. On the other hand, a state on a constant time slice of
the cylinder can be propagated backwards in time until it corresponds to
a boundary condition on a arbitrarily small sphere around the origin of
Rd , which defines a local operator. Furthermore, time translations on the
cylinder correspond to dilatations on Rd . This teaches us that the spectrum
of the dilatation generator on Rd is the same as the energy spectrum for
the theory on R ⇥ S d 1 . i
2.5. Operator Product Expansion

The Operator Product Expansion (OPE) between two scalar primary op-
erators takes the following form
2 3
X 6 7
Oi (x)Oj (0) = Cijk |x| k i j 4Ok (0) + xµ @µ Ok (0) + . . .5 (28)
k
| {z }
descendants
where denotes a number determined by conformal symmetry. For sim-

plicity we show only the contribution of a scalar operator Ok . In general,
in the OPE of two scalars there are primary operators of all spins.
Exercise 2.7. Compute by using this OPE inside a three-point function.
The OPE has a finite radius of convergence inside correlation functions.
This follows from the state operator map with an appropriate choice of
origin for radial quantization.
iMore precisely, there can be a constant shift equal to the Casimir energy of the vacuum
on S d 1 , which is related with the Weyl anomaly. In d = 2, this gives the usual energy
c 1
spectrum 12 L
where c is the central charge and L is the radius of S 1 .
10 João Penedones
2.6. Conformal Bootstrap
Using the OPE successively one can reduce any n point function to a
sum of one-point functions, which all vanish except for the identity oper-
ator. Thus, knowing the operator content of the theory, i.e. the scaling
dimensions and SO(d) irreps R of all primary operators, and the OPE
coefficients Cijk ,j one can determine all correlation functions of local oper-
ators. This set of data is called CFT data because it essentially defines the
theory. k The CFT data is not arbitrary, it must satisfy several constraints:
• OPE associativity - Di↵erent ways of using the OPE to compute

a correlation function must give the same result. This leads to the
conformal bootstrap equations described below.
• Existence of stress-energy tensor - The stress-energy tensor
Tµ⌫ is a conserved primary operator (with = d) whose correlation
functions obey the conformal Ward identities.
• Unitarity - In our Euclidean context this corresponds to reflection
positivity and it implies lower bounds on the scaling dimensions.
It also implies that one can choose a basis of real operators where
all OPE coefficients are real. In the context of statistical physics,
there are interesting non-unitary CFTs.
It is sufficient to impose OPE associativity for all four-point functions of

the theory. For a four-point function of scalar operators, the bootstrap
equation reads
X (12)(34)
X (13)(24)
C12k Ck34 G k ,lk
(x1 , . . . , x4 ) = C13q Cq24 G q ,lq
(x1 , . . . , x4 ) ,
k q
where G ,l are conformal blocks, which encode the contribution from a

primary operator of dimension and spin l and all its descendants.
j Forprimary operators O1 , O2 , O3 transforming in non-trivial irreps of SO(d) there are

several OPE coefficients C123 . The number of OPE coefficients C123 is given by the
number of symmetric traceless tensor representations that appear in the tensor product
of the 3 irreps of SO(d) associated to O1 , O2 and O3 .
k However, there are observables besides the vacuum correlation functions of local opera-
tors. It is also interesting to study non-local operators (line operators, surface operators,
boundary conditions, etc) and correlation functions in spaces with non-trivial topology
(for example, correlators at finite temperature).
2.7. Embedding Space Formalism

The conformal group SO(d + 1, 1) acts naturally on the space of light rays
through the origin of Rd+1,1 ,
2 2 2
P0 + P1 + · · · + P d+1 =0. (29)
A section of this light-cone is a d dimensional manifold where the CFT
lives. For example, it is easy to see that the Poincaré section P 0 +P d+1 = 1
is just Rd . To see this parametrize this section using
1 + x2 1 x2
P 0 (x) = , P µ (x) = xµ , P d+1 (x) = , (30)
2 2
with µ = 1, . . . , d and xµ 2 Rd and compute the induced metric. In fact, any
conformally flat manifold can be obtained as a section of the light-cone in
the embedding space Rd+1,1 . Using the parametrization P A = ⌦(x)P A (x)
with xµ 2 Rd , one can easily show that the induced metric is simply given
by ds2 = ⌦2 (x) µ⌫ dxµ dx⌫ . With this is mind, it is natural to extend a
primary operator from the physical section to the full light-cone with the
following homogeneity property
O( P ) = O(P ) , 2R. (31)
This implements the Weyl transformation property (19). One can then
compute correlation functions directly in the embedding space, where the
constraints of conformal symmetry are just homogeneity and SO(d + 1, 1)
Lorentz invariance. Physical correlators are simply obtained by restricting
to the section of the light-cone associated with the physical space of interest.
This idea goes back to Dirac [18] and has been further develop by many
authors [19–25].
Exercise 2.8. Rederive the form of two and three point functions of scalar
primary operators in Rd using the embedding space formalism.
Vector primary operators can also be extended to the embedding space.

In this case, we impose
P A OA (P ) = 0 , OA ( P ) = OA (P ) , 2R, (32)
and the physical operator is obtained by projecting the indices to the sec-
tion,
@P A
Oµ (x) = OA (P ) . (33)
@xµ P A =P A (x)
12 João Penedones
Notice that this implies a redundancy: OA (P ) ! OA (P ) + PA ⇤(P ) gives

rise to the same physical operator O(x), for any scalar function ⇤(P ) such
1
that ⇤( P ) = ⇤(P ). This redundancy together with the constraint
P A OA (P ) = 0 remove 2 degrees of freedom of the (d+2)-dimensional vector
OA .
Exercise 2.9. Show that the two-point function of vector primary operators
is given by
⌦ ↵ ⌘ AB (P1 · P2 ) P2A P1B
OA (P1 )OB (P2 ) = const +1
, (34)
( 2P1 · P2 )
up to redundant terms.
Exercise 2.10. Consider the parametrization P A = P 0 , P µ , P d+1 =

µ 0 2 d+1 2
(cosh ⌧, ⌦ , sinh ⌧ ) of the global section P P = 1, where ⌦µ
(µ = 1, . . . , d) parametrizes a unit (d 1) dimensional sphere, ⌦ · ⌦ = 1.
Show that this section has the geometry of a cylinder exactly like the one
used for the state-operator map.
Conformal correlation functions extended to the light-cone of R1,d+1 are

annihilated by the generators of SO(1, d + 1)
n
X (i)
JAB hO1 (P1 ) . . . On (Pn )i = 0 , (35)
i=1
(i)
where JAB is the generator
✓ ◆
@ @
JAB = i PA PB , (36)
@P B @P A
acting on the point Pi . For a given choice of light cone section, some
generators will preserve the section and some will not. The first are Killing
vectors (isometry generators) and the second are conformal Killing vectors.
The commutation relations give the usual Lorentz algebra
[JAB , JCD ] = i (⌘AC JBD + ⌘BD JAC ⌘BC JAD ⌘AD JBC ) . (37)
Exercise 2.11. Check that the conformal algebra (8) follows from (37) and
D= i J0,d+1 , Pµ = Jµ0 Jµ,d+1 ,
Mµ⌫ = Jµ⌫ , Kµ = Jµ0 + Jµ,d+1 . (38)
Exercise 2.12. Show that equation (35) for JAB = J0,d+1 implies time
translation invariance on the cylinder
Xn
@
hO1 (⌧1 , ⌦1 ) . . . On (⌧n , ⌦n )i = 0 , (39)
i=1
@⌧ i
and dilatation invariance on Rd

Xn ✓ ◆
µ @
i + xi hO1 (x1 ) . . . On (xn )i = 0 . (40)
i=1
@xµi
In this case, you will need to use the di↵erential form of the homogeneity
property P A @P@ A Oi (P ) = i Oi (P ). It is instructive to do this exercise
for the other generators as well.
2.8. Large N Factorization

Consider a U (N ) gauge theory with fields valued in the adjoint representa-
tion. Schematically, we can write the action as
N
ˆ h i
2
S= dx Tr (D ) + c3 3 + c4 4 + . . . (41)
where we introduced the ’t Hooft coupling = gY2 M N and ci are other

coupling constants independent of N . Following ’t Hooft [26], we consider
the limit of large N with kept fixed. The propagator of an adjoint field
obeys
⌦ i k↵ i k
j l / (42)
N l j
where we used the fact that the adjoint representation can be represented
as the direct product of the fundamental and the anti-fundamental repre-
sentation. This suggests that one can represent a propagator by a double
line, where each line denotes the flow of a fundamental index. Start by con-
sidering the vacuum diagrams in this language. A diagram with V vertices,
E propagators (or edges) and F lines (or faces) scales as
✓ ◆V ✓ ◆E ✓ ◆
N F N F
N = , (43)
N
where = V + F E = 2 2g is the minimal Euler character of the two
dimensional surface where the double line diagram can be embedded and
g is the number of handles of this surface. Therefore, the large N limit is
dominated by diagrams that can be drawn on a sphere (g = 0). These dia-
grams are called planar diagrams. For a given topology, there is an infinite
14 João Penedones
number of diagrams that contribute with increasing powers of the coupling

, corresponding to tesselating the surface with more and more faces. Fig-
ure 1 shows two examples of vacuum diagrams in the double line notation.
This topological expansion has the structure of string perturbation theory
with /N playing the role of the string coupling. As we shall see this is
precisely realized in maximally supersymmetric Yang-Mills theory (SYM).
V =2 V =4
E=3 E=6
F =3 F =2
g=0 g=1
Fig. 1. Vacuum diagrams in the double line notation. Interaction vertices are marked
with a small blue dot. The left diagram is planar while the diagram on the right has the
topology of a torus (genus 1 surface).
Let us now consider single-trace local operators of the form O =

cJ Tr J , where cJ is a normalization constant independent of N . Adapt-
ing the argument above, it is easyV to
= 2concludeVthat
= 4the connected correlators
are given by a large N expansion E =of3the formE = 6
F =X 31 F =2
hO1 . . . On ic = N 2 n 2g fg ( ) , (44)
g=0
which is dominated by the planars diagrams (g = 0). Moreover, we see

that the planar two-point function is independent of N while connected
higher point functions are suppressed by powers of N . This is large N
factorization. In particular it implies that the two-point function of a multi-
trace operator Õ(x) =: O1 (x) . . . Ok (x) : is dominated by the product of
the two-point functions of its single-trace constituents
D E Y 1
Õ(x)Õ(y) ⇡ hOi (x)Oi (y)i = P
2 i i
, (45)
i
(x y)
where we assumed that the single-trace operators were scalar conformal
primaries properly normalized. We conclude that the scaling dimension
P 2
of the multi-trace operator Õ is given by i i + O(1/N ) . In other
words, the space of local operators in a large N CFT has the structure of
a Fock space with single-trace operators playing the role of single particle
states of a weakly coupled theory. This is the form of large N factorization
relevant for AdS/CFT. However, notice that conformal invariance was not
important for the argument. It is well known that large N factorization

also occurs in confining gauge theories. Physically, it means that colour
singlets (like glueballs or mesons) interact weakly in large N gauge theories
(see [27] for a clear summary).
The stress tensor has a natural normalization that follows from the
action, Tµ⌫ ⇠ N Tr (@µ @⌫ ). This leads to the large N scaling
hTµ1 ⌫1 (x1 ) . . . Tµn ⌫n (xn )ic ⇠ N 2 , (46)
which will be important below. This normalization of Tµ⌫ is also fixed by
the Ward identities.
3. Anti-de Sitter Spacetime
Euclidean AdS spacetime is the hyperboloid

2 2 2
X0 + X1 + · · · + X d+1 = R2 , X0 > 0 , (47)
embedded in R d+1,1 0
. For large values of X this hyperboloid approaches
the light-cone of the embedding space that we discussed in section 2.7. It is
clear from the definition that Euclidean AdS is invariant under SO(d+1, 1).
The generators are given by
✓ ◆
@ @
JAB = i XA XB . (48)
@X B @X A
Poincaré coordinates are defined by
1 + x2 + z 2
X0 = R
2z
µ
x
Xµ = R (49)
z
d+1 1 x2 z 2
X =R
2z
where x 2 R and z > 0. In these coordinates, the metric reads
µ d
dz 2 + µ⌫ dx
µ
dx⌫
ds2 = R2 . (50)
z2
This shows that AdS is conformal to R+ ⇥ Rd whose boundary at z = 0 is
just Rd . These coordinates make explicit the subgroup SO(1, 1)⇥ISO(d) of
the full isometry group of AdS. These correspond to dilatation and Poincaré
symmetries inside the d dimensional conformal group. In particular, the
dilatation generator is
@ @ @ @
D = i J0,d+1 = X0 + Xd+1 = z xµ µ . (51)
@X d+1 @X 0 @z @x
16 João Penedones
Another useful coordinate system is

X 0 = R cosh ⌧ cosh ⇢
X µ = R ⌦µ sinh ⇢ (52)
X d+1 = R sinh ⌧ cosh ⇢
where ⌦µ (µ = 1, . . . , d) parametrizes a unit (d 1) dimensional sphere,
⌦ · ⌦ = 1. The metric is given by
⇥ ⇤
ds2 = R2 cosh2 ⇢ d⌧ 2 + d⇢2 + sinh2 ⇢ d⌦2d 1 . (53)
To understand the global structure of this spacetime it is convenient to
change the radial coordinate via tanh ⇢ = sin r so that r 2 [0, ⇡2 [. Then,
the metric becomes
R2 ⇥ 2 ⇤
ds2 = d⌧ + dr2 + sin2 r d⌦2d 1 , (54)
cos2 r
which is conformal to a solid cylinder whose boundary at r = ⇡2 is R⇥S d 1 .
@
In these coordinates, the dilatation generator D = i J0,d+1 = @⌧ is the
hamiltonian conjugate to global time.
3.1. Particle dynamics in AdS

For most purposes it is more convenient to work in Euclidean signature
and analytically continue to Lorentzian signature only at the end of the
calculation. However, it is important to discuss the Lorentzian signature to
gain some intuition about real time dynamics. In this case, AdS is defined
by the universal cover of the manifold
2 2 2 2
X0 + X1 + · · · + Xd X d+1 = R2 , (55)
embedded in Rd,2 . The universal cover means that we should unroll the
non-contractible (timelike) cycle. To see this explicitly it is convenient to
introduce global coordinatesl
X 0 = R cos t cosh ⇢
X µ = R ⌦µ sinh ⇢ (56)
d+1
X = R sin t cosh ⇢
where ⌦µ (µ = 1, . . . , d) parametrizes a unit (d 1) dimensional sphere.
The original hyperboloid is covered with t 2 [0, 2⇡[ but we consider t 2 R.
l Noticethat this is just the analytic continuation ⌧ ! i t and X d+1 ! i X d+1 of the
Euclidean global coordinates (52).
The metric is given by

⇥ ⇤
ds2 = R2 cosh2 ⇢ dt2 + d⇢2 + sinh2 ⇢ d⌦2d 1 . (57)
To understand the global structure of this spacetime it is convenient to

change the radial coordinate via tanh ⇢ = sin r so that r 2 [0, ⇡2 [. Then,
the metric becomes
R2 ⇥ ⇤
ds2 = dt2 + dr2 + sin2 r d⌦2d 1 , (58)
cos2 r
which is conformal to a solid cylinder whose boundary at r = ⇡2 is R⇥S d 1 .
Geodesics are given by the intersection of AdS with 2-planes through the
origin of the embedding space. In global coordinates, the simplest timelike
geodesic describes a particle sitting at ⇢ = 0. This corresponds to (the
universal cover of) the intersection of X µ = 0 for µ = 1, . . . , d with the
hyperboloid (55). Performing a boost in the X 1 , X d+1 plane we can obtain
an equivalent timelike geodesic X 1 cosh = X d+1 sinh and X µ = 0 for
µ = 2, . . . , d. In global coordinates, this gives an oscillating trajectory
tanh ⇢ = tanh sin t , (59)
with period 2⇡. In fact, all timelike geodesics oscillate with period 2⇡ in
global time. One can say AdS acts like a box that confines massive particles.
However, it is a very symmetric box that does not have a center because
all points are equivalent.
Null geodesics in AdS are also null geodesics in the embedding space.
For example, the null ray X d+1 X 1 = X 0 R = X µ = 0 for µ = 2, . . . , d is
1
a null ray in AdS which in global coordinates is given by cosh ⇢ = cos t . This
describes a light ray that passes through the origin at t = 0 and reaches
the conformal boundary ⇢ = 1 at t = ± ⇡2 . All light rays in AdS start and
end at the conformal boundary traveling for a global time interval equal to
⇡.
One can also introduce Poincaré coordinates
xµ
Xµ = R
z
R 1 x2 z 2
Xd = (60)
2 z
R 1 + x2 + z 2
X d+1 =
2 z
where now µ = 0, 1, . . . , d 1 and x2 = ⌘µ⌫ xµ x⌫ . However, in Lorentzian
signature, Poincaré coordinates do not cover the entire spacetime. Surfaces
18 João Penedones
of constant z approach the light-like surface X d + X d+1 = 0 when z ! 1.

This null surface is often called the Poincaré horizon.
We have seen that AdS acts like a box for classical massive particles.
Quantum mechanically, this confining potential gives rise to a discrete en-
ergy spectrum. Consider the Klein-Gordon equation
r2 = m2 , (61)
in global coordinates (57). In order to emphasize the correspondence with

CFT we will solve this problem using an indirect route. Firstly, consider
the action of the quadratic Casimir of the AdS isometry group on a scalar
field
1 ⇥ ⇤
JAB J BA = X 2 @X
2
+ X · @X (d + X · @X ) . (62)
2
Formally, we are extending the function from AdS, defined by the hyper-
surface X 2 = R2 , to the embedding space. However, the action of the
quadratic Casimir is independent
⇥ of this extension
⇤ because the generators
2 2
JAB are interior to AdS, i.e. JAB , X + R = 0. If we foliate the embed-
ding space Rd,2 with AdS surfaces of di↵erent radii R, we obtain that the
laplacian in the embedding space can be written as
2 1 @ d+1 @
@X = R + r2AdS . (63)
Rd+1 @R @R
Substituting this in (62) and noticing that X · @X = R@R we conclude that
1
JAB J BA = R2 r2AdS . (64)
2
Therefore, we should identify m2 R2 with the quadratic Casimir of the con-

formal group.
The Lorentzian version of the conformal generators (38) is
D= J0,d+1 , Pµ = Jµ0 + i Jµ,d+1 , (65)

Mµ⌫ = Jµ⌫ , Kµ = Jµ0 i Jµ,d+1 . (66)
Exercise 3.1. Show that, in global coordinates, the conformal generators

take the form

✓ ◆
@ @ @
D=i , Mµ⌫ = i ⌦µ ⌫ ⌦⌫ ,
@t @⌦ @⌦µ

1
Pµ = ie it ⌦µ (@⇢ i tanh ⇢ @t ) + rµ ,
tanh ⇢

1
Kµ = ieit ⌦µ ( @⇢ i tanh ⇢ @t ) rµ ,
tanh ⇢
@ @
where rµ = @⌦µ ⌦µ ⌦⌫ @⌦ ⌫ is the covariant derivative on the unit sphere
Sd 1.
In analogy with the CFT construction we can look for primary states,
which are annihilated by Kµ and are eigenstates of the hamiltonian, D =
. The condition Kµ = 0 splits in one term proportional to ⌦µ and one
term orthogonal to ⌦µ . The second term implies that is independent of
the angular variables ⌦µ . The first term gives (@⇢ + tanh ⇢) = 0, which
implies that
✓ ◆ ✓ ◆
e it R
/ = . (67)
cosh ⇢ X 0 X d+1
This is the lowest energy state. One can get excited states acting with Pµ .
Notice that all this states will have the same value of the quadratic Casimir
1
JAB J BA = ( d) . (68)
2
This way one can generate all normalizable solutions of r2 = m2 with
m2 R2 = ( d). This shows that the one-particle energy spectrum is
given by ! = + l + 2n where l = 0, 1, 2, . . . is the spin, generated by
acting with Pµ1 . . . Pµl traces , and n = 0, 1, 2, . . . is generated by acting
n
with P 2 .
Exercise 3.2. Given the symmetry of the metric (54) we can look for so-
lutions of the form
= ei!t Yl (⌦)F (r) , (69)
where Yl (⌦) is a spherical harmonic with eigenvalue l(l + d 2) of the
laplacian on the unit sphere S d 1 . Derive a di↵erential equation for F (r)
and show that it is solved by
✓ ◆
l l+ ! l+ +! d
F (r) = (cos r) (sin r) 2 F1 , , l + , sin r , (70)
2 2 2
20 João Penedones
p
with 2 = d + d2 + 4(mR)2 . We chose this solution because it is smooth
at r = 0. Now we also need to impose another boundary condition at the
boundary of AdS r = ⇡2 . Imposing that there is no energy flux through the
boundary leads to the quantization of the energies |!| = + l + 2n with
n = 0, 1, 2, . . . (see reference [8]).
If there are no interactions between the particles in AdS, then the
Hilbert space is a Fock space and the energy of a multi-particle state is
just the sum of the energies of each particle. Turning on small interactions
leads to small energy shifts of the multi-particle states. This structure is
very similar to the space of local operators in large N CFTs if we identify
single-particle states with single-trace operators.
3.2. Quantum Field Theory in AdS

Let us now return to Euclidean signature and consider QFT on the AdS
background. For simplicity, consider a free scalar field with action

1 1
ˆ
2
S= dX (r ) + m2 2 . (71)
AdS 2 2
The two-point function h (X) (Y )i is given by the propagator ⇧(X, Y ),
which obeys
⇥ 2 ⇤
rX m2 ⇧(X, Y ) = (X, Y ) . (72)
From the symmetry of the problem it is clear that the propagator can
only depend on the invariant X · Y or equivalently on the chordal distance
⇣ = (X Y )2 /R2 . From now on we will set R = 1 and all lengths will be
expressed in units of the AdS radius.
Exercise 3.3. Use (62) and (64) to show that
✓ ◆
C d 1 4
⇧(X, Y ) = F
2 1 , + ,2 d + 1, , (73)
⇣ 2 2 ⇣
p
where 2 = d + d2 + (2m)2 and
( )
C = d d
. (74)
2⇡ 2 2 +1
For a free field, higher point functions are simply given by Wick con-
tractions. For example,
h (X1 ) (X2 ) (X3 ) (X4 )i = ⇧(X1 , X2 )⇧(X3 , X4 ) + ⇧(X1 , X3 )⇧(X2 , X4 )
+ ⇧(X1 , X4 )⇧(X2 , X3 ) . (75)
Weak interactions of can be treated perturbatively. Suppose the action

includes a cubic term,

1 1 1
ˆ
2
S= dX (r ) + m2 2 + g 3 . (76)
AdS 2 2 3!
Then, there is a non-vanishing three-point function
ˆ
h (X1 ) (X2 ) (X3 )i = g dY ⇧(X1 , Y )⇧(X2 , Y )⇧(X3 , Y ) + O(g 3 ) ,
AdS
and a connected part of the four-point function of order g 2 . This is very

similar to perturbative QFT in flat space.
Given a correlation function in AdS we can consider the limit where we
send all points to infinity. More precisely, we introduce
1
O(P ) = p lim (X = P + . . . ) , (77)
C !1
where P is a future directed null vector in Rd+1,1 and the . . . denote terms
that do not grow with whose only purpose is to enforce the AdS condition
X 2 = 1. In other words, the operator O(P ) is the limit of the field (X)
when X approaches the boundary point P of AdS. Notice that, by defini-
tion, the operator O(P ) obeys the homogeneity condition (31). Correlation
functions of O are naturally defined by the limit of correlation functions of
in AdS. For example, the two-point function is given by
1
hO(P1 )O(P2 )i = + O(g 2 ) , (78)
( 2P1 · P2 )
which is exactly the CFT two-point function of a primary operator of di-
mension . The three-point function hO(P1 )O(P2 )O(P3 )i is given by
ˆ
3
gC 2
dX ⇧(X, P1 )⇧(X, P2 )⇧(X, P3 ) + O(g 3 ) , (79)
AdS
where
C
⇧(X, P ) = lim ⇧ (X, Y = P + . . . ) = (80)
!1 ( 2P · X)
is the bulk to boundary propagator.
Exercise 3.4. Write the bulk to boundary propagator in Poincaré coordi-

nates.
22 João Penedones
Exercise 3.5. Compute the following generalization of the integral in (79),

Y3
1
ˆ
dX i
, (81)
AdS i=1 ( 2Pi · X)
and show that it reproduces the expected formula for the CFT three-point
function hO1 (P1 )O2 (P2 )O3 (P3 )i. It is helpful to use the integral represen-
tation
ˆ 1
1 1 ds
= s e2sP ·X (82)
( 2P · X) ( ) 0 s
to bring the AdS integral to the form
ˆ
dXe2X·Q (83)
AdS
with Q a future directed timelike vector. Choosing the X 0 direction along

Q and using the Poincaré coordinates (49) it is easy to show that
ˆ 1
dz
ˆ
d d 2
dXe2X·Q = ⇡ 2 z 2 e z+Q /z . (84)
AdS 0 z
To factorize the remaining integrals over s1 , s2 , s3 and z it is helpful to
change to the variables t1 , t2 , t3 and z using
p
z t1 t2 t3
si = . (85)
ti
3.2.1. State-Operator Map

We have seen that the correlation functions of the boundary operator (77)
have the correct homogeneity property and SO(d+1, 1) invariance expected
of CFT correlators of a primary scalar operator with scaling dimension .
We will now argue that they also obey an associative OPE. The argument
is very similar to the one used in CFT. We think of the correlation functions
as vacuum expectation values of time ordered products
h (X1 ) (X2 ) (X3 ) . . . i = h0| . . . ˆ(⌧3 , ⇢3 , ⌦3 ) ˆ(⌧2 , ⇢2 , ⌦2 ) ˆ(⌧1 , ⇢1 , ⌦1 ) |0i ,
where we assumed ⌧1 < ⌧2 < 0 < ⌧3 < . . . . We then insert a complete
basis of states at ⌧ = 0,
h (X1 ) (X2 ) (X3 ) . . . i (86)
X
= h0| . . . ˆ(⌧3 , ⇢3 , ⌦3 ) | i h | ˆ(⌧2 , ⇢2 , ⌦2 ) ˆ(⌧1 , ⇢1 , ⌦1 ) |0i .
Using ˆ(⌧, ⇢, ⌦) = e⌧ D ˆ(0, ⇢, ⌦)e ⌧ D and choosing an eigenbasis of the

@
Hamiltonian D = @⌧ it is clear that the sum converges for the assumed
time ordering. The next step, is to establish a one-to-one map between
the states | i and boundary operators. It is clear that every boundary
operator (77) defines a state. Inserting the boundary operator at P A =
P 0 , P µ , P d+1 = 12 , 0, 12 , which is the boundary point defined by ⌧ !
1 in global coordinates, we can write
h. . . (X3 )O(P )i = h0| . . . ˆ(⌧3 , ⇢3 , ⌦3 ) |Oi , (87)
where
|Oi = lim e ⌧ cosh ⇢ ˆ(⌧, ⇢, ⌦) |0i (88)

⌧! 1
X
)ˆ
= | i (cosh ⇢) lim h | e⌧ (D (0, ⇢, ⌦) |0i .
⌧! 1
The limit ⌧ ! 1 projects onto the primary state with wave function
(67).
The map from states to boundary operators can be established using
global time translation invariance,
h0| . . . ˆ(⌧3 , ⇢3 , ⌦3 ) | (0)i (89)

= lim h0| . . . ˆ(⌧3 , ⇢3 , ⌦3 )e⌧ D | (⌧ )i ⌘ h. . . (X3 )O (P )i
⌧! 1
where | (⌧ )i = e ⌧ D | i and P A = 21 , 0, 12 is again the boundary point

defined by ⌧ ! 1 in global coordinates. The idea is that | (⌧ )i prepares a
boundary condition for the path integral on a surface of constant ⌧ and this
surface converges to a small cap around the boundary point P A = 12 , 0, 12
when ⌧ ! 1. This is depicted in figure 2.
The Hilbert space of the bulk theory can be decomposed in irreducible
representations of the isometry group SO(d + 1, 1). These are the high-
est weight representations of the conformal group, labelled by the scaling
dimension and SO(d) irrep of the the primary state. Therefore, the CFT
conformal block decomposition of correlators follows from the partial wave
decomposition in AdS, i.e. the decomposition in intermediate eigenstates
of the Hamiltonian organized in irreps of the isometry group SO(d + 1, 1).
For example, the conformal block decomposition of the disconnected part
of the four-point function,
1 1 1
hO(P1 ) . . . O(P4 )i = + + , (90)
(P12 P34 ) (P13 P24 ) (P14 P23 )
24 João Penedones
⌧= 1 ⌧ =1
Fig. 2. Curves of constant ⌧ (in blue) and constant ⇢ (in red) for AdS2 stereographically
projected to the unit disk (Poincaré disk). This shows how surfaces of constant ⌧ converge
to a boundary bound when ⌧ ! 1. The cartesian coordinates in the plane of the figure
(cosh ⇢ sinh ⌧,sinh ⇢) ~2
~ = 1+cosh ⇢ cosh ⌧ which puts the AdS2 metric in the form ds2 = 14dw
are given by w ~2
w
.
where Pij = 2Pi · Pj , is given by a sum of conformal blocks associated

with the vacuum and two-particle intermediate states
1 X
X 1
hO(P1 ) . . . O(P4 )i = G0,0 (P1 , . . . , P4 )+ cn,l G2 +2n+l,l (P1 , . . . , P4 ) .
l=0 n=0
even
Exercise 3.6. Check this statement in d = 2 using the formula [28]

k(E + l, z)k(E l, z̄) + k(E l, z)k(E + l, z̄)
GE,l (P1 , P2 , P3 , P4 ) = (91)
( 2P1 · P2 ) ( 2P3 · P4 ) (1 + l,0 )
where
k(2 , z) = ( z) 2 F1 ( , , 2 , z) . (92)
Determine the coefficients cn,l for n  1 by matching the Taylor series

expansion around z = z̄ = 0. Extra: using a computer you can compute
many coefficients and guess the general formula.
3.2.2. Generating function

There is an equivalent way of defining CFT correlation functions from QFT
in AdS. We introduce the generating function
D ´ E
W [ b ] = e @AdS dP b (P )O(P ) , (93)
where the integral over @AdS denotes an integral over a chosen section
of the null cone in Rd+1,1 with its induced measure. We impose that the
d
source obeys b ( P ) = b (P ) so that the integral is invariant under
a change of section, i.e. conformal
´ dinvariant. For example, in the Poincaré
section the integral reduces to d x b (x)O(x). Correlation functions are
easily obtained with functional derivatives
hO(P1 ) . . . O(Pn )i = ... W [ b] . (94)

b (P1 ) b (Pn ) b =0
If we set the generating function to be equal to the path integral over the
field in AdS
[d ] e S[ ]
´
! b
W [ b] = ´ , (95)
!0
[d ] e S[ ]
with the boundary condition that it approaches the source b at the bound-
ary,
1 1
lim d (X = P + . . . ) = p b (P ) , (96)
!1 2 d C
then we recover the correlation functions of O defined above as limits of
the correlation functions of .
For a quadratic bulk action, tha ratio of path intagrals in (95) is given
e S computed on the classical solution obeying the required boundary con-
ditions. A natural guess for this solution is
p ˆ
b (P )
(X) = C dP . (97)
@AdS ( 2P · X)
This automatically solves the AdS equation of motion r2 = m2 , because
it is an homogeneous function of weight and it obeys @A @ A = 0 in the
embedding space (see equations (62) and (64)). To see that it also obeys
the boundary condition (96) it is convenient to use the Poincaré section.
Exercise 3.7. In the Poincaré section (30) and using Poincaré coordinates
(49), formula (97) reads
p ˆ
z b (y)
(z, x) = C dd y (98)
(z + (x y)2 )
2
and (96) reads

d 1 1
lim z (z, x) = p b (x) . (99)
z!0 2 d C
Show that (99) follows from (98). You can assume 2 > d.
26 João Penedones
1
The cubic term 3! g 3 in the action will lead to (calculable) corrections of
order g in the classical solution (97). To determine the generating function
W [ b ] in the classical limit we just have to compute the value of the bulk
action (76) on the classical solution. However, before doing that, we have
to address a small subtlety. We need to add a boundary term to the action
(76) in order to have a well posed variational problem.
Exercise 3.8. The coefficient should be chosen such that the quadratic
action m

ˆ
1 p 2 1
ˆ p
S2 = dw G (r ) + m2 2
+ dw G r↵ ( r↵ ) (100)
AdS 2 2 AdS
is stationary around a classical solution obeying (99) for any variation

that preserves the boundary condition, i.e.
(z, x) = z [f (x) + O(z)] . (101)

d
Show that = d and that the on-shell action is given by a boundary
term
2 d p
ˆ
S2 = dw g r↵ ( r↵ ) . (102)
2d AdS
Finally, show that for the classical solution (98) this action is given by n
1
ˆ
S2 = dd y1 dd y2 b (y1 ) b (y2 )K(y1 , y2 ) , (103)
2
where
2 d dd x z z
ˆ
K(y1 , y2 ) = C lim @z
d z!0 z d 1 (z 2 + (x y1 )2 ) (z 2 + (x y2 ) 2 )
1
= (104)
(y1 y2 ) 2
is the CFT two point function (78).

m Here w stands for a generic coordinate in AdS and the index ↵ runs over the d + 1
dimensions of AdS.
n This integral is divergent if the source
b is a smooth function and > d2 . The
divergence comes from the short distance limit y1 ! y2 and does not a↵ect the value of
correlation functions at separate points. Notice that a small value of z > 0 provides a
UV regulator.
Exercise 3.9. Using = 0 + O(g) with 0 given by (97), show that the
complete on-shell action is given by
1 1
ˆ ˆ
3
S= dd y1 dd y2 b (y1 ) b (y2 )K(y1 , y2 )+ g dX [ 0 (X)] +O(g 2 ) ,
2 3! AdS
and that this leads to the three-point function (79). Extra: Compute the
terms of O(g 2 ) in the on-shell action.
We have seen that QFT on an AdS background naturally defines con-

formal correlation functions living on the boundary of AdS. Moreover, we
saw that a weakly coupled theory in AdS gives rise to factorization of CFT
correlators like in a large N expansion. However, there is one missing in-
gredient to obtain a full-fledged CFT: a stress-energy tensor. In the next
section, we will see that this requires dynamical gravity in AdS. The next
exercise also shows that a free QFT in AdSd+1 can not be dual to a local
CFTd .
Exercise 3.10. Compute the free-energy of a gas of free scalar particles in

AdS. Since particles are free and bosonic one can create multi-particle states
by populating each single particle state an arbitrary number of times. That
means that the total partition function is a product over all single particle
states and it is entirely determined by the single particle partition function.
More precisely, show that
1
! 1
Y X X 1
kE sp
F = T log Z = T log q = T Z1 (q n ) , (105)
n=1
n
sp k=0
X q
Z1 (q) = qE sp = , (106)
(1 q)d
sp
1
where q = e RT and we have used the single-particle spectrum of the hamil-
@
tonian D = @⌧ of AdS in global coordinates. Show that
F ⇡ ⇣(d + 1)Rd T d+1 (107)
in the high temperature regime and compute the entropy using the thermo-
dynamic relation S = @F @T . Compare this result with the expectation
S ⇠ (RT )d 1
, (108)
for the high temperature behaviour of the entropy of a CFT on a sphere

S d 1 of radius R. See section 4.3 of reference [29] for more details.
28 João Penedones
3.3. Gravity with AdS boundary conditions

Consider general relativity in the presence of a negative cosmological con-
stant
1
ˆ p
I[G] = d 1 dd+1 w G [R 2⇤] . (109)
`P
The AdS geometry
dz 2 + dxµ dxµ
ds2 = R2 , (110)
z2
is a maximally symmetric classical solution with ⇤ = d(d 1)
2R2 . When
the AdS radius R is much larger than the Planck length `P the metric
fluctuations are weakly coupled and form an approximate Fock space of
graviton states. One can compute the single graviton states and verify that
they are in one-to-one correspondence with the CFT stress-tensor operator
and its descendants (with AdS energies matching scaling dimensions). One
can also obtain CFT correlation functions of the stress-energy tensor using
Witten diagrams in AdS. The new ingredients are the bulk to boundary
and bulk to bulk graviton propagators [30–34].
In the gravitational context, it is nicer to use the partition function
formulation
ˆ
Z[gµ⌫ , b ] = [dG] [d ] e I[G, ] (111)
G!g
! b
where

1
ˆ p 1 2 1
I[G, ] = dd+1 w G R 2⇤ + (r ) + m2 2
(112)
`dP 1 2 2
and the boundary condition are
dz 2 + dxµ dx⌫ [gµ⌫ (x) + O(z)]
ds2 = G↵ dw↵ dw = R2 , (113)
z2
zd
= [ b (x) + O(z)] .
2 d
By construction the partition function is invariant under di↵eomorphisms
of the boundary metric gµ⌫ . Therefore, this definition implies the Ward
identity (18). The generating function is also invariant under Weyl trans-
formations
⇥ ⇤
Z ⌦2 gµ⌫ , ⌦ d b = Z [gµ⌫ , b ] (naive) (114)
This follows from the fact that the boundary condition

2 µ ⌫
⇥ 2 ⇤
2 2 dz + dx dx ⌦ (x)gµ⌫ (x) + O(z)
ds = R (115)
z2
z d ⇥ ⇤
= ⌦ d (x) b (x) + O(z)
2 d
can be mapped to (113) by the following coordinate transformation
1 3 2
z ! z⌦ z ⌦ (@µ log ⌦) + O(z 5 ) (116)
4
1 2 µ
xµ ! xµ z @ log ⌦ + O(z 4 )
2
where indices are raised and contracted using the metric gµ⌫ and its in-
verse. In other words, a bulk geometry that satisfies (113) also satisfies
(115) with an appropriate choice of coordinates. If the partition function
(111) was a finite quantity this would be the end of the story. However,
even in the classical limit, where Z ⇡ e I , the partition function needs to
be regulated. The divergences originate from the z ! 0 region and can
be regulated by cutting o↵ the bulk integrals at z = ✏ (as it happened
for the scalar case discussed above). Since the coordinate transformation
(116) does not preserve the cuto↵, the regulated partition function is not
obviously Weyl invariant. This has been studied in great detail in the con-
text of holographic renormalization [35, 36]. In particular, it leads to the
Weyl anomaly g µ⌫ Tµ⌫ 6= 0 when d is even. The crucial point is that this
is a UV e↵ect that does not a↵ect the connected correlation functions of
operators at separate points. In particular, the integrated form (21)=(23)
of the conformal Ward identity is valid.
We do not now how to define the quantum gravity path integral in
(111). The best we can do is a semiclassical expansion when `P ⌧ R.
This semiclassical expansion gives rise to connected correlators of the stress
tensor Tµ⌫ that scale as
✓ ◆d 1
R
hTµ1 ⌫1 (x1 ) . . . Tµn ⌫n (xn )ic ⇠ . (117)
`P
This is exactly the scaling (46) we found from large N factorization if we

⇣ ⌘d 1
identify N 2 ⇠ `RP . This suggests that CFTs related to semiclassical
Einstein gravity in AdS, should have a large number of local degrees of
freedom. This can be made more precise. The two-point function of the
30 João Penedones
stress tensor in a CFT is given by


CT 1 1 1 1
hTµ⌫ (x)T ⇢ (0)i = 2 2d Iµ I⌫⇢ + Iµ⇢ I⌫ µ⌫ ⇢ , (118)
Sd x 2 2 d
2⇡ d/2
where Sd = (d/2) is the volume of a (d 1)-dimensional unit sphere and
xµ x⌫
Iµ⌫ = µ⌫ . 2 (119)
x2
The constant CT provides an (approximate) measure of the number of
degrees of freedom.o For instance, for n' free scalar fields and n free
Dirac fields we find [37]
d
+ n 2[ 2 ]
d
1
C T = n' d, (120)
d 1
where [x] is the integer part of x. If the CFT is described by Einstein
gravity in AdS, we find [30]
d
d + 1 ⇡ 2 (d + 1) Rd 1
CT = 8 3 d
, (121)
d 1 2 `dP 1
which shows that the CFT dual of a semiclassical gravitational theory with
R `P , must have a very large number of degrees of freedom.
In summary, semiclassical gravity with AdS boundary conditions gives
rise to a set of correlation functions that have all the properties (conformal
invariance, Ward identities, large N factorization) expected for the correla-
tion functions of the stress tensor of a large N CFT. Therefore, it is natural
to ask if a CFT with finite N is a quantum theory of gravity.
4. The AdS/CFT Correspondence
4.1. Quantum Gravity as CFT

What is quantum gravity? The most conservative answer is a standard
quantum mechanical theory whose low energy dynamics around a weakly
curved background is well described by general relativity (or some other
theory with a dynamical metric). This viewpoint is particularly useful
with asymptotically AdS boundary conditions. In this case, we can view
the AdS geometry with a radius much larger than the Planck length as a
background for excitations (gravitons) that are weakly coupled at low en-
ergies. Therefore, we just need to find a quantum system that reproduces
o However,for d > 2, CT is not a c-function that always decreases under Renormalization
Group flow.
the dynamics of low energy gravitons in a large AdS. In fact, we should be

more precise about the word “reproduces”. We should define observables in
quantum gravity that our quantum system must reproduce. This is not so
easy because the spacetime geometry is dynamical and we can not define
local operators. In fact, the only well defined observables are defined at
the (conformal) boundary like the partition function (111) and the associ-
ated correlation functions obtained by taking functional derivatives. But in
the previous section we saw that these observables have all the properties
expected for the correlation functions of a large N CFT. Thus, quantum
gravity with AdS boundary conditions is equivalent to a CFT.
There are many CFTs and not all of them are useful theories of quantum
gravity. Firstly, it is convenient to consider a family of CFTs labeled by
N , so that we can match the bulk semiclassical expansion using N 2 ⇠
⇣ ⌘d 1
R
`P . In the large N limit, every CFT single-trace primary operator of
scaling dimension gives rise to a weakly coupled field in AdS with mass
m ⇠ /R. Therefore, if are looking for a UV completion of pure gravity
in AdS without any other low energy fields, then we need to find a CFT
where all single-trace operators have parametrically large dimension, except
the stress tensor. This requires strong coupling and seems rather hard to
achieve. Notice that a weakly coupled CFT with gauge group SU (N ) and
fields in the adjoint representation has an infinite number of primary single-
trace operators with order 1 scaling dimension. It is natural to conjecture
that large N factorization and correct spectrum of single-trace operators
are sufficient conditions for a CFT to provide a UV completion of General
Relativity (GR) [38]. However, this is not obvious because we still have to
check if the CFT correlation functions of Tµ⌫ match the prediction from
GR in AdS. For example, the stress tensor three-point function is fixed by
conformal symmetry to be a linear combination of 3 independent conformal
invariant structures. p On the other hand, the action (109) predicts a
specific linear combination. It is not obvious that all large N CFTs with
the correct spectrum will automatically give rise to the same three-point
function. There has been some recent progress in this respect. The authors
of [39] used causality to show that this is the case. Unfortunately, their
argument uses the bulk theory and can not be formulated entirely in CFT
language. In any case, this is just the three-point fuction and GR predicts
the leading large N behaviour of all n-point functions. It is an important
open problem to prove the following conjecture:
p Here we are assuming d 4. For d = 3 there are only 2 independent structures.

32 João Penedones
Any large N CFT where all single-trace operators, except the stress ten-
sor, have parametrically large scaling dimensions, has the stress tensor cor-
relation functions predicted by General Relativity in AdS.
Perhaps the most pressing question is if such CFTs exist at all. At the
moment, we do not know the answer to this question but in the next section
we will discuss closely related examples that are realized in the context of
string theory.
If some CFTs are theories of quantum gravity, it is natural to ask if
there are other CFT observables with a nice gravitational interpretation.
One interesting example that will be extensively discussed in this school is
the entanglement entropy of a subsystem. In section 4.3, we will discuss how
CFT thermodynamics compares with black hole thermodynamics in AdS.
In addition, in section 4.4 we will give several examples of QFT phenomena
that have beautiful geometric meaning in the holographic dual.
4.2. String Theory
The logical flow presented above is very di↵erent from the historical route
that led to the AdS/CFT correspondence. Moreover, from what we said so
far AdS/CFT looks like a very abstract construction without any concrete
examples of CFTs that have simple gravitational duals. If this was the full
story probably I would not be writing these lecture notes. The problem
is that we have stated properties that we want for our CFTs but we have
said nothing about how to construct these CFTs besides the fact that they
should be strongly coupled and obey large N factorization. Remarkably,
string theory provides a “method” to find explicit examples of CFTs and
their dual gravitational theories.
The basic idea is to consider the low energy description of D-brane
systems from the perspective of open and closed strings. Let us illus-
trate the argument by quickly summarizing the prototypical example of
AdS/CFT [1]. Consider N coincident D3-branes of type IIB string the-
ory in 10 dimensional Minkowski spacetime. Closed strings propagating in
10 dimensions can interact with the D3-branes. This interaction can be
described in two equivalent ways:
(a) D3-branes can be defined as a submanifold where open strings can
end. This means that a closed string interacts with the D3-branes by break-
ing the string loop into an open string with endpoints attached to the D3-
branes.
(b) D3-branes can be defined as solitons of closed string theory. In other
words, they create a non-trivial curved background where closed strings

propagate.
(a) (b)
Fig. 3. (a) Closed string scattering o↵ branes in flat space. (b) Closed string propagat-
ing in a curved background.
These two alternatives are depicted in figure 3. Their equivalence is

called open/closed duality. The AdS/CFT correspondence follows from
the low-energy limit of open/closed duality. We implement this low-energy
limit by taking the string length `s ! 0, keeping the string coupling gs ,
the number of branes N and the energy fixed. In description (a), the
low energy excitations of the system form two decoupled sectors: massless
closed strings propagating in 10 dimensional Minkoski spacetime and mass-
less open strings attached to the D3-branes, which at low energies are well
described by N = 4 Supersymmetric Yang-Mills (SYM) with gauge group
SU (N ). In description (b), the massless closed strings propagate in the
following geometry
1 p ⇥ ⇤
ds2 = p ⌘µ⌫ dxµ dx⌫ + H(r) dr2 + r2 d⌦25 , (122)
H(r)
where ⌘µ⌫ is the metric of the 4 dimensional Minkowski space along the
branes and
R4
H(r) = 1 + , R4 = 4⇡gs N `4s . (123)
r4
Naively, the limit `s ! 0 just produces 10 dimensional Minkowski space-
time. However, one has to be careful with the region close to the branes at
r = 0. A local high energy excitation in this region will be very redshifted
34 João Penedones
from the point of view of the observer at infinity. In order to determine

the correct low-energy limit in the region around r = 0 we introduce a new
coordinate z = R2 /r, which we keep fixed as `s ! 0. This leads to
dz 2 + ⌘µ⌫ dxµ dx⌫
ds2 = R2 + R2 d⌦25 , (124)
z2
which is the metric of AdS5 ⇥S 5 both with radius R. Therefore, description
(b) also leads to 2 decoupled sectors of low energy excitations: massless
closed strings in 10D and full type IIB string theory on AdS5 ⇥ S 5 . This
led Maldacena to conjecture that
SU (N ) SYM IIB strings on AdS5 ⇥ S 5

, R4
gY2 M = 4⇡gs 2
`4 = g Y M N ⌘
s
SYM is conformal for any value of N and the coupling constant gY2 M .
The lagrangian of the theory involves the field strength
Fµ⌫ = @µ A⌫ @ ⌫ Aµ i [Aµ , A⌫ ] , (125)
6 scalars fields m and 4 Weyl fermions a , which are all valued in the
adjoint representation of SU (N ). The lagrangian is given by

1 1 1 2
Tr F µ⌫ Fµ⌫ + (Dµ m ) + ¯ a µ Dµ a (126)
gY2 M 4 2
1 m n2 ⇥ ⇤
[ , ] Cmab
a[
m
, b ] C̄mab ¯ a m , ¯ b ,
4
ab
where Dµ is the gauge covariant derivative and Cm and C̄mab are constants
fixed by the SO(6) = SU (4) global symmetry of the theory. Notice that the
isometry group of AdS5 ⇥ S 5 is SO(5, 1) ⇥ SO(6), which matches precisely
the bosonic symmetries of SYM: conformal group ⇥ global SO(6). There
are many interesting things to say about SYM. In some sense, SYM is the
simplest interacting QFT in 4 dimensions [40]. However, this is not the
focus of these lectures and we refer the reader to the numerous existing
reviews about SYM [10, 41].
The remarkable conjecture of Maldacena has been extensively tested
since it was first proposed in 1997 [1]. To test this conjecture one has to
be able to compute the same observable on both sides of the duality. This
is usually a very difficult task. On the SYM side, the regime accessible
to perturbation theory is gY2 M N ⌧ 1. This implies gs ⌧ 1, which on the
string theory side suppresses string loops. However, it also implies that the
AdS radius of curvature R is much smaller than the string length `s . This
means that the string worldsheet theory is very strongly coupled. In fact,
the easy regime on the string theory side is gs ⌧ 1 and R `s , so that
(locally) strings propagate in an approximately flat space. Thus, directly
testing the conjecture is a formidable task. There are three situations where
a direct check can be made analitycally.
The first situation arises when some observable is independent of the
coupling constant. In this case, one can compute it at weak coupling ⌧
1 using the field theory description and at strong coupling 1 using
the string theory description. Usually this involves completely di↵erent
techniques but in the end the results agree. Due to the large supersymmetry
of SYM there are many observables that do not depend on the coupling
constant. Notable examples include the scaling dimensions of half BPS
single-trace operators and their three-point functions [42].
The second situation involves observables that depend on the coupling
constant but preserve enough supersymmetry that can be computed at
any value of using a technique called localization. Important examples
of this type are the sphere partition function and the expectation value of
circular Wilson loops [43, 44].
Finally, the third situation follows from the conjectured integrability
of SYM in the planar limit. Assuming integrability one can compute the
scaling dimension of non-protected single-trace operators at any value of
and match this result with SYM perturbative calculations for ⌧ 1 and
with weakly coupled string theory for 1 (see figure 1 from [45]). Planar
scattering amplitudes an three-point functions of single-trace operators can
also be computed using integrability [46, 47].
There are also numerical tests of the gauge/gravity duality. The most
impressive study in this context, was the Monte-Carlo simulation of the
BFSS matrix model [48] at finite temperature that reproduced the predic-
tions from its dual black hole geometry [49–55].
How does the Maldacena conjecture fit into the general discussion of the
previous sections? One important novelty is the presence of a large internal
sphere on the gravitational side. We can perform a Kaluza-Klein reduction
on S 5 and obtain an e↵ective action for AdS5
1 10 p R5 p
ˆ ˆ
2
d x g10 e [R10 + . . . ] ! d5 x g5 [R5 + . . . ] .
(2⇡)7 `8s 8(2⇡)4 gs2 `8s
This allows us to identify the 5 dimensional Planck length
8(2⇡)4 gs2 `8s
`3P = (127)
R5
36 João Penedones
and verify the general prediction N 2 ⇠ R3 /`3P . Remarkably, at strong

coupling 1 all single-trace non-protected operators of SYM have para-
metrically large scaling dimensions. This is simple to understand from the
string point of view. Massive string states have masses m ⇠ 1/`s . But we
saw in the previous sections that the dual operator to an AdS field of mass
1
m has a scaling dimension ⇠ mR ⇠ R/`s ⇠ 4 . The only CFT opera-
tors that have small scaling dimension for 1 are dual to massless string
states that constitute the fields of type IIB supergravity (SUGRA). There-
fore, one can say that SYM (with N 1) provides a UV completion
of IIB SUGRA with AdS5 ⇥ S 5 boundary conditions.
String theory provides more concrete examples of AdS/CFT dual
pairs. These examples usually involve SCFTs (or closely related non-
supersymmetry theories). This is surprising because SUSY played no role in
our general discussion. At the moment, it is not known if SUSY is an essen-
tial ingredient of AdS/CFT or if it is only a useful property that simplifies
the calculations. The latter seems more likely but notice that SUSY might
be essential to stabilize very strong coupling and allow the phenomena of
large scaling dimensions for almost all single-trace operators. Another ob-
servation is that it turns out to be very difficult to construct AdS duals with
small internal spaces (for SYM we got a 5-sphere with the same radius of
AdS5 ). It is an open problem to find CFTs with gravity duals in less than
10 dimensions (see [56, 57] for attempts in this direction).
Another interesting class of examples are the dualities between vector
models and Higher Spin Theories (HST) [58, 59]. Consider for simplicity
the free O(n) model in 3 dimensions
Xn
1
ˆ
S = d3 x @ µ 'i @ µ 'i . (128)
i=1
2
In this case, the analogue of single-trace operators are the O(n) singlets
P
Ol = i 'i @µ1 . . . @µl 'i with even spin l and dimension = 1 + l. At large
n, the correlation functions of these operators factorize with n playing the
role of N 2 in a SU (N ) gauge theory with adjoint fields. The AdS dual
of this CFT is a theory with one massless field for each even spin. These
theories are rather non-local and they can not be defined in flat spacetime.
2
Even if we introduce the relevant interaction 'i 'i and flow to the IR
fixed point (Wilson-Fisher fixed point), the operators Ol with l > 2 get
anomalous dimensions of order n1 and therefore the classical AdS theory
still contains the same number of massless higher spin fields. This duality
has been extended to theories with fermions and to theories where the global
O(n) symmetry is gauged using Chern-Simons gauge fields. It is remarkable

that HST in AdS seems to have the correct structure to reproduce the
CFT observables that have been computed so far. Notice that in these
examples of AdS/CFT supersymmetry plays no role. However, it is unclear
if the AdS description is really useful in this case.q In practice, the large
n limit of these vector models is solvable and the dual HST in AdS is
rather complicated to work with even at the classical level. There are also
analogous models in AdS3 /CFT2 duality [62].
4.3. Finite Temperature

In section 3.3, we argued that holographic CFTs must have a large number
of local degrees of freedom, using the two-point function of the stress tensor.
Another way of counting degrees of freedom is to look at the entropy density
when the system is put at finite temperature. For a CFT in flat space and
infinite volume, the temperature dependence of the entropy density is fixed
by dimensional analysis because there is no other scale available,
s = cs T d 1
. (129)
The constant cs is a physical measure of the number of degrees of freedom.

The gravitational dual of the system at finite temperature is a black
brane in asymptotically AdS space. The Euclidean metric is given by
 ✓ ◆
2 R2 dz 2 zd
ds = 2 + 1 d
d⌧ 2 + ij dxi dxj . (130)
z 1 (z/zH )d zH
Exercise 4.1. Show that in order to avoid a conical defect at the horizon
z = zH , we need to identify Euclidean time ⌧ with period 4⇡z
d . This fixes
H
d
the Hawking temperature T = 4⇡zH .
d
The formula T = 4⇡z H
illustrates a general phenomena in holography: high
energy corresponds to the region close to the boundary and low energy
corresponds to the deep interior of the dual geometry.
The entropy of the system is given by the Bekenstein-Hawking formula
AH 4⇡ Rd 1 (4⇡)d Rd 1
ˆ
S= = d 1 d 1 dd 1 x ) cs = d 1 d 1 . (131)
4GN ` P zH d `P
q In practice it was very useful because it led to an intensive study of Chern-Simons
matter theories, which gave rise to the remarkable conjecture of fermion/boson duality
in 3 dimensions [60, 61].
38 João Penedones
As expected cs is very large in the bulk classical limit R `P . Interestingly,

the ratio
d ✓ ◆d
cs ⇡2 4 d 1 3 d2
= (132)
CT 8 d d + 1 (d)
only depends on the spacetime dimension d if the CFT has a classical bulk
dual [63]. It would be very nice to prove that all large N CFTs where all
single-trace operators, except the stress tensor, have parametrically large
scaling dimensions, satisfy (132). Notice that (132) is automatic in d = 2
because CT = 2c and cs = ⇡3 c are uniquely fixed in terms of the central
charge c. In planar SYM, CT = 40N 2 is independent of the ’t Hooft cou-
pling but cs varies with (although not much, cs ( = 1) = 34 cs ( = 0)).
In this case, (132) is only satisfied at strong coupling, when all primary
operators with spin greater than 2 have parametrically large scaling dimen-
sions.
Exercise 4.2. Consider a CFT on a sphere of radius L and at temperature

T . In this case, the entropy is a non-trivial function of the dimensioless
combination LT . Let us compute this function assuming the CFT is well
described by Einstein gravity with asymptotically AdS boundary conditions.
There are two possible bulk geometries that asymptote to the Euclidean
boundary S 1 ⇥ S d 1 . The first is pure AdS

dr2
ds2 = R2 + 1 + r2 d⌧ 2 + r2 d⌦2d 1 (133)
1 + r2
with Euclidean time periodically identified and the second is Schwarzschild-
AdS
 2
2 2 dr
ds = R ⌧ 2 + r2 d⌦2d 1 ,
+ f (r)d˜ (134)
f (r)
where f (r) = 1 + r2 rdm 2 . At the boundary r = rmax 1, both solutions

should be conformal to S 1 ⇥ S d 1 with the correct radii. Show that this fixes
the periodicities
1 r 1 r
⌧= p max , ⌧˜ = p max . (135)
T L 1 + rmax
2 T L f (rmax )
Show also that regularity of the metric (134) implies the periodicity
4⇡ 4⇡
⌧˜ = = , (136)
f 0 (rH ) rH d + drH2
where r = rH is the largest zero of f (r). Notice that thispimplies a minimal

d(d 2)
temperature for Schwarzschild black holes in AdS, T > 2⇡L .
Both (133) and (134) are stationary points of the Euclidean action
(109). Therefore, we must compute the value of the on-shell action in order
to decide which one dominates the path integral. Show that the di↵erence
of the on-shell actions is given by
Rd 1 ⇥ d d d
⇤
IBH IAdS = 2Sd rmax ⌧ rmax rH ⌧˜ (137)
`dP 1
Rd 1
1 d 2 2
! Sd r (1 rH ) (138)
`dP 1 TL H
where Sd is the area of a unit (d 1)-dimensional sphere and in the last step
we took the limit rmax ! 1. Conclude that the black hole only dominates
the bulk path integral when rH > 1, which corresponds to T > d2⇡L1 . This is
the Hawking-Page phase transition [64]. It is natural to set the free-energy
of the AdS phase to zero because this phase corresponds to a gas of gravitons
around the AdS background whose free energy does not scale with the large
parameter (R/`P )d 1 . Therefore, the free energy of the black hole phase is
given by
1 Rd 1
d 2 2
FBH = Sd rH (1 rH ). (139)
L `dP 1
Verify that the thermodynamic relation @F @T = S agrees with the

Bekenstein-Hawking formula for the black hole entropy. Since this a first
order phase transition you can also compute its latent heat.
In the last exercise, we saw that for a holographic CFT on a sphere of

radius L, the entropy is a discontinuous function of the temperature. In
fact, we found that for sufficiently high temperatures T > d2⇡L1 , the entropy
was very large S ⇠ CT , while for lower temperatures the entropy was small
because it did not scale with CT . This can be interpreted as deconfinement
of the numerous degrees of freedom measured by CT 1 which do not
d 1
contribute to the entropy below the transition temperature Tc = 2⇡L . How
can this bevavior be understood from the point of view of a large N gauge
CFT?
4.4. Applications
The AdS/CFT correspondence (or the gauge/gravity duality more gener-
ally) is a useful framework for thinking about strong coupling phenomena
40 João Penedones
in QFT. Besides the specific examples of strongly coupled CFTs that can be
studied in great detail using the gravitational dual description, AdS/CFT
provides a geometric reformulation of many e↵ects in QFT. Usually, we do
not know the precise gravitational dual of a given QFT of interest (like
QCD) but it is still very useful to study gravitational toy models that
preserve the main features we are interested in. These models enlarge our
intuition because they are very di↵erent from QFT models based on weakly
interacting quasi-particles. There are many examples of QFT observables
that have a nice geometric interpretation in the dual gravitational descrip-
tion. Perhaps the most striking one is the computation of entanglement
entropy as the area of a minimal surface in the dual geometry [65]. Let us
illustrate this approach in the context of confinig gauge theories like pure
Yang-Mills theory.
Confinement means that the quark anti-quark potential between static
quarks grows linearly with the distance L at large distances
V (L) ⇡ L , L!1, (140)
where is the tension of the flux tube or e↵ective string. This potential
can be defined through the expectation value of a Wilson loop (in the
fundamental representation)
˛
W [C] = Tr P exp Aµ dxµ , (141)
C
for a rectangular contour C with sides T ⇥ L,

T V (L)
hW [C]i ⇠ e , T !1. (142)
This is equivalent to the area law hW [C]i ⇠ e Area[C] for large contours.
In the gauge/string duality there is a simple geometric rule to compute
expectation values of Wilson loops [66]. One should evaluate the path
integral
ˆ
hW [C]i = [d⌃]e Ss [⌃] (143)
@⌃=C
summing over all surfaces ⌃ in the dual geometry that end at the contour
C at the boundary. The path integral is weighted using the dual string
world-sheet action. At large N , we expect that the dominant contribution
comes from surfaces ⌃ with disk topology. In specific examples, like SYM,
this can be made very precise. For example, at large ’t Hooft coupling the
r
world-sheet action reduces to
1
Ss [⌃] = Area[⌃] . (144)
4⇡`2s
In this case, since the theory is conformal, there is no confinement and the
quark anti-quark potential is Coulomb like,
a(N, )
V (L) = . (145)
L
For most confining gauge theories (e.g. pure Yang-Mills theory) we do
not know neither the dual geometry nor the dual string world-sheet action.
However, we can get a nice qualitative picture if we assume (144) and only
change the background geometry. The most general (d + 1)-dimensional
geometry that preserves d-dimensional Poincaré invariance can be written
as
 2
dz
ds2 = R2 + A2 (z)dxµ dxµ . (146)
z2
The profile of the function A2 (z) encodes many properties of the dual QFT.
For a CFT, scale invariance fixes A(z) / z 1 . For asymptotically free gauge
theories, we still expect that A(z) diverges for z ! 0 however the function
should be very di↵erent for larger values of z. In particular, it should have
a minimum for some value z = z? > 0. Let us see what this implies for the
expectation value of a large Wilson loop. The string path integral (143)
will be dominated by the surface ⌃ with minimal area. For large contours
C, this surface will sink inside AdS until the value z = z? that minimizes
A2 (z) and the worldsheet area will be given by
R2 A2 (z? )Area[C] + O(Length[C]) . (147)
Therefore, we find a confining potential with flux tube tension
A2 (z? ) R2
= . (148)
4⇡ `2s
What happens if we put the QFT at finite temperature? In this case,
we can probe confinement by computing
⌦ ↵
W (Cx )W̄ (Cx+L ) = e Fqq̄ ( ,L) (149)
where Cx is the contour around the Euclidean time circle at the spatial po-
sition x (Polyakov loop). Fqq̄ ( , L) denotes the free energy of a static quark
r Infact, the total area of ⌃ is infinite but the divergence comes from the region close to
the boundary of AdS. This can be regulated by cutting of AdS at z = ✏, and renormalized
by subtracting a divergent piece proportional to the length of the contour C.
42 João Penedones
anti-quark pair at distance L and temperature 1/ . If Fqq̄ ( , L) ! 1 as we

separate the pair, then we are in the confined phase. On the other hand, if
Fqq̄ ( , L) remains finite when L ! 1, we are in the deconfined phase. Let
us see how this works in the holographic dual. For low temperatures, the
dual geometry is simply given by (146) with Euclidean time identified with
period . Therefore, the bulk minimal surface that ends on Cx and Cx+L
will have a cylindrical topology and its area will scale linearly with L at
large L. In fact, we find Fqq̄ ( , L) ⇡ L like in the vacuum. On the other
hand, for high enough temperature we expect the bulk path integral to be
dominated by a black hole geometry (see exercise 4.2 about Hawking-Page
phase transition). The metric can then be written as

dz 2
ds2 = R2 2 + f (z)d⌧ 2 + g(z)dxi dxi , (150)
z f (z)
where f (z) vanishes for some value z = zH . This means that the Euclidean
time circle is contractible in the bulk. Therefore, for large L, the minimal
surface has two disconnected pieces with disk topology ending on Cx and
Cx+L whose area remains finite when L ! 1. This means deconfinement
⌦ ↵ 2
lim W (Cx )W̄ (Cx+L ) = hW (Cx )i = e 2 Fq ( ) > 0 . (151)
L!1
Another feature of a confining gauge theory is a mass gap and a dis-

crete spectrum of mesons and glueballs. To compute this spectrum using
the bulk dual one should study fluctuations around the vacuum geometry
(146). Consider for simplicity, a scalar field obeying r2 = m2 . Since
we are interested in finding the spectrum of the operator Pµ P µ we look for
solutions of the form = eik·x (z), which leads to
z k2
@z zAd (z)@z = m2 R 2 . (152)
Ad (z) A2 (z)
The main idea is that this equation will only have solutions that obey the
boundary conditions (0) = (1) = 0 for special discrete values of k 2 . In
other words, we obtain a discrete mass spectrum as expected for a confining
gauge theory.
Exercise 4.3. Consider the simplest holographic model of a confining gauge

theory: the hard wall model. This is just a slice of AdS, i.e. we take
A(z) = 1/z and cuto↵ space at z = z? . Show that (152) reduces to the
Bessel equation
⇥ 2 2 ⇤
z @z + z@z ↵2 k 2 z 2 h(z) = 0 , (153)
d
where ↵2 = m2 R2 + d2 /4 and h(z) = z 2 (z). Finally, show that the
boundary conditions h(0) = h(z? ) = 0, lead to the quantization
✓ ◆
z u2↵,n
hn (z) = J↵ u↵,n , m2n = k 2 = 2 , n = 1, 2, . . . (154)
z? z?
where u↵,n is the nth zero of the Bessel function J↵ .
It is instructive to compare the lightest glueball mass m1 with the flux

1 R2 R2
tube tension = 4⇡z 2 2 in the hard wall model. We find that m2 ⇠ `2 .
? `s 1 s
The fact that this ratio is of order 1 in pure Yang-Mills theory is another
indication that its holographic dual must be very stringy (curvature radius
of the same order of the string length).
Above the deconfinement temperature, the system is described by
a plasma of deconfined partons (quarks and gluons in QCD). The
gauge/gravity duality is also very useful to describe this strongly coupled
plasma. The idea is that the hydrodynamic behavior of the plasma is dual
to the long wavelength fluctuations of the black hole horizon. This map can
be made very precise and has led to significant developments in the theory
of relativistic hydrodynamics. One important feature of the gravitational
description is that dissipation is built in because black hole horizons natu-
rally relax to equilibrium. A famous result from this line of work was the
discovery of a universal ratio of shear viscosity ⌘ to entropy density s. Any
CFT dual to Einstein gravity in AdS has ⌘s = 4⇡ 1
. This is a rather small
⌘
number (water at room temperature has s ⇠ 30) but remarkably it is of
the same order of magnitude of that observed in the quark-gluon plasma
produced in heavy ion collisions [67].
There are also many interesting applications of the gauge/gravity du-
ality to Condensed Matter physics [9, 14]. There are many materials that
are not well described by weakly coupled quasi-particles. In this case, it
is useful to have alternative models based on gravitational theories in AdS
that share the same qualitative features. This can give geometric intuition
about the system in question.
The study of holographic models is also very useful for the discovery of
general properties of CFT (and QFT more generally). If one observes that a
given property holds both in weakly coupled and in holographic CFTs, it is
natural to conjecture that such property holds in all CFTs. This reasoning
has led to the discovery (and sometimes proof) of several important facts
about CFTs, like the generalization of Zamolodchikov’s c-theorem to d > 2
(known as F-theorem in d = 3 and a-theorem in d = 4) [68–71] or the
44 João Penedones
existence of universal bounds on the three-point function of the stress tensor

and its relation to the idea of energy correlators [72–74].
Onother example along this line is the existence of “double-trace” op-
erators with large spin in any CFT. The precise statement is that in the
OPE of two operators O1 and O2 there is an infinite number of operators
On,l of spin l 1 and scaling dimension
n
n,l ⇡ 1 + 2 + 2n + l + (155)
l⌧min
where ⌧min is the minimal twist (dimension minus spin) of all the operators
that appear in both OPEs O1 ⇥ O1 and O2 ⇥ O2 . In a generic CFT, this
will be the stress tensor with ⌧min = d 2 and one can derive explicit
formulas for n [75–78]. This statement has been proven using the confor-
mal bootstrap equations but its physical meaning is more intuitive in the
dual AdS language. Consider two particle primary states in AdS. Without
interactions the energy of such states is given by 1 + 2 + 2n + l where
n = 0, 1, 2, . . . is a radial quantum number and l is the spin. Turning on
interactions will change the energies of these two-particle states. However,
the states with large spin and fixed n correspond to two particles orbitating
each other at large distances and therefore they will su↵er a small energy
shift due to the gravitational long range force. At large spin, all other in-
teractions (corresponding to operators with higher twist) give subdominant
contributions to this energy shift. In other words, the general result (155) is
the CFT reflection of the simple fact that interactions decay with distance
in the dual AdS picture.
5. Mellin amplitudes
Correlation functions of local operators in CFT are rather complicated func-

tions of the cross-ratios. Since these are crucial observables in AdS/CFT
it is useful to find simpler representations. This is the motivation to study
Mellin amplitudes. They were introduced by G. Mack in 2009 [79, 80]
following earlier work [81, 82]. Mellin amplitudes share many of the prop-
erties of scattering amplitudes of dual resonance models. In particular,
they are crossing symmetric and have a simple analytic structure (related
to the OPE). As we shall see, in the case of holographic CFTs, we can
take this analogy further and obtain bulk flat space scattering amplitudes
as a limit of the dual CFT Mellin amplitudes. Independently of AdS/CFT
applications, Mellin amplitudes can be useful to describe CFTs in general.
5.1. Definition
Consider the n-point function of scalar primary operators s
ˆ Y ( ij )
hO1 (P1 ) . . . On (Pn )i = [d ]M ( ij ) (156)
( 2Pi · Pj ) ij
1i<jn
Conformal invariance requires weight i in each Pi . This leads to con-

straints in the Mellin variables which can be conveniently written as
n
X
ij =0, ij = ji , ii = i . (157)
j=1
Notice that for n = 2 and n = 3 the Mellin variables are entirely fixed
by these constraints. In these cases, there is no integral to do and the
Mellin representation just gives the known form of the conformal two and
three point function. The integration measure [d ] is over the n(n 3)/2
1
independent Mellin variables (including a factor of 2⇡i for each variable)
and the integration contours run parallel to the imaginary axis. The precise
contour in the complex plane is dictated by the requirement that it should
pass to the right/left of the semi-infinite sequences of poles of the integrand
that run to the left/right. This will become clear in the following example.
12
3 2 1 0 14
Fig. 4. Integration contour for the Mellin variable 12 . The crosses represent (double)
poles of the -functions given by (160) and (161). In general, the Mellin amplitude has
several semi-infinite sequence of poles. Each sequence should stay entirely on one side
of the contour.
Consider the case of a four-point function of a scalar operator of dimen-

sion . In this case, there are two independent Mellin variables which we
s We shall use the notation M ( ij ) to denote a function M ( 12 , 13 , . . . ) of all Mellin
variables.
46 João Penedones
can choose to be 12 and 14 .

This leads to
ˆ i1
1 d 12 14
hO(P1 ) . . . O(P4 )i = 2
M̂ ( 12 , 14 )u
12
v 14
,
(P13 P24 ) i1 (2⇡i)
(158)
where u and v are the cross ratios (11) and
2 2 2
M̂ ( 12 , 14 ) = M( 12 , 14 ) ( 12 ) ( 14 ) ( 12 14 ) . (159)
Consider the first the complex plane of 12 depicted in figure 4. The -
functions give rise to semi-infinite sequences of (double) poles at
12 = 0, 1, 2, . . . (160)
12 = 14 , 14 + 1, 14 + 2, . . . (161)
As we shall see in the next section, the Mellin amplitude M ( ij ) also has
the same type of semi-infinite sequences of poles. The integration contour
should pass in the middle of these sequences of poles as shown in figure
4. Invariance of the four-point function under permutation of the insertion
points Pi , leads to crossing symmetry of the Mellin amplitude
M( 12 , 13 , 14 ) = M( 13 , 12 , 14 ) = M( 14 , 13 , 12 ) , (162)
where we used 3 variables obeying a single constraint 12 + 13 + 14 = .
This is reminiscent of crossing symmetry of scattering amplitudes written
in terms of Mandelstam invariants.
It is convenient to introduce fictitious momenta pi such that ij = pi ·pj .
Pn
Imposing momentum conservation i=1 pi = 0 and the on-shell condition
p2i = i automatically leads to the constraints (157). These fictitious
momenta are a convenient trick but we do not know how to define them
directly. In all formulas, we will only use their inner products ij = pi · pj .
In particular, it is not clear in what vector space do the momenta pi live. t
Let us be more precise about the number of independent cross ratios.
The correct formula is
n(n 3)
, nd+2 (163)
2
(d + 1)(d + 2)
nd , n d+2 (164)
2
In fact, for n > d + 2 one can write identities like
det Pi · Pj = 0 (165)
i,j
t Theflat space limit of AdS discussed in section 5.3.2, suggests a d + 1 dimensional space
but this is unclear before the limit.
using d + 3 embedding space vectors. Notice that this makes the Mellin
representation non-unique. We can shift the Mellin amplitude by the Mellin
transform of
F (P1 , . . . , Pn ) det Pi · Pj = 0 (166)
i,j
where F is any scalar function with the appropriate homogeneity prop-

erties. This non-uniqueness of the Mellin amplitude is analogous to the
non-uniqueness of the n-particle scattering amplitudes (as functions of the
invariants ki · kj ) in (d + 1)-dimensional spacetime if n > d + 2.
5.2. OPE ) Factorization

Consider the OPE
X k 1 2 ⇥ ⇤
O1 (x1 )O1 (x2 ) = C12k x212 2
Ok (x2 ) + c x212 @ 2 Ok (x2 ) + . . .
k
(167)
where the sum is over primary operators Ok and, for simplicity, we wrote the
contribution of a scalar operator. The term proportional to the constant c
is a descendant and is fixed by conformal symmetry like all the other terms
represented by . . . . Let us compare this with the Mellin representation.
When x212 ! 0, it is convenient to integrate over 12 closing the contour to
the left in the 12 -complex plane. This gives
X ˆ 0
Y
2 ¯12 0
x2ij
ij
hO1 (x1 )O1 (x2 ) . . . i = x12 [d ] Res ¯12 M̂ ( ij )
¯12
(168)
Q0
where [d ]0 and stand for the integration measure and product excluding
ij = 12. Comparing the two expressions we conclude that M̂ must have
poles at
1 + 2
k 2m
12 = , m = 0, 1, 2, . . . (169)
2
where the poles with m > 0 correspond to descendant contributions. If
the CFT has a discrete spectrum of scaling dimensions then its Mellin
amplitudes are analytic functions with single poles as its only singularities
(meromorphic functions). It is also clear that the residues of these poles will
be proportional to the product of the OPE coefficient C12k and the Mellin
amplitude of the lower point correlator hOk . . . i. The precise formulas
are derived in [79, 83]. Here we shall just list the main results without
derivation.
48 João Penedones
5.2.1. Four-point function

In the case of the four-point function it is convenient to write the Mellin
amplitude in terms of ‘Mandelstam invariants’
s= (p1 + p2 )2 = 1 + 2 2 12 (170)
2
t= (p1 + p3 ) = 1 + 3 2 13 (171)
Then, the poles and residues of the Mellin amplitude take the following
form [79]
Qlk ,m (t)
M (s, t) ⇡ C12k C34k , m = 0, 1, 2, . . . (172)
s k + lk 2m
where Ql,m (t) is a kinematical polynomial of degree l in the variable t.
This strengthens the analogy with scattering amplitudes. Each operator
of spin l in the OPE O1 ⇥O2 gives rise to poles in the Mellin amplitude very
similar to the poles in the scattering amplitude associated to the exchange
of a particle of the same spin.
5.2.2. Planar correlators

Notice that the polynomial behaviour of the residues requires the inclusion
of the -functions in the definition (156) of Mellin amplitudes. On the
other hand, the -functions themselves have poles at fixed positions. For
example, ( 12 ) gives rise to poles at s = 1 + 2 +2m with m = 0, 1, 2, . . . .
In a generic CFT, there are no operators with these scaling dimensions and
therefore the Mellin amplitude must have zeros at these values to cancel
these unwanted OPE contributions. However, in correlation functions of
single-trace operators in large N CFTs we expect precisely this type of
contributions. At the planar level, the -functions account for all multi-
trace OPE contributions and the Mellin amplitude only has poles associated
to single-trace operators.
5.2.3. n-point function

Considering the OPE of k scalar operators, one can derive more general
factorization formulas [83]. For example, for each primary operator O of
dimension and spin l that appears in the OPEs O1 ⇥ · · · ⇥ Ok and
Ok+1 ⇥ · · · ⇥ On , we obtain the following sequence of poles in the n-point
Mellin amplitude,
Qm
Mn ⇡ , m = 0, 1, 2, . . . (173)
LR +l 2m
where
k
!2 k X
n
X X
LR = pi = ij . (174)
i=1 i=1 j>k
In general, the residue can be written in terms of lower point Mellin ampli-
tudes. For example, if l = 0 the residue factorizes
L
Q0 = 2 ( )Mk+1 MnR k+1 , (175)
L
with Mk+1 the Mellin amplitude of hO1 . . . Ok Oi and MnR k+1 the Mellin
amplitude of hOOk+1 . . . On i . The satellite poles also factorize but give
rise to more complicated formulae
2 ( )m!
Qm = d
L m Rm , (176)
2 +1 m
with
X Y ( ab )nab
Lm = M L( ab + nab ) (177)
nab !
n
P ab
0 1a<bk
nab =m
and similarly for Rm .

There also factorization formulas for the residues associated with oper-
ators with non-zero spin [83]. However, the general case including external
operators with spin has not been worked out.
5.3. Holographic CFTs

As discussed in section 4.1, holographic CFTs have two special properties:
large N factorization and a small number of low dimension single-trace
operators. Therefore, one should expect that the corresponding Mellin
amplitudes are particularly simple, at least at the planar level. We shall
now confirm this expectation with a few simple examples.
5.3.1. Witten diagrams

Consider the contact Witten diagram of figure 5. It corresponds to an
u
interaction vertex 1 . . . n in the bulk lagrangian and it contributes
n p
ˆ Y C i
hO1 (P1 ) . . . On (Pn )i = dX (178)
AdS i=1
( 2Pi · X) i
u We are using CFT operators Oi normalized to have unit two point function.
50 João Penedones
O2 (P2 )
X
O1 (P1 )
On (Pn )
Fig. 5. Witten diagram for a n-point contact interaction in AdS. The interior of the disk
represents the bulk of AdS and the circumference represents its conformal boundary. The
lines connecting the boundary points Pi to the bulk point X represent bulk to boundary
propagators.
to the dual CFT correlation function. One can show that this corresponds
to a constant Mellin amplitude,
✓P ◆ n p
1 d i d Y C i
M= ⇡ 2 . (179)
2 2 i=1
( i)
Exercise 5.1. Check the last statement. Start by using the integral repre-
sentation of the bulk to boundary propagators and performing the integral
over AdS using Poincare coordinates as explained in exercise 3.5. This
turns (178) into
✓P ◆ˆ 1 P Yn p
d i d s s P C i i 1
⇡2 e i<j i j ij
s dsi . (180)
2 0 i=1
( i) i
Next, use the Mellin representation (c > 0)

ˆ c+i1
d ij
e si sj Pij = ( ij )(si sj Pij ) ij
(181)
c i1 2⇡i
for n(n 3)/2 exponential factors. A good choice is to keep n factors,

corresponding to the exponential
Pn
s1 si P1i s2 s3 P23
e i=2 . (182)
The integrals over s4 , . . . , sn can be easily done in terms of -functions.

Finally, do the integrals over s1 , s2 , s3 using the same type of change of
variables as in exercise 3.5.
This result can be easily generalized to interaction vertices with deriva-

tives. For example, the vertex (r↵ 1 r↵ 2 ) 3 . . . n gives rise to
n p
ˆ Y C i
hO1 (P1 ) . . . On (Pn )i = dX ⇥ (183)
AdS i=3
( 2P i · X) i
p p
AB A B @ C 1 @ C 2
⇥ (⌘ +X X ) .
@X A ( 2P1 · X) 1 @X B ( 2P2 · X) 2
Here we have used the fact that covariant derivatives in AdS can be com-
puted as partial derivatives in the embedding space projected to the AdS
sub-manifold.v This gives
n
Y p
1 2 ( 2P12 D 1 +1, 2 +1, 3 ,..., n
+D 1, 2, 3 ,..., n
) C i
(184)
i=1
where we introduced the D-function [33]

Yn
1
ˆ
D 1 ,..., n ⌘ dX . (185)
AdS i=1
( 2Pi · X) i
More generally, it is clear that the contact Witten diagram associated with
a generic vertex r . . . r 1 r . . . r 2 . . . r . . . r n with all derivatives con-
tracted among di↵erent fields, gives rise to a linear combination of terms
of the form
Y n
D 1 +⇤1 ,..., n +⇤n Pijij (186)
i<j
P
where ij are non-negative integers and ⇤i = j6=i ij . As we will see in the
next exercise, the Mellin amplitude of (186) is a polynomial in the Mellin
variables. Therefore, the Mellin amplitude associated to contact Witten
diagrams is polynomial. The absence of poles in the Mellin amplitude
means that the conformal block decomposition of the contact diagram only
contains multi-trace operators, in agreement with previous results [85, 86].
Exercise 5.2. If the vertex r . . . r 1 r . . . r 2 . . . r . . . r n has 2N =

P
2 i<j ↵ij derivatives with ↵ij contractions of derivatives acting on i and
j , show that the contact Witten diagram is given by
n
! n
Y p Y
C i D 1 +⇤1 ,..., n +⇤n ( 2Pij )↵ij + . . . (187)
i=1 i<j
v Seeappendix F.1 of [84] for a derivation of this statement in the analogous case of a
sphere embedded in Euclidean space.
52 João Penedones
P
where ⇤i = j6=i ↵ij and the . . . represent similar terms with less Pij
factors. Hint: use the trick of writing covariant derivatives in AdS as partial
derivatives in the embedding space projected to the AdS sub-manifold.
The Mellin representation of the D-functions is very simple. As we saw
in exercise 5.1, the Mellin amplitude associated to D 1 ,..., n is simply
⇣P ⌘
d i d
⇡2 2
Qn . (188)
2 i=1 ( i )
Show that the Mellin amplitude associated to the correlation function (187)
is given by the polynomial
! d ⇣ P i +2N d ⌘ n
Yn
p ⇡2 2 Y
C i Qn ( 2 ij )↵ij + . . . (189)
i=1
2 i=1 ( i + ⇤ )
i i<j
where the . . . represent terms of lower degree in ij . Hint: this follows

easily from shifting the integration variables in the Mellin representation
(156).
O2 (P2 ) O3 (P3 )
,l
O1 (P1 ) O4 (P4 )
Fig. 6. Witten diagram describing the exchange of a bulk field dual to an operator of
dimension and spin l.
Consider now the Witten diagram shown in figure 6 describing the ex-
change of a bulk field dual to a single-trace boundary operator O of dimen-
sion and spin l. The conformal block decomposition of this diagram in the
(12)(34) channel contains the single-trace operator O plus double-trace op-
erators schematically of the form O1 (@ 2 )n @µ1 ...µj O2 and O3 (@ 2 )n @µ1 ...µj O4 .
Moreover, the OPE in the crossed channels only contains double-trace op-
erators. This means that the Mellin amplitude is of the form
1
X Ql,m (t)
M = C12O C34O + R(s, t) (190)
m=0
s + l 2m
where the OPE coefficients C12O and C34O are proportional to the bulk
cubic couplings and R(s, t) is an analytic function. The residues are pro-
portional to degree l Mack polynomials Ql,m (t) which are entirely fixed by
conformal symmetry as we saw in 5.2.1. If we choose minimal coupling
between the spin l bulk field and the external scalars, then R(s, t) is a
polynomial of degree  l 1. This is particularly simple in the case of
a scalar exchange (l = 0). Then the residues are independent of t and
R = 0 [87]. Notice that this simple looking Mellin amplitude gives rise to
a rather involved function of the cross-ratios in position space. This ex-
ample illustrates clearly the advantage of using the Mellin reprsentation to
describe Witten diagrams.
The Mellin amplitude of a general tree-level scalar Witten diagrams was
determined in [88–91]. The final result can be summarized in the following
Feynman rules:
• Associate a momentum pj to every line (propagator) in the Witten

diagram. External lines have incoming momentum pi satisfying
p2i = i . Momentum is conserved at every vertex of the diagram.
• Assign an integer mj to every line. External lines have mi = 0.
• Every internal line (bulk-to-bulk propagator) contributes a factorw
Smjj
(192)
p2j + j + 2mj
where j is the dimension of the propagating scalar field.

• Every vertex, g 1 . . . k joining k lines, contributes a factorx
... k
g Vm11...m k
(193)
• Sum over all integers mj associated with internal lines. Each sum
runs from 0 to 1.
w The propagator numerator is given by
⇣ ⌘
d
2
+1+m
Sm = ⇣ ⌘. (191)
2 d
2(m!) 2
+1
x The vertex factor is given by

m1 mk P ! k
...
X X j( j + 2nj ) d Y ( m j )n j
Vm11...mkk = ··· ⇣ ⌘ .
n1 =0 nk =0
2 j=1 nj ! j
d
+1
2 nj
54 João Penedones
• Multiply by
n p
⇡2 Y C i
d
N = (194)
2 i=1 ( i )
to get the n-point Mellin amplitude in our normalization of the
external operators.
O1 O5
5, 0
p1 p5
1,
0
,0 7 4,
2 6,
m ,m 0
p2 p6 6 7p7 p4
O2 O4
3, 0
p3
O3
Fig. 7. A tree level scalar Witten diagram contributing to a 5-point function. The
auxiliary momenta pi is conserved at each vertex, i.e. p6 = p1 + p2 and p7 = p4 + p5 .
As an example, the Witten diagram in figure 7 gives rise to the following

Mellin amplitude
1
X X1
Sm66 Sm77
N V0 01m62 6 2 Vm660 m3 7
V 7 4 5
m =0 m =0
p6 + 6 + 2m6 7
p7 + 7 + 2m7 m7 0 0
2
6 7
where p26 = (p1 +p2 )2 = 2 12 1

2 2
2 and p7 = (p4 +p5 ) = 2 45 4 5.
These Feynman rules suggest that we should think of the Mellin amplitude
as an amputated amplitude because the bulk to boundary propagators do
not contribute. In the case of scalar tree level diagrams (with non-derivative
interaction vertices), the only dependence in the Mellin variables ij comes
from the bulk-to-bulk propagators. It is not known how to generalize these
Feynman rules for loop diagrams or tree-level diagrams involving fields with
spin. There are partial results in literature [83, 89] but nothing systematic.
Mellin amplitudes are also useful in the context of weakly coupled CFTs.
The associated Feynman rules for tree level diagrams were given in [92].
Exercise 5.3. Consider the residue of the Mellin amplitude at the first
pole (m = 0) associated to a bulk-to-bulk propagator. Show that the Feyn-
man rules above are compatible with the factorization property (175) of this
residue. Extra: check the factorization formula (176) for the satellite poles
with m > 0.
5.3.2. Flat space limit of AdS

If we consider a scattering process where all length scales are much smaller
than the AdS radius R then the curvature e↵ects should be negligible. Con-
sider a relativistic invariant theory in flat spacetime with a characteristic
length scale `s (this scale could come from a mass or from a dimensionful
coupling). Then, a scattering amplitude Tn of n massless scalar particles
in this theory will depend on `s and on the relativistic invariants ki · kj ,
where ki are the momenta of the external particles. On the other hand,
this theory in AdS will give rise to Mellin amplitudes that depend on the
dimensionless parameter ✓ = R/`s and the Mellin variables ij . We claim
that these two quantities are related by
⇣ ⌘
✓2 2
Tn (`s , ki ) 1
ˆ
d↵ d
P
i
M n ✓, ij = 2↵ ` s k i · k j
= lim ↵ 2 e↵ (195)
n d2 1 d 1 ✓!1 N 2⇡i ✓ n 1 d
+d+1
` s
2
where the contour runs parallel to the imaginary axis and passes to the
right of the branch point at ↵ = 0 and to the left of all poles of Mn .
The powers of `s where introduced to make both sides of the equation
dimensionless and the constant N was given in (194). The external particles
are massless in flat space but in AdS they can have any scaling dimension
i of order 1. We expect this equation to hold when both sides of the
equation are well defined. In case the flat space scattering amplitude Tn is
IR divergent, we expect that the limit ✓ ! 1 of the Mellin amplitude will
not be finite.y
Exercise 5.4. Consider the vertex r . . . r 1 . . . r . . . r n discussed in

exercise 5.2 in d + 1 spacetime dimensions. Start by writing the coupling
constant as a power (`s )q of a characteristic length scale `s and determine
the value of q. Then, use the Mellin amplitude (189) in the flat space limit
formula (195) and obtain the expected n-particle scattering amplitude
Y
Tn = ( ki · kj )↵ij . (196)
i<j
The last exercise can be seen as a derivation of the flat space limit
formula (195). The point is that a generic Feynman diagram can be written
y It might be useful to think of large ✓ as an IR regulator for the scattering amplitude.
56 João Penedones
as a (infinite) sum of contact diagrams with any number of derivatives.

This corresponds to integrating out the internal particles and replacing
there e↵ect by contact vertices among the external particles. Since formula
(195) works for any contact diagram it should work in general. This has
been tested in several explicit examples, including 1-loop diagrams [87, 88,
93]. In addition, the same formula was derived in [90] using a wave-packet
construction where the scattering region was limited to a small flat region
of AdS.
In principle, formula (195) provides a non-perturbative definition of
string theory scattering amplitudes in terms of SYM correlation functions.
However, we do not know how to directly compute SYM correlators at
strong coupling. In practice, what we can do is to use formula (195) in
the opposite direction, i.e. we can use known string scattering amplitudes
in flat space to obtain information about the strong coupling expansion of
SYM correlators [87, 94]. If the external particles are massive in flat space
then formula (195) is not adequate. This case was studied in [95].
5.4. Open questions
The study of Mellin amplitudes is still very incomplete. Firstly, it is impor-

tant to understand in what conditions do we have a well defined analytic
Mellin amplitude. For example, in free CFTs the Mellin representation
requires some form of regularization. This might be a technical detail but
it would be useful to understand in general the status of the Mellin repre-
sentation. Another important question is the asymptotic behavior of the
Mellin amplitude when the Mellin variables are large. In the case of the
four-point Mellin amplitude discussed in 5.2.1, the limit of large s with
fixed t is called the Regge limit in analogy with flat space scattering am-
plitudes. In [96], we studied this limit using Regge theory techniques and
making some reasonable assumptions about the large spin behaviour of the
conformal partial amplitudes. Proving these assumptions is an important
open question. The bound on chaos [97] is another possible approach to the
Regge limit of Mellin amplitudes. Notice that if we can tame the asymp-
totic behaviour of M (s, t) when s ! 1, then we can write a dispersion
relation that expresses M (s, t) in terms of its poles in s, which are given
by (172). This could provide a reformulation of the conformal bootstrap
approach.
In the holographic context, it would be interesting to establish more
general Feynman rules for Mellin amplitudes associated to Witten diagrams
involving loops and particles with spin. It would also be useful to generalize
more modern approaches to scattering amplitudes, like BCFW [98] or CHY
[99], to Mellin amplitudes.
Acknowledgements
I would like to thank the students of TASI 2015 for their questions and en-
thusiasm. I am also grateful to the organizers Joe Polchinski, Pedro Vieira,
Tom DeGrand and Oliver DeWolfe for their invitation and their patience
with my delays in finishing these notes. Centro de Fı́sica do Porto is par-
tially funded by FCT. The research leading to these results has received
funding from the People Programme (Marie Curie Actions) of the Euro-
pean Union’s Seventh Framework Programme FP7/2007-2013/ under REA
Grant Agreement No 317089 and CERN/FIS-NUC/0045/2015.
References
[1] J. M. Maldacena, “The Large N limit of superconformal field theories and

supergravity,” Int.J.Theor.Phys. 38 (1999) 1113–1133,
arXiv:hep-th/9711200 [hep-th].
[2] S. S. Gubser, I. R. Klebanov, and A. M. Polyakov, “Gauge theory
correlators from noncritical string theory,” Phys. Lett. B428 (1998)
105–114, arXiv:hep-th/9802109 [hep-th].
[3] E. Witten, “Anti-de Sitter space and holography,” Adv. Theor. Math.
Phys. 2 (1998) 253–291, arXiv:hep-th/9802150 [hep-th].
[4] A. Almheiri, D. Marolf, J. Polchinski, and J. Sully, “Black Holes:
Complementarity or Firewalls?,” JHEP 02 (2013) 062, arXiv:1207.3123
[hep-th].
[5] D. Simmons-Duffin, “TASI Lectures on the Conformal Bootstrap,”
arXiv:1602.07982 [hep-th].
[6] S. Rychkov, “EPFL Lectures on Conformal Field Theory in D 3
Dimensions,” arXiv:1601.05000 [hep-th].
[7] J. D. Qualls, “Lectures on Conformal Field Theory,” arXiv:1511.04074
[hep-th].
[8] O. Aharony, S. S. Gubser, J. M. Maldacena, H. Ooguri, and Y. Oz, “Large
N field theories, string theory and gravity,” Phys. Rept. 323 (2000)
183–386, arXiv:hep-th/9905111 [hep-th].
[9] J. McGreevy, “Holographic duality with a view toward many-body
physics,” Adv. High Energy Phys. 2010 (2010) 723105, arXiv:0909.0518
[hep-th].
[10] E. D’Hoker and D. Z. Freedman, “Supersymmetric gauge theories and the
AdS / CFT correspondence,” arXiv:hep-th/0201253 [hep-th].
58 João Penedones
[11] D. Mateos, “String Theory and Quantum Chromodynamics,” Class.

Quant. Grav. 24 (2007) S713–S740, arXiv:0709.1523 [hep-th].
[12] H. Nastase, “Introduction to AdS-CFT,” arXiv:0712.0689 [hep-th].
[13] S. S. Gubser and A. Karch, “From gauge-string duality to strong
interactions: A Pedestrian’s Guide,” Ann. Rev. Nucl. Part. Sci. 59 (2009)
145–168, arXiv:0901.0935 [hep-th].
[14] S. A. Hartnoll, “Lectures on holographic methods for condensed matter
physics,” Class. Quant. Grav. 26 (2009) 224002, arXiv:0903.3246
[hep-th].
[15] J. Polchinski, “Introduction to Gauge/Gravity Duality,” arXiv:1010.6134
[hep-th].
[16] S. A. Hartnoll, “Horizons, holography and condensed matter,”
arXiv:1106.4324 [hep-th].
[17] J. Kaplan, “Lectures on AdS/CFT from the Bottom Up.”
http://www.pha.jhu.edu/~jaredk/AdSCFTCourseNotesPublic.pdf.
[18] P. A. M. Dirac, “Wave equations in conformal space,” Annals Math. 37
(1936) 429–442.
[19] G. Mack and A. Salam, “Finite component field representations of the
conformal group,” Annals Phys. 53 (1969) 174–202.
[20] D. G. Boulware, L. S. Brown, and R. D. Peccei, “Deep-inelastic
electroproduction and conformal symmetry,” Phys. Rev. D2 (1970)
293–298.
[21] S. Ferrara, R. Gatto, and A. F. Grillo, “Conformal algebra in space-time
and operator product expansion,” Springer Tracts Mod. Phys. 67 (1973)
1–64.
[22] S. Ferrara, A. F. Grillo, and R. Gatto, “Tensor representations of
conformal algebra and conformally covariant operator product expansion,”
Annals Phys. 76 (1973) 161–188.
[23] L. Cornalba, M. S. Costa, and J. Penedones, “Deep Inelastic Scattering in
Conformal QCD,” JHEP 03 (2010) 133, arXiv:0911.0043 [hep-th].
[24] S. Weinberg, “Six-dimensional Methods for Four-dimensional Conformal
Field Theories,” Phys. Rev. D82 (2010) 045031, arXiv:1006.3480
[hep-th].
[25] M. S. Costa, J. Penedones, D. Poland, and S. Rychkov, “Spinning
Conformal Correlators,” JHEP 11 (2011) 071, arXiv:1107.3554 [hep-th].
[26] G. ’t Hooft, “A Planar Diagram Theory for Strong Interactions,”
Nucl.Phys. B72 (1974) 461.
[27] E. Witten, “Baryons in the 1/n Expansion,” Nucl. Phys. B160 (1979)
57–115.
[28] F. A. Dolan and H. Osborn, “Conformal four point functions and the
operator product expansion,” Nucl. Phys. B599 (2001) 459–496,
[29] S. El-Showk and K. Papadodimas, “Emergent Spacetime and Holographic
CFTs,” JHEP 10 (2012) 106, arXiv:1101.4163 [hep-th].
[30] H. Liu and A. A. Tseytlin, “D = 4 superYang-Mills, D = 5 gauged
supergravity, and D = 4 conformal supergravity,” Nucl. Phys. B533
(1998) 88–108, arXiv:hep-th/9804083 [hep-th].

[31] H. Liu and A. A. Tseytlin, “On four point functions in the CFT / AdS
correspondence,” Phys. Rev. D59 (1999) 086002, arXiv:hep-th/9807097
[hep-th].
[32] E. D’Hoker, D. Z. Freedman, S. D. Mathur, A. Matusis, and L. Rastelli,
“Graviton and gauge boson propagators in AdS(d+1),” Nucl. Phys. B562
(1999) 330–352, arXiv:hep-th/9902042 [hep-th].
[33] E. D’Hoker, D. Z. Freedman, S. D. Mathur, A. Matusis, and L. Rastelli,
“Graviton exchange and complete four point functions in the AdS / CFT
correspondence,” Nucl. Phys. B562 (1999) 353–394,
[34] M. S. Costa, V. Gonçalves, and J. Penedones, “Spinning AdS
Propagators,” JHEP 09 (2014) 064, arXiv:1404.5625 [hep-th].
[35] V. Balasubramanian and P. Kraus, “A Stress tensor for Anti-de Sitter
gravity,” Commun. Math. Phys. 208 (1999) 413–428,
[36] K. Skenderis, “Lecture notes on holographic renormalization,” Class.
Quant. Grav. 19 (2002) 5849–5876, arXiv:hep-th/0209067 [hep-th].
[37] H. Osborn and A. C. Petkou, “Implications of conformal invariance in field
theories for general dimensions,” Annals Phys. 231 (1994) 311–362,
[38] I. Heemskerk, J. Penedones, J. Polchinski, and J. Sully, “Holography from
Conformal Field Theory,” JHEP 10 (2009) 079, arXiv:0907.0151
[hep-th].
[39] X. O. Camanho, J. D. Edelstein, J. Maldacena, and A. Zhiboedov,
“Causality Constraints on Corrections to the Graviton Three-Point
Coupling,” JHEP 02 (2016) 020, arXiv:1407.5597 [hep-th].
[40] N. Arkani-Hamed, F. Cachazo, and J. Kaplan, “What is the Simplest
Quantum Field Theory?,” JHEP 09 (2010) 016, arXiv:0808.1446
[hep-th].
[41] N. Beisert et al., “Review of AdS/CFT Integrability: An Overview,” Lett.
Math. Phys. 99 (2012) 3–32, arXiv:1012.3982 [hep-th].
[42] S. Lee, S. Minwalla, M. Rangamani, and N. Seiberg, “Three point
functions of chiral operators in D = 4, N=4 SYM at large N,” Adv. Theor.
Math. Phys. 2 (1998) 697–718, arXiv:hep-th/9806074 [hep-th].
[43] V. Pestun, “Localization of gauge theory on a four-sphere and
supersymmetric Wilson loops,” Commun. Math. Phys. 313 (2012) 71–129,
[44] J. K. Erickson, G. W. Semeno↵, and K. Zarembo, “Wilson loops in N=4
supersymmetric Yang-Mills theory,” Nucl. Phys. B582 (2000) 155–175,
[45] N. Gromov, V. Kazakov, and P. Vieira, “Exact Spectrum of Planar N = 4
Supersymmetric Yang-Mills Theory: Konishi Dimension at Any Coupling,”
Phys. Rev. Lett. 104 (2010) 211601, arXiv:0906.4240 [hep-th].
[46] B. Basso, A. Sever, and P. Vieira, “Spacetime and Flux Tube S-Matrices at
Finite Coupling for N=4 Supersymmetric Yang-Mills Theory,” Phys. Rev.
60 João Penedones
Lett. 111 no. 9, (2013) 091602, arXiv:1303.1396 [hep-th].

[47] B. Basso, S. Komatsu, and P. Vieira, “Structure Constants and Integrable
Bootstrap in Planar N=4 SYM Theory,” arXiv:1505.06745 [hep-th].
[48] T. Banks, W. Fischler, S. H. Shenker, and L. Susskind, “M theory as a
matrix model: A Conjecture,” Phys. Rev. D55 (1997) 5112–5128,
[49] S. Catterall and T. Wiseman, “Towards lattice simulation of the gauge
theory duals to black holes and hot strings,” JHEP 12 (2007) 104,
arXiv:0706.3518 [hep-lat].
[50] K. N. Anagnostopoulos, M. Hanada, J. Nishimura, and S. Takeuchi,
“Monte Carlo studies of supersymmetric matrix quantum mechanics with
sixteen supercharges at finite temperature,” Phys. Rev. Lett. 100 (2008)
021601, arXiv:0707.4454 [hep-th].
[51] S. Catterall and T. Wiseman, “Black hole thermodynamics from
simulations of lattice Yang-Mills theory,” Phys. Rev. D78 (2008) 041502,
[52] M. Hanada, Y. Hyakutake, J. Nishimura, and S. Takeuchi, “Higher
derivative corrections to black hole thermodynamics from supersymmetric
matrix quantum mechanics,” Phys. Rev. Lett. 102 (2009) 191602,
[53] S. Catterall and T. Wiseman, “Extracting black hole physics from the
lattice,” JHEP 04 (2010) 077, arXiv:0909.4947 [hep-th].
[54] M. Hanada, J. Nishimura, Y. Sekino, and T. Yoneya, “Direct test of the
gauge-gravity correspondence for Matrix theory correlation functions,”
JHEP 12 (2011) 020, arXiv:1108.5153 [hep-th].
[55] M. Hanada, Y. Hyakutake, G. Ishiki, and J. Nishimura, “Holographic
description of quantum black hole on a computer,” Science 344 (2014)
882–885, arXiv:1311.5607 [hep-th].
[56] J. Polchinski and E. Silverstein, “Dual Purpose Landscaping Tools: Small
Extra Dimensions in AdS/CFT,” arXiv:0908.0756 [hep-th].
[57] S. de Alwis, R. K. Gupta, F. Quevedo, and R. Valandro, “On KKLT/CFT
and LVS/CFT Dualities,” JHEP 07 (2015) 036, arXiv:1412.6999
[hep-th].
[58] I. R. Klebanov and A. M. Polyakov, “AdS dual of the critical O(N) vector
model,” Phys. Lett. B550 (2002) 213–219, arXiv:hep-th/0210114
[hep-th].
[59] E. Sezgin and P. Sundell, “Massless higher spins and holography,” Nucl.
Phys. B644 (2002) 303–370, arXiv:hep-th/0205131 [hep-th]. [Erratum:
Nucl. Phys.B660,403(2003)].
[60] S. Giombi, S. Minwalla, S. Prakash, S. P. Trivedi, S. R. Wadia, and X. Yin,
“Chern-Simons Theory with Vector Fermion Matter,” Eur. Phys. J. C72
(2012) 2112, arXiv:1110.4386 [hep-th].
[61] O. Aharony, G. Gur-Ari, and R. Yacoby, “Correlation Functions of Large
N Chern-Simons-Matter Theories and Bosonization in Three Dimensions,”
JHEP 12 (2012) 028, arXiv:1207.4593 [hep-th].
[62] M. R. Gaberdiel and R. Gopakumar, “Minimal Model Holography,” J.
Phys. A46 (2013) 214002, arXiv:1207.6697 [hep-th].

[63] P. Kovtun and A. Ritz, “Black holes and universality classes of critical
points,” Phys. Rev. Lett. 100 (2008) 171606, arXiv:0801.2785 [hep-th].
[64] S. W. Hawking and D. N. Page, “Thermodynamics of Black Holes in
anti-De Sitter Space,” Commun. Math. Phys. 87 (1983) 577.
[65] S. Ryu and T. Takayanagi, “Holographic derivation of entanglement
entropy from AdS/CFT,” Phys. Rev. Lett. 96 (2006) 181602,
[66] J. M. Maldacena, “Wilson loops in large N field theories,” Phys. Rev. Lett.
80 (1998) 4859–4862, arXiv:hep-th/9803002 [hep-th].
[67] P. Romatschke and U. Romatschke, “Viscosity Information from
Relativistic Nuclear Collisions: How Perfect is the Fluid Observed at
RHIC?,” Phys. Rev. Lett. 99 (2007) 172301, arXiv:0706.1522 [nucl-th].
[68] A. B. Zamolodchikov, “Irreversibility of the Flux of the Renormalization
Group in a 2D Field Theory,” JETP Lett. 43 (1986) 730–732. [Pisma Zh.
Eksp. Teor. Fiz.43,565(1986)].
[69] R. C. Myers and A. Sinha, “Holographic c-theorems in arbitrary
dimensions,” JHEP 01 (2011) 125, arXiv:1011.5819 [hep-th].
[70] Z. Komargodski and A. Schwimmer, “On Renormalization Group Flows in
Four Dimensions,” JHEP 12 (2011) 099, arXiv:1107.3987 [hep-th].
[71] H. Casini and M. Huerta, “On the RG running of the entanglement entropy
of a circle,” Phys. Rev. D85 (2012) 125016, arXiv:1202.5650 [hep-th].
[72] D. M. Hofman and J. Maldacena, “Conformal collider physics: Energy and
charge correlations,” JHEP 05 (2008) 012, arXiv:0803.1467 [hep-th].
[73] D. M. Hofman, D. Li, D. Meltzer, D. Poland, and F. Rejon-Barrera, “A
Proof of the Conformal Collider Bounds,” JHEP 06 (2016) 111,
arXiv:1603.03771 [hep-th].
[74] T. Faulkner, R. G. Leigh, O. Parrikar, and H. Wang, “Modular
Hamiltonians for Deformed Half-Spaces and the Averaged Null Energy
Condition,” arXiv:1605.08072 [hep-th].
[75] A. L. Fitzpatrick, J. Kaplan, D. Poland, and D. Simmons-Duffin, “The
Analytic Bootstrap and AdS Superhorizon Locality,” JHEP 12 (2013) 004,
[76] Z. Komargodski and A. Zhiboedov, “Convexity and Liberation at Large
Spin,” JHEP 11 (2013) 140, arXiv:1212.4103 [hep-th].
[77] A. Kaviraj, K. Sen, and A. Sinha, “Analytic bootstrap at large spin,”
JHEP 11 (2015) 083, arXiv:1502.01437 [hep-th].
[78] A. Kaviraj, K. Sen, and A. Sinha, “Universal anomalous dimensions at
large spin and large twist,” JHEP 07 (2015) 026, arXiv:1504.00772
[hep-th].
[79] G. Mack, “D-independent representation of Conformal Field Theories in D
dimensions via transformation to auxiliary Dual Resonance Models. Scalar
amplitudes,” arXiv:0907.2407 [hep-th].
[80] G. Mack, “D-dimensional Conformal Field Theories with anomalous
dimensions as Dual Resonance Models,” Bulg. J. Phys. 36 (2009) 214–226,
62 João Penedones
[81] K. Symanzik, “On Calculations in conformal invariant field theories,” Lett.

Nuovo Cim. 3 (1972) 734–738.
[82] V. K. Dobrev, V. B. Petkova, S. G. Petrova, and I. T. Todorov,
“Dynamical Derivation of Vacuum Operator Product Expansion in
Euclidean Conformal Quantum Field Theory,” Phys. Rev. D13 (1976) 887.
[83] V. Gonçalves, J. Penedones, and E. Trevisani, “Factorization of Mellin
amplitudes,” JHEP 10 (2015) 040, arXiv:1410.4185 [hep-th].
[84] M. S. Costa, T. Hansen, J. Penedones, and E. Trevisani, “Projectors and
seed conformal blocks for traceless mixed-symmetry tensors,” JHEP 07
(2016) 018, arXiv:1603.05551 [hep-th].
[85] H. Liu, “Scattering in anti-de Sitter space and operator product
expansion,” Phys. Rev. D60 (1999) 106005, arXiv:hep-th/9811152
[hep-th].
[86] E. D’Hoker and D. Z. Freedman, “General scalar exchange in AdS(d+1),”
Nucl. Phys. B550 (1999) 261–288, arXiv:hep-th/9811257 [hep-th].
[87] J. Penedones, “Writing CFT correlation functions as AdS scattering
amplitudes,” JHEP 03 (2011) 025, arXiv:1011.1485 [hep-th].
[88] A. L. Fitzpatrick, J. Kaplan, J. Penedones, S. Raju, and B. C. van Rees,
“A Natural Language for AdS/CFT Correlators,” JHEP 11 (2011) 095,
[89] M. F. Paulos, “Towards Feynman rules for Mellin amplitudes,” JHEP 10
(2011) 074, arXiv:1107.1504 [hep-th].
[90] A. L. Fitzpatrick and J. Kaplan, “Analyticity and the Holographic
S-Matrix,” JHEP 10 (2012) 127, arXiv:1111.6972 [hep-th].
[91] D. Nandan, A. Volovich, and C. Wen, “On Feynman Rules for Mellin
Amplitudes in AdS/CFT,” JHEP 05 (2012) 129, arXiv:1112.0305
[hep-th].
[92] A. A. Nizami, A. Rudra, S. Sarkar, and M. Verma, “Exploring Perturbative
Conformal Field Theory in Mellin space,” arXiv:1607.07334 [hep-th].
[93] A. L. Fitzpatrick and J. Kaplan, “Unitarity and the Holographic
S-Matrix,” JHEP 10 (2012) 032, arXiv:1112.4845 [hep-th].
[94] V. Gonçalves, “Four point function of N = 4 stress-tensor multiplet at
strong coupling,” JHEP 04 (2015) 150, arXiv:1411.1675 [hep-th].
[95] M. F. Paulos, J. Penedones, J. Toledo, B. C. van Rees, and P. Vieira, “The
S-matrix Bootstrap I: QFT in AdS,” arXiv:1607.06109 [hep-th].
[96] M. S. Costa, V. Goncalves, and J. Penedones, “Conformal Regge theory,”
JHEP 12 (2012) 091, arXiv:1209.4355 [hep-th].
[97] J. Maldacena, S. H. Shenker, and D. Stanford, “A bound on chaos,”
arXiv:1503.01409 [hep-th].
[98] R. Britto, F. Cachazo, B. Feng, and E. Witten, “Direct proof of tree-level
recursion relation in Yang-Mills theory,” Phys. Rev. Lett. 94 (2005)
181602, arXiv:hep-th/0501052 [hep-th].
[99] F. Cachazo, S. He, and E. Y. Yuan, “Scattering of Massless Particles in
Arbitrary Dimensions,” Phys. Rev. Lett. 113 no. 17, (2014) 171601,

Joao Penedones PDF

Uploaded by

Copyright:

Available Formats

Joao Penedones PDF

Uploaded by

Document Information

Original Title

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

Joao Penedones PDF

Uploaded by

Copyright:

Available Formats

TASI Lectures on AdS/CFT

arXiv:1608.04948v1 [hep-th] 17 Aug 2016

We introduce the AdS/CFT correspondence as a natural extension of

The AdS/CFT correspondence [1–3] is a well established general approach

• System in a box - we work with Anti-de Sitter (AdS) boundary con-

Starting from these 3 facts it is entirely natural to define quantum gravity

gauge/gravity duality is that it gives us a geometric description of QFT

2. Conformal Field Theory

This section briefly describes the basic concepts necessary to formulate a

is a CFT observable [4].

2.1. Conformal Transformations

For simplicity, in most formulas, we will consider Euclidean signature. We

A conformal transformation is a coordinate transformation that preserves

✏µ (x) = aµ + xµ + mµ⌫ x⌫ + x2 bµ 2x↵ b↵ xµ . (3)

In spacetime dimension d > 2, conformal transformations form the

Exercise 2.2. Verify that inversion is a conformal transformation.

The form of the generators of the conformal algebra acting on functions

[D, Pµ ] = Pµ , [D, Kµ ] = Kµ , [Kµ , P⌫ ] = 2 µ⌫ D 2i Mµ⌫ ,

2.2. Local Operators

Correlation functions of scalar primary operators obey

for all conformal transformations x ! x̃. As explained above, it is sufficient

fixes the form of the two and three point functions,

2.3. Ward identities

metric gµ⌫ . Recalling the classical definition

Under an infinitesimal coordinate transformation x̃µ = xµ + ✏µ (x), the

hO1 (x̃1 ) . . . On (x̃n )ig̃ = hO1 (x1 ) . . . On (xn )ig , (17)

for all ✏µ (x) that decays sufficiently fast at infinity. Thus rµ T µ⌫ = 0 up to

hO1 (x1 ) . . . On (xn )ig

Consider now an infinitesimal Weyl transformation ⌦ = 1 + !, which

anomaly Z[⌦2 g] = Z[g]e SW eyl [⌦,g] .

Consider the following codimension 1 integral over the boundary of a

Exercise 2.5. Conformal symmetry fixes the three-point function of a spin

2.4. State-Operator Map

2.5. Operator Product Expansion

where denotes a number determined by conformal symmetry. For sim-

2.6. Conformal Bootstrap

• OPE associativity - Di↵erent ways of using the OPE to compute

It is sufficient to impose OPE associativity for all four-point functions of

where G ,l are conformal blocks, which encode the contribution from a

j Forprimary operators O1 , O2 , O3 transforming in non-trivial irreps of SO(d) there are

2.7. Embedding Space Formalism

Vector primary operators can also be extended to the embedding space.

Notice that this implies a redundancy: OA (P ) ! OA (P ) + PA ⇤(P ) gives

Exercise 2.10. Consider the parametrization P A = P 0 , P µ , P d+1 =

Conformal correlation functions extended to the light-cone of R1,d+1 are

and dilatation invariance on Rd

2.8. Large N Factorization

where we introduced the ’t Hooft coupling = gY2 M N and ci are other

number of diagrams that contribute with increasing powers of the coupling

Let us now consider single-trace local operators of the form O =

which is dominated by the planars diagrams (g = 0). Moreover, we see

important for the argument. It is well known that large N factorization

3. Anti-de Sitter Spacetime

Euclidean AdS spacetime is the hyperboloid

Another useful coordinate system is

3.1. Particle dynamics in AdS

The metric is given by

To understand the global structure of this spacetime it is convenient to

tanh ⇢ = tanh sin t , (59)