Simmons Duffin Notes

arXiv:1602.
07982v1 [hep-th] 25 Feb 2016
TASI Lectures on the Conformal Bootstrap
David Simmons-Duffin
School of Natural Sciences, Institute for Advanced Study
Princeton, NJ 08540,
[email protected]
These notes are from courses given at TASI and the Advanced Strings
School in summer 2015. Starting from principles of quantum field theory
and the assumption of a traceless stress tensor, we develop the basics
of conformal field theory, including conformal Ward identities, radial
quantization, reflection positivity, the operator product expansion, and
conformal blocks. We end with an introduction to numerical bootstrap
methods, focusing on the 2d and 3d Ising models.
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1. Landmarks in the Space of QFTs . . . . . . . . . . . . .
1.2. Critical Universality . . . . . . . . . . . . . . . . . . . .
1.3. The Bootstrap Philosophy . . . . . . . . . . . . . . . . .
2. QFT Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1. The Stress Tensor . . . . . . . . . . . . . . . . . . . . .
2.2. Quantization . . . . . . . . . . . . . . . . . . . . . . . .
2.3. Topological Operators and Symmetries . . . . . . . . . .
2.4. More Symmetries . . . . . . . . . . . . . . . . . . . . . .
3. Conformal Symmetry . . . . . . . . . . . . . . . . . . . . . .
3.1. Finite Conformal Transformations . . . . . . . . . . . .
3.2. The Conformal Algebra . . . . . . . . . . . . . . . . . .
4. Primaries and Descendants . . . . . . . . . . . . . . . . . . .
4.1. Poincare Representations . . . . . . . . . . . . . . . . .
4.2. Scale+Poincare Representations . . . . . . . . . . . . .
4.3. Conformal Representations . . . . . . . . . . . . . . . .
4.4. Finite Conformal Transformations . . . . . . . . . . . .
5. Conformal Correlators . . . . . . . . . . . . . . . . . . . . . .
5.1. Scalar Operators . . . . . . . . . . . . . . . . . . . . . .
5.2. Spinning Operators . . . . . . . . . . . . . . . . . . . . .
6. Radial Quantization and the State-Operator Correspondence
6.1. Operator = State . . . . . . . . . . . . . . . . . . . . .
6.2. Operator = State . . . . . . . . . . . . . . . . . . . . .
6.3. Operator State . . . . . . . . . . . . . . . . . . . .
6.4. Another View of Radial Quantization . . . . . . . . . .
7. Reflection Positivity and Unitarity Bounds . . . . . . . . . .
7.1. Reflection Positivity . . . . . . . . . . . . . . . . . . . .
7.2. Reflection Positivity on the Cylinder . . . . . . . . . . .
7.3. Unitarity Bounds . . . . . . . . . . . . . . . . . . . . . .
7.4. Only Primaries and Descendants . . . . . . . . . . . . .
8. The Operator Product Expansion . . . . . . . . . . . . . . .
8.1. Consistency with Conformal Invariance . . . . . . . . .
8.2. Computing Correlators with the OPE . . . . . . . . . .
9. Conformal Blocks . . . . . . . . . . . . . . . . . . . . . . . .
9.1. Using the OPE . . . . . . . . . . . . . . . . . . . . . . .
9.2. In Radial Quantization . . . . . . . . . . . . . . . . . .
9.3. From the Conformal Casimir . . . . . . . . . . . . . . .
9.4. Series Expansion . . . . . . . . . . . . . . . . . . . . . .
10. The Conformal Bootstrap . . . . . . . . . . . . . . . . . . . .
10.1. OPE Associativity and Crossing Symmetry . . . . . . .
10.2. Crossing Symmetry for Identical Scalars . . . . . . . . .
10.3. An Infinite Number of Primaries . . . . . . . . . . . . .
10.4. Bounds on CFT Data . . . . . . . . . . . . . . . . . . .
10.5. An Example Bound . . . . . . . . . . . . . . . . . . . .
10.6. Numerical Techniques . . . . . . . . . . . . . . . . . . .
10.7. Improving on our Hand-Computed Bound . . . . . . . .
10.8. Numerical Results in 3d . . . . . . . . . . . . . . . . . .
10.9. Open Questions . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
4
5
6
6
8
9
11
12
12
13
16
16
17
18
20
21
21
23
24
26
27
28
29
31
31
34
37
39
40
41
43
43
43
45
46
47
49
49
52
53
54
55
58
60
61
63
Appendix A. Quantization of the Lattice Ising Model . . . . . . .

Appendix B. Euclidean vs. Lorentzian and Analytic Continuation
Appendix C. Semidefinite Programming . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
67
68
69
Other Resources
This course is heavily inspired by Slava Rychkovs EPFL Lectures on Conformal Field Theory in d 3 Dimensions [1]. His notes cover similar topics,
plus additional material that we wont have time for here, including conformal invariance in perturbation theory, the embedding formalism, and
some analytical bootstrap bounds. By contrast, these lectures spend more
time on QFT basics and numerical bootstrap methods. See also lectures
by Sheer El-Showk [2] and Joshua Qualls [3].
Our discussion of symmetries and quantization is based on Polchinskis
String Theory, Vol. 1 [4]: mostly Chapter 2 on 2d CFTs and Appendix
A on path integrals. Appendix A is required reading for any high energy
theory student.
The book Conformal Field Theory by Di Francesco, Mathieu, and
Senechal [5] is also a useful reference. It starts with a discussion of CFTs
in general spacetime dimensions, but includes much more about 2d CFTs,
a topic we unfortunately neglect in this course.
1. Introduction
1.1. Landmarks in the Space of QFTs
Quantum field theories generically become scale-invariant at long distances.
Often, invariance under rescaling actually implies invariance under the
larger conformal group, which consists of transformations that locally look
like a rescaling and a rotation.a These extra symmetries are powerful tools
for organizing a theory. Because their emergence requires no special structure beyond the long distance limit, they are present in a huge variety of
physical systems.
We can think of a UV-complete QFT as a renormalization group (RG)
flow between conformal field theories (CFTs),b
CFTU V
QFT.
(1)
CFTIR
Studying CFTs lets us map out the possible endpoints of RG flows, and
thus understand the space of QFTs.
Many of the most interesting RG flows are nonperturbative. A simple
example is 4 theory in 3 dimensions, which has the Euclidean action

Z
1 2 2
1 4
1
2
3
() + m + g .
(2)
S= d x
2
2
4!
This theory is free in the UV, since m and g have mass dimension 1 and can
be ignored at high energies. The behavior of the theory in the IR depends
on the ratio g 2 /m2 . If m2 is large and positive, the IR theory is massive
and preserves the Z2 symmetry . If m2 is large and negative, the IR
theory is again massive but spontaneously breaks Z2 . For a special value of
g 2 /m2 , in between these two regimes, the IR theory becomes gapless and
is described by a nontrivial interacting CFT.c
It is hard to study this CFT with Feynman diagrams. By dimensional
analysis, naive perturbation theory leads to an expansion in gx, where x is a
a The
question of when scale invariance implies conformal invariance is an important

foundational problem in quantum field theory that is still under active study. In 2d and
4d, it has been proven that Lorentz-invariance and unitarity are sufficient conditions
[6, 7]. In 3d or d 5, the appropriate conditions are not known, but conformal invariance
appears in many examples.
b Having a CFT in the IR is generic. We do not necessarily have a CFT in the UV, but
assuming one is sometimes a useful framework.
c The precise value of g 2 /m2 that gives a CFT is scheme-dependent: it depends on how
one regulates UV divergences.
distance scale characterizing the observable were computing. At distances

x 1/g, the expansion breaks down. Instead, the best perturbative tool
we have is the -expansion, where we compute Feynman diagrams in 4
dimensions and afterwards continue 1. This works surprisingly well,
but is conceptually a little shaky.
1.2. Critical Universality

In the example above, the UV theory was a continuum QFT: the free boson.
However, IR CFTs can also arise from very different microscopic systems
[8]. An example is the 3d Ising model, which is a lattice of classical spins
si {1} with nearest-neighbor interactions. The partition function is
ZIsing =
X
{si }
exp J
si sj ,
(3)
hiji
where i, j label lattice points and hiji indicates that i and j are nearest
neighbors. We can think of this sum as a discrete path integral, where the
integration variable is the space of functions
s : Lattice {1}.
(4)
For a special value of J, this theory also becomes a nontrivial CFT at long
distances. Actually it is the same CFT as the one appearing in 4 theory!
The Ising CFT also arises in water (and other liquids) at the critical point on its phase diagram, and in uni-axial magnets at their critical
temperatures [9]. We say that 4 theory, the Ising model, water, and uniaxial magnets are IR equivalent at their critical points (figure 1), and that
they are in the same universality class. IR equivalences show up all over
high-energy and condensed-matter physics, where they are sometimes called
dualities. The ubiquity of IR equivalences is the phenomenon of critical
universality.
The above examples are Euclidean field theories. But Lorentzian CFTs
also appear in nature, describing quantum critical points. For example,
the Lorentzian O(2) model describes thin-film superconductors [10, 11],
while its Wick-rotation, the Euclidean O(2) model, describes the superfluid
transition in 4 He [12]. Amazingly, the critical exponents of these theories
agree, allowing us to see Wick rotation in nature!
Fig. 1. Many microscopic theories can flow to the same IR CFT. We say that the
theories are IR equivalent, or IR dual. The UV can even be something exotic like a stack
of M5-branes in M-theory.
1.3. The Bootstrap Philosophy

Critical universality means we can study the 3d Ising CFT by doing computations in any of its microscopic realizations. This is a powerful tool.
For example, we can model critical water by simulating classical spins on a
computer, without ever worrying about 1023 bouncing water molecules! For
analytical results, we can use the -expansion. But all of these approaches
fail to exploit the emergent symmetries of the IR theory.
The conformal bootstrap philosophy is to:
0.
1.
2.
3.
focus on the CFT itself and not a specific microscopic realization,

determine the full consequences of symmetries,
impose consistency conditions,
combine (1) and (2) to constrain or even solve the theory.
This strategy was first articulated by Ferrara, Gatto, and Grillo [13]
and Polyakov [14] in the 70s. Importantly, it only uses nonperturbative
structures, and thus has a hope of working for strongly-coupled theories. Its
effectiveness for studying the 3d Ising model will become clear during this
course. In addition, sometimes bootstrapping is the only known strategy
for understanding the full dynamics of a theory. An example is the 6d
N = (2, 0) supersymmetric CFT describing the IR limit of a stack of M5
branes in M-theory. This theory has no known Lagrangian description, but
is amenable to bootstrap analysis [15].d
A beautiful and ambitious goal of the bootstrap program is to eventud At
large central charge, this theory is solved by the AdS/CFT correspondence [16].
Supersymmetry also lets one compute a variety of protected quantities (at any central
charge). However, the bootstrap is currently the only known tool for studying nonprotected quantities at small central charge.
ally provide a fully nonperturbative formulation of quantum field theory,

removing the need for a Lagrangian. We are not there yet, but you can
help!
2. QFT Basics
The first step of the conformal bootstrap is to determine the full consequences of symmetries. In this section, we quickly review symmetries in
quantum field theory, phrasing the discussion in language that will be useful later. We work in Euclidean signature throughout.
2.1. The Stress Tensor
A local quantum field theory has a conserved stress tensor,
T (x) = 0
(operator equation).
(5)
This holds as an operator equation, meaning it is true away from other

operator insertions. In the presence of other operators, (5) gets modified
to include contact terms on the right-hand side,
hT (x)O1 (x1 ) . . . On (xn )i =
X
i
(x xi )i hO1 (x1 ) . . . On (xn )i.

(6)
Exercise 2.1. Consider a QFT coupled to a background metric g. For

concreteness, suppose correlators are given by the path integral
Z
hO1 (x1 ) . . . On (xn )ig = D O1 (x1 ) . . . On (xn ) eS[g,] .
(7)
A stress tensor insertion is the response to a small metric perturbation,e
2
hO1 (x1 ) . . . On (xn )ig .
hT (x)O1 (x1 ) . . . On (xn )ig =
g g (x)
(8)
Derive (6) by demanding that S[g, ] be diffeomorphism invariant near flat

space. Find how to modify (6) when the Oi have spin.
e This
definition of the stress tensor works in a continuum field theory. If the UV is a

lattice model, we must assume (or prove) that a stress tensor emerges in the IR.
Consider the integral of T over a closed surface ,f g

Z
P ()
dS T (x).
(9)
The Ward identity (6) implies that a correlator of P () with other operators is unchanged as we move , as long as doesnt cross any operator
insertions (figure 2). We say that P () is a topological surface operator.
Fig. 2. A surface supporting the operator P () can be freely deformed 0

without changing the correlation function, as long as it doesnt cross any operator insertions.
Let = B be the boundary of a ball B containing x and no other

insertions. Integrating (6) over B gives
hP ()O(x) . . .i = hO(x) . . .i.
(10)
In other words, surrounding O(x) with the topological surface operator P

is equivalent to taking a derivative (figure 3).
Fig. 3.
f The
Surrounding O(x) with P gives a derivative.
word surface usually refers to a 2-manifold, but we will abuse terminology and
use it to refer to a codimension-1 manifold.
g Our definition of P differs from the usual one by a factor of i. This convention is
much nicer for Euclidean field theories, but it has the effect of modifying some familiar
formulae, and also changing the properties of symmetry generators under Hermitian
conjugation. More on this in section 7.1.
In quantum field theory, having a topological codimension-1 operator is

the same as having a symmetry.h This may be unfamiliar language, so to
connect to something more familiar, let us revisit the relation between the
path integral and Hamiltonian formalisms.
2.2. Quantization
A single path integral can be interpreted in terms of different time evolutions in different Hilbert spaces. For example, in a rotationally-invariant
Euclidean theory on Rd , we can choose any direction as time and think
of states living on slices orthogonal to the time direction (figure 4). We call
each interpretation a quantization of the theory.
Fig. 4.
In a rotationally invariant Euclidean theory, we can choose any direction as
time. States live on slices orthogonal to the time direction.
To specify a quantization, we foliate spacetime by slices related by an

isometry t . A slice has an associated Hilbert space of states. A correlation
function hO1 (x1 ) On (x2 )i gets interpreted as a time-ordered expectation
value
b1 (t1 , x1 ) O
bn (tn , xn )}|0i.
hO1 (x1 ) On (xn )i = h0|T {O
(11)
Here, the time ordering T {. . . } is with respect to our foliation, |0i is the
bi (x) : H H
vacuum in the Hilbert space H living on a spatial slice,i and O
are quantum operators corresponding to the path integral insertions Oi (x).
A different quantization of the theory would give a completely different
Hilbert space H0 , a completely different time-ordering, and completely difb0 . However, some equations satisfied by these
ferent quantum operators O
i
h Topological
operators with support on other types of manifolds give generalized symmetries [17].
i Other choices of initial and final state correspond to different boundary conditions for
the path integral.
new operators on this new Hilbert space would be unchanged. For example, if we arrange the operators as shown on the right-hand side of (11), we
always get the correlator on the left-hand side.
We demonstrate these ideas explicitly in appendix A, where we show
how to (discretely) quantize the lattice Ising model in different ways.
2.3. Topological Operators and Symmetries
Fig. 5.
The charge P (t ) can be moved from one time to another t t0 without
changing the correlation function.
Let t be a spatial slice at time t and consider the operator P (t ).

Because P () is topological, we are free to move it forward or backward in
time from one spatial slice to another as long as it doesnt cross any operator
insertions (figure 5). In fact, we often neglect to specify the surface t and
just write P (though we should keep in mind where the surface lives with
respect to other operators). We call P momentum, and the fact that its
topological is the path integral version of the statement that momentum is
conserved.
Let us understand what happens when we move P past an operator
insertion. Consider a local operator O(x) at time t. If 1 , 2 are spatial
surfaces at times t1 < t < t2 , then when we quantize our theory, the difference P (2 ) P (1 ) becomes a commutator because of time ordering,
b
h(P (2 ) P (1 ))O(x) . . .i = h0|T {[Pb , O(x)]
. . . }|0i.
(12)
(We assume that the other insertions . . . are not between times t1 and
t2 .) Because P is topological, we can deform 2 1 to a sphere S
surrounding O(x), as in figure 6. Then using the Ward identity (10), we
10
find
b
h0|T {[Pb , O(x)]
. . . }|0i = h(P (2 ) P (1 ))O(x) . . .i
= hP (S)O(x) . . .i
= hO(x) . . .i
b
= h0|T {O(x)
. . . }|0i,
(13)
in other words,
b
b
[Pb , O(x)]
= O(x).
(14)
Fig. 6.
For any charge Q(), we can deform Q(2 ) Q(1 ) = Q(2 1 ) to an
insertion of Q(S). Here, arrows indicate the orientation of the surface.
Figure 6 makes it clear that this result is independent of how we quantize

our theory, since we always obtain a sphere surrounding O(x) no matter
which direction we choose as time. Thus, we often write
[P , O(x)] = O(x),
(15)
without specifying a quantization. In fact, from now on, we will no longer

distinguish between path integral insertions O(x) and quantum operators
b
O(x).
The expression [Q, O(x)] can be interpreted as either an actual
b O(x)]
b
commutator [Q,
in any quantization of the theory, or in path-integral
language as surrounding O(x) with a topological surface operator Q(S).
Figure 6 also explains why the commutator [Q, O(x)] of a charge Q with
a local operator O(x) is local, even though Q is nonlocal (it is the integral
of a current). The reason is that the support of Q can be deformed to an
arbitrarily small sphere S around x, so that the insertion Q(S)O(x) only
affects the path integral in an infinitesimal neighborhood of x. In general,
the way local operators transform under symmetry is always insensitive to
IR details like spontaneous symmetry breaking or finite temperature. This
is because commutators with conserved charges can be computed at short
distances.
11
Equation (15) integrates to

O(x) = exP O(0)exP .
(16)
This statement is also true in any quantization of the theory. In path

integral language, exP () is another type of topological surface operator. When we surround O(0) with exP (), it becomes conjugation
b b
bx
P
exP ()O(0) eP x O(0)e
in any quantization.
Consider the time-ordered correlator (11) with tn > > t1 . Using

(16), it becomes
hO1 (x1 ) On (xn )i
0
= h0|etn P On (0, xn )etn P et1 P O1 (0, x1 )et1 P |0i

0
= h0|On (0, xn )e(tn tn1 )P e(t2 t1 )P O1 (0, x1 )|0i.
(17)
In other words, the path integral between spatial slices separated by time
0
t computes the time evolution operator U (t) = etP . In unitary theories
(defined in more detail in section 7.1), P 0 has a positive real spectrum, so
U (t) causes damping at large time separations.
2.4. More Symmetries
Given a conserved current J = 0 (operator equation), we can always define a topological surface operator by integration.j For P , the corresponding currents are T (x). More generally, given a vector field = (x) ,
the charge
Z
Q () =
dS (x)T (x)
(18)
will be conserved whenever

0 = ( T )
= T + T
1
= ( + )T ,
2
(19)
or
+ = 0.
j It
(20)
is an interesting question whether the converse is true. When a theory has a Lagrangian description, the Noether procedure gives a conserved current for any continuous
symmetry (that is manifest in the Lagrangian). Proving Noethers theorem without a
Lagrangian is an open problem.
12
This is the Killing equation. In flat space, it has solutions

p =
(translations),
m = x x
(rotations).
(21)
The corresponding charges are momentum P = Qp and angular momentum M = Qm .

3. Conformal Symmetry
In a conformal theory, the stress tensor satisfies an additional condition: it
is traceless,
T (x) = 0
(operator equation).
(22)
This is equivalent to the statement that the theory is insensitive to positiondependent rescalings of the metric g = (x)g near flat space.k When
the stress tensor is traceless, we can relax the requirement (20) further to
+ = c(x) ,
(23)
where c(x) is a scalar function. Contracting both sides with gives

c(x) = d2 (x). Equation (23) is the conformal Killing equation. It has
two additional types of solutions in Rd ,
d = x
(dilatations),
2
k = 2x (x ) x
(special conformal transformations).
(24)
The corresponding symmetry charges are D = Qd and K = Qk .l

3.1. Finite Conformal Transformations
Before discussing the charges P , M , D, K, let us take a moment to understand the geometrical meaning of the conformal Killing vectors (21) and
(24). Consider an infinitesimal transformation x x0 = x + (x). If
satisfies the conformal Killing equation, then

x0
1
1
= + =
1 + ( )
+ ( ) . (25)
x
d
2
k In
curved space, there can by Weyl anomalies.

above solutions are present in any spacetime dimension. In two dimensions, there
exists an infinite set of additional solutions to the conformal Killing equation, leading to
an infinite set of additional conserved quantities [18]. This is an extremely interesting
subject that we unfortunately wont have time for in this course.
l The
13
The right-hand side is an infinitesimal rescaling times an infinitesimal rotation. Exponentiating gives a coordinate transformation x x0 such that
x0
= (x)R (x),
x
RT R = Idd ,
(26)
where (x) and R (x) are finite position-dependent rescalings and rotations. Equivalently, the transformation x x0 rescales the metric by a
scalar factor,
x0 x0
= (x)2 .
x x
(27)
Such transformations are called conformal. They comprise the conformal

group, a finite-dimensional subgroup of the diffeomorphism group of Rd .
The exponentials of p and m are translations and rotations. Exponentiating d gives a scale transformation x x. We can understand the
exponential of k by first considering an inversion
I : x
x
.
x2
(28)
I is a conformal transformation, but it is not continuously connected to

the identity, so it cant be obtained by exponentiating a conformal Killing
vector. This means that a CFT need not have I as a symmetry.
Exercise 3.1. Show that I is continuously connected to a reflection x0
x0 . Conclude that a CFT is invariant under I if and only if it is invariant
under reflections.
Exercise 3.2. Show that k = Ip I. Conclude that eak implements the
transformation
x x0 (x) =
x a x2
.
1 2(a x) + a2 x2
(29)
We can think of k as a translation that moves infinity and fixes the

origin in the same sense that the usual translations move the origin and
fix infinity, see figure 7.
3.2. The Conformal Algebra
The charges Q give a representation of the conformal algebra
[Q1 , Q2 ] = Q[1 ,2 ] ,
(30)
14
Fig. 7.
k is analogous to p , with the origin and the point at infinity swapped by an
inversion.
where [1 , 2 ] is a commutator of vector fields.m This is not obvious and

deserves proof. In fact, it is not true in 2-dimensional CFTs, where the
algebra of charges is a central extension of the the algebra of conformal
killing vectors.
Exercise 3.3. Show that in d 3,
[Q , T ] = T + ( )T T + T .
(31)
Argue as follows. Assume that only the stress tensor appears on the righthand side.n Using linearity in , dimensional analysis, and the conformal
Killing equation, show that (31) contains all terms that could possibly appear.o Fix the relative coefficients using conservation, tracelessness, and
symmetry under . Fix the overall coefficient by matching with (15).
Exercise 3.4. Using (31), prove the commutation relation (30).
Exercise 3.5. When d = 2, its possible to add an extra term in (31) proportional to the unit operator that is consistent with dimensional analysis,
conservation, and tracelessness. Find this term (up to an overall coefficient),p and show how it modifies the commutation relations (30). This is
the Virasoro algebra!
m The minus sign in (30) comes from the fact that when charges Q are represented
i
by differential operators Di , repeated action reverses the order [Q1 , [Q2 , O]] = D2 D1 O.
Alternatively, we could have introduced an extra minus sign in the Qs so that [Q, O] =
D and then Q, D would have the same commutation relations.
n Bonus exercise: can other operators appear?
o The terms on the right-hand side are local in because we can evaluate [Q , T (x)] in

an arbitrarily small neighborhood of x. Assuming the singularity as two T s coincide
is bounded, we can then replace by its Taylor expansion around x.
p The coefficient can be fixed by comparing with the OPE, see e.g. [4]. It is proportional
to the central charge c.
15
As usual, (30) is true in any quantization of the theory. In path integral

language, it tells us how to move the topological surface operators Q ()
through each other.
Exercise 3.6. Show that
[M , P ] = P P ,
[M , K ] = K K ,
[M , M ] = M M + M M ,
(32)
(33)
(34)
[D, P ] = P ,
(35)
[D, K ] = K ,
(36)
[K , P ] = 2 D 2M ,
(37)
and all other commutators vanish.

The first three commutation relations say that M generates the algebra
of Euclidean rotations SO(d) and that P , K transform as vectors. The
last three are more interesting. Equations (35) and (36) say that P and
K can be thought of as raising and lowering operators for D. We will
return to this idea shortly.
Exercise 3.7. Define the generators
L = M ,
L1,0 = D,
1
L0, = (P + K ),
2
1
L1, = (P K ),
2
(38)
where Lab = Lba and a, b {1, 0, 1, . . . , d}. Show that Lab satisfy the
commutation relations of SO(d + 1, 1).
The fact that the conformal algebra is SO(d + 1, 1) suggests that it might
be good to think about its action in terms of Rd+1,1 instead of Rd . This is
the idea behind the embedding space formalism [1924], which provides a
simple and powerful way to understand the constraints of conformal invariance. We will be more pedestrian in this course, but I recommend reading
about the embedding space formalism in the lecture notes by Penedones [25]
or Rychkov [1].
16
4. Primaries and Descendants

Now that we have our conserved charges, we can classify operators into
representations of those charges. We do this in steps. First we classify operators into Poincare representations, then scale+Poincare representations,
and finally conformal representations.
4.1. Poincare Representations
In a rotationally-invariant QFT, local operators at the origin transform in
irreducible representations of the rotation group SO(d),
[M , Oa (0)] = (S )b a Ob (0),
(39)
where S are matrices satisfying the same algebra as M , and a, b are

indices for the SO(d) representation of O.qr We often suppress spin indices
and write the right-hand side as simply S O(0). The action (39), together
with the commutation relations of the Poincare group, determines how
rotations act away from the origin.
Fig. 8. The shorthand notation QO stands for surrounding O with a surface operator
Q(). Equivalently, it stands for [Q, O] in any quantization of the theory.
To see this, it is convenient to adopt shorthand notation where commutators of charges with local operators are implicit, [Q, O] QO, see
figure 8. This notation is valid because of the Jacobi identity (more formally, the fact that adjoint action gives a representation of a Lie algebra).
In path integral language, Qn Q1 O(x) means surrounding O(x) with
topological surface operators where Qn is the outermost surface and Q1 is
q The funny index contractions in (39) ensure that M
and S have the same commutation relations (exercise!).
r Because our commutation relations (34) for SO(d) differ from the usual conventions by
a factor of i, the generators S will be anti-hermitian, S = S.
17
the innermost. The conformal commutation relations tell us how to re-order

these surfaces.
Acting with a rotation on O(x), we have
M O(x) = M exP O(0)
= exP (exP M exP )O(0)
= exP (x P + x P + M )O(0)
= (x x + S )exP O(0)
= (m + S )O(x).
(40)
In the third line, weve used the Poincare algebra and the Hausdorff formula
eA BeA = e[A,] B = B + [A, B] +
1
[A, [A, B]] + . . . .
2!
(41)
4.2. Scale+Poincare Representations

In a scale-invariant theory, its also natural to diagonalize the dilatation
operator acting on operators at the origin,
[D, O(0)] = O(0).
(42)
The eigenvalue is the dimension of O.

Exercise 4.1. Mimic the computation (40) to derive the action of dilatation on O(x) away from the origin,
[D, O(x)] = (x + )O(x).
(43)
Equation (43) is constraining enough to fix two-point functions of scalars

up to a constant. Firstly, by rotation and translation invariance, we must
have
hO1 (x)O2 (y)i = f (|x y|),
(44)
for some function f .

In a scale-invariant theory with scale-invariant boundary conditions, the
simultaneous action of D on all operators in a correlator must vanish, as
illustrated in figure 9. Moving D to the boundary gives zero.s On the
other hand, shrinking D to the interior gives the sum of its actions on the
individual operators. By the Ward identity (43), this is
s It
0 = (x + 1 + y + 2 ) f (|x y|).
(45)
is also interesting to consider non-scale-invariant boundary conditions. These can be

interpreted as having a nontrivial operator at .
18
We could alternatively derive (45) by working in some quantization, where

it follows from trivial algebra and the fact that D|0i = 0,
0 = h0|[D, O1 (x)O2 (y)]|0i
= h0|[D, O1 (x)]O2 (y) + O1 (x)[D, O2 (y)]|0i
= (x + 1 + y + 2 ) h0|O1 (x)O2 (y)|0i.
(46)
Either way, the solution is

f (|x y|) =
C
.
|x y|1 +2
(47)
If we had an operator with negative scaling dimension, then its correlators would grow with distance, violating cluster decomposition. This
is unphysical, so we expect dimensions to be positive. Shortly, we will
prove this fact for unitary conformal theories (and derive even stronger
constraints on ).
Fig. 9.
The Ward identity for scale invariance of a two-point function.
4.3. Conformal Representations

Note that K is a lowering operator for dimension,
DK O(0) = ([D, K ] + K D)O(0)
= ( 1)K O(0).
(48)
(Again, were using shorthand notation [Q, O] QO.) Thus, given

an operator O(0), we can repeatedly act with K to obtain operators
K1 . . . Kn O(0) with arbitrarily low dimension. Because dimensions are
bounded from below in physically sensible theories, this process must eventually terminate. That is, there must exist operators such that
[K , O(0)] = 0
(primary operator).
(49)
19
Such operators are called primary. Given a primary, we can construct

operators of higher dimension, called descendants, by acting with momentum generators, which act like raising operators for dimension,
O(0) P1 Pn O(0)
(descendant operators)
+ n.
(50)
For example, O(x) = exP O(0) is an (infinite) linear combination of descendant operators. The conditions (39, 42, 49) are enough to determine how
K acts on any descendant using the conformal algebra. For example,
Exercise 4.2. Let O(0) be a primary operator with rotation representation
matrices S and dimension . Using the conformal algebra, show
[K , O(x)] = (k + 2x 2x S )O(x),
(51)
where k is the conformal Killing vector defined in (24).

To summarize, a primary operator satisfies
[D, O(0)] = O(0)
[M , O(0)] = S O(0)
[K , O(0)] = 0.
(52)
From these conditions, we can construct a representation of the conformal

algebra out of O(0) and its descendants,
operator dimension
..
.
P1 P2 O(0) + 2
P1 O(0)
+1
O(0)
.
(53)
The action of conformal generators on each state follows from the conformal
algebra. This should remind you of the construction of irreducible representations of SU(2) starting from a highest-weight state. In this case, our
primary is a lowest-weight state of D, but the representation is built in an
analogous way.t It turns out that any local operator in a unitary CFT is
t Generically, the representation (53) is an induced representation IndG (R ), where H
H
H
is the subgroup of the conformal group generated by D, M , K (called the isotropy
subgroup), RH is the finite-dimensional representation of H defined by (52), and G is the
20
a linear combination of primaries and descendants. We will prove this in

section 7.4.
Exercise 4.3. Show that (15), (40), (43), and (51) can be summarized as

1
[Q , O(x)] = + ( ) ( )S O(x).
(54)
d
2
Exercise 4.4. Deduce that T is primary by comparing (54) with (31).
4.4. Finite Conformal Transformations
An exponentiated charge U = eQ implements a finite conformal transformation. Denote the corresponding diffeomorphism e by x 7 x0 (x). By
comparing with (25) and (26), we find that (54) exponentiates to
U Oa (x)U 1 = (x0 ) D(R(x0 ))b a Ob (x0 ),
(55)
where as before
x0
= (x0 )R (x0 ),
x
R (x0 ) SO(d).
(56)
Here, D(R)b a is a matrix implementing the action of R in the SO(d) representation of O, for example
D(R) = 1
D(R) = R
(scalar representation),
(vector representation),
(57)
and so on.
We could have started the whole course by taking (55) as the definition
of a primary operator. But the connection to the underlying conformal
algebra will be crucial in what follows, so we have chosen to derive it.
Exercise 4.5. Show that the transformation (55) composes correctly to give
a representation of the conformal group. That is, show
Ug1 Ug2 Oa (x)Ug1
Ug1
= Ug1 g2 Oa (x)Ug1
2
1
1 g2
(58)
where x 7 gi (x) are conformal transformations, g1 g2 denotes composition

x 7 g1 (g2 (x)), and Ug is the unitary operator associated to g.
full conformal group. It is also called a parabolic Verma module. Sometimes the operator
O satisfies shortening conditions where a linear combination of descendants vanishes.
(A conserved current is an example.) In this case, the Verma module is reducible and
the actual conformal multiplet of O is one of the irreducible components.
21
5. Conformal Correlators
5.1. Scalar Operators
We have already seen that scale invariance fixes two-point functions of
scalars up to a constant
hO1 (x1 )O2 (x2 )i =
C
|x1 x2 |1 +2
(SFT).
(59)
For primary scalars in a CFT, the correlators must satisfy a stronger

Ward identity,
hO1 (x1 ) . . . On (xn )i = h(U O1 (x1 )U 1 ) (U On (xn )U 1 )i
= (x01 )1 (x0n )n hO1 (x0n ) On (x0n )i. (60)
Let us check whether this holds for (59).

Exercise 5.1. Show that for a conformal transformation,
(x y)2 =
(x0 y 0 )2
.
(x0 )(y 0 )
(61)
Hint: This is obviously true for translations, rotations, and scale transformations. It suffices to check it for inversions I : x xx2 (why?).
Using (61), we find
1 +2
1 +2
C
C
= (x01 ) 2 (x02 ) 2
.
+
1
2
|x1 x2 |
|x01 x02 |1 +2
(62)
Consistency with (60) then requires 1 = 2 or C = 0. In other words,

hO1 (x1 )O2 (x2 )i =
C1 2
1
x2
12
(CFT, primary operators),
(63)
where x12 x1 x2 .
Exercise 5.2. Recover the same result using the Ward identity for K
h[K , O1 (x1 )]O2 (x2 )i + hO1 (x1 )[K , O2 (x2 )]i = 0.
(64)
Conformal invariance is also powerful enough to fix a three-point function of primary scalars, up to an overall coefficient. Using (61), its easy to
check that the famous formula [8]
hO1 (x1 )O2 (x2 )O3 (x3 )i =
f123
,
1 +2 3 2 +3 1 3 +1 2
x
x
x31
12
23
with f123 constant, satisfies the Ward identity (60).
(65)
22
With four points, there are nontrivial conformally invariant combinations of the points called conformal cross-ratios,
u=
x212 x234
,
x213 x224
v=
x223 x214
.
x213 x224
(66)
The reason that there are exactly two independent cross-ratios can be understood as follows.
Using
Using
Using
Using
special conformal transformations, we can move x4 to infinity.

translations, we can move x1 to zero.
rotations and dilatations, we can move x3 to (1, 0, . . . , 0).
rotations that fix x3 , we can move x2 to (x, y, 0, . . . , 0).
x1
x2
x3
1
x4
Fig. 10.
Using conformal transformations, we can place four points on a plane in the
configuration shown above (figure from [26]).
This procedure leaves exactly two undetermined quantities x, y, giving

two independent conformal invariants. Evaluating u and v for this special
configuration of points (figure 10) gives
u = z z,
v = (1 z)(1 z),
(67)
where z x + iy.
Four-point functions can depend nontrivially on the cross-ratios. For a
scalar with dimension , the formula
h(x1 )(x2 )(x3 )(x4 )i =
g(u, v)
2
2
x12 x34
(68)
satisfies the Ward identity (60) for any function g(u, v).
Exercise 5.3. Generalize (68) to the case of non-identical scalars i (x)
with dimensions i .
23
The left-hand side of (68) is manifestly invariant under permutations of

the points xi . This leads to consistency conditions on g(u, v),
g(u, v) = g(u/v, 1/v)
(from swapping 1 2 or 3 4),
(69)
u
g(v, u)
(from swapping 1 3 or 2 4).
(70)
g(u, v) =
v
All other permutations can be generated from the ones above. We will
see shortly that g(u, v) is actually determined in terms of the dimensions
i and three-point coefficients fijk of the theory. Equation (69) will be
satisfied for trivial reasons. However (70) will lead to powerful constraints
on the i , fijk .
5.2. Spinning Operators
The story is similar for operators with spin. For brevity, we give the answers
without doing any computations. The embedding space formalism provides
a transparent and practical way to derive all of these results [24], so its not
worth dwelling on them here.
Two-point functions of spinning operators are fixed by conformal invariance. They are nonzero only if the operators have identical dimensions and
spins. For example, a two-point function of spin-1 operators with dimension
is given by
I (x y)
,
(71)
(x y)2
x x
I (x) 2 2 ,
(72)
x
where CJ is a constant. Note that I (x) is the orthogonal matrix associ0
ated with an inversion, x

x = (x)I (x).
hJ (x)J (y)i = CJ
Exercise 5.4. Check that (71) is consistent with conformal symmetry.

Hint: it is enough to check inversions.
Two-point functions of operators in more general spin representations
can be constructed from the above. For spin-` traceless symmetric tensors,
(1

I 1 (x) I ` ) ` (x)
1 ...`
(x)J1 ...` (0)i = CJ
hJ
traces , (73)
x2
where we can symmetrize either the s or s (or both). Subtracting traces
means adding terms proportional to i j and i j so that the result is
separately traceless in the indices and the indices (not necessarily under
- contractions).
24
It is sometimes conventional to normalize J so that CJ = 1 in (71), (73).

An exception is if J already has a natural normalization. For example, the
normalization of the stress tensor is fixed by demanding that T satisfy
the appropriate Ward identities. In this case, CT is physically meaningful.
Three-point functions are fixed up to a finite number of coefficients.
For example, a three-point function of scalars 1 , 2 and a spin-` operator
J1 ...` is determined up to a single coefficient f1 2 J ,
f1 2 J (Z 1 Z ` traces)
,
1 +2 3 +` 2 +3 1 ` 3 +1 2 `
x31
x23
x12
x
x23
Z 13
.
(74)
x213
x223
h1 (x1 )2 (x2 )J 1 ...` (x3 )i =
When multiple operators have spin, there can be more than one linearly
independent structure consistent with conformal invariance.
Formula (74) applies when J is the stress tensor. In that case, the
coefficient f1 2 T is fixed by demanding that integrals of T give the correct action of the conformal charges Q (see the exercise in Joao Penedones
notes [25]). The result is
f1 2 T =
d1 1
C12 ,
d 1 Sd
(75)
where Sd is the volume of the unit sphere S d1 and C12 is the coefficient in
the two-point function h1 (x)2 (0)i = C12 x21 (note C12 vanishes unless
1 = 2 ). The coefficient f1 2 J is fixed by Ward identities whenever J is
a conserved current.
6. Radial Quantization and the State-Operator Correspondence
So far, weve written lots of commutation relations, and carefully pointed
out that they are true in any quantization of the theory. Now well really put
that idea to use. In general, we should to choose quantizations that respect
symmetries. In a scale-invariant theory, its natural to foliate spacetime
with spheres around the origin and consider evolving states from smaller
spheres to larger spheres using the dilatation operator (figure 11). This
is called radial quantization. The sphere S d1 has an associated Hilbert
space H. We can act on H by inserting operators on the surface of the
sphere. For example, to act with a symmetry generator Q, we insert the
surface operator Q(S d1 ) into the path integral (figure 12).
25
Fig. 11. In radial quantization, states live on spheres, and we evolve from one state to
another with the dilatation operator.
Fig. 12. We act with a charge in radial quantization by inserting Q(S d1 ) just outside
the sphere on which the state is defined.
In radial quantization, a correlation function gets interpreted as a radially ordered product,

hO1 (x1 ) On (xn )i = h0|R{O1 (x1 ) On (xn )}|0i
(|xn | |xn1 ) (|x2 | |x1 |)h0|O(xn ) O(x1 )|0i

+permutations.
(76)
Of course, we can perform radial quantization around different points. The

same correlation function then gets interpreted as an expectation value
of differently ordered operators acting on different states in different (but
isomorphic) Hilbert spaces (figure 13). This is completely analogous to
changing reference frames in Lorentz invariant theories. The radial ordering prescription is consistent because operators at the same radius but different angles on the sphere commute, just as spacelike-separated operators
commute in the usual quantization.
26
Fig. 13. When we perform radial quantization around different points, the same correlator gets interpreted as a product of operators with different orderings.
6.1. Operator = State

The simplest way to prepare a state in radial quantization is to perform
the path integral over the interior B of the sphere, with no operator insertions inside B. This gives the vacuum state |0i on the boundary B
(figure 14). Its easy to see that |0i is invariant under all symmetries because a topological surface on the boundary of B can be shrunk to zero
inside B (figure 15).
Fig. 14. The vacuum in radial quantization is given by the path integral over the interior
of the sphere, with no operator insertions.
Fig. 15.
The vacuum is automatically invariant under all symmetries.
To be explicit, suppose our CFT is given by the path integral over a

scalar field . The Hilbert space in radial quantization is spanned by field
eigenstates |b i, where b (n) is a field configuration on the sphere n B.
27
The subscript b indicates that b is defined only on the boundary B

and not in the interior. A general state is a linear combination of field
eigenstates
Z
|i Db |b ihb |i.
(77)
R
Here, Db represents a d 1-dimensional path integral over fields on B.
For the vacuum, the coefficients hb |0i are given by the path integral
over the interior with boundary conditions b and no operator insertions,
Z
hb |0i =
D(r, n)eS[] .
(78)
(1,n)=b (n)
r1
A more exciting possibility is to insert an operator O(x) inside B and

then perform the path integral,
Z
hb |O(x)|0i =
D(r, n)O(x)eS[] .
(79)
(1,n)=b (n)
r1
This defines a state called O(x)|0i, see figure 16. By inserting different
operators inside B, we can prepare a variety of states on the boundary B.
In this language, |0i is prepared by inserting the unit operator.
Fig. 16. The state O(x)|0i is given by inserting O(x) inside the sphere and performing
the path integral over the interior.
6.2. Operator = State

This construction also works backwards. Let |Oi i be eigenstates of the
dilatation operator
D|Oi i = i |Oi i.
(80)
The |Oi i can themselves be used as operators: we cut spherical holes Bi

out of the path integral centered around positions xi and glue in the states
|Oi i at the boundary of the holes, as in figure 17. This gives a quantity
28
Fig. 17. A correlator of states is defined by cutting holes out of the path integral and
gluing states into the holes.
that behaves exactly like a correlator of local operators. In the scalar field
example, the gluing procedure gives
Z Y
Z
hO1 (x1 ) On (xn )i =
Dbi hbi |Oi i
D(x) eS , (81)
i =bi
xB
/ i
where the path integral D(x) is performed over the region outside the balls
Bi , and the integrals Dbi are over field configurations on the boundaries
Bi . Here, i denotes the restriction of the bulk field (x) to the i-th
boundary Bi .
This construction only works when the xi are far enough apart that the
balls Bi dont overlap. If theyre too close together, we can use
P
hO(x1 ) O(xn )i =
hO1 (x1 ) On (xn )i,
(82)
with sufficiently large to define the correlator. Since the xi can now be
arbitrarily close together, we have defined local operators.u
6.3. Operator State
So far Ive been vague about what I mean by a local operator. But now,
we can give a rigorous definition: we will simply define a local operator to
be an eigenstate of D in radial quantization.v With this definition, the two
uA
more careful construction of the state = operator map that doesnt require this
rescaling trick is given in Polchinski [4] volume 1, chapter 2.
v The dilatation operator is diagonalizable in unitary (reflection positive) CFTs. However, there exist interesting non-unitary theories where D has a nontrivial Jordan block
29
constructions above are inverse to each other, with the identification

O(0)
O(0)|0i |Oi.
(83)
This is the state-operator correspondence.

It is straightforward to see how the conformal group acts on states in
radial quantization. A primary operator creates a state that is killed by
K and transforms in a finite-dimensional representation of D and M ,
[K , O(0)] = 0
[D, O(0)] = O(0)
[M , O(0)] = S O(0)
K |Oi = 0,
(84)
D|Oi = |Oi,
(85)
M |Oi = S |Oi.
(86)
This follows by acting on |0i with the operator equations above and using
the fact that |0i is killed by K, D, and M .
A conformal multiplet in radial quantization is given by acting with
momentum generators on a primary state
|Oi, P |Oi, P P |Oi, . . .
(conformal multiplet).
(87)
This is equivalent to acting with derivatives of O(x) at the origin, for example
O(x)|x=0 |0i = [P , O(0)]|0i = P |Oi.
(88)
The operator O(x) creates an infinite linear combination of descendants,

O(x)|0i = exP O(0)exP |0i = exP |Oi =
X
1
(x P )n |Oi.
n!
n=0
(89)
As with the classification of operators, the action of the conformal algebra on a multiplet in radial quantization is determined by the commutation relations of the algebra. In fact the required computations look exactly
identical to the computations we did to determine the action of conformal
generators on operators (40, 43, 51). This is because by surrounding operators with charges supported on spheres, we were secretly doing radial
quantization all along!
6.4. Another View of Radial Quantization
To study a conformal Killing vector , it is often helpful to perform a Weyl
rescaling of the metric g (x)2 g so that becomes a regular Killing
decomposition. In these cases, we define a local operator as a finite-dimensional representation of D.
30
vector, i.e. an isometry. We can turn a dilatation into an isometry by

performing a Weyl rescaling from Rd to the cylinder R S d1 ,
ds2Rd = dr2 + r2 ds2S d1

2
dr
2
+
ds
= r2
d1
S
r2
= e2 (d 2 + ds2S d1 ) = e2 ds2RS d1 ,
(90)
where r = e .
Dilatations r r become shifts of radial time + log . Radial
quantization in flat space is equivalent to the usual quantization on the
cylinder. States live on spheres and time evolution is generated by acting
with eD . While the development of radial quantization in the previous
sections relied only on scale invariance, the cylinder picture relies on conformal invariance because we have performed a nontrivial Weyl rescaling.
Let us build a more detailed dictionary between the two pictures. Under
a Weyl rescaling, correlation functions of local operators transform asw
!
Y
hO1 (x1 ) On (xn )i2 g
hO1 (x1 ) On (xn )ig
i
. (91)
=
(xi )
h1ig
h1i2 g
i
This is a nontrivial claim if we implement the Ising model in flat space,
compute expectation values and take the continuum limit, its not obvious
that the answer should be simply related to the same lattice theory on the
cylinder.x In general it isnt, but at the critical value of the coupling when
the theory becomes conformal, tracelessness of the stress tensor implies
insensitivity to Weyl rescalings, and the answers become related.
Exercise 6.1. By integrating by parts in (43), show that
T (x)O(y) = (x y)O(y).
(92)
An insertion of T is the response of the theory to an infinitesimal Weyl

transformation g e2 g. Derive (91) by exponentiation.y
w In
even dimensions, the partition function itself can transform with a Weyl anomaly
h1ig = h1i2 g eSWeyl [g] . This will not be important for our discussion, so we have divided
through by the partition function.
x Comparing the flat and cylindrical Ising models is relatively easy in 2d, but harder in
3d since S 2 is curved. See [27] for a recent attempt.
y We cheated here by only deriving (92) in flat space. In curved space there is an additional contribution to T coming from the Weyl anomaly. This factor cancels in (91).
There could also be modifications to the contact term (92). However, in a conformally
flat metric, we can simply define the curved space operator O(x) so that it satisfies (92).
For instance, we may modify the Weyl factor so that it is constant in a tiny neighborhood
31
Thus, given an operator O(x) in Rd , it is natural to define a cylinder

operator
Ocyl. (, n) e Oflat (x = e n).
(93)
We often omit the subscripts cyl. and flat, relying on the coordinates
to indicate which type of operator were discussing.
Exercise 6.2. Using (91), compute a two-point function of cylinder operators
hO(1 , n1 )O(2 , n2 )i.
(94)
Verify that it is time-translationally invariant on the cylinder. Show that in

the limit of large time separation = 2 1 1, the two-point function has
an expansion in terms of the form e(+n) with integer n 0. Interpret
these as coming from the exchange of states in the conformal multiplet of
O.
7. Reflection Positivity and Unitarity Bounds
7.1. Reflection Positivity
In Lorentzian signature, we are interested in unitary theories: theories
where the conserved charges (including the Hamiltonian) are Hermitian
operators so that they generate unitary transformations. Unitarity in
Lorentzian signature is equivalent to a property called reflection positivity in Euclidean signature.z
Consider a Lorentzian theory with a local operator OL and Hermitian
energy-momentum generators (H, PL ) (L is for Lorentzian). We have
the textbook formula
OL (t, x) = eiHtixPL OL (0, 0)eiHt+ixPL .
(95)
Let OL (0, 0) be Hermitian. It follows from (95) that OL (t, x) is Hermitian

too.
Now, let us Wick-rotate to Euclidean signature,
OE (tE , x) OL (itE , x) = eHtE ixPL OL (0, 0)eHtE +ixPL .
(96)
of O(x) and the flat-space calculation applies. This definition might not be consistent
with other independent definitions. For instance, if O(x) is the stress tensor, it gives a
different answer from the canonical definition (8) because of the Weyl anomaly.
z We make some brief comments about Euclidean vs. Lorentzian field theory and analytic
continuation in appendix B.
32
The Euclidean operator satisfies

OE (tE , x) = OE (tE , x).
(97)
To Wick-rotate an operator with spin, we conventionally add factors of i
to the time components,aa e.g. for a vector operator OL

,
0
0
OE
(tE , x) = iOL
(itE , x),
i
i
OE
(tE , x) = OL
(itE , x).
(98)
This leads to
1 ...`
1 `
(tE , x),
OE
(tE , x) = 1 1 ` ` OE
(99)
where = 20 0 is a reflection in the time-direction.

Thus, the way Hermitian conjugation acts on a Euclidean operator depends on which direction we call time. Whether an operator is Hermitian
or not depends on how we quantize the theory! This is very different from
Lorentzian signature, where the conjugation properties of operators dont
depend on a choice of reference frame.
As an example, consider the momentum generators
Z
P = dd1 x T 0 (0, x).
(100)
(From now on, we work in the Euclidean theory and omit the E subscripts.)
Using (99), we have
T i0 (0, x) = T i0 (0, x),
T 00 (0, x) = T 00 (0, x).

It follows that P
write
(101)
is Hermitian, and the P are antihermitian. We may

P 0 = H,
P j = iPLj ,
(102)
with H, PL Hermitian, and then (16) agrees with the formula we got from
Wick rotation (96). If we had quantized with a different time direction,
say the x1 -direction, then we would conclude that P 1 is Hermitian, while
P 0 , P 2 , . . . , P d1 are antihermitian.
To reiterate, the way conjugation acts on operators depends on how we
quantize our theory. This makes sense, because Hermitian conjugation is
something you do to operators on Hilbert spaces, and different quantizations have different Hilbert spaces.
aa These
factors are needed to make Euclidean correlation functions manifestly covariant

under SO(d) rotations.
33
This raises the question: given a Euclidean path integral, how do we

know if it computes the Wick-rotation of a unitary Lorentzian theory? One
important condition is that norms of states should be positive. Consider
some in-state |i given by acting on the vacuum with a bunch of operators
at negative Euclidean time
|i = O(tE1 ) O(tEn )|0i.
(103)
For brevity, we suppress the spatial positions of the operators. The conjugate state is given by
h| = (O(tE1 ) O(tEn )|0i)
= h0|O(tEn ) O(tE1 ).
(104)
That is, h| is given by taking the vacuum in the future and positioning
operators in a time-reflected way. Thus, the condition
h|i 0
(105)
says that a time-reflection symmetric configuration should have a positive

path integral, see figure 18. This is called reflection positivity.
Fig. 18.
Reflection positivity.
If a Euclidean theory is the Wick-rotation of a unitary Lorentzian theory,

then it will be reflection positive. However, some theories are more naturally
defined in Euclidean signature. In this case, reflection positivity must be
checked. It often suffices to check it in any microscopic theory in the same
universality class as the CFT were interested in.
Exercise 7.1. Consider the 2d Ising lattice correlator shown in figure 19.
Show that it can be written as a sum of squares, and is hence positive.
(Hint: first sum over spins off the line L, and then sum over spins on L.)
Generalize your proof to argue that the 2d Ising model is reflection-positive.
34
Fig. 19. A two point function on a 4 5 Ising lattice with free boundary conditions,
with spin operators inserted at the sites marked with an X.
The Osterwalder-Schrader reconstruction theorem says that, given a

collection of Euclidean correlators satisfying reflection positivity (and some
additional technical assumptions), we can reconstruct a unitary Lorentzian
quantum field theory by analytic continuation [28]. So reflection positivity
in Euclidean signature and unitarity in Lorentzian signature are essentially
equivalent, and we will use the terms interchangeably.
7.1.1. Real vs. Complex Operators
Because Hermitian conjugation is tricky in Euclidean signature, it is helpful
introduce some extra terminology. We call a local operator real if it is
Hermitian in Lorentzian signature. In Euclidean signature, real operators
= OL
, we have
satisfy (99). By contrast, for a complex operator OL
1 ...`
1 `
OE
(tE , x) = 1 1 ` ` OE
(tE , x).
(106)
Later we will need the following result. If 1 , 2 are real scalars and
O is a real operator with spin ` in a unitary theory, then the three-point
coefficient f1 2 O is real. This is easiest to see in Lorentzian signature when
the operators are spacelike separated x2ij > 0. Because local operators
commute at spacelike separation, we have
h0|1 (x1 )2 (x2 )O1 ` (x3 )|0i = h0|1 (x1 )2 (x2 )O1 ` (x3 )|0i. (107)
Substituting (74) gives f1 2 O = f1 2 O .
7.2. Reflection Positivity on the Cylinder
Reflection-positivity (or unitarity) has interesting consequences for CFTs
on the cylinder. The Hermitian conjugate of a real cylinder operator is
Ocyl. (, n)rad = Ocyl. (, n).
(108)
35
Using (93), this becomes

Oflat (x)rad = x2 Oflat
x
x2

.
(109)
Above, we have written rad to emphasize that Hermitian conjugation in

radial quantization is different from Hermitian conjugation in the usual P 0
quantization. From now on we write simply , and hope that the meaning
will be clear from context.
The right-hand side of (109) is just the image of O(x) under an inversion
I : x xx2 . The same is true for operators with spin, where the full
formula (55) gives
x
,
O1 ` (x) = I 1 1 (x) I ` ` (x)x2 O1 ...`
x2
2x x
.
(110)
I (x) =
x2
Exercise 7.2. Check that the two-point function of spin-1 operators (71)
satisfies reflection-positivity on the cylinder if CJ > 0.
Applying (110) to the stress tensor, we find that the action of conjugation
on the conformal charges in radial quantization is
Q = QII .
(111)
In particular, we have
M
= M ,
D = D,
P = K .
(112)
These facts let us calculate properties of correlation functions purely

algebraically. As an example, consider a two-point function. Letting ye =
y/y 2 , we have
hO(y)O(x)i = h0|(y 2 O(e
y )) O(x)|0i
= y 2 h0|(eyeP O(0)eeyP ) exP O(0)exP |0i
= y 2 h0|eeyK O(0) eyeK exP O(0)exP |0i
= y 2 h0|O(0) eyeK exP O(0)|0i

= y 2 hO|eyeK exP |Oi,
(113)
36
where weve defined

hO| h0|O(0) = lim y 2 h0|O(y).
y
(114)
By expanding the exponentials, we can evaluate (113) using the conformal

algebra. For example, the first couple terms are

y
2
hO(y)O(x)i = y
hO|Oi + 2 x hO|K P |Oi + . . . , (115)
y
where weve used that K|Oi = hO|P = 0 because O is primary. Using the
conformal commutation relations,
hO|K P |Oi = hO|[K , P ]|Oi
= hO|(2D 2M )|Oi
= 2 hO|Oi.
(116)
Thus,

yx
hO(y)O(x)i = y 2 hO|Oi 1 + 2 2 + . . . .
y
(117)
This exactly matches the expansion of hO|Oi/(x y)2 in small |x|/|y|!

You can imagine computing all the higher terms and matching the whole
series expansion.
Let us also prove our earlier claim that a two-point function of operators in different irreducible spin representations must vanish. Consider a
primary operator Oa transforming in a nontrivial unitary representation of
SO(d). The dual state transforms in the dual representation, so we will
write it with a lowered index (|Oa i) = hOa |. Consider the matrix element
hOa |M |Ob i. Using that M is antihermitian (112), we can act with it
on both the bra and the ket:
((S )c a ) hOc |Ob i = hOa |M |Ob i = hOa |Oc i(S )c b .
(118)
But S is antihermitian as well, so as a matrix equation this says

S N = N S ,
(119)
where Na b hOa |Ob i. By Schurs lemma, Na b must vanish if a and b are

indices of different irreducible representations. If a, b are indices for a single
irreducible representation, then N is proportional to the identity.
Exercise 7.3. This computation is not directly relevant to the course, but
it is instructive for getting used to radial ordering. Consider a three-point
37
function of scalars
hOi (x1 )Oj (x2 )Ok (x3 )i = h0|R{Oi (x1 )Oj (x2 )Ok (x3 )}|0i
= (|x3 | |x2 | |x1 |)h0|Ok (x3 )Oj (x2 )Oi (x1 )|0i
+permutations.
(120)
Consider the operator e2iD1 where

D1 = x1 1 + 1 .
(121)
Using the fact that e2iD1 O1 (x1 ) = e2iD O1 (x1 )e2iD , compute the action
of e2iD1 on each of the terms above. You will get different answers for each
of the different operator orderings.
Now determine the action of e2iD1 on the known answer for the scalar
three-point function (65). Check that the two answers agree.
7.3. Unitarity Bounds
Thinking about the theory on the cylinder gives a natural inner product
on states in radial quantization. Unitarity (or reflection positivity) implies
that the norms of states must be nonnegative. By demanding positivity for
every state in a conformal multiplet, we obtain bounds on dimensions of
primary operators [2931]. We have already seen an example in (116). We
found
|P0 |Oi|2 = hO|K0 P0 |Oi = 2hO|Oi.
(122)
Unitarity implies 0.
Let us do the same exercise, this time for an operator Oa in a nontrivial
irreducible representation RO of SO(d). We normalize O so that
hOb |Oa i = ba .
(123)
Taking inner products between first-level descendants and using the conformal algebra, we find
(P |Oa i) P |Ob i = hOa |K P |Ob i = 2 ab 2(S )a b .
(124)
The state P |Ob i lives in the representation V RO of SO(d), where V is

the vector representation. Unitarity implies that (124) must be positivedefinite as a matrix acting on this space. This implies
max-eigenvalue((S )a b ).
(125)
38
Let us write
1
(L ) (S )a b
2
,
(S )a b =
(L )
(126)
where (L ) is the generator of rotations in the vector representation V .

Writing A = for an adjoint index of SO(d), and thinking of LA , SA as
operators on V RO , this becomes

1
LA SA =
(L + S)2 L2 S 2
2
1
= (Cas(V RO ) + Cas(V ) + Cas(RO )) ,
(127)
2
where weve used the familiar trick from basic quantum mechanics to get a
linear combination of Casimir operators.bb
Lets specialize to the case where RO = V` is the spin-` traceless symmetric tensor representation. V` has Casimir `(` + d 2). To get the
maximal eigenvalue of L S, we need the minimal Casimir of
V V` = V`1 . . .
(` > 0).
(128)
Here the . . . are irreducible representations with larger Casimirs. Thus,

1
(Cas(V`1 ) + Cas(V ) + Cas(V` ))
2
= ` + d 2.
(129)
This computation was valid only for ` > 0, since for scalars V V`=0 = V .
One can also consider more complicated descendants.
Exercise 7.4. Compute the norm of P P |Oi, where O is a scalar. Show
that unitarity implies either = 0 or d2
2 . This gives a stronger
condition than what we derived above ( 0) for scalars.
For traceless symmetric tensors in general conformal field theories, these
inequalities are the best you can do (other descendants give no new information). In theories with more symmetry, like supersymmetric theories or
2d CFTs, unitarity bounds can be more interesting. A classic reference
for unitarity bound computations is [31]. In the math literature, unitarity
bounds for higher-dimensional CFTs were essentially computed long ago by
Jantzen [30], though the relevance of that work for physics has only been
emphasized recently [32, 33].
bb The
quadratic Casimir is L2 because our generators are antihermitian and differ

from the conventional ones by a factor of i.
39
In summary, we have the unitarity bounds

= 0 (unit operator), or
(
d2
` = 0,
2
` + d 2 ` > 0.
(130)
7.3.1. Null States and Conserved Currents

If saturates the bounds (130), the conformal multiplet will have a null
state. For the unit operator, all descendants are null. For a scalar with
dimension d2
2 , the null state is
P 2 |Oi = 0.
(131)
In operator language, this says 2 O(x) = 0, which means O satisfies the

Klein-Gordon equation, and is thus a free scalar that decouples from the
rest of the CFT.
For a spin-` operator, the null state iscc
P |O2 ` i = 0.
(132)
In operator language, this becomes the equation for a conserved current

O2 ` (x) = 0.
(133)
The implication also works the other way, so

=`+d2
if and only if
O is a conserved current. (134)
Some important examples are global symmetry currents (` = 1, = d 1)

and the stress tensor (` = 2, = d). For CFTs in d 3, the presence of
currents with spin ` 3 implies that the theory is free [34, 35].dd
7.4. Only Primaries and Descendants
With a positive-definite inner product, we can now prove that all operators
in unitary CFTs are linear combinations of primaries and descendants. We
cc The null state has spin ` 1 because the unitarity bound came from V
`1 V V` .
Something special happens for vectors in 2d, where V V = 1 1 V2 , with the extra
1 appearing because of the antisymmetric symbol. Unitarity then implies that J
and J are each separately conserved.
dd One must also assume the existence of exactly one stress tensor, since otherwise the
theory could contain a free subsector, decoupled from the rest.
40
will use one additional physical assumption: that the partition function of
the theory on S d1 S1 is finite,
ZS d1 S1 = Tr(eD ) < .
(135)
This means that eD is trace-class, and hence diagonalizable with a discrete spectrum (by the spectral theorem).ee It follows that D is also diagonalizable, with real eigenvalues because D is Hermitian.
Now consider a local operator O, and assume for simplicity it is an
eigenvector of dilatation with dimension . By finiteness of the partition
function, there are a finite number of primary operators Op with dimension
less than or equal to . Using the inner product, we may subtract off
the projections of O onto the conformal multiplets of Op to get O0 . Now
suppose (for a contradiction) that O0 6= 0. Acting on it with K s, we must
eventually get zero (again by finiteness of the partition function), which
means there is a new primary with dimension below , a contradiction.
Thus O0 = 0, and O is a linear combination of states in the multiplets Op .
8. The Operator Product Expansion
If we insert two operators Oi (x)Oj (0) inside a ball and perform the path
integral over the interior, we get some state on the boundary. Because
every state is a linear combination of primaries and descendants, we can
decompose this state as
X
Oi (x)Oj (0)|0i =
Cijk (x, P )Ok (0)|0i,
(136)
k
where k runs over primary operators and Cijk (x, P ) is an operator that
packages together primaries and descendants in the k-th conformal multiplet (figure 20).
Eq. (136) is an exact equation that can be used in the path integral, as
long as all other operators are outside the sphere with radius |x|. Using the
state-operator correspondence, we can write
X
Oi (x1 )Oj (x2 ) =
Cijk (x12 , 2 )Ok (x2 ),
(OPE)
(137)
k
ee Assuming
eD
is trace-class may be too strong for some applications. Boundedness

of eD suffices for D to be diagonalizable (with a possibly continuous spectrum). An
interesting example is Liouville theory, which has a divergent partition function and
continuous spectrum, but still has many properties of a sensible CFT, like an OPE.
41
Fig. 20. A state created by two operator insertions can be expanded as a sum of primary
and descendant states.
where it is understood that (137) is valid inside any correlation function

where the other operators On (xn ) have |x2n | |x12 |. Eq. (137) is called
the Operator Product Expansion (OPE).
We could alternatively perform radial quantization around a different
point x3 , giving
X
0
Oi (x1 )Oj (x2 ) =
Cijk
(x13 , x23 , 3 )Ok (x3 ),
(138)
k
0
where Cijk
(x13 , x23 , 3 ) is some other differential operator (figure 21). The
form (137) is usually more convenient for computations, but the existence
of (138) is important. It shows that we can do the OPE between two
operators whenever its possible to draw any sphere that separates the two
operators from all the others.
Fig. 21.
It isnt necessary for one of the operators to be at the origin.
We are being a bit schematic in writing the above equations. Its possible for all the operators to have spin. In this case, the OPE looks like
X
ab
c
Oia (x1 )Ojb (x2 ) =
Cijk
(139)
c (x12 , 2 )Ok (x2 ),
k
where a, b, c are indices for (possibly different) representations of SO(d).

8.1. Consistency with Conformal Invariance
Conformal invariance strongly restricts the form of the OPE. For simplicity,
suppose Oi , Oj , and Ok are scalars.
42
Exercise 8.1. By acting on both sides of (136) with D, prove that

Cijk (x, ) has an expansion of the form

Cijk (x, ) |x|k i j 1 + #x + #x x + #x2 2 + . . . .
(140)
This is just a fancy way of saying we can do dimensional analysis and that
Oi has length-dimension i . Were also implicitly using rotational invariance by contracting all the indices appropriately. We could have proved this
too by acting with M .
We get a more interesting constraint by acting with K . In fact, consistency with K completely fixes Cijk up to an overall coefficient. In this
way, we can determine the coefficients in (140).
This computation is a little annoying (exercise!), so heres a simpler way
to see why the form of the OPE is fixed, and to get the coefficients in (140).
Take the correlation function of both sides of (137) with a third operator
Ok (x3 ) (we will assume |x23 | |x12 |, so that the OPE is valid),
X
hOi (x1 )Oj (x2 )Ok (x3 )i =
Cijk0 (x12 , 2 )hOk0 (x2 )Ok (x3 )i. (141)
k0
The three-point function on the left-hand side is fixed by conformal invariance, and is given in (65). We can choose an orthonormal basis of primary
k
operators, so that hOk (x2 )Ok0 (x3 )i = kk0 x2
. The sum then collapses
23
to a single term, giving
fijk
i +j k j +k i k +i j
x12
x23
x31
k
= Cijk (x12 , 2 )x2
.
23
(142)
This determines Cijk to be proportional to fijk , times a differential operator

that depends only on the i s. The operator can be obtained by matching
the small |x12 |/|x23 | expansion of both sides of (142).
Exercise 8.2. Consider the special case i = j = , and k = .
Show

1
Cijk (x, ) = fijk x2 1 + x + x x + x2 2 + . . . ,
2
(143)
where
=
+2
,
8( + 1)
and
16(
d2
2 )(
+ 1)
(144)
43
8.2. Computing Correlators with the OPE

Equation (141) gives an example of using the OPE to reduce a three-point
function to a sum of two-point functions. In general, we can use the OPE
to reduce any n-point function to a sum of n 1-point functions,
X
hO1 (x1 )O2 (x2 ) On (xn )i =
C12k (x12 , 2 )hOk (x2 ) On (xn )i.
k
(145)
Recursing, we reduce everything to a sum of one-point functions, which are
fixed by dimensional analysis,
(
1 if O is the unit operator,
hO(x)i =
(146)
0 otherwise.
This gives an algorithm for computing any flat-space correlation function
using the OPE. It shows that all these correlators are determined by dimensions i , spins, and OPE coefficients fijk .ff
9. Conformal Blocks
9.1. Using the OPE
Let us use the OPE to compute a four-point function of identical scalars.
Recall that Ward identities imply
h(x1 )(x2 )(x3 )(x4 )i =
g(u, v)
x12 x34
where the cross-ratios u, v are given by (66).

The OPE takes the form
X
(x1 )(x2 ) =
fO Ca (x12 , 2 )Oa (x2 ),
(147)
(148)
where Oa can have nonzero spin in general. For Oa to appear in the OPE
of two scalars, it must transform in a spin-` traceless symmetric tensor
representation of SO(d).
Exercise 9.1. Prove this as follows. Show that hOa |(x)|i vanishes unless Oa is a symmetric tensor. (Tracelessness comes from restricting to
ff The
OPE is also valid on any conformally flat manifold. The difference is that on
nontrivial manifolds, non-unit operators can have nonzero one-point functions. An example is Rd1 S1 , which has the interpretation as a CFT at finite temperature. By
dimensional analysis, we have hOiRd1 S 1 O T O .
44
irreducible representations of SO(d).) Argue that if hOa |(x)|i vanishes,

then for any descendent |i = P P |Oi, the matrix element h|(x)|i
vanishes as well.
Exercise 9.2. Using (65), show that fO vanishes unless ` is even.
Assuming the points are configured appropriately, we can pair up the
operators (12) (34) and perform the OPE between them,gg
h(x1 )(x2 )(x3 )(x4 )i
X
fO fO0 Ca (x12 , 2 )Cb (x34 , 4 )hOa (x2 )O0b (x4 )i
=
O,O 0
2
fO
Ca (x12 , 2 )Cb (x34 , 4 )
x12 x34
I ab (x24 )
O
x2
24
2
fO
gO ,Ò (xi ),
(149)
where
g,` (xi ) x12 x34 Ca (x12 , 2 )Cb (x34 , 4 )
I ab (x24 )
.
x2
24
(150)
In (149), we have chosen an orthonormal basis of operators and used that

hOa (x)O0b (0)i = OO0
I ab (x)
,
x2O
(151)
where I ab (x) = I 1 ` ,1 ` (x) is the tensor in (73).

The functions g,` (xi ) are called conformal blocks. Although its not
obvious from the way we defined them, it turns out they are actually functions of the conformal cross-ratios u, v alone. We thus have the conformal
block decomposition
X
2
g(u, v) =
fO
gO ,Ò (u, v).
(152)
O
Exercise 9.3. Using the differential operator (143), show

g,0 (u, v) = u/2 (1 + . . . ) .
gg Although
(153)
our computation will make it look like we need x3,4 to be sufficiently far
from x1,2 , we will see shortly that the answer will be correct whenever we can draw any
sphere separating x1 , x2 from x3 , x4 .
45
Exercise 9.4. Using (74), argue that x2 CO (x, ) is independent of

for any spin of O. Conclude that g,` (u, v) is independent of . (This
is a special property of conformal blocks for operators with identical scaling
dimensions.)
9.2. In Radial Quantization
A conformal block represents the contribution of a single conformal multiplet to a four-point function. It is instructive to understand it in radial
quantization. Along the way, well explain why the blocks are functions of
the cross-ratios u, v alone.
Let us pick an origin such that |x3,4 | |x1,2 |, so that
h(x1 )(x2 )(x3 )(x4 )i = h0|R{(x3 )(x4 )}R{(x1 )(x2 )}|0i. (154)
For a primary operator O, let |O| be the projector onto the conformal
multiplet of O,
X
1
|O|
|iN
h|,
N h|i.
(155)
,=O,P O,P P O,...
The identity is the sum of these projectors over all primary operators.
X
1=
|O|.
(156)
O
Inserting this into (154) gives

X
h(x1 )(x2 )(x3 )(x4 )i =
h0|R{(x3 )(x4 )}|O|R{(x1 )(x2 )}|0i.
O
(157)
Each term in the sum is a conformal block times a squared OPE coefficient
and some conventional powers of xij ,
h0|R{(x3 )(x4 )}|O|R{(x1 )(x2 )}|0i =
2
fO
2
x12 x34
gO ,Ò (u, v). (158)
Exercise 9.5. Verify the equivalence between (158) and (150) by performing the OPE between (x3 )(x4 ) and (x1 )(x2 ).
This expression makes it clear why g,` (u, v) is a function of u and v:
the projector |O| commutes with all conformal generators (by construction).
Thus, the object above satisfies all the same Ward identities as a four-point
function of primaries, and it must take the form (68). In path integral
46
language, we can think of |O| as a new type of surface operator. Here,

weve inserted it on a sphere separating x1,2 from x3,4 .
9.3. From the Conformal Casimir
We can now give a simple and elegant way to compute the conformal block,
due to Dolan & Osborn [36]. Recall that the conformal algebra is isomorphic
to SO(d + 1, 1), with generators Lab given by (38). The Casimir C =
21 Lab Lab acts with the same eigenvalue on every state in an irreducible
representation.
Exercise 9.6. Show that this eigenvalue is given by
C|Oi = ,` |Oi,
,` ( d) + `(` + d 2).
(159)
It follows that C gives this same eigenvalue when acting on the projection
operator |O| from either the left or right,
C|O| = |O|C = ,` |O|.
(160)
Let Lab,i be the differential operator giving the action of Lab on the
operator (xi ). Note that
(Lab,1 + Lab,2 )(x1 )(x2 )|0i = ([Lab , (x1 )](x2 ) + (x1 )[Lab , (x2 )]) |0i
= Lab (x1 )(x2 )|0i.
(161)
Thus,
C(x1 )(x2 )|0i = D1,2 (x1 )(x2 )|0i,
1
+ Lab
where
D1,2 (Lab
2 )(Lab,1 + Lab,2 ).
2 1
We then have
(162)
D1,2 h0|R{(x3 )(x4 )}|O|R{(x1 )(x2 )}|0i
= h0|R{(x3 )(x4 )}|O|CR{(x1 )(x2 )}|0i
= ,` h0|R{(x3 )(x4 )}|O|R{(x1 )(x2 )}|0i.
(163)
Plugging in (158), we find that g,` satisfies the differential equation

Dg,` (u, v) = ,` g,` (u, v),
(164)
where the second-order differential operator D is given by
D = 2(z 2 (1 z)z2 z 2 z ) + 2(
z 2 (1 z)z2 z2 z)
z z
+2(d 2)
((1 z)z (1 z)z).
z z
(165)
47
Eq. (164), together with the boundary condition (153) (and its generalization to nonzero spin, which we give shortly), then determines the
conformal block g,` (u, v). In even dimensions, the Casimir equation can
be solved analytically. For example, in 2d and 4d [36, 37],
(2d)
g,` (u, v) = k+` (z)k` (

z ) + k` (z)k+` (
z ),
(166)
z z
(4d)
(k+` (z)k`2 (
z ) k`2 (z)k+` (
z )) , (167)
g,` (u, v) =
z z

k (x) x/2 2 F1
, , , x .
(168)
2 2
In odd dimensions, no explicit formula in terms of elementary functions
is known. However the blocks can still be computed in a series expansion
using the Casimir equation or alternative techniques like recursion relations.
9.4. Series Expansion
It will be helpful to understand the series expansion of the conformal blocks
in more detail. The radial coordinates of [26, 38] are ideal for this purpose. Using conformal transformations, we can place all four operators on
a plane in the configuration shown in figure 22. This makes it clear that
the conformal block expansion is valid whenever || < 1.
x2 =
x3 = 1
x4 = 1
x1 =
Fig. 22. Any four points can be brought to the above configuration using conformal
transformations. (Figure from [26].)
Exercise 9.7. Show that = rei is related to z via

=
(1 +
,
1 z)2
(and similarly for = rei and z).
z=
4
(1 + )2
(169)
48
In radial quantization, this corresponds to placing cylinder operators

(93) at diametrically opposite points n and n0 on S d1 , with cos =
n n0 , and with the pairs separated by time = log r (figure 23). The
conformal block is then
g,` (u, v) = h(n)||O|e D |(n0 )i,
(170)
where weve defined the statehh

|(n)i
Fig. 23.
2
cyl. (0, n)cyl. (0, n)|0i.
fO
(171)
Configuration on the cylinder corresponding to (170).
A descendant P 1 P n |Oi has energy + n on the cylinder. Within

the n-th energy level, the SO(d) spins that appear are
j {` + n, ` + n 2, . . . , max(` n, ` + n mod 2)}.
(172)
Consider a set of descendent states |n, ji1 j with energy + n and spin
j. They contribute
r+n h(n)|n, ji1 j 1 j hn, j|(n0 )i.
(173)
By rotational invariance,
h(n)|n, ji1 j n1 nj traces.
(174)
Because |(n)i = |(n)i, j must be even (and thus n is even). The

contraction of two traceless symmetric tensors is a Gegenbauer polynomial,
d2
2
Cj
(n n0 ) (n1 nj traces)(n01 n0j traces),

2
factor 2 = hcyl. (0, n)cyl. (0, n)i1 comes from transforming x12
cylinder (exercise!).
hh The
(175)
to the
49
so (173) becomes
d2
2
r+n h(n)|n, ji1 j 1 j hn, j|(n0 )i r+n Cj
Summing over descendants, we find

X
d2
g,` (u, v) =
Bn,j r+n Cj 2 (cos ),
(cos ). (176)
(177)
n=0,2,...
j
where j ranges over the values in (172) and Bn,j are constants. Notice a
few properties:
The leading term in the r-expansion comes from the primary state |Oi
with n = 0 and j = `. This can be used as a boundary condition in the
Casimir equation to determine the higher coefficients Bn,j .
The Bn,j are positive in a unitary theory because they are given by
norms of projections of |i onto energy and spin eigenstates.
The Bn,j are rational functions of . This follows because the Casimir
eigenvalue ,` is polynomial in , or alternatively from the fact that
the differential operators Ca (x, ) appearing in the OPE (148) have a
series expansion in x with rational coefficients, see exercise 8.2.
(2d)
(4d)
Exercise 9.8. Expand g,` (u, v) and g,` (u, v) to the first few orders in r,
and check these properties. Verify that some of the coefficients Bn,j become
negative when violates the unitarity bound.
Exercise 9.9. By rewriting in terms of r, and using (177), show that
even spin blocks are invariant under x1 x2 or x3 x4 ,

u 1
,
g,` (u, v) = g,`
,
(` even).
(178)
v v
10. The Conformal Bootstrap
10.1. OPE Associativity and Crossing Symmetry
Weve gotten pretty far using symmetries and basic principles of quantum
field theory. We classified operators into primaries and descendants. We
established the OPE, which determines n-point functions as sums of (n1)point functions,
X
hO1 (x1 )O2 (x2 ) On (xn )i =
C12k (x12 , 2 )hO2 (x2 ) On (xn )i.
k
(179)
50
And we showed that the differential operators Cijk (x, ) are determined by
conformal symmetry in terms of dimensions i , spins, and OPE coefficients
fijk .
Now its time to implement the last step of the bootstrap program:
impose consistency conditions and derive constraints. Using the OPE, all
correlation functions can be written in terms of the CFT data i , fijk .
Now suppose someone hands you a random set of numbers i , fijk . Does
that define a consistent CFT?
Fig. 24. Two different ways of evaluating a five-point function using the OPE. Dots
represent operators in the correlator, and vertices represent the OPE. The two ways differ
by a crossing symmetry transformation (182) applied to the left part of the diagram.
The answer is: not always. By doing the OPE (179) between different
pairs of operators in different orders (see figure 24), we get naively different
expressions for the same correlator in terms of CFT data. These expressions
should agree. This means the OPE should be associative,
O1 O2 O3 = O1 O2 O3 ,
(180)
or more explicitly,
C12i (x12 , 2 )Ci3j (x23 , 3 )Oj (x3 ) = C23i (x23 , 3 )C1ij (x13 , 3 )Oj (x3 ).
(181)
(We suppress spin indices for simplicity.) Taking the correlator of both
sides with a fourth operator O4 (x4 ) gives the crossing symmetry equation
1
X A Oi
A

A

i
A

2
H

H
=
Oi
H
H

(182)
51
The left-hand side is the conformal block expansion of hO1 O2 O3 O4 i in the

12 34 channel, while the right-hand side is the expansion in the 14 23
channel.
Exercise 10.1. Argue that by choosing different operators O4 and taking
linear combinations of derivatives, one can recover OPE associativity (181)
from the crossing equation (182). Conclude that crossing symmetry of all
four-point functions implies crossing symmetry of all n-point functions (i.e.
that any way of computing an n-point function using the OPE gives the
same result).
The crossing equation (182) is a powerful but complicated constraint
on the CFT data. The rest of this course will be devoted to studying its
implications for the simplest possible case: a four-point function of identical
scalars h(x1 )(x2 )(x3 )(x4 )i.
10.1.1. Additional Structures and Consistency Conditions
Before jumping in, let us reflect on the implications of exercise 10.1: A
solution to the crossing equations (182) gives a completely nonperturbative definition of flat-space correlation functions of local operators, without
the need for a Lagrangian. This is most of the way towards a full theory. However, some structures associated with local QFTs are missing, and
additional structures might bring new consistency conditions.
Firstly, CFTs can admit extended objects like line operators, surface
operators, boundaries, and interfaces. These objects have additional data
associated with them, and its possible to write down OPEs and crossing equations that relate this data to itself and the usual CFT data, see
e.g. [39, 40]. It is also interesting to consider correlation functions on manifolds not conformally equivalent to flat space. An example includes the
theory at finite temperature (discussed in footnote ff). This introduces
more data, for example the one-point functions of local and extended operators on nontrivial manifolds.ii Other interesting constraints come from
studying CFTs in Lorentzian signature. Examples include bounds from
energy positivity [43], dispersion relations [4447], and causality [48, 49].
The full set of data and consistency conditions associated with a CFT
is not known in general. However, we do have examples of constraints on
ii It
is known that this data is not determined by the local operator spectrum. For
example, pure Chern-Simons theory has no local operators at all, but has interesting
nonlocal observables that depend on the gauge group and level [41]. Also, 4d conformal
gauge theories admit different sets of line operators for the same set of local operators [42].
52
local operators beyond the OPE and crossing equations. The most famous
is modular invariance: the requirement that the partition function of a 2d
CFT on the torus T 2 be invariant (or covariant) under large diffeomorphisms. Imposing modular invariance is an additional step that must be
performed after solving the crossing equations in 2d CFTs [50].jj
10.2. Crossing Symmetry for Identical Scalars
For the rest of this course, we study the crossing equation for a four-point
function of identical real scalars h(x1 )(x2 )(x3 )(x4 )i. Let us summarize
the consequences of conformal symmetry and unitarity for this case.
We have the OPE
(x1 )(x2 ) =
fO C1 ` (x12 , 2 )O1 ` (x2 ).
(183)
We denote the dimension of O by and the spin by `. By exercise 9.2,

` must be even.
We can choose a basis of operators such that the Os are real and
orthonormal, as in (151). Unitarity implies that the three-point coefficients fO are real in this basis (section 7.1.1).
Each O satisfies the unitarity bounds
= 0 (unit operator), or
d2
(` = 0),
2
` + d 2 (` > 0).
(184)
We have the conformal block expansion

h(x1 )(x2 )(x3 )(x4 )i =
g(u, v) =
g(u, v)
2
2
x12 x34
X
2
fO
g,` (u, v),
O
(185)
(186)
where g,` (u, v) are conformal blocks, and the cross ratios are
u = z z =
jj 2d
x212 x234
,
x213 x224
v = (1 z)(1 z) =
x223 x214
.
x213 x224
(187)
is special because the space of states on a spatial slice S 1 T 2 is the same as

the space of states in radial quantization, and thus modular invariance on T 2 directly
constrains local operators. This is not true in d 3, so it is not clear how modular
invariance on T d constrains local operators in that case.
53
Crossing symmetry is equivalent to the condition (70) that our fourpoint function is invariant under 1 3 or 2 4,
u
g(u, v) =
g(v, u).
(188)
v
Eq. (178) shows that invariance of the four-point function under 1
2 or 3 4 is true block-by-block. All other permutations can be
generated from these.
We know at least two operators present in the OPE: the unit
operator and the stress tensor. Normalizing so that h(x)(0)i = x2 ,
we have f1 = 1. The stress tensor three-point coefficient is set by Ward
identities to be fT / CT , where CT is the coefficient of the twopoint function of the canonically normalized stress tensor (73). The factor
of 1/ CT relative to (75) comes from choosing the basis of operators O to

be orthonormal.
10.3. An Infinite Number of Primaries
To get a feel for the crossing equation (188), let us consider a simple limit:
z 0 with z = z. This corresponds to x2 x1 with all four operators
collinear.
Recall that the blocks go like g,` (u, v) (z z)/2 in this limit, so the
left-hand side of (188) is dominated by the smallest dimension operator in
the OPE, the unit operator:
LHS : 1 + . . .
(z 0).
(189)
Crossing u v corresponds to (z, z) (1 z, 1 z). In the limit

z 0, the crossed conformal blocks g,` (1 z, 1 z) go like log z.
Exercise 10.2. Check this for the explicit formulae (166) and (167).
Thus, each term on the right-hand side goes like
each term, RHS : z 2 log z + . . .
(z 0).
(190)
As z 0, any finite sum of terms of the form (190) vanishes. Thus, for a
sum of operators on the right-hand side to reproduce the unit operator on
the left-hand side, we need an infinite number of primary operators.kk
kk This
doesnt contradict the textbook statement that rational 2d CFTs contain a finite
number of primary operators. In that context, primary refers to primary operators
with respect to the Virasoro algebra. Here, we are discussing primaries with respect
to the global conformal group, which is SL(2, R) SL(2, R) in 2d. A single Virasoro
representation contains an infinite number of global conformal representations.
54
One can prove that as z 0, the sum on the right-hand side is dom
inated by operators of dimension 1/ z [38]. In other words, the
unit operator on the left-hand side maps to the large- asymptotics of the
sum over operators on the right-hand side. This is a general feature of the
crossing equation it cannot be satisfied block-by-block.
One can also show [38] that the conformal block expansion converges
exponentially in whenever || 1, where is defined in (169). In particular, this means that both sides of the crossing equation converge exponentially in a finite neighborhood of the point z = z = 12 , which will play
an important role in the next section.
Analyzing different limits of the crossing equation can give other information about the CFT spectrum. For example, the limit z 0 with z
fixed gives information about operators with large spin [46, 5153].
10.4. Bounds on CFT Data
The crossing equation (188) has been known for decades. However, little progress was made in solving it for CFTs in d 3 until 2008, in a
breakthrough paper by Rattazzi, Rychkov, Tonni, and Vichi [54]. Instead
of trying to solve the crossing equation exactly, their insight was to derive bounds on CFT data by studying the crossing equation geometrically.
Crucially, their methods let one make rigorous statements about some of
the CFT data (for example, the lowest few operator dimensions), without
having to compute all of it.
The basic idea is simple. Let us write the crossing equation as
X

2
(191)
fO
v g,` (u, v) u g,` (v, u) = 0.
{z
}
|
O
F,`
(u,v)
Abstractly, we can think of the functions F,` (u, v) as vectors F~,` in the
(infinite-dimensional) vector space of functions of u and v. Recall that the
2
are positive, so (191) has the form
coefficients fO
X
p,` F~,` = 0,
p,` 0,
(192)
,`
where , ` run over dimensions and spins of operators in the OPE.

Equation (192) says that a bunch of vectors sum to zero with positive
coefficients. This may or may not be possible, depending on the vectors.
The left-hand side of figure 25 shows a case where its possible, and the
right-hand side shows a case where its impossible. The way to distinguish
55
Fig. 25. On the left, a bunch of vectors that can sum to zero with positive coefficients.
On the right, a bunch of vectors that cant. In the latter case, its possible to find a
separating plane .
these cases is to search for a separating plane through the origin such that
all the vectors F~,` lie on one side of . If exists, then the F~,` cannot
2
satisfy crossing, for any choice of coefficients p,` = fO . This suggests
the following procedure for bounding CFT data.
Algorithm 1 (Bounding Operator Dimensions).
(1) Make a hypothesis for which dimensions and spins , ` appear in the
OPE.
(2) Search for a linear functional that is nonnegative acting on all F~,`
satisfying the hypothesis,
(F~,` ) 0,
(193)
and strictly positive on at least one operator (usually taken to be the

unit operator).
(3) If exists, the hypothesis is wrong. We see this by applying to both
sides of (192) and finding a contradiction.
A slight modification of this algorithm also lets one bound OPE coefficients [55].
10.5. An Example Bound
Lets work through an example.ll Consider a 2d CFT with a real scalar
primary of dimension = 81 . Project the crossing equation onto a
ll An
early version of this example is due to Sheer El-Showk, and this specific implementation is due to Jo
ao Penedones and Pedro Vieira.
56
two-dimensional subspace with the linear map

1 3
1 1
1 3
1 1
~v (F ) = H
,
H
,
,H
,
H
,
R2 ,
2 5
2 3
2 5
3 4
F (u, v)
where H(z, z) =
,
u v
u = z z,
v = (1 z)(1 z).
By linearity, the vectors
~v (F,` )
(194)
also sum to zero with positive coefficients,
p,`~v (F,` ) = 0.
(195)
,`
In figure 26, we plot ~v (F,` ) for all , ` satisfying the unitarity bounds
(184), where the conformal blocks are given by (166). We have normalized
the vectors so that they are easy to see, since changes in normalization can
be absorbed into the coefficients p,` .
As varies from ` (the unitarity bound) to , ~v (F,` ) sweeps out a

curve. The curves for higher spin operators ` 2 are extremely simple,
converging quickly at large . The scalar curve is more interesting. It
circles counterclockwise partway around the origin before circling back and
converging as . The region [0.161, 1.04] of the scalar curve
lies in a different half space from the other curves. To satisfy (195), we
must include at least one vector from this region. Thus, we immediately
conclude
Claim 1. In a unitary 2d CFT with a real operator of dimension = 81 ,
there must exist a scalar in the OPE with dimension [0.161, 1.04].
Proof. We have already given the proof, but let us rephrase it in terms of
Algorithm 1.
Suppose (for a contradiction) that there are no scalars with
[0.161, 1.04] in the OPE.
Let
(F ) = ~u ~v (F ),
(196)
where ~v (F ) is defined in (194) and ~u R2 is normal to the dashed line,
pointing to the bottom right in figure 26. Note that (F,` ) 0 for
all , ` satisfying our hypothesis. Further, is strictly positive on at
57
=
-
Fig. 26.
Vectors ~v (F,`
) for all values of , ` satisfying the 2d unitarity bound `,
with ` even. Dots represent vectors at the unitarity bound = `. As varies, ~v (F,`
)
sweeps out a curve starting at the dot and approaching the negative y-axis as .
The curves for spins ` = 16, 18, . . . look similar and converge quickly as ` , so we
have not included them in the figure. All vectors are normalized for visual simplicity,
except for the unit operator ~v (F0,0 ) = ~0. The dashed line splits the figure into two
half-spaces with the stress tensor ~v (F2,2 ) on the boundary. The thicker region of the
` = 0 curve, in a different half-space from the rest of the figure, corresponds to scalars
with dimension [0.161, 1.04].
least one vector appearing in the OPE. (To establish this, we

could rotate ~u slightly so that is strictly positive on the stress tensor
vector. Alternatively, we could use the fact that at least one higher
dimension operator must appear in the OPE.)
Applying to both sides of (192), we find a contradiction: 0 > 0.
58
Exercise 10.3. Check that (F,` ) 0 is true asymptotically as

and ` . Convince yourself that the proof of Claim 1 could be made
rigorous to a mathematicians standards.
Fiddling around with two-dimensional vectors has yielded a surprisingly
strong result. The 2d Ising CFT is an example of a unitary theory with
a real scalar (the spin operator) with dimension = 81 . The lowest
dimension scalar in the OPE is the energy operator , which has
= 1. So our upper bound scalar 1.04 is within 4 percent of being
saturated by an actual theory!
10.6. Numerical Techniques
The bound scalar [0.161, 1.04] is not particularly special. If we had
picked a different two-dimensional subspace (194), we would have gotten
different numbers. We might also consider higher-dimensional subspaces
and derive even stronger results. Although it is possible to prove bounds
by hand as we did in the previous subsection, computerized searches are the
current state-of-the-art. In this section, we describe some of the techniques
involved.
The hard part of Algorithm 1 is the middle step: finding a functional
such that
(F~,` ) 0,
for all , ` satisfying our hypothesis.
(197)
If we want to use a computer, we have two immediate difficulties:

(1) The space of possible s is infinite-dimensional.
(2) There are an infinite number of positivity constraints (197) one for
each , ` satisfying our hypothesis. Our hypothesis usually allows `
to range from 0 to , and to vary continuously (aside from a few
discrete values).mm
The first difficulty is easy to fix. Instead of searching the infinitedimensional space of all functionals, we restrict to a finite-dimensional subspace. If we find in our subspace that satisfies the positivity constraints,
we can immediately rule out our hypothesis about the spectrum. If we
mm This
is due to ignorance about the spectrum. Although physical CFT spectra should
be discrete, we dont know exactly which discrete values takes, and so we must include
positivity constraints for continuously varying .
59
cant find , then we cant conclude anything about the spectrum: either
no functional exists, or we just werent searching a big enough subspace.
In the example from section 10.5, we restricted to linear combinations
of the components of ~v (F ) in (194). For numerical computations, we usually
take linear combinations of derivatives around the crossing-symmetric point
z = z = 21 ,
(F ) =
amn zm zn F (z, z)|z=z= 21 ,
(198)
m+n
where is some cutoff. The functional is now parameterized by a finite

number of coefficients amn , and a computer can search over these coefficients.nnoo
Getting around the second difficulty takes more care. To solve the
inequalities (197) on a computer, we must encode them with a finite amount
of data. It is usually sufficient to restrict ` `max for some large cutoff
`max . After we find , we can go back and check afterwards that it satisfies
(F,` ) 0 for ` > `max , as in exercise 10.3.

To deal with the continuous infinity of s, three techniques have been
used in the literature:
Discretize with a small spacing and impose a cutoff max . We then
have a finite set of linear inequalities for amn , which can be solved using
linear programming. This is the approach taken in the original paper
on CFT bounds [54].
Use a version of the simplex algorithm (underlying many linear programming solvers) that is customized to handle continuously varying
constraints, see [56, 57].
Approximate the constraints (197) as positivity conditions on polynomials and use semidefinite programming [5861]. Appendix C explains
the basic idea.
nn Note
that F (z, z) is symmetric under z z (because u and v are), so we can restrict

m n. Also, F (z, z) is odd under (z, z) (1 z, 1 z), so we can restrict to m + n
odd. This gives 21 b +1
c(b +1
c + 1) coefficients.
2
2
oo Sometimes these bounds appear to converge as increases, justifying post-hoc the
choice of subspace (198). However, this subspace is not always obviously the best choice.
New results might come from studying different points in the z, z plane, integrating
against kernels K(z, z), or doing something more exotic. For example, the limit z
0, with z fixed is known to encode interesting information about high spin operators.
Finding the optimal space of functionals is an open problem.
60
10.7. Improving on our Hand-Computed Bound

Let us compute an upper bound on the lowest-dimension scalar in the
OPE using a computer search. We will assume a Z2 symmetry under which
is odd so that doesnt appear in its own OPE. The procedure is as
follows
(1) Pick a value 0 and assume that all scalars in the OPE have
dimension 0 .
(2) Use a computer to search for amn such that
X
amn zm zn F,` (z, z)|z=z= 12 0,

m+n
for all
` = 0, 2, . . . , `max ,
(
0
`+d2
(` = 0),
(` > 0).
(199)
(3) If (199) is solvable, there must exist a scalar with dimension below 0 .
The best bound is the critical value crit.
above which (199) is solvable
0
and below which it is not. To find it, we can perform a binary search in
0 , running the algorithm above at each step. By additionally varying ,
we obtain a -dependent upper bound on the lowest-dimension scalar in
the OPE.
An implementation of this procedure is included with the semidefinite
program solver SDPB [61].pp See also [62] for a Python interface to SDPB
and [57] for another user-friendly bootstrap package. Running the code for
= 6, 8, 12, 16, 20, 28 gives the bounds shown in figure 27.qq
As the cutoff on the number of derivatives increases, the bounds
crit.
0 ( ) get stronger. Remarkably, the strongest bounds are nearly saturated by interesting physical theories. The most obvious feature of figure 27
is a kink near the location of the 2d Ising model ( , 0 ) = ( 81 , 1). (Other
exactly soluble unitary minimal models Mm,m+1 also lie near the bound.)
The bounds for different at the 2d Ising point = 81 are given in table 1. Taking = 28 gives a bound crit.
( 18 ) 1.0000005, within
0
7
5 10 of the correct value.
Table 1.
Upper bounds on in the 2d Ising model, computed with different cutoffs
on the number of derivatives.
6
8
12
16
20
28
1
crit.
(
=
)
1.020
1.0027
1.00053
1.000043
1.0000070
1.0000005
0
8
pp See
qq We
mathematica/Bootstrap2dExample.m at https://github.com/davidsd/sdpb.
use the SDPB parameters listed in the appendix of [61].
61
0
m,m+1 , m>3
Disallowed
1.5
=6
=8
..
.
=28
2d Ising
1.0
Allowed
0.5
Free
0.05
0.10
0.15
0.20
0.25
0.30
Fig. 27.
Upper bounds on the dimension 0 of the lowest dimension scalar in the
OPE as a function of , for 2d CFTs with a Z2 symmetry. The bounds are
computed using SDPB for = 6, 8, 12, 16, 20, 28, with the strongest bound (darkest blue
curve) corresponding to = 28 (a 105-dimensional space of functionals). The black dots
3
4
, 2 m+1
)
represent the unitary minimal models Mm,m+1 with ( , 0 ) = ( 12 2(m+1)
for m = 3, 4, 5, 6, of which the 2d Ising model is the case m = 3. The dashed line
represents the lowest dimension scalar in an OPE of operators cos(k) in the free boson
theory. These bounds first appeared in [63]. It should be possible to improve on the
lower bound in section 10.5 as well, though we have not attempted this.
10.8. Numerical Results in 3d

It is helpful to compare to exact solutions in 2d, but the above results
are remarkable because the methods are so general. We input information
about 2d global conformal symmetry (nothing about the Virasoro algebra!)
and unitarity, and the 2d Ising model pops out. Wonderfully, the same thing
happens in 3d! Again, we compute an upper bound on the lowest dimension
scalar in the OPE, this time using the 3d conformal blocks and the
3d unitarity bound. The resulting bound, shown in figure 28, has a kink
at ( , 0 ) (0.518, 1.412) close to the values realized in the 3d Ising
CFT [64].
All the results discussed so far come from studying crossing symmetry
of a single four-point function. However, the techniques can be generalized
to systems of correlation functions. The system hi, hi, hi in
62
Disallowed
1.8
1.6
3d Ising?
1.4
Allowed
1.2
1.0
0.50
0.52
0.54
0.56
0.58
0.60
0.62
0.64
Fig. 28.
Upper bound on the dimension of the lowest dimension scalar in the
OPE, where is a real scalar primary in a unitary 3d CFT with a Z2 symmetry,
from [56]. This bound is computed with = 24 (78-dimensional space of derivatives).
the 3d Ising CFT was studied in [60]. To get interesting new bounds in
this case, its necessary to input an additional fact: that and are the
only relevant scalars in the theory.rr In practice, this roughly means that
we impose positivity conditions (F,` ) 0 for = , , and 3.
The resulting bound in figure 29 now restricts ( , ) to a small island in
the space of operator dimensions.
The same multiple correlator bound, computed with = 43 using SDPB,
is shown in figure 30 [61]. The island has shrunk substantially, giving a
precise determination of the 3d Ising operator dimensions,
( , ) = (0.518151(6), 1.41264(6)).
(200)
Figures 29 and 30 are conceptually interesting. Firstly, the striking agreement between Monte Carlo simulations and the conformal bootstrap is
rr This
is an obvious experimental fact about the 3d Ising CFT. (It would be interesting to prove mathematically.) Relevant scalars are in one-to-one correspondence with
parameters that must be tuned to reach the critical point in some microscopic theory.
The fact that the phase diagram of water is 2-dimensional (the axes are temperature
and pressure) tells us that the critical point of water has two relevant operators.
63
1.7
1.6
1.5
1.4
1.3
1.2
1.1
1
0.5
0.51
0.52
0.53
0.54
0.55
0.56
0.57
0.58
Fig. 29.
Bound on ( , ) in a unitary 3d CFT with a Z2 symmetry and two relevant
scalars , with Z2 charges , +. The bound comes from studying crossing symmetry of
hi, hi, hi, and is computed with = 12. The allowed region is now a small
island near the 3d Ising point (black cross), with an additional bulk region to the right.
strong evidence that the critical 3d Ising model actually does flow to a
conformal fixed-point, as originally conjectured by Polyakov [8].
Secondly, figures 29 and 30 give a way to understand the phenomenon
of critical universality discussed at the beginning of this course. If a theory
flows to a unitary 3d CFT with a Z2 -symmetry and two relevant scalars ,
and if , dont live in the bulk region in figure 29 then the IR
theory must live in the 3d Ising island! Perhaps future bootstrap studies
will shrink the 3d Ising island to a point, proving the IR equivalence of
these theories.
10.9. Open Questions
The techniques above have been applied to numerous theories in different
spacetime dimensions, with different amounts of supersymmetry [15, 39,
40, 5461, 6393]. Because we dont start with a Lagrangian, theres no
guarantee when and how a particular physical theory will show up in the
bounds. Its an open question which correlators to study to isolate different
CFTs.
64
1.4131
1.413
Monte Carlo
1.4129
1.4128
1.4127
1.4126
1.4125
1.4124
4135
0.51808
0.5181
0.51812
0.51814
0.51816
0.51818
Fig. 30.
Bound on ( , ) in a unitary 3d CFT with a Z2 symmetry and two relevant
scalars , with Z2 charges , +. The bound comes from studying crossing symmetry
of hi, hi, hi, and is computed with = 43 using SDPB. The allowed region
is the blue sliver. The dashed rectangle shows the 68% confidence region for the current
best Monte Carlo determinations.
Other open questions include the following:

How do the bounds behave in the limit ? Does the Ising island
shrink to a point, still using a finite number of correlation functions, or
must we study larger systems of crossing equations?
How does one efficiently compute higher operator dimensions and OPE
coefficients? The extremal functional method [56, 65, 94] is one way,
but it is hard to estimate the associated errors.
Can the requirement of unitarity be relaxed? Gliozzis method of determinants [95] has shown success analyzing the crossing equation in
nonunitary theories and other situations where positivity is not obviously present [9698]. Can it be made rigorous?
What information is hidden in correlators of higher-spin operators like
stress tensors?
What can we prove analytically about the crossing equations? Progress
has been made in certain limits, for example large-N [99], large dimension [38], large spin [46, 5153], and combinations thereof [100102].
Another approach is to study the implications of slightly broken sym-
65
metries [103106]. It would be extremely interesting to prove analytical

results about the small-N , small , ` regime.
What additional structures and consistency conditions should we incorporate into the bootstrap? (See section 10.1.1.)
What protected information can be computed using supersymmetry?
Bootstrap studies recently led to the discovery of beautiful new algebraic structures in the OPE algebra of supersymmetric theories in 3, 4, 6
dimensions [82, 107, 108]. How do these structures interact with the
full non-protected bootstrap?
Thats a lot of open questions, and there are certainly many more. I
hope some of you will help find the answers!
Acknowledgements
I am grateful to Joe Polchinski and Pedro Vieira for inviting me to give
this course, and Tom DeGrand, Oliver DeWolfe, and Sherry Namburi for
helping make TASI such a fun experience. I am also grateful to Justin
David, Chethan Krishnan and Gautam Mandal for organizing the Advanced
Strings School at ICTS, Bangalore. A special thanks to the spectacular students at TASI and the Strings School, who asked so many good questions.
Thanks to Chris Beem, Joanna Huey, Filip Kos, Petr Kravchuk, David
Poland, and Slava Rychkov for comments. Thanks to Sheer El-Showk,
Jo
ao Penedones, and Pedro Vieira for the nice example in section 10.5.
I am supported by DOE grant number DE-SC0009988 and a William D.
Loughlin Membership at the Institute for Advanced Study.
Appendix A. Quantization of the Lattice Ising Model

In this section, we show how to interpret the partition function of the
Ising model on a square lattice in terms of Hilbert spaces and discrete time
evolution. This is a textbook trick,ss but we review it because it clearly
illustrates several ideas from section 2.2.
Consider the 2d Ising model on an m n lattice with periodic boundary
conditions. The spins are given by si,j {1}, where i Z/mZ and
ss It
is the starting point for Onsagers exact solution of the 2d Ising model [109].
66
j Z/nZ. The partition function is

X
Z=
exp (JSh (s) JSv (s)) ,
{s}
Sh (s)
Sv (s)
si,j si+1,j ,
i,j
si,j si,j+1 ,
(A.1)
i,j
where we have separated the action into contributions from horizontal and
vertical bonds.
We will think of the j-direction as time, and introduce a Hilbert space
Hm associated with a slice of m lattice sites at constant time. The space
Hm has a basis state for each spin configuration on the slice,
|s1 , . . . , sm i,
si {1}.
(A.2)
These are the analogs of the field eigenstates |b i in section 6.1. The Pauli
spin matrices
bi , = x, y, z act on the i-th site.
The operator
!
m
X
z
U exp J
biz
bi+1
(A.3)
i=1
encodes the contribution to the partition function from horizontal bonds

between m spins in a line. For example on an m 1 lattice, we would have
U |s1 , . . . , sm i = eJSh (s) |s1 , . . . , sm i.
(A.4)
The operator
V
Y
(eJ + eJ
bix )
(A.5)
encodes the effects of vertical bonds. For each site, it either preserves the
spin, giving a factor eJ associated with aligned spins, or flips it, giving a
factor eJ associated with anti-aligned spins. Defining the transfer matrix
T V U , its easy to check that
Z = TrHm (T n ).
(A.6)
We have interpreted the discrete path integral (A.1) in terms of operators on a Hilbert space. The transfer matrix is a discrete analogue of the
time-evolution operator etH . The path integral variable si,j maps to the
quantum operator
si,j T j iz T j ,
(A.7)
67
and correlation functions become traces of time-ordered products, e.g.tt

hi1 ,j1 i2 ,j2 i = TrHm (T n+j2 j1
biz1 T j1 j2
biz2 )(j1 j2 ) + (1 2).
(A.8)
We could instead have quantized the theory with the horizontal direction
as time. This would give a different Hilbert space Hn with dimension 2n
instead of 2m , a new transfer matrix T 0 (acting on Hn ), and a different
formula for the same path integral
Z = TrHn (T 0m ) = TrHm (T n ).
(A.9)
The new quantization map would be

si,j T 0i
bjz T 0i .
(A.10)
Let us emphasize that the operators (A.7) and (A.10) are truly different,
even though they represent the same path integral variable. They even
act on different-dimensional Hilbert spaces (2m vs. 2n )! Thus, its not
surprising that properties associated to a particular quantization, like their
behavior under Hermitian conjugation (section 7.1), could be different.
Appendix B. Euclidean vs. Lorentzian and Analytic Continuation
Here we make some brief comments about Euclidean and Lorentzian correlation functions and analytic continuation between them.
The first comment is that in Euclidean quantum field theory, out-oftime-order correlators dont make sense. For example, consider a Euclidean
two-point function,
h0|O1 (t1 )O2 (t2 )|0i = h0|O1 (0)eH(t2 t1 ) O2 (0)|0i.
(B.1)
In QFT, the Hamiltonian H is bounded from below and has arbitrarily

large positive eigenvalues. If we take t2 > t1 , then the operator eH(t2 t1 )
is unbounded. Generically, local operators O1,2 (0) have nonzero amplitude
to create arbitrarily high energy states. Thus, (B.1) is formally infinite.
Because the Euclidean path integral gives a time-ordered product
hO1 (t1 )O2 (t2 )i = h0|O1 (0)eH(t2 t1 ) O2 (0)|0i(t1 t2 ) +
h0|O2 (0)eH(t1 t2 ) O1 (0)|0i(t2 t1 ),
tt We
use the convention (0) =
1
.
2
(B.2)
68
it is well-defined for any ordering of the time coordinates. Specifically, the

operators eH(ti tj ) in (B.2) are always bounded.
In Lorentzian quantum field theory, however, non-time-ordered correlators (Wightman functions) are interesting observables. They can be obtained from time-ordered Euclidean correlators as follows. First set the
Euclidean times equal to small values tEi = i , increasing in the same order
as the operator ordering we want. For example, to place O1 later than O2 ,
consider
hO1 ()O2 (0)i = h0|O1 (0)eH O2 (0)|0i,
> 0.
(B.3)
Now continue tEi in the pure imaginary direction to the desired Lorentzian
times itLi . Because eH(tEi tEj ) never becomes unbounded, the operators
remain in the same order,
hO1 ( + itL1 )O2 (itL2 )i = h0|O1 (0)eHiH(tL1 tL2 ) O2 (0)|0i. (B.4)
Finally, take 0 to get the desired Wightman function.
To get a time-ordered Lorentzian correlator, there is a simple trick:
just simultaneously rotate all Euclidean times t i(1 i)t. Because the
ordering of the real parts of t are preserved, the order of the operators will
be too. This is Wick rotation.
Many properties of correlators under analytic continuation are clearer
when thinking about states and Hamiltonians, as opposed to path integrals.
Appendix C. Semidefinite Programming
For our purposes, a semidefinite program solver is an oracle that can solve
the following problem:
Find ~a such that ~a P~i (x) 0 for all x 0, i = 1, . . . , N ,
(C.1)
where P~i (x) are vector-valued polynomials. There are many freely-available
semidefinite program solvers. SDPB [61] in particular was written for application to the conformal bootstrap.
We would like to write our search in the form (C.1). After restricting
to the subspace (198), our positivity constraints become
X
(C.2)
amn zm zn F,` (z, z)|z=z= 21 0.
m+n
The trick is to find an approximation
zm zn F,` (z, z)|z=z= 12 ` ()P`mn (),
(C.3)
69
where ` () 0 are positive and P`mn () are polynomials. Then, dividing

(C.2) by ` () and writing = min,` + x, our inequality becomes
X
amn P`mn (min,` + x) 0.
(C.4)
m+n
This has the right form if we group the coefficients amn into a vector ~a and
identify ` i, `max N . The value min,` depends on the calculation at
hand, see for example (199).
To get a positive-times-polynomial approximation, we can start with
the series expansion (177),
X
d2
(C.5)
g,` (u, v) =
Bn,j r+n Cj 2 (cos ).
n,j
Recall that the coefficients Bn,j are positive rational functions of . The
1
crossing-symmetric
point z = z = 2 corresponds to a very small value of
r = r = 3 2 2 0.17. Thus, truncating the series at some large nmax
gives a good approximation,
ra b g,` (u, v)|r=r ,=0 r
Pàb ()
+ O(r+nmax ),
Q` ()
(C.6)
where Pàb and Q` are polynomials and Q` () is positive. Since deriva

tives of F,` (z, z) are linear combinations of derivatives of g,` (u, v), this
establishes (C.3) with
` () =
r
.
Q` ()
(C.7)
When exact formulae for conformal blocks are not available (for example, in odd dimensions), the polynomials Pàb () can be computed using
recursion relations [32, 33, 59, 60, 110112] or differential equations [113].
References
[1] S. Rychkov, EPFL Lectures on Conformal Field Theory in D 3 Dimensions (2016). arXiv:1601.05000.
[2] S. El-Showk, Solving Conformal Theories with the Bootstrap: Overview
and Recent Results, Lectures at the 9th Asian Winter School on Strings,
Particles and Cosmology, January 2015, Busan, South Korea. http://
home.kias.re.kr/MKG/h/AWSSPC2015/?pageNo=1006.
[3] J. D. Qualls, Lectures on Conformal Field Theory (2015).
arXiv:1511.04074.
[4] J. Polchinski, String theory. Vol. 1: An introduction to the bosonic string.
Cambridge, UK: Univ. Pr. (1998). 402 p.
70
[5] P. Di Francesco, P. Mathieu, and D. Senchal, Conformal Field Theory.

Springer-Verlag New York (1997).
[6] J. Polchinski, Scale and Conformal Invariance in Quantum Field Theory,
Nucl. Phys. B303, 226 (1988). doi: 10.1016/0550-3213(88)90179-4.
[7] A. Dymarsky, Z. Komargodski, A. Schwimmer, and S. Theisen, On Scale
and Conformal Invariance in Four Dimensions, JHEP. 10, 171 (2015). doi:
10.1007/JHEP10(2015)171. arXiv:1309.2921.
[8] A. M. Polyakov, Conformal symmetry of critical fluctuations, JETP Lett.
12, 381383 (1970).
[9] J. L. Cardy, Scaling and renormalization in statistical physics. Cambridge
lecture notes in physics: 3, Cambridge, UK: Univ. Pr. (1996). 238 p.
[10] M.-C. Cha, M. P. A. Fisher, S. M. Girvin, M. Wallin, and A. P. Young,
Universal conductivity of two-dimensional films at the superconductorinsulator transition, Phys. Rev. B. 44, 68836902 (Oct, 1991). doi: 10.1103/
PhysRevB.44.6883. URL http://link.aps.org/doi/10.1103/PhysRevB.
44.6883.
[11] J. Smakov
and E. Srensen, Universal scaling of the conductivity at the
superfluid-insulator phase transition, Phys. Rev. Lett. 95, 180603 (Oct,
2005). doi: 10.1103/PhysRevLett.95.180603. URL http://link.aps.org/
doi/10.1103/PhysRevLett.95.180603.
[12] J. Lipa, J. Nissen, D. Stricker, D. Swanson, and T. Chui, Specific heat of
liquid helium in zero gravity very near the lambda point, Phys.Rev. B68,
174518 (2003). doi: 10.1103/PhysRevB.68.174518.
[13] S. Ferrara, A. F. Grillo, and R. Gatto, Tensor representations of conformal algebra and conformally covariant operator product expansion, Annals
Phys. 76, 161188 (1973). doi: 10.1016/0003-4916(73)90446-6.
[14] A. M. Polyakov, Nonhamiltonian approach to conformal quantum field theory, Zh. Eksp. Teor. Fiz. 66, 2342 (1974).
[15] C. Beem, M. Lemos, L. Rastelli, and B. C. van Rees, The (2, 0) superconformal bootstrap (2015). arXiv:1507.05637.
[16] J. M. Maldacena, The Large N limit of superconformal field theories and
supergravity, Int. J. Theor. Phys. 38, 11131133 (1999). doi: 10.1023/A:
1026654312961. [Adv. Theor. Math. Phys.2,231(1998)].
[17] D. Gaiotto, A. Kapustin, N. Seiberg, and B. Willett, Generalized Global
Symmetries, JHEP. 02, 172 (2015). doi: 10.1007/JHEP02(2015)172.
arXiv:1412.5148.
[18] A. A. Belavin, A. M. Polyakov, and A. B. Zamolodchikov, Infinite conformal
symmetry in two-dimensional quantum field theory, Nucl. Phys. B241,
333380 (1984). doi: 10.1016/0550-3213(84)90052-X.
[19] P. A. Dirac, Wave equations in conformal space, Annals Math. 37, 429442
(1936). doi: 10.2307/1968455.
[20] G. Mack and A. Salam, Finite component field representations of the conformal group, Ann. Phys. 53, 174202 (1969). doi: 10.1016/0003-4916(69)
90278-4.
[21] D. Boulware, L. Brown, and R. Peccei, Deep-inelastic electroproduction
and conformal symmetry, Phys.Rev. D2, 293298 (1970). doi: 10.1103/
71
PhysRevD.2.293.
[22] S. Ferrara, R. Gatto, and A. Grillo, Conformal algebra in space-time and
operator product expansion, Springer Tracts Mod.Phys. 67, 164 (1973).
doi: 10.1007/BFb0111104.
[23] S. Weinberg, Six-dimensional Methods for Four-dimensional Conformal
Field Theories, Phys.Rev. D82, 045031 (2010). doi: 10.1103/PhysRevD.
82.045031. arXiv:1006.3480.
[24] M. S. Costa, J. Penedones, D. Poland, and S. Rychkov, Spinning Conformal Correlators, JHEP. 1111, 071 (2011). doi: 10.1007/JHEP11(2011)071.
arXiv:1107.3554.
[25] J. Penedones, TASI Lectures on AdS/CFT (2015).
[26] M. Hogervorst and S. Rychkov, Radial Coordinates for Conformal Blocks,
Phys.Rev. D87(10), 106004 (2013). doi: 10.1103/PhysRevD.87.106004.
arXiv:1303.1111.
[27] R. C. Brower, M. Cheng, and G. T. Fleming, Improved Lattice Radial
Quantization, PoS. LATTICE2013, 335 (2014).
[28] J. Glimm and A. Jaffe, Quantum Physics: A Functional Integral Point of
View. Springer-Verlag New York (1987).
[29] G. Mack, All Unitary Ray Representations of the Conformal Group
SU (2, 2) with Positive Energy, Commun.Math.Phys. 55, 1 (1977). doi:
10.1007/BF01613145.
[30] J. C. Jantzen, Kontravariante formen auf induzierten darstellungen halbeinfacher lie-algebren, Mathematische Annalen. 226(1), 5365 ISSN
1432-1807. doi: 10.1007/BF01391218. URL http://dx.doi.org/10.1007/
BF01391218.
[31] S. Minwalla, Restrictions Imposed by Superconformal Invariance on
Quantum Field Theories, Adv. Theor. Math. Phys. 2, 781846 (1998).
hep-th/9712074.
[32] M. Yamazaki, Comments on Determinant Formulas for General CFTs
(2016). arXiv:1601.04072.
[33] J. Penedones, E. Trevisani, and M. Yamazaki, Recursion Relations for Conformal Blocks (2015). arXiv:1509.00428.
[34] J. Maldacena and A. Zhiboedov, Constraining Conformal Field Theories
with A Higher Spin Symmetry (2011). arXiv:1112.1016.
[35] V. Alba and K. Diab, Constraining conformal field theories with a higher
spin symmetry in d > 3 dimensions (2015). arXiv:1510.02535.
[36] F. Dolan and H. Osborn, Conformal partial waves and the operator product
expansion, Nucl.Phys. B678, 491507 (2004). doi: 10.1016/j.nuclphysb.
2003.11.016. hep-th/0309180.
[37] F. Dolan and H. Osborn, Conformal four point functions and the operator product expansion, Nucl.Phys. B599, 459496 (2001). doi: 10.1016/
S0550-3213(01)00013-X. hep-th/0011040.
[38] D. Pappadopulo, S. Rychkov, J. Espin, and R. Rattazzi, OPE Convergence
in Conformal Field Theory, Phys.Rev. D86, 105043 (2012). doi: 10.1103/
PhysRevD.86.105043. arXiv:1208.6449.
[39] P. Liendo, L. Rastelli, and B. C. van Rees, The Bootstrap Program for
72
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
Boundary CFTd , JHEP. 1307, 113 (2013). doi: 10.1007/JHEP07(2013)

113. arXiv:1210.4258.
D. Gaiotto, D. Mazac, and M. F. Paulos, Bootstrapping the 3d Ising
twist defect, JHEP. 1403, 100 (2014). doi: 10.1007/JHEP03(2014)100.
arXiv:1310.5078.
E. Witten, Quantum Field Theory and the Jones Polynomial, Commun.
Math. Phys. 121, 351399 (1989). doi: 10.1007/BF01217730.
O. Aharony, N. Seiberg, and Y. Tachikawa, Reading between the lines
of four-dimensional gauge theories, JHEP. 08, 115 (2013). doi: 10.1007/
JHEP08(2013)115. arXiv:1305.0318.
D. M. Hofman and J. Maldacena, Conformal collider physics: Energy and
charge correlations, JHEP. 05, 012 (2008). doi: 10.1088/1126-6708/2008/
05/012. arXiv:0803.1467.
Z. Komargodski and A. Schwimmer, On Renormalization Group Flows in
Four Dimensions, JHEP. 12, 099 (2011). doi: 10.1007/JHEP12(2011)099.
arXiv:1107.3987.
O. Nachtmann, Positivity constraints for anomalous dimensions,
Nucl.Phys. B63, 237247 (1973). doi: 10.1016/0550-3213(73)90144-2.
Z. Komargodski and A. Zhiboedov, Convexity and Liberation at Large
Spin, JHEP. 1311, 140 (2013). doi: 10.1007/JHEP11(2013)140.
Z. Komargodski, M. Kulaxizi, A. Parnachev, and A. Zhiboedov, Conformal
Field Theories and Deep Inelastic Scattering (2016). arXiv:1601.05453.
J. Maldacena, S. H. Shenker, and D. Stanford, A bound on chaos (2015).
arXiv:1503.01409.
T. Hartman, S. Jain, and S. Kundu, Causality Constraints in Conformal
Field Theory (2015). arXiv:1509.00014.
G. W. Moore and N. Seiberg, Polynomial Equations for Rational Conformal
Field Theories, Phys. Lett. B212, 451 (1988). doi: 10.1016/0370-2693(88)
91796-0.
L. F. Alday and J. M. Maldacena, Comments on operators with
large spin, JHEP. 11, 019 (2007). doi: 10.1088/1126-6708/2007/11/019.
arXiv:0708.0672.
A. L. Fitzpatrick, J. Kaplan, D. Poland, and D. Simmons-Duffin, The Analytic Bootstrap and AdS Superhorizon Locality, JHEP. 1312, 004 (2013).
doi: 10.1007/JHEP12(2013)004. arXiv:1212.3616.
L. F. Alday and A. Zhiboedov, An Algebraic Approach to the Analytic
Bootstrap (2015). arXiv:1510.08091.
R. Rattazzi, V. S. Rychkov, E. Tonni, and A. Vichi, Bounding scalar operator dimensions in 4D CFT, JHEP. 12, 031 (2008). doi: 10.1088/1126-6708/
2008/12/031. arXiv:0807.0004.
F. Caracciolo and V. S. Rychkov, Rigorous Limits on the Interaction
Strength in Quantum Field Theory, Phys. Rev. D81, 085037 (2010). doi:
10.1103/PhysRevD.81.085037. arXiv:0912.2726.
S. El-Showk, M. F. Paulos, D. Poland, S. Rychkov, D. Simmons-Duffin,
and A. Vichi, Solving the 3d Ising Model with the Conformal Bootstrap
II. c-Minimization and Precise Critical Exponents, J.Stat.Phys. 157, 869
73
(June, 2014). doi: 10.1007/s10955-014-1042-7. arXiv:1403.4545.

[57] M. F. Paulos, JuliBootS: a hands-on guide to the conformal bootstrap
(2014). arXiv:1412.4127.
[58] D. Poland, D. Simmons-Duffin, and A. Vichi, Carving Out the Space
of 4D CFTs, JHEP. 1205, 110 (2012). doi: 10.1007/JHEP05(2012)110.
arXiv:1109.5176.
[59] F. Kos, D. Poland, and D. Simmons-Duffin, Bootstrapping the O(N ) vector models, JHEP. 1406, 091 (2014). doi: 10.1007/JHEP06(2014)091.
arXiv:1307.6856.
[60] F. Kos, D. Poland, and D. Simmons-Duffin, Bootstrapping Mixed Correlators in the 3D Ising Model, JHEP. 1411, 109 (2014). doi: 10.1007/
JHEP11(2014)109. arXiv:1406.4858.
[61] D. Simmons-Duffin, A Semidefinite Program Solver for the Conformal Bootstrap, JHEP. 06, 174 (2015). doi: 10.1007/JHEP06(2015)174.
arXiv:1502.02033.
[62] C. Behan, PyCFTBoot: A flexible interface for the conformal bootstrap
(2016). arXiv:1602.02810.
[63] V. S. Rychkov and A. Vichi, Universal Constraints on Conformal Operator
Dimensions, Phys. Rev. D80, 045006 (2009). doi: 10.1103/PhysRevD.80.
045006. arXiv:0905.2211.
[64] S. El-Showk, M. F. Paulos, D. Poland, S. Rychkov, D. Simmons-Duffin,
and A. Vichi, Solving the 3D Ising Model with the Conformal Bootstrap, Phys.Rev. D86, 025022 (2012). doi: 10.1103/PhysRevD.86.025022.
arXiv:1203.6064.
[65] D. Poland and D. Simmons-Duffin, Bounds on 4D Conformal and Superconformal Field Theories, JHEP. 1105, 017 (2011). doi: 10.1007/
JHEP05(2011)017. arXiv:1009.2087.
[66] R. Rattazzi, S. Rychkov, and A. Vichi, Central Charge Bounds in 4D
Conformal Field Theory, Phys. Rev. D83, 046011 (2011). doi: 10.1103/
PhysRevD.83.046011. arXiv:1009.2725.
[67] R. Rattazzi, S. Rychkov, and A. Vichi, Bounds in 4D Conformal Field
Theories with Global Symmetry, J. Phys. A44, 035402 (2011). doi: 10.
1088/1751-8113/44/3/035402. arXiv:1009.5985.
[68] A. Vichi, Improved bounds for CFTs with global symmetries, JHEP. 1201,
162 (2012). doi: 10.1007/JHEP01(2012)162. arXiv:1106.4037.
[69] S. Rychkov, Conformal Bootstrap in Three Dimensions?
(2011).
arXiv:1111.2115.
[70] C. Beem, L. Rastelli, and B. C. van Rees, The N = 4 Superconformal
Bootstrap, Phys.Rev.Lett. 111, 071601 (2013). doi: 10.1103/PhysRevLett.
111.071601. arXiv:1304.1803.
[71] S. El-Showk, M. Paulos, D. Poland, S. Rychkov, D. Simmons-Duffin,
and A. Vichi, Conformal field theories in fractional dimensions, Phys.
Rev. Lett. 112, 141601 (Apr, 2014). doi: 10.1103/PhysRevLett.112.141601.
arXiv:1309.5089.
[72] L. F. Alday and A. Bissi, The superconformal bootstrap for structure constants, JHEP. 09, 144 (2014). doi: 10.1007/JHEP09(2014)144.
74
arXiv:1310.3757.
[73] D. Bashkirov, Bootstrapping the N = 1 SCFT in three dimensions (2013).
arXiv:1310.8255.
[74] M. Berkooz, R. Yacoby, and A. Zait, Bounds on N = 1 superconformal
theories with global symmetries, JHEP. 1408, 008 (2014). doi: 10.1007/
JHEP01(2015)132,10.1007/JHEP08(2014)008. arXiv:1402.6068.
[75] Y. Nakayama and T. Ohtsuki, Approaching conformal window of
O(n) O(m) symmetric Landau-Ginzburg models from conformal bootstrap, Phys.Rev. D89, 126009 (2014). doi: 10.1103/PhysRevD.89.126009.
arXiv:1404.0489.
[76] Y. Nakayama and T. Ohtsuki, Five dimensional O(N )-symmetric CFTs
from conformal bootstrap, Phys. Lett. B734, 193197 (2014). doi: 10.1016/
j.physletb.2014.05.058. arXiv:1404.5201.
[77] L. F. Alday and A. Bissi, Generalized bootstrap equations for N = 4 SCFT,
JHEP. 02, 101 (2015). doi: 10.1007/JHEP02(2015)101. arXiv:1404.5864.
[78] S. M. Chester, J. Lee, S. S. Pufu, and R. Yacoby, The N = 8 superconformal
bootstrap in three dimensions, JHEP. 1409, 143 (2014). doi: 10.1007/
JHEP09(2014)143. arXiv:1406.4814.
[79] F. Caracciolo, A. C. Echeverri, B. von Harling, and M. Serone, Bounds on
OPE Coefficients in 4D Conformal Field Theories, JHEP. 10, 20 (2014).
doi: 10.1007/JHEP10(2014)020. arXiv:1406.7845.
[80] Y. Nakayama and T. Ohtsuki, Bootstrapping phase transitions in QCD and
frustrated spin systems, Phys. Rev. D91(2), 021901 (2015). doi: 10.1103/
PhysRevD.91.021901. arXiv:1407.6195.
[81] J. Golden and M. F. Paulos, No unitary bootstrap for the fractal
Ising model, JHEP. 03, 167 (2015). doi: 10.1007/JHEP03(2015)167.
arXiv:1411.7932.
[82] S. M. Chester, J. Lee, S. S. Pufu, and R. Yacoby, Exact Correlators of
BPS Operators from the 3d Superconformal Bootstrap, JHEP. 1503, 130
(2015). doi: 10.1007/JHEP03(2015)130. arXiv:1412.0334.
[83] C. Beem, M. Lemos, P. Liendo, L. Rastelli, and B. C. van Rees, The N = 2
superconformal bootstrap (2014). arXiv:1412.7541.
[84] N. Bobev, S. El-Showk, D. Mazac, and M. F. Paulos, Bootstrapping the
Three-Dimensional Supersymmetric Ising Model, Phys. Rev. Lett. 115(5),
051601 (2015). doi: 10.1103/PhysRevLett.115.051601. arXiv:1502.04124.
[85] N. Bobev, S. El-Showk, D. Mazac, and M. F. Paulos, Bootstrapping
SCFTs with Four Supercharges, JHEP. 08, 142 (2015). doi: 10.1007/
JHEP08(2015)142. arXiv:1503.02081.
[86] F. Kos, D. Poland, D. Simmons-Duffin, and A. Vichi, Bootstrapping the
O(N) Archipelago, JHEP. 11, 106 (2015). doi: 10.1007/JHEP11(2015)106.
arXiv:1504.07997.
[87] S. M. Chester, S. Giombi, L. V. Iliesiu, I. R. Klebanov, S. S. Pufu, and
R. Yacoby, Accidental Symmetries and the Conformal Bootstrap (2015).
arXiv:1507.04424.
[88] L. Iliesiu, F. Kos, D. Poland, S. S. Pufu, D. Simmons-Duffin, and R. Yacoby,
Bootstrapping 3D Fermions (2015). arXiv:1508.00012.
75
[89] D. Poland and A. Stergiou, Exploring the Minimal 4D N = 1 SCFT (2015).

arXiv:1509.06368.
[90] M. Lemos and P. Liendo, Bootstrapping N = 2 chiral correlators (2015).
arXiv:1510.03866.
[91] Y.-H. Lin, S.-H. Shao, D. Simmons-Duffin, Y. Wang, and X. Yin, N=4
Superconformal Bootstrap of the K3 CFT (2015). arXiv:1511.04065.
[92] S. M. Chester, L. V. Iliesiu, S. S. Pufu, and R. Yacoby, Bootstrapping
O(N ) Vector Models with Four Supercharges in 3 d 4 (2015).
arXiv:1511.07552.
[93] S. M. Chester and S. S. Pufu, Towards Bootstrapping QED3 (2016).
arXiv:1601.03476.
[94] S. El-Showk and M. F. Paulos, Bootstrapping Conformal Field Theories
with the Extremal Functional Method, Phys.Rev.Lett. 111(24), 241601
(2013). doi: 10.1103/PhysRevLett.111.241601. arXiv:1211.2810.
[95] F. Gliozzi, More constraining conformal bootstrap, Phys.Rev.Lett. 111,
161602 (2013). doi: 10.1103/PhysRevLett.111.161602. arXiv:1307.3111.
[96] F. Gliozzi and A. Rago, Critical exponents of the 3d Ising and related
models from Conformal Bootstrap (2014). arXiv:1403.6003.
[97] F. Gliozzi, P. Liendo, M. Meineri, and A. Rago, Boundary and Interface
CFTs from the Conformal Bootstrap, JHEP. 05, 036 (2015). doi: 10.1007/
JHEP05(2015)036. arXiv:1502.07217.
[98] Y. Nakayama, Bootstrapping critical Ising model on three-dimensional real
projective space (2016). arXiv:1601.06851.
[99] I. Heemskerk, J. Penedones, J. Polchinski, and J. Sully, Holography from
Conformal Field Theory, JHEP. 0910, 079 (2009). doi: 10.1088/1126-6708/
2009/10/079. arXiv:0907.0151.
[100] A. L. Fitzpatrick, J. Kaplan, and M. T. Walters, Universality of LongDistance AdS Physics from the CFT Bootstrap (2014). arXiv:1403.6829.
[101] A. L. Fitzpatrick, J. Kaplan, M. T. Walters, and J. Wang, Eikonalization of
Conformal Blocks, JHEP. 09, 019 (2015). doi: 10.1007/JHEP09(2015)019.
arXiv:1504.01737.
[102] A. L. Fitzpatrick, J. Kaplan, and M. T. Walters, Virasoro Conformal Blocks
and Thermality from Classical Background Fields, JHEP. 11, 200 (2015).
doi: 10.1007/JHEP11(2015)200. arXiv:1501.05315.
[103] J. Maldacena and A. Zhiboedov, Constraining conformal field theories with
a slightly broken higher spin symmetry, Class.Quant.Grav. 30, 104003
(2013). doi: 10.1088/0264-9381/30/10/104003. arXiv:1204.3882.
[104] S. Rychkov and Z. M. Tan, The -expansion from conformal field theory,
J. Phys. A48(29), 29FT01 (2015). doi: 10.1088/1751-8113/48/29/29FT01.
arXiv:1505.00963.
[105] L. F. Alday and A. Zhiboedov, Conformal Bootstrap With Slightly Broken
Higher Spin Symmetry (2015). arXiv:1506.04659.
[106] S. Giombi and V. Kirilin, Anomalous dimensions in CFT with weakly broken higher spin symmetry (2016). arXiv:1601.01310.
[107] C. Beem, M. Lemos, P. Liendo, W. Peelaers, L. Rastelli, and B. van Rees,
Infinite Chiral Symmetry in Four Dimensions (2013). arXiv:1312.5344.
76
[108] C. Beem, W. Peelaers, and L. Rastelli, Deformation quantization and superconformal symmetry in three dimensions (2016). arXiv:1601.05378.
[109] L. Onsager, Crystal statistics. i. a two-dimensional model with an orderdisorder transition, Phys. Rev. 65, 117149 (Feb, 1944). doi: 10.1103/
PhysRev.65.117. URL http://link.aps.org/doi/10.1103/PhysRev.65.
117.
[110] Al.B. Zamolodchikov, Conformal Symmetry in Two-Dimensions: an Explicit Recurrence Formula for the Conformal Partial Wave Amplitude,
Commun.Math.Phys. 96, 419422 (1984). doi: 10.1007/BF01214585.
[111] Al.B. Zamolodchikov, Conformal symmetry in two-dimensional space: Recursion representation of conformal block, Theoretical and Mathematical Physics. 73(1), 10881093 (1987). ISSN 0040-5779. doi: 10.1007/
BF01022967.
[112] L. Iliesiu, F. Kos, D. Poland, S. S. Pufu, D. Simmons-Duffin, and R. Yacoby,
Fermion-Scalar Conformal Blocks (2015). arXiv:1511.01497.
[113] M. Hogervorst, H. Osborn, and S. Rychkov, Diagonal Limit for Conformal Blocks in d Dimensions, JHEP. 1308, 014 (2013). doi: 10.1007/
JHEP08(2013)014. arXiv:1305.1321.

Simmons Duffin Notes

Uploaded by

Copyright:

Available Formats

Simmons Duffin Notes

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Simmons Duffin Notes

Uploaded by

Copyright:

Available Formats

arXiv:1602.

07982v1 [hep-th] 25 Feb 2016

TASI Lectures on the Conformal Bootstrap

TASI Lectures on the Conformal Bootstrap

Appendix A. Quantization of the Lattice Ising Model . . . . . . .

TASI Lectures on the Conformal Bootstrap

question of when scale invariance implies conformal invariance is an important

distance scale characterizing the observable were computing. At distances

1.2. Critical Universality

TASI Lectures on the Conformal Bootstrap

1.3. The Bootstrap Philosophy

focus on the CFT itself and not a specific microscopic realization,

ally provide a fully nonperturbative formulation of quantum field theory,

This holds as an operator equation, meaning it is true away from other

(x xi )i hO1 (x1 ) . . . On (xn )i.

Exercise 2.1. Consider a QFT coupled to a background metric g. For

Derive (6) by demanding that S[g, ] be diffeomorphism invariant near flat

definition of the stress tensor works in a continuum field theory. If the UV is a

TASI Lectures on the Conformal Bootstrap

Consider the integral of T over a closed surface ,f g

Fig. 2. A surface supporting the operator P () can be freely deformed 0

Let = B be the boundary of a ball B containing x and no other

In other words, surrounding O(x) with the topological surface operator P

Surrounding O(x) with P gives a derivative.

In quantum field theory, having a topological codimension-1 operator is

To specify a quantization, we foliate spacetime by slices related by an

TASI Lectures on the Conformal Bootstrap

2.3. Topological Operators and Symmetries

Let t be a spatial slice at time t and consider the operator P (t ).

Figure 6 makes it clear that this result is independent of how we quantize

without specifying a quantization. In fact, from now on, we will no longer

TASI Lectures on the Conformal Bootstrap

Equation (15) integrates to

This statement is also true in any quantization of the theory. In path

Consider the time-ordered correlator (11) with tn > > t1 . Using

= h0|etn P On (0, xn )etn P et1 P O1 (0, x1 )et1 P |0i

= h0|On (0, xn )e(tn tn1 )P e(t2 t1 )P O1 (0, x1 )|0i.

will be conserved whenever

This is the Killing equation. In flat space, it has solutions

The corresponding charges are momentum P = Qp and angular momentum M = Qm .

where c(x) is a scalar function. Contracting both sides with gives

(special conformal transformations).

The corresponding symmetry charges are D = Qd and K = Qk .l

curved space, there can by Weyl anomalies.

TASI Lectures on the Conformal Bootstrap

Such transformations are called conformal. They comprise the conformal

I is a conformal transformation, but it is not continuously connected to

We can think of k as a translation that moves infinity and fixes the

where [1 , 2 ] is a commutator of vector fields.m This is not obvious and

TASI Lectures on the Conformal Bootstrap

As usual, (30) is true in any quantization of the theory. In path integral

and all other commutators vanish.

4. Primaries and Descendants

where S are matrices satisfying the same algebra as M , and a, b are

TASI Lectures on the Conformal Bootstrap

the innermost. The conformal commutation relations tell us how to re-order

= exP (exP M exP )O(0)

4.2. Scale+Poincare Representations

The eigenvalue is the dimension of O.

where [1 , 2 ] is a commutator of vector fields.m This is not obvious and